Skip to content

P
Paddle

机器未来 / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 1
Fork 0
P
Paddle
提交 f01c9668 编写于 作者: T Tao Luo
浏览文件
操作

Merge branch 'develop' into mm_dnn

上级 22c71398 51a9fca3
显示空白变更内容
内联 并排
Showing with 1937 addition and 184 deletion
CMakeLists.txt
浏览文件 @ f01c9668
......@@ -126,16 +126,12 @@ if(ANDROID OR IOS)
add_definitions(-DPADDLE_MOBILE_INFERENCE)
endif()
if (APPLE OR WIN32)
if (APPLE)
set(WITH_MKL OFF CACHE STRING
"Disable MKL for building on mac and windows" FORCE)
"Disable MKL for building on mac" FORCE)
endif()
if (WIN32)
set(WITH_DSO OFF CACHE STRING
"Disable DSO when compiling for Windows" FORCE)
set(WITH_MKL OFF CACHE STRING
"Disable MKL when compiling for Windows" FORCE)
set(WITH_DISTRIBUTE OFF CACHE STRING
"Disable DISTRIBUTE when compiling for Windows" FORCE)
set(WITH_C_API OFF CACHE STRING
......
cmake/cudnn.cmake
浏览文件 @ f01c9668
......@@ -44,9 +44,9 @@ if(WIN32)
set(CUDNN_LIB_NAME "cudnn.lib" "cudnn64_7.dll")
endif(WIN32)
if(Apple)
if(APPLE)
set(CUDNN_LIB_NAME "libcudnn.dylib" "libcudnn.so")
endif(Apple)
endif(APPLE)
find_library(CUDNN_LIBRARY NAMES ${CUDNN_LIB_NAME} # libcudnn_static.a
PATHS ${CUDNN_CHECK_LIBRARY_DIRS} ${CUDNN_INCLUDE_DIR} ${__libpath_hist}
......
cmake/external/mkldnn.cmake
浏览文件 @ f01c9668
......@@ -23,15 +23,14 @@ SET(MKLDNN_SOURCES_DIR ${THIRD_PARTY_PATH}/mkldnn)
SET(MKLDNN_INSTALL_DIR ${THIRD_PARTY_PATH}/install/mkldnn)
SET(MKLDNN_INC_DIR "${MKLDNN_INSTALL_DIR}/include" CACHE PATH "mkldnn include directory." FORCE)
IF(WIN32 OR APPLE)
IF(APPLE)
MESSAGE(WARNING
"Windows or Mac is not supported with MKLDNN in Paddle yet."
"Mac is not supported with MKLDNN in Paddle yet."
"Force WITH_MKLDNN=OFF")
SET(WITH_MKLDNN OFF CACHE STRING "Disable MKLDNN in Windows and MacOS" FORCE)
SET(WITH_MKLDNN OFF CACHE STRING "Disable MKLDNN in MacOS" FORCE)
return()
ENDIF()
SET(MKLDNN_LIB "${MKLDNN_INSTALL_DIR}/lib/libmkldnn.so" CACHE FILEPATH "mkldnn library." FORCE)
MESSAGE(STATUS "Set ${MKLDNN_INSTALL_DIR}/lib to runtime path")
SET(CMAKE_INSTALL_RPATH_USE_LINK_PATH TRUE)
SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${MKLDNN_INSTALL_DIR}/lib")
......@@ -44,10 +43,14 @@ IF(${CBLAS_PROVIDER} STREQUAL "MKLML")
ELSE()
MESSAGE(FATAL_ERROR "Should enable MKLML when build MKLDNN")
ENDIF()
SET(MKLDNN_FLAG "-Wno-error=strict-overflow -Wno-error=unused-result -Wno-error=array-bounds")
SET(MKLDNN_FLAG "${MKLDNN_FLAG} -Wno-unused-result -Wno-unused-value")
SET(MKLDNN_CFLAG "${CMAKE_C_FLAGS} ${MKLDNN_FLAG}")
SET(MKLDNN_CXXFLAG "${CMAKE_CXX_FLAGS} ${MKLDNN_FLAG}")
IF(NOT WIN32)
SET(MKLDNN_FLAG "-Wno-error=strict-overflow -Wno-error=unused-result -Wno-error=array-bounds")
SET(MKLDNN_FLAG "${MKLDNN_FLAG} -Wno-unused-result -Wno-unused-value")
SET(MKLDNN_CFLAG "${CMAKE_C_FLAGS} ${MKLDNN_FLAG}")
SET(MKLDNN_CXXFLAG "${CMAKE_CXX_FLAGS} ${MKLDNN_FLAG}")
ENDIF(NOT WIN32)
ExternalProject_Add(
${MKLDNN_PROJECT}
${EXTERNAL_PROJECT_LOG_ARGS}
......@@ -58,8 +61,15 @@ ExternalProject_Add(
UPDATE_COMMAND ""
CMAKE_ARGS -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
CMAKE_ARGS -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
CMAKE_ARGS -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
CMAKE_ARGS -DCMAKE_CXX_FLAGS_RELEASE=${CMAKE_CXX_FLAGS_RELEASE}
CMAKE_ARGS -DCMAKE_CXX_FLAGS_DEBUG=${CMAKE_CXX_FLAGS_DEBUG}
CMAKE_ARGS -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}
CMAKE_ARGS -DCMAKE_C_FLAGS_DEBUG=${CMAKE_C_FLAGS_DEBUG}
CMAKE_ARGS -DCMAKE_C_FLAGS_RELEASE=${CMAKE_C_FLAGS_RELEASE}
CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${MKLDNN_INSTALL_DIR}
CMAKE_ARGS -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}
CMAKE_ARGS -DCMAKE_POSITION_INDEPENDENT_CODE=ON
CMAKE_ARGS -DMKLROOT=${MKLML_ROOT}
CMAKE_ARGS -DCMAKE_C_FLAGS=${MKLDNN_CFLAG}
CMAKE_ARGS -DCMAKE_CXX_FLAGS=${MKLDNN_CXXFLAG}
......@@ -67,6 +77,11 @@ ExternalProject_Add(
CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${MKLDNN_INSTALL_DIR}
-DMKLROOT:PATH=${MKLML_ROOT}
)
if(WIN32)
SET(MKLDNN_LIB "${MKLDNN_INSTALL_DIR}/lib/mkldnn.lib" CACHE FILEPATH "mkldnn library." FORCE)
else(WIN32)
SET(MKLDNN_LIB "${MKLDNN_INSTALL_DIR}/lib/libmkldnn.so" CACHE FILEPATH "mkldnn library." FORCE)
endif(WIN32)
ADD_LIBRARY(shared_mkldnn SHARED IMPORTED GLOBAL)
SET_PROPERTY(TARGET shared_mkldnn PROPERTY IMPORTED_LOCATION ${MKLDNN_LIB})
......@@ -85,10 +100,14 @@ ADD_DEPENDENCIES(mkldnn ${MKLDNN_PROJECT})
# copy the real so.0 lib to install dir
# it can be directly contained in wheel or capi
SET(MKLDNN_SHARED_LIB ${MKLDNN_INSTALL_DIR}/libmkldnn.so.0)
ADD_CUSTOM_COMMAND(OUTPUT ${MKLDNN_SHARED_LIB}
COMMAND cp ${MKLDNN_LIB} ${MKLDNN_SHARED_LIB}
if(WIN32)
SET(MKLDNN_SHARED_LIB ${MKLDNN_INSTALL_DIR}/lib/mkldnn.dll)
else(WIN32)
SET(MKLDNN_SHARED_LIB ${MKLDNN_INSTALL_DIR}/libmkldnn.so.0)
ADD_CUSTOM_COMMAND(OUTPUT ${MKLDNN_SHARED_LIB}
COMMAND ${CMAKE_COMMAND} -E copy ${MKLDNN_LIB} ${MKLDNN_SHARED_LIB}
DEPENDS mkldnn)
endif(WIN32)
ADD_CUSTOM_TARGET(mkldnn_shared_lib ALL DEPENDS ${MKLDNN_SHARED_LIB})
IF(WITH_C_API)
......
cmake/external/mklml.cmake
浏览文件 @ f01c9668
......@@ -16,56 +16,67 @@ IF(NOT ${WITH_MKLML})
return()
ENDIF(NOT ${WITH_MKLML})
IF(WIN32 OR APPLE)
IF(APPLE)
MESSAGE(WARNING
"Windows or Mac is not supported with MKLML in Paddle yet."
"Mac is not supported with MKLML in Paddle yet."
