diff --git a/cpp/apidoc/Doxyfile b/cpp/apidoc/Doxyfile
index de7777a6c1ca5..baf848ca237dd 100644
--- a/cpp/apidoc/Doxyfile
+++ b/cpp/apidoc/Doxyfile
@@ -1095,6 +1095,7 @@ EXCLUDE_PATTERNS       = *-test.cc \
                          *test* \
                          *_generated.h \
                          *-benchmark.cc \
+                         *_codegen.py \
                          *internal*
 
 # The EXCLUDE_SYMBOLS tag can be used to specify one or more symbol names
diff --git a/cpp/src/arrow/CMakeLists.txt b/cpp/src/arrow/CMakeLists.txt
index 9acd708343725..ce5d616bcb9ab 100644
--- a/cpp/src/arrow/CMakeLists.txt
+++ b/cpp/src/arrow/CMakeLists.txt
@@ -490,6 +490,8 @@ set(ARROW_UTIL_SRCS
     util/bitmap_builders.cc
     util/bitmap_ops.cc
     util/bpacking.cc
+    util/bpacking_scalar.cc
+    util/bpacking_simd_min.cc
     util/byte_size.cc
     util/byte_stream_split_internal.cc
     util/cancel.cc
@@ -533,11 +535,8 @@ set(ARROW_UTIL_SRCS
 
 append_runtime_avx2_src(ARROW_UTIL_SRCS util/byte_stream_split_internal_avx2.cc)
 
-append_runtime_avx2_src(ARROW_UTIL_SRCS util/bpacking_avx2.cc)
-append_runtime_avx512_src(ARROW_UTIL_SRCS util/bpacking_avx512.cc)
-if(ARROW_HAVE_NEON)
-  list(APPEND ARROW_UTIL_SRCS util/bpacking_neon.cc)
-endif()
+append_runtime_avx2_src(ARROW_UTIL_SRCS util/bpacking_simd_avx2.cc)
+append_runtime_avx512_src(ARROW_UTIL_SRCS util/bpacking_simd_avx512.cc)
 
 if(ARROW_WITH_BROTLI)
   list(APPEND ARROW_UTIL_SRCS util/compression_brotli.cc)
diff --git a/cpp/src/arrow/util/bit_stream_utils_internal.h b/cpp/src/arrow/util/bit_stream_utils_internal.h
index 1f3b699e1ac6b..cccc2cb129528 100644
--- a/cpp/src/arrow/util/bit_stream_utils_internal.h
+++ b/cpp/src/arrow/util/bit_stream_utils_internal.h
@@ -123,6 +123,34 @@ inline uint64_t ReadLittleEndianWord(const uint8_t* buffer, int bytes_remaining)
 /// bytes in one read (e.g. encoded int).
 class BitReader {
  public:
+  template <typename T, typename = void>
+  struct UnpackFnDetect;
+
+  template <typename T>
+  struct UnpackFnDetect<T, std::enable_if_t<(sizeof(T) >= sizeof(int))>> {
+    using type = internal::UnpackFn<std::make_unsigned_t<T>>;
+  };
+
+  template <typename T>
+  struct UnpackFnDetect<T, std::enable_if_t<(sizeof(T) < sizeof(int))>> {
+    using type = internal::UnpackFn<uint32_t>;
+  };
+
+  /// The type for a function that can extract bit-packed integers.
+  template <typename T>
+  using UnpackFn = typename UnpackFnDetect<T>::type;
+
+  /// Get the unack function most appropriated for this type and bit width.
+  template <typename T>
+  static UnpackFn<T> get_unpack_fn(int num_bits) {
+    // This is intimately linked to the GetBatch implementation
+    if constexpr (sizeof(T) >= sizeof(int)) {
+      return internal::get_unpack_fn<std::make_unsigned_t<T>>(num_bits);
+    } else {
+      return internal::get_unpack_fn<uint32_t>(num_bits);
+    }
+  }
+
   BitReader() noexcept = default;
 
   /// 'buffer' is the buffer to read from.  The buffer's length is 'buffer_len'.
@@ -148,6 +176,11 @@ class BitReader {
   template <typename T>
   int GetBatch(int num_bits, T* v, int batch_size);
 
+  /// Get a number of values from the buffer. Return the number of values actually read.
+  /// @param unpack Function pointer to the unpack function for the correct bit width.
+  template <typename T>
+  int GetBatch(int num_bits, T* v, int batch_size, UnpackFn<T> unpack);
+
   /// Reads a 'num_bytes'-sized value from the buffer and stores it in 'v'. T
   /// needs to be a little-endian native type and big enough to store
   /// 'num_bytes'. The value is assumed to be byte-aligned so the stream will
@@ -297,7 +330,7 @@ inline bool BitReader::GetValue(int num_bits, T* v) {
 }
 
 template <typename T>
-inline int BitReader::GetBatch(int num_bits, T* v, int batch_size) {
+int BitReader::GetBatch(int num_bits, T* v, int batch_size, UnpackFn<T> unpack) {
   ARROW_DCHECK(buffer_ != NULL);
   ARROW_DCHECK_LE(num_bits, static_cast<int>(sizeof(T) * 8)) << "num_bits: " << num_bits;
 
@@ -323,19 +356,11 @@ inline int BitReader::GetBatch(int num_bits, T* v, int batch_size) {
     }
   }
 
-  if (sizeof(T) == 4) {
+  // unpack for uint16_t not as fast as unpack for uint32_t + memcpy.
+  if constexpr (sizeof(T) >= sizeof(32)) {
     int num_unpacked =
-        internal::unpack32(buffer + byte_offset, reinterpret_cast<uint32_t*>(v + i),
-                           batch_size - i, num_bits);
-    i += num_unpacked;
-    byte_offset += num_unpacked * num_bits / 8;
-  } else if (sizeof(T) == 8 && num_bits > 32) {
-    // Use unpack64 only if num_bits is larger than 32
-    // TODO (ARROW-13677): improve the performance of internal::unpack64
-    // and remove the restriction of num_bits
-    int num_unpacked =
-        internal::unpack64(buffer + byte_offset, reinterpret_cast<uint64_t*>(v + i),
-                           batch_size - i, num_bits);
+        unpack(buffer + byte_offset, reinterpret_cast<std::make_unsigned_t<T>*>(v + i),
+               batch_size - i);
     i += num_unpacked;
     byte_offset += num_unpacked * num_bits / 8;
   } else {
@@ -345,8 +370,7 @@ inline int BitReader::GetBatch(int num_bits, T* v, int batch_size) {
     uint32_t unpack_buffer[buffer_size];
     while (i < batch_size) {
       int unpack_size = std::min(buffer_size, batch_size - i);
-      int num_unpacked =
-          internal::unpack32(buffer + byte_offset, unpack_buffer, unpack_size, num_bits);
+      int num_unpacked = unpack(buffer + byte_offset, unpack_buffer, unpack_size);
       if (num_unpacked == 0) {
         break;
       }
@@ -380,6 +404,11 @@ inline int BitReader::GetBatch(int num_bits, T* v, int batch_size) {
   return batch_size;
 }
 
+template <typename T>
+inline int BitReader::GetBatch(int num_bits, T* v, int batch_size) {
+  return GetBatch(num_bits, v, batch_size, get_unpack_fn<T>(num_bits));
+}
+
 template <typename T>
 inline bool BitReader::GetAligned(int num_bytes, T* v) {
   if (ARROW_PREDICT_FALSE(num_bytes > static_cast<int>(sizeof(T)))) {
diff --git a/cpp/src/arrow/util/bpacking.cc b/cpp/src/arrow/util/bpacking.cc
index 990f76875aafd..4c3ec54a135a0 100644
--- a/cpp/src/arrow/util/bpacking.cc
+++ b/cpp/src/arrow/util/bpacking.cc
@@ -15,154 +15,76 @@
 // specific language governing permissions and limitations
 // under the License.
 
-#include "arrow/util/bpacking_internal.h"
+#include <array>
 
-#include "arrow/util/bpacking64_default_internal.h"
-#include "arrow/util/bpacking_default_internal.h"
-#include "arrow/util/cpu_info.h"
+#include "arrow/util/bpacking_dispatch_internal.h"
+#include "arrow/util/bpacking_internal.h"
+#include "arrow/util/bpacking_scalar_internal.h"
+#include "arrow/util/bpacking_simd_internal.h"
 #include "arrow/util/dispatch_internal.h"
-#include "arrow/util/logging_internal.h"
 
+namespace arrow::internal {
+
+namespace {
+
+template <typename Uint>
+struct UnpackDynamicFunction {
+  using FunctionType = decltype(&unpack_scalar<Uint>);
+  using Implementation = std::pair<DispatchLevel, FunctionType>;
+
+  static auto implementations() {
+    return std::array {
+      // Current SIMD unpack algorithm works terribly on SSE4.2 due to lack of variable
+      // rhsift and poor xsimd fallback.
+      Implementation{DispatchLevel::NONE, &unpack_scalar<Uint>},
 #if defined(ARROW_HAVE_RUNTIME_AVX2)
-#  include "arrow/util/bpacking_avx2_internal.h"
+          Implementation{DispatchLevel::AVX2, &unpack_avx2<Uint>},
 #endif
 #if defined(ARROW_HAVE_RUNTIME_AVX512)
-#  include "arrow/util/bpacking_avx512_internal.h"
+          Implementation{DispatchLevel::AVX512, &unpack_avx512<Uint>},
 #endif
-#if defined(ARROW_HAVE_NEON)
-#  include "arrow/util/bpacking_neon_internal.h"
-#endif
-
-namespace arrow {
-namespace internal {
-
-int unpack32_scalar(const uint8_t* in_, uint32_t* out, int batch_size, int num_bits) {
-  const uint32_t* in = reinterpret_cast<const uint32_t*>(in_);
+    };
+  }
+};
 
-  batch_size = batch_size / 32 * 32;
-  int num_loops = batch_size / 32;
+}  // namespace
 
-  switch (num_bits) {
-    case 0:
-      for (int i = 0; i < num_loops; ++i) in = nullunpacker32(in, out + i * 32);
-      break;
-    case 1:
-      for (int i = 0; i < num_loops; ++i) in = unpack1_32(in, out + i * 32);
-      break;
-    case 2:
-      for (int i = 0; i < num_loops; ++i) in = unpack2_32(in, out + i * 32);
-      break;
-    case 3:
-      for (int i = 0; i < num_loops; ++i) in = unpack3_32(in, out + i * 32);
-      break;
-    case 4:
-      for (int i = 0; i < num_loops; ++i) in = unpack4_32(in, out + i * 32);
-      break;
-    case 5:
-      for (int i = 0; i < num_loops; ++i) in = unpack5_32(in, out + i * 32);
-      break;
-    case 6:
-      for (int i = 0; i < num_loops; ++i) in = unpack6_32(in, out + i * 32);
-      break;
-    case 7:
-      for (int i = 0; i < num_loops; ++i) in = unpack7_32(in, out + i * 32);
-      break;
-    case 8:
-      for (int i = 0; i < num_loops; ++i) in = unpack8_32(in, out + i * 32);
-      break;
-    case 9:
-      for (int i = 0; i < num_loops; ++i) in = unpack9_32(in, out + i * 32);
-      break;
-    case 10:
-      for (int i = 0; i < num_loops; ++i) in = unpack10_32(in, out + i * 32);
-      break;
-    case 11:
-      for (int i = 0; i < num_loops; ++i) in = unpack11_32(in, out + i * 32);
-      break;
-    case 12:
-      for (int i = 0; i < num_loops; ++i) in = unpack12_32(in, out + i * 32);
-      break;
-    case 13:
-      for (int i = 0; i < num_loops; ++i) in = unpack13_32(in, out + i * 32);
-      break;
-    case 14:
-      for (int i = 0; i < num_loops; ++i) in = unpack14_32(in, out + i * 32);
-      break;
-    case 15:
-      for (int i = 0; i < num_loops; ++i) in = unpack15_32(in, out + i * 32);
-      break;
-    case 16:
-      for (int i = 0; i < num_loops; ++i) in = unpack16_32(in, out + i * 32);
-      break;
-    case 17:
-      for (int i = 0; i < num_loops; ++i) in = unpack17_32(in, out + i * 32);
-      break;
-    case 18:
-      for (int i = 0; i < num_loops; ++i) in = unpack18_32(in, out + i * 32);
-      break;
-    case 19:
-      for (int i = 0; i < num_loops; ++i) in = unpack19_32(in, out + i * 32);
-      break;
-    case 20:
-      for (int i = 0; i < num_loops; ++i) in = unpack20_32(in, out + i * 32);
-      break;
-    case 21:
-      for (int i = 0; i < num_loops; ++i) in = unpack21_32(in, out + i * 32);
-      break;
-    case 22:
-      for (int i = 0; i < num_loops; ++i) in = unpack22_32(in, out + i * 32);
-      break;
-    case 23:
-      for (int i = 0; i < num_loops; ++i) in = unpack23_32(in, out + i * 32);
-      break;
-    case 24:
-      for (int i = 0; i < num_loops; ++i) in = unpack24_32(in, out + i * 32);
-      break;
-    case 25:
-      for (int i = 0; i < num_loops; ++i) in = unpack25_32(in, out + i * 32);
-      break;
-    case 26:
-      for (int i = 0; i < num_loops; ++i) in = unpack26_32(in, out + i * 32);
-      break;
-    case 27:
-      for (int i = 0; i < num_loops; ++i) in = unpack27_32(in, out + i * 32);
-      break;
-    case 28:
-      for (int i = 0; i < num_loops; ++i) in = unpack28_32(in, out + i * 32);
-      break;
-    case 29:
-      for (int i = 0; i < num_loops; ++i) in = unpack29_32(in, out + i * 32);
-      break;
-    case 30:
-      for (int i = 0; i < num_loops; ++i) in = unpack30_32(in, out + i * 32);
-      break;
-    case 31:
-      for (int i = 0; i < num_loops; ++i) in = unpack31_32(in, out + i * 32);
-      break;
-    case 32:
-      for (int i = 0; i < num_loops; ++i) in = unpack32_32(in, out + i * 32);
-      break;
-    default:
-      DCHECK(false) << "Unsupported num_bits";
+template <typename Uint>
+int unpack(const uint8_t* in, Uint* out, int batch_size, int num_bits) {
+  if constexpr (std::is_same_v<Uint, uint16_t>) {
+    // Current SIMD unpack function do not out beat scalar implementation for uin16_t
+    return unpack_scalar<uint16_t>(in, out, batch_size, num_bits);
+  } else {
+#if defined(ARROW_HAVE_NEON)
+    return unpack_neon(in, out, batch_size, num_bits);
+#else
+    static DynamicDispatch<UnpackDynamicFunction<Uint> > dispatch;
+    return dispatch.func(in, out, batch_size, num_bits);
+#endif
   }
-
-  return batch_size;
 }
 
+template int unpack<uint16_t>(const uint8_t*, uint16_t*, int, int);
+template int unpack<uint32_t>(const uint8_t*, uint32_t*, int, int);
+template int unpack<uint64_t>(const uint8_t*, uint64_t*, int, int);
+
 namespace {
 
-struct Unpack32DynamicFunction {
-  using FunctionType = decltype(&unpack32_scalar);
+template <typename Uint>
+struct GetUnpackDynamicFunction {
+  using FunctionType = decltype(&get_unpack_fn_scalar<Uint>);
+  using Implementation = std::pair<DispatchLevel, FunctionType>;
 
-  static std::vector<std::pair<DispatchLevel, FunctionType>> implementations() {
-    return {{DispatchLevel::NONE, unpack32_scalar}
+  static auto implementations() {
+    return std::array {
+      // Current SIMD unpack algorithm works terribly on SSE4.2 due to lack of variable
+      // rhsift and poor xsimd fallback.
+      Implementation{DispatchLevel::NONE, &get_unpack_fn_scalar<Uint>},
 #if defined(ARROW_HAVE_RUNTIME_AVX2)
-            ,
-            {DispatchLevel::AVX2, unpack32_avx2}
+          Implementation{DispatchLevel::AVX2, &get_unpack_fn_avx2<Uint>},
 #endif
 #if defined(ARROW_HAVE_RUNTIME_AVX512)
-            ,
-            {DispatchLevel::AVX512, unpack32_avx512}
+          Implementation{DispatchLevel::AVX512, &get_unpack_fn_avx512<Uint>},
 #endif
     };
   }
@@ -170,226 +92,23 @@ struct Unpack32DynamicFunction {
 
 }  // namespace
 
-int unpack32(const uint8_t* in, uint32_t* out, int batch_size, int num_bits) {
+template <typename Uint>
+UnpackFn<Uint> get_unpack_fn(int num_bits) {
+  if constexpr (std::is_same_v<Uint, uint16_t>) {
+    // Current SIMD unpack function do not out beat scalar implementation for uin16_t
+    return get_unpack_fn_scalar<uint16_t>(num_bits);
+  } else {
 #if defined(ARROW_HAVE_NEON)
-  return unpack32_neon(in, out, batch_size, num_bits);
+    return get_unpack_fn_neon<Uint>(num_bits);
 #else
-  static DynamicDispatch<Unpack32DynamicFunction> dispatch;
-  return dispatch.func(in, out, batch_size, num_bits);
+    static DynamicDispatch<GetUnpackDynamicFunction<Uint> > dispatch;
+    return dispatch.func(num_bits);
 #endif
-}
-
-int unpack64_scalar(const uint8_t* in, uint64_t* out, int batch_size, int num_bits) {
-  batch_size = batch_size / 32 * 32;
-  int num_loops = batch_size / 32;
-
-  switch (num_bits) {
-    case 0:
-      for (int i = 0; i < num_loops; ++i) in = unpack0_64(in, out + i * 32);
-      break;
-    case 1:
-      for (int i = 0; i < num_loops; ++i) in = unpack1_64(in, out + i * 32);
-      break;
-    case 2:
-      for (int i = 0; i < num_loops; ++i) in = unpack2_64(in, out + i * 32);
-      break;
-    case 3:
-      for (int i = 0; i < num_loops; ++i) in = unpack3_64(in, out + i * 32);
-      break;
-    case 4:
-      for (int i = 0; i < num_loops; ++i) in = unpack4_64(in, out + i * 32);
-      break;
-    case 5:
-      for (int i = 0; i < num_loops; ++i) in = unpack5_64(in, out + i * 32);
-      break;
-    case 6:
-      for (int i = 0; i < num_loops; ++i) in = unpack6_64(in, out + i * 32);
-      break;
-    case 7:
-      for (int i = 0; i < num_loops; ++i) in = unpack7_64(in, out + i * 32);
-      break;
-    case 8:
-      for (int i = 0; i < num_loops; ++i) in = unpack8_64(in, out + i * 32);
-      break;
-    case 9:
-      for (int i = 0; i < num_loops; ++i) in = unpack9_64(in, out + i * 32);
-      break;
-    case 10:
-      for (int i = 0; i < num_loops; ++i) in = unpack10_64(in, out + i * 32);
-      break;
-    case 11:
-      for (int i = 0; i < num_loops; ++i) in = unpack11_64(in, out + i * 32);
-      break;
-    case 12:
-      for (int i = 0; i < num_loops; ++i) in = unpack12_64(in, out + i * 32);
-      break;
-    case 13:
-      for (int i = 0; i < num_loops; ++i) in = unpack13_64(in, out + i * 32);
-      break;
-    case 14:
-      for (int i = 0; i < num_loops; ++i) in = unpack14_64(in, out + i * 32);
-      break;
-    case 15:
-      for (int i = 0; i < num_loops; ++i) in = unpack15_64(in, out + i * 32);
-      break;
-    case 16:
-      for (int i = 0; i < num_loops; ++i) in = unpack16_64(in, out + i * 32);
-      break;
-    case 17:
-      for (int i = 0; i < num_loops; ++i) in = unpack17_64(in, out + i * 32);
-      break;
-    case 18:
-      for (int i = 0; i < num_loops; ++i) in = unpack18_64(in, out + i * 32);
-      break;
-    case 19:
-      for (int i = 0; i < num_loops; ++i) in = unpack19_64(in, out + i * 32);
-      break;
-    case 20:
-      for (int i = 0; i < num_loops; ++i) in = unpack20_64(in, out + i * 32);
-      break;
-    case 21:
-      for (int i = 0; i < num_loops; ++i) in = unpack21_64(in, out + i * 32);
-      break;
-    case 22:
-      for (int i = 0; i < num_loops; ++i) in = unpack22_64(in, out + i * 32);
-      break;
-    case 23:
-      for (int i = 0; i < num_loops; ++i) in = unpack23_64(in, out + i * 32);
-      break;
-    case 24:
-      for (int i = 0; i < num_loops; ++i) in = unpack24_64(in, out + i * 32);
-      break;
-    case 25:
-      for (int i = 0; i < num_loops; ++i) in = unpack25_64(in, out + i * 32);
-      break;
-    case 26:
-      for (int i = 0; i < num_loops; ++i) in = unpack26_64(in, out + i * 32);
-      break;
-    case 27:
-      for (int i = 0; i < num_loops; ++i) in = unpack27_64(in, out + i * 32);
-      break;
-    case 28:
-      for (int i = 0; i < num_loops; ++i) in = unpack28_64(in, out + i * 32);
-      break;
-    case 29:
-      for (int i = 0; i < num_loops; ++i) in = unpack29_64(in, out + i * 32);
-      break;
-    case 30:
-      for (int i = 0; i < num_loops; ++i) in = unpack30_64(in, out + i * 32);
-      break;
-    case 31:
-      for (int i = 0; i < num_loops; ++i) in = unpack31_64(in, out + i * 32);
-      break;
-    case 32:
-      for (int i = 0; i < num_loops; ++i) in = unpack32_64(in, out + i * 32);
-      break;
-    case 33:
-      for (int i = 0; i < num_loops; ++i) in = unpack33_64(in, out + i * 32);
-      break;
-    case 34:
-      for (int i = 0; i < num_loops; ++i) in = unpack34_64(in, out + i * 32);
-      break;
-    case 35:
-      for (int i = 0; i < num_loops; ++i) in = unpack35_64(in, out + i * 32);
-      break;
-    case 36:
-      for (int i = 0; i < num_loops; ++i) in = unpack36_64(in, out + i * 32);
-      break;
-    case 37:
-      for (int i = 0; i < num_loops; ++i) in = unpack37_64(in, out + i * 32);
-      break;
-    case 38:
-      for (int i = 0; i < num_loops; ++i) in = unpack38_64(in, out + i * 32);
-      break;
-    case 39:
-      for (int i = 0; i < num_loops; ++i) in = unpack39_64(in, out + i * 32);
-      break;
-    case 40:
-      for (int i = 0; i < num_loops; ++i) in = unpack40_64(in, out + i * 32);
-      break;
-    case 41:
-      for (int i = 0; i < num_loops; ++i) in = unpack41_64(in, out + i * 32);
-      break;
-    case 42:
-      for (int i = 0; i < num_loops; ++i) in = unpack42_64(in, out + i * 32);
-      break;
-    case 43:
-      for (int i = 0; i < num_loops; ++i) in = unpack43_64(in, out + i * 32);
-      break;
-    case 44:
-      for (int i = 0; i < num_loops; ++i) in = unpack44_64(in, out + i * 32);
-      break;
-    case 45:
-      for (int i = 0; i < num_loops; ++i) in = unpack45_64(in, out + i * 32);
-      break;
-    case 46:
-      for (int i = 0; i < num_loops; ++i) in = unpack46_64(in, out + i * 32);
-      break;
-    case 47:
-      for (int i = 0; i < num_loops; ++i) in = unpack47_64(in, out + i * 32);
-      break;
-    case 48:
-      for (int i = 0; i < num_loops; ++i) in = unpack48_64(in, out + i * 32);
-      break;
-    case 49:
-      for (int i = 0; i < num_loops; ++i) in = unpack49_64(in, out + i * 32);
-      break;
-    case 50:
-      for (int i = 0; i < num_loops; ++i) in = unpack50_64(in, out + i * 32);
-      break;
-    case 51:
-      for (int i = 0; i < num_loops; ++i) in = unpack51_64(in, out + i * 32);
-      break;
-    case 52:
-      for (int i = 0; i < num_loops; ++i) in = unpack52_64(in, out + i * 32);
-      break;
-    case 53:
-      for (int i = 0; i < num_loops; ++i) in = unpack53_64(in, out + i * 32);
-      break;
-    case 54:
-      for (int i = 0; i < num_loops; ++i) in = unpack54_64(in, out + i * 32);
-      break;
-    case 55:
-      for (int i = 0; i < num_loops; ++i) in = unpack55_64(in, out + i * 32);
-      break;
-    case 56:
-      for (int i = 0; i < num_loops; ++i) in = unpack56_64(in, out + i * 32);
-      break;
-    case 57:
-      for (int i = 0; i < num_loops; ++i) in = unpack57_64(in, out + i * 32);
-      break;
-    case 58:
-      for (int i = 0; i < num_loops; ++i) in = unpack58_64(in, out + i * 32);
-      break;
-    case 59:
-      for (int i = 0; i < num_loops; ++i) in = unpack59_64(in, out + i * 32);
-      break;
-    case 60:
-      for (int i = 0; i < num_loops; ++i) in = unpack60_64(in, out + i * 32);
-      break;
-    case 61:
-      for (int i = 0; i < num_loops; ++i) in = unpack61_64(in, out + i * 32);
-      break;
-    case 62:
-      for (int i = 0; i < num_loops; ++i) in = unpack62_64(in, out + i * 32);
-      break;
-    case 63:
-      for (int i = 0; i < num_loops; ++i) in = unpack63_64(in, out + i * 32);
-      break;
-    case 64:
-      for (int i = 0; i < num_loops; ++i) in = unpack64_64(in, out + i * 32);
-      break;
-    default:
-      DCHECK(false) << "Unsupported num_bits";
   }
-
-  return batch_size;
 }
 
-int unpack64(const uint8_t* in, uint64_t* out, int batch_size, int num_bits) {
-  // TODO: unpack64_neon, unpack64_avx2 and unpack64_avx512
-  return unpack64_scalar(in, out, batch_size, num_bits);
-}
+template UnpackFn<uint16_t> get_unpack_fn<uint16_t>(int);
+template UnpackFn<uint32_t> get_unpack_fn<uint32_t>(int);
+template UnpackFn<uint64_t> get_unpack_fn<uint64_t>(int);
 
-}  // namespace internal
-}  // namespace arrow
+}  // namespace arrow::internal
diff --git a/cpp/src/arrow/util/bpacking64_codegen.py b/cpp/src/arrow/util/bpacking64_codegen.py
deleted file mode 100644
index 22135fcbb2358..0000000000000
--- a/cpp/src/arrow/util/bpacking64_codegen.py
+++ /dev/null
@@ -1,128 +0,0 @@
-#!/bin/python
-
-# Licensed to the Apache Software Foundation (ASF) under one
-# or more contributor license agreements.  See the NOTICE file
-# distributed with this work for additional information
-# regarding copyright ownership.  The ASF licenses this file
-# to you under the Apache License, Version 2.0 (the
-# "License"); you may not use this file except in compliance
-# with the License.  You may obtain a copy of the License at
-#
-#   http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing,
-# software distributed under the License is distributed on an
-# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
-# KIND, either express or implied.  See the License for the
-# specific language governing permissions and limitations
-# under the License.
-
-# This script is modified from its original version in GitHub. Original source:
-# https://github.com/lemire/FrameOfReference/blob/146948b6058a976bc7767262ad3a2ce201486b93/scripts/turbopacking64.py
-
-# Usage:
-#   python bpacking64_codegen.py > bpacking64_default_internal.h
-
-def howmany(bit):
-    """ how many values are we going to pack? """
-    return 32
-
-
-def howmanywords(bit):
-    return (howmany(bit) * bit + 63)//64
-
-
-def howmanybytes(bit):
-    return (howmany(bit) * bit + 7)//8
-
-
-print('''// Licensed to the Apache Software Foundation (ASF) under one
-// or more contributor license agreements.  See the NOTICE file
-// distributed with this work for additional information
-// regarding copyright ownership.  The ASF licenses this file
-// to you under the Apache License, Version 2.0 (the
-// "License"); you may not use this file except in compliance
-// with the License.  You may obtain a copy of the License at
-//
-//   http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing,
-// software distributed under the License is distributed on an
-// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
-// KIND, either express or implied.  See the License for the
-// specific language governing permissions and limitations
-// under the License.
-
-// This file was generated by script which is modified from its original version in GitHub.
-// Original source:
-// https://github.com/lemire/FrameOfReference/blob/master/scripts/turbopacking64.py
-// The original copyright notice follows.
-
-// This code is released under the
-// Apache License Version 2.0 http://www.apache.org/licenses/.
-// (c) Daniel Lemire 2013
-
-#pragma once
-
-#include "arrow/util/bit_util.h"
-#include "arrow/util/ubsan.h"
-
-namespace arrow {
-namespace internal {
-''')
-
-
-print("inline const uint8_t* unpack0_64(const uint8_t* in, uint64_t* out) {")
-print(f"  for(int k = 0; k < {howmany(0)} ; k += 1) {{")
-print("    out[k] = 0;")
-print("  }")
-print("  return in;")
-print("}")
-
-for bit in range(1, 65):
-    print("")
-    print(f"inline const uint8_t* unpack{bit}_64(const uint8_t* in, uint64_t* out) {{")
-
-    if(bit < 64):
-        print(f"  const uint64_t mask = {((1 << bit)-1)}ULL;")
-    maskstr = " & mask"
-    if (bit == 64):
-        maskstr = ""  # no need
-
-    for k in range(howmanywords(bit)-1):
-        print(f"  uint64_t w{k} = util::SafeLoadAs<uint64_t>(in);")
-        print(f"  w{k} = arrow::BitUtil::FromLittleEndian(w{k});")
-        print("  in += 8;")
-    k = howmanywords(bit) - 1
-    if (bit % 2 == 0):
-        print(f"  uint64_t w{k} = util::SafeLoadAs<uint64_t>(in);")
-        print(f"  w{k} = arrow::BitUtil::FromLittleEndian(w{k});")
-        print("  in += 8;")
-    else:
-        print(f"  uint64_t w{k} = util::SafeLoadAs<uint32_t>(in);")
-        print(f"  w{k} = arrow::BitUtil::FromLittleEndian(w{k});")
-        print("  in += 4;")
-
-    for j in range(howmany(bit)):
-        firstword = j * bit // 64
-        secondword = (j * bit + bit - 1)//64
-        firstshift = (j*bit) % 64
-        firstshiftstr = f" >> {firstshift}"
-        if(firstshift == 0):
-            firstshiftstr = ""  # no need
-        if(firstword == secondword):
-            if(firstshift + bit == 64):
-                print(f"  out[{j}] = w{firstword}{firstshiftstr};")
-            else:
-                print(f"  out[{j}] = (w{firstword}{firstshiftstr}){maskstr};")
-        else:
-            secondshift = (64-firstshift)
-            print(f"  out[{j}] = ((w{firstword}{firstshiftstr}) | "
-                  f"(w{firstword+1} << {secondshift})){maskstr};")
-    print("")
-    print("  return in;")
-    print("}")
-
-print('''
-}  // namespace internal
-}  // namespace arrow''')
diff --git a/cpp/src/arrow/util/bpacking64_default_internal.h b/cpp/src/arrow/util/bpacking64_default_internal.h
deleted file mode 100644
index 256cdda87e34f..0000000000000
--- a/cpp/src/arrow/util/bpacking64_default_internal.h
+++ /dev/null
@@ -1,5640 +0,0 @@
-// Licensed to the Apache Software Foundation (ASF) under one
-// or more contributor license agreements.  See the NOTICE file
-// distributed with this work for additional information
-// regarding copyright ownership.  The ASF licenses this file
-// to you under the Apache License, Version 2.0 (the
-// "License"); you may not use this file except in compliance
-// with the License.  You may obtain a copy of the License at
-//
-//   http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing,
-// software distributed under the License is distributed on an
-// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
-// KIND, either express or implied.  See the License for the
-// specific language governing permissions and limitations
-// under the License.
-
-// This file was generated by script which is modified from its original version in
-// GitHub. Original source:
-// https://github.com/lemire/FrameOfReference/blob/146948b6058a976bc7767262ad3a2ce201486b93/scripts/turbopacking64.py
-// The original copyright notice follows.
-
-// This code is released under the
-// Apache License Version 2.0 http://www.apache.org/licenses/.
-// (c) Daniel Lemire 2013
-
-#pragma once
-
-#include "arrow/util/endian.h"
-#include "arrow/util/ubsan.h"
-
-namespace arrow::internal {
-
-inline const uint8_t* unpack0_64(const uint8_t* in, uint64_t* out) {
-  for (int k = 0; k < 32; k += 1) {
-    out[k] = 0;
-  }
-  return in;
-}
-
-inline const uint8_t* unpack1_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 1ULL;
-  uint64_t w0 = util::SafeLoadAs<uint32_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 4;
-  out[0] = (w0)&mask;
-  out[1] = (w0 >> 1) & mask;
-  out[2] = (w0 >> 2) & mask;
-  out[3] = (w0 >> 3) & mask;
-  out[4] = (w0 >> 4) & mask;
-  out[5] = (w0 >> 5) & mask;
-  out[6] = (w0 >> 6) & mask;
-  out[7] = (w0 >> 7) & mask;
-  out[8] = (w0 >> 8) & mask;
-  out[9] = (w0 >> 9) & mask;
-  out[10] = (w0 >> 10) & mask;
-  out[11] = (w0 >> 11) & mask;
-  out[12] = (w0 >> 12) & mask;
-  out[13] = (w0 >> 13) & mask;
-  out[14] = (w0 >> 14) & mask;
-  out[15] = (w0 >> 15) & mask;
-  out[16] = (w0 >> 16) & mask;
-  out[17] = (w0 >> 17) & mask;
-  out[18] = (w0 >> 18) & mask;
-  out[19] = (w0 >> 19) & mask;
-  out[20] = (w0 >> 20) & mask;
-  out[21] = (w0 >> 21) & mask;
-  out[22] = (w0 >> 22) & mask;
-  out[23] = (w0 >> 23) & mask;
-  out[24] = (w0 >> 24) & mask;
-  out[25] = (w0 >> 25) & mask;
-  out[26] = (w0 >> 26) & mask;
-  out[27] = (w0 >> 27) & mask;
-  out[28] = (w0 >> 28) & mask;
-  out[29] = (w0 >> 29) & mask;
-  out[30] = (w0 >> 30) & mask;
-  out[31] = (w0 >> 31) & mask;
-
-  return in;
-}
-
-inline const uint8_t* unpack2_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 3ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  out[0] = (w0)&mask;
-  out[1] = (w0 >> 2) & mask;
-  out[2] = (w0 >> 4) & mask;
-  out[3] = (w0 >> 6) & mask;
-  out[4] = (w0 >> 8) & mask;
-  out[5] = (w0 >> 10) & mask;
-  out[6] = (w0 >> 12) & mask;
-  out[7] = (w0 >> 14) & mask;
-  out[8] = (w0 >> 16) & mask;
-  out[9] = (w0 >> 18) & mask;
-  out[10] = (w0 >> 20) & mask;
-  out[11] = (w0 >> 22) & mask;
-  out[12] = (w0 >> 24) & mask;
-  out[13] = (w0 >> 26) & mask;
-  out[14] = (w0 >> 28) & mask;
-  out[15] = (w0 >> 30) & mask;
-  out[16] = (w0 >> 32) & mask;
-  out[17] = (w0 >> 34) & mask;
-  out[18] = (w0 >> 36) & mask;
-  out[19] = (w0 >> 38) & mask;
-  out[20] = (w0 >> 40) & mask;
-  out[21] = (w0 >> 42) & mask;
-  out[22] = (w0 >> 44) & mask;
-  out[23] = (w0 >> 46) & mask;
-  out[24] = (w0 >> 48) & mask;
-  out[25] = (w0 >> 50) & mask;
-  out[26] = (w0 >> 52) & mask;
-  out[27] = (w0 >> 54) & mask;
-  out[28] = (w0 >> 56) & mask;
-  out[29] = (w0 >> 58) & mask;
-  out[30] = (w0 >> 60) & mask;
-  out[31] = w0 >> 62;
-
-  return in;
-}
-
-inline const uint8_t* unpack3_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 7ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint32_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 4;
-  out[0] = (w0)&mask;
-  out[1] = (w0 >> 3) & mask;
-  out[2] = (w0 >> 6) & mask;
-  out[3] = (w0 >> 9) & mask;
-  out[4] = (w0 >> 12) & mask;
-  out[5] = (w0 >> 15) & mask;
-  out[6] = (w0 >> 18) & mask;
-  out[7] = (w0 >> 21) & mask;
-  out[8] = (w0 >> 24) & mask;
-  out[9] = (w0 >> 27) & mask;
-  out[10] = (w0 >> 30) & mask;
-  out[11] = (w0 >> 33) & mask;
-  out[12] = (w0 >> 36) & mask;
-  out[13] = (w0 >> 39) & mask;
-  out[14] = (w0 >> 42) & mask;
-  out[15] = (w0 >> 45) & mask;
-  out[16] = (w0 >> 48) & mask;
-  out[17] = (w0 >> 51) & mask;
-  out[18] = (w0 >> 54) & mask;
-  out[19] = (w0 >> 57) & mask;
-  out[20] = (w0 >> 60) & mask;
-  out[21] = ((w0 >> 63) | (w1 << 1)) & mask;
-  out[22] = (w1 >> 2) & mask;
-  out[23] = (w1 >> 5) & mask;
-  out[24] = (w1 >> 8) & mask;
-  out[25] = (w1 >> 11) & mask;
-  out[26] = (w1 >> 14) & mask;
-  out[27] = (w1 >> 17) & mask;
-  out[28] = (w1 >> 20) & mask;
-  out[29] = (w1 >> 23) & mask;
-  out[30] = (w1 >> 26) & mask;
-  out[31] = (w1 >> 29) & mask;
-
-  return in;
-}
-
-inline const uint8_t* unpack4_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 15ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  out[0] = (w0)&mask;
-  out[1] = (w0 >> 4) & mask;
-  out[2] = (w0 >> 8) & mask;
-  out[3] = (w0 >> 12) & mask;
-  out[4] = (w0 >> 16) & mask;
-  out[5] = (w0 >> 20) & mask;
-  out[6] = (w0 >> 24) & mask;
-  out[7] = (w0 >> 28) & mask;
-  out[8] = (w0 >> 32) & mask;
-  out[9] = (w0 >> 36) & mask;
-  out[10] = (w0 >> 40) & mask;
-  out[11] = (w0 >> 44) & mask;
-  out[12] = (w0 >> 48) & mask;
-  out[13] = (w0 >> 52) & mask;
-  out[14] = (w0 >> 56) & mask;
-  out[15] = w0 >> 60;
-  out[16] = (w1)&mask;
-  out[17] = (w1 >> 4) & mask;
-  out[18] = (w1 >> 8) & mask;
-  out[19] = (w1 >> 12) & mask;
-  out[20] = (w1 >> 16) & mask;
-  out[21] = (w1 >> 20) & mask;
-  out[22] = (w1 >> 24) & mask;
-  out[23] = (w1 >> 28) & mask;
-  out[24] = (w1 >> 32) & mask;
-  out[25] = (w1 >> 36) & mask;
-  out[26] = (w1 >> 40) & mask;
-  out[27] = (w1 >> 44) & mask;
-  out[28] = (w1 >> 48) & mask;
-  out[29] = (w1 >> 52) & mask;
-  out[30] = (w1 >> 56) & mask;
-  out[31] = w1 >> 60;
-
-  return in;
-}
-
-inline const uint8_t* unpack5_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 31ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint32_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 4;
-  out[0] = (w0)&mask;
-  out[1] = (w0 >> 5) & mask;
-  out[2] = (w0 >> 10) & mask;
-  out[3] = (w0 >> 15) & mask;
-  out[4] = (w0 >> 20) & mask;
-  out[5] = (w0 >> 25) & mask;
-  out[6] = (w0 >> 30) & mask;
-  out[7] = (w0 >> 35) & mask;
-  out[8] = (w0 >> 40) & mask;
-  out[9] = (w0 >> 45) & mask;
-  out[10] = (w0 >> 50) & mask;
-  out[11] = (w0 >> 55) & mask;
-  out[12] = ((w0 >> 60) | (w1 << 4)) & mask;
-  out[13] = (w1 >> 1) & mask;
-  out[14] = (w1 >> 6) & mask;
-  out[15] = (w1 >> 11) & mask;
-  out[16] = (w1 >> 16) & mask;
-  out[17] = (w1 >> 21) & mask;
-  out[18] = (w1 >> 26) & mask;
-  out[19] = (w1 >> 31) & mask;
-  out[20] = (w1 >> 36) & mask;
-  out[21] = (w1 >> 41) & mask;
-  out[22] = (w1 >> 46) & mask;
-  out[23] = (w1 >> 51) & mask;
-  out[24] = (w1 >> 56) & mask;
-  out[25] = ((w1 >> 61) | (w2 << 3)) & mask;
-  out[26] = (w2 >> 2) & mask;
-  out[27] = (w2 >> 7) & mask;
-  out[28] = (w2 >> 12) & mask;
-  out[29] = (w2 >> 17) & mask;
-  out[30] = (w2 >> 22) & mask;
-  out[31] = (w2 >> 27) & mask;
-
-  return in;
-}
-
-inline const uint8_t* unpack6_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 63ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  out[0] = (w0)&mask;
-  out[1] = (w0 >> 6) & mask;
-  out[2] = (w0 >> 12) & mask;
-  out[3] = (w0 >> 18) & mask;
-  out[4] = (w0 >> 24) & mask;
-  out[5] = (w0 >> 30) & mask;
-  out[6] = (w0 >> 36) & mask;
-  out[7] = (w0 >> 42) & mask;
-  out[8] = (w0 >> 48) & mask;
-  out[9] = (w0 >> 54) & mask;
-  out[10] = ((w0 >> 60) | (w1 << 4)) & mask;
-  out[11] = (w1 >> 2) & mask;
-  out[12] = (w1 >> 8) & mask;
-  out[13] = (w1 >> 14) & mask;
-  out[14] = (w1 >> 20) & mask;
-  out[15] = (w1 >> 26) & mask;
-  out[16] = (w1 >> 32) & mask;
-  out[17] = (w1 >> 38) & mask;
-  out[18] = (w1 >> 44) & mask;
-  out[19] = (w1 >> 50) & mask;
-  out[20] = (w1 >> 56) & mask;
-  out[21] = ((w1 >> 62) | (w2 << 2)) & mask;
-  out[22] = (w2 >> 4) & mask;
-  out[23] = (w2 >> 10) & mask;
-  out[24] = (w2 >> 16) & mask;
-  out[25] = (w2 >> 22) & mask;
-  out[26] = (w2 >> 28) & mask;
-  out[27] = (w2 >> 34) & mask;
-  out[28] = (w2 >> 40) & mask;
-  out[29] = (w2 >> 46) & mask;
-  out[30] = (w2 >> 52) & mask;
-  out[31] = w2 >> 58;
-
-  return in;
-}
-
-inline const uint8_t* unpack7_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 127ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint32_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 4;
-  out[0] = (w0)&mask;
-  out[1] = (w0 >> 7) & mask;
-  out[2] = (w0 >> 14) & mask;
-  out[3] = (w0 >> 21) & mask;
-  out[4] = (w0 >> 28) & mask;
-  out[5] = (w0 >> 35) & mask;
-  out[6] = (w0 >> 42) & mask;
-  out[7] = (w0 >> 49) & mask;
-  out[8] = (w0 >> 56) & mask;
-  out[9] = ((w0 >> 63) | (w1 << 1)) & mask;
-  out[10] = (w1 >> 6) & mask;
-  out[11] = (w1 >> 13) & mask;
-  out[12] = (w1 >> 20) & mask;
-  out[13] = (w1 >> 27) & mask;
-  out[14] = (w1 >> 34) & mask;
-  out[15] = (w1 >> 41) & mask;
-  out[16] = (w1 >> 48) & mask;
-  out[17] = (w1 >> 55) & mask;
-  out[18] = ((w1 >> 62) | (w2 << 2)) & mask;
-  out[19] = (w2 >> 5) & mask;
-  out[20] = (w2 >> 12) & mask;
-  out[21] = (w2 >> 19) & mask;
-  out[22] = (w2 >> 26) & mask;
-  out[23] = (w2 >> 33) & mask;
-  out[24] = (w2 >> 40) & mask;
-  out[25] = (w2 >> 47) & mask;
-  out[26] = (w2 >> 54) & mask;
-  out[27] = ((w2 >> 61) | (w3 << 3)) & mask;
-  out[28] = (w3 >> 4) & mask;
-  out[29] = (w3 >> 11) & mask;
-  out[30] = (w3 >> 18) & mask;
-  out[31] = (w3 >> 25) & mask;
-
-  return in;
-}
-
-inline const uint8_t* unpack8_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 255ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  out[0] = (w0)&mask;
-  out[1] = (w0 >> 8) & mask;
-  out[2] = (w0 >> 16) & mask;
-  out[3] = (w0 >> 24) & mask;
-  out[4] = (w0 >> 32) & mask;
-  out[5] = (w0 >> 40) & mask;
-  out[6] = (w0 >> 48) & mask;
-  out[7] = w0 >> 56;
-  out[8] = (w1)&mask;
-  out[9] = (w1 >> 8) & mask;
-  out[10] = (w1 >> 16) & mask;
-  out[11] = (w1 >> 24) & mask;
-  out[12] = (w1 >> 32) & mask;
-  out[13] = (w1 >> 40) & mask;
-  out[14] = (w1 >> 48) & mask;
-  out[15] = w1 >> 56;
-  out[16] = (w2)&mask;
-  out[17] = (w2 >> 8) & mask;
-  out[18] = (w2 >> 16) & mask;
-  out[19] = (w2 >> 24) & mask;
-  out[20] = (w2 >> 32) & mask;
-  out[21] = (w2 >> 40) & mask;
-  out[22] = (w2 >> 48) & mask;
-  out[23] = w2 >> 56;
-  out[24] = (w3)&mask;
-  out[25] = (w3 >> 8) & mask;
-  out[26] = (w3 >> 16) & mask;
-  out[27] = (w3 >> 24) & mask;
-  out[28] = (w3 >> 32) & mask;
-  out[29] = (w3 >> 40) & mask;
-  out[30] = (w3 >> 48) & mask;
-  out[31] = w3 >> 56;
-
-  return in;
-}
-
-inline const uint8_t* unpack9_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 511ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint32_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 4;
-  out[0] = (w0)&mask;
-  out[1] = (w0 >> 9) & mask;
-  out[2] = (w0 >> 18) & mask;
-  out[3] = (w0 >> 27) & mask;
-  out[4] = (w0 >> 36) & mask;
-  out[5] = (w0 >> 45) & mask;
-  out[6] = (w0 >> 54) & mask;
-  out[7] = ((w0 >> 63) | (w1 << 1)) & mask;
-  out[8] = (w1 >> 8) & mask;
-  out[9] = (w1 >> 17) & mask;
-  out[10] = (w1 >> 26) & mask;
-  out[11] = (w1 >> 35) & mask;
-  out[12] = (w1 >> 44) & mask;
-  out[13] = (w1 >> 53) & mask;
-  out[14] = ((w1 >> 62) | (w2 << 2)) & mask;
-  out[15] = (w2 >> 7) & mask;
-  out[16] = (w2 >> 16) & mask;
-  out[17] = (w2 >> 25) & mask;
-  out[18] = (w2 >> 34) & mask;
-  out[19] = (w2 >> 43) & mask;
-  out[20] = (w2 >> 52) & mask;
-  out[21] = ((w2 >> 61) | (w3 << 3)) & mask;
-  out[22] = (w3 >> 6) & mask;
-  out[23] = (w3 >> 15) & mask;
-  out[24] = (w3 >> 24) & mask;
-  out[25] = (w3 >> 33) & mask;
-  out[26] = (w3 >> 42) & mask;
-  out[27] = (w3 >> 51) & mask;
-  out[28] = ((w3 >> 60) | (w4 << 4)) & mask;
-  out[29] = (w4 >> 5) & mask;
-  out[30] = (w4 >> 14) & mask;
-  out[31] = (w4 >> 23) & mask;
-
-  return in;
-}
-
-inline const uint8_t* unpack10_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 1023ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  out[0] = (w0)&mask;
-  out[1] = (w0 >> 10) & mask;
-  out[2] = (w0 >> 20) & mask;
-  out[3] = (w0 >> 30) & mask;
-  out[4] = (w0 >> 40) & mask;
-  out[5] = (w0 >> 50) & mask;
-  out[6] = ((w0 >> 60) | (w1 << 4)) & mask;
-  out[7] = (w1 >> 6) & mask;
-  out[8] = (w1 >> 16) & mask;
-  out[9] = (w1 >> 26) & mask;
-  out[10] = (w1 >> 36) & mask;
-  out[11] = (w1 >> 46) & mask;
-  out[12] = ((w1 >> 56) | (w2 << 8)) & mask;
-  out[13] = (w2 >> 2) & mask;
-  out[14] = (w2 >> 12) & mask;
-  out[15] = (w2 >> 22) & mask;
-  out[16] = (w2 >> 32) & mask;
-  out[17] = (w2 >> 42) & mask;
-  out[18] = (w2 >> 52) & mask;
-  out[19] = ((w2 >> 62) | (w3 << 2)) & mask;
-  out[20] = (w3 >> 8) & mask;
-  out[21] = (w3 >> 18) & mask;
-  out[22] = (w3 >> 28) & mask;
-  out[23] = (w3 >> 38) & mask;
-  out[24] = (w3 >> 48) & mask;
-  out[25] = ((w3 >> 58) | (w4 << 6)) & mask;
-  out[26] = (w4 >> 4) & mask;
-  out[27] = (w4 >> 14) & mask;
-  out[28] = (w4 >> 24) & mask;
-  out[29] = (w4 >> 34) & mask;
-  out[30] = (w4 >> 44) & mask;
-  out[31] = w4 >> 54;
-
-  return in;
-}
-
-inline const uint8_t* unpack11_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 2047ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint32_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 4;
-  out[0] = (w0)&mask;
-  out[1] = (w0 >> 11) & mask;
-  out[2] = (w0 >> 22) & mask;
-  out[3] = (w0 >> 33) & mask;
-  out[4] = (w0 >> 44) & mask;
-  out[5] = ((w0 >> 55) | (w1 << 9)) & mask;
-  out[6] = (w1 >> 2) & mask;
-  out[7] = (w1 >> 13) & mask;
-  out[8] = (w1 >> 24) & mask;
-  out[9] = (w1 >> 35) & mask;
-  out[10] = (w1 >> 46) & mask;
-  out[11] = ((w1 >> 57) | (w2 << 7)) & mask;
-  out[12] = (w2 >> 4) & mask;
-  out[13] = (w2 >> 15) & mask;
-  out[14] = (w2 >> 26) & mask;
-  out[15] = (w2 >> 37) & mask;
-  out[16] = (w2 >> 48) & mask;
-  out[17] = ((w2 >> 59) | (w3 << 5)) & mask;
-  out[18] = (w3 >> 6) & mask;
-  out[19] = (w3 >> 17) & mask;
-  out[20] = (w3 >> 28) & mask;
-  out[21] = (w3 >> 39) & mask;
-  out[22] = (w3 >> 50) & mask;
-  out[23] = ((w3 >> 61) | (w4 << 3)) & mask;
-  out[24] = (w4 >> 8) & mask;
-  out[25] = (w4 >> 19) & mask;
-  out[26] = (w4 >> 30) & mask;
-  out[27] = (w4 >> 41) & mask;
-  out[28] = (w4 >> 52) & mask;
-  out[29] = ((w4 >> 63) | (w5 << 1)) & mask;
-  out[30] = (w5 >> 10) & mask;
-  out[31] = (w5 >> 21) & mask;
-
-  return in;
-}
-
-inline const uint8_t* unpack12_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 4095ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  out[0] = (w0)&mask;
-  out[1] = (w0 >> 12) & mask;
-  out[2] = (w0 >> 24) & mask;
-  out[3] = (w0 >> 36) & mask;
-  out[4] = (w0 >> 48) & mask;
-  out[5] = ((w0 >> 60) | (w1 << 4)) & mask;
-  out[6] = (w1 >> 8) & mask;
-  out[7] = (w1 >> 20) & mask;
-  out[8] = (w1 >> 32) & mask;
-  out[9] = (w1 >> 44) & mask;
-  out[10] = ((w1 >> 56) | (w2 << 8)) & mask;
-  out[11] = (w2 >> 4) & mask;
-  out[12] = (w2 >> 16) & mask;
-  out[13] = (w2 >> 28) & mask;
-  out[14] = (w2 >> 40) & mask;
-  out[15] = w2 >> 52;
-  out[16] = (w3)&mask;
-  out[17] = (w3 >> 12) & mask;
-  out[18] = (w3 >> 24) & mask;
-  out[19] = (w3 >> 36) & mask;
-  out[20] = (w3 >> 48) & mask;
-  out[21] = ((w3 >> 60) | (w4 << 4)) & mask;
-  out[22] = (w4 >> 8) & mask;
-  out[23] = (w4 >> 20) & mask;
-  out[24] = (w4 >> 32) & mask;
-  out[25] = (w4 >> 44) & mask;
-  out[26] = ((w4 >> 56) | (w5 << 8)) & mask;
-  out[27] = (w5 >> 4) & mask;
-  out[28] = (w5 >> 16) & mask;
-  out[29] = (w5 >> 28) & mask;
-  out[30] = (w5 >> 40) & mask;
-  out[31] = w5 >> 52;
-
-  return in;
-}
-
-inline const uint8_t* unpack13_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 8191ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint32_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 4;
-  out[0] = (w0)&mask;
-  out[1] = (w0 >> 13) & mask;
-  out[2] = (w0 >> 26) & mask;
-  out[3] = (w0 >> 39) & mask;
-  out[4] = ((w0 >> 52) | (w1 << 12)) & mask;
-  out[5] = (w1 >> 1) & mask;
-  out[6] = (w1 >> 14) & mask;
-  out[7] = (w1 >> 27) & mask;
-  out[8] = (w1 >> 40) & mask;
-  out[9] = ((w1 >> 53) | (w2 << 11)) & mask;
-  out[10] = (w2 >> 2) & mask;
-  out[11] = (w2 >> 15) & mask;
-  out[12] = (w2 >> 28) & mask;
-  out[13] = (w2 >> 41) & mask;
-  out[14] = ((w2 >> 54) | (w3 << 10)) & mask;
-  out[15] = (w3 >> 3) & mask;
-  out[16] = (w3 >> 16) & mask;
-  out[17] = (w3 >> 29) & mask;
-  out[18] = (w3 >> 42) & mask;
-  out[19] = ((w3 >> 55) | (w4 << 9)) & mask;
-  out[20] = (w4 >> 4) & mask;
-  out[21] = (w4 >> 17) & mask;
-  out[22] = (w4 >> 30) & mask;
-  out[23] = (w4 >> 43) & mask;
-  out[24] = ((w4 >> 56) | (w5 << 8)) & mask;
-  out[25] = (w5 >> 5) & mask;
-  out[26] = (w5 >> 18) & mask;
-  out[27] = (w5 >> 31) & mask;
-  out[28] = (w5 >> 44) & mask;
-  out[29] = ((w5 >> 57) | (w6 << 7)) & mask;
-  out[30] = (w6 >> 6) & mask;
-  out[31] = (w6 >> 19) & mask;
-
-  return in;
-}
-
-inline const uint8_t* unpack14_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 16383ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  out[0] = (w0)&mask;
-  out[1] = (w0 >> 14) & mask;
-  out[2] = (w0 >> 28) & mask;
-  out[3] = (w0 >> 42) & mask;
-  out[4] = ((w0 >> 56) | (w1 << 8)) & mask;
-  out[5] = (w1 >> 6) & mask;
-  out[6] = (w1 >> 20) & mask;
-  out[7] = (w1 >> 34) & mask;
-  out[8] = (w1 >> 48) & mask;
-  out[9] = ((w1 >> 62) | (w2 << 2)) & mask;
-  out[10] = (w2 >> 12) & mask;
-  out[11] = (w2 >> 26) & mask;
-  out[12] = (w2 >> 40) & mask;
-  out[13] = ((w2 >> 54) | (w3 << 10)) & mask;
-  out[14] = (w3 >> 4) & mask;
-  out[15] = (w3 >> 18) & mask;
-  out[16] = (w3 >> 32) & mask;
-  out[17] = (w3 >> 46) & mask;
-  out[18] = ((w3 >> 60) | (w4 << 4)) & mask;
-  out[19] = (w4 >> 10) & mask;
-  out[20] = (w4 >> 24) & mask;
-  out[21] = (w4 >> 38) & mask;
-  out[22] = ((w4 >> 52) | (w5 << 12)) & mask;
-  out[23] = (w5 >> 2) & mask;
-  out[24] = (w5 >> 16) & mask;
-  out[25] = (w5 >> 30) & mask;
-  out[26] = (w5 >> 44) & mask;
-  out[27] = ((w5 >> 58) | (w6 << 6)) & mask;
-  out[28] = (w6 >> 8) & mask;
-  out[29] = (w6 >> 22) & mask;
-  out[30] = (w6 >> 36) & mask;
-  out[31] = w6 >> 50;
-
-  return in;
-}
-
-inline const uint8_t* unpack15_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 32767ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint32_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 4;
-  out[0] = (w0)&mask;
-  out[1] = (w0 >> 15) & mask;
-  out[2] = (w0 >> 30) & mask;
-  out[3] = (w0 >> 45) & mask;
-  out[4] = ((w0 >> 60) | (w1 << 4)) & mask;
-  out[5] = (w1 >> 11) & mask;
-  out[6] = (w1 >> 26) & mask;
-  out[7] = (w1 >> 41) & mask;
-  out[8] = ((w1 >> 56) | (w2 << 8)) & mask;
-  out[9] = (w2 >> 7) & mask;
-  out[10] = (w2 >> 22) & mask;
-  out[11] = (w2 >> 37) & mask;
-  out[12] = ((w2 >> 52) | (w3 << 12)) & mask;
-  out[13] = (w3 >> 3) & mask;
-  out[14] = (w3 >> 18) & mask;
-  out[15] = (w3 >> 33) & mask;
-  out[16] = (w3 >> 48) & mask;
-  out[17] = ((w3 >> 63) | (w4 << 1)) & mask;
-  out[18] = (w4 >> 14) & mask;
-  out[19] = (w4 >> 29) & mask;
-  out[20] = (w4 >> 44) & mask;
-  out[21] = ((w4 >> 59) | (w5 << 5)) & mask;
-  out[22] = (w5 >> 10) & mask;
-  out[23] = (w5 >> 25) & mask;
-  out[24] = (w5 >> 40) & mask;
-  out[25] = ((w5 >> 55) | (w6 << 9)) & mask;
-  out[26] = (w6 >> 6) & mask;
-  out[27] = (w6 >> 21) & mask;
-  out[28] = (w6 >> 36) & mask;
-  out[29] = ((w6 >> 51) | (w7 << 13)) & mask;
-  out[30] = (w7 >> 2) & mask;
-  out[31] = (w7 >> 17) & mask;
-
-  return in;
-}
-
-inline const uint8_t* unpack16_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 65535ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  out[0] = (w0)&mask;
-  out[1] = (w0 >> 16) & mask;
-  out[2] = (w0 >> 32) & mask;
-  out[3] = w0 >> 48;
-  out[4] = (w1)&mask;
-  out[5] = (w1 >> 16) & mask;
-  out[6] = (w1 >> 32) & mask;
-  out[7] = w1 >> 48;
-  out[8] = (w2)&mask;
-  out[9] = (w2 >> 16) & mask;
-  out[10] = (w2 >> 32) & mask;
-  out[11] = w2 >> 48;
-  out[12] = (w3)&mask;
-  out[13] = (w3 >> 16) & mask;
-  out[14] = (w3 >> 32) & mask;
-  out[15] = w3 >> 48;
-  out[16] = (w4)&mask;
-  out[17] = (w4 >> 16) & mask;
-  out[18] = (w4 >> 32) & mask;
-  out[19] = w4 >> 48;
-  out[20] = (w5)&mask;
-  out[21] = (w5 >> 16) & mask;
-  out[22] = (w5 >> 32) & mask;
-  out[23] = w5 >> 48;
-  out[24] = (w6)&mask;
-  out[25] = (w6 >> 16) & mask;
-  out[26] = (w6 >> 32) & mask;
-  out[27] = w6 >> 48;
-  out[28] = (w7)&mask;
-  out[29] = (w7 >> 16) & mask;
-  out[30] = (w7 >> 32) & mask;
-  out[31] = w7 >> 48;
-
-  return in;
-}
-
-inline const uint8_t* unpack17_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 131071ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint32_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 4;
-  out[0] = (w0)&mask;
-  out[1] = (w0 >> 17) & mask;
-  out[2] = (w0 >> 34) & mask;
-  out[3] = ((w0 >> 51) | (w1 << 13)) & mask;
-  out[4] = (w1 >> 4) & mask;
-  out[5] = (w1 >> 21) & mask;
-  out[6] = (w1 >> 38) & mask;
-  out[7] = ((w1 >> 55) | (w2 << 9)) & mask;
-  out[8] = (w2 >> 8) & mask;
-  out[9] = (w2 >> 25) & mask;
-  out[10] = (w2 >> 42) & mask;
-  out[11] = ((w2 >> 59) | (w3 << 5)) & mask;
-  out[12] = (w3 >> 12) & mask;
-  out[13] = (w3 >> 29) & mask;
-  out[14] = (w3 >> 46) & mask;
-  out[15] = ((w3 >> 63) | (w4 << 1)) & mask;
-  out[16] = (w4 >> 16) & mask;
-  out[17] = (w4 >> 33) & mask;
-  out[18] = ((w4 >> 50) | (w5 << 14)) & mask;
-  out[19] = (w5 >> 3) & mask;
-  out[20] = (w5 >> 20) & mask;
-  out[21] = (w5 >> 37) & mask;
-  out[22] = ((w5 >> 54) | (w6 << 10)) & mask;
-  out[23] = (w6 >> 7) & mask;
-  out[24] = (w6 >> 24) & mask;
-  out[25] = (w6 >> 41) & mask;
-  out[26] = ((w6 >> 58) | (w7 << 6)) & mask;
-  out[27] = (w7 >> 11) & mask;
-  out[28] = (w7 >> 28) & mask;
-  out[29] = (w7 >> 45) & mask;
-  out[30] = ((w7 >> 62) | (w8 << 2)) & mask;
-  out[31] = (w8 >> 15) & mask;
-
-  return in;
-}
-
-inline const uint8_t* unpack18_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 262143ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint64_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 8;
-  out[0] = (w0)&mask;
-  out[1] = (w0 >> 18) & mask;
-  out[2] = (w0 >> 36) & mask;
-  out[3] = ((w0 >> 54) | (w1 << 10)) & mask;
-  out[4] = (w1 >> 8) & mask;
-  out[5] = (w1 >> 26) & mask;
-  out[6] = (w1 >> 44) & mask;
-  out[7] = ((w1 >> 62) | (w2 << 2)) & mask;
-  out[8] = (w2 >> 16) & mask;
-  out[9] = (w2 >> 34) & mask;
-  out[10] = ((w2 >> 52) | (w3 << 12)) & mask;
-  out[11] = (w3 >> 6) & mask;
-  out[12] = (w3 >> 24) & mask;
-  out[13] = (w3 >> 42) & mask;
-  out[14] = ((w3 >> 60) | (w4 << 4)) & mask;
-  out[15] = (w4 >> 14) & mask;
-  out[16] = (w4 >> 32) & mask;
-  out[17] = ((w4 >> 50) | (w5 << 14)) & mask;
-  out[18] = (w5 >> 4) & mask;
-  out[19] = (w5 >> 22) & mask;
-  out[20] = (w5 >> 40) & mask;
-  out[21] = ((w5 >> 58) | (w6 << 6)) & mask;
-  out[22] = (w6 >> 12) & mask;
-  out[23] = (w6 >> 30) & mask;
-  out[24] = ((w6 >> 48) | (w7 << 16)) & mask;
-  out[25] = (w7 >> 2) & mask;
-  out[26] = (w7 >> 20) & mask;
-  out[27] = (w7 >> 38) & mask;
-  out[28] = ((w7 >> 56) | (w8 << 8)) & mask;
-  out[29] = (w8 >> 10) & mask;
-  out[30] = (w8 >> 28) & mask;
-  out[31] = w8 >> 46;
-
-  return in;
-}
-
-inline const uint8_t* unpack19_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 524287ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint64_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 8;
-  uint64_t w9 = util::SafeLoadAs<uint32_t>(in);
-  w9 = arrow::bit_util::FromLittleEndian(w9);
-  in += 4;
-  out[0] = (w0)&mask;
-  out[1] = (w0 >> 19) & mask;
-  out[2] = (w0 >> 38) & mask;
-  out[3] = ((w0 >> 57) | (w1 << 7)) & mask;
-  out[4] = (w1 >> 12) & mask;
-  out[5] = (w1 >> 31) & mask;
-  out[6] = ((w1 >> 50) | (w2 << 14)) & mask;
-  out[7] = (w2 >> 5) & mask;
-  out[8] = (w2 >> 24) & mask;
-  out[9] = (w2 >> 43) & mask;
-  out[10] = ((w2 >> 62) | (w3 << 2)) & mask;
-  out[11] = (w3 >> 17) & mask;
-  out[12] = (w3 >> 36) & mask;
-  out[13] = ((w3 >> 55) | (w4 << 9)) & mask;
-  out[14] = (w4 >> 10) & mask;
-  out[15] = (w4 >> 29) & mask;
-  out[16] = ((w4 >> 48) | (w5 << 16)) & mask;
-  out[17] = (w5 >> 3) & mask;
-  out[18] = (w5 >> 22) & mask;
-  out[19] = (w5 >> 41) & mask;
-  out[20] = ((w5 >> 60) | (w6 << 4)) & mask;
-  out[21] = (w6 >> 15) & mask;
-  out[22] = (w6 >> 34) & mask;
-  out[23] = ((w6 >> 53) | (w7 << 11)) & mask;
-  out[24] = (w7 >> 8) & mask;
-  out[25] = (w7 >> 27) & mask;
-  out[26] = ((w7 >> 46) | (w8 << 18)) & mask;
-  out[27] = (w8 >> 1) & mask;
-  out[28] = (w8 >> 20) & mask;
-  out[29] = (w8 >> 39) & mask;
-  out[30] = ((w8 >> 58) | (w9 << 6)) & mask;
-  out[31] = (w9 >> 13) & mask;
-
-  return in;
-}
-
-inline const uint8_t* unpack20_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 1048575ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint64_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 8;
-  uint64_t w9 = util::SafeLoadAs<uint64_t>(in);
-  w9 = arrow::bit_util::FromLittleEndian(w9);
-  in += 8;
-  out[0] = (w0)&mask;
-  out[1] = (w0 >> 20) & mask;
-  out[2] = (w0 >> 40) & mask;
-  out[3] = ((w0 >> 60) | (w1 << 4)) & mask;
-  out[4] = (w1 >> 16) & mask;
-  out[5] = (w1 >> 36) & mask;
-  out[6] = ((w1 >> 56) | (w2 << 8)) & mask;
-  out[7] = (w2 >> 12) & mask;
-  out[8] = (w2 >> 32) & mask;
-  out[9] = ((w2 >> 52) | (w3 << 12)) & mask;
-  out[10] = (w3 >> 8) & mask;
-  out[11] = (w3 >> 28) & mask;
-  out[12] = ((w3 >> 48) | (w4 << 16)) & mask;
-  out[13] = (w4 >> 4) & mask;
-  out[14] = (w4 >> 24) & mask;
-  out[15] = w4 >> 44;
-  out[16] = (w5)&mask;
-  out[17] = (w5 >> 20) & mask;
-  out[18] = (w5 >> 40) & mask;
-  out[19] = ((w5 >> 60) | (w6 << 4)) & mask;
-  out[20] = (w6 >> 16) & mask;
-  out[21] = (w6 >> 36) & mask;
-  out[22] = ((w6 >> 56) | (w7 << 8)) & mask;
-  out[23] = (w7 >> 12) & mask;
-  out[24] = (w7 >> 32) & mask;
-  out[25] = ((w7 >> 52) | (w8 << 12)) & mask;
-  out[26] = (w8 >> 8) & mask;
-  out[27] = (w8 >> 28) & mask;
-  out[28] = ((w8 >> 48) | (w9 << 16)) & mask;
-  out[29] = (w9 >> 4) & mask;
-  out[30] = (w9 >> 24) & mask;
-  out[31] = w9 >> 44;
-
-  return in;
-}
-
-inline const uint8_t* unpack21_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 2097151ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint64_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 8;
-  uint64_t w9 = util::SafeLoadAs<uint64_t>(in);
-  w9 = arrow::bit_util::FromLittleEndian(w9);
-  in += 8;
-  uint64_t w10 = util::SafeLoadAs<uint32_t>(in);
-  w10 = arrow::bit_util::FromLittleEndian(w10);
-  in += 4;
-  out[0] = (w0)&mask;
-  out[1] = (w0 >> 21) & mask;
-  out[2] = (w0 >> 42) & mask;
-  out[3] = ((w0 >> 63) | (w1 << 1)) & mask;
-  out[4] = (w1 >> 20) & mask;
-  out[5] = (w1 >> 41) & mask;
-  out[6] = ((w1 >> 62) | (w2 << 2)) & mask;
-  out[7] = (w2 >> 19) & mask;
-  out[8] = (w2 >> 40) & mask;
-  out[9] = ((w2 >> 61) | (w3 << 3)) & mask;
-  out[10] = (w3 >> 18) & mask;
-  out[11] = (w3 >> 39) & mask;
-  out[12] = ((w3 >> 60) | (w4 << 4)) & mask;
-  out[13] = (w4 >> 17) & mask;
-  out[14] = (w4 >> 38) & mask;
-  out[15] = ((w4 >> 59) | (w5 << 5)) & mask;
-  out[16] = (w5 >> 16) & mask;
-  out[17] = (w5 >> 37) & mask;
-  out[18] = ((w5 >> 58) | (w6 << 6)) & mask;
-  out[19] = (w6 >> 15) & mask;
-  out[20] = (w6 >> 36) & mask;
-  out[21] = ((w6 >> 57) | (w7 << 7)) & mask;
-  out[22] = (w7 >> 14) & mask;
-  out[23] = (w7 >> 35) & mask;
-  out[24] = ((w7 >> 56) | (w8 << 8)) & mask;
-  out[25] = (w8 >> 13) & mask;
-  out[26] = (w8 >> 34) & mask;
-  out[27] = ((w8 >> 55) | (w9 << 9)) & mask;
-  out[28] = (w9 >> 12) & mask;
-  out[29] = (w9 >> 33) & mask;
-  out[30] = ((w9 >> 54) | (w10 << 10)) & mask;
-  out[31] = (w10 >> 11) & mask;
-
-  return in;
-}
-
-inline const uint8_t* unpack22_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 4194303ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint64_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 8;
-  uint64_t w9 = util::SafeLoadAs<uint64_t>(in);
-  w9 = arrow::bit_util::FromLittleEndian(w9);
-  in += 8;
-  uint64_t w10 = util::SafeLoadAs<uint64_t>(in);
-  w10 = arrow::bit_util::FromLittleEndian(w10);
-  in += 8;
-  out[0] = (w0)&mask;
-  out[1] = (w0 >> 22) & mask;
-  out[2] = ((w0 >> 44) | (w1 << 20)) & mask;
-  out[3] = (w1 >> 2) & mask;
-  out[4] = (w1 >> 24) & mask;
-  out[5] = ((w1 >> 46) | (w2 << 18)) & mask;
-  out[6] = (w2 >> 4) & mask;
-  out[7] = (w2 >> 26) & mask;
-  out[8] = ((w2 >> 48) | (w3 << 16)) & mask;
-  out[9] = (w3 >> 6) & mask;
-  out[10] = (w3 >> 28) & mask;
-  out[11] = ((w3 >> 50) | (w4 << 14)) & mask;
-  out[12] = (w4 >> 8) & mask;
-  out[13] = (w4 >> 30) & mask;
-  out[14] = ((w4 >> 52) | (w5 << 12)) & mask;
-  out[15] = (w5 >> 10) & mask;
-  out[16] = (w5 >> 32) & mask;
-  out[17] = ((w5 >> 54) | (w6 << 10)) & mask;
-  out[18] = (w6 >> 12) & mask;
-  out[19] = (w6 >> 34) & mask;
-  out[20] = ((w6 >> 56) | (w7 << 8)) & mask;
-  out[21] = (w7 >> 14) & mask;
-  out[22] = (w7 >> 36) & mask;
-  out[23] = ((w7 >> 58) | (w8 << 6)) & mask;
-  out[24] = (w8 >> 16) & mask;
-  out[25] = (w8 >> 38) & mask;
-  out[26] = ((w8 >> 60) | (w9 << 4)) & mask;
-  out[27] = (w9 >> 18) & mask;
-  out[28] = (w9 >> 40) & mask;
-  out[29] = ((w9 >> 62) | (w10 << 2)) & mask;
-  out[30] = (w10 >> 20) & mask;
-  out[31] = w10 >> 42;
-
-  return in;
-}
-
-inline const uint8_t* unpack23_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 8388607ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint64_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 8;
-  uint64_t w9 = util::SafeLoadAs<uint64_t>(in);
-  w9 = arrow::bit_util::FromLittleEndian(w9);
-  in += 8;
-  uint64_t w10 = util::SafeLoadAs<uint64_t>(in);
-  w10 = arrow::bit_util::FromLittleEndian(w10);
-  in += 8;
-  uint64_t w11 = util::SafeLoadAs<uint32_t>(in);
-  w11 = arrow::bit_util::FromLittleEndian(w11);
-  in += 4;
-  out[0] = (w0)&mask;
-  out[1] = (w0 >> 23) & mask;
-  out[2] = ((w0 >> 46) | (w1 << 18)) & mask;
-  out[3] = (w1 >> 5) & mask;
-  out[4] = (w1 >> 28) & mask;
-  out[5] = ((w1 >> 51) | (w2 << 13)) & mask;
-  out[6] = (w2 >> 10) & mask;
-  out[7] = (w2 >> 33) & mask;
-  out[8] = ((w2 >> 56) | (w3 << 8)) & mask;
-  out[9] = (w3 >> 15) & mask;
-  out[10] = (w3 >> 38) & mask;
-  out[11] = ((w3 >> 61) | (w4 << 3)) & mask;
-  out[12] = (w4 >> 20) & mask;
-  out[13] = ((w4 >> 43) | (w5 << 21)) & mask;
-  out[14] = (w5 >> 2) & mask;
-  out[15] = (w5 >> 25) & mask;
-  out[16] = ((w5 >> 48) | (w6 << 16)) & mask;
-  out[17] = (w6 >> 7) & mask;
-  out[18] = (w6 >> 30) & mask;
-  out[19] = ((w6 >> 53) | (w7 << 11)) & mask;
-  out[20] = (w7 >> 12) & mask;
-  out[21] = (w7 >> 35) & mask;
-  out[22] = ((w7 >> 58) | (w8 << 6)) & mask;
-  out[23] = (w8 >> 17) & mask;
-  out[24] = (w8 >> 40) & mask;
-  out[25] = ((w8 >> 63) | (w9 << 1)) & mask;
-  out[26] = (w9 >> 22) & mask;
-  out[27] = ((w9 >> 45) | (w10 << 19)) & mask;
-  out[28] = (w10 >> 4) & mask;
-  out[29] = (w10 >> 27) & mask;
-  out[30] = ((w10 >> 50) | (w11 << 14)) & mask;
-  out[31] = (w11 >> 9) & mask;
-
-  return in;
-}
-
-inline const uint8_t* unpack24_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 16777215ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint64_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 8;
-  uint64_t w9 = util::SafeLoadAs<uint64_t>(in);
-  w9 = arrow::bit_util::FromLittleEndian(w9);
-  in += 8;
-  uint64_t w10 = util::SafeLoadAs<uint64_t>(in);
-  w10 = arrow::bit_util::FromLittleEndian(w10);
-  in += 8;
-  uint64_t w11 = util::SafeLoadAs<uint64_t>(in);
-  w11 = arrow::bit_util::FromLittleEndian(w11);
-  in += 8;
-  out[0] = (w0)&mask;
-  out[1] = (w0 >> 24) & mask;
-  out[2] = ((w0 >> 48) | (w1 << 16)) & mask;
-  out[3] = (w1 >> 8) & mask;
-  out[4] = (w1 >> 32) & mask;
-  out[5] = ((w1 >> 56) | (w2 << 8)) & mask;
-  out[6] = (w2 >> 16) & mask;
-  out[7] = w2 >> 40;
-  out[8] = (w3)&mask;
-  out[9] = (w3 >> 24) & mask;
-  out[10] = ((w3 >> 48) | (w4 << 16)) & mask;
-  out[11] = (w4 >> 8) & mask;
-  out[12] = (w4 >> 32) & mask;
-  out[13] = ((w4 >> 56) | (w5 << 8)) & mask;
-  out[14] = (w5 >> 16) & mask;
-  out[15] = w5 >> 40;
-  out[16] = (w6)&mask;
-  out[17] = (w6 >> 24) & mask;
-  out[18] = ((w6 >> 48) | (w7 << 16)) & mask;
-  out[19] = (w7 >> 8) & mask;
-  out[20] = (w7 >> 32) & mask;
-  out[21] = ((w7 >> 56) | (w8 << 8)) & mask;
-  out[22] = (w8 >> 16) & mask;
-  out[23] = w8 >> 40;
-  out[24] = (w9)&mask;
-  out[25] = (w9 >> 24) & mask;
-  out[26] = ((w9 >> 48) | (w10 << 16)) & mask;
-  out[27] = (w10 >> 8) & mask;
-  out[28] = (w10 >> 32) & mask;
-  out[29] = ((w10 >> 56) | (w11 << 8)) & mask;
-  out[30] = (w11 >> 16) & mask;
-  out[31] = w11 >> 40;
-
-  return in;
-}
-
-inline const uint8_t* unpack25_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 33554431ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint64_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 8;
-  uint64_t w9 = util::SafeLoadAs<uint64_t>(in);
-  w9 = arrow::bit_util::FromLittleEndian(w9);
-  in += 8;
-  uint64_t w10 = util::SafeLoadAs<uint64_t>(in);
-  w10 = arrow::bit_util::FromLittleEndian(w10);
-  in += 8;
-  uint64_t w11 = util::SafeLoadAs<uint64_t>(in);
-  w11 = arrow::bit_util::FromLittleEndian(w11);
-  in += 8;
-  uint64_t w12 = util::SafeLoadAs<uint32_t>(in);
-  w12 = arrow::bit_util::FromLittleEndian(w12);
-  in += 4;
-  out[0] = (w0)&mask;
-  out[1] = (w0 >> 25) & mask;
-  out[2] = ((w0 >> 50) | (w1 << 14)) & mask;
-  out[3] = (w1 >> 11) & mask;
-  out[4] = (w1 >> 36) & mask;
-  out[5] = ((w1 >> 61) | (w2 << 3)) & mask;
-  out[6] = (w2 >> 22) & mask;
-  out[7] = ((w2 >> 47) | (w3 << 17)) & mask;
-  out[8] = (w3 >> 8) & mask;
-  out[9] = (w3 >> 33) & mask;
-  out[10] = ((w3 >> 58) | (w4 << 6)) & mask;
-  out[11] = (w4 >> 19) & mask;
-  out[12] = ((w4 >> 44) | (w5 << 20)) & mask;
-  out[13] = (w5 >> 5) & mask;
-  out[14] = (w5 >> 30) & mask;
-  out[15] = ((w5 >> 55) | (w6 << 9)) & mask;
-  out[16] = (w6 >> 16) & mask;
-  out[17] = ((w6 >> 41) | (w7 << 23)) & mask;
-  out[18] = (w7 >> 2) & mask;
-  out[19] = (w7 >> 27) & mask;
-  out[20] = ((w7 >> 52) | (w8 << 12)) & mask;
-  out[21] = (w8 >> 13) & mask;
-  out[22] = (w8 >> 38) & mask;
-  out[23] = ((w8 >> 63) | (w9 << 1)) & mask;
-  out[24] = (w9 >> 24) & mask;
-  out[25] = ((w9 >> 49) | (w10 << 15)) & mask;
-  out[26] = (w10 >> 10) & mask;
-  out[27] = (w10 >> 35) & mask;
-  out[28] = ((w10 >> 60) | (w11 << 4)) & mask;
-  out[29] = (w11 >> 21) & mask;
-  out[30] = ((w11 >> 46) | (w12 << 18)) & mask;
-  out[31] = (w12 >> 7) & mask;
-
-  return in;
-}
-
-inline const uint8_t* unpack26_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 67108863ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint64_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 8;
-  uint64_t w9 = util::SafeLoadAs<uint64_t>(in);
-  w9 = arrow::bit_util::FromLittleEndian(w9);
-  in += 8;
-  uint64_t w10 = util::SafeLoadAs<uint64_t>(in);
-  w10 = arrow::bit_util::FromLittleEndian(w10);
-  in += 8;
-  uint64_t w11 = util::SafeLoadAs<uint64_t>(in);
-  w11 = arrow::bit_util::FromLittleEndian(w11);
-  in += 8;
-  uint64_t w12 = util::SafeLoadAs<uint64_t>(in);
-  w12 = arrow::bit_util::FromLittleEndian(w12);
-  in += 8;
-  out[0] = (w0)&mask;
-  out[1] = (w0 >> 26) & mask;
-  out[2] = ((w0 >> 52) | (w1 << 12)) & mask;
-  out[3] = (w1 >> 14) & mask;
-  out[4] = ((w1 >> 40) | (w2 << 24)) & mask;
-  out[5] = (w2 >> 2) & mask;
-  out[6] = (w2 >> 28) & mask;
-  out[7] = ((w2 >> 54) | (w3 << 10)) & mask;
-  out[8] = (w3 >> 16) & mask;
-  out[9] = ((w3 >> 42) | (w4 << 22)) & mask;
-  out[10] = (w4 >> 4) & mask;
-  out[11] = (w4 >> 30) & mask;
-  out[12] = ((w4 >> 56) | (w5 << 8)) & mask;
-  out[13] = (w5 >> 18) & mask;
-  out[14] = ((w5 >> 44) | (w6 << 20)) & mask;
-  out[15] = (w6 >> 6) & mask;
-  out[16] = (w6 >> 32) & mask;
-  out[17] = ((w6 >> 58) | (w7 << 6)) & mask;
-  out[18] = (w7 >> 20) & mask;
-  out[19] = ((w7 >> 46) | (w8 << 18)) & mask;
-  out[20] = (w8 >> 8) & mask;
-  out[21] = (w8 >> 34) & mask;
-  out[22] = ((w8 >> 60) | (w9 << 4)) & mask;
-  out[23] = (w9 >> 22) & mask;
-  out[24] = ((w9 >> 48) | (w10 << 16)) & mask;
-  out[25] = (w10 >> 10) & mask;
-  out[26] = (w10 >> 36) & mask;
-  out[27] = ((w10 >> 62) | (w11 << 2)) & mask;
-  out[28] = (w11 >> 24) & mask;
-  out[29] = ((w11 >> 50) | (w12 << 14)) & mask;
-  out[30] = (w12 >> 12) & mask;
-  out[31] = w12 >> 38;
-
-  return in;
-}
-
-inline const uint8_t* unpack27_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 134217727ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint64_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 8;
-  uint64_t w9 = util::SafeLoadAs<uint64_t>(in);
-  w9 = arrow::bit_util::FromLittleEndian(w9);
-  in += 8;
-  uint64_t w10 = util::SafeLoadAs<uint64_t>(in);
-  w10 = arrow::bit_util::FromLittleEndian(w10);
-  in += 8;
-  uint64_t w11 = util::SafeLoadAs<uint64_t>(in);
-  w11 = arrow::bit_util::FromLittleEndian(w11);
-  in += 8;
-  uint64_t w12 = util::SafeLoadAs<uint64_t>(in);
-  w12 = arrow::bit_util::FromLittleEndian(w12);
-  in += 8;
-  uint64_t w13 = util::SafeLoadAs<uint32_t>(in);
-  w13 = arrow::bit_util::FromLittleEndian(w13);
-  in += 4;
-  out[0] = (w0)&mask;
-  out[1] = (w0 >> 27) & mask;
-  out[2] = ((w0 >> 54) | (w1 << 10)) & mask;
-  out[3] = (w1 >> 17) & mask;
-  out[4] = ((w1 >> 44) | (w2 << 20)) & mask;
-  out[5] = (w2 >> 7) & mask;
-  out[6] = (w2 >> 34) & mask;
-  out[7] = ((w2 >> 61) | (w3 << 3)) & mask;
-  out[8] = (w3 >> 24) & mask;
-  out[9] = ((w3 >> 51) | (w4 << 13)) & mask;
-  out[10] = (w4 >> 14) & mask;
-  out[11] = ((w4 >> 41) | (w5 << 23)) & mask;
-  out[12] = (w5 >> 4) & mask;
-  out[13] = (w5 >> 31) & mask;
-  out[14] = ((w5 >> 58) | (w6 << 6)) & mask;
-  out[15] = (w6 >> 21) & mask;
-  out[16] = ((w6 >> 48) | (w7 << 16)) & mask;
-  out[17] = (w7 >> 11) & mask;
-  out[18] = ((w7 >> 38) | (w8 << 26)) & mask;
-  out[19] = (w8 >> 1) & mask;
-  out[20] = (w8 >> 28) & mask;
-  out[21] = ((w8 >> 55) | (w9 << 9)) & mask;
-  out[22] = (w9 >> 18) & mask;
-  out[23] = ((w9 >> 45) | (w10 << 19)) & mask;
-  out[24] = (w10 >> 8) & mask;
-  out[25] = (w10 >> 35) & mask;
-  out[26] = ((w10 >> 62) | (w11 << 2)) & mask;
-  out[27] = (w11 >> 25) & mask;
-  out[28] = ((w11 >> 52) | (w12 << 12)) & mask;
-  out[29] = (w12 >> 15) & mask;
-  out[30] = ((w12 >> 42) | (w13 << 22)) & mask;
-  out[31] = (w13 >> 5) & mask;
-
-  return in;
-}
-
-inline const uint8_t* unpack28_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 268435455ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint64_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 8;
-  uint64_t w9 = util::SafeLoadAs<uint64_t>(in);
-  w9 = arrow::bit_util::FromLittleEndian(w9);
-  in += 8;
-  uint64_t w10 = util::SafeLoadAs<uint64_t>(in);
-  w10 = arrow::bit_util::FromLittleEndian(w10);
-  in += 8;
-  uint64_t w11 = util::SafeLoadAs<uint64_t>(in);
-  w11 = arrow::bit_util::FromLittleEndian(w11);
-  in += 8;
-  uint64_t w12 = util::SafeLoadAs<uint64_t>(in);
-  w12 = arrow::bit_util::FromLittleEndian(w12);
-  in += 8;
-  uint64_t w13 = util::SafeLoadAs<uint64_t>(in);
-  w13 = arrow::bit_util::FromLittleEndian(w13);
-  in += 8;
-  out[0] = (w0)&mask;
-  out[1] = (w0 >> 28) & mask;
-  out[2] = ((w0 >> 56) | (w1 << 8)) & mask;
-  out[3] = (w1 >> 20) & mask;
-  out[4] = ((w1 >> 48) | (w2 << 16)) & mask;
-  out[5] = (w2 >> 12) & mask;
-  out[6] = ((w2 >> 40) | (w3 << 24)) & mask;
-  out[7] = (w3 >> 4) & mask;
-  out[8] = (w3 >> 32) & mask;
-  out[9] = ((w3 >> 60) | (w4 << 4)) & mask;
-  out[10] = (w4 >> 24) & mask;
-  out[11] = ((w4 >> 52) | (w5 << 12)) & mask;
-  out[12] = (w5 >> 16) & mask;
-  out[13] = ((w5 >> 44) | (w6 << 20)) & mask;
-  out[14] = (w6 >> 8) & mask;
-  out[15] = w6 >> 36;
-  out[16] = (w7)&mask;
-  out[17] = (w7 >> 28) & mask;
-  out[18] = ((w7 >> 56) | (w8 << 8)) & mask;
-  out[19] = (w8 >> 20) & mask;
-  out[20] = ((w8 >> 48) | (w9 << 16)) & mask;
-  out[21] = (w9 >> 12) & mask;
-  out[22] = ((w9 >> 40) | (w10 << 24)) & mask;
-  out[23] = (w10 >> 4) & mask;
-  out[24] = (w10 >> 32) & mask;
-  out[25] = ((w10 >> 60) | (w11 << 4)) & mask;
-  out[26] = (w11 >> 24) & mask;
-  out[27] = ((w11 >> 52) | (w12 << 12)) & mask;
-  out[28] = (w12 >> 16) & mask;
-  out[29] = ((w12 >> 44) | (w13 << 20)) & mask;
-  out[30] = (w13 >> 8) & mask;
-  out[31] = w13 >> 36;
-
-  return in;
-}
-
-inline const uint8_t* unpack29_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 536870911ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint64_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 8;
-  uint64_t w9 = util::SafeLoadAs<uint64_t>(in);
-  w9 = arrow::bit_util::FromLittleEndian(w9);
-  in += 8;
-  uint64_t w10 = util::SafeLoadAs<uint64_t>(in);
-  w10 = arrow::bit_util::FromLittleEndian(w10);
-  in += 8;
-  uint64_t w11 = util::SafeLoadAs<uint64_t>(in);
-  w11 = arrow::bit_util::FromLittleEndian(w11);
-  in += 8;
-  uint64_t w12 = util::SafeLoadAs<uint64_t>(in);
-  w12 = arrow::bit_util::FromLittleEndian(w12);
-  in += 8;
-  uint64_t w13 = util::SafeLoadAs<uint64_t>(in);
-  w13 = arrow::bit_util::FromLittleEndian(w13);
-  in += 8;
-  uint64_t w14 = util::SafeLoadAs<uint32_t>(in);
-  w14 = arrow::bit_util::FromLittleEndian(w14);
-  in += 4;
-  out[0] = (w0)&mask;
-  out[1] = (w0 >> 29) & mask;
-  out[2] = ((w0 >> 58) | (w1 << 6)) & mask;
-  out[3] = (w1 >> 23) & mask;
-  out[4] = ((w1 >> 52) | (w2 << 12)) & mask;
-  out[5] = (w2 >> 17) & mask;
-  out[6] = ((w2 >> 46) | (w3 << 18)) & mask;
-  out[7] = (w3 >> 11) & mask;
-  out[8] = ((w3 >> 40) | (w4 << 24)) & mask;
-  out[9] = (w4 >> 5) & mask;
-  out[10] = (w4 >> 34) & mask;
-  out[11] = ((w4 >> 63) | (w5 << 1)) & mask;
-  out[12] = (w5 >> 28) & mask;
-  out[13] = ((w5 >> 57) | (w6 << 7)) & mask;
-  out[14] = (w6 >> 22) & mask;
-  out[15] = ((w6 >> 51) | (w7 << 13)) & mask;
-  out[16] = (w7 >> 16) & mask;
-  out[17] = ((w7 >> 45) | (w8 << 19)) & mask;
-  out[18] = (w8 >> 10) & mask;
-  out[19] = ((w8 >> 39) | (w9 << 25)) & mask;
-  out[20] = (w9 >> 4) & mask;
-  out[21] = (w9 >> 33) & mask;
-  out[22] = ((w9 >> 62) | (w10 << 2)) & mask;
-  out[23] = (w10 >> 27) & mask;
-  out[24] = ((w10 >> 56) | (w11 << 8)) & mask;
-  out[25] = (w11 >> 21) & mask;
-  out[26] = ((w11 >> 50) | (w12 << 14)) & mask;
-  out[27] = (w12 >> 15) & mask;
-  out[28] = ((w12 >> 44) | (w13 << 20)) & mask;
-  out[29] = (w13 >> 9) & mask;
-  out[30] = ((w13 >> 38) | (w14 << 26)) & mask;
-  out[31] = (w14 >> 3) & mask;
-
-  return in;
-}
-
-inline const uint8_t* unpack30_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 1073741823ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint64_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 8;
-  uint64_t w9 = util::SafeLoadAs<uint64_t>(in);
-  w9 = arrow::bit_util::FromLittleEndian(w9);
-  in += 8;
-  uint64_t w10 = util::SafeLoadAs<uint64_t>(in);
-  w10 = arrow::bit_util::FromLittleEndian(w10);
-  in += 8;
-  uint64_t w11 = util::SafeLoadAs<uint64_t>(in);
-  w11 = arrow::bit_util::FromLittleEndian(w11);
-  in += 8;
-  uint64_t w12 = util::SafeLoadAs<uint64_t>(in);
-  w12 = arrow::bit_util::FromLittleEndian(w12);
-  in += 8;
-  uint64_t w13 = util::SafeLoadAs<uint64_t>(in);
-  w13 = arrow::bit_util::FromLittleEndian(w13);
-  in += 8;
-  uint64_t w14 = util::SafeLoadAs<uint64_t>(in);
-  w14 = arrow::bit_util::FromLittleEndian(w14);
-  in += 8;
-  out[0] = (w0)&mask;
-  out[1] = (w0 >> 30) & mask;
-  out[2] = ((w0 >> 60) | (w1 << 4)) & mask;
-  out[3] = (w1 >> 26) & mask;
-  out[4] = ((w1 >> 56) | (w2 << 8)) & mask;
-  out[5] = (w2 >> 22) & mask;
-  out[6] = ((w2 >> 52) | (w3 << 12)) & mask;
-  out[7] = (w3 >> 18) & mask;
-  out[8] = ((w3 >> 48) | (w4 << 16)) & mask;
-  out[9] = (w4 >> 14) & mask;
-  out[10] = ((w4 >> 44) | (w5 << 20)) & mask;
-  out[11] = (w5 >> 10) & mask;
-  out[12] = ((w5 >> 40) | (w6 << 24)) & mask;
-  out[13] = (w6 >> 6) & mask;
-  out[14] = ((w6 >> 36) | (w7 << 28)) & mask;
-  out[15] = (w7 >> 2) & mask;
-  out[16] = (w7 >> 32) & mask;
-  out[17] = ((w7 >> 62) | (w8 << 2)) & mask;
-  out[18] = (w8 >> 28) & mask;
-  out[19] = ((w8 >> 58) | (w9 << 6)) & mask;
-  out[20] = (w9 >> 24) & mask;
-  out[21] = ((w9 >> 54) | (w10 << 10)) & mask;
-  out[22] = (w10 >> 20) & mask;
-  out[23] = ((w10 >> 50) | (w11 << 14)) & mask;
-  out[24] = (w11 >> 16) & mask;
-  out[25] = ((w11 >> 46) | (w12 << 18)) & mask;
-  out[26] = (w12 >> 12) & mask;
-  out[27] = ((w12 >> 42) | (w13 << 22)) & mask;
-  out[28] = (w13 >> 8) & mask;
-  out[29] = ((w13 >> 38) | (w14 << 26)) & mask;
-  out[30] = (w14 >> 4) & mask;
-  out[31] = w14 >> 34;
-
-  return in;
-}
-
-inline const uint8_t* unpack31_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 2147483647ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint64_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 8;
-  uint64_t w9 = util::SafeLoadAs<uint64_t>(in);
-  w9 = arrow::bit_util::FromLittleEndian(w9);
-  in += 8;
-  uint64_t w10 = util::SafeLoadAs<uint64_t>(in);
-  w10 = arrow::bit_util::FromLittleEndian(w10);
-  in += 8;
-  uint64_t w11 = util::SafeLoadAs<uint64_t>(in);
-  w11 = arrow::bit_util::FromLittleEndian(w11);
-  in += 8;
-  uint64_t w12 = util::SafeLoadAs<uint64_t>(in);
-  w12 = arrow::bit_util::FromLittleEndian(w12);
-  in += 8;
-  uint64_t w13 = util::SafeLoadAs<uint64_t>(in);
-  w13 = arrow::bit_util::FromLittleEndian(w13);
-  in += 8;
-  uint64_t w14 = util::SafeLoadAs<uint64_t>(in);
-  w14 = arrow::bit_util::FromLittleEndian(w14);
-  in += 8;
-  uint64_t w15 = util::SafeLoadAs<uint32_t>(in);
-  w15 = arrow::bit_util::FromLittleEndian(w15);
-  in += 4;
-  out[0] = (w0)&mask;
-  out[1] = (w0 >> 31) & mask;
-  out[2] = ((w0 >> 62) | (w1 << 2)) & mask;
-  out[3] = (w1 >> 29) & mask;
-  out[4] = ((w1 >> 60) | (w2 << 4)) & mask;
-  out[5] = (w2 >> 27) & mask;
-  out[6] = ((w2 >> 58) | (w3 << 6)) & mask;
-  out[7] = (w3 >> 25) & mask;
-  out[8] = ((w3 >> 56) | (w4 << 8)) & mask;
-  out[9] = (w4 >> 23) & mask;
-  out[10] = ((w4 >> 54) | (w5 << 10)) & mask;
-  out[11] = (w5 >> 21) & mask;
-  out[12] = ((w5 >> 52) | (w6 << 12)) & mask;
-  out[13] = (w6 >> 19) & mask;
-  out[14] = ((w6 >> 50) | (w7 << 14)) & mask;
-  out[15] = (w7 >> 17) & mask;
-  out[16] = ((w7 >> 48) | (w8 << 16)) & mask;
-  out[17] = (w8 >> 15) & mask;
-  out[18] = ((w8 >> 46) | (w9 << 18)) & mask;
-  out[19] = (w9 >> 13) & mask;
-  out[20] = ((w9 >> 44) | (w10 << 20)) & mask;
-  out[21] = (w10 >> 11) & mask;
-  out[22] = ((w10 >> 42) | (w11 << 22)) & mask;
-  out[23] = (w11 >> 9) & mask;
-  out[24] = ((w11 >> 40) | (w12 << 24)) & mask;
-  out[25] = (w12 >> 7) & mask;
-  out[26] = ((w12 >> 38) | (w13 << 26)) & mask;
-  out[27] = (w13 >> 5) & mask;
-  out[28] = ((w13 >> 36) | (w14 << 28)) & mask;
-  out[29] = (w14 >> 3) & mask;
-  out[30] = ((w14 >> 34) | (w15 << 30)) & mask;
-  out[31] = (w15 >> 1) & mask;
-
-  return in;
-}
-
-inline const uint8_t* unpack32_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 4294967295ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint64_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 8;
-  uint64_t w9 = util::SafeLoadAs<uint64_t>(in);
-  w9 = arrow::bit_util::FromLittleEndian(w9);
-  in += 8;
-  uint64_t w10 = util::SafeLoadAs<uint64_t>(in);
-  w10 = arrow::bit_util::FromLittleEndian(w10);
-  in += 8;
-  uint64_t w11 = util::SafeLoadAs<uint64_t>(in);
-  w11 = arrow::bit_util::FromLittleEndian(w11);
-  in += 8;
-  uint64_t w12 = util::SafeLoadAs<uint64_t>(in);
-  w12 = arrow::bit_util::FromLittleEndian(w12);
-  in += 8;
-  uint64_t w13 = util::SafeLoadAs<uint64_t>(in);
-  w13 = arrow::bit_util::FromLittleEndian(w13);
-  in += 8;
-  uint64_t w14 = util::SafeLoadAs<uint64_t>(in);
-  w14 = arrow::bit_util::FromLittleEndian(w14);
-  in += 8;
-  uint64_t w15 = util::SafeLoadAs<uint64_t>(in);
-  w15 = arrow::bit_util::FromLittleEndian(w15);
-  in += 8;
-  out[0] = (w0)&mask;
-  out[1] = w0 >> 32;
-  out[2] = (w1)&mask;
-  out[3] = w1 >> 32;
-  out[4] = (w2)&mask;
-  out[5] = w2 >> 32;
-  out[6] = (w3)&mask;
-  out[7] = w3 >> 32;
-  out[8] = (w4)&mask;
-  out[9] = w4 >> 32;
-  out[10] = (w5)&mask;
-  out[11] = w5 >> 32;
-  out[12] = (w6)&mask;
-  out[13] = w6 >> 32;
-  out[14] = (w7)&mask;
-  out[15] = w7 >> 32;
-  out[16] = (w8)&mask;
-  out[17] = w8 >> 32;
-  out[18] = (w9)&mask;
-  out[19] = w9 >> 32;
-  out[20] = (w10)&mask;
-  out[21] = w10 >> 32;
-  out[22] = (w11)&mask;
-  out[23] = w11 >> 32;
-  out[24] = (w12)&mask;
-  out[25] = w12 >> 32;
-  out[26] = (w13)&mask;
-  out[27] = w13 >> 32;
-  out[28] = (w14)&mask;
-  out[29] = w14 >> 32;
-  out[30] = (w15)&mask;
-  out[31] = w15 >> 32;
-
-  return in;
-}
-
-inline const uint8_t* unpack33_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 8589934591ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint64_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 8;
-  uint64_t w9 = util::SafeLoadAs<uint64_t>(in);
-  w9 = arrow::bit_util::FromLittleEndian(w9);
-  in += 8;
-  uint64_t w10 = util::SafeLoadAs<uint64_t>(in);
-  w10 = arrow::bit_util::FromLittleEndian(w10);
-  in += 8;
-  uint64_t w11 = util::SafeLoadAs<uint64_t>(in);
-  w11 = arrow::bit_util::FromLittleEndian(w11);
-  in += 8;
-  uint64_t w12 = util::SafeLoadAs<uint64_t>(in);
-  w12 = arrow::bit_util::FromLittleEndian(w12);
-  in += 8;
-  uint64_t w13 = util::SafeLoadAs<uint64_t>(in);
-  w13 = arrow::bit_util::FromLittleEndian(w13);
-  in += 8;
-  uint64_t w14 = util::SafeLoadAs<uint64_t>(in);
-  w14 = arrow::bit_util::FromLittleEndian(w14);
-  in += 8;
-  uint64_t w15 = util::SafeLoadAs<uint64_t>(in);
-  w15 = arrow::bit_util::FromLittleEndian(w15);
-  in += 8;
-  uint64_t w16 = util::SafeLoadAs<uint32_t>(in);
-  w16 = arrow::bit_util::FromLittleEndian(w16);
-  in += 4;
-  out[0] = (w0)&mask;
-  out[1] = ((w0 >> 33) | (w1 << 31)) & mask;
-  out[2] = (w1 >> 2) & mask;
-  out[3] = ((w1 >> 35) | (w2 << 29)) & mask;
-  out[4] = (w2 >> 4) & mask;
-  out[5] = ((w2 >> 37) | (w3 << 27)) & mask;
-  out[6] = (w3 >> 6) & mask;
-  out[7] = ((w3 >> 39) | (w4 << 25)) & mask;
-  out[8] = (w4 >> 8) & mask;
-  out[9] = ((w4 >> 41) | (w5 << 23)) & mask;
-  out[10] = (w5 >> 10) & mask;
-  out[11] = ((w5 >> 43) | (w6 << 21)) & mask;
-  out[12] = (w6 >> 12) & mask;
-  out[13] = ((w6 >> 45) | (w7 << 19)) & mask;
-  out[14] = (w7 >> 14) & mask;
-  out[15] = ((w7 >> 47) | (w8 << 17)) & mask;
-  out[16] = (w8 >> 16) & mask;
-  out[17] = ((w8 >> 49) | (w9 << 15)) & mask;
-  out[18] = (w9 >> 18) & mask;
-  out[19] = ((w9 >> 51) | (w10 << 13)) & mask;
-  out[20] = (w10 >> 20) & mask;
-  out[21] = ((w10 >> 53) | (w11 << 11)) & mask;
-  out[22] = (w11 >> 22) & mask;
-  out[23] = ((w11 >> 55) | (w12 << 9)) & mask;
-  out[24] = (w12 >> 24) & mask;
-  out[25] = ((w12 >> 57) | (w13 << 7)) & mask;
-  out[26] = (w13 >> 26) & mask;
-  out[27] = ((w13 >> 59) | (w14 << 5)) & mask;
-  out[28] = (w14 >> 28) & mask;
-  out[29] = ((w14 >> 61) | (w15 << 3)) & mask;
-  out[30] = (w15 >> 30) & mask;
-  out[31] = ((w15 >> 63) | (w16 << 1)) & mask;
-
-  return in;
-}
-
-inline const uint8_t* unpack34_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 17179869183ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint64_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 8;
-  uint64_t w9 = util::SafeLoadAs<uint64_t>(in);
-  w9 = arrow::bit_util::FromLittleEndian(w9);
-  in += 8;
-  uint64_t w10 = util::SafeLoadAs<uint64_t>(in);
-  w10 = arrow::bit_util::FromLittleEndian(w10);
-  in += 8;
-  uint64_t w11 = util::SafeLoadAs<uint64_t>(in);
-  w11 = arrow::bit_util::FromLittleEndian(w11);
-  in += 8;
-  uint64_t w12 = util::SafeLoadAs<uint64_t>(in);
-  w12 = arrow::bit_util::FromLittleEndian(w12);
-  in += 8;
-  uint64_t w13 = util::SafeLoadAs<uint64_t>(in);
-  w13 = arrow::bit_util::FromLittleEndian(w13);
-  in += 8;
-  uint64_t w14 = util::SafeLoadAs<uint64_t>(in);
-  w14 = arrow::bit_util::FromLittleEndian(w14);
-  in += 8;
-  uint64_t w15 = util::SafeLoadAs<uint64_t>(in);
-  w15 = arrow::bit_util::FromLittleEndian(w15);
-  in += 8;
-  uint64_t w16 = util::SafeLoadAs<uint64_t>(in);
-  w16 = arrow::bit_util::FromLittleEndian(w16);
-  in += 8;
-  out[0] = (w0)&mask;
-  out[1] = ((w0 >> 34) | (w1 << 30)) & mask;
-  out[2] = (w1 >> 4) & mask;
-  out[3] = ((w1 >> 38) | (w2 << 26)) & mask;
-  out[4] = (w2 >> 8) & mask;
-  out[5] = ((w2 >> 42) | (w3 << 22)) & mask;
-  out[6] = (w3 >> 12) & mask;
-  out[7] = ((w3 >> 46) | (w4 << 18)) & mask;
-  out[8] = (w4 >> 16) & mask;
-  out[9] = ((w4 >> 50) | (w5 << 14)) & mask;
-  out[10] = (w5 >> 20) & mask;
-  out[11] = ((w5 >> 54) | (w6 << 10)) & mask;
-  out[12] = (w6 >> 24) & mask;
-  out[13] = ((w6 >> 58) | (w7 << 6)) & mask;
-  out[14] = (w7 >> 28) & mask;
-  out[15] = ((w7 >> 62) | (w8 << 2)) & mask;
-  out[16] = ((w8 >> 32) | (w9 << 32)) & mask;
-  out[17] = (w9 >> 2) & mask;
-  out[18] = ((w9 >> 36) | (w10 << 28)) & mask;
-  out[19] = (w10 >> 6) & mask;
-  out[20] = ((w10 >> 40) | (w11 << 24)) & mask;
-  out[21] = (w11 >> 10) & mask;
-  out[22] = ((w11 >> 44) | (w12 << 20)) & mask;
-  out[23] = (w12 >> 14) & mask;
-  out[24] = ((w12 >> 48) | (w13 << 16)) & mask;
-  out[25] = (w13 >> 18) & mask;
-  out[26] = ((w13 >> 52) | (w14 << 12)) & mask;
-  out[27] = (w14 >> 22) & mask;
-  out[28] = ((w14 >> 56) | (w15 << 8)) & mask;
-  out[29] = (w15 >> 26) & mask;
-  out[30] = ((w15 >> 60) | (w16 << 4)) & mask;
-  out[31] = w16 >> 30;
-
-  return in;
-}
-
-inline const uint8_t* unpack35_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 34359738367ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint64_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 8;
-  uint64_t w9 = util::SafeLoadAs<uint64_t>(in);
-  w9 = arrow::bit_util::FromLittleEndian(w9);
-  in += 8;
-  uint64_t w10 = util::SafeLoadAs<uint64_t>(in);
-  w10 = arrow::bit_util::FromLittleEndian(w10);
-  in += 8;
-  uint64_t w11 = util::SafeLoadAs<uint64_t>(in);
-  w11 = arrow::bit_util::FromLittleEndian(w11);
-  in += 8;
-  uint64_t w12 = util::SafeLoadAs<uint64_t>(in);
-  w12 = arrow::bit_util::FromLittleEndian(w12);
-  in += 8;
-  uint64_t w13 = util::SafeLoadAs<uint64_t>(in);
-  w13 = arrow::bit_util::FromLittleEndian(w13);
-  in += 8;
-  uint64_t w14 = util::SafeLoadAs<uint64_t>(in);
-  w14 = arrow::bit_util::FromLittleEndian(w14);
-  in += 8;
-  uint64_t w15 = util::SafeLoadAs<uint64_t>(in);
-  w15 = arrow::bit_util::FromLittleEndian(w15);
-  in += 8;
-  uint64_t w16 = util::SafeLoadAs<uint64_t>(in);
-  w16 = arrow::bit_util::FromLittleEndian(w16);
-  in += 8;
-  uint64_t w17 = util::SafeLoadAs<uint32_t>(in);
-  w17 = arrow::bit_util::FromLittleEndian(w17);
-  in += 4;
-  out[0] = (w0)&mask;
-  out[1] = ((w0 >> 35) | (w1 << 29)) & mask;
-  out[2] = (w1 >> 6) & mask;
-  out[3] = ((w1 >> 41) | (w2 << 23)) & mask;
-  out[4] = (w2 >> 12) & mask;
-  out[5] = ((w2 >> 47) | (w3 << 17)) & mask;
-  out[6] = (w3 >> 18) & mask;
-  out[7] = ((w3 >> 53) | (w4 << 11)) & mask;
-  out[8] = (w4 >> 24) & mask;
-  out[9] = ((w4 >> 59) | (w5 << 5)) & mask;
-  out[10] = ((w5 >> 30) | (w6 << 34)) & mask;
-  out[11] = (w6 >> 1) & mask;
-  out[12] = ((w6 >> 36) | (w7 << 28)) & mask;
-  out[13] = (w7 >> 7) & mask;
-  out[14] = ((w7 >> 42) | (w8 << 22)) & mask;
-  out[15] = (w8 >> 13) & mask;
-  out[16] = ((w8 >> 48) | (w9 << 16)) & mask;
-  out[17] = (w9 >> 19) & mask;
-  out[18] = ((w9 >> 54) | (w10 << 10)) & mask;
-  out[19] = (w10 >> 25) & mask;
-  out[20] = ((w10 >> 60) | (w11 << 4)) & mask;
-  out[21] = ((w11 >> 31) | (w12 << 33)) & mask;
-  out[22] = (w12 >> 2) & mask;
-  out[23] = ((w12 >> 37) | (w13 << 27)) & mask;
-  out[24] = (w13 >> 8) & mask;
-  out[25] = ((w13 >> 43) | (w14 << 21)) & mask;
-  out[26] = (w14 >> 14) & mask;
-  out[27] = ((w14 >> 49) | (w15 << 15)) & mask;
-  out[28] = (w15 >> 20) & mask;
-  out[29] = ((w15 >> 55) | (w16 << 9)) & mask;
-  out[30] = (w16 >> 26) & mask;
-  out[31] = ((w16 >> 61) | (w17 << 3)) & mask;
-
-  return in;
-}
-
-inline const uint8_t* unpack36_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 68719476735ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint64_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 8;
-  uint64_t w9 = util::SafeLoadAs<uint64_t>(in);
-  w9 = arrow::bit_util::FromLittleEndian(w9);
-  in += 8;
-  uint64_t w10 = util::SafeLoadAs<uint64_t>(in);
-  w10 = arrow::bit_util::FromLittleEndian(w10);
-  in += 8;
-  uint64_t w11 = util::SafeLoadAs<uint64_t>(in);
-  w11 = arrow::bit_util::FromLittleEndian(w11);
-  in += 8;
-  uint64_t w12 = util::SafeLoadAs<uint64_t>(in);
-  w12 = arrow::bit_util::FromLittleEndian(w12);
-  in += 8;
-  uint64_t w13 = util::SafeLoadAs<uint64_t>(in);
-  w13 = arrow::bit_util::FromLittleEndian(w13);
-  in += 8;
-  uint64_t w14 = util::SafeLoadAs<uint64_t>(in);
-  w14 = arrow::bit_util::FromLittleEndian(w14);
-  in += 8;
-  uint64_t w15 = util::SafeLoadAs<uint64_t>(in);
-  w15 = arrow::bit_util::FromLittleEndian(w15);
-  in += 8;
-  uint64_t w16 = util::SafeLoadAs<uint64_t>(in);
-  w16 = arrow::bit_util::FromLittleEndian(w16);
-  in += 8;
-  uint64_t w17 = util::SafeLoadAs<uint64_t>(in);
-  w17 = arrow::bit_util::FromLittleEndian(w17);
-  in += 8;
-  out[0] = (w0)&mask;
-  out[1] = ((w0 >> 36) | (w1 << 28)) & mask;
-  out[2] = (w1 >> 8) & mask;
-  out[3] = ((w1 >> 44) | (w2 << 20)) & mask;
-  out[4] = (w2 >> 16) & mask;
-  out[5] = ((w2 >> 52) | (w3 << 12)) & mask;
-  out[6] = (w3 >> 24) & mask;
-  out[7] = ((w3 >> 60) | (w4 << 4)) & mask;
-  out[8] = ((w4 >> 32) | (w5 << 32)) & mask;
-  out[9] = (w5 >> 4) & mask;
-  out[10] = ((w5 >> 40) | (w6 << 24)) & mask;
-  out[11] = (w6 >> 12) & mask;
-  out[12] = ((w6 >> 48) | (w7 << 16)) & mask;
-  out[13] = (w7 >> 20) & mask;
-  out[14] = ((w7 >> 56) | (w8 << 8)) & mask;
-  out[15] = w8 >> 28;
-  out[16] = (w9)&mask;
-  out[17] = ((w9 >> 36) | (w10 << 28)) & mask;
-  out[18] = (w10 >> 8) & mask;
-  out[19] = ((w10 >> 44) | (w11 << 20)) & mask;
-  out[20] = (w11 >> 16) & mask;
-  out[21] = ((w11 >> 52) | (w12 << 12)) & mask;
-  out[22] = (w12 >> 24) & mask;
-  out[23] = ((w12 >> 60) | (w13 << 4)) & mask;
-  out[24] = ((w13 >> 32) | (w14 << 32)) & mask;
-  out[25] = (w14 >> 4) & mask;
-  out[26] = ((w14 >> 40) | (w15 << 24)) & mask;
-  out[27] = (w15 >> 12) & mask;
-  out[28] = ((w15 >> 48) | (w16 << 16)) & mask;
-  out[29] = (w16 >> 20) & mask;
-  out[30] = ((w16 >> 56) | (w17 << 8)) & mask;
-  out[31] = w17 >> 28;
-
-  return in;
-}
-
-inline const uint8_t* unpack37_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 137438953471ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint64_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 8;
-  uint64_t w9 = util::SafeLoadAs<uint64_t>(in);
-  w9 = arrow::bit_util::FromLittleEndian(w9);
-  in += 8;
-  uint64_t w10 = util::SafeLoadAs<uint64_t>(in);
-  w10 = arrow::bit_util::FromLittleEndian(w10);
-  in += 8;
-  uint64_t w11 = util::SafeLoadAs<uint64_t>(in);
-  w11 = arrow::bit_util::FromLittleEndian(w11);
-  in += 8;
-  uint64_t w12 = util::SafeLoadAs<uint64_t>(in);
-  w12 = arrow::bit_util::FromLittleEndian(w12);
-  in += 8;
-  uint64_t w13 = util::SafeLoadAs<uint64_t>(in);
-  w13 = arrow::bit_util::FromLittleEndian(w13);
-  in += 8;
-  uint64_t w14 = util::SafeLoadAs<uint64_t>(in);
-  w14 = arrow::bit_util::FromLittleEndian(w14);
-  in += 8;
-  uint64_t w15 = util::SafeLoadAs<uint64_t>(in);
-  w15 = arrow::bit_util::FromLittleEndian(w15);
-  in += 8;
-  uint64_t w16 = util::SafeLoadAs<uint64_t>(in);
-  w16 = arrow::bit_util::FromLittleEndian(w16);
-  in += 8;
-  uint64_t w17 = util::SafeLoadAs<uint64_t>(in);
-  w17 = arrow::bit_util::FromLittleEndian(w17);
-  in += 8;
-  uint64_t w18 = util::SafeLoadAs<uint32_t>(in);
-  w18 = arrow::bit_util::FromLittleEndian(w18);
-  in += 4;
-  out[0] = (w0)&mask;
-  out[1] = ((w0 >> 37) | (w1 << 27)) & mask;
-  out[2] = (w1 >> 10) & mask;
-  out[3] = ((w1 >> 47) | (w2 << 17)) & mask;
-  out[4] = (w2 >> 20) & mask;
-  out[5] = ((w2 >> 57) | (w3 << 7)) & mask;
-  out[6] = ((w3 >> 30) | (w4 << 34)) & mask;
-  out[7] = (w4 >> 3) & mask;
-  out[8] = ((w4 >> 40) | (w5 << 24)) & mask;
-  out[9] = (w5 >> 13) & mask;
-  out[10] = ((w5 >> 50) | (w6 << 14)) & mask;
-  out[11] = (w6 >> 23) & mask;
-  out[12] = ((w6 >> 60) | (w7 << 4)) & mask;
-  out[13] = ((w7 >> 33) | (w8 << 31)) & mask;
-  out[14] = (w8 >> 6) & mask;
-  out[15] = ((w8 >> 43) | (w9 << 21)) & mask;
-  out[16] = (w9 >> 16) & mask;
-  out[17] = ((w9 >> 53) | (w10 << 11)) & mask;
-  out[18] = (w10 >> 26) & mask;
-  out[19] = ((w10 >> 63) | (w11 << 1)) & mask;
-  out[20] = ((w11 >> 36) | (w12 << 28)) & mask;
-  out[21] = (w12 >> 9) & mask;
-  out[22] = ((w12 >> 46) | (w13 << 18)) & mask;
-  out[23] = (w13 >> 19) & mask;
-  out[24] = ((w13 >> 56) | (w14 << 8)) & mask;
-  out[25] = ((w14 >> 29) | (w15 << 35)) & mask;
-  out[26] = (w15 >> 2) & mask;
-  out[27] = ((w15 >> 39) | (w16 << 25)) & mask;
-  out[28] = (w16 >> 12) & mask;
-  out[29] = ((w16 >> 49) | (w17 << 15)) & mask;
-  out[30] = (w17 >> 22) & mask;
-  out[31] = ((w17 >> 59) | (w18 << 5)) & mask;
-
-  return in;
-}
-
-inline const uint8_t* unpack38_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 274877906943ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint64_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 8;
-  uint64_t w9 = util::SafeLoadAs<uint64_t>(in);
-  w9 = arrow::bit_util::FromLittleEndian(w9);
-  in += 8;
-  uint64_t w10 = util::SafeLoadAs<uint64_t>(in);
-  w10 = arrow::bit_util::FromLittleEndian(w10);
-  in += 8;
-  uint64_t w11 = util::SafeLoadAs<uint64_t>(in);
-  w11 = arrow::bit_util::FromLittleEndian(w11);
-  in += 8;
-  uint64_t w12 = util::SafeLoadAs<uint64_t>(in);
-  w12 = arrow::bit_util::FromLittleEndian(w12);
-  in += 8;
-  uint64_t w13 = util::SafeLoadAs<uint64_t>(in);
-  w13 = arrow::bit_util::FromLittleEndian(w13);
-  in += 8;
-  uint64_t w14 = util::SafeLoadAs<uint64_t>(in);
-  w14 = arrow::bit_util::FromLittleEndian(w14);
-  in += 8;
-  uint64_t w15 = util::SafeLoadAs<uint64_t>(in);
-  w15 = arrow::bit_util::FromLittleEndian(w15);
-  in += 8;
-  uint64_t w16 = util::SafeLoadAs<uint64_t>(in);
-  w16 = arrow::bit_util::FromLittleEndian(w16);
-  in += 8;
-  uint64_t w17 = util::SafeLoadAs<uint64_t>(in);
-  w17 = arrow::bit_util::FromLittleEndian(w17);
-  in += 8;
-  uint64_t w18 = util::SafeLoadAs<uint64_t>(in);
-  w18 = arrow::bit_util::FromLittleEndian(w18);
-  in += 8;
-  out[0] = (w0)&mask;
-  out[1] = ((w0 >> 38) | (w1 << 26)) & mask;
-  out[2] = (w1 >> 12) & mask;
-  out[3] = ((w1 >> 50) | (w2 << 14)) & mask;
-  out[4] = (w2 >> 24) & mask;
-  out[5] = ((w2 >> 62) | (w3 << 2)) & mask;
-  out[6] = ((w3 >> 36) | (w4 << 28)) & mask;
-  out[7] = (w4 >> 10) & mask;
-  out[8] = ((w4 >> 48) | (w5 << 16)) & mask;
-  out[9] = (w5 >> 22) & mask;
-  out[10] = ((w5 >> 60) | (w6 << 4)) & mask;
-  out[11] = ((w6 >> 34) | (w7 << 30)) & mask;
-  out[12] = (w7 >> 8) & mask;
-  out[13] = ((w7 >> 46) | (w8 << 18)) & mask;
-  out[14] = (w8 >> 20) & mask;
-  out[15] = ((w8 >> 58) | (w9 << 6)) & mask;
-  out[16] = ((w9 >> 32) | (w10 << 32)) & mask;
-  out[17] = (w10 >> 6) & mask;
-  out[18] = ((w10 >> 44) | (w11 << 20)) & mask;
-  out[19] = (w11 >> 18) & mask;
-  out[20] = ((w11 >> 56) | (w12 << 8)) & mask;
-  out[21] = ((w12 >> 30) | (w13 << 34)) & mask;
-  out[22] = (w13 >> 4) & mask;
-  out[23] = ((w13 >> 42) | (w14 << 22)) & mask;
-  out[24] = (w14 >> 16) & mask;
-  out[25] = ((w14 >> 54) | (w15 << 10)) & mask;
-  out[26] = ((w15 >> 28) | (w16 << 36)) & mask;
-  out[27] = (w16 >> 2) & mask;
-  out[28] = ((w16 >> 40) | (w17 << 24)) & mask;
-  out[29] = (w17 >> 14) & mask;
-  out[30] = ((w17 >> 52) | (w18 << 12)) & mask;
-  out[31] = w18 >> 26;
-
-  return in;
-}
-
-inline const uint8_t* unpack39_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 549755813887ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint64_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 8;
-  uint64_t w9 = util::SafeLoadAs<uint64_t>(in);
-  w9 = arrow::bit_util::FromLittleEndian(w9);
-  in += 8;
-  uint64_t w10 = util::SafeLoadAs<uint64_t>(in);
-  w10 = arrow::bit_util::FromLittleEndian(w10);
-  in += 8;
-  uint64_t w11 = util::SafeLoadAs<uint64_t>(in);
-  w11 = arrow::bit_util::FromLittleEndian(w11);
-  in += 8;
-  uint64_t w12 = util::SafeLoadAs<uint64_t>(in);
-  w12 = arrow::bit_util::FromLittleEndian(w12);
-  in += 8;
-  uint64_t w13 = util::SafeLoadAs<uint64_t>(in);
-  w13 = arrow::bit_util::FromLittleEndian(w13);
-  in += 8;
-  uint64_t w14 = util::SafeLoadAs<uint64_t>(in);
-  w14 = arrow::bit_util::FromLittleEndian(w14);
-  in += 8;
-  uint64_t w15 = util::SafeLoadAs<uint64_t>(in);
-  w15 = arrow::bit_util::FromLittleEndian(w15);
-  in += 8;
-  uint64_t w16 = util::SafeLoadAs<uint64_t>(in);
-  w16 = arrow::bit_util::FromLittleEndian(w16);
-  in += 8;
-  uint64_t w17 = util::SafeLoadAs<uint64_t>(in);
-  w17 = arrow::bit_util::FromLittleEndian(w17);
-  in += 8;
-  uint64_t w18 = util::SafeLoadAs<uint64_t>(in);
-  w18 = arrow::bit_util::FromLittleEndian(w18);
-  in += 8;
-  uint64_t w19 = util::SafeLoadAs<uint32_t>(in);
-  w19 = arrow::bit_util::FromLittleEndian(w19);
-  in += 4;
-  out[0] = (w0)&mask;
-  out[1] = ((w0 >> 39) | (w1 << 25)) & mask;
-  out[2] = (w1 >> 14) & mask;
