Merge pull request #60 from sterrettm2/avx2-32bit

Adds support for AVX2 for 32-bit types for quicksort and quickselect

Author: Raghuveer Devulapalli

_clang-format

@@ -74,7 +74,7 @@ PenaltyExcessCharacter: 1000000
 PenaltyReturnTypeOnItsOwnLine: 60
 PointerAlignment: Right
 ReflowComments:  false
-SortIncludes:    true
+SortIncludes:    false
 SortUsingDeclarations: true
 SpaceAfterCStyleCast: false
 SpaceAfterTemplateKeyword: true

examples/Makefile

@@ -1,6 +1,6 @@
 CXX ?= g++-12
 CFLAGS = -I../src -std=c++17 -O3
-EXE = argsort kvsort qsortfp16 qsort16 qsort32 qsort64
+EXE = qsort32avx2 argsort kvsort qsortfp16 qsort16 qsort32 qsort64
 
 default: all
 all : $(EXE)
@@ -14,6 +14,9 @@ qsort16: avx512-16bit-qsort.cpp
 qsort32: avx512-32bit-qsort.cpp
 	     $(CXX) -o qsort32 -march=skylake-avx512 $(CFLAGS) avx512-32bit-qsort.cpp
 
+qsort32avx2: avx2-32bit-qsort.cpp
+	     $(CXX) -o qsort32avx2 -march=haswell $(CFLAGS) avx2-32bit-qsort.cpp
+
 qsort64: avx512-64bit-qsort.cpp
 	     $(CXX) -o qsort64 -march=skylake-avx512 $(CFLAGS) avx512-64bit-qsort.cpp
 

examples/avx2-32bit-qsort.cpp

@@ -0,0 +1,10 @@
+#include "avx2-32bit-qsort.hpp"
+
+int main() {
+    const int size = 1000;
+    float arr[size];
+    avx2_qsort(arr, size);
+    avx2_qselect(arr, 10, size);
+    avx2_partial_qsort(arr, 10, size);
+    return 0;
+}

lib/meson.build

@@ -1,5 +1,15 @@
 libtargets = []
 
+if cpp.has_argument('-march=haswell')
+  libtargets += static_library('libavx',
+    files(
+      'x86simdsort-avx2.cpp',
+      ),
+    include_directories : [src],
+    cpp_args : ['-march=haswell', flags_hide_symbols],
+    )
+endif
+
 if cpp.has_argument('-march=skylake-avx512')
   libtargets += static_library('libskx',
     files(

lib/x86simdsort-avx2.cpp

@@ -0,0 +1,28 @@
+// AVX2 specific routines:
+#include "avx2-32bit-qsort.hpp"
+#include "x86simdsort-internal.h"
+
+#define DEFINE_ALL_METHODS(type) \
+    template <> \
+    void qsort(type *arr, size_t arrsize) \
+    { \
+        avx2_qsort(arr, arrsize); \
+    } \
+    template <> \
+    void qselect(type *arr, size_t k, size_t arrsize, bool hasnan) \
+    { \
+        avx2_qselect(arr, k, arrsize, hasnan); \
+    } \
+    template <> \
+    void partial_qsort(type *arr, size_t k, size_t arrsize, bool hasnan) \
+    { \
+        avx2_partial_qsort(arr, k, arrsize, hasnan); \
+    }
+
+namespace xss {
+namespace avx2 {
+    DEFINE_ALL_METHODS(uint32_t)
+    DEFINE_ALL_METHODS(int32_t)
+    DEFINE_ALL_METHODS(float)
+} // namespace avx512
+} // namespace xss

lib/x86simdsort.cpp

@@ -7,7 +7,7 @@
 
 static int check_cpu_feature_support(std::string_view cpufeature)
 {
-    const char* disable_avx512 = std::getenv("XSS_DISABLE_AVX512");
+    const char *disable_avx512 = std::getenv("XSS_DISABLE_AVX512");
 
     if ((cpufeature == "avx512_spr") && (!disable_avx512))
 #ifdef __FLT16_MAX__
@@ -100,34 +100,40 @@ dispatch_requested(std::string_view cpurequested,
     }
 