"Force WITH_MKLML=OFF")
SET(WITH_MKLML OFF CACHE STRING "Disable MKLML package in Windows and MacOS" FORCE)
return()
ENDIF()
INCLUDE(ExternalProject)
SET(MKLML_PROJECT "extern_mklml")
IF((NOT DEFINED MKLML_VER) OR (NOT DEFINED MKLML_URL))
MESSAGE(STATUS "use pre defined download url")
SET(MKLML_VER "mklml_lnx_2019.0.20180710" CACHE STRING "" FORCE)
SET(MKLML_URL "http://paddlepaddledeps.cdn.bcebos.com/${MKLML_VER}.tgz" CACHE STRING "" FORCE)
ENDIF()
MESSAGE(STATUS "MKLML_VER: ${MKLML_VER}, MKLML_URL: ${MKLML_URL}")
SET(MKLML_SOURCE_DIR "${THIRD_PARTY_PATH}/mklml")
SET(MKLML_DOWNLOAD_DIR "${MKLML_SOURCE_DIR}/src/${MKLML_PROJECT}")
SET(MKLML_DST_DIR "mklml")
SET(MKLML_INSTALL_ROOT "${THIRD_PARTY_PATH}/install")
SET(MKLML_INSTALL_DIR ${MKLML_INSTALL_ROOT}/${MKLML_DST_DIR})
SET(MKLML_ROOT ${MKLML_INSTALL_DIR})
SET(MKLML_INC_DIR ${MKLML_ROOT}/include)
SET(MKLML_LIB_DIR ${MKLML_ROOT}/lib)
SET(MKLML_LIB ${MKLML_LIB_DIR}/libmklml_intel.so)
SET(MKLML_IOMP_LIB ${MKLML_LIB_DIR}/libiomp5.so)
if(WIN32)
SET(MKLML_LIB ${MKLML_LIB_DIR}/mklml.lib)
SET(MKLML_IOMP_LIB ${MKLML_LIB_DIR}/libiomp5md.lib)
SET(MKLML_SHARED_LIB ${MKLML_LIB_DIR}/mklml.dll)
SET(MKLML_SHARED_IOMP_LIB ${MKLML_LIB_DIR}/libiomp5md.dll)
else()
SET(MKLML_LIB ${MKLML_LIB_DIR}/libmklml_intel.so)
SET(MKLML_IOMP_LIB ${MKLML_LIB_DIR}/libiomp5.so)
SET(MKLML_SHARED_LIB ${MKLML_LIB_DIR}/libmklml_intel.so)
SET(MKLML_SHARED_IOMP_LIB ${MKLML_LIB_DIR}/libiomp5.so)
endif()
SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${MKLML_ROOT}/lib")
INCLUDE_DIRECTORIES(${MKLML_INC_DIR})
IF((NOT DEFINED MKLML_VER) OR (NOT DEFINED MKLML_URL))
MESSAGE(STATUS "use pre defined download url")
if(WIN32)
SET(MKLML_VER "mklml_win_2019.0.20180710" CACHE STRING "" FORCE)
SET(MKLML_URL "https://paddlepaddledeps.cdn.bcebos.com/${MKLML_VER}.zip" CACHE STRING "" FORCE)
else()
SET(MKLML_VER "mklml_lnx_2019.0.20180710" CACHE STRING "" FORCE)
SET(MKLML_URL "http://paddlepaddledeps.cdn.bcebos.com/${MKLML_VER}.tgz" CACHE STRING "" FORCE)
ENDIF()
endif()
FILE(WRITE ${MKLML_DOWNLOAD_DIR}/CMakeLists.txt
"PROJECT(MKLML)\n"
"cmake_minimum_required(VERSION 3.0)\n"
"install(DIRECTORY ${MKLML_VER}/include ${MKLML_VER}/lib \n"
" DESTINATION ${MKLML_DST_DIR})\n")
SET(MKLML_PROJECT "extern_mklml")
MESSAGE(STATUS "MKLML_VER: ${MKLML_VER}, MKLML_URL: ${MKLML_URL}")
SET(MKLML_SOURCE_DIR "${THIRD_PARTY_PATH}/mklml")
SET(MKLML_DOWNLOAD_DIR "${MKLML_SOURCE_DIR}/src/${MKLML_PROJECT}")
ExternalProject_Add(
${MKLML_PROJECT}
${EXTERNAL_PROJECT_LOG_ARGS}
PREFIX ${MKLML_SOURCE_DIR}
URL ${MKLML_URL}
DOWNLOAD_DIR ${MKLML_DOWNLOAD_DIR}
DOWNLOAD_COMMAND wget --no-check-certificate ${MKLML_URL} -c -q -O ${MKLML_VER}.tgz
&& tar zxf ${MKLML_VER}.tgz
DOWNLOAD_NO_PROGRESS 1
CONFIGURE_COMMAND ""
BUILD_COMMAND ""
UPDATE_COMMAND ""
CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${MKLML_INSTALL_ROOT}
CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${MKLML_INSTALL_ROOT}
INSTALL_COMMAND
${CMAKE_COMMAND} -E copy_directory ${MKLML_DOWNLOAD_DIR}/include ${MKLML_INC_DIR} &&
${CMAKE_COMMAND} -E copy_directory ${MKLML_DOWNLOAD_DIR}/lib ${MKLML_LIB_DIR}
)
INCLUDE_DIRECTORIES(${MKLML_INC_DIR})
ADD_LIBRARY(mklml SHARED IMPORTED GLOBAL)
SET_PROPERTY(TARGET mklml PROPERTY IMPORTED_LOCATION ${MKLML_LIB})
ADD_DEPENDENCIES(mklml ${MKLML_PROJECT})
......
cmake/generic.cmake
浏览文件 @ f01c9668
......@@ -267,7 +267,11 @@ function(cc_library TARGET_NAME)
list(APPEND cc_library_DEPS dynload_mklml)
endif()
add_dependencies(${TARGET_NAME} mklml)
if(WIN32)
target_link_libraries(${TARGET_NAME} ${MKLML_IOMP_LIB})
else(WIN32)
target_link_libraries(${TARGET_NAME} "-L${MKLML_LIB_DIR} -liomp5 -Wl,--as-needed")
endif(WIN32)
endif()
# remove link to python, see notes at:
# https://github.com/pybind/pybind11/blob/master/docs/compiling.rst#building-manually
......
cmake/inference_lib.cmake
浏览文件 @ f01c9668
......@@ -115,20 +115,20 @@ if (NOT PROTOBUF_FOUND OR WIN32)
)
endif ()
if (NOT CBLAS_FOUND)
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/openblas")
copy(openblas_lib
SRCS ${CBLAS_INSTALL_DIR}/lib ${CBLAS_INSTALL_DIR}/include
DSTS ${dst_dir} ${dst_dir}
DEPS extern_openblas
)
elseif (WITH_MKLML)
if (WITH_MKLML)
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/mklml")
copy(mklml_lib
SRCS ${MKLML_LIB} ${MKLML_IOMP_LIB} ${MKLML_INC_DIR}
DSTS ${dst_dir}/lib ${dst_dir}/lib ${dst_dir}
DEPS mklml
)
elseif (NOT CBLAS_FOUND OR WIN32)
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/openblas")
copy(openblas_lib
SRCS ${CBLAS_INSTALL_DIR}/lib ${CBLAS_INSTALL_DIR}/include
DSTS ${dst_dir} ${dst_dir}
DEPS extern_openblas
)
endif ()
if (WITH_MKLDNN)
......
paddle/fluid/API.spec
浏览文件 @ f01c9668
......@@ -351,6 +351,23 @@ paddle.fluid.contrib.QuantizeTranspiler.__init__ ArgSpec(args=['self', 'weight_b
paddle.fluid.contrib.QuantizeTranspiler.convert_to_int8 ArgSpec(args=['self', 'program', 'place', 'scope'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.contrib.QuantizeTranspiler.freeze_program ArgSpec(args=['self', 'program', 'place', 'fuse_bn', 'scope'], varargs=None, keywords=None, defaults=(False, None))
paddle.fluid.contrib.QuantizeTranspiler.training_transpile ArgSpec(args=['self', 'program', 'startup_program'], varargs=None, keywords=None, defaults=(None, None))
paddle.fluid.contrib.build_compressor ArgSpec(args=['place', 'data_reader', 'data_feeder', 'scope', 'metrics', 'epoch', 'config'], varargs=None, keywords=None, defaults=(None, None, None, None, None, None, None))
paddle.fluid.contrib.CompressPass.__init__ ArgSpec(args=['self', 'place', 'data_reader', 'data_feeder', 'scope', 'metrics', 'epoch', 'program_exe'], varargs=None, keywords=None, defaults=(None, None, None, None, None, None, None))
paddle.fluid.contrib.CompressPass.add_strategy ArgSpec(args=['self', 'strategy'], varargs=None, keywords=None, defaults=None)
paddle.fluid.contrib.CompressPass.apply ArgSpec(args=['self', 'graph'], varargs=None, keywords=None, defaults=None)
paddle.fluid.contrib.ImitationGraph.__init__ ArgSpec(args=['self', 'program'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.contrib.ImitationGraph.all_parameters ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.contrib.SensitivePruneStrategy.__init__ ArgSpec(args=['self', 'pruner', 'start_epoch', 'end_epoch', 'delta_rate', 'acc_loss_threshold', 'sensitivities'], varargs=None, keywords=None, defaults=(None, 0, 10, 0.2, 0.2, None))
paddle.fluid.contrib.SensitivePruneStrategy.on_batch_begin ArgSpec(args=['self', 'context'], varargs=None, keywords=None, defaults=None)
paddle.fluid.contrib.SensitivePruneStrategy.on_batch_end ArgSpec(args=['self', 'context'], varargs=None, keywords=None, defaults=None)
paddle.fluid.contrib.SensitivePruneStrategy.on_compress_begin ArgSpec(args=['self', 'context'], varargs=None, keywords=None, defaults=None)
paddle.fluid.contrib.SensitivePruneStrategy.on_compress_end ArgSpec(args=['self', 'context'], varargs=None, keywords=None, defaults=None)
paddle.fluid.contrib.SensitivePruneStrategy.on_epoch_begin ArgSpec(args=['self', 'context'], varargs=None, keywords=None, defaults=None)
paddle.fluid.contrib.SensitivePruneStrategy.on_epoch_end ArgSpec(args=['self', 'context'], varargs=None, keywords=None, defaults=None)
paddle.fluid.contrib.MagnitudePruner.__init__ ArgSpec(args=['self', 'threshold'], varargs=None, keywords=None, defaults=None)
paddle.fluid.contrib.MagnitudePruner.prune ArgSpec(args=['self', 'param', 'threshold'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.contrib.RatioPruner.__init__ ArgSpec(args=['self', 'ratios'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.contrib.RatioPruner.prune ArgSpec(args=['self', 'param', 'ratio'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.contrib.load_persistables_for_increment ArgSpec(args=['dirname', 'executor', 'program', 'lookup_table_var', 'lookup_table_var_path'], varargs=None, keywords=None, defaults=None)
paddle.fluid.contrib.load_persistables_for_inference ArgSpec(args=['dirname', 'executor', 'program', 'lookup_table_var_name'], varargs=None, keywords=None, defaults=None)
paddle.fluid.contrib.convert_dist_to_sparse_program ArgSpec(args=['program'], varargs=None, keywords=None, defaults=None)
......
paddle/fluid/framework/lod_tensor.cc
浏览文件 @ f01c9668
......@@ -157,13 +157,8 @@ bool CheckLoD(const LoD &in, int tensor_height) {
if (level.size() < 2) return false;
// check: the first offset(the begin offset) of each level should be 0.
if (level.front() != 0) return false;
// check: all the offsets in a level should be ascending(no same items
// allows).
if (!std::is_sorted(level.begin(), level.begin(), [](size_t a, size_t b) {
if (a < b) return true;
return false;
})) {
LOG(INFO) << "ascending error";
// check: all the offsets in a level should be ascending(allow same items)
if (!std::is_sorted(level.begin(), level.end())) {
return false;
}
}
......
paddle/fluid/framework/lod_tensor_test.cc
浏览文件 @ f01c9668
......@@ -217,6 +217,11 @@ TEST(LoD, CheckLoD) {
// check with underlying tensor storage.