-  out[3] = ((w1 >> 53) | (w2 << 11)) & mask;
-  out[4] = ((w2 >> 28) | (w3 << 36)) & mask;
-  out[5] = (w3 >> 3) & mask;
-  out[6] = ((w3 >> 42) | (w4 << 22)) & mask;
-  out[7] = (w4 >> 17) & mask;
-  out[8] = ((w4 >> 56) | (w5 << 8)) & mask;
-  out[9] = ((w5 >> 31) | (w6 << 33)) & mask;
-  out[10] = (w6 >> 6) & mask;
-  out[11] = ((w6 >> 45) | (w7 << 19)) & mask;
-  out[12] = (w7 >> 20) & mask;
-  out[13] = ((w7 >> 59) | (w8 << 5)) & mask;
-  out[14] = ((w8 >> 34) | (w9 << 30)) & mask;
-  out[15] = (w9 >> 9) & mask;
-  out[16] = ((w9 >> 48) | (w10 << 16)) & mask;
-  out[17] = (w10 >> 23) & mask;
-  out[18] = ((w10 >> 62) | (w11 << 2)) & mask;
-  out[19] = ((w11 >> 37) | (w12 << 27)) & mask;
-  out[20] = (w12 >> 12) & mask;
-  out[21] = ((w12 >> 51) | (w13 << 13)) & mask;
-  out[22] = ((w13 >> 26) | (w14 << 38)) & mask;
-  out[23] = (w14 >> 1) & mask;
-  out[24] = ((w14 >> 40) | (w15 << 24)) & mask;
-  out[25] = (w15 >> 15) & mask;
-  out[26] = ((w15 >> 54) | (w16 << 10)) & mask;
-  out[27] = ((w16 >> 29) | (w17 << 35)) & mask;
-  out[28] = (w17 >> 4) & mask;
-  out[29] = ((w17 >> 43) | (w18 << 21)) & mask;
-  out[30] = (w18 >> 18) & mask;
-  out[31] = ((w18 >> 57) | (w19 << 7)) & mask;
-
-  return in;
-}
-
-inline const uint8_t* unpack40_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 1099511627775ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint64_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 8;
-  uint64_t w9 = util::SafeLoadAs<uint64_t>(in);
-  w9 = arrow::bit_util::FromLittleEndian(w9);
-  in += 8;
-  uint64_t w10 = util::SafeLoadAs<uint64_t>(in);
-  w10 = arrow::bit_util::FromLittleEndian(w10);
-  in += 8;
-  uint64_t w11 = util::SafeLoadAs<uint64_t>(in);
-  w11 = arrow::bit_util::FromLittleEndian(w11);
-  in += 8;
-  uint64_t w12 = util::SafeLoadAs<uint64_t>(in);
-  w12 = arrow::bit_util::FromLittleEndian(w12);
-  in += 8;
-  uint64_t w13 = util::SafeLoadAs<uint64_t>(in);
-  w13 = arrow::bit_util::FromLittleEndian(w13);
-  in += 8;
-  uint64_t w14 = util::SafeLoadAs<uint64_t>(in);
-  w14 = arrow::bit_util::FromLittleEndian(w14);
-  in += 8;
-  uint64_t w15 = util::SafeLoadAs<uint64_t>(in);
-  w15 = arrow::bit_util::FromLittleEndian(w15);
-  in += 8;
-  uint64_t w16 = util::SafeLoadAs<uint64_t>(in);
-  w16 = arrow::bit_util::FromLittleEndian(w16);
-  in += 8;
-  uint64_t w17 = util::SafeLoadAs<uint64_t>(in);
-  w17 = arrow::bit_util::FromLittleEndian(w17);
-  in += 8;
-  uint64_t w18 = util::SafeLoadAs<uint64_t>(in);
-  w18 = arrow::bit_util::FromLittleEndian(w18);
-  in += 8;
-  uint64_t w19 = util::SafeLoadAs<uint64_t>(in);
-  w19 = arrow::bit_util::FromLittleEndian(w19);
-  in += 8;
-  out[0] = (w0)&mask;
-  out[1] = ((w0 >> 40) | (w1 << 24)) & mask;
-  out[2] = (w1 >> 16) & mask;
-  out[3] = ((w1 >> 56) | (w2 << 8)) & mask;
-  out[4] = ((w2 >> 32) | (w3 << 32)) & mask;
-  out[5] = (w3 >> 8) & mask;
-  out[6] = ((w3 >> 48) | (w4 << 16)) & mask;
-  out[7] = w4 >> 24;
-  out[8] = (w5)&mask;
-  out[9] = ((w5 >> 40) | (w6 << 24)) & mask;
-  out[10] = (w6 >> 16) & mask;
-  out[11] = ((w6 >> 56) | (w7 << 8)) & mask;
-  out[12] = ((w7 >> 32) | (w8 << 32)) & mask;
-  out[13] = (w8 >> 8) & mask;
-  out[14] = ((w8 >> 48) | (w9 << 16)) & mask;
-  out[15] = w9 >> 24;
-  out[16] = (w10)&mask;
-  out[17] = ((w10 >> 40) | (w11 << 24)) & mask;
-  out[18] = (w11 >> 16) & mask;
-  out[19] = ((w11 >> 56) | (w12 << 8)) & mask;
-  out[20] = ((w12 >> 32) | (w13 << 32)) & mask;
-  out[21] = (w13 >> 8) & mask;
-  out[22] = ((w13 >> 48) | (w14 << 16)) & mask;
-  out[23] = w14 >> 24;
-  out[24] = (w15)&mask;
-  out[25] = ((w15 >> 40) | (w16 << 24)) & mask;
-  out[26] = (w16 >> 16) & mask;
-  out[27] = ((w16 >> 56) | (w17 << 8)) & mask;
-  out[28] = ((w17 >> 32) | (w18 << 32)) & mask;
-  out[29] = (w18 >> 8) & mask;
-  out[30] = ((w18 >> 48) | (w19 << 16)) & mask;
-  out[31] = w19 >> 24;
-
-  return in;
-}
-
-inline const uint8_t* unpack41_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 2199023255551ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint64_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 8;
-  uint64_t w9 = util::SafeLoadAs<uint64_t>(in);
-  w9 = arrow::bit_util::FromLittleEndian(w9);
-  in += 8;
-  uint64_t w10 = util::SafeLoadAs<uint64_t>(in);
-  w10 = arrow::bit_util::FromLittleEndian(w10);
-  in += 8;
-  uint64_t w11 = util::SafeLoadAs<uint64_t>(in);
-  w11 = arrow::bit_util::FromLittleEndian(w11);
-  in += 8;
-  uint64_t w12 = util::SafeLoadAs<uint64_t>(in);
-  w12 = arrow::bit_util::FromLittleEndian(w12);
-  in += 8;
-  uint64_t w13 = util::SafeLoadAs<uint64_t>(in);
-  w13 = arrow::bit_util::FromLittleEndian(w13);
-  in += 8;
-  uint64_t w14 = util::SafeLoadAs<uint64_t>(in);
-  w14 = arrow::bit_util::FromLittleEndian(w14);
-  in += 8;
-  uint64_t w15 = util::SafeLoadAs<uint64_t>(in);
-  w15 = arrow::bit_util::FromLittleEndian(w15);
-  in += 8;
-  uint64_t w16 = util::SafeLoadAs<uint64_t>(in);
-  w16 = arrow::bit_util::FromLittleEndian(w16);
-  in += 8;
-  uint64_t w17 = util::SafeLoadAs<uint64_t>(in);
-  w17 = arrow::bit_util::FromLittleEndian(w17);
-  in += 8;
-  uint64_t w18 = util::SafeLoadAs<uint64_t>(in);
-  w18 = arrow::bit_util::FromLittleEndian(w18);
-  in += 8;
-  uint64_t w19 = util::SafeLoadAs<uint64_t>(in);
-  w19 = arrow::bit_util::FromLittleEndian(w19);
-  in += 8;
-  uint64_t w20 = util::SafeLoadAs<uint32_t>(in);
-  w20 = arrow::bit_util::FromLittleEndian(w20);
-  in += 4;
-  out[0] = (w0)&mask;
-  out[1] = ((w0 >> 41) | (w1 << 23)) & mask;
-  out[2] = (w1 >> 18) & mask;
-  out[3] = ((w1 >> 59) | (w2 << 5)) & mask;
-  out[4] = ((w2 >> 36) | (w3 << 28)) & mask;
-  out[5] = (w3 >> 13) & mask;
-  out[6] = ((w3 >> 54) | (w4 << 10)) & mask;
-  out[7] = ((w4 >> 31) | (w5 << 33)) & mask;
-  out[8] = (w5 >> 8) & mask;
-  out[9] = ((w5 >> 49) | (w6 << 15)) & mask;
-  out[10] = ((w6 >> 26) | (w7 << 38)) & mask;
-  out[11] = (w7 >> 3) & mask;
-  out[12] = ((w7 >> 44) | (w8 << 20)) & mask;
-  out[13] = (w8 >> 21) & mask;
-  out[14] = ((w8 >> 62) | (w9 << 2)) & mask;
-  out[15] = ((w9 >> 39) | (w10 << 25)) & mask;
-  out[16] = (w10 >> 16) & mask;
-  out[17] = ((w10 >> 57) | (w11 << 7)) & mask;
-  out[18] = ((w11 >> 34) | (w12 << 30)) & mask;
-  out[19] = (w12 >> 11) & mask;
-  out[20] = ((w12 >> 52) | (w13 << 12)) & mask;
-  out[21] = ((w13 >> 29) | (w14 << 35)) & mask;
-  out[22] = (w14 >> 6) & mask;
-  out[23] = ((w14 >> 47) | (w15 << 17)) & mask;
-  out[24] = ((w15 >> 24) | (w16 << 40)) & mask;
-  out[25] = (w16 >> 1) & mask;
-  out[26] = ((w16 >> 42) | (w17 << 22)) & mask;
-  out[27] = (w17 >> 19) & mask;
-  out[28] = ((w17 >> 60) | (w18 << 4)) & mask;
-  out[29] = ((w18 >> 37) | (w19 << 27)) & mask;
-  out[30] = (w19 >> 14) & mask;
-  out[31] = ((w19 >> 55) | (w20 << 9)) & mask;
-
-  return in;
-}
-
-inline const uint8_t* unpack42_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 4398046511103ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint64_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 8;
-  uint64_t w9 = util::SafeLoadAs<uint64_t>(in);
-  w9 = arrow::bit_util::FromLittleEndian(w9);
-  in += 8;
-  uint64_t w10 = util::SafeLoadAs<uint64_t>(in);
-  w10 = arrow::bit_util::FromLittleEndian(w10);
-  in += 8;
-  uint64_t w11 = util::SafeLoadAs<uint64_t>(in);
-  w11 = arrow::bit_util::FromLittleEndian(w11);
-  in += 8;
-  uint64_t w12 = util::SafeLoadAs<uint64_t>(in);
-  w12 = arrow::bit_util::FromLittleEndian(w12);
-  in += 8;
-  uint64_t w13 = util::SafeLoadAs<uint64_t>(in);
-  w13 = arrow::bit_util::FromLittleEndian(w13);
-  in += 8;
-  uint64_t w14 = util::SafeLoadAs<uint64_t>(in);
-  w14 = arrow::bit_util::FromLittleEndian(w14);
-  in += 8;
-  uint64_t w15 = util::SafeLoadAs<uint64_t>(in);
-  w15 = arrow::bit_util::FromLittleEndian(w15);
-  in += 8;
-  uint64_t w16 = util::SafeLoadAs<uint64_t>(in);
-  w16 = arrow::bit_util::FromLittleEndian(w16);
-  in += 8;
-  uint64_t w17 = util::SafeLoadAs<uint64_t>(in);
-  w17 = arrow::bit_util::FromLittleEndian(w17);
-  in += 8;
-  uint64_t w18 = util::SafeLoadAs<uint64_t>(in);
-  w18 = arrow::bit_util::FromLittleEndian(w18);
-  in += 8;
-  uint64_t w19 = util::SafeLoadAs<uint64_t>(in);
-  w19 = arrow::bit_util::FromLittleEndian(w19);
-  in += 8;
-  uint64_t w20 = util::SafeLoadAs<uint64_t>(in);
-  w20 = arrow::bit_util::FromLittleEndian(w20);
-  in += 8;
-  out[0] = (w0)&mask;
-  out[1] = ((w0 >> 42) | (w1 << 22)) & mask;
-  out[2] = (w1 >> 20) & mask;
-  out[3] = ((w1 >> 62) | (w2 << 2)) & mask;
-  out[4] = ((w2 >> 40) | (w3 << 24)) & mask;
-  out[5] = (w3 >> 18) & mask;
-  out[6] = ((w3 >> 60) | (w4 << 4)) & mask;
-  out[7] = ((w4 >> 38) | (w5 << 26)) & mask;
-  out[8] = (w5 >> 16) & mask;
-  out[9] = ((w5 >> 58) | (w6 << 6)) & mask;
-  out[10] = ((w6 >> 36) | (w7 << 28)) & mask;
-  out[11] = (w7 >> 14) & mask;
-  out[12] = ((w7 >> 56) | (w8 << 8)) & mask;
-  out[13] = ((w8 >> 34) | (w9 << 30)) & mask;
-  out[14] = (w9 >> 12) & mask;
-  out[15] = ((w9 >> 54) | (w10 << 10)) & mask;
-  out[16] = ((w10 >> 32) | (w11 << 32)) & mask;
-  out[17] = (w11 >> 10) & mask;
-  out[18] = ((w11 >> 52) | (w12 << 12)) & mask;
-  out[19] = ((w12 >> 30) | (w13 << 34)) & mask;
-  out[20] = (w13 >> 8) & mask;
-  out[21] = ((w13 >> 50) | (w14 << 14)) & mask;
-  out[22] = ((w14 >> 28) | (w15 << 36)) & mask;
-  out[23] = (w15 >> 6) & mask;
-  out[24] = ((w15 >> 48) | (w16 << 16)) & mask;
-  out[25] = ((w16 >> 26) | (w17 << 38)) & mask;
-  out[26] = (w17 >> 4) & mask;
-  out[27] = ((w17 >> 46) | (w18 << 18)) & mask;
-  out[28] = ((w18 >> 24) | (w19 << 40)) & mask;
-  out[29] = (w19 >> 2) & mask;
-  out[30] = ((w19 >> 44) | (w20 << 20)) & mask;
-  out[31] = w20 >> 22;
-
-  return in;
-}
-
-inline const uint8_t* unpack43_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 8796093022207ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint64_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 8;
-  uint64_t w9 = util::SafeLoadAs<uint64_t>(in);
-  w9 = arrow::bit_util::FromLittleEndian(w9);
-  in += 8;
-  uint64_t w10 = util::SafeLoadAs<uint64_t>(in);
-  w10 = arrow::bit_util::FromLittleEndian(w10);
-  in += 8;
-  uint64_t w11 = util::SafeLoadAs<uint64_t>(in);
-  w11 = arrow::bit_util::FromLittleEndian(w11);
-  in += 8;
-  uint64_t w12 = util::SafeLoadAs<uint64_t>(in);
-  w12 = arrow::bit_util::FromLittleEndian(w12);
-  in += 8;
-  uint64_t w13 = util::SafeLoadAs<uint64_t>(in);
-  w13 = arrow::bit_util::FromLittleEndian(w13);
-  in += 8;
-  uint64_t w14 = util::SafeLoadAs<uint64_t>(in);
-  w14 = arrow::bit_util::FromLittleEndian(w14);
-  in += 8;
-  uint64_t w15 = util::SafeLoadAs<uint64_t>(in);
-  w15 = arrow::bit_util::FromLittleEndian(w15);
-  in += 8;
-  uint64_t w16 = util::SafeLoadAs<uint64_t>(in);
-  w16 = arrow::bit_util::FromLittleEndian(w16);
-  in += 8;
-  uint64_t w17 = util::SafeLoadAs<uint64_t>(in);
-  w17 = arrow::bit_util::FromLittleEndian(w17);
-  in += 8;
-  uint64_t w18 = util::SafeLoadAs<uint64_t>(in);
-  w18 = arrow::bit_util::FromLittleEndian(w18);
-  in += 8;
-  uint64_t w19 = util::SafeLoadAs<uint64_t>(in);
-  w19 = arrow::bit_util::FromLittleEndian(w19);
-  in += 8;
-  uint64_t w20 = util::SafeLoadAs<uint64_t>(in);
-  w20 = arrow::bit_util::FromLittleEndian(w20);
-  in += 8;
-  uint64_t w21 = util::SafeLoadAs<uint32_t>(in);
-  w21 = arrow::bit_util::FromLittleEndian(w21);
-  in += 4;
-  out[0] = (w0)&mask;
-  out[1] = ((w0 >> 43) | (w1 << 21)) & mask;
-  out[2] = ((w1 >> 22) | (w2 << 42)) & mask;
-  out[3] = (w2 >> 1) & mask;
-  out[4] = ((w2 >> 44) | (w3 << 20)) & mask;
-  out[5] = ((w3 >> 23) | (w4 << 41)) & mask;
-  out[6] = (w4 >> 2) & mask;
-  out[7] = ((w4 >> 45) | (w5 << 19)) & mask;
-  out[8] = ((w5 >> 24) | (w6 << 40)) & mask;
-  out[9] = (w6 >> 3) & mask;
-  out[10] = ((w6 >> 46) | (w7 << 18)) & mask;
-  out[11] = ((w7 >> 25) | (w8 << 39)) & mask;
-  out[12] = (w8 >> 4) & mask;
-  out[13] = ((w8 >> 47) | (w9 << 17)) & mask;
-  out[14] = ((w9 >> 26) | (w10 << 38)) & mask;
-  out[15] = (w10 >> 5) & mask;
-  out[16] = ((w10 >> 48) | (w11 << 16)) & mask;
-  out[17] = ((w11 >> 27) | (w12 << 37)) & mask;
-  out[18] = (w12 >> 6) & mask;
-  out[19] = ((w12 >> 49) | (w13 << 15)) & mask;
-  out[20] = ((w13 >> 28) | (w14 << 36)) & mask;
-  out[21] = (w14 >> 7) & mask;
-  out[22] = ((w14 >> 50) | (w15 << 14)) & mask;
-  out[23] = ((w15 >> 29) | (w16 << 35)) & mask;
-  out[24] = (w16 >> 8) & mask;
-  out[25] = ((w16 >> 51) | (w17 << 13)) & mask;
-  out[26] = ((w17 >> 30) | (w18 << 34)) & mask;
-  out[27] = (w18 >> 9) & mask;
-  out[28] = ((w18 >> 52) | (w19 << 12)) & mask;
-  out[29] = ((w19 >> 31) | (w20 << 33)) & mask;
-  out[30] = (w20 >> 10) & mask;
-  out[31] = ((w20 >> 53) | (w21 << 11)) & mask;
-
-  return in;
-}
-
-inline const uint8_t* unpack44_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 17592186044415ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint64_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 8;
-  uint64_t w9 = util::SafeLoadAs<uint64_t>(in);
-  w9 = arrow::bit_util::FromLittleEndian(w9);
-  in += 8;
-  uint64_t w10 = util::SafeLoadAs<uint64_t>(in);
-  w10 = arrow::bit_util::FromLittleEndian(w10);
-  in += 8;
-  uint64_t w11 = util::SafeLoadAs<uint64_t>(in);
-  w11 = arrow::bit_util::FromLittleEndian(w11);
-  in += 8;
-  uint64_t w12 = util::SafeLoadAs<uint64_t>(in);
-  w12 = arrow::bit_util::FromLittleEndian(w12);
-  in += 8;
-  uint64_t w13 = util::SafeLoadAs<uint64_t>(in);
-  w13 = arrow::bit_util::FromLittleEndian(w13);
-  in += 8;
-  uint64_t w14 = util::SafeLoadAs<uint64_t>(in);
-  w14 = arrow::bit_util::FromLittleEndian(w14);
-  in += 8;
-  uint64_t w15 = util::SafeLoadAs<uint64_t>(in);
-  w15 = arrow::bit_util::FromLittleEndian(w15);
-  in += 8;
-  uint64_t w16 = util::SafeLoadAs<uint64_t>(in);
-  w16 = arrow::bit_util::FromLittleEndian(w16);
-  in += 8;
-  uint64_t w17 = util::SafeLoadAs<uint64_t>(in);
-  w17 = arrow::bit_util::FromLittleEndian(w17);
-  in += 8;
-  uint64_t w18 = util::SafeLoadAs<uint64_t>(in);
-  w18 = arrow::bit_util::FromLittleEndian(w18);
-  in += 8;
-  uint64_t w19 = util::SafeLoadAs<uint64_t>(in);
-  w19 = arrow::bit_util::FromLittleEndian(w19);
-  in += 8;
-  uint64_t w20 = util::SafeLoadAs<uint64_t>(in);
-  w20 = arrow::bit_util::FromLittleEndian(w20);
-  in += 8;
-  uint64_t w21 = util::SafeLoadAs<uint64_t>(in);
-  w21 = arrow::bit_util::FromLittleEndian(w21);
-  in += 8;
-  out[0] = (w0)&mask;
-  out[1] = ((w0 >> 44) | (w1 << 20)) & mask;
-  out[2] = ((w1 >> 24) | (w2 << 40)) & mask;
-  out[3] = (w2 >> 4) & mask;
-  out[4] = ((w2 >> 48) | (w3 << 16)) & mask;
-  out[5] = ((w3 >> 28) | (w4 << 36)) & mask;
-  out[6] = (w4 >> 8) & mask;
-  out[7] = ((w4 >> 52) | (w5 << 12)) & mask;
-  out[8] = ((w5 >> 32) | (w6 << 32)) & mask;
-  out[9] = (w6 >> 12) & mask;
-  out[10] = ((w6 >> 56) | (w7 << 8)) & mask;
-  out[11] = ((w7 >> 36) | (w8 << 28)) & mask;
-  out[12] = (w8 >> 16) & mask;
-  out[13] = ((w8 >> 60) | (w9 << 4)) & mask;
-  out[14] = ((w9 >> 40) | (w10 << 24)) & mask;
-  out[15] = w10 >> 20;
-  out[16] = (w11)&mask;
-  out[17] = ((w11 >> 44) | (w12 << 20)) & mask;
-  out[18] = ((w12 >> 24) | (w13 << 40)) & mask;
-  out[19] = (w13 >> 4) & mask;
-  out[20] = ((w13 >> 48) | (w14 << 16)) & mask;
-  out[21] = ((w14 >> 28) | (w15 << 36)) & mask;
-  out[22] = (w15 >> 8) & mask;
-  out[23] = ((w15 >> 52) | (w16 << 12)) & mask;
-  out[24] = ((w16 >> 32) | (w17 << 32)) & mask;
-  out[25] = (w17 >> 12) & mask;
-  out[26] = ((w17 >> 56) | (w18 << 8)) & mask;
-  out[27] = ((w18 >> 36) | (w19 << 28)) & mask;
-  out[28] = (w19 >> 16) & mask;
-  out[29] = ((w19 >> 60) | (w20 << 4)) & mask;
-  out[30] = ((w20 >> 40) | (w21 << 24)) & mask;
-  out[31] = w21 >> 20;
-
-  return in;
-}
-
-inline const uint8_t* unpack45_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 35184372088831ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint64_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 8;
-  uint64_t w9 = util::SafeLoadAs<uint64_t>(in);
-  w9 = arrow::bit_util::FromLittleEndian(w9);
-  in += 8;
-  uint64_t w10 = util::SafeLoadAs<uint64_t>(in);
-  w10 = arrow::bit_util::FromLittleEndian(w10);
-  in += 8;
-  uint64_t w11 = util::SafeLoadAs<uint64_t>(in);
-  w11 = arrow::bit_util::FromLittleEndian(w11);
-  in += 8;
-  uint64_t w12 = util::SafeLoadAs<uint64_t>(in);
-  w12 = arrow::bit_util::FromLittleEndian(w12);
-  in += 8;
-  uint64_t w13 = util::SafeLoadAs<uint64_t>(in);
-  w13 = arrow::bit_util::FromLittleEndian(w13);
-  in += 8;
-  uint64_t w14 = util::SafeLoadAs<uint64_t>(in);
-  w14 = arrow::bit_util::FromLittleEndian(w14);
-  in += 8;
-  uint64_t w15 = util::SafeLoadAs<uint64_t>(in);
-  w15 = arrow::bit_util::FromLittleEndian(w15);
-  in += 8;
-  uint64_t w16 = util::SafeLoadAs<uint64_t>(in);
-  w16 = arrow::bit_util::FromLittleEndian(w16);
-  in += 8;
-  uint64_t w17 = util::SafeLoadAs<uint64_t>(in);
-  w17 = arrow::bit_util::FromLittleEndian(w17);
-  in += 8;
-  uint64_t w18 = util::SafeLoadAs<uint64_t>(in);
-  w18 = arrow::bit_util::FromLittleEndian(w18);
-  in += 8;
-  uint64_t w19 = util::SafeLoadAs<uint64_t>(in);
-  w19 = arrow::bit_util::FromLittleEndian(w19);
-  in += 8;
-  uint64_t w20 = util::SafeLoadAs<uint64_t>(in);
-  w20 = arrow::bit_util::FromLittleEndian(w20);
-  in += 8;
-  uint64_t w21 = util::SafeLoadAs<uint64_t>(in);
-  w21 = arrow::bit_util::FromLittleEndian(w21);
-  in += 8;
-  uint64_t w22 = util::SafeLoadAs<uint32_t>(in);
-  w22 = arrow::bit_util::FromLittleEndian(w22);
-  in += 4;
-  out[0] = (w0)&mask;
-  out[1] = ((w0 >> 45) | (w1 << 19)) & mask;
-  out[2] = ((w1 >> 26) | (w2 << 38)) & mask;
-  out[3] = (w2 >> 7) & mask;
-  out[4] = ((w2 >> 52) | (w3 << 12)) & mask;
-  out[5] = ((w3 >> 33) | (w4 << 31)) & mask;
-  out[6] = (w4 >> 14) & mask;
-  out[7] = ((w4 >> 59) | (w5 << 5)) & mask;
-  out[8] = ((w5 >> 40) | (w6 << 24)) & mask;
-  out[9] = ((w6 >> 21) | (w7 << 43)) & mask;
-  out[10] = (w7 >> 2) & mask;
-  out[11] = ((w7 >> 47) | (w8 << 17)) & mask;
-  out[12] = ((w8 >> 28) | (w9 << 36)) & mask;
-  out[13] = (w9 >> 9) & mask;
-  out[14] = ((w9 >> 54) | (w10 << 10)) & mask;
-  out[15] = ((w10 >> 35) | (w11 << 29)) & mask;
-  out[16] = (w11 >> 16) & mask;
-  out[17] = ((w11 >> 61) | (w12 << 3)) & mask;
-  out[18] = ((w12 >> 42) | (w13 << 22)) & mask;
-  out[19] = ((w13 >> 23) | (w14 << 41)) & mask;
-  out[20] = (w14 >> 4) & mask;
-  out[21] = ((w14 >> 49) | (w15 << 15)) & mask;
-  out[22] = ((w15 >> 30) | (w16 << 34)) & mask;
-  out[23] = (w16 >> 11) & mask;
-  out[24] = ((w16 >> 56) | (w17 << 8)) & mask;
-  out[25] = ((w17 >> 37) | (w18 << 27)) & mask;
-  out[26] = (w18 >> 18) & mask;
-  out[27] = ((w18 >> 63) | (w19 << 1)) & mask;
-  out[28] = ((w19 >> 44) | (w20 << 20)) & mask;
-  out[29] = ((w20 >> 25) | (w21 << 39)) & mask;
-  out[30] = (w21 >> 6) & mask;
-  out[31] = ((w21 >> 51) | (w22 << 13)) & mask;
-
-  return in;
-}
-
-inline const uint8_t* unpack46_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 70368744177663ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint64_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 8;
-  uint64_t w9 = util::SafeLoadAs<uint64_t>(in);
-  w9 = arrow::bit_util::FromLittleEndian(w9);
-  in += 8;
-  uint64_t w10 = util::SafeLoadAs<uint64_t>(in);
-  w10 = arrow::bit_util::FromLittleEndian(w10);
-  in += 8;
-  uint64_t w11 = util::SafeLoadAs<uint64_t>(in);
-  w11 = arrow::bit_util::FromLittleEndian(w11);
-  in += 8;
-  uint64_t w12 = util::SafeLoadAs<uint64_t>(in);
-  w12 = arrow::bit_util::FromLittleEndian(w12);
-  in += 8;
-  uint64_t w13 = util::SafeLoadAs<uint64_t>(in);
-  w13 = arrow::bit_util::FromLittleEndian(w13);
-  in += 8;
-  uint64_t w14 = util::SafeLoadAs<uint64_t>(in);
-  w14 = arrow::bit_util::FromLittleEndian(w14);
-  in += 8;
-  uint64_t w15 = util::SafeLoadAs<uint64_t>(in);
-  w15 = arrow::bit_util::FromLittleEndian(w15);
-  in += 8;
-  uint64_t w16 = util::SafeLoadAs<uint64_t>(in);
-  w16 = arrow::bit_util::FromLittleEndian(w16);
-  in += 8;
-  uint64_t w17 = util::SafeLoadAs<uint64_t>(in);
-  w17 = arrow::bit_util::FromLittleEndian(w17);
-  in += 8;
-  uint64_t w18 = util::SafeLoadAs<uint64_t>(in);
-  w18 = arrow::bit_util::FromLittleEndian(w18);
-  in += 8;
-  uint64_t w19 = util::SafeLoadAs<uint64_t>(in);
-  w19 = arrow::bit_util::FromLittleEndian(w19);
-  in += 8;
-  uint64_t w20 = util::SafeLoadAs<uint64_t>(in);
-  w20 = arrow::bit_util::FromLittleEndian(w20);
-  in += 8;
-  uint64_t w21 = util::SafeLoadAs<uint64_t>(in);
-  w21 = arrow::bit_util::FromLittleEndian(w21);
-  in += 8;
-  uint64_t w22 = util::SafeLoadAs<uint64_t>(in);
-  w22 = arrow::bit_util::FromLittleEndian(w22);
-  in += 8;
-  out[0] = (w0)&mask;
-  out[1] = ((w0 >> 46) | (w1 << 18)) & mask;
-  out[2] = ((w1 >> 28) | (w2 << 36)) & mask;
-  out[3] = (w2 >> 10) & mask;
-  out[4] = ((w2 >> 56) | (w3 << 8)) & mask;
-  out[5] = ((w3 >> 38) | (w4 << 26)) & mask;
-  out[6] = ((w4 >> 20) | (w5 << 44)) & mask;
-  out[7] = (w5 >> 2) & mask;
-  out[8] = ((w5 >> 48) | (w6 << 16)) & mask;
-  out[9] = ((w6 >> 30) | (w7 << 34)) & mask;
-  out[10] = (w7 >> 12) & mask;
-  out[11] = ((w7 >> 58) | (w8 << 6)) & mask;
-  out[12] = ((w8 >> 40) | (w9 << 24)) & mask;
-  out[13] = ((w9 >> 22) | (w10 << 42)) & mask;
-  out[14] = (w10 >> 4) & mask;
-  out[15] = ((w10 >> 50) | (w11 << 14)) & mask;
-  out[16] = ((w11 >> 32) | (w12 << 32)) & mask;
-  out[17] = (w12 >> 14) & mask;
-  out[18] = ((w12 >> 60) | (w13 << 4)) & mask;
-  out[19] = ((w13 >> 42) | (w14 << 22)) & mask;
-  out[20] = ((w14 >> 24) | (w15 << 40)) & mask;
-  out[21] = (w15 >> 6) & mask;
-  out[22] = ((w15 >> 52) | (w16 << 12)) & mask;
-  out[23] = ((w16 >> 34) | (w17 << 30)) & mask;
-  out[24] = (w17 >> 16) & mask;
-  out[25] = ((w17 >> 62) | (w18 << 2)) & mask;
-  out[26] = ((w18 >> 44) | (w19 << 20)) & mask;
-  out[27] = ((w19 >> 26) | (w20 << 38)) & mask;
-  out[28] = (w20 >> 8) & mask;
-  out[29] = ((w20 >> 54) | (w21 << 10)) & mask;
-  out[30] = ((w21 >> 36) | (w22 << 28)) & mask;
-  out[31] = w22 >> 18;
-
-  return in;
-}
-
-inline const uint8_t* unpack47_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 140737488355327ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint64_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 8;
-  uint64_t w9 = util::SafeLoadAs<uint64_t>(in);
-  w9 = arrow::bit_util::FromLittleEndian(w9);
-  in += 8;
-  uint64_t w10 = util::SafeLoadAs<uint64_t>(in);
-  w10 = arrow::bit_util::FromLittleEndian(w10);
-  in += 8;
-  uint64_t w11 = util::SafeLoadAs<uint64_t>(in);
-  w11 = arrow::bit_util::FromLittleEndian(w11);
-  in += 8;
-  uint64_t w12 = util::SafeLoadAs<uint64_t>(in);
-  w12 = arrow::bit_util::FromLittleEndian(w12);
-  in += 8;
-  uint64_t w13 = util::SafeLoadAs<uint64_t>(in);
-  w13 = arrow::bit_util::FromLittleEndian(w13);
-  in += 8;
-  uint64_t w14 = util::SafeLoadAs<uint64_t>(in);
-  w14 = arrow::bit_util::FromLittleEndian(w14);
-  in += 8;
-  uint64_t w15 = util::SafeLoadAs<uint64_t>(in);
-  w15 = arrow::bit_util::FromLittleEndian(w15);
-  in += 8;
-  uint64_t w16 = util::SafeLoadAs<uint64_t>(in);
-  w16 = arrow::bit_util::FromLittleEndian(w16);
-  in += 8;
-  uint64_t w17 = util::SafeLoadAs<uint64_t>(in);
-  w17 = arrow::bit_util::FromLittleEndian(w17);
-  in += 8;
-  uint64_t w18 = util::SafeLoadAs<uint64_t>(in);
-  w18 = arrow::bit_util::FromLittleEndian(w18);
-  in += 8;
-  uint64_t w19 = util::SafeLoadAs<uint64_t>(in);
-  w19 = arrow::bit_util::FromLittleEndian(w19);
-  in += 8;
-  uint64_t w20 = util::SafeLoadAs<uint64_t>(in);
-  w20 = arrow::bit_util::FromLittleEndian(w20);
-  in += 8;
-  uint64_t w21 = util::SafeLoadAs<uint64_t>(in);
-  w21 = arrow::bit_util::FromLittleEndian(w21);
-  in += 8;
-  uint64_t w22 = util::SafeLoadAs<uint64_t>(in);
-  w22 = arrow::bit_util::FromLittleEndian(w22);
-  in += 8;
-  uint64_t w23 = util::SafeLoadAs<uint32_t>(in);
-  w23 = arrow::bit_util::FromLittleEndian(w23);
-  in += 4;
-  out[0] = (w0)&mask;
-  out[1] = ((w0 >> 47) | (w1 << 17)) & mask;
-  out[2] = ((w1 >> 30) | (w2 << 34)) & mask;
-  out[3] = (w2 >> 13) & mask;
-  out[4] = ((w2 >> 60) | (w3 << 4)) & mask;
-  out[5] = ((w3 >> 43) | (w4 << 21)) & mask;
-  out[6] = ((w4 >> 26) | (w5 << 38)) & mask;
-  out[7] = (w5 >> 9) & mask;
-  out[8] = ((w5 >> 56) | (w6 << 8)) & mask;
-  out[9] = ((w6 >> 39) | (w7 << 25)) & mask;
-  out[10] = ((w7 >> 22) | (w8 << 42)) & mask;
-  out[11] = (w8 >> 5) & mask;
-  out[12] = ((w8 >> 52) | (w9 << 12)) & mask;
-  out[13] = ((w9 >> 35) | (w10 << 29)) & mask;
-  out[14] = ((w10 >> 18) | (w11 << 46)) & mask;
-  out[15] = (w11 >> 1) & mask;
-  out[16] = ((w11 >> 48) | (w12 << 16)) & mask;
-  out[17] = ((w12 >> 31) | (w13 << 33)) & mask;
-  out[18] = (w13 >> 14) & mask;
-  out[19] = ((w13 >> 61) | (w14 << 3)) & mask;
-  out[20] = ((w14 >> 44) | (w15 << 20)) & mask;
-  out[21] = ((w15 >> 27) | (w16 << 37)) & mask;
-  out[22] = (w16 >> 10) & mask;
-  out[23] = ((w16 >> 57) | (w17 << 7)) & mask;
-  out[24] = ((w17 >> 40) | (w18 << 24)) & mask;
-  out[25] = ((w18 >> 23) | (w19 << 41)) & mask;
-  out[26] = (w19 >> 6) & mask;
-  out[27] = ((w19 >> 53) | (w20 << 11)) & mask;
-  out[28] = ((w20 >> 36) | (w21 << 28)) & mask;
-  out[29] = ((w21 >> 19) | (w22 << 45)) & mask;
-  out[30] = (w22 >> 2) & mask;
-  out[31] = ((w22 >> 49) | (w23 << 15)) & mask;
-
-  return in;
-}
-
-inline const uint8_t* unpack48_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 281474976710655ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint64_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 8;
-  uint64_t w9 = util::SafeLoadAs<uint64_t>(in);
-  w9 = arrow::bit_util::FromLittleEndian(w9);
-  in += 8;
-  uint64_t w10 = util::SafeLoadAs<uint64_t>(in);
-  w10 = arrow::bit_util::FromLittleEndian(w10);
-  in += 8;
-  uint64_t w11 = util::SafeLoadAs<uint64_t>(in);
-  w11 = arrow::bit_util::FromLittleEndian(w11);
-  in += 8;
-  uint64_t w12 = util::SafeLoadAs<uint64_t>(in);
-  w12 = arrow::bit_util::FromLittleEndian(w12);
-  in += 8;
-  uint64_t w13 = util::SafeLoadAs<uint64_t>(in);
-  w13 = arrow::bit_util::FromLittleEndian(w13);
-  in += 8;
-  uint64_t w14 = util::SafeLoadAs<uint64_t>(in);
-  w14 = arrow::bit_util::FromLittleEndian(w14);
-  in += 8;
-  uint64_t w15 = util::SafeLoadAs<uint64_t>(in);
-  w15 = arrow::bit_util::FromLittleEndian(w15);
-  in += 8;
-  uint64_t w16 = util::SafeLoadAs<uint64_t>(in);
-  w16 = arrow::bit_util::FromLittleEndian(w16);
-  in += 8;
-  uint64_t w17 = util::SafeLoadAs<uint64_t>(in);
-  w17 = arrow::bit_util::FromLittleEndian(w17);
-  in += 8;
-  uint64_t w18 = util::SafeLoadAs<uint64_t>(in);
-  w18 = arrow::bit_util::FromLittleEndian(w18);
-  in += 8;
-  uint64_t w19 = util::SafeLoadAs<uint64_t>(in);
-  w19 = arrow::bit_util::FromLittleEndian(w19);
-  in += 8;
-  uint64_t w20 = util::SafeLoadAs<uint64_t>(in);
-  w20 = arrow::bit_util::FromLittleEndian(w20);
-  in += 8;
-  uint64_t w21 = util::SafeLoadAs<uint64_t>(in);
-  w21 = arrow::bit_util::FromLittleEndian(w21);
-  in += 8;
-  uint64_t w22 = util::SafeLoadAs<uint64_t>(in);
-  w22 = arrow::bit_util::FromLittleEndian(w22);
-  in += 8;
-  uint64_t w23 = util::SafeLoadAs<uint64_t>(in);
-  w23 = arrow::bit_util::FromLittleEndian(w23);
-  in += 8;
-  out[0] = (w0)&mask;
-  out[1] = ((w0 >> 48) | (w1 << 16)) & mask;
-  out[2] = ((w1 >> 32) | (w2 << 32)) & mask;
-  out[3] = w2 >> 16;
-  out[4] = (w3)&mask;
-  out[5] = ((w3 >> 48) | (w4 << 16)) & mask;
-  out[6] = ((w4 >> 32) | (w5 << 32)) & mask;
-  out[7] = w5 >> 16;
-  out[8] = (w6)&mask;
-  out[9] = ((w6 >> 48) | (w7 << 16)) & mask;
-  out[10] = ((w7 >> 32) | (w8 << 32)) & mask;
-  out[11] = w8 >> 16;
-  out[12] = (w9)&mask;
-  out[13] = ((w9 >> 48) | (w10 << 16)) & mask;
-  out[14] = ((w10 >> 32) | (w11 << 32)) & mask;
-  out[15] = w11 >> 16;
-  out[16] = (w12)&mask;
-  out[17] = ((w12 >> 48) | (w13 << 16)) & mask;
-  out[18] = ((w13 >> 32) | (w14 << 32)) & mask;
-  out[19] = w14 >> 16;
-  out[20] = (w15)&mask;
-  out[21] = ((w15 >> 48) | (w16 << 16)) & mask;
-  out[22] = ((w16 >> 32) | (w17 << 32)) & mask;
-  out[23] = w17 >> 16;
-  out[24] = (w18)&mask;
-  out[25] = ((w18 >> 48) | (w19 << 16)) & mask;
-  out[26] = ((w19 >> 32) | (w20 << 32)) & mask;
-  out[27] = w20 >> 16;
-  out[28] = (w21)&mask;
-  out[29] = ((w21 >> 48) | (w22 << 16)) & mask;
-  out[30] = ((w22 >> 32) | (w23 << 32)) & mask;
-  out[31] = w23 >> 16;
-
-  return in;
-}
-
-inline const uint8_t* unpack49_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 562949953421311ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint64_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 8;
-  uint64_t w9 = util::SafeLoadAs<uint64_t>(in);
-  w9 = arrow::bit_util::FromLittleEndian(w9);
-  in += 8;
-  uint64_t w10 = util::SafeLoadAs<uint64_t>(in);
-  w10 = arrow::bit_util::FromLittleEndian(w10);
-  in += 8;
-  uint64_t w11 = util::SafeLoadAs<uint64_t>(in);
-  w11 = arrow::bit_util::FromLittleEndian(w11);
-  in += 8;
-  uint64_t w12 = util::SafeLoadAs<uint64_t>(in);
-  w12 = arrow::bit_util::FromLittleEndian(w12);
-  in += 8;
-  uint64_t w13 = util::SafeLoadAs<uint64_t>(in);
-  w13 = arrow::bit_util::FromLittleEndian(w13);
-  in += 8;
-  uint64_t w14 = util::SafeLoadAs<uint64_t>(in);
-  w14 = arrow::bit_util::FromLittleEndian(w14);
-  in += 8;
-  uint64_t w15 = util::SafeLoadAs<uint64_t>(in);
-  w15 = arrow::bit_util::FromLittleEndian(w15);
-  in += 8;
-  uint64_t w16 = util::SafeLoadAs<uint64_t>(in);
-  w16 = arrow::bit_util::FromLittleEndian(w16);
-  in += 8;
-  uint64_t w17 = util::SafeLoadAs<uint64_t>(in);
-  w17 = arrow::bit_util::FromLittleEndian(w17);
-  in += 8;
-  uint64_t w18 = util::SafeLoadAs<uint64_t>(in);
-  w18 = arrow::bit_util::FromLittleEndian(w18);
-  in += 8;
-  uint64_t w19 = util::SafeLoadAs<uint64_t>(in);
-  w19 = arrow::bit_util::FromLittleEndian(w19);
-  in += 8;
-  uint64_t w20 = util::SafeLoadAs<uint64_t>(in);
-  w20 = arrow::bit_util::FromLittleEndian(w20);
-  in += 8;
-  uint64_t w21 = util::SafeLoadAs<uint64_t>(in);
-  w21 = arrow::bit_util::FromLittleEndian(w21);
-  in += 8;
-  uint64_t w22 = util::SafeLoadAs<uint64_t>(in);
-  w22 = arrow::bit_util::FromLittleEndian(w22);
-  in += 8;
-  uint64_t w23 = util::SafeLoadAs<uint64_t>(in);
-  w23 = arrow::bit_util::FromLittleEndian(w23);
-  in += 8;
-  uint64_t w24 = util::SafeLoadAs<uint32_t>(in);
-  w24 = arrow::bit_util::FromLittleEndian(w24);
-  in += 4;
-  out[0] = (w0)&mask;
-  out[1] = ((w0 >> 49) | (w1 << 15)) & mask;
-  out[2] = ((w1 >> 34) | (w2 << 30)) & mask;
-  out[3] = ((w2 >> 19) | (w3 << 45)) & mask;
-  out[4] = (w3 >> 4) & mask;
-  out[5] = ((w3 >> 53) | (w4 << 11)) & mask;
-  out[6] = ((w4 >> 38) | (w5 << 26)) & mask;
-  out[7] = ((w5 >> 23) | (w6 << 41)) & mask;
-  out[8] = (w6 >> 8) & mask;
-  out[9] = ((w6 >> 57) | (w7 << 7)) & mask;
-  out[10] = ((w7 >> 42) | (w8 << 22)) & mask;
-  out[11] = ((w8 >> 27) | (w9 << 37)) & mask;
-  out[12] = (w9 >> 12) & mask;
-  out[13] = ((w9 >> 61) | (w10 << 3)) & mask;
-  out[14] = ((w10 >> 46) | (w11 << 18)) & mask;
-  out[15] = ((w11 >> 31) | (w12 << 33)) & mask;
-  out[16] = ((w12 >> 16) | (w13 << 48)) & mask;
-  out[17] = (w13 >> 1) & mask;
-  out[18] = ((w13 >> 50) | (w14 << 14)) & mask;
-  out[19] = ((w14 >> 35) | (w15 << 29)) & mask;
-  out[20] = ((w15 >> 20) | (w16 << 44)) & mask;
-  out[21] = (w16 >> 5) & mask;
-  out[22] = ((w16 >> 54) | (w17 << 10)) & mask;
-  out[23] = ((w17 >> 39) | (w18 << 25)) & mask;
-  out[24] = ((w18 >> 24) | (w19 << 40)) & mask;
-  out[25] = (w19 >> 9) & mask;
-  out[26] = ((w19 >> 58) | (w20 << 6)) & mask;
-  out[27] = ((w20 >> 43) | (w21 << 21)) & mask;
-  out[28] = ((w21 >> 28) | (w22 << 36)) & mask;
-  out[29] = (w22 >> 13) & mask;
-  out[30] = ((w22 >> 62) | (w23 << 2)) & mask;
-  out[31] = ((w23 >> 47) | (w24 << 17)) & mask;
-
-  return in;
-}
-
-inline const uint8_t* unpack50_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 1125899906842623ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint64_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 8;
-  uint64_t w9 = util::SafeLoadAs<uint64_t>(in);
-  w9 = arrow::bit_util::FromLittleEndian(w9);
-  in += 8;
-  uint64_t w10 = util::SafeLoadAs<uint64_t>(in);
-  w10 = arrow::bit_util::FromLittleEndian(w10);
-  in += 8;
-  uint64_t w11 = util::SafeLoadAs<uint64_t>(in);
-  w11 = arrow::bit_util::FromLittleEndian(w11);
-  in += 8;
-  uint64_t w12 = util::SafeLoadAs<uint64_t>(in);
-  w12 = arrow::bit_util::FromLittleEndian(w12);
-  in += 8;
-  uint64_t w13 = util::SafeLoadAs<uint64_t>(in);
-  w13 = arrow::bit_util::FromLittleEndian(w13);
-  in += 8;
-  uint64_t w14 = util::SafeLoadAs<uint64_t>(in);
-  w14 = arrow::bit_util::FromLittleEndian(w14);
-  in += 8;
-  uint64_t w15 = util::SafeLoadAs<uint64_t>(in);
-  w15 = arrow::bit_util::FromLittleEndian(w15);
-  in += 8;
-  uint64_t w16 = util::SafeLoadAs<uint64_t>(in);
-  w16 = arrow::bit_util::FromLittleEndian(w16);
-  in += 8;
-  uint64_t w17 = util::SafeLoadAs<uint64_t>(in);
-  w17 = arrow::bit_util::FromLittleEndian(w17);
-  in += 8;
-  uint64_t w18 = util::SafeLoadAs<uint64_t>(in);
-  w18 = arrow::bit_util::FromLittleEndian(w18);
-  in += 8;
-  uint64_t w19 = util::SafeLoadAs<uint64_t>(in);
-  w19 = arrow::bit_util::FromLittleEndian(w19);
-  in += 8;
-  uint64_t w20 = util::SafeLoadAs<uint64_t>(in);
-  w20 = arrow::bit_util::FromLittleEndian(w20);
-  in += 8;
-  uint64_t w21 = util::SafeLoadAs<uint64_t>(in);
-  w21 = arrow::bit_util::FromLittleEndian(w21);
-  in += 8;
-  uint64_t w22 = util::SafeLoadAs<uint64_t>(in);
-  w22 = arrow::bit_util::FromLittleEndian(w22);
-  in += 8;
-  uint64_t w23 = util::SafeLoadAs<uint64_t>(in);
-  w23 = arrow::bit_util::FromLittleEndian(w23);
-  in += 8;
-  uint64_t w24 = util::SafeLoadAs<uint64_t>(in);
-  w24 = arrow::bit_util::FromLittleEndian(w24);
-  in += 8;
-  out[0] = (w0)&mask;
-  out[1] = ((w0 >> 50) | (w1 << 14)) & mask;
-  out[2] = ((w1 >> 36) | (w2 << 28)) & mask;
-  out[3] = ((w2 >> 22) | (w3 << 42)) & mask;
-  out[4] = (w3 >> 8) & mask;
-  out[5] = ((w3 >> 58) | (w4 << 6)) & mask;
-  out[6] = ((w4 >> 44) | (w5 << 20)) & mask;
-  out[7] = ((w5 >> 30) | (w6 << 34)) & mask;
-  out[8] = ((w6 >> 16) | (w7 << 48)) & mask;
-  out[9] = (w7 >> 2) & mask;
-  out[10] = ((w7 >> 52) | (w8 << 12)) & mask;
-  out[11] = ((w8 >> 38) | (w9 << 26)) & mask;
-  out[12] = ((w9 >> 24) | (w10 << 40)) & mask;
-  out[13] = (w10 >> 10) & mask;
-  out[14] = ((w10 >> 60) | (w11 << 4)) & mask;
-  out[15] = ((w11 >> 46) | (w12 << 18)) & mask;
-  out[16] = ((w12 >> 32) | (w13 << 32)) & mask;
-  out[17] = ((w13 >> 18) | (w14 << 46)) & mask;
-  out[18] = (w14 >> 4) & mask;
-  out[19] = ((w14 >> 54) | (w15 << 10)) & mask;
-  out[20] = ((w15 >> 40) | (w16 << 24)) & mask;
-  out[21] = ((w16 >> 26) | (w17 << 38)) & mask;
-  out[22] = (w17 >> 12) & mask;
-  out[23] = ((w17 >> 62) | (w18 << 2)) & mask;
-  out[24] = ((w18 >> 48) | (w19 << 16)) & mask;
-  out[25] = ((w19 >> 34) | (w20 << 30)) & mask;
-  out[26] = ((w20 >> 20) | (w21 << 44)) & mask;
-  out[27] = (w21 >> 6) & mask;
-  out[28] = ((w21 >> 56) | (w22 << 8)) & mask;
-  out[29] = ((w22 >> 42) | (w23 << 22)) & mask;
-  out[30] = ((w23 >> 28) | (w24 << 36)) & mask;
-  out[31] = w24 >> 14;
-
-  return in;
-}
-
-inline const uint8_t* unpack51_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 2251799813685247ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint64_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 8;
-  uint64_t w9 = util::SafeLoadAs<uint64_t>(in);
-  w9 = arrow::bit_util::FromLittleEndian(w9);
-  in += 8;
-  uint64_t w10 = util::SafeLoadAs<uint64_t>(in);
-  w10 = arrow::bit_util::FromLittleEndian(w10);
-  in += 8;
-  uint64_t w11 = util::SafeLoadAs<uint64_t>(in);
-  w11 = arrow::bit_util::FromLittleEndian(w11);
-  in += 8;
-  uint64_t w12 = util::SafeLoadAs<uint64_t>(in);
-  w12 = arrow::bit_util::FromLittleEndian(w12);
-  in += 8;
-  uint64_t w13 = util::SafeLoadAs<uint64_t>(in);
-  w13 = arrow::bit_util::FromLittleEndian(w13);
-  in += 8;
-  uint64_t w14 = util::SafeLoadAs<uint64_t>(in);
-  w14 = arrow::bit_util::FromLittleEndian(w14);
-  in += 8;
-  uint64_t w15 = util::SafeLoadAs<uint64_t>(in);
-  w15 = arrow::bit_util::FromLittleEndian(w15);
-  in += 8;
-  uint64_t w16 = util::SafeLoadAs<uint64_t>(in);
-  w16 = arrow::bit_util::FromLittleEndian(w16);
-  in += 8;
-  uint64_t w17 = util::SafeLoadAs<uint64_t>(in);
-  w17 = arrow::bit_util::FromLittleEndian(w17);
-  in += 8;
-  uint64_t w18 = util::SafeLoadAs<uint64_t>(in);
-  w18 = arrow::bit_util::FromLittleEndian(w18);
-  in += 8;
-  uint64_t w19 = util::SafeLoadAs<uint64_t>(in);
-  w19 = arrow::bit_util::FromLittleEndian(w19);
-  in += 8;
-  uint64_t w20 = util::SafeLoadAs<uint64_t>(in);
-  w20 = arrow::bit_util::FromLittleEndian(w20);
-  in += 8;
-  uint64_t w21 = util::SafeLoadAs<uint64_t>(in);
-  w21 = arrow::bit_util::FromLittleEndian(w21);
-  in += 8;
-  uint64_t w22 = util::SafeLoadAs<uint64_t>(in);
-  w22 = arrow::bit_util::FromLittleEndian(w22);
-  in += 8;
-  uint64_t w23 = util::SafeLoadAs<uint64_t>(in);
-  w23 = arrow::bit_util::FromLittleEndian(w23);
-  in += 8;
-  uint64_t w24 = util::SafeLoadAs<uint64_t>(in);
-  w24 = arrow::bit_util::FromLittleEndian(w24);
-  in += 8;
-  uint64_t w25 = util::SafeLoadAs<uint32_t>(in);
-  w25 = arrow::bit_util::FromLittleEndian(w25);
-  in += 4;
-  out[0] = (w0)&mask;
-  out[1] = ((w0 >> 51) | (w1 << 13)) & mask;
-  out[2] = ((w1 >> 38) | (w2 << 26)) & mask;
-  out[3] = ((w2 >> 25) | (w3 << 39)) & mask;
-  out[4] = (w3 >> 12) & mask;
-  out[5] = ((w3 >> 63) | (w4 << 1)) & mask;
-  out[6] = ((w4 >> 50) | (w5 << 14)) & mask;
-  out[7] = ((w5 >> 37) | (w6 << 27)) & mask;
-  out[8] = ((w6 >> 24) | (w7 << 40)) & mask;
-  out[9] = (w7 >> 11) & mask;
-  out[10] = ((w7 >> 62) | (w8 << 2)) & mask;
-  out[11] = ((w8 >> 49) | (w9 << 15)) & mask;
-  out[12] = ((w9 >> 36) | (w10 << 28)) & mask;
-  out[13] = ((w10 >> 23) | (w11 << 41)) & mask;
-  out[14] = (w11 >> 10) & mask;
-  out[15] = ((w11 >> 61) | (w12 << 3)) & mask;
-  out[16] = ((w12 >> 48) | (w13 << 16)) & mask;
-  out[17] = ((w13 >> 35) | (w14 << 29)) & mask;
-  out[18] = ((w14 >> 22) | (w15 << 42)) & mask;
-  out[19] = (w15 >> 9) & mask;
-  out[20] = ((w15 >> 60) | (w16 << 4)) & mask;
-  out[21] = ((w16 >> 47) | (w17 << 17)) & mask;
-  out[22] = ((w17 >> 34) | (w18 << 30)) & mask;
-  out[23] = ((w18 >> 21) | (w19 << 43)) & mask;
-  out[24] = (w19 >> 8) & mask;
-  out[25] = ((w19 >> 59) | (w20 << 5)) & mask;
-  out[26] = ((w20 >> 46) | (w21 << 18)) & mask;
-  out[27] = ((w21 >> 33) | (w22 << 31)) & mask;
-  out[28] = ((w22 >> 20) | (w23 << 44)) & mask;
-  out[29] = (w23 >> 7) & mask;
-  out[30] = ((w23 >> 58) | (w24 << 6)) & mask;
-  out[31] = ((w24 >> 45) | (w25 << 19)) & mask;
-
-  return in;
-}
-
-inline const uint8_t* unpack52_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 4503599627370495ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint64_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 8;
-  uint64_t w9 = util::SafeLoadAs<uint64_t>(in);
-  w9 = arrow::bit_util::FromLittleEndian(w9);
-  in += 8;
-  uint64_t w10 = util::SafeLoadAs<uint64_t>(in);
-  w10 = arrow::bit_util::FromLittleEndian(w10);
-  in += 8;
-  uint64_t w11 = util::SafeLoadAs<uint64_t>(in);
-  w11 = arrow::bit_util::FromLittleEndian(w11);
-  in += 8;
-  uint64_t w12 = util::SafeLoadAs<uint64_t>(in);
-  w12 = arrow::bit_util::FromLittleEndian(w12);
-  in += 8;
-  uint64_t w13 = util::SafeLoadAs<uint64_t>(in);
-  w13 = arrow::bit_util::FromLittleEndian(w13);
-  in += 8;
-  uint64_t w14 = util::SafeLoadAs<uint64_t>(in);
-  w14 = arrow::bit_util::FromLittleEndian(w14);
-  in += 8;
-  uint64_t w15 = util::SafeLoadAs<uint64_t>(in);
-  w15 = arrow::bit_util::FromLittleEndian(w15);
-  in += 8;
-  uint64_t w16 = util::SafeLoadAs<uint64_t>(in);
-  w16 = arrow::bit_util::FromLittleEndian(w16);
-  in += 8;
-  uint64_t w17 = util::SafeLoadAs<uint64_t>(in);
-  w17 = arrow::bit_util::FromLittleEndian(w17);
-  in += 8;
-  uint64_t w18 = util::SafeLoadAs<uint64_t>(in);
-  w18 = arrow::bit_util::FromLittleEndian(w18);
-  in += 8;
-  uint64_t w19 = util::SafeLoadAs<uint64_t>(in);
-  w19 = arrow::bit_util::FromLittleEndian(w19);
-  in += 8;
-  uint64_t w20 = util::SafeLoadAs<uint64_t>(in);
-  w20 = arrow::bit_util::FromLittleEndian(w20);
-  in += 8;
-  uint64_t w21 = util::SafeLoadAs<uint64_t>(in);
-  w21 = arrow::bit_util::FromLittleEndian(w21);
-  in += 8;
-  uint64_t w22 = util::SafeLoadAs<uint64_t>(in);
-  w22 = arrow::bit_util::FromLittleEndian(w22);
-  in += 8;
-  uint64_t w23 = util::SafeLoadAs<uint64_t>(in);
-  w23 = arrow::bit_util::FromLittleEndian(w23);
-  in += 8;
-  uint64_t w24 = util::SafeLoadAs<uint64_t>(in);
-  w24 = arrow::bit_util::FromLittleEndian(w24);
-  in += 8;
-  uint64_t w25 = util::SafeLoadAs<uint64_t>(in);
-  w25 = arrow::bit_util::FromLittleEndian(w25);
-  in += 8;
-  out[0] = (w0)&mask;
-  out[1] = ((w0 >> 52) | (w1 << 12)) & mask;
-  out[2] = ((w1 >> 40) | (w2 << 24)) & mask;
-  out[3] = ((w2 >> 28) | (w3 << 36)) & mask;
-  out[4] = ((w3 >> 16) | (w4 << 48)) & mask;
-  out[5] = (w4 >> 4) & mask;
-  out[6] = ((w4 >> 56) | (w5 << 8)) & mask;
-  out[7] = ((w5 >> 44) | (w6 << 20)) & mask;
-  out[8] = ((w6 >> 32) | (w7 << 32)) & mask;
-  out[9] = ((w7 >> 20) | (w8 << 44)) & mask;
-  out[10] = (w8 >> 8) & mask;
-  out[11] = ((w8 >> 60) | (w9 << 4)) & mask;
-  out[12] = ((w9 >> 48) | (w10 << 16)) & mask;
-  out[13] = ((w10 >> 36) | (w11 << 28)) & mask;
-  out[14] = ((w11 >> 24) | (w12 << 40)) & mask;
-  out[15] = w12 >> 12;
-  out[16] = (w13)&mask;
-  out[17] = ((w13 >> 52) | (w14 << 12)) & mask;
-  out[18] = ((w14 >> 40) | (w15 << 24)) & mask;
-  out[19] = ((w15 >> 28) | (w16 << 36)) & mask;
-  out[20] = ((w16 >> 16) | (w17 << 48)) & mask;
-  out[21] = (w17 >> 4) & mask;
-  out[22] = ((w17 >> 56) | (w18 << 8)) & mask;
-  out[23] = ((w18 >> 44) | (w19 << 20)) & mask;
-  out[24] = ((w19 >> 32) | (w20 << 32)) & mask;
-  out[25] = ((w20 >> 20) | (w21 << 44)) & mask;
-  out[26] = (w21 >> 8) & mask;
-  out[27] = ((w21 >> 60) | (w22 << 4)) & mask;
-  out[28] = ((w22 >> 48) | (w23 << 16)) & mask;
-  out[29] = ((w23 >> 36) | (w24 << 28)) & mask;
-  out[30] = ((w24 >> 24) | (w25 << 40)) & mask;
-  out[31] = w25 >> 12;
-
-  return in;
-}
-
-inline const uint8_t* unpack53_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 9007199254740991ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint64_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 8;
-  uint64_t w9 = util::SafeLoadAs<uint64_t>(in);
-  w9 = arrow::bit_util::FromLittleEndian(w9);
-  in += 8;
-  uint64_t w10 = util::SafeLoadAs<uint64_t>(in);
-  w10 = arrow::bit_util::FromLittleEndian(w10);
-  in += 8;
-  uint64_t w11 = util::SafeLoadAs<uint64_t>(in);
-  w11 = arrow::bit_util::FromLittleEndian(w11);
-  in += 8;
-  uint64_t w12 = util::SafeLoadAs<uint64_t>(in);
-  w12 = arrow::bit_util::FromLittleEndian(w12);
-  in += 8;
-  uint64_t w13 = util::SafeLoadAs<uint64_t>(in);
-  w13 = arrow::bit_util::FromLittleEndian(w13);
-  in += 8;
-  uint64_t w14 = util::SafeLoadAs<uint64_t>(in);
-  w14 = arrow::bit_util::FromLittleEndian(w14);
-  in += 8;
-  uint64_t w15 = util::SafeLoadAs<uint64_t>(in);
-  w15 = arrow::bit_util::FromLittleEndian(w15);
-  in += 8;
-  uint64_t w16 = util::SafeLoadAs<uint64_t>(in);
-  w16 = arrow::bit_util::FromLittleEndian(w16);
-  in += 8;
-  uint64_t w17 = util::SafeLoadAs<uint64_t>(in);
-  w17 = arrow::bit_util::FromLittleEndian(w17);
-  in += 8;
-  uint64_t w18 = util::SafeLoadAs<uint64_t>(in);
-  w18 = arrow::bit_util::FromLittleEndian(w18);
-  in += 8;
-  uint64_t w19 = util::SafeLoadAs<uint64_t>(in);
-  w19 = arrow::bit_util::FromLittleEndian(w19);
-  in += 8;
-  uint64_t w20 = util::SafeLoadAs<uint64_t>(in);
-  w20 = arrow::bit_util::FromLittleEndian(w20);
-  in += 8;
-  uint64_t w21 = util::SafeLoadAs<uint64_t>(in);
-  w21 = arrow::bit_util::FromLittleEndian(w21);
-  in += 8;
-  uint64_t w22 = util::SafeLoadAs<uint64_t>(in);
-  w22 = arrow::bit_util::FromLittleEndian(w22);
-  in += 8;
-  uint64_t w23 = util::SafeLoadAs<uint64_t>(in);
-  w23 = arrow::bit_util::FromLittleEndian(w23);
-  in += 8;
-  uint64_t w24 = util::SafeLoadAs<uint64_t>(in);
-  w24 = arrow::bit_util::FromLittleEndian(w24);
-  in += 8;
-  uint64_t w25 = util::SafeLoadAs<uint64_t>(in);
-  w25 = arrow::bit_util::FromLittleEndian(w25);
-  in += 8;
-  uint64_t w26 = util::SafeLoadAs<uint32_t>(in);
-  w26 = arrow::bit_util::FromLittleEndian(w26);
-  in += 4;
-  out[0] = (w0)&mask;
-  out[1] = ((w0 >> 53) | (w1 << 11)) & mask;
-  out[2] = ((w1 >> 42) | (w2 << 22)) & mask;
-  out[3] = ((w2 >> 31) | (w3 << 33)) & mask;
-  out[4] = ((w3 >> 20) | (w4 << 44)) & mask;
-  out[5] = (w4 >> 9) & mask;
-  out[6] = ((w4 >> 62) | (w5 << 2)) & mask;
-  out[7] = ((w5 >> 51) | (w6 << 13)) & mask;
-  out[8] = ((w6 >> 40) | (w7 << 24)) & mask;
-  out[9] = ((w7 >> 29) | (w8 << 35)) & mask;
-  out[10] = ((w8 >> 18) | (w9 << 46)) & mask;
-  out[11] = (w9 >> 7) & mask;
-  out[12] = ((w9 >> 60) | (w10 << 4)) & mask;
-  out[13] = ((w10 >> 49) | (w11 << 15)) & mask;
-  out[14] = ((w11 >> 38) | (w12 << 26)) & mask;
-  out[15] = ((w12 >> 27) | (w13 << 37)) & mask;
-  out[16] = ((w13 >> 16) | (w14 << 48)) & mask;
-  out[17] = (w14 >> 5) & mask;
-  out[18] = ((w14 >> 58) | (w15 << 6)) & mask;
-  out[19] = ((w15 >> 47) | (w16 << 17)) & mask;
-  out[20] = ((w16 >> 36) | (w17 << 28)) & mask;
-  out[21] = ((w17 >> 25) | (w18 << 39)) & mask;
-  out[22] = ((w18 >> 14) | (w19 << 50)) & mask;
-  out[23] = (w19 >> 3) & mask;
-  out[24] = ((w19 >> 56) | (w20 << 8)) & mask;
-  out[25] = ((w20 >> 45) | (w21 << 19)) & mask;
-  out[26] = ((w21 >> 34) | (w22 << 30)) & mask;
-  out[27] = ((w22 >> 23) | (w23 << 41)) & mask;
-  out[28] = ((w23 >> 12) | (w24 << 52)) & mask;
-  out[29] = (w24 >> 1) & mask;
-  out[30] = ((w24 >> 54) | (w25 << 10)) & mask;
-  out[31] = ((w25 >> 43) | (w26 << 21)) & mask;
-
-  return in;
-}
-
-inline const uint8_t* unpack54_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 18014398509481983ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint64_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 8;
-  uint64_t w9 = util::SafeLoadAs<uint64_t>(in);
-  w9 = arrow::bit_util::FromLittleEndian(w9);
-  in += 8;
-  uint64_t w10 = util::SafeLoadAs<uint64_t>(in);
-  w10 = arrow::bit_util::FromLittleEndian(w10);
-  in += 8;
-  uint64_t w11 = util::SafeLoadAs<uint64_t>(in);
-  w11 = arrow::bit_util::FromLittleEndian(w11);
-  in += 8;
-  uint64_t w12 = util::SafeLoadAs<uint64_t>(in);
-  w12 = arrow::bit_util::FromLittleEndian(w12);
-  in += 8;
-  uint64_t w13 = util::SafeLoadAs<uint64_t>(in);
-  w13 = arrow::bit_util::FromLittleEndian(w13);
-  in += 8;
-  uint64_t w14 = util::SafeLoadAs<uint64_t>(in);
-  w14 = arrow::bit_util::FromLittleEndian(w14);
-  in += 8;
-  uint64_t w15 = util::SafeLoadAs<uint64_t>(in);
-  w15 = arrow::bit_util::FromLittleEndian(w15);
-  in += 8;
-  uint64_t w16 = util::SafeLoadAs<uint64_t>(in);
-  w16 = arrow::bit_util::FromLittleEndian(w16);
-  in += 8;
-  uint64_t w17 = util::SafeLoadAs<uint64_t>(in);
-  w17 = arrow::bit_util::FromLittleEndian(w17);
-  in += 8;
-  uint64_t w18 = util::SafeLoadAs<uint64_t>(in);
-  w18 = arrow::bit_util::FromLittleEndian(w18);
-  in += 8;
-  uint64_t w19 = util::SafeLoadAs<uint64_t>(in);
-  w19 = arrow::bit_util::FromLittleEndian(w19);
-  in += 8;
-  uint64_t w20 = util::SafeLoadAs<uint64_t>(in);
-  w20 = arrow::bit_util::FromLittleEndian(w20);
-  in += 8;
-  uint64_t w21 = util::SafeLoadAs<uint64_t>(in);
-  w21 = arrow::bit_util::FromLittleEndian(w21);
-  in += 8;
-  uint64_t w22 = util::SafeLoadAs<uint64_t>(in);
-  w22 = arrow::bit_util::FromLittleEndian(w22);
-  in += 8;
-  uint64_t w23 = util::SafeLoadAs<uint64_t>(in);
-  w23 = arrow::bit_util::FromLittleEndian(w23);
-  in += 8;
-  uint64_t w24 = util::SafeLoadAs<uint64_t>(in);
-  w24 = arrow::bit_util::FromLittleEndian(w24);
-  in += 8;
-  uint64_t w25 = util::SafeLoadAs<uint64_t>(in);
-  w25 = arrow::bit_util::FromLittleEndian(w25);
-  in += 8;
-  uint64_t w26 = util::SafeLoadAs<uint64_t>(in);
-  w26 = arrow::bit_util::FromLittleEndian(w26);
-  in += 8;
-  out[0] = (w0)&mask;
-  out[1] = ((w0 >> 54) | (w1 << 10)) & mask;
-  out[2] = ((w1 >> 44) | (w2 << 20)) & mask;
-  out[3] = ((w2 >> 34) | (w3 << 30)) & mask;
-  out[4] = ((w3 >> 24) | (w4 << 40)) & mask;
-  out[5] = ((w4 >> 14) | (w5 << 50)) & mask;
-  out[6] = (w5 >> 4) & mask;
-  out[7] = ((w5 >> 58) | (w6 << 6)) & mask;
-  out[8] = ((w6 >> 48) | (w7 << 16)) & mask;
-  out[9] = ((w7 >> 38) | (w8 << 26)) & mask;
-  out[10] = ((w8 >> 28) | (w9 << 36)) & mask;
-  out[11] = ((w9 >> 18) | (w10 << 46)) & mask;
-  out[12] = (w10 >> 8) & mask;
-  out[13] = ((w10 >> 62) | (w11 << 2)) & mask;
-  out[14] = ((w11 >> 52) | (w12 << 12)) & mask;
-  out[15] = ((w12 >> 42) | (w13 << 22)) & mask;
-  out[16] = ((w13 >> 32) | (w14 << 32)) & mask;
-  out[17] = ((w14 >> 22) | (w15 << 42)) & mask;
-  out[18] = ((w15 >> 12) | (w16 << 52)) & mask;
-  out[19] = (w16 >> 2) & mask;
-  out[20] = ((w16 >> 56) | (w17 << 8)) & mask;
-  out[21] = ((w17 >> 46) | (w18 << 18)) & mask;
-  out[22] = ((w18 >> 36) | (w19 << 28)) & mask;
-  out[23] = ((w19 >> 26) | (w20 << 38)) & mask;
-  out[24] = ((w20 >> 16) | (w21 << 48)) & mask;
-  out[25] = (w21 >> 6) & mask;
-  out[26] = ((w21 >> 60) | (w22 << 4)) & mask;
-  out[27] = ((w22 >> 50) | (w23 << 14)) & mask;
-  out[28] = ((w23 >> 40) | (w24 << 24)) & mask;
-  out[29] = ((w24 >> 30) | (w25 << 34)) & mask;
-  out[30] = ((w25 >> 20) | (w26 << 44)) & mask;
-  out[31] = w26 >> 10;
-
-  return in;
-}
-
-inline const uint8_t* unpack55_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 36028797018963967ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint64_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 8;
-  uint64_t w9 = util::SafeLoadAs<uint64_t>(in);
-  w9 = arrow::bit_util::FromLittleEndian(w9);
-  in += 8;
-  uint64_t w10 = util::SafeLoadAs<uint64_t>(in);
-  w10 = arrow::bit_util::FromLittleEndian(w10);
-  in += 8;
-  uint64_t w11 = util::SafeLoadAs<uint64_t>(in);
-  w11 = arrow::bit_util::FromLittleEndian(w11);
-  in += 8;
-  uint64_t w12 = util::SafeLoadAs<uint64_t>(in);
-  w12 = arrow::bit_util::FromLittleEndian(w12);
-  in += 8;
-  uint64_t w13 = util::SafeLoadAs<uint64_t>(in);
-  w13 = arrow::bit_util::FromLittleEndian(w13);
-  in += 8;
-  uint64_t w14 = util::SafeLoadAs<uint64_t>(in);
-  w14 = arrow::bit_util::FromLittleEndian(w14);
-  in += 8;
-  uint64_t w15 = util::SafeLoadAs<uint64_t>(in);
-  w15 = arrow::bit_util::FromLittleEndian(w15);
-  in += 8;
-  uint64_t w16 = util::SafeLoadAs<uint64_t>(in);
-  w16 = arrow::bit_util::FromLittleEndian(w16);
-  in += 8;
-  uint64_t w17 = util::SafeLoadAs<uint64_t>(in);
-  w17 = arrow::bit_util::FromLittleEndian(w17);
-  in += 8;
-  uint64_t w18 = util::SafeLoadAs<uint64_t>(in);
-  w18 = arrow::bit_util::FromLittleEndian(w18);
-  in += 8;
-  uint64_t w19 = util::SafeLoadAs<uint64_t>(in);
-  w19 = arrow::bit_util::FromLittleEndian(w19);
-  in += 8;
-  uint64_t w20 = util::SafeLoadAs<uint64_t>(in);
-  w20 = arrow::bit_util::FromLittleEndian(w20);
-  in += 8;
-  uint64_t w21 = util::SafeLoadAs<uint64_t>(in);
-  w21 = arrow::bit_util::FromLittleEndian(w21);
-  in += 8;
-  uint64_t w22 = util::SafeLoadAs<uint64_t>(in);
-  w22 = arrow::bit_util::FromLittleEndian(w22);
-  in += 8;
-  uint64_t w23 = util::SafeLoadAs<uint64_t>(in);
-  w23 = arrow::bit_util::FromLittleEndian(w23);
-  in += 8;
-  uint64_t w24 = util::SafeLoadAs<uint64_t>(in);
-  w24 = arrow::bit_util::FromLittleEndian(w24);
-  in += 8;
-  uint64_t w25 = util::SafeLoadAs<uint64_t>(in);
-  w25 = arrow::bit_util::FromLittleEndian(w25);
-  in += 8;
-  uint64_t w26 = util::SafeLoadAs<uint64_t>(in);
-  w26 = arrow::bit_util::FromLittleEndian(w26);
-  in += 8;
-  uint64_t w27 = util::SafeLoadAs<uint32_t>(in);
-  w27 = arrow::bit_util::FromLittleEndian(w27);
-  in += 4;
-  out[0] = (w0)&mask;
-  out[1] = ((w0 >> 55) | (w1 << 9)) & mask;
-  out[2] = ((w1 >> 46) | (w2 << 18)) & mask;
-  out[3] = ((w2 >> 37) | (w3 << 27)) & mask;
-  out[4] = ((w3 >> 28) | (w4 << 36)) & mask;
-  out[5] = ((w4 >> 19) | (w5 << 45)) & mask;
-  out[6] = ((w5 >> 10) | (w6 << 54)) & mask;
-  out[7] = (w6 >> 1) & mask;
-  out[8] = ((w6 >> 56) | (w7 << 8)) & mask;
-  out[9] = ((w7 >> 47) | (w8 << 17)) & mask;
-  out[10] = ((w8 >> 38) | (w9 << 26)) & mask;
-  out[11] = ((w9 >> 29) | (w10 << 35)) & mask;
-  out[12] = ((w10 >> 20) | (w11 << 44)) & mask;
-  out[13] = ((w11 >> 11) | (w12 << 53)) & mask;
-  out[14] = (w12 >> 2) & mask;
-  out[15] = ((w12 >> 57) | (w13 << 7)) & mask;
-  out[16] = ((w13 >> 48) | (w14 << 16)) & mask;
-  out[17] = ((w14 >> 39) | (w15 << 25)) & mask;
-  out[18] = ((w15 >> 30) | (w16 << 34)) & mask;
-  out[19] = ((w16 >> 21) | (w17 << 43)) & mask;
-  out[20] = ((w17 >> 12) | (w18 << 52)) & mask;
-  out[21] = (w18 >> 3) & mask;
-  out[22] = ((w18 >> 58) | (w19 << 6)) & mask;
-  out[23] = ((w19 >> 49) | (w20 << 15)) & mask;
-  out[24] = ((w20 >> 40) | (w21 << 24)) & mask;
-  out[25] = ((w21 >> 31) | (w22 << 33)) & mask;
-  out[26] = ((w22 >> 22) | (w23 << 42)) & mask;
-  out[27] = ((w23 >> 13) | (w24 << 51)) & mask;
-  out[28] = (w24 >> 4) & mask;
-  out[29] = ((w24 >> 59) | (w25 << 5)) & mask;
-  out[30] = ((w25 >> 50) | (w26 << 14)) & mask;
-  out[31] = ((w26 >> 41) | (w27 << 23)) & mask;
-
-  return in;
-}
-
-inline const uint8_t* unpack56_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 72057594037927935ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint64_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 8;
-  uint64_t w9 = util::SafeLoadAs<uint64_t>(in);
-  w9 = arrow::bit_util::FromLittleEndian(w9);
-  in += 8;
-  uint64_t w10 = util::SafeLoadAs<uint64_t>(in);
-  w10 = arrow::bit_util::FromLittleEndian(w10);
-  in += 8;
-  uint64_t w11 = util::SafeLoadAs<uint64_t>(in);
-  w11 = arrow::bit_util::FromLittleEndian(w11);
-  in += 8;
-  uint64_t w12 = util::SafeLoadAs<uint64_t>(in);
-  w12 = arrow::bit_util::FromLittleEndian(w12);
-  in += 8;
-  uint64_t w13 = util::SafeLoadAs<uint64_t>(in);
-  w13 = arrow::bit_util::FromLittleEndian(w13);
-  in += 8;
-  uint64_t w14 = util::SafeLoadAs<uint64_t>(in);
-  w14 = arrow::bit_util::FromLittleEndian(w14);
-  in += 8;
-  uint64_t w15 = util::SafeLoadAs<uint64_t>(in);
-  w15 = arrow::bit_util::FromLittleEndian(w15);
-  in += 8;
-  uint64_t w16 = util::SafeLoadAs<uint64_t>(in);
-  w16 = arrow::bit_util::FromLittleEndian(w16);
-  in += 8;
-  uint64_t w17 = util::SafeLoadAs<uint64_t>(in);
-  w17 = arrow::bit_util::FromLittleEndian(w17);
-  in += 8;
-  uint64_t w18 = util::SafeLoadAs<uint64_t>(in);
-  w18 = arrow::bit_util::FromLittleEndian(w18);
-  in += 8;
-  uint64_t w19 = util::SafeLoadAs<uint64_t>(in);
-  w19 = arrow::bit_util::FromLittleEndian(w19);
-  in += 8;
-  uint64_t w20 = util::SafeLoadAs<uint64_t>(in);
-  w20 = arrow::bit_util::FromLittleEndian(w20);
-  in += 8;
-  uint64_t w21 = util::SafeLoadAs<uint64_t>(in);
-  w21 = arrow::bit_util::FromLittleEndian(w21);
-  in += 8;
-  uint64_t w22 = util::SafeLoadAs<uint64_t>(in);
-  w22 = arrow::bit_util::FromLittleEndian(w22);
-  in += 8;
-  uint64_t w23 = util::SafeLoadAs<uint64_t>(in);
-  w23 = arrow::bit_util::FromLittleEndian(w23);
-  in += 8;
-  uint64_t w24 = util::SafeLoadAs<uint64_t>(in);
-  w24 = arrow::bit_util::FromLittleEndian(w24);
-  in += 8;
-  uint64_t w25 = util::SafeLoadAs<uint64_t>(in);
-  w25 = arrow::bit_util::FromLittleEndian(w25);
-  in += 8;
-  uint64_t w26 = util::SafeLoadAs<uint64_t>(in);
-  w26 = arrow::bit_util::FromLittleEndian(w26);
-  in += 8;
-  uint64_t w27 = util::SafeLoadAs<uint64_t>(in);
-  w27 = arrow::bit_util::FromLittleEndian(w27);
-  in += 8;
-  out[0] = (w0)&mask;
-  out[1] = ((w0 >> 56) | (w1 << 8)) & mask;
-  out[2] = ((w1 >> 48) | (w2 << 16)) & mask;
-  out[3] = ((w2 >> 40) | (w3 << 24)) & mask;
-  out[4] = ((w3 >> 32) | (w4 << 32)) & mask;
-  out[5] = ((w4 >> 24) | (w5 << 40)) & mask;
-  out[6] = ((w5 >> 16) | (w6 << 48)) & mask;
-  out[7] = w6 >> 8;
-  out[8] = (w7)&mask;
-  out[9] = ((w7 >> 56) | (w8 << 8)) & mask;
-  out[10] = ((w8 >> 48) | (w9 << 16)) & mask;
-  out[11] = ((w9 >> 40) | (w10 << 24)) & mask;
-  out[12] = ((w10 >> 32) | (w11 << 32)) & mask;
-  out[13] = ((w11 >> 24) | (w12 << 40)) & mask;
-  out[14] = ((w12 >> 16) | (w13 << 48)) & mask;
-  out[15] = w13 >> 8;
-  out[16] = (w14)&mask;
-  out[17] = ((w14 >> 56) | (w15 << 8)) & mask;
-  out[18] = ((w15 >> 48) | (w16 << 16)) & mask;
-  out[19] = ((w16 >> 40) | (w17 << 24)) & mask;
-  out[20] = ((w17 >> 32) | (w18 << 32)) & mask;
-  out[21] = ((w18 >> 24) | (w19 << 40)) & mask;
-  out[22] = ((w19 >> 16) | (w20 << 48)) & mask;
-  out[23] = w20 >> 8;
-  out[24] = (w21)&mask;
-  out[25] = ((w21 >> 56) | (w22 << 8)) & mask;
-  out[26] = ((w22 >> 48) | (w23 << 16)) & mask;
-  out[27] = ((w23 >> 40) | (w24 << 24)) & mask;
-  out[28] = ((w24 >> 32) | (w25 << 32)) & mask;
-  out[29] = ((w25 >> 24) | (w26 << 40)) & mask;
-  out[30] = ((w26 >> 16) | (w27 << 48)) & mask;
-  out[31] = w27 >> 8;
-
-  return in;
-}
-
-inline const uint8_t* unpack57_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 144115188075855871ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint64_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 8;
-  uint64_t w9 = util::SafeLoadAs<uint64_t>(in);
-  w9 = arrow::bit_util::FromLittleEndian(w9);
-  in += 8;
-  uint64_t w10 = util::SafeLoadAs<uint64_t>(in);
-  w10 = arrow::bit_util::FromLittleEndian(w10);
-  in += 8;
-  uint64_t w11 = util::SafeLoadAs<uint64_t>(in);
-  w11 = arrow::bit_util::FromLittleEndian(w11);
-  in += 8;
-  uint64_t w12 = util::SafeLoadAs<uint64_t>(in);
-  w12 = arrow::bit_util::FromLittleEndian(w12);
-  in += 8;
-  uint64_t w13 = util::SafeLoadAs<uint64_t>(in);
-  w13 = arrow::bit_util::FromLittleEndian(w13);
-  in += 8;
-  uint64_t w14 = util::SafeLoadAs<uint64_t>(in);
-  w14 = arrow::bit_util::FromLittleEndian(w14);
-  in += 8;
-  uint64_t w15 = util::SafeLoadAs<uint64_t>(in);
-  w15 = arrow::bit_util::FromLittleEndian(w15);
-  in += 8;
-  uint64_t w16 = util::SafeLoadAs<uint64_t>(in);
-  w16 = arrow::bit_util::FromLittleEndian(w16);
-  in += 8;
-  uint64_t w17 = util::SafeLoadAs<uint64_t>(in);
-  w17 = arrow::bit_util::FromLittleEndian(w17);
-  in += 8;
-  uint64_t w18 = util::SafeLoadAs<uint64_t>(in);
-  w18 = arrow::bit_util::FromLittleEndian(w18);
-  in += 8;
-  uint64_t w19 = util::SafeLoadAs<uint64_t>(in);
-  w19 = arrow::bit_util::FromLittleEndian(w19);
-  in += 8;
-  uint64_t w20 = util::SafeLoadAs<uint64_t>(in);
-  w20 = arrow::bit_util::FromLittleEndian(w20);
-  in += 8;
-  uint64_t w21 = util::SafeLoadAs<uint64_t>(in);
-  w21 = arrow::bit_util::FromLittleEndian(w21);
-  in += 8;
-  uint64_t w22 = util::SafeLoadAs<uint64_t>(in);
-  w22 = arrow::bit_util::FromLittleEndian(w22);
-  in += 8;
-  uint64_t w23 = util::SafeLoadAs<uint64_t>(in);
-  w23 = arrow::bit_util::FromLittleEndian(w23);
-  in += 8;
-  uint64_t w24 = util::SafeLoadAs<uint64_t>(in);
-  w24 = arrow::bit_util::FromLittleEndian(w24);
-  in += 8;
-  uint64_t w25 = util::SafeLoadAs<uint64_t>(in);
-  w25 = arrow::bit_util::FromLittleEndian(w25);
-  in += 8;
-  uint64_t w26 = util::SafeLoadAs<uint64_t>(in);
-  w26 = arrow::bit_util::FromLittleEndian(w26);
-  in += 8;
-  uint64_t w27 = util::SafeLoadAs<uint64_t>(in);
-  w27 = arrow::bit_util::FromLittleEndian(w27);
-  in += 8;
-  uint64_t w28 = util::SafeLoadAs<uint32_t>(in);
-  w28 = arrow::bit_util::FromLittleEndian(w28);
-  in += 4;
-  out[0] = (w0)&mask;
-  out[1] = ((w0 >> 57) | (w1 << 7)) & mask;
-  out[2] = ((w1 >> 50) | (w2 << 14)) & mask;
-  out[3] = ((w2 >> 43) | (w3 << 21)) & mask;
-  out[4] = ((w3 >> 36) | (w4 << 28)) & mask;
-  out[5] = ((w4 >> 29) | (w5 << 35)) & mask;
-  out[6] = ((w5 >> 22) | (w6 << 42)) & mask;
-  out[7] = ((w6 >> 15) | (w7 << 49)) & mask;
-  out[8] = ((w7 >> 8) | (w8 << 56)) & mask;
-  out[9] = (w8 >> 1) & mask;
-  out[10] = ((w8 >> 58) | (w9 << 6)) & mask;
-  out[11] = ((w9 >> 51) | (w10 << 13)) & mask;
-  out[12] = ((w10 >> 44) | (w11 << 20)) & mask;
-  out[13] = ((w11 >> 37) | (w12 << 27)) & mask;
-  out[14] = ((w12 >> 30) | (w13 << 34)) & mask;
-  out[15] = ((w13 >> 23) | (w14 << 41)) & mask;
-  out[16] = ((w14 >> 16) | (w15 << 48)) & mask;
-  out[17] = ((w15 >> 9) | (w16 << 55)) & mask;
-  out[18] = (w16 >> 2) & mask;
-  out[19] = ((w16 >> 59) | (w17 << 5)) & mask;
-  out[20] = ((w17 >> 52) | (w18 << 12)) & mask;
-  out[21] = ((w18 >> 45) | (w19 << 19)) & mask;
-  out[22] = ((w19 >> 38) | (w20 << 26)) & mask;
-  out[23] = ((w20 >> 31) | (w21 << 33)) & mask;
-  out[24] = ((w21 >> 24) | (w22 << 40)) & mask;
-  out[25] = ((w22 >> 17) | (w23 << 47)) & mask;
-  out[26] = ((w23 >> 10) | (w24 << 54)) & mask;
-  out[27] = (w24 >> 3) & mask;
-  out[28] = ((w24 >> 60) | (w25 << 4)) & mask;
-  out[29] = ((w25 >> 53) | (w26 << 11)) & mask;
-  out[30] = ((w26 >> 46) | (w27 << 18)) & mask;
-  out[31] = ((w27 >> 39) | (w28 << 25)) & mask;
-
-  return in;
-}
-
-inline const uint8_t* unpack58_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 288230376151711743ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint64_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 8;
-  uint64_t w9 = util::SafeLoadAs<uint64_t>(in);
-  w9 = arrow::bit_util::FromLittleEndian(w9);
-  in += 8;
-  uint64_t w10 = util::SafeLoadAs<uint64_t>(in);
-  w10 = arrow::bit_util::FromLittleEndian(w10);
-  in += 8;
-  uint64_t w11 = util::SafeLoadAs<uint64_t>(in);
-  w11 = arrow::bit_util::FromLittleEndian(w11);
-  in += 8;
-  uint64_t w12 = util::SafeLoadAs<uint64_t>(in);
-  w12 = arrow::bit_util::FromLittleEndian(w12);
-  in += 8;
-  uint64_t w13 = util::SafeLoadAs<uint64_t>(in);
-  w13 = arrow::bit_util::FromLittleEndian(w13);
-  in += 8;
-  uint64_t w14 = util::SafeLoadAs<uint64_t>(in);
-  w14 = arrow::bit_util::FromLittleEndian(w14);
-  in += 8;
-  uint64_t w15 = util::SafeLoadAs<uint64_t>(in);
-  w15 = arrow::bit_util::FromLittleEndian(w15);
-  in += 8;
-  uint64_t w16 = util::SafeLoadAs<uint64_t>(in);
-  w16 = arrow::bit_util::FromLittleEndian(w16);
-  in += 8;
-  uint64_t w17 = util::SafeLoadAs<uint64_t>(in);
-  w17 = arrow::bit_util::FromLittleEndian(w17);
-  in += 8;
-  uint64_t w18 = util::SafeLoadAs<uint64_t>(in);
-  w18 = arrow::bit_util::FromLittleEndian(w18);
-  in += 8;
-  uint64_t w19 = util::SafeLoadAs<uint64_t>(in);
-  w19 = arrow::bit_util::FromLittleEndian(w19);
-  in += 8;
-  uint64_t w20 = util::SafeLoadAs<uint64_t>(in);
-  w20 = arrow::bit_util::FromLittleEndian(w20);
-  in += 8;
-  uint64_t w21 = util::SafeLoadAs<uint64_t>(in);
-  w21 = arrow::bit_util::FromLittleEndian(w21);
-  in += 8;
-  uint64_t w22 = util::SafeLoadAs<uint64_t>(in);
-  w22 = arrow::bit_util::FromLittleEndian(w22);
-  in += 8;
-  uint64_t w23 = util::SafeLoadAs<uint64_t>(in);
-  w23 = arrow::bit_util::FromLittleEndian(w23);
-  in += 8;
-  uint64_t w24 = util::SafeLoadAs<uint64_t>(in);
-  w24 = arrow::bit_util::FromLittleEndian(w24);
-  in += 8;
-  uint64_t w25 = util::SafeLoadAs<uint64_t>(in);
-  w25 = arrow::bit_util::FromLittleEndian(w25);
-  in += 8;
-  uint64_t w26 = util::SafeLoadAs<uint64_t>(in);
-  w26 = arrow::bit_util::FromLittleEndian(w26);
-  in += 8;
-  uint64_t w27 = util::SafeLoadAs<uint64_t>(in);
-  w27 = arrow::bit_util::FromLittleEndian(w27);
-  in += 8;
-  uint64_t w28 = util::SafeLoadAs<uint64_t>(in);
-  w28 = arrow::bit_util::FromLittleEndian(w28);
-  in += 8;
-  out[0] = (w0)&mask;
-  out[1] = ((w0 >> 58) | (w1 << 6)) & mask;
-  out[2] = ((w1 >> 52) | (w2 << 12)) & mask;
-  out[3] = ((w2 >> 46) | (w3 << 18)) & mask;
-  out[4] = ((w3 >> 40) | (w4 << 24)) & mask;
-  out[5] = ((w4 >> 34) | (w5 << 30)) & mask;
-  out[6] = ((w5 >> 28) | (w6 << 36)) & mask;
-  out[7] = ((w6 >> 22) | (w7 << 42)) & mask;
-  out[8] = ((w7 >> 16) | (w8 << 48)) & mask;
-  out[9] = ((w8 >> 10) | (w9 << 54)) & mask;
-  out[10] = (w9 >> 4) & mask;
-  out[11] = ((w9 >> 62) | (w10 << 2)) & mask;
-  out[12] = ((w10 >> 56) | (w11 << 8)) & mask;
-  out[13] = ((w11 >> 50) | (w12 << 14)) & mask;
-  out[14] = ((w12 >> 44) | (w13 << 20)) & mask;
-  out[15] = ((w13 >> 38) | (w14 << 26)) & mask;
-  out[16] = ((w14 >> 32) | (w15 << 32)) & mask;
-  out[17] = ((w15 >> 26) | (w16 << 38)) & mask;
-  out[18] = ((w16 >> 20) | (w17 << 44)) & mask;
-  out[19] = ((w17 >> 14) | (w18 << 50)) & mask;
-  out[20] = ((w18 >> 8) | (w19 << 56)) & mask;
-  out[21] = (w19 >> 2) & mask;
-  out[22] = ((w19 >> 60) | (w20 << 4)) & mask;
-  out[23] = ((w20 >> 54) | (w21 << 10)) & mask;
-  out[24] = ((w21 >> 48) | (w22 << 16)) & mask;
-  out[25] = ((w22 >> 42) | (w23 << 22)) & mask;
-  out[26] = ((w23 >> 36) | (w24 << 28)) & mask;
-  out[27] = ((w24 >> 30) | (w25 << 34)) & mask;
-  out[28] = ((w25 >> 24) | (w26 << 40)) & mask;
-  out[29] = ((w26 >> 18) | (w27 << 46)) & mask;
-  out[30] = ((w27 >> 12) | (w28 << 52)) & mask;
-  out[31] = w28 >> 6;
-
-  return in;
-}
-
-inline const uint8_t* unpack59_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 576460752303423487ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint64_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 8;
-  uint64_t w9 = util::SafeLoadAs<uint64_t>(in);
-  w9 = arrow::bit_util::FromLittleEndian(w9);
-  in += 8;
-  uint64_t w10 = util::SafeLoadAs<uint64_t>(in);
-  w10 = arrow::bit_util::FromLittleEndian(w10);
-  in += 8;
-  uint64_t w11 = util::SafeLoadAs<uint64_t>(in);
-  w11 = arrow::bit_util::FromLittleEndian(w11);
-  in += 8;
-  uint64_t w12 = util::SafeLoadAs<uint64_t>(in);
-  w12 = arrow::bit_util::FromLittleEndian(w12);
-  in += 8;
-  uint64_t w13 = util::SafeLoadAs<uint64_t>(in);
-  w13 = arrow::bit_util::FromLittleEndian(w13);
-  in += 8;
-  uint64_t w14 = util::SafeLoadAs<uint64_t>(in);
-  w14 = arrow::bit_util::FromLittleEndian(w14);
-  in += 8;
-  uint64_t w15 = util::SafeLoadAs<uint64_t>(in);
-  w15 = arrow::bit_util::FromLittleEndian(w15);
-  in += 8;
-  uint64_t w16 = util::SafeLoadAs<uint64_t>(in);
-  w16 = arrow::bit_util::FromLittleEndian(w16);
-  in += 8;
-  uint64_t w17 = util::SafeLoadAs<uint64_t>(in);
-  w17 = arrow::bit_util::FromLittleEndian(w17);
-  in += 8;
-  uint64_t w18 = util::SafeLoadAs<uint64_t>(in);
-  w18 = arrow::bit_util::FromLittleEndian(w18);
-  in += 8;
-  uint64_t w19 = util::SafeLoadAs<uint64_t>(in);
-  w19 = arrow::bit_util::FromLittleEndian(w19);
-  in += 8;
-  uint64_t w20 = util::SafeLoadAs<uint64_t>(in);
-  w20 = arrow::bit_util::FromLittleEndian(w20);
-  in += 8;
-  uint64_t w21 = util::SafeLoadAs<uint64_t>(in);
-  w21 = arrow::bit_util::FromLittleEndian(w21);
-  in += 8;
-  uint64_t w22 = util::SafeLoadAs<uint64_t>(in);
-  w22 = arrow::bit_util::FromLittleEndian(w22);
-  in += 8;
-  uint64_t w23 = util::SafeLoadAs<uint64_t>(in);
-  w23 = arrow::bit_util::FromLittleEndian(w23);
-  in += 8;
-  uint64_t w24 = util::SafeLoadAs<uint64_t>(in);
-  w24 = arrow::bit_util::FromLittleEndian(w24);
-  in += 8;
-  uint64_t w25 = util::SafeLoadAs<uint64_t>(in);
-  w25 = arrow::bit_util::FromLittleEndian(w25);
-  in += 8;
-  uint64_t w26 = util::SafeLoadAs<uint64_t>(in);
-  w26 = arrow::bit_util::FromLittleEndian(w26);
-  in += 8;
-  uint64_t w27 = util::SafeLoadAs<uint64_t>(in);
-  w27 = arrow::bit_util::FromLittleEndian(w27);
-  in += 8;
-  uint64_t w28 = util::SafeLoadAs<uint64_t>(in);
-  w28 = arrow::bit_util::FromLittleEndian(w28);
-  in += 8;
-  uint64_t w29 = util::SafeLoadAs<uint32_t>(in);
-  w29 = arrow::bit_util::FromLittleEndian(w29);
-  in += 4;
-  out[0] = (w0)&mask;
-  out[1] = ((w0 >> 59) | (w1 << 5)) & mask;
-  out[2] = ((w1 >> 54) | (w2 << 10)) & mask;
-  out[3] = ((w2 >> 49) | (w3 << 15)) & mask;
-  out[4] = ((w3 >> 44) | (w4 << 20)) & mask;
-  out[5] = ((w4 >> 39) | (w5 << 25)) & mask;
-  out[6] = ((w5 >> 34) | (w6 << 30)) & mask;
-  out[7] = ((w6 >> 29) | (w7 << 35)) & mask;
-  out[8] = ((w7 >> 24) | (w8 << 40)) & mask;
-  out[9] = ((w8 >> 19) | (w9 << 45)) & mask;
-  out[10] = ((w9 >> 14) | (w10 << 50)) & mask;
-  out[11] = ((w10 >> 9) | (w11 << 55)) & mask;
-  out[12] = (w11 >> 4) & mask;
-  out[13] = ((w11 >> 63) | (w12 << 1)) & mask;
-  out[14] = ((w12 >> 58) | (w13 << 6)) & mask;
-  out[15] = ((w13 >> 53) | (w14 << 11)) & mask;
-  out[16] = ((w14 >> 48) | (w15 << 16)) & mask;
-  out[17] = ((w15 >> 43) | (w16 << 21)) & mask;
-  out[18] = ((w16 >> 38) | (w17 << 26)) & mask;
-  out[19] = ((w17 >> 33) | (w18 << 31)) & mask;
-  out[20] = ((w18 >> 28) | (w19 << 36)) & mask;
-  out[21] = ((w19 >> 23) | (w20 << 41)) & mask;
-  out[22] = ((w20 >> 18) | (w21 << 46)) & mask;
-  out[23] = ((w21 >> 13) | (w22 << 51)) & mask;
-  out[24] = ((w22 >> 8) | (w23 << 56)) & mask;
-  out[25] = (w23 >> 3) & mask;
-  out[26] = ((w23 >> 62) | (w24 << 2)) & mask;
-  out[27] = ((w24 >> 57) | (w25 << 7)) & mask;
-  out[28] = ((w25 >> 52) | (w26 << 12)) & mask;
-  out[29] = ((w26 >> 47) | (w27 << 17)) & mask;
-  out[30] = ((w27 >> 42) | (w28 << 22)) & mask;
-  out[31] = ((w28 >> 37) | (w29 << 27)) & mask;
-
-  return in;
-}
-
-inline const uint8_t* unpack60_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 1152921504606846975ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint64_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 8;
-  uint64_t w9 = util::SafeLoadAs<uint64_t>(in);
-  w9 = arrow::bit_util::FromLittleEndian(w9);
-  in += 8;
-  uint64_t w10 = util::SafeLoadAs<uint64_t>(in);
-  w10 = arrow::bit_util::FromLittleEndian(w10);
-  in += 8;
-  uint64_t w11 = util::SafeLoadAs<uint64_t>(in);
-  w11 = arrow::bit_util::FromLittleEndian(w11);
-  in += 8;
-  uint64_t w12 = util::SafeLoadAs<uint64_t>(in);
-  w12 = arrow::bit_util::FromLittleEndian(w12);
-  in += 8;
-  uint64_t w13 = util::SafeLoadAs<uint64_t>(in);
-  w13 = arrow::bit_util::FromLittleEndian(w13);
-  in += 8;
-  uint64_t w14 = util::SafeLoadAs<uint64_t>(in);
-  w14 = arrow::bit_util::FromLittleEndian(w14);
-  in += 8;
-  uint64_t w15 = util::SafeLoadAs<uint64_t>(in);
-  w15 = arrow::bit_util::FromLittleEndian(w15);
-  in += 8;
-  uint64_t w16 = util::SafeLoadAs<uint64_t>(in);
-  w16 = arrow::bit_util::FromLittleEndian(w16);
-  in += 8;
-  uint64_t w17 = util::SafeLoadAs<uint64_t>(in);
-  w17 = arrow::bit_util::FromLittleEndian(w17);
-  in += 8;
-  uint64_t w18 = util::SafeLoadAs<uint64_t>(in);
-  w18 = arrow::bit_util::FromLittleEndian(w18);
-  in += 8;
-  uint64_t w19 = util::SafeLoadAs<uint64_t>(in);
-  w19 = arrow::bit_util::FromLittleEndian(w19);
-  in += 8;
-  uint64_t w20 = util::SafeLoadAs<uint64_t>(in);
-  w20 = arrow::bit_util::FromLittleEndian(w20);
-  in += 8;
-  uint64_t w21 = util::SafeLoadAs<uint64_t>(in);
-  w21 = arrow::bit_util::FromLittleEndian(w21);
-  in += 8;
-  uint64_t w22 = util::SafeLoadAs<uint64_t>(in);
-  w22 = arrow::bit_util::FromLittleEndian(w22);
-  in += 8;
-  uint64_t w23 = util::SafeLoadAs<uint64_t>(in);
-  w23 = arrow::bit_util::FromLittleEndian(w23);
-  in += 8;
-  uint64_t w24 = util::SafeLoadAs<uint64_t>(in);
-  w24 = arrow::bit_util::FromLittleEndian(w24);
-  in += 8;
-  uint64_t w25 = util::SafeLoadAs<uint64_t>(in);
-  w25 = arrow::bit_util::FromLittleEndian(w25);
-  in += 8;
-  uint64_t w26 = util::SafeLoadAs<uint64_t>(in);
-  w26 = arrow::bit_util::FromLittleEndian(w26);
-  in += 8;
-  uint64_t w27 = util::SafeLoadAs<uint64_t>(in);
-  w27 = arrow::bit_util::FromLittleEndian(w27);
-  in += 8;
-  uint64_t w28 = util::SafeLoadAs<uint64_t>(in);
-  w28 = arrow::bit_util::FromLittleEndian(w28);
-  in += 8;
-  uint64_t w29 = util::SafeLoadAs<uint64_t>(in);
-  w29 = arrow::bit_util::FromLittleEndian(w29);
-  in += 8;
-  out[0] = (w0)&mask;
-  out[1] = ((w0 >> 60) | (w1 << 4)) & mask;
-  out[2] = ((w1 >> 56) | (w2 << 8)) & mask;
-  out[3] = ((w2 >> 52) | (w3 << 12)) & mask;
-  out[4] = ((w3 >> 48) | (w4 << 16)) & mask;
-  out[5] = ((w4 >> 44) | (w5 << 20)) & mask;
-  out[6] = ((w5 >> 40) | (w6 << 24)) & mask;
-  out[7] = ((w6 >> 36) | (w7 << 28)) & mask;
-  out[8] = ((w7 >> 32) | (w8 << 32)) & mask;
-  out[9] = ((w8 >> 28) | (w9 << 36)) & mask;
-  out[10] = ((w9 >> 24) | (w10 << 40)) & mask;
-  out[11] = ((w10 >> 20) | (w11 << 44)) & mask;
-  out[12] = ((w11 >> 16) | (w12 << 48)) & mask;
-  out[13] = ((w12 >> 12) | (w13 << 52)) & mask;
-  out[14] = ((w13 >> 8) | (w14 << 56)) & mask;
-  out[15] = w14 >> 4;
-  out[16] = (w15)&mask;
-  out[17] = ((w15 >> 60) | (w16 << 4)) & mask;
-  out[18] = ((w16 >> 56) | (w17 << 8)) & mask;
-  out[19] = ((w17 >> 52) | (w18 << 12)) & mask;
-  out[20] = ((w18 >> 48) | (w19 << 16)) & mask;
-  out[21] = ((w19 >> 44) | (w20 << 20)) & mask;
-  out[22] = ((w20 >> 40) | (w21 << 24)) & mask;
-  out[23] = ((w21 >> 36) | (w22 << 28)) & mask;
-  out[24] = ((w22 >> 32) | (w23 << 32)) & mask;
-  out[25] = ((w23 >> 28) | (w24 << 36)) & mask;
-  out[26] = ((w24 >> 24) | (w25 << 40)) & mask;
-  out[27] = ((w25 >> 20) | (w26 << 44)) & mask;
-  out[28] = ((w26 >> 16) | (w27 << 48)) & mask;
-  out[29] = ((w27 >> 12) | (w28 << 52)) & mask;
-  out[30] = ((w28 >> 8) | (w29 << 56)) & mask;
-  out[31] = w29 >> 4;
-
-  return in;
-}
-
-inline const uint8_t* unpack61_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 2305843009213693951ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint64_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 8;
-  uint64_t w9 = util::SafeLoadAs<uint64_t>(in);
-  w9 = arrow::bit_util::FromLittleEndian(w9);
-  in += 8;
-  uint64_t w10 = util::SafeLoadAs<uint64_t>(in);
-  w10 = arrow::bit_util::FromLittleEndian(w10);
-  in += 8;
-  uint64_t w11 = util::SafeLoadAs<uint64_t>(in);
-  w11 = arrow::bit_util::FromLittleEndian(w11);
-  in += 8;
-  uint64_t w12 = util::SafeLoadAs<uint64_t>(in);
-  w12 = arrow::bit_util::FromLittleEndian(w12);
-  in += 8;
-  uint64_t w13 = util::SafeLoadAs<uint64_t>(in);
-  w13 = arrow::bit_util::FromLittleEndian(w13);
-  in += 8;
-  uint64_t w14 = util::SafeLoadAs<uint64_t>(in);
-  w14 = arrow::bit_util::FromLittleEndian(w14);
-  in += 8;
-  uint64_t w15 = util::SafeLoadAs<uint64_t>(in);
-  w15 = arrow::bit_util::FromLittleEndian(w15);
-  in += 8;
-  uint64_t w16 = util::SafeLoadAs<uint64_t>(in);
-  w16 = arrow::bit_util::FromLittleEndian(w16);
-  in += 8;
-  uint64_t w17 = util::SafeLoadAs<uint64_t>(in);
-  w17 = arrow::bit_util::FromLittleEndian(w17);
-  in += 8;
-  uint64_t w18 = util::SafeLoadAs<uint64_t>(in);
-  w18 = arrow::bit_util::FromLittleEndian(w18);
-  in += 8;
-  uint64_t w19 = util::SafeLoadAs<uint64_t>(in);
-  w19 = arrow::bit_util::FromLittleEndian(w19);
-  in += 8;
-  uint64_t w20 = util::SafeLoadAs<uint64_t>(in);
-  w20 = arrow::bit_util::FromLittleEndian(w20);
-  in += 8;
-  uint64_t w21 = util::SafeLoadAs<uint64_t>(in);
-  w21 = arrow::bit_util::FromLittleEndian(w21);
-  in += 8;
-  uint64_t w22 = util::SafeLoadAs<uint64_t>(in);
-  w22 = arrow::bit_util::FromLittleEndian(w22);
-  in += 8;
-  uint64_t w23 = util::SafeLoadAs<uint64_t>(in);
-  w23 = arrow::bit_util::FromLittleEndian(w23);
-  in += 8;
-  uint64_t w24 = util::SafeLoadAs<uint64_t>(in);
-  w24 = arrow::bit_util::FromLittleEndian(w24);
-  in += 8;
-  uint64_t w25 = util::SafeLoadAs<uint64_t>(in);
-  w25 = arrow::bit_util::FromLittleEndian(w25);
-  in += 8;
-  uint64_t w26 = util::SafeLoadAs<uint64_t>(in);
-  w26 = arrow::bit_util::FromLittleEndian(w26);
-  in += 8;
-  uint64_t w27 = util::SafeLoadAs<uint64_t>(in);
-  w27 = arrow::bit_util::FromLittleEndian(w27);
-  in += 8;
-  uint64_t w28 = util::SafeLoadAs<uint64_t>(in);
-  w28 = arrow::bit_util::FromLittleEndian(w28);
-  in += 8;
-  uint64_t w29 = util::SafeLoadAs<uint64_t>(in);
-  w29 = arrow::bit_util::FromLittleEndian(w29);
-  in += 8;
-  uint64_t w30 = util::SafeLoadAs<uint32_t>(in);
-  w30 = arrow::bit_util::FromLittleEndian(w30);
-  in += 4;
-  out[0] = (w0)&mask;
-  out[1] = ((w0 >> 61) | (w1 << 3)) & mask;
-  out[2] = ((w1 >> 58) | (w2 << 6)) & mask;
-  out[3] = ((w2 >> 55) | (w3 << 9)) & mask;
-  out[4] = ((w3 >> 52) | (w4 << 12)) & mask;
-  out[5] = ((w4 >> 49) | (w5 << 15)) & mask;
-  out[6] = ((w5 >> 46) | (w6 << 18)) & mask;
-  out[7] = ((w6 >> 43) | (w7 << 21)) & mask;
-  out[8] = ((w7 >> 40) | (w8 << 24)) & mask;
-  out[9] = ((w8 >> 37) | (w9 << 27)) & mask;
-  out[10] = ((w9 >> 34) | (w10 << 30)) & mask;
-  out[11] = ((w10 >> 31) | (w11 << 33)) & mask;
-  out[12] = ((w11 >> 28) | (w12 << 36)) & mask;
-  out[13] = ((w12 >> 25) | (w13 << 39)) & mask;
-  out[14] = ((w13 >> 22) | (w14 << 42)) & mask;
-  out[15] = ((w14 >> 19) | (w15 << 45)) & mask;
-  out[16] = ((w15 >> 16) | (w16 << 48)) & mask;
-  out[17] = ((w16 >> 13) | (w17 << 51)) & mask;
-  out[18] = ((w17 >> 10) | (w18 << 54)) & mask;
-  out[19] = ((w18 >> 7) | (w19 << 57)) & mask;
-  out[20] = ((w19 >> 4) | (w20 << 60)) & mask;
-  out[21] = (w20 >> 1) & mask;
-  out[22] = ((w20 >> 62) | (w21 << 2)) & mask;
-  out[23] = ((w21 >> 59) | (w22 << 5)) & mask;
-  out[24] = ((w22 >> 56) | (w23 << 8)) & mask;
-  out[25] = ((w23 >> 53) | (w24 << 11)) & mask;
-  out[26] = ((w24 >> 50) | (w25 << 14)) & mask;
-  out[27] = ((w25 >> 47) | (w26 << 17)) & mask;
-  out[28] = ((w26 >> 44) | (w27 << 20)) & mask;
-  out[29] = ((w27 >> 41) | (w28 << 23)) & mask;
-  out[30] = ((w28 >> 38) | (w29 << 26)) & mask;
-  out[31] = ((w29 >> 35) | (w30 << 29)) & mask;
-
-  return in;
-}
-
-inline const uint8_t* unpack62_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 4611686018427387903ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint64_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 8;
-  uint64_t w9 = util::SafeLoadAs<uint64_t>(in);
-  w9 = arrow::bit_util::FromLittleEndian(w9);
-  in += 8;
-  uint64_t w10 = util::SafeLoadAs<uint64_t>(in);
-  w10 = arrow::bit_util::FromLittleEndian(w10);
-  in += 8;
-  uint64_t w11 = util::SafeLoadAs<uint64_t>(in);
-  w11 = arrow::bit_util::FromLittleEndian(w11);
-  in += 8;
-  uint64_t w12 = util::SafeLoadAs<uint64_t>(in);
-  w12 = arrow::bit_util::FromLittleEndian(w12);
-  in += 8;
-  uint64_t w13 = util::SafeLoadAs<uint64_t>(in);
-  w13 = arrow::bit_util::FromLittleEndian(w13);
-  in += 8;
-  uint64_t w14 = util::SafeLoadAs<uint64_t>(in);
-  w14 = arrow::bit_util::FromLittleEndian(w14);
-  in += 8;
-  uint64_t w15 = util::SafeLoadAs<uint64_t>(in);
-  w15 = arrow::bit_util::FromLittleEndian(w15);
-  in += 8;
-  uint64_t w16 = util::SafeLoadAs<uint64_t>(in);
-  w16 = arrow::bit_util::FromLittleEndian(w16);
-  in += 8;
-  uint64_t w17 = util::SafeLoadAs<uint64_t>(in);
-  w17 = arrow::bit_util::FromLittleEndian(w17);
-  in += 8;
-  uint64_t w18 = util::SafeLoadAs<uint64_t>(in);
-  w18 = arrow::bit_util::FromLittleEndian(w18);
-  in += 8;
-  uint64_t w19 = util::SafeLoadAs<uint64_t>(in);
-  w19 = arrow::bit_util::FromLittleEndian(w19);
-  in += 8;
-  uint64_t w20 = util::SafeLoadAs<uint64_t>(in);
-  w20 = arrow::bit_util::FromLittleEndian(w20);
-  in += 8;
-  uint64_t w21 = util::SafeLoadAs<uint64_t>(in);
-  w21 = arrow::bit_util::FromLittleEndian(w21);
-  in += 8;
-  uint64_t w22 = util::SafeLoadAs<uint64_t>(in);
-  w22 = arrow::bit_util::FromLittleEndian(w22);
-  in += 8;
-  uint64_t w23 = util::SafeLoadAs<uint64_t>(in);
-  w23 = arrow::bit_util::FromLittleEndian(w23);
-  in += 8;
-  uint64_t w24 = util::SafeLoadAs<uint64_t>(in);
-  w24 = arrow::bit_util::FromLittleEndian(w24);
-  in += 8;
-  uint64_t w25 = util::SafeLoadAs<uint64_t>(in);
-  w25 = arrow::bit_util::FromLittleEndian(w25);
-  in += 8;
-  uint64_t w26 = util::SafeLoadAs<uint64_t>(in);
-  w26 = arrow::bit_util::FromLittleEndian(w26);
-  in += 8;
-  uint64_t w27 = util::SafeLoadAs<uint64_t>(in);
-  w27 = arrow::bit_util::FromLittleEndian(w27);
-  in += 8;
-  uint64_t w28 = util::SafeLoadAs<uint64_t>(in);
-  w28 = arrow::bit_util::FromLittleEndian(w28);
-  in += 8;
-  uint64_t w29 = util::SafeLoadAs<uint64_t>(in);
-  w29 = arrow::bit_util::FromLittleEndian(w29);
-  in += 8;
-  uint64_t w30 = util::SafeLoadAs<uint64_t>(in);
-  w30 = arrow::bit_util::FromLittleEndian(w30);
-  in += 8;
-  out[0] = (w0)&mask;
-  out[1] = ((w0 >> 62) | (w1 << 2)) & mask;
-  out[2] = ((w1 >> 60) | (w2 << 4)) & mask;
-  out[3] = ((w2 >> 58) | (w3 << 6)) & mask;
-  out[4] = ((w3 >> 56) | (w4 << 8)) & mask;
-  out[5] = ((w4 >> 54) | (w5 << 10)) & mask;
-  out[6] = ((w5 >> 52) | (w6 << 12)) & mask;
-  out[7] = ((w6 >> 50) | (w7 << 14)) & mask;
-  out[8] = ((w7 >> 48) | (w8 << 16)) & mask;
-  out[9] = ((w8 >> 46) | (w9 << 18)) & mask;
-  out[10] = ((w9 >> 44) | (w10 << 20)) & mask;
-  out[11] = ((w10 >> 42) | (w11 << 22)) & mask;
-  out[12] = ((w11 >> 40) | (w12 << 24)) & mask;
-  out[13] = ((w12 >> 38) | (w13 << 26)) & mask;
-  out[14] = ((w13 >> 36) | (w14 << 28)) & mask;
-  out[15] = ((w14 >> 34) | (w15 << 30)) & mask;
-  out[16] = ((w15 >> 32) | (w16 << 32)) & mask;
-  out[17] = ((w16 >> 30) | (w17 << 34)) & mask;
-  out[18] = ((w17 >> 28) | (w18 << 36)) & mask;
-  out[19] = ((w18 >> 26) | (w19 << 38)) & mask;
-  out[20] = ((w19 >> 24) | (w20 << 40)) & mask;
-  out[21] = ((w20 >> 22) | (w21 << 42)) & mask;
-  out[22] = ((w21 >> 20) | (w22 << 44)) & mask;
-  out[23] = ((w22 >> 18) | (w23 << 46)) & mask;
-  out[24] = ((w23 >> 16) | (w24 << 48)) & mask;
-  out[25] = ((w24 >> 14) | (w25 << 50)) & mask;
-  out[26] = ((w25 >> 12) | (w26 << 52)) & mask;
-  out[27] = ((w26 >> 10) | (w27 << 54)) & mask;
-  out[28] = ((w27 >> 8) | (w28 << 56)) & mask;
-  out[29] = ((w28 >> 6) | (w29 << 58)) & mask;
-  out[30] = ((w29 >> 4) | (w30 << 60)) & mask;
-  out[31] = w30 >> 2;
-
-  return in;
-}
-
-inline const uint8_t* unpack63_64(const uint8_t* in, uint64_t* out) {
-  const uint64_t mask = 9223372036854775807ULL;
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint64_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 8;
-  uint64_t w9 = util::SafeLoadAs<uint64_t>(in);
-  w9 = arrow::bit_util::FromLittleEndian(w9);
-  in += 8;
-  uint64_t w10 = util::SafeLoadAs<uint64_t>(in);
-  w10 = arrow::bit_util::FromLittleEndian(w10);
-  in += 8;
-  uint64_t w11 = util::SafeLoadAs<uint64_t>(in);
-  w11 = arrow::bit_util::FromLittleEndian(w11);
-  in += 8;
-  uint64_t w12 = util::SafeLoadAs<uint64_t>(in);
-  w12 = arrow::bit_util::FromLittleEndian(w12);
-  in += 8;
-  uint64_t w13 = util::SafeLoadAs<uint64_t>(in);
-  w13 = arrow::bit_util::FromLittleEndian(w13);
-  in += 8;
-  uint64_t w14 = util::SafeLoadAs<uint64_t>(in);
-  w14 = arrow::bit_util::FromLittleEndian(w14);
-  in += 8;
-  uint64_t w15 = util::SafeLoadAs<uint64_t>(in);
-  w15 = arrow::bit_util::FromLittleEndian(w15);
-  in += 8;
-  uint64_t w16 = util::SafeLoadAs<uint64_t>(in);
-  w16 = arrow::bit_util::FromLittleEndian(w16);
-  in += 8;
-  uint64_t w17 = util::SafeLoadAs<uint64_t>(in);
-  w17 = arrow::bit_util::FromLittleEndian(w17);
-  in += 8;
-  uint64_t w18 = util::SafeLoadAs<uint64_t>(in);
-  w18 = arrow::bit_util::FromLittleEndian(w18);
-  in += 8;
-  uint64_t w19 = util::SafeLoadAs<uint64_t>(in);
-  w19 = arrow::bit_util::FromLittleEndian(w19);
-  in += 8;
-  uint64_t w20 = util::SafeLoadAs<uint64_t>(in);
-  w20 = arrow::bit_util::FromLittleEndian(w20);
-  in += 8;
-  uint64_t w21 = util::SafeLoadAs<uint64_t>(in);
-  w21 = arrow::bit_util::FromLittleEndian(w21);
-  in += 8;
-  uint64_t w22 = util::SafeLoadAs<uint64_t>(in);
-  w22 = arrow::bit_util::FromLittleEndian(w22);
-  in += 8;
-  uint64_t w23 = util::SafeLoadAs<uint64_t>(in);
-  w23 = arrow::bit_util::FromLittleEndian(w23);
-  in += 8;
-  uint64_t w24 = util::SafeLoadAs<uint64_t>(in);
-  w24 = arrow::bit_util::FromLittleEndian(w24);
-  in += 8;
-  uint64_t w25 = util::SafeLoadAs<uint64_t>(in);
-  w25 = arrow::bit_util::FromLittleEndian(w25);
-  in += 8;
-  uint64_t w26 = util::SafeLoadAs<uint64_t>(in);
-  w26 = arrow::bit_util::FromLittleEndian(w26);
-  in += 8;
-  uint64_t w27 = util::SafeLoadAs<uint64_t>(in);
-  w27 = arrow::bit_util::FromLittleEndian(w27);
-  in += 8;
-  uint64_t w28 = util::SafeLoadAs<uint64_t>(in);
-  w28 = arrow::bit_util::FromLittleEndian(w28);
-  in += 8;
-  uint64_t w29 = util::SafeLoadAs<uint64_t>(in);
-  w29 = arrow::bit_util::FromLittleEndian(w29);
-  in += 8;
-  uint64_t w30 = util::SafeLoadAs<uint64_t>(in);
-  w30 = arrow::bit_util::FromLittleEndian(w30);
-  in += 8;
-  uint64_t w31 = util::SafeLoadAs<uint32_t>(in);
-  w31 = arrow::bit_util::FromLittleEndian(w31);
-  in += 4;
-  out[0] = (w0)&mask;
-  out[1] = ((w0 >> 63) | (w1 << 1)) & mask;
-  out[2] = ((w1 >> 62) | (w2 << 2)) & mask;
-  out[3] = ((w2 >> 61) | (w3 << 3)) & mask;
-  out[4] = ((w3 >> 60) | (w4 << 4)) & mask;
-  out[5] = ((w4 >> 59) | (w5 << 5)) & mask;
-  out[6] = ((w5 >> 58) | (w6 << 6)) & mask;
-  out[7] = ((w6 >> 57) | (w7 << 7)) & mask;
-  out[8] = ((w7 >> 56) | (w8 << 8)) & mask;
-  out[9] = ((w8 >> 55) | (w9 << 9)) & mask;
-  out[10] = ((w9 >> 54) | (w10 << 10)) & mask;
-  out[11] = ((w10 >> 53) | (w11 << 11)) & mask;
-  out[12] = ((w11 >> 52) | (w12 << 12)) & mask;
-  out[13] = ((w12 >> 51) | (w13 << 13)) & mask;
-  out[14] = ((w13 >> 50) | (w14 << 14)) & mask;
-  out[15] = ((w14 >> 49) | (w15 << 15)) & mask;
-  out[16] = ((w15 >> 48) | (w16 << 16)) & mask;
-  out[17] = ((w16 >> 47) | (w17 << 17)) & mask;
-  out[18] = ((w17 >> 46) | (w18 << 18)) & mask;
-  out[19] = ((w18 >> 45) | (w19 << 19)) & mask;
-  out[20] = ((w19 >> 44) | (w20 << 20)) & mask;
-  out[21] = ((w20 >> 43) | (w21 << 21)) & mask;
-  out[22] = ((w21 >> 42) | (w22 << 22)) & mask;
-  out[23] = ((w22 >> 41) | (w23 << 23)) & mask;
-  out[24] = ((w23 >> 40) | (w24 << 24)) & mask;
-  out[25] = ((w24 >> 39) | (w25 << 25)) & mask;
-  out[26] = ((w25 >> 38) | (w26 << 26)) & mask;
-  out[27] = ((w26 >> 37) | (w27 << 27)) & mask;
-  out[28] = ((w27 >> 36) | (w28 << 28)) & mask;
-  out[29] = ((w28 >> 35) | (w29 << 29)) & mask;
-  out[30] = ((w29 >> 34) | (w30 << 30)) & mask;
-  out[31] = ((w30 >> 33) | (w31 << 31)) & mask;
-
-  return in;
-}
-
-inline const uint8_t* unpack64_64(const uint8_t* in, uint64_t* out) {
-  uint64_t w0 = util::SafeLoadAs<uint64_t>(in);
-  w0 = arrow::bit_util::FromLittleEndian(w0);
-  in += 8;
-  uint64_t w1 = util::SafeLoadAs<uint64_t>(in);
-  w1 = arrow::bit_util::FromLittleEndian(w1);
-  in += 8;
-  uint64_t w2 = util::SafeLoadAs<uint64_t>(in);
-  w2 = arrow::bit_util::FromLittleEndian(w2);
-  in += 8;
-  uint64_t w3 = util::SafeLoadAs<uint64_t>(in);
-  w3 = arrow::bit_util::FromLittleEndian(w3);
-  in += 8;
-  uint64_t w4 = util::SafeLoadAs<uint64_t>(in);
-  w4 = arrow::bit_util::FromLittleEndian(w4);
-  in += 8;
-  uint64_t w5 = util::SafeLoadAs<uint64_t>(in);
-  w5 = arrow::bit_util::FromLittleEndian(w5);
-  in += 8;
-  uint64_t w6 = util::SafeLoadAs<uint64_t>(in);
-  w6 = arrow::bit_util::FromLittleEndian(w6);
-  in += 8;
-  uint64_t w7 = util::SafeLoadAs<uint64_t>(in);
-  w7 = arrow::bit_util::FromLittleEndian(w7);
-  in += 8;
-  uint64_t w8 = util::SafeLoadAs<uint64_t>(in);
-  w8 = arrow::bit_util::FromLittleEndian(w8);
-  in += 8;
-  uint64_t w9 = util::SafeLoadAs<uint64_t>(in);
-  w9 = arrow::bit_util::FromLittleEndian(w9);
-  in += 8;
-  uint64_t w10 = util::SafeLoadAs<uint64_t>(in);
-  w10 = arrow::bit_util::FromLittleEndian(w10);
-  in += 8;
-  uint64_t w11 = util::SafeLoadAs<uint64_t>(in);
-  w11 = arrow::bit_util::FromLittleEndian(w11);
-  in += 8;
-  uint64_t w12 = util::SafeLoadAs<uint64_t>(in);
-  w12 = arrow::bit_util::FromLittleEndian(w12);
-  in += 8;
-  uint64_t w13 = util::SafeLoadAs<uint64_t>(in);
-  w13 = arrow::bit_util::FromLittleEndian(w13);
-  in += 8;
-  uint64_t w14 = util::SafeLoadAs<uint64_t>(in);
-  w14 = arrow::bit_util::FromLittleEndian(w14);
-  in += 8;
-  uint64_t w15 = util::SafeLoadAs<uint64_t>(in);
-  w15 = arrow::bit_util::FromLittleEndian(w15);
-  in += 8;
-  uint64_t w16 = util::SafeLoadAs<uint64_t>(in);
-  w16 = arrow::bit_util::FromLittleEndian(w16);
-  in += 8;
-  uint64_t w17 = util::SafeLoadAs<uint64_t>(in);
-  w17 = arrow::bit_util::FromLittleEndian(w17);
-  in += 8;
-  uint64_t w18 = util::SafeLoadAs<uint64_t>(in);
-  w18 = arrow::bit_util::FromLittleEndian(w18);
-  in += 8;
-  uint64_t w19 = util::SafeLoadAs<uint64_t>(in);
-  w19 = arrow::bit_util::FromLittleEndian(w19);
-  in += 8;
-  uint64_t w20 = util::SafeLoadAs<uint64_t>(in);
-  w20 = arrow::bit_util::FromLittleEndian(w20);
-  in += 8;
-  uint64_t w21 = util::SafeLoadAs<uint64_t>(in);
-  w21 = arrow::bit_util::FromLittleEndian(w21);
-  in += 8;
-  uint64_t w22 = util::SafeLoadAs<uint64_t>(in);
-  w22 = arrow::bit_util::FromLittleEndian(w22);
-  in += 8;
-  uint64_t w23 = util::SafeLoadAs<uint64_t>(in);
-  w23 = arrow::bit_util::FromLittleEndian(w23);
-  in += 8;
-  uint64_t w24 = util::SafeLoadAs<uint64_t>(in);
-  w24 = arrow::bit_util::FromLittleEndian(w24);
-  in += 8;
-  uint64_t w25 = util::SafeLoadAs<uint64_t>(in);
-  w25 = arrow::bit_util::FromLittleEndian(w25);
-  in += 8;
-  uint64_t w26 = util::SafeLoadAs<uint64_t>(in);
-  w26 = arrow::bit_util::FromLittleEndian(w26);
-  in += 8;
-  uint64_t w27 = util::SafeLoadAs<uint64_t>(in);
-  w27 = arrow::bit_util::FromLittleEndian(w27);
-  in += 8;
-  uint64_t w28 = util::SafeLoadAs<uint64_t>(in);
-  w28 = arrow::bit_util::FromLittleEndian(w28);
-  in += 8;
-  uint64_t w29 = util::SafeLoadAs<uint64_t>(in);
-  w29 = arrow::bit_util::FromLittleEndian(w29);
-  in += 8;
-  uint64_t w30 = util::SafeLoadAs<uint64_t>(in);
-  w30 = arrow::bit_util::FromLittleEndian(w30);
-  in += 8;
-  uint64_t w31 = util::SafeLoadAs<uint64_t>(in);
-  w31 = arrow::bit_util::FromLittleEndian(w31);
-  in += 8;
-  out[0] = w0;
-  out[1] = w1;
-  out[2] = w2;
-  out[3] = w3;
-  out[4] = w4;
-  out[5] = w5;
-  out[6] = w6;
-  out[7] = w7;
-  out[8] = w8;
-  out[9] = w9;
-  out[10] = w10;
-  out[11] = w11;
-  out[12] = w12;
-  out[13] = w13;
-  out[14] = w14;
-  out[15] = w15;
-  out[16] = w16;
-  out[17] = w17;
-  out[18] = w18;
-  out[19] = w19;
-  out[20] = w20;
-  out[21] = w21;
-  out[22] = w22;
-  out[23] = w23;
-  out[24] = w24;
-  out[25] = w25;
-  out[26] = w26;
-  out[27] = w27;
-  out[28] = w28;
-  out[29] = w29;
-  out[30] = w30;
-  out[31] = w31;
-
-  return in;
-}
-
-}  // namespace arrow::internal
diff --git a/cpp/src/arrow/util/bpacking_avx2_internal.h b/cpp/src/arrow/util/bpacking_avx2_internal.h
deleted file mode 100644
index b2c213fe2aac4..0000000000000
--- a/cpp/src/arrow/util/bpacking_avx2_internal.h
+++ /dev/null
@@ -1,29 +0,0 @@
-// Licensed to the Apache Software Foundation (ASF) under one
-// or more contributor license agreements.  See the NOTICE file
-// distributed with this work for additional information
-// regarding copyright ownership.  The ASF licenses this file
-// to you under the Apache License, Version 2.0 (the
-// "License"); you may not use this file except in compliance
-// with the License.  You may obtain a copy of the License at
-//
-//   http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing,
-// software distributed under the License is distributed on an
-// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
-// KIND, either express or implied.  See the License for the
-// specific language governing permissions and limitations
-// under the License.
-
-#pragma once
-
-#include "arrow/util/visibility.h"
-
-#include <cstdint>
-
-namespace arrow::internal {
-
-ARROW_EXPORT int unpack32_avx2(const uint8_t* in, uint32_t* out, int batch_size,
-                               int num_bits);
-
-}  // namespace arrow::internal
diff --git a/cpp/src/arrow/util/bpacking_avx512_internal.h b/cpp/src/arrow/util/bpacking_avx512_internal.h
deleted file mode 100644
index 847aa9814336f..0000000000000
--- a/cpp/src/arrow/util/bpacking_avx512_internal.h
+++ /dev/null
@@ -1,29 +0,0 @@
-// Licensed to the Apache Software Foundation (ASF) under one
-// or more contributor license agreements.  See the NOTICE file
-// distributed with this work for additional information
-// regarding copyright ownership.  The ASF licenses this file
-// to you under the Apache License, Version 2.0 (the
-// "License"); you may not use this file except in compliance
-// with the License.  You may obtain a copy of the License at
-//
-//   http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing,
-// software distributed under the License is distributed on an
-// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
-// KIND, either express or implied.  See the License for the
-// specific language governing permissions and limitations
-// under the License.
-
-#pragma once
-
-#include "arrow/util/visibility.h"
-
-#include <cstdint>
-
-namespace arrow::internal {
-
-ARROW_EXPORT int unpack32_avx512(const uint8_t* in, uint32_t* out, int batch_size,
-                                 int num_bits);
-
-}  // namespace arrow::internal
diff --git a/cpp/src/arrow/util/bpacking_benchmark.cc b/cpp/src/arrow/util/bpacking_benchmark.cc
index f0ac22910c657..f281fdf17d369 100644
--- a/cpp/src/arrow/util/bpacking_benchmark.cc
+++ b/cpp/src/arrow/util/bpacking_benchmark.cc
@@ -22,17 +22,12 @@
 
