ROCm · iotamudelta · Aug 2, 2018 · Jul 30, 2018 · Jul 30, 2018 · Jul 30, 2018
diff --git a/.clang-tidy b/.clang-tidy
@@ -2,13 +2,15 @@
 # NOTE: there must be no spaces before the '-', so put the comma first.
 Checks: '
   *
+  ,clang-analyzer-*
   ,modernize-*
   ,-cert-err58-cpp
   ,-cert-err60-cpp
   ,-clang-diagnostic-*
   ,-cppcoreguidelines-owning-memory
   ,-cppcoreguidelines-pro-bounds-array-to-pointer-decay
   ,-cppcoreguidelines-pro-bounds-constant-array-index
+  ,-cppcoreguidelines-pro-type-member-init
   ,-cppcoreguidelines-pro-type-static-cast-downcast
   ,-cppcoreguidelines-pro-type-vararg
   ,-cppcoreguidelines-special-member-functions
@@ -23,9 +25,11 @@ Checks: '
   ,-hicpp-braces-around-statements
   ,-hicpp-explicit-conversions
   ,-hicpp-no-array-decay
+  ,-hicpp-signed-bitwise
   ,-hicpp-special-member-functions
   ,-hicpp-vararg
   ,-llvm-header-guard
+  ,-llvm-include-order
   ,-llvm-namespace-comment
   ,-misc-unused-parameters
   ,-modernize-make-unique
@@ -34,7 +38,6 @@ Checks: '
   ,-readability-braces-around-statements
   ,-readability-else-after-return
   ,-readability-named-parameter
-  ,clang-analyzer-*
   '
 WarningsAsErrors: ''
 HeaderFilterRegex: 'torch/csrc/'

diff --git a/.jenkins/caffe2/test.sh b/.jenkins/caffe2/test.sh
@@ -64,7 +64,13 @@ for test in $(find "${INSTALL_PREFIX}/test" -executable -type f); do
       ;;
     */aten/*)
       # ATen uses test framework Catch2
-      "$test" -r=xml -o "${junit_reports_dir}/$(basename $test).xml"
+      # NB: We do NOT use the xml test reporter, because
+      # Catch doesn't support multiple reporters
+      # c.f. https://github.com/catchorg/Catch2/blob/master/docs/release-notes.md#223
+      # which means that enabling XML output means you lose useful stdout
+      # output for Jenkins.  It's more important to have useful console
+      # output than it is to have XML output for Jenkins.
+      "$test"
       ;;
     *)
       "$test" --gtest_output=xml:"$gtest_reports_dir/$(basename $test).xml"

diff --git a/.jenkins/pytorch/build.sh b/.jenkins/pytorch/build.sh
@@ -43,12 +43,9 @@ if [[ "$BUILD_ENVIRONMENT" == *rocm* ]]; then
   # https://github.com/RadeonOpenCompute/hcc#hcc-with-thinlto-linking
   export KMTHINLTO=1
 
-  sudo chown -R jenkins:jenkins /usr/local
-  rm -rf "$(dirname "${BASH_SOURCE[0]}")/../../../pytorch_amd/" || true
-  python "$(dirname "${BASH_SOURCE[0]}")/../../tools/amd_build/build_pytorch_amd.py"
-
-  USE_ROCM=1 python setup.py install
-  exit
+  python tools/amd_build/build_pytorch_amd.py
+  USE_ROCM=1 python setup.py install --user
+  exit 0
 fi
 
 # TODO: Don't install this here

diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -284,6 +284,8 @@ include_directories(BEFORE ${PROJECT_SOURCE_DIR})
 # in PROJECT_SOURCE_DIR.
 include_directories(BEFORE ${PROJECT_BINARY_DIR})
 