ASSERT_TRUE(CheckLoD(relative_lod, 5));
ASSERT_FALSE(CheckLoD(relative_lod, 9));
// check whether lod is ascending-sorted (allow same items)
ASSERT_TRUE(CheckLoD({{0, 1, 2, 3, 4, 5}}, 5));
ASSERT_TRUE(CheckLoD({{0, 1, 3, 3, 4, 5}}, 5));
ASSERT_FALSE(CheckLoD({{0, 1, 3, 2, 5}}, 5));
}
TEST(LoD, CheckAbsLoD) {
......
paddle/fluid/framework/operator.cc
浏览文件 @ f01c9668
......@@ -476,6 +476,28 @@ const Tensor* ExecutionContext::LegacyInput<Tensor>(
template <>
const std::vector<const Tensor*> ExecutionContext::MultiInput<Tensor>(
const std::string& name) const {
auto it = ctx_.inputs.find(name);
if (it == ctx_.inputs.end()) {
return {};
}
const std::vector<Variable*>& vars = it->second;
std::vector<const Tensor*> res;
res.reserve(vars.size());
std::transform(vars.begin(), vars.end(), std::back_inserter(res),
[&](Variable* var) -> const Tensor* {
if (var == nullptr) return nullptr;
PADDLE_ENFORCE(
var->IsType<LoDTensor>(),
"should be LoDTensor, but the received type is %s",
var->Type().name());
return &(var->Get<LoDTensor>());
});
return res;
}
template <>
const std::vector<const Tensor*> ExecutionContext::LegacyMultiInput<Tensor>(
const std::string& name) const {
auto names = op().Inputs(name);
std::vector<const Tensor*> res;
res.reserve(names.size());
......
paddle/fluid/framework/operator.h
浏览文件 @ f01c9668
......@@ -197,8 +197,31 @@ class ExecutionContext {
const std::vector<const Variable*> MultiInputVar(
const std::string& name) const {
auto names = op_.Inputs(name);
auto it = ctx_.inputs.find(name);
if (it == ctx_.inputs.end()) {
return {};
}
std::vector<const Variable*> res;
res.reserve(it->second.size());
std::transform(it->second.begin(), it->second.end(),
std::back_inserter(res),
[this](Variable* var) { return var; });
return res;
}
std::vector<Variable*> MultiOutputVar(const std::string& name) const {
auto names = op_.Outputs(name);
auto it = ctx_.outputs.find(name);
if (it == ctx_.outputs.end()) {
return {};
}
return it->second;
}
const std::vector<Variable*> LegacyMultiInputVar(
const std::string& name) const {
auto names = op_.Inputs(name);
std::vector<Variable*> res;
res.reserve(names.size());
std::transform(names.begin(), names.end(), std::back_inserter(res),
[this](const std::string& name) {
......@@ -208,7 +231,7 @@ class ExecutionContext {
return res;
}
std::vector<Variable*> MultiOutputVar(const std::string& name) const {
std::vector<Variable*> LegacyMultiOutputVar(const std::string& name) const {
auto names = op_.Outputs(name);
std::vector<Variable*> res;
res.reserve(names.size());
......@@ -250,6 +273,38 @@ class ExecutionContext {
template <typename T>
const std::vector<const T*> MultiInput(const std::string& name) const {
auto it = ctx_.inputs.find(name);
if (it == ctx_.inputs.end()) {
return {};
}
const std::vector<Variable*>& vars = it->second;
std::vector<const T*> res;
res.reserve(vars.size());
std::transform(vars.begin(), vars.end(), std::back_inserter(res),
[&](Variable* var) -> const T* {
return var == nullptr ? nullptr : &var->Get<T>();
});
return res;
}
template <typename T>
std::vector<T*> MultiOutput(const std::string& name) const {
auto it = ctx_.outputs.find(name);
if (it == ctx_.outputs.end()) {
return {};
}
const std::vector<Variable*>& vars = it->second;
std::vector<T*> res;
res.reserve(vars.size());
std::transform(vars.begin(), vars.end(), std::back_inserter(res),
[&](Variable* var) -> T* {
return var == nullptr ? nullptr : var->GetMutable<T>();
});
return res;
}
template <typename T>
const std::vector<const T*> LegacyMultiInput(const std::string& name) const {
auto names = op_.Inputs(name);
std::vector<const T*> res;
res.reserve(names.size());
......@@ -262,7 +317,7 @@ class ExecutionContext {
}
template <typename T>
std::vector<T*> MultiOutput(const std::string& name) const {
std::vector<T*> LegacyMultiOutput(const std::string& name) const {
auto names = op_.Outputs(name);
std::vector<T*> res;
res.reserve(names.size());
......@@ -321,6 +376,10 @@ template <>
const std::vector<const Tensor*> ExecutionContext::MultiInput<Tensor>(
const std::string& name) const;
template <>
const std::vector<const Tensor*> ExecutionContext::LegacyMultiInput<Tensor>(
const std::string& name) const;
template <>
Tensor* ExecutionContext::Output<Tensor>(const std::string& name) const;
......
paddle/fluid/framework/tensor.cc
浏览文件 @ f01c9668
......@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/framework/tensor.h"
#include "paddle/fluid/framework/var_type.h"
namespace paddle {
namespace framework {
......@@ -27,6 +28,9 @@ void Tensor::check_memory_size() const {
"or maybe the required data-type mismatches the data already stored.");
}
Tensor::Tensor(std::type_index type)
: type_(framework::ToDataType(type)), offset_(0) {}
size_t Tensor::memory_size() const {
return holder_ == nullptr ? 0UL : holder_->size() - offset_;
}
......@@ -101,5 +105,12 @@ const DDim& Tensor::dims() const { return dims_; }
int64_t Tensor::numel() const { return product(dims_); }
void Tensor::ResetHolder(std::shared_ptr<memory::Allocation> holder) {
if (holder_) {
PADDLE_ENFORCE_EQ(numel() * SizeOfType(type()), holder->size());
}
holder_ = holder;
}
} // namespace framework
} // namespace paddle
paddle/fluid/framework/tensor.h
浏览文件 @ f01c9668
......@@ -69,6 +69,8 @@ class Tensor {
public:
Tensor() : type_(proto::VarType::FP32), offset_(0) {}
explicit Tensor(std::type_index type);
/*! Return a pointer to mutable memory block. */
template <typename T>
T* data();
......@@ -162,6 +164,8 @@ class Tensor {
return std::move(holder_);
}
void ResetHolder(std::shared_ptr<memory::Allocation> holder);
private:
/*! holds the memory block if allocated. */
std::shared_ptr<memory::Allocation> holder_;
......
paddle/fluid/inference/api/analysis_predictor.cc
浏览文件 @ f01c9668
......@@ -231,11 +231,14 @@ bool AnalysisPredictor::SetFeed(const std::vector<PaddleTensor> &inputs,
inputs[i].data.length());
} else {
#ifdef PADDLE_WITH_CUDA
platform::DeviceContextPool &pool =
platform::DeviceContextPool::Instance();
auto *dev_ctx =
static_cast<const platform::CUDADeviceContext *>(pool.Get(place_));
auto dst_gpu_place = boost::get<platform::CUDAPlace>(place_);
memory::Copy(dst_gpu_place, static_cast<void *>(input_ptr),
platform::CPUPlace(), inputs[i].data.data(),
inputs[i].data.length(),
0); // stream 0 for sync copy
inputs[i].data.length(), dev_ctx->stream());
#else
PADDLE_THROW("Not compile with CUDA, should not reach here.");
#endif
......
paddle/fluid/inference/api/api_impl.cc
浏览文件 @ f01c9668
......@@ -208,11 +208,14 @@ bool NativePaddlePredictor::SetFeed(const std::vector<PaddleTensor> &inputs,
inputs[i].data.length());
} else {
#ifdef PADDLE_WITH_CUDA
platform::DeviceContextPool &pool =
platform::DeviceContextPool::Instance();
auto *dev_ctx =
static_cast<const platform::CUDADeviceContext *>(pool.Get(place_));
auto dst_gpu_place = boost::get<platform::CUDAPlace>(place_);
memory::Copy(dst_gpu_place, static_cast<void *>(input_ptr),
platform::CPUPlace(), inputs[i].data.data(),
inputs[i].data.length(),
0); // stream 0 for sync copy
inputs[i].data.length(), dev_ctx->stream());
#else
PADDLE_THROW("Not compile with CUDA, should not reach here.");
#endif
......
paddle/fluid/inference/api/demo_ci/CMakeLists.txt
浏览文件 @ f01c9668
......@@ -89,12 +89,21 @@ endif()
if(WITH_MKL)
include_directories("${PADDLE_LIB}/third_party/install/mklml/include")
if(NOT WIN32)
set(MATH_LIB ${PADDLE_LIB}/third_party/install/mklml/lib/libmklml_intel${CMAKE_SHARED_LIBRARY_SUFFIX}
${PADDLE_LIB}/third_party/install/mklml/lib/libiomp5${CMAKE_SHARED_LIBRARY_SUFFIX})
else(WIN32)
set(MATH_LIB ${PADDLE_LIB}/third_party/install/mklml/lib/libmklml${CMAKE_SHARED_LIBRARY_SUFFIX}
${PADDLE_LIB}/third_party/install/mklml/lib/libiomp5md${CMAKE_SHARED_LIBRARY_SUFFIX})
endif(WIN32)
set(MKLDNN_PATH "${PADDLE_LIB}/third_party/install/mkldnn")
if(EXISTS ${MKLDNN_PATH})
include_directories("${MKLDNN_PATH}/include")
if(WIN32)
set(MKLDNN_LIB ${MKLDNN_PATH}/lib/mkldnn.lib)
else(WIN32)
set(MKLDNN_LIB ${MKLDNN_PATH}/lib/libmkldnn.so.0)
endif(WIN32)
endif()
else()
set(MATH_LIB ${PADDLE_LIB}/third_party/install/openblas/lib/libopenblas${CMAKE_STATIC_LIBRARY_SUFFIX})
......
paddle/fluid/operators/conv_op.h
浏览文件 @ f01c9668
......@@ -22,6 +22,7 @@ limitations under the License. */
#include "paddle/fluid/operators/math/depthwise_conv.h"
#include "paddle/fluid/operators/math/im2col.h"
#include "paddle/fluid/operators/math/vol2col.h"
#include "paddle/fluid/platform/create_tensor_with_allocationptr.h"
namespace paddle {
namespace operators {
......@@ -123,6 +124,8 @@ class GemmConvKernel : public framework::OpKernel<T> {
std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
std::vector<int> dilations = context.Attr<std::vector<int>>("dilations");
auto& dev_ctx = context.template device_context<DeviceContext>();
const int batch_size = static_cast<int>(input->dims()[0]);
// filter_shape_vec: {k_o, k_i, k_h, k_w} or {k_o, k_i, k_d, k_h, k_w}
......@@ -155,13 +158,19 @@ class GemmConvKernel : public framework::OpKernel<T> {
// to call the matrix multiplication interface.
Tensor col_matrix;
if (is_expand) {
col.mutable_data<T>(col_shape, context.GetPlace());
auto tmp_allocation_ptr =
platform::DeviceTemporaryAllocator::Instance().Get(dev_ctx).Allocate(
framework::product(col_shape) * sizeof(T));
Tensor tep_tensor =
platform::GetTensor<T>(std::move(tmp_allocation_ptr), col_shape);
col.ShareDataWith(tep_tensor);
col_matrix.ShareDataWith(col);
col_matrix.Resize(col_matrix_shape);
}
framework::DDim input_shape = framework::slice_ddim(
input->dims(), 1, static_cast<int>(input->dims().size()));
framework::DDim input_shape =
framework::slice_ddim(input->dims(), 1, input->dims().size());
framework::DDim filter_matrix_shape = {filter.dims()[0],
filter.numel() / filter.dims()[0]};
......@@ -178,7 +187,6 @@ class GemmConvKernel : public framework::OpKernel<T> {
math::Vol2ColFunctor<DeviceContext, T> vol2col;
math::Im2ColFunctor<math::ColFormat::kCFO, DeviceContext, T> im2col;
auto& dev_ctx = context.template device_context<DeviceContext>();
auto blas = math::GetBlas<DeviceContext, T>(dev_ctx);
for (int i = 0; i < batch_size; i++) {
Tensor in_batch = input->Slice(i, i + 1).Resize(input_shape);
......@@ -237,6 +245,8 @@ class GemmConvGradKernel : public framework::OpKernel<T> {
const int batch_size = static_cast<int>(input->dims()[0]);
auto& dev_ctx = context.template device_context<DeviceContext>();
// filter_shape_vec: {k_o, k_i, k_h, k_w} or {k_o, k_i, k_d, k_h, k_w}
std::vector<int64_t> filter_shape_vec(framework::vectorize(filter.dims()));
// output_shape_vec: {o_n, o_c, o_h, o_w} or {o_n, o_c, o_d, o_h, o_w}
......@@ -262,8 +272,8 @@ class GemmConvGradKernel : public framework::OpKernel<T> {
framework::DDim col_matrix_shape =
framework::flatten_to_2d(col_shape, data_dim + 1);
framework::DDim input_shape = framework::slice_ddim(
input->dims(), 1, static_cast<int>(input->dims().size()));
framework::DDim input_shape =
framework::slice_ddim(input->dims(), 1, input->dims().size());
framework::DDim filter_matrix_shape = {filter.dims()[0],
filter.numel() / filter.dims()[0]};
......@@ -286,13 +296,18 @@ class GemmConvGradKernel : public framework::OpKernel<T> {
// to call the matrix multiplication interface.