 #include "arrow/testing/util.h"
 #include "arrow/util/bpacking_internal.h"
+#include "arrow/util/bpacking_scalar_internal.h"
+#include "arrow/util/bpacking_simd_internal.h"
 
 #if defined(ARROW_HAVE_RUNTIME_AVX2)
-#  include "arrow/util/bpacking_avx2_internal.h"
 #  include "arrow/util/cpu_info.h"
 #endif
-#if defined(ARROW_HAVE_RUNTIME_AVX512)
-#  include "arrow/util/bpacking_avx512_internal.h"
-#endif
-#if defined(ARROW_HAVE_NEON)
-#  include "arrow/util/bpacking_neon_internal.h"
-#endif
 
 namespace arrow::internal {
 namespace {
@@ -102,8 +97,13 @@ void BM_Unpack(benchmark::State& state, bool aligned, UnpackFunc<Int> unpack, bo
 
 constexpr int32_t kMinRange = 64;
 constexpr int32_t kMaxRange = 32768;
+constexpr std::initializer_list<int64_t> kBitWidths16 = {1, 2, 8, 13};
 constexpr std::initializer_list<int64_t> kBitWidths32 = {1, 2, 8, 20};
 constexpr std::initializer_list<int64_t> kBitWidths64 = {1, 2, 8, 20, 47};
+static const std::vector<std::vector<int64_t>> kBitWidthsNumValues16 = {
+    kBitWidths16,
+    benchmark::CreateRange(kMinRange, kMaxRange, /*multi=*/32),
+};
 static const std::vector<std::vector<int64_t>> kBitWidthsNumValues32 = {
     kBitWidths32,
     benchmark::CreateRange(kMinRange, kMaxRange, /*multi=*/32),
@@ -113,6 +113,11 @@ static const std::vector<std::vector<int64_t>> kBitWidthsNumValues64 = {
     benchmark::CreateRange(kMinRange, kMaxRange, /*multi=*/32),
 };
 
+/// Nudge for MSVC template inside BENCHMARK_CAPTURE macro.
+void BM_UnpackUint16(benchmark::State& state, bool aligned, UnpackFunc<uint16_t> unpack,
+                     bool skip = false, std::string skip_msg = "") {
+  return BM_Unpack<uint16_t>(state, aligned, unpack, skip, std::move(skip_msg));
+}
 /// Nudge for MSVC template inside BENCHMARK_CAPTURE macro.
 void BM_UnpackUint32(benchmark::State& state, bool aligned, UnpackFunc<uint32_t> unpack,
                      bool skip = false, std::string skip_msg = "") {
@@ -124,38 +129,70 @@ void BM_UnpackUint64(benchmark::State& state, bool aligned, UnpackFunc<uint64_t>
   return BM_Unpack<uint64_t>(state, aligned, unpack, skip, std::move(skip_msg));
 }
 
-BENCHMARK_CAPTURE(BM_UnpackUint32, ScalarUnaligned, false, unpack32_scalar)
+BENCHMARK_CAPTURE(BM_UnpackUint16, ScalarUnaligned, false, &unpack_scalar<uint16_t>)
+    ->ArgsProduct(kBitWidthsNumValues16);
+BENCHMARK_CAPTURE(BM_UnpackUint32, ScalarUnaligned, false, &unpack_scalar<uint32_t>)
     ->ArgsProduct(kBitWidthsNumValues32);
-BENCHMARK_CAPTURE(BM_UnpackUint64, ScalarUnaligned, false, unpack64_scalar)
+BENCHMARK_CAPTURE(BM_UnpackUint64, ScalarUnaligned, false, &unpack_scalar<uint64_t>)
     ->ArgsProduct(kBitWidthsNumValues64);
 
+#if defined(ARROW_HAVE_SSE4_2)
+BENCHMARK_CAPTURE(BM_UnpackUint16, Sse42Unaligned, false, &unpack_sse4_2<uint16_t>)
+    ->ArgsProduct(kBitWidthsNumValues16);
+BENCHMARK_CAPTURE(BM_UnpackUint32, Sse42Unaligned, false, &unpack_sse4_2<uint32_t>)
+    ->ArgsProduct(kBitWidthsNumValues32);
+BENCHMARK_CAPTURE(BM_UnpackUint64, Sse42Unaligned, false, &unpack_sse4_2<uint64_t>)
+    ->ArgsProduct(kBitWidthsNumValues64);
+#endif
+
 #if defined(ARROW_HAVE_RUNTIME_AVX2)
-BENCHMARK_CAPTURE(BM_UnpackUint32, Avx2Unaligned, false, unpack32_avx2,
+BENCHMARK_CAPTURE(BM_UnpackUint16, Avx2Unaligned, false, &unpack_avx2<uint16_t>,
+                  !CpuInfo::GetInstance()->IsSupported(CpuInfo::AVX2),
+                  "Avx2 not available")
+    ->ArgsProduct(kBitWidthsNumValues16);
+BENCHMARK_CAPTURE(BM_UnpackUint32, Avx2Unaligned, false, &unpack_avx2<uint32_t>,
                   !CpuInfo::GetInstance()->IsSupported(CpuInfo::AVX2),
                   "Avx2 not available")
     ->ArgsProduct(kBitWidthsNumValues32);
+BENCHMARK_CAPTURE(BM_UnpackUint64, Avx2Unaligned, false, &unpack_avx2<uint64_t>,
+                  !CpuInfo::GetInstance()->IsSupported(CpuInfo::AVX2),
+                  "Avx2 not available")
+    ->ArgsProduct(kBitWidthsNumValues64);
 #endif
 
 #if defined(ARROW_HAVE_RUNTIME_AVX512)
-BENCHMARK_CAPTURE(BM_UnpackUint32, Avx512Unaligned, false, unpack32_avx512,
+BENCHMARK_CAPTURE(BM_UnpackUint32, Avx512Unaligned, false, &unpack_avx512<uint32_t>,
                   !CpuInfo::GetInstance()->IsSupported(CpuInfo::AVX512),
                   "Avx512 not available")
     ->ArgsProduct(kBitWidthsNumValues32);
+BENCHMARK_CAPTURE(BM_UnpackUint64, Avx512Unaligned, false, &unpack_avx512<uint64_t>,
+                  !CpuInfo::GetInstance()->IsSupported(CpuInfo::AVX512),
+                  "Avx512 not available")
+    ->ArgsProduct(kBitWidthsNumValues64);
 #endif
 
 #if defined(ARROW_HAVE_NEON)
-BENCHMARK_CAPTURE(BM_UnpackUint32, NeonUnaligned, false, unpack32_neon)
+BENCHMARK_CAPTURE(BM_UnpackUint16, NeonUnaligned, false, &unpack_neon<uint16_t>)
+    ->ArgsProduct(kBitWidthsNumValues16);
+BENCHMARK_CAPTURE(BM_UnpackUint32, NeonUnaligned, false, &unpack_neon<uint32_t>)
     ->ArgsProduct(kBitWidthsNumValues32);
+BENCHMARK_CAPTURE(BM_UnpackUint64, NeonUnaligned, false, &unpack_neon<uint64_t>)
+    ->ArgsProduct(kBitWidthsNumValues64);
 #endif
 
-BENCHMARK_CAPTURE(BM_UnpackUint32, DynamicAligned, true, unpack32)
+BENCHMARK_CAPTURE(BM_UnpackUint16, DynamicAligned, true, &unpack<uint16_t>)
+    ->ArgsProduct(kBitWidthsNumValues16);
+BENCHMARK_CAPTURE(BM_UnpackUint16, DynamicUnaligned, false, &unpack<uint16_t>)
+    ->ArgsProduct(kBitWidthsNumValues16);
+
+BENCHMARK_CAPTURE(BM_UnpackUint32, DynamicAligned, true, &unpack<uint32_t>)
     ->ArgsProduct(kBitWidthsNumValues32);
-BENCHMARK_CAPTURE(BM_UnpackUint32, DynamicUnaligned, false, unpack32)
+BENCHMARK_CAPTURE(BM_UnpackUint32, DynamicUnaligned, false, &unpack<uint32_t>)
     ->ArgsProduct(kBitWidthsNumValues32);
 
-BENCHMARK_CAPTURE(BM_UnpackUint64, DynamicAligned, true, unpack64)
+BENCHMARK_CAPTURE(BM_UnpackUint64, DynamicAligned, true, &unpack<uint64_t>)
     ->ArgsProduct(kBitWidthsNumValues64);
-BENCHMARK_CAPTURE(BM_UnpackUint64, DynamicUnaligned, false, unpack64)
+BENCHMARK_CAPTURE(BM_UnpackUint64, DynamicUnaligned, false, &unpack<uint64_t>)
     ->ArgsProduct(kBitWidthsNumValues64);
 
 }  // namespace
diff --git a/cpp/src/arrow/util/bpacking_default_internal.h b/cpp/src/arrow/util/bpacking_default_internal.h
deleted file mode 100644
index 4c661dcce3798..0000000000000
--- a/cpp/src/arrow/util/bpacking_default_internal.h
+++ /dev/null
@@ -1,4251 +0,0 @@
-// Licensed to the Apache Software Foundation (ASF) under one
-// or more contributor license agreements.  See the NOTICE file
-// distributed with this work for additional information
-// regarding copyright ownership.  The ASF licenses this file
-// to you under the Apache License, Version 2.0 (the
-// "License"); you may not use this file except in compliance
-// with the License.  You may obtain a copy of the License at
-//
-//   http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing,
-// software distributed under the License is distributed on an
-// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
-// KIND, either express or implied.  See the License for the
-// specific language governing permissions and limitations
-// under the License.
-
-// This file was modified from its original version for inclusion in parquet-cpp.
-// Original source:
-// https://github.com/lemire/FrameOfReference/blob/6ccaf9e97160f9a3b299e23a8ef739e711ef0c71/src/bpacking.cpp
-// The original copyright notice follows.
-
-// This code is released under the
-// Apache License Version 2.0 http://www.apache.org/licenses/.
-// (c) Daniel Lemire 2013
-
-#pragma once
-
-#include "arrow/util/bit_util.h"
-#include "arrow/util/ubsan.h"
-
-namespace arrow {
-namespace internal {
-
-inline const uint32_t* unpack1_32(const uint32_t* in, uint32_t* out) {
-  uint32_t inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out = (inl >> 0) & 1;
-  out++;
-  *out = (inl >> 1) & 1;
-  out++;
-  *out = (inl >> 2) & 1;
-  out++;
-  *out = (inl >> 3) & 1;
-  out++;
-  *out = (inl >> 4) & 1;
-  out++;
-  *out = (inl >> 5) & 1;
-  out++;
-  *out = (inl >> 6) & 1;
-  out++;
-  *out = (inl >> 7) & 1;
-  out++;
-  *out = (inl >> 8) & 1;
-  out++;
-  *out = (inl >> 9) & 1;
-  out++;
-  *out = (inl >> 10) & 1;
-  out++;
-  *out = (inl >> 11) & 1;
-  out++;
-  *out = (inl >> 12) & 1;
-  out++;
-  *out = (inl >> 13) & 1;
-  out++;
-  *out = (inl >> 14) & 1;
-  out++;
-  *out = (inl >> 15) & 1;
-  out++;
-  *out = (inl >> 16) & 1;
-  out++;
-  *out = (inl >> 17) & 1;
-  out++;
-  *out = (inl >> 18) & 1;
-  out++;
-  *out = (inl >> 19) & 1;
-  out++;
-  *out = (inl >> 20) & 1;
-  out++;
-  *out = (inl >> 21) & 1;
-  out++;
-  *out = (inl >> 22) & 1;
-  out++;
-  *out = (inl >> 23) & 1;
-  out++;
-  *out = (inl >> 24) & 1;
-  out++;
-  *out = (inl >> 25) & 1;
-  out++;
-  *out = (inl >> 26) & 1;
-  out++;
-  *out = (inl >> 27) & 1;
-  out++;
-  *out = (inl >> 28) & 1;
-  out++;
-  *out = (inl >> 29) & 1;
-  out++;
-  *out = (inl >> 30) & 1;
-  out++;
-  *out = (inl >> 31);
-  ++in;
-  out++;
-
-  return in;
-}
-
-inline const uint32_t* unpack2_32(const uint32_t* in, uint32_t* out) {
-  uint32_t inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out = (inl >> 0) % (1U << 2);
-  out++;
-  *out = (inl >> 2) % (1U << 2);
-  out++;
-  *out = (inl >> 4) % (1U << 2);
-  out++;
-  *out = (inl >> 6) % (1U << 2);
-  out++;
-  *out = (inl >> 8) % (1U << 2);
-  out++;
-  *out = (inl >> 10) % (1U << 2);
-  out++;
-  *out = (inl >> 12) % (1U << 2);
-  out++;
-  *out = (inl >> 14) % (1U << 2);
-  out++;
-  *out = (inl >> 16) % (1U << 2);
-  out++;
-  *out = (inl >> 18) % (1U << 2);
-  out++;
-  *out = (inl >> 20) % (1U << 2);
-  out++;
-  *out = (inl >> 22) % (1U << 2);
-  out++;
-  *out = (inl >> 24) % (1U << 2);
-  out++;
-  *out = (inl >> 26) % (1U << 2);
-  out++;
-  *out = (inl >> 28) % (1U << 2);
-  out++;
-  *out = (inl >> 30);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 2);
-  out++;
-  *out = (inl >> 2) % (1U << 2);
-  out++;
-  *out = (inl >> 4) % (1U << 2);
-  out++;
-  *out = (inl >> 6) % (1U << 2);
-  out++;
-  *out = (inl >> 8) % (1U << 2);
-  out++;
-  *out = (inl >> 10) % (1U << 2);
-  out++;
-  *out = (inl >> 12) % (1U << 2);
-  out++;
-  *out = (inl >> 14) % (1U << 2);
-  out++;
-  *out = (inl >> 16) % (1U << 2);
-  out++;
-  *out = (inl >> 18) % (1U << 2);
-  out++;
-  *out = (inl >> 20) % (1U << 2);
-  out++;
-  *out = (inl >> 22) % (1U << 2);
-  out++;
-  *out = (inl >> 24) % (1U << 2);
-  out++;
-  *out = (inl >> 26) % (1U << 2);
-  out++;
-  *out = (inl >> 28) % (1U << 2);
-  out++;
-  *out = (inl >> 30);
-  ++in;
-  out++;
-
-  return in;
-}
-
-inline const uint32_t* unpack3_32(const uint32_t* in, uint32_t* out) {
-  uint32_t inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out = (inl >> 0) % (1U << 3);
-  out++;
-  *out = (inl >> 3) % (1U << 3);
-  out++;
-  *out = (inl >> 6) % (1U << 3);
-  out++;
-  *out = (inl >> 9) % (1U << 3);
-  out++;
-  *out = (inl >> 12) % (1U << 3);
-  out++;
-  *out = (inl >> 15) % (1U << 3);
-  out++;
-  *out = (inl >> 18) % (1U << 3);
-  out++;
-  *out = (inl >> 21) % (1U << 3);
-  out++;
-  *out = (inl >> 24) % (1U << 3);
-  out++;
-  *out = (inl >> 27) % (1U << 3);
-  out++;
-  *out = (inl >> 30);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 1)) << (3 - 1);
-  out++;
-  *out = (inl >> 1) % (1U << 3);
-  out++;
-  *out = (inl >> 4) % (1U << 3);
-  out++;
-  *out = (inl >> 7) % (1U << 3);
-  out++;
-  *out = (inl >> 10) % (1U << 3);
-  out++;
-  *out = (inl >> 13) % (1U << 3);
-  out++;
-  *out = (inl >> 16) % (1U << 3);
-  out++;
-  *out = (inl >> 19) % (1U << 3);
-  out++;
-  *out = (inl >> 22) % (1U << 3);
-  out++;
-  *out = (inl >> 25) % (1U << 3);
-  out++;
-  *out = (inl >> 28) % (1U << 3);
-  out++;
-  *out = (inl >> 31);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 2)) << (3 - 2);
-  out++;
-  *out = (inl >> 2) % (1U << 3);
-  out++;
-  *out = (inl >> 5) % (1U << 3);
-  out++;
-  *out = (inl >> 8) % (1U << 3);
-  out++;
-  *out = (inl >> 11) % (1U << 3);
-  out++;
-  *out = (inl >> 14) % (1U << 3);
-  out++;
-  *out = (inl >> 17) % (1U << 3);
-  out++;
-  *out = (inl >> 20) % (1U << 3);
-  out++;
-  *out = (inl >> 23) % (1U << 3);
-  out++;
-  *out = (inl >> 26) % (1U << 3);
-  out++;
-  *out = (inl >> 29);
-  ++in;
-  out++;
-
-  return in;
-}
-
-inline const uint32_t* unpack4_32(const uint32_t* in, uint32_t* out) {
-  uint32_t inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out = (inl >> 0) % (1U << 4);
-  out++;
-  *out = (inl >> 4) % (1U << 4);
-  out++;
-  *out = (inl >> 8) % (1U << 4);
-  out++;
-  *out = (inl >> 12) % (1U << 4);
-  out++;
-  *out = (inl >> 16) % (1U << 4);
-  out++;
-  *out = (inl >> 20) % (1U << 4);
-  out++;
-  *out = (inl >> 24) % (1U << 4);
-  out++;
-  *out = (inl >> 28);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 4);
-  out++;
-  *out = (inl >> 4) % (1U << 4);
-  out++;
-  *out = (inl >> 8) % (1U << 4);
-  out++;
-  *out = (inl >> 12) % (1U << 4);
-  out++;
-  *out = (inl >> 16) % (1U << 4);
-  out++;
-  *out = (inl >> 20) % (1U << 4);
-  out++;
-  *out = (inl >> 24) % (1U << 4);
-  out++;
-  *out = (inl >> 28);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 4);
-  out++;
-  *out = (inl >> 4) % (1U << 4);
-  out++;
-  *out = (inl >> 8) % (1U << 4);
-  out++;
-  *out = (inl >> 12) % (1U << 4);
-  out++;
-  *out = (inl >> 16) % (1U << 4);
-  out++;
-  *out = (inl >> 20) % (1U << 4);
-  out++;
-  *out = (inl >> 24) % (1U << 4);
-  out++;
-  *out = (inl >> 28);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 4);
-  out++;
-  *out = (inl >> 4) % (1U << 4);
-  out++;
-  *out = (inl >> 8) % (1U << 4);
-  out++;
-  *out = (inl >> 12) % (1U << 4);
-  out++;
-  *out = (inl >> 16) % (1U << 4);
-  out++;
-  *out = (inl >> 20) % (1U << 4);
-  out++;
-  *out = (inl >> 24) % (1U << 4);
-  out++;
-  *out = (inl >> 28);
-  ++in;
-  out++;
-
-  return in;
-}
-
-inline const uint32_t* unpack5_32(const uint32_t* in, uint32_t* out) {
-  uint32_t inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out = (inl >> 0) % (1U << 5);
-  out++;
-  *out = (inl >> 5) % (1U << 5);
-  out++;
-  *out = (inl >> 10) % (1U << 5);
-  out++;
-  *out = (inl >> 15) % (1U << 5);
-  out++;
-  *out = (inl >> 20) % (1U << 5);
-  out++;
-  *out = (inl >> 25) % (1U << 5);
-  out++;
-  *out = (inl >> 30);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 3)) << (5 - 3);
-  out++;
-  *out = (inl >> 3) % (1U << 5);
-  out++;
-  *out = (inl >> 8) % (1U << 5);
-  out++;
-  *out = (inl >> 13) % (1U << 5);
-  out++;
-  *out = (inl >> 18) % (1U << 5);
-  out++;
-  *out = (inl >> 23) % (1U << 5);
-  out++;
-  *out = (inl >> 28);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 1)) << (5 - 1);
-  out++;
-  *out = (inl >> 1) % (1U << 5);
-  out++;
-  *out = (inl >> 6) % (1U << 5);
-  out++;
-  *out = (inl >> 11) % (1U << 5);
-  out++;
-  *out = (inl >> 16) % (1U << 5);
-  out++;
-  *out = (inl >> 21) % (1U << 5);
-  out++;
-  *out = (inl >> 26) % (1U << 5);
-  out++;
-  *out = (inl >> 31);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 4)) << (5 - 4);
-  out++;
-  *out = (inl >> 4) % (1U << 5);
-  out++;
-  *out = (inl >> 9) % (1U << 5);
-  out++;
-  *out = (inl >> 14) % (1U << 5);
-  out++;
-  *out = (inl >> 19) % (1U << 5);
-  out++;
-  *out = (inl >> 24) % (1U << 5);
-  out++;
-  *out = (inl >> 29);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 2)) << (5 - 2);
-  out++;
-  *out = (inl >> 2) % (1U << 5);
-  out++;
-  *out = (inl >> 7) % (1U << 5);
-  out++;
-  *out = (inl >> 12) % (1U << 5);
-  out++;
-  *out = (inl >> 17) % (1U << 5);
-  out++;
-  *out = (inl >> 22) % (1U << 5);
-  out++;
-  *out = (inl >> 27);
-  ++in;
-  out++;
-
-  return in;
-}
-
-inline const uint32_t* unpack6_32(const uint32_t* in, uint32_t* out) {
-  uint32_t inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out = (inl >> 0) % (1U << 6);
-  out++;
-  *out = (inl >> 6) % (1U << 6);
-  out++;
-  *out = (inl >> 12) % (1U << 6);
-  out++;
-  *out = (inl >> 18) % (1U << 6);
-  out++;
-  *out = (inl >> 24) % (1U << 6);
-  out++;
-  *out = (inl >> 30);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 4)) << (6 - 4);
-  out++;
-  *out = (inl >> 4) % (1U << 6);
-  out++;
-  *out = (inl >> 10) % (1U << 6);
-  out++;
-  *out = (inl >> 16) % (1U << 6);
-  out++;
-  *out = (inl >> 22) % (1U << 6);
-  out++;
-  *out = (inl >> 28);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 2)) << (6 - 2);
-  out++;
-  *out = (inl >> 2) % (1U << 6);
-  out++;
-  *out = (inl >> 8) % (1U << 6);
-  out++;
-  *out = (inl >> 14) % (1U << 6);
-  out++;
-  *out = (inl >> 20) % (1U << 6);
-  out++;
-  *out = (inl >> 26);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 6);
-  out++;
-  *out = (inl >> 6) % (1U << 6);
-  out++;
-  *out = (inl >> 12) % (1U << 6);
-  out++;
-  *out = (inl >> 18) % (1U << 6);
-  out++;
-  *out = (inl >> 24) % (1U << 6);
-  out++;
-  *out = (inl >> 30);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 4)) << (6 - 4);
-  out++;
-  *out = (inl >> 4) % (1U << 6);
-  out++;
-  *out = (inl >> 10) % (1U << 6);
-  out++;
-  *out = (inl >> 16) % (1U << 6);
-  out++;
-  *out = (inl >> 22) % (1U << 6);
-  out++;
-  *out = (inl >> 28);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 2)) << (6 - 2);
-  out++;
-  *out = (inl >> 2) % (1U << 6);
-  out++;
-  *out = (inl >> 8) % (1U << 6);
-  out++;
-  *out = (inl >> 14) % (1U << 6);
-  out++;
-  *out = (inl >> 20) % (1U << 6);
-  out++;
-  *out = (inl >> 26);
-  ++in;
-  out++;
-
-  return in;
-}
-
-inline const uint32_t* unpack7_32(const uint32_t* in, uint32_t* out) {
-  uint32_t inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out = (inl >> 0) % (1U << 7);
-  out++;
-  *out = (inl >> 7) % (1U << 7);
-  out++;
-  *out = (inl >> 14) % (1U << 7);
-  out++;
-  *out = (inl >> 21) % (1U << 7);
-  out++;
-  *out = (inl >> 28);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 3)) << (7 - 3);
-  out++;
-  *out = (inl >> 3) % (1U << 7);
-  out++;
-  *out = (inl >> 10) % (1U << 7);
-  out++;
-  *out = (inl >> 17) % (1U << 7);
-  out++;
-  *out = (inl >> 24) % (1U << 7);
-  out++;
-  *out = (inl >> 31);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 6)) << (7 - 6);
-  out++;
-  *out = (inl >> 6) % (1U << 7);
-  out++;
-  *out = (inl >> 13) % (1U << 7);
-  out++;
-  *out = (inl >> 20) % (1U << 7);
-  out++;
-  *out = (inl >> 27);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 2)) << (7 - 2);
-  out++;
-  *out = (inl >> 2) % (1U << 7);
-  out++;
-  *out = (inl >> 9) % (1U << 7);
-  out++;
-  *out = (inl >> 16) % (1U << 7);
-  out++;
-  *out = (inl >> 23) % (1U << 7);
-  out++;
-  *out = (inl >> 30);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 5)) << (7 - 5);
-  out++;
-  *out = (inl >> 5) % (1U << 7);
-  out++;
-  *out = (inl >> 12) % (1U << 7);
-  out++;
-  *out = (inl >> 19) % (1U << 7);
-  out++;
-  *out = (inl >> 26);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 1)) << (7 - 1);
-  out++;
-  *out = (inl >> 1) % (1U << 7);
-  out++;
-  *out = (inl >> 8) % (1U << 7);
-  out++;
-  *out = (inl >> 15) % (1U << 7);
-  out++;
-  *out = (inl >> 22) % (1U << 7);
-  out++;
-  *out = (inl >> 29);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 4)) << (7 - 4);
-  out++;
-  *out = (inl >> 4) % (1U << 7);
-  out++;
-  *out = (inl >> 11) % (1U << 7);
-  out++;
-  *out = (inl >> 18) % (1U << 7);
-  out++;
-  *out = (inl >> 25);
-  ++in;
-  out++;
-
-  return in;
-}
-
-inline const uint32_t* unpack8_32(const uint32_t* in, uint32_t* out) {
-  uint32_t inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out = (inl >> 0) % (1U << 8);
-  out++;
-  *out = (inl >> 8) % (1U << 8);
-  out++;
-  *out = (inl >> 16) % (1U << 8);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 8);
-  out++;
-  *out = (inl >> 8) % (1U << 8);
-  out++;
-  *out = (inl >> 16) % (1U << 8);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 8);
-  out++;
-  *out = (inl >> 8) % (1U << 8);
-  out++;
-  *out = (inl >> 16) % (1U << 8);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 8);
-  out++;
-  *out = (inl >> 8) % (1U << 8);
-  out++;
-  *out = (inl >> 16) % (1U << 8);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 8);
-  out++;
-  *out = (inl >> 8) % (1U << 8);
-  out++;
-  *out = (inl >> 16) % (1U << 8);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 8);
-  out++;
-  *out = (inl >> 8) % (1U << 8);
-  out++;
-  *out = (inl >> 16) % (1U << 8);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 8);
-  out++;
-  *out = (inl >> 8) % (1U << 8);
-  out++;
-  *out = (inl >> 16) % (1U << 8);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 8);
-  out++;
-  *out = (inl >> 8) % (1U << 8);
-  out++;
-  *out = (inl >> 16) % (1U << 8);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  out++;
-
-  return in;
-}
-
-inline const uint32_t* unpack9_32(const uint32_t* in, uint32_t* out) {
-  uint32_t inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out = (inl >> 0) % (1U << 9);
-  out++;
-  *out = (inl >> 9) % (1U << 9);
-  out++;
-  *out = (inl >> 18) % (1U << 9);
-  out++;
-  *out = (inl >> 27);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 4)) << (9 - 4);
-  out++;
-  *out = (inl >> 4) % (1U << 9);
-  out++;
-  *out = (inl >> 13) % (1U << 9);
-  out++;
-  *out = (inl >> 22) % (1U << 9);
-  out++;
-  *out = (inl >> 31);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 8)) << (9 - 8);
-  out++;
-  *out = (inl >> 8) % (1U << 9);
-  out++;
-  *out = (inl >> 17) % (1U << 9);
-  out++;
-  *out = (inl >> 26);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 3)) << (9 - 3);
-  out++;
-  *out = (inl >> 3) % (1U << 9);
-  out++;
-  *out = (inl >> 12) % (1U << 9);
-  out++;
-  *out = (inl >> 21) % (1U << 9);
-  out++;
-  *out = (inl >> 30);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 7)) << (9 - 7);
-  out++;
-  *out = (inl >> 7) % (1U << 9);
-  out++;
-  *out = (inl >> 16) % (1U << 9);
-  out++;
-  *out = (inl >> 25);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 2)) << (9 - 2);
-  out++;
-  *out = (inl >> 2) % (1U << 9);
-  out++;
-  *out = (inl >> 11) % (1U << 9);
-  out++;
-  *out = (inl >> 20) % (1U << 9);
-  out++;
-  *out = (inl >> 29);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 6)) << (9 - 6);
-  out++;
-  *out = (inl >> 6) % (1U << 9);
-  out++;
-  *out = (inl >> 15) % (1U << 9);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 1)) << (9 - 1);
-  out++;
-  *out = (inl >> 1) % (1U << 9);
-  out++;
-  *out = (inl >> 10) % (1U << 9);
-  out++;
-  *out = (inl >> 19) % (1U << 9);
-  out++;
-  *out = (inl >> 28);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 5)) << (9 - 5);
-  out++;
-  *out = (inl >> 5) % (1U << 9);
-  out++;
-  *out = (inl >> 14) % (1U << 9);
-  out++;
-  *out = (inl >> 23);
-  ++in;
-  out++;
-
-  return in;
-}
-
-inline const uint32_t* unpack10_32(const uint32_t* in, uint32_t* out) {
-  uint32_t inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out = (inl >> 0) % (1U << 10);
-  out++;
-  *out = (inl >> 10) % (1U << 10);
-  out++;
-  *out = (inl >> 20) % (1U << 10);
-  out++;
-  *out = (inl >> 30);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 8)) << (10 - 8);
-  out++;
-  *out = (inl >> 8) % (1U << 10);
-  out++;
-  *out = (inl >> 18) % (1U << 10);
-  out++;
-  *out = (inl >> 28);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 6)) << (10 - 6);
-  out++;
-  *out = (inl >> 6) % (1U << 10);
-  out++;
-  *out = (inl >> 16) % (1U << 10);
-  out++;
-  *out = (inl >> 26);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 4)) << (10 - 4);
-  out++;
-  *out = (inl >> 4) % (1U << 10);
-  out++;
-  *out = (inl >> 14) % (1U << 10);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 2)) << (10 - 2);
-  out++;
-  *out = (inl >> 2) % (1U << 10);
-  out++;
-  *out = (inl >> 12) % (1U << 10);
-  out++;
-  *out = (inl >> 22);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 10);
-  out++;
-  *out = (inl >> 10) % (1U << 10);
-  out++;
-  *out = (inl >> 20) % (1U << 10);
-  out++;
-  *out = (inl >> 30);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 8)) << (10 - 8);
-  out++;
-  *out = (inl >> 8) % (1U << 10);
-  out++;
-  *out = (inl >> 18) % (1U << 10);
-  out++;
-  *out = (inl >> 28);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 6)) << (10 - 6);
-  out++;
-  *out = (inl >> 6) % (1U << 10);
-  out++;
-  *out = (inl >> 16) % (1U << 10);
-  out++;
-  *out = (inl >> 26);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 4)) << (10 - 4);
-  out++;
-  *out = (inl >> 4) % (1U << 10);
-  out++;
-  *out = (inl >> 14) % (1U << 10);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 2)) << (10 - 2);
-  out++;
-  *out = (inl >> 2) % (1U << 10);
-  out++;
-  *out = (inl >> 12) % (1U << 10);
-  out++;
-  *out = (inl >> 22);
-  ++in;
-  out++;
-
-  return in;
-}
-
-inline const uint32_t* unpack11_32(const uint32_t* in, uint32_t* out) {
-  uint32_t inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out = (inl >> 0) % (1U << 11);
-  out++;
-  *out = (inl >> 11) % (1U << 11);
-  out++;
-  *out = (inl >> 22);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 1)) << (11 - 1);
-  out++;
-  *out = (inl >> 1) % (1U << 11);
-  out++;
-  *out = (inl >> 12) % (1U << 11);
-  out++;
-  *out = (inl >> 23);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 2)) << (11 - 2);
-  out++;
-  *out = (inl >> 2) % (1U << 11);
-  out++;
-  *out = (inl >> 13) % (1U << 11);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 3)) << (11 - 3);
-  out++;
-  *out = (inl >> 3) % (1U << 11);
-  out++;
-  *out = (inl >> 14) % (1U << 11);
-  out++;
-  *out = (inl >> 25);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 4)) << (11 - 4);
-  out++;
-  *out = (inl >> 4) % (1U << 11);
-  out++;
-  *out = (inl >> 15) % (1U << 11);
-  out++;
-  *out = (inl >> 26);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 5)) << (11 - 5);
-  out++;
-  *out = (inl >> 5) % (1U << 11);
-  out++;
-  *out = (inl >> 16) % (1U << 11);
-  out++;
-  *out = (inl >> 27);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 6)) << (11 - 6);
-  out++;
-  *out = (inl >> 6) % (1U << 11);
-  out++;
-  *out = (inl >> 17) % (1U << 11);
-  out++;
-  *out = (inl >> 28);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 7)) << (11 - 7);
-  out++;
-  *out = (inl >> 7) % (1U << 11);
-  out++;
-  *out = (inl >> 18) % (1U << 11);
-  out++;
-  *out = (inl >> 29);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 8)) << (11 - 8);
-  out++;
-  *out = (inl >> 8) % (1U << 11);
-  out++;
-  *out = (inl >> 19) % (1U << 11);
-  out++;
-  *out = (inl >> 30);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 9)) << (11 - 9);
-  out++;
-  *out = (inl >> 9) % (1U << 11);
-  out++;
-  *out = (inl >> 20) % (1U << 11);
-  out++;
-  *out = (inl >> 31);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 10)) << (11 - 10);
-  out++;
-  *out = (inl >> 10) % (1U << 11);
-  out++;
-  *out = (inl >> 21);
-  ++in;
-  out++;
-
-  return in;
-}
-
-inline const uint32_t* unpack12_32(const uint32_t* in, uint32_t* out) {
-  uint32_t inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out = (inl >> 0) % (1U << 12);
-  out++;
-  *out = (inl >> 12) % (1U << 12);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 4)) << (12 - 4);
-  out++;
-  *out = (inl >> 4) % (1U << 12);
-  out++;
-  *out = (inl >> 16) % (1U << 12);
-  out++;
-  *out = (inl >> 28);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 8)) << (12 - 8);
-  out++;
-  *out = (inl >> 8) % (1U << 12);
-  out++;
-  *out = (inl >> 20);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 12);
-  out++;
-  *out = (inl >> 12) % (1U << 12);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 4)) << (12 - 4);
-  out++;
-  *out = (inl >> 4) % (1U << 12);
-  out++;
-  *out = (inl >> 16) % (1U << 12);
-  out++;
-  *out = (inl >> 28);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 8)) << (12 - 8);
-  out++;
-  *out = (inl >> 8) % (1U << 12);
-  out++;
-  *out = (inl >> 20);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 12);
-  out++;
-  *out = (inl >> 12) % (1U << 12);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 4)) << (12 - 4);
-  out++;
-  *out = (inl >> 4) % (1U << 12);
-  out++;
-  *out = (inl >> 16) % (1U << 12);
-  out++;
-  *out = (inl >> 28);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 8)) << (12 - 8);
-  out++;
-  *out = (inl >> 8) % (1U << 12);
-  out++;
-  *out = (inl >> 20);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 12);
-  out++;
-  *out = (inl >> 12) % (1U << 12);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 4)) << (12 - 4);
-  out++;
-  *out = (inl >> 4) % (1U << 12);
-  out++;
-  *out = (inl >> 16) % (1U << 12);
-  out++;
-  *out = (inl >> 28);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 8)) << (12 - 8);
-  out++;
-  *out = (inl >> 8) % (1U << 12);
-  out++;
-  *out = (inl >> 20);
-  ++in;
-  out++;
-
-  return in;
-}
-
-inline const uint32_t* unpack13_32(const uint32_t* in, uint32_t* out) {
-  uint32_t inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out = (inl >> 0) % (1U << 13);
-  out++;
-  *out = (inl >> 13) % (1U << 13);
-  out++;
-  *out = (inl >> 26);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 7)) << (13 - 7);
-  out++;
-  *out = (inl >> 7) % (1U << 13);
-  out++;
-  *out = (inl >> 20);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 1)) << (13 - 1);
-  out++;
-  *out = (inl >> 1) % (1U << 13);
-  out++;
-  *out = (inl >> 14) % (1U << 13);
-  out++;
-  *out = (inl >> 27);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 8)) << (13 - 8);
-  out++;
-  *out = (inl >> 8) % (1U << 13);
-  out++;
-  *out = (inl >> 21);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 2)) << (13 - 2);
-  out++;
-  *out = (inl >> 2) % (1U << 13);
-  out++;
-  *out = (inl >> 15) % (1U << 13);
-  out++;
-  *out = (inl >> 28);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 9)) << (13 - 9);
-  out++;
-  *out = (inl >> 9) % (1U << 13);
-  out++;
-  *out = (inl >> 22);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 3)) << (13 - 3);
-  out++;
-  *out = (inl >> 3) % (1U << 13);
-  out++;
-  *out = (inl >> 16) % (1U << 13);
-  out++;
-  *out = (inl >> 29);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 10)) << (13 - 10);
-  out++;
-  *out = (inl >> 10) % (1U << 13);
-  out++;
-  *out = (inl >> 23);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 4)) << (13 - 4);
-  out++;
-  *out = (inl >> 4) % (1U << 13);
-  out++;
-  *out = (inl >> 17) % (1U << 13);
-  out++;
-  *out = (inl >> 30);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 11)) << (13 - 11);
-  out++;
-  *out = (inl >> 11) % (1U << 13);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 5)) << (13 - 5);
-  out++;
-  *out = (inl >> 5) % (1U << 13);
-  out++;
-  *out = (inl >> 18) % (1U << 13);
-  out++;
-  *out = (inl >> 31);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 12)) << (13 - 12);
-  out++;
-  *out = (inl >> 12) % (1U << 13);
-  out++;
-  *out = (inl >> 25);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 6)) << (13 - 6);
-  out++;
-  *out = (inl >> 6) % (1U << 13);
-  out++;
-  *out = (inl >> 19);
-  ++in;
-  out++;
-
-  return in;
-}
-
-inline const uint32_t* unpack14_32(const uint32_t* in, uint32_t* out) {
-  uint32_t inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out = (inl >> 0) % (1U << 14);
-  out++;
-  *out = (inl >> 14) % (1U << 14);
-  out++;
-  *out = (inl >> 28);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 10)) << (14 - 10);
-  out++;
-  *out = (inl >> 10) % (1U << 14);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 6)) << (14 - 6);
-  out++;
-  *out = (inl >> 6) % (1U << 14);
-  out++;
-  *out = (inl >> 20);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 2)) << (14 - 2);
-  out++;
-  *out = (inl >> 2) % (1U << 14);
-  out++;
-  *out = (inl >> 16) % (1U << 14);
-  out++;
-  *out = (inl >> 30);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 12)) << (14 - 12);
-  out++;
-  *out = (inl >> 12) % (1U << 14);
-  out++;
-  *out = (inl >> 26);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 8)) << (14 - 8);
-  out++;
-  *out = (inl >> 8) % (1U << 14);
-  out++;
-  *out = (inl >> 22);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 4)) << (14 - 4);
-  out++;
-  *out = (inl >> 4) % (1U << 14);
-  out++;
-  *out = (inl >> 18);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 14);
-  out++;
-  *out = (inl >> 14) % (1U << 14);
-  out++;
-  *out = (inl >> 28);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 10)) << (14 - 10);
-  out++;
-  *out = (inl >> 10) % (1U << 14);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 6)) << (14 - 6);
-  out++;
-  *out = (inl >> 6) % (1U << 14);
-  out++;
-  *out = (inl >> 20);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 2)) << (14 - 2);
-  out++;
-  *out = (inl >> 2) % (1U << 14);
-  out++;
-  *out = (inl >> 16) % (1U << 14);
-  out++;
-  *out = (inl >> 30);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 12)) << (14 - 12);
-  out++;
-  *out = (inl >> 12) % (1U << 14);
-  out++;
-  *out = (inl >> 26);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 8)) << (14 - 8);
-  out++;
-  *out = (inl >> 8) % (1U << 14);
-  out++;
-  *out = (inl >> 22);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 4)) << (14 - 4);
-  out++;
-  *out = (inl >> 4) % (1U << 14);
-  out++;
-  *out = (inl >> 18);
-  ++in;
-  out++;
-
-  return in;
-}
-
-inline const uint32_t* unpack15_32(const uint32_t* in, uint32_t* out) {
-  uint32_t inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out = (inl >> 0) % (1U << 15);
-  out++;
-  *out = (inl >> 15) % (1U << 15);
-  out++;
-  *out = (inl >> 30);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 13)) << (15 - 13);
-  out++;
-  *out = (inl >> 13) % (1U << 15);
-  out++;
-  *out = (inl >> 28);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 11)) << (15 - 11);
-  out++;
-  *out = (inl >> 11) % (1U << 15);
-  out++;
-  *out = (inl >> 26);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 9)) << (15 - 9);
-  out++;
-  *out = (inl >> 9) % (1U << 15);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 7)) << (15 - 7);
-  out++;
-  *out = (inl >> 7) % (1U << 15);
-  out++;
-  *out = (inl >> 22);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 5)) << (15 - 5);
-  out++;
-  *out = (inl >> 5) % (1U << 15);
-  out++;
-  *out = (inl >> 20);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 3)) << (15 - 3);
-  out++;
-  *out = (inl >> 3) % (1U << 15);
-  out++;
-  *out = (inl >> 18);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 1)) << (15 - 1);
-  out++;
-  *out = (inl >> 1) % (1U << 15);
-  out++;
-  *out = (inl >> 16) % (1U << 15);
-  out++;
-  *out = (inl >> 31);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 14)) << (15 - 14);
-  out++;
-  *out = (inl >> 14) % (1U << 15);
-  out++;
-  *out = (inl >> 29);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 12)) << (15 - 12);
-  out++;
-  *out = (inl >> 12) % (1U << 15);
-  out++;
-  *out = (inl >> 27);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 10)) << (15 - 10);
-  out++;
-  *out = (inl >> 10) % (1U << 15);
-  out++;
-  *out = (inl >> 25);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 8)) << (15 - 8);
-  out++;
-  *out = (inl >> 8) % (1U << 15);
-  out++;
-  *out = (inl >> 23);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 6)) << (15 - 6);
-  out++;
-  *out = (inl >> 6) % (1U << 15);
-  out++;
-  *out = (inl >> 21);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 4)) << (15 - 4);
-  out++;
-  *out = (inl >> 4) % (1U << 15);
-  out++;
-  *out = (inl >> 19);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 2)) << (15 - 2);
-  out++;
-  *out = (inl >> 2) % (1U << 15);
-  out++;
-  *out = (inl >> 17);
-  ++in;
-  out++;
-
-  return in;
-}
-
-inline const uint32_t* unpack16_32(const uint32_t* in, uint32_t* out) {
-  uint32_t inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out = (inl >> 0) % (1U << 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  out++;
-
-  return in;
-}
-
-inline const uint32_t* unpack17_32(const uint32_t* in, uint32_t* out) {
-  uint32_t inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out = (inl >> 0) % (1U << 17);
-  out++;
-  *out = (inl >> 17);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 2)) << (17 - 2);
-  out++;
-  *out = (inl >> 2) % (1U << 17);
-  out++;
-  *out = (inl >> 19);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 4)) << (17 - 4);
-  out++;
-  *out = (inl >> 4) % (1U << 17);
-  out++;
-  *out = (inl >> 21);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 6)) << (17 - 6);
-  out++;
-  *out = (inl >> 6) % (1U << 17);
-  out++;
-  *out = (inl >> 23);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 8)) << (17 - 8);
-  out++;
-  *out = (inl >> 8) % (1U << 17);
-  out++;
-  *out = (inl >> 25);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 10)) << (17 - 10);
-  out++;
-  *out = (inl >> 10) % (1U << 17);
-  out++;
-  *out = (inl >> 27);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 12)) << (17 - 12);
-  out++;
-  *out = (inl >> 12) % (1U << 17);
-  out++;
-  *out = (inl >> 29);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 14)) << (17 - 14);
-  out++;
-  *out = (inl >> 14) % (1U << 17);
-  out++;
-  *out = (inl >> 31);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 16)) << (17 - 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 1)) << (17 - 1);
-  out++;
-  *out = (inl >> 1) % (1U << 17);
-  out++;
-  *out = (inl >> 18);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 3)) << (17 - 3);
-  out++;
-  *out = (inl >> 3) % (1U << 17);
-  out++;
-  *out = (inl >> 20);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 5)) << (17 - 5);
-  out++;
-  *out = (inl >> 5) % (1U << 17);
-  out++;
-  *out = (inl >> 22);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 7)) << (17 - 7);
-  out++;
-  *out = (inl >> 7) % (1U << 17);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 9)) << (17 - 9);
-  out++;
-  *out = (inl >> 9) % (1U << 17);
-  out++;
-  *out = (inl >> 26);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 11)) << (17 - 11);
-  out++;
-  *out = (inl >> 11) % (1U << 17);
-  out++;
-  *out = (inl >> 28);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 13)) << (17 - 13);
-  out++;
-  *out = (inl >> 13) % (1U << 17);
-  out++;
-  *out = (inl >> 30);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 15)) << (17 - 15);
-  out++;
-  *out = (inl >> 15);
-  ++in;
-  out++;
-
-  return in;
-}
-
-inline const uint32_t* unpack18_32(const uint32_t* in, uint32_t* out) {
-  uint32_t inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out = (inl >> 0) % (1U << 18);
-  out++;
-  *out = (inl >> 18);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 4)) << (18 - 4);
-  out++;
-  *out = (inl >> 4) % (1U << 18);
-  out++;
-  *out = (inl >> 22);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 8)) << (18 - 8);
-  out++;
-  *out = (inl >> 8) % (1U << 18);
-  out++;
-  *out = (inl >> 26);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 12)) << (18 - 12);
-  out++;
-  *out = (inl >> 12) % (1U << 18);
-  out++;
-  *out = (inl >> 30);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 16)) << (18 - 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 2)) << (18 - 2);
-  out++;
-  *out = (inl >> 2) % (1U << 18);
-  out++;
-  *out = (inl >> 20);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 6)) << (18 - 6);
-  out++;
-  *out = (inl >> 6) % (1U << 18);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 10)) << (18 - 10);
-  out++;
-  *out = (inl >> 10) % (1U << 18);
-  out++;
-  *out = (inl >> 28);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 14)) << (18 - 14);
-  out++;
-  *out = (inl >> 14);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 18);
-  out++;
-  *out = (inl >> 18);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 4)) << (18 - 4);
-  out++;
-  *out = (inl >> 4) % (1U << 18);
-  out++;
-  *out = (inl >> 22);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 8)) << (18 - 8);
-  out++;
-  *out = (inl >> 8) % (1U << 18);
-  out++;
-  *out = (inl >> 26);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 12)) << (18 - 12);
-  out++;
-  *out = (inl >> 12) % (1U << 18);
-  out++;
-  *out = (inl >> 30);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 16)) << (18 - 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 2)) << (18 - 2);
-  out++;
-  *out = (inl >> 2) % (1U << 18);
-  out++;
-  *out = (inl >> 20);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 6)) << (18 - 6);
-  out++;
-  *out = (inl >> 6) % (1U << 18);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 10)) << (18 - 10);
-  out++;
-  *out = (inl >> 10) % (1U << 18);
-  out++;
-  *out = (inl >> 28);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 14)) << (18 - 14);
-  out++;
-  *out = (inl >> 14);
-  ++in;
-  out++;
-
-  return in;
-}
-
-inline const uint32_t* unpack19_32(const uint32_t* in, uint32_t* out) {
-  uint32_t inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out = (inl >> 0) % (1U << 19);
-  out++;
-  *out = (inl >> 19);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 6)) << (19 - 6);
-  out++;
-  *out = (inl >> 6) % (1U << 19);
-  out++;
-  *out = (inl >> 25);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 12)) << (19 - 12);
-  out++;
-  *out = (inl >> 12) % (1U << 19);
-  out++;
-  *out = (inl >> 31);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 18)) << (19 - 18);
-  out++;
-  *out = (inl >> 18);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 5)) << (19 - 5);
-  out++;
-  *out = (inl >> 5) % (1U << 19);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 11)) << (19 - 11);
-  out++;
-  *out = (inl >> 11) % (1U << 19);
-  out++;
-  *out = (inl >> 30);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 17)) << (19 - 17);
-  out++;
-  *out = (inl >> 17);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 4)) << (19 - 4);
-  out++;
-  *out = (inl >> 4) % (1U << 19);
-  out++;
-  *out = (inl >> 23);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 10)) << (19 - 10);
-  out++;
-  *out = (inl >> 10) % (1U << 19);
-  out++;
-  *out = (inl >> 29);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 16)) << (19 - 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 3)) << (19 - 3);
-  out++;
-  *out = (inl >> 3) % (1U << 19);
-  out++;
-  *out = (inl >> 22);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 9)) << (19 - 9);
-  out++;
-  *out = (inl >> 9) % (1U << 19);
-  out++;
-  *out = (inl >> 28);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 15)) << (19 - 15);
-  out++;
-  *out = (inl >> 15);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 2)) << (19 - 2);
-  out++;
-  *out = (inl >> 2) % (1U << 19);
-  out++;
-  *out = (inl >> 21);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 8)) << (19 - 8);
-  out++;
-  *out = (inl >> 8) % (1U << 19);
-  out++;
-  *out = (inl >> 27);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 14)) << (19 - 14);
-  out++;
-  *out = (inl >> 14);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 1)) << (19 - 1);
-  out++;
-  *out = (inl >> 1) % (1U << 19);
-  out++;
-  *out = (inl >> 20);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 7)) << (19 - 7);
-  out++;
-  *out = (inl >> 7) % (1U << 19);
-  out++;
-  *out = (inl >> 26);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 13)) << (19 - 13);
-  out++;
-  *out = (inl >> 13);
-  ++in;
-  out++;
-
-  return in;
-}
-
-inline const uint32_t* unpack20_32(const uint32_t* in, uint32_t* out) {
-  uint32_t inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out = (inl >> 0) % (1U << 20);
-  out++;
-  *out = (inl >> 20);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 8)) << (20 - 8);
-  out++;
-  *out = (inl >> 8) % (1U << 20);
-  out++;
-  *out = (inl >> 28);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 16)) << (20 - 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 4)) << (20 - 4);
-  out++;
-  *out = (inl >> 4) % (1U << 20);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 12)) << (20 - 12);
-  out++;
-  *out = (inl >> 12);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 20);
-  out++;
-  *out = (inl >> 20);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 8)) << (20 - 8);
-  out++;
-  *out = (inl >> 8) % (1U << 20);
-  out++;
-  *out = (inl >> 28);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 16)) << (20 - 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 4)) << (20 - 4);
-  out++;
-  *out = (inl >> 4) % (1U << 20);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 12)) << (20 - 12);
-  out++;
-  *out = (inl >> 12);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 20);
-  out++;
-  *out = (inl >> 20);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 8)) << (20 - 8);
-  out++;
-  *out = (inl >> 8) % (1U << 20);
-  out++;
-  *out = (inl >> 28);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 16)) << (20 - 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 4)) << (20 - 4);
-  out++;
-  *out = (inl >> 4) % (1U << 20);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 12)) << (20 - 12);
-  out++;
-  *out = (inl >> 12);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 20);
-  out++;
-  *out = (inl >> 20);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 8)) << (20 - 8);
-  out++;
-  *out = (inl >> 8) % (1U << 20);
-  out++;
-  *out = (inl >> 28);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 16)) << (20 - 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 4)) << (20 - 4);
-  out++;
-  *out = (inl >> 4) % (1U << 20);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 12)) << (20 - 12);
-  out++;
-  *out = (inl >> 12);
-  ++in;
-  out++;
-
-  return in;
-}
-
-inline const uint32_t* unpack21_32(const uint32_t* in, uint32_t* out) {
-  uint32_t inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out = (inl >> 0) % (1U << 21);
-  out++;
-  *out = (inl >> 21);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 10)) << (21 - 10);
-  out++;
-  *out = (inl >> 10) % (1U << 21);
-  out++;
-  *out = (inl >> 31);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 20)) << (21 - 20);
-  out++;
-  *out = (inl >> 20);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 9)) << (21 - 9);
-  out++;
-  *out = (inl >> 9) % (1U << 21);
-  out++;
-  *out = (inl >> 30);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 19)) << (21 - 19);
-  out++;
-  *out = (inl >> 19);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 8)) << (21 - 8);
-  out++;
-  *out = (inl >> 8) % (1U << 21);
-  out++;
-  *out = (inl >> 29);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 18)) << (21 - 18);
-  out++;
-  *out = (inl >> 18);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 7)) << (21 - 7);
-  out++;
-  *out = (inl >> 7) % (1U << 21);
-  out++;
-  *out = (inl >> 28);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 17)) << (21 - 17);
-  out++;
-  *out = (inl >> 17);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 6)) << (21 - 6);
-  out++;
-  *out = (inl >> 6) % (1U << 21);
-  out++;
-  *out = (inl >> 27);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 16)) << (21 - 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 5)) << (21 - 5);
-  out++;
-  *out = (inl >> 5) % (1U << 21);
-  out++;
-  *out = (inl >> 26);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 15)) << (21 - 15);
-  out++;
-  *out = (inl >> 15);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 4)) << (21 - 4);
-  out++;
-  *out = (inl >> 4) % (1U << 21);
-  out++;
-  *out = (inl >> 25);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 14)) << (21 - 14);
-  out++;
-  *out = (inl >> 14);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 3)) << (21 - 3);
-  out++;
-  *out = (inl >> 3) % (1U << 21);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 13)) << (21 - 13);
-  out++;
-  *out = (inl >> 13);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 2)) << (21 - 2);
-  out++;
-  *out = (inl >> 2) % (1U << 21);
-  out++;
-  *out = (inl >> 23);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 12)) << (21 - 12);
-  out++;
-  *out = (inl >> 12);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 1)) << (21 - 1);
-  out++;
-  *out = (inl >> 1) % (1U << 21);
-  out++;
-  *out = (inl >> 22);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 11)) << (21 - 11);
-  out++;
-  *out = (inl >> 11);
-  ++in;
-  out++;
-
-  return in;
-}
-
-inline const uint32_t* unpack22_32(const uint32_t* in, uint32_t* out) {
-  uint32_t inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out = (inl >> 0) % (1U << 22);
-  out++;
-  *out = (inl >> 22);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 12)) << (22 - 12);
-  out++;
-  *out = (inl >> 12);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 2)) << (22 - 2);
-  out++;
-  *out = (inl >> 2) % (1U << 22);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 14)) << (22 - 14);
-  out++;
-  *out = (inl >> 14);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 4)) << (22 - 4);
-  out++;
-  *out = (inl >> 4) % (1U << 22);
-  out++;
-  *out = (inl >> 26);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 16)) << (22 - 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 6)) << (22 - 6);
-  out++;
-  *out = (inl >> 6) % (1U << 22);
-  out++;
-  *out = (inl >> 28);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 18)) << (22 - 18);
-  out++;
-  *out = (inl >> 18);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 8)) << (22 - 8);
-  out++;
-  *out = (inl >> 8) % (1U << 22);
-  out++;
-  *out = (inl >> 30);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 20)) << (22 - 20);
-  out++;
-  *out = (inl >> 20);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 10)) << (22 - 10);
-  out++;
-  *out = (inl >> 10);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 22);
-  out++;
-  *out = (inl >> 22);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 12)) << (22 - 12);
-  out++;
-  *out = (inl >> 12);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 2)) << (22 - 2);
-  out++;
-  *out = (inl >> 2) % (1U << 22);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 14)) << (22 - 14);
-  out++;
-  *out = (inl >> 14);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 4)) << (22 - 4);
-  out++;
-  *out = (inl >> 4) % (1U << 22);
-  out++;
-  *out = (inl >> 26);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 16)) << (22 - 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 6)) << (22 - 6);
-  out++;
-  *out = (inl >> 6) % (1U << 22);
-  out++;
-  *out = (inl >> 28);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 18)) << (22 - 18);
-  out++;
-  *out = (inl >> 18);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 8)) << (22 - 8);
-  out++;
-  *out = (inl >> 8) % (1U << 22);
-  out++;
-  *out = (inl >> 30);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 20)) << (22 - 20);
-  out++;
-  *out = (inl >> 20);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 10)) << (22 - 10);
-  out++;
-  *out = (inl >> 10);
-  ++in;
-  out++;
-
-  return in;
-}
-
-inline const uint32_t* unpack23_32(const uint32_t* in, uint32_t* out) {
-  uint32_t inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out = (inl >> 0) % (1U << 23);
-  out++;
-  *out = (inl >> 23);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 14)) << (23 - 14);
-  out++;
-  *out = (inl >> 14);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 5)) << (23 - 5);
-  out++;
-  *out = (inl >> 5) % (1U << 23);
-  out++;
-  *out = (inl >> 28);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 19)) << (23 - 19);
-  out++;
-  *out = (inl >> 19);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 10)) << (23 - 10);
-  out++;
-  *out = (inl >> 10);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 1)) << (23 - 1);
-  out++;
-  *out = (inl >> 1) % (1U << 23);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 15)) << (23 - 15);
-  out++;
-  *out = (inl >> 15);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 6)) << (23 - 6);
-  out++;
-  *out = (inl >> 6) % (1U << 23);
-  out++;
-  *out = (inl >> 29);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 20)) << (23 - 20);
-  out++;
-  *out = (inl >> 20);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 11)) << (23 - 11);
-  out++;
-  *out = (inl >> 11);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 2)) << (23 - 2);
-  out++;
-  *out = (inl >> 2) % (1U << 23);
-  out++;
-  *out = (inl >> 25);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 16)) << (23 - 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 7)) << (23 - 7);
-  out++;
-  *out = (inl >> 7) % (1U << 23);
-  out++;
-  *out = (inl >> 30);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 21)) << (23 - 21);
-  out++;
-  *out = (inl >> 21);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 12)) << (23 - 12);
-  out++;
-  *out = (inl >> 12);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 3)) << (23 - 3);
-  out++;
-  *out = (inl >> 3) % (1U << 23);
-  out++;
-  *out = (inl >> 26);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 17)) << (23 - 17);
-  out++;
-  *out = (inl >> 17);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 8)) << (23 - 8);
-  out++;
-  *out = (inl >> 8) % (1U << 23);
-  out++;
-  *out = (inl >> 31);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 22)) << (23 - 22);
-  out++;
-  *out = (inl >> 22);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 13)) << (23 - 13);
-  out++;
-  *out = (inl >> 13);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 4)) << (23 - 4);
-  out++;
-  *out = (inl >> 4) % (1U << 23);
-  out++;
-  *out = (inl >> 27);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 18)) << (23 - 18);
-  out++;
-  *out = (inl >> 18);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 9)) << (23 - 9);
-  out++;
-  *out = (inl >> 9);
-  ++in;
-  out++;
-
-  return in;
-}
-
-inline const uint32_t* unpack24_32(const uint32_t* in, uint32_t* out) {
-  uint32_t inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out = (inl >> 0) % (1U << 24);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 16)) << (24 - 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 8)) << (24 - 8);
-  out++;
-  *out = (inl >> 8);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 24);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 16)) << (24 - 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 8)) << (24 - 8);
-  out++;
-  *out = (inl >> 8);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 24);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 16)) << (24 - 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 8)) << (24 - 8);
-  out++;
-  *out = (inl >> 8);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 24);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 16)) << (24 - 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 8)) << (24 - 8);
-  out++;
-  *out = (inl >> 8);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 24);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 16)) << (24 - 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 8)) << (24 - 8);
-  out++;
-  *out = (inl >> 8);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 24);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 16)) << (24 - 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 8)) << (24 - 8);
-  out++;
-  *out = (inl >> 8);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 24);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 16)) << (24 - 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 8)) << (24 - 8);
-  out++;
-  *out = (inl >> 8);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 24);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 16)) << (24 - 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 8)) << (24 - 8);
-  out++;
-  *out = (inl >> 8);
-  ++in;
-  out++;
-
-  return in;
-}
-
-inline const uint32_t* unpack25_32(const uint32_t* in, uint32_t* out) {
-  uint32_t inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out = (inl >> 0) % (1U << 25);
-  out++;
-  *out = (inl >> 25);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 18)) << (25 - 18);
-  out++;
-  *out = (inl >> 18);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 11)) << (25 - 11);
-  out++;
-  *out = (inl >> 11);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 4)) << (25 - 4);
-  out++;
-  *out = (inl >> 4) % (1U << 25);
-  out++;
-  *out = (inl >> 29);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 22)) << (25 - 22);
-  out++;
-  *out = (inl >> 22);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 15)) << (25 - 15);
-  out++;
-  *out = (inl >> 15);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 8)) << (25 - 8);
-  out++;
-  *out = (inl >> 8);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 1)) << (25 - 1);
-  out++;
-  *out = (inl >> 1) % (1U << 25);
-  out++;
-  *out = (inl >> 26);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 19)) << (25 - 19);
-  out++;
-  *out = (inl >> 19);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 12)) << (25 - 12);
-  out++;
-  *out = (inl >> 12);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 5)) << (25 - 5);
-  out++;
-  *out = (inl >> 5) % (1U << 25);
-  out++;
-  *out = (inl >> 30);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 23)) << (25 - 23);
-  out++;
-  *out = (inl >> 23);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 16)) << (25 - 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 9)) << (25 - 9);
-  out++;
-  *out = (inl >> 9);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 2)) << (25 - 2);
-  out++;
-  *out = (inl >> 2) % (1U << 25);
-  out++;
-  *out = (inl >> 27);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 20)) << (25 - 20);
-  out++;
-  *out = (inl >> 20);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 13)) << (25 - 13);
-  out++;
-  *out = (inl >> 13);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 6)) << (25 - 6);
-  out++;
-  *out = (inl >> 6) % (1U << 25);
-  out++;
-  *out = (inl >> 31);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 24)) << (25 - 24);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 17)) << (25 - 17);
-  out++;
-  *out = (inl >> 17);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 10)) << (25 - 10);
-  out++;
-  *out = (inl >> 10);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 3)) << (25 - 3);
-  out++;
-  *out = (inl >> 3) % (1U << 25);
-  out++;
-  *out = (inl >> 28);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 21)) << (25 - 21);
-  out++;
-  *out = (inl >> 21);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 14)) << (25 - 14);
-  out++;
-  *out = (inl >> 14);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 7)) << (25 - 7);
-  out++;
-  *out = (inl >> 7);
-  ++in;
-  out++;
-
-  return in;
-}
-
-inline const uint32_t* unpack26_32(const uint32_t* in, uint32_t* out) {
-  uint32_t inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out = (inl >> 0) % (1U << 26);
-  out++;
-  *out = (inl >> 26);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 20)) << (26 - 20);
-  out++;
-  *out = (inl >> 20);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 14)) << (26 - 14);
-  out++;
-  *out = (inl >> 14);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 8)) << (26 - 8);
-  out++;
-  *out = (inl >> 8);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 2)) << (26 - 2);
-  out++;
-  *out = (inl >> 2) % (1U << 26);
-  out++;
-  *out = (inl >> 28);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 22)) << (26 - 22);
-  out++;
-  *out = (inl >> 22);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 16)) << (26 - 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 10)) << (26 - 10);
-  out++;
-  *out = (inl >> 10);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 4)) << (26 - 4);
-  out++;
-  *out = (inl >> 4) % (1U << 26);
-  out++;
-  *out = (inl >> 30);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 24)) << (26 - 24);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 18)) << (26 - 18);
-  out++;
-  *out = (inl >> 18);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 12)) << (26 - 12);
-  out++;
-  *out = (inl >> 12);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 6)) << (26 - 6);
-  out++;
-  *out = (inl >> 6);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 26);
-  out++;
-  *out = (inl >> 26);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 20)) << (26 - 20);
-  out++;
-  *out = (inl >> 20);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 14)) << (26 - 14);
-  out++;
-  *out = (inl >> 14);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 8)) << (26 - 8);
-  out++;
-  *out = (inl >> 8);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 2)) << (26 - 2);
-  out++;
-  *out = (inl >> 2) % (1U << 26);
-  out++;
-  *out = (inl >> 28);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 22)) << (26 - 22);
-  out++;
-  *out = (inl >> 22);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 16)) << (26 - 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 10)) << (26 - 10);
-  out++;
-  *out = (inl >> 10);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 4)) << (26 - 4);
-  out++;
-  *out = (inl >> 4) % (1U << 26);
-  out++;
-  *out = (inl >> 30);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 24)) << (26 - 24);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 18)) << (26 - 18);
-  out++;
-  *out = (inl >> 18);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 12)) << (26 - 12);
-  out++;
-  *out = (inl >> 12);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 6)) << (26 - 6);
-  out++;
-  *out = (inl >> 6);
-  ++in;
-  out++;
-
-  return in;
-}
-
-inline const uint32_t* unpack27_32(const uint32_t* in, uint32_t* out) {
-  uint32_t inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out = (inl >> 0) % (1U << 27);
-  out++;
-  *out = (inl >> 27);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 22)) << (27 - 22);
-  out++;
-  *out = (inl >> 22);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 17)) << (27 - 17);
-  out++;
-  *out = (inl >> 17);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 12)) << (27 - 12);
-  out++;
-  *out = (inl >> 12);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 7)) << (27 - 7);
-  out++;
-  *out = (inl >> 7);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 2)) << (27 - 2);
-  out++;
-  *out = (inl >> 2) % (1U << 27);
-  out++;
-  *out = (inl >> 29);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 24)) << (27 - 24);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 19)) << (27 - 19);
-  out++;
-  *out = (inl >> 19);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 14)) << (27 - 14);
-  out++;
-  *out = (inl >> 14);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 9)) << (27 - 9);
-  out++;
-  *out = (inl >> 9);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 4)) << (27 - 4);
-  out++;
-  *out = (inl >> 4) % (1U << 27);
-  out++;
-  *out = (inl >> 31);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 26)) << (27 - 26);
-  out++;
-  *out = (inl >> 26);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 21)) << (27 - 21);
-  out++;
-  *out = (inl >> 21);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 16)) << (27 - 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 11)) << (27 - 11);
-  out++;
-  *out = (inl >> 11);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 6)) << (27 - 6);
-  out++;
-  *out = (inl >> 6);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 1)) << (27 - 1);
-  out++;
-  *out = (inl >> 1) % (1U << 27);
-  out++;
-  *out = (inl >> 28);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 23)) << (27 - 23);
-  out++;
-  *out = (inl >> 23);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 18)) << (27 - 18);
-  out++;
-  *out = (inl >> 18);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 13)) << (27 - 13);
-  out++;
-  *out = (inl >> 13);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 8)) << (27 - 8);
-  out++;
-  *out = (inl >> 8);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 3)) << (27 - 3);
-  out++;
-  *out = (inl >> 3) % (1U << 27);
-  out++;
-  *out = (inl >> 30);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 25)) << (27 - 25);
-  out++;
-  *out = (inl >> 25);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 20)) << (27 - 20);
-  out++;
-  *out = (inl >> 20);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 15)) << (27 - 15);
-  out++;
-  *out = (inl >> 15);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 10)) << (27 - 10);
-  out++;
-  *out = (inl >> 10);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 5)) << (27 - 5);
-  out++;
-  *out = (inl >> 5);
-  ++in;
-  out++;
-
-  return in;
-}
-
-inline const uint32_t* unpack28_32(const uint32_t* in, uint32_t* out) {
-  uint32_t inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out = (inl >> 0) % (1U << 28);
-  out++;
-  *out = (inl >> 28);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 24)) << (28 - 24);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 20)) << (28 - 20);
-  out++;
-  *out = (inl >> 20);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 16)) << (28 - 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 12)) << (28 - 12);
-  out++;
-  *out = (inl >> 12);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 8)) << (28 - 8);
-  out++;
-  *out = (inl >> 8);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 4)) << (28 - 4);
-  out++;
-  *out = (inl >> 4);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 28);
-  out++;
-  *out = (inl >> 28);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 24)) << (28 - 24);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 20)) << (28 - 20);
-  out++;
-  *out = (inl >> 20);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 16)) << (28 - 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 12)) << (28 - 12);
-  out++;
-  *out = (inl >> 12);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 8)) << (28 - 8);
-  out++;
-  *out = (inl >> 8);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 4)) << (28 - 4);
-  out++;
-  *out = (inl >> 4);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 28);
-  out++;
-  *out = (inl >> 28);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 24)) << (28 - 24);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 20)) << (28 - 20);
-  out++;
-  *out = (inl >> 20);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 16)) << (28 - 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 12)) << (28 - 12);
-  out++;
-  *out = (inl >> 12);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 8)) << (28 - 8);
-  out++;
-  *out = (inl >> 8);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 4)) << (28 - 4);
-  out++;
-  *out = (inl >> 4);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 28);
-  out++;
-  *out = (inl >> 28);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 24)) << (28 - 24);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 20)) << (28 - 20);
-  out++;
-  *out = (inl >> 20);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 16)) << (28 - 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 12)) << (28 - 12);
-  out++;
-  *out = (inl >> 12);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 8)) << (28 - 8);
-  out++;
-  *out = (inl >> 8);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 4)) << (28 - 4);
-  out++;
-  *out = (inl >> 4);
-  ++in;
-  out++;
-
-  return in;
-}
-
-inline const uint32_t* unpack29_32(const uint32_t* in, uint32_t* out) {
-  uint32_t inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out = (inl >> 0) % (1U << 29);
-  out++;
-  *out = (inl >> 29);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 26)) << (29 - 26);
-  out++;
-  *out = (inl >> 26);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 23)) << (29 - 23);
-  out++;
-  *out = (inl >> 23);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 20)) << (29 - 20);
-  out++;
-  *out = (inl >> 20);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 17)) << (29 - 17);
-  out++;
-  *out = (inl >> 17);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 14)) << (29 - 14);
-  out++;
-  *out = (inl >> 14);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 11)) << (29 - 11);
-  out++;
-  *out = (inl >> 11);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 8)) << (29 - 8);
-  out++;
-  *out = (inl >> 8);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 5)) << (29 - 5);
-  out++;
-  *out = (inl >> 5);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 2)) << (29 - 2);
-  out++;
-  *out = (inl >> 2) % (1U << 29);
-  out++;
-  *out = (inl >> 31);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 28)) << (29 - 28);
-  out++;
-  *out = (inl >> 28);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 25)) << (29 - 25);
-  out++;
-  *out = (inl >> 25);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 22)) << (29 - 22);
-  out++;
-  *out = (inl >> 22);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 19)) << (29 - 19);
-  out++;
-  *out = (inl >> 19);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 16)) << (29 - 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 13)) << (29 - 13);
-  out++;
-  *out = (inl >> 13);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 10)) << (29 - 10);
-  out++;
-  *out = (inl >> 10);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 7)) << (29 - 7);
-  out++;
-  *out = (inl >> 7);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 4)) << (29 - 4);
-  out++;
-  *out = (inl >> 4);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 1)) << (29 - 1);
-  out++;
-  *out = (inl >> 1) % (1U << 29);
-  out++;
-  *out = (inl >> 30);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 27)) << (29 - 27);
-  out++;
-  *out = (inl >> 27);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 24)) << (29 - 24);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 21)) << (29 - 21);
-  out++;
-  *out = (inl >> 21);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 18)) << (29 - 18);
-  out++;
-  *out = (inl >> 18);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 15)) << (29 - 15);
-  out++;
-  *out = (inl >> 15);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 12)) << (29 - 12);
-  out++;
-  *out = (inl >> 12);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 9)) << (29 - 9);
-  out++;
-  *out = (inl >> 9);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 6)) << (29 - 6);
-  out++;
-  *out = (inl >> 6);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 3)) << (29 - 3);
-  out++;
-  *out = (inl >> 3);
-  ++in;
-  out++;
-
-  return in;
-}
-
-inline const uint32_t* unpack30_32(const uint32_t* in, uint32_t* out) {
-  uint32_t inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out = (inl >> 0) % (1U << 30);
-  out++;
-  *out = (inl >> 30);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 28)) << (30 - 28);
-  out++;
-  *out = (inl >> 28);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 26)) << (30 - 26);
-  out++;
-  *out = (inl >> 26);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 24)) << (30 - 24);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 22)) << (30 - 22);
-  out++;
-  *out = (inl >> 22);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 20)) << (30 - 20);
-  out++;
-  *out = (inl >> 20);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 18)) << (30 - 18);
-  out++;
-  *out = (inl >> 18);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 16)) << (30 - 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 14)) << (30 - 14);
-  out++;
-  *out = (inl >> 14);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 12)) << (30 - 12);
-  out++;
-  *out = (inl >> 12);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 10)) << (30 - 10);
-  out++;
-  *out = (inl >> 10);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 8)) << (30 - 8);
-  out++;
-  *out = (inl >> 8);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 6)) << (30 - 6);
-  out++;
-  *out = (inl >> 6);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 4)) << (30 - 4);
-  out++;
-  *out = (inl >> 4);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 2)) << (30 - 2);
-  out++;
-  *out = (inl >> 2);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0) % (1U << 30);
-  out++;
-  *out = (inl >> 30);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 28)) << (30 - 28);
-  out++;
-  *out = (inl >> 28);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 26)) << (30 - 26);
-  out++;
-  *out = (inl >> 26);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 24)) << (30 - 24);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 22)) << (30 - 22);
-  out++;
-  *out = (inl >> 22);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 20)) << (30 - 20);
-  out++;
-  *out = (inl >> 20);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 18)) << (30 - 18);
-  out++;
-  *out = (inl >> 18);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 16)) << (30 - 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 14)) << (30 - 14);
-  out++;
-  *out = (inl >> 14);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 12)) << (30 - 12);
-  out++;
-  *out = (inl >> 12);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 10)) << (30 - 10);
-  out++;
-  *out = (inl >> 10);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 8)) << (30 - 8);
-  out++;
-  *out = (inl >> 8);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 6)) << (30 - 6);
-  out++;
-  *out = (inl >> 6);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 4)) << (30 - 4);
-  out++;
-  *out = (inl >> 4);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 2)) << (30 - 2);
-  out++;
-  *out = (inl >> 2);
-  ++in;
-  out++;
-
-  return in;
-}
-
-inline const uint32_t* unpack31_32(const uint32_t* in, uint32_t* out) {
-  uint32_t inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out = (inl >> 0) % (1U << 31);
-  out++;
-  *out = (inl >> 31);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 30)) << (31 - 30);
-  out++;
-  *out = (inl >> 30);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 29)) << (31 - 29);
-  out++;
-  *out = (inl >> 29);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 28)) << (31 - 28);
-  out++;
-  *out = (inl >> 28);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 27)) << (31 - 27);
-  out++;
-  *out = (inl >> 27);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 26)) << (31 - 26);
-  out++;
-  *out = (inl >> 26);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 25)) << (31 - 25);
-  out++;
-  *out = (inl >> 25);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 24)) << (31 - 24);
-  out++;
-  *out = (inl >> 24);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 23)) << (31 - 23);
-  out++;
-  *out = (inl >> 23);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 22)) << (31 - 22);
-  out++;
-  *out = (inl >> 22);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 21)) << (31 - 21);
-  out++;
-  *out = (inl >> 21);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 20)) << (31 - 20);
-  out++;
-  *out = (inl >> 20);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 19)) << (31 - 19);
-  out++;
-  *out = (inl >> 19);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 18)) << (31 - 18);
-  out++;
-  *out = (inl >> 18);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 17)) << (31 - 17);
-  out++;
-  *out = (inl >> 17);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 16)) << (31 - 16);
-  out++;
-  *out = (inl >> 16);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 15)) << (31 - 15);
-  out++;
-  *out = (inl >> 15);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 14)) << (31 - 14);
-  out++;
-  *out = (inl >> 14);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 13)) << (31 - 13);
-  out++;
-  *out = (inl >> 13);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 12)) << (31 - 12);
-  out++;
-  *out = (inl >> 12);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 11)) << (31 - 11);
-  out++;
-  *out = (inl >> 11);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 10)) << (31 - 10);
-  out++;
-  *out = (inl >> 10);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 9)) << (31 - 9);
-  out++;
-  *out = (inl >> 9);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 8)) << (31 - 8);
-  out++;
-  *out = (inl >> 8);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 7)) << (31 - 7);
-  out++;
-  *out = (inl >> 7);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 6)) << (31 - 6);
-  out++;
-  *out = (inl >> 6);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 5)) << (31 - 5);
-  out++;
-  *out = (inl >> 5);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 4)) << (31 - 4);
-  out++;
-  *out = (inl >> 4);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 3)) << (31 - 3);
-  out++;
-  *out = (inl >> 3);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 2)) << (31 - 2);
-  out++;
-  *out = (inl >> 2);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out |= (inl % (1U << 1)) << (31 - 1);
-  out++;
-  *out = (inl >> 1);
-  ++in;
-  out++;
-
-  return in;
-}
-
-inline const uint32_t* unpack32_32(const uint32_t* in, uint32_t* out) {
-  uint32_t inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  *out = (inl >> 0);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0);
-  ++in;
-  inl = util::SafeLoad(in);
-  inl = arrow::bit_util::FromLittleEndian(inl);
-  out++;
-  *out = (inl >> 0);
-  ++in;
-  out++;
-
-  return in;
-}
-
-inline const uint32_t* nullunpacker32(const uint32_t* in, uint32_t* out) {
-  for (int k = 0; k < 32; ++k) {
-    out[k] = 0;
-  }
-  return in;
-}
-
-}  // namespace internal
-}  // namespace arrow
diff --git a/cpp/src/arrow/util/bpacking_dispatch_internal.h b/cpp/src/arrow/util/bpacking_dispatch_internal.h
new file mode 100644
index 0000000000000..d7eefb6cb1edf
--- /dev/null
+++ b/cpp/src/arrow/util/bpacking_dispatch_internal.h
@@ -0,0 +1,104 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include <cstring>
+#include <utility>
+
+#include "arrow/util/endian.h"
+#include "arrow/util/logging.h"
+#include "arrow/util/ubsan.h"
+
+namespace arrow::internal {
+
+/// Unpack a zero bit packed array.
+template <typename Uint>
+int unpack_null(const uint8_t* in, Uint* out, int batch_size) {
+  std::memset(out, 0, batch_size * sizeof(Uint));
+  return batch_size;
+}
+
+/// Unpack a packed array where packed and unpacked values have exactly the same number of
+/// bits.
+template <typename Uint>
+int unpack_full(const uint8_t* in, Uint* out, int batch_size) {
+  if constexpr (ARROW_LITTLE_ENDIAN == 1) {
+    std::memcpy(out, in, batch_size * sizeof(Uint));
+  } else {
+    using bit_util::FromLittleEndian;
+    using util::SafeLoadAs;
+
+    for (int k = 0; k < batch_size; k += 1) {
+      out[k] = FromLittleEndian(SafeLoadAs<Uint>(in + (k * sizeof(Uint))));
+    }
+  }
+  return batch_size;
+}
+
+/// Unpack a packed array, delegating to a Unpacker struct.
+///
+/// @tparam kPackedBitWidth The width in bits of the values in the packed array.
+/// @tparam Unpacker The struct providing information and an ``unpack`` method to unpack a
+///                  fixed amount of values (usually constrained by SIMD batch sizes and
+///                  byte alignment).
+/// @tparam UnpackedUInt The type in which we unpack the values.
+template <int kPackedBitWidth, template <typename, int> typename Unpacker,
+          typename UnpackedUInt>
+int unpack_width(const uint8_t* in, UnpackedUInt* out, int batch_size) {
+  using UnpackerForWidth = Unpacker<UnpackedUInt, kPackedBitWidth>;
+
+  constexpr auto kValuesUnpacked = UnpackerForWidth::kValuesUnpacked;
+  batch_size = batch_size / kValuesUnpacked * kValuesUnpacked;
+  int num_loops = batch_size / kValuesUnpacked;
+
+  for (int i = 0; i < num_loops; ++i) {
+    in = UnpackerForWidth::unpack(in, out + i * kValuesUnpacked);
+  }
+
+  return batch_size;
+}
+
+template <template <typename, int> typename Unpacker, typename UnpackedUint, int... Is>
+constexpr auto make_unpack_jump_table_impl(std::integer_sequence<int, Is...>) {
+  return std::array{
+      &unpack_null<UnpackedUint>,
+      &unpack_width<Is + 1, Unpacker, UnpackedUint>...,
+      &unpack_full<UnpackedUint>,
+  };
+}
+
+template <template <typename, int> typename Unpacker, typename UnpackedUint>
+constexpr auto make_unpack_jump_table() {
+  return make_unpack_jump_table_impl<Unpacker, UnpackedUint>(
+      std::make_integer_sequence<int, 8 * sizeof(UnpackedUint) - 1>());
+}
+
+template <template <typename, int> typename Unpacker, typename UnpackedUint>
+constexpr auto get_unpack_fn(int packed_width) {
+  constexpr auto kJumpTable = make_unpack_jump_table<Unpacker, UnpackedUint>();
+  return kJumpTable[packed_width];
+}
+
+template <template <typename, int> typename Unpacker, typename UnpackedUint>
+static int unpack_jump(const uint8_t* in, UnpackedUint* out, int batch_size,
+                       int num_bits) {
+  ARROW_DCHECK_GE(num_bits, 0);
+  ARROW_DCHECK_LE(num_bits, static_cast<int>(8 * sizeof(UnpackedUint)));
+  return get_unpack_fn<Unpacker, UnpackedUint>(num_bits)(in, out, batch_size);
+}
+}  // namespace arrow::internal
diff --git a/cpp/src/arrow/util/bpacking_internal.h b/cpp/src/arrow/util/bpacking_internal.h
index e003cd8c0c6a7..8695beeb3528f 100644
--- a/cpp/src/arrow/util/bpacking_internal.h
+++ b/cpp/src/arrow/util/bpacking_internal.h
@@ -23,17 +23,28 @@
 
 namespace arrow::internal {
 
-/// The scalar 32 bit unpacking.
-ARROW_EXPORT int unpack32_scalar(const uint8_t* in, uint32_t* out, int batch_size,
-                                 int num_bits);
-
-/// The scalar 64 bit unpacking.
-ARROW_EXPORT int unpack64_scalar(const uint8_t* in, uint64_t* out, int batch_size,
-                                 int num_bits);
-
-ARROW_EXPORT
-int unpack32(const uint8_t* in, uint32_t* out, int batch_size, int num_bits);
-ARROW_EXPORT
-int unpack64(const uint8_t* in, uint64_t* out, int batch_size, int num_bits);
+template <typename Uint>
+ARROW_EXPORT int unpack(const uint8_t* in, Uint* out, int batch_size, int num_bits);
+
+extern template ARROW_TEMPLATE_EXPORT int unpack<uint16_t>(const uint8_t*, uint16_t*, int,
+                                                           int);
+
+extern template ARROW_TEMPLATE_EXPORT int unpack<uint32_t>(const uint8_t*, uint32_t*, int,
+                                                           int);
+
+extern template ARROW_TEMPLATE_EXPORT int unpack<uint64_t>(const uint8_t*, uint64_t*, int,
+                                                           int);
+
+template <typename Uint>
+using UnpackFn = int (*)(const uint8_t* in, Uint* out, int batch_size);
+
+template <typename Uint>
+ARROW_EXPORT UnpackFn<Uint> get_unpack_fn(int num_bits);
+
+extern template ARROW_TEMPLATE_EXPORT UnpackFn<uint16_t> get_unpack_fn<uint16_t>(int);
+
+extern template ARROW_TEMPLATE_EXPORT UnpackFn<uint32_t> get_unpack_fn<uint32_t>(int);
+
+extern template ARROW_TEMPLATE_EXPORT UnpackFn<uint64_t> get_unpack_fn<uint64_t>(int);
 