 /* runtime dispatch mechanism */
-#define DISPATCH(func, TYPE, ...) \
+#define DISPATCH(func, TYPE, ISA) \
     DECLARE_INTERNAL_##func(TYPE) static __attribute__((constructor)) void \
     CAT(CAT(resolve_, func), TYPE)(void) \
     { \
         CAT(CAT(internal_, func), TYPE) = &xss::scalar::func<TYPE>; \
         __builtin_cpu_init(); \
-        std::string_view preferred_cpu = find_preferred_cpu({__VA_ARGS__}); \
-        if constexpr (dispatch_requested("avx512", {__VA_ARGS__})) { \
+        std::string_view preferred_cpu = find_preferred_cpu(ISA); \
+        if constexpr (dispatch_requested("avx512", ISA)) { \
             if (preferred_cpu.find("avx512") != std::string_view::npos) { \
                 CAT(CAT(internal_, func), TYPE) = &xss::avx512::func<TYPE>; \
                 return; \
             } \
         } \
-        else if constexpr (dispatch_requested("avx2", {__VA_ARGS__})) { \
+        if constexpr (dispatch_requested("avx2", ISA)) { \
             if (preferred_cpu.find("avx2") != std::string_view::npos) { \
                 CAT(CAT(internal_, func), TYPE) = &xss::avx2::func<TYPE>; \
                 return; \
             } \
         } \
     }
 
+#define ISA_LIST(...) \
+    std::initializer_list<std::string_view> \
+    { \
+        __VA_ARGS__ \
+    }
+
 namespace x86simdsort {
 #ifdef __FLT16_MAX__
-DISPATCH(qsort, _Float16, "avx512_spr")
-DISPATCH(qselect, _Float16, "avx512_spr")
-DISPATCH(partial_qsort, _Float16, "avx512_spr")
-DISPATCH(argsort, _Float16, "none")
-DISPATCH(argselect, _Float16, "none")
+DISPATCH(qsort, _Float16, ISA_LIST("avx512_spr"))
+DISPATCH(qselect, _Float16, ISA_LIST("avx512_spr"))
+DISPATCH(partial_qsort, _Float16, ISA_LIST("avx512_spr"))
+DISPATCH(argsort, _Float16, ISA_LIST("none"))
+DISPATCH(argselect, _Float16, ISA_LIST("none"))
 #endif
 
 #define DISPATCH_ALL(func, ISA_16BIT, ISA_32BIT, ISA_64BIT) \
@@ -140,10 +146,25 @@ DISPATCH(argselect, _Float16, "none")
     DISPATCH(func, uint64_t, ISA_64BIT) \
     DISPATCH(func, double, ISA_64BIT)
 
-DISPATCH_ALL(qsort, ("avx512_icl"), ("avx512_skx"), ("avx512_skx"))
-DISPATCH_ALL(qselect, ("avx512_icl"), ("avx512_skx"), ("avx512_skx"))
-DISPATCH_ALL(partial_qsort, ("avx512_icl"), ("avx512_skx"), ("avx512_skx"))
-DISPATCH_ALL(argsort, "none", "avx512_skx", "avx512_skx")
-DISPATCH_ALL(argselect, "none", "avx512_skx", "avx512_skx")
+DISPATCH_ALL(qsort,
+             (ISA_LIST("avx512_icl")),
+             (ISA_LIST("avx512_skx", "avx2")),
+             (ISA_LIST("avx512_skx")))
+DISPATCH_ALL(qselect,
+             (ISA_LIST("avx512_icl")),
+             (ISA_LIST("avx512_skx", "avx2")),
+             (ISA_LIST("avx512_skx")))
+DISPATCH_ALL(partial_qsort,
+             (ISA_LIST("avx512_icl")),
+             (ISA_LIST("avx512_skx", "avx2")),
+             (ISA_LIST("avx512_skx")))
+DISPATCH_ALL(argsort,
+             (ISA_LIST("none")),
+             (ISA_LIST("avx512_skx")),
+             (ISA_LIST("avx512_skx")))
+DISPATCH_ALL(argselect,
+             (ISA_LIST("none")),
+             (ISA_LIST("avx512_skx")),
+             (ISA_LIST("avx512_skx")))
 
 } // namespace x86simdsort