Tensor col_matrix;
if (is_expand) {
col.mutable_data<T>(col_shape, context.GetPlace());
auto tmp_allocation_ptr =
platform::DeviceTemporaryAllocator::Instance().Get(dev_ctx).Allocate(
framework::product(col_shape) * sizeof(T));
Tensor tep_tensor =
platform::GetTensor<T>(std::move(tmp_allocation_ptr), col_shape);
col.ShareDataWith(tep_tensor);
col_matrix.ShareDataWith(col);
col_matrix.Resize(col_matrix_shape);
}
math::SetConstant<DeviceContext, T> set_zero;
auto& dev_ctx = context.template device_context<DeviceContext>();
auto blas = math::GetBlas<DeviceContext, T>(dev_ctx);
if (input_grad) {
......
paddle/fluid/operators/detection/density_prior_box_op.cu
浏览文件 @ f01c9668
......@@ -142,12 +142,13 @@ class DensityPriorBoxOpCUDAKernel : public framework::OpKernel<T> {
vars->mutable_data<T>(ctx.GetPlace());
framework::Tensor d_temp;
framework::TensorCopySync(h_temp, ctx.GetPlace(), &d_temp);
framework::TensorCopy(h_temp, ctx.GetPlace(), &d_temp);
// At least use 32 threads, at most 512 threads.
// blockx is multiple of 32.
int blockx = std::min(
static_cast<long>(((feature_width * num_priors + 31) >> 5) << 5), 512L);
static_cast<int64_t>(((feature_width * num_priors + 31) >> 5) << 5),
512L);
int gridx = (feature_width * num_priors + blockx - 1) / blockx;
dim3 threads(blockx, 1);
dim3 grids(gridx, feature_height);
......
paddle/fluid/operators/elementwise/elementwise_mul_mkldnn_op.cc
浏览文件 @ f01c9668
......@@ -16,11 +16,14 @@ limitations under the License. */
#include "paddle/fluid/operators/elementwise/elementwise_op.h"
#include "paddle/fluid/operators/elementwise/elementwise_op_function.h"
#include "paddle/fluid/operators/jit/kernels.h"
#include "paddle/fluid/platform/cpu_info.h"
#include "paddle/fluid/platform/mkldnn_helper.h"
#include "paddle/fluid/operators/jit/kernels.h"
#ifdef PADDLE_WITH_XBYAK
#include "xbyak/xbyak.h"
#include "xbyak/xbyak_util.h"
#endif
namespace paddle {
namespace operators {
......@@ -81,8 +84,7 @@ class ElementwiseMulMKLDNNKernel : public framework::OpKernel<T> {
UpdateDataFormat(ctx, const_cast<Tensor*>(x), "x_data_format");
UpdateDataFormat(ctx, const_cast<Tensor*>(y), "y_data_format");
Xbyak::util::Cpu cpu;
const bool is_avx512_enabled = cpu.has(Xbyak::util::Cpu::tAVX512F);
const bool is_avx512_enabled = platform::MayIUse(platform::avx512f);
const bool are_dims_divisable = !(x_int_dims[1] % 16);
const bool is_x_format_correct = x->format() == memory::format::nChw16c;
const bool is_y_format_correct = y->format() == memory::format::nc;
......
paddle/fluid/operators/jit/CMakeLists.txt
浏览文件 @ f01c9668
......@@ -21,5 +21,5 @@ endif()
cc_library(jit_kernel_helper SRCS ${jit_kernel_cc_srcs} DEPS ${JIT_KERNEL_DEPS})
cc_test(jit_kernel_test SRCS test.cc DEPS jit_kernel_helper)
if(NOT WIN32)
cc_binary(jit_kernel_benchmark SRCS benchmark.cc DEPS jit_kernel_helper)
cc_binary(jit_kernel_benchmark SRCS benchmark.cc DEPS jit_kernel_helper device_tracer)
endif()
paddle/fluid/operators/jit/README.en.md 0 → 100644
浏览文件 @ f01c9668
# JIT Kernel
JIT(Just In Time) Kernel contains actually generated code and some other implemenations with the same logic.
Each implementations has its own condition to use, defined in `UseMe`.
They are combined together to get the best performance of one single independent function.
They could be some very simple functions like vector multiply, or some complicated functions like LSTM.
And they can be composed with some other exited jit kernels to build up a complex function.
Currently it's only supported on CPU yet.
## Contents
```txt
PaddlePaddle/Paddle/paddle/fluid/
├── ...
└── operators/
├── .../
└── jit/
├── ...
├── gen/
│ └── ...
|── more/
│ ├── ...
│ ├── mkl/
│ │ └── ...
│ ├── mkldnn/
│ │ └── ...
│ ├── mix/
│ │ └── ...
│ ├── intrinsic/
│ │ └── ...
│ └── openblas/
│ └── ...
└── refer/
└── ...
```
All basical definations of jit kernels are addressed in `paddle/fluid/operators/jit` including these three key folders `refer`, `gen`, `more`. There is only one unique name for each kernel while may have seraval implementations with same functionality.
- `refer`: Each kernel must have one reference implementation on CPU, and it should only focus on the correctness and should not depends on any third-party libraries.
- `gen`: The code generated should be kept here. They should be designed focusing on the best performance, which depends on Xbyak.
- `more`: All other implementations should be kept in this folder with one directory corresponding to one library kind or method kind, such as mkl, mkldnn, openblas or intrinsic code. Each implementation should have it advantage.
## How to use
One simple function `jit::Get`, which is very easy to use, is supported to get the kernel.
It can automatically return the expected function with best performance under the given attributes.
All kernels are inlcuded in `paddle/fluid/operators/jit/kernels.h`, you can only include this one header to get all the registered kernels.
## Solid Test
- Unit Test
All functions should be compared with the corresponding reference functions, including data tyep `float` and `double`.
- Benchmark
All functions should be tested, and make sure the `jit::Get` function obtain the best performance with all attributes.
# How to add new kernel
## Required
1. Add `your_key` at `KernelType`.
2. Add reference function of `your_key`.
Note:
- this should be run on CPU and do not depend on any third-party.
- Add `USE_JITKERNEL_REFER(your_key)` in `refer/CmakeLists.txt` to make sure this code can be used.
3. Add unit test in `test.cc`, and verfiy at least `float` and `double`.
Test more data type for some special functions if necessary, for example `int8`.
4. Add functions in `benchmark.cc` to test all function of same `KernelType`. Make sure `jit::Get` always get the best one.
## Optional
Add more implementations of `your_kery` for performance enhancement.
1. Add functions based on generated code in `gen`. It should be derived from `JitCode` and should have corepsonding creator from `JitCodeCreator` which will be registered on the `your_key`.
Note: Add new `KernelTuples` if necessary，your can refer to `XYZNTuples`.
Specialie method `JitCodeKey` when add new attribute type。
2. Add more functions in `more`，you can use any third party you wish, like mkl, mkldnn or intrinsic code to reach the best performance.
paddle/fluid/operators/jit/README.md
浏览文件 @ f01c9668
......@@ -10,9 +10,9 @@
```txt
PaddlePaddle/Paddle/paddle/fluid/
├── ...
├── operator/
│ ├── .../
└── jit/
└── operators/
├── .../
└── jit/
├── ...
├── gen/
│ └── ...
......
paddle/fluid/operators/jit/gen_base.h
浏览文件 @ f01c9668
......@@ -36,6 +36,8 @@ class GenBase : public Kernel {
if (FLAGS_dump_jitcode) {
this->dumpCode(code);
}
// Note: failed to cast with reinterpret_cast<const Func> on Mac clang,
// then workaround with const_cast. Any better idea is appreciated.
return reinterpret_cast<Func>(const_cast<unsigned char*>(code));
}
......