 }  // namespace arrow::internal
diff --git a/cpp/src/arrow/util/bpacking_neon.cc b/cpp/src/arrow/util/bpacking_neon.cc
deleted file mode 100644
index 407b309b7e8e4..0000000000000
--- a/cpp/src/arrow/util/bpacking_neon.cc
+++ /dev/null
@@ -1,29 +0,0 @@
-// Licensed to the Apache Software Foundation (ASF) under one
-// or more contributor license agreements.  See the NOTICE file
-// distributed with this work for additional information
-// regarding copyright ownership.  The ASF licenses this file
-// to you under the Apache License, Version 2.0 (the
-// "License"); you may not use this file except in compliance
-// with the License.  You may obtain a copy of the License at
-//
-//   http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing,
-// software distributed under the License is distributed on an
-// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
-// KIND, either express or implied.  See the License for the
-// specific language governing permissions and limitations
-// under the License.
-
-#include "arrow/util/bpacking_neon_internal.h"
-#include "arrow/util/bpacking_simd128_generated_internal.h"
-#include "arrow/util/bpacking_simd_internal.h"
-
-namespace arrow::internal {
-
-int unpack32_neon(const uint8_t* in, uint32_t* out, int batch_size, int num_bits) {
-  return unpack32_specialized<UnpackBits128<DispatchLevel::NEON>>(
-      reinterpret_cast<const uint32_t*>(in), out, batch_size, num_bits);
-}
-
-}  // namespace arrow::internal
diff --git a/cpp/src/arrow/util/bpacking_neon_internal.h b/cpp/src/arrow/util/bpacking_neon_internal.h
deleted file mode 100644
index 683aa5cbc4780..0000000000000
--- a/cpp/src/arrow/util/bpacking_neon_internal.h
+++ /dev/null
@@ -1,29 +0,0 @@
-// Licensed to the Apache Software Foundation (ASF) under one
-// or more contributor license agreements.  See the NOTICE file
-// distributed with this work for additional information
-// regarding copyright ownership.  The ASF licenses this file
-// to you under the Apache License, Version 2.0 (the
-// "License"); you may not use this file except in compliance
-// with the License.  You may obtain a copy of the License at
-//
-//   http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing,
-// software distributed under the License is distributed on an
-// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
-// KIND, either express or implied.  See the License for the
-// specific language governing permissions and limitations
-// under the License.
-
-#pragma once
-
-#include "arrow/util/visibility.h"
-
-#include <cstdint>
-
-namespace arrow::internal {
-
-ARROW_EXPORT int unpack32_neon(const uint8_t* in, uint32_t* out, int batch_size,
-                               int num_bits);
-
-}  // namespace arrow::internal
diff --git a/cpp/src/arrow/util/bpacking_scalar.cc b/cpp/src/arrow/util/bpacking_scalar.cc
new file mode 100644
index 0000000000000..632b399e74e89
--- /dev/null
+++ b/cpp/src/arrow/util/bpacking_scalar.cc
@@ -0,0 +1,42 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#include "arrow/util/bpacking_dispatch_internal.h"
+#include "arrow/util/bpacking_scalar_generated_internal.h"
+#include "arrow/util/bpacking_scalar_internal.h"
+
+namespace arrow::internal {
+
+template <typename Uint>
+int unpack_scalar(const uint8_t* in, Uint* out, int batch_size, int num_bits) {
+  return unpack_jump<ScalarUnpackerForWidth>(in, out, batch_size, num_bits);
+}
+
+template int unpack_scalar<uint16_t>(const uint8_t*, uint16_t*, int, int);
+template int unpack_scalar<uint32_t>(const uint8_t*, uint32_t*, int, int);
+template int unpack_scalar<uint64_t>(const uint8_t*, uint64_t*, int, int);
+
+template <typename Uint>
+UnpackFn<Uint> get_unpack_fn_scalar(int num_bits) {
+  return get_unpack_fn<ScalarUnpackerForWidth, Uint>(num_bits);
+}
+
+template UnpackFn<uint16_t> get_unpack_fn_scalar<uint16_t>(int);
+template UnpackFn<uint32_t> get_unpack_fn_scalar<uint32_t>(int);
+template UnpackFn<uint64_t> get_unpack_fn_scalar<uint64_t>(int);
+
+}  // namespace arrow::internal
diff --git a/cpp/src/arrow/util/bpacking_scalar_codegen.py b/cpp/src/arrow/util/bpacking_scalar_codegen.py
new file mode 100644
index 0000000000000..3fe6b0dd71ee6
--- /dev/null
+++ b/cpp/src/arrow/util/bpacking_scalar_codegen.py
@@ -0,0 +1,222 @@
+#!/bin/python
+
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements.  See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership.  The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License.  You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied.  See the License for the
+# specific language governing permissions and limitations
+# under the License.
+
+# This script is modified from its original version in GitHub. Original source:
+# https://github.com/lemire/FrameOfReference/blob/146948b6058a976bc7767262ad3a2ce201486b93/scripts/turbopacking64.py
+
+# Usage:
+#   python bpacking64_codegen.py > bpacking64_default_internal.h
+
+
+import dataclasses
+import sys
+
+
+LICENSE = """// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+// This file was generated by script which is modified from its original version in GitHub.
+// Original source:
+// https://github.com/lemire/FrameOfReference/blob/master/scripts/turbopacking64.py
+// The original copyright notice follows.
+
+// This code is released under the
+// Apache License Version 2.0 http://www.apache.org/licenses/.
+// (c) Daniel Lemire 2013
+"""
+
+HEADER = """
+#pragma once
+
+#include <cstdint>
+#include <cstring>
+
+#include "arrow/util/endian.h"
+#include "arrow/util/ubsan.h"
+
+namespace arrow::internal {
+
+template <typename Int>
+Int LoadInt(const uint8_t* in) {
+  return bit_util::FromLittleEndian(util::SafeLoadAs<Int>(in));
+}
+"""
+
+FOOTER = """
+}  // namespace arrow::internal
+"""
+
+
+@dataclasses.dataclass
+class ScalarUnpackGenerator:
+    out_bit_width: int
+    smart_halve: bool
+
+    @property
+    def out_byte_width(self) -> int:
+        return self.out_bit_width // 8
+
+    @property
+    def out_type(self) -> str:
+        return f"uint{self.out_bit_width}_t"
+
+    @property
+    def out_type_half(self) -> str:
+        return f"uint{self.out_bit_width // 2}_t"
+
+    @property
+    def struct_name(self) -> str:
+        return "ScalarUnpackerForWidth"
+
+    def struct_specialization(self, bit: int) -> str:
+        return f"{self.struct_name}<{self.out_type}, {bit}>"
+
+    def print_struct_declaration(self):
+        print("template<typename Uint, int BitWidth>")
+        print(f"struct {self.struct_name};")
+
+    @property
+    def howmany(self) -> int:
+        """How many values are we going to pack?"""
+        if self.smart_halve:
+            return self.out_bit_width // 2
+        return self.out_bit_width
+
+    def howmanywords(self, bit: int) -> int:
+        return (self.howmany * bit + self.out_bit_width - 1) // self.out_bit_width
+
+    def howmanybytes(self, bit: int) -> int:
+        return (self.howmany * bit + self.out_byte_width - 1) // self.out_byte_width
+
+    def print_unpack_k(self, bit: int) -> None:
+        print(
+            f"  static const uint8_t* unpack(const uint8_t* in, {self.out_type}* out) {{"
+        )
+
+        print(
+            f"    constexpr {self.out_type} mask = "
+            f"(({self.out_type}{{1}} << {bit}) - {self.out_type}{{1}});"
+        )
+        print()
+        maskstr = " & mask"
+
+        for k in range(self.howmanywords(bit) - 1):
+            print(
+                f"    const auto w{k} = LoadInt<{self.out_type}>("
+                f"in + {k} * {self.out_byte_width});"
+            )
+
+        k = self.howmanywords(bit) - 1
+        use_smart_halving = self.smart_halve and bit % 2 == 1
+        if use_smart_halving:
+            print(
+                f"    const auto w{k} = static_cast<{self.out_type}>(LoadInt<{self.out_type_half}>("
+                f"in + {k} * {self.out_byte_width}));"
+            )
+        else:
+            print(
+                f"    const auto w{k} = LoadInt<{self.out_type}>("
+                f"in + {k} * {self.out_byte_width});"
+            )
+
+        for j in range(self.howmany):
+            firstword = j * bit // self.out_bit_width
+            secondword = (j * bit + bit - 1) // self.out_bit_width
+            firstshift = (j * bit) % self.out_bit_width
+            firstshiftstr = f" >> {firstshift}"
+            if firstshift == 0:
+                firstshiftstr = ""  # no need
+            if firstword == secondword:
+                if firstshift + bit == self.out_bit_width:
+                    print(f"    out[{j}] = w{firstword}{firstshiftstr};")
+                else:
+                    print(f"    out[{j}] = (w{firstword}{firstshiftstr}){maskstr};")
+            else:
+                secondshift = self.out_bit_width - firstshift
+                print(
+                    f"    out[{j}] = ((w{firstword}{firstshiftstr}) | "
+                    f"(w{firstword + 1} << {secondshift})){maskstr};"
+                )
+        print()
+
+        if use_smart_halving:
+            print(
+                f"    return in + ({self.howmanywords(bit) - 1} * {self.out_byte_width}"
+                f" + {self.out_byte_width // 2});"
+            )
+        else:
+            print(f"    return in + ({self.howmanywords(bit)} * {self.out_byte_width});")
+        print("  }")
+
+    def print_struct_k(self, bit: int) -> None:
+        print("template<>")
+        print(f"struct {self.struct_specialization(bit)} {{")
+        print()
+        print(f"  static constexpr int kValuesUnpacked = {self.howmany};")
+        print()
+        self.print_unpack_k(bit)
+        print("};")
+
+    def print_structs(self):
+        for bit in range(1, self.out_bit_width):
+            self.print_struct_k(bit)
+            print()
+
+
+def print_note():
+    print("// WARNING: this file is generated, DO NOT EDIT.")
+    print("// Usage:")
+    print(f"//   python {' '.join(sys.orig_argv[1:])}")
+
+
+if __name__ == "__main__":
+    print(LICENSE)
+    print_note()
+    print(HEADER)
+
+    gen = ScalarUnpackGenerator(16, smart_halve=False)
+    gen.print_struct_declaration()
+    print()
+    gen.print_structs()
+    print()
+
+    gen = ScalarUnpackGenerator(32, smart_halve=False)
+    print()
+
+    gen.print_structs()
+    print()
+
+    gen = ScalarUnpackGenerator(64, smart_halve=True)
+    gen.print_structs()
+
+    print(FOOTER)
diff --git a/cpp/src/arrow/util/bpacking_scalar_generated_internal.h b/cpp/src/arrow/util/bpacking_scalar_generated_internal.h
new file mode 100644
index 0000000000000..ded68a6113413
--- /dev/null
+++ b/cpp/src/arrow/util/bpacking_scalar_generated_internal.h
@@ -0,0 +1,6359 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+// This file was generated by script which is modified from its original version in GitHub.
+// Original source:
+// https://github.com/lemire/FrameOfReference/blob/master/scripts/turbopacking64.py
+// The original copyright notice follows.
+
+// This code is released under the
+// Apache License Version 2.0 http://www.apache.org/licenses/.
+// (c) Daniel Lemire 2013
+
+// WARNING: this file is generated, DO NOT EDIT.
+// Usage:
+//   python cpp/src/arrow/util/bpacking_scalar_codegen.py
+
+#pragma once
+
+#include <cstdint>
+#include <cstring>
+
+#include "arrow/util/endian.h"
+#include "arrow/util/ubsan.h"
+
+namespace arrow::internal {
+
+template <typename Int>
+Int LoadInt(const uint8_t* in) {
+  return bit_util::FromLittleEndian(util::SafeLoadAs<Int>(in));
+}
+
+template<typename Uint, int BitWidth>
+struct ScalarUnpackerForWidth;
+
+template<>
+struct ScalarUnpackerForWidth<uint16_t, 1> {
+
+  static constexpr int kValuesUnpacked = 16;
+
+  static const uint8_t* unpack(const uint8_t* in, uint16_t* out) {
+    constexpr uint16_t mask = ((uint16_t{1} << 1) - uint16_t{1});
+
+    const auto w0 = LoadInt<uint16_t>(in + 0 * 2);
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 1) & mask;
+    out[2] = (w0 >> 2) & mask;
+    out[3] = (w0 >> 3) & mask;
+    out[4] = (w0 >> 4) & mask;
+    out[5] = (w0 >> 5) & mask;
+    out[6] = (w0 >> 6) & mask;
+    out[7] = (w0 >> 7) & mask;
+    out[8] = (w0 >> 8) & mask;
+    out[9] = (w0 >> 9) & mask;
+    out[10] = (w0 >> 10) & mask;
+    out[11] = (w0 >> 11) & mask;
+    out[12] = (w0 >> 12) & mask;
+    out[13] = (w0 >> 13) & mask;
+    out[14] = (w0 >> 14) & mask;
+    out[15] = w0 >> 15;
+
+    return in + (1 * 2);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint16_t, 2> {
+
+  static constexpr int kValuesUnpacked = 16;
+
+  static const uint8_t* unpack(const uint8_t* in, uint16_t* out) {
+    constexpr uint16_t mask = ((uint16_t{1} << 2) - uint16_t{1});
+
+    const auto w0 = LoadInt<uint16_t>(in + 0 * 2);
+    const auto w1 = LoadInt<uint16_t>(in + 1 * 2);
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 2) & mask;
+    out[2] = (w0 >> 4) & mask;
+    out[3] = (w0 >> 6) & mask;
+    out[4] = (w0 >> 8) & mask;
+    out[5] = (w0 >> 10) & mask;
+    out[6] = (w0 >> 12) & mask;
+    out[7] = w0 >> 14;
+    out[8] = (w1) & mask;
+    out[9] = (w1 >> 2) & mask;
+    out[10] = (w1 >> 4) & mask;
+    out[11] = (w1 >> 6) & mask;
+    out[12] = (w1 >> 8) & mask;
+    out[13] = (w1 >> 10) & mask;
+    out[14] = (w1 >> 12) & mask;
+    out[15] = w1 >> 14;
+
+    return in + (2 * 2);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint16_t, 3> {
+
+  static constexpr int kValuesUnpacked = 16;
+
+  static const uint8_t* unpack(const uint8_t* in, uint16_t* out) {
+    constexpr uint16_t mask = ((uint16_t{1} << 3) - uint16_t{1});
+
+    const auto w0 = LoadInt<uint16_t>(in + 0 * 2);
+    const auto w1 = LoadInt<uint16_t>(in + 1 * 2);
+    const auto w2 = LoadInt<uint16_t>(in + 2 * 2);
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 3) & mask;
+    out[2] = (w0 >> 6) & mask;
+    out[3] = (w0 >> 9) & mask;
+    out[4] = (w0 >> 12) & mask;
+    out[5] = ((w0 >> 15) | (w1 << 1)) & mask;
+    out[6] = (w1 >> 2) & mask;
+    out[7] = (w1 >> 5) & mask;
+    out[8] = (w1 >> 8) & mask;
+    out[9] = (w1 >> 11) & mask;
+    out[10] = ((w1 >> 14) | (w2 << 2)) & mask;
+    out[11] = (w2 >> 1) & mask;
+    out[12] = (w2 >> 4) & mask;
+    out[13] = (w2 >> 7) & mask;
+    out[14] = (w2 >> 10) & mask;
+    out[15] = w2 >> 13;
+
+    return in + (3 * 2);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint16_t, 4> {
+
+  static constexpr int kValuesUnpacked = 16;
+
+  static const uint8_t* unpack(const uint8_t* in, uint16_t* out) {
+    constexpr uint16_t mask = ((uint16_t{1} << 4) - uint16_t{1});
+
+    const auto w0 = LoadInt<uint16_t>(in + 0 * 2);
+    const auto w1 = LoadInt<uint16_t>(in + 1 * 2);
+    const auto w2 = LoadInt<uint16_t>(in + 2 * 2);
+    const auto w3 = LoadInt<uint16_t>(in + 3 * 2);
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 4) & mask;
+    out[2] = (w0 >> 8) & mask;
+    out[3] = w0 >> 12;
+    out[4] = (w1) & mask;
+    out[5] = (w1 >> 4) & mask;
+    out[6] = (w1 >> 8) & mask;
+    out[7] = w1 >> 12;
+    out[8] = (w2) & mask;
+    out[9] = (w2 >> 4) & mask;
+    out[10] = (w2 >> 8) & mask;
+    out[11] = w2 >> 12;
+    out[12] = (w3) & mask;
+    out[13] = (w3 >> 4) & mask;
+    out[14] = (w3 >> 8) & mask;
+    out[15] = w3 >> 12;
+
+    return in + (4 * 2);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint16_t, 5> {
+
+  static constexpr int kValuesUnpacked = 16;
+
+  static const uint8_t* unpack(const uint8_t* in, uint16_t* out) {
+    constexpr uint16_t mask = ((uint16_t{1} << 5) - uint16_t{1});
+
+    const auto w0 = LoadInt<uint16_t>(in + 0 * 2);
+    const auto w1 = LoadInt<uint16_t>(in + 1 * 2);
+    const auto w2 = LoadInt<uint16_t>(in + 2 * 2);
+    const auto w3 = LoadInt<uint16_t>(in + 3 * 2);
+    const auto w4 = LoadInt<uint16_t>(in + 4 * 2);
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 5) & mask;
+    out[2] = (w0 >> 10) & mask;
+    out[3] = ((w0 >> 15) | (w1 << 1)) & mask;
+    out[4] = (w1 >> 4) & mask;
+    out[5] = (w1 >> 9) & mask;
+    out[6] = ((w1 >> 14) | (w2 << 2)) & mask;
+    out[7] = (w2 >> 3) & mask;
+    out[8] = (w2 >> 8) & mask;
+    out[9] = ((w2 >> 13) | (w3 << 3)) & mask;
+    out[10] = (w3 >> 2) & mask;
+    out[11] = (w3 >> 7) & mask;
+    out[12] = ((w3 >> 12) | (w4 << 4)) & mask;
+    out[13] = (w4 >> 1) & mask;
+    out[14] = (w4 >> 6) & mask;
+    out[15] = w4 >> 11;
+
+    return in + (5 * 2);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint16_t, 6> {
+
+  static constexpr int kValuesUnpacked = 16;
+
+  static const uint8_t* unpack(const uint8_t* in, uint16_t* out) {
+    constexpr uint16_t mask = ((uint16_t{1} << 6) - uint16_t{1});
+
+    const auto w0 = LoadInt<uint16_t>(in + 0 * 2);
+    const auto w1 = LoadInt<uint16_t>(in + 1 * 2);
+    const auto w2 = LoadInt<uint16_t>(in + 2 * 2);
+    const auto w3 = LoadInt<uint16_t>(in + 3 * 2);
+    const auto w4 = LoadInt<uint16_t>(in + 4 * 2);
+    const auto w5 = LoadInt<uint16_t>(in + 5 * 2);
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 6) & mask;
+    out[2] = ((w0 >> 12) | (w1 << 4)) & mask;
+    out[3] = (w1 >> 2) & mask;
+    out[4] = (w1 >> 8) & mask;
+    out[5] = ((w1 >> 14) | (w2 << 2)) & mask;
+    out[6] = (w2 >> 4) & mask;
+    out[7] = w2 >> 10;
+    out[8] = (w3) & mask;
+    out[9] = (w3 >> 6) & mask;
+    out[10] = ((w3 >> 12) | (w4 << 4)) & mask;
+    out[11] = (w4 >> 2) & mask;
+    out[12] = (w4 >> 8) & mask;
+    out[13] = ((w4 >> 14) | (w5 << 2)) & mask;
+    out[14] = (w5 >> 4) & mask;
+    out[15] = w5 >> 10;
+
+    return in + (6 * 2);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint16_t, 7> {
+
+  static constexpr int kValuesUnpacked = 16;
+
+  static const uint8_t* unpack(const uint8_t* in, uint16_t* out) {
+    constexpr uint16_t mask = ((uint16_t{1} << 7) - uint16_t{1});
+
+    const auto w0 = LoadInt<uint16_t>(in + 0 * 2);
+    const auto w1 = LoadInt<uint16_t>(in + 1 * 2);
+    const auto w2 = LoadInt<uint16_t>(in + 2 * 2);
+    const auto w3 = LoadInt<uint16_t>(in + 3 * 2);
+    const auto w4 = LoadInt<uint16_t>(in + 4 * 2);
+    const auto w5 = LoadInt<uint16_t>(in + 5 * 2);
+    const auto w6 = LoadInt<uint16_t>(in + 6 * 2);
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 7) & mask;
+    out[2] = ((w0 >> 14) | (w1 << 2)) & mask;
+    out[3] = (w1 >> 5) & mask;
+    out[4] = ((w1 >> 12) | (w2 << 4)) & mask;
+    out[5] = (w2 >> 3) & mask;
+    out[6] = ((w2 >> 10) | (w3 << 6)) & mask;
+    out[7] = (w3 >> 1) & mask;
+    out[8] = (w3 >> 8) & mask;
+    out[9] = ((w3 >> 15) | (w4 << 1)) & mask;
+    out[10] = (w4 >> 6) & mask;
+    out[11] = ((w4 >> 13) | (w5 << 3)) & mask;
+    out[12] = (w5 >> 4) & mask;
+    out[13] = ((w5 >> 11) | (w6 << 5)) & mask;
+    out[14] = (w6 >> 2) & mask;
+    out[15] = w6 >> 9;
+
+    return in + (7 * 2);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint16_t, 8> {
+
+  static constexpr int kValuesUnpacked = 16;
+
+  static const uint8_t* unpack(const uint8_t* in, uint16_t* out) {
+    constexpr uint16_t mask = ((uint16_t{1} << 8) - uint16_t{1});
+
+    const auto w0 = LoadInt<uint16_t>(in + 0 * 2);
+    const auto w1 = LoadInt<uint16_t>(in + 1 * 2);
+    const auto w2 = LoadInt<uint16_t>(in + 2 * 2);
+    const auto w3 = LoadInt<uint16_t>(in + 3 * 2);
+    const auto w4 = LoadInt<uint16_t>(in + 4 * 2);
+    const auto w5 = LoadInt<uint16_t>(in + 5 * 2);
+    const auto w6 = LoadInt<uint16_t>(in + 6 * 2);
+    const auto w7 = LoadInt<uint16_t>(in + 7 * 2);
+    out[0] = (w0) & mask;
+    out[1] = w0 >> 8;
+    out[2] = (w1) & mask;
+    out[3] = w1 >> 8;
+    out[4] = (w2) & mask;
+    out[5] = w2 >> 8;
+    out[6] = (w3) & mask;
+    out[7] = w3 >> 8;
+    out[8] = (w4) & mask;
+    out[9] = w4 >> 8;
+    out[10] = (w5) & mask;
+    out[11] = w5 >> 8;
+    out[12] = (w6) & mask;
+    out[13] = w6 >> 8;
+    out[14] = (w7) & mask;
+    out[15] = w7 >> 8;
+
+    return in + (8 * 2);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint16_t, 9> {
+
+  static constexpr int kValuesUnpacked = 16;
+
+  static const uint8_t* unpack(const uint8_t* in, uint16_t* out) {
+    constexpr uint16_t mask = ((uint16_t{1} << 9) - uint16_t{1});
+
+    const auto w0 = LoadInt<uint16_t>(in + 0 * 2);
+    const auto w1 = LoadInt<uint16_t>(in + 1 * 2);
+    const auto w2 = LoadInt<uint16_t>(in + 2 * 2);
+    const auto w3 = LoadInt<uint16_t>(in + 3 * 2);
+    const auto w4 = LoadInt<uint16_t>(in + 4 * 2);
+    const auto w5 = LoadInt<uint16_t>(in + 5 * 2);
+    const auto w6 = LoadInt<uint16_t>(in + 6 * 2);
+    const auto w7 = LoadInt<uint16_t>(in + 7 * 2);
+    const auto w8 = LoadInt<uint16_t>(in + 8 * 2);
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 9) | (w1 << 7)) & mask;
+    out[2] = (w1 >> 2) & mask;
+    out[3] = ((w1 >> 11) | (w2 << 5)) & mask;
+    out[4] = (w2 >> 4) & mask;
+    out[5] = ((w2 >> 13) | (w3 << 3)) & mask;
+    out[6] = (w3 >> 6) & mask;
+    out[7] = ((w3 >> 15) | (w4 << 1)) & mask;
+    out[8] = ((w4 >> 8) | (w5 << 8)) & mask;
+    out[9] = (w5 >> 1) & mask;
+    out[10] = ((w5 >> 10) | (w6 << 6)) & mask;
+    out[11] = (w6 >> 3) & mask;
+    out[12] = ((w6 >> 12) | (w7 << 4)) & mask;
+    out[13] = (w7 >> 5) & mask;
+    out[14] = ((w7 >> 14) | (w8 << 2)) & mask;
+    out[15] = w8 >> 7;
+
+    return in + (9 * 2);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint16_t, 10> {
+
+  static constexpr int kValuesUnpacked = 16;
+
+  static const uint8_t* unpack(const uint8_t* in, uint16_t* out) {
+    constexpr uint16_t mask = ((uint16_t{1} << 10) - uint16_t{1});
+
+    const auto w0 = LoadInt<uint16_t>(in + 0 * 2);
+    const auto w1 = LoadInt<uint16_t>(in + 1 * 2);
+    const auto w2 = LoadInt<uint16_t>(in + 2 * 2);
+    const auto w3 = LoadInt<uint16_t>(in + 3 * 2);
+    const auto w4 = LoadInt<uint16_t>(in + 4 * 2);
+    const auto w5 = LoadInt<uint16_t>(in + 5 * 2);
+    const auto w6 = LoadInt<uint16_t>(in + 6 * 2);
+    const auto w7 = LoadInt<uint16_t>(in + 7 * 2);
+    const auto w8 = LoadInt<uint16_t>(in + 8 * 2);
+    const auto w9 = LoadInt<uint16_t>(in + 9 * 2);
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 10) | (w1 << 6)) & mask;
+    out[2] = (w1 >> 4) & mask;
+    out[3] = ((w1 >> 14) | (w2 << 2)) & mask;
+    out[4] = ((w2 >> 8) | (w3 << 8)) & mask;
+    out[5] = (w3 >> 2) & mask;
+    out[6] = ((w3 >> 12) | (w4 << 4)) & mask;
+    out[7] = w4 >> 6;
+    out[8] = (w5) & mask;
+    out[9] = ((w5 >> 10) | (w6 << 6)) & mask;
+    out[10] = (w6 >> 4) & mask;
+    out[11] = ((w6 >> 14) | (w7 << 2)) & mask;
+    out[12] = ((w7 >> 8) | (w8 << 8)) & mask;
+    out[13] = (w8 >> 2) & mask;
+    out[14] = ((w8 >> 12) | (w9 << 4)) & mask;
+    out[15] = w9 >> 6;
+
+    return in + (10 * 2);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint16_t, 11> {
+
+  static constexpr int kValuesUnpacked = 16;
+
+  static const uint8_t* unpack(const uint8_t* in, uint16_t* out) {
+    constexpr uint16_t mask = ((uint16_t{1} << 11) - uint16_t{1});
+
+    const auto w0 = LoadInt<uint16_t>(in + 0 * 2);
+    const auto w1 = LoadInt<uint16_t>(in + 1 * 2);
+    const auto w2 = LoadInt<uint16_t>(in + 2 * 2);
+    const auto w3 = LoadInt<uint16_t>(in + 3 * 2);
+    const auto w4 = LoadInt<uint16_t>(in + 4 * 2);
+    const auto w5 = LoadInt<uint16_t>(in + 5 * 2);
+    const auto w6 = LoadInt<uint16_t>(in + 6 * 2);
+    const auto w7 = LoadInt<uint16_t>(in + 7 * 2);
+    const auto w8 = LoadInt<uint16_t>(in + 8 * 2);
+    const auto w9 = LoadInt<uint16_t>(in + 9 * 2);
+    const auto w10 = LoadInt<uint16_t>(in + 10 * 2);
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 11) | (w1 << 5)) & mask;
+    out[2] = ((w1 >> 6) | (w2 << 10)) & mask;
+    out[3] = (w2 >> 1) & mask;
+    out[4] = ((w2 >> 12) | (w3 << 4)) & mask;
+    out[5] = ((w3 >> 7) | (w4 << 9)) & mask;
+    out[6] = (w4 >> 2) & mask;
+    out[7] = ((w4 >> 13) | (w5 << 3)) & mask;
+    out[8] = ((w5 >> 8) | (w6 << 8)) & mask;
+    out[9] = (w6 >> 3) & mask;
+    out[10] = ((w6 >> 14) | (w7 << 2)) & mask;
+    out[11] = ((w7 >> 9) | (w8 << 7)) & mask;
+    out[12] = (w8 >> 4) & mask;
+    out[13] = ((w8 >> 15) | (w9 << 1)) & mask;
+    out[14] = ((w9 >> 10) | (w10 << 6)) & mask;
+    out[15] = w10 >> 5;
+
+    return in + (11 * 2);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint16_t, 12> {
+
+  static constexpr int kValuesUnpacked = 16;
+
+  static const uint8_t* unpack(const uint8_t* in, uint16_t* out) {
+    constexpr uint16_t mask = ((uint16_t{1} << 12) - uint16_t{1});
+
+    const auto w0 = LoadInt<uint16_t>(in + 0 * 2);
+    const auto w1 = LoadInt<uint16_t>(in + 1 * 2);
+    const auto w2 = LoadInt<uint16_t>(in + 2 * 2);
+    const auto w3 = LoadInt<uint16_t>(in + 3 * 2);
+    const auto w4 = LoadInt<uint16_t>(in + 4 * 2);
+    const auto w5 = LoadInt<uint16_t>(in + 5 * 2);
+    const auto w6 = LoadInt<uint16_t>(in + 6 * 2);
+    const auto w7 = LoadInt<uint16_t>(in + 7 * 2);
+    const auto w8 = LoadInt<uint16_t>(in + 8 * 2);
+    const auto w9 = LoadInt<uint16_t>(in + 9 * 2);
+    const auto w10 = LoadInt<uint16_t>(in + 10 * 2);
+    const auto w11 = LoadInt<uint16_t>(in + 11 * 2);
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 12) | (w1 << 4)) & mask;
+    out[2] = ((w1 >> 8) | (w2 << 8)) & mask;
+    out[3] = w2 >> 4;
+    out[4] = (w3) & mask;
+    out[5] = ((w3 >> 12) | (w4 << 4)) & mask;
+    out[6] = ((w4 >> 8) | (w5 << 8)) & mask;
+    out[7] = w5 >> 4;
+    out[8] = (w6) & mask;
+    out[9] = ((w6 >> 12) | (w7 << 4)) & mask;
+    out[10] = ((w7 >> 8) | (w8 << 8)) & mask;
+    out[11] = w8 >> 4;
+    out[12] = (w9) & mask;
+    out[13] = ((w9 >> 12) | (w10 << 4)) & mask;
+    out[14] = ((w10 >> 8) | (w11 << 8)) & mask;
+    out[15] = w11 >> 4;
+
+    return in + (12 * 2);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint16_t, 13> {
+
+  static constexpr int kValuesUnpacked = 16;
+
+  static const uint8_t* unpack(const uint8_t* in, uint16_t* out) {
+    constexpr uint16_t mask = ((uint16_t{1} << 13) - uint16_t{1});
+
+    const auto w0 = LoadInt<uint16_t>(in + 0 * 2);
+    const auto w1 = LoadInt<uint16_t>(in + 1 * 2);
+    const auto w2 = LoadInt<uint16_t>(in + 2 * 2);
+    const auto w3 = LoadInt<uint16_t>(in + 3 * 2);
+    const auto w4 = LoadInt<uint16_t>(in + 4 * 2);
+    const auto w5 = LoadInt<uint16_t>(in + 5 * 2);
+    const auto w6 = LoadInt<uint16_t>(in + 6 * 2);
+    const auto w7 = LoadInt<uint16_t>(in + 7 * 2);
+    const auto w8 = LoadInt<uint16_t>(in + 8 * 2);
+    const auto w9 = LoadInt<uint16_t>(in + 9 * 2);
+    const auto w10 = LoadInt<uint16_t>(in + 10 * 2);
+    const auto w11 = LoadInt<uint16_t>(in + 11 * 2);
+    const auto w12 = LoadInt<uint16_t>(in + 12 * 2);
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 13) | (w1 << 3)) & mask;
+    out[2] = ((w1 >> 10) | (w2 << 6)) & mask;
+    out[3] = ((w2 >> 7) | (w3 << 9)) & mask;
+    out[4] = ((w3 >> 4) | (w4 << 12)) & mask;
+    out[5] = (w4 >> 1) & mask;
+    out[6] = ((w4 >> 14) | (w5 << 2)) & mask;
+    out[7] = ((w5 >> 11) | (w6 << 5)) & mask;
+    out[8] = ((w6 >> 8) | (w7 << 8)) & mask;
+    out[9] = ((w7 >> 5) | (w8 << 11)) & mask;
+    out[10] = (w8 >> 2) & mask;
+    out[11] = ((w8 >> 15) | (w9 << 1)) & mask;
+    out[12] = ((w9 >> 12) | (w10 << 4)) & mask;
+    out[13] = ((w10 >> 9) | (w11 << 7)) & mask;
+    out[14] = ((w11 >> 6) | (w12 << 10)) & mask;
+    out[15] = w12 >> 3;
+
+    return in + (13 * 2);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint16_t, 14> {
+
+  static constexpr int kValuesUnpacked = 16;
+
+  static const uint8_t* unpack(const uint8_t* in, uint16_t* out) {
+    constexpr uint16_t mask = ((uint16_t{1} << 14) - uint16_t{1});
+
+    const auto w0 = LoadInt<uint16_t>(in + 0 * 2);
+    const auto w1 = LoadInt<uint16_t>(in + 1 * 2);
+    const auto w2 = LoadInt<uint16_t>(in + 2 * 2);
+    const auto w3 = LoadInt<uint16_t>(in + 3 * 2);
+    const auto w4 = LoadInt<uint16_t>(in + 4 * 2);
+    const auto w5 = LoadInt<uint16_t>(in + 5 * 2);
+    const auto w6 = LoadInt<uint16_t>(in + 6 * 2);
+    const auto w7 = LoadInt<uint16_t>(in + 7 * 2);
+    const auto w8 = LoadInt<uint16_t>(in + 8 * 2);
+    const auto w9 = LoadInt<uint16_t>(in + 9 * 2);
+    const auto w10 = LoadInt<uint16_t>(in + 10 * 2);
+    const auto w11 = LoadInt<uint16_t>(in + 11 * 2);
+    const auto w12 = LoadInt<uint16_t>(in + 12 * 2);
+    const auto w13 = LoadInt<uint16_t>(in + 13 * 2);
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 14) | (w1 << 2)) & mask;
+    out[2] = ((w1 >> 12) | (w2 << 4)) & mask;
+    out[3] = ((w2 >> 10) | (w3 << 6)) & mask;
+    out[4] = ((w3 >> 8) | (w4 << 8)) & mask;
+    out[5] = ((w4 >> 6) | (w5 << 10)) & mask;
+    out[6] = ((w5 >> 4) | (w6 << 12)) & mask;
+    out[7] = w6 >> 2;
+    out[8] = (w7) & mask;
+    out[9] = ((w7 >> 14) | (w8 << 2)) & mask;
+    out[10] = ((w8 >> 12) | (w9 << 4)) & mask;
+    out[11] = ((w9 >> 10) | (w10 << 6)) & mask;
+    out[12] = ((w10 >> 8) | (w11 << 8)) & mask;
+    out[13] = ((w11 >> 6) | (w12 << 10)) & mask;
+    out[14] = ((w12 >> 4) | (w13 << 12)) & mask;
+    out[15] = w13 >> 2;
+
+    return in + (14 * 2);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint16_t, 15> {
+
+  static constexpr int kValuesUnpacked = 16;
+
+  static const uint8_t* unpack(const uint8_t* in, uint16_t* out) {
+    constexpr uint16_t mask = ((uint16_t{1} << 15) - uint16_t{1});
+
+    const auto w0 = LoadInt<uint16_t>(in + 0 * 2);
+    const auto w1 = LoadInt<uint16_t>(in + 1 * 2);
+    const auto w2 = LoadInt<uint16_t>(in + 2 * 2);
+    const auto w3 = LoadInt<uint16_t>(in + 3 * 2);
+    const auto w4 = LoadInt<uint16_t>(in + 4 * 2);
+    const auto w5 = LoadInt<uint16_t>(in + 5 * 2);
+    const auto w6 = LoadInt<uint16_t>(in + 6 * 2);
+    const auto w7 = LoadInt<uint16_t>(in + 7 * 2);
+    const auto w8 = LoadInt<uint16_t>(in + 8 * 2);
+    const auto w9 = LoadInt<uint16_t>(in + 9 * 2);
+    const auto w10 = LoadInt<uint16_t>(in + 10 * 2);
+    const auto w11 = LoadInt<uint16_t>(in + 11 * 2);
+    const auto w12 = LoadInt<uint16_t>(in + 12 * 2);
+    const auto w13 = LoadInt<uint16_t>(in + 13 * 2);
+    const auto w14 = LoadInt<uint16_t>(in + 14 * 2);
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 15) | (w1 << 1)) & mask;
+    out[2] = ((w1 >> 14) | (w2 << 2)) & mask;
+    out[3] = ((w2 >> 13) | (w3 << 3)) & mask;
+    out[4] = ((w3 >> 12) | (w4 << 4)) & mask;
+    out[5] = ((w4 >> 11) | (w5 << 5)) & mask;
+    out[6] = ((w5 >> 10) | (w6 << 6)) & mask;
+    out[7] = ((w6 >> 9) | (w7 << 7)) & mask;
+    out[8] = ((w7 >> 8) | (w8 << 8)) & mask;
+    out[9] = ((w8 >> 7) | (w9 << 9)) & mask;
+    out[10] = ((w9 >> 6) | (w10 << 10)) & mask;
+    out[11] = ((w10 >> 5) | (w11 << 11)) & mask;
+    out[12] = ((w11 >> 4) | (w12 << 12)) & mask;
+    out[13] = ((w12 >> 3) | (w13 << 13)) & mask;
+    out[14] = ((w13 >> 2) | (w14 << 14)) & mask;
+    out[15] = w14 >> 1;
+
+    return in + (15 * 2);
+  }
+};
+
+
+
+template<>
+struct ScalarUnpackerForWidth<uint32_t, 1> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t mask = ((uint32_t{1} << 1) - uint32_t{1});
+
+    const auto w0 = LoadInt<uint32_t>(in + 0 * 4);
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 1) & mask;
+    out[2] = (w0 >> 2) & mask;
+    out[3] = (w0 >> 3) & mask;
+    out[4] = (w0 >> 4) & mask;
+    out[5] = (w0 >> 5) & mask;
+    out[6] = (w0 >> 6) & mask;
+    out[7] = (w0 >> 7) & mask;
+    out[8] = (w0 >> 8) & mask;
+    out[9] = (w0 >> 9) & mask;
+    out[10] = (w0 >> 10) & mask;
+    out[11] = (w0 >> 11) & mask;
+    out[12] = (w0 >> 12) & mask;
+    out[13] = (w0 >> 13) & mask;
+    out[14] = (w0 >> 14) & mask;
+    out[15] = (w0 >> 15) & mask;
+    out[16] = (w0 >> 16) & mask;
+    out[17] = (w0 >> 17) & mask;
+    out[18] = (w0 >> 18) & mask;
+    out[19] = (w0 >> 19) & mask;
+    out[20] = (w0 >> 20) & mask;
+    out[21] = (w0 >> 21) & mask;
+    out[22] = (w0 >> 22) & mask;
+    out[23] = (w0 >> 23) & mask;
+    out[24] = (w0 >> 24) & mask;
+    out[25] = (w0 >> 25) & mask;
+    out[26] = (w0 >> 26) & mask;
+    out[27] = (w0 >> 27) & mask;
+    out[28] = (w0 >> 28) & mask;
+    out[29] = (w0 >> 29) & mask;
+    out[30] = (w0 >> 30) & mask;
+    out[31] = w0 >> 31;
+
+    return in + (1 * 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint32_t, 2> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t mask = ((uint32_t{1} << 2) - uint32_t{1});
+
+    const auto w0 = LoadInt<uint32_t>(in + 0 * 4);
+    const auto w1 = LoadInt<uint32_t>(in + 1 * 4);
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 2) & mask;
+    out[2] = (w0 >> 4) & mask;
+    out[3] = (w0 >> 6) & mask;
+    out[4] = (w0 >> 8) & mask;
+    out[5] = (w0 >> 10) & mask;
+    out[6] = (w0 >> 12) & mask;
+    out[7] = (w0 >> 14) & mask;
+    out[8] = (w0 >> 16) & mask;
+    out[9] = (w0 >> 18) & mask;
+    out[10] = (w0 >> 20) & mask;
+    out[11] = (w0 >> 22) & mask;
+    out[12] = (w0 >> 24) & mask;
+    out[13] = (w0 >> 26) & mask;
+    out[14] = (w0 >> 28) & mask;
+    out[15] = w0 >> 30;
+    out[16] = (w1) & mask;
+    out[17] = (w1 >> 2) & mask;
+    out[18] = (w1 >> 4) & mask;
+    out[19] = (w1 >> 6) & mask;
+    out[20] = (w1 >> 8) & mask;
+    out[21] = (w1 >> 10) & mask;
+    out[22] = (w1 >> 12) & mask;
+    out[23] = (w1 >> 14) & mask;
+    out[24] = (w1 >> 16) & mask;
+    out[25] = (w1 >> 18) & mask;
+    out[26] = (w1 >> 20) & mask;
+    out[27] = (w1 >> 22) & mask;
+    out[28] = (w1 >> 24) & mask;
+    out[29] = (w1 >> 26) & mask;
+    out[30] = (w1 >> 28) & mask;
+    out[31] = w1 >> 30;
+
+    return in + (2 * 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint32_t, 3> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t mask = ((uint32_t{1} << 3) - uint32_t{1});
+
+    const auto w0 = LoadInt<uint32_t>(in + 0 * 4);
+    const auto w1 = LoadInt<uint32_t>(in + 1 * 4);
+    const auto w2 = LoadInt<uint32_t>(in + 2 * 4);
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 3) & mask;
+    out[2] = (w0 >> 6) & mask;
+    out[3] = (w0 >> 9) & mask;
+    out[4] = (w0 >> 12) & mask;
+    out[5] = (w0 >> 15) & mask;
+    out[6] = (w0 >> 18) & mask;
+    out[7] = (w0 >> 21) & mask;
+    out[8] = (w0 >> 24) & mask;
+    out[9] = (w0 >> 27) & mask;
+    out[10] = ((w0 >> 30) | (w1 << 2)) & mask;
+    out[11] = (w1 >> 1) & mask;
+    out[12] = (w1 >> 4) & mask;
+    out[13] = (w1 >> 7) & mask;
+    out[14] = (w1 >> 10) & mask;
+    out[15] = (w1 >> 13) & mask;
+    out[16] = (w1 >> 16) & mask;
+    out[17] = (w1 >> 19) & mask;
+    out[18] = (w1 >> 22) & mask;
+    out[19] = (w1 >> 25) & mask;
+    out[20] = (w1 >> 28) & mask;
+    out[21] = ((w1 >> 31) | (w2 << 1)) & mask;
+    out[22] = (w2 >> 2) & mask;
+    out[23] = (w2 >> 5) & mask;
+    out[24] = (w2 >> 8) & mask;
+    out[25] = (w2 >> 11) & mask;
+    out[26] = (w2 >> 14) & mask;
+    out[27] = (w2 >> 17) & mask;
+    out[28] = (w2 >> 20) & mask;
+    out[29] = (w2 >> 23) & mask;
+    out[30] = (w2 >> 26) & mask;
+    out[31] = w2 >> 29;
+
+    return in + (3 * 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint32_t, 4> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t mask = ((uint32_t{1} << 4) - uint32_t{1});
+
+    const auto w0 = LoadInt<uint32_t>(in + 0 * 4);
+    const auto w1 = LoadInt<uint32_t>(in + 1 * 4);
+    const auto w2 = LoadInt<uint32_t>(in + 2 * 4);
+    const auto w3 = LoadInt<uint32_t>(in + 3 * 4);
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 4) & mask;
+    out[2] = (w0 >> 8) & mask;
+    out[3] = (w0 >> 12) & mask;
+    out[4] = (w0 >> 16) & mask;
+    out[5] = (w0 >> 20) & mask;
+    out[6] = (w0 >> 24) & mask;
+    out[7] = w0 >> 28;
+    out[8] = (w1) & mask;
+    out[9] = (w1 >> 4) & mask;
+    out[10] = (w1 >> 8) & mask;
+    out[11] = (w1 >> 12) & mask;
+    out[12] = (w1 >> 16) & mask;
+    out[13] = (w1 >> 20) & mask;
+    out[14] = (w1 >> 24) & mask;
+    out[15] = w1 >> 28;
+    out[16] = (w2) & mask;
+    out[17] = (w2 >> 4) & mask;
+    out[18] = (w2 >> 8) & mask;
+    out[19] = (w2 >> 12) & mask;
+    out[20] = (w2 >> 16) & mask;
+    out[21] = (w2 >> 20) & mask;
+    out[22] = (w2 >> 24) & mask;
+    out[23] = w2 >> 28;
+    out[24] = (w3) & mask;
+    out[25] = (w3 >> 4) & mask;
+    out[26] = (w3 >> 8) & mask;
+    out[27] = (w3 >> 12) & mask;
+    out[28] = (w3 >> 16) & mask;
+    out[29] = (w3 >> 20) & mask;
+    out[30] = (w3 >> 24) & mask;
+    out[31] = w3 >> 28;
+
+    return in + (4 * 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint32_t, 5> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t mask = ((uint32_t{1} << 5) - uint32_t{1});
+
+    const auto w0 = LoadInt<uint32_t>(in + 0 * 4);
+    const auto w1 = LoadInt<uint32_t>(in + 1 * 4);
+    const auto w2 = LoadInt<uint32_t>(in + 2 * 4);
+    const auto w3 = LoadInt<uint32_t>(in + 3 * 4);
+    const auto w4 = LoadInt<uint32_t>(in + 4 * 4);
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 5) & mask;
+    out[2] = (w0 >> 10) & mask;
+    out[3] = (w0 >> 15) & mask;
+    out[4] = (w0 >> 20) & mask;
+    out[5] = (w0 >> 25) & mask;
+    out[6] = ((w0 >> 30) | (w1 << 2)) & mask;
+    out[7] = (w1 >> 3) & mask;
+    out[8] = (w1 >> 8) & mask;
+    out[9] = (w1 >> 13) & mask;
+    out[10] = (w1 >> 18) & mask;
+    out[11] = (w1 >> 23) & mask;
+    out[12] = ((w1 >> 28) | (w2 << 4)) & mask;
+    out[13] = (w2 >> 1) & mask;
+    out[14] = (w2 >> 6) & mask;
+    out[15] = (w2 >> 11) & mask;
+    out[16] = (w2 >> 16) & mask;
+    out[17] = (w2 >> 21) & mask;
+    out[18] = (w2 >> 26) & mask;
+    out[19] = ((w2 >> 31) | (w3 << 1)) & mask;
+    out[20] = (w3 >> 4) & mask;
+    out[21] = (w3 >> 9) & mask;
+    out[22] = (w3 >> 14) & mask;
+    out[23] = (w3 >> 19) & mask;
+    out[24] = (w3 >> 24) & mask;
+    out[25] = ((w3 >> 29) | (w4 << 3)) & mask;
+    out[26] = (w4 >> 2) & mask;
+    out[27] = (w4 >> 7) & mask;
+    out[28] = (w4 >> 12) & mask;
+    out[29] = (w4 >> 17) & mask;
+    out[30] = (w4 >> 22) & mask;
+    out[31] = w4 >> 27;
+
+    return in + (5 * 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint32_t, 6> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t mask = ((uint32_t{1} << 6) - uint32_t{1});
+
+    const auto w0 = LoadInt<uint32_t>(in + 0 * 4);
+    const auto w1 = LoadInt<uint32_t>(in + 1 * 4);
+    const auto w2 = LoadInt<uint32_t>(in + 2 * 4);
+    const auto w3 = LoadInt<uint32_t>(in + 3 * 4);
+    const auto w4 = LoadInt<uint32_t>(in + 4 * 4);
+    const auto w5 = LoadInt<uint32_t>(in + 5 * 4);
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 6) & mask;
+    out[2] = (w0 >> 12) & mask;
+    out[3] = (w0 >> 18) & mask;
+    out[4] = (w0 >> 24) & mask;
+    out[5] = ((w0 >> 30) | (w1 << 2)) & mask;
+    out[6] = (w1 >> 4) & mask;
+    out[7] = (w1 >> 10) & mask;
+    out[8] = (w1 >> 16) & mask;
+    out[9] = (w1 >> 22) & mask;
+    out[10] = ((w1 >> 28) | (w2 << 4)) & mask;
+    out[11] = (w2 >> 2) & mask;
+    out[12] = (w2 >> 8) & mask;
+    out[13] = (w2 >> 14) & mask;
+    out[14] = (w2 >> 20) & mask;
+    out[15] = w2 >> 26;
+    out[16] = (w3) & mask;
+    out[17] = (w3 >> 6) & mask;
+    out[18] = (w3 >> 12) & mask;
+    out[19] = (w3 >> 18) & mask;
+    out[20] = (w3 >> 24) & mask;
+    out[21] = ((w3 >> 30) | (w4 << 2)) & mask;
+    out[22] = (w4 >> 4) & mask;
+    out[23] = (w4 >> 10) & mask;
+    out[24] = (w4 >> 16) & mask;
+    out[25] = (w4 >> 22) & mask;
+    out[26] = ((w4 >> 28) | (w5 << 4)) & mask;
+    out[27] = (w5 >> 2) & mask;
+    out[28] = (w5 >> 8) & mask;
+    out[29] = (w5 >> 14) & mask;
+    out[30] = (w5 >> 20) & mask;
+    out[31] = w5 >> 26;
+
+    return in + (6 * 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint32_t, 7> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t mask = ((uint32_t{1} << 7) - uint32_t{1});
+
+    const auto w0 = LoadInt<uint32_t>(in + 0 * 4);
+    const auto w1 = LoadInt<uint32_t>(in + 1 * 4);
+    const auto w2 = LoadInt<uint32_t>(in + 2 * 4);
+    const auto w3 = LoadInt<uint32_t>(in + 3 * 4);
+    const auto w4 = LoadInt<uint32_t>(in + 4 * 4);
+    const auto w5 = LoadInt<uint32_t>(in + 5 * 4);
+    const auto w6 = LoadInt<uint32_t>(in + 6 * 4);
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 7) & mask;
+    out[2] = (w0 >> 14) & mask;
+    out[3] = (w0 >> 21) & mask;
+    out[4] = ((w0 >> 28) | (w1 << 4)) & mask;
+    out[5] = (w1 >> 3) & mask;
+    out[6] = (w1 >> 10) & mask;
+    out[7] = (w1 >> 17) & mask;
+    out[8] = (w1 >> 24) & mask;
+    out[9] = ((w1 >> 31) | (w2 << 1)) & mask;
+    out[10] = (w2 >> 6) & mask;
+    out[11] = (w2 >> 13) & mask;
+    out[12] = (w2 >> 20) & mask;
+    out[13] = ((w2 >> 27) | (w3 << 5)) & mask;
+    out[14] = (w3 >> 2) & mask;
+    out[15] = (w3 >> 9) & mask;
+    out[16] = (w3 >> 16) & mask;
+    out[17] = (w3 >> 23) & mask;
+    out[18] = ((w3 >> 30) | (w4 << 2)) & mask;
+    out[19] = (w4 >> 5) & mask;
+    out[20] = (w4 >> 12) & mask;
+    out[21] = (w4 >> 19) & mask;
+    out[22] = ((w4 >> 26) | (w5 << 6)) & mask;
+    out[23] = (w5 >> 1) & mask;
+    out[24] = (w5 >> 8) & mask;
+    out[25] = (w5 >> 15) & mask;
+    out[26] = (w5 >> 22) & mask;
+    out[27] = ((w5 >> 29) | (w6 << 3)) & mask;
+    out[28] = (w6 >> 4) & mask;
+    out[29] = (w6 >> 11) & mask;
+    out[30] = (w6 >> 18) & mask;
+    out[31] = w6 >> 25;
+
+    return in + (7 * 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint32_t, 8> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t mask = ((uint32_t{1} << 8) - uint32_t{1});
+
+    const auto w0 = LoadInt<uint32_t>(in + 0 * 4);
+    const auto w1 = LoadInt<uint32_t>(in + 1 * 4);
+    const auto w2 = LoadInt<uint32_t>(in + 2 * 4);
+    const auto w3 = LoadInt<uint32_t>(in + 3 * 4);
+    const auto w4 = LoadInt<uint32_t>(in + 4 * 4);
+    const auto w5 = LoadInt<uint32_t>(in + 5 * 4);
+    const auto w6 = LoadInt<uint32_t>(in + 6 * 4);
+    const auto w7 = LoadInt<uint32_t>(in + 7 * 4);
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 8) & mask;
+    out[2] = (w0 >> 16) & mask;
+    out[3] = w0 >> 24;
+    out[4] = (w1) & mask;
+    out[5] = (w1 >> 8) & mask;
+    out[6] = (w1 >> 16) & mask;
+    out[7] = w1 >> 24;
+    out[8] = (w2) & mask;
+    out[9] = (w2 >> 8) & mask;
+    out[10] = (w2 >> 16) & mask;
+    out[11] = w2 >> 24;
+    out[12] = (w3) & mask;
+    out[13] = (w3 >> 8) & mask;
+    out[14] = (w3 >> 16) & mask;
+    out[15] = w3 >> 24;
+    out[16] = (w4) & mask;
+    out[17] = (w4 >> 8) & mask;
+    out[18] = (w4 >> 16) & mask;
+    out[19] = w4 >> 24;
+    out[20] = (w5) & mask;
+    out[21] = (w5 >> 8) & mask;
+    out[22] = (w5 >> 16) & mask;
+    out[23] = w5 >> 24;
+    out[24] = (w6) & mask;
+    out[25] = (w6 >> 8) & mask;
+    out[26] = (w6 >> 16) & mask;
+    out[27] = w6 >> 24;
+    out[28] = (w7) & mask;
+    out[29] = (w7 >> 8) & mask;
+    out[30] = (w7 >> 16) & mask;
+    out[31] = w7 >> 24;
+
+    return in + (8 * 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint32_t, 9> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t mask = ((uint32_t{1} << 9) - uint32_t{1});
+
+    const auto w0 = LoadInt<uint32_t>(in + 0 * 4);
+    const auto w1 = LoadInt<uint32_t>(in + 1 * 4);
+    const auto w2 = LoadInt<uint32_t>(in + 2 * 4);
+    const auto w3 = LoadInt<uint32_t>(in + 3 * 4);
+    const auto w4 = LoadInt<uint32_t>(in + 4 * 4);
+    const auto w5 = LoadInt<uint32_t>(in + 5 * 4);
+    const auto w6 = LoadInt<uint32_t>(in + 6 * 4);
+    const auto w7 = LoadInt<uint32_t>(in + 7 * 4);
+    const auto w8 = LoadInt<uint32_t>(in + 8 * 4);
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 9) & mask;
+    out[2] = (w0 >> 18) & mask;
+    out[3] = ((w0 >> 27) | (w1 << 5)) & mask;
+    out[4] = (w1 >> 4) & mask;
+    out[5] = (w1 >> 13) & mask;
+    out[6] = (w1 >> 22) & mask;
+    out[7] = ((w1 >> 31) | (w2 << 1)) & mask;
+    out[8] = (w2 >> 8) & mask;
+    out[9] = (w2 >> 17) & mask;
+    out[10] = ((w2 >> 26) | (w3 << 6)) & mask;
+    out[11] = (w3 >> 3) & mask;
+    out[12] = (w3 >> 12) & mask;
+    out[13] = (w3 >> 21) & mask;
+    out[14] = ((w3 >> 30) | (w4 << 2)) & mask;
+    out[15] = (w4 >> 7) & mask;
+    out[16] = (w4 >> 16) & mask;
+    out[17] = ((w4 >> 25) | (w5 << 7)) & mask;
+    out[18] = (w5 >> 2) & mask;
+    out[19] = (w5 >> 11) & mask;
+    out[20] = (w5 >> 20) & mask;
+    out[21] = ((w5 >> 29) | (w6 << 3)) & mask;
+    out[22] = (w6 >> 6) & mask;
+    out[23] = (w6 >> 15) & mask;
+    out[24] = ((w6 >> 24) | (w7 << 8)) & mask;
+    out[25] = (w7 >> 1) & mask;
+    out[26] = (w7 >> 10) & mask;
+    out[27] = (w7 >> 19) & mask;
+    out[28] = ((w7 >> 28) | (w8 << 4)) & mask;
+    out[29] = (w8 >> 5) & mask;
+    out[30] = (w8 >> 14) & mask;
+    out[31] = w8 >> 23;
+
+    return in + (9 * 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint32_t, 10> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t mask = ((uint32_t{1} << 10) - uint32_t{1});
+
+    const auto w0 = LoadInt<uint32_t>(in + 0 * 4);
+    const auto w1 = LoadInt<uint32_t>(in + 1 * 4);
+    const auto w2 = LoadInt<uint32_t>(in + 2 * 4);
+    const auto w3 = LoadInt<uint32_t>(in + 3 * 4);
+    const auto w4 = LoadInt<uint32_t>(in + 4 * 4);
+    const auto w5 = LoadInt<uint32_t>(in + 5 * 4);
+    const auto w6 = LoadInt<uint32_t>(in + 6 * 4);
+    const auto w7 = LoadInt<uint32_t>(in + 7 * 4);
+    const auto w8 = LoadInt<uint32_t>(in + 8 * 4);
+    const auto w9 = LoadInt<uint32_t>(in + 9 * 4);
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 10) & mask;
+    out[2] = (w0 >> 20) & mask;
+    out[3] = ((w0 >> 30) | (w1 << 2)) & mask;
+    out[4] = (w1 >> 8) & mask;
+    out[5] = (w1 >> 18) & mask;
+    out[6] = ((w1 >> 28) | (w2 << 4)) & mask;
+    out[7] = (w2 >> 6) & mask;
+    out[8] = (w2 >> 16) & mask;
+    out[9] = ((w2 >> 26) | (w3 << 6)) & mask;
+    out[10] = (w3 >> 4) & mask;
+    out[11] = (w3 >> 14) & mask;
+    out[12] = ((w3 >> 24) | (w4 << 8)) & mask;
+    out[13] = (w4 >> 2) & mask;
+    out[14] = (w4 >> 12) & mask;
+    out[15] = w4 >> 22;
+    out[16] = (w5) & mask;
+    out[17] = (w5 >> 10) & mask;
+    out[18] = (w5 >> 20) & mask;
+    out[19] = ((w5 >> 30) | (w6 << 2)) & mask;
+    out[20] = (w6 >> 8) & mask;
+    out[21] = (w6 >> 18) & mask;
+    out[22] = ((w6 >> 28) | (w7 << 4)) & mask;
+    out[23] = (w7 >> 6) & mask;
+    out[24] = (w7 >> 16) & mask;
+    out[25] = ((w7 >> 26) | (w8 << 6)) & mask;
+    out[26] = (w8 >> 4) & mask;
+    out[27] = (w8 >> 14) & mask;
+    out[28] = ((w8 >> 24) | (w9 << 8)) & mask;
+    out[29] = (w9 >> 2) & mask;
+    out[30] = (w9 >> 12) & mask;
+    out[31] = w9 >> 22;
+
+    return in + (10 * 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint32_t, 11> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t mask = ((uint32_t{1} << 11) - uint32_t{1});
+
+    const auto w0 = LoadInt<uint32_t>(in + 0 * 4);
+    const auto w1 = LoadInt<uint32_t>(in + 1 * 4);
+    const auto w2 = LoadInt<uint32_t>(in + 2 * 4);
+    const auto w3 = LoadInt<uint32_t>(in + 3 * 4);
+    const auto w4 = LoadInt<uint32_t>(in + 4 * 4);
+    const auto w5 = LoadInt<uint32_t>(in + 5 * 4);
+    const auto w6 = LoadInt<uint32_t>(in + 6 * 4);
+    const auto w7 = LoadInt<uint32_t>(in + 7 * 4);
+    const auto w8 = LoadInt<uint32_t>(in + 8 * 4);
+    const auto w9 = LoadInt<uint32_t>(in + 9 * 4);
+    const auto w10 = LoadInt<uint32_t>(in + 10 * 4);
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 11) & mask;
+    out[2] = ((w0 >> 22) | (w1 << 10)) & mask;
+    out[3] = (w1 >> 1) & mask;
+    out[4] = (w1 >> 12) & mask;
+    out[5] = ((w1 >> 23) | (w2 << 9)) & mask;
+    out[6] = (w2 >> 2) & mask;
+    out[7] = (w2 >> 13) & mask;
+    out[8] = ((w2 >> 24) | (w3 << 8)) & mask;
+    out[9] = (w3 >> 3) & mask;
+    out[10] = (w3 >> 14) & mask;
+    out[11] = ((w3 >> 25) | (w4 << 7)) & mask;
+    out[12] = (w4 >> 4) & mask;
+    out[13] = (w4 >> 15) & mask;
+    out[14] = ((w4 >> 26) | (w5 << 6)) & mask;
+    out[15] = (w5 >> 5) & mask;
+    out[16] = (w5 >> 16) & mask;
+    out[17] = ((w5 >> 27) | (w6 << 5)) & mask;
+    out[18] = (w6 >> 6) & mask;
+    out[19] = (w6 >> 17) & mask;
+    out[20] = ((w6 >> 28) | (w7 << 4)) & mask;
+    out[21] = (w7 >> 7) & mask;
+    out[22] = (w7 >> 18) & mask;
+    out[23] = ((w7 >> 29) | (w8 << 3)) & mask;
+    out[24] = (w8 >> 8) & mask;
+    out[25] = (w8 >> 19) & mask;
+    out[26] = ((w8 >> 30) | (w9 << 2)) & mask;
+    out[27] = (w9 >> 9) & mask;
+    out[28] = (w9 >> 20) & mask;
+    out[29] = ((w9 >> 31) | (w10 << 1)) & mask;
+    out[30] = (w10 >> 10) & mask;
+    out[31] = w10 >> 21;
+
+    return in + (11 * 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint32_t, 12> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t mask = ((uint32_t{1} << 12) - uint32_t{1});
+
+    const auto w0 = LoadInt<uint32_t>(in + 0 * 4);
+    const auto w1 = LoadInt<uint32_t>(in + 1 * 4);
+    const auto w2 = LoadInt<uint32_t>(in + 2 * 4);
+    const auto w3 = LoadInt<uint32_t>(in + 3 * 4);
+    const auto w4 = LoadInt<uint32_t>(in + 4 * 4);
+    const auto w5 = LoadInt<uint32_t>(in + 5 * 4);
+    const auto w6 = LoadInt<uint32_t>(in + 6 * 4);
+    const auto w7 = LoadInt<uint32_t>(in + 7 * 4);
+    const auto w8 = LoadInt<uint32_t>(in + 8 * 4);
+    const auto w9 = LoadInt<uint32_t>(in + 9 * 4);
+    const auto w10 = LoadInt<uint32_t>(in + 10 * 4);
+    const auto w11 = LoadInt<uint32_t>(in + 11 * 4);
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 12) & mask;
+    out[2] = ((w0 >> 24) | (w1 << 8)) & mask;
+    out[3] = (w1 >> 4) & mask;
+    out[4] = (w1 >> 16) & mask;
+    out[5] = ((w1 >> 28) | (w2 << 4)) & mask;
+    out[6] = (w2 >> 8) & mask;
+    out[7] = w2 >> 20;
+    out[8] = (w3) & mask;
+    out[9] = (w3 >> 12) & mask;
+    out[10] = ((w3 >> 24) | (w4 << 8)) & mask;
+    out[11] = (w4 >> 4) & mask;
+    out[12] = (w4 >> 16) & mask;
+    out[13] = ((w4 >> 28) | (w5 << 4)) & mask;
+    out[14] = (w5 >> 8) & mask;
+    out[15] = w5 >> 20;
+    out[16] = (w6) & mask;
+    out[17] = (w6 >> 12) & mask;
+    out[18] = ((w6 >> 24) | (w7 << 8)) & mask;
+    out[19] = (w7 >> 4) & mask;
+    out[20] = (w7 >> 16) & mask;
+    out[21] = ((w7 >> 28) | (w8 << 4)) & mask;
+    out[22] = (w8 >> 8) & mask;
+    out[23] = w8 >> 20;
+    out[24] = (w9) & mask;
+    out[25] = (w9 >> 12) & mask;
+    out[26] = ((w9 >> 24) | (w10 << 8)) & mask;
+    out[27] = (w10 >> 4) & mask;
+    out[28] = (w10 >> 16) & mask;
+    out[29] = ((w10 >> 28) | (w11 << 4)) & mask;
+    out[30] = (w11 >> 8) & mask;
+    out[31] = w11 >> 20;
+
+    return in + (12 * 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint32_t, 13> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t mask = ((uint32_t{1} << 13) - uint32_t{1});
+
+    const auto w0 = LoadInt<uint32_t>(in + 0 * 4);
+    const auto w1 = LoadInt<uint32_t>(in + 1 * 4);
+    const auto w2 = LoadInt<uint32_t>(in + 2 * 4);
+    const auto w3 = LoadInt<uint32_t>(in + 3 * 4);
+    const auto w4 = LoadInt<uint32_t>(in + 4 * 4);
+    const auto w5 = LoadInt<uint32_t>(in + 5 * 4);
+    const auto w6 = LoadInt<uint32_t>(in + 6 * 4);
+    const auto w7 = LoadInt<uint32_t>(in + 7 * 4);
+    const auto w8 = LoadInt<uint32_t>(in + 8 * 4);
+    const auto w9 = LoadInt<uint32_t>(in + 9 * 4);
+    const auto w10 = LoadInt<uint32_t>(in + 10 * 4);
+    const auto w11 = LoadInt<uint32_t>(in + 11 * 4);
+    const auto w12 = LoadInt<uint32_t>(in + 12 * 4);
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 13) & mask;
+    out[2] = ((w0 >> 26) | (w1 << 6)) & mask;
+    out[3] = (w1 >> 7) & mask;
+    out[4] = ((w1 >> 20) | (w2 << 12)) & mask;
+    out[5] = (w2 >> 1) & mask;
+    out[6] = (w2 >> 14) & mask;
+    out[7] = ((w2 >> 27) | (w3 << 5)) & mask;
+    out[8] = (w3 >> 8) & mask;
+    out[9] = ((w3 >> 21) | (w4 << 11)) & mask;
+    out[10] = (w4 >> 2) & mask;
+    out[11] = (w4 >> 15) & mask;
+    out[12] = ((w4 >> 28) | (w5 << 4)) & mask;
+    out[13] = (w5 >> 9) & mask;
+    out[14] = ((w5 >> 22) | (w6 << 10)) & mask;
+    out[15] = (w6 >> 3) & mask;
+    out[16] = (w6 >> 16) & mask;
+    out[17] = ((w6 >> 29) | (w7 << 3)) & mask;
+    out[18] = (w7 >> 10) & mask;
+    out[19] = ((w7 >> 23) | (w8 << 9)) & mask;
+    out[20] = (w8 >> 4) & mask;
+    out[21] = (w8 >> 17) & mask;
+    out[22] = ((w8 >> 30) | (w9 << 2)) & mask;
+    out[23] = (w9 >> 11) & mask;
+    out[24] = ((w9 >> 24) | (w10 << 8)) & mask;
+    out[25] = (w10 >> 5) & mask;
+    out[26] = (w10 >> 18) & mask;
+    out[27] = ((w10 >> 31) | (w11 << 1)) & mask;
+    out[28] = (w11 >> 12) & mask;
+    out[29] = ((w11 >> 25) | (w12 << 7)) & mask;
+    out[30] = (w12 >> 6) & mask;
+    out[31] = w12 >> 19;
+
+    return in + (13 * 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint32_t, 14> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t mask = ((uint32_t{1} << 14) - uint32_t{1});
+
+    const auto w0 = LoadInt<uint32_t>(in + 0 * 4);
+    const auto w1 = LoadInt<uint32_t>(in + 1 * 4);
+    const auto w2 = LoadInt<uint32_t>(in + 2 * 4);
+    const auto w3 = LoadInt<uint32_t>(in + 3 * 4);
+    const auto w4 = LoadInt<uint32_t>(in + 4 * 4);
+    const auto w5 = LoadInt<uint32_t>(in + 5 * 4);
+    const auto w6 = LoadInt<uint32_t>(in + 6 * 4);
+    const auto w7 = LoadInt<uint32_t>(in + 7 * 4);
+    const auto w8 = LoadInt<uint32_t>(in + 8 * 4);
+    const auto w9 = LoadInt<uint32_t>(in + 9 * 4);
+    const auto w10 = LoadInt<uint32_t>(in + 10 * 4);
+    const auto w11 = LoadInt<uint32_t>(in + 11 * 4);
+    const auto w12 = LoadInt<uint32_t>(in + 12 * 4);
+    const auto w13 = LoadInt<uint32_t>(in + 13 * 4);
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 14) & mask;
+    out[2] = ((w0 >> 28) | (w1 << 4)) & mask;
+    out[3] = (w1 >> 10) & mask;
+    out[4] = ((w1 >> 24) | (w2 << 8)) & mask;
+    out[5] = (w2 >> 6) & mask;
+    out[6] = ((w2 >> 20) | (w3 << 12)) & mask;
+    out[7] = (w3 >> 2) & mask;
+    out[8] = (w3 >> 16) & mask;
+    out[9] = ((w3 >> 30) | (w4 << 2)) & mask;
+    out[10] = (w4 >> 12) & mask;
+    out[11] = ((w4 >> 26) | (w5 << 6)) & mask;
+    out[12] = (w5 >> 8) & mask;
+    out[13] = ((w5 >> 22) | (w6 << 10)) & mask;
+    out[14] = (w6 >> 4) & mask;
+    out[15] = w6 >> 18;
+    out[16] = (w7) & mask;
+    out[17] = (w7 >> 14) & mask;
+    out[18] = ((w7 >> 28) | (w8 << 4)) & mask;
+    out[19] = (w8 >> 10) & mask;
+    out[20] = ((w8 >> 24) | (w9 << 8)) & mask;
+    out[21] = (w9 >> 6) & mask;
+    out[22] = ((w9 >> 20) | (w10 << 12)) & mask;
+    out[23] = (w10 >> 2) & mask;
+    out[24] = (w10 >> 16) & mask;
+    out[25] = ((w10 >> 30) | (w11 << 2)) & mask;
+    out[26] = (w11 >> 12) & mask;
+    out[27] = ((w11 >> 26) | (w12 << 6)) & mask;
+    out[28] = (w12 >> 8) & mask;
+    out[29] = ((w12 >> 22) | (w13 << 10)) & mask;
+    out[30] = (w13 >> 4) & mask;
+    out[31] = w13 >> 18;
+
+    return in + (14 * 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint32_t, 15> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t mask = ((uint32_t{1} << 15) - uint32_t{1});
+
+    const auto w0 = LoadInt<uint32_t>(in + 0 * 4);
+    const auto w1 = LoadInt<uint32_t>(in + 1 * 4);
+    const auto w2 = LoadInt<uint32_t>(in + 2 * 4);
+    const auto w3 = LoadInt<uint32_t>(in + 3 * 4);
+    const auto w4 = LoadInt<uint32_t>(in + 4 * 4);
+    const auto w5 = LoadInt<uint32_t>(in + 5 * 4);
+    const auto w6 = LoadInt<uint32_t>(in + 6 * 4);
+    const auto w7 = LoadInt<uint32_t>(in + 7 * 4);
+    const auto w8 = LoadInt<uint32_t>(in + 8 * 4);
+    const auto w9 = LoadInt<uint32_t>(in + 9 * 4);
+    const auto w10 = LoadInt<uint32_t>(in + 10 * 4);
+    const auto w11 = LoadInt<uint32_t>(in + 11 * 4);
+    const auto w12 = LoadInt<uint32_t>(in + 12 * 4);
+    const auto w13 = LoadInt<uint32_t>(in + 13 * 4);
+    const auto w14 = LoadInt<uint32_t>(in + 14 * 4);
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 15) & mask;
+    out[2] = ((w0 >> 30) | (w1 << 2)) & mask;
+    out[3] = (w1 >> 13) & mask;
+    out[4] = ((w1 >> 28) | (w2 << 4)) & mask;
+    out[5] = (w2 >> 11) & mask;
+    out[6] = ((w2 >> 26) | (w3 << 6)) & mask;
+    out[7] = (w3 >> 9) & mask;
+    out[8] = ((w3 >> 24) | (w4 << 8)) & mask;
+    out[9] = (w4 >> 7) & mask;
+    out[10] = ((w4 >> 22) | (w5 << 10)) & mask;
+    out[11] = (w5 >> 5) & mask;
+    out[12] = ((w5 >> 20) | (w6 << 12)) & mask;
+    out[13] = (w6 >> 3) & mask;
+    out[14] = ((w6 >> 18) | (w7 << 14)) & mask;
+    out[15] = (w7 >> 1) & mask;
+    out[16] = (w7 >> 16) & mask;
+    out[17] = ((w7 >> 31) | (w8 << 1)) & mask;
+    out[18] = (w8 >> 14) & mask;
+    out[19] = ((w8 >> 29) | (w9 << 3)) & mask;
+    out[20] = (w9 >> 12) & mask;
+    out[21] = ((w9 >> 27) | (w10 << 5)) & mask;
+    out[22] = (w10 >> 10) & mask;
+    out[23] = ((w10 >> 25) | (w11 << 7)) & mask;
+    out[24] = (w11 >> 8) & mask;
+    out[25] = ((w11 >> 23) | (w12 << 9)) & mask;
+    out[26] = (w12 >> 6) & mask;
+    out[27] = ((w12 >> 21) | (w13 << 11)) & mask;
+    out[28] = (w13 >> 4) & mask;
+    out[29] = ((w13 >> 19) | (w14 << 13)) & mask;
+    out[30] = (w14 >> 2) & mask;
+    out[31] = w14 >> 17;
+
+    return in + (15 * 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint32_t, 16> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t mask = ((uint32_t{1} << 16) - uint32_t{1});
+
+    const auto w0 = LoadInt<uint32_t>(in + 0 * 4);
+    const auto w1 = LoadInt<uint32_t>(in + 1 * 4);
+    const auto w2 = LoadInt<uint32_t>(in + 2 * 4);
+    const auto w3 = LoadInt<uint32_t>(in + 3 * 4);
+    const auto w4 = LoadInt<uint32_t>(in + 4 * 4);
+    const auto w5 = LoadInt<uint32_t>(in + 5 * 4);
+    const auto w6 = LoadInt<uint32_t>(in + 6 * 4);
+    const auto w7 = LoadInt<uint32_t>(in + 7 * 4);
+    const auto w8 = LoadInt<uint32_t>(in + 8 * 4);
+    const auto w9 = LoadInt<uint32_t>(in + 9 * 4);
+    const auto w10 = LoadInt<uint32_t>(in + 10 * 4);
+    const auto w11 = LoadInt<uint32_t>(in + 11 * 4);
+    const auto w12 = LoadInt<uint32_t>(in + 12 * 4);
+    const auto w13 = LoadInt<uint32_t>(in + 13 * 4);
+    const auto w14 = LoadInt<uint32_t>(in + 14 * 4);
+    const auto w15 = LoadInt<uint32_t>(in + 15 * 4);
+    out[0] = (w0) & mask;
+    out[1] = w0 >> 16;
+    out[2] = (w1) & mask;
+    out[3] = w1 >> 16;
+    out[4] = (w2) & mask;
+    out[5] = w2 >> 16;
+    out[6] = (w3) & mask;
+    out[7] = w3 >> 16;
+    out[8] = (w4) & mask;
+    out[9] = w4 >> 16;
+    out[10] = (w5) & mask;
+    out[11] = w5 >> 16;
+    out[12] = (w6) & mask;
+    out[13] = w6 >> 16;
+    out[14] = (w7) & mask;
+    out[15] = w7 >> 16;
+    out[16] = (w8) & mask;
+    out[17] = w8 >> 16;
+    out[18] = (w9) & mask;
+    out[19] = w9 >> 16;
+    out[20] = (w10) & mask;
+    out[21] = w10 >> 16;
+    out[22] = (w11) & mask;
+    out[23] = w11 >> 16;
+    out[24] = (w12) & mask;
+    out[25] = w12 >> 16;
+    out[26] = (w13) & mask;
+    out[27] = w13 >> 16;
+    out[28] = (w14) & mask;
+    out[29] = w14 >> 16;
+    out[30] = (w15) & mask;
+    out[31] = w15 >> 16;
+
+    return in + (16 * 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint32_t, 17> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t mask = ((uint32_t{1} << 17) - uint32_t{1});
+
+    const auto w0 = LoadInt<uint32_t>(in + 0 * 4);
+    const auto w1 = LoadInt<uint32_t>(in + 1 * 4);
+    const auto w2 = LoadInt<uint32_t>(in + 2 * 4);
+    const auto w3 = LoadInt<uint32_t>(in + 3 * 4);
+    const auto w4 = LoadInt<uint32_t>(in + 4 * 4);
+    const auto w5 = LoadInt<uint32_t>(in + 5 * 4);
+    const auto w6 = LoadInt<uint32_t>(in + 6 * 4);
+    const auto w7 = LoadInt<uint32_t>(in + 7 * 4);
+    const auto w8 = LoadInt<uint32_t>(in + 8 * 4);
+    const auto w9 = LoadInt<uint32_t>(in + 9 * 4);
+    const auto w10 = LoadInt<uint32_t>(in + 10 * 4);
+    const auto w11 = LoadInt<uint32_t>(in + 11 * 4);
+    const auto w12 = LoadInt<uint32_t>(in + 12 * 4);
+    const auto w13 = LoadInt<uint32_t>(in + 13 * 4);
+    const auto w14 = LoadInt<uint32_t>(in + 14 * 4);
+    const auto w15 = LoadInt<uint32_t>(in + 15 * 4);
+    const auto w16 = LoadInt<uint32_t>(in + 16 * 4);
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 17) | (w1 << 15)) & mask;
+    out[2] = (w1 >> 2) & mask;
+    out[3] = ((w1 >> 19) | (w2 << 13)) & mask;
+    out[4] = (w2 >> 4) & mask;
+    out[5] = ((w2 >> 21) | (w3 << 11)) & mask;
+    out[6] = (w3 >> 6) & mask;
+    out[7] = ((w3 >> 23) | (w4 << 9)) & mask;
+    out[8] = (w4 >> 8) & mask;
+    out[9] = ((w4 >> 25) | (w5 << 7)) & mask;
+    out[10] = (w5 >> 10) & mask;
+    out[11] = ((w5 >> 27) | (w6 << 5)) & mask;
+    out[12] = (w6 >> 12) & mask;
+    out[13] = ((w6 >> 29) | (w7 << 3)) & mask;
+    out[14] = (w7 >> 14) & mask;
+    out[15] = ((w7 >> 31) | (w8 << 1)) & mask;
+    out[16] = ((w8 >> 16) | (w9 << 16)) & mask;
+    out[17] = (w9 >> 1) & mask;
+    out[18] = ((w9 >> 18) | (w10 << 14)) & mask;
+    out[19] = (w10 >> 3) & mask;
+    out[20] = ((w10 >> 20) | (w11 << 12)) & mask;
+    out[21] = (w11 >> 5) & mask;
+    out[22] = ((w11 >> 22) | (w12 << 10)) & mask;
+    out[23] = (w12 >> 7) & mask;
+    out[24] = ((w12 >> 24) | (w13 << 8)) & mask;
+    out[25] = (w13 >> 9) & mask;
+    out[26] = ((w13 >> 26) | (w14 << 6)) & mask;
+    out[27] = (w14 >> 11) & mask;
+    out[28] = ((w14 >> 28) | (w15 << 4)) & mask;
+    out[29] = (w15 >> 13) & mask;
+    out[30] = ((w15 >> 30) | (w16 << 2)) & mask;
+    out[31] = w16 >> 15;
+
+    return in + (17 * 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint32_t, 18> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t mask = ((uint32_t{1} << 18) - uint32_t{1});
+
+    const auto w0 = LoadInt<uint32_t>(in + 0 * 4);
+    const auto w1 = LoadInt<uint32_t>(in + 1 * 4);
+    const auto w2 = LoadInt<uint32_t>(in + 2 * 4);
+    const auto w3 = LoadInt<uint32_t>(in + 3 * 4);
+    const auto w4 = LoadInt<uint32_t>(in + 4 * 4);
+    const auto w5 = LoadInt<uint32_t>(in + 5 * 4);
+    const auto w6 = LoadInt<uint32_t>(in + 6 * 4);
+    const auto w7 = LoadInt<uint32_t>(in + 7 * 4);
+    const auto w8 = LoadInt<uint32_t>(in + 8 * 4);
+    const auto w9 = LoadInt<uint32_t>(in + 9 * 4);
+    const auto w10 = LoadInt<uint32_t>(in + 10 * 4);
+    const auto w11 = LoadInt<uint32_t>(in + 11 * 4);
+    const auto w12 = LoadInt<uint32_t>(in + 12 * 4);
+    const auto w13 = LoadInt<uint32_t>(in + 13 * 4);
+    const auto w14 = LoadInt<uint32_t>(in + 14 * 4);
+    const auto w15 = LoadInt<uint32_t>(in + 15 * 4);
+    const auto w16 = LoadInt<uint32_t>(in + 16 * 4);
+    const auto w17 = LoadInt<uint32_t>(in + 17 * 4);
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 18) | (w1 << 14)) & mask;
+    out[2] = (w1 >> 4) & mask;
+    out[3] = ((w1 >> 22) | (w2 << 10)) & mask;
+    out[4] = (w2 >> 8) & mask;
+    out[5] = ((w2 >> 26) | (w3 << 6)) & mask;
+    out[6] = (w3 >> 12) & mask;
+    out[7] = ((w3 >> 30) | (w4 << 2)) & mask;
+    out[8] = ((w4 >> 16) | (w5 << 16)) & mask;
+    out[9] = (w5 >> 2) & mask;
+    out[10] = ((w5 >> 20) | (w6 << 12)) & mask;
+    out[11] = (w6 >> 6) & mask;
+    out[12] = ((w6 >> 24) | (w7 << 8)) & mask;
+    out[13] = (w7 >> 10) & mask;
+    out[14] = ((w7 >> 28) | (w8 << 4)) & mask;
+    out[15] = w8 >> 14;
+    out[16] = (w9) & mask;
+    out[17] = ((w9 >> 18) | (w10 << 14)) & mask;
+    out[18] = (w10 >> 4) & mask;
+    out[19] = ((w10 >> 22) | (w11 << 10)) & mask;
+    out[20] = (w11 >> 8) & mask;
+    out[21] = ((w11 >> 26) | (w12 << 6)) & mask;
+    out[22] = (w12 >> 12) & mask;
+    out[23] = ((w12 >> 30) | (w13 << 2)) & mask;
+    out[24] = ((w13 >> 16) | (w14 << 16)) & mask;
+    out[25] = (w14 >> 2) & mask;
+    out[26] = ((w14 >> 20) | (w15 << 12)) & mask;
+    out[27] = (w15 >> 6) & mask;
+    out[28] = ((w15 >> 24) | (w16 << 8)) & mask;
+    out[29] = (w16 >> 10) & mask;
+    out[30] = ((w16 >> 28) | (w17 << 4)) & mask;
+    out[31] = w17 >> 14;
+
+    return in + (18 * 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint32_t, 19> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t mask = ((uint32_t{1} << 19) - uint32_t{1});
+
+    const auto w0 = LoadInt<uint32_t>(in + 0 * 4);
+    const auto w1 = LoadInt<uint32_t>(in + 1 * 4);
+    const auto w2 = LoadInt<uint32_t>(in + 2 * 4);
+    const auto w3 = LoadInt<uint32_t>(in + 3 * 4);
+    const auto w4 = LoadInt<uint32_t>(in + 4 * 4);
+    const auto w5 = LoadInt<uint32_t>(in + 5 * 4);
+    const auto w6 = LoadInt<uint32_t>(in + 6 * 4);
+    const auto w7 = LoadInt<uint32_t>(in + 7 * 4);
+    const auto w8 = LoadInt<uint32_t>(in + 8 * 4);
+    const auto w9 = LoadInt<uint32_t>(in + 9 * 4);
+    const auto w10 = LoadInt<uint32_t>(in + 10 * 4);
+    const auto w11 = LoadInt<uint32_t>(in + 11 * 4);
+    const auto w12 = LoadInt<uint32_t>(in + 12 * 4);
+    const auto w13 = LoadInt<uint32_t>(in + 13 * 4);
+    const auto w14 = LoadInt<uint32_t>(in + 14 * 4);
+    const auto w15 = LoadInt<uint32_t>(in + 15 * 4);
+    const auto w16 = LoadInt<uint32_t>(in + 16 * 4);
+    const auto w17 = LoadInt<uint32_t>(in + 17 * 4);
+    const auto w18 = LoadInt<uint32_t>(in + 18 * 4);
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 19) | (w1 << 13)) & mask;
+    out[2] = (w1 >> 6) & mask;
+    out[3] = ((w1 >> 25) | (w2 << 7)) & mask;
+    out[4] = (w2 >> 12) & mask;
+    out[5] = ((w2 >> 31) | (w3 << 1)) & mask;
+    out[6] = ((w3 >> 18) | (w4 << 14)) & mask;
+    out[7] = (w4 >> 5) & mask;
+    out[8] = ((w4 >> 24) | (w5 << 8)) & mask;
+    out[9] = (w5 >> 11) & mask;
+    out[10] = ((w5 >> 30) | (w6 << 2)) & mask;
+    out[11] = ((w6 >> 17) | (w7 << 15)) & mask;
+    out[12] = (w7 >> 4) & mask;
+    out[13] = ((w7 >> 23) | (w8 << 9)) & mask;
+    out[14] = (w8 >> 10) & mask;
+    out[15] = ((w8 >> 29) | (w9 << 3)) & mask;
+    out[16] = ((w9 >> 16) | (w10 << 16)) & mask;
+    out[17] = (w10 >> 3) & mask;
+    out[18] = ((w10 >> 22) | (w11 << 10)) & mask;
+    out[19] = (w11 >> 9) & mask;
+    out[20] = ((w11 >> 28) | (w12 << 4)) & mask;
+    out[21] = ((w12 >> 15) | (w13 << 17)) & mask;
+    out[22] = (w13 >> 2) & mask;
+    out[23] = ((w13 >> 21) | (w14 << 11)) & mask;
+    out[24] = (w14 >> 8) & mask;
+    out[25] = ((w14 >> 27) | (w15 << 5)) & mask;
+    out[26] = ((w15 >> 14) | (w16 << 18)) & mask;
+    out[27] = (w16 >> 1) & mask;
+    out[28] = ((w16 >> 20) | (w17 << 12)) & mask;
+    out[29] = (w17 >> 7) & mask;
+    out[30] = ((w17 >> 26) | (w18 << 6)) & mask;
+    out[31] = w18 >> 13;
+
+    return in + (19 * 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint32_t, 20> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t mask = ((uint32_t{1} << 20) - uint32_t{1});
+
+    const auto w0 = LoadInt<uint32_t>(in + 0 * 4);
+    const auto w1 = LoadInt<uint32_t>(in + 1 * 4);
+    const auto w2 = LoadInt<uint32_t>(in + 2 * 4);
+    const auto w3 = LoadInt<uint32_t>(in + 3 * 4);
+    const auto w4 = LoadInt<uint32_t>(in + 4 * 4);
+    const auto w5 = LoadInt<uint32_t>(in + 5 * 4);
+    const auto w6 = LoadInt<uint32_t>(in + 6 * 4);
+    const auto w7 = LoadInt<uint32_t>(in + 7 * 4);
+    const auto w8 = LoadInt<uint32_t>(in + 8 * 4);
+    const auto w9 = LoadInt<uint32_t>(in + 9 * 4);
+    const auto w10 = LoadInt<uint32_t>(in + 10 * 4);
+    const auto w11 = LoadInt<uint32_t>(in + 11 * 4);
+    const auto w12 = LoadInt<uint32_t>(in + 12 * 4);
+    const auto w13 = LoadInt<uint32_t>(in + 13 * 4);
+    const auto w14 = LoadInt<uint32_t>(in + 14 * 4);
+    const auto w15 = LoadInt<uint32_t>(in + 15 * 4);
+    const auto w16 = LoadInt<uint32_t>(in + 16 * 4);
+    const auto w17 = LoadInt<uint32_t>(in + 17 * 4);
+    const auto w18 = LoadInt<uint32_t>(in + 18 * 4);
+    const auto w19 = LoadInt<uint32_t>(in + 19 * 4);
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 20) | (w1 << 12)) & mask;
+    out[2] = (w1 >> 8) & mask;
+    out[3] = ((w1 >> 28) | (w2 << 4)) & mask;
+    out[4] = ((w2 >> 16) | (w3 << 16)) & mask;
+    out[5] = (w3 >> 4) & mask;
+    out[6] = ((w3 >> 24) | (w4 << 8)) & mask;
+    out[7] = w4 >> 12;
+    out[8] = (w5) & mask;
+    out[9] = ((w5 >> 20) | (w6 << 12)) & mask;
+    out[10] = (w6 >> 8) & mask;
+    out[11] = ((w6 >> 28) | (w7 << 4)) & mask;
+    out[12] = ((w7 >> 16) | (w8 << 16)) & mask;
+    out[13] = (w8 >> 4) & mask;
+    out[14] = ((w8 >> 24) | (w9 << 8)) & mask;
+    out[15] = w9 >> 12;
+    out[16] = (w10) & mask;
+    out[17] = ((w10 >> 20) | (w11 << 12)) & mask;
+    out[18] = (w11 >> 8) & mask;
+    out[19] = ((w11 >> 28) | (w12 << 4)) & mask;
+    out[20] = ((w12 >> 16) | (w13 << 16)) & mask;
+    out[21] = (w13 >> 4) & mask;
+    out[22] = ((w13 >> 24) | (w14 << 8)) & mask;
+    out[23] = w14 >> 12;
+    out[24] = (w15) & mask;
+    out[25] = ((w15 >> 20) | (w16 << 12)) & mask;
+    out[26] = (w16 >> 8) & mask;
+    out[27] = ((w16 >> 28) | (w17 << 4)) & mask;
+    out[28] = ((w17 >> 16) | (w18 << 16)) & mask;
+    out[29] = (w18 >> 4) & mask;
+    out[30] = ((w18 >> 24) | (w19 << 8)) & mask;
+    out[31] = w19 >> 12;
+
+    return in + (20 * 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint32_t, 21> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t mask = ((uint32_t{1} << 21) - uint32_t{1});
+
+    const auto w0 = LoadInt<uint32_t>(in + 0 * 4);
+    const auto w1 = LoadInt<uint32_t>(in + 1 * 4);
+    const auto w2 = LoadInt<uint32_t>(in + 2 * 4);
+    const auto w3 = LoadInt<uint32_t>(in + 3 * 4);
+    const auto w4 = LoadInt<uint32_t>(in + 4 * 4);
+    const auto w5 = LoadInt<uint32_t>(in + 5 * 4);
+    const auto w6 = LoadInt<uint32_t>(in + 6 * 4);
+    const auto w7 = LoadInt<uint32_t>(in + 7 * 4);
+    const auto w8 = LoadInt<uint32_t>(in + 8 * 4);
+    const auto w9 = LoadInt<uint32_t>(in + 9 * 4);
+    const auto w10 = LoadInt<uint32_t>(in + 10 * 4);
+    const auto w11 = LoadInt<uint32_t>(in + 11 * 4);
+    const auto w12 = LoadInt<uint32_t>(in + 12 * 4);
+    const auto w13 = LoadInt<uint32_t>(in + 13 * 4);
+    const auto w14 = LoadInt<uint32_t>(in + 14 * 4);
+    const auto w15 = LoadInt<uint32_t>(in + 15 * 4);
+    const auto w16 = LoadInt<uint32_t>(in + 16 * 4);
+    const auto w17 = LoadInt<uint32_t>(in + 17 * 4);
+    const auto w18 = LoadInt<uint32_t>(in + 18 * 4);
+    const auto w19 = LoadInt<uint32_t>(in + 19 * 4);
+    const auto w20 = LoadInt<uint32_t>(in + 20 * 4);
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 21) | (w1 << 11)) & mask;
+    out[2] = (w1 >> 10) & mask;
+    out[3] = ((w1 >> 31) | (w2 << 1)) & mask;
+    out[4] = ((w2 >> 20) | (w3 << 12)) & mask;
+    out[5] = (w3 >> 9) & mask;
+    out[6] = ((w3 >> 30) | (w4 << 2)) & mask;
+    out[7] = ((w4 >> 19) | (w5 << 13)) & mask;
+    out[8] = (w5 >> 8) & mask;
+    out[9] = ((w5 >> 29) | (w6 << 3)) & mask;
+    out[10] = ((w6 >> 18) | (w7 << 14)) & mask;
+    out[11] = (w7 >> 7) & mask;
+    out[12] = ((w7 >> 28) | (w8 << 4)) & mask;
+    out[13] = ((w8 >> 17) | (w9 << 15)) & mask;
+    out[14] = (w9 >> 6) & mask;
+    out[15] = ((w9 >> 27) | (w10 << 5)) & mask;
+    out[16] = ((w10 >> 16) | (w11 << 16)) & mask;
+    out[17] = (w11 >> 5) & mask;
+    out[18] = ((w11 >> 26) | (w12 << 6)) & mask;
+    out[19] = ((w12 >> 15) | (w13 << 17)) & mask;
+    out[20] = (w13 >> 4) & mask;
+    out[21] = ((w13 >> 25) | (w14 << 7)) & mask;
+    out[22] = ((w14 >> 14) | (w15 << 18)) & mask;
+    out[23] = (w15 >> 3) & mask;
+    out[24] = ((w15 >> 24) | (w16 << 8)) & mask;
+    out[25] = ((w16 >> 13) | (w17 << 19)) & mask;
+    out[26] = (w17 >> 2) & mask;
+    out[27] = ((w17 >> 23) | (w18 << 9)) & mask;
+    out[28] = ((w18 >> 12) | (w19 << 20)) & mask;
+    out[29] = (w19 >> 1) & mask;
+    out[30] = ((w19 >> 22) | (w20 << 10)) & mask;
+    out[31] = w20 >> 11;
+
+    return in + (21 * 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint32_t, 22> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t mask = ((uint32_t{1} << 22) - uint32_t{1});
+
+    const auto w0 = LoadInt<uint32_t>(in + 0 * 4);
+    const auto w1 = LoadInt<uint32_t>(in + 1 * 4);
+    const auto w2 = LoadInt<uint32_t>(in + 2 * 4);
+    const auto w3 = LoadInt<uint32_t>(in + 3 * 4);
+    const auto w4 = LoadInt<uint32_t>(in + 4 * 4);
+    const auto w5 = LoadInt<uint32_t>(in + 5 * 4);
+    const auto w6 = LoadInt<uint32_t>(in + 6 * 4);
+    const auto w7 = LoadInt<uint32_t>(in + 7 * 4);
+    const auto w8 = LoadInt<uint32_t>(in + 8 * 4);
+    const auto w9 = LoadInt<uint32_t>(in + 9 * 4);
+    const auto w10 = LoadInt<uint32_t>(in + 10 * 4);
+    const auto w11 = LoadInt<uint32_t>(in + 11 * 4);
+    const auto w12 = LoadInt<uint32_t>(in + 12 * 4);
+    const auto w13 = LoadInt<uint32_t>(in + 13 * 4);
+    const auto w14 = LoadInt<uint32_t>(in + 14 * 4);
+    const auto w15 = LoadInt<uint32_t>(in + 15 * 4);
+    const auto w16 = LoadInt<uint32_t>(in + 16 * 4);
+    const auto w17 = LoadInt<uint32_t>(in + 17 * 4);
+    const auto w18 = LoadInt<uint32_t>(in + 18 * 4);
+    const auto w19 = LoadInt<uint32_t>(in + 19 * 4);
+    const auto w20 = LoadInt<uint32_t>(in + 20 * 4);
+    const auto w21 = LoadInt<uint32_t>(in + 21 * 4);
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 22) | (w1 << 10)) & mask;
+    out[2] = ((w1 >> 12) | (w2 << 20)) & mask;
+    out[3] = (w2 >> 2) & mask;
+    out[4] = ((w2 >> 24) | (w3 << 8)) & mask;
+    out[5] = ((w3 >> 14) | (w4 << 18)) & mask;
+    out[6] = (w4 >> 4) & mask;
+    out[7] = ((w4 >> 26) | (w5 << 6)) & mask;
+    out[8] = ((w5 >> 16) | (w6 << 16)) & mask;
+    out[9] = (w6 >> 6) & mask;
+    out[10] = ((w6 >> 28) | (w7 << 4)) & mask;
+    out[11] = ((w7 >> 18) | (w8 << 14)) & mask;
+    out[12] = (w8 >> 8) & mask;
+    out[13] = ((w8 >> 30) | (w9 << 2)) & mask;
+    out[14] = ((w9 >> 20) | (w10 << 12)) & mask;
+    out[15] = w10 >> 10;
+    out[16] = (w11) & mask;
+    out[17] = ((w11 >> 22) | (w12 << 10)) & mask;
+    out[18] = ((w12 >> 12) | (w13 << 20)) & mask;
+    out[19] = (w13 >> 2) & mask;
+    out[20] = ((w13 >> 24) | (w14 << 8)) & mask;
+    out[21] = ((w14 >> 14) | (w15 << 18)) & mask;
+    out[22] = (w15 >> 4) & mask;
+    out[23] = ((w15 >> 26) | (w16 << 6)) & mask;
+    out[24] = ((w16 >> 16) | (w17 << 16)) & mask;
+    out[25] = (w17 >> 6) & mask;
+    out[26] = ((w17 >> 28) | (w18 << 4)) & mask;
+    out[27] = ((w18 >> 18) | (w19 << 14)) & mask;
+    out[28] = (w19 >> 8) & mask;
+    out[29] = ((w19 >> 30) | (w20 << 2)) & mask;
+    out[30] = ((w20 >> 20) | (w21 << 12)) & mask;
+    out[31] = w21 >> 10;
+
+    return in + (22 * 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint32_t, 23> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t mask = ((uint32_t{1} << 23) - uint32_t{1});
+
+    const auto w0 = LoadInt<uint32_t>(in + 0 * 4);
+    const auto w1 = LoadInt<uint32_t>(in + 1 * 4);
+    const auto w2 = LoadInt<uint32_t>(in + 2 * 4);
+    const auto w3 = LoadInt<uint32_t>(in + 3 * 4);
+    const auto w4 = LoadInt<uint32_t>(in + 4 * 4);
+    const auto w5 = LoadInt<uint32_t>(in + 5 * 4);
+    const auto w6 = LoadInt<uint32_t>(in + 6 * 4);
+    const auto w7 = LoadInt<uint32_t>(in + 7 * 4);
+    const auto w8 = LoadInt<uint32_t>(in + 8 * 4);
+    const auto w9 = LoadInt<uint32_t>(in + 9 * 4);
+    const auto w10 = LoadInt<uint32_t>(in + 10 * 4);
+    const auto w11 = LoadInt<uint32_t>(in + 11 * 4);
+    const auto w12 = LoadInt<uint32_t>(in + 12 * 4);
+    const auto w13 = LoadInt<uint32_t>(in + 13 * 4);
+    const auto w14 = LoadInt<uint32_t>(in + 14 * 4);
+    const auto w15 = LoadInt<uint32_t>(in + 15 * 4);
+    const auto w16 = LoadInt<uint32_t>(in + 16 * 4);
+    const auto w17 = LoadInt<uint32_t>(in + 17 * 4);
+    const auto w18 = LoadInt<uint32_t>(in + 18 * 4);
+    const auto w19 = LoadInt<uint32_t>(in + 19 * 4);
+    const auto w20 = LoadInt<uint32_t>(in + 20 * 4);
+    const auto w21 = LoadInt<uint32_t>(in + 21 * 4);
+    const auto w22 = LoadInt<uint32_t>(in + 22 * 4);
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 23) | (w1 << 9)) & mask;
+    out[2] = ((w1 >> 14) | (w2 << 18)) & mask;
+    out[3] = (w2 >> 5) & mask;
+    out[4] = ((w2 >> 28) | (w3 << 4)) & mask;
+    out[5] = ((w3 >> 19) | (w4 << 13)) & mask;
+    out[6] = ((w4 >> 10) | (w5 << 22)) & mask;
+    out[7] = (w5 >> 1) & mask;
+    out[8] = ((w5 >> 24) | (w6 << 8)) & mask;
+    out[9] = ((w6 >> 15) | (w7 << 17)) & mask;
+    out[10] = (w7 >> 6) & mask;
+    out[11] = ((w7 >> 29) | (w8 << 3)) & mask;
+    out[12] = ((w8 >> 20) | (w9 << 12)) & mask;
+    out[13] = ((w9 >> 11) | (w10 << 21)) & mask;
+    out[14] = (w10 >> 2) & mask;
+    out[15] = ((w10 >> 25) | (w11 << 7)) & mask;
+    out[16] = ((w11 >> 16) | (w12 << 16)) & mask;
+    out[17] = (w12 >> 7) & mask;
+    out[18] = ((w12 >> 30) | (w13 << 2)) & mask;
+    out[19] = ((w13 >> 21) | (w14 << 11)) & mask;
+    out[20] = ((w14 >> 12) | (w15 << 20)) & mask;
+    out[21] = (w15 >> 3) & mask;
+    out[22] = ((w15 >> 26) | (w16 << 6)) & mask;
+    out[23] = ((w16 >> 17) | (w17 << 15)) & mask;
+    out[24] = (w17 >> 8) & mask;
+    out[25] = ((w17 >> 31) | (w18 << 1)) & mask;
+    out[26] = ((w18 >> 22) | (w19 << 10)) & mask;
+    out[27] = ((w19 >> 13) | (w20 << 19)) & mask;
+    out[28] = (w20 >> 4) & mask;
+    out[29] = ((w20 >> 27) | (w21 << 5)) & mask;
+    out[30] = ((w21 >> 18) | (w22 << 14)) & mask;
+    out[31] = w22 >> 9;
+
+    return in + (23 * 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint32_t, 24> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t mask = ((uint32_t{1} << 24) - uint32_t{1});
+
+    const auto w0 = LoadInt<uint32_t>(in + 0 * 4);
+    const auto w1 = LoadInt<uint32_t>(in + 1 * 4);
+    const auto w2 = LoadInt<uint32_t>(in + 2 * 4);
+    const auto w3 = LoadInt<uint32_t>(in + 3 * 4);
+    const auto w4 = LoadInt<uint32_t>(in + 4 * 4);
+    const auto w5 = LoadInt<uint32_t>(in + 5 * 4);
+    const auto w6 = LoadInt<uint32_t>(in + 6 * 4);
+    const auto w7 = LoadInt<uint32_t>(in + 7 * 4);
+    const auto w8 = LoadInt<uint32_t>(in + 8 * 4);
+    const auto w9 = LoadInt<uint32_t>(in + 9 * 4);
+    const auto w10 = LoadInt<uint32_t>(in + 10 * 4);
+    const auto w11 = LoadInt<uint32_t>(in + 11 * 4);
+    const auto w12 = LoadInt<uint32_t>(in + 12 * 4);
+    const auto w13 = LoadInt<uint32_t>(in + 13 * 4);
+    const auto w14 = LoadInt<uint32_t>(in + 14 * 4);
+    const auto w15 = LoadInt<uint32_t>(in + 15 * 4);
+    const auto w16 = LoadInt<uint32_t>(in + 16 * 4);
+    const auto w17 = LoadInt<uint32_t>(in + 17 * 4);
+    const auto w18 = LoadInt<uint32_t>(in + 18 * 4);
+    const auto w19 = LoadInt<uint32_t>(in + 19 * 4);
+    const auto w20 = LoadInt<uint32_t>(in + 20 * 4);
+    const auto w21 = LoadInt<uint32_t>(in + 21 * 4);
+    const auto w22 = LoadInt<uint32_t>(in + 22 * 4);
+    const auto w23 = LoadInt<uint32_t>(in + 23 * 4);
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 24) | (w1 << 8)) & mask;
+    out[2] = ((w1 >> 16) | (w2 << 16)) & mask;
+    out[3] = w2 >> 8;
+    out[4] = (w3) & mask;
+    out[5] = ((w3 >> 24) | (w4 << 8)) & mask;
+    out[6] = ((w4 >> 16) | (w5 << 16)) & mask;
+    out[7] = w5 >> 8;
+    out[8] = (w6) & mask;
+    out[9] = ((w6 >> 24) | (w7 << 8)) & mask;
+    out[10] = ((w7 >> 16) | (w8 << 16)) & mask;
+    out[11] = w8 >> 8;
+    out[12] = (w9) & mask;
+    out[13] = ((w9 >> 24) | (w10 << 8)) & mask;
+    out[14] = ((w10 >> 16) | (w11 << 16)) & mask;
+    out[15] = w11 >> 8;
+    out[16] = (w12) & mask;
+    out[17] = ((w12 >> 24) | (w13 << 8)) & mask;
+    out[18] = ((w13 >> 16) | (w14 << 16)) & mask;
+    out[19] = w14 >> 8;
+    out[20] = (w15) & mask;
+    out[21] = ((w15 >> 24) | (w16 << 8)) & mask;
+    out[22] = ((w16 >> 16) | (w17 << 16)) & mask;
+    out[23] = w17 >> 8;
+    out[24] = (w18) & mask;
+    out[25] = ((w18 >> 24) | (w19 << 8)) & mask;
+    out[26] = ((w19 >> 16) | (w20 << 16)) & mask;
+    out[27] = w20 >> 8;
+    out[28] = (w21) & mask;
+    out[29] = ((w21 >> 24) | (w22 << 8)) & mask;
+    out[30] = ((w22 >> 16) | (w23 << 16)) & mask;
+    out[31] = w23 >> 8;
+
+    return in + (24 * 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint32_t, 25> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t mask = ((uint32_t{1} << 25) - uint32_t{1});
+
+    const auto w0 = LoadInt<uint32_t>(in + 0 * 4);
+    const auto w1 = LoadInt<uint32_t>(in + 1 * 4);
+    const auto w2 = LoadInt<uint32_t>(in + 2 * 4);
+    const auto w3 = LoadInt<uint32_t>(in + 3 * 4);
+    const auto w4 = LoadInt<uint32_t>(in + 4 * 4);
+    const auto w5 = LoadInt<uint32_t>(in + 5 * 4);
+    const auto w6 = LoadInt<uint32_t>(in + 6 * 4);
+    const auto w7 = LoadInt<uint32_t>(in + 7 * 4);
+    const auto w8 = LoadInt<uint32_t>(in + 8 * 4);
+    const auto w9 = LoadInt<uint32_t>(in + 9 * 4);
+    const auto w10 = LoadInt<uint32_t>(in + 10 * 4);
+    const auto w11 = LoadInt<uint32_t>(in + 11 * 4);
+    const auto w12 = LoadInt<uint32_t>(in + 12 * 4);
+    const auto w13 = LoadInt<uint32_t>(in + 13 * 4);
+    const auto w14 = LoadInt<uint32_t>(in + 14 * 4);
+    const auto w15 = LoadInt<uint32_t>(in + 15 * 4);
+    const auto w16 = LoadInt<uint32_t>(in + 16 * 4);
+    const auto w17 = LoadInt<uint32_t>(in + 17 * 4);
+    const auto w18 = LoadInt<uint32_t>(in + 18 * 4);
+    const auto w19 = LoadInt<uint32_t>(in + 19 * 4);
+    const auto w20 = LoadInt<uint32_t>(in + 20 * 4);
+    const auto w21 = LoadInt<uint32_t>(in + 21 * 4);
+    const auto w22 = LoadInt<uint32_t>(in + 22 * 4);
+    const auto w23 = LoadInt<uint32_t>(in + 23 * 4);
+    const auto w24 = LoadInt<uint32_t>(in + 24 * 4);
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 25) | (w1 << 7)) & mask;
+    out[2] = ((w1 >> 18) | (w2 << 14)) & mask;
+    out[3] = ((w2 >> 11) | (w3 << 21)) & mask;
+    out[4] = (w3 >> 4) & mask;
+    out[5] = ((w3 >> 29) | (w4 << 3)) & mask;
+    out[6] = ((w4 >> 22) | (w5 << 10)) & mask;
+    out[7] = ((w5 >> 15) | (w6 << 17)) & mask;
+    out[8] = ((w6 >> 8) | (w7 << 24)) & mask;
+    out[9] = (w7 >> 1) & mask;
+    out[10] = ((w7 >> 26) | (w8 << 6)) & mask;
+    out[11] = ((w8 >> 19) | (w9 << 13)) & mask;
+    out[12] = ((w9 >> 12) | (w10 << 20)) & mask;
+    out[13] = (w10 >> 5) & mask;
+    out[14] = ((w10 >> 30) | (w11 << 2)) & mask;
+    out[15] = ((w11 >> 23) | (w12 << 9)) & mask;
+    out[16] = ((w12 >> 16) | (w13 << 16)) & mask;
+    out[17] = ((w13 >> 9) | (w14 << 23)) & mask;
+    out[18] = (w14 >> 2) & mask;
+    out[19] = ((w14 >> 27) | (w15 << 5)) & mask;
+    out[20] = ((w15 >> 20) | (w16 << 12)) & mask;
+    out[21] = ((w16 >> 13) | (w17 << 19)) & mask;
+    out[22] = (w17 >> 6) & mask;
+    out[23] = ((w17 >> 31) | (w18 << 1)) & mask;
+    out[24] = ((w18 >> 24) | (w19 << 8)) & mask;
+    out[25] = ((w19 >> 17) | (w20 << 15)) & mask;
+    out[26] = ((w20 >> 10) | (w21 << 22)) & mask;
+    out[27] = (w21 >> 3) & mask;
+    out[28] = ((w21 >> 28) | (w22 << 4)) & mask;
+    out[29] = ((w22 >> 21) | (w23 << 11)) & mask;
+    out[30] = ((w23 >> 14) | (w24 << 18)) & mask;
+    out[31] = w24 >> 7;
+
+    return in + (25 * 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint32_t, 26> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t mask = ((uint32_t{1} << 26) - uint32_t{1});
+
+    const auto w0 = LoadInt<uint32_t>(in + 0 * 4);
+    const auto w1 = LoadInt<uint32_t>(in + 1 * 4);
+    const auto w2 = LoadInt<uint32_t>(in + 2 * 4);
+    const auto w3 = LoadInt<uint32_t>(in + 3 * 4);
+    const auto w4 = LoadInt<uint32_t>(in + 4 * 4);
+    const auto w5 = LoadInt<uint32_t>(in + 5 * 4);
+    const auto w6 = LoadInt<uint32_t>(in + 6 * 4);
+    const auto w7 = LoadInt<uint32_t>(in + 7 * 4);
+    const auto w8 = LoadInt<uint32_t>(in + 8 * 4);
+    const auto w9 = LoadInt<uint32_t>(in + 9 * 4);
+    const auto w10 = LoadInt<uint32_t>(in + 10 * 4);
+    const auto w11 = LoadInt<uint32_t>(in + 11 * 4);
+    const auto w12 = LoadInt<uint32_t>(in + 12 * 4);
+    const auto w13 = LoadInt<uint32_t>(in + 13 * 4);
+    const auto w14 = LoadInt<uint32_t>(in + 14 * 4);
+    const auto w15 = LoadInt<uint32_t>(in + 15 * 4);
+    const auto w16 = LoadInt<uint32_t>(in + 16 * 4);
+    const auto w17 = LoadInt<uint32_t>(in + 17 * 4);
+    const auto w18 = LoadInt<uint32_t>(in + 18 * 4);
+    const auto w19 = LoadInt<uint32_t>(in + 19 * 4);
+    const auto w20 = LoadInt<uint32_t>(in + 20 * 4);
+    const auto w21 = LoadInt<uint32_t>(in + 21 * 4);
+    const auto w22 = LoadInt<uint32_t>(in + 22 * 4);
+    const auto w23 = LoadInt<uint32_t>(in + 23 * 4);
+    const auto w24 = LoadInt<uint32_t>(in + 24 * 4);
+    const auto w25 = LoadInt<uint32_t>(in + 25 * 4);
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 26) | (w1 << 6)) & mask;
+    out[2] = ((w1 >> 20) | (w2 << 12)) & mask;
+    out[3] = ((w2 >> 14) | (w3 << 18)) & mask;
+    out[4] = ((w3 >> 8) | (w4 << 24)) & mask;
+    out[5] = (w4 >> 2) & mask;
+    out[6] = ((w4 >> 28) | (w5 << 4)) & mask;
+    out[7] = ((w5 >> 22) | (w6 << 10)) & mask;
+    out[8] = ((w6 >> 16) | (w7 << 16)) & mask;
+    out[9] = ((w7 >> 10) | (w8 << 22)) & mask;
+    out[10] = (w8 >> 4) & mask;
+    out[11] = ((w8 >> 30) | (w9 << 2)) & mask;
+    out[12] = ((w9 >> 24) | (w10 << 8)) & mask;
+    out[13] = ((w10 >> 18) | (w11 << 14)) & mask;
+    out[14] = ((w11 >> 12) | (w12 << 20)) & mask;
+    out[15] = w12 >> 6;
+    out[16] = (w13) & mask;
+    out[17] = ((w13 >> 26) | (w14 << 6)) & mask;
+    out[18] = ((w14 >> 20) | (w15 << 12)) & mask;
+    out[19] = ((w15 >> 14) | (w16 << 18)) & mask;
+    out[20] = ((w16 >> 8) | (w17 << 24)) & mask;
+    out[21] = (w17 >> 2) & mask;
+    out[22] = ((w17 >> 28) | (w18 << 4)) & mask;
+    out[23] = ((w18 >> 22) | (w19 << 10)) & mask;
+    out[24] = ((w19 >> 16) | (w20 << 16)) & mask;
+    out[25] = ((w20 >> 10) | (w21 << 22)) & mask;
+    out[26] = (w21 >> 4) & mask;
+    out[27] = ((w21 >> 30) | (w22 << 2)) & mask;
+    out[28] = ((w22 >> 24) | (w23 << 8)) & mask;
+    out[29] = ((w23 >> 18) | (w24 << 14)) & mask;
+    out[30] = ((w24 >> 12) | (w25 << 20)) & mask;
+    out[31] = w25 >> 6;
+
+    return in + (26 * 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint32_t, 27> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t mask = ((uint32_t{1} << 27) - uint32_t{1});
+
+    const auto w0 = LoadInt<uint32_t>(in + 0 * 4);
+    const auto w1 = LoadInt<uint32_t>(in + 1 * 4);
+    const auto w2 = LoadInt<uint32_t>(in + 2 * 4);
+    const auto w3 = LoadInt<uint32_t>(in + 3 * 4);
+    const auto w4 = LoadInt<uint32_t>(in + 4 * 4);
+    const auto w5 = LoadInt<uint32_t>(in + 5 * 4);
+    const auto w6 = LoadInt<uint32_t>(in + 6 * 4);
+    const auto w7 = LoadInt<uint32_t>(in + 7 * 4);
+    const auto w8 = LoadInt<uint32_t>(in + 8 * 4);
+    const auto w9 = LoadInt<uint32_t>(in + 9 * 4);
+    const auto w10 = LoadInt<uint32_t>(in + 10 * 4);
+    const auto w11 = LoadInt<uint32_t>(in + 11 * 4);
+    const auto w12 = LoadInt<uint32_t>(in + 12 * 4);
+    const auto w13 = LoadInt<uint32_t>(in + 13 * 4);
+    const auto w14 = LoadInt<uint32_t>(in + 14 * 4);
+    const auto w15 = LoadInt<uint32_t>(in + 15 * 4);
+    const auto w16 = LoadInt<uint32_t>(in + 16 * 4);
+    const auto w17 = LoadInt<uint32_t>(in + 17 * 4);
+    const auto w18 = LoadInt<uint32_t>(in + 18 * 4);
+    const auto w19 = LoadInt<uint32_t>(in + 19 * 4);
+    const auto w20 = LoadInt<uint32_t>(in + 20 * 4);
+    const auto w21 = LoadInt<uint32_t>(in + 21 * 4);
+    const auto w22 = LoadInt<uint32_t>(in + 22 * 4);
+    const auto w23 = LoadInt<uint32_t>(in + 23 * 4);
+    const auto w24 = LoadInt<uint32_t>(in + 24 * 4);
+    const auto w25 = LoadInt<uint32_t>(in + 25 * 4);
+    const auto w26 = LoadInt<uint32_t>(in + 26 * 4);
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 27) | (w1 << 5)) & mask;
+    out[2] = ((w1 >> 22) | (w2 << 10)) & mask;
+    out[3] = ((w2 >> 17) | (w3 << 15)) & mask;
+    out[4] = ((w3 >> 12) | (w4 << 20)) & mask;
+    out[5] = ((w4 >> 7) | (w5 << 25)) & mask;
+    out[6] = (w5 >> 2) & mask;
+    out[7] = ((w5 >> 29) | (w6 << 3)) & mask;
+    out[8] = ((w6 >> 24) | (w7 << 8)) & mask;
+    out[9] = ((w7 >> 19) | (w8 << 13)) & mask;
+    out[10] = ((w8 >> 14) | (w9 << 18)) & mask;
+    out[11] = ((w9 >> 9) | (w10 << 23)) & mask;
+    out[12] = (w10 >> 4) & mask;
+    out[13] = ((w10 >> 31) | (w11 << 1)) & mask;
+    out[14] = ((w11 >> 26) | (w12 << 6)) & mask;
+    out[15] = ((w12 >> 21) | (w13 << 11)) & mask;
+    out[16] = ((w13 >> 16) | (w14 << 16)) & mask;
+    out[17] = ((w14 >> 11) | (w15 << 21)) & mask;
+    out[18] = ((w15 >> 6) | (w16 << 26)) & mask;
+    out[19] = (w16 >> 1) & mask;
+    out[20] = ((w16 >> 28) | (w17 << 4)) & mask;
+    out[21] = ((w17 >> 23) | (w18 << 9)) & mask;
+    out[22] = ((w18 >> 18) | (w19 << 14)) & mask;
+    out[23] = ((w19 >> 13) | (w20 << 19)) & mask;
+    out[24] = ((w20 >> 8) | (w21 << 24)) & mask;
+    out[25] = (w21 >> 3) & mask;
+    out[26] = ((w21 >> 30) | (w22 << 2)) & mask;
+    out[27] = ((w22 >> 25) | (w23 << 7)) & mask;
+    out[28] = ((w23 >> 20) | (w24 << 12)) & mask;
+    out[29] = ((w24 >> 15) | (w25 << 17)) & mask;
+    out[30] = ((w25 >> 10) | (w26 << 22)) & mask;
+    out[31] = w26 >> 5;
+
+    return in + (27 * 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint32_t, 28> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t mask = ((uint32_t{1} << 28) - uint32_t{1});
+
+    const auto w0 = LoadInt<uint32_t>(in + 0 * 4);
+    const auto w1 = LoadInt<uint32_t>(in + 1 * 4);
+    const auto w2 = LoadInt<uint32_t>(in + 2 * 4);
+    const auto w3 = LoadInt<uint32_t>(in + 3 * 4);
+    const auto w4 = LoadInt<uint32_t>(in + 4 * 4);
+    const auto w5 = LoadInt<uint32_t>(in + 5 * 4);
+    const auto w6 = LoadInt<uint32_t>(in + 6 * 4);
+    const auto w7 = LoadInt<uint32_t>(in + 7 * 4);
+    const auto w8 = LoadInt<uint32_t>(in + 8 * 4);
+    const auto w9 = LoadInt<uint32_t>(in + 9 * 4);
+    const auto w10 = LoadInt<uint32_t>(in + 10 * 4);
+    const auto w11 = LoadInt<uint32_t>(in + 11 * 4);
+    const auto w12 = LoadInt<uint32_t>(in + 12 * 4);
+    const auto w13 = LoadInt<uint32_t>(in + 13 * 4);
+    const auto w14 = LoadInt<uint32_t>(in + 14 * 4);
+    const auto w15 = LoadInt<uint32_t>(in + 15 * 4);
+    const auto w16 = LoadInt<uint32_t>(in + 16 * 4);
+    const auto w17 = LoadInt<uint32_t>(in + 17 * 4);
+    const auto w18 = LoadInt<uint32_t>(in + 18 * 4);
+    const auto w19 = LoadInt<uint32_t>(in + 19 * 4);
+    const auto w20 = LoadInt<uint32_t>(in + 20 * 4);
+    const auto w21 = LoadInt<uint32_t>(in + 21 * 4);
+    const auto w22 = LoadInt<uint32_t>(in + 22 * 4);
+    const auto w23 = LoadInt<uint32_t>(in + 23 * 4);
+    const auto w24 = LoadInt<uint32_t>(in + 24 * 4);
+    const auto w25 = LoadInt<uint32_t>(in + 25 * 4);
+    const auto w26 = LoadInt<uint32_t>(in + 26 * 4);
+    const auto w27 = LoadInt<uint32_t>(in + 27 * 4);
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 28) | (w1 << 4)) & mask;
+    out[2] = ((w1 >> 24) | (w2 << 8)) & mask;
+    out[3] = ((w2 >> 20) | (w3 << 12)) & mask;
+    out[4] = ((w3 >> 16) | (w4 << 16)) & mask;
+    out[5] = ((w4 >> 12) | (w5 << 20)) & mask;
+    out[6] = ((w5 >> 8) | (w6 << 24)) & mask;
+    out[7] = w6 >> 4;
+    out[8] = (w7) & mask;
+    out[9] = ((w7 >> 28) | (w8 << 4)) & mask;
+    out[10] = ((w8 >> 24) | (w9 << 8)) & mask;
+    out[11] = ((w9 >> 20) | (w10 << 12)) & mask;
+    out[12] = ((w10 >> 16) | (w11 << 16)) & mask;
+    out[13] = ((w11 >> 12) | (w12 << 20)) & mask;
+    out[14] = ((w12 >> 8) | (w13 << 24)) & mask;
+    out[15] = w13 >> 4;
+    out[16] = (w14) & mask;
+    out[17] = ((w14 >> 28) | (w15 << 4)) & mask;
+    out[18] = ((w15 >> 24) | (w16 << 8)) & mask;
+    out[19] = ((w16 >> 20) | (w17 << 12)) & mask;
+    out[20] = ((w17 >> 16) | (w18 << 16)) & mask;
+    out[21] = ((w18 >> 12) | (w19 << 20)) & mask;
+    out[22] = ((w19 >> 8) | (w20 << 24)) & mask;
+    out[23] = w20 >> 4;
+    out[24] = (w21) & mask;
+    out[25] = ((w21 >> 28) | (w22 << 4)) & mask;
+    out[26] = ((w22 >> 24) | (w23 << 8)) & mask;
+    out[27] = ((w23 >> 20) | (w24 << 12)) & mask;
+    out[28] = ((w24 >> 16) | (w25 << 16)) & mask;
+    out[29] = ((w25 >> 12) | (w26 << 20)) & mask;
+    out[30] = ((w26 >> 8) | (w27 << 24)) & mask;
+    out[31] = w27 >> 4;
+
+    return in + (28 * 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint32_t, 29> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t mask = ((uint32_t{1} << 29) - uint32_t{1});
+
+    const auto w0 = LoadInt<uint32_t>(in + 0 * 4);
+    const auto w1 = LoadInt<uint32_t>(in + 1 * 4);
+    const auto w2 = LoadInt<uint32_t>(in + 2 * 4);
+    const auto w3 = LoadInt<uint32_t>(in + 3 * 4);
+    const auto w4 = LoadInt<uint32_t>(in + 4 * 4);
+    const auto w5 = LoadInt<uint32_t>(in + 5 * 4);
+    const auto w6 = LoadInt<uint32_t>(in + 6 * 4);
+    const auto w7 = LoadInt<uint32_t>(in + 7 * 4);
+    const auto w8 = LoadInt<uint32_t>(in + 8 * 4);
+    const auto w9 = LoadInt<uint32_t>(in + 9 * 4);
+    const auto w10 = LoadInt<uint32_t>(in + 10 * 4);
+    const auto w11 = LoadInt<uint32_t>(in + 11 * 4);
+    const auto w12 = LoadInt<uint32_t>(in + 12 * 4);
+    const auto w13 = LoadInt<uint32_t>(in + 13 * 4);
+    const auto w14 = LoadInt<uint32_t>(in + 14 * 4);
+    const auto w15 = LoadInt<uint32_t>(in + 15 * 4);
+    const auto w16 = LoadInt<uint32_t>(in + 16 * 4);
+    const auto w17 = LoadInt<uint32_t>(in + 17 * 4);
+    const auto w18 = LoadInt<uint32_t>(in + 18 * 4);
+    const auto w19 = LoadInt<uint32_t>(in + 19 * 4);
+    const auto w20 = LoadInt<uint32_t>(in + 20 * 4);
+    const auto w21 = LoadInt<uint32_t>(in + 21 * 4);
+    const auto w22 = LoadInt<uint32_t>(in + 22 * 4);
+    const auto w23 = LoadInt<uint32_t>(in + 23 * 4);
+    const auto w24 = LoadInt<uint32_t>(in + 24 * 4);
+    const auto w25 = LoadInt<uint32_t>(in + 25 * 4);
+    const auto w26 = LoadInt<uint32_t>(in + 26 * 4);
+    const auto w27 = LoadInt<uint32_t>(in + 27 * 4);
+    const auto w28 = LoadInt<uint32_t>(in + 28 * 4);
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 29) | (w1 << 3)) & mask;
+    out[2] = ((w1 >> 26) | (w2 << 6)) & mask;
+    out[3] = ((w2 >> 23) | (w3 << 9)) & mask;
+    out[4] = ((w3 >> 20) | (w4 << 12)) & mask;
+    out[5] = ((w4 >> 17) | (w5 << 15)) & mask;
+    out[6] = ((w5 >> 14) | (w6 << 18)) & mask;
+    out[7] = ((w6 >> 11) | (w7 << 21)) & mask;
+    out[8] = ((w7 >> 8) | (w8 << 24)) & mask;
+    out[9] = ((w8 >> 5) | (w9 << 27)) & mask;
+    out[10] = (w9 >> 2) & mask;
+    out[11] = ((w9 >> 31) | (w10 << 1)) & mask;
+    out[12] = ((w10 >> 28) | (w11 << 4)) & mask;
+    out[13] = ((w11 >> 25) | (w12 << 7)) & mask;
+    out[14] = ((w12 >> 22) | (w13 << 10)) & mask;
+    out[15] = ((w13 >> 19) | (w14 << 13)) & mask;
+    out[16] = ((w14 >> 16) | (w15 << 16)) & mask;
+    out[17] = ((w15 >> 13) | (w16 << 19)) & mask;
+    out[18] = ((w16 >> 10) | (w17 << 22)) & mask;
+    out[19] = ((w17 >> 7) | (w18 << 25)) & mask;
+    out[20] = ((w18 >> 4) | (w19 << 28)) & mask;
+    out[21] = (w19 >> 1) & mask;
+    out[22] = ((w19 >> 30) | (w20 << 2)) & mask;
+    out[23] = ((w20 >> 27) | (w21 << 5)) & mask;
+    out[24] = ((w21 >> 24) | (w22 << 8)) & mask;
+    out[25] = ((w22 >> 21) | (w23 << 11)) & mask;
+    out[26] = ((w23 >> 18) | (w24 << 14)) & mask;
+    out[27] = ((w24 >> 15) | (w25 << 17)) & mask;
+    out[28] = ((w25 >> 12) | (w26 << 20)) & mask;
+    out[29] = ((w26 >> 9) | (w27 << 23)) & mask;
+    out[30] = ((w27 >> 6) | (w28 << 26)) & mask;
+    out[31] = w28 >> 3;
+
+    return in + (29 * 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint32_t, 30> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t mask = ((uint32_t{1} << 30) - uint32_t{1});
+
+    const auto w0 = LoadInt<uint32_t>(in + 0 * 4);
+    const auto w1 = LoadInt<uint32_t>(in + 1 * 4);
+    const auto w2 = LoadInt<uint32_t>(in + 2 * 4);
+    const auto w3 = LoadInt<uint32_t>(in + 3 * 4);
+    const auto w4 = LoadInt<uint32_t>(in + 4 * 4);
+    const auto w5 = LoadInt<uint32_t>(in + 5 * 4);
+    const auto w6 = LoadInt<uint32_t>(in + 6 * 4);
+    const auto w7 = LoadInt<uint32_t>(in + 7 * 4);
+    const auto w8 = LoadInt<uint32_t>(in + 8 * 4);
+    const auto w9 = LoadInt<uint32_t>(in + 9 * 4);
+    const auto w10 = LoadInt<uint32_t>(in + 10 * 4);
+    const auto w11 = LoadInt<uint32_t>(in + 11 * 4);
+    const auto w12 = LoadInt<uint32_t>(in + 12 * 4);
+    const auto w13 = LoadInt<uint32_t>(in + 13 * 4);
+    const auto w14 = LoadInt<uint32_t>(in + 14 * 4);
+    const auto w15 = LoadInt<uint32_t>(in + 15 * 4);
+    const auto w16 = LoadInt<uint32_t>(in + 16 * 4);
+    const auto w17 = LoadInt<uint32_t>(in + 17 * 4);
+    const auto w18 = LoadInt<uint32_t>(in + 18 * 4);
+    const auto w19 = LoadInt<uint32_t>(in + 19 * 4);
+    const auto w20 = LoadInt<uint32_t>(in + 20 * 4);
+    const auto w21 = LoadInt<uint32_t>(in + 21 * 4);
+    const auto w22 = LoadInt<uint32_t>(in + 22 * 4);
+    const auto w23 = LoadInt<uint32_t>(in + 23 * 4);
+    const auto w24 = LoadInt<uint32_t>(in + 24 * 4);
+    const auto w25 = LoadInt<uint32_t>(in + 25 * 4);
+    const auto w26 = LoadInt<uint32_t>(in + 26 * 4);
+    const auto w27 = LoadInt<uint32_t>(in + 27 * 4);
+    const auto w28 = LoadInt<uint32_t>(in + 28 * 4);
+    const auto w29 = LoadInt<uint32_t>(in + 29 * 4);
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 30) | (w1 << 2)) & mask;
+    out[2] = ((w1 >> 28) | (w2 << 4)) & mask;
+    out[3] = ((w2 >> 26) | (w3 << 6)) & mask;
+    out[4] = ((w3 >> 24) | (w4 << 8)) & mask;
+    out[5] = ((w4 >> 22) | (w5 << 10)) & mask;
+    out[6] = ((w5 >> 20) | (w6 << 12)) & mask;
+    out[7] = ((w6 >> 18) | (w7 << 14)) & mask;
+    out[8] = ((w7 >> 16) | (w8 << 16)) & mask;
+    out[9] = ((w8 >> 14) | (w9 << 18)) & mask;
+    out[10] = ((w9 >> 12) | (w10 << 20)) & mask;
+    out[11] = ((w10 >> 10) | (w11 << 22)) & mask;
+    out[12] = ((w11 >> 8) | (w12 << 24)) & mask;
+    out[13] = ((w12 >> 6) | (w13 << 26)) & mask;
+    out[14] = ((w13 >> 4) | (w14 << 28)) & mask;
+    out[15] = w14 >> 2;
+    out[16] = (w15) & mask;
+    out[17] = ((w15 >> 30) | (w16 << 2)) & mask;
+    out[18] = ((w16 >> 28) | (w17 << 4)) & mask;
+    out[19] = ((w17 >> 26) | (w18 << 6)) & mask;
+    out[20] = ((w18 >> 24) | (w19 << 8)) & mask;
+    out[21] = ((w19 >> 22) | (w20 << 10)) & mask;
+    out[22] = ((w20 >> 20) | (w21 << 12)) & mask;
+    out[23] = ((w21 >> 18) | (w22 << 14)) & mask;
+    out[24] = ((w22 >> 16) | (w23 << 16)) & mask;
+    out[25] = ((w23 >> 14) | (w24 << 18)) & mask;
+    out[26] = ((w24 >> 12) | (w25 << 20)) & mask;
+    out[27] = ((w25 >> 10) | (w26 << 22)) & mask;
+    out[28] = ((w26 >> 8) | (w27 << 24)) & mask;
+    out[29] = ((w27 >> 6) | (w28 << 26)) & mask;
+    out[30] = ((w28 >> 4) | (w29 << 28)) & mask;
+    out[31] = w29 >> 2;
+
+    return in + (30 * 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint32_t, 31> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t mask = ((uint32_t{1} << 31) - uint32_t{1});
+
+    const auto w0 = LoadInt<uint32_t>(in + 0 * 4);
+    const auto w1 = LoadInt<uint32_t>(in + 1 * 4);
+    const auto w2 = LoadInt<uint32_t>(in + 2 * 4);
+    const auto w3 = LoadInt<uint32_t>(in + 3 * 4);
+    const auto w4 = LoadInt<uint32_t>(in + 4 * 4);
+    const auto w5 = LoadInt<uint32_t>(in + 5 * 4);
+    const auto w6 = LoadInt<uint32_t>(in + 6 * 4);
+    const auto w7 = LoadInt<uint32_t>(in + 7 * 4);
+    const auto w8 = LoadInt<uint32_t>(in + 8 * 4);
+    const auto w9 = LoadInt<uint32_t>(in + 9 * 4);
+    const auto w10 = LoadInt<uint32_t>(in + 10 * 4);
+    const auto w11 = LoadInt<uint32_t>(in + 11 * 4);
+    const auto w12 = LoadInt<uint32_t>(in + 12 * 4);
+    const auto w13 = LoadInt<uint32_t>(in + 13 * 4);
+    const auto w14 = LoadInt<uint32_t>(in + 14 * 4);
+    const auto w15 = LoadInt<uint32_t>(in + 15 * 4);
+    const auto w16 = LoadInt<uint32_t>(in + 16 * 4);
+    const auto w17 = LoadInt<uint32_t>(in + 17 * 4);
+    const auto w18 = LoadInt<uint32_t>(in + 18 * 4);
+    const auto w19 = LoadInt<uint32_t>(in + 19 * 4);
+    const auto w20 = LoadInt<uint32_t>(in + 20 * 4);
+    const auto w21 = LoadInt<uint32_t>(in + 21 * 4);
+    const auto w22 = LoadInt<uint32_t>(in + 22 * 4);
+    const auto w23 = LoadInt<uint32_t>(in + 23 * 4);
+    const auto w24 = LoadInt<uint32_t>(in + 24 * 4);
+    const auto w25 = LoadInt<uint32_t>(in + 25 * 4);
+    const auto w26 = LoadInt<uint32_t>(in + 26 * 4);
+    const auto w27 = LoadInt<uint32_t>(in + 27 * 4);
+    const auto w28 = LoadInt<uint32_t>(in + 28 * 4);
+    const auto w29 = LoadInt<uint32_t>(in + 29 * 4);
+    const auto w30 = LoadInt<uint32_t>(in + 30 * 4);
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 31) | (w1 << 1)) & mask;
+    out[2] = ((w1 >> 30) | (w2 << 2)) & mask;
+    out[3] = ((w2 >> 29) | (w3 << 3)) & mask;
+    out[4] = ((w3 >> 28) | (w4 << 4)) & mask;
+    out[5] = ((w4 >> 27) | (w5 << 5)) & mask;
+    out[6] = ((w5 >> 26) | (w6 << 6)) & mask;
+    out[7] = ((w6 >> 25) | (w7 << 7)) & mask;
+    out[8] = ((w7 >> 24) | (w8 << 8)) & mask;
+    out[9] = ((w8 >> 23) | (w9 << 9)) & mask;
+    out[10] = ((w9 >> 22) | (w10 << 10)) & mask;
+    out[11] = ((w10 >> 21) | (w11 << 11)) & mask;
+    out[12] = ((w11 >> 20) | (w12 << 12)) & mask;
+    out[13] = ((w12 >> 19) | (w13 << 13)) & mask;
+    out[14] = ((w13 >> 18) | (w14 << 14)) & mask;
+    out[15] = ((w14 >> 17) | (w15 << 15)) & mask;
+    out[16] = ((w15 >> 16) | (w16 << 16)) & mask;
+    out[17] = ((w16 >> 15) | (w17 << 17)) & mask;
+    out[18] = ((w17 >> 14) | (w18 << 18)) & mask;
+    out[19] = ((w18 >> 13) | (w19 << 19)) & mask;
+    out[20] = ((w19 >> 12) | (w20 << 20)) & mask;
+    out[21] = ((w20 >> 11) | (w21 << 21)) & mask;
+    out[22] = ((w21 >> 10) | (w22 << 22)) & mask;
+    out[23] = ((w22 >> 9) | (w23 << 23)) & mask;
+    out[24] = ((w23 >> 8) | (w24 << 24)) & mask;
+    out[25] = ((w24 >> 7) | (w25 << 25)) & mask;
+    out[26] = ((w25 >> 6) | (w26 << 26)) & mask;
+    out[27] = ((w26 >> 5) | (w27 << 27)) & mask;
+    out[28] = ((w27 >> 4) | (w28 << 28)) & mask;
+    out[29] = ((w28 >> 3) | (w29 << 29)) & mask;
+    out[30] = ((w29 >> 2) | (w30 << 30)) & mask;
+    out[31] = w30 >> 1;
+
+    return in + (31 * 4);
+  }
+};
+
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 1> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 1) - uint64_t{1});
+
+    const auto w0 = static_cast<uint64_t>(LoadInt<uint32_t>(in + 0 * 8));
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 1) & mask;
+    out[2] = (w0 >> 2) & mask;
+    out[3] = (w0 >> 3) & mask;
+    out[4] = (w0 >> 4) & mask;
+    out[5] = (w0 >> 5) & mask;
+    out[6] = (w0 >> 6) & mask;
+    out[7] = (w0 >> 7) & mask;
+    out[8] = (w0 >> 8) & mask;
+    out[9] = (w0 >> 9) & mask;
+    out[10] = (w0 >> 10) & mask;
+    out[11] = (w0 >> 11) & mask;
+    out[12] = (w0 >> 12) & mask;
+    out[13] = (w0 >> 13) & mask;
+    out[14] = (w0 >> 14) & mask;
+    out[15] = (w0 >> 15) & mask;
+    out[16] = (w0 >> 16) & mask;
+    out[17] = (w0 >> 17) & mask;
+    out[18] = (w0 >> 18) & mask;
+    out[19] = (w0 >> 19) & mask;
+    out[20] = (w0 >> 20) & mask;
+    out[21] = (w0 >> 21) & mask;
+    out[22] = (w0 >> 22) & mask;
+    out[23] = (w0 >> 23) & mask;
+    out[24] = (w0 >> 24) & mask;
+    out[25] = (w0 >> 25) & mask;
+    out[26] = (w0 >> 26) & mask;
+    out[27] = (w0 >> 27) & mask;
+    out[28] = (w0 >> 28) & mask;
+    out[29] = (w0 >> 29) & mask;
+    out[30] = (w0 >> 30) & mask;
+    out[31] = (w0 >> 31) & mask;
+
+    return in + (0 * 8 + 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 2> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 2) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 2) & mask;
+    out[2] = (w0 >> 4) & mask;
+    out[3] = (w0 >> 6) & mask;
+    out[4] = (w0 >> 8) & mask;
+    out[5] = (w0 >> 10) & mask;
+    out[6] = (w0 >> 12) & mask;
+    out[7] = (w0 >> 14) & mask;
+    out[8] = (w0 >> 16) & mask;
+    out[9] = (w0 >> 18) & mask;
+    out[10] = (w0 >> 20) & mask;
+    out[11] = (w0 >> 22) & mask;
+    out[12] = (w0 >> 24) & mask;
+    out[13] = (w0 >> 26) & mask;
+    out[14] = (w0 >> 28) & mask;
+    out[15] = (w0 >> 30) & mask;
+    out[16] = (w0 >> 32) & mask;
+    out[17] = (w0 >> 34) & mask;
+    out[18] = (w0 >> 36) & mask;
+    out[19] = (w0 >> 38) & mask;
+    out[20] = (w0 >> 40) & mask;
+    out[21] = (w0 >> 42) & mask;
+    out[22] = (w0 >> 44) & mask;
+    out[23] = (w0 >> 46) & mask;
+    out[24] = (w0 >> 48) & mask;
+    out[25] = (w0 >> 50) & mask;
+    out[26] = (w0 >> 52) & mask;
+    out[27] = (w0 >> 54) & mask;
+    out[28] = (w0 >> 56) & mask;
+    out[29] = (w0 >> 58) & mask;
+    out[30] = (w0 >> 60) & mask;
+    out[31] = w0 >> 62;
+
+    return in + (1 * 8);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 3> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 3) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = static_cast<uint64_t>(LoadInt<uint32_t>(in + 1 * 8));
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 3) & mask;
+    out[2] = (w0 >> 6) & mask;
+    out[3] = (w0 >> 9) & mask;
+    out[4] = (w0 >> 12) & mask;
+    out[5] = (w0 >> 15) & mask;
+    out[6] = (w0 >> 18) & mask;
+    out[7] = (w0 >> 21) & mask;
+    out[8] = (w0 >> 24) & mask;
+    out[9] = (w0 >> 27) & mask;
+    out[10] = (w0 >> 30) & mask;
+    out[11] = (w0 >> 33) & mask;
+    out[12] = (w0 >> 36) & mask;
+    out[13] = (w0 >> 39) & mask;
+    out[14] = (w0 >> 42) & mask;
+    out[15] = (w0 >> 45) & mask;
+    out[16] = (w0 >> 48) & mask;
+    out[17] = (w0 >> 51) & mask;
+    out[18] = (w0 >> 54) & mask;
+    out[19] = (w0 >> 57) & mask;
+    out[20] = (w0 >> 60) & mask;
+    out[21] = ((w0 >> 63) | (w1 << 1)) & mask;
+    out[22] = (w1 >> 2) & mask;
+    out[23] = (w1 >> 5) & mask;
+    out[24] = (w1 >> 8) & mask;
+    out[25] = (w1 >> 11) & mask;
+    out[26] = (w1 >> 14) & mask;
+    out[27] = (w1 >> 17) & mask;
+    out[28] = (w1 >> 20) & mask;
+    out[29] = (w1 >> 23) & mask;
+    out[30] = (w1 >> 26) & mask;
+    out[31] = (w1 >> 29) & mask;
+
+    return in + (1 * 8 + 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 4> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 4) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 4) & mask;
+    out[2] = (w0 >> 8) & mask;
+    out[3] = (w0 >> 12) & mask;
+    out[4] = (w0 >> 16) & mask;
+    out[5] = (w0 >> 20) & mask;
+    out[6] = (w0 >> 24) & mask;
+    out[7] = (w0 >> 28) & mask;
+    out[8] = (w0 >> 32) & mask;
+    out[9] = (w0 >> 36) & mask;
+    out[10] = (w0 >> 40) & mask;
+    out[11] = (w0 >> 44) & mask;
+    out[12] = (w0 >> 48) & mask;
+    out[13] = (w0 >> 52) & mask;
+    out[14] = (w0 >> 56) & mask;
+    out[15] = w0 >> 60;
+    out[16] = (w1) & mask;
+    out[17] = (w1 >> 4) & mask;
+    out[18] = (w1 >> 8) & mask;
+    out[19] = (w1 >> 12) & mask;
+    out[20] = (w1 >> 16) & mask;
+    out[21] = (w1 >> 20) & mask;
+    out[22] = (w1 >> 24) & mask;
+    out[23] = (w1 >> 28) & mask;
+    out[24] = (w1 >> 32) & mask;
+    out[25] = (w1 >> 36) & mask;
+    out[26] = (w1 >> 40) & mask;
+    out[27] = (w1 >> 44) & mask;
+    out[28] = (w1 >> 48) & mask;
+    out[29] = (w1 >> 52) & mask;
+    out[30] = (w1 >> 56) & mask;
+    out[31] = w1 >> 60;
+
+    return in + (2 * 8);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 5> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 5) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = static_cast<uint64_t>(LoadInt<uint32_t>(in + 2 * 8));
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 5) & mask;
+    out[2] = (w0 >> 10) & mask;
+    out[3] = (w0 >> 15) & mask;
+    out[4] = (w0 >> 20) & mask;
+    out[5] = (w0 >> 25) & mask;
+    out[6] = (w0 >> 30) & mask;
+    out[7] = (w0 >> 35) & mask;
+    out[8] = (w0 >> 40) & mask;
+    out[9] = (w0 >> 45) & mask;
+    out[10] = (w0 >> 50) & mask;
+    out[11] = (w0 >> 55) & mask;
+    out[12] = ((w0 >> 60) | (w1 << 4)) & mask;
+    out[13] = (w1 >> 1) & mask;
+    out[14] = (w1 >> 6) & mask;
+    out[15] = (w1 >> 11) & mask;
+    out[16] = (w1 >> 16) & mask;
+    out[17] = (w1 >> 21) & mask;
+    out[18] = (w1 >> 26) & mask;
+    out[19] = (w1 >> 31) & mask;
+    out[20] = (w1 >> 36) & mask;
+    out[21] = (w1 >> 41) & mask;
+    out[22] = (w1 >> 46) & mask;
+    out[23] = (w1 >> 51) & mask;
+    out[24] = (w1 >> 56) & mask;
+    out[25] = ((w1 >> 61) | (w2 << 3)) & mask;
+    out[26] = (w2 >> 2) & mask;
+    out[27] = (w2 >> 7) & mask;
+    out[28] = (w2 >> 12) & mask;
+    out[29] = (w2 >> 17) & mask;
+    out[30] = (w2 >> 22) & mask;
+    out[31] = (w2 >> 27) & mask;
+
+    return in + (2 * 8 + 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 6> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 6) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 6) & mask;
+    out[2] = (w0 >> 12) & mask;
+    out[3] = (w0 >> 18) & mask;
+    out[4] = (w0 >> 24) & mask;
+    out[5] = (w0 >> 30) & mask;
+    out[6] = (w0 >> 36) & mask;
+    out[7] = (w0 >> 42) & mask;
+    out[8] = (w0 >> 48) & mask;
+    out[9] = (w0 >> 54) & mask;
+    out[10] = ((w0 >> 60) | (w1 << 4)) & mask;
+    out[11] = (w1 >> 2) & mask;
+    out[12] = (w1 >> 8) & mask;
+    out[13] = (w1 >> 14) & mask;
+    out[14] = (w1 >> 20) & mask;
+    out[15] = (w1 >> 26) & mask;
+    out[16] = (w1 >> 32) & mask;
+    out[17] = (w1 >> 38) & mask;
+    out[18] = (w1 >> 44) & mask;
+    out[19] = (w1 >> 50) & mask;
+    out[20] = (w1 >> 56) & mask;
+    out[21] = ((w1 >> 62) | (w2 << 2)) & mask;
+    out[22] = (w2 >> 4) & mask;
+    out[23] = (w2 >> 10) & mask;
+    out[24] = (w2 >> 16) & mask;
+    out[25] = (w2 >> 22) & mask;
+    out[26] = (w2 >> 28) & mask;
+    out[27] = (w2 >> 34) & mask;
+    out[28] = (w2 >> 40) & mask;
+    out[29] = (w2 >> 46) & mask;
+    out[30] = (w2 >> 52) & mask;
+    out[31] = w2 >> 58;
+
+    return in + (3 * 8);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 7> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 7) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = static_cast<uint64_t>(LoadInt<uint32_t>(in + 3 * 8));
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 7) & mask;
+    out[2] = (w0 >> 14) & mask;
+    out[3] = (w0 >> 21) & mask;
+    out[4] = (w0 >> 28) & mask;
+    out[5] = (w0 >> 35) & mask;
+    out[6] = (w0 >> 42) & mask;
+    out[7] = (w0 >> 49) & mask;
+    out[8] = (w0 >> 56) & mask;
+    out[9] = ((w0 >> 63) | (w1 << 1)) & mask;
+    out[10] = (w1 >> 6) & mask;
+    out[11] = (w1 >> 13) & mask;
+    out[12] = (w1 >> 20) & mask;
+    out[13] = (w1 >> 27) & mask;
+    out[14] = (w1 >> 34) & mask;
+    out[15] = (w1 >> 41) & mask;
+    out[16] = (w1 >> 48) & mask;
+    out[17] = (w1 >> 55) & mask;
+    out[18] = ((w1 >> 62) | (w2 << 2)) & mask;
+    out[19] = (w2 >> 5) & mask;
+    out[20] = (w2 >> 12) & mask;
+    out[21] = (w2 >> 19) & mask;
+    out[22] = (w2 >> 26) & mask;
+    out[23] = (w2 >> 33) & mask;
+    out[24] = (w2 >> 40) & mask;
+    out[25] = (w2 >> 47) & mask;
+    out[26] = (w2 >> 54) & mask;
+    out[27] = ((w2 >> 61) | (w3 << 3)) & mask;
+    out[28] = (w3 >> 4) & mask;
+    out[29] = (w3 >> 11) & mask;
+    out[30] = (w3 >> 18) & mask;
+    out[31] = (w3 >> 25) & mask;
+
+    return in + (3 * 8 + 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 8> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 8) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 8) & mask;
+    out[2] = (w0 >> 16) & mask;
+    out[3] = (w0 >> 24) & mask;
+    out[4] = (w0 >> 32) & mask;
+    out[5] = (w0 >> 40) & mask;
+    out[6] = (w0 >> 48) & mask;
+    out[7] = w0 >> 56;
+    out[8] = (w1) & mask;
+    out[9] = (w1 >> 8) & mask;
+    out[10] = (w1 >> 16) & mask;
+    out[11] = (w1 >> 24) & mask;
+    out[12] = (w1 >> 32) & mask;
+    out[13] = (w1 >> 40) & mask;
+    out[14] = (w1 >> 48) & mask;
+    out[15] = w1 >> 56;
+    out[16] = (w2) & mask;
+    out[17] = (w2 >> 8) & mask;
+    out[18] = (w2 >> 16) & mask;
+    out[19] = (w2 >> 24) & mask;
+    out[20] = (w2 >> 32) & mask;
+    out[21] = (w2 >> 40) & mask;
+    out[22] = (w2 >> 48) & mask;
+    out[23] = w2 >> 56;
+    out[24] = (w3) & mask;
+    out[25] = (w3 >> 8) & mask;
+    out[26] = (w3 >> 16) & mask;
+    out[27] = (w3 >> 24) & mask;
+    out[28] = (w3 >> 32) & mask;
+    out[29] = (w3 >> 40) & mask;
+    out[30] = (w3 >> 48) & mask;
+    out[31] = w3 >> 56;
+
+    return in + (4 * 8);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 9> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 9) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = static_cast<uint64_t>(LoadInt<uint32_t>(in + 4 * 8));
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 9) & mask;
+    out[2] = (w0 >> 18) & mask;
+    out[3] = (w0 >> 27) & mask;
+    out[4] = (w0 >> 36) & mask;
+    out[5] = (w0 >> 45) & mask;
+    out[6] = (w0 >> 54) & mask;
+    out[7] = ((w0 >> 63) | (w1 << 1)) & mask;
+    out[8] = (w1 >> 8) & mask;
+    out[9] = (w1 >> 17) & mask;
+    out[10] = (w1 >> 26) & mask;
+    out[11] = (w1 >> 35) & mask;
+    out[12] = (w1 >> 44) & mask;
+    out[13] = (w1 >> 53) & mask;
+    out[14] = ((w1 >> 62) | (w2 << 2)) & mask;
+    out[15] = (w2 >> 7) & mask;
+    out[16] = (w2 >> 16) & mask;
+    out[17] = (w2 >> 25) & mask;
+    out[18] = (w2 >> 34) & mask;
+    out[19] = (w2 >> 43) & mask;
+    out[20] = (w2 >> 52) & mask;
+    out[21] = ((w2 >> 61) | (w3 << 3)) & mask;
+    out[22] = (w3 >> 6) & mask;
+    out[23] = (w3 >> 15) & mask;
+    out[24] = (w3 >> 24) & mask;
+    out[25] = (w3 >> 33) & mask;
+    out[26] = (w3 >> 42) & mask;
+    out[27] = (w3 >> 51) & mask;
+    out[28] = ((w3 >> 60) | (w4 << 4)) & mask;
+    out[29] = (w4 >> 5) & mask;
+    out[30] = (w4 >> 14) & mask;
+    out[31] = (w4 >> 23) & mask;
+
+    return in + (4 * 8 + 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 10> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 10) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 10) & mask;
+    out[2] = (w0 >> 20) & mask;
+    out[3] = (w0 >> 30) & mask;
+    out[4] = (w0 >> 40) & mask;
+    out[5] = (w0 >> 50) & mask;
+    out[6] = ((w0 >> 60) | (w1 << 4)) & mask;
+    out[7] = (w1 >> 6) & mask;
+    out[8] = (w1 >> 16) & mask;
+    out[9] = (w1 >> 26) & mask;
+    out[10] = (w1 >> 36) & mask;
+    out[11] = (w1 >> 46) & mask;
+    out[12] = ((w1 >> 56) | (w2 << 8)) & mask;
+    out[13] = (w2 >> 2) & mask;
+    out[14] = (w2 >> 12) & mask;
+    out[15] = (w2 >> 22) & mask;
+    out[16] = (w2 >> 32) & mask;
+    out[17] = (w2 >> 42) & mask;
+    out[18] = (w2 >> 52) & mask;
+    out[19] = ((w2 >> 62) | (w3 << 2)) & mask;
+    out[20] = (w3 >> 8) & mask;
+    out[21] = (w3 >> 18) & mask;
+    out[22] = (w3 >> 28) & mask;
+    out[23] = (w3 >> 38) & mask;
+    out[24] = (w3 >> 48) & mask;
+    out[25] = ((w3 >> 58) | (w4 << 6)) & mask;
+    out[26] = (w4 >> 4) & mask;
+    out[27] = (w4 >> 14) & mask;
+    out[28] = (w4 >> 24) & mask;
+    out[29] = (w4 >> 34) & mask;
+    out[30] = (w4 >> 44) & mask;
+    out[31] = w4 >> 54;
+