+include_directories(BEFORE ${PROJECT_SOURCE_DIR}/aten/src/)
+
 # ---[ Old caffe protobuf
 if(BUILD_CAFFE2)
   add_subdirectory(caffe/proto)

diff --git a/aten/src/ATen/CMakeLists.txt b/aten/src/ATen/CMakeLists.txt
@@ -44,6 +44,7 @@ CONFIGURE_FILE(cuda/CUDAConfig.h.in "${CMAKE_CURRENT_SOURCE_DIR}/cuda/CUDAConfig
 # NB: If you edit these globs, you'll have to update setup.py package_data as well
 FILE(GLOB base_h "*.h" "detail/*.h")
 FILE(GLOB base_cpp "*.cpp" "detail/*.cpp")
+add_subdirectory(core)
 FILE(GLOB cuda_h "cuda/*.h" "cuda/detail/*.h" "cuda/*.cuh" "cuda/detail/*.cuh")
 FILE(GLOB cuda_cpp "cuda/*.cpp" "cuda/detail/*.cpp")
 FILE(GLOB cuda_cu "cuda/*.cu" "cuda/detail/*.cu")
@@ -62,7 +63,7 @@ FILE(GLOB native_cuda_cpp "native/cuda/*.cpp")
 FILE(GLOB native_mkl_cpp "native/mkl/*.cpp")
 FILE(GLOB native_mkldnn_cpp "native/mkldnn/*.cpp")
 
-set(all_cpu_cpp ${base_cpp} ${native_cpp} ${native_sparse_cpp} ${native_mkl_cpp} ${native_mkldnn_cpp} ${generated_cpp} ${ATen_CPU_SRCS} ${cpu_kernel_cpp})
+set(all_cpu_cpp ${base_cpp} ${ATen_CORE_SRCS} ${native_cpp} ${native_sparse_cpp} ${native_mkl_cpp} ${native_mkldnn_cpp} ${generated_cpp} ${ATen_CPU_SRCS} ${cpu_kernel_cpp})
 if(AT_MKL_ENABLED)
   set(all_cpu_cpp ${all_cpu_cpp} ${mkl_cpp})
 endif()
@@ -393,7 +394,7 @@ INSTALL(FILES "${CMAKE_CURRENT_BINARY_DIR}/cmake-exports/ATenConfig.cmake"
   DESTINATION "${AT_INSTALL_SHARE_DIR}/cmake/ATen")
 
 # https://stackoverflow.com/questions/11096471/how-can-i-install-a-hierarchy-of-files-using-cmake
-FOREACH(HEADER ${base_h} ${cuda_h} ${cudnn_h})
+FOREACH(HEADER ${base_h} ${ATen_CORE_HEADERS} ${cuda_h} ${cudnn_h})
   string(REPLACE "${CMAKE_CURRENT_SOURCE_DIR}/" "" HEADER_SUB ${HEADER})
   GET_FILENAME_COMPONENT(DIR ${HEADER_SUB} DIRECTORY)
   INSTALL(FILES ${HEADER} DESTINATION ${AT_INSTALL_INCLUDE_DIR}/ATen/${DIR})

diff --git a/aten/src/ATen/Context.h b/aten/src/ATen/Context.h
@@ -9,6 +9,9 @@
 #include "ATen/detail/CUDAHooksInterface.h"
 #include "ATen/CUDAStream.h"
 
+// This is temporary
+#include "ATen/core/ATenCoreTest.h"
+
 #include <memory>
 #include <mutex>
 #include <cstdint>

diff --git a/aten/src/ATen/Storage.cpp b/aten/src/ATen/Storage.cpp
@@ -4,20 +4,30 @@
 
 namespace at {
 
+Storage::Storage(at::ScalarType scalar_type, size_t size, Allocator* allocator)
+    : storage_impl_(new StorageImpl(
+          scalar_type,
+          size,
+          allocator,
+          /* resizable */ false)) {}
+
+Storage::Storage(
+    at::ScalarType scalar_type,
+    at::DataPtr data_ptr,
+    size_t size,
+    const std::function<void(void*)>& deleter)
+    : storage_impl_(new StorageImpl(
+          scalar_type,
+          size,
+          std::move(data_ptr),
+          /* allocator */ nullptr,
+          /* resizable */ false)) {}
+
 Storage::~Storage() {
   if (!storage_impl_) {
     return;
   }
-  if (--storage_impl_->refcount == 0) {
-    if (storage_impl_->finalizer) {
-      (*storage_impl_->finalizer)();
-    }
-    storage_impl_->finalizer = nullptr;
-    storage_impl_->data_ptr.clear();
-    if (storage_impl_ && --storage_impl_->weakcount == 0) {
-      delete storage_impl_;
-    }
-  }
+  storage_impl_->release();
 }
 