paddle/fluid/operators/math/concat_and_split.cu
浏览文件 @ f01c9668
......@@ -131,9 +131,8 @@ class ConcatFunctor<platform::CUDADeviceContext, T> {
int in_col = input[0].numel() / in_row;
int out_row = in_row, out_col = 0;
framework::Vector<int16_t> inputs_data(in_num * sizeof(T*) / 2);
framework::Vector<int> inputs_col(in_num + 1);
T** inputs_ptr = reinterpret_cast<T**>(inputs_data.data());
std::vector<T*> inputs_data(in_num);
std::vector<int> inputs_col(in_num + 1);
inputs_col[0] = 0;
bool sameShape = true;
......@@ -144,12 +143,9 @@ class ConcatFunctor<platform::CUDADeviceContext, T> {
}
out_col += t_cols;
inputs_col[i + 1] = out_col;
inputs_ptr[i] = const_cast<T*>(input[i].data<T>());
inputs_data[i] = const_cast<T*>(input[i].data<T>());
}
T** dev_ins_data =
reinterpret_cast<T**>(inputs_data.CUDAMutableData(context.GetPlace()));
// computation
// set the thread block and grid according to CurrentDeviceId
const int kThreadsPerBlock = 1024;
......@@ -169,18 +165,32 @@ class ConcatFunctor<platform::CUDADeviceContext, T> {
std::min(max_blocks / grid_cols, std::max(out_row / block_rows, 1));
dim3 grid_size = dim3(grid_cols, grid_rows, 1);
auto tmp_dev_ins_data =
platform::DeviceTemporaryAllocator::Instance().Get(context).Allocate(
inputs_data.size() * sizeof(T*));
memory::Copy(boost::get<platform::CUDAPlace>(context.GetPlace()),
tmp_dev_ins_data->ptr(), platform::CPUPlace(),
static_cast<void*>(inputs_data.data()),
inputs_data.size() * sizeof(T*), context.stream());
T** dev_ins_data = reinterpret_cast<T**>(tmp_dev_ins_data->ptr());
if (sameShape) {
ConcatKernel<<<grid_size, block_size, 0, context.stream()>>>(
dev_ins_data, in_col, out_row, out_col, output->data<T>());
} else {
const int* dev_ins_col_data = inputs_col.CUDAData(context.GetPlace());
auto tmp_dev_ins_col_data =
platform::DeviceTemporaryAllocator::Instance().Get(context).Allocate(
inputs_col.size() * sizeof(int));
memory::Copy(boost::get<platform::CUDAPlace>(context.GetPlace()),
tmp_dev_ins_col_data->ptr(), platform::CPUPlace(),
static_cast<void*>(inputs_col.data()),
inputs_col.size() * sizeof(int), context.stream());
int* dev_ins_col_data = static_cast<int*>(tmp_dev_ins_col_data->ptr());
ConcatKernel<<<grid_size, block_size, 0, context.stream()>>>(
dev_ins_data, dev_ins_col_data, static_cast<int>(inputs_col.size()),
out_row, out_col, output->data<T>());
}
// Wait() must be called because `inputs_data` may be destructed before
// kernel ends
context.Wait();
}
};
......@@ -207,9 +217,8 @@ class SplitFunctor<platform::CUDADeviceContext, T> {
int in_col = 0, in_row = out_row;
bool sameShape = true;
framework::Vector<int16_t> outputs_data(o_num * sizeof(T*) / 2);
framework::Vector<int> outputs_cols(o_num + 1);
T** outputs_ptr = reinterpret_cast<T**>(outputs_data.data());
std::vector<T*> outputs_data(o_num);
std::vector<int> outputs_cols(o_num + 1);
outputs_cols[0] = 0;
for (int i = 0; i < o_num; ++i) {
......@@ -220,15 +229,12 @@ class SplitFunctor<platform::CUDADeviceContext, T> {
in_col += t_col;
outputs_cols[i + 1] = in_col;
if (outputs->at(i) != nullptr) {
outputs_ptr[i] = outputs->at(i)->data<T>();
outputs_data[i] = outputs->at(i)->data<T>();
} else {
outputs_ptr[i] = nullptr;
outputs_data[i] = nullptr;
}
}
T** dev_out_gpu_data =
reinterpret_cast<T**>(outputs_data.CUDAMutableData(context.GetPlace()));
// computation
const int kThreadsPerBlock = 1024;
int block_cols = kThreadsPerBlock;
......@@ -247,18 +253,33 @@ class SplitFunctor<platform::CUDADeviceContext, T> {
std::min(max_blocks / grid_cols, std::max(out_row / block_rows, 1));
dim3 grid_size = dim3(grid_cols, grid_rows, 1);
auto tmp_dev_outs_data =
platform::DeviceTemporaryAllocator::Instance().Get(context).Allocate(
outputs_data.size() * sizeof(T*));
memory::Copy(boost::get<platform::CUDAPlace>(context.GetPlace()),
tmp_dev_outs_data->ptr(), platform::CPUPlace(),
reinterpret_cast<void*>(outputs_data.data()),
outputs_data.size() * sizeof(T*), context.stream());
T** dev_out_gpu_data = reinterpret_cast<T**>(tmp_dev_outs_data->ptr());
if (sameShape) {
SplitKernel<<<grid_size, block_size, 0, context.stream()>>>(
input.data<T>(), in_row, in_col, out0_col, dev_out_gpu_data);
} else {
const int* dev_outs_col_data = outputs_cols.CUDAData(context.GetPlace());
auto tmp_dev_ins_col_data =
platform::DeviceTemporaryAllocator::Instance().Get(context).Allocate(
outputs_cols.size() * sizeof(int));
memory::Copy(boost::get<platform::CUDAPlace>(context.GetPlace()),
tmp_dev_ins_col_data->ptr(), platform::CPUPlace(),
reinterpret_cast<void*>(outputs_cols.data()),
outputs_cols.size() * sizeof(int), context.stream());
int* dev_outs_col_data =
reinterpret_cast<int*>(tmp_dev_ins_col_data->ptr());
SplitKernel<<<grid_size, block_size, 0, context.stream()>>>(
input.data<T>(), in_row, in_col, dev_outs_col_data,
static_cast<int>(outputs_cols.size()), dev_out_gpu_data);
}
// Wait() must be called because `outputs_data` may be destructed before
// kernel ends
context.Wait();
}
};
......
paddle/fluid/operators/math/detail/lstm_cpu_kernel.h
浏览文件 @ f01c9668
......@@ -17,6 +17,12 @@ limitations under the License. */
#include "paddle/fluid/operators/math/detail/activation_functions.h"
#include "paddle/fluid/operators/math/lstm_compute.h"
#if defined(_WIN32)
#if defined(__AVX2__) || defined(__AVX__)
inline __m256 operator+=(__m256 a, __m256 b) { return _mm256_add_ps(a, b); }
#endif
#endif
namespace paddle {
namespace operators {
namespace math {
......
paddle/fluid/operators/mean_iou_op.cu
浏览文件 @ f01c9668
......@@ -92,8 +92,8 @@ template <typename T>
class MeanIoUCUDAOpKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
auto& place = *ctx.template device_context<platform::CUDADeviceContext>()
.eigen_device();
auto& dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
auto& place = *dev_ctx.eigen_device();
// get input and output tensor
auto* predictions = ctx.Input<Tensor>("Predictions");
auto* labels = ctx.Input<Tensor>("Labels");
......@@ -115,11 +115,11 @@ class MeanIoUCUDAOpKernel : public framework::OpKernel<T> {
auto out_wrong_t = EigenTensor<int, 1>::From(*out_wrong);
auto out_correct_t = EigenTensor<int, 1>::From(*out_correct);
// Temporary tensor
Tensor ious;
float* ious_data = ious.mutable_data<float>(
{static_cast<int64_t>(num_classes)}, ctx.GetPlace());
auto ious_t = EigenTensor<float, 1>::From(ious);
// Temporary memory
auto& allocator =
platform::DeviceTemporaryAllocator::Instance().Get(dev_ctx);
auto tmp_ious_data = allocator.Allocate(num_classes * sizeof(float));
float* ious_data = static_cast<float*>(tmp_ious_data->ptr());
// Init out_wrong, out_correct and out_mean_iou
out_wrong_t.device(place) = out_wrong_t.constant(0);
......@@ -148,7 +148,7 @@ class MeanIoUCUDAOpKernel : public framework::OpKernel<T> {
CountCUDAKernel<T><<<grid, block, cache_size, stream>>>(
num_classes, predictions->numel(), predictions_data, labels_data,
out_wrong_data, out_correct_data);
ctx.device_context().Wait();
ComputeIoUCUDAKernel<<<1, block, 0, stream>>>(num_classes, out_wrong_data,
out_correct_data, ious_data,
out_mean_iou_data);
......
paddle/fluid/operators/transpose_mkldnn_op.cc
浏览文件 @ f01c9668
......@@ -29,10 +29,6 @@ class TransposeMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
void Compute(const paddle::framework::ExecutionContext& ctx) const override {
PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()),
"It must use CPUPlace.");
const bool is_test = ctx.Attr<bool>("is_test");
PADDLE_ENFORCE(
is_test == true,
"TransposeMKLDNN works only for inference!. Set is_test = True");
auto& dev_ctx =
ctx.template device_context<paddle::platform::MKLDNNDeviceContext>();
const auto& mkldnn_engine = dev_ctx.GetEngine();
......@@ -68,6 +64,57 @@ class TransposeMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
}
};
template <typename T>
class TransposeMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
public:
void Compute(const paddle::framework::ExecutionContext& ctx) const override {
PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()),
"It must use CPUPlace.");
auto* out_grad =
ctx.Input<framework::Tensor>(framework::GradVarName("Out"));
auto* x_grad = ctx.Output<framework::Tensor>(framework::GradVarName("X"));
if (!x_grad) return;
auto& dev_ctx =
ctx.template device_context<paddle::platform::MKLDNNDeviceContext>();
const auto& mkldnn_engine = dev_ctx.GetEngine();
std::vector<int> axis = ctx.Attr<std::vector<int>>("axis");
std::vector<int> reversed_axis(axis);
int ndims = axis.size();
if (ndims == 1) {
x_grad->ShareDataWith(*out_grad);
return;
}
for (size_t i = 0; i < axis.size(); i++) {
reversed_axis[axis[i]] = i;
}
const T* out_grad_data = out_grad->data<T>();
x_grad->mutable_data<T>(ctx.GetPlace());
std::vector<int> nchw_tz =
paddle::framework::vectorize2int(out_grad->dims());
const std::string key = platform::TransposeMKLDNNHandler::GetHash(
nchw_tz, axis, ctx.op().Output(framework::GradVarName("X")));
platform::TransposeMKLDNNHandler handler(nchw_tz, reversed_axis, dev_ctx,
mkldnn_engine, key);
auto transpose_src_memory_p = handler.AcquireSrcMemory(
out_grad->format(), platform::to_void_cast<T>(out_grad_data));
auto transpose_dst_memory_p =
handler.AcquireDstMemory(x_grad, ctx.GetPlace());
auto transpose_p = handler.AcquireTranspose(transpose_dst_memory_p,
transpose_src_memory_p);
std::vector<mkldnn::primitive> pipeline;
pipeline.push_back(*transpose_p);
mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait();
}
};
} // namespace operators
} // namespace paddle
......@@ -77,3 +124,8 @@ REGISTER_OP_KERNEL(transpose2, MKLDNN, ::paddle::platform::CPUPlace,
ops::TransposeMKLDNNOpKernel<float>);
REGISTER_OP_KERNEL(transpose, MKLDNN, ::paddle::platform::CPUPlace,
ops::TransposeMKLDNNOpKernel<float>);
REGISTER_OP_KERNEL(transpose_grad, MKLDNN, ::paddle::platform::CPUPlace,
ops::TransposeMKLDNNGradOpKernel<float>);
REGISTER_OP_KERNEL(transpose2_grad, MKLDNN, ::paddle::platform::CPUPlace,
ops::TransposeMKLDNNGradOpKernel<float>);
paddle/fluid/operators/transpose_op.cc
浏览文件 @ f01c9668
......@@ -79,10 +79,6 @@ class TransposeOp : public framework::OperatorWithKernel {
class TransposeOpMaker : public framework::OpProtoAndCheckerMaker {
public:
void Make() override {
AddAttr<bool>("is_test",
"(bool, default false) Set to true for inference only, false "
"for training. Some layers may run faster when this is true.")