+    return in + (5 * 8);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 11> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 11) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = static_cast<uint64_t>(LoadInt<uint32_t>(in + 5 * 8));
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 11) & mask;
+    out[2] = (w0 >> 22) & mask;
+    out[3] = (w0 >> 33) & mask;
+    out[4] = (w0 >> 44) & mask;
+    out[5] = ((w0 >> 55) | (w1 << 9)) & mask;
+    out[6] = (w1 >> 2) & mask;
+    out[7] = (w1 >> 13) & mask;
+    out[8] = (w1 >> 24) & mask;
+    out[9] = (w1 >> 35) & mask;
+    out[10] = (w1 >> 46) & mask;
+    out[11] = ((w1 >> 57) | (w2 << 7)) & mask;
+    out[12] = (w2 >> 4) & mask;
+    out[13] = (w2 >> 15) & mask;
+    out[14] = (w2 >> 26) & mask;
+    out[15] = (w2 >> 37) & mask;
+    out[16] = (w2 >> 48) & mask;
+    out[17] = ((w2 >> 59) | (w3 << 5)) & mask;
+    out[18] = (w3 >> 6) & mask;
+    out[19] = (w3 >> 17) & mask;
+    out[20] = (w3 >> 28) & mask;
+    out[21] = (w3 >> 39) & mask;
+    out[22] = (w3 >> 50) & mask;
+    out[23] = ((w3 >> 61) | (w4 << 3)) & mask;
+    out[24] = (w4 >> 8) & mask;
+    out[25] = (w4 >> 19) & mask;
+    out[26] = (w4 >> 30) & mask;
+    out[27] = (w4 >> 41) & mask;
+    out[28] = (w4 >> 52) & mask;
+    out[29] = ((w4 >> 63) | (w5 << 1)) & mask;
+    out[30] = (w5 >> 10) & mask;
+    out[31] = (w5 >> 21) & mask;
+
+    return in + (5 * 8 + 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 12> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 12) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 12) & mask;
+    out[2] = (w0 >> 24) & mask;
+    out[3] = (w0 >> 36) & mask;
+    out[4] = (w0 >> 48) & mask;
+    out[5] = ((w0 >> 60) | (w1 << 4)) & mask;
+    out[6] = (w1 >> 8) & mask;
+    out[7] = (w1 >> 20) & mask;
+    out[8] = (w1 >> 32) & mask;
+    out[9] = (w1 >> 44) & mask;
+    out[10] = ((w1 >> 56) | (w2 << 8)) & mask;
+    out[11] = (w2 >> 4) & mask;
+    out[12] = (w2 >> 16) & mask;
+    out[13] = (w2 >> 28) & mask;
+    out[14] = (w2 >> 40) & mask;
+    out[15] = w2 >> 52;
+    out[16] = (w3) & mask;
+    out[17] = (w3 >> 12) & mask;
+    out[18] = (w3 >> 24) & mask;
+    out[19] = (w3 >> 36) & mask;
+    out[20] = (w3 >> 48) & mask;
+    out[21] = ((w3 >> 60) | (w4 << 4)) & mask;
+    out[22] = (w4 >> 8) & mask;
+    out[23] = (w4 >> 20) & mask;
+    out[24] = (w4 >> 32) & mask;
+    out[25] = (w4 >> 44) & mask;
+    out[26] = ((w4 >> 56) | (w5 << 8)) & mask;
+    out[27] = (w5 >> 4) & mask;
+    out[28] = (w5 >> 16) & mask;
+    out[29] = (w5 >> 28) & mask;
+    out[30] = (w5 >> 40) & mask;
+    out[31] = w5 >> 52;
+
+    return in + (6 * 8);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 13> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 13) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = static_cast<uint64_t>(LoadInt<uint32_t>(in + 6 * 8));
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 13) & mask;
+    out[2] = (w0 >> 26) & mask;
+    out[3] = (w0 >> 39) & mask;
+    out[4] = ((w0 >> 52) | (w1 << 12)) & mask;
+    out[5] = (w1 >> 1) & mask;
+    out[6] = (w1 >> 14) & mask;
+    out[7] = (w1 >> 27) & mask;
+    out[8] = (w1 >> 40) & mask;
+    out[9] = ((w1 >> 53) | (w2 << 11)) & mask;
+    out[10] = (w2 >> 2) & mask;
+    out[11] = (w2 >> 15) & mask;
+    out[12] = (w2 >> 28) & mask;
+    out[13] = (w2 >> 41) & mask;
+    out[14] = ((w2 >> 54) | (w3 << 10)) & mask;
+    out[15] = (w3 >> 3) & mask;
+    out[16] = (w3 >> 16) & mask;
+    out[17] = (w3 >> 29) & mask;
+    out[18] = (w3 >> 42) & mask;
+    out[19] = ((w3 >> 55) | (w4 << 9)) & mask;
+    out[20] = (w4 >> 4) & mask;
+    out[21] = (w4 >> 17) & mask;
+    out[22] = (w4 >> 30) & mask;
+    out[23] = (w4 >> 43) & mask;
+    out[24] = ((w4 >> 56) | (w5 << 8)) & mask;
+    out[25] = (w5 >> 5) & mask;
+    out[26] = (w5 >> 18) & mask;
+    out[27] = (w5 >> 31) & mask;
+    out[28] = (w5 >> 44) & mask;
+    out[29] = ((w5 >> 57) | (w6 << 7)) & mask;
+    out[30] = (w6 >> 6) & mask;
+    out[31] = (w6 >> 19) & mask;
+
+    return in + (6 * 8 + 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 14> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 14) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 14) & mask;
+    out[2] = (w0 >> 28) & mask;
+    out[3] = (w0 >> 42) & mask;
+    out[4] = ((w0 >> 56) | (w1 << 8)) & mask;
+    out[5] = (w1 >> 6) & mask;
+    out[6] = (w1 >> 20) & mask;
+    out[7] = (w1 >> 34) & mask;
+    out[8] = (w1 >> 48) & mask;
+    out[9] = ((w1 >> 62) | (w2 << 2)) & mask;
+    out[10] = (w2 >> 12) & mask;
+    out[11] = (w2 >> 26) & mask;
+    out[12] = (w2 >> 40) & mask;
+    out[13] = ((w2 >> 54) | (w3 << 10)) & mask;
+    out[14] = (w3 >> 4) & mask;
+    out[15] = (w3 >> 18) & mask;
+    out[16] = (w3 >> 32) & mask;
+    out[17] = (w3 >> 46) & mask;
+    out[18] = ((w3 >> 60) | (w4 << 4)) & mask;
+    out[19] = (w4 >> 10) & mask;
+    out[20] = (w4 >> 24) & mask;
+    out[21] = (w4 >> 38) & mask;
+    out[22] = ((w4 >> 52) | (w5 << 12)) & mask;
+    out[23] = (w5 >> 2) & mask;
+    out[24] = (w5 >> 16) & mask;
+    out[25] = (w5 >> 30) & mask;
+    out[26] = (w5 >> 44) & mask;
+    out[27] = ((w5 >> 58) | (w6 << 6)) & mask;
+    out[28] = (w6 >> 8) & mask;
+    out[29] = (w6 >> 22) & mask;
+    out[30] = (w6 >> 36) & mask;
+    out[31] = w6 >> 50;
+
+    return in + (7 * 8);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 15> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 15) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = static_cast<uint64_t>(LoadInt<uint32_t>(in + 7 * 8));
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 15) & mask;
+    out[2] = (w0 >> 30) & mask;
+    out[3] = (w0 >> 45) & mask;
+    out[4] = ((w0 >> 60) | (w1 << 4)) & mask;
+    out[5] = (w1 >> 11) & mask;
+    out[6] = (w1 >> 26) & mask;
+    out[7] = (w1 >> 41) & mask;
+    out[8] = ((w1 >> 56) | (w2 << 8)) & mask;
+    out[9] = (w2 >> 7) & mask;
+    out[10] = (w2 >> 22) & mask;
+    out[11] = (w2 >> 37) & mask;
+    out[12] = ((w2 >> 52) | (w3 << 12)) & mask;
+    out[13] = (w3 >> 3) & mask;
+    out[14] = (w3 >> 18) & mask;
+    out[15] = (w3 >> 33) & mask;
+    out[16] = (w3 >> 48) & mask;
+    out[17] = ((w3 >> 63) | (w4 << 1)) & mask;
+    out[18] = (w4 >> 14) & mask;
+    out[19] = (w4 >> 29) & mask;
+    out[20] = (w4 >> 44) & mask;
+    out[21] = ((w4 >> 59) | (w5 << 5)) & mask;
+    out[22] = (w5 >> 10) & mask;
+    out[23] = (w5 >> 25) & mask;
+    out[24] = (w5 >> 40) & mask;
+    out[25] = ((w5 >> 55) | (w6 << 9)) & mask;
+    out[26] = (w6 >> 6) & mask;
+    out[27] = (w6 >> 21) & mask;
+    out[28] = (w6 >> 36) & mask;
+    out[29] = ((w6 >> 51) | (w7 << 13)) & mask;
+    out[30] = (w7 >> 2) & mask;
+    out[31] = (w7 >> 17) & mask;
+
+    return in + (7 * 8 + 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 16> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 16) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 16) & mask;
+    out[2] = (w0 >> 32) & mask;
+    out[3] = w0 >> 48;
+    out[4] = (w1) & mask;
+    out[5] = (w1 >> 16) & mask;
+    out[6] = (w1 >> 32) & mask;
+    out[7] = w1 >> 48;
+    out[8] = (w2) & mask;
+    out[9] = (w2 >> 16) & mask;
+    out[10] = (w2 >> 32) & mask;
+    out[11] = w2 >> 48;
+    out[12] = (w3) & mask;
+    out[13] = (w3 >> 16) & mask;
+    out[14] = (w3 >> 32) & mask;
+    out[15] = w3 >> 48;
+    out[16] = (w4) & mask;
+    out[17] = (w4 >> 16) & mask;
+    out[18] = (w4 >> 32) & mask;
+    out[19] = w4 >> 48;
+    out[20] = (w5) & mask;
+    out[21] = (w5 >> 16) & mask;
+    out[22] = (w5 >> 32) & mask;
+    out[23] = w5 >> 48;
+    out[24] = (w6) & mask;
+    out[25] = (w6 >> 16) & mask;
+    out[26] = (w6 >> 32) & mask;
+    out[27] = w6 >> 48;
+    out[28] = (w7) & mask;
+    out[29] = (w7 >> 16) & mask;
+    out[30] = (w7 >> 32) & mask;
+    out[31] = w7 >> 48;
+
+    return in + (8 * 8);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 17> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 17) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    const auto w8 = static_cast<uint64_t>(LoadInt<uint32_t>(in + 8 * 8));
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 17) & mask;
+    out[2] = (w0 >> 34) & mask;
+    out[3] = ((w0 >> 51) | (w1 << 13)) & mask;
+    out[4] = (w1 >> 4) & mask;
+    out[5] = (w1 >> 21) & mask;
+    out[6] = (w1 >> 38) & mask;
+    out[7] = ((w1 >> 55) | (w2 << 9)) & mask;
+    out[8] = (w2 >> 8) & mask;
+    out[9] = (w2 >> 25) & mask;
+    out[10] = (w2 >> 42) & mask;
+    out[11] = ((w2 >> 59) | (w3 << 5)) & mask;
+    out[12] = (w3 >> 12) & mask;
+    out[13] = (w3 >> 29) & mask;
+    out[14] = (w3 >> 46) & mask;
+    out[15] = ((w3 >> 63) | (w4 << 1)) & mask;
+    out[16] = (w4 >> 16) & mask;
+    out[17] = (w4 >> 33) & mask;
+    out[18] = ((w4 >> 50) | (w5 << 14)) & mask;
+    out[19] = (w5 >> 3) & mask;
+    out[20] = (w5 >> 20) & mask;
+    out[21] = (w5 >> 37) & mask;
+    out[22] = ((w5 >> 54) | (w6 << 10)) & mask;
+    out[23] = (w6 >> 7) & mask;
+    out[24] = (w6 >> 24) & mask;
+    out[25] = (w6 >> 41) & mask;
+    out[26] = ((w6 >> 58) | (w7 << 6)) & mask;
+    out[27] = (w7 >> 11) & mask;
+    out[28] = (w7 >> 28) & mask;
+    out[29] = (w7 >> 45) & mask;
+    out[30] = ((w7 >> 62) | (w8 << 2)) & mask;
+    out[31] = (w8 >> 15) & mask;
+
+    return in + (8 * 8 + 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 18> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 18) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    const auto w8 = LoadInt<uint64_t>(in + 8 * 8);
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 18) & mask;
+    out[2] = (w0 >> 36) & mask;
+    out[3] = ((w0 >> 54) | (w1 << 10)) & mask;
+    out[4] = (w1 >> 8) & mask;
+    out[5] = (w1 >> 26) & mask;
+    out[6] = (w1 >> 44) & mask;
+    out[7] = ((w1 >> 62) | (w2 << 2)) & mask;
+    out[8] = (w2 >> 16) & mask;
+    out[9] = (w2 >> 34) & mask;
+    out[10] = ((w2 >> 52) | (w3 << 12)) & mask;
+    out[11] = (w3 >> 6) & mask;
+    out[12] = (w3 >> 24) & mask;
+    out[13] = (w3 >> 42) & mask;
+    out[14] = ((w3 >> 60) | (w4 << 4)) & mask;
+    out[15] = (w4 >> 14) & mask;
+    out[16] = (w4 >> 32) & mask;
+    out[17] = ((w4 >> 50) | (w5 << 14)) & mask;
+    out[18] = (w5 >> 4) & mask;
+    out[19] = (w5 >> 22) & mask;
+    out[20] = (w5 >> 40) & mask;
+    out[21] = ((w5 >> 58) | (w6 << 6)) & mask;
+    out[22] = (w6 >> 12) & mask;
+    out[23] = (w6 >> 30) & mask;
+    out[24] = ((w6 >> 48) | (w7 << 16)) & mask;
+    out[25] = (w7 >> 2) & mask;
+    out[26] = (w7 >> 20) & mask;
+    out[27] = (w7 >> 38) & mask;
+    out[28] = ((w7 >> 56) | (w8 << 8)) & mask;
+    out[29] = (w8 >> 10) & mask;
+    out[30] = (w8 >> 28) & mask;
+    out[31] = w8 >> 46;
+
+    return in + (9 * 8);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 19> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 19) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    const auto w8 = LoadInt<uint64_t>(in + 8 * 8);
+    const auto w9 = static_cast<uint64_t>(LoadInt<uint32_t>(in + 9 * 8));
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 19) & mask;
+    out[2] = (w0 >> 38) & mask;
+    out[3] = ((w0 >> 57) | (w1 << 7)) & mask;
+    out[4] = (w1 >> 12) & mask;
+    out[5] = (w1 >> 31) & mask;
+    out[6] = ((w1 >> 50) | (w2 << 14)) & mask;
+    out[7] = (w2 >> 5) & mask;
+    out[8] = (w2 >> 24) & mask;
+    out[9] = (w2 >> 43) & mask;
+    out[10] = ((w2 >> 62) | (w3 << 2)) & mask;
+    out[11] = (w3 >> 17) & mask;
+    out[12] = (w3 >> 36) & mask;
+    out[13] = ((w3 >> 55) | (w4 << 9)) & mask;
+    out[14] = (w4 >> 10) & mask;
+    out[15] = (w4 >> 29) & mask;
+    out[16] = ((w4 >> 48) | (w5 << 16)) & mask;
+    out[17] = (w5 >> 3) & mask;
+    out[18] = (w5 >> 22) & mask;
+    out[19] = (w5 >> 41) & mask;
+    out[20] = ((w5 >> 60) | (w6 << 4)) & mask;
+    out[21] = (w6 >> 15) & mask;
+    out[22] = (w6 >> 34) & mask;
+    out[23] = ((w6 >> 53) | (w7 << 11)) & mask;
+    out[24] = (w7 >> 8) & mask;
+    out[25] = (w7 >> 27) & mask;
+    out[26] = ((w7 >> 46) | (w8 << 18)) & mask;
+    out[27] = (w8 >> 1) & mask;
+    out[28] = (w8 >> 20) & mask;
+    out[29] = (w8 >> 39) & mask;
+    out[30] = ((w8 >> 58) | (w9 << 6)) & mask;
+    out[31] = (w9 >> 13) & mask;
+
+    return in + (9 * 8 + 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 20> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 20) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    const auto w8 = LoadInt<uint64_t>(in + 8 * 8);
+    const auto w9 = LoadInt<uint64_t>(in + 9 * 8);
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 20) & mask;
+    out[2] = (w0 >> 40) & mask;
+    out[3] = ((w0 >> 60) | (w1 << 4)) & mask;
+    out[4] = (w1 >> 16) & mask;
+    out[5] = (w1 >> 36) & mask;
+    out[6] = ((w1 >> 56) | (w2 << 8)) & mask;
+    out[7] = (w2 >> 12) & mask;
+    out[8] = (w2 >> 32) & mask;
+    out[9] = ((w2 >> 52) | (w3 << 12)) & mask;
+    out[10] = (w3 >> 8) & mask;
+    out[11] = (w3 >> 28) & mask;
+    out[12] = ((w3 >> 48) | (w4 << 16)) & mask;
+    out[13] = (w4 >> 4) & mask;
+    out[14] = (w4 >> 24) & mask;
+    out[15] = w4 >> 44;
+    out[16] = (w5) & mask;
+    out[17] = (w5 >> 20) & mask;
+    out[18] = (w5 >> 40) & mask;
+    out[19] = ((w5 >> 60) | (w6 << 4)) & mask;
+    out[20] = (w6 >> 16) & mask;
+    out[21] = (w6 >> 36) & mask;
+    out[22] = ((w6 >> 56) | (w7 << 8)) & mask;
+    out[23] = (w7 >> 12) & mask;
+    out[24] = (w7 >> 32) & mask;
+    out[25] = ((w7 >> 52) | (w8 << 12)) & mask;
+    out[26] = (w8 >> 8) & mask;
+    out[27] = (w8 >> 28) & mask;
+    out[28] = ((w8 >> 48) | (w9 << 16)) & mask;
+    out[29] = (w9 >> 4) & mask;
+    out[30] = (w9 >> 24) & mask;
+    out[31] = w9 >> 44;
+
+    return in + (10 * 8);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 21> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 21) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    const auto w8 = LoadInt<uint64_t>(in + 8 * 8);
+    const auto w9 = LoadInt<uint64_t>(in + 9 * 8);
+    const auto w10 = static_cast<uint64_t>(LoadInt<uint32_t>(in + 10 * 8));
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 21) & mask;
+    out[2] = (w0 >> 42) & mask;
+    out[3] = ((w0 >> 63) | (w1 << 1)) & mask;
+    out[4] = (w1 >> 20) & mask;
+    out[5] = (w1 >> 41) & mask;
+    out[6] = ((w1 >> 62) | (w2 << 2)) & mask;
+    out[7] = (w2 >> 19) & mask;
+    out[8] = (w2 >> 40) & mask;
+    out[9] = ((w2 >> 61) | (w3 << 3)) & mask;
+    out[10] = (w3 >> 18) & mask;
+    out[11] = (w3 >> 39) & mask;
+    out[12] = ((w3 >> 60) | (w4 << 4)) & mask;
+    out[13] = (w4 >> 17) & mask;
+    out[14] = (w4 >> 38) & mask;
+    out[15] = ((w4 >> 59) | (w5 << 5)) & mask;
+    out[16] = (w5 >> 16) & mask;
+    out[17] = (w5 >> 37) & mask;
+    out[18] = ((w5 >> 58) | (w6 << 6)) & mask;
+    out[19] = (w6 >> 15) & mask;
+    out[20] = (w6 >> 36) & mask;
+    out[21] = ((w6 >> 57) | (w7 << 7)) & mask;
+    out[22] = (w7 >> 14) & mask;
+    out[23] = (w7 >> 35) & mask;
+    out[24] = ((w7 >> 56) | (w8 << 8)) & mask;
+    out[25] = (w8 >> 13) & mask;
+    out[26] = (w8 >> 34) & mask;
+    out[27] = ((w8 >> 55) | (w9 << 9)) & mask;
+    out[28] = (w9 >> 12) & mask;
+    out[29] = (w9 >> 33) & mask;
+    out[30] = ((w9 >> 54) | (w10 << 10)) & mask;
+    out[31] = (w10 >> 11) & mask;
+
+    return in + (10 * 8 + 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 22> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 22) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    const auto w8 = LoadInt<uint64_t>(in + 8 * 8);
+    const auto w9 = LoadInt<uint64_t>(in + 9 * 8);
+    const auto w10 = LoadInt<uint64_t>(in + 10 * 8);
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 22) & mask;
+    out[2] = ((w0 >> 44) | (w1 << 20)) & mask;
+    out[3] = (w1 >> 2) & mask;
+    out[4] = (w1 >> 24) & mask;
+    out[5] = ((w1 >> 46) | (w2 << 18)) & mask;
+    out[6] = (w2 >> 4) & mask;
+    out[7] = (w2 >> 26) & mask;
+    out[8] = ((w2 >> 48) | (w3 << 16)) & mask;
+    out[9] = (w3 >> 6) & mask;
+    out[10] = (w3 >> 28) & mask;
+    out[11] = ((w3 >> 50) | (w4 << 14)) & mask;
+    out[12] = (w4 >> 8) & mask;
+    out[13] = (w4 >> 30) & mask;
+    out[14] = ((w4 >> 52) | (w5 << 12)) & mask;
+    out[15] = (w5 >> 10) & mask;
+    out[16] = (w5 >> 32) & mask;
+    out[17] = ((w5 >> 54) | (w6 << 10)) & mask;
+    out[18] = (w6 >> 12) & mask;
+    out[19] = (w6 >> 34) & mask;
+    out[20] = ((w6 >> 56) | (w7 << 8)) & mask;
+    out[21] = (w7 >> 14) & mask;
+    out[22] = (w7 >> 36) & mask;
+    out[23] = ((w7 >> 58) | (w8 << 6)) & mask;
+    out[24] = (w8 >> 16) & mask;
+    out[25] = (w8 >> 38) & mask;
+    out[26] = ((w8 >> 60) | (w9 << 4)) & mask;
+    out[27] = (w9 >> 18) & mask;
+    out[28] = (w9 >> 40) & mask;
+    out[29] = ((w9 >> 62) | (w10 << 2)) & mask;
+    out[30] = (w10 >> 20) & mask;
+    out[31] = w10 >> 42;
+
+    return in + (11 * 8);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 23> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 23) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    const auto w8 = LoadInt<uint64_t>(in + 8 * 8);
+    const auto w9 = LoadInt<uint64_t>(in + 9 * 8);
+    const auto w10 = LoadInt<uint64_t>(in + 10 * 8);
+    const auto w11 = static_cast<uint64_t>(LoadInt<uint32_t>(in + 11 * 8));
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 23) & mask;
+    out[2] = ((w0 >> 46) | (w1 << 18)) & mask;
+    out[3] = (w1 >> 5) & mask;
+    out[4] = (w1 >> 28) & mask;
+    out[5] = ((w1 >> 51) | (w2 << 13)) & mask;
+    out[6] = (w2 >> 10) & mask;
+    out[7] = (w2 >> 33) & mask;
+    out[8] = ((w2 >> 56) | (w3 << 8)) & mask;
+    out[9] = (w3 >> 15) & mask;
+    out[10] = (w3 >> 38) & mask;
+    out[11] = ((w3 >> 61) | (w4 << 3)) & mask;
+    out[12] = (w4 >> 20) & mask;
+    out[13] = ((w4 >> 43) | (w5 << 21)) & mask;
+    out[14] = (w5 >> 2) & mask;
+    out[15] = (w5 >> 25) & mask;
+    out[16] = ((w5 >> 48) | (w6 << 16)) & mask;
+    out[17] = (w6 >> 7) & mask;
+    out[18] = (w6 >> 30) & mask;
+    out[19] = ((w6 >> 53) | (w7 << 11)) & mask;
+    out[20] = (w7 >> 12) & mask;
+    out[21] = (w7 >> 35) & mask;
+    out[22] = ((w7 >> 58) | (w8 << 6)) & mask;
+    out[23] = (w8 >> 17) & mask;
+    out[24] = (w8 >> 40) & mask;
+    out[25] = ((w8 >> 63) | (w9 << 1)) & mask;
+    out[26] = (w9 >> 22) & mask;
+    out[27] = ((w9 >> 45) | (w10 << 19)) & mask;
+    out[28] = (w10 >> 4) & mask;
+    out[29] = (w10 >> 27) & mask;
+    out[30] = ((w10 >> 50) | (w11 << 14)) & mask;
+    out[31] = (w11 >> 9) & mask;
+
+    return in + (11 * 8 + 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 24> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 24) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    const auto w8 = LoadInt<uint64_t>(in + 8 * 8);
+    const auto w9 = LoadInt<uint64_t>(in + 9 * 8);
+    const auto w10 = LoadInt<uint64_t>(in + 10 * 8);
+    const auto w11 = LoadInt<uint64_t>(in + 11 * 8);
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 24) & mask;
+    out[2] = ((w0 >> 48) | (w1 << 16)) & mask;
+    out[3] = (w1 >> 8) & mask;
+    out[4] = (w1 >> 32) & mask;
+    out[5] = ((w1 >> 56) | (w2 << 8)) & mask;
+    out[6] = (w2 >> 16) & mask;
+    out[7] = w2 >> 40;
+    out[8] = (w3) & mask;
+    out[9] = (w3 >> 24) & mask;
+    out[10] = ((w3 >> 48) | (w4 << 16)) & mask;
+    out[11] = (w4 >> 8) & mask;
+    out[12] = (w4 >> 32) & mask;
+    out[13] = ((w4 >> 56) | (w5 << 8)) & mask;
+    out[14] = (w5 >> 16) & mask;
+    out[15] = w5 >> 40;
+    out[16] = (w6) & mask;
+    out[17] = (w6 >> 24) & mask;
+    out[18] = ((w6 >> 48) | (w7 << 16)) & mask;
+    out[19] = (w7 >> 8) & mask;
+    out[20] = (w7 >> 32) & mask;
+    out[21] = ((w7 >> 56) | (w8 << 8)) & mask;
+    out[22] = (w8 >> 16) & mask;
+    out[23] = w8 >> 40;
+    out[24] = (w9) & mask;
+    out[25] = (w9 >> 24) & mask;
+    out[26] = ((w9 >> 48) | (w10 << 16)) & mask;
+    out[27] = (w10 >> 8) & mask;
+    out[28] = (w10 >> 32) & mask;
+    out[29] = ((w10 >> 56) | (w11 << 8)) & mask;
+    out[30] = (w11 >> 16) & mask;
+    out[31] = w11 >> 40;
+
+    return in + (12 * 8);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 25> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 25) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    const auto w8 = LoadInt<uint64_t>(in + 8 * 8);
+    const auto w9 = LoadInt<uint64_t>(in + 9 * 8);
+    const auto w10 = LoadInt<uint64_t>(in + 10 * 8);
+    const auto w11 = LoadInt<uint64_t>(in + 11 * 8);
+    const auto w12 = static_cast<uint64_t>(LoadInt<uint32_t>(in + 12 * 8));
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 25) & mask;
+    out[2] = ((w0 >> 50) | (w1 << 14)) & mask;
+    out[3] = (w1 >> 11) & mask;
+    out[4] = (w1 >> 36) & mask;
+    out[5] = ((w1 >> 61) | (w2 << 3)) & mask;
+    out[6] = (w2 >> 22) & mask;
+    out[7] = ((w2 >> 47) | (w3 << 17)) & mask;
+    out[8] = (w3 >> 8) & mask;
+    out[9] = (w3 >> 33) & mask;
+    out[10] = ((w3 >> 58) | (w4 << 6)) & mask;
+    out[11] = (w4 >> 19) & mask;
+    out[12] = ((w4 >> 44) | (w5 << 20)) & mask;
+    out[13] = (w5 >> 5) & mask;
+    out[14] = (w5 >> 30) & mask;
+    out[15] = ((w5 >> 55) | (w6 << 9)) & mask;
+    out[16] = (w6 >> 16) & mask;
+    out[17] = ((w6 >> 41) | (w7 << 23)) & mask;
+    out[18] = (w7 >> 2) & mask;
+    out[19] = (w7 >> 27) & mask;
+    out[20] = ((w7 >> 52) | (w8 << 12)) & mask;
+    out[21] = (w8 >> 13) & mask;
+    out[22] = (w8 >> 38) & mask;
+    out[23] = ((w8 >> 63) | (w9 << 1)) & mask;
+    out[24] = (w9 >> 24) & mask;
+    out[25] = ((w9 >> 49) | (w10 << 15)) & mask;
+    out[26] = (w10 >> 10) & mask;
+    out[27] = (w10 >> 35) & mask;
+    out[28] = ((w10 >> 60) | (w11 << 4)) & mask;
+    out[29] = (w11 >> 21) & mask;
+    out[30] = ((w11 >> 46) | (w12 << 18)) & mask;
+    out[31] = (w12 >> 7) & mask;
+
+    return in + (12 * 8 + 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 26> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 26) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    const auto w8 = LoadInt<uint64_t>(in + 8 * 8);
+    const auto w9 = LoadInt<uint64_t>(in + 9 * 8);
+    const auto w10 = LoadInt<uint64_t>(in + 10 * 8);
+    const auto w11 = LoadInt<uint64_t>(in + 11 * 8);
+    const auto w12 = LoadInt<uint64_t>(in + 12 * 8);
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 26) & mask;
+    out[2] = ((w0 >> 52) | (w1 << 12)) & mask;
+    out[3] = (w1 >> 14) & mask;
+    out[4] = ((w1 >> 40) | (w2 << 24)) & mask;
+    out[5] = (w2 >> 2) & mask;
+    out[6] = (w2 >> 28) & mask;
+    out[7] = ((w2 >> 54) | (w3 << 10)) & mask;
+    out[8] = (w3 >> 16) & mask;
+    out[9] = ((w3 >> 42) | (w4 << 22)) & mask;
+    out[10] = (w4 >> 4) & mask;
+    out[11] = (w4 >> 30) & mask;
+    out[12] = ((w4 >> 56) | (w5 << 8)) & mask;
+    out[13] = (w5 >> 18) & mask;
+    out[14] = ((w5 >> 44) | (w6 << 20)) & mask;
+    out[15] = (w6 >> 6) & mask;
+    out[16] = (w6 >> 32) & mask;
+    out[17] = ((w6 >> 58) | (w7 << 6)) & mask;
+    out[18] = (w7 >> 20) & mask;
+    out[19] = ((w7 >> 46) | (w8 << 18)) & mask;
+    out[20] = (w8 >> 8) & mask;
+    out[21] = (w8 >> 34) & mask;
+    out[22] = ((w8 >> 60) | (w9 << 4)) & mask;
+    out[23] = (w9 >> 22) & mask;
+    out[24] = ((w9 >> 48) | (w10 << 16)) & mask;
+    out[25] = (w10 >> 10) & mask;
+    out[26] = (w10 >> 36) & mask;
+    out[27] = ((w10 >> 62) | (w11 << 2)) & mask;
+    out[28] = (w11 >> 24) & mask;
+    out[29] = ((w11 >> 50) | (w12 << 14)) & mask;
+    out[30] = (w12 >> 12) & mask;
+    out[31] = w12 >> 38;
+
+    return in + (13 * 8);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 27> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 27) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    const auto w8 = LoadInt<uint64_t>(in + 8 * 8);
+    const auto w9 = LoadInt<uint64_t>(in + 9 * 8);
+    const auto w10 = LoadInt<uint64_t>(in + 10 * 8);
+    const auto w11 = LoadInt<uint64_t>(in + 11 * 8);
+    const auto w12 = LoadInt<uint64_t>(in + 12 * 8);
+    const auto w13 = static_cast<uint64_t>(LoadInt<uint32_t>(in + 13 * 8));
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 27) & mask;
+    out[2] = ((w0 >> 54) | (w1 << 10)) & mask;
+    out[3] = (w1 >> 17) & mask;
+    out[4] = ((w1 >> 44) | (w2 << 20)) & mask;
+    out[5] = (w2 >> 7) & mask;
+    out[6] = (w2 >> 34) & mask;
+    out[7] = ((w2 >> 61) | (w3 << 3)) & mask;
+    out[8] = (w3 >> 24) & mask;
+    out[9] = ((w3 >> 51) | (w4 << 13)) & mask;
+    out[10] = (w4 >> 14) & mask;
+    out[11] = ((w4 >> 41) | (w5 << 23)) & mask;
+    out[12] = (w5 >> 4) & mask;
+    out[13] = (w5 >> 31) & mask;
+    out[14] = ((w5 >> 58) | (w6 << 6)) & mask;
+    out[15] = (w6 >> 21) & mask;
+    out[16] = ((w6 >> 48) | (w7 << 16)) & mask;
+    out[17] = (w7 >> 11) & mask;
+    out[18] = ((w7 >> 38) | (w8 << 26)) & mask;
+    out[19] = (w8 >> 1) & mask;
+    out[20] = (w8 >> 28) & mask;
+    out[21] = ((w8 >> 55) | (w9 << 9)) & mask;
+    out[22] = (w9 >> 18) & mask;
+    out[23] = ((w9 >> 45) | (w10 << 19)) & mask;
+    out[24] = (w10 >> 8) & mask;
+    out[25] = (w10 >> 35) & mask;
+    out[26] = ((w10 >> 62) | (w11 << 2)) & mask;
+    out[27] = (w11 >> 25) & mask;
+    out[28] = ((w11 >> 52) | (w12 << 12)) & mask;
+    out[29] = (w12 >> 15) & mask;
+    out[30] = ((w12 >> 42) | (w13 << 22)) & mask;
+    out[31] = (w13 >> 5) & mask;
+
+    return in + (13 * 8 + 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 28> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 28) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    const auto w8 = LoadInt<uint64_t>(in + 8 * 8);
+    const auto w9 = LoadInt<uint64_t>(in + 9 * 8);
+    const auto w10 = LoadInt<uint64_t>(in + 10 * 8);
+    const auto w11 = LoadInt<uint64_t>(in + 11 * 8);
+    const auto w12 = LoadInt<uint64_t>(in + 12 * 8);
+    const auto w13 = LoadInt<uint64_t>(in + 13 * 8);
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 28) & mask;
+    out[2] = ((w0 >> 56) | (w1 << 8)) & mask;
+    out[3] = (w1 >> 20) & mask;
+    out[4] = ((w1 >> 48) | (w2 << 16)) & mask;
+    out[5] = (w2 >> 12) & mask;
+    out[6] = ((w2 >> 40) | (w3 << 24)) & mask;
+    out[7] = (w3 >> 4) & mask;
+    out[8] = (w3 >> 32) & mask;
+    out[9] = ((w3 >> 60) | (w4 << 4)) & mask;
+    out[10] = (w4 >> 24) & mask;
+    out[11] = ((w4 >> 52) | (w5 << 12)) & mask;
+    out[12] = (w5 >> 16) & mask;
+    out[13] = ((w5 >> 44) | (w6 << 20)) & mask;
+    out[14] = (w6 >> 8) & mask;
+    out[15] = w6 >> 36;
+    out[16] = (w7) & mask;
+    out[17] = (w7 >> 28) & mask;
+    out[18] = ((w7 >> 56) | (w8 << 8)) & mask;
+    out[19] = (w8 >> 20) & mask;
+    out[20] = ((w8 >> 48) | (w9 << 16)) & mask;
+    out[21] = (w9 >> 12) & mask;
+    out[22] = ((w9 >> 40) | (w10 << 24)) & mask;
+    out[23] = (w10 >> 4) & mask;
+    out[24] = (w10 >> 32) & mask;
+    out[25] = ((w10 >> 60) | (w11 << 4)) & mask;
+    out[26] = (w11 >> 24) & mask;
+    out[27] = ((w11 >> 52) | (w12 << 12)) & mask;
+    out[28] = (w12 >> 16) & mask;
+    out[29] = ((w12 >> 44) | (w13 << 20)) & mask;
+    out[30] = (w13 >> 8) & mask;
+    out[31] = w13 >> 36;
+
+    return in + (14 * 8);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 29> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 29) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    const auto w8 = LoadInt<uint64_t>(in + 8 * 8);
+    const auto w9 = LoadInt<uint64_t>(in + 9 * 8);
+    const auto w10 = LoadInt<uint64_t>(in + 10 * 8);
+    const auto w11 = LoadInt<uint64_t>(in + 11 * 8);
+    const auto w12 = LoadInt<uint64_t>(in + 12 * 8);
+    const auto w13 = LoadInt<uint64_t>(in + 13 * 8);
+    const auto w14 = static_cast<uint64_t>(LoadInt<uint32_t>(in + 14 * 8));
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 29) & mask;
+    out[2] = ((w0 >> 58) | (w1 << 6)) & mask;
+    out[3] = (w1 >> 23) & mask;
+    out[4] = ((w1 >> 52) | (w2 << 12)) & mask;
+    out[5] = (w2 >> 17) & mask;
+    out[6] = ((w2 >> 46) | (w3 << 18)) & mask;
+    out[7] = (w3 >> 11) & mask;
+    out[8] = ((w3 >> 40) | (w4 << 24)) & mask;
+    out[9] = (w4 >> 5) & mask;
+    out[10] = (w4 >> 34) & mask;
+    out[11] = ((w4 >> 63) | (w5 << 1)) & mask;
+    out[12] = (w5 >> 28) & mask;
+    out[13] = ((w5 >> 57) | (w6 << 7)) & mask;
+    out[14] = (w6 >> 22) & mask;
+    out[15] = ((w6 >> 51) | (w7 << 13)) & mask;
+    out[16] = (w7 >> 16) & mask;
+    out[17] = ((w7 >> 45) | (w8 << 19)) & mask;
+    out[18] = (w8 >> 10) & mask;
+    out[19] = ((w8 >> 39) | (w9 << 25)) & mask;
+    out[20] = (w9 >> 4) & mask;
+    out[21] = (w9 >> 33) & mask;
+    out[22] = ((w9 >> 62) | (w10 << 2)) & mask;
+    out[23] = (w10 >> 27) & mask;
+    out[24] = ((w10 >> 56) | (w11 << 8)) & mask;
+    out[25] = (w11 >> 21) & mask;
+    out[26] = ((w11 >> 50) | (w12 << 14)) & mask;
+    out[27] = (w12 >> 15) & mask;
+    out[28] = ((w12 >> 44) | (w13 << 20)) & mask;
+    out[29] = (w13 >> 9) & mask;
+    out[30] = ((w13 >> 38) | (w14 << 26)) & mask;
+    out[31] = (w14 >> 3) & mask;
+
+    return in + (14 * 8 + 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 30> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 30) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    const auto w8 = LoadInt<uint64_t>(in + 8 * 8);
+    const auto w9 = LoadInt<uint64_t>(in + 9 * 8);
+    const auto w10 = LoadInt<uint64_t>(in + 10 * 8);
+    const auto w11 = LoadInt<uint64_t>(in + 11 * 8);
+    const auto w12 = LoadInt<uint64_t>(in + 12 * 8);
+    const auto w13 = LoadInt<uint64_t>(in + 13 * 8);
+    const auto w14 = LoadInt<uint64_t>(in + 14 * 8);
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 30) & mask;
+    out[2] = ((w0 >> 60) | (w1 << 4)) & mask;
+    out[3] = (w1 >> 26) & mask;
+    out[4] = ((w1 >> 56) | (w2 << 8)) & mask;
+    out[5] = (w2 >> 22) & mask;
+    out[6] = ((w2 >> 52) | (w3 << 12)) & mask;
+    out[7] = (w3 >> 18) & mask;
+    out[8] = ((w3 >> 48) | (w4 << 16)) & mask;
+    out[9] = (w4 >> 14) & mask;
+    out[10] = ((w4 >> 44) | (w5 << 20)) & mask;
+    out[11] = (w5 >> 10) & mask;
+    out[12] = ((w5 >> 40) | (w6 << 24)) & mask;
+    out[13] = (w6 >> 6) & mask;
+    out[14] = ((w6 >> 36) | (w7 << 28)) & mask;
+    out[15] = (w7 >> 2) & mask;
+    out[16] = (w7 >> 32) & mask;
+    out[17] = ((w7 >> 62) | (w8 << 2)) & mask;
+    out[18] = (w8 >> 28) & mask;
+    out[19] = ((w8 >> 58) | (w9 << 6)) & mask;
+    out[20] = (w9 >> 24) & mask;
+    out[21] = ((w9 >> 54) | (w10 << 10)) & mask;
+    out[22] = (w10 >> 20) & mask;
+    out[23] = ((w10 >> 50) | (w11 << 14)) & mask;
+    out[24] = (w11 >> 16) & mask;
+    out[25] = ((w11 >> 46) | (w12 << 18)) & mask;
+    out[26] = (w12 >> 12) & mask;
+    out[27] = ((w12 >> 42) | (w13 << 22)) & mask;
+    out[28] = (w13 >> 8) & mask;
+    out[29] = ((w13 >> 38) | (w14 << 26)) & mask;
+    out[30] = (w14 >> 4) & mask;
+    out[31] = w14 >> 34;
+
+    return in + (15 * 8);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 31> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 31) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    const auto w8 = LoadInt<uint64_t>(in + 8 * 8);
+    const auto w9 = LoadInt<uint64_t>(in + 9 * 8);
+    const auto w10 = LoadInt<uint64_t>(in + 10 * 8);
+    const auto w11 = LoadInt<uint64_t>(in + 11 * 8);
+    const auto w12 = LoadInt<uint64_t>(in + 12 * 8);
+    const auto w13 = LoadInt<uint64_t>(in + 13 * 8);
+    const auto w14 = LoadInt<uint64_t>(in + 14 * 8);
+    const auto w15 = static_cast<uint64_t>(LoadInt<uint32_t>(in + 15 * 8));
+    out[0] = (w0) & mask;
+    out[1] = (w0 >> 31) & mask;
+    out[2] = ((w0 >> 62) | (w1 << 2)) & mask;
+    out[3] = (w1 >> 29) & mask;
+    out[4] = ((w1 >> 60) | (w2 << 4)) & mask;
+    out[5] = (w2 >> 27) & mask;
+    out[6] = ((w2 >> 58) | (w3 << 6)) & mask;
+    out[7] = (w3 >> 25) & mask;
+    out[8] = ((w3 >> 56) | (w4 << 8)) & mask;
+    out[9] = (w4 >> 23) & mask;
+    out[10] = ((w4 >> 54) | (w5 << 10)) & mask;
+    out[11] = (w5 >> 21) & mask;
+    out[12] = ((w5 >> 52) | (w6 << 12)) & mask;
+    out[13] = (w6 >> 19) & mask;
+    out[14] = ((w6 >> 50) | (w7 << 14)) & mask;
+    out[15] = (w7 >> 17) & mask;
+    out[16] = ((w7 >> 48) | (w8 << 16)) & mask;
+    out[17] = (w8 >> 15) & mask;
+    out[18] = ((w8 >> 46) | (w9 << 18)) & mask;
+    out[19] = (w9 >> 13) & mask;
+    out[20] = ((w9 >> 44) | (w10 << 20)) & mask;
+    out[21] = (w10 >> 11) & mask;
+    out[22] = ((w10 >> 42) | (w11 << 22)) & mask;
+    out[23] = (w11 >> 9) & mask;
+    out[24] = ((w11 >> 40) | (w12 << 24)) & mask;
+    out[25] = (w12 >> 7) & mask;
+    out[26] = ((w12 >> 38) | (w13 << 26)) & mask;
+    out[27] = (w13 >> 5) & mask;
+    out[28] = ((w13 >> 36) | (w14 << 28)) & mask;
+    out[29] = (w14 >> 3) & mask;
+    out[30] = ((w14 >> 34) | (w15 << 30)) & mask;
+    out[31] = (w15 >> 1) & mask;
+
+    return in + (15 * 8 + 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 32> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 32) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    const auto w8 = LoadInt<uint64_t>(in + 8 * 8);
+    const auto w9 = LoadInt<uint64_t>(in + 9 * 8);
+    const auto w10 = LoadInt<uint64_t>(in + 10 * 8);
+    const auto w11 = LoadInt<uint64_t>(in + 11 * 8);
+    const auto w12 = LoadInt<uint64_t>(in + 12 * 8);
+    const auto w13 = LoadInt<uint64_t>(in + 13 * 8);
+    const auto w14 = LoadInt<uint64_t>(in + 14 * 8);
+    const auto w15 = LoadInt<uint64_t>(in + 15 * 8);
+    out[0] = (w0) & mask;
+    out[1] = w0 >> 32;
+    out[2] = (w1) & mask;
+    out[3] = w1 >> 32;
+    out[4] = (w2) & mask;
+    out[5] = w2 >> 32;
+    out[6] = (w3) & mask;
+    out[7] = w3 >> 32;
+    out[8] = (w4) & mask;
+    out[9] = w4 >> 32;
+    out[10] = (w5) & mask;
+    out[11] = w5 >> 32;
+    out[12] = (w6) & mask;
+    out[13] = w6 >> 32;
+    out[14] = (w7) & mask;
+    out[15] = w7 >> 32;
+    out[16] = (w8) & mask;
+    out[17] = w8 >> 32;
+    out[18] = (w9) & mask;
+    out[19] = w9 >> 32;
+    out[20] = (w10) & mask;
+    out[21] = w10 >> 32;
+    out[22] = (w11) & mask;
+    out[23] = w11 >> 32;
+    out[24] = (w12) & mask;
+    out[25] = w12 >> 32;
+    out[26] = (w13) & mask;
+    out[27] = w13 >> 32;
+    out[28] = (w14) & mask;
+    out[29] = w14 >> 32;
+    out[30] = (w15) & mask;
+    out[31] = w15 >> 32;
+
+    return in + (16 * 8);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 33> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 33) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    const auto w8 = LoadInt<uint64_t>(in + 8 * 8);
+    const auto w9 = LoadInt<uint64_t>(in + 9 * 8);
+    const auto w10 = LoadInt<uint64_t>(in + 10 * 8);
+    const auto w11 = LoadInt<uint64_t>(in + 11 * 8);
+    const auto w12 = LoadInt<uint64_t>(in + 12 * 8);
+    const auto w13 = LoadInt<uint64_t>(in + 13 * 8);
+    const auto w14 = LoadInt<uint64_t>(in + 14 * 8);
+    const auto w15 = LoadInt<uint64_t>(in + 15 * 8);
+    const auto w16 = static_cast<uint64_t>(LoadInt<uint32_t>(in + 16 * 8));
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 33) | (w1 << 31)) & mask;
+    out[2] = (w1 >> 2) & mask;
+    out[3] = ((w1 >> 35) | (w2 << 29)) & mask;
+    out[4] = (w2 >> 4) & mask;
+    out[5] = ((w2 >> 37) | (w3 << 27)) & mask;
+    out[6] = (w3 >> 6) & mask;
+    out[7] = ((w3 >> 39) | (w4 << 25)) & mask;
+    out[8] = (w4 >> 8) & mask;
+    out[9] = ((w4 >> 41) | (w5 << 23)) & mask;
+    out[10] = (w5 >> 10) & mask;
+    out[11] = ((w5 >> 43) | (w6 << 21)) & mask;
+    out[12] = (w6 >> 12) & mask;
+    out[13] = ((w6 >> 45) | (w7 << 19)) & mask;
+    out[14] = (w7 >> 14) & mask;
+    out[15] = ((w7 >> 47) | (w8 << 17)) & mask;
+    out[16] = (w8 >> 16) & mask;
+    out[17] = ((w8 >> 49) | (w9 << 15)) & mask;
+    out[18] = (w9 >> 18) & mask;
+    out[19] = ((w9 >> 51) | (w10 << 13)) & mask;
+    out[20] = (w10 >> 20) & mask;
+    out[21] = ((w10 >> 53) | (w11 << 11)) & mask;
+    out[22] = (w11 >> 22) & mask;
+    out[23] = ((w11 >> 55) | (w12 << 9)) & mask;
+    out[24] = (w12 >> 24) & mask;
+    out[25] = ((w12 >> 57) | (w13 << 7)) & mask;
+    out[26] = (w13 >> 26) & mask;
+    out[27] = ((w13 >> 59) | (w14 << 5)) & mask;
+    out[28] = (w14 >> 28) & mask;
+    out[29] = ((w14 >> 61) | (w15 << 3)) & mask;
+    out[30] = (w15 >> 30) & mask;
+    out[31] = ((w15 >> 63) | (w16 << 1)) & mask;
+
+    return in + (16 * 8 + 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 34> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 34) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    const auto w8 = LoadInt<uint64_t>(in + 8 * 8);
+    const auto w9 = LoadInt<uint64_t>(in + 9 * 8);
+    const auto w10 = LoadInt<uint64_t>(in + 10 * 8);
+    const auto w11 = LoadInt<uint64_t>(in + 11 * 8);
+    const auto w12 = LoadInt<uint64_t>(in + 12 * 8);
+    const auto w13 = LoadInt<uint64_t>(in + 13 * 8);
+    const auto w14 = LoadInt<uint64_t>(in + 14 * 8);
+    const auto w15 = LoadInt<uint64_t>(in + 15 * 8);
+    const auto w16 = LoadInt<uint64_t>(in + 16 * 8);
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 34) | (w1 << 30)) & mask;
+    out[2] = (w1 >> 4) & mask;
+    out[3] = ((w1 >> 38) | (w2 << 26)) & mask;
+    out[4] = (w2 >> 8) & mask;
+    out[5] = ((w2 >> 42) | (w3 << 22)) & mask;
+    out[6] = (w3 >> 12) & mask;
+    out[7] = ((w3 >> 46) | (w4 << 18)) & mask;
+    out[8] = (w4 >> 16) & mask;
+    out[9] = ((w4 >> 50) | (w5 << 14)) & mask;
+    out[10] = (w5 >> 20) & mask;
+    out[11] = ((w5 >> 54) | (w6 << 10)) & mask;
+    out[12] = (w6 >> 24) & mask;
+    out[13] = ((w6 >> 58) | (w7 << 6)) & mask;
+    out[14] = (w7 >> 28) & mask;
+    out[15] = ((w7 >> 62) | (w8 << 2)) & mask;
+    out[16] = ((w8 >> 32) | (w9 << 32)) & mask;
+    out[17] = (w9 >> 2) & mask;
+    out[18] = ((w9 >> 36) | (w10 << 28)) & mask;
+    out[19] = (w10 >> 6) & mask;
+    out[20] = ((w10 >> 40) | (w11 << 24)) & mask;
+    out[21] = (w11 >> 10) & mask;
+    out[22] = ((w11 >> 44) | (w12 << 20)) & mask;
+    out[23] = (w12 >> 14) & mask;
+    out[24] = ((w12 >> 48) | (w13 << 16)) & mask;
+    out[25] = (w13 >> 18) & mask;
+    out[26] = ((w13 >> 52) | (w14 << 12)) & mask;
+    out[27] = (w14 >> 22) & mask;
+    out[28] = ((w14 >> 56) | (w15 << 8)) & mask;
+    out[29] = (w15 >> 26) & mask;
+    out[30] = ((w15 >> 60) | (w16 << 4)) & mask;
+    out[31] = w16 >> 30;
+
+    return in + (17 * 8);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 35> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 35) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    const auto w8 = LoadInt<uint64_t>(in + 8 * 8);
+    const auto w9 = LoadInt<uint64_t>(in + 9 * 8);
+    const auto w10 = LoadInt<uint64_t>(in + 10 * 8);
+    const auto w11 = LoadInt<uint64_t>(in + 11 * 8);
+    const auto w12 = LoadInt<uint64_t>(in + 12 * 8);
+    const auto w13 = LoadInt<uint64_t>(in + 13 * 8);
+    const auto w14 = LoadInt<uint64_t>(in + 14 * 8);
+    const auto w15 = LoadInt<uint64_t>(in + 15 * 8);
+    const auto w16 = LoadInt<uint64_t>(in + 16 * 8);
+    const auto w17 = static_cast<uint64_t>(LoadInt<uint32_t>(in + 17 * 8));
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 35) | (w1 << 29)) & mask;
+    out[2] = (w1 >> 6) & mask;
+    out[3] = ((w1 >> 41) | (w2 << 23)) & mask;
+    out[4] = (w2 >> 12) & mask;
+    out[5] = ((w2 >> 47) | (w3 << 17)) & mask;
+    out[6] = (w3 >> 18) & mask;
+    out[7] = ((w3 >> 53) | (w4 << 11)) & mask;
+    out[8] = (w4 >> 24) & mask;
+    out[9] = ((w4 >> 59) | (w5 << 5)) & mask;
+    out[10] = ((w5 >> 30) | (w6 << 34)) & mask;
+    out[11] = (w6 >> 1) & mask;
+    out[12] = ((w6 >> 36) | (w7 << 28)) & mask;
+    out[13] = (w7 >> 7) & mask;
+    out[14] = ((w7 >> 42) | (w8 << 22)) & mask;
+    out[15] = (w8 >> 13) & mask;
+    out[16] = ((w8 >> 48) | (w9 << 16)) & mask;
+    out[17] = (w9 >> 19) & mask;
+    out[18] = ((w9 >> 54) | (w10 << 10)) & mask;
+    out[19] = (w10 >> 25) & mask;
+    out[20] = ((w10 >> 60) | (w11 << 4)) & mask;
+    out[21] = ((w11 >> 31) | (w12 << 33)) & mask;
+    out[22] = (w12 >> 2) & mask;
+    out[23] = ((w12 >> 37) | (w13 << 27)) & mask;
+    out[24] = (w13 >> 8) & mask;
+    out[25] = ((w13 >> 43) | (w14 << 21)) & mask;
+    out[26] = (w14 >> 14) & mask;
+    out[27] = ((w14 >> 49) | (w15 << 15)) & mask;
+    out[28] = (w15 >> 20) & mask;
+    out[29] = ((w15 >> 55) | (w16 << 9)) & mask;
+    out[30] = (w16 >> 26) & mask;
+    out[31] = ((w16 >> 61) | (w17 << 3)) & mask;
+
+    return in + (17 * 8 + 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 36> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 36) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    const auto w8 = LoadInt<uint64_t>(in + 8 * 8);
+    const auto w9 = LoadInt<uint64_t>(in + 9 * 8);
+    const auto w10 = LoadInt<uint64_t>(in + 10 * 8);
+    const auto w11 = LoadInt<uint64_t>(in + 11 * 8);
+    const auto w12 = LoadInt<uint64_t>(in + 12 * 8);
+    const auto w13 = LoadInt<uint64_t>(in + 13 * 8);
+    const auto w14 = LoadInt<uint64_t>(in + 14 * 8);
+    const auto w15 = LoadInt<uint64_t>(in + 15 * 8);
+    const auto w16 = LoadInt<uint64_t>(in + 16 * 8);
+    const auto w17 = LoadInt<uint64_t>(in + 17 * 8);
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 36) | (w1 << 28)) & mask;
+    out[2] = (w1 >> 8) & mask;
+    out[3] = ((w1 >> 44) | (w2 << 20)) & mask;
+    out[4] = (w2 >> 16) & mask;
+    out[5] = ((w2 >> 52) | (w3 << 12)) & mask;
+    out[6] = (w3 >> 24) & mask;
+    out[7] = ((w3 >> 60) | (w4 << 4)) & mask;
+    out[8] = ((w4 >> 32) | (w5 << 32)) & mask;
+    out[9] = (w5 >> 4) & mask;
+    out[10] = ((w5 >> 40) | (w6 << 24)) & mask;
+    out[11] = (w6 >> 12) & mask;
+    out[12] = ((w6 >> 48) | (w7 << 16)) & mask;
+    out[13] = (w7 >> 20) & mask;
+    out[14] = ((w7 >> 56) | (w8 << 8)) & mask;
+    out[15] = w8 >> 28;
+    out[16] = (w9) & mask;
+    out[17] = ((w9 >> 36) | (w10 << 28)) & mask;
+    out[18] = (w10 >> 8) & mask;
+    out[19] = ((w10 >> 44) | (w11 << 20)) & mask;
+    out[20] = (w11 >> 16) & mask;
+    out[21] = ((w11 >> 52) | (w12 << 12)) & mask;
+    out[22] = (w12 >> 24) & mask;
+    out[23] = ((w12 >> 60) | (w13 << 4)) & mask;
+    out[24] = ((w13 >> 32) | (w14 << 32)) & mask;
+    out[25] = (w14 >> 4) & mask;
+    out[26] = ((w14 >> 40) | (w15 << 24)) & mask;
+    out[27] = (w15 >> 12) & mask;
+    out[28] = ((w15 >> 48) | (w16 << 16)) & mask;
+    out[29] = (w16 >> 20) & mask;
+    out[30] = ((w16 >> 56) | (w17 << 8)) & mask;
+    out[31] = w17 >> 28;
+
+    return in + (18 * 8);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 37> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 37) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    const auto w8 = LoadInt<uint64_t>(in + 8 * 8);
+    const auto w9 = LoadInt<uint64_t>(in + 9 * 8);
+    const auto w10 = LoadInt<uint64_t>(in + 10 * 8);
+    const auto w11 = LoadInt<uint64_t>(in + 11 * 8);
+    const auto w12 = LoadInt<uint64_t>(in + 12 * 8);
+    const auto w13 = LoadInt<uint64_t>(in + 13 * 8);
+    const auto w14 = LoadInt<uint64_t>(in + 14 * 8);
+    const auto w15 = LoadInt<uint64_t>(in + 15 * 8);
+    const auto w16 = LoadInt<uint64_t>(in + 16 * 8);
+    const auto w17 = LoadInt<uint64_t>(in + 17 * 8);
+    const auto w18 = static_cast<uint64_t>(LoadInt<uint32_t>(in + 18 * 8));
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 37) | (w1 << 27)) & mask;
+    out[2] = (w1 >> 10) & mask;
+    out[3] = ((w1 >> 47) | (w2 << 17)) & mask;
+    out[4] = (w2 >> 20) & mask;
+    out[5] = ((w2 >> 57) | (w3 << 7)) & mask;
+    out[6] = ((w3 >> 30) | (w4 << 34)) & mask;
+    out[7] = (w4 >> 3) & mask;
+    out[8] = ((w4 >> 40) | (w5 << 24)) & mask;
+    out[9] = (w5 >> 13) & mask;
+    out[10] = ((w5 >> 50) | (w6 << 14)) & mask;
+    out[11] = (w6 >> 23) & mask;
+    out[12] = ((w6 >> 60) | (w7 << 4)) & mask;
+    out[13] = ((w7 >> 33) | (w8 << 31)) & mask;
+    out[14] = (w8 >> 6) & mask;
+    out[15] = ((w8 >> 43) | (w9 << 21)) & mask;
+    out[16] = (w9 >> 16) & mask;
+    out[17] = ((w9 >> 53) | (w10 << 11)) & mask;
+    out[18] = (w10 >> 26) & mask;
+    out[19] = ((w10 >> 63) | (w11 << 1)) & mask;
+    out[20] = ((w11 >> 36) | (w12 << 28)) & mask;
+    out[21] = (w12 >> 9) & mask;
+    out[22] = ((w12 >> 46) | (w13 << 18)) & mask;
+    out[23] = (w13 >> 19) & mask;
+    out[24] = ((w13 >> 56) | (w14 << 8)) & mask;
+    out[25] = ((w14 >> 29) | (w15 << 35)) & mask;
+    out[26] = (w15 >> 2) & mask;
+    out[27] = ((w15 >> 39) | (w16 << 25)) & mask;
+    out[28] = (w16 >> 12) & mask;
+    out[29] = ((w16 >> 49) | (w17 << 15)) & mask;
+    out[30] = (w17 >> 22) & mask;
+    out[31] = ((w17 >> 59) | (w18 << 5)) & mask;
+
+    return in + (18 * 8 + 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 38> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 38) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    const auto w8 = LoadInt<uint64_t>(in + 8 * 8);
+    const auto w9 = LoadInt<uint64_t>(in + 9 * 8);
+    const auto w10 = LoadInt<uint64_t>(in + 10 * 8);
+    const auto w11 = LoadInt<uint64_t>(in + 11 * 8);
+    const auto w12 = LoadInt<uint64_t>(in + 12 * 8);
+    const auto w13 = LoadInt<uint64_t>(in + 13 * 8);
+    const auto w14 = LoadInt<uint64_t>(in + 14 * 8);
+    const auto w15 = LoadInt<uint64_t>(in + 15 * 8);
+    const auto w16 = LoadInt<uint64_t>(in + 16 * 8);
+    const auto w17 = LoadInt<uint64_t>(in + 17 * 8);
+    const auto w18 = LoadInt<uint64_t>(in + 18 * 8);
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 38) | (w1 << 26)) & mask;
+    out[2] = (w1 >> 12) & mask;
+    out[3] = ((w1 >> 50) | (w2 << 14)) & mask;
+    out[4] = (w2 >> 24) & mask;
+    out[5] = ((w2 >> 62) | (w3 << 2)) & mask;
+    out[6] = ((w3 >> 36) | (w4 << 28)) & mask;
+    out[7] = (w4 >> 10) & mask;
+    out[8] = ((w4 >> 48) | (w5 << 16)) & mask;
+    out[9] = (w5 >> 22) & mask;
+    out[10] = ((w5 >> 60) | (w6 << 4)) & mask;
+    out[11] = ((w6 >> 34) | (w7 << 30)) & mask;
+    out[12] = (w7 >> 8) & mask;
+    out[13] = ((w7 >> 46) | (w8 << 18)) & mask;
+    out[14] = (w8 >> 20) & mask;
+    out[15] = ((w8 >> 58) | (w9 << 6)) & mask;
+    out[16] = ((w9 >> 32) | (w10 << 32)) & mask;
+    out[17] = (w10 >> 6) & mask;
+    out[18] = ((w10 >> 44) | (w11 << 20)) & mask;
+    out[19] = (w11 >> 18) & mask;
+    out[20] = ((w11 >> 56) | (w12 << 8)) & mask;
+    out[21] = ((w12 >> 30) | (w13 << 34)) & mask;
+    out[22] = (w13 >> 4) & mask;
+    out[23] = ((w13 >> 42) | (w14 << 22)) & mask;
+    out[24] = (w14 >> 16) & mask;
+    out[25] = ((w14 >> 54) | (w15 << 10)) & mask;
+    out[26] = ((w15 >> 28) | (w16 << 36)) & mask;
+    out[27] = (w16 >> 2) & mask;
+    out[28] = ((w16 >> 40) | (w17 << 24)) & mask;
+    out[29] = (w17 >> 14) & mask;
+    out[30] = ((w17 >> 52) | (w18 << 12)) & mask;
+    out[31] = w18 >> 26;
+
+    return in + (19 * 8);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 39> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 39) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    const auto w8 = LoadInt<uint64_t>(in + 8 * 8);
+    const auto w9 = LoadInt<uint64_t>(in + 9 * 8);
+    const auto w10 = LoadInt<uint64_t>(in + 10 * 8);
+    const auto w11 = LoadInt<uint64_t>(in + 11 * 8);
+    const auto w12 = LoadInt<uint64_t>(in + 12 * 8);
+    const auto w13 = LoadInt<uint64_t>(in + 13 * 8);
+    const auto w14 = LoadInt<uint64_t>(in + 14 * 8);
+    const auto w15 = LoadInt<uint64_t>(in + 15 * 8);
+    const auto w16 = LoadInt<uint64_t>(in + 16 * 8);
+    const auto w17 = LoadInt<uint64_t>(in + 17 * 8);
+    const auto w18 = LoadInt<uint64_t>(in + 18 * 8);
+    const auto w19 = static_cast<uint64_t>(LoadInt<uint32_t>(in + 19 * 8));
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 39) | (w1 << 25)) & mask;
+    out[2] = (w1 >> 14) & mask;
+    out[3] = ((w1 >> 53) | (w2 << 11)) & mask;
+    out[4] = ((w2 >> 28) | (w3 << 36)) & mask;
+    out[5] = (w3 >> 3) & mask;
+    out[6] = ((w3 >> 42) | (w4 << 22)) & mask;
+    out[7] = (w4 >> 17) & mask;
+    out[8] = ((w4 >> 56) | (w5 << 8)) & mask;
+    out[9] = ((w5 >> 31) | (w6 << 33)) & mask;
+    out[10] = (w6 >> 6) & mask;
+    out[11] = ((w6 >> 45) | (w7 << 19)) & mask;
+    out[12] = (w7 >> 20) & mask;
+    out[13] = ((w7 >> 59) | (w8 << 5)) & mask;
+    out[14] = ((w8 >> 34) | (w9 << 30)) & mask;
+    out[15] = (w9 >> 9) & mask;
+    out[16] = ((w9 >> 48) | (w10 << 16)) & mask;
+    out[17] = (w10 >> 23) & mask;
+    out[18] = ((w10 >> 62) | (w11 << 2)) & mask;
+    out[19] = ((w11 >> 37) | (w12 << 27)) & mask;
+    out[20] = (w12 >> 12) & mask;
+    out[21] = ((w12 >> 51) | (w13 << 13)) & mask;
+    out[22] = ((w13 >> 26) | (w14 << 38)) & mask;
+    out[23] = (w14 >> 1) & mask;
+    out[24] = ((w14 >> 40) | (w15 << 24)) & mask;
+    out[25] = (w15 >> 15) & mask;
+    out[26] = ((w15 >> 54) | (w16 << 10)) & mask;
+    out[27] = ((w16 >> 29) | (w17 << 35)) & mask;
+    out[28] = (w17 >> 4) & mask;
+    out[29] = ((w17 >> 43) | (w18 << 21)) & mask;
+    out[30] = (w18 >> 18) & mask;
+    out[31] = ((w18 >> 57) | (w19 << 7)) & mask;
+
+    return in + (19 * 8 + 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 40> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 40) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    const auto w8 = LoadInt<uint64_t>(in + 8 * 8);
+    const auto w9 = LoadInt<uint64_t>(in + 9 * 8);
+    const auto w10 = LoadInt<uint64_t>(in + 10 * 8);
+    const auto w11 = LoadInt<uint64_t>(in + 11 * 8);
+    const auto w12 = LoadInt<uint64_t>(in + 12 * 8);
+    const auto w13 = LoadInt<uint64_t>(in + 13 * 8);
+    const auto w14 = LoadInt<uint64_t>(in + 14 * 8);
+    const auto w15 = LoadInt<uint64_t>(in + 15 * 8);
+    const auto w16 = LoadInt<uint64_t>(in + 16 * 8);
+    const auto w17 = LoadInt<uint64_t>(in + 17 * 8);
+    const auto w18 = LoadInt<uint64_t>(in + 18 * 8);
+    const auto w19 = LoadInt<uint64_t>(in + 19 * 8);
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 40) | (w1 << 24)) & mask;
+    out[2] = (w1 >> 16) & mask;
+    out[3] = ((w1 >> 56) | (w2 << 8)) & mask;
+    out[4] = ((w2 >> 32) | (w3 << 32)) & mask;
+    out[5] = (w3 >> 8) & mask;
+    out[6] = ((w3 >> 48) | (w4 << 16)) & mask;
+    out[7] = w4 >> 24;
+    out[8] = (w5) & mask;
+    out[9] = ((w5 >> 40) | (w6 << 24)) & mask;
+    out[10] = (w6 >> 16) & mask;
+    out[11] = ((w6 >> 56) | (w7 << 8)) & mask;
+    out[12] = ((w7 >> 32) | (w8 << 32)) & mask;
+    out[13] = (w8 >> 8) & mask;
+    out[14] = ((w8 >> 48) | (w9 << 16)) & mask;
+    out[15] = w9 >> 24;
+    out[16] = (w10) & mask;
+    out[17] = ((w10 >> 40) | (w11 << 24)) & mask;
+    out[18] = (w11 >> 16) & mask;
+    out[19] = ((w11 >> 56) | (w12 << 8)) & mask;
+    out[20] = ((w12 >> 32) | (w13 << 32)) & mask;
+    out[21] = (w13 >> 8) & mask;
+    out[22] = ((w13 >> 48) | (w14 << 16)) & mask;
+    out[23] = w14 >> 24;
+    out[24] = (w15) & mask;
+    out[25] = ((w15 >> 40) | (w16 << 24)) & mask;
+    out[26] = (w16 >> 16) & mask;
+    out[27] = ((w16 >> 56) | (w17 << 8)) & mask;
+    out[28] = ((w17 >> 32) | (w18 << 32)) & mask;
+    out[29] = (w18 >> 8) & mask;
+    out[30] = ((w18 >> 48) | (w19 << 16)) & mask;
+    out[31] = w19 >> 24;
+
+    return in + (20 * 8);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 41> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 41) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    const auto w8 = LoadInt<uint64_t>(in + 8 * 8);
+    const auto w9 = LoadInt<uint64_t>(in + 9 * 8);
+    const auto w10 = LoadInt<uint64_t>(in + 10 * 8);
+    const auto w11 = LoadInt<uint64_t>(in + 11 * 8);
+    const auto w12 = LoadInt<uint64_t>(in + 12 * 8);
+    const auto w13 = LoadInt<uint64_t>(in + 13 * 8);
+    const auto w14 = LoadInt<uint64_t>(in + 14 * 8);
+    const auto w15 = LoadInt<uint64_t>(in + 15 * 8);
+    const auto w16 = LoadInt<uint64_t>(in + 16 * 8);
+    const auto w17 = LoadInt<uint64_t>(in + 17 * 8);
+    const auto w18 = LoadInt<uint64_t>(in + 18 * 8);
+    const auto w19 = LoadInt<uint64_t>(in + 19 * 8);
+    const auto w20 = static_cast<uint64_t>(LoadInt<uint32_t>(in + 20 * 8));
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 41) | (w1 << 23)) & mask;
+    out[2] = (w1 >> 18) & mask;
+    out[3] = ((w1 >> 59) | (w2 << 5)) & mask;
+    out[4] = ((w2 >> 36) | (w3 << 28)) & mask;
+    out[5] = (w3 >> 13) & mask;
+    out[6] = ((w3 >> 54) | (w4 << 10)) & mask;
+    out[7] = ((w4 >> 31) | (w5 << 33)) & mask;
+    out[8] = (w5 >> 8) & mask;
+    out[9] = ((w5 >> 49) | (w6 << 15)) & mask;
+    out[10] = ((w6 >> 26) | (w7 << 38)) & mask;
+    out[11] = (w7 >> 3) & mask;
+    out[12] = ((w7 >> 44) | (w8 << 20)) & mask;
+    out[13] = (w8 >> 21) & mask;
+    out[14] = ((w8 >> 62) | (w9 << 2)) & mask;
+    out[15] = ((w9 >> 39) | (w10 << 25)) & mask;
+    out[16] = (w10 >> 16) & mask;
+    out[17] = ((w10 >> 57) | (w11 << 7)) & mask;
+    out[18] = ((w11 >> 34) | (w12 << 30)) & mask;
+    out[19] = (w12 >> 11) & mask;
+    out[20] = ((w12 >> 52) | (w13 << 12)) & mask;
+    out[21] = ((w13 >> 29) | (w14 << 35)) & mask;
+    out[22] = (w14 >> 6) & mask;
+    out[23] = ((w14 >> 47) | (w15 << 17)) & mask;
+    out[24] = ((w15 >> 24) | (w16 << 40)) & mask;
+    out[25] = (w16 >> 1) & mask;
+    out[26] = ((w16 >> 42) | (w17 << 22)) & mask;
+    out[27] = (w17 >> 19) & mask;
+    out[28] = ((w17 >> 60) | (w18 << 4)) & mask;
+    out[29] = ((w18 >> 37) | (w19 << 27)) & mask;
+    out[30] = (w19 >> 14) & mask;
+    out[31] = ((w19 >> 55) | (w20 << 9)) & mask;
+
+    return in + (20 * 8 + 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 42> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 42) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    const auto w8 = LoadInt<uint64_t>(in + 8 * 8);
+    const auto w9 = LoadInt<uint64_t>(in + 9 * 8);
+    const auto w10 = LoadInt<uint64_t>(in + 10 * 8);
+    const auto w11 = LoadInt<uint64_t>(in + 11 * 8);
+    const auto w12 = LoadInt<uint64_t>(in + 12 * 8);
+    const auto w13 = LoadInt<uint64_t>(in + 13 * 8);
+    const auto w14 = LoadInt<uint64_t>(in + 14 * 8);
+    const auto w15 = LoadInt<uint64_t>(in + 15 * 8);
+    const auto w16 = LoadInt<uint64_t>(in + 16 * 8);
+    const auto w17 = LoadInt<uint64_t>(in + 17 * 8);
+    const auto w18 = LoadInt<uint64_t>(in + 18 * 8);
+    const auto w19 = LoadInt<uint64_t>(in + 19 * 8);
+    const auto w20 = LoadInt<uint64_t>(in + 20 * 8);
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 42) | (w1 << 22)) & mask;
+    out[2] = (w1 >> 20) & mask;
+    out[3] = ((w1 >> 62) | (w2 << 2)) & mask;
+    out[4] = ((w2 >> 40) | (w3 << 24)) & mask;
+    out[5] = (w3 >> 18) & mask;
+    out[6] = ((w3 >> 60) | (w4 << 4)) & mask;
+    out[7] = ((w4 >> 38) | (w5 << 26)) & mask;
+    out[8] = (w5 >> 16) & mask;
+    out[9] = ((w5 >> 58) | (w6 << 6)) & mask;
+    out[10] = ((w6 >> 36) | (w7 << 28)) & mask;
+    out[11] = (w7 >> 14) & mask;
+    out[12] = ((w7 >> 56) | (w8 << 8)) & mask;
+    out[13] = ((w8 >> 34) | (w9 << 30)) & mask;
+    out[14] = (w9 >> 12) & mask;
+    out[15] = ((w9 >> 54) | (w10 << 10)) & mask;
+    out[16] = ((w10 >> 32) | (w11 << 32)) & mask;
+    out[17] = (w11 >> 10) & mask;
+    out[18] = ((w11 >> 52) | (w12 << 12)) & mask;
+    out[19] = ((w12 >> 30) | (w13 << 34)) & mask;
+    out[20] = (w13 >> 8) & mask;
+    out[21] = ((w13 >> 50) | (w14 << 14)) & mask;
+    out[22] = ((w14 >> 28) | (w15 << 36)) & mask;
+    out[23] = (w15 >> 6) & mask;
+    out[24] = ((w15 >> 48) | (w16 << 16)) & mask;
+    out[25] = ((w16 >> 26) | (w17 << 38)) & mask;
+    out[26] = (w17 >> 4) & mask;
+    out[27] = ((w17 >> 46) | (w18 << 18)) & mask;
+    out[28] = ((w18 >> 24) | (w19 << 40)) & mask;
+    out[29] = (w19 >> 2) & mask;
+    out[30] = ((w19 >> 44) | (w20 << 20)) & mask;
+    out[31] = w20 >> 22;
+
+    return in + (21 * 8);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 43> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 43) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    const auto w8 = LoadInt<uint64_t>(in + 8 * 8);
+    const auto w9 = LoadInt<uint64_t>(in + 9 * 8);
+    const auto w10 = LoadInt<uint64_t>(in + 10 * 8);
+    const auto w11 = LoadInt<uint64_t>(in + 11 * 8);
+    const auto w12 = LoadInt<uint64_t>(in + 12 * 8);
+    const auto w13 = LoadInt<uint64_t>(in + 13 * 8);
+    const auto w14 = LoadInt<uint64_t>(in + 14 * 8);
+    const auto w15 = LoadInt<uint64_t>(in + 15 * 8);
+    const auto w16 = LoadInt<uint64_t>(in + 16 * 8);
+    const auto w17 = LoadInt<uint64_t>(in + 17 * 8);
+    const auto w18 = LoadInt<uint64_t>(in + 18 * 8);
+    const auto w19 = LoadInt<uint64_t>(in + 19 * 8);
+    const auto w20 = LoadInt<uint64_t>(in + 20 * 8);
+    const auto w21 = static_cast<uint64_t>(LoadInt<uint32_t>(in + 21 * 8));
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 43) | (w1 << 21)) & mask;
+    out[2] = ((w1 >> 22) | (w2 << 42)) & mask;
+    out[3] = (w2 >> 1) & mask;
+    out[4] = ((w2 >> 44) | (w3 << 20)) & mask;
+    out[5] = ((w3 >> 23) | (w4 << 41)) & mask;
+    out[6] = (w4 >> 2) & mask;
+    out[7] = ((w4 >> 45) | (w5 << 19)) & mask;
+    out[8] = ((w5 >> 24) | (w6 << 40)) & mask;
+    out[9] = (w6 >> 3) & mask;
+    out[10] = ((w6 >> 46) | (w7 << 18)) & mask;
+    out[11] = ((w7 >> 25) | (w8 << 39)) & mask;
+    out[12] = (w8 >> 4) & mask;
+    out[13] = ((w8 >> 47) | (w9 << 17)) & mask;
+    out[14] = ((w9 >> 26) | (w10 << 38)) & mask;
+    out[15] = (w10 >> 5) & mask;
+    out[16] = ((w10 >> 48) | (w11 << 16)) & mask;
+    out[17] = ((w11 >> 27) | (w12 << 37)) & mask;
+    out[18] = (w12 >> 6) & mask;
+    out[19] = ((w12 >> 49) | (w13 << 15)) & mask;
+    out[20] = ((w13 >> 28) | (w14 << 36)) & mask;
+    out[21] = (w14 >> 7) & mask;
+    out[22] = ((w14 >> 50) | (w15 << 14)) & mask;
+    out[23] = ((w15 >> 29) | (w16 << 35)) & mask;
+    out[24] = (w16 >> 8) & mask;
+    out[25] = ((w16 >> 51) | (w17 << 13)) & mask;
+    out[26] = ((w17 >> 30) | (w18 << 34)) & mask;
+    out[27] = (w18 >> 9) & mask;
+    out[28] = ((w18 >> 52) | (w19 << 12)) & mask;
+    out[29] = ((w19 >> 31) | (w20 << 33)) & mask;
+    out[30] = (w20 >> 10) & mask;
+    out[31] = ((w20 >> 53) | (w21 << 11)) & mask;
+
+    return in + (21 * 8 + 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 44> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 44) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    const auto w8 = LoadInt<uint64_t>(in + 8 * 8);
+    const auto w9 = LoadInt<uint64_t>(in + 9 * 8);
+    const auto w10 = LoadInt<uint64_t>(in + 10 * 8);
+    const auto w11 = LoadInt<uint64_t>(in + 11 * 8);
+    const auto w12 = LoadInt<uint64_t>(in + 12 * 8);
+    const auto w13 = LoadInt<uint64_t>(in + 13 * 8);
+    const auto w14 = LoadInt<uint64_t>(in + 14 * 8);
+    const auto w15 = LoadInt<uint64_t>(in + 15 * 8);
+    const auto w16 = LoadInt<uint64_t>(in + 16 * 8);
+    const auto w17 = LoadInt<uint64_t>(in + 17 * 8);
+    const auto w18 = LoadInt<uint64_t>(in + 18 * 8);
+    const auto w19 = LoadInt<uint64_t>(in + 19 * 8);
+    const auto w20 = LoadInt<uint64_t>(in + 20 * 8);
+    const auto w21 = LoadInt<uint64_t>(in + 21 * 8);
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 44) | (w1 << 20)) & mask;
+    out[2] = ((w1 >> 24) | (w2 << 40)) & mask;
+    out[3] = (w2 >> 4) & mask;
+    out[4] = ((w2 >> 48) | (w3 << 16)) & mask;
+    out[5] = ((w3 >> 28) | (w4 << 36)) & mask;
+    out[6] = (w4 >> 8) & mask;
+    out[7] = ((w4 >> 52) | (w5 << 12)) & mask;
+    out[8] = ((w5 >> 32) | (w6 << 32)) & mask;
+    out[9] = (w6 >> 12) & mask;
+    out[10] = ((w6 >> 56) | (w7 << 8)) & mask;
+    out[11] = ((w7 >> 36) | (w8 << 28)) & mask;
+    out[12] = (w8 >> 16) & mask;
+    out[13] = ((w8 >> 60) | (w9 << 4)) & mask;
+    out[14] = ((w9 >> 40) | (w10 << 24)) & mask;
+    out[15] = w10 >> 20;
+    out[16] = (w11) & mask;
+    out[17] = ((w11 >> 44) | (w12 << 20)) & mask;
+    out[18] = ((w12 >> 24) | (w13 << 40)) & mask;
+    out[19] = (w13 >> 4) & mask;
+    out[20] = ((w13 >> 48) | (w14 << 16)) & mask;
+    out[21] = ((w14 >> 28) | (w15 << 36)) & mask;
+    out[22] = (w15 >> 8) & mask;
+    out[23] = ((w15 >> 52) | (w16 << 12)) & mask;
+    out[24] = ((w16 >> 32) | (w17 << 32)) & mask;
+    out[25] = (w17 >> 12) & mask;
+    out[26] = ((w17 >> 56) | (w18 << 8)) & mask;
+    out[27] = ((w18 >> 36) | (w19 << 28)) & mask;
+    out[28] = (w19 >> 16) & mask;
+    out[29] = ((w19 >> 60) | (w20 << 4)) & mask;
+    out[30] = ((w20 >> 40) | (w21 << 24)) & mask;
+    out[31] = w21 >> 20;
+
+    return in + (22 * 8);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 45> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 45) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    const auto w8 = LoadInt<uint64_t>(in + 8 * 8);
+    const auto w9 = LoadInt<uint64_t>(in + 9 * 8);
+    const auto w10 = LoadInt<uint64_t>(in + 10 * 8);
+    const auto w11 = LoadInt<uint64_t>(in + 11 * 8);
+    const auto w12 = LoadInt<uint64_t>(in + 12 * 8);
+    const auto w13 = LoadInt<uint64_t>(in + 13 * 8);
+    const auto w14 = LoadInt<uint64_t>(in + 14 * 8);
+    const auto w15 = LoadInt<uint64_t>(in + 15 * 8);
+    const auto w16 = LoadInt<uint64_t>(in + 16 * 8);
+    const auto w17 = LoadInt<uint64_t>(in + 17 * 8);
+    const auto w18 = LoadInt<uint64_t>(in + 18 * 8);
+    const auto w19 = LoadInt<uint64_t>(in + 19 * 8);
+    const auto w20 = LoadInt<uint64_t>(in + 20 * 8);
+    const auto w21 = LoadInt<uint64_t>(in + 21 * 8);
+    const auto w22 = static_cast<uint64_t>(LoadInt<uint32_t>(in + 22 * 8));
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 45) | (w1 << 19)) & mask;
+    out[2] = ((w1 >> 26) | (w2 << 38)) & mask;
+    out[3] = (w2 >> 7) & mask;
+    out[4] = ((w2 >> 52) | (w3 << 12)) & mask;
+    out[5] = ((w3 >> 33) | (w4 << 31)) & mask;
+    out[6] = (w4 >> 14) & mask;
+    out[7] = ((w4 >> 59) | (w5 << 5)) & mask;
+    out[8] = ((w5 >> 40) | (w6 << 24)) & mask;
+    out[9] = ((w6 >> 21) | (w7 << 43)) & mask;
+    out[10] = (w7 >> 2) & mask;
+    out[11] = ((w7 >> 47) | (w8 << 17)) & mask;
+    out[12] = ((w8 >> 28) | (w9 << 36)) & mask;
+    out[13] = (w9 >> 9) & mask;
+    out[14] = ((w9 >> 54) | (w10 << 10)) & mask;
+    out[15] = ((w10 >> 35) | (w11 << 29)) & mask;
+    out[16] = (w11 >> 16) & mask;
+    out[17] = ((w11 >> 61) | (w12 << 3)) & mask;
+    out[18] = ((w12 >> 42) | (w13 << 22)) & mask;
+    out[19] = ((w13 >> 23) | (w14 << 41)) & mask;
+    out[20] = (w14 >> 4) & mask;
+    out[21] = ((w14 >> 49) | (w15 << 15)) & mask;
+    out[22] = ((w15 >> 30) | (w16 << 34)) & mask;
+    out[23] = (w16 >> 11) & mask;
+    out[24] = ((w16 >> 56) | (w17 << 8)) & mask;
+    out[25] = ((w17 >> 37) | (w18 << 27)) & mask;
+    out[26] = (w18 >> 18) & mask;
+    out[27] = ((w18 >> 63) | (w19 << 1)) & mask;
+    out[28] = ((w19 >> 44) | (w20 << 20)) & mask;
+    out[29] = ((w20 >> 25) | (w21 << 39)) & mask;
+    out[30] = (w21 >> 6) & mask;
+    out[31] = ((w21 >> 51) | (w22 << 13)) & mask;
+
+    return in + (22 * 8 + 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 46> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 46) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    const auto w8 = LoadInt<uint64_t>(in + 8 * 8);
+    const auto w9 = LoadInt<uint64_t>(in + 9 * 8);
+    const auto w10 = LoadInt<uint64_t>(in + 10 * 8);
+    const auto w11 = LoadInt<uint64_t>(in + 11 * 8);
+    const auto w12 = LoadInt<uint64_t>(in + 12 * 8);
+    const auto w13 = LoadInt<uint64_t>(in + 13 * 8);
+    const auto w14 = LoadInt<uint64_t>(in + 14 * 8);
+    const auto w15 = LoadInt<uint64_t>(in + 15 * 8);
+    const auto w16 = LoadInt<uint64_t>(in + 16 * 8);
+    const auto w17 = LoadInt<uint64_t>(in + 17 * 8);
+    const auto w18 = LoadInt<uint64_t>(in + 18 * 8);
+    const auto w19 = LoadInt<uint64_t>(in + 19 * 8);
+    const auto w20 = LoadInt<uint64_t>(in + 20 * 8);
+    const auto w21 = LoadInt<uint64_t>(in + 21 * 8);
+    const auto w22 = LoadInt<uint64_t>(in + 22 * 8);
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 46) | (w1 << 18)) & mask;
+    out[2] = ((w1 >> 28) | (w2 << 36)) & mask;
+    out[3] = (w2 >> 10) & mask;
+    out[4] = ((w2 >> 56) | (w3 << 8)) & mask;
+    out[5] = ((w3 >> 38) | (w4 << 26)) & mask;
+    out[6] = ((w4 >> 20) | (w5 << 44)) & mask;
+    out[7] = (w5 >> 2) & mask;
+    out[8] = ((w5 >> 48) | (w6 << 16)) & mask;
+    out[9] = ((w6 >> 30) | (w7 << 34)) & mask;
+    out[10] = (w7 >> 12) & mask;
+    out[11] = ((w7 >> 58) | (w8 << 6)) & mask;
+    out[12] = ((w8 >> 40) | (w9 << 24)) & mask;
+    out[13] = ((w9 >> 22) | (w10 << 42)) & mask;
+    out[14] = (w10 >> 4) & mask;
+    out[15] = ((w10 >> 50) | (w11 << 14)) & mask;
+    out[16] = ((w11 >> 32) | (w12 << 32)) & mask;
+    out[17] = (w12 >> 14) & mask;
+    out[18] = ((w12 >> 60) | (w13 << 4)) & mask;
+    out[19] = ((w13 >> 42) | (w14 << 22)) & mask;
+    out[20] = ((w14 >> 24) | (w15 << 40)) & mask;
+    out[21] = (w15 >> 6) & mask;
+    out[22] = ((w15 >> 52) | (w16 << 12)) & mask;
+    out[23] = ((w16 >> 34) | (w17 << 30)) & mask;
+    out[24] = (w17 >> 16) & mask;
+    out[25] = ((w17 >> 62) | (w18 << 2)) & mask;
+    out[26] = ((w18 >> 44) | (w19 << 20)) & mask;
+    out[27] = ((w19 >> 26) | (w20 << 38)) & mask;
+    out[28] = (w20 >> 8) & mask;
+    out[29] = ((w20 >> 54) | (w21 << 10)) & mask;
+    out[30] = ((w21 >> 36) | (w22 << 28)) & mask;
+    out[31] = w22 >> 18;
+
+    return in + (23 * 8);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 47> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 47) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    const auto w8 = LoadInt<uint64_t>(in + 8 * 8);
+    const auto w9 = LoadInt<uint64_t>(in + 9 * 8);
+    const auto w10 = LoadInt<uint64_t>(in + 10 * 8);
+    const auto w11 = LoadInt<uint64_t>(in + 11 * 8);
+    const auto w12 = LoadInt<uint64_t>(in + 12 * 8);
+    const auto w13 = LoadInt<uint64_t>(in + 13 * 8);
+    const auto w14 = LoadInt<uint64_t>(in + 14 * 8);
+    const auto w15 = LoadInt<uint64_t>(in + 15 * 8);
+    const auto w16 = LoadInt<uint64_t>(in + 16 * 8);
+    const auto w17 = LoadInt<uint64_t>(in + 17 * 8);
+    const auto w18 = LoadInt<uint64_t>(in + 18 * 8);
+    const auto w19 = LoadInt<uint64_t>(in + 19 * 8);
+    const auto w20 = LoadInt<uint64_t>(in + 20 * 8);
+    const auto w21 = LoadInt<uint64_t>(in + 21 * 8);
+    const auto w22 = LoadInt<uint64_t>(in + 22 * 8);
+    const auto w23 = static_cast<uint64_t>(LoadInt<uint32_t>(in + 23 * 8));
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 47) | (w1 << 17)) & mask;
+    out[2] = ((w1 >> 30) | (w2 << 34)) & mask;
+    out[3] = (w2 >> 13) & mask;
+    out[4] = ((w2 >> 60) | (w3 << 4)) & mask;
+    out[5] = ((w3 >> 43) | (w4 << 21)) & mask;
+    out[6] = ((w4 >> 26) | (w5 << 38)) & mask;
+    out[7] = (w5 >> 9) & mask;
+    out[8] = ((w5 >> 56) | (w6 << 8)) & mask;
+    out[9] = ((w6 >> 39) | (w7 << 25)) & mask;
+    out[10] = ((w7 >> 22) | (w8 << 42)) & mask;
+    out[11] = (w8 >> 5) & mask;
+    out[12] = ((w8 >> 52) | (w9 << 12)) & mask;
+    out[13] = ((w9 >> 35) | (w10 << 29)) & mask;
+    out[14] = ((w10 >> 18) | (w11 << 46)) & mask;
+    out[15] = (w11 >> 1) & mask;
+    out[16] = ((w11 >> 48) | (w12 << 16)) & mask;
+    out[17] = ((w12 >> 31) | (w13 << 33)) & mask;
+    out[18] = (w13 >> 14) & mask;
+    out[19] = ((w13 >> 61) | (w14 << 3)) & mask;
+    out[20] = ((w14 >> 44) | (w15 << 20)) & mask;
+    out[21] = ((w15 >> 27) | (w16 << 37)) & mask;
+    out[22] = (w16 >> 10) & mask;
+    out[23] = ((w16 >> 57) | (w17 << 7)) & mask;
+    out[24] = ((w17 >> 40) | (w18 << 24)) & mask;
+    out[25] = ((w18 >> 23) | (w19 << 41)) & mask;
+    out[26] = (w19 >> 6) & mask;
+    out[27] = ((w19 >> 53) | (w20 << 11)) & mask;
+    out[28] = ((w20 >> 36) | (w21 << 28)) & mask;
+    out[29] = ((w21 >> 19) | (w22 << 45)) & mask;
+    out[30] = (w22 >> 2) & mask;
+    out[31] = ((w22 >> 49) | (w23 << 15)) & mask;
+
+    return in + (23 * 8 + 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 48> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 48) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    const auto w8 = LoadInt<uint64_t>(in + 8 * 8);
+    const auto w9 = LoadInt<uint64_t>(in + 9 * 8);
+    const auto w10 = LoadInt<uint64_t>(in + 10 * 8);
+    const auto w11 = LoadInt<uint64_t>(in + 11 * 8);
+    const auto w12 = LoadInt<uint64_t>(in + 12 * 8);
+    const auto w13 = LoadInt<uint64_t>(in + 13 * 8);
+    const auto w14 = LoadInt<uint64_t>(in + 14 * 8);
+    const auto w15 = LoadInt<uint64_t>(in + 15 * 8);
+    const auto w16 = LoadInt<uint64_t>(in + 16 * 8);
+    const auto w17 = LoadInt<uint64_t>(in + 17 * 8);
+    const auto w18 = LoadInt<uint64_t>(in + 18 * 8);
+    const auto w19 = LoadInt<uint64_t>(in + 19 * 8);
+    const auto w20 = LoadInt<uint64_t>(in + 20 * 8);
+    const auto w21 = LoadInt<uint64_t>(in + 21 * 8);
+    const auto w22 = LoadInt<uint64_t>(in + 22 * 8);
+    const auto w23 = LoadInt<uint64_t>(in + 23 * 8);
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 48) | (w1 << 16)) & mask;
+    out[2] = ((w1 >> 32) | (w2 << 32)) & mask;
+    out[3] = w2 >> 16;
+    out[4] = (w3) & mask;
+    out[5] = ((w3 >> 48) | (w4 << 16)) & mask;
+    out[6] = ((w4 >> 32) | (w5 << 32)) & mask;
+    out[7] = w5 >> 16;
+    out[8] = (w6) & mask;
+    out[9] = ((w6 >> 48) | (w7 << 16)) & mask;
+    out[10] = ((w7 >> 32) | (w8 << 32)) & mask;
+    out[11] = w8 >> 16;
+    out[12] = (w9) & mask;
+    out[13] = ((w9 >> 48) | (w10 << 16)) & mask;
+    out[14] = ((w10 >> 32) | (w11 << 32)) & mask;
+    out[15] = w11 >> 16;
+    out[16] = (w12) & mask;
+    out[17] = ((w12 >> 48) | (w13 << 16)) & mask;
+    out[18] = ((w13 >> 32) | (w14 << 32)) & mask;
+    out[19] = w14 >> 16;
+    out[20] = (w15) & mask;
+    out[21] = ((w15 >> 48) | (w16 << 16)) & mask;
+    out[22] = ((w16 >> 32) | (w17 << 32)) & mask;
+    out[23] = w17 >> 16;
+    out[24] = (w18) & mask;
+    out[25] = ((w18 >> 48) | (w19 << 16)) & mask;
+    out[26] = ((w19 >> 32) | (w20 << 32)) & mask;
+    out[27] = w20 >> 16;
+    out[28] = (w21) & mask;
+    out[29] = ((w21 >> 48) | (w22 << 16)) & mask;
+    out[30] = ((w22 >> 32) | (w23 << 32)) & mask;
+    out[31] = w23 >> 16;
+
+    return in + (24 * 8);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 49> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 49) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    const auto w8 = LoadInt<uint64_t>(in + 8 * 8);
+    const auto w9 = LoadInt<uint64_t>(in + 9 * 8);
+    const auto w10 = LoadInt<uint64_t>(in + 10 * 8);
+    const auto w11 = LoadInt<uint64_t>(in + 11 * 8);
+    const auto w12 = LoadInt<uint64_t>(in + 12 * 8);
+    const auto w13 = LoadInt<uint64_t>(in + 13 * 8);
+    const auto w14 = LoadInt<uint64_t>(in + 14 * 8);
+    const auto w15 = LoadInt<uint64_t>(in + 15 * 8);
+    const auto w16 = LoadInt<uint64_t>(in + 16 * 8);
+    const auto w17 = LoadInt<uint64_t>(in + 17 * 8);
+    const auto w18 = LoadInt<uint64_t>(in + 18 * 8);
+    const auto w19 = LoadInt<uint64_t>(in + 19 * 8);
+    const auto w20 = LoadInt<uint64_t>(in + 20 * 8);
+    const auto w21 = LoadInt<uint64_t>(in + 21 * 8);
+    const auto w22 = LoadInt<uint64_t>(in + 22 * 8);
+    const auto w23 = LoadInt<uint64_t>(in + 23 * 8);
+    const auto w24 = static_cast<uint64_t>(LoadInt<uint32_t>(in + 24 * 8));
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 49) | (w1 << 15)) & mask;
+    out[2] = ((w1 >> 34) | (w2 << 30)) & mask;
+    out[3] = ((w2 >> 19) | (w3 << 45)) & mask;
+    out[4] = (w3 >> 4) & mask;
+    out[5] = ((w3 >> 53) | (w4 << 11)) & mask;
+    out[6] = ((w4 >> 38) | (w5 << 26)) & mask;
+    out[7] = ((w5 >> 23) | (w6 << 41)) & mask;
+    out[8] = (w6 >> 8) & mask;
+    out[9] = ((w6 >> 57) | (w7 << 7)) & mask;
+    out[10] = ((w7 >> 42) | (w8 << 22)) & mask;
+    out[11] = ((w8 >> 27) | (w9 << 37)) & mask;
+    out[12] = (w9 >> 12) & mask;
+    out[13] = ((w9 >> 61) | (w10 << 3)) & mask;
+    out[14] = ((w10 >> 46) | (w11 << 18)) & mask;
+    out[15] = ((w11 >> 31) | (w12 << 33)) & mask;
+    out[16] = ((w12 >> 16) | (w13 << 48)) & mask;
+    out[17] = (w13 >> 1) & mask;
+    out[18] = ((w13 >> 50) | (w14 << 14)) & mask;
+    out[19] = ((w14 >> 35) | (w15 << 29)) & mask;
+    out[20] = ((w15 >> 20) | (w16 << 44)) & mask;
+    out[21] = (w16 >> 5) & mask;
+    out[22] = ((w16 >> 54) | (w17 << 10)) & mask;
+    out[23] = ((w17 >> 39) | (w18 << 25)) & mask;
+    out[24] = ((w18 >> 24) | (w19 << 40)) & mask;
+    out[25] = (w19 >> 9) & mask;
+    out[26] = ((w19 >> 58) | (w20 << 6)) & mask;
+    out[27] = ((w20 >> 43) | (w21 << 21)) & mask;
+    out[28] = ((w21 >> 28) | (w22 << 36)) & mask;
+    out[29] = (w22 >> 13) & mask;
+    out[30] = ((w22 >> 62) | (w23 << 2)) & mask;
+    out[31] = ((w23 >> 47) | (w24 << 17)) & mask;
+
+    return in + (24 * 8 + 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 50> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 50) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    const auto w8 = LoadInt<uint64_t>(in + 8 * 8);
+    const auto w9 = LoadInt<uint64_t>(in + 9 * 8);
+    const auto w10 = LoadInt<uint64_t>(in + 10 * 8);
+    const auto w11 = LoadInt<uint64_t>(in + 11 * 8);
+    const auto w12 = LoadInt<uint64_t>(in + 12 * 8);
+    const auto w13 = LoadInt<uint64_t>(in + 13 * 8);
+    const auto w14 = LoadInt<uint64_t>(in + 14 * 8);
+    const auto w15 = LoadInt<uint64_t>(in + 15 * 8);
+    const auto w16 = LoadInt<uint64_t>(in + 16 * 8);
+    const auto w17 = LoadInt<uint64_t>(in + 17 * 8);
+    const auto w18 = LoadInt<uint64_t>(in + 18 * 8);
+    const auto w19 = LoadInt<uint64_t>(in + 19 * 8);
+    const auto w20 = LoadInt<uint64_t>(in + 20 * 8);
+    const auto w21 = LoadInt<uint64_t>(in + 21 * 8);
+    const auto w22 = LoadInt<uint64_t>(in + 22 * 8);
+    const auto w23 = LoadInt<uint64_t>(in + 23 * 8);
+    const auto w24 = LoadInt<uint64_t>(in + 24 * 8);
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 50) | (w1 << 14)) & mask;
+    out[2] = ((w1 >> 36) | (w2 << 28)) & mask;
+    out[3] = ((w2 >> 22) | (w3 << 42)) & mask;
+    out[4] = (w3 >> 8) & mask;
+    out[5] = ((w3 >> 58) | (w4 << 6)) & mask;
+    out[6] = ((w4 >> 44) | (w5 << 20)) & mask;
+    out[7] = ((w5 >> 30) | (w6 << 34)) & mask;
+    out[8] = ((w6 >> 16) | (w7 << 48)) & mask;
+    out[9] = (w7 >> 2) & mask;
+    out[10] = ((w7 >> 52) | (w8 << 12)) & mask;
+    out[11] = ((w8 >> 38) | (w9 << 26)) & mask;
+    out[12] = ((w9 >> 24) | (w10 << 40)) & mask;
+    out[13] = (w10 >> 10) & mask;
+    out[14] = ((w10 >> 60) | (w11 << 4)) & mask;
+    out[15] = ((w11 >> 46) | (w12 << 18)) & mask;
+    out[16] = ((w12 >> 32) | (w13 << 32)) & mask;
+    out[17] = ((w13 >> 18) | (w14 << 46)) & mask;
+    out[18] = (w14 >> 4) & mask;
+    out[19] = ((w14 >> 54) | (w15 << 10)) & mask;
+    out[20] = ((w15 >> 40) | (w16 << 24)) & mask;
+    out[21] = ((w16 >> 26) | (w17 << 38)) & mask;
+    out[22] = (w17 >> 12) & mask;
+    out[23] = ((w17 >> 62) | (w18 << 2)) & mask;
+    out[24] = ((w18 >> 48) | (w19 << 16)) & mask;
+    out[25] = ((w19 >> 34) | (w20 << 30)) & mask;
+    out[26] = ((w20 >> 20) | (w21 << 44)) & mask;
+    out[27] = (w21 >> 6) & mask;
+    out[28] = ((w21 >> 56) | (w22 << 8)) & mask;
+    out[29] = ((w22 >> 42) | (w23 << 22)) & mask;
+    out[30] = ((w23 >> 28) | (w24 << 36)) & mask;
+    out[31] = w24 >> 14;
+
+    return in + (25 * 8);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 51> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 51) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    const auto w8 = LoadInt<uint64_t>(in + 8 * 8);
+    const auto w9 = LoadInt<uint64_t>(in + 9 * 8);
+    const auto w10 = LoadInt<uint64_t>(in + 10 * 8);
+    const auto w11 = LoadInt<uint64_t>(in + 11 * 8);
+    const auto w12 = LoadInt<uint64_t>(in + 12 * 8);
+    const auto w13 = LoadInt<uint64_t>(in + 13 * 8);
+    const auto w14 = LoadInt<uint64_t>(in + 14 * 8);
+    const auto w15 = LoadInt<uint64_t>(in + 15 * 8);
+    const auto w16 = LoadInt<uint64_t>(in + 16 * 8);
+    const auto w17 = LoadInt<uint64_t>(in + 17 * 8);
+    const auto w18 = LoadInt<uint64_t>(in + 18 * 8);
+    const auto w19 = LoadInt<uint64_t>(in + 19 * 8);
+    const auto w20 = LoadInt<uint64_t>(in + 20 * 8);
+    const auto w21 = LoadInt<uint64_t>(in + 21 * 8);
+    const auto w22 = LoadInt<uint64_t>(in + 22 * 8);
+    const auto w23 = LoadInt<uint64_t>(in + 23 * 8);
+    const auto w24 = LoadInt<uint64_t>(in + 24 * 8);
+    const auto w25 = static_cast<uint64_t>(LoadInt<uint32_t>(in + 25 * 8));
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 51) | (w1 << 13)) & mask;
+    out[2] = ((w1 >> 38) | (w2 << 26)) & mask;
+    out[3] = ((w2 >> 25) | (w3 << 39)) & mask;
+    out[4] = (w3 >> 12) & mask;
+    out[5] = ((w3 >> 63) | (w4 << 1)) & mask;
+    out[6] = ((w4 >> 50) | (w5 << 14)) & mask;
+    out[7] = ((w5 >> 37) | (w6 << 27)) & mask;
+    out[8] = ((w6 >> 24) | (w7 << 40)) & mask;
+    out[9] = (w7 >> 11) & mask;
+    out[10] = ((w7 >> 62) | (w8 << 2)) & mask;
+    out[11] = ((w8 >> 49) | (w9 << 15)) & mask;
+    out[12] = ((w9 >> 36) | (w10 << 28)) & mask;
+    out[13] = ((w10 >> 23) | (w11 << 41)) & mask;
+    out[14] = (w11 >> 10) & mask;
+    out[15] = ((w11 >> 61) | (w12 << 3)) & mask;
+    out[16] = ((w12 >> 48) | (w13 << 16)) & mask;
+    out[17] = ((w13 >> 35) | (w14 << 29)) & mask;
+    out[18] = ((w14 >> 22) | (w15 << 42)) & mask;
+    out[19] = (w15 >> 9) & mask;
+    out[20] = ((w15 >> 60) | (w16 << 4)) & mask;
+    out[21] = ((w16 >> 47) | (w17 << 17)) & mask;
+    out[22] = ((w17 >> 34) | (w18 << 30)) & mask;
+    out[23] = ((w18 >> 21) | (w19 << 43)) & mask;
+    out[24] = (w19 >> 8) & mask;
+    out[25] = ((w19 >> 59) | (w20 << 5)) & mask;
+    out[26] = ((w20 >> 46) | (w21 << 18)) & mask;
+    out[27] = ((w21 >> 33) | (w22 << 31)) & mask;
+    out[28] = ((w22 >> 20) | (w23 << 44)) & mask;
+    out[29] = (w23 >> 7) & mask;
+    out[30] = ((w23 >> 58) | (w24 << 6)) & mask;
+    out[31] = ((w24 >> 45) | (w25 << 19)) & mask;
+
+    return in + (25 * 8 + 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 52> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 52) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    const auto w8 = LoadInt<uint64_t>(in + 8 * 8);
+    const auto w9 = LoadInt<uint64_t>(in + 9 * 8);
+    const auto w10 = LoadInt<uint64_t>(in + 10 * 8);
+    const auto w11 = LoadInt<uint64_t>(in + 11 * 8);
+    const auto w12 = LoadInt<uint64_t>(in + 12 * 8);
+    const auto w13 = LoadInt<uint64_t>(in + 13 * 8);
+    const auto w14 = LoadInt<uint64_t>(in + 14 * 8);
+    const auto w15 = LoadInt<uint64_t>(in + 15 * 8);
+    const auto w16 = LoadInt<uint64_t>(in + 16 * 8);
+    const auto w17 = LoadInt<uint64_t>(in + 17 * 8);
+    const auto w18 = LoadInt<uint64_t>(in + 18 * 8);
+    const auto w19 = LoadInt<uint64_t>(in + 19 * 8);
+    const auto w20 = LoadInt<uint64_t>(in + 20 * 8);
+    const auto w21 = LoadInt<uint64_t>(in + 21 * 8);
+    const auto w22 = LoadInt<uint64_t>(in + 22 * 8);
+    const auto w23 = LoadInt<uint64_t>(in + 23 * 8);
+    const auto w24 = LoadInt<uint64_t>(in + 24 * 8);
+    const auto w25 = LoadInt<uint64_t>(in + 25 * 8);
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 52) | (w1 << 12)) & mask;
+    out[2] = ((w1 >> 40) | (w2 << 24)) & mask;
+    out[3] = ((w2 >> 28) | (w3 << 36)) & mask;
+    out[4] = ((w3 >> 16) | (w4 << 48)) & mask;
+    out[5] = (w4 >> 4) & mask;
+    out[6] = ((w4 >> 56) | (w5 << 8)) & mask;
+    out[7] = ((w5 >> 44) | (w6 << 20)) & mask;
+    out[8] = ((w6 >> 32) | (w7 << 32)) & mask;
+    out[9] = ((w7 >> 20) | (w8 << 44)) & mask;
+    out[10] = (w8 >> 8) & mask;
+    out[11] = ((w8 >> 60) | (w9 << 4)) & mask;
+    out[12] = ((w9 >> 48) | (w10 << 16)) & mask;
+    out[13] = ((w10 >> 36) | (w11 << 28)) & mask;
+    out[14] = ((w11 >> 24) | (w12 << 40)) & mask;
+    out[15] = w12 >> 12;
+    out[16] = (w13) & mask;
+    out[17] = ((w13 >> 52) | (w14 << 12)) & mask;
+    out[18] = ((w14 >> 40) | (w15 << 24)) & mask;
+    out[19] = ((w15 >> 28) | (w16 << 36)) & mask;
+    out[20] = ((w16 >> 16) | (w17 << 48)) & mask;
+    out[21] = (w17 >> 4) & mask;
+    out[22] = ((w17 >> 56) | (w18 << 8)) & mask;
+    out[23] = ((w18 >> 44) | (w19 << 20)) & mask;
+    out[24] = ((w19 >> 32) | (w20 << 32)) & mask;
+    out[25] = ((w20 >> 20) | (w21 << 44)) & mask;
+    out[26] = (w21 >> 8) & mask;
+    out[27] = ((w21 >> 60) | (w22 << 4)) & mask;
+    out[28] = ((w22 >> 48) | (w23 << 16)) & mask;
+    out[29] = ((w23 >> 36) | (w24 << 28)) & mask;
+    out[30] = ((w24 >> 24) | (w25 << 40)) & mask;
+    out[31] = w25 >> 12;
+
+    return in + (26 * 8);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 53> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 53) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    const auto w8 = LoadInt<uint64_t>(in + 8 * 8);
+    const auto w9 = LoadInt<uint64_t>(in + 9 * 8);
+    const auto w10 = LoadInt<uint64_t>(in + 10 * 8);
+    const auto w11 = LoadInt<uint64_t>(in + 11 * 8);
+    const auto w12 = LoadInt<uint64_t>(in + 12 * 8);
+    const auto w13 = LoadInt<uint64_t>(in + 13 * 8);
+    const auto w14 = LoadInt<uint64_t>(in + 14 * 8);
+    const auto w15 = LoadInt<uint64_t>(in + 15 * 8);
+    const auto w16 = LoadInt<uint64_t>(in + 16 * 8);
+    const auto w17 = LoadInt<uint64_t>(in + 17 * 8);
+    const auto w18 = LoadInt<uint64_t>(in + 18 * 8);
+    const auto w19 = LoadInt<uint64_t>(in + 19 * 8);
+    const auto w20 = LoadInt<uint64_t>(in + 20 * 8);
+    const auto w21 = LoadInt<uint64_t>(in + 21 * 8);
+    const auto w22 = LoadInt<uint64_t>(in + 22 * 8);
+    const auto w23 = LoadInt<uint64_t>(in + 23 * 8);
+    const auto w24 = LoadInt<uint64_t>(in + 24 * 8);
+    const auto w25 = LoadInt<uint64_t>(in + 25 * 8);
+    const auto w26 = static_cast<uint64_t>(LoadInt<uint32_t>(in + 26 * 8));
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 53) | (w1 << 11)) & mask;
+    out[2] = ((w1 >> 42) | (w2 << 22)) & mask;
+    out[3] = ((w2 >> 31) | (w3 << 33)) & mask;
+    out[4] = ((w3 >> 20) | (w4 << 44)) & mask;
+    out[5] = (w4 >> 9) & mask;
+    out[6] = ((w4 >> 62) | (w5 << 2)) & mask;
+    out[7] = ((w5 >> 51) | (w6 << 13)) & mask;
+    out[8] = ((w6 >> 40) | (w7 << 24)) & mask;
+    out[9] = ((w7 >> 29) | (w8 << 35)) & mask;
+    out[10] = ((w8 >> 18) | (w9 << 46)) & mask;
+    out[11] = (w9 >> 7) & mask;
+    out[12] = ((w9 >> 60) | (w10 << 4)) & mask;
+    out[13] = ((w10 >> 49) | (w11 << 15)) & mask;
+    out[14] = ((w11 >> 38) | (w12 << 26)) & mask;
+    out[15] = ((w12 >> 27) | (w13 << 37)) & mask;
+    out[16] = ((w13 >> 16) | (w14 << 48)) & mask;
+    out[17] = (w14 >> 5) & mask;
+    out[18] = ((w14 >> 58) | (w15 << 6)) & mask;
+    out[19] = ((w15 >> 47) | (w16 << 17)) & mask;
+    out[20] = ((w16 >> 36) | (w17 << 28)) & mask;
+    out[21] = ((w17 >> 25) | (w18 << 39)) & mask;
+    out[22] = ((w18 >> 14) | (w19 << 50)) & mask;
+    out[23] = (w19 >> 3) & mask;
+    out[24] = ((w19 >> 56) | (w20 << 8)) & mask;
+    out[25] = ((w20 >> 45) | (w21 << 19)) & mask;
+    out[26] = ((w21 >> 34) | (w22 << 30)) & mask;
+    out[27] = ((w22 >> 23) | (w23 << 41)) & mask;
+    out[28] = ((w23 >> 12) | (w24 << 52)) & mask;
+    out[29] = (w24 >> 1) & mask;
+    out[30] = ((w24 >> 54) | (w25 << 10)) & mask;
+    out[31] = ((w25 >> 43) | (w26 << 21)) & mask;
+
+    return in + (26 * 8 + 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 54> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 54) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    const auto w8 = LoadInt<uint64_t>(in + 8 * 8);
+    const auto w9 = LoadInt<uint64_t>(in + 9 * 8);
+    const auto w10 = LoadInt<uint64_t>(in + 10 * 8);
+    const auto w11 = LoadInt<uint64_t>(in + 11 * 8);
+    const auto w12 = LoadInt<uint64_t>(in + 12 * 8);
+    const auto w13 = LoadInt<uint64_t>(in + 13 * 8);
+    const auto w14 = LoadInt<uint64_t>(in + 14 * 8);
+    const auto w15 = LoadInt<uint64_t>(in + 15 * 8);
+    const auto w16 = LoadInt<uint64_t>(in + 16 * 8);
+    const auto w17 = LoadInt<uint64_t>(in + 17 * 8);
+    const auto w18 = LoadInt<uint64_t>(in + 18 * 8);
+    const auto w19 = LoadInt<uint64_t>(in + 19 * 8);
+    const auto w20 = LoadInt<uint64_t>(in + 20 * 8);
+    const auto w21 = LoadInt<uint64_t>(in + 21 * 8);
+    const auto w22 = LoadInt<uint64_t>(in + 22 * 8);
+    const auto w23 = LoadInt<uint64_t>(in + 23 * 8);
+    const auto w24 = LoadInt<uint64_t>(in + 24 * 8);
+    const auto w25 = LoadInt<uint64_t>(in + 25 * 8);
+    const auto w26 = LoadInt<uint64_t>(in + 26 * 8);
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 54) | (w1 << 10)) & mask;
+    out[2] = ((w1 >> 44) | (w2 << 20)) & mask;
+    out[3] = ((w2 >> 34) | (w3 << 30)) & mask;
+    out[4] = ((w3 >> 24) | (w4 << 40)) & mask;
+    out[5] = ((w4 >> 14) | (w5 << 50)) & mask;
+    out[6] = (w5 >> 4) & mask;
+    out[7] = ((w5 >> 58) | (w6 << 6)) & mask;
+    out[8] = ((w6 >> 48) | (w7 << 16)) & mask;
+    out[9] = ((w7 >> 38) | (w8 << 26)) & mask;
+    out[10] = ((w8 >> 28) | (w9 << 36)) & mask;
+    out[11] = ((w9 >> 18) | (w10 << 46)) & mask;
+    out[12] = (w10 >> 8) & mask;
+    out[13] = ((w10 >> 62) | (w11 << 2)) & mask;
+    out[14] = ((w11 >> 52) | (w12 << 12)) & mask;
+    out[15] = ((w12 >> 42) | (w13 << 22)) & mask;
+    out[16] = ((w13 >> 32) | (w14 << 32)) & mask;
+    out[17] = ((w14 >> 22) | (w15 << 42)) & mask;
+    out[18] = ((w15 >> 12) | (w16 << 52)) & mask;
+    out[19] = (w16 >> 2) & mask;
+    out[20] = ((w16 >> 56) | (w17 << 8)) & mask;
+    out[21] = ((w17 >> 46) | (w18 << 18)) & mask;
+    out[22] = ((w18 >> 36) | (w19 << 28)) & mask;
+    out[23] = ((w19 >> 26) | (w20 << 38)) & mask;
+    out[24] = ((w20 >> 16) | (w21 << 48)) & mask;
+    out[25] = (w21 >> 6) & mask;
+    out[26] = ((w21 >> 60) | (w22 << 4)) & mask;
+    out[27] = ((w22 >> 50) | (w23 << 14)) & mask;
+    out[28] = ((w23 >> 40) | (w24 << 24)) & mask;
+    out[29] = ((w24 >> 30) | (w25 << 34)) & mask;
+    out[30] = ((w25 >> 20) | (w26 << 44)) & mask;
+    out[31] = w26 >> 10;
+
+    return in + (27 * 8);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 55> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 55) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    const auto w8 = LoadInt<uint64_t>(in + 8 * 8);
+    const auto w9 = LoadInt<uint64_t>(in + 9 * 8);
+    const auto w10 = LoadInt<uint64_t>(in + 10 * 8);
+    const auto w11 = LoadInt<uint64_t>(in + 11 * 8);
+    const auto w12 = LoadInt<uint64_t>(in + 12 * 8);
+    const auto w13 = LoadInt<uint64_t>(in + 13 * 8);
+    const auto w14 = LoadInt<uint64_t>(in + 14 * 8);
+    const auto w15 = LoadInt<uint64_t>(in + 15 * 8);
+    const auto w16 = LoadInt<uint64_t>(in + 16 * 8);
+    const auto w17 = LoadInt<uint64_t>(in + 17 * 8);
+    const auto w18 = LoadInt<uint64_t>(in + 18 * 8);
+    const auto w19 = LoadInt<uint64_t>(in + 19 * 8);
+    const auto w20 = LoadInt<uint64_t>(in + 20 * 8);
+    const auto w21 = LoadInt<uint64_t>(in + 21 * 8);
+    const auto w22 = LoadInt<uint64_t>(in + 22 * 8);
+    const auto w23 = LoadInt<uint64_t>(in + 23 * 8);
+    const auto w24 = LoadInt<uint64_t>(in + 24 * 8);
+    const auto w25 = LoadInt<uint64_t>(in + 25 * 8);
+    const auto w26 = LoadInt<uint64_t>(in + 26 * 8);
+    const auto w27 = static_cast<uint64_t>(LoadInt<uint32_t>(in + 27 * 8));
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 55) | (w1 << 9)) & mask;
+    out[2] = ((w1 >> 46) | (w2 << 18)) & mask;
+    out[3] = ((w2 >> 37) | (w3 << 27)) & mask;
+    out[4] = ((w3 >> 28) | (w4 << 36)) & mask;
+    out[5] = ((w4 >> 19) | (w5 << 45)) & mask;
+    out[6] = ((w5 >> 10) | (w6 << 54)) & mask;
+    out[7] = (w6 >> 1) & mask;
+    out[8] = ((w6 >> 56) | (w7 << 8)) & mask;
+    out[9] = ((w7 >> 47) | (w8 << 17)) & mask;
+    out[10] = ((w8 >> 38) | (w9 << 26)) & mask;
+    out[11] = ((w9 >> 29) | (w10 << 35)) & mask;
+    out[12] = ((w10 >> 20) | (w11 << 44)) & mask;
+    out[13] = ((w11 >> 11) | (w12 << 53)) & mask;
+    out[14] = (w12 >> 2) & mask;
+    out[15] = ((w12 >> 57) | (w13 << 7)) & mask;
+    out[16] = ((w13 >> 48) | (w14 << 16)) & mask;
+    out[17] = ((w14 >> 39) | (w15 << 25)) & mask;
+    out[18] = ((w15 >> 30) | (w16 << 34)) & mask;
+    out[19] = ((w16 >> 21) | (w17 << 43)) & mask;
+    out[20] = ((w17 >> 12) | (w18 << 52)) & mask;
+    out[21] = (w18 >> 3) & mask;
+    out[22] = ((w18 >> 58) | (w19 << 6)) & mask;
+    out[23] = ((w19 >> 49) | (w20 << 15)) & mask;
+    out[24] = ((w20 >> 40) | (w21 << 24)) & mask;
+    out[25] = ((w21 >> 31) | (w22 << 33)) & mask;
+    out[26] = ((w22 >> 22) | (w23 << 42)) & mask;
+    out[27] = ((w23 >> 13) | (w24 << 51)) & mask;
+    out[28] = (w24 >> 4) & mask;
+    out[29] = ((w24 >> 59) | (w25 << 5)) & mask;
+    out[30] = ((w25 >> 50) | (w26 << 14)) & mask;
+    out[31] = ((w26 >> 41) | (w27 << 23)) & mask;
+
+    return in + (27 * 8 + 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 56> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 56) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    const auto w8 = LoadInt<uint64_t>(in + 8 * 8);
+    const auto w9 = LoadInt<uint64_t>(in + 9 * 8);
+    const auto w10 = LoadInt<uint64_t>(in + 10 * 8);
+    const auto w11 = LoadInt<uint64_t>(in + 11 * 8);
+    const auto w12 = LoadInt<uint64_t>(in + 12 * 8);
+    const auto w13 = LoadInt<uint64_t>(in + 13 * 8);
+    const auto w14 = LoadInt<uint64_t>(in + 14 * 8);
+    const auto w15 = LoadInt<uint64_t>(in + 15 * 8);
+    const auto w16 = LoadInt<uint64_t>(in + 16 * 8);
+    const auto w17 = LoadInt<uint64_t>(in + 17 * 8);
+    const auto w18 = LoadInt<uint64_t>(in + 18 * 8);
+    const auto w19 = LoadInt<uint64_t>(in + 19 * 8);
+    const auto w20 = LoadInt<uint64_t>(in + 20 * 8);
+    const auto w21 = LoadInt<uint64_t>(in + 21 * 8);
+    const auto w22 = LoadInt<uint64_t>(in + 22 * 8);
+    const auto w23 = LoadInt<uint64_t>(in + 23 * 8);
+    const auto w24 = LoadInt<uint64_t>(in + 24 * 8);
+    const auto w25 = LoadInt<uint64_t>(in + 25 * 8);
+    const auto w26 = LoadInt<uint64_t>(in + 26 * 8);
+    const auto w27 = LoadInt<uint64_t>(in + 27 * 8);
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 56) | (w1 << 8)) & mask;
+    out[2] = ((w1 >> 48) | (w2 << 16)) & mask;
+    out[3] = ((w2 >> 40) | (w3 << 24)) & mask;
+    out[4] = ((w3 >> 32) | (w4 << 32)) & mask;
+    out[5] = ((w4 >> 24) | (w5 << 40)) & mask;
+    out[6] = ((w5 >> 16) | (w6 << 48)) & mask;
+    out[7] = w6 >> 8;
+    out[8] = (w7) & mask;
+    out[9] = ((w7 >> 56) | (w8 << 8)) & mask;
+    out[10] = ((w8 >> 48) | (w9 << 16)) & mask;
+    out[11] = ((w9 >> 40) | (w10 << 24)) & mask;
+    out[12] = ((w10 >> 32) | (w11 << 32)) & mask;
+    out[13] = ((w11 >> 24) | (w12 << 40)) & mask;
+    out[14] = ((w12 >> 16) | (w13 << 48)) & mask;
+    out[15] = w13 >> 8;
+    out[16] = (w14) & mask;
+    out[17] = ((w14 >> 56) | (w15 << 8)) & mask;
+    out[18] = ((w15 >> 48) | (w16 << 16)) & mask;
+    out[19] = ((w16 >> 40) | (w17 << 24)) & mask;
+    out[20] = ((w17 >> 32) | (w18 << 32)) & mask;
+    out[21] = ((w18 >> 24) | (w19 << 40)) & mask;
+    out[22] = ((w19 >> 16) | (w20 << 48)) & mask;
+    out[23] = w20 >> 8;
+    out[24] = (w21) & mask;
+    out[25] = ((w21 >> 56) | (w22 << 8)) & mask;
+    out[26] = ((w22 >> 48) | (w23 << 16)) & mask;
+    out[27] = ((w23 >> 40) | (w24 << 24)) & mask;
+    out[28] = ((w24 >> 32) | (w25 << 32)) & mask;
+    out[29] = ((w25 >> 24) | (w26 << 40)) & mask;
+    out[30] = ((w26 >> 16) | (w27 << 48)) & mask;
+    out[31] = w27 >> 8;
+
+    return in + (28 * 8);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 57> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 57) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    const auto w8 = LoadInt<uint64_t>(in + 8 * 8);
+    const auto w9 = LoadInt<uint64_t>(in + 9 * 8);
+    const auto w10 = LoadInt<uint64_t>(in + 10 * 8);
+    const auto w11 = LoadInt<uint64_t>(in + 11 * 8);
+    const auto w12 = LoadInt<uint64_t>(in + 12 * 8);
+    const auto w13 = LoadInt<uint64_t>(in + 13 * 8);
+    const auto w14 = LoadInt<uint64_t>(in + 14 * 8);
+    const auto w15 = LoadInt<uint64_t>(in + 15 * 8);
+    const auto w16 = LoadInt<uint64_t>(in + 16 * 8);
+    const auto w17 = LoadInt<uint64_t>(in + 17 * 8);
+    const auto w18 = LoadInt<uint64_t>(in + 18 * 8);
+    const auto w19 = LoadInt<uint64_t>(in + 19 * 8);
+    const auto w20 = LoadInt<uint64_t>(in + 20 * 8);
+    const auto w21 = LoadInt<uint64_t>(in + 21 * 8);
+    const auto w22 = LoadInt<uint64_t>(in + 22 * 8);
+    const auto w23 = LoadInt<uint64_t>(in + 23 * 8);
+    const auto w24 = LoadInt<uint64_t>(in + 24 * 8);
+    const auto w25 = LoadInt<uint64_t>(in + 25 * 8);
+    const auto w26 = LoadInt<uint64_t>(in + 26 * 8);
+    const auto w27 = LoadInt<uint64_t>(in + 27 * 8);
+    const auto w28 = static_cast<uint64_t>(LoadInt<uint32_t>(in + 28 * 8));
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 57) | (w1 << 7)) & mask;
+    out[2] = ((w1 >> 50) | (w2 << 14)) & mask;
+    out[3] = ((w2 >> 43) | (w3 << 21)) & mask;
+    out[4] = ((w3 >> 36) | (w4 << 28)) & mask;
+    out[5] = ((w4 >> 29) | (w5 << 35)) & mask;
+    out[6] = ((w5 >> 22) | (w6 << 42)) & mask;
+    out[7] = ((w6 >> 15) | (w7 << 49)) & mask;
+    out[8] = ((w7 >> 8) | (w8 << 56)) & mask;
+    out[9] = (w8 >> 1) & mask;
+    out[10] = ((w8 >> 58) | (w9 << 6)) & mask;
+    out[11] = ((w9 >> 51) | (w10 << 13)) & mask;
+    out[12] = ((w10 >> 44) | (w11 << 20)) & mask;
+    out[13] = ((w11 >> 37) | (w12 << 27)) & mask;
+    out[14] = ((w12 >> 30) | (w13 << 34)) & mask;
+    out[15] = ((w13 >> 23) | (w14 << 41)) & mask;
+    out[16] = ((w14 >> 16) | (w15 << 48)) & mask;
+    out[17] = ((w15 >> 9) | (w16 << 55)) & mask;
+    out[18] = (w16 >> 2) & mask;
+    out[19] = ((w16 >> 59) | (w17 << 5)) & mask;
+    out[20] = ((w17 >> 52) | (w18 << 12)) & mask;
+    out[21] = ((w18 >> 45) | (w19 << 19)) & mask;
+    out[22] = ((w19 >> 38) | (w20 << 26)) & mask;
+    out[23] = ((w20 >> 31) | (w21 << 33)) & mask;
+    out[24] = ((w21 >> 24) | (w22 << 40)) & mask;
+    out[25] = ((w22 >> 17) | (w23 << 47)) & mask;
+    out[26] = ((w23 >> 10) | (w24 << 54)) & mask;
+    out[27] = (w24 >> 3) & mask;
+    out[28] = ((w24 >> 60) | (w25 << 4)) & mask;
+    out[29] = ((w25 >> 53) | (w26 << 11)) & mask;
+    out[30] = ((w26 >> 46) | (w27 << 18)) & mask;
+    out[31] = ((w27 >> 39) | (w28 << 25)) & mask;
+
+    return in + (28 * 8 + 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 58> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 58) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    const auto w8 = LoadInt<uint64_t>(in + 8 * 8);
+    const auto w9 = LoadInt<uint64_t>(in + 9 * 8);
+    const auto w10 = LoadInt<uint64_t>(in + 10 * 8);
+    const auto w11 = LoadInt<uint64_t>(in + 11 * 8);
+    const auto w12 = LoadInt<uint64_t>(in + 12 * 8);
+    const auto w13 = LoadInt<uint64_t>(in + 13 * 8);
+    const auto w14 = LoadInt<uint64_t>(in + 14 * 8);
+    const auto w15 = LoadInt<uint64_t>(in + 15 * 8);
+    const auto w16 = LoadInt<uint64_t>(in + 16 * 8);
+    const auto w17 = LoadInt<uint64_t>(in + 17 * 8);
+    const auto w18 = LoadInt<uint64_t>(in + 18 * 8);
+    const auto w19 = LoadInt<uint64_t>(in + 19 * 8);
+    const auto w20 = LoadInt<uint64_t>(in + 20 * 8);
+    const auto w21 = LoadInt<uint64_t>(in + 21 * 8);
+    const auto w22 = LoadInt<uint64_t>(in + 22 * 8);
+    const auto w23 = LoadInt<uint64_t>(in + 23 * 8);
+    const auto w24 = LoadInt<uint64_t>(in + 24 * 8);
+    const auto w25 = LoadInt<uint64_t>(in + 25 * 8);
+    const auto w26 = LoadInt<uint64_t>(in + 26 * 8);
+    const auto w27 = LoadInt<uint64_t>(in + 27 * 8);
+    const auto w28 = LoadInt<uint64_t>(in + 28 * 8);
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 58) | (w1 << 6)) & mask;
+    out[2] = ((w1 >> 52) | (w2 << 12)) & mask;
+    out[3] = ((w2 >> 46) | (w3 << 18)) & mask;
+    out[4] = ((w3 >> 40) | (w4 << 24)) & mask;
+    out[5] = ((w4 >> 34) | (w5 << 30)) & mask;
+    out[6] = ((w5 >> 28) | (w6 << 36)) & mask;
+    out[7] = ((w6 >> 22) | (w7 << 42)) & mask;
+    out[8] = ((w7 >> 16) | (w8 << 48)) & mask;
+    out[9] = ((w8 >> 10) | (w9 << 54)) & mask;
+    out[10] = (w9 >> 4) & mask;
+    out[11] = ((w9 >> 62) | (w10 << 2)) & mask;
+    out[12] = ((w10 >> 56) | (w11 << 8)) & mask;
+    out[13] = ((w11 >> 50) | (w12 << 14)) & mask;
+    out[14] = ((w12 >> 44) | (w13 << 20)) & mask;
+    out[15] = ((w13 >> 38) | (w14 << 26)) & mask;
+    out[16] = ((w14 >> 32) | (w15 << 32)) & mask;
+    out[17] = ((w15 >> 26) | (w16 << 38)) & mask;
+    out[18] = ((w16 >> 20) | (w17 << 44)) & mask;
+    out[19] = ((w17 >> 14) | (w18 << 50)) & mask;
+    out[20] = ((w18 >> 8) | (w19 << 56)) & mask;
+    out[21] = (w19 >> 2) & mask;
+    out[22] = ((w19 >> 60) | (w20 << 4)) & mask;
+    out[23] = ((w20 >> 54) | (w21 << 10)) & mask;
+    out[24] = ((w21 >> 48) | (w22 << 16)) & mask;
+    out[25] = ((w22 >> 42) | (w23 << 22)) & mask;
+    out[26] = ((w23 >> 36) | (w24 << 28)) & mask;
+    out[27] = ((w24 >> 30) | (w25 << 34)) & mask;
+    out[28] = ((w25 >> 24) | (w26 << 40)) & mask;
+    out[29] = ((w26 >> 18) | (w27 << 46)) & mask;
+    out[30] = ((w27 >> 12) | (w28 << 52)) & mask;
+    out[31] = w28 >> 6;
+
+    return in + (29 * 8);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 59> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 59) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    const auto w8 = LoadInt<uint64_t>(in + 8 * 8);
+    const auto w9 = LoadInt<uint64_t>(in + 9 * 8);
+    const auto w10 = LoadInt<uint64_t>(in + 10 * 8);
+    const auto w11 = LoadInt<uint64_t>(in + 11 * 8);
+    const auto w12 = LoadInt<uint64_t>(in + 12 * 8);
+    const auto w13 = LoadInt<uint64_t>(in + 13 * 8);
+    const auto w14 = LoadInt<uint64_t>(in + 14 * 8);
+    const auto w15 = LoadInt<uint64_t>(in + 15 * 8);
+    const auto w16 = LoadInt<uint64_t>(in + 16 * 8);
+    const auto w17 = LoadInt<uint64_t>(in + 17 * 8);
+    const auto w18 = LoadInt<uint64_t>(in + 18 * 8);
+    const auto w19 = LoadInt<uint64_t>(in + 19 * 8);
+    const auto w20 = LoadInt<uint64_t>(in + 20 * 8);
+    const auto w21 = LoadInt<uint64_t>(in + 21 * 8);
+    const auto w22 = LoadInt<uint64_t>(in + 22 * 8);
+    const auto w23 = LoadInt<uint64_t>(in + 23 * 8);
+    const auto w24 = LoadInt<uint64_t>(in + 24 * 8);
+    const auto w25 = LoadInt<uint64_t>(in + 25 * 8);
+    const auto w26 = LoadInt<uint64_t>(in + 26 * 8);
+    const auto w27 = LoadInt<uint64_t>(in + 27 * 8);
+    const auto w28 = LoadInt<uint64_t>(in + 28 * 8);
+    const auto w29 = static_cast<uint64_t>(LoadInt<uint32_t>(in + 29 * 8));
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 59) | (w1 << 5)) & mask;
+    out[2] = ((w1 >> 54) | (w2 << 10)) & mask;
+    out[3] = ((w2 >> 49) | (w3 << 15)) & mask;
+    out[4] = ((w3 >> 44) | (w4 << 20)) & mask;
+    out[5] = ((w4 >> 39) | (w5 << 25)) & mask;
+    out[6] = ((w5 >> 34) | (w6 << 30)) & mask;
+    out[7] = ((w6 >> 29) | (w7 << 35)) & mask;
+    out[8] = ((w7 >> 24) | (w8 << 40)) & mask;
+    out[9] = ((w8 >> 19) | (w9 << 45)) & mask;
+    out[10] = ((w9 >> 14) | (w10 << 50)) & mask;
+    out[11] = ((w10 >> 9) | (w11 << 55)) & mask;
+    out[12] = (w11 >> 4) & mask;
+    out[13] = ((w11 >> 63) | (w12 << 1)) & mask;
+    out[14] = ((w12 >> 58) | (w13 << 6)) & mask;
+    out[15] = ((w13 >> 53) | (w14 << 11)) & mask;
+    out[16] = ((w14 >> 48) | (w15 << 16)) & mask;
+    out[17] = ((w15 >> 43) | (w16 << 21)) & mask;
+    out[18] = ((w16 >> 38) | (w17 << 26)) & mask;
+    out[19] = ((w17 >> 33) | (w18 << 31)) & mask;
+    out[20] = ((w18 >> 28) | (w19 << 36)) & mask;
+    out[21] = ((w19 >> 23) | (w20 << 41)) & mask;
+    out[22] = ((w20 >> 18) | (w21 << 46)) & mask;
+    out[23] = ((w21 >> 13) | (w22 << 51)) & mask;
+    out[24] = ((w22 >> 8) | (w23 << 56)) & mask;
+    out[25] = (w23 >> 3) & mask;
+    out[26] = ((w23 >> 62) | (w24 << 2)) & mask;
+    out[27] = ((w24 >> 57) | (w25 << 7)) & mask;
+    out[28] = ((w25 >> 52) | (w26 << 12)) & mask;
+    out[29] = ((w26 >> 47) | (w27 << 17)) & mask;
+    out[30] = ((w27 >> 42) | (w28 << 22)) & mask;
+    out[31] = ((w28 >> 37) | (w29 << 27)) & mask;
+
+    return in + (29 * 8 + 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 60> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 60) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    const auto w8 = LoadInt<uint64_t>(in + 8 * 8);
+    const auto w9 = LoadInt<uint64_t>(in + 9 * 8);
+    const auto w10 = LoadInt<uint64_t>(in + 10 * 8);
+    const auto w11 = LoadInt<uint64_t>(in + 11 * 8);
+    const auto w12 = LoadInt<uint64_t>(in + 12 * 8);
+    const auto w13 = LoadInt<uint64_t>(in + 13 * 8);
+    const auto w14 = LoadInt<uint64_t>(in + 14 * 8);
+    const auto w15 = LoadInt<uint64_t>(in + 15 * 8);
+    const auto w16 = LoadInt<uint64_t>(in + 16 * 8);
+    const auto w17 = LoadInt<uint64_t>(in + 17 * 8);
+    const auto w18 = LoadInt<uint64_t>(in + 18 * 8);
+    const auto w19 = LoadInt<uint64_t>(in + 19 * 8);
+    const auto w20 = LoadInt<uint64_t>(in + 20 * 8);
+    const auto w21 = LoadInt<uint64_t>(in + 21 * 8);
+    const auto w22 = LoadInt<uint64_t>(in + 22 * 8);
+    const auto w23 = LoadInt<uint64_t>(in + 23 * 8);
+    const auto w24 = LoadInt<uint64_t>(in + 24 * 8);
+    const auto w25 = LoadInt<uint64_t>(in + 25 * 8);
+    const auto w26 = LoadInt<uint64_t>(in + 26 * 8);
+    const auto w27 = LoadInt<uint64_t>(in + 27 * 8);
+    const auto w28 = LoadInt<uint64_t>(in + 28 * 8);
+    const auto w29 = LoadInt<uint64_t>(in + 29 * 8);
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 60) | (w1 << 4)) & mask;
+    out[2] = ((w1 >> 56) | (w2 << 8)) & mask;
+    out[3] = ((w2 >> 52) | (w3 << 12)) & mask;
+    out[4] = ((w3 >> 48) | (w4 << 16)) & mask;
+    out[5] = ((w4 >> 44) | (w5 << 20)) & mask;
+    out[6] = ((w5 >> 40) | (w6 << 24)) & mask;
+    out[7] = ((w6 >> 36) | (w7 << 28)) & mask;
+    out[8] = ((w7 >> 32) | (w8 << 32)) & mask;
+    out[9] = ((w8 >> 28) | (w9 << 36)) & mask;
+    out[10] = ((w9 >> 24) | (w10 << 40)) & mask;
+    out[11] = ((w10 >> 20) | (w11 << 44)) & mask;
+    out[12] = ((w11 >> 16) | (w12 << 48)) & mask;
+    out[13] = ((w12 >> 12) | (w13 << 52)) & mask;
+    out[14] = ((w13 >> 8) | (w14 << 56)) & mask;
+    out[15] = w14 >> 4;
+    out[16] = (w15) & mask;
+    out[17] = ((w15 >> 60) | (w16 << 4)) & mask;
+    out[18] = ((w16 >> 56) | (w17 << 8)) & mask;
+    out[19] = ((w17 >> 52) | (w18 << 12)) & mask;
+    out[20] = ((w18 >> 48) | (w19 << 16)) & mask;
+    out[21] = ((w19 >> 44) | (w20 << 20)) & mask;
+    out[22] = ((w20 >> 40) | (w21 << 24)) & mask;
+    out[23] = ((w21 >> 36) | (w22 << 28)) & mask;
+    out[24] = ((w22 >> 32) | (w23 << 32)) & mask;
+    out[25] = ((w23 >> 28) | (w24 << 36)) & mask;
+    out[26] = ((w24 >> 24) | (w25 << 40)) & mask;
+    out[27] = ((w25 >> 20) | (w26 << 44)) & mask;
+    out[28] = ((w26 >> 16) | (w27 << 48)) & mask;
+    out[29] = ((w27 >> 12) | (w28 << 52)) & mask;
+    out[30] = ((w28 >> 8) | (w29 << 56)) & mask;
+    out[31] = w29 >> 4;
+
+    return in + (30 * 8);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 61> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 61) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    const auto w8 = LoadInt<uint64_t>(in + 8 * 8);
+    const auto w9 = LoadInt<uint64_t>(in + 9 * 8);
+    const auto w10 = LoadInt<uint64_t>(in + 10 * 8);
+    const auto w11 = LoadInt<uint64_t>(in + 11 * 8);
+    const auto w12 = LoadInt<uint64_t>(in + 12 * 8);
+    const auto w13 = LoadInt<uint64_t>(in + 13 * 8);
+    const auto w14 = LoadInt<uint64_t>(in + 14 * 8);
+    const auto w15 = LoadInt<uint64_t>(in + 15 * 8);
+    const auto w16 = LoadInt<uint64_t>(in + 16 * 8);
+    const auto w17 = LoadInt<uint64_t>(in + 17 * 8);
+    const auto w18 = LoadInt<uint64_t>(in + 18 * 8);
+    const auto w19 = LoadInt<uint64_t>(in + 19 * 8);
+    const auto w20 = LoadInt<uint64_t>(in + 20 * 8);
+    const auto w21 = LoadInt<uint64_t>(in + 21 * 8);
+    const auto w22 = LoadInt<uint64_t>(in + 22 * 8);
+    const auto w23 = LoadInt<uint64_t>(in + 23 * 8);
+    const auto w24 = LoadInt<uint64_t>(in + 24 * 8);
+    const auto w25 = LoadInt<uint64_t>(in + 25 * 8);
+    const auto w26 = LoadInt<uint64_t>(in + 26 * 8);
+    const auto w27 = LoadInt<uint64_t>(in + 27 * 8);
+    const auto w28 = LoadInt<uint64_t>(in + 28 * 8);
+    const auto w29 = LoadInt<uint64_t>(in + 29 * 8);
+    const auto w30 = static_cast<uint64_t>(LoadInt<uint32_t>(in + 30 * 8));
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 61) | (w1 << 3)) & mask;
+    out[2] = ((w1 >> 58) | (w2 << 6)) & mask;
+    out[3] = ((w2 >> 55) | (w3 << 9)) & mask;
+    out[4] = ((w3 >> 52) | (w4 << 12)) & mask;
+    out[5] = ((w4 >> 49) | (w5 << 15)) & mask;
+    out[6] = ((w5 >> 46) | (w6 << 18)) & mask;
+    out[7] = ((w6 >> 43) | (w7 << 21)) & mask;
+    out[8] = ((w7 >> 40) | (w8 << 24)) & mask;
+    out[9] = ((w8 >> 37) | (w9 << 27)) & mask;
+    out[10] = ((w9 >> 34) | (w10 << 30)) & mask;
+    out[11] = ((w10 >> 31) | (w11 << 33)) & mask;
+    out[12] = ((w11 >> 28) | (w12 << 36)) & mask;
+    out[13] = ((w12 >> 25) | (w13 << 39)) & mask;
+    out[14] = ((w13 >> 22) | (w14 << 42)) & mask;
+    out[15] = ((w14 >> 19) | (w15 << 45)) & mask;
+    out[16] = ((w15 >> 16) | (w16 << 48)) & mask;
+    out[17] = ((w16 >> 13) | (w17 << 51)) & mask;
+    out[18] = ((w17 >> 10) | (w18 << 54)) & mask;
+    out[19] = ((w18 >> 7) | (w19 << 57)) & mask;
+    out[20] = ((w19 >> 4) | (w20 << 60)) & mask;
+    out[21] = (w20 >> 1) & mask;
+    out[22] = ((w20 >> 62) | (w21 << 2)) & mask;
+    out[23] = ((w21 >> 59) | (w22 << 5)) & mask;
+    out[24] = ((w22 >> 56) | (w23 << 8)) & mask;
+    out[25] = ((w23 >> 53) | (w24 << 11)) & mask;
+    out[26] = ((w24 >> 50) | (w25 << 14)) & mask;
+    out[27] = ((w25 >> 47) | (w26 << 17)) & mask;
+    out[28] = ((w26 >> 44) | (w27 << 20)) & mask;
+    out[29] = ((w27 >> 41) | (w28 << 23)) & mask;
+    out[30] = ((w28 >> 38) | (w29 << 26)) & mask;
+    out[31] = ((w29 >> 35) | (w30 << 29)) & mask;
+
+    return in + (30 * 8 + 4);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 62> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 62) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    const auto w8 = LoadInt<uint64_t>(in + 8 * 8);
+    const auto w9 = LoadInt<uint64_t>(in + 9 * 8);
+    const auto w10 = LoadInt<uint64_t>(in + 10 * 8);
+    const auto w11 = LoadInt<uint64_t>(in + 11 * 8);
+    const auto w12 = LoadInt<uint64_t>(in + 12 * 8);
+    const auto w13 = LoadInt<uint64_t>(in + 13 * 8);
+    const auto w14 = LoadInt<uint64_t>(in + 14 * 8);
+    const auto w15 = LoadInt<uint64_t>(in + 15 * 8);
+    const auto w16 = LoadInt<uint64_t>(in + 16 * 8);
+    const auto w17 = LoadInt<uint64_t>(in + 17 * 8);
+    const auto w18 = LoadInt<uint64_t>(in + 18 * 8);
+    const auto w19 = LoadInt<uint64_t>(in + 19 * 8);
+    const auto w20 = LoadInt<uint64_t>(in + 20 * 8);
+    const auto w21 = LoadInt<uint64_t>(in + 21 * 8);
+    const auto w22 = LoadInt<uint64_t>(in + 22 * 8);
+    const auto w23 = LoadInt<uint64_t>(in + 23 * 8);
+    const auto w24 = LoadInt<uint64_t>(in + 24 * 8);
+    const auto w25 = LoadInt<uint64_t>(in + 25 * 8);
+    const auto w26 = LoadInt<uint64_t>(in + 26 * 8);
+    const auto w27 = LoadInt<uint64_t>(in + 27 * 8);
+    const auto w28 = LoadInt<uint64_t>(in + 28 * 8);
+    const auto w29 = LoadInt<uint64_t>(in + 29 * 8);
+    const auto w30 = LoadInt<uint64_t>(in + 30 * 8);
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 62) | (w1 << 2)) & mask;
+    out[2] = ((w1 >> 60) | (w2 << 4)) & mask;
+    out[3] = ((w2 >> 58) | (w3 << 6)) & mask;
+    out[4] = ((w3 >> 56) | (w4 << 8)) & mask;
+    out[5] = ((w4 >> 54) | (w5 << 10)) & mask;
+    out[6] = ((w5 >> 52) | (w6 << 12)) & mask;
+    out[7] = ((w6 >> 50) | (w7 << 14)) & mask;
+    out[8] = ((w7 >> 48) | (w8 << 16)) & mask;
+    out[9] = ((w8 >> 46) | (w9 << 18)) & mask;
+    out[10] = ((w9 >> 44) | (w10 << 20)) & mask;
+    out[11] = ((w10 >> 42) | (w11 << 22)) & mask;
+    out[12] = ((w11 >> 40) | (w12 << 24)) & mask;
+    out[13] = ((w12 >> 38) | (w13 << 26)) & mask;
+    out[14] = ((w13 >> 36) | (w14 << 28)) & mask;
+    out[15] = ((w14 >> 34) | (w15 << 30)) & mask;
+    out[16] = ((w15 >> 32) | (w16 << 32)) & mask;
+    out[17] = ((w16 >> 30) | (w17 << 34)) & mask;
+    out[18] = ((w17 >> 28) | (w18 << 36)) & mask;
+    out[19] = ((w18 >> 26) | (w19 << 38)) & mask;
+    out[20] = ((w19 >> 24) | (w20 << 40)) & mask;
+    out[21] = ((w20 >> 22) | (w21 << 42)) & mask;
+    out[22] = ((w21 >> 20) | (w22 << 44)) & mask;
+    out[23] = ((w22 >> 18) | (w23 << 46)) & mask;
+    out[24] = ((w23 >> 16) | (w24 << 48)) & mask;
+    out[25] = ((w24 >> 14) | (w25 << 50)) & mask;
+    out[26] = ((w25 >> 12) | (w26 << 52)) & mask;
+    out[27] = ((w26 >> 10) | (w27 << 54)) & mask;
+    out[28] = ((w27 >> 8) | (w28 << 56)) & mask;
+    out[29] = ((w28 >> 6) | (w29 << 58)) & mask;
+    out[30] = ((w29 >> 4) | (w30 << 60)) & mask;
+    out[31] = w30 >> 2;
+
+    return in + (31 * 8);
+  }
+};
+
+template<>
+struct ScalarUnpackerForWidth<uint64_t, 63> {
+
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t mask = ((uint64_t{1} << 63) - uint64_t{1});
+
+    const auto w0 = LoadInt<uint64_t>(in + 0 * 8);
+    const auto w1 = LoadInt<uint64_t>(in + 1 * 8);
+    const auto w2 = LoadInt<uint64_t>(in + 2 * 8);
+    const auto w3 = LoadInt<uint64_t>(in + 3 * 8);
+    const auto w4 = LoadInt<uint64_t>(in + 4 * 8);
+    const auto w5 = LoadInt<uint64_t>(in + 5 * 8);
+    const auto w6 = LoadInt<uint64_t>(in + 6 * 8);
+    const auto w7 = LoadInt<uint64_t>(in + 7 * 8);
+    const auto w8 = LoadInt<uint64_t>(in + 8 * 8);
+    const auto w9 = LoadInt<uint64_t>(in + 9 * 8);
+    const auto w10 = LoadInt<uint64_t>(in + 10 * 8);
+    const auto w11 = LoadInt<uint64_t>(in + 11 * 8);
+    const auto w12 = LoadInt<uint64_t>(in + 12 * 8);
+    const auto w13 = LoadInt<uint64_t>(in + 13 * 8);
+    const auto w14 = LoadInt<uint64_t>(in + 14 * 8);
+    const auto w15 = LoadInt<uint64_t>(in + 15 * 8);
+    const auto w16 = LoadInt<uint64_t>(in + 16 * 8);
+    const auto w17 = LoadInt<uint64_t>(in + 17 * 8);
+    const auto w18 = LoadInt<uint64_t>(in + 18 * 8);
+    const auto w19 = LoadInt<uint64_t>(in + 19 * 8);
+    const auto w20 = LoadInt<uint64_t>(in + 20 * 8);
+    const auto w21 = LoadInt<uint64_t>(in + 21 * 8);
+    const auto w22 = LoadInt<uint64_t>(in + 22 * 8);
+    const auto w23 = LoadInt<uint64_t>(in + 23 * 8);
+    const auto w24 = LoadInt<uint64_t>(in + 24 * 8);
+    const auto w25 = LoadInt<uint64_t>(in + 25 * 8);
+    const auto w26 = LoadInt<uint64_t>(in + 26 * 8);
+    const auto w27 = LoadInt<uint64_t>(in + 27 * 8);
+    const auto w28 = LoadInt<uint64_t>(in + 28 * 8);
+    const auto w29 = LoadInt<uint64_t>(in + 29 * 8);
+    const auto w30 = LoadInt<uint64_t>(in + 30 * 8);
+    const auto w31 = static_cast<uint64_t>(LoadInt<uint32_t>(in + 31 * 8));
+    out[0] = (w0) & mask;
+    out[1] = ((w0 >> 63) | (w1 << 1)) & mask;
+    out[2] = ((w1 >> 62) | (w2 << 2)) & mask;
+    out[3] = ((w2 >> 61) | (w3 << 3)) & mask;
+    out[4] = ((w3 >> 60) | (w4 << 4)) & mask;
+    out[5] = ((w4 >> 59) | (w5 << 5)) & mask;
+    out[6] = ((w5 >> 58) | (w6 << 6)) & mask;
+    out[7] = ((w6 >> 57) | (w7 << 7)) & mask;
+    out[8] = ((w7 >> 56) | (w8 << 8)) & mask;
+    out[9] = ((w8 >> 55) | (w9 << 9)) & mask;
+    out[10] = ((w9 >> 54) | (w10 << 10)) & mask;
+    out[11] = ((w10 >> 53) | (w11 << 11)) & mask;
+    out[12] = ((w11 >> 52) | (w12 << 12)) & mask;
+    out[13] = ((w12 >> 51) | (w13 << 13)) & mask;
+    out[14] = ((w13 >> 50) | (w14 << 14)) & mask;
+    out[15] = ((w14 >> 49) | (w15 << 15)) & mask;
+    out[16] = ((w15 >> 48) | (w16 << 16)) & mask;
+    out[17] = ((w16 >> 47) | (w17 << 17)) & mask;
+    out[18] = ((w17 >> 46) | (w18 << 18)) & mask;
+    out[19] = ((w18 >> 45) | (w19 << 19)) & mask;
+    out[20] = ((w19 >> 44) | (w20 << 20)) & mask;
+    out[21] = ((w20 >> 43) | (w21 << 21)) & mask;
+    out[22] = ((w21 >> 42) | (w22 << 22)) & mask;
+    out[23] = ((w22 >> 41) | (w23 << 23)) & mask;
+    out[24] = ((w23 >> 40) | (w24 << 24)) & mask;
+    out[25] = ((w24 >> 39) | (w25 << 25)) & mask;
+    out[26] = ((w25 >> 38) | (w26 << 26)) & mask;
+    out[27] = ((w26 >> 37) | (w27 << 27)) & mask;
+    out[28] = ((w27 >> 36) | (w28 << 28)) & mask;
+    out[29] = ((w28 >> 35) | (w29 << 29)) & mask;
+    out[30] = ((w29 >> 34) | (w30 << 30)) & mask;
+    out[31] = ((w30 >> 33) | (w31 << 31)) & mask;
+
+    return in + (31 * 8 + 4);
+  }
+};
+
+
+}  // namespace arrow::internal
+
diff --git a/cpp/src/arrow/util/bpacking_scalar_internal.h b/cpp/src/arrow/util/bpacking_scalar_internal.h
new file mode 100644
index 0000000000000..0ae3de10c2f18
--- /dev/null
+++ b/cpp/src/arrow/util/bpacking_scalar_internal.h
@@ -0,0 +1,52 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#pragma once
+
+#include "arrow/util/bpacking_internal.h"
+#include "arrow/util/visibility.h"
+
+#include <cstdint>
+
+namespace arrow::internal {
+
+template <typename Uint>
+ARROW_EXPORT int unpack_scalar(const uint8_t* in, Uint* out, int batch_size,
+                               int num_bits);
+
+extern template ARROW_TEMPLATE_EXPORT int unpack_scalar<uint16_t>(const uint8_t*,
+                                                                  uint16_t*, int, int);
+
+extern template ARROW_TEMPLATE_EXPORT int unpack_scalar<uint32_t>(const uint8_t*,
+                                                                  uint32_t*, int, int);
+
+extern template ARROW_TEMPLATE_EXPORT int unpack_scalar<uint64_t>(const uint8_t*,
+                                                                  uint64_t*, int, int);
+
+template <typename Uint>
+ARROW_EXPORT UnpackFn<Uint> get_unpack_fn_scalar(int num_bits);
+
+extern template ARROW_TEMPLATE_EXPORT UnpackFn<uint16_t> get_unpack_fn_scalar<uint16_t>(
+    int);
+
+extern template ARROW_TEMPLATE_EXPORT UnpackFn<uint32_t> get_unpack_fn_scalar<uint32_t>(
+    int);
+
+extern template ARROW_TEMPLATE_EXPORT UnpackFn<uint64_t> get_unpack_fn_scalar<uint64_t>(
+    int);
+
+}  // namespace arrow::internal
diff --git a/cpp/src/arrow/util/bpacking_simd128_generated_internal.h b/cpp/src/arrow/util/bpacking_simd128_generated_internal.h
index 5beecad421033..7e22db4d28dc4 100644
--- a/cpp/src/arrow/util/bpacking_simd128_generated_internal.h
+++ b/cpp/src/arrow/util/bpacking_simd128_generated_internal.h
@@ -15,7 +15,9 @@
 // specific language governing permissions and limitations
 // under the License.
 