 } // namespace at
diff --git a/aten/src/ATen/Storage.h b/aten/src/ATen/Storage.h
@@ -8,6 +8,12 @@ struct AT_API Storage {
 public:
   Storage() = delete;
   Storage(StorageImpl* storage_impl) : storage_impl_(storage_impl) {}
+  Storage(at::ScalarType, size_t size, Allocator* allocator);
+  Storage(
+      at::ScalarType,
+      at::DataPtr,
+      size_t size,
+      const std::function<void(void*)>& deleter);
   ~Storage();
   // There are reasonable interpretations of these constructors, but they're to
   // be implemented on demand.

diff --git a/aten/src/ATen/StorageImpl.cpp b/aten/src/ATen/StorageImpl.cpp
@@ -12,8 +12,6 @@ StorageImpl::StorageImpl(
     : scalar_type(scalar_type),
       data_ptr(std::move(data_ptr)),
       size(size),
-      refcount(1),
-      weakcount(1), // from the strong reference
       resizable(resizable),
       allocator(allocator),
       finalizer(nullptr) {}

diff --git a/aten/src/ATen/StorageImpl.h b/aten/src/ATen/StorageImpl.h
@@ -5,6 +5,7 @@
 #include <ATen/Allocator.h>
 #include <ATen/ScalarType.h>
 #include <ATen/ScalarTypeUtils.h>
+#include <ATen/Retainable.h>
 #include <TH/THTypeConversion.hpp>
 #include <atomic>
 
@@ -39,7 +40,7 @@ namespace at {
 
 struct Type;
 
-struct TH_CPP_API StorageImpl {
+struct TH_CPP_API StorageImpl : public Retainable {
 
   StorageImpl() = delete;
   virtual ~StorageImpl() {};
@@ -48,8 +49,6 @@ struct TH_CPP_API StorageImpl {
   at::ScalarType scalar_type;
   at::DataPtr data_ptr;
   ptrdiff_t size;
-  std::atomic<int> refcount;
-  std::atomic<int> weakcount;
   bool resizable;
   at::Allocator* allocator;
   std::unique_ptr<THFinalizer> finalizer;
@@ -76,6 +75,14 @@ struct TH_CPP_API StorageImpl {
     return static_cast<T*>(this->data_ptr.get());
   }
 
+  void release_resources() {
+    if (finalizer) {
+      (*finalizer)();
+    }
+    finalizer = nullptr;
+    data_ptr.clear();
+  }
+
   void operator=(const StorageImpl&) = delete;
 
   virtual size_t elementSize() const {
@@ -94,9 +101,6 @@ struct TH_CPP_API StorageImpl {
   const void* data() const {
     return data_ptr.get();
   };
-  void retain() {
-    ++refcount;
-  }
 
   int getDevice() const {
     return data_ptr.device().index();

diff --git a/aten/src/ATen/THLongStorageView.h b/aten/src/ATen/THLongStorageView.h
@@ -64,7 +64,6 @@ class THLongStorageView {
       storage.size = ref.size();
     }
     storage.scalar_type = at::CTypeToScalarType<th::from_type<int64_t>>::to();
-    storage.refcount = 0;
     storage.set_resizable(false);
   }
 private:

diff --git a/aten/src/ATen/core/ATenCoreTest.cpp b/aten/src/ATen/core/ATenCoreTest.cpp
@@ -0,0 +1,10 @@
+#include <ATen/core/ATenCoreTest.h>
+
+namespace at {
+
+static int CoreTestGlobal = 0;
+int CoreTest() {
+  return CoreTestGlobal++;
+}
+
+}
diff --git a/aten/src/ATen/core/ATenCoreTest.h b/aten/src/ATen/core/ATenCoreTest.h
@@ -0,0 +1,18 @@
+#pragma once
+
+// TODO: Move this to something like ATenCoreGeneral.h
+#ifdef _WIN32
+# if defined(ATen_cpu_EXPORTS) || defined(caffe2_EXPORTS)
+#  define AT_CORE_API __declspec(dllexport)
+# else
+#  define AT_CORE_API __declspec(dllimport)
+# endif
+#else
+# define AT_CORE_API
+#endif
+
+namespace at {
+
+AT_CORE_API int CoreTest();
+
+}
diff --git a/aten/src/ATen/core/CMakeLists.txt b/aten/src/ATen/core/CMakeLists.txt
@@ -0,0 +1,13 @@
+# This file solely exists to let Caffe2 Android build get at the list
+# of core files without having to trundle through all of ATen's CMakeLists.txt
+
+FILE(GLOB ATen_CORE_HEADERS "*.h")
+FILE(GLOB ATen_CORE_SRCS "*.cpp")
+
+# Pass to parent
+set(ATen_CORE_HEADERS ${ATen_CORE_HEADERS} PARENT_SCOPE)
+set(ATen_CORE_SRCS ${ATen_CORE_SRCS} PARENT_SCOPE)
+# This is a little dodgy, because it means ALL ATen headers are made
+# visible.  Fortunately, you should just get a lot of undefined symbol
+# errors if you go outside core
+set(ATen_CORE_INCLUDE ${CMAKE_CURRENT_SOURCE_DIR}/../.. PARENT_SCOPE)
diff --git a/aten/src/ATen/core/README.md b/aten/src/ATen/core/README.md
@@ -0,0 +1,5 @@
+ATen Core
+---------
+
+ATen Core is a minimal subset of ATen which is suitable for deployment
+on mobile.  Binary size of files in this folder is an important constraint.
diff --git a/aten/src/ATen/cudnn/Descriptors.h b/aten/src/ATen/cudnn/Descriptors.h
@@ -319,6 +319,20 @@ struct AT_CUDA_API RNNDescriptor
   }
 };
 
+#if CUDNN_VERSION >= 7000
+
+struct AT_CUDA_API CTCLossDescriptor
+  : public Descriptor<cudnnCTCLossStruct,
+                      &cudnnCreateCTCLossDescriptor,
+                      &cudnnDestroyCTCLossDescriptor>
+{
+  void set(cudnnDataType_t datatype) {
+    AT_CUDNN_CHECK(cudnnSetCTCLossDescriptor(mut_desc(), datatype));
+  }
+};
+
+#endif
+
 union Constant
 {
   float f;

diff --git a/aten/src/ATen/function_wrapper.py b/aten/src/ATen/function_wrapper.py
@@ -290,7 +290,7 @@ def __init__(self, reason):
             'Backend::${DenseBackend}, ScalarType::Long)'),
     'THStorage*':
         CodeTemplate(
-            'checked_cast_storage<${Storage}>('
+            'checked_cast_storage<Storage>('
             '&${arg_name},"${arg_name}",${arg_pos}, '
             'Backend::${Backend}, ScalarType::${ScalarName})'),
     'THGenerator*':