.SetDefault(false);
AddInput(
"X",
"(Tensor) The input tensor, tensors with rank up to 6 are supported.");
......@@ -147,6 +143,24 @@ class TransposeOpGrad : public framework::OperatorWithKernel {
ctx->SetOutputDim(framework::GradVarName("X"), x_dims);
}
}
protected:
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext &ctx) const override {
framework::LibraryType library_{framework::LibraryType::kPlain};
std::string data_format = ctx.Attr<std::string>("data_format");
framework::DataLayout layout_ = framework::StringToDataLayout(data_format);
#ifdef PADDLE_WITH_MKLDNN
if (library_ == framework::LibraryType::kPlain &&
platform::CanMKLDNNBeUsed(ctx)) {
library_ = framework::LibraryType::kMKLDNN;
layout_ = framework::DataLayout::kMKLDNN;
}
#endif
return framework::OpKernelType(
ctx.Input<framework::LoDTensor>(framework::GradVarName("Out"))->type(),
ctx.GetPlace(), layout_, library_);
}
};
// FIXME(zcd): transpose2 adds an intermediate output(XShape) based on
......@@ -237,9 +251,19 @@ class Transpose2OpGrad : public framework::OperatorWithKernel {
protected:
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext &ctx) const override {
framework::LibraryType library_{framework::LibraryType::kPlain};
std::string data_format = ctx.Attr<std::string>("data_format");
framework::DataLayout layout_ = framework::StringToDataLayout(data_format);
#ifdef PADDLE_WITH_MKLDNN
if (library_ == framework::LibraryType::kPlain &&
platform::CanMKLDNNBeUsed(ctx)) {
library_ = framework::LibraryType::kMKLDNN;
layout_ = framework::DataLayout::kMKLDNN;
}
#endif
return framework::OpKernelType(
ctx.Input<framework::LoDTensor>(framework::GradVarName("Out"))->type(),
ctx.device_context());
ctx.GetPlace(), layout_, library_);
}
};
......
paddle/fluid/platform/CMakeLists.txt
浏览文件 @ f01c9668
......@@ -56,6 +56,8 @@ ELSE()
set(MKLDNN_CTX_DEPS)
ENDIF()
cc_library(temp_allocator SRCS temporary_allocator.cc DEPS allocator_facade)
nv_library(stream_callback_manager SRCS stream_callback_manager.cc DEPS simple_threadpool enforce)
IF(WITH_GPU)
set(STREAM_CALLBACK_DEPS stream_callback_manager)
......@@ -66,7 +68,8 @@ ENDIF()
# memcpy depends on device_context, here add deps individually for
# avoiding cycle dependencies
cc_library(device_context SRCS device_context.cc init.cc DEPS simple_threadpool malloc ${STREAM_CALLBACK_DEPS}
place eigen3 stringpiece cpu_helper cpu_info framework_proto ${GPU_CTX_DEPS} ${MKLDNN_CTX_DEPS})
place eigen3 stringpiece cpu_helper cpu_info framework_proto ${GPU_CTX_DEPS} ${MKLDNN_CTX_DEPS} temp_allocator)
if(WIN32)
if(WITH_GPU AND NOT WITH_DSO)
get_property(cuda_modules GLOBAL PROPERTY CUDA_MODULES)
......@@ -92,3 +95,9 @@ IF(WITH_GPU)
nv_test(cuda_helper_test SRCS cuda_helper_test.cu)
ENDIF()
nv_library(cuda_device_guard SRCS cuda_device_guard.cc DEPS gpu_info)
if(WITH_GPU)
nv_test(temporal_allocator_test SRCS temporary_allocator_test.cc DEPS temp_allocator tensor)
else()
cc_test(temporal_allocator_test SRCS temporary_allocator_test.cc DEPS temp_allocator tensor)
endif()
paddle/fluid/platform/cpu_info.cc
浏览文件 @ f01c9668
......@@ -22,7 +22,6 @@ limitations under the License. */
#ifdef __APPLE__
#include <sys/sysctl.h>
#include <sys/types.h>
#elif defined(_WIN32)
#define NOMINMAX // msvc max/min macro conflict with std::min/max
#include <windows.h>
......
paddle/fluid/platform/create_tensor_with_allocationptr.h 0 → 100644
浏览文件 @ f01c9668
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed 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 "paddle/fluid/framework/tensor.h"
#include "paddle/fluid/platform/temporary_allocator.h"
namespace paddle {
namespace platform {
template <typename T>
paddle::framework::Tensor GetTensor(
memory::allocation::AllocationPtr temp_allocation_ptr,
const framework::DDim &dim) {
auto &deleter = temp_allocation_ptr.get_deleter();
auto *allocation_ptr = temp_allocation_ptr.release();
auto shared_allocation =
std::shared_ptr<memory::allocation::Allocation>(allocation_ptr, deleter);
PADDLE_ENFORCE(dynamic_cast<TemporaryAllocation *>(allocation_ptr) != nullptr,
"The AllocationPtr must be TemporaryAllocation.");
PADDLE_ENFORCE_EQ(allocation_ptr->size(),
framework::product(dim) * sizeof(T));
paddle::framework::Tensor temp_tensor(std::type_index(typeid(T)));
temp_tensor.Resize(dim);
temp_tensor.ResetHolder(std::move(shared_allocation));
return temp_tensor;
}
} // namespace platform
} // namespace paddle
paddle/fluid/platform/device_context.cc
浏览文件 @ f01c9668
......@@ -85,6 +85,49 @@ DeviceContextPool::DeviceContextPool(
}
}
DeviceTemporaryAllocator* DeviceTemporaryAllocator::allocators = nullptr;
#ifdef PADDLE_WITH_CUDA
platform::TemporaryAllocator& DeviceTemporaryAllocator::Get(
const platform::Place& place, const cudaStream_t& stream) {
PADDLE_ENFORCE(platform::is_gpu_place(place));
auto place_stream = std::make_pair(place, stream);
{
std::unique_lock<std::mutex> lock(mtx_);
if (!device_allocator_.count(place_stream)) {
device_allocator_[place_stream].reset(new TemporaryAllocator(place));
device_allocator_[place_stream]->SetCallback([stream]() {
PADDLE_ENFORCE(cudaStreamSynchronize(stream));
PADDLE_ENFORCE(cudaGetLastError());
});
}
}
return *device_allocator_.at(place_stream);
}
template <>
platform::TemporaryAllocator& DeviceTemporaryAllocator::Get(
const platform::CUDADeviceContext& dev_ctx) {
auto place_stream = std::make_pair(dev_ctx.GetPlace(), dev_ctx.stream());
if (device_allocator_.count(place_stream)) {
return *device_allocator_.at(place_stream);
}
return Get(dev_ctx.GetPlace(), dev_ctx.stream());
}
#endif
template <>
platform::TemporaryAllocator& DeviceTemporaryAllocator::Get(
const platform::CPUDeviceContext& dev_ctx) {
return cpu_allocator_;
}
platform::TemporaryAllocator& DeviceTemporaryAllocator::Get(
const platform::Place& place) {
PADDLE_ENFORCE(platform::is_cpu_place(place), "You should pass CPUPlace");
return cpu_allocator_;
}
CPUDeviceContext::CPUDeviceContext() {
eigen_device_.reset(new Eigen::DefaultDevice());
}
......@@ -271,8 +314,12 @@ CUDADeviceContext::~CUDADeviceContext() {
Place CUDADeviceContext::GetPlace() const { return place_; }
void CUDADeviceContext::Wait() const {
auto& allocator =
DeviceTemporaryAllocator::Instance().Get<CUDADeviceContext>(*this);
allocator.Release([=]() {
PADDLE_ENFORCE(cudaStreamSynchronize(stream_));
PADDLE_ENFORCE(cudaGetLastError());
});
}
int CUDADeviceContext::GetComputeCapability() const {
......
paddle/fluid/platform/device_context.h
浏览文件 @ f01c9668
......@@ -15,8 +15,10 @@ limitations under the License. */
#include <mutex> // NOLINT
#include <string>
#include <unordered_map>
#include <utility>
#include <vector>
#include "paddle/fluid/memory/malloc.h"
#include "paddle/fluid/platform/temporary_allocator.h"
#ifdef PADDLE_WITH_CUDA
#include "paddle/fluid/platform/dynload/cublas.h"
#include "paddle/fluid/platform/dynload/cudnn.h"
......@@ -39,6 +41,50 @@ limitations under the License. */
namespace paddle {
namespace platform {
/*! \brief device temporary allocator singleton */
class DeviceTemporaryAllocator {
public:
static DeviceTemporaryAllocator& Instance() {
PADDLE_ENFORCE_NOT_NULL(allocators,
"Need to Create DeviceTemporaryAllocator first!");
return *allocators;
}
static DeviceTemporaryAllocator& Init() {
if (allocators == nullptr) {
allocators = new DeviceTemporaryAllocator();
}
return *allocators;
}
/*! \brief Return handle of single temporary allocator. */
#ifdef PADDLE_WITH_CUDA
platform::TemporaryAllocator& Get(const platform::Place& place,
const cudaStream_t& stream);
#endif
template <typename DeviceContext>
platform::TemporaryAllocator& Get(const DeviceContext& dev_ctx);
platform::TemporaryAllocator& Get(const platform::Place& place);
private:
DeviceTemporaryAllocator() : cpu_allocator_(platform::CPUPlace()) {}
static DeviceTemporaryAllocator* allocators;
platform::TemporaryAllocator cpu_allocator_;
#ifdef PADDLE_WITH_CUDA
std::map<std::pair<platform::Place, cudaStream_t>,
std::unique_ptr<platform::TemporaryAllocator>>
device_allocator_;
#endif
std::mutex mtx_;
DISABLE_COPY_AND_ASSIGN(DeviceTemporaryAllocator);
};
class DeviceContext {
public:
virtual ~DeviceContext() {}
......
paddle/fluid/platform/dynload/dynamic_loader.cc
浏览文件 @ f01c9668
......@@ -227,6 +227,8 @@ void* GetTensorRtDsoHandle() {
void* GetMKLMLDsoHandle() {
#if defined(__APPLE__) || defined(__OSX__)
return GetDsoHandleFromSearchPath(FLAGS_mklml_dir, "libmklml_intel.dylib");
#elif defined(_WIN32)
return GetDsoHandleFromSearchPath(FLAGS_mklml_dir, "mklml.dll");
#else
return GetDsoHandleFromSearchPath(FLAGS_mklml_dir, "libmklml_intel.so");
#endif
......
paddle/fluid/platform/init.cc
浏览文件 @ f01c9668
......@@ -110,7 +110,7 @@ void InitDevices(bool init_p2p, const std::vector<int> devices) {
}
places.emplace_back(platform::CPUPlace());
platform::DeviceContextPool::Init(places);
platform::DeviceTemporaryAllocator::Init();
#ifndef PADDLE_WITH_MKLDNN
platform::SetNumThreads(FLAGS_paddle_num_threads);
#endif
......