-// Automatically generated file; DO NOT EDIT.
+// WARNING: this file is generated, DO NOT EDIT.
+// Usage:
+//   python cpp/src/arrow/util/bpacking_simd_codegen.py 128
 
 #pragma once
 
@@ -24,2115 +26,25485 @@
 
 #include <xsimd/xsimd.hpp>
 
-#include "arrow/util/dispatch_internal.h"
 #include "arrow/util/ubsan.h"
 
-namespace arrow {
-namespace internal {
-namespace {
-
-using ::arrow::util::SafeLoad;
-
-template <DispatchLevel level>
-struct UnpackBits128 {
-
-using simd_batch = xsimd::make_sized_batch_t<uint32_t, 4>;
-
-inline static const uint32_t* unpack0_32(const uint32_t* in, uint32_t* out) {
-  memset(out, 0x0, 32 * sizeof(*out));
-  out += 32;
-
-  return in;
-}
-
-inline static const uint32_t* unpack1_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x1;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 1-bit bundles 0 to 3
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) };
-  shifts = simd_batch{ 0, 1, 2, 3 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 1-bit bundles 4 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) };
-  shifts = simd_batch{ 4, 5, 6, 7 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 1-bit bundles 8 to 11
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) };
-  shifts = simd_batch{ 8, 9, 10, 11 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 1-bit bundles 12 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) };
-  shifts = simd_batch{ 12, 13, 14, 15 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 1-bit bundles 16 to 19
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) };
-  shifts = simd_batch{ 16, 17, 18, 19 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 1-bit bundles 20 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) };
-  shifts = simd_batch{ 20, 21, 22, 23 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 1-bit bundles 24 to 27
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) };
-  shifts = simd_batch{ 24, 25, 26, 27 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 1-bit bundles 28 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) };
-  shifts = simd_batch{ 28, 29, 30, 31 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  in += 1;
-  return in;
-}
-
-inline static const uint32_t* unpack2_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x3;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 2-bit bundles 0 to 3
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) };
-  shifts = simd_batch{ 0, 2, 4, 6 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 2-bit bundles 4 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) };
-  shifts = simd_batch{ 8, 10, 12, 14 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 2-bit bundles 8 to 11
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) };
-  shifts = simd_batch{ 16, 18, 20, 22 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 2-bit bundles 12 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) };
-  shifts = simd_batch{ 24, 26, 28, 30 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 2-bit bundles 16 to 19
-  words = simd_batch{ SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) };
-  shifts = simd_batch{ 0, 2, 4, 6 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 2-bit bundles 20 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) };
-  shifts = simd_batch{ 8, 10, 12, 14 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 2-bit bundles 24 to 27
-  words = simd_batch{ SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) };
-  shifts = simd_batch{ 16, 18, 20, 22 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 2-bit bundles 28 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) };
-  shifts = simd_batch{ 24, 26, 28, 30 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  in += 2;
-  return in;
-}
-
-inline static const uint32_t* unpack3_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x7;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 3-bit bundles 0 to 3
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) };
-  shifts = simd_batch{ 0, 3, 6, 9 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 3-bit bundles 4 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) };
-  shifts = simd_batch{ 12, 15, 18, 21 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 3-bit bundles 8 to 11
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 30 | SafeLoad<uint32_t>(in + 1) << 2, SafeLoad<uint32_t>(in + 1) };
-  shifts = simd_batch{ 24, 27, 0, 1 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 3-bit bundles 12 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) };
-  shifts = simd_batch{ 4, 7, 10, 13 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 3-bit bundles 16 to 19
-  words = simd_batch{ SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) };
-  shifts = simd_batch{ 16, 19, 22, 25 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 3-bit bundles 20 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) >> 31 | SafeLoad<uint32_t>(in + 2) << 1, SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) };
-  shifts = simd_batch{ 28, 0, 2, 5 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 3-bit bundles 24 to 27
-  words = simd_batch{ SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) };
-  shifts = simd_batch{ 8, 11, 14, 17 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 3-bit bundles 28 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) };
-  shifts = simd_batch{ 20, 23, 26, 29 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  in += 3;
-  return in;
-}
-
-inline static const uint32_t* unpack4_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0xf;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 4-bit bundles 0 to 3
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) };
-  shifts = simd_batch{ 0, 4, 8, 12 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 4-bit bundles 4 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) };
-  shifts = simd_batch{ 16, 20, 24, 28 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 4-bit bundles 8 to 11
-  words = simd_batch{ SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) };
-  shifts = simd_batch{ 0, 4, 8, 12 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 4-bit bundles 12 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) };
-  shifts = simd_batch{ 16, 20, 24, 28 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 4-bit bundles 16 to 19
-  words = simd_batch{ SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) };
-  shifts = simd_batch{ 0, 4, 8, 12 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 4-bit bundles 20 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) };
-  shifts = simd_batch{ 16, 20, 24, 28 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 4-bit bundles 24 to 27
-  words = simd_batch{ SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) };
-  shifts = simd_batch{ 0, 4, 8, 12 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 4-bit bundles 28 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) };
-  shifts = simd_batch{ 16, 20, 24, 28 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  in += 4;
-  return in;
-}
-
-inline static const uint32_t* unpack5_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x1f;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 5-bit bundles 0 to 3
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) };
-  shifts = simd_batch{ 0, 5, 10, 15 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 5-bit bundles 4 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 30 | SafeLoad<uint32_t>(in + 1) << 2, SafeLoad<uint32_t>(in + 1) };
-  shifts = simd_batch{ 20, 25, 0, 3 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 5-bit bundles 8 to 11
-  words = simd_batch{ SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) };
-  shifts = simd_batch{ 8, 13, 18, 23 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 5-bit bundles 12 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 1) >> 28 | SafeLoad<uint32_t>(in + 2) << 4, SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) };
-  shifts = simd_batch{ 0, 1, 6, 11 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 5-bit bundles 16 to 19
-  words = simd_batch{ SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) >> 31 | SafeLoad<uint32_t>(in + 3) << 1 };
-  shifts = simd_batch{ 16, 21, 26, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 5-bit bundles 20 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) };
-  shifts = simd_batch{ 4, 9, 14, 19 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 5-bit bundles 24 to 27
-  words = simd_batch{ SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) >> 29 | SafeLoad<uint32_t>(in + 4) << 3, SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4) };
-  shifts = simd_batch{ 24, 0, 2, 7 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 5-bit bundles 28 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4) };
-  shifts = simd_batch{ 12, 17, 22, 27 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  in += 5;
-  return in;
-}
-
-inline static const uint32_t* unpack6_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x3f;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 6-bit bundles 0 to 3
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) };
-  shifts = simd_batch{ 0, 6, 12, 18 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 6-bit bundles 4 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 30 | SafeLoad<uint32_t>(in + 1) << 2, SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) };
-  shifts = simd_batch{ 24, 0, 4, 10 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 6-bit bundles 8 to 11
-  words = simd_batch{ SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) >> 28 | SafeLoad<uint32_t>(in + 2) << 4, SafeLoad<uint32_t>(in + 2) };
-  shifts = simd_batch{ 16, 22, 0, 2 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 6-bit bundles 12 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) };
-  shifts = simd_batch{ 8, 14, 20, 26 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 6-bit bundles 16 to 19
-  words = simd_batch{ SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) };
-  shifts = simd_batch{ 0, 6, 12, 18 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 6-bit bundles 20 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) >> 30 | SafeLoad<uint32_t>(in + 4) << 2, SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4) };
-  shifts = simd_batch{ 24, 0, 4, 10 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 6-bit bundles 24 to 27
-  words = simd_batch{ SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4) >> 28 | SafeLoad<uint32_t>(in + 5) << 4, SafeLoad<uint32_t>(in + 5) };
-  shifts = simd_batch{ 16, 22, 0, 2 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 6-bit bundles 28 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) };
-  shifts = simd_batch{ 8, 14, 20, 26 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  in += 6;
-  return in;
-}
-
-inline static const uint32_t* unpack7_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x7f;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 7-bit bundles 0 to 3
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) };
-  shifts = simd_batch{ 0, 7, 14, 21 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 7-bit bundles 4 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0) >> 28 | SafeLoad<uint32_t>(in + 1) << 4, SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) };
-  shifts = simd_batch{ 0, 3, 10, 17 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 7-bit bundles 8 to 11
-  words = simd_batch{ SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) >> 31 | SafeLoad<uint32_t>(in + 2) << 1, SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) };
-  shifts = simd_batch{ 24, 0, 6, 13 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 7-bit bundles 12 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) >> 27 | SafeLoad<uint32_t>(in + 3) << 5, SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) };
-  shifts = simd_batch{ 20, 0, 2, 9 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 7-bit bundles 16 to 19
-  words = simd_batch{ SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) >> 30 | SafeLoad<uint32_t>(in + 4) << 2, SafeLoad<uint32_t>(in + 4) };
-  shifts = simd_batch{ 16, 23, 0, 5 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 7-bit bundles 20 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4) >> 26 | SafeLoad<uint32_t>(in + 5) << 6, SafeLoad<uint32_t>(in + 5) };
-  shifts = simd_batch{ 12, 19, 0, 1 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 7-bit bundles 24 to 27
-  words = simd_batch{ SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) >> 29 | SafeLoad<uint32_t>(in + 6) << 3 };
-  shifts = simd_batch{ 8, 15, 22, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 7-bit bundles 28 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6) };
-  shifts = simd_batch{ 4, 11, 18, 25 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  in += 7;
-  return in;
-}
-
-inline static const uint32_t* unpack8_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0xff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 8-bit bundles 0 to 3
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) };
-  shifts = simd_batch{ 0, 8, 16, 24 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 8-bit bundles 4 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) };
-  shifts = simd_batch{ 0, 8, 16, 24 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 8-bit bundles 8 to 11
-  words = simd_batch{ SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) };
-  shifts = simd_batch{ 0, 8, 16, 24 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 8-bit bundles 12 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) };
-  shifts = simd_batch{ 0, 8, 16, 24 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 8-bit bundles 16 to 19
-  words = simd_batch{ SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4) };
-  shifts = simd_batch{ 0, 8, 16, 24 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 8-bit bundles 20 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) };
-  shifts = simd_batch{ 0, 8, 16, 24 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 8-bit bundles 24 to 27
-  words = simd_batch{ SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6) };
-  shifts = simd_batch{ 0, 8, 16, 24 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 8-bit bundles 28 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7) };
-  shifts = simd_batch{ 0, 8, 16, 24 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  in += 8;
-  return in;
-}
-
-inline static const uint32_t* unpack9_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x1ff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 9-bit bundles 0 to 3
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 27 | SafeLoad<uint32_t>(in + 1) << 5 };
-  shifts = simd_batch{ 0, 9, 18, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 9-bit bundles 4 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) >> 31 | SafeLoad<uint32_t>(in + 2) << 1 };
-  shifts = simd_batch{ 4, 13, 22, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 9-bit bundles 8 to 11
-  words = simd_batch{ SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) >> 26 | SafeLoad<uint32_t>(in + 3) << 6, SafeLoad<uint32_t>(in + 3) };
-  shifts = simd_batch{ 8, 17, 0, 3 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 9-bit bundles 12 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) >> 30 | SafeLoad<uint32_t>(in + 4) << 2, SafeLoad<uint32_t>(in + 4) };
-  shifts = simd_batch{ 12, 21, 0, 7 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 9-bit bundles 16 to 19
-  words = simd_batch{ SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4) >> 25 | SafeLoad<uint32_t>(in + 5) << 7, SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) };
-  shifts = simd_batch{ 16, 0, 2, 11 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 9-bit bundles 20 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) >> 29 | SafeLoad<uint32_t>(in + 6) << 3, SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6) };
-  shifts = simd_batch{ 20, 0, 6, 15 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 9-bit bundles 24 to 27
-  words = simd_batch{ SafeLoad<uint32_t>(in + 6) >> 24 | SafeLoad<uint32_t>(in + 7) << 8, SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7) };
-  shifts = simd_batch{ 0, 1, 10, 19 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 9-bit bundles 28 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 7) >> 28 | SafeLoad<uint32_t>(in + 8) << 4, SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 8) };
-  shifts = simd_batch{ 0, 5, 14, 23 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  in += 9;
-  return in;
-}
-
-inline static const uint32_t* unpack10_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x3ff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 10-bit bundles 0 to 3
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 30 | SafeLoad<uint32_t>(in + 1) << 2 };
-  shifts = simd_batch{ 0, 10, 20, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 10-bit bundles 4 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) >> 28 | SafeLoad<uint32_t>(in + 2) << 4, SafeLoad<uint32_t>(in + 2) };
-  shifts = simd_batch{ 8, 18, 0, 6 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 10-bit bundles 8 to 11
-  words = simd_batch{ SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) >> 26 | SafeLoad<uint32_t>(in + 3) << 6, SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) };
-  shifts = simd_batch{ 16, 0, 4, 14 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 10-bit bundles 12 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 3) >> 24 | SafeLoad<uint32_t>(in + 4) << 8, SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4) };
-  shifts = simd_batch{ 0, 2, 12, 22 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 10-bit bundles 16 to 19
-  words = simd_batch{ SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) >> 30 | SafeLoad<uint32_t>(in + 6) << 2 };
-  shifts = simd_batch{ 0, 10, 20, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 10-bit bundles 20 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6) >> 28 | SafeLoad<uint32_t>(in + 7) << 4, SafeLoad<uint32_t>(in + 7) };
-  shifts = simd_batch{ 8, 18, 0, 6 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 10-bit bundles 24 to 27
-  words = simd_batch{ SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7) >> 26 | SafeLoad<uint32_t>(in + 8) << 6, SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 8) };
-  shifts = simd_batch{ 16, 0, 4, 14 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 10-bit bundles 28 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 8) >> 24 | SafeLoad<uint32_t>(in + 9) << 8, SafeLoad<uint32_t>(in + 9), SafeLoad<uint32_t>(in + 9), SafeLoad<uint32_t>(in + 9) };
-  shifts = simd_batch{ 0, 2, 12, 22 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  in += 10;
-  return in;
-}
-
-inline static const uint32_t* unpack11_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x7ff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 11-bit bundles 0 to 3
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 22 | SafeLoad<uint32_t>(in + 1) << 10, SafeLoad<uint32_t>(in + 1) };
-  shifts = simd_batch{ 0, 11, 0, 1 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 11-bit bundles 4 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) >> 23 | SafeLoad<uint32_t>(in + 2) << 9, SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) };
-  shifts = simd_batch{ 12, 0, 2, 13 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 11-bit bundles 8 to 11
-  words = simd_batch{ SafeLoad<uint32_t>(in + 2) >> 24 | SafeLoad<uint32_t>(in + 3) << 8, SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) >> 25 | SafeLoad<uint32_t>(in + 4) << 7 };
-  shifts = simd_batch{ 0, 3, 14, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 11-bit bundles 12 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4) >> 26 | SafeLoad<uint32_t>(in + 5) << 6, SafeLoad<uint32_t>(in + 5) };
-  shifts = simd_batch{ 4, 15, 0, 5 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 11-bit bundles 16 to 19
-  words = simd_batch{ SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) >> 27 | SafeLoad<uint32_t>(in + 6) << 5, SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6) };
-  shifts = simd_batch{ 16, 0, 6, 17 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 11-bit bundles 20 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 6) >> 28 | SafeLoad<uint32_t>(in + 7) << 4, SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7) >> 29 | SafeLoad<uint32_t>(in + 8) << 3 };
-  shifts = simd_batch{ 0, 7, 18, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 11-bit bundles 24 to 27
-  words = simd_batch{ SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 8) >> 30 | SafeLoad<uint32_t>(in + 9) << 2, SafeLoad<uint32_t>(in + 9) };
-  shifts = simd_batch{ 8, 19, 0, 9 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 11-bit bundles 28 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 9), SafeLoad<uint32_t>(in + 9) >> 31 | SafeLoad<uint32_t>(in + 10) << 1, SafeLoad<uint32_t>(in + 10), SafeLoad<uint32_t>(in + 10) };
-  shifts = simd_batch{ 20, 0, 10, 21 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  in += 11;
-  return in;
-}
-
-inline static const uint32_t* unpack12_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0xfff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 12-bit bundles 0 to 3
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 24 | SafeLoad<uint32_t>(in + 1) << 8, SafeLoad<uint32_t>(in + 1) };
-  shifts = simd_batch{ 0, 12, 0, 4 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 12-bit bundles 4 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) >> 28 | SafeLoad<uint32_t>(in + 2) << 4, SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) };
-  shifts = simd_batch{ 16, 0, 8, 20 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 12-bit bundles 8 to 11
-  words = simd_batch{ SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) >> 24 | SafeLoad<uint32_t>(in + 4) << 8, SafeLoad<uint32_t>(in + 4) };
-  shifts = simd_batch{ 0, 12, 0, 4 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 12-bit bundles 12 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4) >> 28 | SafeLoad<uint32_t>(in + 5) << 4, SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) };
-  shifts = simd_batch{ 16, 0, 8, 20 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 12-bit bundles 16 to 19
-  words = simd_batch{ SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6) >> 24 | SafeLoad<uint32_t>(in + 7) << 8, SafeLoad<uint32_t>(in + 7) };
-  shifts = simd_batch{ 0, 12, 0, 4 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 12-bit bundles 20 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7) >> 28 | SafeLoad<uint32_t>(in + 8) << 4, SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 8) };
-  shifts = simd_batch{ 16, 0, 8, 20 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 12-bit bundles 24 to 27
-  words = simd_batch{ SafeLoad<uint32_t>(in + 9), SafeLoad<uint32_t>(in + 9), SafeLoad<uint32_t>(in + 9) >> 24 | SafeLoad<uint32_t>(in + 10) << 8, SafeLoad<uint32_t>(in + 10) };
-  shifts = simd_batch{ 0, 12, 0, 4 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 12-bit bundles 28 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 10), SafeLoad<uint32_t>(in + 10) >> 28 | SafeLoad<uint32_t>(in + 11) << 4, SafeLoad<uint32_t>(in + 11), SafeLoad<uint32_t>(in + 11) };
-  shifts = simd_batch{ 16, 0, 8, 20 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  in += 12;
-  return in;
-}
-
-inline static const uint32_t* unpack13_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x1fff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 13-bit bundles 0 to 3
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 26 | SafeLoad<uint32_t>(in + 1) << 6, SafeLoad<uint32_t>(in + 1) };
-  shifts = simd_batch{ 0, 13, 0, 7 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 13-bit bundles 4 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 1) >> 20 | SafeLoad<uint32_t>(in + 2) << 12, SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) >> 27 | SafeLoad<uint32_t>(in + 3) << 5 };
-  shifts = simd_batch{ 0, 1, 14, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 13-bit bundles 8 to 11
-  words = simd_batch{ SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) >> 21 | SafeLoad<uint32_t>(in + 4) << 11, SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4) };
-  shifts = simd_batch{ 8, 0, 2, 15 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 13-bit bundles 12 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 4) >> 28 | SafeLoad<uint32_t>(in + 5) << 4, SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) >> 22 | SafeLoad<uint32_t>(in + 6) << 10, SafeLoad<uint32_t>(in + 6) };
-  shifts = simd_batch{ 0, 9, 0, 3 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 13-bit bundles 16 to 19
-  words = simd_batch{ SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6) >> 29 | SafeLoad<uint32_t>(in + 7) << 3, SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7) >> 23 | SafeLoad<uint32_t>(in + 8) << 9 };
-  shifts = simd_batch{ 16, 0, 10, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 13-bit bundles 20 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 8) >> 30 | SafeLoad<uint32_t>(in + 9) << 2, SafeLoad<uint32_t>(in + 9) };
-  shifts = simd_batch{ 4, 17, 0, 11 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 13-bit bundles 24 to 27
-  words = simd_batch{ SafeLoad<uint32_t>(in + 9) >> 24 | SafeLoad<uint32_t>(in + 10) << 8, SafeLoad<uint32_t>(in + 10), SafeLoad<uint32_t>(in + 10), SafeLoad<uint32_t>(in + 10) >> 31 | SafeLoad<uint32_t>(in + 11) << 1 };
-  shifts = simd_batch{ 0, 5, 18, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 13-bit bundles 28 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 11), SafeLoad<uint32_t>(in + 11) >> 25 | SafeLoad<uint32_t>(in + 12) << 7, SafeLoad<uint32_t>(in + 12), SafeLoad<uint32_t>(in + 12) };
-  shifts = simd_batch{ 12, 0, 6, 19 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  in += 13;
-  return in;
-}
-
-inline static const uint32_t* unpack14_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x3fff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 14-bit bundles 0 to 3
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 28 | SafeLoad<uint32_t>(in + 1) << 4, SafeLoad<uint32_t>(in + 1) };
-  shifts = simd_batch{ 0, 14, 0, 10 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 14-bit bundles 4 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 1) >> 24 | SafeLoad<uint32_t>(in + 2) << 8, SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) >> 20 | SafeLoad<uint32_t>(in + 3) << 12, SafeLoad<uint32_t>(in + 3) };
-  shifts = simd_batch{ 0, 6, 0, 2 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 14-bit bundles 8 to 11
-  words = simd_batch{ SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) >> 30 | SafeLoad<uint32_t>(in + 4) << 2, SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4) >> 26 | SafeLoad<uint32_t>(in + 5) << 6 };
-  shifts = simd_batch{ 16, 0, 12, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 14-bit bundles 12 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) >> 22 | SafeLoad<uint32_t>(in + 6) << 10, SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6) };
-  shifts = simd_batch{ 8, 0, 4, 18 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 14-bit bundles 16 to 19
-  words = simd_batch{ SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7) >> 28 | SafeLoad<uint32_t>(in + 8) << 4, SafeLoad<uint32_t>(in + 8) };
-  shifts = simd_batch{ 0, 14, 0, 10 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 14-bit bundles 20 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 8) >> 24 | SafeLoad<uint32_t>(in + 9) << 8, SafeLoad<uint32_t>(in + 9), SafeLoad<uint32_t>(in + 9) >> 20 | SafeLoad<uint32_t>(in + 10) << 12, SafeLoad<uint32_t>(in + 10) };
-  shifts = simd_batch{ 0, 6, 0, 2 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 14-bit bundles 24 to 27
-  words = simd_batch{ SafeLoad<uint32_t>(in + 10), SafeLoad<uint32_t>(in + 10) >> 30 | SafeLoad<uint32_t>(in + 11) << 2, SafeLoad<uint32_t>(in + 11), SafeLoad<uint32_t>(in + 11) >> 26 | SafeLoad<uint32_t>(in + 12) << 6 };
-  shifts = simd_batch{ 16, 0, 12, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 14-bit bundles 28 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 12), SafeLoad<uint32_t>(in + 12) >> 22 | SafeLoad<uint32_t>(in + 13) << 10, SafeLoad<uint32_t>(in + 13), SafeLoad<uint32_t>(in + 13) };
-  shifts = simd_batch{ 8, 0, 4, 18 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  in += 14;
-  return in;
-}
-
-inline static const uint32_t* unpack15_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x7fff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 15-bit bundles 0 to 3
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 30 | SafeLoad<uint32_t>(in + 1) << 2, SafeLoad<uint32_t>(in + 1) };
-  shifts = simd_batch{ 0, 15, 0, 13 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 15-bit bundles 4 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 1) >> 28 | SafeLoad<uint32_t>(in + 2) << 4, SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) >> 26 | SafeLoad<uint32_t>(in + 3) << 6, SafeLoad<uint32_t>(in + 3) };
-  shifts = simd_batch{ 0, 11, 0, 9 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 15-bit bundles 8 to 11
-  words = simd_batch{ SafeLoad<uint32_t>(in + 3) >> 24 | SafeLoad<uint32_t>(in + 4) << 8, SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4) >> 22 | SafeLoad<uint32_t>(in + 5) << 10, SafeLoad<uint32_t>(in + 5) };
-  shifts = simd_batch{ 0, 7, 0, 5 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 15-bit bundles 12 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 5) >> 20 | SafeLoad<uint32_t>(in + 6) << 12, SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6) >> 18 | SafeLoad<uint32_t>(in + 7) << 14, SafeLoad<uint32_t>(in + 7) };
-  shifts = simd_batch{ 0, 3, 0, 1 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 15-bit bundles 16 to 19
-  words = simd_batch{ SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7) >> 31 | SafeLoad<uint32_t>(in + 8) << 1, SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 8) >> 29 | SafeLoad<uint32_t>(in + 9) << 3 };
-  shifts = simd_batch{ 16, 0, 14, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 15-bit bundles 20 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 9), SafeLoad<uint32_t>(in + 9) >> 27 | SafeLoad<uint32_t>(in + 10) << 5, SafeLoad<uint32_t>(in + 10), SafeLoad<uint32_t>(in + 10) >> 25 | SafeLoad<uint32_t>(in + 11) << 7 };
-  shifts = simd_batch{ 12, 0, 10, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 15-bit bundles 24 to 27
-  words = simd_batch{ SafeLoad<uint32_t>(in + 11), SafeLoad<uint32_t>(in + 11) >> 23 | SafeLoad<uint32_t>(in + 12) << 9, SafeLoad<uint32_t>(in + 12), SafeLoad<uint32_t>(in + 12) >> 21 | SafeLoad<uint32_t>(in + 13) << 11 };
-  shifts = simd_batch{ 8, 0, 6, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 15-bit bundles 28 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 13), SafeLoad<uint32_t>(in + 13) >> 19 | SafeLoad<uint32_t>(in + 14) << 13, SafeLoad<uint32_t>(in + 14), SafeLoad<uint32_t>(in + 14) };
-  shifts = simd_batch{ 4, 0, 2, 17 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  in += 15;
-  return in;
-}
-
-inline static const uint32_t* unpack16_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0xffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 16-bit bundles 0 to 3
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) };
-  shifts = simd_batch{ 0, 16, 0, 16 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 16-bit bundles 4 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) };
-  shifts = simd_batch{ 0, 16, 0, 16 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 16-bit bundles 8 to 11
-  words = simd_batch{ SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) };
-  shifts = simd_batch{ 0, 16, 0, 16 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 16-bit bundles 12 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7) };
-  shifts = simd_batch{ 0, 16, 0, 16 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 16-bit bundles 16 to 19
-  words = simd_batch{ SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 9), SafeLoad<uint32_t>(in + 9) };
-  shifts = simd_batch{ 0, 16, 0, 16 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 16-bit bundles 20 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 10), SafeLoad<uint32_t>(in + 10), SafeLoad<uint32_t>(in + 11), SafeLoad<uint32_t>(in + 11) };
-  shifts = simd_batch{ 0, 16, 0, 16 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 16-bit bundles 24 to 27
-  words = simd_batch{ SafeLoad<uint32_t>(in + 12), SafeLoad<uint32_t>(in + 12), SafeLoad<uint32_t>(in + 13), SafeLoad<uint32_t>(in + 13) };
-  shifts = simd_batch{ 0, 16, 0, 16 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 16-bit bundles 28 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 14), SafeLoad<uint32_t>(in + 14), SafeLoad<uint32_t>(in + 15), SafeLoad<uint32_t>(in + 15) };
-  shifts = simd_batch{ 0, 16, 0, 16 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  in += 16;
-  return in;
-}
-
-inline static const uint32_t* unpack17_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x1ffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 17-bit bundles 0 to 3
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 17 | SafeLoad<uint32_t>(in + 1) << 15, SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) >> 19 | SafeLoad<uint32_t>(in + 2) << 13 };
-  shifts = simd_batch{ 0, 0, 2, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 17-bit bundles 4 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) >> 21 | SafeLoad<uint32_t>(in + 3) << 11, SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) >> 23 | SafeLoad<uint32_t>(in + 4) << 9 };
-  shifts = simd_batch{ 4, 0, 6, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 17-bit bundles 8 to 11
-  words = simd_batch{ SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4) >> 25 | SafeLoad<uint32_t>(in + 5) << 7, SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) >> 27 | SafeLoad<uint32_t>(in + 6) << 5 };
-  shifts = simd_batch{ 8, 0, 10, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 17-bit bundles 12 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6) >> 29 | SafeLoad<uint32_t>(in + 7) << 3, SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7) >> 31 | SafeLoad<uint32_t>(in + 8) << 1 };
-  shifts = simd_batch{ 12, 0, 14, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 17-bit bundles 16 to 19
-  words = simd_batch{ SafeLoad<uint32_t>(in + 8) >> 16 | SafeLoad<uint32_t>(in + 9) << 16, SafeLoad<uint32_t>(in + 9), SafeLoad<uint32_t>(in + 9) >> 18 | SafeLoad<uint32_t>(in + 10) << 14, SafeLoad<uint32_t>(in + 10) };
-  shifts = simd_batch{ 0, 1, 0, 3 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 17-bit bundles 20 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 10) >> 20 | SafeLoad<uint32_t>(in + 11) << 12, SafeLoad<uint32_t>(in + 11), SafeLoad<uint32_t>(in + 11) >> 22 | SafeLoad<uint32_t>(in + 12) << 10, SafeLoad<uint32_t>(in + 12) };
-  shifts = simd_batch{ 0, 5, 0, 7 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 17-bit bundles 24 to 27
-  words = simd_batch{ SafeLoad<uint32_t>(in + 12) >> 24 | SafeLoad<uint32_t>(in + 13) << 8, SafeLoad<uint32_t>(in + 13), SafeLoad<uint32_t>(in + 13) >> 26 | SafeLoad<uint32_t>(in + 14) << 6, SafeLoad<uint32_t>(in + 14) };
-  shifts = simd_batch{ 0, 9, 0, 11 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 17-bit bundles 28 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 14) >> 28 | SafeLoad<uint32_t>(in + 15) << 4, SafeLoad<uint32_t>(in + 15), SafeLoad<uint32_t>(in + 15) >> 30 | SafeLoad<uint32_t>(in + 16) << 2, SafeLoad<uint32_t>(in + 16) };
-  shifts = simd_batch{ 0, 13, 0, 15 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  in += 17;
-  return in;
-}
-
-inline static const uint32_t* unpack18_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x3ffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 18-bit bundles 0 to 3
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 18 | SafeLoad<uint32_t>(in + 1) << 14, SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) >> 22 | SafeLoad<uint32_t>(in + 2) << 10 };
-  shifts = simd_batch{ 0, 0, 4, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 18-bit bundles 4 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) >> 26 | SafeLoad<uint32_t>(in + 3) << 6, SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) >> 30 | SafeLoad<uint32_t>(in + 4) << 2 };
-  shifts = simd_batch{ 8, 0, 12, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 18-bit bundles 8 to 11
-  words = simd_batch{ SafeLoad<uint32_t>(in + 4) >> 16 | SafeLoad<uint32_t>(in + 5) << 16, SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) >> 20 | SafeLoad<uint32_t>(in + 6) << 12, SafeLoad<uint32_t>(in + 6) };
-  shifts = simd_batch{ 0, 2, 0, 6 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 18-bit bundles 12 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 6) >> 24 | SafeLoad<uint32_t>(in + 7) << 8, SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7) >> 28 | SafeLoad<uint32_t>(in + 8) << 4, SafeLoad<uint32_t>(in + 8) };
-  shifts = simd_batch{ 0, 10, 0, 14 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 18-bit bundles 16 to 19
-  words = simd_batch{ SafeLoad<uint32_t>(in + 9), SafeLoad<uint32_t>(in + 9) >> 18 | SafeLoad<uint32_t>(in + 10) << 14, SafeLoad<uint32_t>(in + 10), SafeLoad<uint32_t>(in + 10) >> 22 | SafeLoad<uint32_t>(in + 11) << 10 };
-  shifts = simd_batch{ 0, 0, 4, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 18-bit bundles 20 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 11), SafeLoad<uint32_t>(in + 11) >> 26 | SafeLoad<uint32_t>(in + 12) << 6, SafeLoad<uint32_t>(in + 12), SafeLoad<uint32_t>(in + 12) >> 30 | SafeLoad<uint32_t>(in + 13) << 2 };
-  shifts = simd_batch{ 8, 0, 12, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 18-bit bundles 24 to 27
-  words = simd_batch{ SafeLoad<uint32_t>(in + 13) >> 16 | SafeLoad<uint32_t>(in + 14) << 16, SafeLoad<uint32_t>(in + 14), SafeLoad<uint32_t>(in + 14) >> 20 | SafeLoad<uint32_t>(in + 15) << 12, SafeLoad<uint32_t>(in + 15) };
-  shifts = simd_batch{ 0, 2, 0, 6 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 18-bit bundles 28 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 15) >> 24 | SafeLoad<uint32_t>(in + 16) << 8, SafeLoad<uint32_t>(in + 16), SafeLoad<uint32_t>(in + 16) >> 28 | SafeLoad<uint32_t>(in + 17) << 4, SafeLoad<uint32_t>(in + 17) };
-  shifts = simd_batch{ 0, 10, 0, 14 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  in += 18;
-  return in;
-}
-
-inline static const uint32_t* unpack19_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x7ffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 19-bit bundles 0 to 3
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 19 | SafeLoad<uint32_t>(in + 1) << 13, SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) >> 25 | SafeLoad<uint32_t>(in + 2) << 7 };
-  shifts = simd_batch{ 0, 0, 6, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 19-bit bundles 4 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) >> 31 | SafeLoad<uint32_t>(in + 3) << 1, SafeLoad<uint32_t>(in + 3) >> 18 | SafeLoad<uint32_t>(in + 4) << 14, SafeLoad<uint32_t>(in + 4) };
-  shifts = simd_batch{ 12, 0, 0, 5 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 19-bit bundles 8 to 11
-  words = simd_batch{ SafeLoad<uint32_t>(in + 4) >> 24 | SafeLoad<uint32_t>(in + 5) << 8, SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) >> 30 | SafeLoad<uint32_t>(in + 6) << 2, SafeLoad<uint32_t>(in + 6) >> 17 | SafeLoad<uint32_t>(in + 7) << 15 };
-  shifts = simd_batch{ 0, 11, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 19-bit bundles 12 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7) >> 23 | SafeLoad<uint32_t>(in + 8) << 9, SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 8) >> 29 | SafeLoad<uint32_t>(in + 9) << 3 };
-  shifts = simd_batch{ 4, 0, 10, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 19-bit bundles 16 to 19
-  words = simd_batch{ SafeLoad<uint32_t>(in + 9) >> 16 | SafeLoad<uint32_t>(in + 10) << 16, SafeLoad<uint32_t>(in + 10), SafeLoad<uint32_t>(in + 10) >> 22 | SafeLoad<uint32_t>(in + 11) << 10, SafeLoad<uint32_t>(in + 11) };
-  shifts = simd_batch{ 0, 3, 0, 9 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 19-bit bundles 20 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 11) >> 28 | SafeLoad<uint32_t>(in + 12) << 4, SafeLoad<uint32_t>(in + 12) >> 15 | SafeLoad<uint32_t>(in + 13) << 17, SafeLoad<uint32_t>(in + 13), SafeLoad<uint32_t>(in + 13) >> 21 | SafeLoad<uint32_t>(in + 14) << 11 };
-  shifts = simd_batch{ 0, 0, 2, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 19-bit bundles 24 to 27
-  words = simd_batch{ SafeLoad<uint32_t>(in + 14), SafeLoad<uint32_t>(in + 14) >> 27 | SafeLoad<uint32_t>(in + 15) << 5, SafeLoad<uint32_t>(in + 15) >> 14 | SafeLoad<uint32_t>(in + 16) << 18, SafeLoad<uint32_t>(in + 16) };
-  shifts = simd_batch{ 8, 0, 0, 1 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 19-bit bundles 28 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 16) >> 20 | SafeLoad<uint32_t>(in + 17) << 12, SafeLoad<uint32_t>(in + 17), SafeLoad<uint32_t>(in + 17) >> 26 | SafeLoad<uint32_t>(in + 18) << 6, SafeLoad<uint32_t>(in + 18) };
-  shifts = simd_batch{ 0, 7, 0, 13 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  in += 19;
-  return in;
-}
-
-inline static const uint32_t* unpack20_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0xfffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 20-bit bundles 0 to 3
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 20 | SafeLoad<uint32_t>(in + 1) << 12, SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) >> 28 | SafeLoad<uint32_t>(in + 2) << 4 };
-  shifts = simd_batch{ 0, 0, 8, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 20-bit bundles 4 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 2) >> 16 | SafeLoad<uint32_t>(in + 3) << 16, SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) >> 24 | SafeLoad<uint32_t>(in + 4) << 8, SafeLoad<uint32_t>(in + 4) };
-  shifts = simd_batch{ 0, 4, 0, 12 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 20-bit bundles 8 to 11
-  words = simd_batch{ SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) >> 20 | SafeLoad<uint32_t>(in + 6) << 12, SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6) >> 28 | SafeLoad<uint32_t>(in + 7) << 4 };
-  shifts = simd_batch{ 0, 0, 8, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 20-bit bundles 12 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 7) >> 16 | SafeLoad<uint32_t>(in + 8) << 16, SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 8) >> 24 | SafeLoad<uint32_t>(in + 9) << 8, SafeLoad<uint32_t>(in + 9) };
-  shifts = simd_batch{ 0, 4, 0, 12 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 20-bit bundles 16 to 19
-  words = simd_batch{ SafeLoad<uint32_t>(in + 10), SafeLoad<uint32_t>(in + 10) >> 20 | SafeLoad<uint32_t>(in + 11) << 12, SafeLoad<uint32_t>(in + 11), SafeLoad<uint32_t>(in + 11) >> 28 | SafeLoad<uint32_t>(in + 12) << 4 };
-  shifts = simd_batch{ 0, 0, 8, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 20-bit bundles 20 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 12) >> 16 | SafeLoad<uint32_t>(in + 13) << 16, SafeLoad<uint32_t>(in + 13), SafeLoad<uint32_t>(in + 13) >> 24 | SafeLoad<uint32_t>(in + 14) << 8, SafeLoad<uint32_t>(in + 14) };
-  shifts = simd_batch{ 0, 4, 0, 12 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 20-bit bundles 24 to 27
-  words = simd_batch{ SafeLoad<uint32_t>(in + 15), SafeLoad<uint32_t>(in + 15) >> 20 | SafeLoad<uint32_t>(in + 16) << 12, SafeLoad<uint32_t>(in + 16), SafeLoad<uint32_t>(in + 16) >> 28 | SafeLoad<uint32_t>(in + 17) << 4 };
-  shifts = simd_batch{ 0, 0, 8, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 20-bit bundles 28 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 17) >> 16 | SafeLoad<uint32_t>(in + 18) << 16, SafeLoad<uint32_t>(in + 18), SafeLoad<uint32_t>(in + 18) >> 24 | SafeLoad<uint32_t>(in + 19) << 8, SafeLoad<uint32_t>(in + 19) };
-  shifts = simd_batch{ 0, 4, 0, 12 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  in += 20;
-  return in;
-}
-
-inline static const uint32_t* unpack21_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x1fffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 21-bit bundles 0 to 3
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 21 | SafeLoad<uint32_t>(in + 1) << 11, SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) >> 31 | SafeLoad<uint32_t>(in + 2) << 1 };
-  shifts = simd_batch{ 0, 0, 10, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 21-bit bundles 4 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 2) >> 20 | SafeLoad<uint32_t>(in + 3) << 12, SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) >> 30 | SafeLoad<uint32_t>(in + 4) << 2, SafeLoad<uint32_t>(in + 4) >> 19 | SafeLoad<uint32_t>(in + 5) << 13 };
-  shifts = simd_batch{ 0, 9, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 21-bit bundles 8 to 11
-  words = simd_batch{ SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) >> 29 | SafeLoad<uint32_t>(in + 6) << 3, SafeLoad<uint32_t>(in + 6) >> 18 | SafeLoad<uint32_t>(in + 7) << 14, SafeLoad<uint32_t>(in + 7) };
-  shifts = simd_batch{ 8, 0, 0, 7 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 21-bit bundles 12 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 7) >> 28 | SafeLoad<uint32_t>(in + 8) << 4, SafeLoad<uint32_t>(in + 8) >> 17 | SafeLoad<uint32_t>(in + 9) << 15, SafeLoad<uint32_t>(in + 9), SafeLoad<uint32_t>(in + 9) >> 27 | SafeLoad<uint32_t>(in + 10) << 5 };
-  shifts = simd_batch{ 0, 0, 6, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 21-bit bundles 16 to 19
-  words = simd_batch{ SafeLoad<uint32_t>(in + 10) >> 16 | SafeLoad<uint32_t>(in + 11) << 16, SafeLoad<uint32_t>(in + 11), SafeLoad<uint32_t>(in + 11) >> 26 | SafeLoad<uint32_t>(in + 12) << 6, SafeLoad<uint32_t>(in + 12) >> 15 | SafeLoad<uint32_t>(in + 13) << 17 };
-  shifts = simd_batch{ 0, 5, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 21-bit bundles 20 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 13), SafeLoad<uint32_t>(in + 13) >> 25 | SafeLoad<uint32_t>(in + 14) << 7, SafeLoad<uint32_t>(in + 14) >> 14 | SafeLoad<uint32_t>(in + 15) << 18, SafeLoad<uint32_t>(in + 15) };
-  shifts = simd_batch{ 4, 0, 0, 3 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 21-bit bundles 24 to 27
-  words = simd_batch{ SafeLoad<uint32_t>(in + 15) >> 24 | SafeLoad<uint32_t>(in + 16) << 8, SafeLoad<uint32_t>(in + 16) >> 13 | SafeLoad<uint32_t>(in + 17) << 19, SafeLoad<uint32_t>(in + 17), SafeLoad<uint32_t>(in + 17) >> 23 | SafeLoad<uint32_t>(in + 18) << 9 };
-  shifts = simd_batch{ 0, 0, 2, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 21-bit bundles 28 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 18) >> 12 | SafeLoad<uint32_t>(in + 19) << 20, SafeLoad<uint32_t>(in + 19), SafeLoad<uint32_t>(in + 19) >> 22 | SafeLoad<uint32_t>(in + 20) << 10, SafeLoad<uint32_t>(in + 20) };
-  shifts = simd_batch{ 0, 1, 0, 11 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  in += 21;
-  return in;
-}
-
-inline static const uint32_t* unpack22_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x3fffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 22-bit bundles 0 to 3
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 22 | SafeLoad<uint32_t>(in + 1) << 10, SafeLoad<uint32_t>(in + 1) >> 12 | SafeLoad<uint32_t>(in + 2) << 20, SafeLoad<uint32_t>(in + 2) };
-  shifts = simd_batch{ 0, 0, 0, 2 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 22-bit bundles 4 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 2) >> 24 | SafeLoad<uint32_t>(in + 3) << 8, SafeLoad<uint32_t>(in + 3) >> 14 | SafeLoad<uint32_t>(in + 4) << 18, SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4) >> 26 | SafeLoad<uint32_t>(in + 5) << 6 };
-  shifts = simd_batch{ 0, 0, 4, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 22-bit bundles 8 to 11
-  words = simd_batch{ SafeLoad<uint32_t>(in + 5) >> 16 | SafeLoad<uint32_t>(in + 6) << 16, SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6) >> 28 | SafeLoad<uint32_t>(in + 7) << 4, SafeLoad<uint32_t>(in + 7) >> 18 | SafeLoad<uint32_t>(in + 8) << 14 };
-  shifts = simd_batch{ 0, 6, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 22-bit bundles 12 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 8) >> 30 | SafeLoad<uint32_t>(in + 9) << 2, SafeLoad<uint32_t>(in + 9) >> 20 | SafeLoad<uint32_t>(in + 10) << 12, SafeLoad<uint32_t>(in + 10) };
-  shifts = simd_batch{ 8, 0, 0, 10 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 22-bit bundles 16 to 19
-  words = simd_batch{ SafeLoad<uint32_t>(in + 11), SafeLoad<uint32_t>(in + 11) >> 22 | SafeLoad<uint32_t>(in + 12) << 10, SafeLoad<uint32_t>(in + 12) >> 12 | SafeLoad<uint32_t>(in + 13) << 20, SafeLoad<uint32_t>(in + 13) };
-  shifts = simd_batch{ 0, 0, 0, 2 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 22-bit bundles 20 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 13) >> 24 | SafeLoad<uint32_t>(in + 14) << 8, SafeLoad<uint32_t>(in + 14) >> 14 | SafeLoad<uint32_t>(in + 15) << 18, SafeLoad<uint32_t>(in + 15), SafeLoad<uint32_t>(in + 15) >> 26 | SafeLoad<uint32_t>(in + 16) << 6 };
-  shifts = simd_batch{ 0, 0, 4, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 22-bit bundles 24 to 27
-  words = simd_batch{ SafeLoad<uint32_t>(in + 16) >> 16 | SafeLoad<uint32_t>(in + 17) << 16, SafeLoad<uint32_t>(in + 17), SafeLoad<uint32_t>(in + 17) >> 28 | SafeLoad<uint32_t>(in + 18) << 4, SafeLoad<uint32_t>(in + 18) >> 18 | SafeLoad<uint32_t>(in + 19) << 14 };
-  shifts = simd_batch{ 0, 6, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 22-bit bundles 28 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 19), SafeLoad<uint32_t>(in + 19) >> 30 | SafeLoad<uint32_t>(in + 20) << 2, SafeLoad<uint32_t>(in + 20) >> 20 | SafeLoad<uint32_t>(in + 21) << 12, SafeLoad<uint32_t>(in + 21) };
-  shifts = simd_batch{ 8, 0, 0, 10 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  in += 22;
-  return in;
-}
-
-inline static const uint32_t* unpack23_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x7fffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 23-bit bundles 0 to 3
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 23 | SafeLoad<uint32_t>(in + 1) << 9, SafeLoad<uint32_t>(in + 1) >> 14 | SafeLoad<uint32_t>(in + 2) << 18, SafeLoad<uint32_t>(in + 2) };
-  shifts = simd_batch{ 0, 0, 0, 5 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 23-bit bundles 4 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 2) >> 28 | SafeLoad<uint32_t>(in + 3) << 4, SafeLoad<uint32_t>(in + 3) >> 19 | SafeLoad<uint32_t>(in + 4) << 13, SafeLoad<uint32_t>(in + 4) >> 10 | SafeLoad<uint32_t>(in + 5) << 22, SafeLoad<uint32_t>(in + 5) };
-  shifts = simd_batch{ 0, 0, 0, 1 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 23-bit bundles 8 to 11
-  words = simd_batch{ SafeLoad<uint32_t>(in + 5) >> 24 | SafeLoad<uint32_t>(in + 6) << 8, SafeLoad<uint32_t>(in + 6) >> 15 | SafeLoad<uint32_t>(in + 7) << 17, SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7) >> 29 | SafeLoad<uint32_t>(in + 8) << 3 };
-  shifts = simd_batch{ 0, 0, 6, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 23-bit bundles 12 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 8) >> 20 | SafeLoad<uint32_t>(in + 9) << 12, SafeLoad<uint32_t>(in + 9) >> 11 | SafeLoad<uint32_t>(in + 10) << 21, SafeLoad<uint32_t>(in + 10), SafeLoad<uint32_t>(in + 10) >> 25 | SafeLoad<uint32_t>(in + 11) << 7 };
-  shifts = simd_batch{ 0, 0, 2, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 23-bit bundles 16 to 19
-  words = simd_batch{ SafeLoad<uint32_t>(in + 11) >> 16 | SafeLoad<uint32_t>(in + 12) << 16, SafeLoad<uint32_t>(in + 12), SafeLoad<uint32_t>(in + 12) >> 30 | SafeLoad<uint32_t>(in + 13) << 2, SafeLoad<uint32_t>(in + 13) >> 21 | SafeLoad<uint32_t>(in + 14) << 11 };
-  shifts = simd_batch{ 0, 7, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 23-bit bundles 20 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 14) >> 12 | SafeLoad<uint32_t>(in + 15) << 20, SafeLoad<uint32_t>(in + 15), SafeLoad<uint32_t>(in + 15) >> 26 | SafeLoad<uint32_t>(in + 16) << 6, SafeLoad<uint32_t>(in + 16) >> 17 | SafeLoad<uint32_t>(in + 17) << 15 };
-  shifts = simd_batch{ 0, 3, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 23-bit bundles 24 to 27
-  words = simd_batch{ SafeLoad<uint32_t>(in + 17), SafeLoad<uint32_t>(in + 17) >> 31 | SafeLoad<uint32_t>(in + 18) << 1, SafeLoad<uint32_t>(in + 18) >> 22 | SafeLoad<uint32_t>(in + 19) << 10, SafeLoad<uint32_t>(in + 19) >> 13 | SafeLoad<uint32_t>(in + 20) << 19 };
-  shifts = simd_batch{ 8, 0, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 23-bit bundles 28 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 20), SafeLoad<uint32_t>(in + 20) >> 27 | SafeLoad<uint32_t>(in + 21) << 5, SafeLoad<uint32_t>(in + 21) >> 18 | SafeLoad<uint32_t>(in + 22) << 14, SafeLoad<uint32_t>(in + 22) };
-  shifts = simd_batch{ 4, 0, 0, 9 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  in += 23;
-  return in;
-}
-
-inline static const uint32_t* unpack24_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0xffffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 24-bit bundles 0 to 3
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 24 | SafeLoad<uint32_t>(in + 1) << 8, SafeLoad<uint32_t>(in + 1) >> 16 | SafeLoad<uint32_t>(in + 2) << 16, SafeLoad<uint32_t>(in + 2) };
-  shifts = simd_batch{ 0, 0, 0, 8 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 24-bit bundles 4 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) >> 24 | SafeLoad<uint32_t>(in + 4) << 8, SafeLoad<uint32_t>(in + 4) >> 16 | SafeLoad<uint32_t>(in + 5) << 16, SafeLoad<uint32_t>(in + 5) };
-  shifts = simd_batch{ 0, 0, 0, 8 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 24-bit bundles 8 to 11
-  words = simd_batch{ SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6) >> 24 | SafeLoad<uint32_t>(in + 7) << 8, SafeLoad<uint32_t>(in + 7) >> 16 | SafeLoad<uint32_t>(in + 8) << 16, SafeLoad<uint32_t>(in + 8) };
-  shifts = simd_batch{ 0, 0, 0, 8 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 24-bit bundles 12 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 9), SafeLoad<uint32_t>(in + 9) >> 24 | SafeLoad<uint32_t>(in + 10) << 8, SafeLoad<uint32_t>(in + 10) >> 16 | SafeLoad<uint32_t>(in + 11) << 16, SafeLoad<uint32_t>(in + 11) };
-  shifts = simd_batch{ 0, 0, 0, 8 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 24-bit bundles 16 to 19
-  words = simd_batch{ SafeLoad<uint32_t>(in + 12), SafeLoad<uint32_t>(in + 12) >> 24 | SafeLoad<uint32_t>(in + 13) << 8, SafeLoad<uint32_t>(in + 13) >> 16 | SafeLoad<uint32_t>(in + 14) << 16, SafeLoad<uint32_t>(in + 14) };
-  shifts = simd_batch{ 0, 0, 0, 8 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 24-bit bundles 20 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 15), SafeLoad<uint32_t>(in + 15) >> 24 | SafeLoad<uint32_t>(in + 16) << 8, SafeLoad<uint32_t>(in + 16) >> 16 | SafeLoad<uint32_t>(in + 17) << 16, SafeLoad<uint32_t>(in + 17) };
-  shifts = simd_batch{ 0, 0, 0, 8 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 24-bit bundles 24 to 27
-  words = simd_batch{ SafeLoad<uint32_t>(in + 18), SafeLoad<uint32_t>(in + 18) >> 24 | SafeLoad<uint32_t>(in + 19) << 8, SafeLoad<uint32_t>(in + 19) >> 16 | SafeLoad<uint32_t>(in + 20) << 16, SafeLoad<uint32_t>(in + 20) };
-  shifts = simd_batch{ 0, 0, 0, 8 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 24-bit bundles 28 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 21), SafeLoad<uint32_t>(in + 21) >> 24 | SafeLoad<uint32_t>(in + 22) << 8, SafeLoad<uint32_t>(in + 22) >> 16 | SafeLoad<uint32_t>(in + 23) << 16, SafeLoad<uint32_t>(in + 23) };
-  shifts = simd_batch{ 0, 0, 0, 8 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  in += 24;
-  return in;
-}
-
-inline static const uint32_t* unpack25_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x1ffffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 25-bit bundles 0 to 3
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 25 | SafeLoad<uint32_t>(in + 1) << 7, SafeLoad<uint32_t>(in + 1) >> 18 | SafeLoad<uint32_t>(in + 2) << 14, SafeLoad<uint32_t>(in + 2) >> 11 | SafeLoad<uint32_t>(in + 3) << 21 };
-  shifts = simd_batch{ 0, 0, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 25-bit bundles 4 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) >> 29 | SafeLoad<uint32_t>(in + 4) << 3, SafeLoad<uint32_t>(in + 4) >> 22 | SafeLoad<uint32_t>(in + 5) << 10, SafeLoad<uint32_t>(in + 5) >> 15 | SafeLoad<uint32_t>(in + 6) << 17 };
-  shifts = simd_batch{ 4, 0, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 25-bit bundles 8 to 11
-  words = simd_batch{ SafeLoad<uint32_t>(in + 6) >> 8 | SafeLoad<uint32_t>(in + 7) << 24, SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7) >> 26 | SafeLoad<uint32_t>(in + 8) << 6, SafeLoad<uint32_t>(in + 8) >> 19 | SafeLoad<uint32_t>(in + 9) << 13 };
-  shifts = simd_batch{ 0, 1, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 25-bit bundles 12 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 9) >> 12 | SafeLoad<uint32_t>(in + 10) << 20, SafeLoad<uint32_t>(in + 10), SafeLoad<uint32_t>(in + 10) >> 30 | SafeLoad<uint32_t>(in + 11) << 2, SafeLoad<uint32_t>(in + 11) >> 23 | SafeLoad<uint32_t>(in + 12) << 9 };
-  shifts = simd_batch{ 0, 5, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 25-bit bundles 16 to 19
-  words = simd_batch{ SafeLoad<uint32_t>(in + 12) >> 16 | SafeLoad<uint32_t>(in + 13) << 16, SafeLoad<uint32_t>(in + 13) >> 9 | SafeLoad<uint32_t>(in + 14) << 23, SafeLoad<uint32_t>(in + 14), SafeLoad<uint32_t>(in + 14) >> 27 | SafeLoad<uint32_t>(in + 15) << 5 };
-  shifts = simd_batch{ 0, 0, 2, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 25-bit bundles 20 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 15) >> 20 | SafeLoad<uint32_t>(in + 16) << 12, SafeLoad<uint32_t>(in + 16) >> 13 | SafeLoad<uint32_t>(in + 17) << 19, SafeLoad<uint32_t>(in + 17), SafeLoad<uint32_t>(in + 17) >> 31 | SafeLoad<uint32_t>(in + 18) << 1 };
-  shifts = simd_batch{ 0, 0, 6, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 25-bit bundles 24 to 27
-  words = simd_batch{ SafeLoad<uint32_t>(in + 18) >> 24 | SafeLoad<uint32_t>(in + 19) << 8, SafeLoad<uint32_t>(in + 19) >> 17 | SafeLoad<uint32_t>(in + 20) << 15, SafeLoad<uint32_t>(in + 20) >> 10 | SafeLoad<uint32_t>(in + 21) << 22, SafeLoad<uint32_t>(in + 21) };
-  shifts = simd_batch{ 0, 0, 0, 3 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 25-bit bundles 28 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 21) >> 28 | SafeLoad<uint32_t>(in + 22) << 4, SafeLoad<uint32_t>(in + 22) >> 21 | SafeLoad<uint32_t>(in + 23) << 11, SafeLoad<uint32_t>(in + 23) >> 14 | SafeLoad<uint32_t>(in + 24) << 18, SafeLoad<uint32_t>(in + 24) };
-  shifts = simd_batch{ 0, 0, 0, 7 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  in += 25;
-  return in;
-}
-
-inline static const uint32_t* unpack26_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x3ffffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 26-bit bundles 0 to 3
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 26 | SafeLoad<uint32_t>(in + 1) << 6, SafeLoad<uint32_t>(in + 1) >> 20 | SafeLoad<uint32_t>(in + 2) << 12, SafeLoad<uint32_t>(in + 2) >> 14 | SafeLoad<uint32_t>(in + 3) << 18 };
-  shifts = simd_batch{ 0, 0, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 26-bit bundles 4 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 3) >> 8 | SafeLoad<uint32_t>(in + 4) << 24, SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4) >> 28 | SafeLoad<uint32_t>(in + 5) << 4, SafeLoad<uint32_t>(in + 5) >> 22 | SafeLoad<uint32_t>(in + 6) << 10 };
-  shifts = simd_batch{ 0, 2, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 26-bit bundles 8 to 11
-  words = simd_batch{ SafeLoad<uint32_t>(in + 6) >> 16 | SafeLoad<uint32_t>(in + 7) << 16, SafeLoad<uint32_t>(in + 7) >> 10 | SafeLoad<uint32_t>(in + 8) << 22, SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 8) >> 30 | SafeLoad<uint32_t>(in + 9) << 2 };
-  shifts = simd_batch{ 0, 0, 4, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 26-bit bundles 12 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 9) >> 24 | SafeLoad<uint32_t>(in + 10) << 8, SafeLoad<uint32_t>(in + 10) >> 18 | SafeLoad<uint32_t>(in + 11) << 14, SafeLoad<uint32_t>(in + 11) >> 12 | SafeLoad<uint32_t>(in + 12) << 20, SafeLoad<uint32_t>(in + 12) };
-  shifts = simd_batch{ 0, 0, 0, 6 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 26-bit bundles 16 to 19
-  words = simd_batch{ SafeLoad<uint32_t>(in + 13), SafeLoad<uint32_t>(in + 13) >> 26 | SafeLoad<uint32_t>(in + 14) << 6, SafeLoad<uint32_t>(in + 14) >> 20 | SafeLoad<uint32_t>(in + 15) << 12, SafeLoad<uint32_t>(in + 15) >> 14 | SafeLoad<uint32_t>(in + 16) << 18 };
-  shifts = simd_batch{ 0, 0, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 26-bit bundles 20 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 16) >> 8 | SafeLoad<uint32_t>(in + 17) << 24, SafeLoad<uint32_t>(in + 17), SafeLoad<uint32_t>(in + 17) >> 28 | SafeLoad<uint32_t>(in + 18) << 4, SafeLoad<uint32_t>(in + 18) >> 22 | SafeLoad<uint32_t>(in + 19) << 10 };
-  shifts = simd_batch{ 0, 2, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 26-bit bundles 24 to 27
-  words = simd_batch{ SafeLoad<uint32_t>(in + 19) >> 16 | SafeLoad<uint32_t>(in + 20) << 16, SafeLoad<uint32_t>(in + 20) >> 10 | SafeLoad<uint32_t>(in + 21) << 22, SafeLoad<uint32_t>(in + 21), SafeLoad<uint32_t>(in + 21) >> 30 | SafeLoad<uint32_t>(in + 22) << 2 };
-  shifts = simd_batch{ 0, 0, 4, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 26-bit bundles 28 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 22) >> 24 | SafeLoad<uint32_t>(in + 23) << 8, SafeLoad<uint32_t>(in + 23) >> 18 | SafeLoad<uint32_t>(in + 24) << 14, SafeLoad<uint32_t>(in + 24) >> 12 | SafeLoad<uint32_t>(in + 25) << 20, SafeLoad<uint32_t>(in + 25) };
-  shifts = simd_batch{ 0, 0, 0, 6 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  in += 26;
-  return in;
-}
-
-inline static const uint32_t* unpack27_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x7ffffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 27-bit bundles 0 to 3
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 27 | SafeLoad<uint32_t>(in + 1) << 5, SafeLoad<uint32_t>(in + 1) >> 22 | SafeLoad<uint32_t>(in + 2) << 10, SafeLoad<uint32_t>(in + 2) >> 17 | SafeLoad<uint32_t>(in + 3) << 15 };
-  shifts = simd_batch{ 0, 0, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 27-bit bundles 4 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 3) >> 12 | SafeLoad<uint32_t>(in + 4) << 20, SafeLoad<uint32_t>(in + 4) >> 7 | SafeLoad<uint32_t>(in + 5) << 25, SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) >> 29 | SafeLoad<uint32_t>(in + 6) << 3 };
-  shifts = simd_batch{ 0, 0, 2, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 27-bit bundles 8 to 11
-  words = simd_batch{ SafeLoad<uint32_t>(in + 6) >> 24 | SafeLoad<uint32_t>(in + 7) << 8, SafeLoad<uint32_t>(in + 7) >> 19 | SafeLoad<uint32_t>(in + 8) << 13, SafeLoad<uint32_t>(in + 8) >> 14 | SafeLoad<uint32_t>(in + 9) << 18, SafeLoad<uint32_t>(in + 9) >> 9 | SafeLoad<uint32_t>(in + 10) << 23 };
-  shifts = simd_batch{ 0, 0, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 27-bit bundles 12 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 10), SafeLoad<uint32_t>(in + 10) >> 31 | SafeLoad<uint32_t>(in + 11) << 1, SafeLoad<uint32_t>(in + 11) >> 26 | SafeLoad<uint32_t>(in + 12) << 6, SafeLoad<uint32_t>(in + 12) >> 21 | SafeLoad<uint32_t>(in + 13) << 11 };
-  shifts = simd_batch{ 4, 0, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 27-bit bundles 16 to 19
-  words = simd_batch{ SafeLoad<uint32_t>(in + 13) >> 16 | SafeLoad<uint32_t>(in + 14) << 16, SafeLoad<uint32_t>(in + 14) >> 11 | SafeLoad<uint32_t>(in + 15) << 21, SafeLoad<uint32_t>(in + 15) >> 6 | SafeLoad<uint32_t>(in + 16) << 26, SafeLoad<uint32_t>(in + 16) };
-  shifts = simd_batch{ 0, 0, 0, 1 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 27-bit bundles 20 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 16) >> 28 | SafeLoad<uint32_t>(in + 17) << 4, SafeLoad<uint32_t>(in + 17) >> 23 | SafeLoad<uint32_t>(in + 18) << 9, SafeLoad<uint32_t>(in + 18) >> 18 | SafeLoad<uint32_t>(in + 19) << 14, SafeLoad<uint32_t>(in + 19) >> 13 | SafeLoad<uint32_t>(in + 20) << 19 };
-  shifts = simd_batch{ 0, 0, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 27-bit bundles 24 to 27
-  words = simd_batch{ SafeLoad<uint32_t>(in + 20) >> 8 | SafeLoad<uint32_t>(in + 21) << 24, SafeLoad<uint32_t>(in + 21), SafeLoad<uint32_t>(in + 21) >> 30 | SafeLoad<uint32_t>(in + 22) << 2, SafeLoad<uint32_t>(in + 22) >> 25 | SafeLoad<uint32_t>(in + 23) << 7 };
-  shifts = simd_batch{ 0, 3, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 27-bit bundles 28 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 23) >> 20 | SafeLoad<uint32_t>(in + 24) << 12, SafeLoad<uint32_t>(in + 24) >> 15 | SafeLoad<uint32_t>(in + 25) << 17, SafeLoad<uint32_t>(in + 25) >> 10 | SafeLoad<uint32_t>(in + 26) << 22, SafeLoad<uint32_t>(in + 26) };
-  shifts = simd_batch{ 0, 0, 0, 5 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  in += 27;
-  return in;
-}
-
-inline static const uint32_t* unpack28_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0xfffffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 28-bit bundles 0 to 3
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 28 | SafeLoad<uint32_t>(in + 1) << 4, SafeLoad<uint32_t>(in + 1) >> 24 | SafeLoad<uint32_t>(in + 2) << 8, SafeLoad<uint32_t>(in + 2) >> 20 | SafeLoad<uint32_t>(in + 3) << 12 };
-  shifts = simd_batch{ 0, 0, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 28-bit bundles 4 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 3) >> 16 | SafeLoad<uint32_t>(in + 4) << 16, SafeLoad<uint32_t>(in + 4) >> 12 | SafeLoad<uint32_t>(in + 5) << 20, SafeLoad<uint32_t>(in + 5) >> 8 | SafeLoad<uint32_t>(in + 6) << 24, SafeLoad<uint32_t>(in + 6) };
-  shifts = simd_batch{ 0, 0, 0, 4 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 28-bit bundles 8 to 11
-  words = simd_batch{ SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7) >> 28 | SafeLoad<uint32_t>(in + 8) << 4, SafeLoad<uint32_t>(in + 8) >> 24 | SafeLoad<uint32_t>(in + 9) << 8, SafeLoad<uint32_t>(in + 9) >> 20 | SafeLoad<uint32_t>(in + 10) << 12 };
-  shifts = simd_batch{ 0, 0, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 28-bit bundles 12 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 10) >> 16 | SafeLoad<uint32_t>(in + 11) << 16, SafeLoad<uint32_t>(in + 11) >> 12 | SafeLoad<uint32_t>(in + 12) << 20, SafeLoad<uint32_t>(in + 12) >> 8 | SafeLoad<uint32_t>(in + 13) << 24, SafeLoad<uint32_t>(in + 13) };
-  shifts = simd_batch{ 0, 0, 0, 4 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 28-bit bundles 16 to 19
-  words = simd_batch{ SafeLoad<uint32_t>(in + 14), SafeLoad<uint32_t>(in + 14) >> 28 | SafeLoad<uint32_t>(in + 15) << 4, SafeLoad<uint32_t>(in + 15) >> 24 | SafeLoad<uint32_t>(in + 16) << 8, SafeLoad<uint32_t>(in + 16) >> 20 | SafeLoad<uint32_t>(in + 17) << 12 };
-  shifts = simd_batch{ 0, 0, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 28-bit bundles 20 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 17) >> 16 | SafeLoad<uint32_t>(in + 18) << 16, SafeLoad<uint32_t>(in + 18) >> 12 | SafeLoad<uint32_t>(in + 19) << 20, SafeLoad<uint32_t>(in + 19) >> 8 | SafeLoad<uint32_t>(in + 20) << 24, SafeLoad<uint32_t>(in + 20) };
-  shifts = simd_batch{ 0, 0, 0, 4 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 28-bit bundles 24 to 27
-  words = simd_batch{ SafeLoad<uint32_t>(in + 21), SafeLoad<uint32_t>(in + 21) >> 28 | SafeLoad<uint32_t>(in + 22) << 4, SafeLoad<uint32_t>(in + 22) >> 24 | SafeLoad<uint32_t>(in + 23) << 8, SafeLoad<uint32_t>(in + 23) >> 20 | SafeLoad<uint32_t>(in + 24) << 12 };
-  shifts = simd_batch{ 0, 0, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 28-bit bundles 28 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 24) >> 16 | SafeLoad<uint32_t>(in + 25) << 16, SafeLoad<uint32_t>(in + 25) >> 12 | SafeLoad<uint32_t>(in + 26) << 20, SafeLoad<uint32_t>(in + 26) >> 8 | SafeLoad<uint32_t>(in + 27) << 24, SafeLoad<uint32_t>(in + 27) };
-  shifts = simd_batch{ 0, 0, 0, 4 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  in += 28;
-  return in;
-}
-
-inline static const uint32_t* unpack29_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x1fffffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 29-bit bundles 0 to 3
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 29 | SafeLoad<uint32_t>(in + 1) << 3, SafeLoad<uint32_t>(in + 1) >> 26 | SafeLoad<uint32_t>(in + 2) << 6, SafeLoad<uint32_t>(in + 2) >> 23 | SafeLoad<uint32_t>(in + 3) << 9 };
-  shifts = simd_batch{ 0, 0, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 29-bit bundles 4 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 3) >> 20 | SafeLoad<uint32_t>(in + 4) << 12, SafeLoad<uint32_t>(in + 4) >> 17 | SafeLoad<uint32_t>(in + 5) << 15, SafeLoad<uint32_t>(in + 5) >> 14 | SafeLoad<uint32_t>(in + 6) << 18, SafeLoad<uint32_t>(in + 6) >> 11 | SafeLoad<uint32_t>(in + 7) << 21 };
-  shifts = simd_batch{ 0, 0, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 29-bit bundles 8 to 11
-  words = simd_batch{ SafeLoad<uint32_t>(in + 7) >> 8 | SafeLoad<uint32_t>(in + 8) << 24, SafeLoad<uint32_t>(in + 8) >> 5 | SafeLoad<uint32_t>(in + 9) << 27, SafeLoad<uint32_t>(in + 9), SafeLoad<uint32_t>(in + 9) >> 31 | SafeLoad<uint32_t>(in + 10) << 1 };
-  shifts = simd_batch{ 0, 0, 2, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 29-bit bundles 12 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 10) >> 28 | SafeLoad<uint32_t>(in + 11) << 4, SafeLoad<uint32_t>(in + 11) >> 25 | SafeLoad<uint32_t>(in + 12) << 7, SafeLoad<uint32_t>(in + 12) >> 22 | SafeLoad<uint32_t>(in + 13) << 10, SafeLoad<uint32_t>(in + 13) >> 19 | SafeLoad<uint32_t>(in + 14) << 13 };
-  shifts = simd_batch{ 0, 0, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 29-bit bundles 16 to 19
-  words = simd_batch{ SafeLoad<uint32_t>(in + 14) >> 16 | SafeLoad<uint32_t>(in + 15) << 16, SafeLoad<uint32_t>(in + 15) >> 13 | SafeLoad<uint32_t>(in + 16) << 19, SafeLoad<uint32_t>(in + 16) >> 10 | SafeLoad<uint32_t>(in + 17) << 22, SafeLoad<uint32_t>(in + 17) >> 7 | SafeLoad<uint32_t>(in + 18) << 25 };
-  shifts = simd_batch{ 0, 0, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 29-bit bundles 20 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 18) >> 4 | SafeLoad<uint32_t>(in + 19) << 28, SafeLoad<uint32_t>(in + 19), SafeLoad<uint32_t>(in + 19) >> 30 | SafeLoad<uint32_t>(in + 20) << 2, SafeLoad<uint32_t>(in + 20) >> 27 | SafeLoad<uint32_t>(in + 21) << 5 };
-  shifts = simd_batch{ 0, 1, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 29-bit bundles 24 to 27
-  words = simd_batch{ SafeLoad<uint32_t>(in + 21) >> 24 | SafeLoad<uint32_t>(in + 22) << 8, SafeLoad<uint32_t>(in + 22) >> 21 | SafeLoad<uint32_t>(in + 23) << 11, SafeLoad<uint32_t>(in + 23) >> 18 | SafeLoad<uint32_t>(in + 24) << 14, SafeLoad<uint32_t>(in + 24) >> 15 | SafeLoad<uint32_t>(in + 25) << 17 };
-  shifts = simd_batch{ 0, 0, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 29-bit bundles 28 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 25) >> 12 | SafeLoad<uint32_t>(in + 26) << 20, SafeLoad<uint32_t>(in + 26) >> 9 | SafeLoad<uint32_t>(in + 27) << 23, SafeLoad<uint32_t>(in + 27) >> 6 | SafeLoad<uint32_t>(in + 28) << 26, SafeLoad<uint32_t>(in + 28) };
-  shifts = simd_batch{ 0, 0, 0, 3 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  in += 29;
-  return in;
-}
-
-inline static const uint32_t* unpack30_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x3fffffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 30-bit bundles 0 to 3
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 30 | SafeLoad<uint32_t>(in + 1) << 2, SafeLoad<uint32_t>(in + 1) >> 28 | SafeLoad<uint32_t>(in + 2) << 4, SafeLoad<uint32_t>(in + 2) >> 26 | SafeLoad<uint32_t>(in + 3) << 6 };
-  shifts = simd_batch{ 0, 0, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 30-bit bundles 4 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 3) >> 24 | SafeLoad<uint32_t>(in + 4) << 8, SafeLoad<uint32_t>(in + 4) >> 22 | SafeLoad<uint32_t>(in + 5) << 10, SafeLoad<uint32_t>(in + 5) >> 20 | SafeLoad<uint32_t>(in + 6) << 12, SafeLoad<uint32_t>(in + 6) >> 18 | SafeLoad<uint32_t>(in + 7) << 14 };
-  shifts = simd_batch{ 0, 0, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 30-bit bundles 8 to 11
-  words = simd_batch{ SafeLoad<uint32_t>(in + 7) >> 16 | SafeLoad<uint32_t>(in + 8) << 16, SafeLoad<uint32_t>(in + 8) >> 14 | SafeLoad<uint32_t>(in + 9) << 18, SafeLoad<uint32_t>(in + 9) >> 12 | SafeLoad<uint32_t>(in + 10) << 20, SafeLoad<uint32_t>(in + 10) >> 10 | SafeLoad<uint32_t>(in + 11) << 22 };
-  shifts = simd_batch{ 0, 0, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 30-bit bundles 12 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 11) >> 8 | SafeLoad<uint32_t>(in + 12) << 24, SafeLoad<uint32_t>(in + 12) >> 6 | SafeLoad<uint32_t>(in + 13) << 26, SafeLoad<uint32_t>(in + 13) >> 4 | SafeLoad<uint32_t>(in + 14) << 28, SafeLoad<uint32_t>(in + 14) };
-  shifts = simd_batch{ 0, 0, 0, 2 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 30-bit bundles 16 to 19
-  words = simd_batch{ SafeLoad<uint32_t>(in + 15), SafeLoad<uint32_t>(in + 15) >> 30 | SafeLoad<uint32_t>(in + 16) << 2, SafeLoad<uint32_t>(in + 16) >> 28 | SafeLoad<uint32_t>(in + 17) << 4, SafeLoad<uint32_t>(in + 17) >> 26 | SafeLoad<uint32_t>(in + 18) << 6 };
-  shifts = simd_batch{ 0, 0, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 30-bit bundles 20 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 18) >> 24 | SafeLoad<uint32_t>(in + 19) << 8, SafeLoad<uint32_t>(in + 19) >> 22 | SafeLoad<uint32_t>(in + 20) << 10, SafeLoad<uint32_t>(in + 20) >> 20 | SafeLoad<uint32_t>(in + 21) << 12, SafeLoad<uint32_t>(in + 21) >> 18 | SafeLoad<uint32_t>(in + 22) << 14 };
-  shifts = simd_batch{ 0, 0, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 30-bit bundles 24 to 27
-  words = simd_batch{ SafeLoad<uint32_t>(in + 22) >> 16 | SafeLoad<uint32_t>(in + 23) << 16, SafeLoad<uint32_t>(in + 23) >> 14 | SafeLoad<uint32_t>(in + 24) << 18, SafeLoad<uint32_t>(in + 24) >> 12 | SafeLoad<uint32_t>(in + 25) << 20, SafeLoad<uint32_t>(in + 25) >> 10 | SafeLoad<uint32_t>(in + 26) << 22 };
-  shifts = simd_batch{ 0, 0, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 30-bit bundles 28 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 26) >> 8 | SafeLoad<uint32_t>(in + 27) << 24, SafeLoad<uint32_t>(in + 27) >> 6 | SafeLoad<uint32_t>(in + 28) << 26, SafeLoad<uint32_t>(in + 28) >> 4 | SafeLoad<uint32_t>(in + 29) << 28, SafeLoad<uint32_t>(in + 29) };
-  shifts = simd_batch{ 0, 0, 0, 2 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  in += 30;
-  return in;
-}
-
-inline static const uint32_t* unpack31_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x7fffffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 31-bit bundles 0 to 3
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 31 | SafeLoad<uint32_t>(in + 1) << 1, SafeLoad<uint32_t>(in + 1) >> 30 | SafeLoad<uint32_t>(in + 2) << 2, SafeLoad<uint32_t>(in + 2) >> 29 | SafeLoad<uint32_t>(in + 3) << 3 };
-  shifts = simd_batch{ 0, 0, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 31-bit bundles 4 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 3) >> 28 | SafeLoad<uint32_t>(in + 4) << 4, SafeLoad<uint32_t>(in + 4) >> 27 | SafeLoad<uint32_t>(in + 5) << 5, SafeLoad<uint32_t>(in + 5) >> 26 | SafeLoad<uint32_t>(in + 6) << 6, SafeLoad<uint32_t>(in + 6) >> 25 | SafeLoad<uint32_t>(in + 7) << 7 };
-  shifts = simd_batch{ 0, 0, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 31-bit bundles 8 to 11
-  words = simd_batch{ SafeLoad<uint32_t>(in + 7) >> 24 | SafeLoad<uint32_t>(in + 8) << 8, SafeLoad<uint32_t>(in + 8) >> 23 | SafeLoad<uint32_t>(in + 9) << 9, SafeLoad<uint32_t>(in + 9) >> 22 | SafeLoad<uint32_t>(in + 10) << 10, SafeLoad<uint32_t>(in + 10) >> 21 | SafeLoad<uint32_t>(in + 11) << 11 };
-  shifts = simd_batch{ 0, 0, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 31-bit bundles 12 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 11) >> 20 | SafeLoad<uint32_t>(in + 12) << 12, SafeLoad<uint32_t>(in + 12) >> 19 | SafeLoad<uint32_t>(in + 13) << 13, SafeLoad<uint32_t>(in + 13) >> 18 | SafeLoad<uint32_t>(in + 14) << 14, SafeLoad<uint32_t>(in + 14) >> 17 | SafeLoad<uint32_t>(in + 15) << 15 };
-  shifts = simd_batch{ 0, 0, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 31-bit bundles 16 to 19
-  words = simd_batch{ SafeLoad<uint32_t>(in + 15) >> 16 | SafeLoad<uint32_t>(in + 16) << 16, SafeLoad<uint32_t>(in + 16) >> 15 | SafeLoad<uint32_t>(in + 17) << 17, SafeLoad<uint32_t>(in + 17) >> 14 | SafeLoad<uint32_t>(in + 18) << 18, SafeLoad<uint32_t>(in + 18) >> 13 | SafeLoad<uint32_t>(in + 19) << 19 };
-  shifts = simd_batch{ 0, 0, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 31-bit bundles 20 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 19) >> 12 | SafeLoad<uint32_t>(in + 20) << 20, SafeLoad<uint32_t>(in + 20) >> 11 | SafeLoad<uint32_t>(in + 21) << 21, SafeLoad<uint32_t>(in + 21) >> 10 | SafeLoad<uint32_t>(in + 22) << 22, SafeLoad<uint32_t>(in + 22) >> 9 | SafeLoad<uint32_t>(in + 23) << 23 };
-  shifts = simd_batch{ 0, 0, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 31-bit bundles 24 to 27
-  words = simd_batch{ SafeLoad<uint32_t>(in + 23) >> 8 | SafeLoad<uint32_t>(in + 24) << 24, SafeLoad<uint32_t>(in + 24) >> 7 | SafeLoad<uint32_t>(in + 25) << 25, SafeLoad<uint32_t>(in + 25) >> 6 | SafeLoad<uint32_t>(in + 26) << 26, SafeLoad<uint32_t>(in + 26) >> 5 | SafeLoad<uint32_t>(in + 27) << 27 };
-  shifts = simd_batch{ 0, 0, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  // extract 31-bit bundles 28 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 27) >> 4 | SafeLoad<uint32_t>(in + 28) << 28, SafeLoad<uint32_t>(in + 28) >> 3 | SafeLoad<uint32_t>(in + 29) << 29, SafeLoad<uint32_t>(in + 29) >> 2 | SafeLoad<uint32_t>(in + 30) << 30, SafeLoad<uint32_t>(in + 30) };
-  shifts = simd_batch{ 0, 0, 0, 1 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 4;
-
-  in += 31;
-  return in;
-}
-
-inline static const uint32_t* unpack32_32(const uint32_t* in, uint32_t* out) {
-  memcpy(out, in, 32 * sizeof(*out));
-  in += 32;
-  out += 32;
-
-  return in;
-}
-
-};  // struct UnpackBits128
-
-}  // namespace
-}  // namespace internal
-}  // namespace arrow
-
+namespace arrow::internal {
+
+using ::arrow::util::SafeLoadAs;
+
+template<typename Uint, int BitWidth>
+struct Simd128UnpackerForWidth;
+
+template<>
+struct Simd128UnpackerForWidth<uint16_t, 1> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint16_t, 8>;
+  static constexpr int kValuesUnpacked = 16;
+
+  static const uint8_t* unpack(const uint8_t* in, uint16_t* out) {
+    constexpr uint16_t kMask = 0x1;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 1-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+    };
+    shifts = simd_batch{ 0, 1, 2, 3, 4, 5, 6, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 1-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+    };
+    shifts = simd_batch{ 8, 9, 10, 11, 12, 13, 14, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 1 * 2;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint16_t, 2> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint16_t, 8>;
+  static constexpr int kValuesUnpacked = 16;
+
+  static const uint8_t* unpack(const uint8_t* in, uint16_t* out) {
+    constexpr uint16_t kMask = 0x3;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 2-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+    };
+    shifts = simd_batch{ 0, 2, 4, 6, 8, 10, 12, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 2-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+    };
+    shifts = simd_batch{ 0, 2, 4, 6, 8, 10, 12, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 2 * 2;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint16_t, 3> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint16_t, 8>;
+  static constexpr int kValuesUnpacked = 16;
+
+  static const uint8_t* unpack(const uint8_t* in, uint16_t* out) {
+    constexpr uint16_t kMask = 0x7;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 3-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 0) >> 15 | SafeLoadAs<uint16_t>(in + 2 * 1) << 1),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+    };
+    shifts = simd_batch{ 0, 3, 6, 9, 12, 0, 2, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 3-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 1) >> 14 | SafeLoadAs<uint16_t>(in + 2 * 2) << 2),
+      SafeLoadAs<uint16_t>(in + 2 * 2),
+      SafeLoadAs<uint16_t>(in + 2 * 2),
+      SafeLoadAs<uint16_t>(in + 2 * 2),
+      SafeLoadAs<uint16_t>(in + 2 * 2),
+      SafeLoadAs<uint16_t>(in + 2 * 2),
+    };
+    shifts = simd_batch{ 8, 11, 0, 1, 4, 7, 10, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 3 * 2;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint16_t, 4> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint16_t, 8>;
+  static constexpr int kValuesUnpacked = 16;
+
+  static const uint8_t* unpack(const uint8_t* in, uint16_t* out) {
+    constexpr uint16_t kMask = 0xf;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 4-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+    };
+    shifts = simd_batch{ 0, 4, 8, 12, 0, 4, 8, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 4-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint16_t>(in + 2 * 2),
+      SafeLoadAs<uint16_t>(in + 2 * 2),
+      SafeLoadAs<uint16_t>(in + 2 * 2),
+      SafeLoadAs<uint16_t>(in + 2 * 2),
+      SafeLoadAs<uint16_t>(in + 2 * 3),
+      SafeLoadAs<uint16_t>(in + 2 * 3),
+      SafeLoadAs<uint16_t>(in + 2 * 3),
+      SafeLoadAs<uint16_t>(in + 2 * 3),
+    };
+    shifts = simd_batch{ 0, 4, 8, 12, 0, 4, 8, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 4 * 2;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint16_t, 5> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint16_t, 8>;
+  static constexpr int kValuesUnpacked = 16;
+
+  static const uint8_t* unpack(const uint8_t* in, uint16_t* out) {
+    constexpr uint16_t kMask = 0x1f;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 5-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 0) >> 15 | SafeLoadAs<uint16_t>(in + 2 * 1) << 1),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 1) >> 14 | SafeLoadAs<uint16_t>(in + 2 * 2) << 2),
+      SafeLoadAs<uint16_t>(in + 2 * 2),
+    };
+    shifts = simd_batch{ 0, 5, 10, 0, 4, 9, 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 5-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint16_t>(in + 2 * 2),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 2) >> 13 | SafeLoadAs<uint16_t>(in + 2 * 3) << 3),
+      SafeLoadAs<uint16_t>(in + 2 * 3),
+      SafeLoadAs<uint16_t>(in + 2 * 3),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 3) >> 12 | SafeLoadAs<uint16_t>(in + 2 * 4) << 4),
+      SafeLoadAs<uint16_t>(in + 2 * 4),
+      SafeLoadAs<uint16_t>(in + 2 * 4),
+      SafeLoadAs<uint16_t>(in + 2 * 4),
+    };
+    shifts = simd_batch{ 8, 0, 2, 7, 0, 1, 6, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 5 * 2;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint16_t, 6> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint16_t, 8>;
+  static constexpr int kValuesUnpacked = 16;
+
+  static const uint8_t* unpack(const uint8_t* in, uint16_t* out) {
+    constexpr uint16_t kMask = 0x3f;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 6-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 0) >> 12 | SafeLoadAs<uint16_t>(in + 2 * 1) << 4),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 1) >> 14 | SafeLoadAs<uint16_t>(in + 2 * 2) << 2),
+      SafeLoadAs<uint16_t>(in + 2 * 2),
+      SafeLoadAs<uint16_t>(in + 2 * 2),
+    };
+    shifts = simd_batch{ 0, 6, 0, 2, 8, 0, 4, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 6-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint16_t>(in + 2 * 3),
+      SafeLoadAs<uint16_t>(in + 2 * 3),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 3) >> 12 | SafeLoadAs<uint16_t>(in + 2 * 4) << 4),
+      SafeLoadAs<uint16_t>(in + 2 * 4),
+      SafeLoadAs<uint16_t>(in + 2 * 4),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 4) >> 14 | SafeLoadAs<uint16_t>(in + 2 * 5) << 2),
+      SafeLoadAs<uint16_t>(in + 2 * 5),
+      SafeLoadAs<uint16_t>(in + 2 * 5),
+    };
+    shifts = simd_batch{ 0, 6, 0, 2, 8, 0, 4, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 6 * 2;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint16_t, 7> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint16_t, 8>;
+  static constexpr int kValuesUnpacked = 16;
+
+  static const uint8_t* unpack(const uint8_t* in, uint16_t* out) {
+    constexpr uint16_t kMask = 0x7f;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 7-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 0) >> 14 | SafeLoadAs<uint16_t>(in + 2 * 1) << 2),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 1) >> 12 | SafeLoadAs<uint16_t>(in + 2 * 2) << 4),
+      SafeLoadAs<uint16_t>(in + 2 * 2),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 2) >> 10 | SafeLoadAs<uint16_t>(in + 2 * 3) << 6),
+      SafeLoadAs<uint16_t>(in + 2 * 3),
+    };
+    shifts = simd_batch{ 0, 7, 0, 5, 0, 3, 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 7-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint16_t>(in + 2 * 3),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 3) >> 15 | SafeLoadAs<uint16_t>(in + 2 * 4) << 1),
+      SafeLoadAs<uint16_t>(in + 2 * 4),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 4) >> 13 | SafeLoadAs<uint16_t>(in + 2 * 5) << 3),
+      SafeLoadAs<uint16_t>(in + 2 * 5),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 5) >> 11 | SafeLoadAs<uint16_t>(in + 2 * 6) << 5),
+      SafeLoadAs<uint16_t>(in + 2 * 6),
+      SafeLoadAs<uint16_t>(in + 2 * 6),
+    };
+    shifts = simd_batch{ 8, 0, 6, 0, 4, 0, 2, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 7 * 2;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint16_t, 8> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint16_t, 8>;
+  static constexpr int kValuesUnpacked = 16;
+
+  static const uint8_t* unpack(const uint8_t* in, uint16_t* out) {
+    constexpr uint16_t kMask = 0xff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 8-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+      SafeLoadAs<uint16_t>(in + 2 * 2),
+      SafeLoadAs<uint16_t>(in + 2 * 2),
+      SafeLoadAs<uint16_t>(in + 2 * 3),
+      SafeLoadAs<uint16_t>(in + 2 * 3),
+    };
+    shifts = simd_batch{ 0, 8, 0, 8, 0, 8, 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 8-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint16_t>(in + 2 * 4),
+      SafeLoadAs<uint16_t>(in + 2 * 4),
+      SafeLoadAs<uint16_t>(in + 2 * 5),
+      SafeLoadAs<uint16_t>(in + 2 * 5),
+      SafeLoadAs<uint16_t>(in + 2 * 6),
+      SafeLoadAs<uint16_t>(in + 2 * 6),
+      SafeLoadAs<uint16_t>(in + 2 * 7),
+      SafeLoadAs<uint16_t>(in + 2 * 7),
+    };
+    shifts = simd_batch{ 0, 8, 0, 8, 0, 8, 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 8 * 2;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint16_t, 9> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint16_t, 8>;
+  static constexpr int kValuesUnpacked = 16;
+
+  static const uint8_t* unpack(const uint8_t* in, uint16_t* out) {
+    constexpr uint16_t kMask = 0x1ff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 9-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 0) >> 9 | SafeLoadAs<uint16_t>(in + 2 * 1) << 7),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 1) >> 11 | SafeLoadAs<uint16_t>(in + 2 * 2) << 5),
+      SafeLoadAs<uint16_t>(in + 2 * 2),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 2) >> 13 | SafeLoadAs<uint16_t>(in + 2 * 3) << 3),
+      SafeLoadAs<uint16_t>(in + 2 * 3),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 3) >> 15 | SafeLoadAs<uint16_t>(in + 2 * 4) << 1),
+    };
+    shifts = simd_batch{ 0, 0, 2, 0, 4, 0, 6, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 9-bit bundles 8 to 15
+    words = simd_batch{
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 4) >> 8 | SafeLoadAs<uint16_t>(in + 2 * 5) << 8),
+      SafeLoadAs<uint16_t>(in + 2 * 5),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 5) >> 10 | SafeLoadAs<uint16_t>(in + 2 * 6) << 6),
+      SafeLoadAs<uint16_t>(in + 2 * 6),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 6) >> 12 | SafeLoadAs<uint16_t>(in + 2 * 7) << 4),
+      SafeLoadAs<uint16_t>(in + 2 * 7),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 7) >> 14 | SafeLoadAs<uint16_t>(in + 2 * 8) << 2),
+      SafeLoadAs<uint16_t>(in + 2 * 8),
+    };
+    shifts = simd_batch{ 0, 1, 0, 3, 0, 5, 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 9 * 2;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint16_t, 10> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint16_t, 8>;
+  static constexpr int kValuesUnpacked = 16;
+
+  static const uint8_t* unpack(const uint8_t* in, uint16_t* out) {
+    constexpr uint16_t kMask = 0x3ff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 10-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 0) >> 10 | SafeLoadAs<uint16_t>(in + 2 * 1) << 6),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 1) >> 14 | SafeLoadAs<uint16_t>(in + 2 * 2) << 2),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 2) >> 8 | SafeLoadAs<uint16_t>(in + 2 * 3) << 8),
+      SafeLoadAs<uint16_t>(in + 2 * 3),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 3) >> 12 | SafeLoadAs<uint16_t>(in + 2 * 4) << 4),
+      SafeLoadAs<uint16_t>(in + 2 * 4),
+    };
+    shifts = simd_batch{ 0, 0, 4, 0, 0, 2, 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 10-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint16_t>(in + 2 * 5),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 5) >> 10 | SafeLoadAs<uint16_t>(in + 2 * 6) << 6),
+      SafeLoadAs<uint16_t>(in + 2 * 6),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 6) >> 14 | SafeLoadAs<uint16_t>(in + 2 * 7) << 2),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 7) >> 8 | SafeLoadAs<uint16_t>(in + 2 * 8) << 8),
+      SafeLoadAs<uint16_t>(in + 2 * 8),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 8) >> 12 | SafeLoadAs<uint16_t>(in + 2 * 9) << 4),
+      SafeLoadAs<uint16_t>(in + 2 * 9),
+    };
+    shifts = simd_batch{ 0, 0, 4, 0, 0, 2, 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 10 * 2;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint16_t, 11> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint16_t, 8>;
+  static constexpr int kValuesUnpacked = 16;
+
+  static const uint8_t* unpack(const uint8_t* in, uint16_t* out) {
+    constexpr uint16_t kMask = 0x7ff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 11-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 0) >> 11 | SafeLoadAs<uint16_t>(in + 2 * 1) << 5),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 1) >> 6 | SafeLoadAs<uint16_t>(in + 2 * 2) << 10),
+      SafeLoadAs<uint16_t>(in + 2 * 2),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 2) >> 12 | SafeLoadAs<uint16_t>(in + 2 * 3) << 4),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 3) >> 7 | SafeLoadAs<uint16_t>(in + 2 * 4) << 9),
+      SafeLoadAs<uint16_t>(in + 2 * 4),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 4) >> 13 | SafeLoadAs<uint16_t>(in + 2 * 5) << 3),
+    };
+    shifts = simd_batch{ 0, 0, 0, 1, 0, 0, 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 11-bit bundles 8 to 15
+    words = simd_batch{
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 5) >> 8 | SafeLoadAs<uint16_t>(in + 2 * 6) << 8),
+      SafeLoadAs<uint16_t>(in + 2 * 6),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 6) >> 14 | SafeLoadAs<uint16_t>(in + 2 * 7) << 2),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 7) >> 9 | SafeLoadAs<uint16_t>(in + 2 * 8) << 7),
+      SafeLoadAs<uint16_t>(in + 2 * 8),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 8) >> 15 | SafeLoadAs<uint16_t>(in + 2 * 9) << 1),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 9) >> 10 | SafeLoadAs<uint16_t>(in + 2 * 10) << 6),
+      SafeLoadAs<uint16_t>(in + 2 * 10),
+    };
+    shifts = simd_batch{ 0, 3, 0, 0, 4, 0, 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 11 * 2;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint16_t, 12> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint16_t, 8>;
+  static constexpr int kValuesUnpacked = 16;
+
+  static const uint8_t* unpack(const uint8_t* in, uint16_t* out) {
+    constexpr uint16_t kMask = 0xfff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 12-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 0) >> 12 | SafeLoadAs<uint16_t>(in + 2 * 1) << 4),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 1) >> 8 | SafeLoadAs<uint16_t>(in + 2 * 2) << 8),
+      SafeLoadAs<uint16_t>(in + 2 * 2),
+      SafeLoadAs<uint16_t>(in + 2 * 3),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 3) >> 12 | SafeLoadAs<uint16_t>(in + 2 * 4) << 4),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 4) >> 8 | SafeLoadAs<uint16_t>(in + 2 * 5) << 8),
+      SafeLoadAs<uint16_t>(in + 2 * 5),
+    };
+    shifts = simd_batch{ 0, 0, 0, 4, 0, 0, 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 12-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint16_t>(in + 2 * 6),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 6) >> 12 | SafeLoadAs<uint16_t>(in + 2 * 7) << 4),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 7) >> 8 | SafeLoadAs<uint16_t>(in + 2 * 8) << 8),
+      SafeLoadAs<uint16_t>(in + 2 * 8),
+      SafeLoadAs<uint16_t>(in + 2 * 9),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 9) >> 12 | SafeLoadAs<uint16_t>(in + 2 * 10) << 4),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 10) >> 8 | SafeLoadAs<uint16_t>(in + 2 * 11) << 8),
+      SafeLoadAs<uint16_t>(in + 2 * 11),
+    };
+    shifts = simd_batch{ 0, 0, 0, 4, 0, 0, 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 12 * 2;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint16_t, 13> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint16_t, 8>;
+  static constexpr int kValuesUnpacked = 16;
+
+  static const uint8_t* unpack(const uint8_t* in, uint16_t* out) {
+    constexpr uint16_t kMask = 0x1fff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 13-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 0) >> 13 | SafeLoadAs<uint16_t>(in + 2 * 1) << 3),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 1) >> 10 | SafeLoadAs<uint16_t>(in + 2 * 2) << 6),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 2) >> 7 | SafeLoadAs<uint16_t>(in + 2 * 3) << 9),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 3) >> 4 | SafeLoadAs<uint16_t>(in + 2 * 4) << 12),
+      SafeLoadAs<uint16_t>(in + 2 * 4),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 4) >> 14 | SafeLoadAs<uint16_t>(in + 2 * 5) << 2),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 5) >> 11 | SafeLoadAs<uint16_t>(in + 2 * 6) << 5),
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 1, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 13-bit bundles 8 to 15
+    words = simd_batch{
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 6) >> 8 | SafeLoadAs<uint16_t>(in + 2 * 7) << 8),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 7) >> 5 | SafeLoadAs<uint16_t>(in + 2 * 8) << 11),
+      SafeLoadAs<uint16_t>(in + 2 * 8),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 8) >> 15 | SafeLoadAs<uint16_t>(in + 2 * 9) << 1),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 9) >> 12 | SafeLoadAs<uint16_t>(in + 2 * 10) << 4),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 10) >> 9 | SafeLoadAs<uint16_t>(in + 2 * 11) << 7),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 11) >> 6 | SafeLoadAs<uint16_t>(in + 2 * 12) << 10),
+      SafeLoadAs<uint16_t>(in + 2 * 12),
+    };
+    shifts = simd_batch{ 0, 0, 2, 0, 0, 0, 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 13 * 2;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint16_t, 14> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint16_t, 8>;
+  static constexpr int kValuesUnpacked = 16;
+
+  static const uint8_t* unpack(const uint8_t* in, uint16_t* out) {
+    constexpr uint16_t kMask = 0x3fff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 14-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 0) >> 14 | SafeLoadAs<uint16_t>(in + 2 * 1) << 2),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 1) >> 12 | SafeLoadAs<uint16_t>(in + 2 * 2) << 4),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 2) >> 10 | SafeLoadAs<uint16_t>(in + 2 * 3) << 6),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 3) >> 8 | SafeLoadAs<uint16_t>(in + 2 * 4) << 8),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 4) >> 6 | SafeLoadAs<uint16_t>(in + 2 * 5) << 10),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 5) >> 4 | SafeLoadAs<uint16_t>(in + 2 * 6) << 12),
+      SafeLoadAs<uint16_t>(in + 2 * 6),
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 14-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint16_t>(in + 2 * 7),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 7) >> 14 | SafeLoadAs<uint16_t>(in + 2 * 8) << 2),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 8) >> 12 | SafeLoadAs<uint16_t>(in + 2 * 9) << 4),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 9) >> 10 | SafeLoadAs<uint16_t>(in + 2 * 10) << 6),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 10) >> 8 | SafeLoadAs<uint16_t>(in + 2 * 11) << 8),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 11) >> 6 | SafeLoadAs<uint16_t>(in + 2 * 12) << 10),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 12) >> 4 | SafeLoadAs<uint16_t>(in + 2 * 13) << 12),
+      SafeLoadAs<uint16_t>(in + 2 * 13),
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 14 * 2;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint16_t, 15> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint16_t, 8>;
+  static constexpr int kValuesUnpacked = 16;
+
+  static const uint8_t* unpack(const uint8_t* in, uint16_t* out) {
+    constexpr uint16_t kMask = 0x7fff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 15-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 0) >> 15 | SafeLoadAs<uint16_t>(in + 2 * 1) << 1),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 1) >> 14 | SafeLoadAs<uint16_t>(in + 2 * 2) << 2),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 2) >> 13 | SafeLoadAs<uint16_t>(in + 2 * 3) << 3),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 3) >> 12 | SafeLoadAs<uint16_t>(in + 2 * 4) << 4),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 4) >> 11 | SafeLoadAs<uint16_t>(in + 2 * 5) << 5),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 5) >> 10 | SafeLoadAs<uint16_t>(in + 2 * 6) << 6),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 6) >> 9 | SafeLoadAs<uint16_t>(in + 2 * 7) << 7),
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 15-bit bundles 8 to 15
+    words = simd_batch{
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 7) >> 8 | SafeLoadAs<uint16_t>(in + 2 * 8) << 8),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 8) >> 7 | SafeLoadAs<uint16_t>(in + 2 * 9) << 9),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 9) >> 6 | SafeLoadAs<uint16_t>(in + 2 * 10) << 10),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 10) >> 5 | SafeLoadAs<uint16_t>(in + 2 * 11) << 11),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 11) >> 4 | SafeLoadAs<uint16_t>(in + 2 * 12) << 12),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 12) >> 3 | SafeLoadAs<uint16_t>(in + 2 * 13) << 13),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 13) >> 2 | SafeLoadAs<uint16_t>(in + 2 * 14) << 14),
+      SafeLoadAs<uint16_t>(in + 2 * 14),
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 15 * 2;
+    return in;
+  }
+};
+
+
+template<>
+struct Simd128UnpackerForWidth<uint32_t, 1> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 4>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x1;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 1-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+    };
+    shifts = simd_batch{ 0, 1, 2, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 1-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+    };
+    shifts = simd_batch{ 4, 5, 6, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 1-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+    };
+    shifts = simd_batch{ 8, 9, 10, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 1-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+    };
+    shifts = simd_batch{ 12, 13, 14, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 1-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+    };
+    shifts = simd_batch{ 16, 17, 18, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 1-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+    };
+    shifts = simd_batch{ 20, 21, 22, 23 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 1-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+    };
+    shifts = simd_batch{ 24, 25, 26, 27 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 1-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+    };
+    shifts = simd_batch{ 28, 29, 30, 31 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 1 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint32_t, 2> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 4>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x3;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 2-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+    };
+    shifts = simd_batch{ 0, 2, 4, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 2-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+    };
+    shifts = simd_batch{ 8, 10, 12, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 2-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+    };
+    shifts = simd_batch{ 16, 18, 20, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 2-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+    };
+    shifts = simd_batch{ 24, 26, 28, 30 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 2-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+    };
+    shifts = simd_batch{ 0, 2, 4, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 2-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+    };
+    shifts = simd_batch{ 8, 10, 12, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 2-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+    };
+    shifts = simd_batch{ 16, 18, 20, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 2-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+    };
+    shifts = simd_batch{ 24, 26, 28, 30 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 2 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint32_t, 3> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 4>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x7;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 3-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+    };
+    shifts = simd_batch{ 0, 3, 6, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 3-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+    };
+    shifts = simd_batch{ 12, 15, 18, 21 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 3-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 1) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+    };
+    shifts = simd_batch{ 24, 27, 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 3-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+    };
+    shifts = simd_batch{ 4, 7, 10, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 3-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+    };
+    shifts = simd_batch{ 16, 19, 22, 25 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 3-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 31 | SafeLoadAs<uint32_t>(in + 4 * 2) << 1,
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+    };
+    shifts = simd_batch{ 28, 0, 2, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 3-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+    };
+    shifts = simd_batch{ 8, 11, 14, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 3-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+    };
+    shifts = simd_batch{ 20, 23, 26, 29 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 3 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint32_t, 4> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 4>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0xf;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 4-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+    };
+    shifts = simd_batch{ 0, 4, 8, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 4-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+    };
+    shifts = simd_batch{ 16, 20, 24, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 4-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+    };
+    shifts = simd_batch{ 0, 4, 8, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 4-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+    };
+    shifts = simd_batch{ 16, 20, 24, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 4-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+    };
+    shifts = simd_batch{ 0, 4, 8, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 4-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+    };
+    shifts = simd_batch{ 16, 20, 24, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 4-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+    };
+    shifts = simd_batch{ 0, 4, 8, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 4-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+    };
+    shifts = simd_batch{ 16, 20, 24, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 4 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint32_t, 5> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 4>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x1f;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 5-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+    };
+    shifts = simd_batch{ 0, 5, 10, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 5-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 1) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+    };
+    shifts = simd_batch{ 20, 25, 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 5-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+    };
+    shifts = simd_batch{ 8, 13, 18, 23 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 5-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 2) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+    };
+    shifts = simd_batch{ 0, 1, 6, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 5-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 31 | SafeLoadAs<uint32_t>(in + 4 * 3) << 1,
+    };
+    shifts = simd_batch{ 16, 21, 26, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 5-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+    };
+    shifts = simd_batch{ 4, 9, 14, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 5-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 29 | SafeLoadAs<uint32_t>(in + 4 * 4) << 3,
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+    };
+    shifts = simd_batch{ 24, 0, 2, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 5-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+    };
+    shifts = simd_batch{ 12, 17, 22, 27 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 5 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint32_t, 6> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 4>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x3f;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 6-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+    };
+    shifts = simd_batch{ 0, 6, 12, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 6-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 1) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+    };
+    shifts = simd_batch{ 24, 0, 4, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 6-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 2) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+    };
+    shifts = simd_batch{ 16, 22, 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 6-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+    };
+    shifts = simd_batch{ 8, 14, 20, 26 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 6-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+    };
+    shifts = simd_batch{ 0, 6, 12, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 6-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 4) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+    };
+    shifts = simd_batch{ 24, 0, 4, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 6-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 5) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+    };
+    shifts = simd_batch{ 16, 22, 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 6-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+    };
+    shifts = simd_batch{ 8, 14, 20, 26 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 6 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint32_t, 7> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 4>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x7f;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 7-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+    };
+    shifts = simd_batch{ 0, 7, 14, 21 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 7-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 1) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+    };
+    shifts = simd_batch{ 0, 3, 10, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 7-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 31 | SafeLoadAs<uint32_t>(in + 4 * 2) << 1,
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+    };
+    shifts = simd_batch{ 24, 0, 6, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 7-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 27 | SafeLoadAs<uint32_t>(in + 4 * 3) << 5,
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+    };
+    shifts = simd_batch{ 20, 0, 2, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 7-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 4) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+    };
+    shifts = simd_batch{ 16, 23, 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 7-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 5) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+    };
+    shifts = simd_batch{ 12, 19, 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 7-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 29 | SafeLoadAs<uint32_t>(in + 4 * 6) << 3,
+    };
+    shifts = simd_batch{ 8, 15, 22, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 7-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+    };
+    shifts = simd_batch{ 4, 11, 18, 25 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 7 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint32_t, 8> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 4>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0xff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 8-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+    };
+    shifts = simd_batch{ 0, 8, 16, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 8-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+    };
+    shifts = simd_batch{ 0, 8, 16, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 8-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+    };
+    shifts = simd_batch{ 0, 8, 16, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 8-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+    };
+    shifts = simd_batch{ 0, 8, 16, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 8-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+    };
+    shifts = simd_batch{ 0, 8, 16, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 8-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+    };
+    shifts = simd_batch{ 0, 8, 16, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 8-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+    };
+    shifts = simd_batch{ 0, 8, 16, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 8-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+    };
+    shifts = simd_batch{ 0, 8, 16, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 8 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint32_t, 9> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 4>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x1ff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 9-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 27 | SafeLoadAs<uint32_t>(in + 4 * 1) << 5,
+    };
+    shifts = simd_batch{ 0, 9, 18, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 9-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 31 | SafeLoadAs<uint32_t>(in + 4 * 2) << 1,
+    };
+    shifts = simd_batch{ 4, 13, 22, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 9-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 3) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+    };
+    shifts = simd_batch{ 8, 17, 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 9-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 4) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+    };
+    shifts = simd_batch{ 12, 21, 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 9-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 25 | SafeLoadAs<uint32_t>(in + 4 * 5) << 7,
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+    };
+    shifts = simd_batch{ 16, 0, 2, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 9-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 29 | SafeLoadAs<uint32_t>(in + 4 * 6) << 3,
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+    };
+    shifts = simd_batch{ 20, 0, 6, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 9-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 7) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+    };
+    shifts = simd_batch{ 0, 1, 10, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 9-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 8) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+    };
+    shifts = simd_batch{ 0, 5, 14, 23 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 9 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint32_t, 10> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 4>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x3ff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 10-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 1) << 2,
+    };
+    shifts = simd_batch{ 0, 10, 20, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 10-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 2) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+    };
+    shifts = simd_batch{ 8, 18, 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 10-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 3) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+    };
+    shifts = simd_batch{ 16, 0, 4, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 10-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 4) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+    };
+    shifts = simd_batch{ 0, 2, 12, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 10-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 6) << 2,
+    };
+    shifts = simd_batch{ 0, 10, 20, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 10-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 7) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+    };
+    shifts = simd_batch{ 8, 18, 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 10-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 8) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+    };
+    shifts = simd_batch{ 16, 0, 4, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 10-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 9) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+    };
+    shifts = simd_batch{ 0, 2, 12, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 10 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint32_t, 11> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 4>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x7ff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 11-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 1) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+    };
+    shifts = simd_batch{ 0, 11, 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 11-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 23 | SafeLoadAs<uint32_t>(in + 4 * 2) << 9,
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+    };
+    shifts = simd_batch{ 12, 0, 2, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 11-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 3) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 25 | SafeLoadAs<uint32_t>(in + 4 * 4) << 7,
+    };
+    shifts = simd_batch{ 0, 3, 14, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 11-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 5) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+    };
+    shifts = simd_batch{ 4, 15, 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 11-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 27 | SafeLoadAs<uint32_t>(in + 4 * 6) << 5,
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+    };
+    shifts = simd_batch{ 16, 0, 6, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 11-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 7) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 29 | SafeLoadAs<uint32_t>(in + 4 * 8) << 3,
+    };
+    shifts = simd_batch{ 0, 7, 18, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 11-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 9) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+    };
+    shifts = simd_batch{ 8, 19, 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 11-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 31 | SafeLoadAs<uint32_t>(in + 4 * 10) << 1,
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+    };
+    shifts = simd_batch{ 20, 0, 10, 21 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 11 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint32_t, 12> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 4>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0xfff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 12-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 1) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+    };
+    shifts = simd_batch{ 0, 12, 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 12-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 2) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+    };
+    shifts = simd_batch{ 16, 0, 8, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 12-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 4) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+    };
+    shifts = simd_batch{ 0, 12, 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 12-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 5) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+    };
+    shifts = simd_batch{ 16, 0, 8, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 12-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 7) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+    };
+    shifts = simd_batch{ 0, 12, 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 12-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 8) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+    };
+    shifts = simd_batch{ 16, 0, 8, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 12-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 10) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+    };
+    shifts = simd_batch{ 0, 12, 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 12-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 11) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 11),
+      SafeLoadAs<uint32_t>(in + 4 * 11),
+    };
+    shifts = simd_batch{ 16, 0, 8, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 12 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint32_t, 13> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 4>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x1fff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 13-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 1) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+    };
+    shifts = simd_batch{ 0, 13, 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 13-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 2) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 27 | SafeLoadAs<uint32_t>(in + 4 * 3) << 5,
+    };
+    shifts = simd_batch{ 0, 1, 14, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 13-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 21 | SafeLoadAs<uint32_t>(in + 4 * 4) << 11,
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+    };
+    shifts = simd_batch{ 8, 0, 2, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 13-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 5) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 6) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+    };
+    shifts = simd_batch{ 0, 9, 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 13-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 29 | SafeLoadAs<uint32_t>(in + 4 * 7) << 3,
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 23 | SafeLoadAs<uint32_t>(in + 4 * 8) << 9,
+    };
+    shifts = simd_batch{ 16, 0, 10, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 13-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 9) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+    };
+    shifts = simd_batch{ 4, 17, 0, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 13-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 10) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 31 | SafeLoadAs<uint32_t>(in + 4 * 11) << 1,
+    };
+    shifts = simd_batch{ 0, 5, 18, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 13-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 11),
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 25 | SafeLoadAs<uint32_t>(in + 4 * 12) << 7,
+      SafeLoadAs<uint32_t>(in + 4 * 12),
+      SafeLoadAs<uint32_t>(in + 4 * 12),
+    };
+    shifts = simd_batch{ 12, 0, 6, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 13 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint32_t, 14> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 4>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x3fff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 14-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 1) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+    };
+    shifts = simd_batch{ 0, 14, 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 14-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 2) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 3) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+    };
+    shifts = simd_batch{ 0, 6, 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 14-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 4) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 5) << 6,
+    };
+    shifts = simd_batch{ 16, 0, 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 14-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 6) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+    };
+    shifts = simd_batch{ 8, 0, 4, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 14-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 8) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+    };
+    shifts = simd_batch{ 0, 14, 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 14-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 9) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 10) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+    };
+    shifts = simd_batch{ 0, 6, 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 14-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 11) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 11),
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 12) << 6,
+    };
+    shifts = simd_batch{ 16, 0, 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 14-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 12),
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 13) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 13),
+      SafeLoadAs<uint32_t>(in + 4 * 13),
+    };
+    shifts = simd_batch{ 8, 0, 4, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 14 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint32_t, 15> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 4>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x7fff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 15-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 1) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+    };
+    shifts = simd_batch{ 0, 15, 0, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 15-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 2) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 3) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+    };
+    shifts = simd_batch{ 0, 11, 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 15-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 4) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 5) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+    };
+    shifts = simd_batch{ 0, 7, 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 15-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 6) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 7) << 14,
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+    };
+    shifts = simd_batch{ 0, 3, 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 15-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 31 | SafeLoadAs<uint32_t>(in + 4 * 8) << 1,
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 29 | SafeLoadAs<uint32_t>(in + 4 * 9) << 3,
+    };
+    shifts = simd_batch{ 16, 0, 14, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 15-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 27 | SafeLoadAs<uint32_t>(in + 4 * 10) << 5,
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 25 | SafeLoadAs<uint32_t>(in + 4 * 11) << 7,
+    };
+    shifts = simd_batch{ 12, 0, 10, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 15-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 11),
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 23 | SafeLoadAs<uint32_t>(in + 4 * 12) << 9,
+      SafeLoadAs<uint32_t>(in + 4 * 12),
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 21 | SafeLoadAs<uint32_t>(in + 4 * 13) << 11,
+    };
+    shifts = simd_batch{ 8, 0, 6, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 15-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 13),
+      SafeLoadAs<uint32_t>(in + 4 * 13) >> 19 | SafeLoadAs<uint32_t>(in + 4 * 14) << 13,
+      SafeLoadAs<uint32_t>(in + 4 * 14),
+      SafeLoadAs<uint32_t>(in + 4 * 14),
+    };
+    shifts = simd_batch{ 4, 0, 2, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 15 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint32_t, 16> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 4>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0xffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 16-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+    };
+    shifts = simd_batch{ 0, 16, 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 16-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+    };
+    shifts = simd_batch{ 0, 16, 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 16-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+    };
+    shifts = simd_batch{ 0, 16, 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 16-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+    };
+    shifts = simd_batch{ 0, 16, 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 16-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+    };
+    shifts = simd_batch{ 0, 16, 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 16-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+      SafeLoadAs<uint32_t>(in + 4 * 11),
+      SafeLoadAs<uint32_t>(in + 4 * 11),
+    };
+    shifts = simd_batch{ 0, 16, 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 16-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 12),
+      SafeLoadAs<uint32_t>(in + 4 * 12),
+      SafeLoadAs<uint32_t>(in + 4 * 13),
+      SafeLoadAs<uint32_t>(in + 4 * 13),
+    };
+    shifts = simd_batch{ 0, 16, 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 16-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 14),
+      SafeLoadAs<uint32_t>(in + 4 * 14),
+      SafeLoadAs<uint32_t>(in + 4 * 15),
+      SafeLoadAs<uint32_t>(in + 4 * 15),
+    };
+    shifts = simd_batch{ 0, 16, 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 16 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint32_t, 17> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 4>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x1ffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 17-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 17 | SafeLoadAs<uint32_t>(in + 4 * 1) << 15,
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 19 | SafeLoadAs<uint32_t>(in + 4 * 2) << 13,
+    };
+    shifts = simd_batch{ 0, 0, 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 17-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 21 | SafeLoadAs<uint32_t>(in + 4 * 3) << 11,
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 23 | SafeLoadAs<uint32_t>(in + 4 * 4) << 9,
+    };
+    shifts = simd_batch{ 4, 0, 6, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 17-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 25 | SafeLoadAs<uint32_t>(in + 4 * 5) << 7,
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 27 | SafeLoadAs<uint32_t>(in + 4 * 6) << 5,
+    };
+    shifts = simd_batch{ 8, 0, 10, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 17-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 29 | SafeLoadAs<uint32_t>(in + 4 * 7) << 3,
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 31 | SafeLoadAs<uint32_t>(in + 4 * 8) << 1,
+    };
+    shifts = simd_batch{ 12, 0, 14, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 17-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 9) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 10) << 14,
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+    };
+    shifts = simd_batch{ 0, 1, 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 17-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 11) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 11),
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 12) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 12),
+    };
+    shifts = simd_batch{ 0, 5, 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 17-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 13) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 13),
+      SafeLoadAs<uint32_t>(in + 4 * 13) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 14) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 14),
+    };
+    shifts = simd_batch{ 0, 9, 0, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 17-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 14) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 15) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 15),
+      SafeLoadAs<uint32_t>(in + 4 * 15) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 16) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 16),
+    };
+    shifts = simd_batch{ 0, 13, 0, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 17 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint32_t, 18> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 4>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x3ffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 18-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 1) << 14,
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 2) << 10,
+    };
+    shifts = simd_batch{ 0, 0, 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 18-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 3) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 4) << 2,
+    };
+    shifts = simd_batch{ 8, 0, 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 18-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 5) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 6) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+    };
+    shifts = simd_batch{ 0, 2, 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 18-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 7) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 8) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+    };
+    shifts = simd_batch{ 0, 10, 0, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 18-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 10) << 14,
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 11) << 10,
+    };
+    shifts = simd_batch{ 0, 0, 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 18-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 11),
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 12) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 12),
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 13) << 2,
+    };
+    shifts = simd_batch{ 8, 0, 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 18-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 13) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 14) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 14),
+      SafeLoadAs<uint32_t>(in + 4 * 14) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 15) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 15),
+    };
+    shifts = simd_batch{ 0, 2, 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 18-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 15) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 16) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 16),
+      SafeLoadAs<uint32_t>(in + 4 * 16) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 17) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 17),
+    };
+    shifts = simd_batch{ 0, 10, 0, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 18 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint32_t, 19> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 4>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x7ffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 19-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 19 | SafeLoadAs<uint32_t>(in + 4 * 1) << 13,
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 25 | SafeLoadAs<uint32_t>(in + 4 * 2) << 7,
+    };
+    shifts = simd_batch{ 0, 0, 6, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 19-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 31 | SafeLoadAs<uint32_t>(in + 4 * 3) << 1,
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 4) << 14,
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+    };
+    shifts = simd_batch{ 12, 0, 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 19-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 5) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 6) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 17 | SafeLoadAs<uint32_t>(in + 4 * 7) << 15,
+    };
+    shifts = simd_batch{ 0, 11, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 19-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 23 | SafeLoadAs<uint32_t>(in + 4 * 8) << 9,
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 29 | SafeLoadAs<uint32_t>(in + 4 * 9) << 3,
+    };
+    shifts = simd_batch{ 4, 0, 10, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 19-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 10) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 11) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 11),
+    };
+    shifts = simd_batch{ 0, 3, 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 19-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 12) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 15 | SafeLoadAs<uint32_t>(in + 4 * 13) << 17,
+      SafeLoadAs<uint32_t>(in + 4 * 13),
+      SafeLoadAs<uint32_t>(in + 4 * 13) >> 21 | SafeLoadAs<uint32_t>(in + 4 * 14) << 11,
+    };
+    shifts = simd_batch{ 0, 0, 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 19-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 14),
+      SafeLoadAs<uint32_t>(in + 4 * 14) >> 27 | SafeLoadAs<uint32_t>(in + 4 * 15) << 5,
+      SafeLoadAs<uint32_t>(in + 4 * 15) >> 14 | SafeLoadAs<uint32_t>(in + 4 * 16) << 18,
+      SafeLoadAs<uint32_t>(in + 4 * 16),
+    };
+    shifts = simd_batch{ 8, 0, 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 19-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 16) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 17) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 17),
+      SafeLoadAs<uint32_t>(in + 4 * 17) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 18) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 18),
+    };
+    shifts = simd_batch{ 0, 7, 0, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 19 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint32_t, 20> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 4>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0xfffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 20-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 1) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 2) << 4,
+    };
+    shifts = simd_batch{ 0, 0, 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 20-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 3) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 4) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+    };
+    shifts = simd_batch{ 0, 4, 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 20-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 6) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 7) << 4,
+    };
+    shifts = simd_batch{ 0, 0, 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 20-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 8) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 9) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+    };
+    shifts = simd_batch{ 0, 4, 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 20-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 11) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 11),
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 12) << 4,
+    };
+    shifts = simd_batch{ 0, 0, 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 20-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 13) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 13),
+      SafeLoadAs<uint32_t>(in + 4 * 13) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 14) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 14),
+    };
+    shifts = simd_batch{ 0, 4, 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 20-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 15),
+      SafeLoadAs<uint32_t>(in + 4 * 15) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 16) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 16),
+      SafeLoadAs<uint32_t>(in + 4 * 16) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 17) << 4,
+    };
+    shifts = simd_batch{ 0, 0, 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 20-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 17) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 18) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 18),
+      SafeLoadAs<uint32_t>(in + 4 * 18) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 19) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 19),
+    };
+    shifts = simd_batch{ 0, 4, 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 20 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint32_t, 21> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 4>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x1fffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 21-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 21 | SafeLoadAs<uint32_t>(in + 4 * 1) << 11,
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 31 | SafeLoadAs<uint32_t>(in + 4 * 2) << 1,
+    };
+    shifts = simd_batch{ 0, 0, 10, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 21-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 3) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 4) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 19 | SafeLoadAs<uint32_t>(in + 4 * 5) << 13,
+    };
+    shifts = simd_batch{ 0, 9, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 21-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 29 | SafeLoadAs<uint32_t>(in + 4 * 6) << 3,
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 7) << 14,
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+    };
+    shifts = simd_batch{ 8, 0, 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 21-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 8) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 17 | SafeLoadAs<uint32_t>(in + 4 * 9) << 15,
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 27 | SafeLoadAs<uint32_t>(in + 4 * 10) << 5,
+    };
+    shifts = simd_batch{ 0, 0, 6, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 21-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 11) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 11),
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 12) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 15 | SafeLoadAs<uint32_t>(in + 4 * 13) << 17,
+    };
+    shifts = simd_batch{ 0, 5, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 21-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 13),
+      SafeLoadAs<uint32_t>(in + 4 * 13) >> 25 | SafeLoadAs<uint32_t>(in + 4 * 14) << 7,
+      SafeLoadAs<uint32_t>(in + 4 * 14) >> 14 | SafeLoadAs<uint32_t>(in + 4 * 15) << 18,
+      SafeLoadAs<uint32_t>(in + 4 * 15),
+    };
+    shifts = simd_batch{ 4, 0, 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 21-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 15) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 16) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 16) >> 13 | SafeLoadAs<uint32_t>(in + 4 * 17) << 19,
+      SafeLoadAs<uint32_t>(in + 4 * 17),
+      SafeLoadAs<uint32_t>(in + 4 * 17) >> 23 | SafeLoadAs<uint32_t>(in + 4 * 18) << 9,
+    };
+    shifts = simd_batch{ 0, 0, 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 21-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 18) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 19) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 19),
+      SafeLoadAs<uint32_t>(in + 4 * 19) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 20) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 20),
+    };
+    shifts = simd_batch{ 0, 1, 0, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 21 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint32_t, 22> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 4>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x3fffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 22-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 1) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 2) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+    };
+    shifts = simd_batch{ 0, 0, 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 22-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 3) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 14 | SafeLoadAs<uint32_t>(in + 4 * 4) << 18,
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 5) << 6,
+    };
+    shifts = simd_batch{ 0, 0, 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 22-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 6) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 7) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 8) << 14,
+    };
+    shifts = simd_batch{ 0, 6, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 22-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 9) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 10) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+    };
+    shifts = simd_batch{ 8, 0, 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 22-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 11),
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 12) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 13) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 13),
+    };
+    shifts = simd_batch{ 0, 0, 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 22-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 13) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 14) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 14) >> 14 | SafeLoadAs<uint32_t>(in + 4 * 15) << 18,
+      SafeLoadAs<uint32_t>(in + 4 * 15),
+      SafeLoadAs<uint32_t>(in + 4 * 15) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 16) << 6,
+    };
+    shifts = simd_batch{ 0, 0, 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 22-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 16) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 17) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 17),
+      SafeLoadAs<uint32_t>(in + 4 * 17) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 18) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 18) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 19) << 14,
+    };
+    shifts = simd_batch{ 0, 6, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 22-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 19),
+      SafeLoadAs<uint32_t>(in + 4 * 19) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 20) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 20) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 21) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 21),
+    };
+    shifts = simd_batch{ 8, 0, 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 22 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint32_t, 23> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 4>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x7fffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 23-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 23 | SafeLoadAs<uint32_t>(in + 4 * 1) << 9,
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 14 | SafeLoadAs<uint32_t>(in + 4 * 2) << 18,
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+    };
+    shifts = simd_batch{ 0, 0, 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 23-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 3) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 19 | SafeLoadAs<uint32_t>(in + 4 * 4) << 13,
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 10 | SafeLoadAs<uint32_t>(in + 4 * 5) << 22,
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+    };
+    shifts = simd_batch{ 0, 0, 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 23-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 6) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 15 | SafeLoadAs<uint32_t>(in + 4 * 7) << 17,
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 29 | SafeLoadAs<uint32_t>(in + 4 * 8) << 3,
+    };
+    shifts = simd_batch{ 0, 0, 6, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 23-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 9) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 11 | SafeLoadAs<uint32_t>(in + 4 * 10) << 21,
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 25 | SafeLoadAs<uint32_t>(in + 4 * 11) << 7,
+    };
+    shifts = simd_batch{ 0, 0, 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 23-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 12) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 12),
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 13) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 13) >> 21 | SafeLoadAs<uint32_t>(in + 4 * 14) << 11,
+    };
+    shifts = simd_batch{ 0, 7, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 23-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 14) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 15) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 15),
+      SafeLoadAs<uint32_t>(in + 4 * 15) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 16) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 16) >> 17 | SafeLoadAs<uint32_t>(in + 4 * 17) << 15,
+    };
+    shifts = simd_batch{ 0, 3, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 23-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 17),
+      SafeLoadAs<uint32_t>(in + 4 * 17) >> 31 | SafeLoadAs<uint32_t>(in + 4 * 18) << 1,
+      SafeLoadAs<uint32_t>(in + 4 * 18) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 19) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 19) >> 13 | SafeLoadAs<uint32_t>(in + 4 * 20) << 19,
+    };
+    shifts = simd_batch{ 8, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 23-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 20),
+      SafeLoadAs<uint32_t>(in + 4 * 20) >> 27 | SafeLoadAs<uint32_t>(in + 4 * 21) << 5,
+      SafeLoadAs<uint32_t>(in + 4 * 21) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 22) << 14,
+      SafeLoadAs<uint32_t>(in + 4 * 22),
+    };
+    shifts = simd_batch{ 4, 0, 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 23 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint32_t, 24> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 4>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0xffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 24-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 1) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 2) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+    };
+    shifts = simd_batch{ 0, 0, 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 24-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 4) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 5) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+    };
+    shifts = simd_batch{ 0, 0, 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 24-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 7) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 8) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+    };
+    shifts = simd_batch{ 0, 0, 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 24-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 10) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 11) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 11),
+    };
+    shifts = simd_batch{ 0, 0, 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 24-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 12),
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 13) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 13) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 14) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 14),
+    };
+    shifts = simd_batch{ 0, 0, 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 24-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 15),
+      SafeLoadAs<uint32_t>(in + 4 * 15) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 16) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 16) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 17) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 17),
+    };
+    shifts = simd_batch{ 0, 0, 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 24-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 18),
+      SafeLoadAs<uint32_t>(in + 4 * 18) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 19) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 19) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 20) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 20),
+    };
+    shifts = simd_batch{ 0, 0, 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 24-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 21),
+      SafeLoadAs<uint32_t>(in + 4 * 21) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 22) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 22) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 23) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 23),
+    };
+    shifts = simd_batch{ 0, 0, 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 24 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint32_t, 25> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 4>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x1ffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 25-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 25 | SafeLoadAs<uint32_t>(in + 4 * 1) << 7,
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 2) << 14,
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 11 | SafeLoadAs<uint32_t>(in + 4 * 3) << 21,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 25-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 29 | SafeLoadAs<uint32_t>(in + 4 * 4) << 3,
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 5) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 15 | SafeLoadAs<uint32_t>(in + 4 * 6) << 17,
+    };
+    shifts = simd_batch{ 4, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 25-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 8 | SafeLoadAs<uint32_t>(in + 4 * 7) << 24,
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 8) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 19 | SafeLoadAs<uint32_t>(in + 4 * 9) << 13,
+    };
+    shifts = simd_batch{ 0, 1, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 25-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 10) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 11) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 23 | SafeLoadAs<uint32_t>(in + 4 * 12) << 9,
+    };
+    shifts = simd_batch{ 0, 5, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 25-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 13) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 13) >> 9 | SafeLoadAs<uint32_t>(in + 4 * 14) << 23,
+      SafeLoadAs<uint32_t>(in + 4 * 14),
+      SafeLoadAs<uint32_t>(in + 4 * 14) >> 27 | SafeLoadAs<uint32_t>(in + 4 * 15) << 5,
+    };
+    shifts = simd_batch{ 0, 0, 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 25-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 15) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 16) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 16) >> 13 | SafeLoadAs<uint32_t>(in + 4 * 17) << 19,
+      SafeLoadAs<uint32_t>(in + 4 * 17),
+      SafeLoadAs<uint32_t>(in + 4 * 17) >> 31 | SafeLoadAs<uint32_t>(in + 4 * 18) << 1,
+    };
+    shifts = simd_batch{ 0, 0, 6, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 25-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 18) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 19) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 19) >> 17 | SafeLoadAs<uint32_t>(in + 4 * 20) << 15,
+      SafeLoadAs<uint32_t>(in + 4 * 20) >> 10 | SafeLoadAs<uint32_t>(in + 4 * 21) << 22,
+      SafeLoadAs<uint32_t>(in + 4 * 21),
+    };
+    shifts = simd_batch{ 0, 0, 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 25-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 21) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 22) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 22) >> 21 | SafeLoadAs<uint32_t>(in + 4 * 23) << 11,
+      SafeLoadAs<uint32_t>(in + 4 * 23) >> 14 | SafeLoadAs<uint32_t>(in + 4 * 24) << 18,
+      SafeLoadAs<uint32_t>(in + 4 * 24),
+    };
+    shifts = simd_batch{ 0, 0, 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 25 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint32_t, 26> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 4>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x3ffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 26-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 1) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 2) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 14 | SafeLoadAs<uint32_t>(in + 4 * 3) << 18,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 26-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 8 | SafeLoadAs<uint32_t>(in + 4 * 4) << 24,
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 5) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 6) << 10,
+    };
+    shifts = simd_batch{ 0, 2, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 26-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 7) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 10 | SafeLoadAs<uint32_t>(in + 4 * 8) << 22,
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 9) << 2,
+    };
+    shifts = simd_batch{ 0, 0, 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 26-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 10) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 11) << 14,
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 12) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 12),
+    };
+    shifts = simd_batch{ 0, 0, 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 26-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 13),
+      SafeLoadAs<uint32_t>(in + 4 * 13) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 14) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 14) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 15) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 15) >> 14 | SafeLoadAs<uint32_t>(in + 4 * 16) << 18,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 26-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 16) >> 8 | SafeLoadAs<uint32_t>(in + 4 * 17) << 24,
+      SafeLoadAs<uint32_t>(in + 4 * 17),
+      SafeLoadAs<uint32_t>(in + 4 * 17) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 18) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 18) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 19) << 10,
+    };
+    shifts = simd_batch{ 0, 2, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 26-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 19) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 20) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 20) >> 10 | SafeLoadAs<uint32_t>(in + 4 * 21) << 22,
+      SafeLoadAs<uint32_t>(in + 4 * 21),
+      SafeLoadAs<uint32_t>(in + 4 * 21) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 22) << 2,
+    };
+    shifts = simd_batch{ 0, 0, 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 26-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 22) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 23) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 23) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 24) << 14,
+      SafeLoadAs<uint32_t>(in + 4 * 24) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 25) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 25),
+    };
+    shifts = simd_batch{ 0, 0, 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 26 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint32_t, 27> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 4>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x7ffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 27-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 27 | SafeLoadAs<uint32_t>(in + 4 * 1) << 5,
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 2) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 17 | SafeLoadAs<uint32_t>(in + 4 * 3) << 15,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 27-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 4) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 7 | SafeLoadAs<uint32_t>(in + 4 * 5) << 25,
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 29 | SafeLoadAs<uint32_t>(in + 4 * 6) << 3,
+    };
+    shifts = simd_batch{ 0, 0, 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 27-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 7) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 19 | SafeLoadAs<uint32_t>(in + 4 * 8) << 13,
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 14 | SafeLoadAs<uint32_t>(in + 4 * 9) << 18,
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 9 | SafeLoadAs<uint32_t>(in + 4 * 10) << 23,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 27-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 31 | SafeLoadAs<uint32_t>(in + 4 * 11) << 1,
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 12) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 21 | SafeLoadAs<uint32_t>(in + 4 * 13) << 11,
+    };
+    shifts = simd_batch{ 4, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 27-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 13) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 14) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 14) >> 11 | SafeLoadAs<uint32_t>(in + 4 * 15) << 21,
+      SafeLoadAs<uint32_t>(in + 4 * 15) >> 6 | SafeLoadAs<uint32_t>(in + 4 * 16) << 26,
+      SafeLoadAs<uint32_t>(in + 4 * 16),
+    };
+    shifts = simd_batch{ 0, 0, 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 27-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 16) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 17) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 17) >> 23 | SafeLoadAs<uint32_t>(in + 4 * 18) << 9,
+      SafeLoadAs<uint32_t>(in + 4 * 18) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 19) << 14,
+      SafeLoadAs<uint32_t>(in + 4 * 19) >> 13 | SafeLoadAs<uint32_t>(in + 4 * 20) << 19,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 27-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 20) >> 8 | SafeLoadAs<uint32_t>(in + 4 * 21) << 24,
+      SafeLoadAs<uint32_t>(in + 4 * 21),
+      SafeLoadAs<uint32_t>(in + 4 * 21) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 22) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 22) >> 25 | SafeLoadAs<uint32_t>(in + 4 * 23) << 7,
+    };
+    shifts = simd_batch{ 0, 3, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 27-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 23) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 24) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 24) >> 15 | SafeLoadAs<uint32_t>(in + 4 * 25) << 17,
+      SafeLoadAs<uint32_t>(in + 4 * 25) >> 10 | SafeLoadAs<uint32_t>(in + 4 * 26) << 22,
+      SafeLoadAs<uint32_t>(in + 4 * 26),
+    };
+    shifts = simd_batch{ 0, 0, 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 27 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint32_t, 28> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 4>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0xfffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 28-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 1) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 2) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 3) << 12,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 28-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 4) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 5) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 8 | SafeLoadAs<uint32_t>(in + 4 * 6) << 24,
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+    };
+    shifts = simd_batch{ 0, 0, 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 28-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 8) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 9) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 10) << 12,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 28-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 11) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 12) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 8 | SafeLoadAs<uint32_t>(in + 4 * 13) << 24,
+      SafeLoadAs<uint32_t>(in + 4 * 13),
+    };
+    shifts = simd_batch{ 0, 0, 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 28-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 14),
+      SafeLoadAs<uint32_t>(in + 4 * 14) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 15) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 15) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 16) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 16) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 17) << 12,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 28-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 17) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 18) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 18) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 19) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 19) >> 8 | SafeLoadAs<uint32_t>(in + 4 * 20) << 24,
+      SafeLoadAs<uint32_t>(in + 4 * 20),
+    };
+    shifts = simd_batch{ 0, 0, 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 28-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 21),
+      SafeLoadAs<uint32_t>(in + 4 * 21) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 22) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 22) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 23) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 23) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 24) << 12,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 28-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 24) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 25) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 25) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 26) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 26) >> 8 | SafeLoadAs<uint32_t>(in + 4 * 27) << 24,
+      SafeLoadAs<uint32_t>(in + 4 * 27),
+    };
+    shifts = simd_batch{ 0, 0, 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 28 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint32_t, 29> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 4>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x1fffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 29-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 29 | SafeLoadAs<uint32_t>(in + 4 * 1) << 3,
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 2) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 23 | SafeLoadAs<uint32_t>(in + 4 * 3) << 9,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 29-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 4) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 17 | SafeLoadAs<uint32_t>(in + 4 * 5) << 15,
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 14 | SafeLoadAs<uint32_t>(in + 4 * 6) << 18,
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 11 | SafeLoadAs<uint32_t>(in + 4 * 7) << 21,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 29-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 8 | SafeLoadAs<uint32_t>(in + 4 * 8) << 24,
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 5 | SafeLoadAs<uint32_t>(in + 4 * 9) << 27,
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 31 | SafeLoadAs<uint32_t>(in + 4 * 10) << 1,
+    };
+    shifts = simd_batch{ 0, 0, 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 29-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 11) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 25 | SafeLoadAs<uint32_t>(in + 4 * 12) << 7,
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 13) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 13) >> 19 | SafeLoadAs<uint32_t>(in + 4 * 14) << 13,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 29-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 14) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 15) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 15) >> 13 | SafeLoadAs<uint32_t>(in + 4 * 16) << 19,
+      SafeLoadAs<uint32_t>(in + 4 * 16) >> 10 | SafeLoadAs<uint32_t>(in + 4 * 17) << 22,
+      SafeLoadAs<uint32_t>(in + 4 * 17) >> 7 | SafeLoadAs<uint32_t>(in + 4 * 18) << 25,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 29-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 18) >> 4 | SafeLoadAs<uint32_t>(in + 4 * 19) << 28,
+      SafeLoadAs<uint32_t>(in + 4 * 19),
+      SafeLoadAs<uint32_t>(in + 4 * 19) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 20) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 20) >> 27 | SafeLoadAs<uint32_t>(in + 4 * 21) << 5,
+    };
+    shifts = simd_batch{ 0, 1, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 29-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 21) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 22) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 22) >> 21 | SafeLoadAs<uint32_t>(in + 4 * 23) << 11,
+      SafeLoadAs<uint32_t>(in + 4 * 23) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 24) << 14,
+      SafeLoadAs<uint32_t>(in + 4 * 24) >> 15 | SafeLoadAs<uint32_t>(in + 4 * 25) << 17,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 29-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 25) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 26) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 26) >> 9 | SafeLoadAs<uint32_t>(in + 4 * 27) << 23,
+      SafeLoadAs<uint32_t>(in + 4 * 27) >> 6 | SafeLoadAs<uint32_t>(in + 4 * 28) << 26,
+      SafeLoadAs<uint32_t>(in + 4 * 28),
+    };
+    shifts = simd_batch{ 0, 0, 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 29 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint32_t, 30> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 4>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x3fffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 30-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 1) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 2) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 3) << 6,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 30-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 4) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 5) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 6) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 7) << 14,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 30-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 8) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 14 | SafeLoadAs<uint32_t>(in + 4 * 9) << 18,
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 10) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 10 | SafeLoadAs<uint32_t>(in + 4 * 11) << 22,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 30-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 8 | SafeLoadAs<uint32_t>(in + 4 * 12) << 24,
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 6 | SafeLoadAs<uint32_t>(in + 4 * 13) << 26,
+      SafeLoadAs<uint32_t>(in + 4 * 13) >> 4 | SafeLoadAs<uint32_t>(in + 4 * 14) << 28,
+      SafeLoadAs<uint32_t>(in + 4 * 14),
+    };
+    shifts = simd_batch{ 0, 0, 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 30-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 15),
+      SafeLoadAs<uint32_t>(in + 4 * 15) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 16) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 16) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 17) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 17) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 18) << 6,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 30-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 18) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 19) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 19) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 20) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 20) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 21) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 21) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 22) << 14,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 30-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 22) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 23) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 23) >> 14 | SafeLoadAs<uint32_t>(in + 4 * 24) << 18,
+      SafeLoadAs<uint32_t>(in + 4 * 24) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 25) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 25) >> 10 | SafeLoadAs<uint32_t>(in + 4 * 26) << 22,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 30-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 26) >> 8 | SafeLoadAs<uint32_t>(in + 4 * 27) << 24,
+      SafeLoadAs<uint32_t>(in + 4 * 27) >> 6 | SafeLoadAs<uint32_t>(in + 4 * 28) << 26,
+      SafeLoadAs<uint32_t>(in + 4 * 28) >> 4 | SafeLoadAs<uint32_t>(in + 4 * 29) << 28,
+      SafeLoadAs<uint32_t>(in + 4 * 29),
+    };
+    shifts = simd_batch{ 0, 0, 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 30 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint32_t, 31> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 4>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x7fffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 31-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 31 | SafeLoadAs<uint32_t>(in + 4 * 1) << 1,
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 2) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 29 | SafeLoadAs<uint32_t>(in + 4 * 3) << 3,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 31-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 4) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 27 | SafeLoadAs<uint32_t>(in + 4 * 5) << 5,
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 6) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 25 | SafeLoadAs<uint32_t>(in + 4 * 7) << 7,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 31-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 8) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 23 | SafeLoadAs<uint32_t>(in + 4 * 9) << 9,
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 10) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 21 | SafeLoadAs<uint32_t>(in + 4 * 11) << 11,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 31-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 12) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 19 | SafeLoadAs<uint32_t>(in + 4 * 13) << 13,
+      SafeLoadAs<uint32_t>(in + 4 * 13) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 14) << 14,
+      SafeLoadAs<uint32_t>(in + 4 * 14) >> 17 | SafeLoadAs<uint32_t>(in + 4 * 15) << 15,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 31-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 15) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 16) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 16) >> 15 | SafeLoadAs<uint32_t>(in + 4 * 17) << 17,
+      SafeLoadAs<uint32_t>(in + 4 * 17) >> 14 | SafeLoadAs<uint32_t>(in + 4 * 18) << 18,
+      SafeLoadAs<uint32_t>(in + 4 * 18) >> 13 | SafeLoadAs<uint32_t>(in + 4 * 19) << 19,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 31-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 19) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 20) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 20) >> 11 | SafeLoadAs<uint32_t>(in + 4 * 21) << 21,
+      SafeLoadAs<uint32_t>(in + 4 * 21) >> 10 | SafeLoadAs<uint32_t>(in + 4 * 22) << 22,
+      SafeLoadAs<uint32_t>(in + 4 * 22) >> 9 | SafeLoadAs<uint32_t>(in + 4 * 23) << 23,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 31-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 23) >> 8 | SafeLoadAs<uint32_t>(in + 4 * 24) << 24,
+      SafeLoadAs<uint32_t>(in + 4 * 24) >> 7 | SafeLoadAs<uint32_t>(in + 4 * 25) << 25,
+      SafeLoadAs<uint32_t>(in + 4 * 25) >> 6 | SafeLoadAs<uint32_t>(in + 4 * 26) << 26,
+      SafeLoadAs<uint32_t>(in + 4 * 26) >> 5 | SafeLoadAs<uint32_t>(in + 4 * 27) << 27,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 31-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 27) >> 4 | SafeLoadAs<uint32_t>(in + 4 * 28) << 28,
+      SafeLoadAs<uint32_t>(in + 4 * 28) >> 3 | SafeLoadAs<uint32_t>(in + 4 * 29) << 29,
+      SafeLoadAs<uint32_t>(in + 4 * 29) >> 2 | SafeLoadAs<uint32_t>(in + 4 * 30) << 30,
+      SafeLoadAs<uint32_t>(in + 4 * 30),
+    };
+    shifts = simd_batch{ 0, 0, 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 31 * 4;
+    return in;
+  }
+};
+
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 1> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 1-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 1-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 2, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 1-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 4, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 1-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 6, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 1-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 8, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 1-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 10, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 1-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 12, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 1-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 14, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 1-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 16, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 1-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 18, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 1-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 20, 21 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 1-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 22, 23 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 1-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 24, 25 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 1-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 26, 27 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 1-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 28, 29 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 1-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 30, 31 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 1-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 32, 33 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 1-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 34, 35 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 1-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 36, 37 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 1-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 38, 39 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 1-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 40, 41 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 1-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 42, 43 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 1-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 44, 45 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 1-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 46, 47 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 1-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 48, 49 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 1-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 50, 51 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 1-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 52, 53 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 1-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 54, 55 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 1-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 56, 57 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 1-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 58, 59 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 1-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 60, 61 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 1-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 62, 63 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 1 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 2> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 2-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 2-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 4, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 2-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 8, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 2-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 12, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 2-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 16, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 2-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 20, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 2-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 24, 26 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 2-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 28, 30 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 2-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 32, 34 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 2-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 36, 38 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 2-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 40, 42 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 2-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 44, 46 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 2-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 48, 50 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 2-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 52, 54 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 2-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 56, 58 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 2-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 60, 62 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 2-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 2-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 4, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 2-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 8, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 2-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 12, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 2-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 16, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 2-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 20, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 2-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 24, 26 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 2-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 28, 30 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 2-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 32, 34 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 2-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 36, 38 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 2-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 40, 42 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 2-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 44, 46 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 2-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 48, 50 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 2-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 52, 54 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 2-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 56, 58 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 2-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 60, 62 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 2 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 3> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 3-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 3-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 6, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 3-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 12, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 3-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 18, 21 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 3-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 24, 27 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 3-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 30, 33 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 3-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 36, 39 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 3-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 42, 45 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 3-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 48, 51 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 3-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 54, 57 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 3-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 1) << 1,
+    };
+    shifts = simd_batch{ 60, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 3-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 2, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 3-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 8, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 3-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 14, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 3-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 20, 23 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 3-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 26, 29 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 3-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 32, 35 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 3-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 38, 41 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 3-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 44, 47 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 3-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 50, 53 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 3-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 56, 59 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 3-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 2) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 3-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 4, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 3-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 10, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 3-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 16, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 3-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 22, 25 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 3-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 28, 31 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 3-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 34, 37 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 3-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 40, 43 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 3-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 46, 49 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 3-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 52, 55 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 3-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 58, 61 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 3 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 4> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0xf;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 4-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 4-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 8, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 4-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 16, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 4-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 24, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 4-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 32, 36 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 4-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 40, 44 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 4-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 48, 52 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 4-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 56, 60 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 4-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 4-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 8, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 4-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 16, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 4-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 24, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 4-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 32, 36 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 4-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 40, 44 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 4-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 48, 52 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 4-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 56, 60 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 4-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 4-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 8, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 4-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 16, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 4-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 24, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 4-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 32, 36 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 4-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 40, 44 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 4-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 48, 52 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 4-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 56, 60 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 4-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 4-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 8, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 4-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 16, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 4-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 24, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 4-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 32, 36 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 4-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 40, 44 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 4-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 48, 52 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 4-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 56, 60 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 4 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 5> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1f;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 5-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 5-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 10, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 5-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 20, 25 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 5-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 30, 35 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 5-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 40, 45 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 5-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 50, 55 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 5-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 1) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 5-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 6, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 5-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 16, 21 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 5-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 26, 31 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 5-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 36, 41 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 5-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 46, 51 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 5-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 2) << 3,
+    };
+    shifts = simd_batch{ 56, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 5-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 2, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 5-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 12, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 5-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 22, 27 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 5-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 32, 37 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 5-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 42, 47 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 5-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 52, 57 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 5-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 3) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 5-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 8, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 5-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 18, 23 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 5-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 28, 33 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 5-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 38, 43 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 5-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 48, 53 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 5-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 4) << 1,
+    };
+    shifts = simd_batch{ 58, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 5-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 4, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 5-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 14, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 5-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 24, 29 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 5-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 34, 39 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 5-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 44, 49 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 5-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 54, 59 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 5 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 6> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3f;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 6-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 6-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 12, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 6-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 24, 30 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 6-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 36, 42 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 6-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 48, 54 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 6-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 1) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 6-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 8, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 6-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 20, 26 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 6-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 32, 38 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 6-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 44, 50 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 6-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 2) << 2,
+    };
+    shifts = simd_batch{ 56, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 6-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 4, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 6-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 16, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 6-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 28, 34 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 6-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 40, 46 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 6-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 52, 58 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 6-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 6-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 12, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 6-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 24, 30 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 6-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 36, 42 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 6-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 48, 54 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 6-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 4) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 6-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 8, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 6-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 20, 26 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 6-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 32, 38 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 6-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 44, 50 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 6-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 5) << 2,
+    };
+    shifts = simd_batch{ 56, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 6-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 4, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 6-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 16, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 6-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 28, 34 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 6-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 40, 46 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 6-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 52, 58 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 6 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 7> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7f;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 7-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 7-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 14, 21 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 7-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 28, 35 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 7-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 42, 49 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 7-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 1) << 1,
+    };
+    shifts = simd_batch{ 56, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 7-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 6, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 7-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 20, 27 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 7-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 34, 41 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 7-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 48, 55 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 7-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 2) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 7-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 12, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 7-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 26, 33 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 7-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 40, 47 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 7-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 3) << 3,
+    };
+    shifts = simd_batch{ 54, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 7-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 4, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 7-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 18, 25 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 7-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 32, 39 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 7-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 46, 53 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 7-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 4) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 7-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 10, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 7-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 24, 31 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 7-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 38, 45 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 7-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 5) << 5,
+    };
+    shifts = simd_batch{ 52, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 7-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 2, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 7-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 16, 23 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 7-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 30, 37 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 7-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 44, 51 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 7-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 6) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 7-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 8, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 7-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 22, 29 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 7-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 36, 43 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 7-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 50, 57 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 7 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 8> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0xff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 8-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 8-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 16, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 8-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 32, 40 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 8-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 48, 56 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 8-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 8-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 16, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 8-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 32, 40 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 8-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 48, 56 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 8-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 8-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 16, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 8-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 32, 40 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 8-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 48, 56 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 8-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 8-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 16, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 8-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 32, 40 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 8-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 48, 56 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 8-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 8-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 16, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 8-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 32, 40 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 8-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 48, 56 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 8-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 8-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 16, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 8-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 32, 40 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 8-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 48, 56 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 8-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 8-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 16, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 8-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 32, 40 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 8-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 48, 56 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 8-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 8-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 16, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 8-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 32, 40 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 8-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 48, 56 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 8 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 9> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1ff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 9-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 9-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 18, 27 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 9-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 36, 45 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 9-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 1) << 1,
+    };
+    shifts = simd_batch{ 54, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 9-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 8, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 9-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 26, 35 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 9-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 44, 53 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 9-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 2) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 9-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 16, 25 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 9-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 34, 43 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 9-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 3) << 3,
+    };
+    shifts = simd_batch{ 52, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 9-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 6, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 9-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 24, 33 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 9-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 42, 51 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 9-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 4) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 9-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 14, 23 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 9-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 32, 41 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 9-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 5) << 5,
+    };
+    shifts = simd_batch{ 50, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 9-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 4, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 9-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 22, 31 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 9-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 40, 49 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 9-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 6) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 9-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 12, 21 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 9-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 30, 39 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 9-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 7) << 7,
+    };
+    shifts = simd_batch{ 48, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 9-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 2, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 9-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 20, 29 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 9-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 38, 47 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 9-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 8) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 9-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 10, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 9-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 28, 37 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 9-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 46, 55 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 9 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 10> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3ff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 10-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 10-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 20, 30 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 10-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 40, 50 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 10-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 1) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 10-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 16, 26 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 10-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 36, 46 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 10-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 2) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 10-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 12, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 10-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 32, 42 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 10-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 3) << 2,
+    };
+    shifts = simd_batch{ 52, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 10-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 8, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 10-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 28, 38 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 10-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 4) << 6,
+    };
+    shifts = simd_batch{ 48, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 10-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 4, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 10-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 24, 34 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 10-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 44, 54 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 10-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 10-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 20, 30 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 10-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 40, 50 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 10-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 6) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 10-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 16, 26 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 10-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 36, 46 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 10-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 7) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 10-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 12, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 10-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 32, 42 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 10-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 8) << 2,
+    };
+    shifts = simd_batch{ 52, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 10-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 8, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 10-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 28, 38 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 10-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 9) << 6,
+    };
+    shifts = simd_batch{ 48, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 10-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 4, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 10-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 24, 34 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 10-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 44, 54 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 10 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 11> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7ff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 11-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 11-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 22, 33 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 11-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 1) << 9,
+    };
+    shifts = simd_batch{ 44, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 11-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 2, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 11-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 24, 35 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 11-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 2) << 7,
+    };
+    shifts = simd_batch{ 46, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 11-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 4, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 11-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 26, 37 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 11-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 3) << 5,
+    };
+    shifts = simd_batch{ 48, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 11-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 6, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 11-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 28, 39 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 11-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 4) << 3,
+    };
+    shifts = simd_batch{ 50, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 11-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 8, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 11-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 30, 41 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 11-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 5) << 1,
+    };
+    shifts = simd_batch{ 52, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 11-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 10, 21 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 11-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 32, 43 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 11-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 6) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 11-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 12, 23 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 11-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 34, 45 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 11-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 7) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 11-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 14, 25 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 11-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 36, 47 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 11-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 8) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 11-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 16, 27 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 11-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 38, 49 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 11-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 9) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 11-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 18, 29 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 11-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 40, 51 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 11-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 10) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 11-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 20, 31 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 11-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 42, 53 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 11 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 12> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0xfff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 12-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 12-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 24, 36 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 12-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 1) << 4,
+    };
+    shifts = simd_batch{ 48, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 12-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 8, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 12-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 32, 44 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 12-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 2) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 12-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 16, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 12-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 40, 52 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 12-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 12-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 24, 36 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 12-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 4) << 4,
+    };
+    shifts = simd_batch{ 48, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 12-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 8, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 12-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 32, 44 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 12-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 5) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 12-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 16, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 12-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 40, 52 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 12-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 12-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 24, 36 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 12-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 7) << 4,
+    };
+    shifts = simd_batch{ 48, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 12-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 8, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 12-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 32, 44 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 12-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 8) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 12-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 16, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 12-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 40, 52 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 12-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 12-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 24, 36 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 12-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 10) << 4,
+    };
+    shifts = simd_batch{ 48, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 12-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 8, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 12-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 32, 44 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 12-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 11) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 12-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 16, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 12-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 40, 52 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 12 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 13> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1fff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 13-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 13-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 26, 39 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 13-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 1) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 13-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 14, 27 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 13-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 2) << 11,
+    };
+    shifts = simd_batch{ 40, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 13-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 2, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 13-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 28, 41 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 13-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 3) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 13-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 16, 29 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 13-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 4) << 9,
+    };
+    shifts = simd_batch{ 42, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 13-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 4, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 13-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 30, 43 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 13-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 5) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 13-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 18, 31 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 13-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 6) << 7,
+    };
+    shifts = simd_batch{ 44, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 13-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 6, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 13-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 32, 45 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 13-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 7) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 13-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 20, 33 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 13-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 8) << 5,
+    };
+    shifts = simd_batch{ 46, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 13-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 8, 21 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 13-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 34, 47 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 13-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 9) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 13-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 22, 35 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 13-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 10) << 3,
+    };
+    shifts = simd_batch{ 48, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 13-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 10, 23 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 13-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 36, 49 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 13-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 11) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 0, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 13-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 24, 37 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 13-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 12) << 1,
+    };
+    shifts = simd_batch{ 50, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 13-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 12, 25 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 13-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 38, 51 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 13 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 14> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3fff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 14-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 14-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 28, 42 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 14-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 1) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 14-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 20, 34 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 14-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 2) << 2,
+    };
+    shifts = simd_batch{ 48, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 14-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 12, 26 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 14-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 3) << 10,
+    };
+    shifts = simd_batch{ 40, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 14-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 4, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 14-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 32, 46 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 14-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 4) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 14-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 24, 38 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 14-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 5) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 14-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 16, 30 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 14-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 6) << 6,
+    };
+    shifts = simd_batch{ 44, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 14-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 8, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 14-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 36, 50 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 14-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 0, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 14-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 28, 42 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 14-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 8) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 14-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 20, 34 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 14-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 9) << 2,
+    };
+    shifts = simd_batch{ 48, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 14-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 12, 26 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 14-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 10) << 10,
+    };
+    shifts = simd_batch{ 40, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 14-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 4, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 14-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 32, 46 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 14-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 11) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 14-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 24, 38 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 14-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 12) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 14-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 16, 30 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 14-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 13) << 6,
+    };
+    shifts = simd_batch{ 44, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 14-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 8, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 14-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 36, 50 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 14 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 15> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7fff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 15-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 15-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 30, 45 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 15-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 1) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 15-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 26, 41 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 15-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 2) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 15-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 22, 37 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 15-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 3) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 15-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 18, 33 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 15-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 4) << 1,
+    };
+    shifts = simd_batch{ 48, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 15-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 14, 29 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 15-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 5) << 5,
+    };
+    shifts = simd_batch{ 44, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 15-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 10, 25 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 15-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 6) << 9,
+    };
+    shifts = simd_batch{ 40, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 15-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 6, 21 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 15-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 7) << 13,
+    };
+    shifts = simd_batch{ 36, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 15-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 2, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 15-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 32, 47 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 15-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 8) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 0, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 15-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 28, 43 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 15-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 9) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 15-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 24, 39 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 15-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 10) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 15-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 20, 35 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 15-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 11) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 15-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 16, 31 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 15-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 12) << 3,
+    };
+    shifts = simd_batch{ 46, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 15-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 12, 27 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 15-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 13) << 7,
+    };
+    shifts = simd_batch{ 42, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 15-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 8, 23 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 15-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 14) << 11,
+    };
+    shifts = simd_batch{ 38, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 15-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 4, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 15-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 34, 49 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 15 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 16> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0xffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 16-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 16-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 32, 48 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 16-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 16-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 32, 48 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 16-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 16-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 32, 48 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 16-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 16-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 32, 48 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 16-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 16-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 32, 48 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 16-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 16-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 32, 48 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 16-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 16-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 32, 48 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 16-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 16-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 32, 48 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 16-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 16-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 32, 48 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 16-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 16-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 32, 48 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 16-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 16-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 32, 48 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 16-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 16-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 32, 48 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 16-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 16-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 32, 48 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 16-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 16-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 32, 48 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 16-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 16-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 32, 48 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 16-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+    };
+    shifts = simd_batch{ 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 16-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+    };
+    shifts = simd_batch{ 32, 48 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 16 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 17> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1ffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 17-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 17-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 1) << 13,
+    };
+    shifts = simd_batch{ 34, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 17-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 4, 21 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 17-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 2) << 9,
+    };
+    shifts = simd_batch{ 38, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 17-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 8, 25 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 17-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 3) << 5,
+    };
+    shifts = simd_batch{ 42, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 17-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 12, 29 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 17-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 4) << 1,
+    };
+    shifts = simd_batch{ 46, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 17-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 16, 33 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 17-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 5) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 17-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 20, 37 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 17-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 6) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 17-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 24, 41 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 17-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 7) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 0, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 17-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 28, 45 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 17-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 8) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 0, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 17-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 9) << 15,
+    };
+    shifts = simd_batch{ 32, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 17-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 2, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 17-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 10) << 11,
+    };
+    shifts = simd_batch{ 36, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 17-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 6, 23 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 17-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 11) << 7,
+    };
+    shifts = simd_batch{ 40, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 17-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 10, 27 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 17-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 12) << 3,
+    };
+    shifts = simd_batch{ 44, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 17-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 14, 31 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 17-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 13) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 17-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 18, 35 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 17-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 14) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 17-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 22, 39 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 17-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 15) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+    };
+    shifts = simd_batch{ 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 17-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+    };
+    shifts = simd_batch{ 26, 43 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 17-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 16) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+    };
+    shifts = simd_batch{ 0, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 17-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+    };
+    shifts = simd_batch{ 30, 47 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 17 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 18> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3ffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 18-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 18-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 1) << 10,
+    };
+    shifts = simd_batch{ 36, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 18-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 8, 26 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 18-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 2) << 2,
+    };
+    shifts = simd_batch{ 44, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 18-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 16, 34 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 18-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 3) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 18-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 24, 42 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 18-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 4) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 0, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 18-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 5) << 14,
+    };
+    shifts = simd_batch{ 32, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 18-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 4, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 18-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 6) << 6,
+    };
+    shifts = simd_batch{ 40, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 18-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 12, 30 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 18-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 7) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 18-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 20, 38 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 18-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 8) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 18-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 28, 46 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 18-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 0, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 18-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 10) << 10,
+    };
+    shifts = simd_batch{ 36, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 18-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 8, 26 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 18-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 11) << 2,
+    };
+    shifts = simd_batch{ 44, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 18-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 16, 34 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 18-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 12) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 18-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 24, 42 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 18-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 13) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 0, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 18-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 14) << 14,
+    };
+    shifts = simd_batch{ 32, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 18-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 4, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 18-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 15) << 6,
+    };
+    shifts = simd_batch{ 40, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 18-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+    };
+    shifts = simd_batch{ 12, 30 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 18-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 16) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+    };
+    shifts = simd_batch{ 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 18-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+    };
+    shifts = simd_batch{ 20, 38 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 18-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 17) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+    };
+    shifts = simd_batch{ 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 18-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+    };
+    shifts = simd_batch{ 28, 46 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 18 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 19> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7ffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 19-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 19-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 1) << 7,
+    };
+    shifts = simd_batch{ 38, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 19-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 12, 31 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 19-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 2) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 19-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 24, 43 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 19-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 3) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 0, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 19-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 4) << 9,
+    };
+    shifts = simd_batch{ 36, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 19-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 10, 29 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 19-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 5) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 19-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 22, 41 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 19-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 6) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 0, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 19-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 7) << 11,
+    };
+    shifts = simd_batch{ 34, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 19-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 8, 27 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 19-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 8) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 19-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 20, 39 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 19-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 9) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 0, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 19-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 10) << 13,
+    };
+    shifts = simd_batch{ 32, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 19-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 6, 25 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 19-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 11) << 1,
+    };
+    shifts = simd_batch{ 44, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 19-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 18, 37 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 19-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 12) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 0, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 19-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 13) << 15,
+    };
+    shifts = simd_batch{ 30, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 19-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 4, 23 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 19-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 14) << 3,
+    };
+    shifts = simd_batch{ 42, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 19-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 16, 35 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 19-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 15) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+    };
+    shifts = simd_batch{ 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 19-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 16) << 17,
+    };
+    shifts = simd_batch{ 28, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 19-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+    };
+    shifts = simd_batch{ 2, 21 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 19-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 17) << 5,
+    };
+    shifts = simd_batch{ 40, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 19-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+    };
+    shifts = simd_batch{ 14, 33 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 19-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 18) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+    };
+    shifts = simd_batch{ 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 19-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+    };
+    shifts = simd_batch{ 26, 45 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 19 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 20> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0xfffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 20-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 20-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 1) << 4,
+    };
+    shifts = simd_batch{ 40, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 20-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 16, 36 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 20-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 2) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 20-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 3) << 12,
+    };
+    shifts = simd_batch{ 32, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 20-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 8, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 20-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 4) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 20-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 24, 44 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 20-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 0, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 20-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 6) << 4,
+    };
+    shifts = simd_batch{ 40, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 20-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 16, 36 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 20-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 7) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 20-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 8) << 12,
+    };
+    shifts = simd_batch{ 32, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 20-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 8, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 20-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 9) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 20-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 24, 44 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 20-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 0, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 20-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 11) << 4,
+    };
+    shifts = simd_batch{ 40, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 20-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 16, 36 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 20-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 12) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 20-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 13) << 12,
+    };
+    shifts = simd_batch{ 32, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 20-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 8, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 20-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 14) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 20-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 24, 44 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 20-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+    };
+    shifts = simd_batch{ 0, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 20-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 16) << 4,
+    };
+    shifts = simd_batch{ 40, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 20-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+    };
+    shifts = simd_batch{ 16, 36 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 20-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 17) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+    };
+    shifts = simd_batch{ 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 20-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 18) << 12,
+    };
+    shifts = simd_batch{ 32, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 20-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+    };
+    shifts = simd_batch{ 8, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 20-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 19) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+    };
+    shifts = simd_batch{ 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 20-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+    };
+    shifts = simd_batch{ 24, 44 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 20 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 21> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1fffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 21-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 21 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 21-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 1) << 1,
+    };
+    shifts = simd_batch{ 42, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 21-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 20, 41 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 21-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 2) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 0, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 21-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 3) << 3,
+    };
+    shifts = simd_batch{ 40, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 21-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 18, 39 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 21-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 4) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 0, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 21-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 5) << 5,
+    };
+    shifts = simd_batch{ 38, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 21-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 16, 37 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 21-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 6) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 0, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 21-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 7) << 7,
+    };
+    shifts = simd_batch{ 36, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 21-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 14, 35 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 21-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 8) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 0, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 21-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 9) << 9,
+    };
+    shifts = simd_batch{ 34, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 21-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 12, 33 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 21-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 10) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 0, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 21-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 11) << 11,
+    };
+    shifts = simd_batch{ 32, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 21-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 10, 31 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 21-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 12) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 21-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 13) << 13,
+    };
+    shifts = simd_batch{ 30, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 21-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 8, 29 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 21-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 14) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 21-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 15) << 15,
+    };
+    shifts = simd_batch{ 28, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 21-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+    };
+    shifts = simd_batch{ 6, 27 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 21-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 16) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+    };
+    shifts = simd_batch{ 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 21-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 17) << 17,
+    };
+    shifts = simd_batch{ 26, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 21-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+    };
+    shifts = simd_batch{ 4, 25 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 21-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 18) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+    };
+    shifts = simd_batch{ 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 21-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 19) << 19,
+    };
+    shifts = simd_batch{ 24, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 21-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+    };
+    shifts = simd_batch{ 2, 23 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 21-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 20) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+    };
+    shifts = simd_batch{ 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 21-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+    };
+    shifts = simd_batch{ 22, 43 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 21 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 22> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3fffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 22-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 22-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 1) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 22-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 2) << 18,
+    };
+    shifts = simd_batch{ 24, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 22-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 4, 26 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 22-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 3) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 22-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 4) << 14,
+    };
+    shifts = simd_batch{ 28, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 22-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 8, 30 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 22-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 5) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 22-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 6) << 10,
+    };
+    shifts = simd_batch{ 32, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 22-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 12, 34 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 22-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 7) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 0, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 22-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 8) << 6,
+    };
+    shifts = simd_batch{ 36, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 22-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 16, 38 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 22-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 9) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 0, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 22-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 10) << 2,
+    };
+    shifts = simd_batch{ 40, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 22-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 20, 42 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 22-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 0, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 22-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 12) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 22-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 13) << 18,
+    };
+    shifts = simd_batch{ 24, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 22-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 4, 26 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 22-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 14) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 22-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 15) << 14,
+    };
+    shifts = simd_batch{ 28, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 22-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+    };
+    shifts = simd_batch{ 8, 30 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 22-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 16) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+    };
+    shifts = simd_batch{ 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 22-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 17) << 10,
+    };
+    shifts = simd_batch{ 32, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 22-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+    };
+    shifts = simd_batch{ 12, 34 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 22-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 18) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+    };
+    shifts = simd_batch{ 0, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 22-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 19) << 6,
+    };
+    shifts = simd_batch{ 36, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 22-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+    };
+    shifts = simd_batch{ 16, 38 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 22-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 20) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+    };
+    shifts = simd_batch{ 0, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 22-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 21) << 2,
+    };
+    shifts = simd_batch{ 40, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 22-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+    };
+    shifts = simd_batch{ 20, 42 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 22 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 23> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7fffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 23-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 23 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 23-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 1) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 23-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 2) << 13,
+    };
+    shifts = simd_batch{ 28, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 23-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 10, 33 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 23-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 3) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 0, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 23-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 4) << 3,
+    };
+    shifts = simd_batch{ 38, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 23-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 5) << 21,
+    };
+    shifts = simd_batch{ 20, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 23-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 2, 25 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 23-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 6) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 23-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 7) << 11,
+    };
+    shifts = simd_batch{ 30, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 23-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 12, 35 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 23-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 8) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 0, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 23-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 9) << 1,
+    };
+    shifts = simd_batch{ 40, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 23-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 10) << 19,
+    };
+    shifts = simd_batch{ 22, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 23-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 4, 27 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 23-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 11) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 23-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 12) << 9,
+    };
+    shifts = simd_batch{ 32, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 23-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 14, 37 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 23-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 13) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 0, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 23-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 14) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 23-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 15) << 17,
+    };
+    shifts = simd_batch{ 24, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 23-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+    };
+    shifts = simd_batch{ 6, 29 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 23-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 16) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+    };
+    shifts = simd_batch{ 0, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 23-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 17) << 7,
+    };
+    shifts = simd_batch{ 34, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 23-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+    };
+    shifts = simd_batch{ 16, 39 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 23-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 18) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+    };
+    shifts = simd_batch{ 0, 21 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 23-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 19) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+    };
+    shifts = simd_batch{ 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 23-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 20) << 15,
+    };
+    shifts = simd_batch{ 26, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 23-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+    };
+    shifts = simd_batch{ 8, 31 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 23-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 21) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+    };
+    shifts = simd_batch{ 0, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 23-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 22) << 5,
+    };
+    shifts = simd_batch{ 36, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 23-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+    };
+    shifts = simd_batch{ 18, 41 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 23 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 24> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0xffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 24-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 24-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 1) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 24-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 2) << 8,
+    };
+    shifts = simd_batch{ 32, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 24-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 16, 40 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 24-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 0, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 24-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 4) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 24-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 5) << 8,
+    };
+    shifts = simd_batch{ 32, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 24-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 16, 40 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 24-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 0, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 24-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 7) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 24-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 8) << 8,
+    };
+    shifts = simd_batch{ 32, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 24-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 16, 40 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 24-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 0, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 24-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 10) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 24-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 11) << 8,
+    };
+    shifts = simd_batch{ 32, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 24-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 16, 40 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 24-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 0, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 24-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 13) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 24-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 14) << 8,
+    };
+    shifts = simd_batch{ 32, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 24-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 16, 40 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 24-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+    };
+    shifts = simd_batch{ 0, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 24-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 16) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+    };
+    shifts = simd_batch{ 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 24-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 17) << 8,
+    };
+    shifts = simd_batch{ 32, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 24-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+    };
+    shifts = simd_batch{ 16, 40 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 24-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+    };
+    shifts = simd_batch{ 0, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 24-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 19) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+    };
+    shifts = simd_batch{ 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 24-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 20) << 8,
+    };
+    shifts = simd_batch{ 32, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 24-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+    };
+    shifts = simd_batch{ 16, 40 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 24-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+    };
+    shifts = simd_batch{ 0, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 24-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 22) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+    };
+    shifts = simd_batch{ 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 24-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 23) << 8,
+    };
+    shifts = simd_batch{ 32, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 24-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+    };
+    shifts = simd_batch{ 16, 40 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 24 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 25> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1ffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 25-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 25 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 25-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 1) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 25-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 2) << 3,
+    };
+    shifts = simd_batch{ 36, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 25-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 3) << 17,
+    };
+    shifts = simd_batch{ 22, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 25-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 8, 33 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 25-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 4) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 0, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 25-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 5) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 25-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 6) << 9,
+    };
+    shifts = simd_batch{ 30, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 25-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 7) << 23,
+    };
+    shifts = simd_batch{ 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 25-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 2, 27 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 25-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 8) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 0, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 25-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 9) << 1,
+    };
+    shifts = simd_batch{ 38, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 25-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 10) << 15,
+    };
+    shifts = simd_batch{ 24, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 25-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 10, 35 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 25-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 11) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 0, 21 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 25-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 12) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 25-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 13) << 7,
+    };
+    shifts = simd_batch{ 32, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 25-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 14) << 21,
+    };
+    shifts = simd_batch{ 18, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 25-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 4, 29 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 25-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 15) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+    };
+    shifts = simd_batch{ 0, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 25-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 16) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+    };
+    shifts = simd_batch{ 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 25-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 17) << 13,
+    };
+    shifts = simd_batch{ 26, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 25-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+    };
+    shifts = simd_batch{ 12, 37 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 25-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 18) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+    };
+    shifts = simd_batch{ 0, 23 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 25-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 19) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+    };
+    shifts = simd_batch{ 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 25-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 20) << 5,
+    };
+    shifts = simd_batch{ 34, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 25-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 21) << 19,
+    };
+    shifts = simd_batch{ 20, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 25-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+    };
+    shifts = simd_batch{ 6, 31 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 25-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 22) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+    };
+    shifts = simd_batch{ 0, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 25-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 23) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+    };
+    shifts = simd_batch{ 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 25-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 24) << 11,
+    };
+    shifts = simd_batch{ 28, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 25-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+    };
+    shifts = simd_batch{ 14, 39 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 25 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 26> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3ffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 26-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 26 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 26-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 1) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 26-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 2) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 26-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 3) << 10,
+    };
+    shifts = simd_batch{ 28, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 26-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 4) << 22,
+    };
+    shifts = simd_batch{ 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 26-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 4, 30 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 26-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 5) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 0, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 26-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 6) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 26-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 7) << 6,
+    };
+    shifts = simd_batch{ 32, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 26-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 8) << 18,
+    };
+    shifts = simd_batch{ 20, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 26-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 8, 34 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 26-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 9) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 0, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 26-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 10) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 26-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 11) << 2,
+    };
+    shifts = simd_batch{ 36, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 26-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 12) << 14,
+    };
+    shifts = simd_batch{ 24, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 26-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 12, 38 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 26-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 0, 26 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 26-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 14) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 0, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 26-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 15) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+    };
+    shifts = simd_batch{ 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 26-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 16) << 10,
+    };
+    shifts = simd_batch{ 28, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 26-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 17) << 22,
+    };
+    shifts = simd_batch{ 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 26-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+    };
+    shifts = simd_batch{ 4, 30 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 26-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 18) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+    };
+    shifts = simd_batch{ 0, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 26-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 19) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+    };
+    shifts = simd_batch{ 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 26-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 20) << 6,
+    };
+    shifts = simd_batch{ 32, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 26-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 21) << 18,
+    };
+    shifts = simd_batch{ 20, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 26-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+    };
+    shifts = simd_batch{ 8, 34 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 26-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 22) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+    };
+    shifts = simd_batch{ 0, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 26-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 23) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+    };
+    shifts = simd_batch{ 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 26-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 24) << 2,
+    };
+    shifts = simd_batch{ 36, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 26-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 25) << 14,
+    };
+    shifts = simd_batch{ 24, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 26-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+    };
+    shifts = simd_batch{ 12, 38 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 26 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 27> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7ffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 27-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 27 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 27-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 1) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 27-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 2) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 27-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 3) << 3,
+    };
+    shifts = simd_batch{ 34, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 27-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 4) << 13,
+    };
+    shifts = simd_batch{ 24, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 27-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 5) << 23,
+    };
+    shifts = simd_batch{ 14, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 27-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 4, 31 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 27-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 6) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 0, 21 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 27-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 7) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 0, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 27-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 8) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 27-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 9) << 9,
+    };
+    shifts = simd_batch{ 28, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 27-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 10) << 19,
+    };
+    shifts = simd_batch{ 18, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 27-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 8, 35 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 27-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 11) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 0, 25 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 27-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 12) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 0, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 27-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 13) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 27-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 14) << 5,
+    };
+    shifts = simd_batch{ 32, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 27-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 15) << 15,
+    };
+    shifts = simd_batch{ 22, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 27-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 16) << 25,
+    };
+    shifts = simd_batch{ 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 27-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+    };
+    shifts = simd_batch{ 2, 29 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 27-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 17) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+    };
+    shifts = simd_batch{ 0, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 27-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 18) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+    };
+    shifts = simd_batch{ 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 27-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 19) << 1,
+    };
+    shifts = simd_batch{ 36, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 27-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 20) << 11,
+    };
+    shifts = simd_batch{ 26, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 27-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 21) << 21,
+    };
+    shifts = simd_batch{ 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 27-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+    };
+    shifts = simd_batch{ 6, 33 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 27-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 22) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+    };
+    shifts = simd_batch{ 0, 23 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 27-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 23) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+    };
+    shifts = simd_batch{ 0, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 27-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 24) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+    };
+    shifts = simd_batch{ 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 27-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 25) << 7,
+    };
+    shifts = simd_batch{ 30, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 27-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 26) << 17,
+    };
+    shifts = simd_batch{ 20, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 27-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+    };
+    shifts = simd_batch{ 10, 37 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 27 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 28> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0xfffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 28-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 28-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 1) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 28-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 2) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 28-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 3) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 28-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 4) << 4,
+    };
+    shifts = simd_batch{ 32, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 28-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 5) << 12,
+    };
+    shifts = simd_batch{ 24, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 28-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 6) << 20,
+    };
+    shifts = simd_batch{ 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 28-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 8, 36 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 28-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 0, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 28-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 8) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 0, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 28-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 9) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 28-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 10) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 28-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 11) << 4,
+    };
+    shifts = simd_batch{ 32, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 28-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 12) << 12,
+    };
+    shifts = simd_batch{ 24, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 28-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 13) << 20,
+    };
+    shifts = simd_batch{ 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 28-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 8, 36 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 28-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 0, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 28-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 15) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+    };
+    shifts = simd_batch{ 0, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 28-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 16) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+    };
+    shifts = simd_batch{ 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 28-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 17) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+    };
+    shifts = simd_batch{ 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 28-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 18) << 4,
+    };
+    shifts = simd_batch{ 32, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 28-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 19) << 12,
+    };
+    shifts = simd_batch{ 24, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 28-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 20) << 20,
+    };
+    shifts = simd_batch{ 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 28-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+    };
+    shifts = simd_batch{ 8, 36 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 28-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+    };
+    shifts = simd_batch{ 0, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 28-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 22) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+    };
+    shifts = simd_batch{ 0, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 28-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 23) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+    };
+    shifts = simd_batch{ 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 28-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 24) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+    };
+    shifts = simd_batch{ 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 28-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 25) << 4,
+    };
+    shifts = simd_batch{ 32, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 28-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 26) << 12,
+    };
+    shifts = simd_batch{ 24, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 28-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 27) << 20,
+    };
+    shifts = simd_batch{ 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 28-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+    };
+    shifts = simd_batch{ 8, 36 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 28 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 29> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1fffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 29-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 29 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 29-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 1) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 23 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 29-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 2) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 0, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 29-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 3) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 0, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 29-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 4) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 29-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 5) << 1,
+    };
+    shifts = simd_batch{ 34, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 29-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 6) << 7,
+    };
+    shifts = simd_batch{ 28, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 29-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 7) << 13,
+    };
+    shifts = simd_batch{ 22, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 29-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 8) << 19,
+    };
+    shifts = simd_batch{ 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 29-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 9) << 25,
+    };
+    shifts = simd_batch{ 10, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 29-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 4, 33 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 29-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 10) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 0, 27 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 29-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 11) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 0, 21 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 29-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 12) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 0, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 29-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 13) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 29-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 14) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 29-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 15) << 3,
+    };
+    shifts = simd_batch{ 32, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 29-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 16) << 9,
+    };
+    shifts = simd_batch{ 26, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 29-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 17) << 15,
+    };
+    shifts = simd_batch{ 20, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 29-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 18) << 21,
+    };
+    shifts = simd_batch{ 14, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 29-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 19) << 27,
+    };
+    shifts = simd_batch{ 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 29-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+    };
+    shifts = simd_batch{ 2, 31 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 29-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 20) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+    };
+    shifts = simd_batch{ 0, 25 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 29-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 21) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+    };
+    shifts = simd_batch{ 0, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 29-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 22) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+    };
+    shifts = simd_batch{ 0, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 29-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 23) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+    };
+    shifts = simd_batch{ 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 29-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 24) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+    };
+    shifts = simd_batch{ 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 29-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 25) << 5,
+    };
+    shifts = simd_batch{ 30, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 29-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 26) << 11,
+    };
+    shifts = simd_batch{ 24, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 29-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 27) << 17,
+    };
+    shifts = simd_batch{ 18, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 29-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 28) << 23,
+    };
+    shifts = simd_batch{ 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 29-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+    };
+    shifts = simd_batch{ 6, 35 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 29 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 30> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3fffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 30-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 30 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 30-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 1) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 26 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 30-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 2) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 0, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 30-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 3) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 0, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 30-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 4) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 0, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 30-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 5) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 30-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 6) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 30-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 7) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 30-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 8) << 2,
+    };
+    shifts = simd_batch{ 32, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 30-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 9) << 6,
+    };
+    shifts = simd_batch{ 28, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 30-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 10) << 10,
+    };
+    shifts = simd_batch{ 24, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 30-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 11) << 14,
+    };
+    shifts = simd_batch{ 20, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 30-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 12) << 18,
+    };
+    shifts = simd_batch{ 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 30-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 13) << 22,
+    };
+    shifts = simd_batch{ 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 30-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 14) << 26,
+    };
+    shifts = simd_batch{ 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 30-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 4, 34 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 30-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+    };
+    shifts = simd_batch{ 0, 30 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 30-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 16) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+    };
+    shifts = simd_batch{ 0, 26 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 30-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 17) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+    };
+    shifts = simd_batch{ 0, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 30-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 18) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+    };
+    shifts = simd_batch{ 0, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 30-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 19) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+    };
+    shifts = simd_batch{ 0, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 30-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 20) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+    };
+    shifts = simd_batch{ 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 30-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 21) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+    };
+    shifts = simd_batch{ 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 30-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 22) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+    };
+    shifts = simd_batch{ 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 30-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 23) << 2,
+    };
+    shifts = simd_batch{ 32, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 30-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 24) << 6,
+    };
+    shifts = simd_batch{ 28, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 30-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 25) << 10,
+    };
+    shifts = simd_batch{ 24, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 30-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 26) << 14,
+    };
+    shifts = simd_batch{ 20, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 30-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 27) << 18,
+    };
+    shifts = simd_batch{ 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 30-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 28) << 22,
+    };
+    shifts = simd_batch{ 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 30-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 29) << 26,
+    };
+    shifts = simd_batch{ 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 30-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 29),
+      SafeLoadAs<uint64_t>(in + 8 * 29),
+    };
+    shifts = simd_batch{ 4, 34 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 30 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 31> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7fffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 31-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 31 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 31-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 1) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 29 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 31-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 2) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 0, 27 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 31-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 3) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 0, 25 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 31-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 4) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 0, 23 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 31-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 5) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 0, 21 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 31-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 6) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 0, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 31-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 7) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 0, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 31-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 8) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 0, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 31-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 9) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 0, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 31-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 10) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 0, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 31-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 11) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 31-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 12) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 31-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 13) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 31-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 14) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 31-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 15) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+    };
+    shifts = simd_batch{ 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 31-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 16) << 1,
+    };
+    shifts = simd_batch{ 32, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 31-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 17) << 3,
+    };
+    shifts = simd_batch{ 30, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 31-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 18) << 5,
+    };
+    shifts = simd_batch{ 28, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 31-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 19) << 7,
+    };
+    shifts = simd_batch{ 26, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 31-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 20) << 9,
+    };
+    shifts = simd_batch{ 24, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 31-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 21) << 11,
+    };
+    shifts = simd_batch{ 22, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 31-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 22) << 13,
+    };
+    shifts = simd_batch{ 20, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 31-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 23) << 15,
+    };
+    shifts = simd_batch{ 18, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 31-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 24) << 17,
+    };
+    shifts = simd_batch{ 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 31-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 25) << 19,
+    };
+    shifts = simd_batch{ 14, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 31-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 26) << 21,
+    };
+    shifts = simd_batch{ 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 31-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 27) << 23,
+    };
+    shifts = simd_batch{ 10, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 31-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 28) << 25,
+    };
+    shifts = simd_batch{ 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 31-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 29) << 27,
+    };
+    shifts = simd_batch{ 6, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 31-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 29),
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 30) << 29,
+    };
+    shifts = simd_batch{ 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 31-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+    };
+    shifts = simd_batch{ 2, 33 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 31 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 32> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0xffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 32-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 32-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 32-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 32-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 32-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 32-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 32-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 32-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 32-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 32-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 32-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 32-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 32-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 32-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 32-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 32-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+    };
+    shifts = simd_batch{ 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 32-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+    };
+    shifts = simd_batch{ 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 32-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+    };
+    shifts = simd_batch{ 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 32-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+    };
+    shifts = simd_batch{ 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 32-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+    };
+    shifts = simd_batch{ 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 32-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+    };
+    shifts = simd_batch{ 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 32-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+    };
+    shifts = simd_batch{ 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 32-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+    };
+    shifts = simd_batch{ 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 32-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+    };
+    shifts = simd_batch{ 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 32-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+    };
+    shifts = simd_batch{ 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 32-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+    };
+    shifts = simd_batch{ 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 32-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+    };
+    shifts = simd_batch{ 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 32-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+    };
+    shifts = simd_batch{ 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 32-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+    };
+    shifts = simd_batch{ 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 32-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 29),
+      SafeLoadAs<uint64_t>(in + 8 * 29),
+    };
+    shifts = simd_batch{ 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 32-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+    };
+    shifts = simd_batch{ 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 32-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 31),
+      SafeLoadAs<uint64_t>(in + 8 * 31),
+    };
+    shifts = simd_batch{ 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 32 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 33> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1ffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 33-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 1) << 31,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 33-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 2) << 29,
+    };
+    shifts = simd_batch{ 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 33-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 3) << 27,
+    };
+    shifts = simd_batch{ 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 33-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 4) << 25,
+    };
+    shifts = simd_batch{ 6, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 33-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 5) << 23,
+    };
+    shifts = simd_batch{ 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 33-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 6) << 21,
+    };
+    shifts = simd_batch{ 10, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 33-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 7) << 19,
+    };
+    shifts = simd_batch{ 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 33-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 8) << 17,
+    };
+    shifts = simd_batch{ 14, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 33-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 9) << 15,
+    };
+    shifts = simd_batch{ 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 33-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 10) << 13,
+    };
+    shifts = simd_batch{ 18, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 33-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 11) << 11,
+    };
+    shifts = simd_batch{ 20, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 33-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 12) << 9,
+    };
+    shifts = simd_batch{ 22, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 33-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 13) << 7,
+    };
+    shifts = simd_batch{ 24, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 33-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 14) << 5,
+    };
+    shifts = simd_batch{ 26, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 33-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 15) << 3,
+    };
+    shifts = simd_batch{ 28, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 33-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 16) << 1,
+    };
+    shifts = simd_batch{ 30, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 33-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 17) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+    };
+    shifts = simd_batch{ 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 33-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 18) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+    };
+    shifts = simd_batch{ 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 33-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 19) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+    };
+    shifts = simd_batch{ 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 33-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 20) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+    };
+    shifts = simd_batch{ 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 33-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 21) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+    };
+    shifts = simd_batch{ 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 33-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 22) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+    };
+    shifts = simd_batch{ 0, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 33-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 23) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+    };
+    shifts = simd_batch{ 0, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 33-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 24) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+    };
+    shifts = simd_batch{ 0, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 33-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 25) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+    };
+    shifts = simd_batch{ 0, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 33-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 26) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+    };
+    shifts = simd_batch{ 0, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 33-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 27) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+    };
+    shifts = simd_batch{ 0, 21 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 33-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 28) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+    };
+    shifts = simd_batch{ 0, 23 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 33-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 29) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 29),
+    };
+    shifts = simd_batch{ 0, 25 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 33-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 30) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+    };
+    shifts = simd_batch{ 0, 27 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 33-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 31) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 31),
+    };
+    shifts = simd_batch{ 0, 29 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 33-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 32) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+    };
+    shifts = simd_batch{ 0, 31 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 33 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 34> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3ffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 34-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 1) << 30,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 34-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 2) << 26,
+    };
+    shifts = simd_batch{ 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 34-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 3) << 22,
+    };
+    shifts = simd_batch{ 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 34-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 4) << 18,
+    };
+    shifts = simd_batch{ 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 34-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 5) << 14,
+    };
+    shifts = simd_batch{ 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 34-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 6) << 10,
+    };
+    shifts = simd_batch{ 20, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 34-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 7) << 6,
+    };
+    shifts = simd_batch{ 24, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 34-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 8) << 2,
+    };
+    shifts = simd_batch{ 28, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 34-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 9) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 34-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 10) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 34-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 11) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 34-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 12) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 0, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 34-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 13) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 0, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 34-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 14) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 0, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 34-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 15) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+    };
+    shifts = simd_batch{ 0, 26 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 34-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 16) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+    };
+    shifts = simd_batch{ 0, 30 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 34-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 18) << 30,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 34-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 19) << 26,
+    };
+    shifts = simd_batch{ 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 34-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 20) << 22,
+    };
+    shifts = simd_batch{ 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 34-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 21) << 18,
+    };
+    shifts = simd_batch{ 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 34-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 22) << 14,
+    };
+    shifts = simd_batch{ 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 34-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 23) << 10,
+    };
+    shifts = simd_batch{ 20, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 34-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 24) << 6,
+    };
+    shifts = simd_batch{ 24, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 34-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 25) << 2,
+    };
+    shifts = simd_batch{ 28, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 34-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 26) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+    };
+    shifts = simd_batch{ 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 34-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 27) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+    };
+    shifts = simd_batch{ 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 34-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 28) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+    };
+    shifts = simd_batch{ 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 34-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 29) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 29),
+    };
+    shifts = simd_batch{ 0, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 34-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 30) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+    };
+    shifts = simd_batch{ 0, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 34-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 31) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 31),
+    };
+    shifts = simd_batch{ 0, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 34-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 32) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+    };
+    shifts = simd_batch{ 0, 26 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 34-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 33) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 33),
+    };
+    shifts = simd_batch{ 0, 30 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 34 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 35> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7ffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 35-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 1) << 29,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 35-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 2) << 23,
+    };
+    shifts = simd_batch{ 6, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 35-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 3) << 17,
+    };
+    shifts = simd_batch{ 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 35-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 4) << 11,
+    };
+    shifts = simd_batch{ 18, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 35-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 5) << 5,
+    };
+    shifts = simd_batch{ 24, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 35-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 6) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 35-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 7) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 35-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 8) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 0, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 35-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 9) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 0, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 35-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 10) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 0, 25 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 35-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 11) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 31 | SafeLoadAs<uint64_t>(in + 8 * 12) << 33,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 35-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 13) << 27,
+    };
+    shifts = simd_batch{ 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 35-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 14) << 21,
+    };
+    shifts = simd_batch{ 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 35-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 15) << 15,
+    };
+    shifts = simd_batch{ 14, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 35-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 16) << 9,
+    };
+    shifts = simd_batch{ 20, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 35-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 17) << 3,
+    };
+    shifts = simd_batch{ 26, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 35-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 18) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+    };
+    shifts = simd_batch{ 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 35-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 19) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+    };
+    shifts = simd_batch{ 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 35-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 20) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+    };
+    shifts = simd_batch{ 0, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 35-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 21) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+    };
+    shifts = simd_batch{ 0, 21 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 35-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 22) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+    };
+    shifts = simd_batch{ 0, 27 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 35-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 23) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 24) << 31,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 35-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 25) << 25,
+    };
+    shifts = simd_batch{ 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 35-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 26) << 19,
+    };
+    shifts = simd_batch{ 10, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 35-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 27) << 13,
+    };
+    shifts = simd_batch{ 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 35-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 28) << 7,
+    };
+    shifts = simd_batch{ 22, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 35-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 29) << 1,
+    };
+    shifts = simd_batch{ 28, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 35-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 30) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+    };
+    shifts = simd_batch{ 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 35-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 31) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 31),
+    };
+    shifts = simd_batch{ 0, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 35-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 32) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+    };
+    shifts = simd_batch{ 0, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 35-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 33) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 33),
+    };
+    shifts = simd_batch{ 0, 23 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 35-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 34) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 34),
+    };
+    shifts = simd_batch{ 0, 29 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 35 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 36> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0xfffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 36-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 1) << 28,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 36-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 2) << 20,
+    };
+    shifts = simd_batch{ 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 36-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 3) << 12,
+    };
+    shifts = simd_batch{ 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 36-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 4) << 4,
+    };
+    shifts = simd_batch{ 24, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 36-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 5) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 36-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 6) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 36-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 7) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 0, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 36-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 8) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 0, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 36-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 10) << 28,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 36-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 11) << 20,
+    };
+    shifts = simd_batch{ 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 36-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 12) << 12,
+    };
+    shifts = simd_batch{ 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 36-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 13) << 4,
+    };
+    shifts = simd_batch{ 24, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 36-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 14) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 36-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 15) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+    };
+    shifts = simd_batch{ 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 36-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 16) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+    };
+    shifts = simd_batch{ 0, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 36-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 17) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+    };
+    shifts = simd_batch{ 0, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 36-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 19) << 28,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 36-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 20) << 20,
+    };
+    shifts = simd_batch{ 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 36-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 21) << 12,
+    };
+    shifts = simd_batch{ 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 36-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 22) << 4,
+    };
+    shifts = simd_batch{ 24, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 36-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 23) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+    };
+    shifts = simd_batch{ 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 36-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 24) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+    };
+    shifts = simd_batch{ 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 36-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 25) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+    };
+    shifts = simd_batch{ 0, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 36-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 26) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+    };
+    shifts = simd_batch{ 0, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 36-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 28) << 28,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 36-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 29) << 20,
+    };
+    shifts = simd_batch{ 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 36-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 29),
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 30) << 12,
+    };
+    shifts = simd_batch{ 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 36-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 31) << 4,
+    };
+    shifts = simd_batch{ 24, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 36-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 32) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+    };
+    shifts = simd_batch{ 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 36-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 33) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 33),
+    };
+    shifts = simd_batch{ 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 36-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 34) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 34),
+    };
+    shifts = simd_batch{ 0, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 36-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 35) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+    };
+    shifts = simd_batch{ 0, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 36 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 37> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1fffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 37-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 1) << 27,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 37-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 2) << 17,
+    };
+    shifts = simd_batch{ 10, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 37-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 3) << 7,
+    };
+    shifts = simd_batch{ 20, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 37-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 4) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 37-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 5) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 0, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 37-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 6) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 0, 23 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 37-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 7) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 8) << 31,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 37-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 9) << 21,
+    };
+    shifts = simd_batch{ 6, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 37-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 10) << 11,
+    };
+    shifts = simd_batch{ 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 37-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 11) << 1,
+    };
+    shifts = simd_batch{ 26, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 37-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 12) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 37-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 13) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 0, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 37-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 14) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 29 | SafeLoadAs<uint64_t>(in + 8 * 15) << 35,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 37-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 16) << 25,
+    };
+    shifts = simd_batch{ 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 37-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 17) << 15,
+    };
+    shifts = simd_batch{ 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 37-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 18) << 5,
+    };
+    shifts = simd_batch{ 22, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 37-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 19) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+    };
+    shifts = simd_batch{ 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 37-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 20) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+    };
+    shifts = simd_batch{ 0, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 37-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 21) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+    };
+    shifts = simd_batch{ 0, 25 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 37-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 22) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 23) << 29,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 37-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 24) << 19,
+    };
+    shifts = simd_batch{ 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 37-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 25) << 9,
+    };
+    shifts = simd_batch{ 18, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 37-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 26) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+    };
+    shifts = simd_batch{ 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 37-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 27) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+    };
+    shifts = simd_batch{ 0, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 37-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 28) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+    };
+    shifts = simd_batch{ 0, 21 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 37-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 29) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 31 | SafeLoadAs<uint64_t>(in + 8 * 30) << 33,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 37-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 31) << 23,
+    };
+    shifts = simd_batch{ 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 37-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 31),
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 32) << 13,
+    };
+    shifts = simd_batch{ 14, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 37-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 33) << 3,
+    };
+    shifts = simd_batch{ 24, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 37-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 34) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 34),
+    };
+    shifts = simd_batch{ 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 37-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 35) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+    };
+    shifts = simd_batch{ 0, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 37-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 36) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 36),
+    };
+    shifts = simd_batch{ 0, 27 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 37 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 38> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3fffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 38-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 1) << 26,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 38-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 2) << 14,
+    };
+    shifts = simd_batch{ 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 38-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 3) << 2,
+    };
+    shifts = simd_batch{ 24, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 38-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 4) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 38-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 5) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 0, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 38-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 6) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 7) << 30,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 38-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 8) << 18,
+    };
+    shifts = simd_batch{ 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 38-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 9) << 6,
+    };
+    shifts = simd_batch{ 20, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 38-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 10) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 38-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 11) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 0, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 38-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 12) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 13) << 34,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 38-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 14) << 22,
+    };
+    shifts = simd_batch{ 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 38-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 15) << 10,
+    };
+    shifts = simd_batch{ 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 38-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 16) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+    };
+    shifts = simd_batch{ 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 38-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 17) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+    };
+    shifts = simd_batch{ 0, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 38-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 18) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+    };
+    shifts = simd_batch{ 0, 26 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 38-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 20) << 26,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 38-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 21) << 14,
+    };
+    shifts = simd_batch{ 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 38-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 22) << 2,
+    };
+    shifts = simd_batch{ 24, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 38-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 23) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+    };
+    shifts = simd_batch{ 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 38-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 24) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+    };
+    shifts = simd_batch{ 0, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 38-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 25) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 26) << 30,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 38-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 27) << 18,
+    };
+    shifts = simd_batch{ 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 38-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 28) << 6,
+    };
+    shifts = simd_batch{ 20, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 38-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 29) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 29),
+    };
+    shifts = simd_batch{ 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 38-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 30) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+    };
+    shifts = simd_batch{ 0, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 38-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 31) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 32) << 34,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 38-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 33) << 22,
+    };
+    shifts = simd_batch{ 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 38-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 33),
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 34) << 10,
+    };
+    shifts = simd_batch{ 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 38-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 35) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+    };
+    shifts = simd_batch{ 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 38-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 36) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 36),
+    };
+    shifts = simd_batch{ 0, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 38-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 37) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 37),
+    };
+    shifts = simd_batch{ 0, 26 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 38 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 39> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7fffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 39-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 1) << 25,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 39-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 2) << 11,
+    };
+    shifts = simd_batch{ 14, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 39-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 3) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 39-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 4) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 0, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 39-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 5) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 31 | SafeLoadAs<uint64_t>(in + 8 * 6) << 33,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 39-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 7) << 19,
+    };
+    shifts = simd_batch{ 6, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 39-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 8) << 5,
+    };
+    shifts = simd_batch{ 20, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 39-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 9) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 39-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 10) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 0, 23 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 39-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 11) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 12) << 27,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 39-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 13) << 13,
+    };
+    shifts = simd_batch{ 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 39-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 14) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 39-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 15) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+    };
+    shifts = simd_batch{ 0, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 39-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 16) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 29 | SafeLoadAs<uint64_t>(in + 8 * 17) << 35,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 39-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 18) << 21,
+    };
+    shifts = simd_batch{ 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 39-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 19) << 7,
+    };
+    shifts = simd_batch{ 18, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 39-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 20) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+    };
+    shifts = simd_batch{ 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 39-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 21) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+    };
+    shifts = simd_batch{ 0, 21 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 39-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 22) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 23) << 29,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 39-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 24) << 15,
+    };
+    shifts = simd_batch{ 10, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 39-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 25) << 1,
+    };
+    shifts = simd_batch{ 24, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 39-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 26) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+    };
+    shifts = simd_batch{ 0, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 39-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 27) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 27 | SafeLoadAs<uint64_t>(in + 8 * 28) << 37,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 39-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 29) << 23,
+    };
+    shifts = simd_batch{ 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 39-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 29),
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 30) << 9,
+    };
+    shifts = simd_batch{ 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 39-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 31) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 31),
+    };
+    shifts = simd_batch{ 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 39-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 32) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+    };
+    shifts = simd_batch{ 0, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 39-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 33) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 34) << 31,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 39-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 34),
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 35) << 17,
+    };
+    shifts = simd_batch{ 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 39-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 36) << 3,
+    };
+    shifts = simd_batch{ 22, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 39-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 37) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 37),
+    };
+    shifts = simd_batch{ 0, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 39-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 38) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 38),
+    };
+    shifts = simd_batch{ 0, 25 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 39 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 40> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0xffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 40-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 1) << 24,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 40-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 2) << 8,
+    };
+    shifts = simd_batch{ 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 40-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 3) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 40-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 4) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 0, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 40-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 6) << 24,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 40-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 7) << 8,
+    };
+    shifts = simd_batch{ 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 40-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 8) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 40-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 9) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 0, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 40-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 11) << 24,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 40-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 12) << 8,
+    };
+    shifts = simd_batch{ 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 40-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 13) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 40-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 14) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 0, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 40-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 16) << 24,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 40-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 17) << 8,
+    };
+    shifts = simd_batch{ 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 40-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 18) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+    };
+    shifts = simd_batch{ 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 40-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 19) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+    };
+    shifts = simd_batch{ 0, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 40-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 21) << 24,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 40-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 22) << 8,
+    };
+    shifts = simd_batch{ 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 40-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 23) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+    };
+    shifts = simd_batch{ 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 40-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 24) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+    };
+    shifts = simd_batch{ 0, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 40-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 26) << 24,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 40-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 27) << 8,
+    };
+    shifts = simd_batch{ 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 40-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 28) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+    };
+    shifts = simd_batch{ 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 40-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 29) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 29),
+    };
+    shifts = simd_batch{ 0, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 40-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 31) << 24,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 40-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 31),
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 32) << 8,
+    };
+    shifts = simd_batch{ 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 40-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 33) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 33),
+    };
+    shifts = simd_batch{ 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 40-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 34) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 34),
+    };
+    shifts = simd_batch{ 0, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 40-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 36) << 24,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 40-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 36),
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 37) << 8,
+    };
+    shifts = simd_batch{ 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 40-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 38) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 38),
+    };
+    shifts = simd_batch{ 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 40-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 39) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 39),
+    };
+    shifts = simd_batch{ 0, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 40 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 41> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1ffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 41-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 1) << 23,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 41-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 2) << 5,
+    };
+    shifts = simd_batch{ 18, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 41-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 3) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 0, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 41-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 4) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 31 | SafeLoadAs<uint64_t>(in + 8 * 5) << 33,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 41-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 6) << 15,
+    };
+    shifts = simd_batch{ 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 41-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 7) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 41-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 8) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 0, 21 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 41-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 9) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 10) << 25,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 41-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 11) << 7,
+    };
+    shifts = simd_batch{ 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 41-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 12) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 0, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 41-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 13) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 29 | SafeLoadAs<uint64_t>(in + 8 * 14) << 35,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 41-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 15) << 17,
+    };
+    shifts = simd_batch{ 6, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 41-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 16) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+    };
+    shifts = simd_batch{ 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 41-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 17) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+    };
+    shifts = simd_batch{ 0, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 41-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 18) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 19) << 27,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 41-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 20) << 9,
+    };
+    shifts = simd_batch{ 14, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 41-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 21) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+    };
+    shifts = simd_batch{ 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 41-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 22) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 27 | SafeLoadAs<uint64_t>(in + 8 * 23) << 37,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 41-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 24) << 19,
+    };
+    shifts = simd_batch{ 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 41-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 25) << 1,
+    };
+    shifts = simd_batch{ 22, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 41-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 26) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+    };
+    shifts = simd_batch{ 0, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 41-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 27) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 28) << 29,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 41-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 29) << 11,
+    };
+    shifts = simd_batch{ 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 41-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 30) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+    };
+    shifts = simd_batch{ 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 41-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 31) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 25 | SafeLoadAs<uint64_t>(in + 8 * 32) << 39,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 41-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 33) << 21,
+    };
+    shifts = simd_batch{ 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 41-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 33),
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 34) << 3,
+    };
+    shifts = simd_batch{ 20, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 41-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 35) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+    };
+    shifts = simd_batch{ 0, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 41-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 36) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 37) << 31,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 41-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 37),
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 38) << 13,
+    };
+    shifts = simd_batch{ 10, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 41-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 39) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 39),
+    };
+    shifts = simd_batch{ 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 41-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 40) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 40),
+    };
+    shifts = simd_batch{ 0, 23 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 41 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 42> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3ffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 42-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 1) << 22,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 42-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 2) << 2,
+    };
+    shifts = simd_batch{ 20, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 42-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 3) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 0, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 42-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 4) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 5) << 26,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 42-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 6) << 6,
+    };
+    shifts = simd_batch{ 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 42-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 7) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 0, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 42-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 8) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 9) << 30,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 42-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 10) << 10,
+    };
+    shifts = simd_batch{ 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 42-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 11) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 42-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 12) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 13) << 34,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 42-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 14) << 14,
+    };
+    shifts = simd_batch{ 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 42-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 15) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+    };
+    shifts = simd_batch{ 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 42-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 16) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 17) << 38,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 42-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 18) << 18,
+    };
+    shifts = simd_batch{ 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 42-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 19) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+    };
+    shifts = simd_batch{ 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 42-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 20) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+    };
+    shifts = simd_batch{ 0, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 42-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 22) << 22,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 42-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 23) << 2,
+    };
+    shifts = simd_batch{ 20, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 42-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 24) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+    };
+    shifts = simd_batch{ 0, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 42-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 25) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 26) << 26,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 42-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 27) << 6,
+    };
+    shifts = simd_batch{ 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 42-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 28) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+    };
+    shifts = simd_batch{ 0, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 42-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 29) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 30) << 30,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 42-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 31) << 10,
+    };
+    shifts = simd_batch{ 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 42-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 32) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+    };
+    shifts = simd_batch{ 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 42-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 33) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 34) << 34,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 42-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 34),
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 35) << 14,
+    };
+    shifts = simd_batch{ 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 42-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 36) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 36),
+    };
+    shifts = simd_batch{ 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 42-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 37) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 38) << 38,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 42-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 38),
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 39) << 18,
+    };
+    shifts = simd_batch{ 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 42-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 40) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 40),
+    };
+    shifts = simd_batch{ 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 42-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 41) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 41),
+    };
+    shifts = simd_batch{ 0, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 42 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 43> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7ffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 43-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 1) << 21,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 43-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 2) << 42,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 43-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 3) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 23 | SafeLoadAs<uint64_t>(in + 8 * 4) << 41,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 43-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 5) << 19,
+    };
+    shifts = simd_batch{ 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 43-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 6) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 43-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 7) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 25 | SafeLoadAs<uint64_t>(in + 8 * 8) << 39,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 43-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 9) << 17,
+    };
+    shifts = simd_batch{ 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 43-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 10) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 43-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 11) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 27 | SafeLoadAs<uint64_t>(in + 8 * 12) << 37,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 43-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 13) << 15,
+    };
+    shifts = simd_batch{ 6, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 43-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 14) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 43-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 15) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 29 | SafeLoadAs<uint64_t>(in + 8 * 16) << 35,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 43-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 17) << 13,
+    };
+    shifts = simd_batch{ 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 43-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 18) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+    };
+    shifts = simd_batch{ 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 43-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 19) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 31 | SafeLoadAs<uint64_t>(in + 8 * 20) << 33,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 43-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 21) << 11,
+    };
+    shifts = simd_batch{ 10, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 43-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 22) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+    };
+    shifts = simd_batch{ 0, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 43-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 23) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 24) << 31,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 43-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 25) << 9,
+    };
+    shifts = simd_batch{ 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 43-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 26) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+    };
+    shifts = simd_batch{ 0, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 43-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 27) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 28) << 29,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 43-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 29) << 7,
+    };
+    shifts = simd_batch{ 14, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 43-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 30) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+    };
+    shifts = simd_batch{ 0, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 43-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 31) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 32) << 27,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 43-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 33) << 5,
+    };
+    shifts = simd_batch{ 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 43-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 34) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 34),
+    };
+    shifts = simd_batch{ 0, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 43-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 35) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 36) << 25,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 43-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 36),
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 37) << 3,
+    };
+    shifts = simd_batch{ 18, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 43-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 38) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 38),
+    };
+    shifts = simd_batch{ 0, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 43-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 39) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 40) << 23,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 43-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 40),
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 41) << 1,
+    };
+    shifts = simd_batch{ 20, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 43-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 42) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 42),
+    };
+    shifts = simd_batch{ 0, 21 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 43 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 44> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0xfffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 44-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 1) << 20,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 44-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 2) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 44-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 3) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 4) << 36,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 44-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 5) << 12,
+    };
+    shifts = simd_batch{ 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 44-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 6) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 44-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 7) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 8) << 28,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 44-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 9) << 4,
+    };
+    shifts = simd_batch{ 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 44-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 10) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 0, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 44-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 12) << 20,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 44-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 13) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 44-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 14) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 15) << 36,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 44-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 16) << 12,
+    };
+    shifts = simd_batch{ 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 44-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 17) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+    };
+    shifts = simd_batch{ 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 44-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 18) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 19) << 28,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 44-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 20) << 4,
+    };
+    shifts = simd_batch{ 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 44-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 21) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+    };
+    shifts = simd_batch{ 0, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 44-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 23) << 20,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 44-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 24) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+    };
+    shifts = simd_batch{ 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 44-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 25) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 26) << 36,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 44-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 27) << 12,
+    };
+    shifts = simd_batch{ 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 44-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 28) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+    };
+    shifts = simd_batch{ 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 44-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 29) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 30) << 28,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 44-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 31) << 4,
+    };
+    shifts = simd_batch{ 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 44-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 32) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+    };
+    shifts = simd_batch{ 0, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 44-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 33),
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 34) << 20,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 44-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 35) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+    };
+    shifts = simd_batch{ 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 44-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 36) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 37) << 36,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 44-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 37),
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 38) << 12,
+    };
+    shifts = simd_batch{ 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 44-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 39) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 39),
+    };
+    shifts = simd_batch{ 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 44-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 40) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 41) << 28,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 44-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 41),
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 42) << 4,
+    };
+    shifts = simd_batch{ 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 44-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 43) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 43),
+    };
+    shifts = simd_batch{ 0, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 44 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 45> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1fffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 45-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 1) << 19,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 45-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 2) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 45-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 3) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 4) << 31,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 45-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 5) << 5,
+    };
+    shifts = simd_batch{ 14, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 45-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 6) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 21 | SafeLoadAs<uint64_t>(in + 8 * 7) << 43,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 45-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 8) << 17,
+    };
+    shifts = simd_batch{ 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 45-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 9) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 45-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 10) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 11) << 29,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 45-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 12) << 3,
+    };
+    shifts = simd_batch{ 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 45-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 13) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 23 | SafeLoadAs<uint64_t>(in + 8 * 14) << 41,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 45-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 15) << 15,
+    };
+    shifts = simd_batch{ 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 45-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 16) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+    };
+    shifts = simd_batch{ 0, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 45-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 17) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 18) << 27,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 45-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 19) << 1,
+    };
+    shifts = simd_batch{ 18, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 45-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 20) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 25 | SafeLoadAs<uint64_t>(in + 8 * 21) << 39,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 45-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 22) << 13,
+    };
+    shifts = simd_batch{ 6, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 45-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 23) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+    };
+    shifts = simd_batch{ 0, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 45-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 24) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 25) << 25,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 45-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 26) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+    };
+    shifts = simd_batch{ 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 45-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 27) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 27 | SafeLoadAs<uint64_t>(in + 8 * 28) << 37,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 45-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 29) << 11,
+    };
+    shifts = simd_batch{ 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 45-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 30) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+    };
+    shifts = simd_batch{ 0, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 45-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 31) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 32) << 23,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 45-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 33) << 42,
+      SafeLoadAs<uint64_t>(in + 8 * 33),
+    };
+    shifts = simd_batch{ 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 45-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 34) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 29 | SafeLoadAs<uint64_t>(in + 8 * 35) << 35,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 45-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 36) << 9,
+    };
+    shifts = simd_batch{ 10, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 45-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 37) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 37),
+    };
+    shifts = simd_batch{ 0, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 45-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 38) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 39) << 21,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 45-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 40) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 40),
+    };
+    shifts = simd_batch{ 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 45-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 41) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 31 | SafeLoadAs<uint64_t>(in + 8 * 42) << 33,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 45-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 42),
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 43) << 7,
+    };
+    shifts = simd_batch{ 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 45-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 44) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 44),
+    };
+    shifts = simd_batch{ 0, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 45 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 46> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3fffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 46-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 1) << 18,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 46-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 2) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 46-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 3) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 4) << 26,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 46-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 5) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 46-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 6) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 7) << 34,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 46-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 8) << 6,
+    };
+    shifts = simd_batch{ 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 46-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 9) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 10) << 42,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 46-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 11) << 14,
+    };
+    shifts = simd_batch{ 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 46-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 12) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 0, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 46-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 13) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 14) << 22,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 46-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 15) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+    };
+    shifts = simd_batch{ 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 46-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 16) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 17) << 30,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 46-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 18) << 2,
+    };
+    shifts = simd_batch{ 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 46-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 19) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 20) << 38,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 46-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 21) << 10,
+    };
+    shifts = simd_batch{ 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 46-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 22) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+    };
+    shifts = simd_batch{ 0, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 46-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 24) << 18,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 46-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 25) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+    };
+    shifts = simd_batch{ 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 46-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 26) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 27) << 26,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 46-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 28) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+    };
+    shifts = simd_batch{ 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 46-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 29) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 30) << 34,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 46-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 31) << 6,
+    };
+    shifts = simd_batch{ 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 46-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 32) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 33) << 42,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 46-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 33),
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 34) << 14,
+    };
+    shifts = simd_batch{ 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 46-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 35) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+    };
+    shifts = simd_batch{ 0, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 46-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 36) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 37) << 22,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 46-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 38) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 38),
+    };
+    shifts = simd_batch{ 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 46-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 39) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 40) << 30,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 46-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 40),
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 41) << 2,
+    };
+    shifts = simd_batch{ 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 46-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 42) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 43) << 38,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 46-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 43),
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 44) << 10,
+    };
+    shifts = simd_batch{ 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 46-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 45) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 45),
+    };
+    shifts = simd_batch{ 0, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 46 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 47> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7fffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 47-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 1) << 17,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 47-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 2) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 0, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 47-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 3) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 4) << 21,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 47-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 5) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 47-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 6) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 7) << 25,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 47-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 8) << 42,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 47-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 9) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 10) << 29,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 47-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 11) << 46,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 47-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 12) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 31 | SafeLoadAs<uint64_t>(in + 8 * 13) << 33,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 47-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 14) << 3,
+    };
+    shifts = simd_batch{ 14, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 47-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 15) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 27 | SafeLoadAs<uint64_t>(in + 8 * 16) << 37,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 47-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 17) << 7,
+    };
+    shifts = simd_batch{ 10, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 47-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 18) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 23 | SafeLoadAs<uint64_t>(in + 8 * 19) << 41,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 47-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 20) << 11,
+    };
+    shifts = simd_batch{ 6, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 47-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 21) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 19 | SafeLoadAs<uint64_t>(in + 8 * 22) << 45,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 47-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 23) << 15,
+    };
+    shifts = simd_batch{ 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 47-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 24) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+    };
+    shifts = simd_batch{ 0, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 47-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 25) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 26) << 19,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 47-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 27) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+    };
+    shifts = simd_batch{ 0, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 47-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 28) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 29) << 23,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 47-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 30) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+    };
+    shifts = simd_batch{ 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 47-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 31) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 32) << 27,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 47-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 33) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 33),
+    };
+    shifts = simd_batch{ 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 47-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 34) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 35) << 31,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 47-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 36) << 1,
+    };
+    shifts = simd_batch{ 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 47-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 37) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 29 | SafeLoadAs<uint64_t>(in + 8 * 38) << 35,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 47-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 38),
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 39) << 5,
+    };
+    shifts = simd_batch{ 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 47-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 40) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 25 | SafeLoadAs<uint64_t>(in + 8 * 41) << 39,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 47-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 41),
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 42) << 9,
+    };
+    shifts = simd_batch{ 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 47-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 43) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 21 | SafeLoadAs<uint64_t>(in + 8 * 44) << 43,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 47-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 44),
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 45) << 13,
+    };
+    shifts = simd_batch{ 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 47-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 46) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 46),
+    };
+    shifts = simd_batch{ 0, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 47 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 48> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0xffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 48-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 1) << 16,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 48-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 2) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 48-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 4) << 16,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 48-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 5) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 48-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 7) << 16,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 48-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 8) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 48-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 10) << 16,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 48-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 11) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 48-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 13) << 16,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 48-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 14) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 48-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 16) << 16,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 48-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 17) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+    };
+    shifts = simd_batch{ 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 48-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 19) << 16,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 48-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 20) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+    };
+    shifts = simd_batch{ 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 48-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 22) << 16,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 48-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 23) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+    };
+    shifts = simd_batch{ 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 48-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 25) << 16,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 48-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 26) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+    };
+    shifts = simd_batch{ 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 48-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 28) << 16,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 48-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 29) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 29),
+    };
+    shifts = simd_batch{ 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 48-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 31) << 16,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 48-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 32) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+    };
+    shifts = simd_batch{ 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 48-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 33),
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 34) << 16,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 48-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 35) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+    };
+    shifts = simd_batch{ 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 48-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 36),
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 37) << 16,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 48-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 38) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 38),
+    };
+    shifts = simd_batch{ 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 48-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 39),
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 40) << 16,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 48-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 41) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 41),
+    };
+    shifts = simd_batch{ 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 48-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 42),
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 43) << 16,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 48-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 44) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 44),
+    };
+    shifts = simd_batch{ 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 48-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 45),
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 46) << 16,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 48-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 47) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 47),
+    };
+    shifts = simd_batch{ 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 48 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 49> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1ffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 49-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 1) << 15,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 49-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 2) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 19 | SafeLoadAs<uint64_t>(in + 8 * 3) << 45,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 49-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 4) << 11,
+    };
+    shifts = simd_batch{ 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 49-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 5) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 23 | SafeLoadAs<uint64_t>(in + 8 * 6) << 41,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 49-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 7) << 7,
+    };
+    shifts = simd_batch{ 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 49-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 8) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 27 | SafeLoadAs<uint64_t>(in + 8 * 9) << 37,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 49-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 10) << 3,
+    };
+    shifts = simd_batch{ 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 49-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 11) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 31 | SafeLoadAs<uint64_t>(in + 8 * 12) << 33,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 49-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 13) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 49-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 14) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 15) << 29,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 49-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 16) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+    };
+    shifts = simd_batch{ 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 49-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 17) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 18) << 25,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 49-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 19) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+    };
+    shifts = simd_batch{ 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 49-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 20) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 21) << 21,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 49-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 22) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+    };
+    shifts = simd_batch{ 0, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 49-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 23) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 24) << 17,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 49-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 25) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 17 | SafeLoadAs<uint64_t>(in + 8 * 26) << 47,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 49-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 27) << 13,
+    };
+    shifts = simd_batch{ 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 49-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 28) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 21 | SafeLoadAs<uint64_t>(in + 8 * 29) << 43,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 49-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 29),
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 30) << 9,
+    };
+    shifts = simd_batch{ 6, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 49-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 31) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 25 | SafeLoadAs<uint64_t>(in + 8 * 32) << 39,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 49-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 33) << 5,
+    };
+    shifts = simd_batch{ 10, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 49-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 34) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 29 | SafeLoadAs<uint64_t>(in + 8 * 35) << 35,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 49-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 36) << 1,
+    };
+    shifts = simd_batch{ 14, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 49-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 37) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 38) << 31,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 49-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 39) << 46,
+      SafeLoadAs<uint64_t>(in + 8 * 39),
+    };
+    shifts = simd_batch{ 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 49-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 40) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 41) << 27,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 49-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 42) << 42,
+      SafeLoadAs<uint64_t>(in + 8 * 42),
+    };
+    shifts = simd_batch{ 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 49-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 43) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 44) << 23,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 49-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 45) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 45),
+    };
+    shifts = simd_batch{ 0, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 49-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 46) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 47) << 19,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 49-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 47) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 48) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 48),
+    };
+    shifts = simd_batch{ 0, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 49 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 50> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3ffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 50-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 1) << 14,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 50-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 2) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 3) << 42,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 50-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 4) << 6,
+    };
+    shifts = simd_batch{ 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 50-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 5) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 6) << 34,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 50-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 7) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 50-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 8) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 9) << 26,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 50-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 10) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 50-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 11) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 12) << 18,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 50-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 13) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 14) << 46,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 50-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 15) << 10,
+    };
+    shifts = simd_batch{ 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 50-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 16) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 17) << 38,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 50-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 18) << 2,
+    };
+    shifts = simd_batch{ 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 50-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 19) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 20) << 30,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 50-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 21) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+    };
+    shifts = simd_batch{ 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 50-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 22) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 23) << 22,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 50-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 24) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+    };
+    shifts = simd_batch{ 0, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 50-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 26) << 14,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 50-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 27) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 28) << 42,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 50-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 29) << 6,
+    };
+    shifts = simd_batch{ 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 50-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 30) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 31) << 34,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 50-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 32) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+    };
+    shifts = simd_batch{ 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 50-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 33) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 34) << 26,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 50-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 35) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+    };
+    shifts = simd_batch{ 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 50-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 36) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 37) << 18,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 50-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 38) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 39) << 46,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 50-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 39),
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 40) << 10,
+    };
+    shifts = simd_batch{ 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 50-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 41) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 42) << 38,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 50-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 42),
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 43) << 2,
+    };
+    shifts = simd_batch{ 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 50-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 44) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 45) << 30,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 50-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 46) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 46),
+    };
+    shifts = simd_batch{ 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 50-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 47) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 47) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 48) << 22,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 50-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 48) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 49) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 49),
+    };
+    shifts = simd_batch{ 0, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 50 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 51> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7ffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 51-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 1) << 13,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 51-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 2) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 25 | SafeLoadAs<uint64_t>(in + 8 * 3) << 39,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 51-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 4) << 1,
+    };
+    shifts = simd_batch{ 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 51-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 5) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 6) << 27,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 51-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 7) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 0, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 51-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 8) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 9) << 15,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 51-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 10) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 23 | SafeLoadAs<uint64_t>(in + 8 * 11) << 41,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 51-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 12) << 3,
+    };
+    shifts = simd_batch{ 10, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 51-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 13) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 14) << 29,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 51-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 15) << 42,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+    };
+    shifts = simd_batch{ 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 51-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 16) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 17) << 17,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 51-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 18) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 21 | SafeLoadAs<uint64_t>(in + 8 * 19) << 43,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 51-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 20) << 5,
+    };
+    shifts = simd_batch{ 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 51-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 21) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 22) << 31,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 51-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 23) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+    };
+    shifts = simd_batch{ 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 51-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 24) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 25) << 19,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 51-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 26) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 19 | SafeLoadAs<uint64_t>(in + 8 * 27) << 45,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 51-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 28) << 7,
+    };
+    shifts = simd_batch{ 6, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 51-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 29) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 31 | SafeLoadAs<uint64_t>(in + 8 * 30) << 33,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 51-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 31) << 46,
+      SafeLoadAs<uint64_t>(in + 8 * 31),
+    };
+    shifts = simd_batch{ 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 51-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 32) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 33) << 21,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 51-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 34) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 17 | SafeLoadAs<uint64_t>(in + 8 * 35) << 47,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 51-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 36) << 9,
+    };
+    shifts = simd_batch{ 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 51-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 37) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 29 | SafeLoadAs<uint64_t>(in + 8 * 38) << 35,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 51-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 39) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 39),
+    };
+    shifts = simd_batch{ 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 51-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 40) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 41) << 23,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 51-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 42) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 15 | SafeLoadAs<uint64_t>(in + 8 * 43) << 49,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 51-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 43),
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 44) << 11,
+    };
+    shifts = simd_batch{ 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 51-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 45) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 27 | SafeLoadAs<uint64_t>(in + 8 * 46) << 37,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 51-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 14 | SafeLoadAs<uint64_t>(in + 8 * 47) << 50,
+      SafeLoadAs<uint64_t>(in + 8 * 47),
+    };
+    shifts = simd_batch{ 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 51-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 47) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 48) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 48) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 49) << 25,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 51-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 49) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 50) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 50),
+    };
+    shifts = simd_batch{ 0, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 51 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 52> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0xfffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 52-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 1) << 12,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 52-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 2) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 3) << 36,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 52-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 4) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 52-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 5) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 6) << 20,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 52-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 7) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 8) << 44,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 52-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 9) << 4,
+    };
+    shifts = simd_batch{ 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 52-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 10) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 11) << 28,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 52-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 12) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 52-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 14) << 12,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 52-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 15) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 16) << 36,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 52-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 17) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+    };
+    shifts = simd_batch{ 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 52-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 18) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 19) << 20,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 52-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 20) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 21) << 44,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 52-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 22) << 4,
+    };
+    shifts = simd_batch{ 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 52-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 23) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 24) << 28,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 52-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 25) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+    };
+    shifts = simd_batch{ 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 52-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 27) << 12,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 52-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 28) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 29) << 36,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 52-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 30) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+    };
+    shifts = simd_batch{ 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 52-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 31) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 32) << 20,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 52-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 33) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 34) << 44,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 52-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 34),
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 35) << 4,
+    };
+    shifts = simd_batch{ 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 52-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 36) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 37) << 28,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 52-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 38) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 38),
+    };
+    shifts = simd_batch{ 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 52-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 39),
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 40) << 12,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 52-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 41) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 42) << 36,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 52-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 43) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 43),
+    };
+    shifts = simd_batch{ 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 52-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 44) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 45) << 20,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 52-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 46) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 47) << 44,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 52-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 47),
+      SafeLoadAs<uint64_t>(in + 8 * 47) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 48) << 4,
+    };
+    shifts = simd_batch{ 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 52-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 48) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 49) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 49) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 50) << 28,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 52-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 50) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 51) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 51),
+    };
+    shifts = simd_batch{ 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 52 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 53> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1fffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 53-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 1) << 11,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 53-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 2) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 31 | SafeLoadAs<uint64_t>(in + 8 * 3) << 33,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 53-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 4) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 53-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 5) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 6) << 13,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 53-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 7) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 29 | SafeLoadAs<uint64_t>(in + 8 * 8) << 35,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 53-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 9) << 46,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 53-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 10) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 11) << 15,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 53-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 12) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 27 | SafeLoadAs<uint64_t>(in + 8 * 13) << 37,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 53-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 14) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 53-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 15) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 16) << 17,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 53-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 17) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 25 | SafeLoadAs<uint64_t>(in + 8 * 18) << 39,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 53-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 14 | SafeLoadAs<uint64_t>(in + 8 * 19) << 50,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+    };
+    shifts = simd_batch{ 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 53-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 20) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 21) << 19,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 53-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 22) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 23 | SafeLoadAs<uint64_t>(in + 8 * 23) << 41,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 53-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 24) << 52,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+    };
+    shifts = simd_batch{ 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 53-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 25) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 26) << 21,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 53-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 27) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 21 | SafeLoadAs<uint64_t>(in + 8 * 28) << 43,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 53-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 29) << 1,
+    };
+    shifts = simd_batch{ 10, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 53-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 30) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 31) << 23,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 53-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 32) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 19 | SafeLoadAs<uint64_t>(in + 8 * 33) << 45,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 53-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 33),
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 34) << 3,
+    };
+    shifts = simd_batch{ 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 53-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 35) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 36) << 25,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 53-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 37) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 17 | SafeLoadAs<uint64_t>(in + 8 * 38) << 47,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 53-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 38),
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 39) << 5,
+    };
+    shifts = simd_batch{ 6, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 53-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 40) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 41) << 27,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 53-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 42) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 15 | SafeLoadAs<uint64_t>(in + 8 * 43) << 49,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 53-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 43),
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 44) << 7,
+    };
+    shifts = simd_batch{ 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 53-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 45) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 46) << 29,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 53-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 47) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 47) >> 13 | SafeLoadAs<uint64_t>(in + 8 * 48) << 51,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 53-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 48),
+      SafeLoadAs<uint64_t>(in + 8 * 48) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 49) << 9,
+    };
+    shifts = simd_batch{ 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 53-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 49) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 50) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 50) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 51) << 31,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 53-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 51) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 52) << 42,
+      SafeLoadAs<uint64_t>(in + 8 * 52),
+    };
+    shifts = simd_batch{ 0, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 53 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 54> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3fffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 54-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 1) << 10,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 54-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 2) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 3) << 30,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 54-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 4) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 14 | SafeLoadAs<uint64_t>(in + 8 * 5) << 50,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 54-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 6) << 6,
+    };
+    shifts = simd_batch{ 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 54-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 7) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 8) << 26,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 54-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 9) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 10) << 46,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 54-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 11) << 2,
+    };
+    shifts = simd_batch{ 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 54-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 12) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 13) << 22,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 54-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 14) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 15) << 42,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 54-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 16) << 52,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+    };
+    shifts = simd_batch{ 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 54-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 17) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 18) << 18,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 54-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 19) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 20) << 38,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 54-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 21) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+    };
+    shifts = simd_batch{ 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 54-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 22) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 23) << 14,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 54-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 24) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 25) << 34,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 54-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 26) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+    };
+    shifts = simd_batch{ 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 54-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 28) << 10,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 54-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 29) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 30) << 30,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 54-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 31) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 14 | SafeLoadAs<uint64_t>(in + 8 * 32) << 50,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 54-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 33) << 6,
+    };
+    shifts = simd_batch{ 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 54-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 34) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 35) << 26,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 54-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 36) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 37) << 46,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 54-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 37),
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 38) << 2,
+    };
+    shifts = simd_batch{ 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 54-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 39) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 40) << 22,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 54-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 41) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 42) << 42,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 54-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 43) << 52,
+      SafeLoadAs<uint64_t>(in + 8 * 43),
+    };
+    shifts = simd_batch{ 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 54-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 44) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 45) << 18,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 54-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 46) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 47) << 38,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 54-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 47) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 48) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 48),
+    };
+    shifts = simd_batch{ 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 54-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 48) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 49) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 49) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 50) << 14,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 54-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 50) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 51) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 51) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 52) << 34,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 54-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 52) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 53) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 53),
+    };
+    shifts = simd_batch{ 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 54 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 55> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7fffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 55-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 1) << 9,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 55-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 2) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 3) << 27,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 55-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 4) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 19 | SafeLoadAs<uint64_t>(in + 8 * 5) << 45,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 55-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 10 | SafeLoadAs<uint64_t>(in + 8 * 6) << 54,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 55-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 7) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 8) << 17,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 55-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 9) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 29 | SafeLoadAs<uint64_t>(in + 8 * 10) << 35,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 55-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 11) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 11 | SafeLoadAs<uint64_t>(in + 8 * 12) << 53,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 55-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 13) << 7,
+    };
+    shifts = simd_batch{ 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 55-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 14) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 15) << 25,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 55-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 16) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 21 | SafeLoadAs<uint64_t>(in + 8 * 17) << 43,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 55-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 18) << 52,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+    };
+    shifts = simd_batch{ 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 55-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 19) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 20) << 15,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 55-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 21) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 31 | SafeLoadAs<uint64_t>(in + 8 * 22) << 33,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 55-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 23) << 42,
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 13 | SafeLoadAs<uint64_t>(in + 8 * 24) << 51,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 55-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 25) << 5,
+    };
+    shifts = simd_batch{ 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 55-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 26) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 27) << 23,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 55-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 28) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 23 | SafeLoadAs<uint64_t>(in + 8 * 29) << 41,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 55-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 14 | SafeLoadAs<uint64_t>(in + 8 * 30) << 50,
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+    };
+    shifts = simd_batch{ 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 55-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 31) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 32) << 13,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 55-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 33) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 34) << 31,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 55-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 35) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 15 | SafeLoadAs<uint64_t>(in + 8 * 36) << 49,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 55-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 36),
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 37) << 3,
+    };
+    shifts = simd_batch{ 6, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 55-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 38) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 39) << 21,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 55-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 40) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 25 | SafeLoadAs<uint64_t>(in + 8 * 41) << 39,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 55-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 42) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 42),
+    };
+    shifts = simd_batch{ 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 55-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 43) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 44) << 11,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 55-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 45) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 46) << 29,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 55-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 47) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 47) >> 17 | SafeLoadAs<uint64_t>(in + 8 * 48) << 47,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 55-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 48),
+      SafeLoadAs<uint64_t>(in + 8 * 48) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 49) << 1,
+    };
+    shifts = simd_batch{ 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 55-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 49) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 50) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 50) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 51) << 19,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 55-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 51) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 52) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 52) >> 27 | SafeLoadAs<uint64_t>(in + 8 * 53) << 37,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 55-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 53) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 54) << 46,
+      SafeLoadAs<uint64_t>(in + 8 * 54),
+    };
+    shifts = simd_batch{ 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 55 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 56> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0xffffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 56-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 1) << 8,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 56-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 2) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 3) << 24,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 56-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 4) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 5) << 40,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 56-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 6) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 56-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 8) << 8,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 56-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 9) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 10) << 24,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 56-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 11) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 12) << 40,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 56-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 13) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 56-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 15) << 8,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 56-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 16) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 17) << 24,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 56-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 18) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 19) << 40,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 56-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 20) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+    };
+    shifts = simd_batch{ 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 56-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 22) << 8,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 56-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 23) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 24) << 24,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 56-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 25) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 26) << 40,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 56-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 27) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+    };
+    shifts = simd_batch{ 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 56-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 29) << 8,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 56-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 30) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 31) << 24,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 56-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 32) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 33) << 40,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 56-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 34) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 34),
+    };
+    shifts = simd_batch{ 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 56-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 36) << 8,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 56-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 37) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 38) << 24,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 56-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 39) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 40) << 40,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 56-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 41) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 41),
+    };
+    shifts = simd_batch{ 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 56-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 42),
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 43) << 8,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 56-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 44) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 45) << 24,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 56-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 46) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 47) << 40,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 56-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 47) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 48) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 48),
+    };
+    shifts = simd_batch{ 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 56-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 49),
+      SafeLoadAs<uint64_t>(in + 8 * 49) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 50) << 8,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 56-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 50) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 51) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 51) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 52) << 24,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 56-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 52) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 53) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 53) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 54) << 40,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 56-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 54) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 55) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 55),
+    };
+    shifts = simd_batch{ 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 56 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 57> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1ffffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 57-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 1) << 7,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 57-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 2) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 3) << 21,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 57-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 4) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 29 | SafeLoadAs<uint64_t>(in + 8 * 5) << 35,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 57-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 6) << 42,
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 15 | SafeLoadAs<uint64_t>(in + 8 * 7) << 49,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 57-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 8 | SafeLoadAs<uint64_t>(in + 8 * 8) << 56,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 57-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 9) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 10) << 13,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 57-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 11) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 12) << 27,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 57-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 13) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 23 | SafeLoadAs<uint64_t>(in + 8 * 14) << 41,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 57-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 15) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 9 | SafeLoadAs<uint64_t>(in + 8 * 16) << 55,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 57-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 17) << 5,
+    };
+    shifts = simd_batch{ 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 57-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 18) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 19) << 19,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 57-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 20) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 31 | SafeLoadAs<uint64_t>(in + 8 * 21) << 33,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 57-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 22) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 17 | SafeLoadAs<uint64_t>(in + 8 * 23) << 47,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 57-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 10 | SafeLoadAs<uint64_t>(in + 8 * 24) << 54,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+    };
+    shifts = simd_batch{ 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 57-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 25) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 26) << 11,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 57-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 27) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 28) << 25,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 57-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 29) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 25 | SafeLoadAs<uint64_t>(in + 8 * 30) << 39,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 57-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 31) << 46,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 11 | SafeLoadAs<uint64_t>(in + 8 * 32) << 53,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 57-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 33) << 3,
+    };
+    shifts = simd_batch{ 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 57-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 34) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 35) << 17,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 57-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 36) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 37) << 31,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 57-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 38) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 19 | SafeLoadAs<uint64_t>(in + 8 * 39) << 45,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 57-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 40) << 52,
+      SafeLoadAs<uint64_t>(in + 8 * 40),
+    };
+    shifts = simd_batch{ 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 57-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 41) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 42) << 9,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 57-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 43) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 44) << 23,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 57-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 45) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 27 | SafeLoadAs<uint64_t>(in + 8 * 46) << 37,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 57-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 47) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 47) >> 13 | SafeLoadAs<uint64_t>(in + 8 * 48) << 51,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 57-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 48),
+      SafeLoadAs<uint64_t>(in + 8 * 48) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 49) << 1,
+    };
+    shifts = simd_batch{ 6, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 57-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 49) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 50) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 50) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 51) << 15,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 57-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 51) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 52) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 52) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 53) << 29,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 57-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 53) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 54) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 54) >> 21 | SafeLoadAs<uint64_t>(in + 8 * 55) << 43,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 57-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 55) >> 14 | SafeLoadAs<uint64_t>(in + 8 * 56) << 50,
+      SafeLoadAs<uint64_t>(in + 8 * 56),
+    };
+    shifts = simd_batch{ 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 57 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 58> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3ffffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 58-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 1) << 6,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 58-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 2) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 3) << 18,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 58-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 4) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 5) << 30,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 58-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 6) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 7) << 42,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 58-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 8) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 10 | SafeLoadAs<uint64_t>(in + 8 * 9) << 54,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 58-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 10) << 2,
+    };
+    shifts = simd_batch{ 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 58-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 11) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 12) << 14,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 58-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 13) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 14) << 26,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 58-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 15) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 16) << 38,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 58-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 17) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 14 | SafeLoadAs<uint64_t>(in + 8 * 18) << 50,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 58-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 8 | SafeLoadAs<uint64_t>(in + 8 * 19) << 56,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+    };
+    shifts = simd_batch{ 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 58-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 20) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 21) << 10,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 58-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 22) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 23) << 22,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 58-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 24) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 25) << 34,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 58-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 26) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 27) << 46,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 58-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 28) << 52,
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+    };
+    shifts = simd_batch{ 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 58-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 29),
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 30) << 6,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 58-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 31) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 32) << 18,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 58-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 33) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 34) << 30,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 58-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 35) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 36) << 42,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 58-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 37) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 10 | SafeLoadAs<uint64_t>(in + 8 * 38) << 54,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 58-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 38),
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 39) << 2,
+    };
+    shifts = simd_batch{ 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 58-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 40) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 41) << 14,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 58-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 42) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 43) << 26,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 58-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 44) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 45) << 38,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 58-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 46) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 14 | SafeLoadAs<uint64_t>(in + 8 * 47) << 50,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 58-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 47) >> 8 | SafeLoadAs<uint64_t>(in + 8 * 48) << 56,
+      SafeLoadAs<uint64_t>(in + 8 * 48),
+    };
+    shifts = simd_batch{ 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 58-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 48) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 49) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 49) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 50) << 10,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 58-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 50) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 51) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 51) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 52) << 22,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 58-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 52) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 53) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 53) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 54) << 34,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 58-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 54) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 55) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 55) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 56) << 46,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 58-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 56) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 57) << 52,
+      SafeLoadAs<uint64_t>(in + 8 * 57),
+    };
+    shifts = simd_batch{ 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 58 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 59> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7ffffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 59-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 1) << 5,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 59-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 2) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 3) << 15,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 59-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 4) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 5) << 25,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 59-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 6) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 29 | SafeLoadAs<uint64_t>(in + 8 * 7) << 35,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 59-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 8) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 19 | SafeLoadAs<uint64_t>(in + 8 * 9) << 45,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 59-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 14 | SafeLoadAs<uint64_t>(in + 8 * 10) << 50,
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 9 | SafeLoadAs<uint64_t>(in + 8 * 11) << 55,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 59-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 12) << 1,
+    };
+    shifts = simd_batch{ 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 59-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 13) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 14) << 11,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 59-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 15) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 16) << 21,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 59-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 17) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 18) << 31,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 59-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 19) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 23 | SafeLoadAs<uint64_t>(in + 8 * 20) << 41,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 59-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 21) << 46,
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 13 | SafeLoadAs<uint64_t>(in + 8 * 22) << 51,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 59-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 8 | SafeLoadAs<uint64_t>(in + 8 * 23) << 56,
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+    };
+    shifts = simd_batch{ 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 59-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 24) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 25) << 7,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 59-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 26) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 27) << 17,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 59-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 28) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 29) << 27,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 59-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 30) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 27 | SafeLoadAs<uint64_t>(in + 8 * 31) << 37,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 59-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 32) << 42,
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 17 | SafeLoadAs<uint64_t>(in + 8 * 33) << 47,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 59-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 34) << 52,
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 7 | SafeLoadAs<uint64_t>(in + 8 * 35) << 57,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 59-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 36) << 3,
+    };
+    shifts = simd_batch{ 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 59-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 37) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 38) << 13,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 59-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 39) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 40) << 23,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 59-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 41) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 31 | SafeLoadAs<uint64_t>(in + 8 * 42) << 33,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 59-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 43) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 21 | SafeLoadAs<uint64_t>(in + 8 * 44) << 43,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 59-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 45) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 11 | SafeLoadAs<uint64_t>(in + 8 * 46) << 53,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 59-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 6 | SafeLoadAs<uint64_t>(in + 8 * 47) << 58,
+      SafeLoadAs<uint64_t>(in + 8 * 47),
+    };
+    shifts = simd_batch{ 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 59-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 47) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 48) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 48) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 49) << 9,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 59-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 49) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 50) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 50) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 51) << 19,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 59-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 51) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 52) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 52) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 53) << 29,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 59-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 53) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 54) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 54) >> 25 | SafeLoadAs<uint64_t>(in + 8 * 55) << 39,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 59-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 55) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 56) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 56) >> 15 | SafeLoadAs<uint64_t>(in + 8 * 57) << 49,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 59-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 57) >> 10 | SafeLoadAs<uint64_t>(in + 8 * 58) << 54,
+      SafeLoadAs<uint64_t>(in + 8 * 58),
+    };
+    shifts = simd_batch{ 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 59 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 60> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0xfffffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 60-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 1) << 4,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 60-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 2) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 3) << 12,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 60-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 4) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 5) << 20,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 60-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 6) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 7) << 28,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 60-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 8) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 9) << 36,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 60-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 10) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 11) << 44,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 60-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 12) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 13) << 52,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 60-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 8 | SafeLoadAs<uint64_t>(in + 8 * 14) << 56,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 60-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 16) << 4,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 60-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 17) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 18) << 12,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 60-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 19) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 20) << 20,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 60-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 21) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 22) << 28,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 60-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 23) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 24) << 36,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 60-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 25) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 26) << 44,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 60-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 27) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 28) << 52,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 60-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 8 | SafeLoadAs<uint64_t>(in + 8 * 29) << 56,
+      SafeLoadAs<uint64_t>(in + 8 * 29),
+    };
+    shifts = simd_batch{ 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 60-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 31) << 4,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 60-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 32) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 33) << 12,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 60-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 34) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 35) << 20,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 60-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 36) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 37) << 28,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 60-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 38) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 39) << 36,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 60-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 40) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 41) << 44,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 60-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 42) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 43) << 52,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 60-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 8 | SafeLoadAs<uint64_t>(in + 8 * 44) << 56,
+      SafeLoadAs<uint64_t>(in + 8 * 44),
+    };
+    shifts = simd_batch{ 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 60-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 45),
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 46) << 4,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 60-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 47) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 47) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 48) << 12,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 60-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 48) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 49) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 49) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 50) << 20,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 60-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 50) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 51) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 51) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 52) << 28,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 60-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 52) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 53) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 53) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 54) << 36,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 60-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 54) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 55) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 55) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 56) << 44,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 60-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 56) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 57) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 57) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 58) << 52,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 60-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 58) >> 8 | SafeLoadAs<uint64_t>(in + 8 * 59) << 56,
+      SafeLoadAs<uint64_t>(in + 8 * 59),
+    };
+    shifts = simd_batch{ 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 60 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 61> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1fffffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 61-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 1) << 3,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 61-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 2) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 3) << 9,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 61-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 4) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 5) << 15,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 61-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 6) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 7) << 21,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 61-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 8) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 9) << 27,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 61-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 10) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 31 | SafeLoadAs<uint64_t>(in + 8 * 11) << 33,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 61-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 12) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 25 | SafeLoadAs<uint64_t>(in + 8 * 13) << 39,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 61-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 14) << 42,
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 19 | SafeLoadAs<uint64_t>(in + 8 * 15) << 45,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 61-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 16) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 13 | SafeLoadAs<uint64_t>(in + 8 * 17) << 51,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 61-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 10 | SafeLoadAs<uint64_t>(in + 8 * 18) << 54,
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 7 | SafeLoadAs<uint64_t>(in + 8 * 19) << 57,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 61-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 4 | SafeLoadAs<uint64_t>(in + 8 * 20) << 60,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+    };
+    shifts = simd_batch{ 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 61-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 21) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 22) << 5,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 61-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 23) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 24) << 11,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 61-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 25) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 26) << 17,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 61-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 27) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 28) << 23,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 61-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 29) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 30) << 29,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 61-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 31) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 29 | SafeLoadAs<uint64_t>(in + 8 * 32) << 35,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 61-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 33) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 23 | SafeLoadAs<uint64_t>(in + 8 * 34) << 41,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 61-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 35) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 17 | SafeLoadAs<uint64_t>(in + 8 * 36) << 47,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 61-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 14 | SafeLoadAs<uint64_t>(in + 8 * 37) << 50,
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 11 | SafeLoadAs<uint64_t>(in + 8 * 38) << 53,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 61-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 8 | SafeLoadAs<uint64_t>(in + 8 * 39) << 56,
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 5 | SafeLoadAs<uint64_t>(in + 8 * 40) << 59,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 61-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 40),
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 41) << 1,
+    };
+    shifts = simd_batch{ 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 61-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 42) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 43) << 7,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 61-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 44) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 45) << 13,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 61-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 46) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 47) << 19,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 61-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 47) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 48) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 48) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 49) << 25,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 61-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 49) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 50) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 50) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 51) << 31,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 61-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 51) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 52) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 52) >> 27 | SafeLoadAs<uint64_t>(in + 8 * 53) << 37,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 61-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 53) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 54) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 54) >> 21 | SafeLoadAs<uint64_t>(in + 8 * 55) << 43,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 61-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 55) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 56) << 46,
+      SafeLoadAs<uint64_t>(in + 8 * 56) >> 15 | SafeLoadAs<uint64_t>(in + 8 * 57) << 49,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 61-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 57) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 58) << 52,
+      SafeLoadAs<uint64_t>(in + 8 * 58) >> 9 | SafeLoadAs<uint64_t>(in + 8 * 59) << 55,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 61-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 59) >> 6 | SafeLoadAs<uint64_t>(in + 8 * 60) << 58,
+      SafeLoadAs<uint64_t>(in + 8 * 60),
+    };
+    shifts = simd_batch{ 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 61 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 62> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3fffffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 62-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 1) << 2,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 62-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 2) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 3) << 6,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 62-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 4) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 5) << 10,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 62-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 6) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 7) << 14,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 62-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 8) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 9) << 18,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 62-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 10) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 11) << 22,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 62-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 12) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 13) << 26,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 62-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 14) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 15) << 30,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 62-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 16) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 17) << 34,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 62-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 18) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 19) << 38,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 62-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 20) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 21) << 42,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 62-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 22) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 23) << 46,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 62-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 24) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 14 | SafeLoadAs<uint64_t>(in + 8 * 25) << 50,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 62-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 26) << 52,
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 10 | SafeLoadAs<uint64_t>(in + 8 * 27) << 54,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 62-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 8 | SafeLoadAs<uint64_t>(in + 8 * 28) << 56,
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 6 | SafeLoadAs<uint64_t>(in + 8 * 29) << 58,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 62-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 4 | SafeLoadAs<uint64_t>(in + 8 * 30) << 60,
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+    };
+    shifts = simd_batch{ 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 62-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 31),
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 32) << 2,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 62-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 33) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 34) << 6,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 62-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 35) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 36) << 10,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 62-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 37) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 38) << 14,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 62-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 39) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 40) << 18,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 62-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 41) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 42) << 22,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 62-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 43) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 44) << 26,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 62-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 45) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 46) << 30,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 62-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 47) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 47) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 48) << 34,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 62-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 48) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 49) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 49) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 50) << 38,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 62-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 50) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 51) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 51) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 52) << 42,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 62-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 52) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 53) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 53) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 54) << 46,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 62-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 54) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 55) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 55) >> 14 | SafeLoadAs<uint64_t>(in + 8 * 56) << 50,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 62-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 56) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 57) << 52,
+      SafeLoadAs<uint64_t>(in + 8 * 57) >> 10 | SafeLoadAs<uint64_t>(in + 8 * 58) << 54,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 62-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 58) >> 8 | SafeLoadAs<uint64_t>(in + 8 * 59) << 56,
+      SafeLoadAs<uint64_t>(in + 8 * 59) >> 6 | SafeLoadAs<uint64_t>(in + 8 * 60) << 58,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 62-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 60) >> 4 | SafeLoadAs<uint64_t>(in + 8 * 61) << 60,
+      SafeLoadAs<uint64_t>(in + 8 * 61),
+    };
+    shifts = simd_batch{ 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 62 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd128UnpackerForWidth<uint64_t, 63> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 2>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7fffffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 63-bit bundles 0 to 1
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 1) << 1,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 63-bit bundles 2 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 2) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 3) << 3,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 63-bit bundles 4 to 5
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 4) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 5) << 5,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 63-bit bundles 6 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 6) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 7) << 7,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 63-bit bundles 8 to 9
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 8) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 9) << 9,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 63-bit bundles 10 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 10) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 11) << 11,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 63-bit bundles 12 to 13
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 12) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 13) << 13,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 63-bit bundles 14 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 14) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 15) << 15,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 63-bit bundles 16 to 17
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 16) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 17) << 17,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 63-bit bundles 18 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 18) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 19) << 19,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 63-bit bundles 20 to 21
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 20) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 21) << 21,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 63-bit bundles 22 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 22) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 23) << 23,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 63-bit bundles 24 to 25
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 24) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 25) << 25,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 63-bit bundles 26 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 26) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 27) << 27,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 63-bit bundles 28 to 29
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 28) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 29) << 29,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 63-bit bundles 30 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 30) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 31) << 31,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 63-bit bundles 32 to 33
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 32) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 31 | SafeLoadAs<uint64_t>(in + 8 * 33) << 33,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 63-bit bundles 34 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 34) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 29 | SafeLoadAs<uint64_t>(in + 8 * 35) << 35,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 63-bit bundles 36 to 37
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 36) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 27 | SafeLoadAs<uint64_t>(in + 8 * 37) << 37,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 63-bit bundles 38 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 38) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 25 | SafeLoadAs<uint64_t>(in + 8 * 39) << 39,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 63-bit bundles 40 to 41
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 40) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 23 | SafeLoadAs<uint64_t>(in + 8 * 41) << 41,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 63-bit bundles 42 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 42) << 42,
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 21 | SafeLoadAs<uint64_t>(in + 8 * 43) << 43,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 63-bit bundles 44 to 45
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 44) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 19 | SafeLoadAs<uint64_t>(in + 8 * 45) << 45,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 63-bit bundles 46 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 46) << 46,
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 17 | SafeLoadAs<uint64_t>(in + 8 * 47) << 47,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 63-bit bundles 48 to 49
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 47) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 48) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 48) >> 15 | SafeLoadAs<uint64_t>(in + 8 * 49) << 49,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 63-bit bundles 50 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 49) >> 14 | SafeLoadAs<uint64_t>(in + 8 * 50) << 50,
+      SafeLoadAs<uint64_t>(in + 8 * 50) >> 13 | SafeLoadAs<uint64_t>(in + 8 * 51) << 51,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 63-bit bundles 52 to 53
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 51) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 52) << 52,
+      SafeLoadAs<uint64_t>(in + 8 * 52) >> 11 | SafeLoadAs<uint64_t>(in + 8 * 53) << 53,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 63-bit bundles 54 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 53) >> 10 | SafeLoadAs<uint64_t>(in + 8 * 54) << 54,
+      SafeLoadAs<uint64_t>(in + 8 * 54) >> 9 | SafeLoadAs<uint64_t>(in + 8 * 55) << 55,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 63-bit bundles 56 to 57
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 55) >> 8 | SafeLoadAs<uint64_t>(in + 8 * 56) << 56,
+      SafeLoadAs<uint64_t>(in + 8 * 56) >> 7 | SafeLoadAs<uint64_t>(in + 8 * 57) << 57,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 63-bit bundles 58 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 57) >> 6 | SafeLoadAs<uint64_t>(in + 8 * 58) << 58,
+      SafeLoadAs<uint64_t>(in + 8 * 58) >> 5 | SafeLoadAs<uint64_t>(in + 8 * 59) << 59,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 63-bit bundles 60 to 61
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 59) >> 4 | SafeLoadAs<uint64_t>(in + 8 * 60) << 60,
+      SafeLoadAs<uint64_t>(in + 8 * 60) >> 3 | SafeLoadAs<uint64_t>(in + 8 * 61) << 61,
+    };
+    shifts = simd_batch{ 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    // extract 63-bit bundles 62 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 61) >> 2 | SafeLoadAs<uint64_t>(in + 8 * 62) << 62,
+      SafeLoadAs<uint64_t>(in + 8 * 62),
+    };
+    shifts = simd_batch{ 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 2;
+
+    in += 63 * 8;
+    return in;
+  }
+};
+
+
+}  // namespace arrow::internal
diff --git a/cpp/src/arrow/util/bpacking_simd256_generated_internal.h b/cpp/src/arrow/util/bpacking_simd256_generated_internal.h
index 3dccb1745f745..4757a4183727e 100644
--- a/cpp/src/arrow/util/bpacking_simd256_generated_internal.h
+++ b/cpp/src/arrow/util/bpacking_simd256_generated_internal.h
@@ -15,7 +15,9 @@
 // specific language governing permissions and limitations
 // under the License.
 