diff --git a/aten/src/ATen/gen.py b/aten/src/ATen/gen.py
@@ -103,10 +103,6 @@ def check_all_files_written(self):
 TEMPLATE_PATH = options.source_path + "/templates"
 GENERATOR_DERIVED = CodeTemplate.from_file(
     TEMPLATE_PATH + "/GeneratorDerived.h")
-STORAGE_DERIVED_CPP = CodeTemplate.from_file(
-    TEMPLATE_PATH + "/StorageDerived.cpp")
-STORAGE_DERIVED_H = CodeTemplate.from_file(TEMPLATE_PATH + "/StorageDerived.h")
-
 TYPE_DERIVED_CPP = CodeTemplate.from_file(TEMPLATE_PATH + "/TypeDerived.cpp")
 SPARSE_TYPE_DERIVED_CPP = CodeTemplate.from_file(TEMPLATE_PATH + "/SparseTypeDerived.cpp")
 TYPE_DERIVED_H = CodeTemplate.from_file(TEMPLATE_PATH + "/TypeDerived.h")
@@ -237,7 +233,6 @@ def generate_storage_type_and_tensor(backend, density, scalar_type, declarations
     env['isFloatingType'] = is_floating_type
     env['isIntegralType'] = not is_floating_type
     if density == 'Dense':
-        env['Storage'] = "{}{}Storage".format(backend, scalar_name)
         env['Tensor'] = "{}{}{}Tensor".format(density_tag, backend, scalar_name)
     env['Type'] = "{}{}{}Type".format(density_tag, backend, scalar_name)
     env['DenseTensor'] = "{}{}Tensor".format(backend, scalar_name)
@@ -246,7 +241,6 @@ def generate_storage_type_and_tensor(backend, density, scalar_type, declarations
     env['storage_tensor_headers'] = []
     if density != 'Sparse':
         env['storage_tensor_headers'] = [
-            '#include "ATen/{}.h"'.format(env['Storage']),
             '#include "ATen/{}.h"'.format(env['Tensor']),
             '#include "ATen/{}ByteTensor.h"'.format(env['Backend']),
             '#include "ATen/{}IntTensor.h"'.format(env['Backend']),
@@ -322,8 +316,6 @@ def generate_storage_type_and_tensor(backend, density, scalar_type, declarations
 
     if density != 'Sparse':
         # there are no storage or tensor types for sparse; it's all uniform
-        fm.write(env['Storage'] + ".cpp", STORAGE_DERIVED_CPP, env)
-        fm.write(env['Storage'] + ".h", STORAGE_DERIVED_H, env)
         env['TensorDenseOrSparse'] = TENSOR_DENSE_CPP.substitute(env)
         fm.write(env['Tensor'] + ".cpp", TENSOR_DERIVED_CPP, env)
         fm.write(env['Tensor'] + ".h", TENSOR_DERIVED_H, env)
@@ -379,7 +371,7 @@ def declare_outputs():
     for backend, density, scalar_types in iterate_types():
         scalar_name = scalar_types[0]
         full_backend = "Sparse" + backend if density == "Sparse" else backend
-        for kind in ["Storage", "Type", "Tensor"]:
+        for kind in ["Type", "Tensor"]:
             if kind != 'Type' and density == "Sparse":
                 # No Storage or Tensor for sparse
                 continue

diff --git a/aten/src/ATen/native/Linear.cpp b/aten/src/ATen/native/Linear.cpp
@@ -1,6 +1,7 @@
 #include "ATen/ATen.h"
 #include "ATen/NativeFunctions.h"
 #include "ATen/WrapDimUtilsMulti.h"
+#include <cctype>
 
 namespace at { namespace native {
 
@@ -136,6 +137,8 @@ Tensor einsum(std::string eqn, TensorList tensors) {
   } else {
     in_eqn = eqn;
   }
+  // remove spaces for einsum compatibility (#9929)
+  in_eqn.erase(std::remove_if(in_eqn.begin(), in_eqn.end(), isspace), in_eqn.end());
 
   // next we parse in_eq (the left hand side) by iterating. It is a string of comma separated terms per index
   int64_t operand = 0;
@@ -212,7 +215,7 @@ Tensor einsum(std::string eqn, TensorList tensors) {
             num_output_dims++;
           }
         }
-      } else {                              // letter (hopefully)
+      } else if (! isspace(c)) {                              // letter (hopefully)
         AT_CHECK((ell_char_count == 0) || (ell_char_count == 3), "'.' must only occur in ellipsis in the right hand side");
         AT_CHECK(('a' <= c) && (c <= 'z'), "only lowercase letters a-z allowed as indices");
         int64_t letter_num = c-'a';