paddle/fluid/platform/temporary_allocator.cc 0 → 100644
浏览文件 @ f01c9668
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed 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 "paddle/fluid/platform/temporary_allocator.h"
#include "paddle/fluid/memory/allocation/allocator_facade.h"
DEFINE_double(limit_of_temporary_allocation, -1,
"The up limit of temporary_allocation size.");
namespace paddle {
namespace platform {
namespace alloc = memory::allocation;
TemporaryAllocation::TemporaryAllocation(
alloc::AllocationPtr &&underlying_allocation)
: Allocation(underlying_allocation->ptr(), underlying_allocation->size(),
underlying_allocation->place()),
underlying_allocation_(std::move(underlying_allocation)) {}
TemporaryAllocator::TemporaryAllocator(platform::Place place) : place_(place) {
temp_mem_queue_.reset(new std::deque<TemporaryAllocation *>());
}
bool TemporaryAllocator::IsAllocThreadSafe() const { return true; }
void TemporaryAllocator::Release(const std::function<void()> &callback) {
std::shared_ptr<std::deque<TemporaryAllocation *>> t_allocations;
{
std::unique_lock<std::mutex> lock(mtx_);
callback();
t_allocations = temp_mem_queue_;
temp_mem_queue_.reset(new std::deque<TemporaryAllocation *>());
wait_delete_mem_ = 0;
}
for (auto tmp : *t_allocations) {
VLOG(10) << "Delete temporary allocation " << tmp->ptr()
<< " size: " << tmp->size();
delete tmp;
}
}
void TemporaryAllocator::Free(alloc::Allocation *allocation) {
auto *temp_allocation = dynamic_cast<TemporaryAllocation *>(allocation);
PADDLE_ENFORCE_NOT_NULL(temp_allocation);
if (platform::is_gpu_place(temp_allocation->place())) {
size_t wait_delete_mem = 0;
{
std::unique_lock<std::mutex> lock(mtx_);
temp_mem_queue_->emplace_back(temp_allocation);
wait_delete_mem_ += temp_allocation->size();
wait_delete_mem = wait_delete_mem_;
VLOG(10) << "Move temporary allocation: " << temp_allocation->ptr()
<< " to delete queue: " << temp_allocation->size() << "; "
<< "wait_delete_mem: " << wait_delete_mem_;
}
if (FLAGS_limit_of_temporary_allocation > 0 &&
wait_delete_mem > FLAGS_limit_of_temporary_allocation) {
Release(callback_);
}
return;
}
delete temp_allocation;
}
size_t TemporaryAllocator::TemporaryAllocationQueueSize() {
std::unique_lock<std::mutex> lock(mtx_);
return temp_mem_queue_ ? temp_mem_queue_->size() : 0;
}
void TemporaryAllocator::SetCallback(const std::function<void()> &callback) {
callback_ = callback;
}
alloc::Allocation *TemporaryAllocator::AllocateImpl(
size_t size, alloc::Allocator::Attr attr) {
auto raw_allocation =
alloc::AllocatorFacade::Instance().Alloc(place_, size, attr);
auto temp_mem = new TemporaryAllocation(std::move(raw_allocation));
VLOG(10) << "Alloc temporary allocation: " << temp_mem->ptr() << ": " << size;
return temp_mem;
}
} // namespace platform
} // namespace paddle
paddle/fluid/platform/temporary_allocator.h 0 → 100644
浏览文件 @ f01c9668
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed 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 <condition_variable> // NOLINT
#include <deque>
#include <mutex> // NOLINT
#include "paddle/fluid/memory/allocation/allocator.h"
#include "paddle/fluid/platform/lock_guard_ptr.h"
namespace paddle {
namespace platform {
class TemporaryAllocation : public memory::allocation::Allocation {
public:
explicit TemporaryAllocation(
memory::allocation::AllocationPtr &&underlying_allocation);
memory::allocation::AllocationPtr underlying_allocation_;
};
class TemporaryAllocator : public memory::allocation::Allocator {
public:
explicit TemporaryAllocator(platform::Place place);
void Release(const std::function<void()> &callback);
size_t TemporaryAllocationQueueSize();
bool IsAllocThreadSafe() const override;
void SetCallback(const std::function<void()> &callback);
protected:
void Free(memory::allocation::Allocation *allocation) override;
memory::allocation::Allocation *AllocateImpl(
size_t size, memory::allocation::Allocator::Attr attr) override;
private:
platform::Place place_;
// When the allocation is not held by any variable, it should be placed
// to temp_mem_queue immediately.
std::shared_ptr<std::deque<TemporaryAllocation *>> temp_mem_queue_{nullptr};
std::mutex mtx_;
size_t wait_delete_mem_{0};
std::function<void()> callback_;
};
} // namespace platform
} // namespace paddle
paddle/fluid/platform/temporary_allocator_test.cc 0 → 100644
浏览文件 @ f01c9668
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed 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 "paddle/fluid/platform/temporary_allocator.h"
#include <gtest/gtest.h>
#include "paddle/fluid/framework/tensor.h"
#include "paddle/fluid/platform/create_tensor_with_allocationptr.h"
DECLARE_double(limit_of_temporary_allocation);
namespace paddle {
namespace platform {
TEST(temporary_allocator, temporary_allocator) {
platform::CPUPlace cpu_place;
TemporaryAllocator alloc(cpu_place);
alloc.Allocate(100);
#ifdef PADDLE_WITH_CUDA
platform::CUDAPlace gpu_place(0);
TemporaryAllocator gpu_alloc(gpu_place);
auto allocation = gpu_alloc.Allocate(101);
PADDLE_ENFORCE_EQ(gpu_alloc.TemporaryAllocationQueueSize(), 0);
gpu_alloc.Release([]() {});
PADDLE_ENFORCE_EQ(gpu_alloc.TemporaryAllocationQueueSize(), 0);
{
auto allocation = gpu_alloc.Allocate(102);
PADDLE_ENFORCE_EQ(gpu_alloc.TemporaryAllocationQueueSize(), 0);
}
PADDLE_ENFORCE_EQ(gpu_alloc.TemporaryAllocationQueueSize(), 1);
gpu_alloc.Release([]() {});
PADDLE_ENFORCE_EQ(gpu_alloc.TemporaryAllocationQueueSize(), 0);
#endif
}
TEST(temporary_allocator, add_callback) {
#ifdef PADDLE_WITH_CUDA
FLAGS_limit_of_temporary_allocation = 10;
platform::CUDAPlace gpu_place(0);
TemporaryAllocator gpu_alloc(gpu_place);
platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
auto* dev_ctx =
static_cast<platform::CUDADeviceContext*>(pool.Get(gpu_place));
auto stream = dev_ctx->stream();
bool deleted = false;
gpu_alloc.SetCallback([stream, &deleted]() {
PADDLE_ENFORCE(cudaStreamSynchronize(stream));
PADDLE_ENFORCE(cudaGetLastError());
deleted = true;
});
{ gpu_alloc.Allocate(100); }
PADDLE_ENFORCE(deleted);
FLAGS_limit_of_temporary_allocation = -1;
#endif
}
TEST(temporary_allocator, create_tensor_with_allocationptr) {
platform::CPUPlace cpu_place;
TemporaryAllocator cpu_alloc(cpu_place);
{
size_t memory_size = 200;
auto allocation = cpu_alloc.Allocate(memory_size);
void* address = allocation->ptr();
int numel = memory_size / sizeof(float);
framework::Tensor tensor =
GetTensor<float>(std::move(allocation), framework::make_ddim({numel}));
PADDLE_ENFORCE_EQ(address, tensor.data<float>());
PADDLE_ENFORCE_EQ(tensor.numel(), numel);
}
#ifdef PADDLE_WITH_CUDA
platform::CUDAPlace gpu_place(0);
TemporaryAllocator gpu_alloc(gpu_place);
{
size_t memory_size = 300;
auto allocation = gpu_alloc.Allocate(memory_size);
void* address = allocation->ptr();
int numel = memory_size / sizeof(float);
framework::Tensor tensor =
GetTensor<float>(std::move(allocation), framework::make_ddim({numel}));
PADDLE_ENFORCE_EQ(address, tensor.data<float>());
PADDLE_ENFORCE_EQ(tensor.numel(), numel);
}
// The allocation is not holded now, it should be placed to
// TemporaryAllocationQueue.
PADDLE_ENFORCE_EQ(gpu_alloc.TemporaryAllocationQueueSize(), 1);
gpu_alloc.Release([]() {});
PADDLE_ENFORCE_EQ(gpu_alloc.TemporaryAllocationQueueSize(), 0);
#endif
}
TEST(temporary_allocator, create_tensor_with_allocationptr2) {
platform::CPUPlace cpu_place;
TemporaryAllocator cpu_alloc(cpu_place);
{
size_t memory_size = 400;
int numel = memory_size / sizeof(float);
framework::Tensor out_side_tensor;
void* address;
{
auto allocation = cpu_alloc.Allocate(memory_size);
address = allocation->ptr();
framework::Tensor tensor = GetTensor<float>(
std::move(allocation), framework::make_ddim({numel}));
PADDLE_ENFORCE_EQ(address, tensor.data<float>());
PADDLE_ENFORCE_EQ(tensor.numel(), numel);
out_side_tensor.ShareDataWith(tensor);
}
PADDLE_ENFORCE_EQ(address, out_side_tensor.data<float>());
PADDLE_ENFORCE_EQ(out_side_tensor.numel(), numel);
}
#ifdef PADDLE_WITH_CUDA
platform::CUDAPlace gpu_place(0);
TemporaryAllocator gpu_alloc(gpu_place);
{
void* address;
size_t memory_size = 500;
int numel = memory_size / sizeof(float);
framework::Tensor out_side_tensor;
{
auto allocation = gpu_alloc.Allocate(memory_size);
address = allocation->ptr();
framework::Tensor tensor = GetTensor<float>(
std::move(allocation), framework::make_ddim({numel}));
PADDLE_ENFORCE_EQ(address, tensor.data<float>());
PADDLE_ENFORCE_EQ(tensor.numel(), numel);
out_side_tensor.ShareDataWith(tensor);
}
PADDLE_ENFORCE_EQ(address, out_side_tensor.data<float>());
PADDLE_ENFORCE_EQ(out_side_tensor.numel(), numel);
// The allocation is holded by out_side_tensor.
PADDLE_ENFORCE_EQ(gpu_alloc.TemporaryAllocationQueueSize(), 0);
gpu_alloc.Release([]() {});
PADDLE_ENFORCE_EQ(gpu_alloc.TemporaryAllocationQueueSize(), 0);
}
// The allocation is not holded now, it should be placed to
// TemporaryAllocationQueue.
PADDLE_ENFORCE_EQ(gpu_alloc.TemporaryAllocationQueueSize(), 1);
gpu_alloc.Release([]() {});
PADDLE_ENFORCE_EQ(gpu_alloc.TemporaryAllocationQueueSize(), 0);
#endif
}
} // namespace platform
} // namespace paddle
paddle/fluid/train/demo/CMakeLists.txt
浏览文件 @ f01c9668
......@@ -35,16 +35,26 @@ add_executable(demo_trainer demo_trainer.cc)
if(WITH_MKLDNN)
include_directories("${PADDLE_LIB}/third_party/install/mkldnn/include")
if(WIN32)
set(MKLDNN_LIB ${PADDLE_LIB}/third_party/install/mkldnn/lib/mkldnn.lib)
else(WIN32)
set(MKLDNN_LIB ${PADDLE_LIB}/third_party/install/mkldnn/lib/libmkldnn.so.0)
endif()
endif(WIN32)
endif(WITH_MKLDNN)
if(WITH_MKL)
include_directories("${PADDLE_LIB}/third_party/install/mklml/include")
if(WIN32)
set(MATH_LIB ${PADDLE_LIB}/third_party/install/mklml/lib/mklml.lib)
else(WIN32)
set(MATH_LIB ${PADDLE_LIB}/third_party/install/mklml/lib/libmklml_intel.so)
endif(WIN32)
else()
if(APPLE)
set(MATH_LIB cblas)
else(APPLE)
elseif(WIN32)
set(MATH_LIB ${PADDLE_LIB}/third_party/install/openblas/lib/libopenblas.lib)
else()
set(MATH_LIB ${PADDLE_LIB}/third_party/install/openblas/lib/libopenblas.a)
endif(APPLE)
endif()
......