-// Automatically generated file; DO NOT EDIT.
+// WARNING: this file is generated, DO NOT EDIT.
+// Usage:
+//   python cpp/src/arrow/util/bpacking_simd_codegen.py 256
 
 #pragma once
 
@@ -24,1247 +26,16309 @@
 
 #include <xsimd/xsimd.hpp>
 
-#include "arrow/util/dispatch_internal.h"
 #include "arrow/util/ubsan.h"
 
-namespace arrow {
-namespace internal {
-namespace {
-
-using ::arrow::util::SafeLoad;
-
-template <DispatchLevel level>
-struct UnpackBits256 {
-
-using simd_batch = xsimd::make_sized_batch_t<uint32_t, 8>;
-
-inline static const uint32_t* unpack0_32(const uint32_t* in, uint32_t* out) {
-  memset(out, 0x0, 32 * sizeof(*out));
-  out += 32;
-
-  return in;
-}
-
-inline static const uint32_t* unpack1_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x1;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 1-bit bundles 0 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) };
-  shifts = simd_batch{ 0, 1, 2, 3, 4, 5, 6, 7 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 1-bit bundles 8 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) };
-  shifts = simd_batch{ 8, 9, 10, 11, 12, 13, 14, 15 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 1-bit bundles 16 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) };
-  shifts = simd_batch{ 16, 17, 18, 19, 20, 21, 22, 23 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 1-bit bundles 24 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) };
-  shifts = simd_batch{ 24, 25, 26, 27, 28, 29, 30, 31 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  in += 1;
-  return in;
-}
-
-inline static const uint32_t* unpack2_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x3;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 2-bit bundles 0 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) };
-  shifts = simd_batch{ 0, 2, 4, 6, 8, 10, 12, 14 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 2-bit bundles 8 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) };
-  shifts = simd_batch{ 16, 18, 20, 22, 24, 26, 28, 30 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 2-bit bundles 16 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) };
-  shifts = simd_batch{ 0, 2, 4, 6, 8, 10, 12, 14 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 2-bit bundles 24 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) };
-  shifts = simd_batch{ 16, 18, 20, 22, 24, 26, 28, 30 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  in += 2;
-  return in;
-}
-
-inline static const uint32_t* unpack3_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x7;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 3-bit bundles 0 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) };
-  shifts = simd_batch{ 0, 3, 6, 9, 12, 15, 18, 21 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 3-bit bundles 8 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 30 | SafeLoad<uint32_t>(in + 1) << 2, SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) };
-  shifts = simd_batch{ 24, 27, 0, 1, 4, 7, 10, 13 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 3-bit bundles 16 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) >> 31 | SafeLoad<uint32_t>(in + 2) << 1, SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) };
-  shifts = simd_batch{ 16, 19, 22, 25, 28, 0, 2, 5 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 3-bit bundles 24 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) };
-  shifts = simd_batch{ 8, 11, 14, 17, 20, 23, 26, 29 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  in += 3;
-  return in;
-}
-
-inline static const uint32_t* unpack4_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0xf;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 4-bit bundles 0 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) };
-  shifts = simd_batch{ 0, 4, 8, 12, 16, 20, 24, 28 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 4-bit bundles 8 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) };
-  shifts = simd_batch{ 0, 4, 8, 12, 16, 20, 24, 28 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 4-bit bundles 16 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) };
-  shifts = simd_batch{ 0, 4, 8, 12, 16, 20, 24, 28 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 4-bit bundles 24 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) };
-  shifts = simd_batch{ 0, 4, 8, 12, 16, 20, 24, 28 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  in += 4;
-  return in;
-}
-
-inline static const uint32_t* unpack5_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x1f;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 5-bit bundles 0 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 30 | SafeLoad<uint32_t>(in + 1) << 2, SafeLoad<uint32_t>(in + 1) };
-  shifts = simd_batch{ 0, 5, 10, 15, 20, 25, 0, 3 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 5-bit bundles 8 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) >> 28 | SafeLoad<uint32_t>(in + 2) << 4, SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) };
-  shifts = simd_batch{ 8, 13, 18, 23, 0, 1, 6, 11 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 5-bit bundles 16 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) >> 31 | SafeLoad<uint32_t>(in + 3) << 1, SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) };
-  shifts = simd_batch{ 16, 21, 26, 0, 4, 9, 14, 19 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 5-bit bundles 24 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) >> 29 | SafeLoad<uint32_t>(in + 4) << 3, SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4) };
-  shifts = simd_batch{ 24, 0, 2, 7, 12, 17, 22, 27 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  in += 5;
-  return in;
-}
-
-inline static const uint32_t* unpack6_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x3f;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 6-bit bundles 0 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 30 | SafeLoad<uint32_t>(in + 1) << 2, SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) };
-  shifts = simd_batch{ 0, 6, 12, 18, 24, 0, 4, 10 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 6-bit bundles 8 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) >> 28 | SafeLoad<uint32_t>(in + 2) << 4, SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) };
-  shifts = simd_batch{ 16, 22, 0, 2, 8, 14, 20, 26 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 6-bit bundles 16 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) >> 30 | SafeLoad<uint32_t>(in + 4) << 2, SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4) };
-  shifts = simd_batch{ 0, 6, 12, 18, 24, 0, 4, 10 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 6-bit bundles 24 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4) >> 28 | SafeLoad<uint32_t>(in + 5) << 4, SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) };
-  shifts = simd_batch{ 16, 22, 0, 2, 8, 14, 20, 26 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  in += 6;
-  return in;
-}
-
-inline static const uint32_t* unpack7_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x7f;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 7-bit bundles 0 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 28 | SafeLoad<uint32_t>(in + 1) << 4, SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) };
-  shifts = simd_batch{ 0, 7, 14, 21, 0, 3, 10, 17 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 7-bit bundles 8 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) >> 31 | SafeLoad<uint32_t>(in + 2) << 1, SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) >> 27 | SafeLoad<uint32_t>(in + 3) << 5, SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) };
-  shifts = simd_batch{ 24, 0, 6, 13, 20, 0, 2, 9 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 7-bit bundles 16 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) >> 30 | SafeLoad<uint32_t>(in + 4) << 2, SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4) >> 26 | SafeLoad<uint32_t>(in + 5) << 6, SafeLoad<uint32_t>(in + 5) };
-  shifts = simd_batch{ 16, 23, 0, 5, 12, 19, 0, 1 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 7-bit bundles 24 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) >> 29 | SafeLoad<uint32_t>(in + 6) << 3, SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6) };
-  shifts = simd_batch{ 8, 15, 22, 0, 4, 11, 18, 25 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  in += 7;
-  return in;
-}
-
-inline static const uint32_t* unpack8_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0xff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 8-bit bundles 0 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) };
-  shifts = simd_batch{ 0, 8, 16, 24, 0, 8, 16, 24 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 8-bit bundles 8 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) };
-  shifts = simd_batch{ 0, 8, 16, 24, 0, 8, 16, 24 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 8-bit bundles 16 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) };
-  shifts = simd_batch{ 0, 8, 16, 24, 0, 8, 16, 24 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 8-bit bundles 24 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7) };
-  shifts = simd_batch{ 0, 8, 16, 24, 0, 8, 16, 24 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  in += 8;
-  return in;
-}
-
-inline static const uint32_t* unpack9_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x1ff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 9-bit bundles 0 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 27 | SafeLoad<uint32_t>(in + 1) << 5, SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) >> 31 | SafeLoad<uint32_t>(in + 2) << 1 };
-  shifts = simd_batch{ 0, 9, 18, 0, 4, 13, 22, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 9-bit bundles 8 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) >> 26 | SafeLoad<uint32_t>(in + 3) << 6, SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) >> 30 | SafeLoad<uint32_t>(in + 4) << 2, SafeLoad<uint32_t>(in + 4) };
-  shifts = simd_batch{ 8, 17, 0, 3, 12, 21, 0, 7 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 9-bit bundles 16 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4) >> 25 | SafeLoad<uint32_t>(in + 5) << 7, SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) >> 29 | SafeLoad<uint32_t>(in + 6) << 3, SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6) };
-  shifts = simd_batch{ 16, 0, 2, 11, 20, 0, 6, 15 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 9-bit bundles 24 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 6) >> 24 | SafeLoad<uint32_t>(in + 7) << 8, SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7) >> 28 | SafeLoad<uint32_t>(in + 8) << 4, SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 8) };
-  shifts = simd_batch{ 0, 1, 10, 19, 0, 5, 14, 23 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  in += 9;
-  return in;
-}
-
-inline static const uint32_t* unpack10_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x3ff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 10-bit bundles 0 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 30 | SafeLoad<uint32_t>(in + 1) << 2, SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) >> 28 | SafeLoad<uint32_t>(in + 2) << 4, SafeLoad<uint32_t>(in + 2) };
-  shifts = simd_batch{ 0, 10, 20, 0, 8, 18, 0, 6 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 10-bit bundles 8 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) >> 26 | SafeLoad<uint32_t>(in + 3) << 6, SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) >> 24 | SafeLoad<uint32_t>(in + 4) << 8, SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4) };
-  shifts = simd_batch{ 16, 0, 4, 14, 0, 2, 12, 22 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 10-bit bundles 16 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) >> 30 | SafeLoad<uint32_t>(in + 6) << 2, SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6) >> 28 | SafeLoad<uint32_t>(in + 7) << 4, SafeLoad<uint32_t>(in + 7) };
-  shifts = simd_batch{ 0, 10, 20, 0, 8, 18, 0, 6 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 10-bit bundles 24 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7) >> 26 | SafeLoad<uint32_t>(in + 8) << 6, SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 8) >> 24 | SafeLoad<uint32_t>(in + 9) << 8, SafeLoad<uint32_t>(in + 9), SafeLoad<uint32_t>(in + 9), SafeLoad<uint32_t>(in + 9) };
-  shifts = simd_batch{ 16, 0, 4, 14, 0, 2, 12, 22 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  in += 10;
-  return in;
-}
-
-inline static const uint32_t* unpack11_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x7ff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 11-bit bundles 0 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 22 | SafeLoad<uint32_t>(in + 1) << 10, SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) >> 23 | SafeLoad<uint32_t>(in + 2) << 9, SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) };
-  shifts = simd_batch{ 0, 11, 0, 1, 12, 0, 2, 13 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 11-bit bundles 8 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 2) >> 24 | SafeLoad<uint32_t>(in + 3) << 8, SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) >> 25 | SafeLoad<uint32_t>(in + 4) << 7, SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4) >> 26 | SafeLoad<uint32_t>(in + 5) << 6, SafeLoad<uint32_t>(in + 5) };
-  shifts = simd_batch{ 0, 3, 14, 0, 4, 15, 0, 5 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 11-bit bundles 16 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) >> 27 | SafeLoad<uint32_t>(in + 6) << 5, SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6) >> 28 | SafeLoad<uint32_t>(in + 7) << 4, SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7) >> 29 | SafeLoad<uint32_t>(in + 8) << 3 };
-  shifts = simd_batch{ 16, 0, 6, 17, 0, 7, 18, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 11-bit bundles 24 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 8) >> 30 | SafeLoad<uint32_t>(in + 9) << 2, SafeLoad<uint32_t>(in + 9), SafeLoad<uint32_t>(in + 9), SafeLoad<uint32_t>(in + 9) >> 31 | SafeLoad<uint32_t>(in + 10) << 1, SafeLoad<uint32_t>(in + 10), SafeLoad<uint32_t>(in + 10) };
-  shifts = simd_batch{ 8, 19, 0, 9, 20, 0, 10, 21 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  in += 11;
-  return in;
-}
-
-inline static const uint32_t* unpack12_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0xfff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 12-bit bundles 0 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 24 | SafeLoad<uint32_t>(in + 1) << 8, SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) >> 28 | SafeLoad<uint32_t>(in + 2) << 4, SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) };
-  shifts = simd_batch{ 0, 12, 0, 4, 16, 0, 8, 20 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 12-bit bundles 8 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) >> 24 | SafeLoad<uint32_t>(in + 4) << 8, SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4) >> 28 | SafeLoad<uint32_t>(in + 5) << 4, SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) };
-  shifts = simd_batch{ 0, 12, 0, 4, 16, 0, 8, 20 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 12-bit bundles 16 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6) >> 24 | SafeLoad<uint32_t>(in + 7) << 8, SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7) >> 28 | SafeLoad<uint32_t>(in + 8) << 4, SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 8) };
-  shifts = simd_batch{ 0, 12, 0, 4, 16, 0, 8, 20 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 12-bit bundles 24 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 9), SafeLoad<uint32_t>(in + 9), SafeLoad<uint32_t>(in + 9) >> 24 | SafeLoad<uint32_t>(in + 10) << 8, SafeLoad<uint32_t>(in + 10), SafeLoad<uint32_t>(in + 10), SafeLoad<uint32_t>(in + 10) >> 28 | SafeLoad<uint32_t>(in + 11) << 4, SafeLoad<uint32_t>(in + 11), SafeLoad<uint32_t>(in + 11) };
-  shifts = simd_batch{ 0, 12, 0, 4, 16, 0, 8, 20 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  in += 12;
-  return in;
-}
-
-inline static const uint32_t* unpack13_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x1fff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 13-bit bundles 0 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 26 | SafeLoad<uint32_t>(in + 1) << 6, SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) >> 20 | SafeLoad<uint32_t>(in + 2) << 12, SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) >> 27 | SafeLoad<uint32_t>(in + 3) << 5 };
-  shifts = simd_batch{ 0, 13, 0, 7, 0, 1, 14, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 13-bit bundles 8 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) >> 21 | SafeLoad<uint32_t>(in + 4) << 11, SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4) >> 28 | SafeLoad<uint32_t>(in + 5) << 4, SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) >> 22 | SafeLoad<uint32_t>(in + 6) << 10, SafeLoad<uint32_t>(in + 6) };
-  shifts = simd_batch{ 8, 0, 2, 15, 0, 9, 0, 3 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 13-bit bundles 16 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6) >> 29 | SafeLoad<uint32_t>(in + 7) << 3, SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7) >> 23 | SafeLoad<uint32_t>(in + 8) << 9, SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 8) >> 30 | SafeLoad<uint32_t>(in + 9) << 2, SafeLoad<uint32_t>(in + 9) };
-  shifts = simd_batch{ 16, 0, 10, 0, 4, 17, 0, 11 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 13-bit bundles 24 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 9) >> 24 | SafeLoad<uint32_t>(in + 10) << 8, SafeLoad<uint32_t>(in + 10), SafeLoad<uint32_t>(in + 10), SafeLoad<uint32_t>(in + 10) >> 31 | SafeLoad<uint32_t>(in + 11) << 1, SafeLoad<uint32_t>(in + 11), SafeLoad<uint32_t>(in + 11) >> 25 | SafeLoad<uint32_t>(in + 12) << 7, SafeLoad<uint32_t>(in + 12), SafeLoad<uint32_t>(in + 12) };
-  shifts = simd_batch{ 0, 5, 18, 0, 12, 0, 6, 19 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  in += 13;
-  return in;
-}
-
-inline static const uint32_t* unpack14_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x3fff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 14-bit bundles 0 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 28 | SafeLoad<uint32_t>(in + 1) << 4, SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) >> 24 | SafeLoad<uint32_t>(in + 2) << 8, SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) >> 20 | SafeLoad<uint32_t>(in + 3) << 12, SafeLoad<uint32_t>(in + 3) };
-  shifts = simd_batch{ 0, 14, 0, 10, 0, 6, 0, 2 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 14-bit bundles 8 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) >> 30 | SafeLoad<uint32_t>(in + 4) << 2, SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4) >> 26 | SafeLoad<uint32_t>(in + 5) << 6, SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) >> 22 | SafeLoad<uint32_t>(in + 6) << 10, SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6) };
-  shifts = simd_batch{ 16, 0, 12, 0, 8, 0, 4, 18 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 14-bit bundles 16 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7) >> 28 | SafeLoad<uint32_t>(in + 8) << 4, SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 8) >> 24 | SafeLoad<uint32_t>(in + 9) << 8, SafeLoad<uint32_t>(in + 9), SafeLoad<uint32_t>(in + 9) >> 20 | SafeLoad<uint32_t>(in + 10) << 12, SafeLoad<uint32_t>(in + 10) };
-  shifts = simd_batch{ 0, 14, 0, 10, 0, 6, 0, 2 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 14-bit bundles 24 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 10), SafeLoad<uint32_t>(in + 10) >> 30 | SafeLoad<uint32_t>(in + 11) << 2, SafeLoad<uint32_t>(in + 11), SafeLoad<uint32_t>(in + 11) >> 26 | SafeLoad<uint32_t>(in + 12) << 6, SafeLoad<uint32_t>(in + 12), SafeLoad<uint32_t>(in + 12) >> 22 | SafeLoad<uint32_t>(in + 13) << 10, SafeLoad<uint32_t>(in + 13), SafeLoad<uint32_t>(in + 13) };
-  shifts = simd_batch{ 16, 0, 12, 0, 8, 0, 4, 18 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  in += 14;
-  return in;
-}
-
-inline static const uint32_t* unpack15_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x7fff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 15-bit bundles 0 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 30 | SafeLoad<uint32_t>(in + 1) << 2, SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) >> 28 | SafeLoad<uint32_t>(in + 2) << 4, SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) >> 26 | SafeLoad<uint32_t>(in + 3) << 6, SafeLoad<uint32_t>(in + 3) };
-  shifts = simd_batch{ 0, 15, 0, 13, 0, 11, 0, 9 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 15-bit bundles 8 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 3) >> 24 | SafeLoad<uint32_t>(in + 4) << 8, SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4) >> 22 | SafeLoad<uint32_t>(in + 5) << 10, SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) >> 20 | SafeLoad<uint32_t>(in + 6) << 12, SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6) >> 18 | SafeLoad<uint32_t>(in + 7) << 14, SafeLoad<uint32_t>(in + 7) };
-  shifts = simd_batch{ 0, 7, 0, 5, 0, 3, 0, 1 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 15-bit bundles 16 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7) >> 31 | SafeLoad<uint32_t>(in + 8) << 1, SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 8) >> 29 | SafeLoad<uint32_t>(in + 9) << 3, SafeLoad<uint32_t>(in + 9), SafeLoad<uint32_t>(in + 9) >> 27 | SafeLoad<uint32_t>(in + 10) << 5, SafeLoad<uint32_t>(in + 10), SafeLoad<uint32_t>(in + 10) >> 25 | SafeLoad<uint32_t>(in + 11) << 7 };
-  shifts = simd_batch{ 16, 0, 14, 0, 12, 0, 10, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 15-bit bundles 24 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 11), SafeLoad<uint32_t>(in + 11) >> 23 | SafeLoad<uint32_t>(in + 12) << 9, SafeLoad<uint32_t>(in + 12), SafeLoad<uint32_t>(in + 12) >> 21 | SafeLoad<uint32_t>(in + 13) << 11, SafeLoad<uint32_t>(in + 13), SafeLoad<uint32_t>(in + 13) >> 19 | SafeLoad<uint32_t>(in + 14) << 13, SafeLoad<uint32_t>(in + 14), SafeLoad<uint32_t>(in + 14) };
-  shifts = simd_batch{ 8, 0, 6, 0, 4, 0, 2, 17 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  in += 15;
-  return in;
-}
-
-inline static const uint32_t* unpack16_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0xffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 16-bit bundles 0 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) };
-  shifts = simd_batch{ 0, 16, 0, 16, 0, 16, 0, 16 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 16-bit bundles 8 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7) };
-  shifts = simd_batch{ 0, 16, 0, 16, 0, 16, 0, 16 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 16-bit bundles 16 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 9), SafeLoad<uint32_t>(in + 9), SafeLoad<uint32_t>(in + 10), SafeLoad<uint32_t>(in + 10), SafeLoad<uint32_t>(in + 11), SafeLoad<uint32_t>(in + 11) };
-  shifts = simd_batch{ 0, 16, 0, 16, 0, 16, 0, 16 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 16-bit bundles 24 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 12), SafeLoad<uint32_t>(in + 12), SafeLoad<uint32_t>(in + 13), SafeLoad<uint32_t>(in + 13), SafeLoad<uint32_t>(in + 14), SafeLoad<uint32_t>(in + 14), SafeLoad<uint32_t>(in + 15), SafeLoad<uint32_t>(in + 15) };
-  shifts = simd_batch{ 0, 16, 0, 16, 0, 16, 0, 16 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  in += 16;
-  return in;
-}
-
-inline static const uint32_t* unpack17_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x1ffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 17-bit bundles 0 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 17 | SafeLoad<uint32_t>(in + 1) << 15, SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) >> 19 | SafeLoad<uint32_t>(in + 2) << 13, SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) >> 21 | SafeLoad<uint32_t>(in + 3) << 11, SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) >> 23 | SafeLoad<uint32_t>(in + 4) << 9 };
-  shifts = simd_batch{ 0, 0, 2, 0, 4, 0, 6, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 17-bit bundles 8 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4) >> 25 | SafeLoad<uint32_t>(in + 5) << 7, SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) >> 27 | SafeLoad<uint32_t>(in + 6) << 5, SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6) >> 29 | SafeLoad<uint32_t>(in + 7) << 3, SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7) >> 31 | SafeLoad<uint32_t>(in + 8) << 1 };
-  shifts = simd_batch{ 8, 0, 10, 0, 12, 0, 14, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 17-bit bundles 16 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 8) >> 16 | SafeLoad<uint32_t>(in + 9) << 16, SafeLoad<uint32_t>(in + 9), SafeLoad<uint32_t>(in + 9) >> 18 | SafeLoad<uint32_t>(in + 10) << 14, SafeLoad<uint32_t>(in + 10), SafeLoad<uint32_t>(in + 10) >> 20 | SafeLoad<uint32_t>(in + 11) << 12, SafeLoad<uint32_t>(in + 11), SafeLoad<uint32_t>(in + 11) >> 22 | SafeLoad<uint32_t>(in + 12) << 10, SafeLoad<uint32_t>(in + 12) };
-  shifts = simd_batch{ 0, 1, 0, 3, 0, 5, 0, 7 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 17-bit bundles 24 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 12) >> 24 | SafeLoad<uint32_t>(in + 13) << 8, SafeLoad<uint32_t>(in + 13), SafeLoad<uint32_t>(in + 13) >> 26 | SafeLoad<uint32_t>(in + 14) << 6, SafeLoad<uint32_t>(in + 14), SafeLoad<uint32_t>(in + 14) >> 28 | SafeLoad<uint32_t>(in + 15) << 4, SafeLoad<uint32_t>(in + 15), SafeLoad<uint32_t>(in + 15) >> 30 | SafeLoad<uint32_t>(in + 16) << 2, SafeLoad<uint32_t>(in + 16) };
-  shifts = simd_batch{ 0, 9, 0, 11, 0, 13, 0, 15 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  in += 17;
-  return in;
-}
-
-inline static const uint32_t* unpack18_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x3ffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 18-bit bundles 0 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 18 | SafeLoad<uint32_t>(in + 1) << 14, SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) >> 22 | SafeLoad<uint32_t>(in + 2) << 10, SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) >> 26 | SafeLoad<uint32_t>(in + 3) << 6, SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) >> 30 | SafeLoad<uint32_t>(in + 4) << 2 };
-  shifts = simd_batch{ 0, 0, 4, 0, 8, 0, 12, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 18-bit bundles 8 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 4) >> 16 | SafeLoad<uint32_t>(in + 5) << 16, SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) >> 20 | SafeLoad<uint32_t>(in + 6) << 12, SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6) >> 24 | SafeLoad<uint32_t>(in + 7) << 8, SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7) >> 28 | SafeLoad<uint32_t>(in + 8) << 4, SafeLoad<uint32_t>(in + 8) };
-  shifts = simd_batch{ 0, 2, 0, 6, 0, 10, 0, 14 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 18-bit bundles 16 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 9), SafeLoad<uint32_t>(in + 9) >> 18 | SafeLoad<uint32_t>(in + 10) << 14, SafeLoad<uint32_t>(in + 10), SafeLoad<uint32_t>(in + 10) >> 22 | SafeLoad<uint32_t>(in + 11) << 10, SafeLoad<uint32_t>(in + 11), SafeLoad<uint32_t>(in + 11) >> 26 | SafeLoad<uint32_t>(in + 12) << 6, SafeLoad<uint32_t>(in + 12), SafeLoad<uint32_t>(in + 12) >> 30 | SafeLoad<uint32_t>(in + 13) << 2 };
-  shifts = simd_batch{ 0, 0, 4, 0, 8, 0, 12, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 18-bit bundles 24 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 13) >> 16 | SafeLoad<uint32_t>(in + 14) << 16, SafeLoad<uint32_t>(in + 14), SafeLoad<uint32_t>(in + 14) >> 20 | SafeLoad<uint32_t>(in + 15) << 12, SafeLoad<uint32_t>(in + 15), SafeLoad<uint32_t>(in + 15) >> 24 | SafeLoad<uint32_t>(in + 16) << 8, SafeLoad<uint32_t>(in + 16), SafeLoad<uint32_t>(in + 16) >> 28 | SafeLoad<uint32_t>(in + 17) << 4, SafeLoad<uint32_t>(in + 17) };
-  shifts = simd_batch{ 0, 2, 0, 6, 0, 10, 0, 14 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  in += 18;
-  return in;
-}
-
-inline static const uint32_t* unpack19_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x7ffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 19-bit bundles 0 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 19 | SafeLoad<uint32_t>(in + 1) << 13, SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) >> 25 | SafeLoad<uint32_t>(in + 2) << 7, SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) >> 31 | SafeLoad<uint32_t>(in + 3) << 1, SafeLoad<uint32_t>(in + 3) >> 18 | SafeLoad<uint32_t>(in + 4) << 14, SafeLoad<uint32_t>(in + 4) };
-  shifts = simd_batch{ 0, 0, 6, 0, 12, 0, 0, 5 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 19-bit bundles 8 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 4) >> 24 | SafeLoad<uint32_t>(in + 5) << 8, SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) >> 30 | SafeLoad<uint32_t>(in + 6) << 2, SafeLoad<uint32_t>(in + 6) >> 17 | SafeLoad<uint32_t>(in + 7) << 15, SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7) >> 23 | SafeLoad<uint32_t>(in + 8) << 9, SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 8) >> 29 | SafeLoad<uint32_t>(in + 9) << 3 };
-  shifts = simd_batch{ 0, 11, 0, 0, 4, 0, 10, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 19-bit bundles 16 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 9) >> 16 | SafeLoad<uint32_t>(in + 10) << 16, SafeLoad<uint32_t>(in + 10), SafeLoad<uint32_t>(in + 10) >> 22 | SafeLoad<uint32_t>(in + 11) << 10, SafeLoad<uint32_t>(in + 11), SafeLoad<uint32_t>(in + 11) >> 28 | SafeLoad<uint32_t>(in + 12) << 4, SafeLoad<uint32_t>(in + 12) >> 15 | SafeLoad<uint32_t>(in + 13) << 17, SafeLoad<uint32_t>(in + 13), SafeLoad<uint32_t>(in + 13) >> 21 | SafeLoad<uint32_t>(in + 14) << 11 };
-  shifts = simd_batch{ 0, 3, 0, 9, 0, 0, 2, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 19-bit bundles 24 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 14), SafeLoad<uint32_t>(in + 14) >> 27 | SafeLoad<uint32_t>(in + 15) << 5, SafeLoad<uint32_t>(in + 15) >> 14 | SafeLoad<uint32_t>(in + 16) << 18, SafeLoad<uint32_t>(in + 16), SafeLoad<uint32_t>(in + 16) >> 20 | SafeLoad<uint32_t>(in + 17) << 12, SafeLoad<uint32_t>(in + 17), SafeLoad<uint32_t>(in + 17) >> 26 | SafeLoad<uint32_t>(in + 18) << 6, SafeLoad<uint32_t>(in + 18) };
-  shifts = simd_batch{ 8, 0, 0, 1, 0, 7, 0, 13 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  in += 19;
-  return in;
-}
-
-inline static const uint32_t* unpack20_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0xfffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 20-bit bundles 0 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 20 | SafeLoad<uint32_t>(in + 1) << 12, SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) >> 28 | SafeLoad<uint32_t>(in + 2) << 4, SafeLoad<uint32_t>(in + 2) >> 16 | SafeLoad<uint32_t>(in + 3) << 16, SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) >> 24 | SafeLoad<uint32_t>(in + 4) << 8, SafeLoad<uint32_t>(in + 4) };
-  shifts = simd_batch{ 0, 0, 8, 0, 0, 4, 0, 12 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 20-bit bundles 8 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) >> 20 | SafeLoad<uint32_t>(in + 6) << 12, SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6) >> 28 | SafeLoad<uint32_t>(in + 7) << 4, SafeLoad<uint32_t>(in + 7) >> 16 | SafeLoad<uint32_t>(in + 8) << 16, SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 8) >> 24 | SafeLoad<uint32_t>(in + 9) << 8, SafeLoad<uint32_t>(in + 9) };
-  shifts = simd_batch{ 0, 0, 8, 0, 0, 4, 0, 12 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 20-bit bundles 16 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 10), SafeLoad<uint32_t>(in + 10) >> 20 | SafeLoad<uint32_t>(in + 11) << 12, SafeLoad<uint32_t>(in + 11), SafeLoad<uint32_t>(in + 11) >> 28 | SafeLoad<uint32_t>(in + 12) << 4, SafeLoad<uint32_t>(in + 12) >> 16 | SafeLoad<uint32_t>(in + 13) << 16, SafeLoad<uint32_t>(in + 13), SafeLoad<uint32_t>(in + 13) >> 24 | SafeLoad<uint32_t>(in + 14) << 8, SafeLoad<uint32_t>(in + 14) };
-  shifts = simd_batch{ 0, 0, 8, 0, 0, 4, 0, 12 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 20-bit bundles 24 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 15), SafeLoad<uint32_t>(in + 15) >> 20 | SafeLoad<uint32_t>(in + 16) << 12, SafeLoad<uint32_t>(in + 16), SafeLoad<uint32_t>(in + 16) >> 28 | SafeLoad<uint32_t>(in + 17) << 4, SafeLoad<uint32_t>(in + 17) >> 16 | SafeLoad<uint32_t>(in + 18) << 16, SafeLoad<uint32_t>(in + 18), SafeLoad<uint32_t>(in + 18) >> 24 | SafeLoad<uint32_t>(in + 19) << 8, SafeLoad<uint32_t>(in + 19) };
-  shifts = simd_batch{ 0, 0, 8, 0, 0, 4, 0, 12 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  in += 20;
-  return in;
-}
-
-inline static const uint32_t* unpack21_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x1fffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 21-bit bundles 0 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 21 | SafeLoad<uint32_t>(in + 1) << 11, SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) >> 31 | SafeLoad<uint32_t>(in + 2) << 1, SafeLoad<uint32_t>(in + 2) >> 20 | SafeLoad<uint32_t>(in + 3) << 12, SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) >> 30 | SafeLoad<uint32_t>(in + 4) << 2, SafeLoad<uint32_t>(in + 4) >> 19 | SafeLoad<uint32_t>(in + 5) << 13 };
-  shifts = simd_batch{ 0, 0, 10, 0, 0, 9, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 21-bit bundles 8 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) >> 29 | SafeLoad<uint32_t>(in + 6) << 3, SafeLoad<uint32_t>(in + 6) >> 18 | SafeLoad<uint32_t>(in + 7) << 14, SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7) >> 28 | SafeLoad<uint32_t>(in + 8) << 4, SafeLoad<uint32_t>(in + 8) >> 17 | SafeLoad<uint32_t>(in + 9) << 15, SafeLoad<uint32_t>(in + 9), SafeLoad<uint32_t>(in + 9) >> 27 | SafeLoad<uint32_t>(in + 10) << 5 };
-  shifts = simd_batch{ 8, 0, 0, 7, 0, 0, 6, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 21-bit bundles 16 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 10) >> 16 | SafeLoad<uint32_t>(in + 11) << 16, SafeLoad<uint32_t>(in + 11), SafeLoad<uint32_t>(in + 11) >> 26 | SafeLoad<uint32_t>(in + 12) << 6, SafeLoad<uint32_t>(in + 12) >> 15 | SafeLoad<uint32_t>(in + 13) << 17, SafeLoad<uint32_t>(in + 13), SafeLoad<uint32_t>(in + 13) >> 25 | SafeLoad<uint32_t>(in + 14) << 7, SafeLoad<uint32_t>(in + 14) >> 14 | SafeLoad<uint32_t>(in + 15) << 18, SafeLoad<uint32_t>(in + 15) };
-  shifts = simd_batch{ 0, 5, 0, 0, 4, 0, 0, 3 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 21-bit bundles 24 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 15) >> 24 | SafeLoad<uint32_t>(in + 16) << 8, SafeLoad<uint32_t>(in + 16) >> 13 | SafeLoad<uint32_t>(in + 17) << 19, SafeLoad<uint32_t>(in + 17), SafeLoad<uint32_t>(in + 17) >> 23 | SafeLoad<uint32_t>(in + 18) << 9, SafeLoad<uint32_t>(in + 18) >> 12 | SafeLoad<uint32_t>(in + 19) << 20, SafeLoad<uint32_t>(in + 19), SafeLoad<uint32_t>(in + 19) >> 22 | SafeLoad<uint32_t>(in + 20) << 10, SafeLoad<uint32_t>(in + 20) };
-  shifts = simd_batch{ 0, 0, 2, 0, 0, 1, 0, 11 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  in += 21;
-  return in;
-}
-
-inline static const uint32_t* unpack22_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x3fffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 22-bit bundles 0 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 22 | SafeLoad<uint32_t>(in + 1) << 10, SafeLoad<uint32_t>(in + 1) >> 12 | SafeLoad<uint32_t>(in + 2) << 20, SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) >> 24 | SafeLoad<uint32_t>(in + 3) << 8, SafeLoad<uint32_t>(in + 3) >> 14 | SafeLoad<uint32_t>(in + 4) << 18, SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4) >> 26 | SafeLoad<uint32_t>(in + 5) << 6 };
-  shifts = simd_batch{ 0, 0, 0, 2, 0, 0, 4, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 22-bit bundles 8 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 5) >> 16 | SafeLoad<uint32_t>(in + 6) << 16, SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6) >> 28 | SafeLoad<uint32_t>(in + 7) << 4, SafeLoad<uint32_t>(in + 7) >> 18 | SafeLoad<uint32_t>(in + 8) << 14, SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 8) >> 30 | SafeLoad<uint32_t>(in + 9) << 2, SafeLoad<uint32_t>(in + 9) >> 20 | SafeLoad<uint32_t>(in + 10) << 12, SafeLoad<uint32_t>(in + 10) };
-  shifts = simd_batch{ 0, 6, 0, 0, 8, 0, 0, 10 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 22-bit bundles 16 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 11), SafeLoad<uint32_t>(in + 11) >> 22 | SafeLoad<uint32_t>(in + 12) << 10, SafeLoad<uint32_t>(in + 12) >> 12 | SafeLoad<uint32_t>(in + 13) << 20, SafeLoad<uint32_t>(in + 13), SafeLoad<uint32_t>(in + 13) >> 24 | SafeLoad<uint32_t>(in + 14) << 8, SafeLoad<uint32_t>(in + 14) >> 14 | SafeLoad<uint32_t>(in + 15) << 18, SafeLoad<uint32_t>(in + 15), SafeLoad<uint32_t>(in + 15) >> 26 | SafeLoad<uint32_t>(in + 16) << 6 };
-  shifts = simd_batch{ 0, 0, 0, 2, 0, 0, 4, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 22-bit bundles 24 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 16) >> 16 | SafeLoad<uint32_t>(in + 17) << 16, SafeLoad<uint32_t>(in + 17), SafeLoad<uint32_t>(in + 17) >> 28 | SafeLoad<uint32_t>(in + 18) << 4, SafeLoad<uint32_t>(in + 18) >> 18 | SafeLoad<uint32_t>(in + 19) << 14, SafeLoad<uint32_t>(in + 19), SafeLoad<uint32_t>(in + 19) >> 30 | SafeLoad<uint32_t>(in + 20) << 2, SafeLoad<uint32_t>(in + 20) >> 20 | SafeLoad<uint32_t>(in + 21) << 12, SafeLoad<uint32_t>(in + 21) };
-  shifts = simd_batch{ 0, 6, 0, 0, 8, 0, 0, 10 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  in += 22;
-  return in;
-}
-
-inline static const uint32_t* unpack23_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x7fffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 23-bit bundles 0 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 23 | SafeLoad<uint32_t>(in + 1) << 9, SafeLoad<uint32_t>(in + 1) >> 14 | SafeLoad<uint32_t>(in + 2) << 18, SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) >> 28 | SafeLoad<uint32_t>(in + 3) << 4, SafeLoad<uint32_t>(in + 3) >> 19 | SafeLoad<uint32_t>(in + 4) << 13, SafeLoad<uint32_t>(in + 4) >> 10 | SafeLoad<uint32_t>(in + 5) << 22, SafeLoad<uint32_t>(in + 5) };
-  shifts = simd_batch{ 0, 0, 0, 5, 0, 0, 0, 1 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 23-bit bundles 8 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 5) >> 24 | SafeLoad<uint32_t>(in + 6) << 8, SafeLoad<uint32_t>(in + 6) >> 15 | SafeLoad<uint32_t>(in + 7) << 17, SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7) >> 29 | SafeLoad<uint32_t>(in + 8) << 3, SafeLoad<uint32_t>(in + 8) >> 20 | SafeLoad<uint32_t>(in + 9) << 12, SafeLoad<uint32_t>(in + 9) >> 11 | SafeLoad<uint32_t>(in + 10) << 21, SafeLoad<uint32_t>(in + 10), SafeLoad<uint32_t>(in + 10) >> 25 | SafeLoad<uint32_t>(in + 11) << 7 };
-  shifts = simd_batch{ 0, 0, 6, 0, 0, 0, 2, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 23-bit bundles 16 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 11) >> 16 | SafeLoad<uint32_t>(in + 12) << 16, SafeLoad<uint32_t>(in + 12), SafeLoad<uint32_t>(in + 12) >> 30 | SafeLoad<uint32_t>(in + 13) << 2, SafeLoad<uint32_t>(in + 13) >> 21 | SafeLoad<uint32_t>(in + 14) << 11, SafeLoad<uint32_t>(in + 14) >> 12 | SafeLoad<uint32_t>(in + 15) << 20, SafeLoad<uint32_t>(in + 15), SafeLoad<uint32_t>(in + 15) >> 26 | SafeLoad<uint32_t>(in + 16) << 6, SafeLoad<uint32_t>(in + 16) >> 17 | SafeLoad<uint32_t>(in + 17) << 15 };
-  shifts = simd_batch{ 0, 7, 0, 0, 0, 3, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 23-bit bundles 24 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 17), SafeLoad<uint32_t>(in + 17) >> 31 | SafeLoad<uint32_t>(in + 18) << 1, SafeLoad<uint32_t>(in + 18) >> 22 | SafeLoad<uint32_t>(in + 19) << 10, SafeLoad<uint32_t>(in + 19) >> 13 | SafeLoad<uint32_t>(in + 20) << 19, SafeLoad<uint32_t>(in + 20), SafeLoad<uint32_t>(in + 20) >> 27 | SafeLoad<uint32_t>(in + 21) << 5, SafeLoad<uint32_t>(in + 21) >> 18 | SafeLoad<uint32_t>(in + 22) << 14, SafeLoad<uint32_t>(in + 22) };
-  shifts = simd_batch{ 8, 0, 0, 0, 4, 0, 0, 9 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  in += 23;
-  return in;
-}
-
-inline static const uint32_t* unpack24_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0xffffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 24-bit bundles 0 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 24 | SafeLoad<uint32_t>(in + 1) << 8, SafeLoad<uint32_t>(in + 1) >> 16 | SafeLoad<uint32_t>(in + 2) << 16, SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) >> 24 | SafeLoad<uint32_t>(in + 4) << 8, SafeLoad<uint32_t>(in + 4) >> 16 | SafeLoad<uint32_t>(in + 5) << 16, SafeLoad<uint32_t>(in + 5) };
-  shifts = simd_batch{ 0, 0, 0, 8, 0, 0, 0, 8 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 24-bit bundles 8 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6) >> 24 | SafeLoad<uint32_t>(in + 7) << 8, SafeLoad<uint32_t>(in + 7) >> 16 | SafeLoad<uint32_t>(in + 8) << 16, SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 9), SafeLoad<uint32_t>(in + 9) >> 24 | SafeLoad<uint32_t>(in + 10) << 8, SafeLoad<uint32_t>(in + 10) >> 16 | SafeLoad<uint32_t>(in + 11) << 16, SafeLoad<uint32_t>(in + 11) };
-  shifts = simd_batch{ 0, 0, 0, 8, 0, 0, 0, 8 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 24-bit bundles 16 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 12), SafeLoad<uint32_t>(in + 12) >> 24 | SafeLoad<uint32_t>(in + 13) << 8, SafeLoad<uint32_t>(in + 13) >> 16 | SafeLoad<uint32_t>(in + 14) << 16, SafeLoad<uint32_t>(in + 14), SafeLoad<uint32_t>(in + 15), SafeLoad<uint32_t>(in + 15) >> 24 | SafeLoad<uint32_t>(in + 16) << 8, SafeLoad<uint32_t>(in + 16) >> 16 | SafeLoad<uint32_t>(in + 17) << 16, SafeLoad<uint32_t>(in + 17) };
-  shifts = simd_batch{ 0, 0, 0, 8, 0, 0, 0, 8 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 24-bit bundles 24 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 18), SafeLoad<uint32_t>(in + 18) >> 24 | SafeLoad<uint32_t>(in + 19) << 8, SafeLoad<uint32_t>(in + 19) >> 16 | SafeLoad<uint32_t>(in + 20) << 16, SafeLoad<uint32_t>(in + 20), SafeLoad<uint32_t>(in + 21), SafeLoad<uint32_t>(in + 21) >> 24 | SafeLoad<uint32_t>(in + 22) << 8, SafeLoad<uint32_t>(in + 22) >> 16 | SafeLoad<uint32_t>(in + 23) << 16, SafeLoad<uint32_t>(in + 23) };
-  shifts = simd_batch{ 0, 0, 0, 8, 0, 0, 0, 8 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  in += 24;
-  return in;
-}
-
-inline static const uint32_t* unpack25_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x1ffffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 25-bit bundles 0 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 25 | SafeLoad<uint32_t>(in + 1) << 7, SafeLoad<uint32_t>(in + 1) >> 18 | SafeLoad<uint32_t>(in + 2) << 14, SafeLoad<uint32_t>(in + 2) >> 11 | SafeLoad<uint32_t>(in + 3) << 21, SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) >> 29 | SafeLoad<uint32_t>(in + 4) << 3, SafeLoad<uint32_t>(in + 4) >> 22 | SafeLoad<uint32_t>(in + 5) << 10, SafeLoad<uint32_t>(in + 5) >> 15 | SafeLoad<uint32_t>(in + 6) << 17 };
-  shifts = simd_batch{ 0, 0, 0, 0, 4, 0, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 25-bit bundles 8 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 6) >> 8 | SafeLoad<uint32_t>(in + 7) << 24, SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7) >> 26 | SafeLoad<uint32_t>(in + 8) << 6, SafeLoad<uint32_t>(in + 8) >> 19 | SafeLoad<uint32_t>(in + 9) << 13, SafeLoad<uint32_t>(in + 9) >> 12 | SafeLoad<uint32_t>(in + 10) << 20, SafeLoad<uint32_t>(in + 10), SafeLoad<uint32_t>(in + 10) >> 30 | SafeLoad<uint32_t>(in + 11) << 2, SafeLoad<uint32_t>(in + 11) >> 23 | SafeLoad<uint32_t>(in + 12) << 9 };
-  shifts = simd_batch{ 0, 1, 0, 0, 0, 5, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 25-bit bundles 16 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 12) >> 16 | SafeLoad<uint32_t>(in + 13) << 16, SafeLoad<uint32_t>(in + 13) >> 9 | SafeLoad<uint32_t>(in + 14) << 23, SafeLoad<uint32_t>(in + 14), SafeLoad<uint32_t>(in + 14) >> 27 | SafeLoad<uint32_t>(in + 15) << 5, SafeLoad<uint32_t>(in + 15) >> 20 | SafeLoad<uint32_t>(in + 16) << 12, SafeLoad<uint32_t>(in + 16) >> 13 | SafeLoad<uint32_t>(in + 17) << 19, SafeLoad<uint32_t>(in + 17), SafeLoad<uint32_t>(in + 17) >> 31 | SafeLoad<uint32_t>(in + 18) << 1 };
-  shifts = simd_batch{ 0, 0, 2, 0, 0, 0, 6, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 25-bit bundles 24 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 18) >> 24 | SafeLoad<uint32_t>(in + 19) << 8, SafeLoad<uint32_t>(in + 19) >> 17 | SafeLoad<uint32_t>(in + 20) << 15, SafeLoad<uint32_t>(in + 20) >> 10 | SafeLoad<uint32_t>(in + 21) << 22, SafeLoad<uint32_t>(in + 21), SafeLoad<uint32_t>(in + 21) >> 28 | SafeLoad<uint32_t>(in + 22) << 4, SafeLoad<uint32_t>(in + 22) >> 21 | SafeLoad<uint32_t>(in + 23) << 11, SafeLoad<uint32_t>(in + 23) >> 14 | SafeLoad<uint32_t>(in + 24) << 18, SafeLoad<uint32_t>(in + 24) };
-  shifts = simd_batch{ 0, 0, 0, 3, 0, 0, 0, 7 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  in += 25;
-  return in;
-}
-
-inline static const uint32_t* unpack26_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x3ffffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 26-bit bundles 0 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 26 | SafeLoad<uint32_t>(in + 1) << 6, SafeLoad<uint32_t>(in + 1) >> 20 | SafeLoad<uint32_t>(in + 2) << 12, SafeLoad<uint32_t>(in + 2) >> 14 | SafeLoad<uint32_t>(in + 3) << 18, SafeLoad<uint32_t>(in + 3) >> 8 | SafeLoad<uint32_t>(in + 4) << 24, SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4) >> 28 | SafeLoad<uint32_t>(in + 5) << 4, SafeLoad<uint32_t>(in + 5) >> 22 | SafeLoad<uint32_t>(in + 6) << 10 };
-  shifts = simd_batch{ 0, 0, 0, 0, 0, 2, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 26-bit bundles 8 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 6) >> 16 | SafeLoad<uint32_t>(in + 7) << 16, SafeLoad<uint32_t>(in + 7) >> 10 | SafeLoad<uint32_t>(in + 8) << 22, SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 8) >> 30 | SafeLoad<uint32_t>(in + 9) << 2, SafeLoad<uint32_t>(in + 9) >> 24 | SafeLoad<uint32_t>(in + 10) << 8, SafeLoad<uint32_t>(in + 10) >> 18 | SafeLoad<uint32_t>(in + 11) << 14, SafeLoad<uint32_t>(in + 11) >> 12 | SafeLoad<uint32_t>(in + 12) << 20, SafeLoad<uint32_t>(in + 12) };
-  shifts = simd_batch{ 0, 0, 4, 0, 0, 0, 0, 6 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 26-bit bundles 16 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 13), SafeLoad<uint32_t>(in + 13) >> 26 | SafeLoad<uint32_t>(in + 14) << 6, SafeLoad<uint32_t>(in + 14) >> 20 | SafeLoad<uint32_t>(in + 15) << 12, SafeLoad<uint32_t>(in + 15) >> 14 | SafeLoad<uint32_t>(in + 16) << 18, SafeLoad<uint32_t>(in + 16) >> 8 | SafeLoad<uint32_t>(in + 17) << 24, SafeLoad<uint32_t>(in + 17), SafeLoad<uint32_t>(in + 17) >> 28 | SafeLoad<uint32_t>(in + 18) << 4, SafeLoad<uint32_t>(in + 18) >> 22 | SafeLoad<uint32_t>(in + 19) << 10 };
-  shifts = simd_batch{ 0, 0, 0, 0, 0, 2, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 26-bit bundles 24 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 19) >> 16 | SafeLoad<uint32_t>(in + 20) << 16, SafeLoad<uint32_t>(in + 20) >> 10 | SafeLoad<uint32_t>(in + 21) << 22, SafeLoad<uint32_t>(in + 21), SafeLoad<uint32_t>(in + 21) >> 30 | SafeLoad<uint32_t>(in + 22) << 2, SafeLoad<uint32_t>(in + 22) >> 24 | SafeLoad<uint32_t>(in + 23) << 8, SafeLoad<uint32_t>(in + 23) >> 18 | SafeLoad<uint32_t>(in + 24) << 14, SafeLoad<uint32_t>(in + 24) >> 12 | SafeLoad<uint32_t>(in + 25) << 20, SafeLoad<uint32_t>(in + 25) };
-  shifts = simd_batch{ 0, 0, 4, 0, 0, 0, 0, 6 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  in += 26;
-  return in;
-}
-
-inline static const uint32_t* unpack27_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x7ffffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 27-bit bundles 0 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 27 | SafeLoad<uint32_t>(in + 1) << 5, SafeLoad<uint32_t>(in + 1) >> 22 | SafeLoad<uint32_t>(in + 2) << 10, SafeLoad<uint32_t>(in + 2) >> 17 | SafeLoad<uint32_t>(in + 3) << 15, SafeLoad<uint32_t>(in + 3) >> 12 | SafeLoad<uint32_t>(in + 4) << 20, SafeLoad<uint32_t>(in + 4) >> 7 | SafeLoad<uint32_t>(in + 5) << 25, SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) >> 29 | SafeLoad<uint32_t>(in + 6) << 3 };
-  shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 2, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 27-bit bundles 8 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 6) >> 24 | SafeLoad<uint32_t>(in + 7) << 8, SafeLoad<uint32_t>(in + 7) >> 19 | SafeLoad<uint32_t>(in + 8) << 13, SafeLoad<uint32_t>(in + 8) >> 14 | SafeLoad<uint32_t>(in + 9) << 18, SafeLoad<uint32_t>(in + 9) >> 9 | SafeLoad<uint32_t>(in + 10) << 23, SafeLoad<uint32_t>(in + 10), SafeLoad<uint32_t>(in + 10) >> 31 | SafeLoad<uint32_t>(in + 11) << 1, SafeLoad<uint32_t>(in + 11) >> 26 | SafeLoad<uint32_t>(in + 12) << 6, SafeLoad<uint32_t>(in + 12) >> 21 | SafeLoad<uint32_t>(in + 13) << 11 };
-  shifts = simd_batch{ 0, 0, 0, 0, 4, 0, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 27-bit bundles 16 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 13) >> 16 | SafeLoad<uint32_t>(in + 14) << 16, SafeLoad<uint32_t>(in + 14) >> 11 | SafeLoad<uint32_t>(in + 15) << 21, SafeLoad<uint32_t>(in + 15) >> 6 | SafeLoad<uint32_t>(in + 16) << 26, SafeLoad<uint32_t>(in + 16), SafeLoad<uint32_t>(in + 16) >> 28 | SafeLoad<uint32_t>(in + 17) << 4, SafeLoad<uint32_t>(in + 17) >> 23 | SafeLoad<uint32_t>(in + 18) << 9, SafeLoad<uint32_t>(in + 18) >> 18 | SafeLoad<uint32_t>(in + 19) << 14, SafeLoad<uint32_t>(in + 19) >> 13 | SafeLoad<uint32_t>(in + 20) << 19 };
-  shifts = simd_batch{ 0, 0, 0, 1, 0, 0, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 27-bit bundles 24 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 20) >> 8 | SafeLoad<uint32_t>(in + 21) << 24, SafeLoad<uint32_t>(in + 21), SafeLoad<uint32_t>(in + 21) >> 30 | SafeLoad<uint32_t>(in + 22) << 2, SafeLoad<uint32_t>(in + 22) >> 25 | SafeLoad<uint32_t>(in + 23) << 7, SafeLoad<uint32_t>(in + 23) >> 20 | SafeLoad<uint32_t>(in + 24) << 12, SafeLoad<uint32_t>(in + 24) >> 15 | SafeLoad<uint32_t>(in + 25) << 17, SafeLoad<uint32_t>(in + 25) >> 10 | SafeLoad<uint32_t>(in + 26) << 22, SafeLoad<uint32_t>(in + 26) };
-  shifts = simd_batch{ 0, 3, 0, 0, 0, 0, 0, 5 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  in += 27;
-  return in;
-}
-
-inline static const uint32_t* unpack28_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0xfffffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 28-bit bundles 0 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 28 | SafeLoad<uint32_t>(in + 1) << 4, SafeLoad<uint32_t>(in + 1) >> 24 | SafeLoad<uint32_t>(in + 2) << 8, SafeLoad<uint32_t>(in + 2) >> 20 | SafeLoad<uint32_t>(in + 3) << 12, SafeLoad<uint32_t>(in + 3) >> 16 | SafeLoad<uint32_t>(in + 4) << 16, SafeLoad<uint32_t>(in + 4) >> 12 | SafeLoad<uint32_t>(in + 5) << 20, SafeLoad<uint32_t>(in + 5) >> 8 | SafeLoad<uint32_t>(in + 6) << 24, SafeLoad<uint32_t>(in + 6) };
-  shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 4 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 28-bit bundles 8 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7) >> 28 | SafeLoad<uint32_t>(in + 8) << 4, SafeLoad<uint32_t>(in + 8) >> 24 | SafeLoad<uint32_t>(in + 9) << 8, SafeLoad<uint32_t>(in + 9) >> 20 | SafeLoad<uint32_t>(in + 10) << 12, SafeLoad<uint32_t>(in + 10) >> 16 | SafeLoad<uint32_t>(in + 11) << 16, SafeLoad<uint32_t>(in + 11) >> 12 | SafeLoad<uint32_t>(in + 12) << 20, SafeLoad<uint32_t>(in + 12) >> 8 | SafeLoad<uint32_t>(in + 13) << 24, SafeLoad<uint32_t>(in + 13) };
-  shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 4 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 28-bit bundles 16 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 14), SafeLoad<uint32_t>(in + 14) >> 28 | SafeLoad<uint32_t>(in + 15) << 4, SafeLoad<uint32_t>(in + 15) >> 24 | SafeLoad<uint32_t>(in + 16) << 8, SafeLoad<uint32_t>(in + 16) >> 20 | SafeLoad<uint32_t>(in + 17) << 12, SafeLoad<uint32_t>(in + 17) >> 16 | SafeLoad<uint32_t>(in + 18) << 16, SafeLoad<uint32_t>(in + 18) >> 12 | SafeLoad<uint32_t>(in + 19) << 20, SafeLoad<uint32_t>(in + 19) >> 8 | SafeLoad<uint32_t>(in + 20) << 24, SafeLoad<uint32_t>(in + 20) };
-  shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 4 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 28-bit bundles 24 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 21), SafeLoad<uint32_t>(in + 21) >> 28 | SafeLoad<uint32_t>(in + 22) << 4, SafeLoad<uint32_t>(in + 22) >> 24 | SafeLoad<uint32_t>(in + 23) << 8, SafeLoad<uint32_t>(in + 23) >> 20 | SafeLoad<uint32_t>(in + 24) << 12, SafeLoad<uint32_t>(in + 24) >> 16 | SafeLoad<uint32_t>(in + 25) << 16, SafeLoad<uint32_t>(in + 25) >> 12 | SafeLoad<uint32_t>(in + 26) << 20, SafeLoad<uint32_t>(in + 26) >> 8 | SafeLoad<uint32_t>(in + 27) << 24, SafeLoad<uint32_t>(in + 27) };
-  shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 4 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  in += 28;
-  return in;
-}
-
-inline static const uint32_t* unpack29_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x1fffffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 29-bit bundles 0 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 29 | SafeLoad<uint32_t>(in + 1) << 3, SafeLoad<uint32_t>(in + 1) >> 26 | SafeLoad<uint32_t>(in + 2) << 6, SafeLoad<uint32_t>(in + 2) >> 23 | SafeLoad<uint32_t>(in + 3) << 9, SafeLoad<uint32_t>(in + 3) >> 20 | SafeLoad<uint32_t>(in + 4) << 12, SafeLoad<uint32_t>(in + 4) >> 17 | SafeLoad<uint32_t>(in + 5) << 15, SafeLoad<uint32_t>(in + 5) >> 14 | SafeLoad<uint32_t>(in + 6) << 18, SafeLoad<uint32_t>(in + 6) >> 11 | SafeLoad<uint32_t>(in + 7) << 21 };
-  shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 29-bit bundles 8 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 7) >> 8 | SafeLoad<uint32_t>(in + 8) << 24, SafeLoad<uint32_t>(in + 8) >> 5 | SafeLoad<uint32_t>(in + 9) << 27, SafeLoad<uint32_t>(in + 9), SafeLoad<uint32_t>(in + 9) >> 31 | SafeLoad<uint32_t>(in + 10) << 1, SafeLoad<uint32_t>(in + 10) >> 28 | SafeLoad<uint32_t>(in + 11) << 4, SafeLoad<uint32_t>(in + 11) >> 25 | SafeLoad<uint32_t>(in + 12) << 7, SafeLoad<uint32_t>(in + 12) >> 22 | SafeLoad<uint32_t>(in + 13) << 10, SafeLoad<uint32_t>(in + 13) >> 19 | SafeLoad<uint32_t>(in + 14) << 13 };
-  shifts = simd_batch{ 0, 0, 2, 0, 0, 0, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 29-bit bundles 16 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 14) >> 16 | SafeLoad<uint32_t>(in + 15) << 16, SafeLoad<uint32_t>(in + 15) >> 13 | SafeLoad<uint32_t>(in + 16) << 19, SafeLoad<uint32_t>(in + 16) >> 10 | SafeLoad<uint32_t>(in + 17) << 22, SafeLoad<uint32_t>(in + 17) >> 7 | SafeLoad<uint32_t>(in + 18) << 25, SafeLoad<uint32_t>(in + 18) >> 4 | SafeLoad<uint32_t>(in + 19) << 28, SafeLoad<uint32_t>(in + 19), SafeLoad<uint32_t>(in + 19) >> 30 | SafeLoad<uint32_t>(in + 20) << 2, SafeLoad<uint32_t>(in + 20) >> 27 | SafeLoad<uint32_t>(in + 21) << 5 };
-  shifts = simd_batch{ 0, 0, 0, 0, 0, 1, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 29-bit bundles 24 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 21) >> 24 | SafeLoad<uint32_t>(in + 22) << 8, SafeLoad<uint32_t>(in + 22) >> 21 | SafeLoad<uint32_t>(in + 23) << 11, SafeLoad<uint32_t>(in + 23) >> 18 | SafeLoad<uint32_t>(in + 24) << 14, SafeLoad<uint32_t>(in + 24) >> 15 | SafeLoad<uint32_t>(in + 25) << 17, SafeLoad<uint32_t>(in + 25) >> 12 | SafeLoad<uint32_t>(in + 26) << 20, SafeLoad<uint32_t>(in + 26) >> 9 | SafeLoad<uint32_t>(in + 27) << 23, SafeLoad<uint32_t>(in + 27) >> 6 | SafeLoad<uint32_t>(in + 28) << 26, SafeLoad<uint32_t>(in + 28) };
-  shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 3 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  in += 29;
-  return in;
-}
-
-inline static const uint32_t* unpack30_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x3fffffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 30-bit bundles 0 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 30 | SafeLoad<uint32_t>(in + 1) << 2, SafeLoad<uint32_t>(in + 1) >> 28 | SafeLoad<uint32_t>(in + 2) << 4, SafeLoad<uint32_t>(in + 2) >> 26 | SafeLoad<uint32_t>(in + 3) << 6, SafeLoad<uint32_t>(in + 3) >> 24 | SafeLoad<uint32_t>(in + 4) << 8, SafeLoad<uint32_t>(in + 4) >> 22 | SafeLoad<uint32_t>(in + 5) << 10, SafeLoad<uint32_t>(in + 5) >> 20 | SafeLoad<uint32_t>(in + 6) << 12, SafeLoad<uint32_t>(in + 6) >> 18 | SafeLoad<uint32_t>(in + 7) << 14 };
-  shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 30-bit bundles 8 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 7) >> 16 | SafeLoad<uint32_t>(in + 8) << 16, SafeLoad<uint32_t>(in + 8) >> 14 | SafeLoad<uint32_t>(in + 9) << 18, SafeLoad<uint32_t>(in + 9) >> 12 | SafeLoad<uint32_t>(in + 10) << 20, SafeLoad<uint32_t>(in + 10) >> 10 | SafeLoad<uint32_t>(in + 11) << 22, SafeLoad<uint32_t>(in + 11) >> 8 | SafeLoad<uint32_t>(in + 12) << 24, SafeLoad<uint32_t>(in + 12) >> 6 | SafeLoad<uint32_t>(in + 13) << 26, SafeLoad<uint32_t>(in + 13) >> 4 | SafeLoad<uint32_t>(in + 14) << 28, SafeLoad<uint32_t>(in + 14) };
-  shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 2 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 30-bit bundles 16 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 15), SafeLoad<uint32_t>(in + 15) >> 30 | SafeLoad<uint32_t>(in + 16) << 2, SafeLoad<uint32_t>(in + 16) >> 28 | SafeLoad<uint32_t>(in + 17) << 4, SafeLoad<uint32_t>(in + 17) >> 26 | SafeLoad<uint32_t>(in + 18) << 6, SafeLoad<uint32_t>(in + 18) >> 24 | SafeLoad<uint32_t>(in + 19) << 8, SafeLoad<uint32_t>(in + 19) >> 22 | SafeLoad<uint32_t>(in + 20) << 10, SafeLoad<uint32_t>(in + 20) >> 20 | SafeLoad<uint32_t>(in + 21) << 12, SafeLoad<uint32_t>(in + 21) >> 18 | SafeLoad<uint32_t>(in + 22) << 14 };
-  shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 30-bit bundles 24 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 22) >> 16 | SafeLoad<uint32_t>(in + 23) << 16, SafeLoad<uint32_t>(in + 23) >> 14 | SafeLoad<uint32_t>(in + 24) << 18, SafeLoad<uint32_t>(in + 24) >> 12 | SafeLoad<uint32_t>(in + 25) << 20, SafeLoad<uint32_t>(in + 25) >> 10 | SafeLoad<uint32_t>(in + 26) << 22, SafeLoad<uint32_t>(in + 26) >> 8 | SafeLoad<uint32_t>(in + 27) << 24, SafeLoad<uint32_t>(in + 27) >> 6 | SafeLoad<uint32_t>(in + 28) << 26, SafeLoad<uint32_t>(in + 28) >> 4 | SafeLoad<uint32_t>(in + 29) << 28, SafeLoad<uint32_t>(in + 29) };
-  shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 2 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  in += 30;
-  return in;
-}
-
-inline static const uint32_t* unpack31_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x7fffffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 31-bit bundles 0 to 7
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 31 | SafeLoad<uint32_t>(in + 1) << 1, SafeLoad<uint32_t>(in + 1) >> 30 | SafeLoad<uint32_t>(in + 2) << 2, SafeLoad<uint32_t>(in + 2) >> 29 | SafeLoad<uint32_t>(in + 3) << 3, SafeLoad<uint32_t>(in + 3) >> 28 | SafeLoad<uint32_t>(in + 4) << 4, SafeLoad<uint32_t>(in + 4) >> 27 | SafeLoad<uint32_t>(in + 5) << 5, SafeLoad<uint32_t>(in + 5) >> 26 | SafeLoad<uint32_t>(in + 6) << 6, SafeLoad<uint32_t>(in + 6) >> 25 | SafeLoad<uint32_t>(in + 7) << 7 };
-  shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 31-bit bundles 8 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 7) >> 24 | SafeLoad<uint32_t>(in + 8) << 8, SafeLoad<uint32_t>(in + 8) >> 23 | SafeLoad<uint32_t>(in + 9) << 9, SafeLoad<uint32_t>(in + 9) >> 22 | SafeLoad<uint32_t>(in + 10) << 10, SafeLoad<uint32_t>(in + 10) >> 21 | SafeLoad<uint32_t>(in + 11) << 11, SafeLoad<uint32_t>(in + 11) >> 20 | SafeLoad<uint32_t>(in + 12) << 12, SafeLoad<uint32_t>(in + 12) >> 19 | SafeLoad<uint32_t>(in + 13) << 13, SafeLoad<uint32_t>(in + 13) >> 18 | SafeLoad<uint32_t>(in + 14) << 14, SafeLoad<uint32_t>(in + 14) >> 17 | SafeLoad<uint32_t>(in + 15) << 15 };
-  shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 31-bit bundles 16 to 23
-  words = simd_batch{ SafeLoad<uint32_t>(in + 15) >> 16 | SafeLoad<uint32_t>(in + 16) << 16, SafeLoad<uint32_t>(in + 16) >> 15 | SafeLoad<uint32_t>(in + 17) << 17, SafeLoad<uint32_t>(in + 17) >> 14 | SafeLoad<uint32_t>(in + 18) << 18, SafeLoad<uint32_t>(in + 18) >> 13 | SafeLoad<uint32_t>(in + 19) << 19, SafeLoad<uint32_t>(in + 19) >> 12 | SafeLoad<uint32_t>(in + 20) << 20, SafeLoad<uint32_t>(in + 20) >> 11 | SafeLoad<uint32_t>(in + 21) << 21, SafeLoad<uint32_t>(in + 21) >> 10 | SafeLoad<uint32_t>(in + 22) << 22, SafeLoad<uint32_t>(in + 22) >> 9 | SafeLoad<uint32_t>(in + 23) << 23 };
-  shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  // extract 31-bit bundles 24 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 23) >> 8 | SafeLoad<uint32_t>(in + 24) << 24, SafeLoad<uint32_t>(in + 24) >> 7 | SafeLoad<uint32_t>(in + 25) << 25, SafeLoad<uint32_t>(in + 25) >> 6 | SafeLoad<uint32_t>(in + 26) << 26, SafeLoad<uint32_t>(in + 26) >> 5 | SafeLoad<uint32_t>(in + 27) << 27, SafeLoad<uint32_t>(in + 27) >> 4 | SafeLoad<uint32_t>(in + 28) << 28, SafeLoad<uint32_t>(in + 28) >> 3 | SafeLoad<uint32_t>(in + 29) << 29, SafeLoad<uint32_t>(in + 29) >> 2 | SafeLoad<uint32_t>(in + 30) << 30, SafeLoad<uint32_t>(in + 30) };
-  shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 1 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 8;
-
-  in += 31;
-  return in;
-}
-
-inline static const uint32_t* unpack32_32(const uint32_t* in, uint32_t* out) {
-  memcpy(out, in, 32 * sizeof(*out));
-  in += 32;
-  out += 32;
-
-  return in;
-}
-
-};  // struct UnpackBits256
-
-}  // namespace
-}  // namespace internal
-}  // namespace arrow
-
+namespace arrow::internal {
+
+using ::arrow::util::SafeLoadAs;
+
+template<typename Uint, int BitWidth>
+struct Simd256UnpackerForWidth;
+
+template<>
+struct Simd256UnpackerForWidth<uint16_t, 1> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint16_t, 16>;
+  static constexpr int kValuesUnpacked = 16;
+
+  static const uint8_t* unpack(const uint8_t* in, uint16_t* out) {
+    constexpr uint16_t kMask = 0x1;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 1-bit bundles 0 to 15
+    words = simd_batch{
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+    };
+    shifts = simd_batch{ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    in += 1 * 2;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint16_t, 2> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint16_t, 16>;
+  static constexpr int kValuesUnpacked = 16;
+
+  static const uint8_t* unpack(const uint8_t* in, uint16_t* out) {
+    constexpr uint16_t kMask = 0x3;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 2-bit bundles 0 to 15
+    words = simd_batch{
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+    };
+    shifts = simd_batch{ 0, 2, 4, 6, 8, 10, 12, 14, 0, 2, 4, 6, 8, 10, 12, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    in += 2 * 2;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint16_t, 3> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint16_t, 16>;
+  static constexpr int kValuesUnpacked = 16;
+
+  static const uint8_t* unpack(const uint8_t* in, uint16_t* out) {
+    constexpr uint16_t kMask = 0x7;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 3-bit bundles 0 to 15
+    words = simd_batch{
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 0) >> 15 | SafeLoadAs<uint16_t>(in + 2 * 1) << 1),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 1) >> 14 | SafeLoadAs<uint16_t>(in + 2 * 2) << 2),
+      SafeLoadAs<uint16_t>(in + 2 * 2),
+      SafeLoadAs<uint16_t>(in + 2 * 2),
+      SafeLoadAs<uint16_t>(in + 2 * 2),
+      SafeLoadAs<uint16_t>(in + 2 * 2),
+      SafeLoadAs<uint16_t>(in + 2 * 2),
+    };
+    shifts = simd_batch{ 0, 3, 6, 9, 12, 0, 2, 5, 8, 11, 0, 1, 4, 7, 10, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    in += 3 * 2;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint16_t, 4> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint16_t, 16>;
+  static constexpr int kValuesUnpacked = 16;
+
+  static const uint8_t* unpack(const uint8_t* in, uint16_t* out) {
+    constexpr uint16_t kMask = 0xf;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 4-bit bundles 0 to 15
+    words = simd_batch{
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+      SafeLoadAs<uint16_t>(in + 2 * 2),
+      SafeLoadAs<uint16_t>(in + 2 * 2),
+      SafeLoadAs<uint16_t>(in + 2 * 2),
+      SafeLoadAs<uint16_t>(in + 2 * 2),
+      SafeLoadAs<uint16_t>(in + 2 * 3),
+      SafeLoadAs<uint16_t>(in + 2 * 3),
+      SafeLoadAs<uint16_t>(in + 2 * 3),
+      SafeLoadAs<uint16_t>(in + 2 * 3),
+    };
+    shifts = simd_batch{ 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    in += 4 * 2;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint16_t, 5> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint16_t, 16>;
+  static constexpr int kValuesUnpacked = 16;
+
+  static const uint8_t* unpack(const uint8_t* in, uint16_t* out) {
+    constexpr uint16_t kMask = 0x1f;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 5-bit bundles 0 to 15
+    words = simd_batch{
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 0) >> 15 | SafeLoadAs<uint16_t>(in + 2 * 1) << 1),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 1) >> 14 | SafeLoadAs<uint16_t>(in + 2 * 2) << 2),
+      SafeLoadAs<uint16_t>(in + 2 * 2),
+      SafeLoadAs<uint16_t>(in + 2 * 2),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 2) >> 13 | SafeLoadAs<uint16_t>(in + 2 * 3) << 3),
+      SafeLoadAs<uint16_t>(in + 2 * 3),
+      SafeLoadAs<uint16_t>(in + 2 * 3),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 3) >> 12 | SafeLoadAs<uint16_t>(in + 2 * 4) << 4),
+      SafeLoadAs<uint16_t>(in + 2 * 4),
+      SafeLoadAs<uint16_t>(in + 2 * 4),
+      SafeLoadAs<uint16_t>(in + 2 * 4),
+    };
+    shifts = simd_batch{ 0, 5, 10, 0, 4, 9, 0, 3, 8, 0, 2, 7, 0, 1, 6, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    in += 5 * 2;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint16_t, 6> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint16_t, 16>;
+  static constexpr int kValuesUnpacked = 16;
+
+  static const uint8_t* unpack(const uint8_t* in, uint16_t* out) {
+    constexpr uint16_t kMask = 0x3f;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 6-bit bundles 0 to 15
+    words = simd_batch{
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 0) >> 12 | SafeLoadAs<uint16_t>(in + 2 * 1) << 4),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 1) >> 14 | SafeLoadAs<uint16_t>(in + 2 * 2) << 2),
+      SafeLoadAs<uint16_t>(in + 2 * 2),
+      SafeLoadAs<uint16_t>(in + 2 * 2),
+      SafeLoadAs<uint16_t>(in + 2 * 3),
+      SafeLoadAs<uint16_t>(in + 2 * 3),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 3) >> 12 | SafeLoadAs<uint16_t>(in + 2 * 4) << 4),
+      SafeLoadAs<uint16_t>(in + 2 * 4),
+      SafeLoadAs<uint16_t>(in + 2 * 4),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 4) >> 14 | SafeLoadAs<uint16_t>(in + 2 * 5) << 2),
+      SafeLoadAs<uint16_t>(in + 2 * 5),
+      SafeLoadAs<uint16_t>(in + 2 * 5),
+    };
+    shifts = simd_batch{ 0, 6, 0, 2, 8, 0, 4, 10, 0, 6, 0, 2, 8, 0, 4, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    in += 6 * 2;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint16_t, 7> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint16_t, 16>;
+  static constexpr int kValuesUnpacked = 16;
+
+  static const uint8_t* unpack(const uint8_t* in, uint16_t* out) {
+    constexpr uint16_t kMask = 0x7f;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 7-bit bundles 0 to 15
+    words = simd_batch{
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 0) >> 14 | SafeLoadAs<uint16_t>(in + 2 * 1) << 2),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 1) >> 12 | SafeLoadAs<uint16_t>(in + 2 * 2) << 4),
+      SafeLoadAs<uint16_t>(in + 2 * 2),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 2) >> 10 | SafeLoadAs<uint16_t>(in + 2 * 3) << 6),
+      SafeLoadAs<uint16_t>(in + 2 * 3),
+      SafeLoadAs<uint16_t>(in + 2 * 3),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 3) >> 15 | SafeLoadAs<uint16_t>(in + 2 * 4) << 1),
+      SafeLoadAs<uint16_t>(in + 2 * 4),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 4) >> 13 | SafeLoadAs<uint16_t>(in + 2 * 5) << 3),
+      SafeLoadAs<uint16_t>(in + 2 * 5),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 5) >> 11 | SafeLoadAs<uint16_t>(in + 2 * 6) << 5),
+      SafeLoadAs<uint16_t>(in + 2 * 6),
+      SafeLoadAs<uint16_t>(in + 2 * 6),
+    };
+    shifts = simd_batch{ 0, 7, 0, 5, 0, 3, 0, 1, 8, 0, 6, 0, 4, 0, 2, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    in += 7 * 2;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint16_t, 8> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint16_t, 16>;
+  static constexpr int kValuesUnpacked = 16;
+
+  static const uint8_t* unpack(const uint8_t* in, uint16_t* out) {
+    constexpr uint16_t kMask = 0xff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 8-bit bundles 0 to 15
+    words = simd_batch{
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+      SafeLoadAs<uint16_t>(in + 2 * 2),
+      SafeLoadAs<uint16_t>(in + 2 * 2),
+      SafeLoadAs<uint16_t>(in + 2 * 3),
+      SafeLoadAs<uint16_t>(in + 2 * 3),
+      SafeLoadAs<uint16_t>(in + 2 * 4),
+      SafeLoadAs<uint16_t>(in + 2 * 4),
+      SafeLoadAs<uint16_t>(in + 2 * 5),
+      SafeLoadAs<uint16_t>(in + 2 * 5),
+      SafeLoadAs<uint16_t>(in + 2 * 6),
+      SafeLoadAs<uint16_t>(in + 2 * 6),
+      SafeLoadAs<uint16_t>(in + 2 * 7),
+      SafeLoadAs<uint16_t>(in + 2 * 7),
+    };
+    shifts = simd_batch{ 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    in += 8 * 2;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint16_t, 9> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint16_t, 16>;
+  static constexpr int kValuesUnpacked = 16;
+
+  static const uint8_t* unpack(const uint8_t* in, uint16_t* out) {
+    constexpr uint16_t kMask = 0x1ff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 9-bit bundles 0 to 15
+    words = simd_batch{
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 0) >> 9 | SafeLoadAs<uint16_t>(in + 2 * 1) << 7),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 1) >> 11 | SafeLoadAs<uint16_t>(in + 2 * 2) << 5),
+      SafeLoadAs<uint16_t>(in + 2 * 2),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 2) >> 13 | SafeLoadAs<uint16_t>(in + 2 * 3) << 3),
+      SafeLoadAs<uint16_t>(in + 2 * 3),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 3) >> 15 | SafeLoadAs<uint16_t>(in + 2 * 4) << 1),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 4) >> 8 | SafeLoadAs<uint16_t>(in + 2 * 5) << 8),
+      SafeLoadAs<uint16_t>(in + 2 * 5),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 5) >> 10 | SafeLoadAs<uint16_t>(in + 2 * 6) << 6),
+      SafeLoadAs<uint16_t>(in + 2 * 6),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 6) >> 12 | SafeLoadAs<uint16_t>(in + 2 * 7) << 4),
+      SafeLoadAs<uint16_t>(in + 2 * 7),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 7) >> 14 | SafeLoadAs<uint16_t>(in + 2 * 8) << 2),
+      SafeLoadAs<uint16_t>(in + 2 * 8),
+    };
+    shifts = simd_batch{ 0, 0, 2, 0, 4, 0, 6, 0, 0, 1, 0, 3, 0, 5, 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    in += 9 * 2;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint16_t, 10> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint16_t, 16>;
+  static constexpr int kValuesUnpacked = 16;
+
+  static const uint8_t* unpack(const uint8_t* in, uint16_t* out) {
+    constexpr uint16_t kMask = 0x3ff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 10-bit bundles 0 to 15
+    words = simd_batch{
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 0) >> 10 | SafeLoadAs<uint16_t>(in + 2 * 1) << 6),
+      SafeLoadAs<uint16_t>(in + 2 * 1),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 1) >> 14 | SafeLoadAs<uint16_t>(in + 2 * 2) << 2),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 2) >> 8 | SafeLoadAs<uint16_t>(in + 2 * 3) << 8),
+      SafeLoadAs<uint16_t>(in + 2 * 3),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 3) >> 12 | SafeLoadAs<uint16_t>(in + 2 * 4) << 4),
+      SafeLoadAs<uint16_t>(in + 2 * 4),
+      SafeLoadAs<uint16_t>(in + 2 * 5),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 5) >> 10 | SafeLoadAs<uint16_t>(in + 2 * 6) << 6),
+      SafeLoadAs<uint16_t>(in + 2 * 6),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 6) >> 14 | SafeLoadAs<uint16_t>(in + 2 * 7) << 2),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 7) >> 8 | SafeLoadAs<uint16_t>(in + 2 * 8) << 8),
+      SafeLoadAs<uint16_t>(in + 2 * 8),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 8) >> 12 | SafeLoadAs<uint16_t>(in + 2 * 9) << 4),
+      SafeLoadAs<uint16_t>(in + 2 * 9),
+    };
+    shifts = simd_batch{ 0, 0, 4, 0, 0, 2, 0, 6, 0, 0, 4, 0, 0, 2, 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    in += 10 * 2;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint16_t, 11> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint16_t, 16>;
+  static constexpr int kValuesUnpacked = 16;
+
+  static const uint8_t* unpack(const uint8_t* in, uint16_t* out) {
+    constexpr uint16_t kMask = 0x7ff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 11-bit bundles 0 to 15
+    words = simd_batch{
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 0) >> 11 | SafeLoadAs<uint16_t>(in + 2 * 1) << 5),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 1) >> 6 | SafeLoadAs<uint16_t>(in + 2 * 2) << 10),
+      SafeLoadAs<uint16_t>(in + 2 * 2),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 2) >> 12 | SafeLoadAs<uint16_t>(in + 2 * 3) << 4),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 3) >> 7 | SafeLoadAs<uint16_t>(in + 2 * 4) << 9),
+      SafeLoadAs<uint16_t>(in + 2 * 4),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 4) >> 13 | SafeLoadAs<uint16_t>(in + 2 * 5) << 3),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 5) >> 8 | SafeLoadAs<uint16_t>(in + 2 * 6) << 8),
+      SafeLoadAs<uint16_t>(in + 2 * 6),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 6) >> 14 | SafeLoadAs<uint16_t>(in + 2 * 7) << 2),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 7) >> 9 | SafeLoadAs<uint16_t>(in + 2 * 8) << 7),
+      SafeLoadAs<uint16_t>(in + 2 * 8),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 8) >> 15 | SafeLoadAs<uint16_t>(in + 2 * 9) << 1),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 9) >> 10 | SafeLoadAs<uint16_t>(in + 2 * 10) << 6),
+      SafeLoadAs<uint16_t>(in + 2 * 10),
+    };
+    shifts = simd_batch{ 0, 0, 0, 1, 0, 0, 2, 0, 0, 3, 0, 0, 4, 0, 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    in += 11 * 2;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint16_t, 12> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint16_t, 16>;
+  static constexpr int kValuesUnpacked = 16;
+
+  static const uint8_t* unpack(const uint8_t* in, uint16_t* out) {
+    constexpr uint16_t kMask = 0xfff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 12-bit bundles 0 to 15
+    words = simd_batch{
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 0) >> 12 | SafeLoadAs<uint16_t>(in + 2 * 1) << 4),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 1) >> 8 | SafeLoadAs<uint16_t>(in + 2 * 2) << 8),
+      SafeLoadAs<uint16_t>(in + 2 * 2),
+      SafeLoadAs<uint16_t>(in + 2 * 3),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 3) >> 12 | SafeLoadAs<uint16_t>(in + 2 * 4) << 4),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 4) >> 8 | SafeLoadAs<uint16_t>(in + 2 * 5) << 8),
+      SafeLoadAs<uint16_t>(in + 2 * 5),
+      SafeLoadAs<uint16_t>(in + 2 * 6),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 6) >> 12 | SafeLoadAs<uint16_t>(in + 2 * 7) << 4),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 7) >> 8 | SafeLoadAs<uint16_t>(in + 2 * 8) << 8),
+      SafeLoadAs<uint16_t>(in + 2 * 8),
+      SafeLoadAs<uint16_t>(in + 2 * 9),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 9) >> 12 | SafeLoadAs<uint16_t>(in + 2 * 10) << 4),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 10) >> 8 | SafeLoadAs<uint16_t>(in + 2 * 11) << 8),
+      SafeLoadAs<uint16_t>(in + 2 * 11),
+    };
+    shifts = simd_batch{ 0, 0, 0, 4, 0, 0, 0, 4, 0, 0, 0, 4, 0, 0, 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    in += 12 * 2;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint16_t, 13> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint16_t, 16>;
+  static constexpr int kValuesUnpacked = 16;
+
+  static const uint8_t* unpack(const uint8_t* in, uint16_t* out) {
+    constexpr uint16_t kMask = 0x1fff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 13-bit bundles 0 to 15
+    words = simd_batch{
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 0) >> 13 | SafeLoadAs<uint16_t>(in + 2 * 1) << 3),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 1) >> 10 | SafeLoadAs<uint16_t>(in + 2 * 2) << 6),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 2) >> 7 | SafeLoadAs<uint16_t>(in + 2 * 3) << 9),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 3) >> 4 | SafeLoadAs<uint16_t>(in + 2 * 4) << 12),
+      SafeLoadAs<uint16_t>(in + 2 * 4),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 4) >> 14 | SafeLoadAs<uint16_t>(in + 2 * 5) << 2),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 5) >> 11 | SafeLoadAs<uint16_t>(in + 2 * 6) << 5),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 6) >> 8 | SafeLoadAs<uint16_t>(in + 2 * 7) << 8),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 7) >> 5 | SafeLoadAs<uint16_t>(in + 2 * 8) << 11),
+      SafeLoadAs<uint16_t>(in + 2 * 8),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 8) >> 15 | SafeLoadAs<uint16_t>(in + 2 * 9) << 1),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 9) >> 12 | SafeLoadAs<uint16_t>(in + 2 * 10) << 4),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 10) >> 9 | SafeLoadAs<uint16_t>(in + 2 * 11) << 7),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 11) >> 6 | SafeLoadAs<uint16_t>(in + 2 * 12) << 10),
+      SafeLoadAs<uint16_t>(in + 2 * 12),
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 2, 0, 0, 0, 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    in += 13 * 2;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint16_t, 14> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint16_t, 16>;
+  static constexpr int kValuesUnpacked = 16;
+
+  static const uint8_t* unpack(const uint8_t* in, uint16_t* out) {
+    constexpr uint16_t kMask = 0x3fff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 14-bit bundles 0 to 15
+    words = simd_batch{
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 0) >> 14 | SafeLoadAs<uint16_t>(in + 2 * 1) << 2),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 1) >> 12 | SafeLoadAs<uint16_t>(in + 2 * 2) << 4),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 2) >> 10 | SafeLoadAs<uint16_t>(in + 2 * 3) << 6),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 3) >> 8 | SafeLoadAs<uint16_t>(in + 2 * 4) << 8),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 4) >> 6 | SafeLoadAs<uint16_t>(in + 2 * 5) << 10),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 5) >> 4 | SafeLoadAs<uint16_t>(in + 2 * 6) << 12),
+      SafeLoadAs<uint16_t>(in + 2 * 6),
+      SafeLoadAs<uint16_t>(in + 2 * 7),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 7) >> 14 | SafeLoadAs<uint16_t>(in + 2 * 8) << 2),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 8) >> 12 | SafeLoadAs<uint16_t>(in + 2 * 9) << 4),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 9) >> 10 | SafeLoadAs<uint16_t>(in + 2 * 10) << 6),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 10) >> 8 | SafeLoadAs<uint16_t>(in + 2 * 11) << 8),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 11) >> 6 | SafeLoadAs<uint16_t>(in + 2 * 12) << 10),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 12) >> 4 | SafeLoadAs<uint16_t>(in + 2 * 13) << 12),
+      SafeLoadAs<uint16_t>(in + 2 * 13),
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    in += 14 * 2;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint16_t, 15> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint16_t, 16>;
+  static constexpr int kValuesUnpacked = 16;
+
+  static const uint8_t* unpack(const uint8_t* in, uint16_t* out) {
+    constexpr uint16_t kMask = 0x7fff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 15-bit bundles 0 to 15
+    words = simd_batch{
+      SafeLoadAs<uint16_t>(in + 2 * 0),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 0) >> 15 | SafeLoadAs<uint16_t>(in + 2 * 1) << 1),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 1) >> 14 | SafeLoadAs<uint16_t>(in + 2 * 2) << 2),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 2) >> 13 | SafeLoadAs<uint16_t>(in + 2 * 3) << 3),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 3) >> 12 | SafeLoadAs<uint16_t>(in + 2 * 4) << 4),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 4) >> 11 | SafeLoadAs<uint16_t>(in + 2 * 5) << 5),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 5) >> 10 | SafeLoadAs<uint16_t>(in + 2 * 6) << 6),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 6) >> 9 | SafeLoadAs<uint16_t>(in + 2 * 7) << 7),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 7) >> 8 | SafeLoadAs<uint16_t>(in + 2 * 8) << 8),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 8) >> 7 | SafeLoadAs<uint16_t>(in + 2 * 9) << 9),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 9) >> 6 | SafeLoadAs<uint16_t>(in + 2 * 10) << 10),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 10) >> 5 | SafeLoadAs<uint16_t>(in + 2 * 11) << 11),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 11) >> 4 | SafeLoadAs<uint16_t>(in + 2 * 12) << 12),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 12) >> 3 | SafeLoadAs<uint16_t>(in + 2 * 13) << 13),
+      static_cast<uint16_t>(SafeLoadAs<uint16_t>(in + 2 * 13) >> 2 | SafeLoadAs<uint16_t>(in + 2 * 14) << 14),
+      SafeLoadAs<uint16_t>(in + 2 * 14),
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    in += 15 * 2;
+    return in;
+  }
+};
+
+
+template<>
+struct Simd256UnpackerForWidth<uint32_t, 1> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 8>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x1;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 1-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+    };
+    shifts = simd_batch{ 0, 1, 2, 3, 4, 5, 6, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 1-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+    };
+    shifts = simd_batch{ 8, 9, 10, 11, 12, 13, 14, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 1-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+    };
+    shifts = simd_batch{ 16, 17, 18, 19, 20, 21, 22, 23 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 1-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+    };
+    shifts = simd_batch{ 24, 25, 26, 27, 28, 29, 30, 31 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 1 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint32_t, 2> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 8>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x3;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 2-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+    };
+    shifts = simd_batch{ 0, 2, 4, 6, 8, 10, 12, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 2-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+    };
+    shifts = simd_batch{ 16, 18, 20, 22, 24, 26, 28, 30 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 2-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+    };
+    shifts = simd_batch{ 0, 2, 4, 6, 8, 10, 12, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 2-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+    };
+    shifts = simd_batch{ 16, 18, 20, 22, 24, 26, 28, 30 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 2 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint32_t, 3> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 8>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x7;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 3-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+    };
+    shifts = simd_batch{ 0, 3, 6, 9, 12, 15, 18, 21 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 3-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 1) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+    };
+    shifts = simd_batch{ 24, 27, 0, 1, 4, 7, 10, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 3-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 31 | SafeLoadAs<uint32_t>(in + 4 * 2) << 1,
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+    };
+    shifts = simd_batch{ 16, 19, 22, 25, 28, 0, 2, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 3-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+    };
+    shifts = simd_batch{ 8, 11, 14, 17, 20, 23, 26, 29 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 3 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint32_t, 4> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 8>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0xf;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 4-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+    };
+    shifts = simd_batch{ 0, 4, 8, 12, 16, 20, 24, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 4-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+    };
+    shifts = simd_batch{ 0, 4, 8, 12, 16, 20, 24, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 4-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+    };
+    shifts = simd_batch{ 0, 4, 8, 12, 16, 20, 24, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 4-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+    };
+    shifts = simd_batch{ 0, 4, 8, 12, 16, 20, 24, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 4 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint32_t, 5> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 8>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x1f;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 5-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 1) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+    };
+    shifts = simd_batch{ 0, 5, 10, 15, 20, 25, 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 5-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 2) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+    };
+    shifts = simd_batch{ 8, 13, 18, 23, 0, 1, 6, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 5-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 31 | SafeLoadAs<uint32_t>(in + 4 * 3) << 1,
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+    };
+    shifts = simd_batch{ 16, 21, 26, 0, 4, 9, 14, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 5-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 29 | SafeLoadAs<uint32_t>(in + 4 * 4) << 3,
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+    };
+    shifts = simd_batch{ 24, 0, 2, 7, 12, 17, 22, 27 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 5 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint32_t, 6> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 8>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x3f;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 6-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 1) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+    };
+    shifts = simd_batch{ 0, 6, 12, 18, 24, 0, 4, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 6-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 2) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+    };
+    shifts = simd_batch{ 16, 22, 0, 2, 8, 14, 20, 26 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 6-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 4) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+    };
+    shifts = simd_batch{ 0, 6, 12, 18, 24, 0, 4, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 6-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 5) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+    };
+    shifts = simd_batch{ 16, 22, 0, 2, 8, 14, 20, 26 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 6 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint32_t, 7> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 8>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x7f;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 7-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 1) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+    };
+    shifts = simd_batch{ 0, 7, 14, 21, 0, 3, 10, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 7-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 31 | SafeLoadAs<uint32_t>(in + 4 * 2) << 1,
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 27 | SafeLoadAs<uint32_t>(in + 4 * 3) << 5,
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+    };
+    shifts = simd_batch{ 24, 0, 6, 13, 20, 0, 2, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 7-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 4) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 5) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+    };
+    shifts = simd_batch{ 16, 23, 0, 5, 12, 19, 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 7-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 29 | SafeLoadAs<uint32_t>(in + 4 * 6) << 3,
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+    };
+    shifts = simd_batch{ 8, 15, 22, 0, 4, 11, 18, 25 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 7 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint32_t, 8> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 8>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0xff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 8-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+    };
+    shifts = simd_batch{ 0, 8, 16, 24, 0, 8, 16, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 8-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+    };
+    shifts = simd_batch{ 0, 8, 16, 24, 0, 8, 16, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 8-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+    };
+    shifts = simd_batch{ 0, 8, 16, 24, 0, 8, 16, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 8-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+    };
+    shifts = simd_batch{ 0, 8, 16, 24, 0, 8, 16, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 8 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint32_t, 9> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 8>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x1ff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 9-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 27 | SafeLoadAs<uint32_t>(in + 4 * 1) << 5,
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 31 | SafeLoadAs<uint32_t>(in + 4 * 2) << 1,
+    };
+    shifts = simd_batch{ 0, 9, 18, 0, 4, 13, 22, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 9-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 3) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 4) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+    };
+    shifts = simd_batch{ 8, 17, 0, 3, 12, 21, 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 9-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 25 | SafeLoadAs<uint32_t>(in + 4 * 5) << 7,
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 29 | SafeLoadAs<uint32_t>(in + 4 * 6) << 3,
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+    };
+    shifts = simd_batch{ 16, 0, 2, 11, 20, 0, 6, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 9-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 7) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 8) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+    };
+    shifts = simd_batch{ 0, 1, 10, 19, 0, 5, 14, 23 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 9 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint32_t, 10> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 8>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x3ff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 10-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 1) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 2) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+    };
+    shifts = simd_batch{ 0, 10, 20, 0, 8, 18, 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 10-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 3) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 4) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+    };
+    shifts = simd_batch{ 16, 0, 4, 14, 0, 2, 12, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 10-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 6) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 7) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+    };
+    shifts = simd_batch{ 0, 10, 20, 0, 8, 18, 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 10-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 8) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 9) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+    };
+    shifts = simd_batch{ 16, 0, 4, 14, 0, 2, 12, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 10 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint32_t, 11> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 8>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x7ff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 11-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 1) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 23 | SafeLoadAs<uint32_t>(in + 4 * 2) << 9,
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+    };
+    shifts = simd_batch{ 0, 11, 0, 1, 12, 0, 2, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 11-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 3) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 25 | SafeLoadAs<uint32_t>(in + 4 * 4) << 7,
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 5) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+    };
+    shifts = simd_batch{ 0, 3, 14, 0, 4, 15, 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 11-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 27 | SafeLoadAs<uint32_t>(in + 4 * 6) << 5,
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 7) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 29 | SafeLoadAs<uint32_t>(in + 4 * 8) << 3,
+    };
+    shifts = simd_batch{ 16, 0, 6, 17, 0, 7, 18, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 11-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 9) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 31 | SafeLoadAs<uint32_t>(in + 4 * 10) << 1,
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+    };
+    shifts = simd_batch{ 8, 19, 0, 9, 20, 0, 10, 21 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 11 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint32_t, 12> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 8>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0xfff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 12-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 1) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 2) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+    };
+    shifts = simd_batch{ 0, 12, 0, 4, 16, 0, 8, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 12-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 4) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 5) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+    };
+    shifts = simd_batch{ 0, 12, 0, 4, 16, 0, 8, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 12-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 7) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 8) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+    };
+    shifts = simd_batch{ 0, 12, 0, 4, 16, 0, 8, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 12-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 10) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 11) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 11),
+      SafeLoadAs<uint32_t>(in + 4 * 11),
+    };
+    shifts = simd_batch{ 0, 12, 0, 4, 16, 0, 8, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 12 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint32_t, 13> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 8>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x1fff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 13-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 1) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 2) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 27 | SafeLoadAs<uint32_t>(in + 4 * 3) << 5,
+    };
+    shifts = simd_batch{ 0, 13, 0, 7, 0, 1, 14, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 13-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 21 | SafeLoadAs<uint32_t>(in + 4 * 4) << 11,
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 5) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 6) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+    };
+    shifts = simd_batch{ 8, 0, 2, 15, 0, 9, 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 13-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 29 | SafeLoadAs<uint32_t>(in + 4 * 7) << 3,
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 23 | SafeLoadAs<uint32_t>(in + 4 * 8) << 9,
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 9) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+    };
+    shifts = simd_batch{ 16, 0, 10, 0, 4, 17, 0, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 13-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 10) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 31 | SafeLoadAs<uint32_t>(in + 4 * 11) << 1,
+      SafeLoadAs<uint32_t>(in + 4 * 11),
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 25 | SafeLoadAs<uint32_t>(in + 4 * 12) << 7,
+      SafeLoadAs<uint32_t>(in + 4 * 12),
+      SafeLoadAs<uint32_t>(in + 4 * 12),
+    };
+    shifts = simd_batch{ 0, 5, 18, 0, 12, 0, 6, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 13 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint32_t, 14> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 8>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x3fff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 14-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 1) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 2) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 3) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+    };
+    shifts = simd_batch{ 0, 14, 0, 10, 0, 6, 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 14-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 4) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 5) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 6) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+    };
+    shifts = simd_batch{ 16, 0, 12, 0, 8, 0, 4, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 14-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 8) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 9) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 10) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+    };
+    shifts = simd_batch{ 0, 14, 0, 10, 0, 6, 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 14-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 11) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 11),
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 12) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 12),
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 13) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 13),
+      SafeLoadAs<uint32_t>(in + 4 * 13),
+    };
+    shifts = simd_batch{ 16, 0, 12, 0, 8, 0, 4, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 14 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint32_t, 15> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 8>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x7fff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 15-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 1) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 2) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 3) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+    };
+    shifts = simd_batch{ 0, 15, 0, 13, 0, 11, 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 15-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 4) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 5) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 6) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 7) << 14,
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+    };
+    shifts = simd_batch{ 0, 7, 0, 5, 0, 3, 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 15-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 31 | SafeLoadAs<uint32_t>(in + 4 * 8) << 1,
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 29 | SafeLoadAs<uint32_t>(in + 4 * 9) << 3,
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 27 | SafeLoadAs<uint32_t>(in + 4 * 10) << 5,
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 25 | SafeLoadAs<uint32_t>(in + 4 * 11) << 7,
+    };
+    shifts = simd_batch{ 16, 0, 14, 0, 12, 0, 10, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 15-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 11),
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 23 | SafeLoadAs<uint32_t>(in + 4 * 12) << 9,
+      SafeLoadAs<uint32_t>(in + 4 * 12),
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 21 | SafeLoadAs<uint32_t>(in + 4 * 13) << 11,
+      SafeLoadAs<uint32_t>(in + 4 * 13),
+      SafeLoadAs<uint32_t>(in + 4 * 13) >> 19 | SafeLoadAs<uint32_t>(in + 4 * 14) << 13,
+      SafeLoadAs<uint32_t>(in + 4 * 14),
+      SafeLoadAs<uint32_t>(in + 4 * 14),
+    };
+    shifts = simd_batch{ 8, 0, 6, 0, 4, 0, 2, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 15 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint32_t, 16> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 8>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0xffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 16-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+    };
+    shifts = simd_batch{ 0, 16, 0, 16, 0, 16, 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 16-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+    };
+    shifts = simd_batch{ 0, 16, 0, 16, 0, 16, 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 16-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+      SafeLoadAs<uint32_t>(in + 4 * 11),
+      SafeLoadAs<uint32_t>(in + 4 * 11),
+    };
+    shifts = simd_batch{ 0, 16, 0, 16, 0, 16, 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 16-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 12),
+      SafeLoadAs<uint32_t>(in + 4 * 12),
+      SafeLoadAs<uint32_t>(in + 4 * 13),
+      SafeLoadAs<uint32_t>(in + 4 * 13),
+      SafeLoadAs<uint32_t>(in + 4 * 14),
+      SafeLoadAs<uint32_t>(in + 4 * 14),
+      SafeLoadAs<uint32_t>(in + 4 * 15),
+      SafeLoadAs<uint32_t>(in + 4 * 15),
+    };
+    shifts = simd_batch{ 0, 16, 0, 16, 0, 16, 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 16 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint32_t, 17> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 8>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x1ffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 17-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 17 | SafeLoadAs<uint32_t>(in + 4 * 1) << 15,
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 19 | SafeLoadAs<uint32_t>(in + 4 * 2) << 13,
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 21 | SafeLoadAs<uint32_t>(in + 4 * 3) << 11,
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 23 | SafeLoadAs<uint32_t>(in + 4 * 4) << 9,
+    };
+    shifts = simd_batch{ 0, 0, 2, 0, 4, 0, 6, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 17-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 25 | SafeLoadAs<uint32_t>(in + 4 * 5) << 7,
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 27 | SafeLoadAs<uint32_t>(in + 4 * 6) << 5,
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 29 | SafeLoadAs<uint32_t>(in + 4 * 7) << 3,
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 31 | SafeLoadAs<uint32_t>(in + 4 * 8) << 1,
+    };
+    shifts = simd_batch{ 8, 0, 10, 0, 12, 0, 14, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 17-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 9) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 10) << 14,
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 11) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 11),
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 12) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 12),
+    };
+    shifts = simd_batch{ 0, 1, 0, 3, 0, 5, 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 17-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 13) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 13),
+      SafeLoadAs<uint32_t>(in + 4 * 13) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 14) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 14),
+      SafeLoadAs<uint32_t>(in + 4 * 14) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 15) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 15),
+      SafeLoadAs<uint32_t>(in + 4 * 15) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 16) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 16),
+    };
+    shifts = simd_batch{ 0, 9, 0, 11, 0, 13, 0, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 17 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint32_t, 18> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 8>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x3ffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 18-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 1) << 14,
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 2) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 3) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 4) << 2,
+    };
+    shifts = simd_batch{ 0, 0, 4, 0, 8, 0, 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 18-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 5) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 6) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 7) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 8) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+    };
+    shifts = simd_batch{ 0, 2, 0, 6, 0, 10, 0, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 18-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 10) << 14,
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 11) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 11),
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 12) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 12),
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 13) << 2,
+    };
+    shifts = simd_batch{ 0, 0, 4, 0, 8, 0, 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 18-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 13) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 14) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 14),
+      SafeLoadAs<uint32_t>(in + 4 * 14) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 15) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 15),
+      SafeLoadAs<uint32_t>(in + 4 * 15) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 16) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 16),
+      SafeLoadAs<uint32_t>(in + 4 * 16) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 17) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 17),
+    };
+    shifts = simd_batch{ 0, 2, 0, 6, 0, 10, 0, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 18 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint32_t, 19> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 8>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x7ffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 19-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 19 | SafeLoadAs<uint32_t>(in + 4 * 1) << 13,
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 25 | SafeLoadAs<uint32_t>(in + 4 * 2) << 7,
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 31 | SafeLoadAs<uint32_t>(in + 4 * 3) << 1,
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 4) << 14,
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+    };
+    shifts = simd_batch{ 0, 0, 6, 0, 12, 0, 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 19-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 5) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 6) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 17 | SafeLoadAs<uint32_t>(in + 4 * 7) << 15,
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 23 | SafeLoadAs<uint32_t>(in + 4 * 8) << 9,
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 29 | SafeLoadAs<uint32_t>(in + 4 * 9) << 3,
+    };
+    shifts = simd_batch{ 0, 11, 0, 0, 4, 0, 10, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 19-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 10) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 11) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 11),
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 12) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 15 | SafeLoadAs<uint32_t>(in + 4 * 13) << 17,
+      SafeLoadAs<uint32_t>(in + 4 * 13),
+      SafeLoadAs<uint32_t>(in + 4 * 13) >> 21 | SafeLoadAs<uint32_t>(in + 4 * 14) << 11,
+    };
+    shifts = simd_batch{ 0, 3, 0, 9, 0, 0, 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 19-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 14),
+      SafeLoadAs<uint32_t>(in + 4 * 14) >> 27 | SafeLoadAs<uint32_t>(in + 4 * 15) << 5,
+      SafeLoadAs<uint32_t>(in + 4 * 15) >> 14 | SafeLoadAs<uint32_t>(in + 4 * 16) << 18,
+      SafeLoadAs<uint32_t>(in + 4 * 16),
+      SafeLoadAs<uint32_t>(in + 4 * 16) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 17) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 17),
+      SafeLoadAs<uint32_t>(in + 4 * 17) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 18) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 18),
+    };
+    shifts = simd_batch{ 8, 0, 0, 1, 0, 7, 0, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 19 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint32_t, 20> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 8>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0xfffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 20-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 1) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 2) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 3) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 4) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+    };
+    shifts = simd_batch{ 0, 0, 8, 0, 0, 4, 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 20-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 6) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 7) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 8) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 9) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+    };
+    shifts = simd_batch{ 0, 0, 8, 0, 0, 4, 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 20-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 11) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 11),
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 12) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 13) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 13),
+      SafeLoadAs<uint32_t>(in + 4 * 13) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 14) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 14),
+    };
+    shifts = simd_batch{ 0, 0, 8, 0, 0, 4, 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 20-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 15),
+      SafeLoadAs<uint32_t>(in + 4 * 15) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 16) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 16),
+      SafeLoadAs<uint32_t>(in + 4 * 16) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 17) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 17) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 18) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 18),
+      SafeLoadAs<uint32_t>(in + 4 * 18) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 19) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 19),
+    };
+    shifts = simd_batch{ 0, 0, 8, 0, 0, 4, 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 20 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint32_t, 21> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 8>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x1fffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 21-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 21 | SafeLoadAs<uint32_t>(in + 4 * 1) << 11,
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 31 | SafeLoadAs<uint32_t>(in + 4 * 2) << 1,
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 3) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 4) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 19 | SafeLoadAs<uint32_t>(in + 4 * 5) << 13,
+    };
+    shifts = simd_batch{ 0, 0, 10, 0, 0, 9, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 21-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 29 | SafeLoadAs<uint32_t>(in + 4 * 6) << 3,
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 7) << 14,
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 8) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 17 | SafeLoadAs<uint32_t>(in + 4 * 9) << 15,
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 27 | SafeLoadAs<uint32_t>(in + 4 * 10) << 5,
+    };
+    shifts = simd_batch{ 8, 0, 0, 7, 0, 0, 6, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 21-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 11) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 11),
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 12) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 15 | SafeLoadAs<uint32_t>(in + 4 * 13) << 17,
+      SafeLoadAs<uint32_t>(in + 4 * 13),
+      SafeLoadAs<uint32_t>(in + 4 * 13) >> 25 | SafeLoadAs<uint32_t>(in + 4 * 14) << 7,
+      SafeLoadAs<uint32_t>(in + 4 * 14) >> 14 | SafeLoadAs<uint32_t>(in + 4 * 15) << 18,
+      SafeLoadAs<uint32_t>(in + 4 * 15),
+    };
+    shifts = simd_batch{ 0, 5, 0, 0, 4, 0, 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 21-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 15) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 16) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 16) >> 13 | SafeLoadAs<uint32_t>(in + 4 * 17) << 19,
+      SafeLoadAs<uint32_t>(in + 4 * 17),
+      SafeLoadAs<uint32_t>(in + 4 * 17) >> 23 | SafeLoadAs<uint32_t>(in + 4 * 18) << 9,
+      SafeLoadAs<uint32_t>(in + 4 * 18) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 19) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 19),
+      SafeLoadAs<uint32_t>(in + 4 * 19) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 20) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 20),
+    };
+    shifts = simd_batch{ 0, 0, 2, 0, 0, 1, 0, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 21 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint32_t, 22> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 8>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x3fffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 22-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 1) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 2) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 3) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 14 | SafeLoadAs<uint32_t>(in + 4 * 4) << 18,
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 5) << 6,
+    };
+    shifts = simd_batch{ 0, 0, 0, 2, 0, 0, 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 22-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 6) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 7) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 8) << 14,
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 9) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 10) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+    };
+    shifts = simd_batch{ 0, 6, 0, 0, 8, 0, 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 22-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 11),
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 12) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 13) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 13),
+      SafeLoadAs<uint32_t>(in + 4 * 13) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 14) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 14) >> 14 | SafeLoadAs<uint32_t>(in + 4 * 15) << 18,
+      SafeLoadAs<uint32_t>(in + 4 * 15),
+      SafeLoadAs<uint32_t>(in + 4 * 15) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 16) << 6,
+    };
+    shifts = simd_batch{ 0, 0, 0, 2, 0, 0, 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 22-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 16) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 17) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 17),
+      SafeLoadAs<uint32_t>(in + 4 * 17) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 18) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 18) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 19) << 14,
+      SafeLoadAs<uint32_t>(in + 4 * 19),
+      SafeLoadAs<uint32_t>(in + 4 * 19) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 20) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 20) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 21) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 21),
+    };
+    shifts = simd_batch{ 0, 6, 0, 0, 8, 0, 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 22 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint32_t, 23> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 8>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x7fffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 23-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 23 | SafeLoadAs<uint32_t>(in + 4 * 1) << 9,
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 14 | SafeLoadAs<uint32_t>(in + 4 * 2) << 18,
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 3) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 19 | SafeLoadAs<uint32_t>(in + 4 * 4) << 13,
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 10 | SafeLoadAs<uint32_t>(in + 4 * 5) << 22,
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+    };
+    shifts = simd_batch{ 0, 0, 0, 5, 0, 0, 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 23-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 6) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 15 | SafeLoadAs<uint32_t>(in + 4 * 7) << 17,
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 29 | SafeLoadAs<uint32_t>(in + 4 * 8) << 3,
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 9) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 11 | SafeLoadAs<uint32_t>(in + 4 * 10) << 21,
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 25 | SafeLoadAs<uint32_t>(in + 4 * 11) << 7,
+    };
+    shifts = simd_batch{ 0, 0, 6, 0, 0, 0, 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 23-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 12) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 12),
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 13) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 13) >> 21 | SafeLoadAs<uint32_t>(in + 4 * 14) << 11,
+      SafeLoadAs<uint32_t>(in + 4 * 14) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 15) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 15),
+      SafeLoadAs<uint32_t>(in + 4 * 15) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 16) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 16) >> 17 | SafeLoadAs<uint32_t>(in + 4 * 17) << 15,
+    };
+    shifts = simd_batch{ 0, 7, 0, 0, 0, 3, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 23-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 17),
+      SafeLoadAs<uint32_t>(in + 4 * 17) >> 31 | SafeLoadAs<uint32_t>(in + 4 * 18) << 1,
+      SafeLoadAs<uint32_t>(in + 4 * 18) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 19) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 19) >> 13 | SafeLoadAs<uint32_t>(in + 4 * 20) << 19,
+      SafeLoadAs<uint32_t>(in + 4 * 20),
+      SafeLoadAs<uint32_t>(in + 4 * 20) >> 27 | SafeLoadAs<uint32_t>(in + 4 * 21) << 5,
+      SafeLoadAs<uint32_t>(in + 4 * 21) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 22) << 14,
+      SafeLoadAs<uint32_t>(in + 4 * 22),
+    };
+    shifts = simd_batch{ 8, 0, 0, 0, 4, 0, 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 23 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint32_t, 24> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 8>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0xffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 24-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 1) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 2) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 4) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 5) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+    };
+    shifts = simd_batch{ 0, 0, 0, 8, 0, 0, 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 24-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 7) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 8) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 10) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 11) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 11),
+    };
+    shifts = simd_batch{ 0, 0, 0, 8, 0, 0, 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 24-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 12),
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 13) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 13) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 14) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 14),
+      SafeLoadAs<uint32_t>(in + 4 * 15),
+      SafeLoadAs<uint32_t>(in + 4 * 15) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 16) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 16) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 17) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 17),
+    };
+    shifts = simd_batch{ 0, 0, 0, 8, 0, 0, 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 24-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 18),
+      SafeLoadAs<uint32_t>(in + 4 * 18) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 19) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 19) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 20) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 20),
+      SafeLoadAs<uint32_t>(in + 4 * 21),
+      SafeLoadAs<uint32_t>(in + 4 * 21) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 22) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 22) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 23) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 23),
+    };
+    shifts = simd_batch{ 0, 0, 0, 8, 0, 0, 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 24 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint32_t, 25> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 8>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x1ffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 25-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 25 | SafeLoadAs<uint32_t>(in + 4 * 1) << 7,
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 2) << 14,
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 11 | SafeLoadAs<uint32_t>(in + 4 * 3) << 21,
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 29 | SafeLoadAs<uint32_t>(in + 4 * 4) << 3,
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 5) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 15 | SafeLoadAs<uint32_t>(in + 4 * 6) << 17,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 4, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 25-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 8 | SafeLoadAs<uint32_t>(in + 4 * 7) << 24,
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 8) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 19 | SafeLoadAs<uint32_t>(in + 4 * 9) << 13,
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 10) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 11) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 23 | SafeLoadAs<uint32_t>(in + 4 * 12) << 9,
+    };
+    shifts = simd_batch{ 0, 1, 0, 0, 0, 5, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 25-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 13) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 13) >> 9 | SafeLoadAs<uint32_t>(in + 4 * 14) << 23,
+      SafeLoadAs<uint32_t>(in + 4 * 14),
+      SafeLoadAs<uint32_t>(in + 4 * 14) >> 27 | SafeLoadAs<uint32_t>(in + 4 * 15) << 5,
+      SafeLoadAs<uint32_t>(in + 4 * 15) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 16) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 16) >> 13 | SafeLoadAs<uint32_t>(in + 4 * 17) << 19,
+      SafeLoadAs<uint32_t>(in + 4 * 17),
+      SafeLoadAs<uint32_t>(in + 4 * 17) >> 31 | SafeLoadAs<uint32_t>(in + 4 * 18) << 1,
+    };
+    shifts = simd_batch{ 0, 0, 2, 0, 0, 0, 6, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 25-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 18) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 19) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 19) >> 17 | SafeLoadAs<uint32_t>(in + 4 * 20) << 15,
+      SafeLoadAs<uint32_t>(in + 4 * 20) >> 10 | SafeLoadAs<uint32_t>(in + 4 * 21) << 22,
+      SafeLoadAs<uint32_t>(in + 4 * 21),
+      SafeLoadAs<uint32_t>(in + 4 * 21) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 22) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 22) >> 21 | SafeLoadAs<uint32_t>(in + 4 * 23) << 11,
+      SafeLoadAs<uint32_t>(in + 4 * 23) >> 14 | SafeLoadAs<uint32_t>(in + 4 * 24) << 18,
+      SafeLoadAs<uint32_t>(in + 4 * 24),
+    };
+    shifts = simd_batch{ 0, 0, 0, 3, 0, 0, 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 25 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint32_t, 26> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 8>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x3ffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 26-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 1) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 2) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 14 | SafeLoadAs<uint32_t>(in + 4 * 3) << 18,
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 8 | SafeLoadAs<uint32_t>(in + 4 * 4) << 24,
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 5) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 6) << 10,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 2, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 26-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 7) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 10 | SafeLoadAs<uint32_t>(in + 4 * 8) << 22,
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 9) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 10) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 11) << 14,
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 12) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 12),
+    };
+    shifts = simd_batch{ 0, 0, 4, 0, 0, 0, 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 26-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 13),
+      SafeLoadAs<uint32_t>(in + 4 * 13) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 14) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 14) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 15) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 15) >> 14 | SafeLoadAs<uint32_t>(in + 4 * 16) << 18,
+      SafeLoadAs<uint32_t>(in + 4 * 16) >> 8 | SafeLoadAs<uint32_t>(in + 4 * 17) << 24,
+      SafeLoadAs<uint32_t>(in + 4 * 17),
+      SafeLoadAs<uint32_t>(in + 4 * 17) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 18) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 18) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 19) << 10,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 2, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 26-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 19) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 20) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 20) >> 10 | SafeLoadAs<uint32_t>(in + 4 * 21) << 22,
+      SafeLoadAs<uint32_t>(in + 4 * 21),
+      SafeLoadAs<uint32_t>(in + 4 * 21) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 22) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 22) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 23) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 23) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 24) << 14,
+      SafeLoadAs<uint32_t>(in + 4 * 24) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 25) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 25),
+    };
+    shifts = simd_batch{ 0, 0, 4, 0, 0, 0, 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 26 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint32_t, 27> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 8>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x7ffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 27-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 27 | SafeLoadAs<uint32_t>(in + 4 * 1) << 5,
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 2) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 17 | SafeLoadAs<uint32_t>(in + 4 * 3) << 15,
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 4) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 7 | SafeLoadAs<uint32_t>(in + 4 * 5) << 25,
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 29 | SafeLoadAs<uint32_t>(in + 4 * 6) << 3,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 27-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 7) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 19 | SafeLoadAs<uint32_t>(in + 4 * 8) << 13,
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 14 | SafeLoadAs<uint32_t>(in + 4 * 9) << 18,
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 9 | SafeLoadAs<uint32_t>(in + 4 * 10) << 23,
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 31 | SafeLoadAs<uint32_t>(in + 4 * 11) << 1,
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 12) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 21 | SafeLoadAs<uint32_t>(in + 4 * 13) << 11,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 4, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 27-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 13) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 14) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 14) >> 11 | SafeLoadAs<uint32_t>(in + 4 * 15) << 21,
+      SafeLoadAs<uint32_t>(in + 4 * 15) >> 6 | SafeLoadAs<uint32_t>(in + 4 * 16) << 26,
+      SafeLoadAs<uint32_t>(in + 4 * 16),
+      SafeLoadAs<uint32_t>(in + 4 * 16) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 17) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 17) >> 23 | SafeLoadAs<uint32_t>(in + 4 * 18) << 9,
+      SafeLoadAs<uint32_t>(in + 4 * 18) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 19) << 14,
+      SafeLoadAs<uint32_t>(in + 4 * 19) >> 13 | SafeLoadAs<uint32_t>(in + 4 * 20) << 19,
+    };
+    shifts = simd_batch{ 0, 0, 0, 1, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 27-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 20) >> 8 | SafeLoadAs<uint32_t>(in + 4 * 21) << 24,
+      SafeLoadAs<uint32_t>(in + 4 * 21),
+      SafeLoadAs<uint32_t>(in + 4 * 21) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 22) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 22) >> 25 | SafeLoadAs<uint32_t>(in + 4 * 23) << 7,
+      SafeLoadAs<uint32_t>(in + 4 * 23) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 24) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 24) >> 15 | SafeLoadAs<uint32_t>(in + 4 * 25) << 17,
+      SafeLoadAs<uint32_t>(in + 4 * 25) >> 10 | SafeLoadAs<uint32_t>(in + 4 * 26) << 22,
+      SafeLoadAs<uint32_t>(in + 4 * 26),
+    };
+    shifts = simd_batch{ 0, 3, 0, 0, 0, 0, 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 27 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint32_t, 28> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 8>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0xfffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 28-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 1) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 2) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 3) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 4) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 5) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 8 | SafeLoadAs<uint32_t>(in + 4 * 6) << 24,
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 28-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 8) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 9) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 10) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 11) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 12) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 8 | SafeLoadAs<uint32_t>(in + 4 * 13) << 24,
+      SafeLoadAs<uint32_t>(in + 4 * 13),
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 28-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 14),
+      SafeLoadAs<uint32_t>(in + 4 * 14) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 15) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 15) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 16) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 16) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 17) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 17) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 18) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 18) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 19) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 19) >> 8 | SafeLoadAs<uint32_t>(in + 4 * 20) << 24,
+      SafeLoadAs<uint32_t>(in + 4 * 20),
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 28-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 21),
+      SafeLoadAs<uint32_t>(in + 4 * 21) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 22) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 22) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 23) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 23) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 24) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 24) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 25) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 25) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 26) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 26) >> 8 | SafeLoadAs<uint32_t>(in + 4 * 27) << 24,
+      SafeLoadAs<uint32_t>(in + 4 * 27),
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 28 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint32_t, 29> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 8>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x1fffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 29-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 29 | SafeLoadAs<uint32_t>(in + 4 * 1) << 3,
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 2) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 23 | SafeLoadAs<uint32_t>(in + 4 * 3) << 9,
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 4) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 17 | SafeLoadAs<uint32_t>(in + 4 * 5) << 15,
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 14 | SafeLoadAs<uint32_t>(in + 4 * 6) << 18,
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 11 | SafeLoadAs<uint32_t>(in + 4 * 7) << 21,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 29-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 8 | SafeLoadAs<uint32_t>(in + 4 * 8) << 24,
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 5 | SafeLoadAs<uint32_t>(in + 4 * 9) << 27,
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 31 | SafeLoadAs<uint32_t>(in + 4 * 10) << 1,
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 11) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 25 | SafeLoadAs<uint32_t>(in + 4 * 12) << 7,
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 13) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 13) >> 19 | SafeLoadAs<uint32_t>(in + 4 * 14) << 13,
+    };
+    shifts = simd_batch{ 0, 0, 2, 0, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 29-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 14) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 15) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 15) >> 13 | SafeLoadAs<uint32_t>(in + 4 * 16) << 19,
+      SafeLoadAs<uint32_t>(in + 4 * 16) >> 10 | SafeLoadAs<uint32_t>(in + 4 * 17) << 22,
+      SafeLoadAs<uint32_t>(in + 4 * 17) >> 7 | SafeLoadAs<uint32_t>(in + 4 * 18) << 25,
+      SafeLoadAs<uint32_t>(in + 4 * 18) >> 4 | SafeLoadAs<uint32_t>(in + 4 * 19) << 28,
+      SafeLoadAs<uint32_t>(in + 4 * 19),
+      SafeLoadAs<uint32_t>(in + 4 * 19) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 20) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 20) >> 27 | SafeLoadAs<uint32_t>(in + 4 * 21) << 5,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 1, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 29-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 21) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 22) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 22) >> 21 | SafeLoadAs<uint32_t>(in + 4 * 23) << 11,
+      SafeLoadAs<uint32_t>(in + 4 * 23) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 24) << 14,
+      SafeLoadAs<uint32_t>(in + 4 * 24) >> 15 | SafeLoadAs<uint32_t>(in + 4 * 25) << 17,
+      SafeLoadAs<uint32_t>(in + 4 * 25) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 26) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 26) >> 9 | SafeLoadAs<uint32_t>(in + 4 * 27) << 23,
+      SafeLoadAs<uint32_t>(in + 4 * 27) >> 6 | SafeLoadAs<uint32_t>(in + 4 * 28) << 26,
+      SafeLoadAs<uint32_t>(in + 4 * 28),
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 29 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint32_t, 30> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 8>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x3fffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 30-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 1) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 2) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 3) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 4) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 5) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 6) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 7) << 14,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 30-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 8) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 14 | SafeLoadAs<uint32_t>(in + 4 * 9) << 18,
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 10) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 10 | SafeLoadAs<uint32_t>(in + 4 * 11) << 22,
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 8 | SafeLoadAs<uint32_t>(in + 4 * 12) << 24,
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 6 | SafeLoadAs<uint32_t>(in + 4 * 13) << 26,
+      SafeLoadAs<uint32_t>(in + 4 * 13) >> 4 | SafeLoadAs<uint32_t>(in + 4 * 14) << 28,
+      SafeLoadAs<uint32_t>(in + 4 * 14),
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 30-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 15),
+      SafeLoadAs<uint32_t>(in + 4 * 15) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 16) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 16) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 17) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 17) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 18) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 18) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 19) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 19) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 20) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 20) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 21) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 21) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 22) << 14,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 30-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 22) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 23) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 23) >> 14 | SafeLoadAs<uint32_t>(in + 4 * 24) << 18,
+      SafeLoadAs<uint32_t>(in + 4 * 24) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 25) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 25) >> 10 | SafeLoadAs<uint32_t>(in + 4 * 26) << 22,
+      SafeLoadAs<uint32_t>(in + 4 * 26) >> 8 | SafeLoadAs<uint32_t>(in + 4 * 27) << 24,
+      SafeLoadAs<uint32_t>(in + 4 * 27) >> 6 | SafeLoadAs<uint32_t>(in + 4 * 28) << 26,
+      SafeLoadAs<uint32_t>(in + 4 * 28) >> 4 | SafeLoadAs<uint32_t>(in + 4 * 29) << 28,
+      SafeLoadAs<uint32_t>(in + 4 * 29),
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 30 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint32_t, 31> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 8>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x7fffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 31-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 31 | SafeLoadAs<uint32_t>(in + 4 * 1) << 1,
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 2) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 29 | SafeLoadAs<uint32_t>(in + 4 * 3) << 3,
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 4) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 27 | SafeLoadAs<uint32_t>(in + 4 * 5) << 5,
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 6) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 25 | SafeLoadAs<uint32_t>(in + 4 * 7) << 7,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 31-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 8) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 23 | SafeLoadAs<uint32_t>(in + 4 * 9) << 9,
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 10) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 21 | SafeLoadAs<uint32_t>(in + 4 * 11) << 11,
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 12) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 19 | SafeLoadAs<uint32_t>(in + 4 * 13) << 13,
+      SafeLoadAs<uint32_t>(in + 4 * 13) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 14) << 14,
+      SafeLoadAs<uint32_t>(in + 4 * 14) >> 17 | SafeLoadAs<uint32_t>(in + 4 * 15) << 15,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 31-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 15) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 16) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 16) >> 15 | SafeLoadAs<uint32_t>(in + 4 * 17) << 17,
+      SafeLoadAs<uint32_t>(in + 4 * 17) >> 14 | SafeLoadAs<uint32_t>(in + 4 * 18) << 18,
+      SafeLoadAs<uint32_t>(in + 4 * 18) >> 13 | SafeLoadAs<uint32_t>(in + 4 * 19) << 19,
+      SafeLoadAs<uint32_t>(in + 4 * 19) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 20) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 20) >> 11 | SafeLoadAs<uint32_t>(in + 4 * 21) << 21,
+      SafeLoadAs<uint32_t>(in + 4 * 21) >> 10 | SafeLoadAs<uint32_t>(in + 4 * 22) << 22,
+      SafeLoadAs<uint32_t>(in + 4 * 22) >> 9 | SafeLoadAs<uint32_t>(in + 4 * 23) << 23,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 31-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 23) >> 8 | SafeLoadAs<uint32_t>(in + 4 * 24) << 24,
+      SafeLoadAs<uint32_t>(in + 4 * 24) >> 7 | SafeLoadAs<uint32_t>(in + 4 * 25) << 25,
+      SafeLoadAs<uint32_t>(in + 4 * 25) >> 6 | SafeLoadAs<uint32_t>(in + 4 * 26) << 26,
+      SafeLoadAs<uint32_t>(in + 4 * 26) >> 5 | SafeLoadAs<uint32_t>(in + 4 * 27) << 27,
+      SafeLoadAs<uint32_t>(in + 4 * 27) >> 4 | SafeLoadAs<uint32_t>(in + 4 * 28) << 28,
+      SafeLoadAs<uint32_t>(in + 4 * 28) >> 3 | SafeLoadAs<uint32_t>(in + 4 * 29) << 29,
+      SafeLoadAs<uint32_t>(in + 4 * 29) >> 2 | SafeLoadAs<uint32_t>(in + 4 * 30) << 30,
+      SafeLoadAs<uint32_t>(in + 4 * 30),
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 31 * 4;
+    return in;
+  }
+};
+
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 1> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 1-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 1, 2, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 1-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 4, 5, 6, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 1-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 8, 9, 10, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 1-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 12, 13, 14, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 1-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 16, 17, 18, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 1-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 20, 21, 22, 23 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 1-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 24, 25, 26, 27 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 1-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 28, 29, 30, 31 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 1-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 32, 33, 34, 35 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 1-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 36, 37, 38, 39 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 1-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 40, 41, 42, 43 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 1-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 44, 45, 46, 47 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 1-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 48, 49, 50, 51 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 1-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 52, 53, 54, 55 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 1-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 56, 57, 58, 59 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 1-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 60, 61, 62, 63 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 1 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 2> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 2-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 2, 4, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 2-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 8, 10, 12, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 2-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 16, 18, 20, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 2-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 24, 26, 28, 30 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 2-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 32, 34, 36, 38 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 2-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 40, 42, 44, 46 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 2-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 48, 50, 52, 54 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 2-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 56, 58, 60, 62 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 2-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 2, 4, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 2-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 8, 10, 12, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 2-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 16, 18, 20, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 2-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 24, 26, 28, 30 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 2-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 32, 34, 36, 38 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 2-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 40, 42, 44, 46 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 2-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 48, 50, 52, 54 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 2-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 56, 58, 60, 62 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 2 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 3> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 3-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 3, 6, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 3-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 12, 15, 18, 21 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 3-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 24, 27, 30, 33 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 3-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 36, 39, 42, 45 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 3-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 48, 51, 54, 57 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 3-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 1) << 1,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 60, 0, 2, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 3-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 8, 11, 14, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 3-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 20, 23, 26, 29 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 3-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 32, 35, 38, 41 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 3-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 44, 47, 50, 53 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 3-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 2) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 56, 59, 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 3-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 4, 7, 10, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 3-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 16, 19, 22, 25 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 3-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 28, 31, 34, 37 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 3-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 40, 43, 46, 49 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 3-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 52, 55, 58, 61 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 3 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 4> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0xf;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 4-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 4, 8, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 4-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 16, 20, 24, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 4-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 32, 36, 40, 44 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 4-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 48, 52, 56, 60 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 4-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 4, 8, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 4-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 16, 20, 24, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 4-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 32, 36, 40, 44 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 4-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 48, 52, 56, 60 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 4-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 0, 4, 8, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 4-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 16, 20, 24, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 4-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 32, 36, 40, 44 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 4-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 48, 52, 56, 60 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 4-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 0, 4, 8, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 4-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 16, 20, 24, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 4-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 32, 36, 40, 44 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 4-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 48, 52, 56, 60 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 4 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 5> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1f;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 5-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 5, 10, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 5-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 20, 25, 30, 35 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 5-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 40, 45, 50, 55 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 5-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 1) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 1, 6, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 5-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 16, 21, 26, 31 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 5-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 36, 41, 46, 51 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 5-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 2) << 3,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 56, 0, 2, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 5-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 12, 17, 22, 27 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 5-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 32, 37, 42, 47 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 5-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 3) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 52, 57, 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 5-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 8, 13, 18, 23 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 5-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 28, 33, 38, 43 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 5-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 4) << 1,
+    };
+    shifts = simd_batch{ 48, 53, 58, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 5-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 4, 9, 14, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 5-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 24, 29, 34, 39 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 5-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 44, 49, 54, 59 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 5 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 6> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3f;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 6-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 6, 12, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 6-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 24, 30, 36, 42 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 6-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 1) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 48, 54, 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 6-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 8, 14, 20, 26 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 6-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 32, 38, 44, 50 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 6-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 2) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 56, 0, 4, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 6-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 16, 22, 28, 34 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 6-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 40, 46, 52, 58 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 6-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 0, 6, 12, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 6-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 24, 30, 36, 42 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 6-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 4) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 48, 54, 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 6-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 8, 14, 20, 26 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 6-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 32, 38, 44, 50 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 6-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 5) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 56, 0, 4, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 6-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 16, 22, 28, 34 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 6-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 40, 46, 52, 58 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 6 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 7> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7f;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 7-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 7, 14, 21 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 7-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 28, 35, 42, 49 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 7-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 1) << 1,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 56, 0, 6, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 7-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 20, 27, 34, 41 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 7-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 2) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 48, 55, 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 7-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 12, 19, 26, 33 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 7-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 3) << 3,
+    };
+    shifts = simd_batch{ 40, 47, 54, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 7-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 4, 11, 18, 25 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 7-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 32, 39, 46, 53 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 7-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 4) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 0, 3, 10, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 7-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 24, 31, 38, 45 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 7-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 5) << 5,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 52, 0, 2, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 7-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 16, 23, 30, 37 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 7-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 6) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 44, 51, 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 7-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 8, 15, 22, 29 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 7-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 36, 43, 50, 57 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 7 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 8> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0xff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 8-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 8, 16, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 8-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 32, 40, 48, 56 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 8-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 8, 16, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 8-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 32, 40, 48, 56 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 8-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 0, 8, 16, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 8-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 32, 40, 48, 56 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 8-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 0, 8, 16, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 8-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 32, 40, 48, 56 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 8-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 0, 8, 16, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 8-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 32, 40, 48, 56 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 8-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 0, 8, 16, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 8-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 32, 40, 48, 56 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 8-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 0, 8, 16, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 8-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 32, 40, 48, 56 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 8-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 0, 8, 16, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 8-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 32, 40, 48, 56 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 8 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 9> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1ff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 9-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 9, 18, 27 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 9-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 1) << 1,
+    };
+    shifts = simd_batch{ 36, 45, 54, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 9-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 8, 17, 26, 35 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 9-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 2) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 44, 53, 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 9-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 16, 25, 34, 43 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 9-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 3) << 3,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 52, 0, 6, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 9-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 24, 33, 42, 51 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 9-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 4) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 0, 5, 14, 23 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 9-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 5) << 5,
+    };
+    shifts = simd_batch{ 32, 41, 50, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 9-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 4, 13, 22, 31 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 9-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 6) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 40, 49, 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 9-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 12, 21, 30, 39 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 9-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 7) << 7,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 48, 0, 2, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 9-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 20, 29, 38, 47 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 9-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 8) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 0, 1, 10, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 9-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 28, 37, 46, 55 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 9 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 10> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3ff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 10-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 10, 20, 30 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 10-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 1) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 40, 50, 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 10-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 16, 26, 36, 46 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 10-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 2) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 0, 2, 12, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 10-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 3) << 2,
+    };
+    shifts = simd_batch{ 32, 42, 52, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 10-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 8, 18, 28, 38 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 10-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 4) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 48, 0, 4, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 10-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 24, 34, 44, 54 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 10-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 0, 10, 20, 30 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 10-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 6) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 40, 50, 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 10-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 16, 26, 36, 46 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 10-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 7) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 0, 2, 12, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 10-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 8) << 2,
+    };
+    shifts = simd_batch{ 32, 42, 52, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 10-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 8, 18, 28, 38 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 10-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 9) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 48, 0, 4, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 10-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 24, 34, 44, 54 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 10 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 11> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7ff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 11-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 11, 22, 33 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 11-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 1) << 9,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 44, 0, 2, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 11-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 2) << 7,
+    };
+    shifts = simd_batch{ 24, 35, 46, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 11-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 4, 15, 26, 37 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 11-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 3) << 5,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 48, 0, 6, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 11-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 4) << 3,
+    };
+    shifts = simd_batch{ 28, 39, 50, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 11-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 8, 19, 30, 41 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 11-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 5) << 1,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 52, 0, 10, 21 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 11-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 6) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 32, 43, 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 11-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 12, 23, 34, 45 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 11-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 7) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 0, 3, 14, 25 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 11-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 8) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 36, 47, 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 11-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 16, 27, 38, 49 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 11-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 9) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 0, 7, 18, 29 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 11-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 10) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 40, 51, 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 11-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 20, 31, 42, 53 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 11 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 12> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0xfff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 12-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 12, 24, 36 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 12-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 1) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 48, 0, 8, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 12-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 2) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 32, 44, 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 12-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 16, 28, 40, 52 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 12-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 0, 12, 24, 36 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 12-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 4) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 48, 0, 8, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 12-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 5) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 32, 44, 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 12-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 16, 28, 40, 52 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 12-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 0, 12, 24, 36 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 12-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 7) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 48, 0, 8, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 12-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 8) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 32, 44, 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 12-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 16, 28, 40, 52 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 12-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 0, 12, 24, 36 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 12-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 10) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 48, 0, 8, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 12-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 11) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 32, 44, 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 12-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 16, 28, 40, 52 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 12 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 13> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1fff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 13-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 13, 26, 39 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 13-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 1) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 1, 14, 27 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 13-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 2) << 11,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 40, 0, 2, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 13-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 3) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 28, 41, 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 13-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 4) << 9,
+    };
+    shifts = simd_batch{ 16, 29, 42, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 13-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 4, 17, 30, 43 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 13-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 5) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 0, 5, 18, 31 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 13-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 6) << 7,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 44, 0, 6, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 13-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 7) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 32, 45, 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 13-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 8) << 5,
+    };
+    shifts = simd_batch{ 20, 33, 46, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 13-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 8, 21, 34, 47 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 13-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 9) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 0, 9, 22, 35 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 13-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 10) << 3,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 48, 0, 10, 23 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 13-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 11) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 36, 49, 0, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 13-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 12) << 1,
+    };
+    shifts = simd_batch{ 24, 37, 50, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 13-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 12, 25, 38, 51 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 13 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 14> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3fff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 14-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 14, 28, 42 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 14-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 1) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 6, 20, 34 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 14-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 2) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 48, 0, 12, 26 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 14-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 3) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 40, 0, 4, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 14-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 4) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 32, 46, 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 14-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 5) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 24, 38, 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 14-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 6) << 6,
+    };
+    shifts = simd_batch{ 16, 30, 44, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 14-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 8, 22, 36, 50 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 14-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 0, 14, 28, 42 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 14-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 8) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 0, 6, 20, 34 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 14-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 9) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 48, 0, 12, 26 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 14-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 10) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 40, 0, 4, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 14-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 11) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 32, 46, 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 14-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 12) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 24, 38, 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 14-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 13) << 6,
+    };
+    shifts = simd_batch{ 16, 30, 44, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 14-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 8, 22, 36, 50 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 14 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 15> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7fff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 15-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 15, 30, 45 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 15-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 1) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 11, 26, 41 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 15-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 2) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 0, 7, 22, 37 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 15-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 3) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 0, 3, 18, 33 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 15-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 4) << 1,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 48, 0, 14, 29 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 15-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 5) << 5,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 44, 0, 10, 25 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 15-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 6) << 9,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 40, 0, 6, 21 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 15-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 7) << 13,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 36, 0, 2, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 15-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 8) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 32, 47, 0, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 15-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 9) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 28, 43, 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 15-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 10) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 24, 39, 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 15-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 11) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 20, 35, 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 15-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 12) << 3,
+    };
+    shifts = simd_batch{ 16, 31, 46, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 15-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 13) << 7,
+    };
+    shifts = simd_batch{ 12, 27, 42, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 15-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 14) << 11,
+    };
+    shifts = simd_batch{ 8, 23, 38, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 15-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 4, 19, 34, 49 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 15 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 16> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0xffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 16-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 16, 32, 48 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 16-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 16, 32, 48 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 16-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 0, 16, 32, 48 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 16-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 0, 16, 32, 48 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 16-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 0, 16, 32, 48 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 16-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 0, 16, 32, 48 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 16-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 0, 16, 32, 48 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 16-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 0, 16, 32, 48 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 16-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 0, 16, 32, 48 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 16-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 0, 16, 32, 48 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 16-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 0, 16, 32, 48 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 16-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 0, 16, 32, 48 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 16-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 0, 16, 32, 48 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 16-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 0, 16, 32, 48 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 16-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 0, 16, 32, 48 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 16-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+    };
+    shifts = simd_batch{ 0, 16, 32, 48 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 16 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 17> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1ffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 17-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 1) << 13,
+    };
+    shifts = simd_batch{ 0, 17, 34, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 17-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 2) << 9,
+    };
+    shifts = simd_batch{ 4, 21, 38, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 17-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 3) << 5,
+    };
+    shifts = simd_batch{ 8, 25, 42, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 17-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 4) << 1,
+    };
+    shifts = simd_batch{ 12, 29, 46, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 17-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 5) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 16, 33, 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 17-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 6) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 20, 37, 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 17-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 7) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 24, 41, 0, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 17-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 8) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 28, 45, 0, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 17-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 9) << 15,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 32, 0, 2, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 17-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 10) << 11,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 36, 0, 6, 23 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 17-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 11) << 7,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 40, 0, 10, 27 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 17-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 12) << 3,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 44, 0, 14, 31 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 17-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 13) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 0, 1, 18, 35 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 17-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 14) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 0, 5, 22, 39 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 17-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 15) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+    };
+    shifts = simd_batch{ 0, 9, 26, 43 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 17-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 16) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+    };
+    shifts = simd_batch{ 0, 13, 30, 47 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 17 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 18> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3ffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 18-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 1) << 10,
+    };
+    shifts = simd_batch{ 0, 18, 36, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 18-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 2) << 2,
+    };
+    shifts = simd_batch{ 8, 26, 44, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 18-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 3) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 16, 34, 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 18-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 4) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 24, 42, 0, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 18-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 5) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 32, 0, 4, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 18-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 6) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 40, 0, 12, 30 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 18-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 7) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 0, 2, 20, 38 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 18-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 8) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 0, 10, 28, 46 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 18-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 10) << 10,
+    };
+    shifts = simd_batch{ 0, 18, 36, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 18-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 11) << 2,
+    };
+    shifts = simd_batch{ 8, 26, 44, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 18-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 12) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 16, 34, 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 18-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 13) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 24, 42, 0, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 18-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 14) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 32, 0, 4, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 18-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 15) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+    };
+    shifts = simd_batch{ 40, 0, 12, 30 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 18-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 16) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+    };
+    shifts = simd_batch{ 0, 2, 20, 38 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 18-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 17) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+    };
+    shifts = simd_batch{ 0, 10, 28, 46 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 18 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 19> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7ffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 19-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 1) << 7,
+    };
+    shifts = simd_batch{ 0, 19, 38, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 19-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 2) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 12, 31, 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 19-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 3) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 24, 43, 0, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 19-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 4) << 9,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 36, 0, 10, 29 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 19-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 5) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 0, 3, 22, 41 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 19-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 6) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 7) << 11,
+    };
+    shifts = simd_batch{ 0, 15, 34, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 19-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 8) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 8, 27, 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 19-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 9) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 20, 39, 0, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 19-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 10) << 13,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 32, 0, 6, 25 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 19-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 11) << 1,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 44, 0, 18, 37 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 19-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 12) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 13) << 15,
+    };
+    shifts = simd_batch{ 0, 11, 30, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 19-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 14) << 3,
+    };
+    shifts = simd_batch{ 4, 23, 42, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 19-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 15) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+    };
+    shifts = simd_batch{ 16, 35, 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 19-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 16) << 17,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+    };
+    shifts = simd_batch{ 28, 0, 2, 21 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 19-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 17) << 5,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+    };
+    shifts = simd_batch{ 40, 0, 14, 33 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 19-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 18) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+    };
+    shifts = simd_batch{ 0, 7, 26, 45 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 19 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 20> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0xfffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 20-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 1) << 4,
+    };
+    shifts = simd_batch{ 0, 20, 40, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 20-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 2) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 16, 36, 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 20-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 3) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 32, 0, 8, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 20-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 4) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 0, 4, 24, 44 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 20-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 6) << 4,
+    };
+    shifts = simd_batch{ 0, 20, 40, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 20-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 7) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 16, 36, 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 20-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 8) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 32, 0, 8, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 20-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 9) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 0, 4, 24, 44 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 20-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 11) << 4,
+    };
+    shifts = simd_batch{ 0, 20, 40, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 20-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 12) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 16, 36, 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 20-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 13) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 32, 0, 8, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 20-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 14) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 0, 4, 24, 44 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 20-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 16) << 4,
+    };
+    shifts = simd_batch{ 0, 20, 40, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 20-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 17) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+    };
+    shifts = simd_batch{ 16, 36, 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 20-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 18) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+    };
+    shifts = simd_batch{ 32, 0, 8, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 20-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 19) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+    };
+    shifts = simd_batch{ 0, 4, 24, 44 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 20 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 21> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1fffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 21-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 1) << 1,
+    };
+    shifts = simd_batch{ 0, 21, 42, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 21-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 2) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 20, 41, 0, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 21-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 3) << 3,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 40, 0, 18, 39 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 21-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 4) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 5) << 5,
+    };
+    shifts = simd_batch{ 0, 17, 38, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 21-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 6) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 16, 37, 0, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 21-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 7) << 7,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 36, 0, 14, 35 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 21-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 8) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 9) << 9,
+    };
+    shifts = simd_batch{ 0, 13, 34, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 21-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 10) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 12, 33, 0, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 21-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 11) << 11,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 32, 0, 10, 31 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 21-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 12) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 13) << 13,
+    };
+    shifts = simd_batch{ 0, 9, 30, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 21-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 14) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 8, 29, 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 21-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 15) << 15,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+    };
+    shifts = simd_batch{ 28, 0, 6, 27 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 21-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 16) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 17) << 17,
+    };
+    shifts = simd_batch{ 0, 5, 26, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 21-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 18) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+    };
+    shifts = simd_batch{ 4, 25, 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 21-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 19) << 19,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+    };
+    shifts = simd_batch{ 24, 0, 2, 23 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 21-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 20) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+    };
+    shifts = simd_batch{ 0, 1, 22, 43 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 21 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 22> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3fffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 22-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 1) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 22, 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 22-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 2) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 24, 0, 4, 26 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 22-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 3) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 4) << 14,
+    };
+    shifts = simd_batch{ 0, 6, 28, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 22-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 5) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 8, 30, 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 22-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 6) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 32, 0, 12, 34 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 22-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 7) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 8) << 6,
+    };
+    shifts = simd_batch{ 0, 14, 36, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 22-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 9) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 16, 38, 0, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 22-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 10) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 40, 0, 20, 42 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 22-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 12) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 0, 22, 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 22-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 13) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 24, 0, 4, 26 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 22-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 14) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 15) << 14,
+    };
+    shifts = simd_batch{ 0, 6, 28, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 22-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 16) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+    };
+    shifts = simd_batch{ 8, 30, 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 22-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 17) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+    };
+    shifts = simd_batch{ 32, 0, 12, 34 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 22-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 18) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 19) << 6,
+    };
+    shifts = simd_batch{ 0, 14, 36, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 22-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 20) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+    };
+    shifts = simd_batch{ 16, 38, 0, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 22-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 21) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+    };
+    shifts = simd_batch{ 40, 0, 20, 42 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 22 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 23> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7fffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 23-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 1) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 23, 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 23-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 2) << 13,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 28, 0, 10, 33 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 23-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 3) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 4) << 3,
+    };
+    shifts = simd_batch{ 0, 15, 38, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 23-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 5) << 21,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 20, 0, 2, 25 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 23-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 6) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 7) << 11,
+    };
+    shifts = simd_batch{ 0, 7, 30, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 23-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 8) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 12, 35, 0, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 23-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 9) << 1,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 10) << 19,
+    };
+    shifts = simd_batch{ 40, 0, 22, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 23-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 11) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 4, 27, 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 23-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 12) << 9,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 32, 0, 14, 37 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 23-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 13) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 14) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 0, 19, 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 23-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 15) << 17,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+    };
+    shifts = simd_batch{ 24, 0, 6, 29 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 23-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 16) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 17) << 7,
+    };
+    shifts = simd_batch{ 0, 11, 34, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 23-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 18) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+    };
+    shifts = simd_batch{ 16, 39, 0, 21 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 23-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 19) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 20) << 15,
+    };
+    shifts = simd_batch{ 0, 3, 26, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 23-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 21) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+    };
+    shifts = simd_batch{ 8, 31, 0, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 23-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 22) << 5,
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+    };
+    shifts = simd_batch{ 36, 0, 18, 41 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 23 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 24> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0xffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 24-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 1) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 24, 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 24-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 2) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 32, 0, 16, 40 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 24-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 4) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 0, 24, 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 24-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 5) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 32, 0, 16, 40 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 24-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 7) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 0, 24, 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 24-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 8) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 32, 0, 16, 40 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 24-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 10) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 0, 24, 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 24-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 11) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 32, 0, 16, 40 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 24-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 13) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 0, 24, 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 24-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 14) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 32, 0, 16, 40 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 24-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 16) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+    };
+    shifts = simd_batch{ 0, 24, 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 24-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 17) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+    };
+    shifts = simd_batch{ 32, 0, 16, 40 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 24-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 19) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+    };
+    shifts = simd_batch{ 0, 24, 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 24-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 20) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+    };
+    shifts = simd_batch{ 32, 0, 16, 40 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 24-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 22) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+    };
+    shifts = simd_batch{ 0, 24, 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 24-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 23) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+    };
+    shifts = simd_batch{ 32, 0, 16, 40 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 24 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 25> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1ffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 25-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 1) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 25, 0, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 25-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 2) << 3,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 3) << 17,
+    };
+    shifts = simd_batch{ 36, 0, 22, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 25-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 4) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 8, 33, 0, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 25-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 5) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 6) << 9,
+    };
+    shifts = simd_batch{ 0, 5, 30, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 25-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 7) << 23,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 16, 0, 2, 27 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 25-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 8) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 9) << 1,
+    };
+    shifts = simd_batch{ 0, 13, 38, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 25-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 10) << 15,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 24, 0, 10, 35 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 25-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 11) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 12) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 0, 21, 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 25-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 13) << 7,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 14) << 21,
+    };
+    shifts = simd_batch{ 32, 0, 18, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 25-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 15) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+    };
+    shifts = simd_batch{ 4, 29, 0, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 25-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 16) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 17) << 13,
+    };
+    shifts = simd_batch{ 0, 1, 26, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 25-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 18) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+    };
+    shifts = simd_batch{ 12, 37, 0, 23 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 25-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 19) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 20) << 5,
+    };
+    shifts = simd_batch{ 0, 9, 34, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 25-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 21) << 19,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+    };
+    shifts = simd_batch{ 20, 0, 6, 31 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 25-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 22) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 23) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+    };
+    shifts = simd_batch{ 0, 17, 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 25-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 24) << 11,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+    };
+    shifts = simd_batch{ 28, 0, 14, 39 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 25 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 26> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3ffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 26-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 1) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 26, 0, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 26-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 2) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 3) << 10,
+    };
+    shifts = simd_batch{ 0, 2, 28, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 26-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 4) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 16, 0, 4, 30 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 26-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 5) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 6) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 0, 18, 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 26-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 7) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 8) << 18,
+    };
+    shifts = simd_batch{ 32, 0, 20, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 26-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 9) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 8, 34, 0, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 26-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 10) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 11) << 2,
+    };
+    shifts = simd_batch{ 0, 10, 36, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 26-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 12) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 24, 0, 12, 38 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 26-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 14) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 0, 26, 0, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 26-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 15) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 16) << 10,
+    };
+    shifts = simd_batch{ 0, 2, 28, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 26-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 17) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+    };
+    shifts = simd_batch{ 16, 0, 4, 30 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 26-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 18) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 19) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+    };
+    shifts = simd_batch{ 0, 18, 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 26-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 20) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 21) << 18,
+    };
+    shifts = simd_batch{ 32, 0, 20, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 26-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 22) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+    };
+    shifts = simd_batch{ 8, 34, 0, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 26-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 23) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 24) << 2,
+    };
+    shifts = simd_batch{ 0, 10, 36, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 26-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 25) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+    };
+    shifts = simd_batch{ 24, 0, 12, 38 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 26 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 27> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7ffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 27-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 1) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 27, 0, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 27-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 2) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 3) << 3,
+    };
+    shifts = simd_batch{ 0, 7, 34, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 27-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 4) << 13,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 5) << 23,
+    };
+    shifts = simd_batch{ 24, 0, 14, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 27-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 6) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 4, 31, 0, 21 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 27-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 7) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 8) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 0, 11, 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 27-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 9) << 9,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 10) << 19,
+    };
+    shifts = simd_batch{ 28, 0, 18, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 27-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 11) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 8, 35, 0, 25 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 27-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 12) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 13) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 0, 15, 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 27-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 14) << 5,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 15) << 15,
+    };
+    shifts = simd_batch{ 32, 0, 22, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 27-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 16) << 25,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+    };
+    shifts = simd_batch{ 12, 0, 2, 29 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 27-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 17) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 18) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+    };
+    shifts = simd_batch{ 0, 19, 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 27-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 19) << 1,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 20) << 11,
+    };
+    shifts = simd_batch{ 36, 0, 26, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 27-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 21) << 21,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+    };
+    shifts = simd_batch{ 16, 0, 6, 33 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 27-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 22) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 23) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+    };
+    shifts = simd_batch{ 0, 23, 0, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 27-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 24) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 25) << 7,
+    };
+    shifts = simd_batch{ 0, 3, 30, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 27-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 26) << 17,
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+    };
+    shifts = simd_batch{ 20, 0, 10, 37 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 27 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 28> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0xfffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 28-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 1) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 28, 0, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 28-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 2) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 3) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 0, 12, 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 28-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 4) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 5) << 12,
+    };
+    shifts = simd_batch{ 32, 0, 24, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 28-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 6) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 16, 0, 8, 36 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 28-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 8) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 0, 28, 0, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 28-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 9) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 10) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 0, 12, 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 28-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 11) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 12) << 12,
+    };
+    shifts = simd_batch{ 32, 0, 24, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 28-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 13) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 16, 0, 8, 36 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 28-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 15) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+    };
+    shifts = simd_batch{ 0, 28, 0, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 28-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 16) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 17) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+    };
+    shifts = simd_batch{ 0, 12, 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 28-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 18) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 19) << 12,
+    };
+    shifts = simd_batch{ 32, 0, 24, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 28-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 20) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+    };
+    shifts = simd_batch{ 16, 0, 8, 36 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 28-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 22) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+    };
+    shifts = simd_batch{ 0, 28, 0, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 28-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 23) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 24) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+    };
+    shifts = simd_batch{ 0, 12, 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 28-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 25) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 26) << 12,
+    };
+    shifts = simd_batch{ 32, 0, 24, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 28-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 27) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+    };
+    shifts = simd_batch{ 16, 0, 8, 36 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 28 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 29> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1fffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 29-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 1) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 29, 0, 23 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 29-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 2) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 3) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 0, 17, 0, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 29-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 4) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 5) << 1,
+    };
+    shifts = simd_batch{ 0, 5, 34, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 29-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 6) << 7,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 7) << 13,
+    };
+    shifts = simd_batch{ 28, 0, 22, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 29-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 8) << 19,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 9) << 25,
+    };
+    shifts = simd_batch{ 16, 0, 10, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 29-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 10) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 4, 33, 0, 27 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 29-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 11) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 12) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 0, 21, 0, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 29-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 13) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 14) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 0, 9, 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 29-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 15) << 3,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 16) << 9,
+    };
+    shifts = simd_batch{ 32, 0, 26, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 29-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 17) << 15,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 18) << 21,
+    };
+    shifts = simd_batch{ 20, 0, 14, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 29-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 19) << 27,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+    };
+    shifts = simd_batch{ 8, 0, 2, 31 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 29-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 20) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 21) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+    };
+    shifts = simd_batch{ 0, 25, 0, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 29-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 22) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 23) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+    };
+    shifts = simd_batch{ 0, 13, 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 29-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 24) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 25) << 5,
+    };
+    shifts = simd_batch{ 0, 1, 30, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 29-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 26) << 11,
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 27) << 17,
+    };
+    shifts = simd_batch{ 24, 0, 18, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 29-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 28) << 23,
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+    };
+    shifts = simd_batch{ 12, 0, 6, 35 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 29 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 30> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3fffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 30-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 1) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 30, 0, 26 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 30-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 2) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 3) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 0, 22, 0, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 30-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 4) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 5) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 0, 14, 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 30-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 6) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 7) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 0, 6, 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 30-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 8) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 9) << 6,
+    };
+    shifts = simd_batch{ 32, 0, 28, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 30-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 10) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 11) << 14,
+    };
+    shifts = simd_batch{ 24, 0, 20, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 30-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 12) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 13) << 22,
+    };
+    shifts = simd_batch{ 16, 0, 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 30-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 14) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 8, 0, 4, 34 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 30-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 16) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+    };
+    shifts = simd_batch{ 0, 30, 0, 26 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 30-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 17) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 18) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+    };
+    shifts = simd_batch{ 0, 22, 0, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 30-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 19) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 20) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+    };
+    shifts = simd_batch{ 0, 14, 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 30-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 21) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 22) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+    };
+    shifts = simd_batch{ 0, 6, 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 30-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 23) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 24) << 6,
+    };
+    shifts = simd_batch{ 32, 0, 28, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 30-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 25) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 26) << 14,
+    };
+    shifts = simd_batch{ 24, 0, 20, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 30-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 27) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 28) << 22,
+    };
+    shifts = simd_batch{ 16, 0, 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 30-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 29) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 29),
+      SafeLoadAs<uint64_t>(in + 8 * 29),
+    };
+    shifts = simd_batch{ 8, 0, 4, 34 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 30 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 31> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7fffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 31-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 1) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 31, 0, 29 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 31-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 2) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 3) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 0, 27, 0, 25 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 31-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 4) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 5) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 0, 23, 0, 21 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 31-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 6) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 7) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 0, 19, 0, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 31-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 8) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 9) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 0, 15, 0, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 31-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 10) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 11) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 0, 11, 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 31-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 12) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 13) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 0, 7, 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 31-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 14) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 15) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+    };
+    shifts = simd_batch{ 0, 3, 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 31-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 16) << 1,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 17) << 3,
+    };
+    shifts = simd_batch{ 32, 0, 30, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 31-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 18) << 5,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 19) << 7,
+    };
+    shifts = simd_batch{ 28, 0, 26, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 31-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 20) << 9,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 21) << 11,
+    };
+    shifts = simd_batch{ 24, 0, 22, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 31-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 22) << 13,
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 23) << 15,
+    };
+    shifts = simd_batch{ 20, 0, 18, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 31-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 24) << 17,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 25) << 19,
+    };
+    shifts = simd_batch{ 16, 0, 14, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 31-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 26) << 21,
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 27) << 23,
+    };
+    shifts = simd_batch{ 12, 0, 10, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 31-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 28) << 25,
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 29) << 27,
+    };
+    shifts = simd_batch{ 8, 0, 6, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 31-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 29),
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 30) << 29,
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+    };
+    shifts = simd_batch{ 4, 0, 2, 33 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 31 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 32> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0xffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 32-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 32, 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 32-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 0, 32, 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 32-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 0, 32, 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 32-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 0, 32, 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 32-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 0, 32, 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 32-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 0, 32, 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 32-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 0, 32, 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 32-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+    };
+    shifts = simd_batch{ 0, 32, 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 32-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+    };
+    shifts = simd_batch{ 0, 32, 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 32-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+    };
+    shifts = simd_batch{ 0, 32, 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 32-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+    };
+    shifts = simd_batch{ 0, 32, 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 32-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+    };
+    shifts = simd_batch{ 0, 32, 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 32-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+    };
+    shifts = simd_batch{ 0, 32, 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 32-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+    };
+    shifts = simd_batch{ 0, 32, 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 32-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 29),
+      SafeLoadAs<uint64_t>(in + 8 * 29),
+    };
+    shifts = simd_batch{ 0, 32, 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 32-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+      SafeLoadAs<uint64_t>(in + 8 * 31),
+      SafeLoadAs<uint64_t>(in + 8 * 31),
+    };
+    shifts = simd_batch{ 0, 32, 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 32 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 33> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1ffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 33-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 1) << 31,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 2) << 29,
+    };
+    shifts = simd_batch{ 0, 0, 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 33-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 3) << 27,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 4) << 25,
+    };
+    shifts = simd_batch{ 4, 0, 6, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 33-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 5) << 23,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 6) << 21,
+    };
+    shifts = simd_batch{ 8, 0, 10, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 33-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 7) << 19,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 8) << 17,
+    };
+    shifts = simd_batch{ 12, 0, 14, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 33-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 9) << 15,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 10) << 13,
+    };
+    shifts = simd_batch{ 16, 0, 18, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 33-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 11) << 11,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 12) << 9,
+    };
+    shifts = simd_batch{ 20, 0, 22, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 33-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 13) << 7,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 14) << 5,
+    };
+    shifts = simd_batch{ 24, 0, 26, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 33-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 15) << 3,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 16) << 1,
+    };
+    shifts = simd_batch{ 28, 0, 30, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 33-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 17) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 18) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+    };
+    shifts = simd_batch{ 0, 1, 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 33-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 19) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 20) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+    };
+    shifts = simd_batch{ 0, 5, 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 33-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 21) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 22) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+    };
+    shifts = simd_batch{ 0, 9, 0, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 33-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 23) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 24) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+    };
+    shifts = simd_batch{ 0, 13, 0, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 33-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 25) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 26) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+    };
+    shifts = simd_batch{ 0, 17, 0, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 33-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 27) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 28) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+    };
+    shifts = simd_batch{ 0, 21, 0, 23 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 33-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 29) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 29),
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 30) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+    };
+    shifts = simd_batch{ 0, 25, 0, 27 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 33-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 31) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 31),
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 32) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+    };
+    shifts = simd_batch{ 0, 29, 0, 31 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 33 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 34> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3ffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 34-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 1) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 2) << 26,
+    };
+    shifts = simd_batch{ 0, 0, 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 34-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 3) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 4) << 18,
+    };
+    shifts = simd_batch{ 8, 0, 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 34-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 5) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 6) << 10,
+    };
+    shifts = simd_batch{ 16, 0, 20, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 34-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 7) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 8) << 2,
+    };
+    shifts = simd_batch{ 24, 0, 28, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 34-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 9) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 10) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 0, 2, 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 34-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 11) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 12) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 0, 10, 0, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 34-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 13) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 14) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 0, 18, 0, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 34-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 15) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 16) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+    };
+    shifts = simd_batch{ 0, 26, 0, 30 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 34-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 18) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 19) << 26,
+    };
+    shifts = simd_batch{ 0, 0, 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 34-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 20) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 21) << 18,
+    };
+    shifts = simd_batch{ 8, 0, 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 34-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 22) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 23) << 10,
+    };
+    shifts = simd_batch{ 16, 0, 20, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 34-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 24) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 25) << 2,
+    };
+    shifts = simd_batch{ 24, 0, 28, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 34-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 26) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 27) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+    };
+    shifts = simd_batch{ 0, 2, 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 34-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 28) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 29) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 29),
+    };
+    shifts = simd_batch{ 0, 10, 0, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 34-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 30) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 31) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 31),
+    };
+    shifts = simd_batch{ 0, 18, 0, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 34-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 32) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 33) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 33),
+    };
+    shifts = simd_batch{ 0, 26, 0, 30 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 34 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 35> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7ffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 35-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 1) << 29,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 2) << 23,
+    };
+    shifts = simd_batch{ 0, 0, 6, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 35-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 3) << 17,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 4) << 11,
+    };
+    shifts = simd_batch{ 12, 0, 18, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 35-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 5) << 5,
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 6) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 24, 0, 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 35-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 7) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 8) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 0, 7, 0, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 35-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 9) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 10) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 0, 19, 0, 25 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 35-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 11) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 31 | SafeLoadAs<uint64_t>(in + 8 * 12) << 33,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 13) << 27,
+    };
+    shifts = simd_batch{ 0, 0, 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 35-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 14) << 21,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 15) << 15,
+    };
+    shifts = simd_batch{ 8, 0, 14, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 35-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 16) << 9,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 17) << 3,
+    };
+    shifts = simd_batch{ 20, 0, 26, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 35-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 18) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 19) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+    };
+    shifts = simd_batch{ 0, 3, 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 35-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 20) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 21) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+    };
+    shifts = simd_batch{ 0, 15, 0, 21 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 35-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 22) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 23) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 24) << 31,
+    };
+    shifts = simd_batch{ 0, 27, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 35-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 25) << 25,
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 26) << 19,
+    };
+    shifts = simd_batch{ 4, 0, 10, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 35-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 27) << 13,
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 28) << 7,
+    };
+    shifts = simd_batch{ 16, 0, 22, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 35-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 29) << 1,
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 30) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+    };
+    shifts = simd_batch{ 28, 0, 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 35-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 31) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 31),
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 32) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+    };
+    shifts = simd_batch{ 0, 11, 0, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 35-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 33) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 33),
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 34) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 34),
+    };
+    shifts = simd_batch{ 0, 23, 0, 29 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 35 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 36> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0xfffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 36-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 1) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 2) << 20,
+    };
+    shifts = simd_batch{ 0, 0, 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 36-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 3) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 4) << 4,
+    };
+    shifts = simd_batch{ 16, 0, 24, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 36-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 5) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 6) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 0, 4, 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 36-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 7) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 8) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 0, 20, 0, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 36-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 10) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 11) << 20,
+    };
+    shifts = simd_batch{ 0, 0, 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 36-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 12) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 13) << 4,
+    };
+    shifts = simd_batch{ 16, 0, 24, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 36-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 14) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 15) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+    };
+    shifts = simd_batch{ 0, 4, 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 36-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 16) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 17) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+    };
+    shifts = simd_batch{ 0, 20, 0, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 36-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 19) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 20) << 20,
+    };
+    shifts = simd_batch{ 0, 0, 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 36-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 21) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 22) << 4,
+    };
+    shifts = simd_batch{ 16, 0, 24, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 36-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 23) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 24) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+    };
+    shifts = simd_batch{ 0, 4, 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 36-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 25) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 26) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+    };
+    shifts = simd_batch{ 0, 20, 0, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 36-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 28) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 29) << 20,
+    };
+    shifts = simd_batch{ 0, 0, 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 36-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 29),
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 30) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 31) << 4,
+    };
+    shifts = simd_batch{ 16, 0, 24, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 36-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 32) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 33) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 33),
+    };
+    shifts = simd_batch{ 0, 4, 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 36-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 34) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 34),
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 35) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+    };
+    shifts = simd_batch{ 0, 20, 0, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 36 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 37> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1fffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 37-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 1) << 27,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 2) << 17,
+    };
+    shifts = simd_batch{ 0, 0, 10, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 37-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 3) << 7,
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 4) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 20, 0, 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 37-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 5) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 6) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 0, 13, 0, 23 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 37-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 7) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 8) << 31,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 9) << 21,
+    };
+    shifts = simd_batch{ 0, 0, 6, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 37-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 10) << 11,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 11) << 1,
+    };
+    shifts = simd_batch{ 16, 0, 26, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 37-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 12) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 13) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 0, 9, 0, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 37-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 14) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 29 | SafeLoadAs<uint64_t>(in + 8 * 15) << 35,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 16) << 25,
+    };
+    shifts = simd_batch{ 0, 0, 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 37-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 17) << 15,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 18) << 5,
+    };
+    shifts = simd_batch{ 12, 0, 22, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 37-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 19) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 20) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+    };
+    shifts = simd_batch{ 0, 5, 0, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 37-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 21) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 22) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 23) << 29,
+    };
+    shifts = simd_batch{ 0, 25, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 37-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 24) << 19,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 25) << 9,
+    };
+    shifts = simd_batch{ 8, 0, 18, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 37-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 26) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 27) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+    };
+    shifts = simd_batch{ 0, 1, 0, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 37-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 28) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 29) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 31 | SafeLoadAs<uint64_t>(in + 8 * 30) << 33,
+    };
+    shifts = simd_batch{ 0, 21, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 37-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 31) << 23,
+      SafeLoadAs<uint64_t>(in + 8 * 31),
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 32) << 13,
+    };
+    shifts = simd_batch{ 4, 0, 14, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 37-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 33) << 3,
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 34) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 34),
+    };
+    shifts = simd_batch{ 24, 0, 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 37-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 35) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 36) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 36),
+    };
+    shifts = simd_batch{ 0, 17, 0, 27 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 37 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 38> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3fffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 38-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 1) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 2) << 14,
+    };
+    shifts = simd_batch{ 0, 0, 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 38-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 3) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 4) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 24, 0, 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 38-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 5) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 6) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 7) << 30,
+    };
+    shifts = simd_batch{ 0, 22, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 38-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 8) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 9) << 6,
+    };
+    shifts = simd_batch{ 8, 0, 20, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 38-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 10) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 11) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 0, 6, 0, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 38-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 12) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 13) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 14) << 22,
+    };
+    shifts = simd_batch{ 0, 0, 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 38-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 15) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 16) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+    };
+    shifts = simd_batch{ 16, 0, 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 38-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 17) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 18) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+    };
+    shifts = simd_batch{ 0, 14, 0, 26 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 38-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 20) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 21) << 14,
+    };
+    shifts = simd_batch{ 0, 0, 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 38-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 22) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 23) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+    };
+    shifts = simd_batch{ 24, 0, 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 38-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 24) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 25) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 26) << 30,
+    };
+    shifts = simd_batch{ 0, 22, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 38-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 27) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 28) << 6,
+    };
+    shifts = simd_batch{ 8, 0, 20, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 38-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 29) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 29),
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 30) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+    };
+    shifts = simd_batch{ 0, 6, 0, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 38-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 31) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 32) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 33) << 22,
+    };
+    shifts = simd_batch{ 0, 0, 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 38-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 33),
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 34) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 35) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+    };
+    shifts = simd_batch{ 16, 0, 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 38-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 36) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 36),
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 37) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 37),
+    };
+    shifts = simd_batch{ 0, 14, 0, 26 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 38 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 39> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7fffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 39-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 1) << 25,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 2) << 11,
+    };
+    shifts = simd_batch{ 0, 0, 14, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 39-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 3) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 4) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 0, 3, 0, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 39-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 5) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 31 | SafeLoadAs<uint64_t>(in + 8 * 6) << 33,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 7) << 19,
+    };
+    shifts = simd_batch{ 0, 0, 6, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 39-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 8) << 5,
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 9) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 20, 0, 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 39-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 10) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 11) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 12) << 27,
+    };
+    shifts = simd_batch{ 0, 23, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 39-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 13) << 13,
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 14) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 12, 0, 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 39-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 15) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 16) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 29 | SafeLoadAs<uint64_t>(in + 8 * 17) << 35,
+    };
+    shifts = simd_batch{ 0, 15, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 39-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 18) << 21,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 19) << 7,
+    };
+    shifts = simd_batch{ 4, 0, 18, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 39-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 20) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 21) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+    };
+    shifts = simd_batch{ 0, 7, 0, 21 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 39-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 22) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 23) << 29,
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 24) << 15,
+    };
+    shifts = simd_batch{ 0, 0, 10, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 39-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 25) << 1,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 26) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+    };
+    shifts = simd_batch{ 24, 0, 0, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 39-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 27) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 27 | SafeLoadAs<uint64_t>(in + 8 * 28) << 37,
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 29) << 23,
+    };
+    shifts = simd_batch{ 0, 0, 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 39-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 29),
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 30) << 9,
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 31) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 31),
+    };
+    shifts = simd_batch{ 16, 0, 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 39-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 32) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 33) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 34) << 31,
+    };
+    shifts = simd_batch{ 0, 19, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 39-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 34),
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 35) << 17,
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 36) << 3,
+    };
+    shifts = simd_batch{ 8, 0, 22, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 39-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 37) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 37),
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 38) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 38),
+    };
+    shifts = simd_batch{ 0, 11, 0, 25 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 39 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 40> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0xffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 40-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 1) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 2) << 8,
+    };
+    shifts = simd_batch{ 0, 0, 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 40-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 3) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 4) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 0, 8, 0, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 40-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 6) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 7) << 8,
+    };
+    shifts = simd_batch{ 0, 0, 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 40-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 8) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 9) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 0, 8, 0, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 40-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 11) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 12) << 8,
+    };
+    shifts = simd_batch{ 0, 0, 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 40-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 13) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 14) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 0, 8, 0, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 40-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 16) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 17) << 8,
+    };
+    shifts = simd_batch{ 0, 0, 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 40-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 18) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 19) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+    };
+    shifts = simd_batch{ 0, 8, 0, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 40-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 21) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 22) << 8,
+    };
+    shifts = simd_batch{ 0, 0, 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 40-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 23) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 24) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+    };
+    shifts = simd_batch{ 0, 8, 0, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 40-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 26) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 27) << 8,
+    };
+    shifts = simd_batch{ 0, 0, 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 40-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 28) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 29) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 29),
+    };
+    shifts = simd_batch{ 0, 8, 0, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 40-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 31) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 31),
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 32) << 8,
+    };
+    shifts = simd_batch{ 0, 0, 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 40-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 33) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 33),
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 34) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 34),
+    };
+    shifts = simd_batch{ 0, 8, 0, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 40-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 36) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 36),
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 37) << 8,
+    };
+    shifts = simd_batch{ 0, 0, 16, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 40-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 38) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 38),
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 39) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 39),
+    };
+    shifts = simd_batch{ 0, 8, 0, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 40 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 41> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1ffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 41-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 1) << 23,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 2) << 5,
+    };
+    shifts = simd_batch{ 0, 0, 18, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 41-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 3) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 4) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 31 | SafeLoadAs<uint64_t>(in + 8 * 5) << 33,
+    };
+    shifts = simd_batch{ 0, 13, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 41-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 6) << 15,
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 7) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 8, 0, 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 41-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 8) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 9) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 10) << 25,
+    };
+    shifts = simd_batch{ 0, 21, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 41-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 11) << 7,
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 12) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 16, 0, 0, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 41-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 13) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 29 | SafeLoadAs<uint64_t>(in + 8 * 14) << 35,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 15) << 17,
+    };
+    shifts = simd_batch{ 0, 0, 6, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 41-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 16) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 17) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+    };
+    shifts = simd_batch{ 0, 1, 0, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 41-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 18) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 19) << 27,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 20) << 9,
+    };
+    shifts = simd_batch{ 0, 0, 14, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 41-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 21) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 22) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 27 | SafeLoadAs<uint64_t>(in + 8 * 23) << 37,
+    };
+    shifts = simd_batch{ 0, 9, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 41-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 24) << 19,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 25) << 1,
+    };
+    shifts = simd_batch{ 4, 0, 22, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 41-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 26) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 27) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 28) << 29,
+    };
+    shifts = simd_batch{ 0, 17, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 41-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 29) << 11,
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 30) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+    };
+    shifts = simd_batch{ 12, 0, 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 41-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 31) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 25 | SafeLoadAs<uint64_t>(in + 8 * 32) << 39,
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 33) << 21,
+    };
+    shifts = simd_batch{ 0, 0, 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 41-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 33),
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 34) << 3,
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 35) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+    };
+    shifts = simd_batch{ 20, 0, 0, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 41-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 36) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 37) << 31,
+      SafeLoadAs<uint64_t>(in + 8 * 37),
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 38) << 13,
+    };
+    shifts = simd_batch{ 0, 0, 10, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 41-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 39) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 39),
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 40) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 40),
+    };
+    shifts = simd_batch{ 0, 5, 0, 23 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 41 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 42> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3ffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 42-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 1) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 2) << 2,
+    };
+    shifts = simd_batch{ 0, 0, 20, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 42-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 3) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 4) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 5) << 26,
+    };
+    shifts = simd_batch{ 0, 18, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 42-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 6) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 7) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 16, 0, 0, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 42-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 8) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 9) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 10) << 10,
+    };
+    shifts = simd_batch{ 0, 0, 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 42-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 11) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 12) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 13) << 34,
+    };
+    shifts = simd_batch{ 0, 10, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 42-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 14) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 15) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+    };
+    shifts = simd_batch{ 8, 0, 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 42-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 16) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 17) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 18) << 18,
+    };
+    shifts = simd_batch{ 0, 0, 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 42-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 19) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 20) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+    };
+    shifts = simd_batch{ 0, 2, 0, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 42-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 22) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 23) << 2,
+    };
+    shifts = simd_batch{ 0, 0, 20, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 42-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 24) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 25) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 26) << 26,
+    };
+    shifts = simd_batch{ 0, 18, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 42-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 27) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 28) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+    };
+    shifts = simd_batch{ 16, 0, 0, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 42-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 29) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 30) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 31) << 10,
+    };
+    shifts = simd_batch{ 0, 0, 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 42-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 32) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 33) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 34) << 34,
+    };
+    shifts = simd_batch{ 0, 10, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 42-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 34),
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 35) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 36) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 36),
+    };
+    shifts = simd_batch{ 8, 0, 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 42-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 37) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 38) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 38),
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 39) << 18,
+    };
+    shifts = simd_batch{ 0, 0, 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 42-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 40) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 40),
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 41) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 41),
+    };
+    shifts = simd_batch{ 0, 2, 0, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 42 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 43> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7ffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 43-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 1) << 21,
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 2) << 42,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 0, 0, 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 43-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 3) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 23 | SafeLoadAs<uint64_t>(in + 8 * 4) << 41,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 5) << 19,
+    };
+    shifts = simd_batch{ 0, 0, 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 43-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 6) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 7) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 25 | SafeLoadAs<uint64_t>(in + 8 * 8) << 39,
+    };
+    shifts = simd_batch{ 0, 3, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 43-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 9) << 17,
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 10) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 4, 0, 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 43-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 11) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 27 | SafeLoadAs<uint64_t>(in + 8 * 12) << 37,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 13) << 15,
+    };
+    shifts = simd_batch{ 0, 0, 6, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 43-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 14) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 15) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 29 | SafeLoadAs<uint64_t>(in + 8 * 16) << 35,
+    };
+    shifts = simd_batch{ 0, 7, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 43-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 17) << 13,
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 18) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+    };
+    shifts = simd_batch{ 8, 0, 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 43-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 19) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 31 | SafeLoadAs<uint64_t>(in + 8 * 20) << 33,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 21) << 11,
+    };
+    shifts = simd_batch{ 0, 0, 10, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 43-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 22) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 23) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 24) << 31,
+    };
+    shifts = simd_batch{ 0, 11, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 43-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 25) << 9,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 26) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+    };
+    shifts = simd_batch{ 12, 0, 0, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 43-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 27) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 28) << 29,
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 29) << 7,
+    };
+    shifts = simd_batch{ 0, 0, 14, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 43-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 30) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 31) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 32) << 27,
+    };
+    shifts = simd_batch{ 0, 15, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 43-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 33) << 5,
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 34) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 34),
+    };
+    shifts = simd_batch{ 16, 0, 0, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 43-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 35) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 36) << 25,
+      SafeLoadAs<uint64_t>(in + 8 * 36),
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 37) << 3,
+    };
+    shifts = simd_batch{ 0, 0, 18, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 43-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 38) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 38),
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 39) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 40) << 23,
+    };
+    shifts = simd_batch{ 0, 19, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 43-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 40),
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 41) << 1,
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 42) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 42),
+    };
+    shifts = simd_batch{ 20, 0, 0, 21 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 43 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 44> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0xfffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 44-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 1) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 2) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 0, 0, 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 44-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 3) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 4) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 5) << 12,
+    };
+    shifts = simd_batch{ 0, 0, 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 44-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 6) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 7) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 8) << 28,
+    };
+    shifts = simd_batch{ 0, 12, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 44-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 9) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 10) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 16, 0, 0, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 44-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 12) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 13) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 0, 0, 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 44-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 14) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 15) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 16) << 12,
+    };
+    shifts = simd_batch{ 0, 0, 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 44-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 17) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 18) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 19) << 28,
+    };
+    shifts = simd_batch{ 0, 12, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 44-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 20) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 21) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+    };
+    shifts = simd_batch{ 16, 0, 0, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 44-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 23) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 24) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+    };
+    shifts = simd_batch{ 0, 0, 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 44-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 25) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 26) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 27) << 12,
+    };
+    shifts = simd_batch{ 0, 0, 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 44-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 28) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 29) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 30) << 28,
+    };
+    shifts = simd_batch{ 0, 12, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 44-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 31) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 32) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+    };
+    shifts = simd_batch{ 16, 0, 0, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 44-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 33),
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 34) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 35) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+    };
+    shifts = simd_batch{ 0, 0, 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 44-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 36) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 37) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 37),
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 38) << 12,
+    };
+    shifts = simd_batch{ 0, 0, 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 44-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 39) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 39),
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 40) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 41) << 28,
+    };
+    shifts = simd_batch{ 0, 12, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 44-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 41),
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 42) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 43) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 43),
+    };
+    shifts = simd_batch{ 16, 0, 0, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 44 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 45> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1fffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 45-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 1) << 19,
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 2) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 0, 0, 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 45-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 3) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 4) << 31,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 5) << 5,
+    };
+    shifts = simd_batch{ 0, 0, 14, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 45-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 6) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 21 | SafeLoadAs<uint64_t>(in + 8 * 7) << 43,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 8) << 17,
+    };
+    shifts = simd_batch{ 0, 0, 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 45-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 9) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 10) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 11) << 29,
+    };
+    shifts = simd_batch{ 0, 9, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 45-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 12) << 3,
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 13) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 23 | SafeLoadAs<uint64_t>(in + 8 * 14) << 41,
+    };
+    shifts = simd_batch{ 16, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 45-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 15) << 15,
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 16) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+    };
+    shifts = simd_batch{ 4, 0, 0, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 45-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 17) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 18) << 27,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 19) << 1,
+    };
+    shifts = simd_batch{ 0, 0, 18, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 45-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 20) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 25 | SafeLoadAs<uint64_t>(in + 8 * 21) << 39,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 22) << 13,
+    };
+    shifts = simd_batch{ 0, 0, 6, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 45-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 23) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 24) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 25) << 25,
+    };
+    shifts = simd_batch{ 0, 13, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 45-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 26) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 27) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 27 | SafeLoadAs<uint64_t>(in + 8 * 28) << 37,
+    };
+    shifts = simd_batch{ 0, 1, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 45-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 29) << 11,
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 30) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+    };
+    shifts = simd_batch{ 8, 0, 0, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 45-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 31) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 32) << 23,
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 33) << 42,
+      SafeLoadAs<uint64_t>(in + 8 * 33),
+    };
+    shifts = simd_batch{ 0, 0, 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 45-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 34) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 29 | SafeLoadAs<uint64_t>(in + 8 * 35) << 35,
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 36) << 9,
+    };
+    shifts = simd_batch{ 0, 0, 10, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 45-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 37) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 37),
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 38) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 39) << 21,
+    };
+    shifts = simd_batch{ 0, 17, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 45-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 40) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 40),
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 41) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 31 | SafeLoadAs<uint64_t>(in + 8 * 42) << 33,
+    };
+    shifts = simd_batch{ 0, 5, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 45-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 42),
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 43) << 7,
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 44) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 44),
+    };
+    shifts = simd_batch{ 12, 0, 0, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 45 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 46> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3fffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 46-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 1) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 2) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 0, 0, 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 46-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 3) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 4) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 5) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 0, 0, 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 46-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 6) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 7) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 8) << 6,
+    };
+    shifts = simd_batch{ 0, 0, 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 46-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 9) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 10) << 42,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 11) << 14,
+    };
+    shifts = simd_batch{ 0, 0, 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 46-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 12) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 13) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 14) << 22,
+    };
+    shifts = simd_batch{ 0, 14, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 46-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 15) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 16) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 17) << 30,
+    };
+    shifts = simd_batch{ 0, 6, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 46-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 18) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 19) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 20) << 38,
+    };
+    shifts = simd_batch{ 16, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 46-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 21) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 22) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+    };
+    shifts = simd_batch{ 8, 0, 0, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 46-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 24) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 25) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+    };
+    shifts = simd_batch{ 0, 0, 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 46-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 26) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 27) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 28) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+    };
+    shifts = simd_batch{ 0, 0, 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 46-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 29) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 30) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 31) << 6,
+    };
+    shifts = simd_batch{ 0, 0, 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 46-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 32) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 33) << 42,
+      SafeLoadAs<uint64_t>(in + 8 * 33),
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 34) << 14,
+    };
+    shifts = simd_batch{ 0, 0, 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 46-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 35) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 36) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 37) << 22,
+    };
+    shifts = simd_batch{ 0, 14, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 46-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 38) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 38),
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 39) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 40) << 30,
+    };
+    shifts = simd_batch{ 0, 6, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 46-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 40),
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 41) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 42) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 43) << 38,
+    };
+    shifts = simd_batch{ 16, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 46-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 43),
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 44) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 45) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 45),
+    };
+    shifts = simd_batch{ 8, 0, 0, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 46 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 47> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7fffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 47-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 1) << 17,
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 2) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 0, 0, 0, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 47-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 3) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 4) << 21,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 5) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 0, 0, 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 47-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 6) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 7) << 25,
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 8) << 42,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 0, 0, 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 47-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 9) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 10) << 29,
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 11) << 46,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 0, 0, 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 47-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 12) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 31 | SafeLoadAs<uint64_t>(in + 8 * 13) << 33,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 14) << 3,
+    };
+    shifts = simd_batch{ 0, 0, 14, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 47-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 15) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 27 | SafeLoadAs<uint64_t>(in + 8 * 16) << 37,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 17) << 7,
+    };
+    shifts = simd_batch{ 0, 0, 10, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 47-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 18) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 23 | SafeLoadAs<uint64_t>(in + 8 * 19) << 41,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 20) << 11,
+    };
+    shifts = simd_batch{ 0, 0, 6, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 47-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 21) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 19 | SafeLoadAs<uint64_t>(in + 8 * 22) << 45,
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 23) << 15,
+    };
+    shifts = simd_batch{ 0, 0, 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 47-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 24) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 25) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 26) << 19,
+    };
+    shifts = simd_batch{ 0, 15, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 47-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 27) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 28) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 29) << 23,
+    };
+    shifts = simd_batch{ 0, 11, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 47-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 30) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 31) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 32) << 27,
+    };
+    shifts = simd_batch{ 0, 7, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 47-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 33) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 33),
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 34) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 35) << 31,
+    };
+    shifts = simd_batch{ 0, 3, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 47-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 36) << 1,
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 37) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 29 | SafeLoadAs<uint64_t>(in + 8 * 38) << 35,
+    };
+    shifts = simd_batch{ 16, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 47-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 38),
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 39) << 5,
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 40) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 25 | SafeLoadAs<uint64_t>(in + 8 * 41) << 39,
+    };
+    shifts = simd_batch{ 12, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 47-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 41),
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 42) << 9,
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 43) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 21 | SafeLoadAs<uint64_t>(in + 8 * 44) << 43,
+    };
+    shifts = simd_batch{ 8, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 47-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 44),
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 45) << 13,
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 46) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 46),
+    };
+    shifts = simd_batch{ 4, 0, 0, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 47 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 48> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0xffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 48-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 1) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 2) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 0, 0, 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 48-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 4) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 5) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 0, 0, 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 48-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 7) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 8) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 0, 0, 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 48-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 10) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 11) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 0, 0, 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 48-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 13) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 14) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 0, 0, 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 48-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 16) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 17) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+    };
+    shifts = simd_batch{ 0, 0, 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 48-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 19) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 20) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+    };
+    shifts = simd_batch{ 0, 0, 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 48-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 22) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 23) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+    };
+    shifts = simd_batch{ 0, 0, 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 48-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 25) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 26) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+    };
+    shifts = simd_batch{ 0, 0, 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 48-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 28) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 29) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 29),
+    };
+    shifts = simd_batch{ 0, 0, 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 48-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 31) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 32) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+    };
+    shifts = simd_batch{ 0, 0, 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 48-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 33),
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 34) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 35) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+    };
+    shifts = simd_batch{ 0, 0, 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 48-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 36),
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 37) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 38) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 38),
+    };
+    shifts = simd_batch{ 0, 0, 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 48-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 39),
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 40) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 41) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 41),
+    };
+    shifts = simd_batch{ 0, 0, 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 48-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 42),
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 43) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 44) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 44),
+    };
+    shifts = simd_batch{ 0, 0, 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 48-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 45),
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 46) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 47) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 47),
+    };
+    shifts = simd_batch{ 0, 0, 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 48 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 49> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1ffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 49-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 1) << 15,
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 2) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 19 | SafeLoadAs<uint64_t>(in + 8 * 3) << 45,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 49-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 4) << 11,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 5) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 23 | SafeLoadAs<uint64_t>(in + 8 * 6) << 41,
+    };
+    shifts = simd_batch{ 4, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 49-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 7) << 7,
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 8) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 27 | SafeLoadAs<uint64_t>(in + 8 * 9) << 37,
+    };
+    shifts = simd_batch{ 8, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 49-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 10) << 3,
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 11) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 31 | SafeLoadAs<uint64_t>(in + 8 * 12) << 33,
+    };
+    shifts = simd_batch{ 12, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 49-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 13) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 14) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 15) << 29,
+    };
+    shifts = simd_batch{ 0, 1, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 49-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 16) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 17) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 18) << 25,
+    };
+    shifts = simd_batch{ 0, 5, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 49-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 19) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 20) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 21) << 21,
+    };
+    shifts = simd_batch{ 0, 9, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 49-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 22) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 23) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 24) << 17,
+    };
+    shifts = simd_batch{ 0, 13, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 49-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 25) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 17 | SafeLoadAs<uint64_t>(in + 8 * 26) << 47,
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 27) << 13,
+    };
+    shifts = simd_batch{ 0, 0, 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 49-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 28) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 21 | SafeLoadAs<uint64_t>(in + 8 * 29) << 43,
+      SafeLoadAs<uint64_t>(in + 8 * 29),
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 30) << 9,
+    };
+    shifts = simd_batch{ 0, 0, 6, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 49-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 31) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 25 | SafeLoadAs<uint64_t>(in + 8 * 32) << 39,
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 33) << 5,
+    };
+    shifts = simd_batch{ 0, 0, 10, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 49-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 34) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 29 | SafeLoadAs<uint64_t>(in + 8 * 35) << 35,
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 36) << 1,
+    };
+    shifts = simd_batch{ 0, 0, 14, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 49-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 37) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 38) << 31,
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 39) << 46,
+      SafeLoadAs<uint64_t>(in + 8 * 39),
+    };
+    shifts = simd_batch{ 0, 0, 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 49-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 40) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 41) << 27,
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 42) << 42,
+      SafeLoadAs<uint64_t>(in + 8 * 42),
+    };
+    shifts = simd_batch{ 0, 0, 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 49-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 43) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 44) << 23,
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 45) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 45),
+    };
+    shifts = simd_batch{ 0, 0, 0, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 49-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 46) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 47) << 19,
+      SafeLoadAs<uint64_t>(in + 8 * 47) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 48) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 48),
+    };
+    shifts = simd_batch{ 0, 0, 0, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 49 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 50> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3ffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 50-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 1) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 2) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 3) << 42,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 50-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 4) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 5) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 6) << 34,
+    };
+    shifts = simd_batch{ 8, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 50-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 7) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 8) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 9) << 26,
+    };
+    shifts = simd_batch{ 0, 2, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 50-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 10) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 11) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 12) << 18,
+    };
+    shifts = simd_batch{ 0, 10, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 50-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 13) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 14) << 46,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 15) << 10,
+    };
+    shifts = simd_batch{ 0, 0, 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 50-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 16) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 17) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 18) << 2,
+    };
+    shifts = simd_batch{ 0, 0, 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 50-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 19) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 20) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 21) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+    };
+    shifts = simd_batch{ 0, 0, 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 50-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 22) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 23) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 24) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+    };
+    shifts = simd_batch{ 0, 0, 0, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 50-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 26) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 27) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 28) << 42,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 50-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 29) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 30) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 31) << 34,
+    };
+    shifts = simd_batch{ 8, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 50-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 32) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 33) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 34) << 26,
+    };
+    shifts = simd_batch{ 0, 2, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 50-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 35) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 36) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 37) << 18,
+    };
+    shifts = simd_batch{ 0, 10, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 50-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 38) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 39) << 46,
+      SafeLoadAs<uint64_t>(in + 8 * 39),
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 40) << 10,
+    };
+    shifts = simd_batch{ 0, 0, 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 50-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 41) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 42) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 42),
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 43) << 2,
+    };
+    shifts = simd_batch{ 0, 0, 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 50-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 44) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 45) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 46) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 46),
+    };
+    shifts = simd_batch{ 0, 0, 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 50-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 47) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 47) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 48) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 48) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 49) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 49),
+    };
+    shifts = simd_batch{ 0, 0, 0, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 50 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 51> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7ffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 51-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 1) << 13,
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 2) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 25 | SafeLoadAs<uint64_t>(in + 8 * 3) << 39,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 51-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 4) << 1,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 5) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 6) << 27,
+    };
+    shifts = simd_batch{ 12, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 51-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 7) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 8) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 9) << 15,
+    };
+    shifts = simd_batch{ 0, 11, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 51-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 10) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 23 | SafeLoadAs<uint64_t>(in + 8 * 11) << 41,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 12) << 3,
+    };
+    shifts = simd_batch{ 0, 0, 10, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 51-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 13) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 14) << 29,
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 15) << 42,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+    };
+    shifts = simd_batch{ 0, 0, 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 51-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 16) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 17) << 17,
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 18) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 21 | SafeLoadAs<uint64_t>(in + 8 * 19) << 43,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 51-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 20) << 5,
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 21) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 22) << 31,
+    };
+    shifts = simd_batch{ 8, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 51-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 23) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 24) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 25) << 19,
+    };
+    shifts = simd_batch{ 0, 7, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 51-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 26) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 19 | SafeLoadAs<uint64_t>(in + 8 * 27) << 45,
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 28) << 7,
+    };
+    shifts = simd_batch{ 0, 0, 6, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 51-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 29) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 31 | SafeLoadAs<uint64_t>(in + 8 * 30) << 33,
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 31) << 46,
+      SafeLoadAs<uint64_t>(in + 8 * 31),
+    };
+    shifts = simd_batch{ 0, 0, 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 51-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 32) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 33) << 21,
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 34) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 17 | SafeLoadAs<uint64_t>(in + 8 * 35) << 47,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 51-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 36) << 9,
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 37) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 29 | SafeLoadAs<uint64_t>(in + 8 * 38) << 35,
+    };
+    shifts = simd_batch{ 4, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 51-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 39) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 39),
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 40) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 41) << 23,
+    };
+    shifts = simd_batch{ 0, 3, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 51-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 42) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 15 | SafeLoadAs<uint64_t>(in + 8 * 43) << 49,
+      SafeLoadAs<uint64_t>(in + 8 * 43),
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 44) << 11,
+    };
+    shifts = simd_batch{ 0, 0, 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 51-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 45) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 27 | SafeLoadAs<uint64_t>(in + 8 * 46) << 37,
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 14 | SafeLoadAs<uint64_t>(in + 8 * 47) << 50,
+      SafeLoadAs<uint64_t>(in + 8 * 47),
+    };
+    shifts = simd_batch{ 0, 0, 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 51-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 47) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 48) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 48) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 49) << 25,
+      SafeLoadAs<uint64_t>(in + 8 * 49) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 50) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 50),
+    };
+    shifts = simd_batch{ 0, 0, 0, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 51 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 52> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0xfffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 52-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 1) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 2) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 3) << 36,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 52-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 4) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 5) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 6) << 20,
+    };
+    shifts = simd_batch{ 0, 4, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 52-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 7) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 8) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 9) << 4,
+    };
+    shifts = simd_batch{ 0, 0, 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 52-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 10) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 11) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 12) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 0, 0, 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 52-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 14) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 15) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 16) << 36,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 52-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 17) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 18) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 19) << 20,
+    };
+    shifts = simd_batch{ 0, 4, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 52-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 20) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 21) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 22) << 4,
+    };
+    shifts = simd_batch{ 0, 0, 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 52-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 23) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 24) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 25) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+    };
+    shifts = simd_batch{ 0, 0, 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 52-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 27) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 28) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 29) << 36,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 52-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 30) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 31) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 32) << 20,
+    };
+    shifts = simd_batch{ 0, 4, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 52-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 33) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 34) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 34),
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 35) << 4,
+    };
+    shifts = simd_batch{ 0, 0, 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 52-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 36) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 37) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 38) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 38),
+    };
+    shifts = simd_batch{ 0, 0, 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 52-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 39),
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 40) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 41) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 42) << 36,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 52-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 43) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 43),
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 44) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 45) << 20,
+    };
+    shifts = simd_batch{ 0, 4, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 52-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 46) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 47) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 47),
+      SafeLoadAs<uint64_t>(in + 8 * 47) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 48) << 4,
+    };
+    shifts = simd_batch{ 0, 0, 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 52-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 48) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 49) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 49) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 50) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 50) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 51) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 51),
+    };
+    shifts = simd_batch{ 0, 0, 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 52 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 53> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1fffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 53-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 1) << 11,
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 2) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 31 | SafeLoadAs<uint64_t>(in + 8 * 3) << 33,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 53-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 4) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 5) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 6) << 13,
+    };
+    shifts = simd_batch{ 0, 9, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 53-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 7) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 29 | SafeLoadAs<uint64_t>(in + 8 * 8) << 35,
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 9) << 46,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 0, 0, 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 53-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 10) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 11) << 15,
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 12) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 27 | SafeLoadAs<uint64_t>(in + 8 * 13) << 37,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 53-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 14) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 15) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 16) << 17,
+    };
+    shifts = simd_batch{ 0, 5, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 53-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 17) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 25 | SafeLoadAs<uint64_t>(in + 8 * 18) << 39,
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 14 | SafeLoadAs<uint64_t>(in + 8 * 19) << 50,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+    };
+    shifts = simd_batch{ 0, 0, 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 53-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 20) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 21) << 19,
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 22) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 23 | SafeLoadAs<uint64_t>(in + 8 * 23) << 41,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 53-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 24) << 52,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 25) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 26) << 21,
+    };
+    shifts = simd_batch{ 0, 1, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 53-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 27) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 21 | SafeLoadAs<uint64_t>(in + 8 * 28) << 43,
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 29) << 1,
+    };
+    shifts = simd_batch{ 0, 0, 10, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 53-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 30) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 31) << 23,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 32) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 19 | SafeLoadAs<uint64_t>(in + 8 * 33) << 45,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 53-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 33),
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 34) << 3,
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 35) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 36) << 25,
+    };
+    shifts = simd_batch{ 8, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 53-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 37) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 17 | SafeLoadAs<uint64_t>(in + 8 * 38) << 47,
+      SafeLoadAs<uint64_t>(in + 8 * 38),
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 39) << 5,
+    };
+    shifts = simd_batch{ 0, 0, 6, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 53-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 40) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 41) << 27,
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 42) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 15 | SafeLoadAs<uint64_t>(in + 8 * 43) << 49,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 53-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 43),
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 44) << 7,
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 45) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 46) << 29,
+    };
+    shifts = simd_batch{ 4, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 53-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 47) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 47) >> 13 | SafeLoadAs<uint64_t>(in + 8 * 48) << 51,
+      SafeLoadAs<uint64_t>(in + 8 * 48),
+      SafeLoadAs<uint64_t>(in + 8 * 48) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 49) << 9,
+    };
+    shifts = simd_batch{ 0, 0, 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 53-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 49) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 50) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 50) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 51) << 31,
+      SafeLoadAs<uint64_t>(in + 8 * 51) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 52) << 42,
+      SafeLoadAs<uint64_t>(in + 8 * 52),
+    };
+    shifts = simd_batch{ 0, 0, 0, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 53 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 54> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3fffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 54-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 1) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 2) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 3) << 30,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 54-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 4) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 14 | SafeLoadAs<uint64_t>(in + 8 * 5) << 50,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 6) << 6,
+    };
+    shifts = simd_batch{ 0, 0, 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 54-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 7) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 8) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 9) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 10) << 46,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 54-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 11) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 12) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 13) << 22,
+    };
+    shifts = simd_batch{ 8, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 54-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 14) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 15) << 42,
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 16) << 52,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+    };
+    shifts = simd_batch{ 0, 0, 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 54-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 17) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 18) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 19) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 20) << 38,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 54-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 21) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 22) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 23) << 14,
+    };
+    shifts = simd_batch{ 0, 6, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 54-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 24) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 25) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 26) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+    };
+    shifts = simd_batch{ 0, 0, 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 54-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 28) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 29) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 30) << 30,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 54-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 31) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 14 | SafeLoadAs<uint64_t>(in + 8 * 32) << 50,
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 33) << 6,
+    };
+    shifts = simd_batch{ 0, 0, 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 54-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 34) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 35) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 36) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 37) << 46,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 54-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 37),
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 38) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 39) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 40) << 22,
+    };
+    shifts = simd_batch{ 8, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 54-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 41) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 42) << 42,
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 43) << 52,
+      SafeLoadAs<uint64_t>(in + 8 * 43),
+    };
+    shifts = simd_batch{ 0, 0, 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 54-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 44) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 45) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 46) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 47) << 38,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 54-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 47) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 48) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 48),
+      SafeLoadAs<uint64_t>(in + 8 * 48) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 49) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 49) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 50) << 14,
+    };
+    shifts = simd_batch{ 0, 6, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 54-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 50) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 51) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 51) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 52) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 52) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 53) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 53),
+    };
+    shifts = simd_batch{ 0, 0, 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 54 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 55> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7fffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 55-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 1) << 9,
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 2) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 3) << 27,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 55-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 4) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 19 | SafeLoadAs<uint64_t>(in + 8 * 5) << 45,
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 10 | SafeLoadAs<uint64_t>(in + 8 * 6) << 54,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 0, 0, 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 55-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 7) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 8) << 17,
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 9) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 29 | SafeLoadAs<uint64_t>(in + 8 * 10) << 35,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 55-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 11) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 11 | SafeLoadAs<uint64_t>(in + 8 * 12) << 53,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 13) << 7,
+    };
+    shifts = simd_batch{ 0, 0, 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 55-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 14) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 15) << 25,
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 16) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 21 | SafeLoadAs<uint64_t>(in + 8 * 17) << 43,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 55-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 18) << 52,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 19) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 20) << 15,
+    };
+    shifts = simd_batch{ 0, 3, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 55-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 21) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 31 | SafeLoadAs<uint64_t>(in + 8 * 22) << 33,
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 23) << 42,
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 13 | SafeLoadAs<uint64_t>(in + 8 * 24) << 51,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 55-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 25) << 5,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 26) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 27) << 23,
+    };
+    shifts = simd_batch{ 4, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 55-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 28) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 23 | SafeLoadAs<uint64_t>(in + 8 * 29) << 41,
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 14 | SafeLoadAs<uint64_t>(in + 8 * 30) << 50,
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+    };
+    shifts = simd_batch{ 0, 0, 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 55-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 31) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 32) << 13,
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 33) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 34) << 31,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 55-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 35) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 15 | SafeLoadAs<uint64_t>(in + 8 * 36) << 49,
+      SafeLoadAs<uint64_t>(in + 8 * 36),
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 37) << 3,
+    };
+    shifts = simd_batch{ 0, 0, 6, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 55-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 38) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 39) << 21,
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 40) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 25 | SafeLoadAs<uint64_t>(in + 8 * 41) << 39,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 55-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 42) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 42),
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 43) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 44) << 11,
+    };
+    shifts = simd_batch{ 0, 7, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 55-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 45) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 46) << 29,
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 47) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 47) >> 17 | SafeLoadAs<uint64_t>(in + 8 * 48) << 47,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 55-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 48),
+      SafeLoadAs<uint64_t>(in + 8 * 48) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 49) << 1,
+      SafeLoadAs<uint64_t>(in + 8 * 49) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 50) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 50) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 51) << 19,
+    };
+    shifts = simd_batch{ 8, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 55-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 51) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 52) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 52) >> 27 | SafeLoadAs<uint64_t>(in + 8 * 53) << 37,
+      SafeLoadAs<uint64_t>(in + 8 * 53) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 54) << 46,
+      SafeLoadAs<uint64_t>(in + 8 * 54),
+    };
+    shifts = simd_batch{ 0, 0, 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 55 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 56> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0xffffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 56-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 1) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 2) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 3) << 24,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 56-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 4) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 5) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 6) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 0, 0, 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 56-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 8) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 9) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 10) << 24,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 56-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 11) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 12) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 13) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 0, 0, 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 56-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 15) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 16) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 17) << 24,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 56-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 18) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 19) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 20) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+    };
+    shifts = simd_batch{ 0, 0, 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 56-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 22) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 23) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 24) << 24,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 56-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 25) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 26) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 27) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+    };
+    shifts = simd_batch{ 0, 0, 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 56-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 29) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 30) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 31) << 24,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 56-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 32) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 33) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 34) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 34),
+    };
+    shifts = simd_batch{ 0, 0, 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 56-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 36) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 37) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 38) << 24,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 56-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 39) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 40) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 41) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 41),
+    };
+    shifts = simd_batch{ 0, 0, 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 56-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 42),
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 43) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 44) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 45) << 24,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 56-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 46) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 47) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 47) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 48) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 48),
+    };
+    shifts = simd_batch{ 0, 0, 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 56-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 49),
+      SafeLoadAs<uint64_t>(in + 8 * 49) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 50) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 50) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 51) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 51) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 52) << 24,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 56-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 52) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 53) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 53) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 54) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 54) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 55) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 55),
+    };
+    shifts = simd_batch{ 0, 0, 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 56 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 57> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1ffffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 57-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 1) << 7,
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 2) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 3) << 21,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 57-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 4) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 29 | SafeLoadAs<uint64_t>(in + 8 * 5) << 35,
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 6) << 42,
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 15 | SafeLoadAs<uint64_t>(in + 8 * 7) << 49,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 57-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 8 | SafeLoadAs<uint64_t>(in + 8 * 8) << 56,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 9) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 10) << 13,
+    };
+    shifts = simd_batch{ 0, 1, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 57-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 11) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 12) << 27,
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 13) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 23 | SafeLoadAs<uint64_t>(in + 8 * 14) << 41,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 57-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 15) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 9 | SafeLoadAs<uint64_t>(in + 8 * 16) << 55,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 17) << 5,
+    };
+    shifts = simd_batch{ 0, 0, 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 57-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 18) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 19) << 19,
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 20) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 31 | SafeLoadAs<uint64_t>(in + 8 * 21) << 33,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 57-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 22) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 17 | SafeLoadAs<uint64_t>(in + 8 * 23) << 47,
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 10 | SafeLoadAs<uint64_t>(in + 8 * 24) << 54,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+    };
+    shifts = simd_batch{ 0, 0, 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 57-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 25) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 26) << 11,
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 27) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 28) << 25,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 57-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 29) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 25 | SafeLoadAs<uint64_t>(in + 8 * 30) << 39,
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 31) << 46,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 11 | SafeLoadAs<uint64_t>(in + 8 * 32) << 53,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 57-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 33) << 3,
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 34) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 35) << 17,
+    };
+    shifts = simd_batch{ 4, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 57-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 36) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 37) << 31,
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 38) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 19 | SafeLoadAs<uint64_t>(in + 8 * 39) << 45,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 57-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 40) << 52,
+      SafeLoadAs<uint64_t>(in + 8 * 40),
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 41) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 42) << 9,
+    };
+    shifts = simd_batch{ 0, 5, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 57-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 43) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 44) << 23,
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 45) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 27 | SafeLoadAs<uint64_t>(in + 8 * 46) << 37,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 57-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 47) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 47) >> 13 | SafeLoadAs<uint64_t>(in + 8 * 48) << 51,
+      SafeLoadAs<uint64_t>(in + 8 * 48),
+      SafeLoadAs<uint64_t>(in + 8 * 48) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 49) << 1,
+    };
+    shifts = simd_batch{ 0, 0, 6, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 57-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 49) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 50) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 50) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 51) << 15,
+      SafeLoadAs<uint64_t>(in + 8 * 51) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 52) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 52) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 53) << 29,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 57-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 53) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 54) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 54) >> 21 | SafeLoadAs<uint64_t>(in + 8 * 55) << 43,
+      SafeLoadAs<uint64_t>(in + 8 * 55) >> 14 | SafeLoadAs<uint64_t>(in + 8 * 56) << 50,
+      SafeLoadAs<uint64_t>(in + 8 * 56),
+    };
+    shifts = simd_batch{ 0, 0, 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 57 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 58> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3ffffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 58-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 1) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 2) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 3) << 18,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 58-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 4) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 5) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 6) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 7) << 42,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 58-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 8) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 10 | SafeLoadAs<uint64_t>(in + 8 * 9) << 54,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 10) << 2,
+    };
+    shifts = simd_batch{ 0, 0, 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 58-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 11) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 12) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 13) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 14) << 26,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 58-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 15) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 16) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 17) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 14 | SafeLoadAs<uint64_t>(in + 8 * 18) << 50,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 58-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 8 | SafeLoadAs<uint64_t>(in + 8 * 19) << 56,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 20) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 21) << 10,
+    };
+    shifts = simd_batch{ 0, 2, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 58-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 22) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 23) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 24) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 25) << 34,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 58-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 26) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 27) << 46,
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 28) << 52,
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+    };
+    shifts = simd_batch{ 0, 0, 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 58-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 29),
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 30) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 31) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 32) << 18,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 58-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 33) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 34) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 35) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 36) << 42,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 58-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 37) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 10 | SafeLoadAs<uint64_t>(in + 8 * 38) << 54,
+      SafeLoadAs<uint64_t>(in + 8 * 38),
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 39) << 2,
+    };
+    shifts = simd_batch{ 0, 0, 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 58-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 40) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 41) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 42) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 43) << 26,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 58-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 44) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 45) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 46) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 14 | SafeLoadAs<uint64_t>(in + 8 * 47) << 50,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 58-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 47) >> 8 | SafeLoadAs<uint64_t>(in + 8 * 48) << 56,
+      SafeLoadAs<uint64_t>(in + 8 * 48),
+      SafeLoadAs<uint64_t>(in + 8 * 48) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 49) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 49) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 50) << 10,
+    };
+    shifts = simd_batch{ 0, 2, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 58-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 50) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 51) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 51) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 52) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 52) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 53) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 53) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 54) << 34,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 58-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 54) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 55) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 55) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 56) << 46,
+      SafeLoadAs<uint64_t>(in + 8 * 56) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 57) << 52,
+      SafeLoadAs<uint64_t>(in + 8 * 57),
+    };
+    shifts = simd_batch{ 0, 0, 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 58 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 59> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7ffffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 59-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 1) << 5,
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 2) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 3) << 15,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 59-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 4) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 5) << 25,
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 6) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 29 | SafeLoadAs<uint64_t>(in + 8 * 7) << 35,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 59-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 8) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 19 | SafeLoadAs<uint64_t>(in + 8 * 9) << 45,
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 14 | SafeLoadAs<uint64_t>(in + 8 * 10) << 50,
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 9 | SafeLoadAs<uint64_t>(in + 8 * 11) << 55,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 59-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 12) << 1,
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 13) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 14) << 11,
+    };
+    shifts = simd_batch{ 4, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 59-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 15) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 16) << 21,
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 17) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 18) << 31,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 59-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 19) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 23 | SafeLoadAs<uint64_t>(in + 8 * 20) << 41,
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 21) << 46,
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 13 | SafeLoadAs<uint64_t>(in + 8 * 22) << 51,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 59-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 8 | SafeLoadAs<uint64_t>(in + 8 * 23) << 56,
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 24) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 25) << 7,
+    };
+    shifts = simd_batch{ 0, 3, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 59-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 26) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 27) << 17,
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 28) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 29) << 27,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 59-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 30) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 27 | SafeLoadAs<uint64_t>(in + 8 * 31) << 37,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 32) << 42,
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 17 | SafeLoadAs<uint64_t>(in + 8 * 33) << 47,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 59-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 34) << 52,
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 7 | SafeLoadAs<uint64_t>(in + 8 * 35) << 57,
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 36) << 3,
+    };
+    shifts = simd_batch{ 0, 0, 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 59-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 37) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 38) << 13,
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 39) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 40) << 23,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 59-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 41) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 31 | SafeLoadAs<uint64_t>(in + 8 * 42) << 33,
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 43) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 21 | SafeLoadAs<uint64_t>(in + 8 * 44) << 43,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 59-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 45) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 11 | SafeLoadAs<uint64_t>(in + 8 * 46) << 53,
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 6 | SafeLoadAs<uint64_t>(in + 8 * 47) << 58,
+      SafeLoadAs<uint64_t>(in + 8 * 47),
+    };
+    shifts = simd_batch{ 0, 0, 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 59-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 47) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 48) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 48) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 49) << 9,
+      SafeLoadAs<uint64_t>(in + 8 * 49) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 50) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 50) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 51) << 19,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 59-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 51) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 52) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 52) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 53) << 29,
+      SafeLoadAs<uint64_t>(in + 8 * 53) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 54) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 54) >> 25 | SafeLoadAs<uint64_t>(in + 8 * 55) << 39,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 59-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 55) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 56) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 56) >> 15 | SafeLoadAs<uint64_t>(in + 8 * 57) << 49,
+      SafeLoadAs<uint64_t>(in + 8 * 57) >> 10 | SafeLoadAs<uint64_t>(in + 8 * 58) << 54,
+      SafeLoadAs<uint64_t>(in + 8 * 58),
+    };
+    shifts = simd_batch{ 0, 0, 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 59 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 60> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0xfffffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 60-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 1) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 2) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 3) << 12,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 60-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 4) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 5) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 6) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 7) << 28,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 60-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 8) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 9) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 10) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 11) << 44,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 60-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 12) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 13) << 52,
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 8 | SafeLoadAs<uint64_t>(in + 8 * 14) << 56,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 0, 0, 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 60-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 16) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 17) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 18) << 12,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 60-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 19) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 20) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 21) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 22) << 28,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 60-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 23) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 24) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 25) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 26) << 44,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 60-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 27) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 28) << 52,
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 8 | SafeLoadAs<uint64_t>(in + 8 * 29) << 56,
+      SafeLoadAs<uint64_t>(in + 8 * 29),
+    };
+    shifts = simd_batch{ 0, 0, 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 60-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 31) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 32) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 33) << 12,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 60-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 34) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 35) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 36) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 37) << 28,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 60-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 38) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 39) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 40) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 41) << 44,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 60-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 42) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 43) << 52,
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 8 | SafeLoadAs<uint64_t>(in + 8 * 44) << 56,
+      SafeLoadAs<uint64_t>(in + 8 * 44),
+    };
+    shifts = simd_batch{ 0, 0, 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 60-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 45),
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 46) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 47) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 47) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 48) << 12,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 60-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 48) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 49) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 49) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 50) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 50) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 51) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 51) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 52) << 28,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 60-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 52) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 53) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 53) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 54) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 54) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 55) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 55) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 56) << 44,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 60-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 56) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 57) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 57) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 58) << 52,
+      SafeLoadAs<uint64_t>(in + 8 * 58) >> 8 | SafeLoadAs<uint64_t>(in + 8 * 59) << 56,
+      SafeLoadAs<uint64_t>(in + 8 * 59),
+    };
+    shifts = simd_batch{ 0, 0, 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 60 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 61> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1fffffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 61-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 1) << 3,
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 2) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 3) << 9,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 61-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 4) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 5) << 15,
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 6) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 7) << 21,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 61-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 8) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 9) << 27,
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 10) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 31 | SafeLoadAs<uint64_t>(in + 8 * 11) << 33,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 61-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 12) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 25 | SafeLoadAs<uint64_t>(in + 8 * 13) << 39,
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 14) << 42,
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 19 | SafeLoadAs<uint64_t>(in + 8 * 15) << 45,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 61-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 16) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 13 | SafeLoadAs<uint64_t>(in + 8 * 17) << 51,
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 10 | SafeLoadAs<uint64_t>(in + 8 * 18) << 54,
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 7 | SafeLoadAs<uint64_t>(in + 8 * 19) << 57,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 61-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 4 | SafeLoadAs<uint64_t>(in + 8 * 20) << 60,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 21) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 22) << 5,
+    };
+    shifts = simd_batch{ 0, 1, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 61-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 23) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 24) << 11,
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 25) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 26) << 17,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 61-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 27) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 28) << 23,
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 29) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 30) << 29,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 61-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 31) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 29 | SafeLoadAs<uint64_t>(in + 8 * 32) << 35,
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 33) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 23 | SafeLoadAs<uint64_t>(in + 8 * 34) << 41,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 61-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 35) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 17 | SafeLoadAs<uint64_t>(in + 8 * 36) << 47,
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 14 | SafeLoadAs<uint64_t>(in + 8 * 37) << 50,
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 11 | SafeLoadAs<uint64_t>(in + 8 * 38) << 53,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 61-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 8 | SafeLoadAs<uint64_t>(in + 8 * 39) << 56,
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 5 | SafeLoadAs<uint64_t>(in + 8 * 40) << 59,
+      SafeLoadAs<uint64_t>(in + 8 * 40),
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 41) << 1,
+    };
+    shifts = simd_batch{ 0, 0, 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 61-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 42) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 43) << 7,
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 44) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 45) << 13,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 61-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 46) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 47) << 19,
+      SafeLoadAs<uint64_t>(in + 8 * 47) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 48) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 48) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 49) << 25,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 61-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 49) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 50) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 50) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 51) << 31,
+      SafeLoadAs<uint64_t>(in + 8 * 51) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 52) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 52) >> 27 | SafeLoadAs<uint64_t>(in + 8 * 53) << 37,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 61-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 53) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 54) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 54) >> 21 | SafeLoadAs<uint64_t>(in + 8 * 55) << 43,
+      SafeLoadAs<uint64_t>(in + 8 * 55) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 56) << 46,
+      SafeLoadAs<uint64_t>(in + 8 * 56) >> 15 | SafeLoadAs<uint64_t>(in + 8 * 57) << 49,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 61-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 57) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 58) << 52,
+      SafeLoadAs<uint64_t>(in + 8 * 58) >> 9 | SafeLoadAs<uint64_t>(in + 8 * 59) << 55,
+      SafeLoadAs<uint64_t>(in + 8 * 59) >> 6 | SafeLoadAs<uint64_t>(in + 8 * 60) << 58,
+      SafeLoadAs<uint64_t>(in + 8 * 60),
+    };
+    shifts = simd_batch{ 0, 0, 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 61 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 62> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3fffffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 62-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 1) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 2) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 3) << 6,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 62-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 4) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 5) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 6) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 7) << 14,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 62-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 8) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 9) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 10) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 11) << 22,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 62-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 12) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 13) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 14) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 15) << 30,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 62-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 16) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 17) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 18) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 19) << 38,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 62-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 20) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 21) << 42,
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 22) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 23) << 46,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 62-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 24) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 14 | SafeLoadAs<uint64_t>(in + 8 * 25) << 50,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 26) << 52,
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 10 | SafeLoadAs<uint64_t>(in + 8 * 27) << 54,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 62-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 8 | SafeLoadAs<uint64_t>(in + 8 * 28) << 56,
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 6 | SafeLoadAs<uint64_t>(in + 8 * 29) << 58,
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 4 | SafeLoadAs<uint64_t>(in + 8 * 30) << 60,
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+    };
+    shifts = simd_batch{ 0, 0, 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 62-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 31),
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 32) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 33) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 34) << 6,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 62-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 35) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 36) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 37) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 38) << 14,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 62-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 39) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 40) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 41) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 42) << 22,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 62-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 43) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 44) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 45) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 46) << 30,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 62-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 47) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 47) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 48) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 48) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 49) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 49) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 50) << 38,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 62-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 50) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 51) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 51) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 52) << 42,
+      SafeLoadAs<uint64_t>(in + 8 * 52) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 53) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 53) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 54) << 46,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 62-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 54) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 55) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 55) >> 14 | SafeLoadAs<uint64_t>(in + 8 * 56) << 50,
+      SafeLoadAs<uint64_t>(in + 8 * 56) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 57) << 52,
+      SafeLoadAs<uint64_t>(in + 8 * 57) >> 10 | SafeLoadAs<uint64_t>(in + 8 * 58) << 54,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 62-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 58) >> 8 | SafeLoadAs<uint64_t>(in + 8 * 59) << 56,
+      SafeLoadAs<uint64_t>(in + 8 * 59) >> 6 | SafeLoadAs<uint64_t>(in + 8 * 60) << 58,
+      SafeLoadAs<uint64_t>(in + 8 * 60) >> 4 | SafeLoadAs<uint64_t>(in + 8 * 61) << 60,
+      SafeLoadAs<uint64_t>(in + 8 * 61),
+    };
+    shifts = simd_batch{ 0, 0, 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 62 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd256UnpackerForWidth<uint64_t, 63> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 4>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7fffffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 63-bit bundles 0 to 3
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 1) << 1,
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 2) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 3) << 3,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 63-bit bundles 4 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 4) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 5) << 5,
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 6) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 7) << 7,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 63-bit bundles 8 to 11
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 8) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 9) << 9,
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 10) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 11) << 11,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 63-bit bundles 12 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 12) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 13) << 13,
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 14) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 15) << 15,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 63-bit bundles 16 to 19
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 16) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 17) << 17,
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 18) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 19) << 19,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 63-bit bundles 20 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 20) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 21) << 21,
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 22) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 23) << 23,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 63-bit bundles 24 to 27
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 24) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 25) << 25,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 26) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 27) << 27,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 63-bit bundles 28 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 28) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 29) << 29,
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 30) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 31) << 31,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 63-bit bundles 32 to 35
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 32) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 31 | SafeLoadAs<uint64_t>(in + 8 * 33) << 33,
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 34) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 29 | SafeLoadAs<uint64_t>(in + 8 * 35) << 35,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 63-bit bundles 36 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 36) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 27 | SafeLoadAs<uint64_t>(in + 8 * 37) << 37,
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 38) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 25 | SafeLoadAs<uint64_t>(in + 8 * 39) << 39,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 63-bit bundles 40 to 43
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 40) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 23 | SafeLoadAs<uint64_t>(in + 8 * 41) << 41,
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 42) << 42,
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 21 | SafeLoadAs<uint64_t>(in + 8 * 43) << 43,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 63-bit bundles 44 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 44) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 19 | SafeLoadAs<uint64_t>(in + 8 * 45) << 45,
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 46) << 46,
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 17 | SafeLoadAs<uint64_t>(in + 8 * 47) << 47,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 63-bit bundles 48 to 51
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 47) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 48) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 48) >> 15 | SafeLoadAs<uint64_t>(in + 8 * 49) << 49,
+      SafeLoadAs<uint64_t>(in + 8 * 49) >> 14 | SafeLoadAs<uint64_t>(in + 8 * 50) << 50,
+      SafeLoadAs<uint64_t>(in + 8 * 50) >> 13 | SafeLoadAs<uint64_t>(in + 8 * 51) << 51,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 63-bit bundles 52 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 51) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 52) << 52,
+      SafeLoadAs<uint64_t>(in + 8 * 52) >> 11 | SafeLoadAs<uint64_t>(in + 8 * 53) << 53,
+      SafeLoadAs<uint64_t>(in + 8 * 53) >> 10 | SafeLoadAs<uint64_t>(in + 8 * 54) << 54,
+      SafeLoadAs<uint64_t>(in + 8 * 54) >> 9 | SafeLoadAs<uint64_t>(in + 8 * 55) << 55,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 63-bit bundles 56 to 59
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 55) >> 8 | SafeLoadAs<uint64_t>(in + 8 * 56) << 56,
+      SafeLoadAs<uint64_t>(in + 8 * 56) >> 7 | SafeLoadAs<uint64_t>(in + 8 * 57) << 57,
+      SafeLoadAs<uint64_t>(in + 8 * 57) >> 6 | SafeLoadAs<uint64_t>(in + 8 * 58) << 58,
+      SafeLoadAs<uint64_t>(in + 8 * 58) >> 5 | SafeLoadAs<uint64_t>(in + 8 * 59) << 59,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    // extract 63-bit bundles 60 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 59) >> 4 | SafeLoadAs<uint64_t>(in + 8 * 60) << 60,
+      SafeLoadAs<uint64_t>(in + 8 * 60) >> 3 | SafeLoadAs<uint64_t>(in + 8 * 61) << 61,
+      SafeLoadAs<uint64_t>(in + 8 * 61) >> 2 | SafeLoadAs<uint64_t>(in + 8 * 62) << 62,
+      SafeLoadAs<uint64_t>(in + 8 * 62),
+    };
+    shifts = simd_batch{ 0, 0, 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 4;
+
+    in += 63 * 8;
+    return in;
+  }
+};
+
+
+}  // namespace arrow::internal
diff --git a/cpp/src/arrow/util/bpacking_simd512_generated_internal.h b/cpp/src/arrow/util/bpacking_simd512_generated_internal.h
index 4f2aeaeeb4b00..99089dc21d3d5 100644
--- a/cpp/src/arrow/util/bpacking_simd512_generated_internal.h
+++ b/cpp/src/arrow/util/bpacking_simd512_generated_internal.h
@@ -15,7 +15,9 @@
 // specific language governing permissions and limitations
 // under the License.
 
-// Automatically generated file; DO NOT EDIT.
+// WARNING: this file is generated, DO NOT EDIT.
+// Usage:
+//   python cpp/src/arrow/util/bpacking_simd_codegen.py 512
 
 #pragma once
 
@@ -24,813 +26,11165 @@
 
 #include <xsimd/xsimd.hpp>
 
-#include "arrow/util/dispatch_internal.h"
 #include "arrow/util/ubsan.h"
 
-namespace arrow {
-namespace internal {
-namespace {
-
-using ::arrow::util::SafeLoad;
-
-template <DispatchLevel level>
-struct UnpackBits512 {
-
-using simd_batch = xsimd::make_sized_batch_t<uint32_t, 16>;
-
-inline static const uint32_t* unpack0_32(const uint32_t* in, uint32_t* out) {
-  memset(out, 0x0, 32 * sizeof(*out));
-  out += 32;
-
-  return in;
-}
-
-inline static const uint32_t* unpack1_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x1;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 1-bit bundles 0 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) };
-  shifts = simd_batch{ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  // extract 1-bit bundles 16 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) };
-  shifts = simd_batch{ 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  in += 1;
-  return in;
-}
-
-inline static const uint32_t* unpack2_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x3;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 2-bit bundles 0 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) };
-  shifts = simd_batch{ 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  // extract 2-bit bundles 16 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) };
-  shifts = simd_batch{ 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  in += 2;
-  return in;
-}
-
-inline static const uint32_t* unpack3_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x7;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 3-bit bundles 0 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 30 | SafeLoad<uint32_t>(in + 1) << 2, SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) };
-  shifts = simd_batch{ 0, 3, 6, 9, 12, 15, 18, 21, 24, 27, 0, 1, 4, 7, 10, 13 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  // extract 3-bit bundles 16 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) >> 31 | SafeLoad<uint32_t>(in + 2) << 1, SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) };
-  shifts = simd_batch{ 16, 19, 22, 25, 28, 0, 2, 5, 8, 11, 14, 17, 20, 23, 26, 29 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  in += 3;
-  return in;
-}
-
-inline static const uint32_t* unpack4_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0xf;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 4-bit bundles 0 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) };
-  shifts = simd_batch{ 0, 4, 8, 12, 16, 20, 24, 28, 0, 4, 8, 12, 16, 20, 24, 28 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  // extract 4-bit bundles 16 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) };
-  shifts = simd_batch{ 0, 4, 8, 12, 16, 20, 24, 28, 0, 4, 8, 12, 16, 20, 24, 28 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  in += 4;
-  return in;
-}
-
-inline static const uint32_t* unpack5_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x1f;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 5-bit bundles 0 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 30 | SafeLoad<uint32_t>(in + 1) << 2, SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) >> 28 | SafeLoad<uint32_t>(in + 2) << 4, SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) };
-  shifts = simd_batch{ 0, 5, 10, 15, 20, 25, 0, 3, 8, 13, 18, 23, 0, 1, 6, 11 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  // extract 5-bit bundles 16 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) >> 31 | SafeLoad<uint32_t>(in + 3) << 1, SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) >> 29 | SafeLoad<uint32_t>(in + 4) << 3, SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4) };
-  shifts = simd_batch{ 16, 21, 26, 0, 4, 9, 14, 19, 24, 0, 2, 7, 12, 17, 22, 27 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  in += 5;
-  return in;
-}
-
-inline static const uint32_t* unpack6_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x3f;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 6-bit bundles 0 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 30 | SafeLoad<uint32_t>(in + 1) << 2, SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) >> 28 | SafeLoad<uint32_t>(in + 2) << 4, SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) };
-  shifts = simd_batch{ 0, 6, 12, 18, 24, 0, 4, 10, 16, 22, 0, 2, 8, 14, 20, 26 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  // extract 6-bit bundles 16 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) >> 30 | SafeLoad<uint32_t>(in + 4) << 2, SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4) >> 28 | SafeLoad<uint32_t>(in + 5) << 4, SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) };
-  shifts = simd_batch{ 0, 6, 12, 18, 24, 0, 4, 10, 16, 22, 0, 2, 8, 14, 20, 26 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  in += 6;
-  return in;
-}
-
-inline static const uint32_t* unpack7_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x7f;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 7-bit bundles 0 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 28 | SafeLoad<uint32_t>(in + 1) << 4, SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) >> 31 | SafeLoad<uint32_t>(in + 2) << 1, SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) >> 27 | SafeLoad<uint32_t>(in + 3) << 5, SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) };
-  shifts = simd_batch{ 0, 7, 14, 21, 0, 3, 10, 17, 24, 0, 6, 13, 20, 0, 2, 9 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  // extract 7-bit bundles 16 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) >> 30 | SafeLoad<uint32_t>(in + 4) << 2, SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4) >> 26 | SafeLoad<uint32_t>(in + 5) << 6, SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) >> 29 | SafeLoad<uint32_t>(in + 6) << 3, SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6) };
-  shifts = simd_batch{ 16, 23, 0, 5, 12, 19, 0, 1, 8, 15, 22, 0, 4, 11, 18, 25 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  in += 7;
-  return in;
-}
-
-inline static const uint32_t* unpack8_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0xff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 8-bit bundles 0 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) };
-  shifts = simd_batch{ 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  // extract 8-bit bundles 16 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7) };
-  shifts = simd_batch{ 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  in += 8;
-  return in;
-}
-
-inline static const uint32_t* unpack9_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x1ff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 9-bit bundles 0 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 27 | SafeLoad<uint32_t>(in + 1) << 5, SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) >> 31 | SafeLoad<uint32_t>(in + 2) << 1, SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) >> 26 | SafeLoad<uint32_t>(in + 3) << 6, SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) >> 30 | SafeLoad<uint32_t>(in + 4) << 2, SafeLoad<uint32_t>(in + 4) };
-  shifts = simd_batch{ 0, 9, 18, 0, 4, 13, 22, 0, 8, 17, 0, 3, 12, 21, 0, 7 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  // extract 9-bit bundles 16 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4) >> 25 | SafeLoad<uint32_t>(in + 5) << 7, SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) >> 29 | SafeLoad<uint32_t>(in + 6) << 3, SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6) >> 24 | SafeLoad<uint32_t>(in + 7) << 8, SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7) >> 28 | SafeLoad<uint32_t>(in + 8) << 4, SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 8) };
-  shifts = simd_batch{ 16, 0, 2, 11, 20, 0, 6, 15, 0, 1, 10, 19, 0, 5, 14, 23 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  in += 9;
-  return in;
-}
-
-inline static const uint32_t* unpack10_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x3ff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 10-bit bundles 0 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 30 | SafeLoad<uint32_t>(in + 1) << 2, SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) >> 28 | SafeLoad<uint32_t>(in + 2) << 4, SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) >> 26 | SafeLoad<uint32_t>(in + 3) << 6, SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) >> 24 | SafeLoad<uint32_t>(in + 4) << 8, SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4) };
-  shifts = simd_batch{ 0, 10, 20, 0, 8, 18, 0, 6, 16, 0, 4, 14, 0, 2, 12, 22 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  // extract 10-bit bundles 16 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) >> 30 | SafeLoad<uint32_t>(in + 6) << 2, SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6) >> 28 | SafeLoad<uint32_t>(in + 7) << 4, SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7) >> 26 | SafeLoad<uint32_t>(in + 8) << 6, SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 8) >> 24 | SafeLoad<uint32_t>(in + 9) << 8, SafeLoad<uint32_t>(in + 9), SafeLoad<uint32_t>(in + 9), SafeLoad<uint32_t>(in + 9) };
-  shifts = simd_batch{ 0, 10, 20, 0, 8, 18, 0, 6, 16, 0, 4, 14, 0, 2, 12, 22 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  in += 10;
-  return in;
-}
-
-inline static const uint32_t* unpack11_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x7ff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 11-bit bundles 0 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 22 | SafeLoad<uint32_t>(in + 1) << 10, SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) >> 23 | SafeLoad<uint32_t>(in + 2) << 9, SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) >> 24 | SafeLoad<uint32_t>(in + 3) << 8, SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) >> 25 | SafeLoad<uint32_t>(in + 4) << 7, SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4) >> 26 | SafeLoad<uint32_t>(in + 5) << 6, SafeLoad<uint32_t>(in + 5) };
-  shifts = simd_batch{ 0, 11, 0, 1, 12, 0, 2, 13, 0, 3, 14, 0, 4, 15, 0, 5 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  // extract 11-bit bundles 16 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) >> 27 | SafeLoad<uint32_t>(in + 6) << 5, SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6) >> 28 | SafeLoad<uint32_t>(in + 7) << 4, SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7) >> 29 | SafeLoad<uint32_t>(in + 8) << 3, SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 8) >> 30 | SafeLoad<uint32_t>(in + 9) << 2, SafeLoad<uint32_t>(in + 9), SafeLoad<uint32_t>(in + 9), SafeLoad<uint32_t>(in + 9) >> 31 | SafeLoad<uint32_t>(in + 10) << 1, SafeLoad<uint32_t>(in + 10), SafeLoad<uint32_t>(in + 10) };
-  shifts = simd_batch{ 16, 0, 6, 17, 0, 7, 18, 0, 8, 19, 0, 9, 20, 0, 10, 21 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  in += 11;
-  return in;
-}
-
-inline static const uint32_t* unpack12_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0xfff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 12-bit bundles 0 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 24 | SafeLoad<uint32_t>(in + 1) << 8, SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) >> 28 | SafeLoad<uint32_t>(in + 2) << 4, SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) >> 24 | SafeLoad<uint32_t>(in + 4) << 8, SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4) >> 28 | SafeLoad<uint32_t>(in + 5) << 4, SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) };
-  shifts = simd_batch{ 0, 12, 0, 4, 16, 0, 8, 20, 0, 12, 0, 4, 16, 0, 8, 20 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  // extract 12-bit bundles 16 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6) >> 24 | SafeLoad<uint32_t>(in + 7) << 8, SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7) >> 28 | SafeLoad<uint32_t>(in + 8) << 4, SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 9), SafeLoad<uint32_t>(in + 9), SafeLoad<uint32_t>(in + 9) >> 24 | SafeLoad<uint32_t>(in + 10) << 8, SafeLoad<uint32_t>(in + 10), SafeLoad<uint32_t>(in + 10), SafeLoad<uint32_t>(in + 10) >> 28 | SafeLoad<uint32_t>(in + 11) << 4, SafeLoad<uint32_t>(in + 11), SafeLoad<uint32_t>(in + 11) };
-  shifts = simd_batch{ 0, 12, 0, 4, 16, 0, 8, 20, 0, 12, 0, 4, 16, 0, 8, 20 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  in += 12;
-  return in;
-}
-
-inline static const uint32_t* unpack13_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x1fff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 13-bit bundles 0 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 26 | SafeLoad<uint32_t>(in + 1) << 6, SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) >> 20 | SafeLoad<uint32_t>(in + 2) << 12, SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) >> 27 | SafeLoad<uint32_t>(in + 3) << 5, SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) >> 21 | SafeLoad<uint32_t>(in + 4) << 11, SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4) >> 28 | SafeLoad<uint32_t>(in + 5) << 4, SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) >> 22 | SafeLoad<uint32_t>(in + 6) << 10, SafeLoad<uint32_t>(in + 6) };
-  shifts = simd_batch{ 0, 13, 0, 7, 0, 1, 14, 0, 8, 0, 2, 15, 0, 9, 0, 3 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  // extract 13-bit bundles 16 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6) >> 29 | SafeLoad<uint32_t>(in + 7) << 3, SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7) >> 23 | SafeLoad<uint32_t>(in + 8) << 9, SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 8) >> 30 | SafeLoad<uint32_t>(in + 9) << 2, SafeLoad<uint32_t>(in + 9), SafeLoad<uint32_t>(in + 9) >> 24 | SafeLoad<uint32_t>(in + 10) << 8, SafeLoad<uint32_t>(in + 10), SafeLoad<uint32_t>(in + 10), SafeLoad<uint32_t>(in + 10) >> 31 | SafeLoad<uint32_t>(in + 11) << 1, SafeLoad<uint32_t>(in + 11), SafeLoad<uint32_t>(in + 11) >> 25 | SafeLoad<uint32_t>(in + 12) << 7, SafeLoad<uint32_t>(in + 12), SafeLoad<uint32_t>(in + 12) };
-  shifts = simd_batch{ 16, 0, 10, 0, 4, 17, 0, 11, 0, 5, 18, 0, 12, 0, 6, 19 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  in += 13;
-  return in;
-}
-
-inline static const uint32_t* unpack14_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x3fff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 14-bit bundles 0 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 28 | SafeLoad<uint32_t>(in + 1) << 4, SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) >> 24 | SafeLoad<uint32_t>(in + 2) << 8, SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) >> 20 | SafeLoad<uint32_t>(in + 3) << 12, SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) >> 30 | SafeLoad<uint32_t>(in + 4) << 2, SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4) >> 26 | SafeLoad<uint32_t>(in + 5) << 6, SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) >> 22 | SafeLoad<uint32_t>(in + 6) << 10, SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6) };
-  shifts = simd_batch{ 0, 14, 0, 10, 0, 6, 0, 2, 16, 0, 12, 0, 8, 0, 4, 18 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  // extract 14-bit bundles 16 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7) >> 28 | SafeLoad<uint32_t>(in + 8) << 4, SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 8) >> 24 | SafeLoad<uint32_t>(in + 9) << 8, SafeLoad<uint32_t>(in + 9), SafeLoad<uint32_t>(in + 9) >> 20 | SafeLoad<uint32_t>(in + 10) << 12, SafeLoad<uint32_t>(in + 10), SafeLoad<uint32_t>(in + 10), SafeLoad<uint32_t>(in + 10) >> 30 | SafeLoad<uint32_t>(in + 11) << 2, SafeLoad<uint32_t>(in + 11), SafeLoad<uint32_t>(in + 11) >> 26 | SafeLoad<uint32_t>(in + 12) << 6, SafeLoad<uint32_t>(in + 12), SafeLoad<uint32_t>(in + 12) >> 22 | SafeLoad<uint32_t>(in + 13) << 10, SafeLoad<uint32_t>(in + 13), SafeLoad<uint32_t>(in + 13) };
-  shifts = simd_batch{ 0, 14, 0, 10, 0, 6, 0, 2, 16, 0, 12, 0, 8, 0, 4, 18 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  in += 14;
-  return in;
-}
-
-inline static const uint32_t* unpack15_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x7fff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 15-bit bundles 0 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 30 | SafeLoad<uint32_t>(in + 1) << 2, SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) >> 28 | SafeLoad<uint32_t>(in + 2) << 4, SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) >> 26 | SafeLoad<uint32_t>(in + 3) << 6, SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) >> 24 | SafeLoad<uint32_t>(in + 4) << 8, SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4) >> 22 | SafeLoad<uint32_t>(in + 5) << 10, SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) >> 20 | SafeLoad<uint32_t>(in + 6) << 12, SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6) >> 18 | SafeLoad<uint32_t>(in + 7) << 14, SafeLoad<uint32_t>(in + 7) };
-  shifts = simd_batch{ 0, 15, 0, 13, 0, 11, 0, 9, 0, 7, 0, 5, 0, 3, 0, 1 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  // extract 15-bit bundles 16 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7) >> 31 | SafeLoad<uint32_t>(in + 8) << 1, SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 8) >> 29 | SafeLoad<uint32_t>(in + 9) << 3, SafeLoad<uint32_t>(in + 9), SafeLoad<uint32_t>(in + 9) >> 27 | SafeLoad<uint32_t>(in + 10) << 5, SafeLoad<uint32_t>(in + 10), SafeLoad<uint32_t>(in + 10) >> 25 | SafeLoad<uint32_t>(in + 11) << 7, SafeLoad<uint32_t>(in + 11), SafeLoad<uint32_t>(in + 11) >> 23 | SafeLoad<uint32_t>(in + 12) << 9, SafeLoad<uint32_t>(in + 12), SafeLoad<uint32_t>(in + 12) >> 21 | SafeLoad<uint32_t>(in + 13) << 11, SafeLoad<uint32_t>(in + 13), SafeLoad<uint32_t>(in + 13) >> 19 | SafeLoad<uint32_t>(in + 14) << 13, SafeLoad<uint32_t>(in + 14), SafeLoad<uint32_t>(in + 14) };
-  shifts = simd_batch{ 16, 0, 14, 0, 12, 0, 10, 0, 8, 0, 6, 0, 4, 0, 2, 17 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  in += 15;
-  return in;
-}
-
-inline static const uint32_t* unpack16_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0xffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 16-bit bundles 0 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7) };
-  shifts = simd_batch{ 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  // extract 16-bit bundles 16 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 9), SafeLoad<uint32_t>(in + 9), SafeLoad<uint32_t>(in + 10), SafeLoad<uint32_t>(in + 10), SafeLoad<uint32_t>(in + 11), SafeLoad<uint32_t>(in + 11), SafeLoad<uint32_t>(in + 12), SafeLoad<uint32_t>(in + 12), SafeLoad<uint32_t>(in + 13), SafeLoad<uint32_t>(in + 13), SafeLoad<uint32_t>(in + 14), SafeLoad<uint32_t>(in + 14), SafeLoad<uint32_t>(in + 15), SafeLoad<uint32_t>(in + 15) };
-  shifts = simd_batch{ 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  in += 16;
-  return in;
-}
-
-inline static const uint32_t* unpack17_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x1ffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 17-bit bundles 0 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 17 | SafeLoad<uint32_t>(in + 1) << 15, SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) >> 19 | SafeLoad<uint32_t>(in + 2) << 13, SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) >> 21 | SafeLoad<uint32_t>(in + 3) << 11, SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) >> 23 | SafeLoad<uint32_t>(in + 4) << 9, SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4) >> 25 | SafeLoad<uint32_t>(in + 5) << 7, SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) >> 27 | SafeLoad<uint32_t>(in + 6) << 5, SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6) >> 29 | SafeLoad<uint32_t>(in + 7) << 3, SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7) >> 31 | SafeLoad<uint32_t>(in + 8) << 1 };
-  shifts = simd_batch{ 0, 0, 2, 0, 4, 0, 6, 0, 8, 0, 10, 0, 12, 0, 14, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  // extract 17-bit bundles 16 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 8) >> 16 | SafeLoad<uint32_t>(in + 9) << 16, SafeLoad<uint32_t>(in + 9), SafeLoad<uint32_t>(in + 9) >> 18 | SafeLoad<uint32_t>(in + 10) << 14, SafeLoad<uint32_t>(in + 10), SafeLoad<uint32_t>(in + 10) >> 20 | SafeLoad<uint32_t>(in + 11) << 12, SafeLoad<uint32_t>(in + 11), SafeLoad<uint32_t>(in + 11) >> 22 | SafeLoad<uint32_t>(in + 12) << 10, SafeLoad<uint32_t>(in + 12), SafeLoad<uint32_t>(in + 12) >> 24 | SafeLoad<uint32_t>(in + 13) << 8, SafeLoad<uint32_t>(in + 13), SafeLoad<uint32_t>(in + 13) >> 26 | SafeLoad<uint32_t>(in + 14) << 6, SafeLoad<uint32_t>(in + 14), SafeLoad<uint32_t>(in + 14) >> 28 | SafeLoad<uint32_t>(in + 15) << 4, SafeLoad<uint32_t>(in + 15), SafeLoad<uint32_t>(in + 15) >> 30 | SafeLoad<uint32_t>(in + 16) << 2, SafeLoad<uint32_t>(in + 16) };
-  shifts = simd_batch{ 0, 1, 0, 3, 0, 5, 0, 7, 0, 9, 0, 11, 0, 13, 0, 15 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  in += 17;
-  return in;
-}
-
-inline static const uint32_t* unpack18_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x3ffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 18-bit bundles 0 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 18 | SafeLoad<uint32_t>(in + 1) << 14, SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) >> 22 | SafeLoad<uint32_t>(in + 2) << 10, SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) >> 26 | SafeLoad<uint32_t>(in + 3) << 6, SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) >> 30 | SafeLoad<uint32_t>(in + 4) << 2, SafeLoad<uint32_t>(in + 4) >> 16 | SafeLoad<uint32_t>(in + 5) << 16, SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) >> 20 | SafeLoad<uint32_t>(in + 6) << 12, SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6) >> 24 | SafeLoad<uint32_t>(in + 7) << 8, SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7) >> 28 | SafeLoad<uint32_t>(in + 8) << 4, SafeLoad<uint32_t>(in + 8) };
-  shifts = simd_batch{ 0, 0, 4, 0, 8, 0, 12, 0, 0, 2, 0, 6, 0, 10, 0, 14 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  // extract 18-bit bundles 16 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 9), SafeLoad<uint32_t>(in + 9) >> 18 | SafeLoad<uint32_t>(in + 10) << 14, SafeLoad<uint32_t>(in + 10), SafeLoad<uint32_t>(in + 10) >> 22 | SafeLoad<uint32_t>(in + 11) << 10, SafeLoad<uint32_t>(in + 11), SafeLoad<uint32_t>(in + 11) >> 26 | SafeLoad<uint32_t>(in + 12) << 6, SafeLoad<uint32_t>(in + 12), SafeLoad<uint32_t>(in + 12) >> 30 | SafeLoad<uint32_t>(in + 13) << 2, SafeLoad<uint32_t>(in + 13) >> 16 | SafeLoad<uint32_t>(in + 14) << 16, SafeLoad<uint32_t>(in + 14), SafeLoad<uint32_t>(in + 14) >> 20 | SafeLoad<uint32_t>(in + 15) << 12, SafeLoad<uint32_t>(in + 15), SafeLoad<uint32_t>(in + 15) >> 24 | SafeLoad<uint32_t>(in + 16) << 8, SafeLoad<uint32_t>(in + 16), SafeLoad<uint32_t>(in + 16) >> 28 | SafeLoad<uint32_t>(in + 17) << 4, SafeLoad<uint32_t>(in + 17) };
-  shifts = simd_batch{ 0, 0, 4, 0, 8, 0, 12, 0, 0, 2, 0, 6, 0, 10, 0, 14 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  in += 18;
-  return in;
-}
-
-inline static const uint32_t* unpack19_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x7ffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 19-bit bundles 0 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 19 | SafeLoad<uint32_t>(in + 1) << 13, SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) >> 25 | SafeLoad<uint32_t>(in + 2) << 7, SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) >> 31 | SafeLoad<uint32_t>(in + 3) << 1, SafeLoad<uint32_t>(in + 3) >> 18 | SafeLoad<uint32_t>(in + 4) << 14, SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4) >> 24 | SafeLoad<uint32_t>(in + 5) << 8, SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) >> 30 | SafeLoad<uint32_t>(in + 6) << 2, SafeLoad<uint32_t>(in + 6) >> 17 | SafeLoad<uint32_t>(in + 7) << 15, SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7) >> 23 | SafeLoad<uint32_t>(in + 8) << 9, SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 8) >> 29 | SafeLoad<uint32_t>(in + 9) << 3 };
-  shifts = simd_batch{ 0, 0, 6, 0, 12, 0, 0, 5, 0, 11, 0, 0, 4, 0, 10, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  // extract 19-bit bundles 16 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 9) >> 16 | SafeLoad<uint32_t>(in + 10) << 16, SafeLoad<uint32_t>(in + 10), SafeLoad<uint32_t>(in + 10) >> 22 | SafeLoad<uint32_t>(in + 11) << 10, SafeLoad<uint32_t>(in + 11), SafeLoad<uint32_t>(in + 11) >> 28 | SafeLoad<uint32_t>(in + 12) << 4, SafeLoad<uint32_t>(in + 12) >> 15 | SafeLoad<uint32_t>(in + 13) << 17, SafeLoad<uint32_t>(in + 13), SafeLoad<uint32_t>(in + 13) >> 21 | SafeLoad<uint32_t>(in + 14) << 11, SafeLoad<uint32_t>(in + 14), SafeLoad<uint32_t>(in + 14) >> 27 | SafeLoad<uint32_t>(in + 15) << 5, SafeLoad<uint32_t>(in + 15) >> 14 | SafeLoad<uint32_t>(in + 16) << 18, SafeLoad<uint32_t>(in + 16), SafeLoad<uint32_t>(in + 16) >> 20 | SafeLoad<uint32_t>(in + 17) << 12, SafeLoad<uint32_t>(in + 17), SafeLoad<uint32_t>(in + 17) >> 26 | SafeLoad<uint32_t>(in + 18) << 6, SafeLoad<uint32_t>(in + 18) };
-  shifts = simd_batch{ 0, 3, 0, 9, 0, 0, 2, 0, 8, 0, 0, 1, 0, 7, 0, 13 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  in += 19;
-  return in;
-}
-
-inline static const uint32_t* unpack20_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0xfffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 20-bit bundles 0 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 20 | SafeLoad<uint32_t>(in + 1) << 12, SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) >> 28 | SafeLoad<uint32_t>(in + 2) << 4, SafeLoad<uint32_t>(in + 2) >> 16 | SafeLoad<uint32_t>(in + 3) << 16, SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) >> 24 | SafeLoad<uint32_t>(in + 4) << 8, SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) >> 20 | SafeLoad<uint32_t>(in + 6) << 12, SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6) >> 28 | SafeLoad<uint32_t>(in + 7) << 4, SafeLoad<uint32_t>(in + 7) >> 16 | SafeLoad<uint32_t>(in + 8) << 16, SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 8) >> 24 | SafeLoad<uint32_t>(in + 9) << 8, SafeLoad<uint32_t>(in + 9) };
-  shifts = simd_batch{ 0, 0, 8, 0, 0, 4, 0, 12, 0, 0, 8, 0, 0, 4, 0, 12 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  // extract 20-bit bundles 16 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 10), SafeLoad<uint32_t>(in + 10) >> 20 | SafeLoad<uint32_t>(in + 11) << 12, SafeLoad<uint32_t>(in + 11), SafeLoad<uint32_t>(in + 11) >> 28 | SafeLoad<uint32_t>(in + 12) << 4, SafeLoad<uint32_t>(in + 12) >> 16 | SafeLoad<uint32_t>(in + 13) << 16, SafeLoad<uint32_t>(in + 13), SafeLoad<uint32_t>(in + 13) >> 24 | SafeLoad<uint32_t>(in + 14) << 8, SafeLoad<uint32_t>(in + 14), SafeLoad<uint32_t>(in + 15), SafeLoad<uint32_t>(in + 15) >> 20 | SafeLoad<uint32_t>(in + 16) << 12, SafeLoad<uint32_t>(in + 16), SafeLoad<uint32_t>(in + 16) >> 28 | SafeLoad<uint32_t>(in + 17) << 4, SafeLoad<uint32_t>(in + 17) >> 16 | SafeLoad<uint32_t>(in + 18) << 16, SafeLoad<uint32_t>(in + 18), SafeLoad<uint32_t>(in + 18) >> 24 | SafeLoad<uint32_t>(in + 19) << 8, SafeLoad<uint32_t>(in + 19) };
-  shifts = simd_batch{ 0, 0, 8, 0, 0, 4, 0, 12, 0, 0, 8, 0, 0, 4, 0, 12 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  in += 20;
-  return in;
-}
-
-inline static const uint32_t* unpack21_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x1fffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 21-bit bundles 0 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 21 | SafeLoad<uint32_t>(in + 1) << 11, SafeLoad<uint32_t>(in + 1), SafeLoad<uint32_t>(in + 1) >> 31 | SafeLoad<uint32_t>(in + 2) << 1, SafeLoad<uint32_t>(in + 2) >> 20 | SafeLoad<uint32_t>(in + 3) << 12, SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) >> 30 | SafeLoad<uint32_t>(in + 4) << 2, SafeLoad<uint32_t>(in + 4) >> 19 | SafeLoad<uint32_t>(in + 5) << 13, SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) >> 29 | SafeLoad<uint32_t>(in + 6) << 3, SafeLoad<uint32_t>(in + 6) >> 18 | SafeLoad<uint32_t>(in + 7) << 14, SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7) >> 28 | SafeLoad<uint32_t>(in + 8) << 4, SafeLoad<uint32_t>(in + 8) >> 17 | SafeLoad<uint32_t>(in + 9) << 15, SafeLoad<uint32_t>(in + 9), SafeLoad<uint32_t>(in + 9) >> 27 | SafeLoad<uint32_t>(in + 10) << 5 };
-  shifts = simd_batch{ 0, 0, 10, 0, 0, 9, 0, 0, 8, 0, 0, 7, 0, 0, 6, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  // extract 21-bit bundles 16 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 10) >> 16 | SafeLoad<uint32_t>(in + 11) << 16, SafeLoad<uint32_t>(in + 11), SafeLoad<uint32_t>(in + 11) >> 26 | SafeLoad<uint32_t>(in + 12) << 6, SafeLoad<uint32_t>(in + 12) >> 15 | SafeLoad<uint32_t>(in + 13) << 17, SafeLoad<uint32_t>(in + 13), SafeLoad<uint32_t>(in + 13) >> 25 | SafeLoad<uint32_t>(in + 14) << 7, SafeLoad<uint32_t>(in + 14) >> 14 | SafeLoad<uint32_t>(in + 15) << 18, SafeLoad<uint32_t>(in + 15), SafeLoad<uint32_t>(in + 15) >> 24 | SafeLoad<uint32_t>(in + 16) << 8, SafeLoad<uint32_t>(in + 16) >> 13 | SafeLoad<uint32_t>(in + 17) << 19, SafeLoad<uint32_t>(in + 17), SafeLoad<uint32_t>(in + 17) >> 23 | SafeLoad<uint32_t>(in + 18) << 9, SafeLoad<uint32_t>(in + 18) >> 12 | SafeLoad<uint32_t>(in + 19) << 20, SafeLoad<uint32_t>(in + 19), SafeLoad<uint32_t>(in + 19) >> 22 | SafeLoad<uint32_t>(in + 20) << 10, SafeLoad<uint32_t>(in + 20) };
-  shifts = simd_batch{ 0, 5, 0, 0, 4, 0, 0, 3, 0, 0, 2, 0, 0, 1, 0, 11 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  in += 21;
-  return in;
-}
-
-inline static const uint32_t* unpack22_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x3fffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 22-bit bundles 0 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 22 | SafeLoad<uint32_t>(in + 1) << 10, SafeLoad<uint32_t>(in + 1) >> 12 | SafeLoad<uint32_t>(in + 2) << 20, SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) >> 24 | SafeLoad<uint32_t>(in + 3) << 8, SafeLoad<uint32_t>(in + 3) >> 14 | SafeLoad<uint32_t>(in + 4) << 18, SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4) >> 26 | SafeLoad<uint32_t>(in + 5) << 6, SafeLoad<uint32_t>(in + 5) >> 16 | SafeLoad<uint32_t>(in + 6) << 16, SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6) >> 28 | SafeLoad<uint32_t>(in + 7) << 4, SafeLoad<uint32_t>(in + 7) >> 18 | SafeLoad<uint32_t>(in + 8) << 14, SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 8) >> 30 | SafeLoad<uint32_t>(in + 9) << 2, SafeLoad<uint32_t>(in + 9) >> 20 | SafeLoad<uint32_t>(in + 10) << 12, SafeLoad<uint32_t>(in + 10) };
-  shifts = simd_batch{ 0, 0, 0, 2, 0, 0, 4, 0, 0, 6, 0, 0, 8, 0, 0, 10 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  // extract 22-bit bundles 16 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 11), SafeLoad<uint32_t>(in + 11) >> 22 | SafeLoad<uint32_t>(in + 12) << 10, SafeLoad<uint32_t>(in + 12) >> 12 | SafeLoad<uint32_t>(in + 13) << 20, SafeLoad<uint32_t>(in + 13), SafeLoad<uint32_t>(in + 13) >> 24 | SafeLoad<uint32_t>(in + 14) << 8, SafeLoad<uint32_t>(in + 14) >> 14 | SafeLoad<uint32_t>(in + 15) << 18, SafeLoad<uint32_t>(in + 15), SafeLoad<uint32_t>(in + 15) >> 26 | SafeLoad<uint32_t>(in + 16) << 6, SafeLoad<uint32_t>(in + 16) >> 16 | SafeLoad<uint32_t>(in + 17) << 16, SafeLoad<uint32_t>(in + 17), SafeLoad<uint32_t>(in + 17) >> 28 | SafeLoad<uint32_t>(in + 18) << 4, SafeLoad<uint32_t>(in + 18) >> 18 | SafeLoad<uint32_t>(in + 19) << 14, SafeLoad<uint32_t>(in + 19), SafeLoad<uint32_t>(in + 19) >> 30 | SafeLoad<uint32_t>(in + 20) << 2, SafeLoad<uint32_t>(in + 20) >> 20 | SafeLoad<uint32_t>(in + 21) << 12, SafeLoad<uint32_t>(in + 21) };
-  shifts = simd_batch{ 0, 0, 0, 2, 0, 0, 4, 0, 0, 6, 0, 0, 8, 0, 0, 10 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  in += 22;
-  return in;
-}
-
-inline static const uint32_t* unpack23_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x7fffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 23-bit bundles 0 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 23 | SafeLoad<uint32_t>(in + 1) << 9, SafeLoad<uint32_t>(in + 1) >> 14 | SafeLoad<uint32_t>(in + 2) << 18, SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 2) >> 28 | SafeLoad<uint32_t>(in + 3) << 4, SafeLoad<uint32_t>(in + 3) >> 19 | SafeLoad<uint32_t>(in + 4) << 13, SafeLoad<uint32_t>(in + 4) >> 10 | SafeLoad<uint32_t>(in + 5) << 22, SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) >> 24 | SafeLoad<uint32_t>(in + 6) << 8, SafeLoad<uint32_t>(in + 6) >> 15 | SafeLoad<uint32_t>(in + 7) << 17, SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7) >> 29 | SafeLoad<uint32_t>(in + 8) << 3, SafeLoad<uint32_t>(in + 8) >> 20 | SafeLoad<uint32_t>(in + 9) << 12, SafeLoad<uint32_t>(in + 9) >> 11 | SafeLoad<uint32_t>(in + 10) << 21, SafeLoad<uint32_t>(in + 10), SafeLoad<uint32_t>(in + 10) >> 25 | SafeLoad<uint32_t>(in + 11) << 7 };
-  shifts = simd_batch{ 0, 0, 0, 5, 0, 0, 0, 1, 0, 0, 6, 0, 0, 0, 2, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  // extract 23-bit bundles 16 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 11) >> 16 | SafeLoad<uint32_t>(in + 12) << 16, SafeLoad<uint32_t>(in + 12), SafeLoad<uint32_t>(in + 12) >> 30 | SafeLoad<uint32_t>(in + 13) << 2, SafeLoad<uint32_t>(in + 13) >> 21 | SafeLoad<uint32_t>(in + 14) << 11, SafeLoad<uint32_t>(in + 14) >> 12 | SafeLoad<uint32_t>(in + 15) << 20, SafeLoad<uint32_t>(in + 15), SafeLoad<uint32_t>(in + 15) >> 26 | SafeLoad<uint32_t>(in + 16) << 6, SafeLoad<uint32_t>(in + 16) >> 17 | SafeLoad<uint32_t>(in + 17) << 15, SafeLoad<uint32_t>(in + 17), SafeLoad<uint32_t>(in + 17) >> 31 | SafeLoad<uint32_t>(in + 18) << 1, SafeLoad<uint32_t>(in + 18) >> 22 | SafeLoad<uint32_t>(in + 19) << 10, SafeLoad<uint32_t>(in + 19) >> 13 | SafeLoad<uint32_t>(in + 20) << 19, SafeLoad<uint32_t>(in + 20), SafeLoad<uint32_t>(in + 20) >> 27 | SafeLoad<uint32_t>(in + 21) << 5, SafeLoad<uint32_t>(in + 21) >> 18 | SafeLoad<uint32_t>(in + 22) << 14, SafeLoad<uint32_t>(in + 22) };
-  shifts = simd_batch{ 0, 7, 0, 0, 0, 3, 0, 0, 8, 0, 0, 0, 4, 0, 0, 9 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  in += 23;
-  return in;
-}
-
-inline static const uint32_t* unpack24_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0xffffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 24-bit bundles 0 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 24 | SafeLoad<uint32_t>(in + 1) << 8, SafeLoad<uint32_t>(in + 1) >> 16 | SafeLoad<uint32_t>(in + 2) << 16, SafeLoad<uint32_t>(in + 2), SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) >> 24 | SafeLoad<uint32_t>(in + 4) << 8, SafeLoad<uint32_t>(in + 4) >> 16 | SafeLoad<uint32_t>(in + 5) << 16, SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 6) >> 24 | SafeLoad<uint32_t>(in + 7) << 8, SafeLoad<uint32_t>(in + 7) >> 16 | SafeLoad<uint32_t>(in + 8) << 16, SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 9), SafeLoad<uint32_t>(in + 9) >> 24 | SafeLoad<uint32_t>(in + 10) << 8, SafeLoad<uint32_t>(in + 10) >> 16 | SafeLoad<uint32_t>(in + 11) << 16, SafeLoad<uint32_t>(in + 11) };
-  shifts = simd_batch{ 0, 0, 0, 8, 0, 0, 0, 8, 0, 0, 0, 8, 0, 0, 0, 8 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  // extract 24-bit bundles 16 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 12), SafeLoad<uint32_t>(in + 12) >> 24 | SafeLoad<uint32_t>(in + 13) << 8, SafeLoad<uint32_t>(in + 13) >> 16 | SafeLoad<uint32_t>(in + 14) << 16, SafeLoad<uint32_t>(in + 14), SafeLoad<uint32_t>(in + 15), SafeLoad<uint32_t>(in + 15) >> 24 | SafeLoad<uint32_t>(in + 16) << 8, SafeLoad<uint32_t>(in + 16) >> 16 | SafeLoad<uint32_t>(in + 17) << 16, SafeLoad<uint32_t>(in + 17), SafeLoad<uint32_t>(in + 18), SafeLoad<uint32_t>(in + 18) >> 24 | SafeLoad<uint32_t>(in + 19) << 8, SafeLoad<uint32_t>(in + 19) >> 16 | SafeLoad<uint32_t>(in + 20) << 16, SafeLoad<uint32_t>(in + 20), SafeLoad<uint32_t>(in + 21), SafeLoad<uint32_t>(in + 21) >> 24 | SafeLoad<uint32_t>(in + 22) << 8, SafeLoad<uint32_t>(in + 22) >> 16 | SafeLoad<uint32_t>(in + 23) << 16, SafeLoad<uint32_t>(in + 23) };
-  shifts = simd_batch{ 0, 0, 0, 8, 0, 0, 0, 8, 0, 0, 0, 8, 0, 0, 0, 8 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  in += 24;
-  return in;
-}
-
-inline static const uint32_t* unpack25_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x1ffffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 25-bit bundles 0 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 25 | SafeLoad<uint32_t>(in + 1) << 7, SafeLoad<uint32_t>(in + 1) >> 18 | SafeLoad<uint32_t>(in + 2) << 14, SafeLoad<uint32_t>(in + 2) >> 11 | SafeLoad<uint32_t>(in + 3) << 21, SafeLoad<uint32_t>(in + 3), SafeLoad<uint32_t>(in + 3) >> 29 | SafeLoad<uint32_t>(in + 4) << 3, SafeLoad<uint32_t>(in + 4) >> 22 | SafeLoad<uint32_t>(in + 5) << 10, SafeLoad<uint32_t>(in + 5) >> 15 | SafeLoad<uint32_t>(in + 6) << 17, SafeLoad<uint32_t>(in + 6) >> 8 | SafeLoad<uint32_t>(in + 7) << 24, SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7) >> 26 | SafeLoad<uint32_t>(in + 8) << 6, SafeLoad<uint32_t>(in + 8) >> 19 | SafeLoad<uint32_t>(in + 9) << 13, SafeLoad<uint32_t>(in + 9) >> 12 | SafeLoad<uint32_t>(in + 10) << 20, SafeLoad<uint32_t>(in + 10), SafeLoad<uint32_t>(in + 10) >> 30 | SafeLoad<uint32_t>(in + 11) << 2, SafeLoad<uint32_t>(in + 11) >> 23 | SafeLoad<uint32_t>(in + 12) << 9 };
-  shifts = simd_batch{ 0, 0, 0, 0, 4, 0, 0, 0, 0, 1, 0, 0, 0, 5, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  // extract 25-bit bundles 16 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 12) >> 16 | SafeLoad<uint32_t>(in + 13) << 16, SafeLoad<uint32_t>(in + 13) >> 9 | SafeLoad<uint32_t>(in + 14) << 23, SafeLoad<uint32_t>(in + 14), SafeLoad<uint32_t>(in + 14) >> 27 | SafeLoad<uint32_t>(in + 15) << 5, SafeLoad<uint32_t>(in + 15) >> 20 | SafeLoad<uint32_t>(in + 16) << 12, SafeLoad<uint32_t>(in + 16) >> 13 | SafeLoad<uint32_t>(in + 17) << 19, SafeLoad<uint32_t>(in + 17), SafeLoad<uint32_t>(in + 17) >> 31 | SafeLoad<uint32_t>(in + 18) << 1, SafeLoad<uint32_t>(in + 18) >> 24 | SafeLoad<uint32_t>(in + 19) << 8, SafeLoad<uint32_t>(in + 19) >> 17 | SafeLoad<uint32_t>(in + 20) << 15, SafeLoad<uint32_t>(in + 20) >> 10 | SafeLoad<uint32_t>(in + 21) << 22, SafeLoad<uint32_t>(in + 21), SafeLoad<uint32_t>(in + 21) >> 28 | SafeLoad<uint32_t>(in + 22) << 4, SafeLoad<uint32_t>(in + 22) >> 21 | SafeLoad<uint32_t>(in + 23) << 11, SafeLoad<uint32_t>(in + 23) >> 14 | SafeLoad<uint32_t>(in + 24) << 18, SafeLoad<uint32_t>(in + 24) };
-  shifts = simd_batch{ 0, 0, 2, 0, 0, 0, 6, 0, 0, 0, 0, 3, 0, 0, 0, 7 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  in += 25;
-  return in;
-}
-
-inline static const uint32_t* unpack26_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x3ffffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 26-bit bundles 0 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 26 | SafeLoad<uint32_t>(in + 1) << 6, SafeLoad<uint32_t>(in + 1) >> 20 | SafeLoad<uint32_t>(in + 2) << 12, SafeLoad<uint32_t>(in + 2) >> 14 | SafeLoad<uint32_t>(in + 3) << 18, SafeLoad<uint32_t>(in + 3) >> 8 | SafeLoad<uint32_t>(in + 4) << 24, SafeLoad<uint32_t>(in + 4), SafeLoad<uint32_t>(in + 4) >> 28 | SafeLoad<uint32_t>(in + 5) << 4, SafeLoad<uint32_t>(in + 5) >> 22 | SafeLoad<uint32_t>(in + 6) << 10, SafeLoad<uint32_t>(in + 6) >> 16 | SafeLoad<uint32_t>(in + 7) << 16, SafeLoad<uint32_t>(in + 7) >> 10 | SafeLoad<uint32_t>(in + 8) << 22, SafeLoad<uint32_t>(in + 8), SafeLoad<uint32_t>(in + 8) >> 30 | SafeLoad<uint32_t>(in + 9) << 2, SafeLoad<uint32_t>(in + 9) >> 24 | SafeLoad<uint32_t>(in + 10) << 8, SafeLoad<uint32_t>(in + 10) >> 18 | SafeLoad<uint32_t>(in + 11) << 14, SafeLoad<uint32_t>(in + 11) >> 12 | SafeLoad<uint32_t>(in + 12) << 20, SafeLoad<uint32_t>(in + 12) };
-  shifts = simd_batch{ 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 4, 0, 0, 0, 0, 6 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  // extract 26-bit bundles 16 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 13), SafeLoad<uint32_t>(in + 13) >> 26 | SafeLoad<uint32_t>(in + 14) << 6, SafeLoad<uint32_t>(in + 14) >> 20 | SafeLoad<uint32_t>(in + 15) << 12, SafeLoad<uint32_t>(in + 15) >> 14 | SafeLoad<uint32_t>(in + 16) << 18, SafeLoad<uint32_t>(in + 16) >> 8 | SafeLoad<uint32_t>(in + 17) << 24, SafeLoad<uint32_t>(in + 17), SafeLoad<uint32_t>(in + 17) >> 28 | SafeLoad<uint32_t>(in + 18) << 4, SafeLoad<uint32_t>(in + 18) >> 22 | SafeLoad<uint32_t>(in + 19) << 10, SafeLoad<uint32_t>(in + 19) >> 16 | SafeLoad<uint32_t>(in + 20) << 16, SafeLoad<uint32_t>(in + 20) >> 10 | SafeLoad<uint32_t>(in + 21) << 22, SafeLoad<uint32_t>(in + 21), SafeLoad<uint32_t>(in + 21) >> 30 | SafeLoad<uint32_t>(in + 22) << 2, SafeLoad<uint32_t>(in + 22) >> 24 | SafeLoad<uint32_t>(in + 23) << 8, SafeLoad<uint32_t>(in + 23) >> 18 | SafeLoad<uint32_t>(in + 24) << 14, SafeLoad<uint32_t>(in + 24) >> 12 | SafeLoad<uint32_t>(in + 25) << 20, SafeLoad<uint32_t>(in + 25) };
-  shifts = simd_batch{ 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 4, 0, 0, 0, 0, 6 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  in += 26;
-  return in;
-}
-
-inline static const uint32_t* unpack27_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x7ffffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 27-bit bundles 0 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 27 | SafeLoad<uint32_t>(in + 1) << 5, SafeLoad<uint32_t>(in + 1) >> 22 | SafeLoad<uint32_t>(in + 2) << 10, SafeLoad<uint32_t>(in + 2) >> 17 | SafeLoad<uint32_t>(in + 3) << 15, SafeLoad<uint32_t>(in + 3) >> 12 | SafeLoad<uint32_t>(in + 4) << 20, SafeLoad<uint32_t>(in + 4) >> 7 | SafeLoad<uint32_t>(in + 5) << 25, SafeLoad<uint32_t>(in + 5), SafeLoad<uint32_t>(in + 5) >> 29 | SafeLoad<uint32_t>(in + 6) << 3, SafeLoad<uint32_t>(in + 6) >> 24 | SafeLoad<uint32_t>(in + 7) << 8, SafeLoad<uint32_t>(in + 7) >> 19 | SafeLoad<uint32_t>(in + 8) << 13, SafeLoad<uint32_t>(in + 8) >> 14 | SafeLoad<uint32_t>(in + 9) << 18, SafeLoad<uint32_t>(in + 9) >> 9 | SafeLoad<uint32_t>(in + 10) << 23, SafeLoad<uint32_t>(in + 10), SafeLoad<uint32_t>(in + 10) >> 31 | SafeLoad<uint32_t>(in + 11) << 1, SafeLoad<uint32_t>(in + 11) >> 26 | SafeLoad<uint32_t>(in + 12) << 6, SafeLoad<uint32_t>(in + 12) >> 21 | SafeLoad<uint32_t>(in + 13) << 11 };
-  shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 4, 0, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  // extract 27-bit bundles 16 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 13) >> 16 | SafeLoad<uint32_t>(in + 14) << 16, SafeLoad<uint32_t>(in + 14) >> 11 | SafeLoad<uint32_t>(in + 15) << 21, SafeLoad<uint32_t>(in + 15) >> 6 | SafeLoad<uint32_t>(in + 16) << 26, SafeLoad<uint32_t>(in + 16), SafeLoad<uint32_t>(in + 16) >> 28 | SafeLoad<uint32_t>(in + 17) << 4, SafeLoad<uint32_t>(in + 17) >> 23 | SafeLoad<uint32_t>(in + 18) << 9, SafeLoad<uint32_t>(in + 18) >> 18 | SafeLoad<uint32_t>(in + 19) << 14, SafeLoad<uint32_t>(in + 19) >> 13 | SafeLoad<uint32_t>(in + 20) << 19, SafeLoad<uint32_t>(in + 20) >> 8 | SafeLoad<uint32_t>(in + 21) << 24, SafeLoad<uint32_t>(in + 21), SafeLoad<uint32_t>(in + 21) >> 30 | SafeLoad<uint32_t>(in + 22) << 2, SafeLoad<uint32_t>(in + 22) >> 25 | SafeLoad<uint32_t>(in + 23) << 7, SafeLoad<uint32_t>(in + 23) >> 20 | SafeLoad<uint32_t>(in + 24) << 12, SafeLoad<uint32_t>(in + 24) >> 15 | SafeLoad<uint32_t>(in + 25) << 17, SafeLoad<uint32_t>(in + 25) >> 10 | SafeLoad<uint32_t>(in + 26) << 22, SafeLoad<uint32_t>(in + 26) };
-  shifts = simd_batch{ 0, 0, 0, 1, 0, 0, 0, 0, 0, 3, 0, 0, 0, 0, 0, 5 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  in += 27;
-  return in;
-}
-
-inline static const uint32_t* unpack28_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0xfffffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 28-bit bundles 0 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 28 | SafeLoad<uint32_t>(in + 1) << 4, SafeLoad<uint32_t>(in + 1) >> 24 | SafeLoad<uint32_t>(in + 2) << 8, SafeLoad<uint32_t>(in + 2) >> 20 | SafeLoad<uint32_t>(in + 3) << 12, SafeLoad<uint32_t>(in + 3) >> 16 | SafeLoad<uint32_t>(in + 4) << 16, SafeLoad<uint32_t>(in + 4) >> 12 | SafeLoad<uint32_t>(in + 5) << 20, SafeLoad<uint32_t>(in + 5) >> 8 | SafeLoad<uint32_t>(in + 6) << 24, SafeLoad<uint32_t>(in + 6), SafeLoad<uint32_t>(in + 7), SafeLoad<uint32_t>(in + 7) >> 28 | SafeLoad<uint32_t>(in + 8) << 4, SafeLoad<uint32_t>(in + 8) >> 24 | SafeLoad<uint32_t>(in + 9) << 8, SafeLoad<uint32_t>(in + 9) >> 20 | SafeLoad<uint32_t>(in + 10) << 12, SafeLoad<uint32_t>(in + 10) >> 16 | SafeLoad<uint32_t>(in + 11) << 16, SafeLoad<uint32_t>(in + 11) >> 12 | SafeLoad<uint32_t>(in + 12) << 20, SafeLoad<uint32_t>(in + 12) >> 8 | SafeLoad<uint32_t>(in + 13) << 24, SafeLoad<uint32_t>(in + 13) };
-  shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 4, 0, 0, 0, 0, 0, 0, 0, 4 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  // extract 28-bit bundles 16 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 14), SafeLoad<uint32_t>(in + 14) >> 28 | SafeLoad<uint32_t>(in + 15) << 4, SafeLoad<uint32_t>(in + 15) >> 24 | SafeLoad<uint32_t>(in + 16) << 8, SafeLoad<uint32_t>(in + 16) >> 20 | SafeLoad<uint32_t>(in + 17) << 12, SafeLoad<uint32_t>(in + 17) >> 16 | SafeLoad<uint32_t>(in + 18) << 16, SafeLoad<uint32_t>(in + 18) >> 12 | SafeLoad<uint32_t>(in + 19) << 20, SafeLoad<uint32_t>(in + 19) >> 8 | SafeLoad<uint32_t>(in + 20) << 24, SafeLoad<uint32_t>(in + 20), SafeLoad<uint32_t>(in + 21), SafeLoad<uint32_t>(in + 21) >> 28 | SafeLoad<uint32_t>(in + 22) << 4, SafeLoad<uint32_t>(in + 22) >> 24 | SafeLoad<uint32_t>(in + 23) << 8, SafeLoad<uint32_t>(in + 23) >> 20 | SafeLoad<uint32_t>(in + 24) << 12, SafeLoad<uint32_t>(in + 24) >> 16 | SafeLoad<uint32_t>(in + 25) << 16, SafeLoad<uint32_t>(in + 25) >> 12 | SafeLoad<uint32_t>(in + 26) << 20, SafeLoad<uint32_t>(in + 26) >> 8 | SafeLoad<uint32_t>(in + 27) << 24, SafeLoad<uint32_t>(in + 27) };
-  shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 4, 0, 0, 0, 0, 0, 0, 0, 4 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  in += 28;
-  return in;
-}
-
-inline static const uint32_t* unpack29_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x1fffffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 29-bit bundles 0 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 29 | SafeLoad<uint32_t>(in + 1) << 3, SafeLoad<uint32_t>(in + 1) >> 26 | SafeLoad<uint32_t>(in + 2) << 6, SafeLoad<uint32_t>(in + 2) >> 23 | SafeLoad<uint32_t>(in + 3) << 9, SafeLoad<uint32_t>(in + 3) >> 20 | SafeLoad<uint32_t>(in + 4) << 12, SafeLoad<uint32_t>(in + 4) >> 17 | SafeLoad<uint32_t>(in + 5) << 15, SafeLoad<uint32_t>(in + 5) >> 14 | SafeLoad<uint32_t>(in + 6) << 18, SafeLoad<uint32_t>(in + 6) >> 11 | SafeLoad<uint32_t>(in + 7) << 21, SafeLoad<uint32_t>(in + 7) >> 8 | SafeLoad<uint32_t>(in + 8) << 24, SafeLoad<uint32_t>(in + 8) >> 5 | SafeLoad<uint32_t>(in + 9) << 27, SafeLoad<uint32_t>(in + 9), SafeLoad<uint32_t>(in + 9) >> 31 | SafeLoad<uint32_t>(in + 10) << 1, SafeLoad<uint32_t>(in + 10) >> 28 | SafeLoad<uint32_t>(in + 11) << 4, SafeLoad<uint32_t>(in + 11) >> 25 | SafeLoad<uint32_t>(in + 12) << 7, SafeLoad<uint32_t>(in + 12) >> 22 | SafeLoad<uint32_t>(in + 13) << 10, SafeLoad<uint32_t>(in + 13) >> 19 | SafeLoad<uint32_t>(in + 14) << 13 };
-  shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  // extract 29-bit bundles 16 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 14) >> 16 | SafeLoad<uint32_t>(in + 15) << 16, SafeLoad<uint32_t>(in + 15) >> 13 | SafeLoad<uint32_t>(in + 16) << 19, SafeLoad<uint32_t>(in + 16) >> 10 | SafeLoad<uint32_t>(in + 17) << 22, SafeLoad<uint32_t>(in + 17) >> 7 | SafeLoad<uint32_t>(in + 18) << 25, SafeLoad<uint32_t>(in + 18) >> 4 | SafeLoad<uint32_t>(in + 19) << 28, SafeLoad<uint32_t>(in + 19), SafeLoad<uint32_t>(in + 19) >> 30 | SafeLoad<uint32_t>(in + 20) << 2, SafeLoad<uint32_t>(in + 20) >> 27 | SafeLoad<uint32_t>(in + 21) << 5, SafeLoad<uint32_t>(in + 21) >> 24 | SafeLoad<uint32_t>(in + 22) << 8, SafeLoad<uint32_t>(in + 22) >> 21 | SafeLoad<uint32_t>(in + 23) << 11, SafeLoad<uint32_t>(in + 23) >> 18 | SafeLoad<uint32_t>(in + 24) << 14, SafeLoad<uint32_t>(in + 24) >> 15 | SafeLoad<uint32_t>(in + 25) << 17, SafeLoad<uint32_t>(in + 25) >> 12 | SafeLoad<uint32_t>(in + 26) << 20, SafeLoad<uint32_t>(in + 26) >> 9 | SafeLoad<uint32_t>(in + 27) << 23, SafeLoad<uint32_t>(in + 27) >> 6 | SafeLoad<uint32_t>(in + 28) << 26, SafeLoad<uint32_t>(in + 28) };
-  shifts = simd_batch{ 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 3 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  in += 29;
-  return in;
-}
-
-inline static const uint32_t* unpack30_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x3fffffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 30-bit bundles 0 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 30 | SafeLoad<uint32_t>(in + 1) << 2, SafeLoad<uint32_t>(in + 1) >> 28 | SafeLoad<uint32_t>(in + 2) << 4, SafeLoad<uint32_t>(in + 2) >> 26 | SafeLoad<uint32_t>(in + 3) << 6, SafeLoad<uint32_t>(in + 3) >> 24 | SafeLoad<uint32_t>(in + 4) << 8, SafeLoad<uint32_t>(in + 4) >> 22 | SafeLoad<uint32_t>(in + 5) << 10, SafeLoad<uint32_t>(in + 5) >> 20 | SafeLoad<uint32_t>(in + 6) << 12, SafeLoad<uint32_t>(in + 6) >> 18 | SafeLoad<uint32_t>(in + 7) << 14, SafeLoad<uint32_t>(in + 7) >> 16 | SafeLoad<uint32_t>(in + 8) << 16, SafeLoad<uint32_t>(in + 8) >> 14 | SafeLoad<uint32_t>(in + 9) << 18, SafeLoad<uint32_t>(in + 9) >> 12 | SafeLoad<uint32_t>(in + 10) << 20, SafeLoad<uint32_t>(in + 10) >> 10 | SafeLoad<uint32_t>(in + 11) << 22, SafeLoad<uint32_t>(in + 11) >> 8 | SafeLoad<uint32_t>(in + 12) << 24, SafeLoad<uint32_t>(in + 12) >> 6 | SafeLoad<uint32_t>(in + 13) << 26, SafeLoad<uint32_t>(in + 13) >> 4 | SafeLoad<uint32_t>(in + 14) << 28, SafeLoad<uint32_t>(in + 14) };
-  shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  // extract 30-bit bundles 16 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 15), SafeLoad<uint32_t>(in + 15) >> 30 | SafeLoad<uint32_t>(in + 16) << 2, SafeLoad<uint32_t>(in + 16) >> 28 | SafeLoad<uint32_t>(in + 17) << 4, SafeLoad<uint32_t>(in + 17) >> 26 | SafeLoad<uint32_t>(in + 18) << 6, SafeLoad<uint32_t>(in + 18) >> 24 | SafeLoad<uint32_t>(in + 19) << 8, SafeLoad<uint32_t>(in + 19) >> 22 | SafeLoad<uint32_t>(in + 20) << 10, SafeLoad<uint32_t>(in + 20) >> 20 | SafeLoad<uint32_t>(in + 21) << 12, SafeLoad<uint32_t>(in + 21) >> 18 | SafeLoad<uint32_t>(in + 22) << 14, SafeLoad<uint32_t>(in + 22) >> 16 | SafeLoad<uint32_t>(in + 23) << 16, SafeLoad<uint32_t>(in + 23) >> 14 | SafeLoad<uint32_t>(in + 24) << 18, SafeLoad<uint32_t>(in + 24) >> 12 | SafeLoad<uint32_t>(in + 25) << 20, SafeLoad<uint32_t>(in + 25) >> 10 | SafeLoad<uint32_t>(in + 26) << 22, SafeLoad<uint32_t>(in + 26) >> 8 | SafeLoad<uint32_t>(in + 27) << 24, SafeLoad<uint32_t>(in + 27) >> 6 | SafeLoad<uint32_t>(in + 28) << 26, SafeLoad<uint32_t>(in + 28) >> 4 | SafeLoad<uint32_t>(in + 29) << 28, SafeLoad<uint32_t>(in + 29) };
-  shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  in += 30;
-  return in;
-}
-
-inline static const uint32_t* unpack31_32(const uint32_t* in, uint32_t* out) {
-  uint32_t mask = 0x7fffffff;
-
-  simd_batch masks(mask);
-  simd_batch words, shifts;
-  simd_batch results;
-
-  // extract 31-bit bundles 0 to 15
-  words = simd_batch{ SafeLoad<uint32_t>(in + 0), SafeLoad<uint32_t>(in + 0) >> 31 | SafeLoad<uint32_t>(in + 1) << 1, SafeLoad<uint32_t>(in + 1) >> 30 | SafeLoad<uint32_t>(in + 2) << 2, SafeLoad<uint32_t>(in + 2) >> 29 | SafeLoad<uint32_t>(in + 3) << 3, SafeLoad<uint32_t>(in + 3) >> 28 | SafeLoad<uint32_t>(in + 4) << 4, SafeLoad<uint32_t>(in + 4) >> 27 | SafeLoad<uint32_t>(in + 5) << 5, SafeLoad<uint32_t>(in + 5) >> 26 | SafeLoad<uint32_t>(in + 6) << 6, SafeLoad<uint32_t>(in + 6) >> 25 | SafeLoad<uint32_t>(in + 7) << 7, SafeLoad<uint32_t>(in + 7) >> 24 | SafeLoad<uint32_t>(in + 8) << 8, SafeLoad<uint32_t>(in + 8) >> 23 | SafeLoad<uint32_t>(in + 9) << 9, SafeLoad<uint32_t>(in + 9) >> 22 | SafeLoad<uint32_t>(in + 10) << 10, SafeLoad<uint32_t>(in + 10) >> 21 | SafeLoad<uint32_t>(in + 11) << 11, SafeLoad<uint32_t>(in + 11) >> 20 | SafeLoad<uint32_t>(in + 12) << 12, SafeLoad<uint32_t>(in + 12) >> 19 | SafeLoad<uint32_t>(in + 13) << 13, SafeLoad<uint32_t>(in + 13) >> 18 | SafeLoad<uint32_t>(in + 14) << 14, SafeLoad<uint32_t>(in + 14) >> 17 | SafeLoad<uint32_t>(in + 15) << 15 };
-  shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  // extract 31-bit bundles 16 to 31
-  words = simd_batch{ SafeLoad<uint32_t>(in + 15) >> 16 | SafeLoad<uint32_t>(in + 16) << 16, SafeLoad<uint32_t>(in + 16) >> 15 | SafeLoad<uint32_t>(in + 17) << 17, SafeLoad<uint32_t>(in + 17) >> 14 | SafeLoad<uint32_t>(in + 18) << 18, SafeLoad<uint32_t>(in + 18) >> 13 | SafeLoad<uint32_t>(in + 19) << 19, SafeLoad<uint32_t>(in + 19) >> 12 | SafeLoad<uint32_t>(in + 20) << 20, SafeLoad<uint32_t>(in + 20) >> 11 | SafeLoad<uint32_t>(in + 21) << 21, SafeLoad<uint32_t>(in + 21) >> 10 | SafeLoad<uint32_t>(in + 22) << 22, SafeLoad<uint32_t>(in + 22) >> 9 | SafeLoad<uint32_t>(in + 23) << 23, SafeLoad<uint32_t>(in + 23) >> 8 | SafeLoad<uint32_t>(in + 24) << 24, SafeLoad<uint32_t>(in + 24) >> 7 | SafeLoad<uint32_t>(in + 25) << 25, SafeLoad<uint32_t>(in + 25) >> 6 | SafeLoad<uint32_t>(in + 26) << 26, SafeLoad<uint32_t>(in + 26) >> 5 | SafeLoad<uint32_t>(in + 27) << 27, SafeLoad<uint32_t>(in + 27) >> 4 | SafeLoad<uint32_t>(in + 28) << 28, SafeLoad<uint32_t>(in + 28) >> 3 | SafeLoad<uint32_t>(in + 29) << 29, SafeLoad<uint32_t>(in + 29) >> 2 | SafeLoad<uint32_t>(in + 30) << 30, SafeLoad<uint32_t>(in + 30) };
-  shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1 };
-  results = (words >> shifts) & masks;
-  results.store_unaligned(out);
-  out += 16;
-
-  in += 31;
-  return in;
-}
-
-inline static const uint32_t* unpack32_32(const uint32_t* in, uint32_t* out) {
-  memcpy(out, in, 32 * sizeof(*out));
-  in += 32;
-  out += 32;
-
-  return in;
-}
-
-};  // struct UnpackBits512
-
-}  // namespace
-}  // namespace internal
-}  // namespace arrow
-
+namespace arrow::internal {
+
+using ::arrow::util::SafeLoadAs;
+
+template<typename Uint, int BitWidth>
+struct Simd512UnpackerForWidth;
+
+template<>
+struct Simd512UnpackerForWidth<uint32_t, 1> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 16>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x1;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 1-bit bundles 0 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+    };
+    shifts = simd_batch{ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    // extract 1-bit bundles 16 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+    };
+    shifts = simd_batch{ 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    in += 1 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint32_t, 2> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 16>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x3;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 2-bit bundles 0 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+    };
+    shifts = simd_batch{ 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    // extract 2-bit bundles 16 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+    };
+    shifts = simd_batch{ 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    in += 2 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint32_t, 3> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 16>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x7;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 3-bit bundles 0 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 1) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+    };
+    shifts = simd_batch{ 0, 3, 6, 9, 12, 15, 18, 21, 24, 27, 0, 1, 4, 7, 10, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    // extract 3-bit bundles 16 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 31 | SafeLoadAs<uint32_t>(in + 4 * 2) << 1,
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+    };
+    shifts = simd_batch{ 16, 19, 22, 25, 28, 0, 2, 5, 8, 11, 14, 17, 20, 23, 26, 29 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    in += 3 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint32_t, 4> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 16>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0xf;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 4-bit bundles 0 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+    };
+    shifts = simd_batch{ 0, 4, 8, 12, 16, 20, 24, 28, 0, 4, 8, 12, 16, 20, 24, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    // extract 4-bit bundles 16 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+    };
+    shifts = simd_batch{ 0, 4, 8, 12, 16, 20, 24, 28, 0, 4, 8, 12, 16, 20, 24, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    in += 4 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint32_t, 5> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 16>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x1f;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 5-bit bundles 0 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 1) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 2) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+    };
+    shifts = simd_batch{ 0, 5, 10, 15, 20, 25, 0, 3, 8, 13, 18, 23, 0, 1, 6, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    // extract 5-bit bundles 16 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 31 | SafeLoadAs<uint32_t>(in + 4 * 3) << 1,
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 29 | SafeLoadAs<uint32_t>(in + 4 * 4) << 3,
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+    };
+    shifts = simd_batch{ 16, 21, 26, 0, 4, 9, 14, 19, 24, 0, 2, 7, 12, 17, 22, 27 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    in += 5 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint32_t, 6> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 16>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x3f;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 6-bit bundles 0 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 1) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 2) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+    };
+    shifts = simd_batch{ 0, 6, 12, 18, 24, 0, 4, 10, 16, 22, 0, 2, 8, 14, 20, 26 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    // extract 6-bit bundles 16 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 4) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 5) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+    };
+    shifts = simd_batch{ 0, 6, 12, 18, 24, 0, 4, 10, 16, 22, 0, 2, 8, 14, 20, 26 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    in += 6 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint32_t, 7> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 16>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x7f;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 7-bit bundles 0 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 1) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 31 | SafeLoadAs<uint32_t>(in + 4 * 2) << 1,
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 27 | SafeLoadAs<uint32_t>(in + 4 * 3) << 5,
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+    };
+    shifts = simd_batch{ 0, 7, 14, 21, 0, 3, 10, 17, 24, 0, 6, 13, 20, 0, 2, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    // extract 7-bit bundles 16 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 4) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 5) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 29 | SafeLoadAs<uint32_t>(in + 4 * 6) << 3,
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+    };
+    shifts = simd_batch{ 16, 23, 0, 5, 12, 19, 0, 1, 8, 15, 22, 0, 4, 11, 18, 25 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    in += 7 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint32_t, 8> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 16>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0xff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 8-bit bundles 0 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+    };
+    shifts = simd_batch{ 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    // extract 8-bit bundles 16 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+    };
+    shifts = simd_batch{ 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    in += 8 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint32_t, 9> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 16>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x1ff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 9-bit bundles 0 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 27 | SafeLoadAs<uint32_t>(in + 4 * 1) << 5,
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 31 | SafeLoadAs<uint32_t>(in + 4 * 2) << 1,
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 3) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 4) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+    };
+    shifts = simd_batch{ 0, 9, 18, 0, 4, 13, 22, 0, 8, 17, 0, 3, 12, 21, 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    // extract 9-bit bundles 16 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 25 | SafeLoadAs<uint32_t>(in + 4 * 5) << 7,
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 29 | SafeLoadAs<uint32_t>(in + 4 * 6) << 3,
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 7) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 8) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+    };
+    shifts = simd_batch{ 16, 0, 2, 11, 20, 0, 6, 15, 0, 1, 10, 19, 0, 5, 14, 23 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    in += 9 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint32_t, 10> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 16>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x3ff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 10-bit bundles 0 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 1) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 2) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 3) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 4) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+    };
+    shifts = simd_batch{ 0, 10, 20, 0, 8, 18, 0, 6, 16, 0, 4, 14, 0, 2, 12, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    // extract 10-bit bundles 16 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 6) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 7) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 8) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 9) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+    };
+    shifts = simd_batch{ 0, 10, 20, 0, 8, 18, 0, 6, 16, 0, 4, 14, 0, 2, 12, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    in += 10 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint32_t, 11> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 16>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x7ff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 11-bit bundles 0 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 1) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 23 | SafeLoadAs<uint32_t>(in + 4 * 2) << 9,
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 3) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 25 | SafeLoadAs<uint32_t>(in + 4 * 4) << 7,
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 5) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+    };
+    shifts = simd_batch{ 0, 11, 0, 1, 12, 0, 2, 13, 0, 3, 14, 0, 4, 15, 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    // extract 11-bit bundles 16 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 27 | SafeLoadAs<uint32_t>(in + 4 * 6) << 5,
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 7) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 29 | SafeLoadAs<uint32_t>(in + 4 * 8) << 3,
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 9) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 31 | SafeLoadAs<uint32_t>(in + 4 * 10) << 1,
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+    };
+    shifts = simd_batch{ 16, 0, 6, 17, 0, 7, 18, 0, 8, 19, 0, 9, 20, 0, 10, 21 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    in += 11 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint32_t, 12> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 16>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0xfff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 12-bit bundles 0 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 1) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 2) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 4) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 5) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+    };
+    shifts = simd_batch{ 0, 12, 0, 4, 16, 0, 8, 20, 0, 12, 0, 4, 16, 0, 8, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    // extract 12-bit bundles 16 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 7) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 8) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 10) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 11) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 11),
+      SafeLoadAs<uint32_t>(in + 4 * 11),
+    };
+    shifts = simd_batch{ 0, 12, 0, 4, 16, 0, 8, 20, 0, 12, 0, 4, 16, 0, 8, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    in += 12 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint32_t, 13> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 16>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x1fff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 13-bit bundles 0 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 1) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 2) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 27 | SafeLoadAs<uint32_t>(in + 4 * 3) << 5,
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 21 | SafeLoadAs<uint32_t>(in + 4 * 4) << 11,
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 5) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 6) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+    };
+    shifts = simd_batch{ 0, 13, 0, 7, 0, 1, 14, 0, 8, 0, 2, 15, 0, 9, 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    // extract 13-bit bundles 16 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 29 | SafeLoadAs<uint32_t>(in + 4 * 7) << 3,
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 23 | SafeLoadAs<uint32_t>(in + 4 * 8) << 9,
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 9) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 10) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 31 | SafeLoadAs<uint32_t>(in + 4 * 11) << 1,
+      SafeLoadAs<uint32_t>(in + 4 * 11),
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 25 | SafeLoadAs<uint32_t>(in + 4 * 12) << 7,
+      SafeLoadAs<uint32_t>(in + 4 * 12),
+      SafeLoadAs<uint32_t>(in + 4 * 12),
+    };
+    shifts = simd_batch{ 16, 0, 10, 0, 4, 17, 0, 11, 0, 5, 18, 0, 12, 0, 6, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    in += 13 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint32_t, 14> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 16>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x3fff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 14-bit bundles 0 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 1) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 2) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 3) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 4) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 5) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 6) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+    };
+    shifts = simd_batch{ 0, 14, 0, 10, 0, 6, 0, 2, 16, 0, 12, 0, 8, 0, 4, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    // extract 14-bit bundles 16 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 8) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 9) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 10) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 11) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 11),
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 12) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 12),
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 13) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 13),
+      SafeLoadAs<uint32_t>(in + 4 * 13),
+    };
+    shifts = simd_batch{ 0, 14, 0, 10, 0, 6, 0, 2, 16, 0, 12, 0, 8, 0, 4, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    in += 14 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint32_t, 15> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 16>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x7fff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 15-bit bundles 0 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 1) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 2) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 3) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 4) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 5) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 6) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 7) << 14,
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+    };
+    shifts = simd_batch{ 0, 15, 0, 13, 0, 11, 0, 9, 0, 7, 0, 5, 0, 3, 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    // extract 15-bit bundles 16 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 31 | SafeLoadAs<uint32_t>(in + 4 * 8) << 1,
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 29 | SafeLoadAs<uint32_t>(in + 4 * 9) << 3,
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 27 | SafeLoadAs<uint32_t>(in + 4 * 10) << 5,
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 25 | SafeLoadAs<uint32_t>(in + 4 * 11) << 7,
+      SafeLoadAs<uint32_t>(in + 4 * 11),
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 23 | SafeLoadAs<uint32_t>(in + 4 * 12) << 9,
+      SafeLoadAs<uint32_t>(in + 4 * 12),
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 21 | SafeLoadAs<uint32_t>(in + 4 * 13) << 11,
+      SafeLoadAs<uint32_t>(in + 4 * 13),
+      SafeLoadAs<uint32_t>(in + 4 * 13) >> 19 | SafeLoadAs<uint32_t>(in + 4 * 14) << 13,
+      SafeLoadAs<uint32_t>(in + 4 * 14),
+      SafeLoadAs<uint32_t>(in + 4 * 14),
+    };
+    shifts = simd_batch{ 16, 0, 14, 0, 12, 0, 10, 0, 8, 0, 6, 0, 4, 0, 2, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    in += 15 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint32_t, 16> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 16>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0xffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 16-bit bundles 0 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+    };
+    shifts = simd_batch{ 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    // extract 16-bit bundles 16 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+      SafeLoadAs<uint32_t>(in + 4 * 11),
+      SafeLoadAs<uint32_t>(in + 4 * 11),
+      SafeLoadAs<uint32_t>(in + 4 * 12),
+      SafeLoadAs<uint32_t>(in + 4 * 12),
+      SafeLoadAs<uint32_t>(in + 4 * 13),
+      SafeLoadAs<uint32_t>(in + 4 * 13),
+      SafeLoadAs<uint32_t>(in + 4 * 14),
+      SafeLoadAs<uint32_t>(in + 4 * 14),
+      SafeLoadAs<uint32_t>(in + 4 * 15),
+      SafeLoadAs<uint32_t>(in + 4 * 15),
+    };
+    shifts = simd_batch{ 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    in += 16 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint32_t, 17> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 16>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x1ffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 17-bit bundles 0 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 17 | SafeLoadAs<uint32_t>(in + 4 * 1) << 15,
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 19 | SafeLoadAs<uint32_t>(in + 4 * 2) << 13,
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 21 | SafeLoadAs<uint32_t>(in + 4 * 3) << 11,
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 23 | SafeLoadAs<uint32_t>(in + 4 * 4) << 9,
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 25 | SafeLoadAs<uint32_t>(in + 4 * 5) << 7,
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 27 | SafeLoadAs<uint32_t>(in + 4 * 6) << 5,
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 29 | SafeLoadAs<uint32_t>(in + 4 * 7) << 3,
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 31 | SafeLoadAs<uint32_t>(in + 4 * 8) << 1,
+    };
+    shifts = simd_batch{ 0, 0, 2, 0, 4, 0, 6, 0, 8, 0, 10, 0, 12, 0, 14, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    // extract 17-bit bundles 16 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 9) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 10) << 14,
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 11) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 11),
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 12) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 12),
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 13) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 13),
+      SafeLoadAs<uint32_t>(in + 4 * 13) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 14) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 14),
+      SafeLoadAs<uint32_t>(in + 4 * 14) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 15) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 15),
+      SafeLoadAs<uint32_t>(in + 4 * 15) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 16) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 16),
+    };
+    shifts = simd_batch{ 0, 1, 0, 3, 0, 5, 0, 7, 0, 9, 0, 11, 0, 13, 0, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    in += 17 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint32_t, 18> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 16>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x3ffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 18-bit bundles 0 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 1) << 14,
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 2) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 3) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 4) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 5) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 6) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 7) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 8) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+    };
+    shifts = simd_batch{ 0, 0, 4, 0, 8, 0, 12, 0, 0, 2, 0, 6, 0, 10, 0, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    // extract 18-bit bundles 16 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 10) << 14,
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 11) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 11),
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 12) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 12),
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 13) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 13) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 14) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 14),
+      SafeLoadAs<uint32_t>(in + 4 * 14) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 15) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 15),
+      SafeLoadAs<uint32_t>(in + 4 * 15) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 16) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 16),
+      SafeLoadAs<uint32_t>(in + 4 * 16) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 17) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 17),
+    };
+    shifts = simd_batch{ 0, 0, 4, 0, 8, 0, 12, 0, 0, 2, 0, 6, 0, 10, 0, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    in += 18 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint32_t, 19> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 16>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x7ffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 19-bit bundles 0 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 19 | SafeLoadAs<uint32_t>(in + 4 * 1) << 13,
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 25 | SafeLoadAs<uint32_t>(in + 4 * 2) << 7,
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 31 | SafeLoadAs<uint32_t>(in + 4 * 3) << 1,
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 4) << 14,
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 5) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 6) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 17 | SafeLoadAs<uint32_t>(in + 4 * 7) << 15,
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 23 | SafeLoadAs<uint32_t>(in + 4 * 8) << 9,
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 29 | SafeLoadAs<uint32_t>(in + 4 * 9) << 3,
+    };
+    shifts = simd_batch{ 0, 0, 6, 0, 12, 0, 0, 5, 0, 11, 0, 0, 4, 0, 10, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    // extract 19-bit bundles 16 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 10) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 11) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 11),
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 12) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 15 | SafeLoadAs<uint32_t>(in + 4 * 13) << 17,
+      SafeLoadAs<uint32_t>(in + 4 * 13),
+      SafeLoadAs<uint32_t>(in + 4 * 13) >> 21 | SafeLoadAs<uint32_t>(in + 4 * 14) << 11,
+      SafeLoadAs<uint32_t>(in + 4 * 14),
+      SafeLoadAs<uint32_t>(in + 4 * 14) >> 27 | SafeLoadAs<uint32_t>(in + 4 * 15) << 5,
+      SafeLoadAs<uint32_t>(in + 4 * 15) >> 14 | SafeLoadAs<uint32_t>(in + 4 * 16) << 18,
+      SafeLoadAs<uint32_t>(in + 4 * 16),
+      SafeLoadAs<uint32_t>(in + 4 * 16) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 17) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 17),
+      SafeLoadAs<uint32_t>(in + 4 * 17) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 18) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 18),
+    };
+    shifts = simd_batch{ 0, 3, 0, 9, 0, 0, 2, 0, 8, 0, 0, 1, 0, 7, 0, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    in += 19 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint32_t, 20> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 16>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0xfffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 20-bit bundles 0 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 1) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 2) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 3) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 4) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 6) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 7) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 8) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 9) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+    };
+    shifts = simd_batch{ 0, 0, 8, 0, 0, 4, 0, 12, 0, 0, 8, 0, 0, 4, 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    // extract 20-bit bundles 16 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 11) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 11),
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 12) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 13) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 13),
+      SafeLoadAs<uint32_t>(in + 4 * 13) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 14) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 14),
+      SafeLoadAs<uint32_t>(in + 4 * 15),
+      SafeLoadAs<uint32_t>(in + 4 * 15) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 16) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 16),
+      SafeLoadAs<uint32_t>(in + 4 * 16) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 17) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 17) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 18) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 18),
+      SafeLoadAs<uint32_t>(in + 4 * 18) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 19) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 19),
+    };
+    shifts = simd_batch{ 0, 0, 8, 0, 0, 4, 0, 12, 0, 0, 8, 0, 0, 4, 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    in += 20 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint32_t, 21> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 16>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x1fffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 21-bit bundles 0 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 21 | SafeLoadAs<uint32_t>(in + 4 * 1) << 11,
+      SafeLoadAs<uint32_t>(in + 4 * 1),
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 31 | SafeLoadAs<uint32_t>(in + 4 * 2) << 1,
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 3) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 4) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 19 | SafeLoadAs<uint32_t>(in + 4 * 5) << 13,
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 29 | SafeLoadAs<uint32_t>(in + 4 * 6) << 3,
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 7) << 14,
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 8) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 17 | SafeLoadAs<uint32_t>(in + 4 * 9) << 15,
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 27 | SafeLoadAs<uint32_t>(in + 4 * 10) << 5,
+    };
+    shifts = simd_batch{ 0, 0, 10, 0, 0, 9, 0, 0, 8, 0, 0, 7, 0, 0, 6, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    // extract 21-bit bundles 16 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 11) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 11),
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 12) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 15 | SafeLoadAs<uint32_t>(in + 4 * 13) << 17,
+      SafeLoadAs<uint32_t>(in + 4 * 13),
+      SafeLoadAs<uint32_t>(in + 4 * 13) >> 25 | SafeLoadAs<uint32_t>(in + 4 * 14) << 7,
+      SafeLoadAs<uint32_t>(in + 4 * 14) >> 14 | SafeLoadAs<uint32_t>(in + 4 * 15) << 18,
+      SafeLoadAs<uint32_t>(in + 4 * 15),
+      SafeLoadAs<uint32_t>(in + 4 * 15) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 16) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 16) >> 13 | SafeLoadAs<uint32_t>(in + 4 * 17) << 19,
+      SafeLoadAs<uint32_t>(in + 4 * 17),
+      SafeLoadAs<uint32_t>(in + 4 * 17) >> 23 | SafeLoadAs<uint32_t>(in + 4 * 18) << 9,
+      SafeLoadAs<uint32_t>(in + 4 * 18) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 19) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 19),
+      SafeLoadAs<uint32_t>(in + 4 * 19) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 20) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 20),
+    };
+    shifts = simd_batch{ 0, 5, 0, 0, 4, 0, 0, 3, 0, 0, 2, 0, 0, 1, 0, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    in += 21 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint32_t, 22> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 16>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x3fffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 22-bit bundles 0 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 1) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 2) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 3) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 14 | SafeLoadAs<uint32_t>(in + 4 * 4) << 18,
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 5) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 6) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 7) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 8) << 14,
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 9) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 10) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+    };
+    shifts = simd_batch{ 0, 0, 0, 2, 0, 0, 4, 0, 0, 6, 0, 0, 8, 0, 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    // extract 22-bit bundles 16 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 11),
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 12) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 13) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 13),
+      SafeLoadAs<uint32_t>(in + 4 * 13) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 14) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 14) >> 14 | SafeLoadAs<uint32_t>(in + 4 * 15) << 18,
+      SafeLoadAs<uint32_t>(in + 4 * 15),
+      SafeLoadAs<uint32_t>(in + 4 * 15) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 16) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 16) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 17) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 17),
+      SafeLoadAs<uint32_t>(in + 4 * 17) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 18) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 18) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 19) << 14,
+      SafeLoadAs<uint32_t>(in + 4 * 19),
+      SafeLoadAs<uint32_t>(in + 4 * 19) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 20) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 20) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 21) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 21),
+    };
+    shifts = simd_batch{ 0, 0, 0, 2, 0, 0, 4, 0, 0, 6, 0, 0, 8, 0, 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    in += 22 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint32_t, 23> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 16>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x7fffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 23-bit bundles 0 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 23 | SafeLoadAs<uint32_t>(in + 4 * 1) << 9,
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 14 | SafeLoadAs<uint32_t>(in + 4 * 2) << 18,
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 3) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 19 | SafeLoadAs<uint32_t>(in + 4 * 4) << 13,
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 10 | SafeLoadAs<uint32_t>(in + 4 * 5) << 22,
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 6) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 15 | SafeLoadAs<uint32_t>(in + 4 * 7) << 17,
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 29 | SafeLoadAs<uint32_t>(in + 4 * 8) << 3,
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 9) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 11 | SafeLoadAs<uint32_t>(in + 4 * 10) << 21,
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 25 | SafeLoadAs<uint32_t>(in + 4 * 11) << 7,
+    };
+    shifts = simd_batch{ 0, 0, 0, 5, 0, 0, 0, 1, 0, 0, 6, 0, 0, 0, 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    // extract 23-bit bundles 16 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 12) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 12),
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 13) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 13) >> 21 | SafeLoadAs<uint32_t>(in + 4 * 14) << 11,
+      SafeLoadAs<uint32_t>(in + 4 * 14) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 15) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 15),
+      SafeLoadAs<uint32_t>(in + 4 * 15) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 16) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 16) >> 17 | SafeLoadAs<uint32_t>(in + 4 * 17) << 15,
+      SafeLoadAs<uint32_t>(in + 4 * 17),
+      SafeLoadAs<uint32_t>(in + 4 * 17) >> 31 | SafeLoadAs<uint32_t>(in + 4 * 18) << 1,
+      SafeLoadAs<uint32_t>(in + 4 * 18) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 19) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 19) >> 13 | SafeLoadAs<uint32_t>(in + 4 * 20) << 19,
+      SafeLoadAs<uint32_t>(in + 4 * 20),
+      SafeLoadAs<uint32_t>(in + 4 * 20) >> 27 | SafeLoadAs<uint32_t>(in + 4 * 21) << 5,
+      SafeLoadAs<uint32_t>(in + 4 * 21) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 22) << 14,
+      SafeLoadAs<uint32_t>(in + 4 * 22),
+    };
+    shifts = simd_batch{ 0, 7, 0, 0, 0, 3, 0, 0, 8, 0, 0, 0, 4, 0, 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    in += 23 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint32_t, 24> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 16>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0xffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 24-bit bundles 0 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 1) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 2) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 2),
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 4) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 5) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 7) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 8) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 10) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 11) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 11),
+    };
+    shifts = simd_batch{ 0, 0, 0, 8, 0, 0, 0, 8, 0, 0, 0, 8, 0, 0, 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    // extract 24-bit bundles 16 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 12),
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 13) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 13) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 14) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 14),
+      SafeLoadAs<uint32_t>(in + 4 * 15),
+      SafeLoadAs<uint32_t>(in + 4 * 15) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 16) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 16) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 17) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 17),
+      SafeLoadAs<uint32_t>(in + 4 * 18),
+      SafeLoadAs<uint32_t>(in + 4 * 18) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 19) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 19) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 20) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 20),
+      SafeLoadAs<uint32_t>(in + 4 * 21),
+      SafeLoadAs<uint32_t>(in + 4 * 21) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 22) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 22) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 23) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 23),
+    };
+    shifts = simd_batch{ 0, 0, 0, 8, 0, 0, 0, 8, 0, 0, 0, 8, 0, 0, 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    in += 24 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint32_t, 25> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 16>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x1ffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 25-bit bundles 0 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 25 | SafeLoadAs<uint32_t>(in + 4 * 1) << 7,
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 2) << 14,
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 11 | SafeLoadAs<uint32_t>(in + 4 * 3) << 21,
+      SafeLoadAs<uint32_t>(in + 4 * 3),
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 29 | SafeLoadAs<uint32_t>(in + 4 * 4) << 3,
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 5) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 15 | SafeLoadAs<uint32_t>(in + 4 * 6) << 17,
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 8 | SafeLoadAs<uint32_t>(in + 4 * 7) << 24,
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 8) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 19 | SafeLoadAs<uint32_t>(in + 4 * 9) << 13,
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 10) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 11) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 23 | SafeLoadAs<uint32_t>(in + 4 * 12) << 9,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 4, 0, 0, 0, 0, 1, 0, 0, 0, 5, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    // extract 25-bit bundles 16 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 13) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 13) >> 9 | SafeLoadAs<uint32_t>(in + 4 * 14) << 23,
+      SafeLoadAs<uint32_t>(in + 4 * 14),
+      SafeLoadAs<uint32_t>(in + 4 * 14) >> 27 | SafeLoadAs<uint32_t>(in + 4 * 15) << 5,
+      SafeLoadAs<uint32_t>(in + 4 * 15) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 16) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 16) >> 13 | SafeLoadAs<uint32_t>(in + 4 * 17) << 19,
+      SafeLoadAs<uint32_t>(in + 4 * 17),
+      SafeLoadAs<uint32_t>(in + 4 * 17) >> 31 | SafeLoadAs<uint32_t>(in + 4 * 18) << 1,
+      SafeLoadAs<uint32_t>(in + 4 * 18) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 19) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 19) >> 17 | SafeLoadAs<uint32_t>(in + 4 * 20) << 15,
+      SafeLoadAs<uint32_t>(in + 4 * 20) >> 10 | SafeLoadAs<uint32_t>(in + 4 * 21) << 22,
+      SafeLoadAs<uint32_t>(in + 4 * 21),
+      SafeLoadAs<uint32_t>(in + 4 * 21) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 22) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 22) >> 21 | SafeLoadAs<uint32_t>(in + 4 * 23) << 11,
+      SafeLoadAs<uint32_t>(in + 4 * 23) >> 14 | SafeLoadAs<uint32_t>(in + 4 * 24) << 18,
+      SafeLoadAs<uint32_t>(in + 4 * 24),
+    };
+    shifts = simd_batch{ 0, 0, 2, 0, 0, 0, 6, 0, 0, 0, 0, 3, 0, 0, 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    in += 25 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint32_t, 26> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 16>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x3ffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 26-bit bundles 0 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 1) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 2) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 14 | SafeLoadAs<uint32_t>(in + 4 * 3) << 18,
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 8 | SafeLoadAs<uint32_t>(in + 4 * 4) << 24,
+      SafeLoadAs<uint32_t>(in + 4 * 4),
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 5) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 6) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 7) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 10 | SafeLoadAs<uint32_t>(in + 4 * 8) << 22,
+      SafeLoadAs<uint32_t>(in + 4 * 8),
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 9) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 10) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 11) << 14,
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 12) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 12),
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 4, 0, 0, 0, 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    // extract 26-bit bundles 16 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 13),
+      SafeLoadAs<uint32_t>(in + 4 * 13) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 14) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 14) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 15) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 15) >> 14 | SafeLoadAs<uint32_t>(in + 4 * 16) << 18,
+      SafeLoadAs<uint32_t>(in + 4 * 16) >> 8 | SafeLoadAs<uint32_t>(in + 4 * 17) << 24,
+      SafeLoadAs<uint32_t>(in + 4 * 17),
+      SafeLoadAs<uint32_t>(in + 4 * 17) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 18) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 18) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 19) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 19) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 20) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 20) >> 10 | SafeLoadAs<uint32_t>(in + 4 * 21) << 22,
+      SafeLoadAs<uint32_t>(in + 4 * 21),
+      SafeLoadAs<uint32_t>(in + 4 * 21) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 22) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 22) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 23) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 23) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 24) << 14,
+      SafeLoadAs<uint32_t>(in + 4 * 24) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 25) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 25),
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 4, 0, 0, 0, 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    in += 26 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint32_t, 27> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 16>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x7ffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 27-bit bundles 0 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 27 | SafeLoadAs<uint32_t>(in + 4 * 1) << 5,
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 2) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 17 | SafeLoadAs<uint32_t>(in + 4 * 3) << 15,
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 4) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 7 | SafeLoadAs<uint32_t>(in + 4 * 5) << 25,
+      SafeLoadAs<uint32_t>(in + 4 * 5),
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 29 | SafeLoadAs<uint32_t>(in + 4 * 6) << 3,
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 7) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 19 | SafeLoadAs<uint32_t>(in + 4 * 8) << 13,
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 14 | SafeLoadAs<uint32_t>(in + 4 * 9) << 18,
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 9 | SafeLoadAs<uint32_t>(in + 4 * 10) << 23,
+      SafeLoadAs<uint32_t>(in + 4 * 10),
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 31 | SafeLoadAs<uint32_t>(in + 4 * 11) << 1,
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 12) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 21 | SafeLoadAs<uint32_t>(in + 4 * 13) << 11,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 4, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    // extract 27-bit bundles 16 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 13) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 14) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 14) >> 11 | SafeLoadAs<uint32_t>(in + 4 * 15) << 21,
+      SafeLoadAs<uint32_t>(in + 4 * 15) >> 6 | SafeLoadAs<uint32_t>(in + 4 * 16) << 26,
+      SafeLoadAs<uint32_t>(in + 4 * 16),
+      SafeLoadAs<uint32_t>(in + 4 * 16) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 17) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 17) >> 23 | SafeLoadAs<uint32_t>(in + 4 * 18) << 9,
+      SafeLoadAs<uint32_t>(in + 4 * 18) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 19) << 14,
+      SafeLoadAs<uint32_t>(in + 4 * 19) >> 13 | SafeLoadAs<uint32_t>(in + 4 * 20) << 19,
+      SafeLoadAs<uint32_t>(in + 4 * 20) >> 8 | SafeLoadAs<uint32_t>(in + 4 * 21) << 24,
+      SafeLoadAs<uint32_t>(in + 4 * 21),
+      SafeLoadAs<uint32_t>(in + 4 * 21) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 22) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 22) >> 25 | SafeLoadAs<uint32_t>(in + 4 * 23) << 7,
+      SafeLoadAs<uint32_t>(in + 4 * 23) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 24) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 24) >> 15 | SafeLoadAs<uint32_t>(in + 4 * 25) << 17,
+      SafeLoadAs<uint32_t>(in + 4 * 25) >> 10 | SafeLoadAs<uint32_t>(in + 4 * 26) << 22,
+      SafeLoadAs<uint32_t>(in + 4 * 26),
+    };
+    shifts = simd_batch{ 0, 0, 0, 1, 0, 0, 0, 0, 0, 3, 0, 0, 0, 0, 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    in += 27 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint32_t, 28> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 16>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0xfffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 28-bit bundles 0 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 1) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 2) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 3) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 4) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 5) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 8 | SafeLoadAs<uint32_t>(in + 4 * 6) << 24,
+      SafeLoadAs<uint32_t>(in + 4 * 6),
+      SafeLoadAs<uint32_t>(in + 4 * 7),
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 8) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 9) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 10) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 11) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 12) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 8 | SafeLoadAs<uint32_t>(in + 4 * 13) << 24,
+      SafeLoadAs<uint32_t>(in + 4 * 13),
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 4, 0, 0, 0, 0, 0, 0, 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    // extract 28-bit bundles 16 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 14),
+      SafeLoadAs<uint32_t>(in + 4 * 14) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 15) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 15) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 16) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 16) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 17) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 17) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 18) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 18) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 19) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 19) >> 8 | SafeLoadAs<uint32_t>(in + 4 * 20) << 24,
+      SafeLoadAs<uint32_t>(in + 4 * 20),
+      SafeLoadAs<uint32_t>(in + 4 * 21),
+      SafeLoadAs<uint32_t>(in + 4 * 21) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 22) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 22) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 23) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 23) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 24) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 24) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 25) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 25) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 26) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 26) >> 8 | SafeLoadAs<uint32_t>(in + 4 * 27) << 24,
+      SafeLoadAs<uint32_t>(in + 4 * 27),
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 4, 0, 0, 0, 0, 0, 0, 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    in += 28 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint32_t, 29> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 16>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x1fffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 29-bit bundles 0 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 29 | SafeLoadAs<uint32_t>(in + 4 * 1) << 3,
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 2) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 23 | SafeLoadAs<uint32_t>(in + 4 * 3) << 9,
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 4) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 17 | SafeLoadAs<uint32_t>(in + 4 * 5) << 15,
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 14 | SafeLoadAs<uint32_t>(in + 4 * 6) << 18,
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 11 | SafeLoadAs<uint32_t>(in + 4 * 7) << 21,
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 8 | SafeLoadAs<uint32_t>(in + 4 * 8) << 24,
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 5 | SafeLoadAs<uint32_t>(in + 4 * 9) << 27,
+      SafeLoadAs<uint32_t>(in + 4 * 9),
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 31 | SafeLoadAs<uint32_t>(in + 4 * 10) << 1,
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 11) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 25 | SafeLoadAs<uint32_t>(in + 4 * 12) << 7,
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 13) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 13) >> 19 | SafeLoadAs<uint32_t>(in + 4 * 14) << 13,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    // extract 29-bit bundles 16 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 14) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 15) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 15) >> 13 | SafeLoadAs<uint32_t>(in + 4 * 16) << 19,
+      SafeLoadAs<uint32_t>(in + 4 * 16) >> 10 | SafeLoadAs<uint32_t>(in + 4 * 17) << 22,
+      SafeLoadAs<uint32_t>(in + 4 * 17) >> 7 | SafeLoadAs<uint32_t>(in + 4 * 18) << 25,
+      SafeLoadAs<uint32_t>(in + 4 * 18) >> 4 | SafeLoadAs<uint32_t>(in + 4 * 19) << 28,
+      SafeLoadAs<uint32_t>(in + 4 * 19),
+      SafeLoadAs<uint32_t>(in + 4 * 19) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 20) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 20) >> 27 | SafeLoadAs<uint32_t>(in + 4 * 21) << 5,
+      SafeLoadAs<uint32_t>(in + 4 * 21) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 22) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 22) >> 21 | SafeLoadAs<uint32_t>(in + 4 * 23) << 11,
+      SafeLoadAs<uint32_t>(in + 4 * 23) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 24) << 14,
+      SafeLoadAs<uint32_t>(in + 4 * 24) >> 15 | SafeLoadAs<uint32_t>(in + 4 * 25) << 17,
+      SafeLoadAs<uint32_t>(in + 4 * 25) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 26) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 26) >> 9 | SafeLoadAs<uint32_t>(in + 4 * 27) << 23,
+      SafeLoadAs<uint32_t>(in + 4 * 27) >> 6 | SafeLoadAs<uint32_t>(in + 4 * 28) << 26,
+      SafeLoadAs<uint32_t>(in + 4 * 28),
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    in += 29 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint32_t, 30> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 16>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x3fffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 30-bit bundles 0 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 1) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 2) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 3) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 4) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 5) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 6) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 7) << 14,
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 8) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 14 | SafeLoadAs<uint32_t>(in + 4 * 9) << 18,
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 10) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 10 | SafeLoadAs<uint32_t>(in + 4 * 11) << 22,
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 8 | SafeLoadAs<uint32_t>(in + 4 * 12) << 24,
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 6 | SafeLoadAs<uint32_t>(in + 4 * 13) << 26,
+      SafeLoadAs<uint32_t>(in + 4 * 13) >> 4 | SafeLoadAs<uint32_t>(in + 4 * 14) << 28,
+      SafeLoadAs<uint32_t>(in + 4 * 14),
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    // extract 30-bit bundles 16 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 15),
+      SafeLoadAs<uint32_t>(in + 4 * 15) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 16) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 16) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 17) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 17) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 18) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 18) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 19) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 19) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 20) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 20) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 21) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 21) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 22) << 14,
+      SafeLoadAs<uint32_t>(in + 4 * 22) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 23) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 23) >> 14 | SafeLoadAs<uint32_t>(in + 4 * 24) << 18,
+      SafeLoadAs<uint32_t>(in + 4 * 24) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 25) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 25) >> 10 | SafeLoadAs<uint32_t>(in + 4 * 26) << 22,
+      SafeLoadAs<uint32_t>(in + 4 * 26) >> 8 | SafeLoadAs<uint32_t>(in + 4 * 27) << 24,
+      SafeLoadAs<uint32_t>(in + 4 * 27) >> 6 | SafeLoadAs<uint32_t>(in + 4 * 28) << 26,
+      SafeLoadAs<uint32_t>(in + 4 * 28) >> 4 | SafeLoadAs<uint32_t>(in + 4 * 29) << 28,
+      SafeLoadAs<uint32_t>(in + 4 * 29),
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    in += 30 * 4;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint32_t, 31> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint32_t, 16>;
+  static constexpr int kValuesUnpacked = 32;
+
+  static const uint8_t* unpack(const uint8_t* in, uint32_t* out) {
+    constexpr uint32_t kMask = 0x7fffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 31-bit bundles 0 to 15
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 0),
+      SafeLoadAs<uint32_t>(in + 4 * 0) >> 31 | SafeLoadAs<uint32_t>(in + 4 * 1) << 1,
+      SafeLoadAs<uint32_t>(in + 4 * 1) >> 30 | SafeLoadAs<uint32_t>(in + 4 * 2) << 2,
+      SafeLoadAs<uint32_t>(in + 4 * 2) >> 29 | SafeLoadAs<uint32_t>(in + 4 * 3) << 3,
+      SafeLoadAs<uint32_t>(in + 4 * 3) >> 28 | SafeLoadAs<uint32_t>(in + 4 * 4) << 4,
+      SafeLoadAs<uint32_t>(in + 4 * 4) >> 27 | SafeLoadAs<uint32_t>(in + 4 * 5) << 5,
+      SafeLoadAs<uint32_t>(in + 4 * 5) >> 26 | SafeLoadAs<uint32_t>(in + 4 * 6) << 6,
+      SafeLoadAs<uint32_t>(in + 4 * 6) >> 25 | SafeLoadAs<uint32_t>(in + 4 * 7) << 7,
+      SafeLoadAs<uint32_t>(in + 4 * 7) >> 24 | SafeLoadAs<uint32_t>(in + 4 * 8) << 8,
+      SafeLoadAs<uint32_t>(in + 4 * 8) >> 23 | SafeLoadAs<uint32_t>(in + 4 * 9) << 9,
+      SafeLoadAs<uint32_t>(in + 4 * 9) >> 22 | SafeLoadAs<uint32_t>(in + 4 * 10) << 10,
+      SafeLoadAs<uint32_t>(in + 4 * 10) >> 21 | SafeLoadAs<uint32_t>(in + 4 * 11) << 11,
+      SafeLoadAs<uint32_t>(in + 4 * 11) >> 20 | SafeLoadAs<uint32_t>(in + 4 * 12) << 12,
+      SafeLoadAs<uint32_t>(in + 4 * 12) >> 19 | SafeLoadAs<uint32_t>(in + 4 * 13) << 13,
+      SafeLoadAs<uint32_t>(in + 4 * 13) >> 18 | SafeLoadAs<uint32_t>(in + 4 * 14) << 14,
+      SafeLoadAs<uint32_t>(in + 4 * 14) >> 17 | SafeLoadAs<uint32_t>(in + 4 * 15) << 15,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    // extract 31-bit bundles 16 to 31
+    words = simd_batch{
+      SafeLoadAs<uint32_t>(in + 4 * 15) >> 16 | SafeLoadAs<uint32_t>(in + 4 * 16) << 16,
+      SafeLoadAs<uint32_t>(in + 4 * 16) >> 15 | SafeLoadAs<uint32_t>(in + 4 * 17) << 17,
+      SafeLoadAs<uint32_t>(in + 4 * 17) >> 14 | SafeLoadAs<uint32_t>(in + 4 * 18) << 18,
+      SafeLoadAs<uint32_t>(in + 4 * 18) >> 13 | SafeLoadAs<uint32_t>(in + 4 * 19) << 19,
+      SafeLoadAs<uint32_t>(in + 4 * 19) >> 12 | SafeLoadAs<uint32_t>(in + 4 * 20) << 20,
+      SafeLoadAs<uint32_t>(in + 4 * 20) >> 11 | SafeLoadAs<uint32_t>(in + 4 * 21) << 21,
+      SafeLoadAs<uint32_t>(in + 4 * 21) >> 10 | SafeLoadAs<uint32_t>(in + 4 * 22) << 22,
+      SafeLoadAs<uint32_t>(in + 4 * 22) >> 9 | SafeLoadAs<uint32_t>(in + 4 * 23) << 23,
+      SafeLoadAs<uint32_t>(in + 4 * 23) >> 8 | SafeLoadAs<uint32_t>(in + 4 * 24) << 24,
+      SafeLoadAs<uint32_t>(in + 4 * 24) >> 7 | SafeLoadAs<uint32_t>(in + 4 * 25) << 25,
+      SafeLoadAs<uint32_t>(in + 4 * 25) >> 6 | SafeLoadAs<uint32_t>(in + 4 * 26) << 26,
+      SafeLoadAs<uint32_t>(in + 4 * 26) >> 5 | SafeLoadAs<uint32_t>(in + 4 * 27) << 27,
+      SafeLoadAs<uint32_t>(in + 4 * 27) >> 4 | SafeLoadAs<uint32_t>(in + 4 * 28) << 28,
+      SafeLoadAs<uint32_t>(in + 4 * 28) >> 3 | SafeLoadAs<uint32_t>(in + 4 * 29) << 29,
+      SafeLoadAs<uint32_t>(in + 4 * 29) >> 2 | SafeLoadAs<uint32_t>(in + 4 * 30) << 30,
+      SafeLoadAs<uint32_t>(in + 4 * 30),
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 16;
+
+    in += 31 * 4;
+    return in;
+  }
+};
+
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 1> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 1-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 1, 2, 3, 4, 5, 6, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 1-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 8, 9, 10, 11, 12, 13, 14, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 1-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 16, 17, 18, 19, 20, 21, 22, 23 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 1-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 24, 25, 26, 27, 28, 29, 30, 31 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 1-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 32, 33, 34, 35, 36, 37, 38, 39 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 1-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 40, 41, 42, 43, 44, 45, 46, 47 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 1-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 48, 49, 50, 51, 52, 53, 54, 55 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 1-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 56, 57, 58, 59, 60, 61, 62, 63 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 1 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 2> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 2-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 2, 4, 6, 8, 10, 12, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 2-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 16, 18, 20, 22, 24, 26, 28, 30 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 2-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 32, 34, 36, 38, 40, 42, 44, 46 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 2-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 48, 50, 52, 54, 56, 58, 60, 62 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 2-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 2, 4, 6, 8, 10, 12, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 2-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 16, 18, 20, 22, 24, 26, 28, 30 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 2-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 32, 34, 36, 38, 40, 42, 44, 46 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 2-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 48, 50, 52, 54, 56, 58, 60, 62 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 2 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 3> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 3-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 3, 6, 9, 12, 15, 18, 21 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 3-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 24, 27, 30, 33, 36, 39, 42, 45 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 3-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 1) << 1,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 48, 51, 54, 57, 60, 0, 2, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 3-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 8, 11, 14, 17, 20, 23, 26, 29 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 3-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 32, 35, 38, 41, 44, 47, 50, 53 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 3-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 2) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 56, 59, 0, 1, 4, 7, 10, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 3-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 16, 19, 22, 25, 28, 31, 34, 37 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 3-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 40, 43, 46, 49, 52, 55, 58, 61 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 3 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 4> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0xf;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 4-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 4, 8, 12, 16, 20, 24, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 4-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 32, 36, 40, 44, 48, 52, 56, 60 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 4-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 4, 8, 12, 16, 20, 24, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 4-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 32, 36, 40, 44, 48, 52, 56, 60 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 4-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 0, 4, 8, 12, 16, 20, 24, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 4-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 32, 36, 40, 44, 48, 52, 56, 60 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 4-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 0, 4, 8, 12, 16, 20, 24, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 4-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 32, 36, 40, 44, 48, 52, 56, 60 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 4 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 5> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1f;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 5-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 5, 10, 15, 20, 25, 30, 35 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 5-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 1) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 40, 45, 50, 55, 0, 1, 6, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 5-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 16, 21, 26, 31, 36, 41, 46, 51 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 5-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 2) << 3,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 56, 0, 2, 7, 12, 17, 22, 27 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 5-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 3) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 32, 37, 42, 47, 52, 57, 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 5-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 8, 13, 18, 23, 28, 33, 38, 43 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 5-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 4) << 1,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 48, 53, 58, 0, 4, 9, 14, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 5-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 24, 29, 34, 39, 44, 49, 54, 59 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 5 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 6> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3f;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 6-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 6, 12, 18, 24, 30, 36, 42 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 6-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 1) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 48, 54, 0, 2, 8, 14, 20, 26 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 6-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 2) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 32, 38, 44, 50, 56, 0, 4, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 6-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 16, 22, 28, 34, 40, 46, 52, 58 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 6-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 0, 6, 12, 18, 24, 30, 36, 42 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 6-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 4) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 48, 54, 0, 2, 8, 14, 20, 26 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 6-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 5) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 32, 38, 44, 50, 56, 0, 4, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 6-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 16, 22, 28, 34, 40, 46, 52, 58 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 6 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 7> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7f;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 7-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 7, 14, 21, 28, 35, 42, 49 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 7-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 1) << 1,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 56, 0, 6, 13, 20, 27, 34, 41 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 7-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 2) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 48, 55, 0, 5, 12, 19, 26, 33 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 7-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 3) << 3,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 40, 47, 54, 0, 4, 11, 18, 25 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 7-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 4) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 32, 39, 46, 53, 0, 3, 10, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 7-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 5) << 5,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 24, 31, 38, 45, 52, 0, 2, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 7-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 6) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 16, 23, 30, 37, 44, 51, 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 7-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 8, 15, 22, 29, 36, 43, 50, 57 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 7 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 8> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0xff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 8-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+    };
+    shifts = simd_batch{ 0, 8, 16, 24, 32, 40, 48, 56 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 8-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 8, 16, 24, 32, 40, 48, 56 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 8-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 0, 8, 16, 24, 32, 40, 48, 56 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 8-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 0, 8, 16, 24, 32, 40, 48, 56 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 8-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 0, 8, 16, 24, 32, 40, 48, 56 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 8-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 0, 8, 16, 24, 32, 40, 48, 56 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 8-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 0, 8, 16, 24, 32, 40, 48, 56 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 8-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 0, 8, 16, 24, 32, 40, 48, 56 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 8 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 9> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1ff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 9-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 1) << 1,
+    };
+    shifts = simd_batch{ 0, 9, 18, 27, 36, 45, 54, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 9-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 2) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 8, 17, 26, 35, 44, 53, 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 9-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 3) << 3,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 16, 25, 34, 43, 52, 0, 6, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 9-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 4) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 24, 33, 42, 51, 0, 5, 14, 23 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 9-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 5) << 5,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 32, 41, 50, 0, 4, 13, 22, 31 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 9-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 6) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 40, 49, 0, 3, 12, 21, 30, 39 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 9-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 7) << 7,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 48, 0, 2, 11, 20, 29, 38, 47 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 9-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 8) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 0, 1, 10, 19, 28, 37, 46, 55 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 9 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 10> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3ff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 10-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 1) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 10, 20, 30, 40, 50, 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 10-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 2) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 16, 26, 36, 46, 0, 2, 12, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 10-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 3) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 32, 42, 52, 0, 8, 18, 28, 38 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 10-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 4) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 48, 0, 4, 14, 24, 34, 44, 54 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 10-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 6) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 0, 10, 20, 30, 40, 50, 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 10-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 7) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 16, 26, 36, 46, 0, 2, 12, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 10-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 8) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 32, 42, 52, 0, 8, 18, 28, 38 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 10-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 9) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 48, 0, 4, 14, 24, 34, 44, 54 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 10 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 11> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7ff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 11-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 1) << 9,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 11, 22, 33, 44, 0, 2, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 11-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 2) << 7,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 24, 35, 46, 0, 4, 15, 26, 37 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 11-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 3) << 5,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 4) << 3,
+    };
+    shifts = simd_batch{ 48, 0, 6, 17, 28, 39, 50, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 11-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 5) << 1,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 8, 19, 30, 41, 52, 0, 10, 21 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 11-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 6) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 32, 43, 0, 1, 12, 23, 34, 45 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 11-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 7) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 8) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 0, 3, 14, 25, 36, 47, 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 11-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 9) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 16, 27, 38, 49, 0, 7, 18, 29 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 11-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 10) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 40, 51, 0, 9, 20, 31, 42, 53 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 11 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 12> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0xfff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 12-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 1) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 12, 24, 36, 48, 0, 8, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 12-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 2) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 32, 44, 0, 4, 16, 28, 40, 52 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 12-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 4) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 0, 12, 24, 36, 48, 0, 8, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 12-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 5) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 32, 44, 0, 4, 16, 28, 40, 52 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 12-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 7) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 0, 12, 24, 36, 48, 0, 8, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 12-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 8) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 32, 44, 0, 4, 16, 28, 40, 52 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 12-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 10) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 0, 12, 24, 36, 48, 0, 8, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 12-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 11) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 32, 44, 0, 4, 16, 28, 40, 52 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 12 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 13> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1fff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 13-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 1) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 13, 26, 39, 0, 1, 14, 27 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 13-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 2) << 11,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 3) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 40, 0, 2, 15, 28, 41, 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 13-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 4) << 9,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 16, 29, 42, 0, 4, 17, 30, 43 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 13-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 5) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 6) << 7,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 0, 5, 18, 31, 44, 0, 6, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 13-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 7) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 8) << 5,
+    };
+    shifts = simd_batch{ 32, 45, 0, 7, 20, 33, 46, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 13-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 9) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 8, 21, 34, 47, 0, 9, 22, 35 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 13-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 10) << 3,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 11) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 48, 0, 10, 23, 36, 49, 0, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 13-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 12) << 1,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 24, 37, 50, 0, 12, 25, 38, 51 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 13 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 14> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3fff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 14-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 1) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 14, 28, 42, 0, 6, 20, 34 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 14-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 2) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 3) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 48, 0, 12, 26, 40, 0, 4, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 14-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 4) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 5) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 32, 46, 0, 10, 24, 38, 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 14-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 6) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 16, 30, 44, 0, 8, 22, 36, 50 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 14-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 8) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 0, 14, 28, 42, 0, 6, 20, 34 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 14-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 9) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 10) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 48, 0, 12, 26, 40, 0, 4, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 14-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 11) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 12) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 32, 46, 0, 10, 24, 38, 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 14-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 13) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 16, 30, 44, 0, 8, 22, 36, 50 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 14 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 15> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7fff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 15-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 1) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 15, 30, 45, 0, 11, 26, 41 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 15-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 2) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 3) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 0, 7, 22, 37, 0, 3, 18, 33 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 15-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 4) << 1,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 5) << 5,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 48, 0, 14, 29, 44, 0, 10, 25 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 15-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 6) << 9,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 7) << 13,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 40, 0, 6, 21, 36, 0, 2, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 15-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 8) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 9) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 32, 47, 0, 13, 28, 43, 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 15-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 10) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 11) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 24, 39, 0, 5, 20, 35, 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 15-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 12) << 3,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 13) << 7,
+    };
+    shifts = simd_batch{ 16, 31, 46, 0, 12, 27, 42, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 15-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 14) << 11,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 8, 23, 38, 0, 4, 19, 34, 49 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 15 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 16> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0xffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 16-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+    };
+    shifts = simd_batch{ 0, 16, 32, 48, 0, 16, 32, 48 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 16-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 0, 16, 32, 48, 0, 16, 32, 48 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 16-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 0, 16, 32, 48, 0, 16, 32, 48 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 16-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 0, 16, 32, 48, 0, 16, 32, 48 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 16-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 0, 16, 32, 48, 0, 16, 32, 48 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 16-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 0, 16, 32, 48, 0, 16, 32, 48 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 16-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 0, 16, 32, 48, 0, 16, 32, 48 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 16-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+    };
+    shifts = simd_batch{ 0, 16, 32, 48, 0, 16, 32, 48 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 16 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 17> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1ffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 17-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 1) << 13,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 2) << 9,
+    };
+    shifts = simd_batch{ 0, 17, 34, 0, 4, 21, 38, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 17-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 3) << 5,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 4) << 1,
+    };
+    shifts = simd_batch{ 8, 25, 42, 0, 12, 29, 46, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 17-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 5) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 6) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 16, 33, 0, 3, 20, 37, 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 17-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 7) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 8) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 24, 41, 0, 11, 28, 45, 0, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 17-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 9) << 15,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 10) << 11,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 32, 0, 2, 19, 36, 0, 6, 23 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 17-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 11) << 7,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 12) << 3,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 40, 0, 10, 27, 44, 0, 14, 31 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 17-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 13) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 14) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 0, 1, 18, 35, 0, 5, 22, 39 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 17-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 15) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 16) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+    };
+    shifts = simd_batch{ 0, 9, 26, 43, 0, 13, 30, 47 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 17 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 18> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3ffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 18-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 1) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 2) << 2,
+    };
+    shifts = simd_batch{ 0, 18, 36, 0, 8, 26, 44, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 18-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 3) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 4) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 16, 34, 0, 6, 24, 42, 0, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 18-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 5) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 6) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 32, 0, 4, 22, 40, 0, 12, 30 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 18-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 7) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 8) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 0, 2, 20, 38, 0, 10, 28, 46 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 18-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 10) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 11) << 2,
+    };
+    shifts = simd_batch{ 0, 18, 36, 0, 8, 26, 44, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 18-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 12) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 13) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 16, 34, 0, 6, 24, 42, 0, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 18-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 14) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 15) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+    };
+    shifts = simd_batch{ 32, 0, 4, 22, 40, 0, 12, 30 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 18-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 16) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 17) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+    };
+    shifts = simd_batch{ 0, 2, 20, 38, 0, 10, 28, 46 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 18 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 19> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7ffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 19-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 1) << 7,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 2) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 0, 19, 38, 0, 12, 31, 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 19-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 3) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 4) << 9,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 24, 43, 0, 17, 36, 0, 10, 29 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 19-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 5) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 6) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 7) << 11,
+    };
+    shifts = simd_batch{ 0, 3, 22, 41, 0, 15, 34, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 19-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 8) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 9) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 8, 27, 0, 1, 20, 39, 0, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 19-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 10) << 13,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 11) << 1,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 32, 0, 6, 25, 44, 0, 18, 37 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 19-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 12) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 13) << 15,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 14) << 3,
+    };
+    shifts = simd_batch{ 0, 11, 30, 0, 4, 23, 42, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 19-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 15) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 16) << 17,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+    };
+    shifts = simd_batch{ 16, 35, 0, 9, 28, 0, 2, 21 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 19-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 17) << 5,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 18) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+    };
+    shifts = simd_batch{ 40, 0, 14, 33, 0, 7, 26, 45 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 19 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 20> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0xfffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 20-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 1) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 2) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 0, 20, 40, 0, 16, 36, 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 20-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 3) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 4) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 32, 0, 8, 28, 0, 4, 24, 44 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 20-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 6) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 7) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 0, 20, 40, 0, 16, 36, 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 20-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 8) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 9) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 32, 0, 8, 28, 0, 4, 24, 44 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 20-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 11) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 12) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 0, 20, 40, 0, 16, 36, 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 20-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 13) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 14) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 32, 0, 8, 28, 0, 4, 24, 44 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 20-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 16) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 17) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+    };
+    shifts = simd_batch{ 0, 20, 40, 0, 16, 36, 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 20-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 18) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 19) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+    };
+    shifts = simd_batch{ 32, 0, 8, 28, 0, 4, 24, 44 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 20 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 21> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1fffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 21-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 1) << 1,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 2) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 0, 21, 42, 0, 20, 41, 0, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 21-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 3) << 3,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 4) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 5) << 5,
+    };
+    shifts = simd_batch{ 40, 0, 18, 39, 0, 17, 38, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 21-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 6) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 7) << 7,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 16, 37, 0, 15, 36, 0, 14, 35 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 21-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 8) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 9) << 9,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 10) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 0, 13, 34, 0, 12, 33, 0, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 21-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 11) << 11,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 12) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 13) << 13,
+    };
+    shifts = simd_batch{ 32, 0, 10, 31, 0, 9, 30, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 21-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 14) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 15) << 15,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+    };
+    shifts = simd_batch{ 8, 29, 0, 7, 28, 0, 6, 27 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 21-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 16) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 17) << 17,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 18) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+    };
+    shifts = simd_batch{ 0, 5, 26, 0, 4, 25, 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 21-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 19) << 19,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 20) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+    };
+    shifts = simd_batch{ 24, 0, 2, 23, 0, 1, 22, 43 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 21 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 22> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3fffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 22-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 1) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 2) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 0, 22, 0, 2, 24, 0, 4, 26 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 22-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 3) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 4) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 5) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 0, 6, 28, 0, 8, 30, 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 22-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 6) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 7) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 8) << 6,
+    };
+    shifts = simd_batch{ 32, 0, 12, 34, 0, 14, 36, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 22-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 9) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 10) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 16, 38, 0, 18, 40, 0, 20, 42 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 22-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 12) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 13) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 0, 22, 0, 2, 24, 0, 4, 26 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 22-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 14) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 15) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 16) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+    };
+    shifts = simd_batch{ 0, 6, 28, 0, 8, 30, 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 22-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 17) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 18) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 19) << 6,
+    };
+    shifts = simd_batch{ 32, 0, 12, 34, 0, 14, 36, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 22-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 20) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 21) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+    };
+    shifts = simd_batch{ 16, 38, 0, 18, 40, 0, 20, 42 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 22 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 23> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7fffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 23-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 1) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 2) << 13,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 0, 23, 0, 5, 28, 0, 10, 33 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 23-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 3) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 4) << 3,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 5) << 21,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 0, 15, 38, 0, 20, 0, 2, 25 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 23-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 6) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 7) << 11,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 8) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 0, 7, 30, 0, 12, 35, 0, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 23-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 9) << 1,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 10) << 19,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 11) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 40, 0, 22, 0, 4, 27, 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 23-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 12) << 9,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 13) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 14) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 32, 0, 14, 37, 0, 19, 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 23-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 15) << 17,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 16) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 17) << 7,
+    };
+    shifts = simd_batch{ 24, 0, 6, 29, 0, 11, 34, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 23-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 18) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 19) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 20) << 15,
+    };
+    shifts = simd_batch{ 16, 39, 0, 21, 0, 3, 26, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 23-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 21) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 22) << 5,
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+    };
+    shifts = simd_batch{ 8, 31, 0, 13, 36, 0, 18, 41 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 23 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 24> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0xffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 24-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 1) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 2) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+    };
+    shifts = simd_batch{ 0, 24, 0, 8, 32, 0, 16, 40 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 24-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 4) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 5) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 0, 24, 0, 8, 32, 0, 16, 40 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 24-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 7) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 8) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 0, 24, 0, 8, 32, 0, 16, 40 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 24-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 10) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 11) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 0, 24, 0, 8, 32, 0, 16, 40 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 24-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 13) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 14) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 0, 24, 0, 8, 32, 0, 16, 40 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 24-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 16) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 17) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+    };
+    shifts = simd_batch{ 0, 24, 0, 8, 32, 0, 16, 40 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 24-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 19) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 20) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+    };
+    shifts = simd_batch{ 0, 24, 0, 8, 32, 0, 16, 40 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 24-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 22) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 23) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+    };
+    shifts = simd_batch{ 0, 24, 0, 8, 32, 0, 16, 40 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 24 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 25> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1ffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 25-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 1) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 2) << 3,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 3) << 17,
+    };
+    shifts = simd_batch{ 0, 25, 0, 11, 36, 0, 22, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 25-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 4) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 5) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 6) << 9,
+    };
+    shifts = simd_batch{ 8, 33, 0, 19, 0, 5, 30, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 25-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 7) << 23,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 8) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 9) << 1,
+    };
+    shifts = simd_batch{ 16, 0, 2, 27, 0, 13, 38, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 25-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 10) << 15,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 11) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 12) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 24, 0, 10, 35, 0, 21, 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 25-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 13) << 7,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 14) << 21,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 15) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+    };
+    shifts = simd_batch{ 32, 0, 18, 0, 4, 29, 0, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 25-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 16) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 17) << 13,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 18) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+    };
+    shifts = simd_batch{ 0, 1, 26, 0, 12, 37, 0, 23 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 25-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 19) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 20) << 5,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 21) << 19,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+    };
+    shifts = simd_batch{ 0, 9, 34, 0, 20, 0, 6, 31 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 25-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 22) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 23) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 24) << 11,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+    };
+    shifts = simd_batch{ 0, 17, 0, 3, 28, 0, 14, 39 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 25 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 26> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3ffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 26-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 1) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 2) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 3) << 10,
+    };
+    shifts = simd_batch{ 0, 26, 0, 14, 0, 2, 28, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 26-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 4) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 5) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 6) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 16, 0, 4, 30, 0, 18, 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 26-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 7) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 8) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 9) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 32, 0, 20, 0, 8, 34, 0, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 26-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 10) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 11) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 12) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 0, 10, 36, 0, 24, 0, 12, 38 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 26-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 14) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 15) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 16) << 10,
+    };
+    shifts = simd_batch{ 0, 26, 0, 14, 0, 2, 28, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 26-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 17) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 18) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 19) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+    };
+    shifts = simd_batch{ 16, 0, 4, 30, 0, 18, 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 26-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 20) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 21) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 22) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+    };
+    shifts = simd_batch{ 32, 0, 20, 0, 8, 34, 0, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 26-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 23) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 24) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 25) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+    };
+    shifts = simd_batch{ 0, 10, 36, 0, 24, 0, 12, 38 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 26 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 27> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7ffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 27-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 1) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 2) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 3) << 3,
+    };
+    shifts = simd_batch{ 0, 27, 0, 17, 0, 7, 34, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 27-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 4) << 13,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 5) << 23,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 6) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 24, 0, 14, 0, 4, 31, 0, 21 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 27-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 7) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 8) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 9) << 9,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 10) << 19,
+    };
+    shifts = simd_batch{ 0, 11, 0, 1, 28, 0, 18, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 27-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 11) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 12) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 13) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 8, 35, 0, 25, 0, 15, 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 27-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 14) << 5,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 15) << 15,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 16) << 25,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+    };
+    shifts = simd_batch{ 32, 0, 22, 0, 12, 0, 2, 29 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 27-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 17) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 18) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 19) << 1,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 20) << 11,
+    };
+    shifts = simd_batch{ 0, 19, 0, 9, 36, 0, 26, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 27-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 21) << 21,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 22) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 23) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+    };
+    shifts = simd_batch{ 16, 0, 6, 33, 0, 23, 0, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 27-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 24) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 25) << 7,
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 26) << 17,
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+    };
+    shifts = simd_batch{ 0, 3, 30, 0, 20, 0, 10, 37 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 27 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 28> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0xfffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 28-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 1) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 2) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 3) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 0, 28, 0, 20, 0, 12, 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 28-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 4) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 5) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 6) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 32, 0, 24, 0, 16, 0, 8, 36 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 28-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 8) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 9) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 10) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 0, 28, 0, 20, 0, 12, 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 28-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 11) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 12) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 13) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 32, 0, 24, 0, 16, 0, 8, 36 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 28-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 15) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 16) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 17) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+    };
+    shifts = simd_batch{ 0, 28, 0, 20, 0, 12, 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 28-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 18) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 19) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 20) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+    };
+    shifts = simd_batch{ 32, 0, 24, 0, 16, 0, 8, 36 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 28-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 22) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 23) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 24) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+    };
+    shifts = simd_batch{ 0, 28, 0, 20, 0, 12, 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 28-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 25) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 26) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 27) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+    };
+    shifts = simd_batch{ 32, 0, 24, 0, 16, 0, 8, 36 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 28 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 29> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1fffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 29-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 1) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 2) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 3) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 0, 29, 0, 23, 0, 17, 0, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 29-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 4) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 5) << 1,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 6) << 7,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 7) << 13,
+    };
+    shifts = simd_batch{ 0, 5, 34, 0, 28, 0, 22, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 29-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 8) << 19,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 9) << 25,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 10) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 16, 0, 10, 0, 4, 33, 0, 27 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 29-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 11) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 12) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 13) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 14) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 0, 21, 0, 15, 0, 9, 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 29-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 15) << 3,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 16) << 9,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 17) << 15,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 18) << 21,
+    };
+    shifts = simd_batch{ 32, 0, 26, 0, 20, 0, 14, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 29-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 19) << 27,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 20) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 21) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+    };
+    shifts = simd_batch{ 8, 0, 2, 31, 0, 25, 0, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 29-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 22) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 23) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 24) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 25) << 5,
+    };
+    shifts = simd_batch{ 0, 13, 0, 7, 0, 1, 30, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 29-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 26) << 11,
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 27) << 17,
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 28) << 23,
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+    };
+    shifts = simd_batch{ 24, 0, 18, 0, 12, 0, 6, 35 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 29 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 30> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3fffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 30-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 1) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 2) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 3) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 0, 30, 0, 26, 0, 22, 0, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 30-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 4) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 5) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 6) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 7) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 0, 14, 0, 10, 0, 6, 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 30-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 8) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 9) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 10) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 11) << 14,
+    };
+    shifts = simd_batch{ 32, 0, 28, 0, 24, 0, 20, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 30-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 12) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 13) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 14) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 16, 0, 12, 0, 8, 0, 4, 34 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 30-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 16) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 17) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 18) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+    };
+    shifts = simd_batch{ 0, 30, 0, 26, 0, 22, 0, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 30-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 19) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 20) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 21) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 22) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+    };
+    shifts = simd_batch{ 0, 14, 0, 10, 0, 6, 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 30-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 23) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 24) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 25) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 26) << 14,
+    };
+    shifts = simd_batch{ 32, 0, 28, 0, 24, 0, 20, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 30-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 27) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 28) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 29) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 29),
+      SafeLoadAs<uint64_t>(in + 8 * 29),
+    };
+    shifts = simd_batch{ 16, 0, 12, 0, 8, 0, 4, 34 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 30 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 31> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7fffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 31-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 1) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 2) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 3) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 0, 31, 0, 29, 0, 27, 0, 25 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 31-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 4) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 5) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 6) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 7) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 0, 23, 0, 21, 0, 19, 0, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 31-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 8) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 9) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 10) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 11) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 0, 15, 0, 13, 0, 11, 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 31-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 12) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 13) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 14) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 15) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+    };
+    shifts = simd_batch{ 0, 7, 0, 5, 0, 3, 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 31-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 16) << 1,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 17) << 3,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 18) << 5,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 19) << 7,
+    };
+    shifts = simd_batch{ 32, 0, 30, 0, 28, 0, 26, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 31-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 20) << 9,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 21) << 11,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 22) << 13,
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 23) << 15,
+    };
+    shifts = simd_batch{ 24, 0, 22, 0, 20, 0, 18, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 31-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 24) << 17,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 25) << 19,
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 26) << 21,
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 27) << 23,
+    };
+    shifts = simd_batch{ 16, 0, 14, 0, 12, 0, 10, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 31-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 28) << 25,
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 29) << 27,
+      SafeLoadAs<uint64_t>(in + 8 * 29),
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 30) << 29,
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+    };
+    shifts = simd_batch{ 8, 0, 6, 0, 4, 0, 2, 33 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 31 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 32> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0xffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 32-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+    };
+    shifts = simd_batch{ 0, 32, 0, 32, 0, 32, 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 32-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+    };
+    shifts = simd_batch{ 0, 32, 0, 32, 0, 32, 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 32-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 0, 32, 0, 32, 0, 32, 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 32-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+    };
+    shifts = simd_batch{ 0, 32, 0, 32, 0, 32, 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 32-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+    };
+    shifts = simd_batch{ 0, 32, 0, 32, 0, 32, 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 32-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+    };
+    shifts = simd_batch{ 0, 32, 0, 32, 0, 32, 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 32-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+    };
+    shifts = simd_batch{ 0, 32, 0, 32, 0, 32, 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 32-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 29),
+      SafeLoadAs<uint64_t>(in + 8 * 29),
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+      SafeLoadAs<uint64_t>(in + 8 * 31),
+      SafeLoadAs<uint64_t>(in + 8 * 31),
+    };
+    shifts = simd_batch{ 0, 32, 0, 32, 0, 32, 0, 32 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 32 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 33> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1ffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 33-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 1) << 31,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 2) << 29,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 3) << 27,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 4) << 25,
+    };
+    shifts = simd_batch{ 0, 0, 2, 0, 4, 0, 6, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 33-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 5) << 23,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 6) << 21,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 7) << 19,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 8) << 17,
+    };
+    shifts = simd_batch{ 8, 0, 10, 0, 12, 0, 14, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 33-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 9) << 15,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 10) << 13,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 11) << 11,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 12) << 9,
+    };
+    shifts = simd_batch{ 16, 0, 18, 0, 20, 0, 22, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 33-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 13) << 7,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 14) << 5,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 15) << 3,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 16) << 1,
+    };
+    shifts = simd_batch{ 24, 0, 26, 0, 28, 0, 30, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 33-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 17) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 18) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 19) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 20) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+    };
+    shifts = simd_batch{ 0, 1, 0, 3, 0, 5, 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 33-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 21) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 22) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 23) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 24) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+    };
+    shifts = simd_batch{ 0, 9, 0, 11, 0, 13, 0, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 33-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 25) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 26) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 27) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 28) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+    };
+    shifts = simd_batch{ 0, 17, 0, 19, 0, 21, 0, 23 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 33-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 29) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 29),
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 30) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 31) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 31),
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 32) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+    };
+    shifts = simd_batch{ 0, 25, 0, 27, 0, 29, 0, 31 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 33 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 34> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3ffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 34-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 1) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 2) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 3) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 4) << 18,
+    };
+    shifts = simd_batch{ 0, 0, 4, 0, 8, 0, 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 34-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 5) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 6) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 7) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 8) << 2,
+    };
+    shifts = simd_batch{ 16, 0, 20, 0, 24, 0, 28, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 34-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 9) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 10) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 11) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 12) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 0, 2, 0, 6, 0, 10, 0, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 34-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 13) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 14) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 15) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 16) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+    };
+    shifts = simd_batch{ 0, 18, 0, 22, 0, 26, 0, 30 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 34-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 18) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 19) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 20) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 21) << 18,
+    };
+    shifts = simd_batch{ 0, 0, 4, 0, 8, 0, 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 34-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 22) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 23) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 24) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 25) << 2,
+    };
+    shifts = simd_batch{ 16, 0, 20, 0, 24, 0, 28, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 34-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 26) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 27) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 28) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 29) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 29),
+    };
+    shifts = simd_batch{ 0, 2, 0, 6, 0, 10, 0, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 34-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 30) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 31) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 31),
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 32) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 33) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 33),
+    };
+    shifts = simd_batch{ 0, 18, 0, 22, 0, 26, 0, 30 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 34 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 35> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7ffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 35-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 1) << 29,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 2) << 23,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 3) << 17,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 4) << 11,
+    };
+    shifts = simd_batch{ 0, 0, 6, 0, 12, 0, 18, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 35-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 5) << 5,
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 6) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 7) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 8) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 24, 0, 0, 1, 0, 7, 0, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 35-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 9) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 10) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 11) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 31 | SafeLoadAs<uint64_t>(in + 8 * 12) << 33,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 13) << 27,
+    };
+    shifts = simd_batch{ 0, 19, 0, 25, 0, 0, 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 35-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 14) << 21,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 15) << 15,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 16) << 9,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 17) << 3,
+    };
+    shifts = simd_batch{ 8, 0, 14, 0, 20, 0, 26, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 35-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 18) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 19) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 20) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 21) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+    };
+    shifts = simd_batch{ 0, 3, 0, 9, 0, 15, 0, 21 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 35-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 22) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 23) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 24) << 31,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 25) << 25,
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 26) << 19,
+    };
+    shifts = simd_batch{ 0, 27, 0, 0, 4, 0, 10, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 35-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 27) << 13,
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 28) << 7,
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 29) << 1,
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 30) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+    };
+    shifts = simd_batch{ 16, 0, 22, 0, 28, 0, 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 35-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 31) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 31),
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 32) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 33) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 33),
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 34) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 34),
+    };
+    shifts = simd_batch{ 0, 11, 0, 17, 0, 23, 0, 29 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 35 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 36> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0xfffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 36-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 1) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 2) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 3) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 4) << 4,
+    };
+    shifts = simd_batch{ 0, 0, 8, 0, 16, 0, 24, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 36-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 5) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 6) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 7) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 8) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+    };
+    shifts = simd_batch{ 0, 4, 0, 12, 0, 20, 0, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 36-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 10) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 11) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 12) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 13) << 4,
+    };
+    shifts = simd_batch{ 0, 0, 8, 0, 16, 0, 24, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 36-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 14) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 15) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 16) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 17) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+    };
+    shifts = simd_batch{ 0, 4, 0, 12, 0, 20, 0, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 36-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 19) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 20) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 21) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 22) << 4,
+    };
+    shifts = simd_batch{ 0, 0, 8, 0, 16, 0, 24, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 36-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 23) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 24) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 25) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 26) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+    };
+    shifts = simd_batch{ 0, 4, 0, 12, 0, 20, 0, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 36-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 28) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 29) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 29),
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 30) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 31) << 4,
+    };
+    shifts = simd_batch{ 0, 0, 8, 0, 16, 0, 24, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 36-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 32) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 33) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 33),
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 34) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 34),
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 35) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+    };
+    shifts = simd_batch{ 0, 4, 0, 12, 0, 20, 0, 28 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 36 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 37> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1fffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 37-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 1) << 27,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 2) << 17,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 3) << 7,
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 4) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 0, 0, 10, 0, 20, 0, 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 37-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 5) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 6) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 7) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 8) << 31,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 9) << 21,
+    };
+    shifts = simd_batch{ 0, 13, 0, 23, 0, 0, 6, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 37-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 10) << 11,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 11) << 1,
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 12) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 13) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 16, 0, 26, 0, 0, 9, 0, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 37-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 14) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 29 | SafeLoadAs<uint64_t>(in + 8 * 15) << 35,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 16) << 25,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 17) << 15,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 18) << 5,
+    };
+    shifts = simd_batch{ 0, 0, 2, 0, 12, 0, 22, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 37-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 19) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 20) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 21) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 22) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 23) << 29,
+    };
+    shifts = simd_batch{ 0, 5, 0, 15, 0, 25, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 37-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 24) << 19,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 25) << 9,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 26) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 27) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+    };
+    shifts = simd_batch{ 8, 0, 18, 0, 0, 1, 0, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 37-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 28) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 29) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 31 | SafeLoadAs<uint64_t>(in + 8 * 30) << 33,
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 31) << 23,
+      SafeLoadAs<uint64_t>(in + 8 * 31),
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 32) << 13,
+    };
+    shifts = simd_batch{ 0, 21, 0, 0, 4, 0, 14, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 37-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 33) << 3,
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 34) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 34),
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 35) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 36) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 36),
+    };
+    shifts = simd_batch{ 24, 0, 0, 7, 0, 17, 0, 27 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 37 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 38> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3fffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 38-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 1) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 2) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 3) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 4) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 0, 0, 12, 0, 24, 0, 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 38-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 5) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 6) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 7) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 8) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 9) << 6,
+    };
+    shifts = simd_batch{ 0, 22, 0, 0, 8, 0, 20, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 38-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 10) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 11) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 12) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 13) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 14) << 22,
+    };
+    shifts = simd_batch{ 0, 6, 0, 18, 0, 0, 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 38-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 15) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 16) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 17) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 18) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+    };
+    shifts = simd_batch{ 16, 0, 0, 2, 0, 14, 0, 26 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 38-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 20) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 21) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 22) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 23) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+    };
+    shifts = simd_batch{ 0, 0, 12, 0, 24, 0, 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 38-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 24) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 25) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 26) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 27) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 28) << 6,
+    };
+    shifts = simd_batch{ 0, 22, 0, 0, 8, 0, 20, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 38-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 29) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 29),
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 30) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 31) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 32) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 33) << 22,
+    };
+    shifts = simd_batch{ 0, 6, 0, 18, 0, 0, 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 38-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 33),
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 34) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 35) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 36) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 36),
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 37) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 37),
+    };
+    shifts = simd_batch{ 16, 0, 0, 2, 0, 14, 0, 26 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 38 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 39> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7fffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 39-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 1) << 25,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 2) << 11,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 3) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 4) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 0, 0, 14, 0, 0, 3, 0, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 39-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 5) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 31 | SafeLoadAs<uint64_t>(in + 8 * 6) << 33,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 7) << 19,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 8) << 5,
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 9) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 0, 0, 6, 0, 20, 0, 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 39-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 10) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 11) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 12) << 27,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 13) << 13,
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 14) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 0, 23, 0, 0, 12, 0, 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 39-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 15) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 16) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 29 | SafeLoadAs<uint64_t>(in + 8 * 17) << 35,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 18) << 21,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 19) << 7,
+    };
+    shifts = simd_batch{ 0, 15, 0, 0, 4, 0, 18, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 39-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 20) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 21) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 22) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 23) << 29,
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 24) << 15,
+    };
+    shifts = simd_batch{ 0, 7, 0, 21, 0, 0, 10, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 39-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 25) << 1,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 26) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 27) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 27 | SafeLoadAs<uint64_t>(in + 8 * 28) << 37,
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 29) << 23,
+    };
+    shifts = simd_batch{ 24, 0, 0, 13, 0, 0, 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 39-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 29),
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 30) << 9,
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 31) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 31),
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 32) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 33) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 34) << 31,
+    };
+    shifts = simd_batch{ 16, 0, 0, 5, 0, 19, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 39-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 34),
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 35) << 17,
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 36) << 3,
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 37) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 37),
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 38) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 38),
+    };
+    shifts = simd_batch{ 8, 0, 22, 0, 0, 11, 0, 25 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 39 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 40> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0xffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 40-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 1) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 2) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 3) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 4) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+    };
+    shifts = simd_batch{ 0, 0, 16, 0, 0, 8, 0, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 40-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 6) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 7) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 8) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 9) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+    };
+    shifts = simd_batch{ 0, 0, 16, 0, 0, 8, 0, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 40-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 11) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 12) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 13) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 14) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 0, 0, 16, 0, 0, 8, 0, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 40-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 16) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 17) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 18) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 19) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+    };
+    shifts = simd_batch{ 0, 0, 16, 0, 0, 8, 0, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 40-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 21) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 22) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 23) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 24) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+    };
+    shifts = simd_batch{ 0, 0, 16, 0, 0, 8, 0, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 40-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 26) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 27) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 28) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 29) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 29),
+    };
+    shifts = simd_batch{ 0, 0, 16, 0, 0, 8, 0, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 40-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 31) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 31),
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 32) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 33) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 33),
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 34) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 34),
+    };
+    shifts = simd_batch{ 0, 0, 16, 0, 0, 8, 0, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 40-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 36) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 36),
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 37) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 38) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 38),
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 39) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 39),
+    };
+    shifts = simd_batch{ 0, 0, 16, 0, 0, 8, 0, 24 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 40 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 41> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1ffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 41-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 1) << 23,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 2) << 5,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 3) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 4) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 31 | SafeLoadAs<uint64_t>(in + 8 * 5) << 33,
+    };
+    shifts = simd_batch{ 0, 0, 18, 0, 0, 13, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 41-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 6) << 15,
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 7) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 8) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 9) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 10) << 25,
+    };
+    shifts = simd_batch{ 8, 0, 0, 3, 0, 21, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 41-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 11) << 7,
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 12) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 13) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 29 | SafeLoadAs<uint64_t>(in + 8 * 14) << 35,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 15) << 17,
+    };
+    shifts = simd_batch{ 16, 0, 0, 11, 0, 0, 6, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 41-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 16) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 17) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 18) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 19) << 27,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 20) << 9,
+    };
+    shifts = simd_batch{ 0, 1, 0, 19, 0, 0, 14, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 41-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 21) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 22) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 27 | SafeLoadAs<uint64_t>(in + 8 * 23) << 37,
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 24) << 19,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 25) << 1,
+    };
+    shifts = simd_batch{ 0, 9, 0, 0, 4, 0, 22, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 41-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 26) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 27) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 28) << 29,
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 29) << 11,
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 30) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+    };
+    shifts = simd_batch{ 0, 17, 0, 0, 12, 0, 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 41-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 31) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 25 | SafeLoadAs<uint64_t>(in + 8 * 32) << 39,
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 33) << 21,
+      SafeLoadAs<uint64_t>(in + 8 * 33),
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 34) << 3,
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 35) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+    };
+    shifts = simd_batch{ 0, 0, 2, 0, 20, 0, 0, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 41-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 36) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 37) << 31,
+      SafeLoadAs<uint64_t>(in + 8 * 37),
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 38) << 13,
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 39) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 39),
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 40) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 40),
+    };
+    shifts = simd_batch{ 0, 0, 10, 0, 0, 5, 0, 23 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 41 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 42> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3ffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 42-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 1) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 1),
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 2) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 3) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 4) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 5) << 26,
+    };
+    shifts = simd_batch{ 0, 0, 20, 0, 0, 18, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 42-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 6) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 7) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 8) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 9) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 10) << 10,
+    };
+    shifts = simd_batch{ 16, 0, 0, 14, 0, 0, 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 42-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 11) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 12) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 13) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 14) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 15) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+    };
+    shifts = simd_batch{ 0, 10, 0, 0, 8, 0, 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 42-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 16) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 17) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 18) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 19) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 20) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+    };
+    shifts = simd_batch{ 0, 0, 4, 0, 0, 2, 0, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 42-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 22) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 23) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 24) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 25) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 26) << 26,
+    };
+    shifts = simd_batch{ 0, 0, 20, 0, 0, 18, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 42-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 27) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 28) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 29) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 30) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 31) << 10,
+    };
+    shifts = simd_batch{ 16, 0, 0, 14, 0, 0, 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 42-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 32) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 33) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 34) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 34),
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 35) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 36) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 36),
+    };
+    shifts = simd_batch{ 0, 10, 0, 0, 8, 0, 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 42-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 37) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 38) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 38),
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 39) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 40) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 40),
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 41) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 41),
+    };
+    shifts = simd_batch{ 0, 0, 4, 0, 0, 2, 0, 22 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 42 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 43> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7ffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 43-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 1) << 21,
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 2) << 42,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 3) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 23 | SafeLoadAs<uint64_t>(in + 8 * 4) << 41,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 5) << 19,
+    };
+    shifts = simd_batch{ 0, 0, 0, 1, 0, 0, 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 43-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 6) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 7) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 25 | SafeLoadAs<uint64_t>(in + 8 * 8) << 39,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 9) << 17,
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 10) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 0, 3, 0, 0, 4, 0, 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 43-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 11) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 27 | SafeLoadAs<uint64_t>(in + 8 * 12) << 37,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 13) << 15,
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 14) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 15) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 29 | SafeLoadAs<uint64_t>(in + 8 * 16) << 35,
+    };
+    shifts = simd_batch{ 0, 0, 6, 0, 0, 7, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 43-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 17) << 13,
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 18) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 19) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 31 | SafeLoadAs<uint64_t>(in + 8 * 20) << 33,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 21) << 11,
+    };
+    shifts = simd_batch{ 8, 0, 0, 9, 0, 0, 10, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 43-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 22) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 23) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 24) << 31,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 25) << 9,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 26) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+    };
+    shifts = simd_batch{ 0, 11, 0, 0, 12, 0, 0, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 43-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 27) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 28) << 29,
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 29) << 7,
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 30) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 31) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 32) << 27,
+    };
+    shifts = simd_batch{ 0, 0, 14, 0, 0, 15, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 43-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 33) << 5,
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 34) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 34),
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 35) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 36) << 25,
+      SafeLoadAs<uint64_t>(in + 8 * 36),
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 37) << 3,
+    };
+    shifts = simd_batch{ 16, 0, 0, 17, 0, 0, 18, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 43-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 38) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 38),
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 39) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 40) << 23,
+      SafeLoadAs<uint64_t>(in + 8 * 40),
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 41) << 1,
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 42) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 42),
+    };
+    shifts = simd_batch{ 0, 19, 0, 0, 20, 0, 0, 21 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 43 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 44> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0xfffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 44-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 1) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 2) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 3) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 4) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 5) << 12,
+    };
+    shifts = simd_batch{ 0, 0, 0, 4, 0, 0, 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 44-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 6) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 7) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 8) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 9) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 10) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+    };
+    shifts = simd_batch{ 0, 12, 0, 0, 16, 0, 0, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 44-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 12) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 13) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 14) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 15) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 16) << 12,
+    };
+    shifts = simd_batch{ 0, 0, 0, 4, 0, 0, 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 44-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 17) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 18) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 19) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 20) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 21) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+    };
+    shifts = simd_batch{ 0, 12, 0, 0, 16, 0, 0, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 44-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 23) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 24) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 25) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 26) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 27) << 12,
+    };
+    shifts = simd_batch{ 0, 0, 0, 4, 0, 0, 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 44-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 28) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 29) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 30) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 31) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 32) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+    };
+    shifts = simd_batch{ 0, 12, 0, 0, 16, 0, 0, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 44-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 33),
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 34) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 35) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 36) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 37) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 37),
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 38) << 12,
+    };
+    shifts = simd_batch{ 0, 0, 0, 4, 0, 0, 8, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 44-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 39) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 39),
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 40) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 41) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 41),
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 42) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 43) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 43),
+    };
+    shifts = simd_batch{ 0, 12, 0, 0, 16, 0, 0, 20 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 44 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 45> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1fffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 45-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 1) << 19,
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 2) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 3) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 4) << 31,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 5) << 5,
+    };
+    shifts = simd_batch{ 0, 0, 0, 7, 0, 0, 14, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 45-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 6) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 21 | SafeLoadAs<uint64_t>(in + 8 * 7) << 43,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 8) << 17,
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 9) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 10) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 11) << 29,
+    };
+    shifts = simd_batch{ 0, 0, 2, 0, 0, 9, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 45-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 12) << 3,
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 13) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 23 | SafeLoadAs<uint64_t>(in + 8 * 14) << 41,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 15) << 15,
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 16) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+    };
+    shifts = simd_batch{ 16, 0, 0, 0, 4, 0, 0, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 45-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 17) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 18) << 27,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 19) << 1,
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 20) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 25 | SafeLoadAs<uint64_t>(in + 8 * 21) << 39,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 22) << 13,
+    };
+    shifts = simd_batch{ 0, 0, 18, 0, 0, 0, 6, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 45-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 23) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 24) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 25) << 25,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 26) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 27) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 27 | SafeLoadAs<uint64_t>(in + 8 * 28) << 37,
+    };
+    shifts = simd_batch{ 0, 13, 0, 0, 0, 1, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 45-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 29) << 11,
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 30) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 31) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 32) << 23,
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 33) << 42,
+      SafeLoadAs<uint64_t>(in + 8 * 33),
+    };
+    shifts = simd_batch{ 8, 0, 0, 15, 0, 0, 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 45-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 34) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 29 | SafeLoadAs<uint64_t>(in + 8 * 35) << 35,
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 36) << 9,
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 37) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 37),
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 38) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 39) << 21,
+    };
+    shifts = simd_batch{ 0, 0, 10, 0, 0, 17, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 45-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 40) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 40),
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 41) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 31 | SafeLoadAs<uint64_t>(in + 8 * 42) << 33,
+      SafeLoadAs<uint64_t>(in + 8 * 42),
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 43) << 7,
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 44) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 44),
+    };
+    shifts = simd_batch{ 0, 5, 0, 0, 12, 0, 0, 19 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 45 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 46> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3fffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 46-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 1) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 2) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 3) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 4) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 5) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 0, 0, 0, 10, 0, 0, 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 46-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 6) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 7) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 8) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 9) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 10) << 42,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 11) << 14,
+    };
+    shifts = simd_batch{ 0, 0, 12, 0, 0, 0, 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 46-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 12) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 13) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 14) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 15) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 16) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 17) << 30,
+    };
+    shifts = simd_batch{ 0, 14, 0, 0, 0, 6, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 46-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 18) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 19) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 20) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 21) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 22) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+    };
+    shifts = simd_batch{ 16, 0, 0, 0, 8, 0, 0, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 46-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 24) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 25) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 26) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 27) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 28) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+    };
+    shifts = simd_batch{ 0, 0, 0, 10, 0, 0, 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 46-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 29) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 30) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 31) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 32) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 33) << 42,
+      SafeLoadAs<uint64_t>(in + 8 * 33),
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 34) << 14,
+    };
+    shifts = simd_batch{ 0, 0, 12, 0, 0, 0, 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 46-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 35) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 36) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 37) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 38) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 38),
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 39) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 40) << 30,
+    };
+    shifts = simd_batch{ 0, 14, 0, 0, 0, 6, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 46-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 40),
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 41) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 42) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 43) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 43),
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 44) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 45) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 45),
+    };
+    shifts = simd_batch{ 16, 0, 0, 0, 8, 0, 0, 18 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 46 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 47> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7fffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 47-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 1) << 17,
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 2) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 3) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 4) << 21,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 5) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 0, 0, 0, 13, 0, 0, 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 47-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 6) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 7) << 25,
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 8) << 42,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 9) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 10) << 29,
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 11) << 46,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 0, 0, 0, 5, 0, 0, 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 47-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 12) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 31 | SafeLoadAs<uint64_t>(in + 8 * 13) << 33,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 14) << 3,
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 15) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 27 | SafeLoadAs<uint64_t>(in + 8 * 16) << 37,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 17) << 7,
+    };
+    shifts = simd_batch{ 0, 0, 14, 0, 0, 0, 10, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 47-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 18) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 23 | SafeLoadAs<uint64_t>(in + 8 * 19) << 41,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 20) << 11,
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 21) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 19 | SafeLoadAs<uint64_t>(in + 8 * 22) << 45,
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 23) << 15,
+    };
+    shifts = simd_batch{ 0, 0, 6, 0, 0, 0, 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 47-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 24) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 25) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 26) << 19,
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 27) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 28) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 29) << 23,
+    };
+    shifts = simd_batch{ 0, 15, 0, 0, 0, 11, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 47-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 30) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 31) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 32) << 27,
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 33) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 33),
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 34) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 35) << 31,
+    };
+    shifts = simd_batch{ 0, 7, 0, 0, 0, 3, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 47-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 36) << 1,
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 37) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 29 | SafeLoadAs<uint64_t>(in + 8 * 38) << 35,
+      SafeLoadAs<uint64_t>(in + 8 * 38),
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 39) << 5,
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 40) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 25 | SafeLoadAs<uint64_t>(in + 8 * 41) << 39,
+    };
+    shifts = simd_batch{ 16, 0, 0, 0, 12, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 47-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 41),
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 42) << 9,
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 43) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 21 | SafeLoadAs<uint64_t>(in + 8 * 44) << 43,
+      SafeLoadAs<uint64_t>(in + 8 * 44),
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 45) << 13,
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 46) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 46),
+    };
+    shifts = simd_batch{ 8, 0, 0, 0, 4, 0, 0, 17 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 47 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 48> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0xffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 48-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 1) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 2) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 2),
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 4) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 5) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+    };
+    shifts = simd_batch{ 0, 0, 0, 16, 0, 0, 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 48-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 7) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 8) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 10) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 11) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+    };
+    shifts = simd_batch{ 0, 0, 0, 16, 0, 0, 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 48-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 13) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 14) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 16) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 17) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+    };
+    shifts = simd_batch{ 0, 0, 0, 16, 0, 0, 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 48-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 19) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 20) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 22) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 23) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+    };
+    shifts = simd_batch{ 0, 0, 0, 16, 0, 0, 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 48-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 25) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 26) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 28) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 29) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 29),
+    };
+    shifts = simd_batch{ 0, 0, 0, 16, 0, 0, 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 48-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 31) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 32) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+      SafeLoadAs<uint64_t>(in + 8 * 33),
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 34) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 35) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+    };
+    shifts = simd_batch{ 0, 0, 0, 16, 0, 0, 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 48-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 36),
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 37) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 38) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 38),
+      SafeLoadAs<uint64_t>(in + 8 * 39),
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 40) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 41) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 41),
+    };
+    shifts = simd_batch{ 0, 0, 0, 16, 0, 0, 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 48-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 42),
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 43) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 44) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 44),
+      SafeLoadAs<uint64_t>(in + 8 * 45),
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 46) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 47) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 47),
+    };
+    shifts = simd_batch{ 0, 0, 0, 16, 0, 0, 0, 16 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 48 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 49> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1ffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 49-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 1) << 15,
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 2) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 19 | SafeLoadAs<uint64_t>(in + 8 * 3) << 45,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 4) << 11,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 5) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 23 | SafeLoadAs<uint64_t>(in + 8 * 6) << 41,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 4, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 49-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 7) << 7,
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 8) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 27 | SafeLoadAs<uint64_t>(in + 8 * 9) << 37,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 10) << 3,
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 11) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 31 | SafeLoadAs<uint64_t>(in + 8 * 12) << 33,
+    };
+    shifts = simd_batch{ 8, 0, 0, 0, 12, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 49-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 13) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 14) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 15) << 29,
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 16) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 17) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 18) << 25,
+    };
+    shifts = simd_batch{ 0, 1, 0, 0, 0, 5, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 49-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 19) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 20) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 21) << 21,
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 22) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 22),
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 23) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 24) << 17,
+    };
+    shifts = simd_batch{ 0, 9, 0, 0, 0, 13, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 49-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 25) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 17 | SafeLoadAs<uint64_t>(in + 8 * 26) << 47,
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 27) << 13,
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 28) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 21 | SafeLoadAs<uint64_t>(in + 8 * 29) << 43,
+      SafeLoadAs<uint64_t>(in + 8 * 29),
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 30) << 9,
+    };
+    shifts = simd_batch{ 0, 0, 2, 0, 0, 0, 6, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 49-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 31) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 25 | SafeLoadAs<uint64_t>(in + 8 * 32) << 39,
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 33) << 5,
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 34) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 29 | SafeLoadAs<uint64_t>(in + 8 * 35) << 35,
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 36) << 1,
+    };
+    shifts = simd_batch{ 0, 0, 10, 0, 0, 0, 14, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 49-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 37) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 38) << 31,
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 39) << 46,
+      SafeLoadAs<uint64_t>(in + 8 * 39),
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 40) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 41) << 27,
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 42) << 42,
+      SafeLoadAs<uint64_t>(in + 8 * 42),
+    };
+    shifts = simd_batch{ 0, 0, 0, 3, 0, 0, 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 49-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 43) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 44) << 23,
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 45) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 45),
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 46) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 47) << 19,
+      SafeLoadAs<uint64_t>(in + 8 * 47) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 48) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 48),
+    };
+    shifts = simd_batch{ 0, 0, 0, 11, 0, 0, 0, 15 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 49 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 50> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3ffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 50-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 1) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 2) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 3) << 42,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 4) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 5) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 6) << 34,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 8, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 50-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 7) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 8) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 9) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 10) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 11) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 12) << 18,
+    };
+    shifts = simd_batch{ 0, 2, 0, 0, 0, 10, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 50-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 13) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 14) << 46,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 15) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 16) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 17) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 18) << 2,
+    };
+    shifts = simd_batch{ 0, 0, 4, 0, 0, 0, 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 50-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 19) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 20) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 21) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 22) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 23) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 24) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+    };
+    shifts = simd_batch{ 0, 0, 0, 6, 0, 0, 0, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 50-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 26) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 27) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 28) << 42,
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 29) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 30) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 31) << 34,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 8, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 50-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 32) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 33) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 34) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 35) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 36) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 37) << 18,
+    };
+    shifts = simd_batch{ 0, 2, 0, 0, 0, 10, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 50-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 38) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 39) << 46,
+      SafeLoadAs<uint64_t>(in + 8 * 39),
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 40) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 41) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 42) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 42),
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 43) << 2,
+    };
+    shifts = simd_batch{ 0, 0, 4, 0, 0, 0, 12, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 50-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 44) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 45) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 46) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 46),
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 47) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 47) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 48) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 48) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 49) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 49),
+    };
+    shifts = simd_batch{ 0, 0, 0, 6, 0, 0, 0, 14 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 50 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 51> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7ffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 51-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 1) << 13,
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 2) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 25 | SafeLoadAs<uint64_t>(in + 8 * 3) << 39,
+      SafeLoadAs<uint64_t>(in + 8 * 3),
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 4) << 1,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 5) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 6) << 27,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 12, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 51-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 7) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 8) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 9) << 15,
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 10) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 23 | SafeLoadAs<uint64_t>(in + 8 * 11) << 41,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 12) << 3,
+    };
+    shifts = simd_batch{ 0, 11, 0, 0, 0, 0, 10, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 51-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 13) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 14) << 29,
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 15) << 42,
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 16) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 17) << 17,
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 18) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 21 | SafeLoadAs<uint64_t>(in + 8 * 19) << 43,
+    };
+    shifts = simd_batch{ 0, 0, 0, 9, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 51-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 20) << 5,
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 21) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 22) << 31,
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 23) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 24) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 25) << 19,
+    };
+    shifts = simd_batch{ 8, 0, 0, 0, 0, 7, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 51-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 26) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 19 | SafeLoadAs<uint64_t>(in + 8 * 27) << 45,
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 28) << 7,
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 29) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 31 | SafeLoadAs<uint64_t>(in + 8 * 30) << 33,
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 31) << 46,
+      SafeLoadAs<uint64_t>(in + 8 * 31),
+    };
+    shifts = simd_batch{ 0, 0, 6, 0, 0, 0, 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 51-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 32) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 33) << 21,
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 34) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 17 | SafeLoadAs<uint64_t>(in + 8 * 35) << 47,
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 36) << 9,
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 37) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 29 | SafeLoadAs<uint64_t>(in + 8 * 38) << 35,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 4, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 51-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 39) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 39),
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 40) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 41) << 23,
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 42) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 15 | SafeLoadAs<uint64_t>(in + 8 * 43) << 49,
+      SafeLoadAs<uint64_t>(in + 8 * 43),
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 44) << 11,
+    };
+    shifts = simd_batch{ 0, 3, 0, 0, 0, 0, 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 51-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 45) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 27 | SafeLoadAs<uint64_t>(in + 8 * 46) << 37,
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 14 | SafeLoadAs<uint64_t>(in + 8 * 47) << 50,
+      SafeLoadAs<uint64_t>(in + 8 * 47),
+      SafeLoadAs<uint64_t>(in + 8 * 47) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 48) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 48) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 49) << 25,
+      SafeLoadAs<uint64_t>(in + 8 * 49) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 50) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 50),
+    };
+    shifts = simd_batch{ 0, 0, 0, 1, 0, 0, 0, 13 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 51 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 52> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0xfffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 52-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 1) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 2) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 3) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 4) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 5) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 6) << 20,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 4, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 52-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 7) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 8) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 9) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 10) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 11) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 12) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+    };
+    shifts = simd_batch{ 0, 0, 8, 0, 0, 0, 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 52-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 14) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 15) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 16) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 17) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 17),
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 18) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 19) << 20,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 4, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 52-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 20) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 21) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 22) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 23) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 24) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 25) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 25),
+    };
+    shifts = simd_batch{ 0, 0, 8, 0, 0, 0, 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 52-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 27) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 28) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 29) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 30) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 31) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 32) << 20,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 4, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 52-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 33) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 34) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 34),
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 35) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 36) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 37) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 38) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 38),
+    };
+    shifts = simd_batch{ 0, 0, 8, 0, 0, 0, 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 52-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 39),
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 40) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 41) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 42) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 43) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 43),
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 44) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 45) << 20,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 4, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 52-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 46) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 47) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 47),
+      SafeLoadAs<uint64_t>(in + 8 * 47) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 48) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 48) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 49) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 49) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 50) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 50) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 51) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 51),
+    };
+    shifts = simd_batch{ 0, 0, 8, 0, 0, 0, 0, 12 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 52 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 53> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1fffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 53-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 1) << 11,
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 2) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 31 | SafeLoadAs<uint64_t>(in + 8 * 3) << 33,
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 4) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 4),
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 5) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 6) << 13,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 9, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 53-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 7) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 29 | SafeLoadAs<uint64_t>(in + 8 * 8) << 35,
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 9) << 46,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 10) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 11) << 15,
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 12) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 27 | SafeLoadAs<uint64_t>(in + 8 * 13) << 37,
+    };
+    shifts = simd_batch{ 0, 0, 0, 7, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 53-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 14) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 15) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 16) << 17,
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 17) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 25 | SafeLoadAs<uint64_t>(in + 8 * 18) << 39,
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 14 | SafeLoadAs<uint64_t>(in + 8 * 19) << 50,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+    };
+    shifts = simd_batch{ 0, 5, 0, 0, 0, 0, 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 53-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 20) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 21) << 19,
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 22) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 23 | SafeLoadAs<uint64_t>(in + 8 * 23) << 41,
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 24) << 52,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 25) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 26) << 21,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 1, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 53-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 27) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 21 | SafeLoadAs<uint64_t>(in + 8 * 28) << 43,
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 29) << 1,
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 30) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 31) << 23,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 32) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 19 | SafeLoadAs<uint64_t>(in + 8 * 33) << 45,
+    };
+    shifts = simd_batch{ 0, 0, 10, 0, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 53-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 33),
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 34) << 3,
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 35) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 36) << 25,
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 37) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 17 | SafeLoadAs<uint64_t>(in + 8 * 38) << 47,
+      SafeLoadAs<uint64_t>(in + 8 * 38),
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 39) << 5,
+    };
+    shifts = simd_batch{ 8, 0, 0, 0, 0, 0, 6, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 53-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 40) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 41) << 27,
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 42) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 15 | SafeLoadAs<uint64_t>(in + 8 * 43) << 49,
+      SafeLoadAs<uint64_t>(in + 8 * 43),
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 44) << 7,
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 45) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 46) << 29,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 4, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 53-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 47) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 47) >> 13 | SafeLoadAs<uint64_t>(in + 8 * 48) << 51,
+      SafeLoadAs<uint64_t>(in + 8 * 48),
+      SafeLoadAs<uint64_t>(in + 8 * 48) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 49) << 9,
+      SafeLoadAs<uint64_t>(in + 8 * 49) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 50) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 50) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 51) << 31,
+      SafeLoadAs<uint64_t>(in + 8 * 51) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 52) << 42,
+      SafeLoadAs<uint64_t>(in + 8 * 52),
+    };
+    shifts = simd_batch{ 0, 0, 2, 0, 0, 0, 0, 11 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 53 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 54> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3fffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 54-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 1) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 2) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 3) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 4) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 14 | SafeLoadAs<uint64_t>(in + 8 * 5) << 50,
+      SafeLoadAs<uint64_t>(in + 8 * 5),
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 6) << 6,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 54-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 7) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 8) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 9) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 10) << 46,
+      SafeLoadAs<uint64_t>(in + 8 * 10),
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 11) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 12) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 13) << 22,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 8, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 54-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 14) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 15) << 42,
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 16) << 52,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 17) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 18) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 19) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 20) << 38,
+    };
+    shifts = simd_batch{ 0, 0, 0, 2, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 54-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 21) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 22) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 23) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 24) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 25) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 26) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 26),
+    };
+    shifts = simd_batch{ 0, 6, 0, 0, 0, 0, 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 54-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 28) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 29) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 30) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 31) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 14 | SafeLoadAs<uint64_t>(in + 8 * 32) << 50,
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 33) << 6,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 4, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 54-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 34) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 35) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 36) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 37) << 46,
+      SafeLoadAs<uint64_t>(in + 8 * 37),
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 38) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 39) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 40) << 22,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 8, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 54-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 41) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 42) << 42,
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 43) << 52,
+      SafeLoadAs<uint64_t>(in + 8 * 43),
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 44) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 45) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 46) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 47) << 38,
+    };
+    shifts = simd_batch{ 0, 0, 0, 2, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 54-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 47) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 48) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 48),
+      SafeLoadAs<uint64_t>(in + 8 * 48) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 49) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 49) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 50) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 50) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 51) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 51) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 52) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 52) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 53) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 53),
+    };
+    shifts = simd_batch{ 0, 6, 0, 0, 0, 0, 0, 10 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 54 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 55> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7fffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 55-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 1) << 9,
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 2) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 3) << 27,
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 4) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 19 | SafeLoadAs<uint64_t>(in + 8 * 5) << 45,
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 10 | SafeLoadAs<uint64_t>(in + 8 * 6) << 54,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 55-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 7) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 8) << 17,
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 9) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 29 | SafeLoadAs<uint64_t>(in + 8 * 10) << 35,
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 11) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 11 | SafeLoadAs<uint64_t>(in + 8 * 12) << 53,
+      SafeLoadAs<uint64_t>(in + 8 * 12),
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 13) << 7,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 55-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 14) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 15) << 25,
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 16) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 21 | SafeLoadAs<uint64_t>(in + 8 * 17) << 43,
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 18) << 52,
+      SafeLoadAs<uint64_t>(in + 8 * 18),
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 19) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 20) << 15,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 3, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 55-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 21) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 31 | SafeLoadAs<uint64_t>(in + 8 * 22) << 33,
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 23) << 42,
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 13 | SafeLoadAs<uint64_t>(in + 8 * 24) << 51,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 25) << 5,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 26) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 27) << 23,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 4, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 55-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 28) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 23 | SafeLoadAs<uint64_t>(in + 8 * 29) << 41,
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 14 | SafeLoadAs<uint64_t>(in + 8 * 30) << 50,
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 31) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 32) << 13,
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 33) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 34) << 31,
+    };
+    shifts = simd_batch{ 0, 0, 0, 5, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 55-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 35) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 15 | SafeLoadAs<uint64_t>(in + 8 * 36) << 49,
+      SafeLoadAs<uint64_t>(in + 8 * 36),
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 37) << 3,
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 38) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 39) << 21,
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 40) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 25 | SafeLoadAs<uint64_t>(in + 8 * 41) << 39,
+    };
+    shifts = simd_batch{ 0, 0, 6, 0, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 55-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 42) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 42),
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 43) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 44) << 11,
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 45) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 46) << 29,
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 47) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 47) >> 17 | SafeLoadAs<uint64_t>(in + 8 * 48) << 47,
+    };
+    shifts = simd_batch{ 0, 7, 0, 0, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 55-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 48),
+      SafeLoadAs<uint64_t>(in + 8 * 48) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 49) << 1,
+      SafeLoadAs<uint64_t>(in + 8 * 49) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 50) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 50) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 51) << 19,
+      SafeLoadAs<uint64_t>(in + 8 * 51) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 52) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 52) >> 27 | SafeLoadAs<uint64_t>(in + 8 * 53) << 37,
+      SafeLoadAs<uint64_t>(in + 8 * 53) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 54) << 46,
+      SafeLoadAs<uint64_t>(in + 8 * 54),
+    };
+    shifts = simd_batch{ 8, 0, 0, 0, 0, 0, 0, 9 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 55 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 56> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0xffffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 56-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 1) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 2) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 3) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 4) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 5) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 6) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 6),
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 56-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7),
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 8) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 9) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 10) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 11) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 12) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 13) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 13),
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 56-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 15) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 16) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 17) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 18) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 19) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 20) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 56-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21),
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 22) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 23) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 24) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 25) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 26) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 27) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 27),
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 56-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 29) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 30) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 31) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 32) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 33) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 34) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 34),
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 56-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 36) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 37) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 38) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 39) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 40) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 41) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 41),
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 56-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 42),
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 43) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 44) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 45) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 46) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 47) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 47) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 48) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 48),
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 56-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 49),
+      SafeLoadAs<uint64_t>(in + 8 * 49) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 50) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 50) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 51) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 51) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 52) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 52) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 53) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 53) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 54) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 54) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 55) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 55),
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 8 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 56 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 57> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1ffffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 57-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 1) << 7,
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 2) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 3) << 21,
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 4) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 29 | SafeLoadAs<uint64_t>(in + 8 * 5) << 35,
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 6) << 42,
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 15 | SafeLoadAs<uint64_t>(in + 8 * 7) << 49,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 57-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 8 | SafeLoadAs<uint64_t>(in + 8 * 8) << 56,
+      SafeLoadAs<uint64_t>(in + 8 * 8),
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 9) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 10) << 13,
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 11) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 12) << 27,
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 13) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 23 | SafeLoadAs<uint64_t>(in + 8 * 14) << 41,
+    };
+    shifts = simd_batch{ 0, 1, 0, 0, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 57-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 15) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 9 | SafeLoadAs<uint64_t>(in + 8 * 16) << 55,
+      SafeLoadAs<uint64_t>(in + 8 * 16),
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 17) << 5,
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 18) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 19) << 19,
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 20) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 31 | SafeLoadAs<uint64_t>(in + 8 * 21) << 33,
+    };
+    shifts = simd_batch{ 0, 0, 2, 0, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 57-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 22) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 17 | SafeLoadAs<uint64_t>(in + 8 * 23) << 47,
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 10 | SafeLoadAs<uint64_t>(in + 8 * 24) << 54,
+      SafeLoadAs<uint64_t>(in + 8 * 24),
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 25) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 26) << 11,
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 27) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 28) << 25,
+    };
+    shifts = simd_batch{ 0, 0, 0, 3, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 57-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 29) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 25 | SafeLoadAs<uint64_t>(in + 8 * 30) << 39,
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 31) << 46,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 11 | SafeLoadAs<uint64_t>(in + 8 * 32) << 53,
+      SafeLoadAs<uint64_t>(in + 8 * 32),
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 33) << 3,
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 34) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 35) << 17,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 4, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 57-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 36) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 37) << 31,
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 38) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 19 | SafeLoadAs<uint64_t>(in + 8 * 39) << 45,
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 40) << 52,
+      SafeLoadAs<uint64_t>(in + 8 * 40),
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 41) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 42) << 9,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 5, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 57-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 43) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 44) << 23,
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 45) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 27 | SafeLoadAs<uint64_t>(in + 8 * 46) << 37,
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 47) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 47) >> 13 | SafeLoadAs<uint64_t>(in + 8 * 48) << 51,
+      SafeLoadAs<uint64_t>(in + 8 * 48),
+      SafeLoadAs<uint64_t>(in + 8 * 48) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 49) << 1,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 6, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 57-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 49) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 50) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 50) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 51) << 15,
+      SafeLoadAs<uint64_t>(in + 8 * 51) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 52) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 52) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 53) << 29,
+      SafeLoadAs<uint64_t>(in + 8 * 53) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 54) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 54) >> 21 | SafeLoadAs<uint64_t>(in + 8 * 55) << 43,
+      SafeLoadAs<uint64_t>(in + 8 * 55) >> 14 | SafeLoadAs<uint64_t>(in + 8 * 56) << 50,
+      SafeLoadAs<uint64_t>(in + 8 * 56),
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 7 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 57 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 58> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3ffffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 58-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 1) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 2) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 3) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 4) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 5) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 6) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 7) << 42,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 58-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 8) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 10 | SafeLoadAs<uint64_t>(in + 8 * 9) << 54,
+      SafeLoadAs<uint64_t>(in + 8 * 9),
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 10) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 11) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 12) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 13) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 14) << 26,
+    };
+    shifts = simd_batch{ 0, 0, 4, 0, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 58-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 15) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 16) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 17) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 14 | SafeLoadAs<uint64_t>(in + 8 * 18) << 50,
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 8 | SafeLoadAs<uint64_t>(in + 8 * 19) << 56,
+      SafeLoadAs<uint64_t>(in + 8 * 19),
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 20) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 21) << 10,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 2, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 58-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 22) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 23) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 24) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 25) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 26) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 27) << 46,
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 28) << 52,
+      SafeLoadAs<uint64_t>(in + 8 * 28),
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 58-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 29),
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 30) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 31) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 32) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 33) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 34) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 35) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 36) << 42,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 58-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 37) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 10 | SafeLoadAs<uint64_t>(in + 8 * 38) << 54,
+      SafeLoadAs<uint64_t>(in + 8 * 38),
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 39) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 40) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 41) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 42) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 43) << 26,
+    };
+    shifts = simd_batch{ 0, 0, 4, 0, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 58-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 44) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 45) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 46) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 14 | SafeLoadAs<uint64_t>(in + 8 * 47) << 50,
+      SafeLoadAs<uint64_t>(in + 8 * 47) >> 8 | SafeLoadAs<uint64_t>(in + 8 * 48) << 56,
+      SafeLoadAs<uint64_t>(in + 8 * 48),
+      SafeLoadAs<uint64_t>(in + 8 * 48) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 49) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 49) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 50) << 10,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 2, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 58-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 50) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 51) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 51) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 52) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 52) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 53) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 53) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 54) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 54) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 55) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 55) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 56) << 46,
+      SafeLoadAs<uint64_t>(in + 8 * 56) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 57) << 52,
+      SafeLoadAs<uint64_t>(in + 8 * 57),
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 6 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 58 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 59> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7ffffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 59-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 1) << 5,
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 2) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 3) << 15,
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 4) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 5) << 25,
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 6) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 29 | SafeLoadAs<uint64_t>(in + 8 * 7) << 35,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 59-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 8) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 19 | SafeLoadAs<uint64_t>(in + 8 * 9) << 45,
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 14 | SafeLoadAs<uint64_t>(in + 8 * 10) << 50,
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 9 | SafeLoadAs<uint64_t>(in + 8 * 11) << 55,
+      SafeLoadAs<uint64_t>(in + 8 * 11),
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 12) << 1,
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 13) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 14) << 11,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 4, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 59-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 15) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 16) << 21,
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 17) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 18) << 31,
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 19) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 23 | SafeLoadAs<uint64_t>(in + 8 * 20) << 41,
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 21) << 46,
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 13 | SafeLoadAs<uint64_t>(in + 8 * 22) << 51,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 59-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 8 | SafeLoadAs<uint64_t>(in + 8 * 23) << 56,
+      SafeLoadAs<uint64_t>(in + 8 * 23),
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 24) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 25) << 7,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 26) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 27) << 17,
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 28) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 29) << 27,
+    };
+    shifts = simd_batch{ 0, 3, 0, 0, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 59-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 30) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 27 | SafeLoadAs<uint64_t>(in + 8 * 31) << 37,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 32) << 42,
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 17 | SafeLoadAs<uint64_t>(in + 8 * 33) << 47,
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 34) << 52,
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 7 | SafeLoadAs<uint64_t>(in + 8 * 35) << 57,
+      SafeLoadAs<uint64_t>(in + 8 * 35),
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 36) << 3,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 2, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 59-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 37) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 38) << 13,
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 39) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 40) << 23,
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 41) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 31 | SafeLoadAs<uint64_t>(in + 8 * 42) << 33,
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 43) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 21 | SafeLoadAs<uint64_t>(in + 8 * 44) << 43,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 59-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 45) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 11 | SafeLoadAs<uint64_t>(in + 8 * 46) << 53,
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 6 | SafeLoadAs<uint64_t>(in + 8 * 47) << 58,
+      SafeLoadAs<uint64_t>(in + 8 * 47),
+      SafeLoadAs<uint64_t>(in + 8 * 47) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 48) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 48) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 49) << 9,
+      SafeLoadAs<uint64_t>(in + 8 * 49) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 50) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 50) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 51) << 19,
+    };
+    shifts = simd_batch{ 0, 0, 0, 1, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 59-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 51) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 52) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 52) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 53) << 29,
+      SafeLoadAs<uint64_t>(in + 8 * 53) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 54) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 54) >> 25 | SafeLoadAs<uint64_t>(in + 8 * 55) << 39,
+      SafeLoadAs<uint64_t>(in + 8 * 55) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 56) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 56) >> 15 | SafeLoadAs<uint64_t>(in + 8 * 57) << 49,
+      SafeLoadAs<uint64_t>(in + 8 * 57) >> 10 | SafeLoadAs<uint64_t>(in + 8 * 58) << 54,
+      SafeLoadAs<uint64_t>(in + 8 * 58),
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 5 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 59 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 60> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0xfffffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 60-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 1) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 2) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 3) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 4) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 5) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 6) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 7) << 28,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 60-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 8) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 9) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 10) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 11) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 12) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 13) << 52,
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 8 | SafeLoadAs<uint64_t>(in + 8 * 14) << 56,
+      SafeLoadAs<uint64_t>(in + 8 * 14),
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 60-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15),
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 16) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 17) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 18) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 19) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 20) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 21) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 22) << 28,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 60-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 23) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 24) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 25) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 26) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 27) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 28) << 52,
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 8 | SafeLoadAs<uint64_t>(in + 8 * 29) << 56,
+      SafeLoadAs<uint64_t>(in + 8 * 29),
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 60-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 31) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 32) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 33) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 34) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 35) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 36) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 37) << 28,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 60-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 38) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 39) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 40) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 41) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 42) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 43) << 52,
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 8 | SafeLoadAs<uint64_t>(in + 8 * 44) << 56,
+      SafeLoadAs<uint64_t>(in + 8 * 44),
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 60-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 45),
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 46) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 47) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 47) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 48) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 48) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 49) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 49) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 50) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 50) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 51) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 51) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 52) << 28,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 60-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 52) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 53) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 53) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 54) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 54) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 55) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 55) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 56) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 56) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 57) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 57) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 58) << 52,
+      SafeLoadAs<uint64_t>(in + 8 * 58) >> 8 | SafeLoadAs<uint64_t>(in + 8 * 59) << 56,
+      SafeLoadAs<uint64_t>(in + 8 * 59),
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 4 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 60 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 61> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x1fffffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 61-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 1) << 3,
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 2) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 3) << 9,
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 4) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 5) << 15,
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 6) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 7) << 21,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 61-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 8) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 9) << 27,
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 10) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 31 | SafeLoadAs<uint64_t>(in + 8 * 11) << 33,
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 12) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 25 | SafeLoadAs<uint64_t>(in + 8 * 13) << 39,
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 14) << 42,
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 19 | SafeLoadAs<uint64_t>(in + 8 * 15) << 45,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 61-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 16) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 13 | SafeLoadAs<uint64_t>(in + 8 * 17) << 51,
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 10 | SafeLoadAs<uint64_t>(in + 8 * 18) << 54,
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 7 | SafeLoadAs<uint64_t>(in + 8 * 19) << 57,
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 4 | SafeLoadAs<uint64_t>(in + 8 * 20) << 60,
+      SafeLoadAs<uint64_t>(in + 8 * 20),
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 21) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 22) << 5,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 1, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 61-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 23) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 24) << 11,
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 25) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 26) << 17,
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 27) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 28) << 23,
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 29) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 30) << 29,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 61-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 31) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 29 | SafeLoadAs<uint64_t>(in + 8 * 32) << 35,
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 33) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 23 | SafeLoadAs<uint64_t>(in + 8 * 34) << 41,
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 35) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 17 | SafeLoadAs<uint64_t>(in + 8 * 36) << 47,
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 14 | SafeLoadAs<uint64_t>(in + 8 * 37) << 50,
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 11 | SafeLoadAs<uint64_t>(in + 8 * 38) << 53,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 61-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 8 | SafeLoadAs<uint64_t>(in + 8 * 39) << 56,
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 5 | SafeLoadAs<uint64_t>(in + 8 * 40) << 59,
+      SafeLoadAs<uint64_t>(in + 8 * 40),
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 41) << 1,
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 42) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 43) << 7,
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 44) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 45) << 13,
+    };
+    shifts = simd_batch{ 0, 0, 2, 0, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 61-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 46) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 47) << 19,
+      SafeLoadAs<uint64_t>(in + 8 * 47) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 48) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 48) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 49) << 25,
+      SafeLoadAs<uint64_t>(in + 8 * 49) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 50) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 50) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 51) << 31,
+      SafeLoadAs<uint64_t>(in + 8 * 51) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 52) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 52) >> 27 | SafeLoadAs<uint64_t>(in + 8 * 53) << 37,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 61-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 53) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 54) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 54) >> 21 | SafeLoadAs<uint64_t>(in + 8 * 55) << 43,
+      SafeLoadAs<uint64_t>(in + 8 * 55) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 56) << 46,
+      SafeLoadAs<uint64_t>(in + 8 * 56) >> 15 | SafeLoadAs<uint64_t>(in + 8 * 57) << 49,
+      SafeLoadAs<uint64_t>(in + 8 * 57) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 58) << 52,
+      SafeLoadAs<uint64_t>(in + 8 * 58) >> 9 | SafeLoadAs<uint64_t>(in + 8 * 59) << 55,
+      SafeLoadAs<uint64_t>(in + 8 * 59) >> 6 | SafeLoadAs<uint64_t>(in + 8 * 60) << 58,
+      SafeLoadAs<uint64_t>(in + 8 * 60),
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 3 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 61 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 62> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x3fffffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 62-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 1) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 2) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 3) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 4) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 5) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 6) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 7) << 14,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 62-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 8) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 9) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 10) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 11) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 12) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 13) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 14) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 15) << 30,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 62-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 16) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 17) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 18) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 19) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 20) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 21) << 42,
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 22) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 23) << 46,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 62-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 24) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 14 | SafeLoadAs<uint64_t>(in + 8 * 25) << 50,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 26) << 52,
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 10 | SafeLoadAs<uint64_t>(in + 8 * 27) << 54,
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 8 | SafeLoadAs<uint64_t>(in + 8 * 28) << 56,
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 6 | SafeLoadAs<uint64_t>(in + 8 * 29) << 58,
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 4 | SafeLoadAs<uint64_t>(in + 8 * 30) << 60,
+      SafeLoadAs<uint64_t>(in + 8 * 30),
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 62-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 31),
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 32) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 33) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 34) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 35) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 36) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 37) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 38) << 14,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 62-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 39) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 40) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 41) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 42) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 43) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 44) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 45) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 46) << 30,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 62-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 47) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 47) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 48) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 48) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 49) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 49) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 50) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 50) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 51) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 51) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 52) << 42,
+      SafeLoadAs<uint64_t>(in + 8 * 52) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 53) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 53) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 54) << 46,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 62-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 54) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 55) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 55) >> 14 | SafeLoadAs<uint64_t>(in + 8 * 56) << 50,
+      SafeLoadAs<uint64_t>(in + 8 * 56) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 57) << 52,
+      SafeLoadAs<uint64_t>(in + 8 * 57) >> 10 | SafeLoadAs<uint64_t>(in + 8 * 58) << 54,
+      SafeLoadAs<uint64_t>(in + 8 * 58) >> 8 | SafeLoadAs<uint64_t>(in + 8 * 59) << 56,
+      SafeLoadAs<uint64_t>(in + 8 * 59) >> 6 | SafeLoadAs<uint64_t>(in + 8 * 60) << 58,
+      SafeLoadAs<uint64_t>(in + 8 * 60) >> 4 | SafeLoadAs<uint64_t>(in + 8 * 61) << 60,
+      SafeLoadAs<uint64_t>(in + 8 * 61),
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 2 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 62 * 8;
+    return in;
+  }
+};
+
+template<>
+struct Simd512UnpackerForWidth<uint64_t, 63> {
+
+  using simd_batch = xsimd::make_sized_batch_t<uint64_t, 8>;
+  static constexpr int kValuesUnpacked = 64;
+
+  static const uint8_t* unpack(const uint8_t* in, uint64_t* out) {
+    constexpr uint64_t kMask = 0x7fffffffffffffff;
+
+    simd_batch masks(kMask);
+    simd_batch words, shifts;
+    simd_batch results;
+    // extract 63-bit bundles 0 to 7
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 0),
+      SafeLoadAs<uint64_t>(in + 8 * 0) >> 63 | SafeLoadAs<uint64_t>(in + 8 * 1) << 1,
+      SafeLoadAs<uint64_t>(in + 8 * 1) >> 62 | SafeLoadAs<uint64_t>(in + 8 * 2) << 2,
+      SafeLoadAs<uint64_t>(in + 8 * 2) >> 61 | SafeLoadAs<uint64_t>(in + 8 * 3) << 3,
+      SafeLoadAs<uint64_t>(in + 8 * 3) >> 60 | SafeLoadAs<uint64_t>(in + 8 * 4) << 4,
+      SafeLoadAs<uint64_t>(in + 8 * 4) >> 59 | SafeLoadAs<uint64_t>(in + 8 * 5) << 5,
+      SafeLoadAs<uint64_t>(in + 8 * 5) >> 58 | SafeLoadAs<uint64_t>(in + 8 * 6) << 6,
+      SafeLoadAs<uint64_t>(in + 8 * 6) >> 57 | SafeLoadAs<uint64_t>(in + 8 * 7) << 7,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 63-bit bundles 8 to 15
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 7) >> 56 | SafeLoadAs<uint64_t>(in + 8 * 8) << 8,
+      SafeLoadAs<uint64_t>(in + 8 * 8) >> 55 | SafeLoadAs<uint64_t>(in + 8 * 9) << 9,
+      SafeLoadAs<uint64_t>(in + 8 * 9) >> 54 | SafeLoadAs<uint64_t>(in + 8 * 10) << 10,
+      SafeLoadAs<uint64_t>(in + 8 * 10) >> 53 | SafeLoadAs<uint64_t>(in + 8 * 11) << 11,
+      SafeLoadAs<uint64_t>(in + 8 * 11) >> 52 | SafeLoadAs<uint64_t>(in + 8 * 12) << 12,
+      SafeLoadAs<uint64_t>(in + 8 * 12) >> 51 | SafeLoadAs<uint64_t>(in + 8 * 13) << 13,
+      SafeLoadAs<uint64_t>(in + 8 * 13) >> 50 | SafeLoadAs<uint64_t>(in + 8 * 14) << 14,
+      SafeLoadAs<uint64_t>(in + 8 * 14) >> 49 | SafeLoadAs<uint64_t>(in + 8 * 15) << 15,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 63-bit bundles 16 to 23
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 15) >> 48 | SafeLoadAs<uint64_t>(in + 8 * 16) << 16,
+      SafeLoadAs<uint64_t>(in + 8 * 16) >> 47 | SafeLoadAs<uint64_t>(in + 8 * 17) << 17,
+      SafeLoadAs<uint64_t>(in + 8 * 17) >> 46 | SafeLoadAs<uint64_t>(in + 8 * 18) << 18,
+      SafeLoadAs<uint64_t>(in + 8 * 18) >> 45 | SafeLoadAs<uint64_t>(in + 8 * 19) << 19,
+      SafeLoadAs<uint64_t>(in + 8 * 19) >> 44 | SafeLoadAs<uint64_t>(in + 8 * 20) << 20,
+      SafeLoadAs<uint64_t>(in + 8 * 20) >> 43 | SafeLoadAs<uint64_t>(in + 8 * 21) << 21,
+      SafeLoadAs<uint64_t>(in + 8 * 21) >> 42 | SafeLoadAs<uint64_t>(in + 8 * 22) << 22,
+      SafeLoadAs<uint64_t>(in + 8 * 22) >> 41 | SafeLoadAs<uint64_t>(in + 8 * 23) << 23,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 63-bit bundles 24 to 31
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 23) >> 40 | SafeLoadAs<uint64_t>(in + 8 * 24) << 24,
+      SafeLoadAs<uint64_t>(in + 8 * 24) >> 39 | SafeLoadAs<uint64_t>(in + 8 * 25) << 25,
+      SafeLoadAs<uint64_t>(in + 8 * 25) >> 38 | SafeLoadAs<uint64_t>(in + 8 * 26) << 26,
+      SafeLoadAs<uint64_t>(in + 8 * 26) >> 37 | SafeLoadAs<uint64_t>(in + 8 * 27) << 27,
+      SafeLoadAs<uint64_t>(in + 8 * 27) >> 36 | SafeLoadAs<uint64_t>(in + 8 * 28) << 28,
+      SafeLoadAs<uint64_t>(in + 8 * 28) >> 35 | SafeLoadAs<uint64_t>(in + 8 * 29) << 29,
+      SafeLoadAs<uint64_t>(in + 8 * 29) >> 34 | SafeLoadAs<uint64_t>(in + 8 * 30) << 30,
+      SafeLoadAs<uint64_t>(in + 8 * 30) >> 33 | SafeLoadAs<uint64_t>(in + 8 * 31) << 31,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 63-bit bundles 32 to 39
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 31) >> 32 | SafeLoadAs<uint64_t>(in + 8 * 32) << 32,
+      SafeLoadAs<uint64_t>(in + 8 * 32) >> 31 | SafeLoadAs<uint64_t>(in + 8 * 33) << 33,
+      SafeLoadAs<uint64_t>(in + 8 * 33) >> 30 | SafeLoadAs<uint64_t>(in + 8 * 34) << 34,
+      SafeLoadAs<uint64_t>(in + 8 * 34) >> 29 | SafeLoadAs<uint64_t>(in + 8 * 35) << 35,
+      SafeLoadAs<uint64_t>(in + 8 * 35) >> 28 | SafeLoadAs<uint64_t>(in + 8 * 36) << 36,
+      SafeLoadAs<uint64_t>(in + 8 * 36) >> 27 | SafeLoadAs<uint64_t>(in + 8 * 37) << 37,
+      SafeLoadAs<uint64_t>(in + 8 * 37) >> 26 | SafeLoadAs<uint64_t>(in + 8 * 38) << 38,
+      SafeLoadAs<uint64_t>(in + 8 * 38) >> 25 | SafeLoadAs<uint64_t>(in + 8 * 39) << 39,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 63-bit bundles 40 to 47
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 39) >> 24 | SafeLoadAs<uint64_t>(in + 8 * 40) << 40,
+      SafeLoadAs<uint64_t>(in + 8 * 40) >> 23 | SafeLoadAs<uint64_t>(in + 8 * 41) << 41,
+      SafeLoadAs<uint64_t>(in + 8 * 41) >> 22 | SafeLoadAs<uint64_t>(in + 8 * 42) << 42,
+      SafeLoadAs<uint64_t>(in + 8 * 42) >> 21 | SafeLoadAs<uint64_t>(in + 8 * 43) << 43,
+      SafeLoadAs<uint64_t>(in + 8 * 43) >> 20 | SafeLoadAs<uint64_t>(in + 8 * 44) << 44,
+      SafeLoadAs<uint64_t>(in + 8 * 44) >> 19 | SafeLoadAs<uint64_t>(in + 8 * 45) << 45,
+      SafeLoadAs<uint64_t>(in + 8 * 45) >> 18 | SafeLoadAs<uint64_t>(in + 8 * 46) << 46,
+      SafeLoadAs<uint64_t>(in + 8 * 46) >> 17 | SafeLoadAs<uint64_t>(in + 8 * 47) << 47,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 63-bit bundles 48 to 55
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 47) >> 16 | SafeLoadAs<uint64_t>(in + 8 * 48) << 48,
+      SafeLoadAs<uint64_t>(in + 8 * 48) >> 15 | SafeLoadAs<uint64_t>(in + 8 * 49) << 49,
+      SafeLoadAs<uint64_t>(in + 8 * 49) >> 14 | SafeLoadAs<uint64_t>(in + 8 * 50) << 50,
+      SafeLoadAs<uint64_t>(in + 8 * 50) >> 13 | SafeLoadAs<uint64_t>(in + 8 * 51) << 51,
+      SafeLoadAs<uint64_t>(in + 8 * 51) >> 12 | SafeLoadAs<uint64_t>(in + 8 * 52) << 52,
+      SafeLoadAs<uint64_t>(in + 8 * 52) >> 11 | SafeLoadAs<uint64_t>(in + 8 * 53) << 53,
+      SafeLoadAs<uint64_t>(in + 8 * 53) >> 10 | SafeLoadAs<uint64_t>(in + 8 * 54) << 54,
+      SafeLoadAs<uint64_t>(in + 8 * 54) >> 9 | SafeLoadAs<uint64_t>(in + 8 * 55) << 55,
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 0 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    // extract 63-bit bundles 56 to 63
+    words = simd_batch{
+      SafeLoadAs<uint64_t>(in + 8 * 55) >> 8 | SafeLoadAs<uint64_t>(in + 8 * 56) << 56,
+      SafeLoadAs<uint64_t>(in + 8 * 56) >> 7 | SafeLoadAs<uint64_t>(in + 8 * 57) << 57,
+      SafeLoadAs<uint64_t>(in + 8 * 57) >> 6 | SafeLoadAs<uint64_t>(in + 8 * 58) << 58,
+      SafeLoadAs<uint64_t>(in + 8 * 58) >> 5 | SafeLoadAs<uint64_t>(in + 8 * 59) << 59,
+      SafeLoadAs<uint64_t>(in + 8 * 59) >> 4 | SafeLoadAs<uint64_t>(in + 8 * 60) << 60,
+      SafeLoadAs<uint64_t>(in + 8 * 60) >> 3 | SafeLoadAs<uint64_t>(in + 8 * 61) << 61,
+      SafeLoadAs<uint64_t>(in + 8 * 61) >> 2 | SafeLoadAs<uint64_t>(in + 8 * 62) << 62,
+      SafeLoadAs<uint64_t>(in + 8 * 62),
+    };
+    shifts = simd_batch{ 0, 0, 0, 0, 0, 0, 0, 1 };
+    results = (words >> shifts) & masks;
+    results.store_unaligned(out);
+    out += 8;
+
+    in += 63 * 8;
+    return in;
+  }
+};
+
+
+}  // namespace arrow::internal
diff --git a/cpp/src/arrow/util/bpacking_avx2.cc b/cpp/src/arrow/util/bpacking_simd_avx2.cc
similarity index 55%
rename from cpp/src/arrow/util/bpacking_avx2.cc
rename to cpp/src/arrow/util/bpacking_simd_avx2.cc
index 84f091594c14b..c698ae3389ed6 100644
--- a/cpp/src/arrow/util/bpacking_avx2.cc
+++ b/cpp/src/arrow/util/bpacking_simd_avx2.cc
@@ -15,15 +15,28 @@
 // specific language governing permissions and limitations
 // under the License.
 
-#include "arrow/util/bpacking_avx2_internal.h"
+#include "arrow/util/bpacking_dispatch_internal.h"
 #include "arrow/util/bpacking_simd256_generated_internal.h"
 #include "arrow/util/bpacking_simd_internal.h"
 
 namespace arrow::internal {
 
-int unpack32_avx2(const uint8_t* in, uint32_t* out, int batch_size, int num_bits) {
-  return unpack32_specialized<UnpackBits256<DispatchLevel::AVX2>>(
-      reinterpret_cast<const uint32_t*>(in), out, batch_size, num_bits);
+template <typename Uint>
+int unpack_avx2(const uint8_t* in, Uint* out, int batch_size, int num_bits) {
+  return unpack_jump<Simd256UnpackerForWidth>(in, out, batch_size, num_bits);
 }
 
+template int unpack_avx2<uint16_t>(const uint8_t*, uint16_t*, int, int);
+template int unpack_avx2<uint32_t>(const uint8_t*, uint32_t*, int, int);
+template int unpack_avx2<uint64_t>(const uint8_t*, uint64_t*, int, int);
+
+template <typename Uint>
+UnpackFn<Uint> get_unpack_fn_avx2(int num_bits) {
+  return get_unpack_fn<Simd256UnpackerForWidth, Uint>(num_bits);
+}
+
+template UnpackFn<uint16_t> get_unpack_fn_avx2<uint16_t>(int);
+template UnpackFn<uint32_t> get_unpack_fn_avx2<uint32_t>(int);
+template UnpackFn<uint64_t> get_unpack_fn_avx2<uint64_t>(int);
+
 }  // namespace arrow::internal
diff --git a/cpp/src/arrow/util/bpacking_avx512.cc b/cpp/src/arrow/util/bpacking_simd_avx512.cc
similarity index 59%
rename from cpp/src/arrow/util/bpacking_avx512.cc
rename to cpp/src/arrow/util/bpacking_simd_avx512.cc
index 35de0dd5b476a..0fa52012ae9e1 100644
--- a/cpp/src/arrow/util/bpacking_avx512.cc
+++ b/cpp/src/arrow/util/bpacking_simd_avx512.cc
@@ -15,15 +15,26 @@
 // specific language governing permissions and limitations
 // under the License.
 
-#include "arrow/util/bpacking_avx512_internal.h"
+#include "arrow/util/bpacking_dispatch_internal.h"
 #include "arrow/util/bpacking_simd512_generated_internal.h"
 #include "arrow/util/bpacking_simd_internal.h"
 
 namespace arrow::internal {
 
-int unpack32_avx512(const uint8_t* in, uint32_t* out, int batch_size, int num_bits) {
-  return unpack32_specialized<UnpackBits512<DispatchLevel::AVX512>>(
-      reinterpret_cast<const uint32_t*>(in), out, batch_size, num_bits);
+template <typename Uint>
+int unpack_avx512(const uint8_t* in, Uint* out, int batch_size, int num_bits) {
+  return unpack_jump<Simd512UnpackerForWidth>(in, out, batch_size, num_bits);
 }
 
+template int unpack_avx512<uint32_t>(const uint8_t*, uint32_t*, int, int);
+template int unpack_avx512<uint64_t>(const uint8_t*, uint64_t*, int, int);
+
+template <typename Uint>
+UnpackFn<Uint> get_unpack_fn_avx512(int num_bits) {
+  return get_unpack_fn<Simd512UnpackerForWidth, Uint>(num_bits);
+}
+
+template UnpackFn<uint32_t> get_unpack_fn_avx512<uint32_t>(int);
+template UnpackFn<uint64_t> get_unpack_fn_avx512<uint64_t>(int);
+
 }  // namespace arrow::internal
diff --git a/cpp/src/arrow/util/bpacking_simd_codegen.py b/cpp/src/arrow/util/bpacking_simd_codegen.py
index 9464908c02175..5bb909de89e20 100755
--- a/cpp/src/arrow/util/bpacking_simd_codegen.py
+++ b/cpp/src/arrow/util/bpacking_simd_codegen.py
@@ -22,182 +22,231 @@
 #   python bpacking_simd_codegen.py 256 > bpacking_simd256_generated_internal.h
 #   python bpacking_simd_codegen.py 512 > bpacking_simd512_generated_internal.h
 
-from functools import partial
+import dataclasses
 import sys
 from textwrap import dedent, indent
 
 
-class UnpackGenerator:
+LICENSE = """// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+"""
+
+
+@dataclasses.dataclass
+class UnpackStructGenerator:
+    out_bit_width: int
+    simd_bit_width: int
+
+    @property
+    def simd_byte_width(self) -> int:
+        return self.simd_bit_width // 8
+
+    @property
+    def simd_value_count(self) -> int:
+        return self.simd_bit_width // self.out_bit_width
+
+    @property
+    def out_byte_width(self) -> int:
+        return self.out_bit_width // 8
+
+    @property
+    def out_type(self) -> str:
+        return f"uint{self.out_bit_width}_t"
+
+    @property
+    def struct_name(self) -> str:
+        return f"Simd{self.simd_bit_width}UnpackerForWidth"
+
+    def struct_specialization(self, bit: int) -> str:
+        return f"{self.struct_name}<{self.out_type}, {bit}>"
+
+    def print_struct_declaration(self):
+        print("template<typename Uint, int BitWidth>")
+        print(f"struct {self.struct_name};")
+
+    def __post_init__(self):
+        # whole logic based on these assumptions
+        if self.simd_value_count > self.out_bit_width:
+            raise ValueError("The code is invalid for this size")
+
+        if self.simd_bit_width % self.out_bit_width != 0:
+            raise ("SIMD bit width should be a multiple of output width")
+
+    def print_unpack_bit_func(self, bit: int):
+        def p(code, level=2):
+            print(indent(code, prefix="  " * level))
+
+        p(
+            f"static const uint8_t* unpack(const uint8_t* in, {self.out_type}* out) {{",
+            level=1,
+        )
 
-    def __init__(self, simd_width):
-        self.simd_width = simd_width
-        if simd_width % 32 != 0:
-            raise("SIMD bit width should be a multiple of 32")
-        self.simd_byte_width = simd_width // 8
-
-    def print_unpack_bit0_func(self):
-        print(
-            "inline static const uint32_t* unpack0_32(const uint32_t* in, uint32_t* out) {")
-        print("  memset(out, 0x0, 32 * sizeof(*out));")
-        print("  out += 32;")
-        print("")
-        print("  return in;")
-        print("}")
+        mask = (1 << bit) - 1
+        p(f"constexpr {self.out_type} kMask = 0x{mask:0x};")
+        print()
+        p("simd_batch masks(kMask);")
+        p("simd_batch words, shifts;")
+        p("simd_batch results;")
 
+        def safe_load(index):
+            return f"SafeLoadAs<{self.out_type}>(in + {self.out_byte_width} * {index})"
 
-    def print_unpack_bit32_func(self):
-        print(
-            "inline static const uint32_t* unpack32_32(const uint32_t* in, uint32_t* out) {")
-        print("  memcpy(out, in, 32 * sizeof(*out));")
-        print("  in += 32;")
-        print("  out += 32;")
-        print("")
-        print("  return in;")
-        print("}")
-
-    def print_unpack_bit_func(self, bit):
-        def p(code):
-            print(indent(code, prefix='  '))
+        def static_cast_as_needed(str):
+            if self.out_bit_width < 32:
+                return f"static_cast<{self.out_type}>({str})"
+            return str
 
         shift = 0
         shifts = []
         in_index = 0
         inls = []
-        mask = (1 << bit) - 1
-        bracket = "{"
 
-        print(f"inline static const uint32_t* unpack{bit}_32(const uint32_t* in, uint32_t* out) {{")
-        p(dedent(f"""\
-            uint32_t mask = 0x{mask:0x};
-
-            simd_batch masks(mask);
-            simd_batch words, shifts;
-            simd_batch results;
-            """))
-
-        def safe_load(index):
-            return f"SafeLoad<uint32_t>(in + {index})"
-
-        for i in range(32):
-            if shift + bit == 32:
+        for i in range(self.out_bit_width):
+            if shift + bit == self.out_bit_width:
                 shifts.append(shift)
                 inls.append(safe_load(in_index))
                 in_index += 1
                 shift = 0
-            elif shift + bit > 32:  # cross the boundary
+            elif shift + bit > self.out_bit_width:  # cross the boundary
                 inls.append(
-                    f"{safe_load(in_index)} >> {shift} | {safe_load(in_index + 1)} << {32 - shift}")
+                    static_cast_as_needed(
+                        f"{safe_load(in_index)} >> {shift} "
+                        f"| {safe_load(in_index + 1)} << {self.out_bit_width - shift}"
+                    )
+                )
                 in_index += 1
-                shift = bit - (32 - shift)
+                shift = bit - (self.out_bit_width - shift)
                 shifts.append(0)  # zero shift
             else:
                 shifts.append(shift)
                 inls.append(safe_load(in_index))
                 shift += bit
 
-        bytes_per_batch = self.simd_byte_width
-        words_per_batch = bytes_per_batch // 4
-
         one_word_template = dedent("""\
-            words = simd_batch{{ {words} }};
             shifts = simd_batch{{ {shifts} }};
             results = (words >> shifts) & masks;
             results.store_unaligned(out);
             out += {words_per_batch};
             """)
 
-        for start in range(0, 32, words_per_batch):
-            stop = start + words_per_batch;
+        for start in range(0, self.out_bit_width, self.simd_value_count):
+            stop = start + self.simd_value_count
             p(f"""// extract {bit}-bit bundles {start} to {stop - 1}""")
-            p(one_word_template.format(
-                words=", ".join(inls[start:stop]),
-                shifts=", ".join(map(str, shifts[start:stop])),
-                words_per_batch=words_per_batch))
-
-        p(dedent(f"""\
-            in += {bit};
-            return in;"""))
-        print("}")
-
-
-def print_copyright():
-    print(dedent("""\
-        // Licensed to the Apache Software Foundation (ASF) under one
-        // or more contributor license agreements.  See the NOTICE file
-        // distributed with this work for additional information
-        // regarding copyright ownership.  The ASF licenses this file
-        // to you under the Apache License, Version 2.0 (the
-        // "License"); you may not use this file except in compliance
-        // with the License.  You may obtain a copy of the License at
-        //
-        //   http://www.apache.org/licenses/LICENSE-2.0
-        //
-        // Unless required by applicable law or agreed to in writing,
-        // software distributed under the License is distributed on an
-        // "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
-        // KIND, either express or implied.  See the License for the
-        // specific language governing permissions and limitations
-        // under the License.
-        """))
-
-
-def print_note():
-    print("// Automatically generated file; DO NOT EDIT.")
-    print()
-
+            p("words = simd_batch{")
+            for word_part in inls[start:stop]:
+                p(f"{word_part},", level=3)
+            p("};")
+            p(
+                one_word_template.format(
+                    shifts=", ".join(map(str, shifts[start:stop])),
+                    words_per_batch=self.simd_value_count,
+                )
+            )
+
+        p(f"in += {bit} * {self.out_byte_width};")
+        p("return in;")
+        p("}", level=1)
+
+    def print_struct_k(self, bit: int):
+        print("template<>")
+        print(f"struct {self.struct_specialization(bit)} {{")
+        print()
+        print(
+            "  using simd_batch = xsimd::make_sized_batch_t<"
+            f"{self.out_type}, {self.simd_value_count}>;"
+        )
+        print(f"  static constexpr int kValuesUnpacked = {self.out_bit_width};")
+        print()
+        self.print_unpack_bit_func(bit)
+        print("};")
 
-def main(simd_width):
-    print_copyright()
-    print_note()
+    def print_structs(self):
+        for bit in range(1, self.out_bit_width):
+            self.print_struct_k(bit)
+            print()
 
-    struct_name = f"UnpackBits{simd_width}"
 
-    # NOTE: templating the UnpackBits struct on the dispatch level avoids
-    # potential name collisions if there are several UnpackBits generations
-    # with the same SIMD width on a given architecture.
+@dataclasses.dataclass
+class UnpackFileGenerator:
+    generators: list[UnpackStructGenerator]
 
-    print(dedent(f"""\
-        #pragma once
+    def print_license(self):
+        print(LICENSE)
 
-        #include <cstdint>
-        #include <cstring>
+    def print_note(self):
+        print("// WARNING: this file is generated, DO NOT EDIT.")
+        print("// Usage:")
+        print(f"//   python {' '.join(sys.orig_argv[1:])}")
 
-        #include <xsimd/xsimd.hpp>
+    def print_headers(self):
+        print("#include <cstdint>")
+        print("#include <cstring>")
+        print()
+        print("#include <xsimd/xsimd.hpp>")
+        print()
+        print('#include "arrow/util/ubsan.h"')
 
-        #include "arrow/util/dispatch_internal.h"
-        #include "arrow/util/ubsan.h"
+    def print_file_top(self):
+        print("#pragma once")
+        print()
+        self.print_headers()
+        print()
+        print("namespace arrow::internal {")
+        print()
+        print("using ::arrow::util::SafeLoadAs;")
 
-        namespace arrow {{
-        namespace internal {{
-        namespace {{
+    def print_file_bottom(self):
+        print("}  // namespace arrow::internal")
 
-        using ::arrow::util::SafeLoad;
+    def print_structs(self):
+        delclared = set()
 
-        template <DispatchLevel level>
-        struct {struct_name} {{
+        for gen in self.generators:
+            if gen.simd_bit_width not in delclared:
+                gen.print_struct_declaration()
+                print()
+                delclared.add(gen.simd_bit_width)
 
-        using simd_batch = xsimd::make_sized_batch_t<uint32_t, {simd_width//32}>;
-        """))
+            gen.print_structs()
+            print()
 
-    gen = UnpackGenerator(simd_width)
-    gen.print_unpack_bit0_func()
-    print()
-    for i in range(1, 32):
-        gen.print_unpack_bit_func(i)
+    def print_file(self):
+        self.print_license()
+        self.print_note()
+        print()
+        self.print_file_top()
         print()
-    gen.print_unpack_bit32_func()
-    print()
+        self.print_structs()
+        self.print_file_bottom()
+
 
-    print(dedent(f"""\
-        }};  // struct {struct_name}
+def main(simd_width, outputs):
+    generators = []
+    for out_width in outputs:
+        try:
+            generators.append(UnpackStructGenerator(out_width, simd_width))
+        except Exception as e:
+            print(f"WARNING skipping bit width {out_width}: {e}", file=sys.stderr)
 
-        }}  // namespace
-        }}  // namespace internal
-        }}  // namespace arrow
-        """))
+    gen = UnpackFileGenerator(generators)
+    gen.print_file()
 
 
-if __name__ == '__main__':
+if __name__ == "__main__":
     usage = f"""Usage: {__file__} <SIMD bit-width>"""
     if len(sys.argv) != 2:
         raise ValueError(usage)
@@ -206,4 +255,4 @@ def main(simd_width):
     except ValueError:
         raise ValueError(usage)
 
-    main(simd_width)
+    main(simd_width, [16, 32, 64])
diff --git a/cpp/src/arrow/util/bpacking_simd_internal.h b/cpp/src/arrow/util/bpacking_simd_internal.h
index 98e192e7cb0bd..dc3cafe20c4ad 100644
--- a/cpp/src/arrow/util/bpacking_simd_internal.h
+++ b/cpp/src/arrow/util/bpacking_simd_internal.h
@@ -15,124 +15,108 @@
 // specific language governing permissions and limitations
 // under the License.
 
-#include "arrow/util/dispatch_internal.h"
-#include "arrow/util/logging.h"
-
-namespace arrow {
-namespace internal {
-
-template <typename UnpackBits>
-static int unpack32_specialized(const uint32_t* in, uint32_t* out, int batch_size,
-                                int num_bits) {
-  batch_size = batch_size / 32 * 32;
-  int num_loops = batch_size / 32;
-
-  switch (num_bits) {
-    case 0:
-      for (int i = 0; i < num_loops; ++i) in = UnpackBits::unpack0_32(in, out + i * 32);
-      break;
-    case 1:
-      for (int i = 0; i < num_loops; ++i) in = UnpackBits::unpack1_32(in, out + i * 32);
-      break;
-    case 2:
-      for (int i = 0; i < num_loops; ++i) in = UnpackBits::unpack2_32(in, out + i * 32);
-      break;
-    case 3:
-      for (int i = 0; i < num_loops; ++i) in = UnpackBits::unpack3_32(in, out + i * 32);
-      break;
-    case 4:
-      for (int i = 0; i < num_loops; ++i) in = UnpackBits::unpack4_32(in, out + i * 32);
-      break;
-    case 5:
-      for (int i = 0; i < num_loops; ++i) in = UnpackBits::unpack5_32(in, out + i * 32);
-      break;
-    case 6:
-      for (int i = 0; i < num_loops; ++i) in = UnpackBits::unpack6_32(in, out + i * 32);
-      break;
-    case 7:
-      for (int i = 0; i < num_loops; ++i) in = UnpackBits::unpack7_32(in, out + i * 32);
-      break;
-    case 8:
-      for (int i = 0; i < num_loops; ++i) in = UnpackBits::unpack8_32(in, out + i * 32);
-      break;
-    case 9:
-      for (int i = 0; i < num_loops; ++i) in = UnpackBits::unpack9_32(in, out + i * 32);
-      break;
-    case 10:
-      for (int i = 0; i < num_loops; ++i) in = UnpackBits::unpack10_32(in, out + i * 32);
-      break;
-    case 11:
-      for (int i = 0; i < num_loops; ++i) in = UnpackBits::unpack11_32(in, out + i * 32);
-      break;
-    case 12:
-      for (int i = 0; i < num_loops; ++i) in = UnpackBits::unpack12_32(in, out + i * 32);
-      break;
-    case 13:
-      for (int i = 0; i < num_loops; ++i) in = UnpackBits::unpack13_32(in, out + i * 32);
-      break;
-    case 14:
-      for (int i = 0; i < num_loops; ++i) in = UnpackBits::unpack14_32(in, out + i * 32);
-      break;
-    case 15:
-      for (int i = 0; i < num_loops; ++i) in = UnpackBits::unpack15_32(in, out + i * 32);
-      break;
-    case 16:
-      for (int i = 0; i < num_loops; ++i) in = UnpackBits::unpack16_32(in, out + i * 32);
-      break;
-    case 17:
-      for (int i = 0; i < num_loops; ++i) in = UnpackBits::unpack17_32(in, out + i * 32);
-      break;
-    case 18:
-      for (int i = 0; i < num_loops; ++i) in = UnpackBits::unpack18_32(in, out + i * 32);
-      break;
-    case 19:
-      for (int i = 0; i < num_loops; ++i) in = UnpackBits::unpack19_32(in, out + i * 32);
-      break;
-    case 20:
-      for (int i = 0; i < num_loops; ++i) in = UnpackBits::unpack20_32(in, out + i * 32);
-      break;
-    case 21:
-      for (int i = 0; i < num_loops; ++i) in = UnpackBits::unpack21_32(in, out + i * 32);
-      break;
-    case 22:
-      for (int i = 0; i < num_loops; ++i) in = UnpackBits::unpack22_32(in, out + i * 32);
-      break;
-    case 23:
-      for (int i = 0; i < num_loops; ++i) in = UnpackBits::unpack23_32(in, out + i * 32);
-      break;
-    case 24:
-      for (int i = 0; i < num_loops; ++i) in = UnpackBits::unpack24_32(in, out + i * 32);
-      break;
-    case 25:
-      for (int i = 0; i < num_loops; ++i) in = UnpackBits::unpack25_32(in, out + i * 32);
-      break;
-    case 26:
-      for (int i = 0; i < num_loops; ++i) in = UnpackBits::unpack26_32(in, out + i * 32);
-      break;
-    case 27:
-      for (int i = 0; i < num_loops; ++i) in = UnpackBits::unpack27_32(in, out + i * 32);
-      break;
-    case 28:
-      for (int i = 0; i < num_loops; ++i) in = UnpackBits::unpack28_32(in, out + i * 32);
-      break;
-    case 29:
-      for (int i = 0; i < num_loops; ++i) in = UnpackBits::unpack29_32(in, out + i * 32);
-      break;
-    case 30:
-      for (int i = 0; i < num_loops; ++i) in = UnpackBits::unpack30_32(in, out + i * 32);
-      break;
-    case 31:
-      for (int i = 0; i < num_loops; ++i) in = UnpackBits::unpack31_32(in, out + i * 32);
-      break;
-    case 32:
-      for (int i = 0; i < num_loops; ++i) in = UnpackBits::unpack32_32(in, out + i * 32);
-      break;
-    default:
-      ARROW_DCHECK(false) << "Unsupported num_bits";
-  }
-
-  return batch_size;
-}
-
-}  // namespace internal
-}  // namespace arrow
+#pragma once
+
+#include "arrow/util/bpacking_internal.h"
+#include "arrow/util/visibility.h"
+
+#include <cstdint>
+
+namespace arrow::internal {
+
+#if defined(ARROW_HAVE_NEON)
+
+template <typename Uint>
+ARROW_EXPORT int unpack_neon(const uint8_t* in, Uint* out, int batch_size, int num_bits);
+
+extern template ARROW_TEMPLATE_EXPORT int unpack_neon<uint16_t>(const uint8_t*, uint16_t*,
+                                                                int, int);
+
+extern template ARROW_TEMPLATE_EXPORT int unpack_neon<uint32_t>(const uint8_t*, uint32_t*,
+                                                                int, int);
+
+extern template ARROW_TEMPLATE_EXPORT int unpack_neon<uint64_t>(const uint8_t*, uint64_t*,
+                                                                int, int);
+
+template <typename Uint>
+ARROW_EXPORT UnpackFn<Uint> get_unpack_fn_neon(int num_bits);
+
+extern template ARROW_TEMPLATE_EXPORT UnpackFn<uint16_t> get_unpack_fn_neon<uint16_t>(
+    int);
+
+extern template ARROW_TEMPLATE_EXPORT UnpackFn<uint32_t> get_unpack_fn_neon<uint32_t>(
+    int);
+
+extern template ARROW_TEMPLATE_EXPORT UnpackFn<uint64_t> get_unpack_fn_neon<uint64_t>(
+    int);
+
+#endif
+
+#if defined(ARROW_HAVE_SSE4_2)
+
+template <typename Uint>
+ARROW_EXPORT int unpack_sse4_2(const uint8_t* in, Uint* out, int batch_size,
+                               int num_bits);
+
+extern template ARROW_TEMPLATE_EXPORT int unpack_sse4_2<uint16_t>(const uint8_t*,
+                                                                  uint16_t*, int, int);
+
+extern template ARROW_TEMPLATE_EXPORT int unpack_sse4_2<uint32_t>(const uint8_t*,
+                                                                  uint32_t*, int, int);
+
+extern template ARROW_TEMPLATE_EXPORT int unpack_sse4_2<uint64_t>(const uint8_t*,
+                                                                  uint64_t*, int, int);
+#endif
+
+#if defined(ARROW_HAVE_AVX2) || defined(ARROW_HAVE_RUNTIME_AVX2)
+
+template <typename Uint>
+ARROW_EXPORT int unpack_avx2(const uint8_t* in, Uint* out, int batch_size, int num_bits);
+
+extern template ARROW_TEMPLATE_EXPORT int unpack_avx2<uint16_t>(const uint8_t*, uint16_t*,
+                                                                int, int);
+
+extern template ARROW_TEMPLATE_EXPORT int unpack_avx2<uint32_t>(const uint8_t*, uint32_t*,
+                                                                int, int);
+
+extern template ARROW_TEMPLATE_EXPORT int unpack_avx2<uint64_t>(const uint8_t*, uint64_t*,
+                                                                int, int);
+
+template <typename Uint>
+ARROW_EXPORT UnpackFn<Uint> get_unpack_fn_avx2(int num_bits);
+
+extern template ARROW_TEMPLATE_EXPORT UnpackFn<uint16_t> get_unpack_fn_avx2<uint16_t>(
+    int);
+
+extern template ARROW_TEMPLATE_EXPORT UnpackFn<uint32_t> get_unpack_fn_avx2<uint32_t>(
+    int);
+
+extern template ARROW_TEMPLATE_EXPORT UnpackFn<uint64_t> get_unpack_fn_avx2<uint64_t>(
+    int);
+
+#endif
+
+#if defined(ARROW_HAVE_AVX512) || defined(ARROW_HAVE_RUNTIME_AVX512)
+
+template <typename Uint>
+ARROW_EXPORT int unpack_avx512(const uint8_t* in, Uint* out, int batch_size,
+                               int num_bits);
+
+extern template ARROW_TEMPLATE_EXPORT int unpack_avx512<uint32_t>(const uint8_t*,
+                                                                  uint32_t*, int, int);
+
+extern template ARROW_TEMPLATE_EXPORT int unpack_avx512<uint64_t>(const uint8_t*,
+                                                                  uint64_t*, int, int);
+
+template <typename Uint>
+ARROW_EXPORT UnpackFn<Uint> get_unpack_fn_avx512(int num_bits);
+
+extern template ARROW_TEMPLATE_EXPORT UnpackFn<uint32_t> get_unpack_fn_avx512<uint32_t>(
+    int);
+
+extern template ARROW_TEMPLATE_EXPORT UnpackFn<uint64_t> get_unpack_fn_avx512<uint64_t>(
+    int);
+
+#endif
+
+}  // namespace arrow::internal
diff --git a/cpp/src/arrow/util/bpacking_simd_min.cc b/cpp/src/arrow/util/bpacking_simd_min.cc
new file mode 100644
index 0000000000000..b95c0352b7565
--- /dev/null
+++ b/cpp/src/arrow/util/bpacking_simd_min.cc
@@ -0,0 +1,59 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#include "arrow/util/bpacking_dispatch_internal.h"
+#include "arrow/util/bpacking_simd128_generated_internal.h"
+#include "arrow/util/bpacking_simd_internal.h"
+
+namespace arrow::internal {
+
+#if defined(ARROW_HAVE_NEON)
+
+template <typename Uint>
+int unpack_neon(const uint8_t* in, Uint* out, int batch_size, int num_bits) {
+  return unpack_jump<Simd128UnpackerForWidth>(in, out, batch_size, num_bits);
+}
+
+template int unpack_neon<uint16_t>(const uint8_t*, uint16_t*, int, int);
+template int unpack_neon<uint32_t>(const uint8_t*, uint32_t*, int, int);
+template int unpack_neon<uint64_t>(const uint8_t*, uint64_t*, int, int);
+
+template <typename Uint>
+UnpackFn<Uint> get_unpack_fn_neon(int num_bits) {
+  return get_unpack_fn<Simd128UnpackerForWidth, Uint>(num_bits);
+}
+
+template UnpackFn<uint16_t> get_unpack_fn_neon<uint16_t>(int);
+template UnpackFn<uint32_t> get_unpack_fn_neon<uint32_t>(int);
+template UnpackFn<uint64_t> get_unpack_fn_neon<uint64_t>(int);
+
+#endif
+
+#if defined(ARROW_HAVE_SSE4_2)
+
+template <typename Uint>
+int unpack_sse4_2(const uint8_t* in, Uint* out, int batch_size, int num_bits) {
+  return unpack_jump<Simd128UnpackerForWidth>(in, out, batch_size, num_bits);
+}
+
+template int unpack_sse4_2<uint16_t>(const uint8_t*, uint16_t*, int, int);
+template int unpack_sse4_2<uint32_t>(const uint8_t*, uint32_t*, int, int);
+template int unpack_sse4_2<uint64_t>(const uint8_t*, uint64_t*, int, int);
+
+#endif
+
+}  // namespace arrow::internal
diff --git a/cpp/src/arrow/util/bpacking_test.cc b/cpp/src/arrow/util/bpacking_test.cc
index c2dd4748a44a3..92566eb5989f3 100644
--- a/cpp/src/arrow/util/bpacking_test.cc
+++ b/cpp/src/arrow/util/bpacking_test.cc
@@ -24,18 +24,13 @@
 #include "arrow/testing/util.h"
 #include "arrow/util/bit_stream_utils_internal.h"
 #include "arrow/util/bpacking_internal.h"
+#include "arrow/util/bpacking_scalar_internal.h"
+#include "arrow/util/bpacking_simd_internal.h"
 #include "arrow/util/logging.h"
 
 #if defined(ARROW_HAVE_RUNTIME_AVX2)
-#  include "arrow/util/bpacking_avx2_internal.h"
 #  include "arrow/util/cpu_info.h"
 #endif
-#if defined(ARROW_HAVE_RUNTIME_AVX512)
-#  include "arrow/util/bpacking_avx512_internal.h"
-#endif
-#if defined(ARROW_HAVE_NEON)
-#  include "arrow/util/bpacking_neon_internal.h"
-#endif
 
 namespace arrow::internal {
 
@@ -57,7 +52,7 @@ std::vector<uint8_t> GenerateRandomPackedValues(int32_t num_values, int32_t bit_
   constexpr uint32_t kSeed = 3214;
   EXPECT_OK_AND_ASSIGN(const auto num_bytes, GetNumBytes(num_values, bit_width));
 
-  std::vector<uint8_t> out(num_bytes);
+  std::vector<uint8_t> out(std::max(1, num_bytes));  // We need a valid pointer for size 0
   random_bytes(num_bytes, kSeed, out.data());
 
   return out;
@@ -119,16 +114,12 @@ const uint8_t* GetNextAlignedByte(const uint8_t* ptr, std::size_t alignment) {
   return ptr + bytes_to_add;
 }
 
-struct TestUnpackSize {
-  int32_t num_values;
-  int32_t bit_width;
-};
-
-class TestUnpack : public ::testing::TestWithParam<TestUnpackSize> {
+class TestUnpack : public ::testing::TestWithParam<int> {
  protected:
   template <typename Int>
-  void TestRoundtripAlignment(UnpackFunc<Int> unpack, std::size_t alignment_offset) {
-    auto [num_values, bit_width] = GetParam();
+  void TestRoundtripAlignment(UnpackFunc<Int> unpack, int bit_width,
+                              std::size_t alignment_offset) {
+    int num_values = GetParam();
 
     // Assume std::vector allocation is likely be aligned for greater than a byte.
     // So we allocate more values than necessary and skip to the next byte with the
@@ -142,8 +133,8 @@ class TestUnpack : public ::testing::TestWithParam<TestUnpackSize> {
   }
 
   template <typename Int>
-  void TestUnpackZeros(UnpackFunc<Int> unpack) {
-    auto [num_values, bit_width] = GetParam();
+  void TestUnpackZeros(UnpackFunc<Int> unpack, int bit_width) {
+    int num_values = GetParam();
     EXPECT_OK_AND_ASSIGN(const auto num_bytes, GetNumBytes(num_values, bit_width));
 
     const std::vector<uint8_t> packed(static_cast<std::size_t>(num_bytes), uint8_t{0});
@@ -154,8 +145,8 @@ class TestUnpack : public ::testing::TestWithParam<TestUnpackSize> {
   }
 
   template <typename Int>
-  void TestUnpackOnes(UnpackFunc<Int> unpack) {
-    auto [num_values, bit_width] = GetParam();
+  void TestUnpackOnes(UnpackFunc<Int> unpack, int bit_width) {
+    int num_values = GetParam();
     EXPECT_OK_AND_ASSIGN(const auto num_bytes, GetNumBytes(num_values, bit_width));
 
     const std::vector<uint8_t> packed(static_cast<std::size_t>(num_bytes), uint8_t{0xFF});
@@ -171,8 +162,8 @@ class TestUnpack : public ::testing::TestWithParam<TestUnpackSize> {
   }
 
   template <typename Int>
-  void TestUnpackAlternating(UnpackFunc<Int> unpack) {
-    const auto [num_values, bit_width] = GetParam();
+  void TestUnpackAlternating(UnpackFunc<Int> unpack, int bit_width) {
+    int num_values = GetParam();
     EXPECT_OK_AND_ASSIGN(const auto num_bytes, GetNumBytes(num_values, bit_width));
 
     const std::vector<uint8_t> packed(static_cast<std::size_t>(num_bytes), uint8_t{0xAA});
@@ -201,35 +192,60 @@ class TestUnpack : public ::testing::TestWithParam<TestUnpackSize> {
 
   template <typename Int>
   void TestAll(UnpackFunc<Int> unpack) {
-    // Known values
-    TestUnpackZeros(unpack);
-    TestUnpackOnes(unpack);
-    TestUnpackAlternating(unpack);
-
-    // Roundtrips
-    TestRoundtripAlignment(unpack, /* alignment_offset= */ 0);
-    TestRoundtripAlignment(unpack, /* alignment_offset= */ 1);
+    constexpr int kMaxBitWidth = 8 * sizeof(Int);
+    // Given how many edge cases there are in unpacking integers, it is best to test all
+    // sizes
+    for (int bit_width = 0; bit_width <= kMaxBitWidth; ++bit_width) {
+      SCOPED_TRACE(::testing::Message() << "Testing bit_width=" << bit_width);
+
+      // Known values
+      TestUnpackZeros(unpack, bit_width);
+      TestUnpackOnes(unpack, bit_width);
+      TestUnpackAlternating(unpack, bit_width);
+
+      // Roundtrips
+      TestRoundtripAlignment(unpack, bit_width, /* alignment_offset= */ 0);
+      TestRoundtripAlignment(unpack, bit_width, /* alignment_offset= */ 1);
+    }
   }
 };
 
-INSTANTIATE_TEST_SUITE_P(
-    UnpackMultiplesOf64Values, TestUnpack,
-    ::testing::Values(TestUnpackSize{64, 1}, TestUnpackSize{128, 1},
-                      TestUnpackSize{2048, 1}, TestUnpackSize{64, 31},
-                      TestUnpackSize{128, 31}, TestUnpackSize{2048, 1},
-                      TestUnpackSize{2048, 8}, TestUnpackSize{2048, 13},
-                      TestUnpackSize{2048, 16}, TestUnpackSize{2048, 31},
-                      TestUnpackSize{2048, 32}));
-
-TEST_P(TestUnpack, Unpack32Scalar) { this->TestAll(&unpack32_scalar); }
-TEST_P(TestUnpack, Unpack64Scalar) { this->TestAll(&unpack64_scalar); }
+// There are actually many differences across the different sizes.
+// It is best to test them all.
+INSTANTIATE_TEST_SUITE_P(UnpackMultiplesOf64Values, TestUnpack,
+                         ::testing::Values(64, 128, 2048),
+                         [](const ::testing::TestParamInfo<TestUnpack::ParamType>& info) {
+                           return "Count" + std::to_string(info.param);
+                         });
+
+TEST_P(TestUnpack, Unpack16Scalar) { this->TestAll(&unpack_scalar<uint16_t>); }
+TEST_P(TestUnpack, Unpack32Scalar) { this->TestAll(&unpack_scalar<uint32_t>); }
+TEST_P(TestUnpack, Unpack64Scalar) { this->TestAll(&unpack_scalar<uint64_t>); }
+
+#if defined(ARROW_HAVE_SSE4_2)
+TEST_P(TestUnpack, Unpack16Sse42) { this->TestAll(&unpack_sse4_2<uint16_t>); }
+TEST_P(TestUnpack, Unpack32Sse42) { this->TestAll(&unpack_sse4_2<uint32_t>); }
+TEST_P(TestUnpack, Unpack64Sse42) { this->TestAll(&unpack_sse4_2<uint64_t>); }
+#endif
 
 #if defined(ARROW_HAVE_RUNTIME_AVX2)
+TEST_P(TestUnpack, Unpack16Avx2) {
+  if (!CpuInfo::GetInstance()->IsSupported(CpuInfo::AVX2)) {
+    GTEST_SKIP() << "Test requires AVX2";
+  }
+  this->TestAll(&unpack_avx2<uint16_t>);
+}
 TEST_P(TestUnpack, Unpack32Avx2) {
   if (!CpuInfo::GetInstance()->IsSupported(CpuInfo::AVX2)) {
     GTEST_SKIP() << "Test requires AVX2";
   }
-  this->TestAll(&unpack32_avx2);
+  this->TestAll(&unpack_avx2<uint32_t>);
+}
+TEST_P(TestUnpack, Unpack64Avx2) {
+  if (!CpuInfo::GetInstance()->IsSupported(CpuInfo::AVX2)) {
+    GTEST_SKIP() << "Test requires AVX2";
+  }
+  this->TestAll(&unpack_avx2<uint64_t>);
 }
 #endif
 
@@ -238,15 +254,24 @@ TEST_P(TestUnpack, Unpack32Avx512) {
   if (!CpuInfo::GetInstance()->IsSupported(CpuInfo::AVX512)) {
     GTEST_SKIP() << "Test requires AVX512";
   }
-  this->TestAll(&unpack32_avx512);
+  this->TestAll(&unpack_avx512<uint32_t>);
+}
+TEST_P(TestUnpack, Unpack64Avx512) {
+  if (!CpuInfo::GetInstance()->IsSupported(CpuInfo::AVX512)) {
+    GTEST_SKIP() << "Test requires AVX512";
+  }
+  this->TestAll(&unpack_avx512<uint64_t>);
 }
 #endif
 
 #if defined(ARROW_HAVE_NEON)
-TEST_P(TestUnpack, Unpack32Neon) { this->TestAll(&unpack32_neon); }
+TEST_P(TestUnpack, Unpack16Neon) { this->TestAll(&unpack_neon<uint16_t>); }
+TEST_P(TestUnpack, Unpack32Neon) { this->TestAll(&unpack_neon<uint32_t>); }
+TEST_P(TestUnpack, Unpack64Neon) { this->TestAll(&unpack_neon<uint64_t>); }
 #endif
 
-TEST_P(TestUnpack, Unpack32) { this->TestAll(&unpack32); }
-TEST_P(TestUnpack, Unpack64) { this->TestAll(&unpack64); }
+TEST_P(TestUnpack, Unpack16) { this->TestAll(&unpack<uint16_t>); }
+TEST_P(TestUnpack, Unpack32) { this->TestAll(&unpack<uint32_t>); }
+TEST_P(TestUnpack, Unpack64) { this->TestAll(&unpack<uint64_t>); }
 
 }  // namespace arrow::internal
diff --git a/cpp/src/arrow/util/rle_encoding_internal.h b/cpp/src/arrow/util/rle_encoding_internal.h
index c231c9a63ebb1..eeb3291b426ac 100644
--- a/cpp/src/arrow/util/rle_encoding_internal.h
+++ b/cpp/src/arrow/util/rle_encoding_internal.h
@@ -316,6 +316,10 @@ class RleRunDecoder {
 };
 
 /// Decoder class for single run of bit-packed encoded data.
+///
+/// The value_bit_width parameter is not part of this class, as it turned oput to be more
+/// performant but it must remain constant for as long as this decoder is tied to a run
+/// (in between calls to ``Reset``).
 template <typename T>
 class BitPackedRunDecoder {
  public:
@@ -323,6 +327,12 @@ class BitPackedRunDecoder {
   using value_type = T;
   /// The type of run that can be decoded.
   using RunType = BitPackedRun;
+  /// The function that can extract packed integer for the ``BitReader``.
+  using UnpackFn = ::arrow::bit_util::BitReader::UnpackFn<value_type>;
+
+  static UnpackFn get_unpack_fn(int value_bit_width) {
+    return bit_util::BitReader::get_unpack_fn<value_type>(value_bit_width);
+  }
 
   BitPackedRunDecoder() noexcept = default;
 
@@ -353,21 +363,22 @@ class BitPackedRunDecoder {
   }
 
   /// Get the next value and return false if there are no more or an error occurred.
-  [[nodiscard]] constexpr bool Get(value_type* out_value, rle_size_t value_bit_width) {
-    return GetBatch(out_value, 1, value_bit_width) == 1;
+  [[nodiscard]] constexpr bool Get(value_type* out_value, rle_size_t value_bit_width,
+                                   UnpackFn unpack) {
+    return GetBatch(out_value, 1, value_bit_width, unpack) == 1;
   }
 
   /// Get a batch of values return the number of decoded elements.
   /// May write fewer elements to the output than requested if there are not enough values
   /// left or if an error occurred.
   [[nodiscard]] rle_size_t GetBatch(value_type* out, rle_size_t batch_size,
-                                    rle_size_t value_bit_width) {
+                                    rle_size_t value_bit_width, UnpackFn unpack) {
     if (ARROW_PREDICT_FALSE(remaining_count_ == 0)) {
       return 0;
     }
 
     const auto to_read = std::min(remaining_count_, batch_size);
-    const auto actual_read = bit_reader_.GetBatch(value_bit_width, out, to_read);
+    const auto actual_read = bit_reader_.GetBatch(value_bit_width, out, to_read, unpack);
     // There should not be any reason why the actual read would be different
     // but this is error resistant.
     remaining_count_ -= actual_read;
@@ -408,6 +419,7 @@ class RleBitPackedDecoder {
     parser_.Reset(data, data_size, value_bit_width);
     decoder_ = {};
     value_bit_width_ = value_bit_width;
+    unpack_ = BitPackedRunDecoder<value_type>::get_unpack_fn(value_bit_width);
   }
 
   /// Whether there is still runs to iterate over.
@@ -463,6 +475,17 @@ class RleBitPackedDecoder {
   RleBitPackedParser parser_ = {};
   std::variant<RleRunDecoder<value_type>, BitPackedRunDecoder<value_type>> decoder_ = {};
   rle_size_t value_bit_width_;
+  typename BitPackedRunDecoder<value_type>::UnpackFn unpack_;
+
+  template <typename Decoder>
+  rle_size_t DecoderGetBatch(Decoder* decoder, value_type* out, rle_size_t batch_size) {
+    if constexpr (std::is_same_v<std::decay_t<Decoder>,
+                                 BitPackedRunDecoder<value_type>>) {
+      return decoder->GetBatch(out, batch_size, value_bit_width_, unpack_);
+    } else {
+      return decoder->GetBatch(out, batch_size, value_bit_width_);
+    }
+  }
 
   /// Return the number of values that are remaining in the current run.
   rle_size_t run_remaining() const {
@@ -471,9 +494,8 @@ class RleBitPackedDecoder {
 
   /// Get a batch of values from the current run and return the number elements read.
   [[nodiscard]] rle_size_t RunGetBatch(value_type* out, rle_size_t batch_size) {
-    return std::visit(
-        [&](auto& dec) { return dec.GetBatch(out, batch_size, value_bit_width_); },
-        decoder_);
+    return std::visit([&](auto& dec) { return DecoderGetBatch(&dec, out, batch_size); },
+                      decoder_);
   }
 
   /// Call the parser with a single callable for all event types.
@@ -746,7 +768,7 @@ auto RleBitPackedDecoder<T>::GetBatch(value_type* out, rle_size_t batch_size)
 
     ARROW_DCHECK_LT(values_read, batch_size);
     RunDecoder decoder(run, value_bit_width_);
-    const auto read = decoder.GetBatch(out, batch_size - values_read, value_bit_width_);
+    const auto read = DecoderGetBatch(&decoder, out, batch_size - values_read);
     ARROW_DCHECK_LE(read, batch_size - values_read);
     values_read += read;
     out += read;
@@ -833,7 +855,7 @@ struct GetSpacedResult {
 
 /// Overload for GetSpaced for a single run in a RleDecoder
 template <typename Converter, typename BitRunReader, typename BitRun, typename value_type>
-auto RunGetSpaced(Converter* converter, typename Converter::out_type* out,
+auto RleGetSpaced(Converter* converter, typename Converter::out_type* out,
                   rle_size_t batch_size, rle_size_t null_count,
                   rle_size_t value_bit_width, BitRunReader* validity_reader,
                   BitRun* validity_run, RleRunDecoder<value_type>* decoder)
@@ -894,10 +916,12 @@ auto RunGetSpaced(Converter* converter, typename Converter::out_type* out,
 }
 
 template <typename Converter, typename BitRunReader, typename BitRun, typename value_type>
-auto RunGetSpaced(Converter* converter, typename Converter::out_type* out,
-                  rle_size_t batch_size, rle_size_t null_count,
-                  rle_size_t value_bit_width, BitRunReader* validity_reader,
-                  BitRun* validity_run, BitPackedRunDecoder<value_type>* decoder)
+auto BitPackedGetSpaced(Converter* converter, typename Converter::out_type* out,
+                        rle_size_t batch_size, rle_size_t null_count,
+                        rle_size_t value_bit_width,
+                        typename BitPackedRunDecoder<value_type>::UnpackFn unpack,
+                        BitRunReader* validity_reader, BitRun* validity_run,
+                        BitPackedRunDecoder<value_type>* decoder)
     -> GetSpacedResult<rle_size_t> {
   ARROW_DCHECK_GT(batch_size, 0);
   // The equality case is handled in the main loop in GetSpaced
@@ -932,7 +956,7 @@ auto RunGetSpaced(Converter* converter, typename Converter::out_type* out,
     // buffer_start is 0 at this point so size is end
     buffer_end = std::min(std::min(run_values_remaining(), batch.values_remaining()),
                           kBufferCapacity);
-    buffer_end = decoder->GetBatch(buffer.data(), buffer_size(), value_bit_width);
+    buffer_end = decoder->GetBatch(buffer.data(), buffer_size(), value_bit_width, unpack);
     ARROW_DCHECK_LE(buffer_size(), kBufferCapacity);
 
     if (ARROW_PREDICT_FALSE(!converter->InputIsValid(buffer.data(), buffer_size()))) {
@@ -980,19 +1004,47 @@ auto RunGetSpaced(Converter* converter, typename Converter::out_type* out,
   return {/* .values_read= */ batch.values_read(), /* .null_read= */ batch.null_read()};
 }
 
+/// Overload for GetSpaced that dispatch the Decoder type
+template <typename Decoder, typename Converter, typename BitRunReader, typename BitRun,
+          typename UnpackFunc>
+auto RunGetSpaced(Converter* converter, typename Converter::out_type* out,
+                  rle_size_t batch_size, rle_size_t null_count,
+                  rle_size_t value_bit_width, UnpackFunc unpack,
+                  BitRunReader* validity_reader, BitRun* validity_run, Decoder* decoder)
+    -> GetSpacedResult<rle_size_t> {
+  using value_type = typename Decoder::value_type;
+
+  if constexpr (std::is_same_v<Decoder, BitPackedRunDecoder<value_type>>) {
+    return BitPackedGetSpaced(converter, out, batch_size, null_count, value_bit_width,
+                              unpack, validity_reader, validity_run, decoder);
+  } else {
+    return RleGetSpaced(converter, out, batch_size, null_count, value_bit_width,
+                        validity_reader, validity_run, decoder);
+  }
+}
+
 /// Overload for GetSpaced for a single run in a decoder variant
 template <typename Converter, typename BitRunReader, typename BitRun, typename value_type>
 auto RunGetSpaced(
     Converter* converter, typename Converter::out_type* out, rle_size_t batch_size,
-    rle_size_t null_count, rle_size_t value_bit_width, BitRunReader* validity_reader,
-    BitRun* validity_run,
+    rle_size_t null_count, rle_size_t value_bit_width,
+    typename BitPackedRunDecoder<value_type>::UnpackFn unpack,
+    BitRunReader* validity_reader, BitRun* validity_run,
     std::variant<RleRunDecoder<value_type>, BitPackedRunDecoder<value_type>>* decoder)
     -> GetSpacedResult<rle_size_t> {
   return std::visit(
       [&](auto& dec) {
         ARROW_DCHECK_GT(dec.remaining(), 0);
-        return RunGetSpaced(converter, out, batch_size, null_count, value_bit_width,
-                            validity_reader, validity_run, &dec);
+        if constexpr (std::is_same_v<std::decay_t<decltype(dec)>,
+                                     BitPackedRunDecoder<value_type>>) {
+          return BitPackedGetSpaced(converter, out, batch_size, null_count,
+                                    value_bit_width, unpack, validity_reader,
+                                    validity_run, &dec);
+
+        } else {
+          return RleGetSpaced(converter, out, batch_size, null_count, value_bit_width,
+                              validity_reader, validity_run, &dec);
+        }
       },
       *decoder);
 }
@@ -1035,9 +1087,9 @@ auto RleBitPackedDecoder<T>::GetSpaced(Converter converter,
 
   // Remaining from a previous call that would have left some unread data from a run.
   if (ARROW_PREDICT_FALSE(run_remaining() > 0)) {
-    const auto read = internal::RunGetSpaced(&converter, out, batch.total_remaining(),
-                                             batch.null_remaining(), value_bit_width_,
-                                             &validity_reader, &validity_run, &decoder_);
+    const auto read = internal::RunGetSpaced(
+        &converter, out, batch.total_remaining(), batch.null_remaining(),
+        value_bit_width_, unpack_, &validity_reader, &validity_run, &decoder_);
 
     batch.AccrueReadNulls(read.null_read);
     batch.AccrueReadValues(read.values_read);
@@ -1059,9 +1111,9 @@ auto RleBitPackedDecoder<T>::GetSpaced(Converter converter,
 
     RunDecoder decoder(run, value_bit_width_);
 
-    const auto read = internal::RunGetSpaced(&converter, out, batch.total_remaining(),
-                                             batch.null_remaining(), value_bit_width_,
-                                             &validity_reader, &validity_run, &decoder);
+    const auto read = internal::RunGetSpaced(
+        &converter, out, batch.total_remaining(), batch.null_remaining(),
+        value_bit_width_, unpack_, &validity_reader, &validity_run, &decoder);
 
     batch.AccrueReadNulls(read.null_read);
     batch.AccrueReadValues(read.values_read);
@@ -1218,9 +1270,10 @@ auto RleBitPackedDecoder<T>::GetBatchWithDict(const V* dictionary,
   };
 
   if (ARROW_PREDICT_FALSE(run_remaining() > 0)) {
-    const auto read = internal::RunGetSpaced(&converter, out, batch_size,
-                                             /* null_count= */ 0, value_bit_width_,
-                                             &validity_reader, &validity_run, &decoder_);
+    const auto read =
+        internal::RunGetSpaced(&converter, out, batch_size,
+                               /* null_count= */ 0, value_bit_width_, unpack_,
+                               &validity_reader, &validity_run, &decoder_);
 
     ARROW_DCHECK_EQ(read.null_read, 0);
     values_read += read.values_read;
@@ -1241,9 +1294,10 @@ auto RleBitPackedDecoder<T>::GetBatchWithDict(const V* dictionary,
 
     RunDecoder decoder(run, value_bit_width_);
 
-    const auto read = internal::RunGetSpaced(&converter, out, batch_values_remaining(),
-                                             /* null_count= */ 0, value_bit_width_,
-                                             &validity_reader, &validity_run, &decoder);
+    const auto read =
+        internal::RunGetSpaced(&converter, out, batch_values_remaining(),
+                               /* null_count= */ 0, value_bit_width_, unpack_,
+                               &validity_reader, &validity_run, &decoder);
 
     ARROW_DCHECK_EQ(read.null_read, 0);
     values_read += read.values_read;
diff --git a/cpp/src/arrow/util/rle_encoding_test.cc b/cpp/src/arrow/util/rle_encoding_test.cc
index c7f4878b741de..929b93d5d7cb2 100644
--- a/cpp/src/arrow/util/rle_encoding_test.cc
+++ b/cpp/src/arrow/util/rle_encoding_test.cc
@@ -334,6 +334,8 @@ TEST(Rle, RleDecoder) {
 template <typename T>
 void TestBitPackedDecoder(std::vector<uint8_t> bytes, rle_size_t value_count,
                           rle_size_t bit_width, std::vector<T> expected) {
+  auto unpack = BitPackedRunDecoder<T>::get_unpack_fn(bit_width);
+
   // Pre-requisite for this test
   EXPECT_GT(value_count, 6);
 
@@ -345,7 +347,7 @@ void TestBitPackedDecoder(std::vector<uint8_t> bytes, rle_size_t value_count,
   EXPECT_EQ(decoder.remaining(), value_count);
 
   rle_size_t read = 0;
-  EXPECT_EQ(decoder.Get(vals.data(), bit_width), 1);
+  EXPECT_EQ(decoder.Get(vals.data(), bit_width, unpack), 1);
   EXPECT_EQ(vals.at(0), expected.at(0 + read));
   read += 1;
   EXPECT_EQ(decoder.remaining(), value_count - read);
@@ -355,7 +357,7 @@ void TestBitPackedDecoder(std::vector<uint8_t> bytes, rle_size_t value_count,
   EXPECT_EQ(decoder.remaining(), value_count - read);
 
   vals = {0, 0};
-  EXPECT_EQ(decoder.GetBatch(vals.data(), 2, bit_width), vals.size());
+  EXPECT_EQ(decoder.GetBatch(vals.data(), 2, bit_width, unpack), vals.size());
   EXPECT_EQ(vals.at(0), expected.at(0 + read));
   EXPECT_EQ(vals.at(1), expected.at(1 + read));
   read += static_cast<decltype(read)>(vals.size());
@@ -366,7 +368,7 @@ void TestBitPackedDecoder(std::vector<uint8_t> bytes, rle_size_t value_count,
   EXPECT_EQ(decoder.remaining(), 0);
   EXPECT_EQ(decoder.Advance(1, bit_width), 0);
   vals = {0, 0};
-  EXPECT_EQ(decoder.Get(vals.data(), bit_width), 0);
+  EXPECT_EQ(decoder.Get(vals.data(), bit_width, unpack), 0);
   EXPECT_EQ(vals.at(0), 0);
 
   // Reset the decoder
@@ -374,7 +376,7 @@ void TestBitPackedDecoder(std::vector<uint8_t> bytes, rle_size_t value_count,
   read = 0;
   EXPECT_EQ(decoder.remaining(), value_count);
   vals = {0, 0};
-  EXPECT_EQ(decoder.GetBatch(vals.data(), 2, bit_width), vals.size());
+  EXPECT_EQ(decoder.GetBatch(vals.data(), 2, bit_width, unpack), vals.size());
   EXPECT_EQ(vals.at(0), expected.at(0 + read));
   EXPECT_EQ(vals.at(1), expected.at(1 + read));
 }
@@ -436,13 +438,15 @@ void TestRleBitPackedParser(std::vector<uint8_t> bytes, rle_size_t bit_width,
     }
 
     auto OnBitPackedRun(BitPackedRun run) {
+      auto unpack = BitPackedRunDecoder<T>::get_unpack_fn(bit_width_);
+
       bit_packed_decoder_ptr_->Reset(run, bit_width_);
 
       const auto n_decoded = decoded_ptr_->size();
       const auto n_to_decode = bit_packed_decoder_ptr_->remaining();
       decoded_ptr_->resize(n_decoded + n_to_decode);
       EXPECT_EQ(bit_packed_decoder_ptr_->GetBatch(decoded_ptr_->data() + n_decoded,
-                                                  n_to_decode, bit_width_),
+                                                  n_to_decode, bit_width_, unpack),
                 n_to_decode);
       EXPECT_EQ(bit_packed_decoder_ptr_->remaining(), 0);