python/CMakeLists.txt
浏览文件 @ f01c9668
......@@ -48,18 +48,13 @@ configure_file(${CMAKE_CURRENT_SOURCE_DIR}/setup.py.in
IF(WIN32)
# Python would use the .pyd by default under Windows series platform
set(FLUID_DST_DIR ${PADDLE_BINARY_DIR}/python/paddle/fluid/)
get_filename_component(openblas_refpath ${CBLAS_LIBRARIES} DIRECTORY)
set(FLUID_CORE ${FLUID_DST_DIR}/core.pyd)
add_custom_command(OUTPUT ${FLUID_CORE}
COMMAND cmake -E copy $<TARGET_FILE:paddle_pybind> ${FLUID_CORE}
COMMAND cmake -E copy ${openblas_refpath}/openblas.dll ${FLUID_DST_DIR}
DEPENDS paddle_pybind)
ELSE()
set(FLUID_CORE ${PADDLE_BINARY_DIR}/python/paddle/fluid/core.so)
add_custom_command(OUTPUT ${FLUID_CORE}
ENDIF()
add_custom_command(OUTPUT ${FLUID_CORE}
COMMAND cmake -E copy $<TARGET_FILE:paddle_pybind> ${FLUID_CORE}
DEPENDS paddle_pybind)
ENDIF()
add_custom_target(copy_paddle_pybind ALL DEPENDS ${FLUID_CORE})
IF(WIN32)
......
python/paddle/fluid/contrib/__init__.py
浏览文件 @ f01c9668
......@@ -22,6 +22,8 @@ from . import op_frequence
from .op_frequence import *
from . import quantize
from .quantize import *
from . import slim
from .slim import *
from . import utils
from .utils import *
......@@ -30,4 +32,5 @@ __all__ += decoder.__all__
__all__ += memory_usage_calc.__all__
__all__ += op_frequence.__all__
__all__ += quantize.__all__
__all__ += slim.__all__
__all__ += utils.__all__
python/paddle/fluid/contrib/slim/__init__.py 0 → 100644
浏览文件 @ f01c9668
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed 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.
from .core import *
from .graph import *
from .prune import *
__all__ = [
'build_compressor',
'CompressPass',
'ImitationGraph',
'SensitivePruneStrategy',
'MagnitudePruner',
'RatioPruner',
]
python/paddle/fluid/contrib/slim/core/__init__.py 0 → 100644
浏览文件 @ f01c9668
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed 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.
from . import config
from .config import *
from . import compress_pass
from .compress_pass import *
from . import strategy
from .strategy import *
from . import pass_builder
from .pass_builder import *
__all__ = config.__all__ + compress_pass.__all__ + strategy.__all__ + pass_builder.__all__
python/paddle/fluid/contrib/slim/core/compress_pass.py 0 → 100644
浏览文件 @ f01c9668
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed 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.
from ....core import CPUPlace
from ..graph import get_executor
__all__ = ['Context', 'CompressPass']
class Context(object):
"""
The context in the process of compression.
Args:
exe: The executor used to execute graph.
graph: The graph to be compressed.
scope: The scope used to execute graph.
program_exe: The program_exe is used to execute the program
created for modifying the variables in scope.
"""
def __init__(self, exe, graph, scope, program_exe=None):
# The total number of epoches to be trained.
self.epoch = 0
# Current epoch
self.epoch_id = 0
# Current batch
self.batch_id = 0
self.exe = exe
self.graph = graph
self.scope = scope
self.program_exe = program_exe
class CompressPass(object):
"""
The pass used to compress model.
Args:
place: The device used in compression.
data_reader: The data_reader used to run graph.
data_feeder: The data_feeder used to run graph.
scope: The scope used to run graph.
metrics: The metrics for evaluating model.
epoch: The total epoches of trainning in compression.
program_exe: The program_exe is used to execute the program
created for modifying the variables in scope.
"""
def __init__(self,
place=None,
data_reader=None,
data_feeder=None,
scope=None,
metrics=None,
epoch=None,
program_exe=None):
self.strategies = []
self.place = CPUPlace() if place is None else place
self.data_reader = data_reader
self.data_feeder = data_feeder
self.scope = scope
self.metrics = metrics
self.epoch = epoch
self.program_exe = program_exe
def add_strategy(self, strategy):
"""
Add a strategy to current compress pass.
Args:
strategy: The strategy to be added into current compress pass.
"""
self.strategies.append(strategy)
self.epoch = max(strategy.end_epoch, self.epoch)
def apply(self, graph):
"""
Compress a model.
Args:
graph: The target graph to be compressed.
"""
self.executor = get_executor(graph, self.place)
context = Context(
self.executor, graph, self.scope, program_exe=self.program_exe)
for strategy in self.strategies:
strategy.on_compress_begin(context)
for epoch in range(self.epoch):
for strategy in self.strategies:
strategy.on_epoch_begin(context)
for data in self.data_reader():
for strategy in self.strategies:
strategy.on_batch_begin(context)
fetches = None
if self.metrics:
fetches = self.metrics.values()
feed = None
if self.data_feeder:
feed = self.data_feeder.feed(data)
results = self.executor.run(graph,
fetches=fetches,
scope=self.scope,
feed=feed)
if results:
print("results: {}".format(
zip(self.metrics.keys(), results)))
for strategy in self.strategies:
strategy.on_batch_end(context)
context.batch_id += 1
for strategy in self.strategies:
strategy.on_epoch_end(context)
context.epoch_id += 1
for strategy in self.strategies:
strategy.on_compress_end(context)
python/paddle/fluid/contrib/slim/core/config.py 0 → 100644
浏览文件 @ f01c9668
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed 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.
import inspect
import funcsigs
import yaml
from collections import OrderedDict
from ..prune import *
from .compress_pass import *
from .strategy import *
__all__ = ['ConfigFactory']
"""This factory is used to create instances by loading and parsing configure file with yaml format.
"""
class ConfigFactory(object):
def __init__(self, config):
"""Init a factory from configure file."""
self.instances = {}
self.version = None
self._parse_config(config)
def get_compress_pass(self):
"""
Get compress pass from factory.
"""
return self.instance('compress_pass')
def instance(self, name):
"""
Get instance from factory.
"""
if name in self.instances:
return self.instances[name]
else:
return None
def _new_instance(self, name, attrs):
if name not in self.instances:
class_ = globals()[attrs['class']]
sig = funcsigs.signature(class_.__init__)
keys = [
param.name for param in sig.parameters.values()
if (param.kind == param.POSITIONAL_OR_KEYWORD)
][1:]
keys = set(attrs.keys()).intersection(set(keys))
args = {}
for key in keys:
value = attrs[key]
if isinstance(value, str) and value in self.instances:
value = self.instances[value]
args[key] = value
self.instances[name] = class_(**args)
return self.instances.get(name)
def _parse_config(self, config):
assert config
with open(config, 'r') as config_file:
key_values = self._ordered_load(config_file)
for key in key_values:
# parse version
if key == 'version' and self.version is None:
self.version = int(key_values['version'])
assert self.version == int(key_values['version'])
# parse pruners
if key == 'pruners' or key == 'strategies':
instances = key_values[key]
for name in instances:
self._new_instance(name, instances[name])
if key == 'compress_pass':
compress_pass = self._new_instance(key, key_values[key])
for name in key_values[key]['strategies']:
strategy = self.instance(name)
compress_pass.add_strategy(strategy)
if key == 'include':
for config_file in key_values[key]:
self._parse_config(config_file.strip())
def _ordered_load(self,
stream,
Loader=yaml.Loader,
object_pairs_hook=OrderedDict):
"""
See: https://stackoverflow.com/questions/5121931/in-python-how-can-you-load-yaml-mappings-as-ordereddicts
"""
class OrderedLoader(Loader):
pass
def construct_mapping(loader, node):
loader.flatten_mapping(node)
return object_pairs_hook(loader.construct_pairs(node))
OrderedLoader.add_constructor(
yaml.resolver.BaseResolver.DEFAULT_MAPPING_TAG, construct_mapping)
return yaml.load(stream, OrderedLoader)
python/paddle/fluid/contrib/slim/core/pass_builder.py 0 → 100644
浏览文件 @ f01c9668
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed 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.
from .compress_pass import CompressPass
from .config import ConfigFactory
__all__ = ['build_compressor']
def build_compressor(place=None,
data_reader=None,
data_feeder=None,
scope=None,
metrics=None,
epoch=None,
config=None):
if config is not None:
factory = ConfigFactory(config)
comp_pass = factory.get_compress_pass()
else:
comp_pass = CompressPass()
comp_pass.place = place
comp_pass.data_reader = data_reader
comp_pass.data_feeder = data_feeder
comp_pass.scope = scope
comp_pass.metrics = metrics
comp_pass.epoch = epoch
return comp_pass
python/paddle/fluid/contrib/slim/core/strategy.py 0 → 100644
浏览文件 @ f01c9668
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed 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.
__all__ = ['Strategy']
class Strategy(object):
"""
Base class for all strategies.
"""
def __init__(self, start_epoch=0, end_epoch=10):
"""
Args:
start_epoch: The first epoch to apply the strategy.
end_epoch: The last epoch to apply the strategy.
"""
self.start_epoch = start_epoch
self.end_epoch = end_epoch
def on_compress_begin(self, context):
pass
def on_epoch_begin(self, context):
pass
def on_epoch_end(self, context):
pass
def on_batch_begin(self, context):
pass
def on_batch_end(self, context):
pass
def on_compress_end(self, context):
pass
python/paddle/fluid/contrib/slim/demo/filter_prune/config.yaml 0 → 100644
浏览文件 @ f01c9668
version: 1.0
pruners:
pruner_1:
class: 'RatioPruner'
ratios:
'conv1_1.w': 0.3
'conv1_2.w': 0.4
'*': 0.9
group_dims:
'*': [1, 2, 3]
criterions:
'*': 'l1-norm'
strategies:
strategy_1:
class: 'SensitivePruneStrategy'
pruner: 'pruner_1'
start_epoch: 0
end_epoch: 10
delta_rate: 0.20
acc_loss_threshold: 0.2
sensitivities:
'conv1_1.w': 0.4
compress_pass:
class: 'CompressPass'
epoch: 100
strategies:
- strategy_1
python/paddle/fluid/contrib/slim/demo/filter_prune/demo.py 0 → 100644
浏览文件 @ f01c9668
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed 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.
import paddle.fluid as fluid
import paddle
import os
import sys
from paddle.fluid.contrib.slim import CompressPass
from paddle.fluid.contrib.slim import build_compressor
from paddle.fluid.contrib.slim import ImitationGraph
class LinearModel(object):
def __init__(slef):
pass
def train(self):
train_program = fluid.Program()
startup_program = fluid.Program()
startup_program.random_seed = 10
with fluid.program_guard(train_program, startup_program):
x = fluid.layers.data(name='x', shape=[13], dtype='float32')
y = fluid.layers.data(name='y', shape=[1], dtype='float32')
predict = fluid.layers.fc(input=x, size=1, act=None)
cost = fluid.layers.square_error_cost(input=predict, label=y)
avg_cost = fluid.layers.mean(cost)
eval_program = train_program.clone()
sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.001)
sgd_optimizer.minimize(avg_cost)
train_reader = paddle.batch(
paddle.dataset.uci_housing.train(), batch_size=1)
eval_reader = paddle.batch(
paddle.dataset.uci_housing.test(), batch_size=1)
place = fluid.CPUPlace()
train_feeder = fluid.DataFeeder(place=place, feed_list=[x, y])
eval_feeder = fluid.DataFeeder(place=place, feed_list=[x, y])
exe = fluid.Executor(place)
exe.run(startup_program)
train_metrics = {"loss": avg_cost.name}
eval_metrics = {"loss": avg_cost.name}
graph = ImitationGraph(train_program)
config = './config.yaml'
comp_pass = build_compressor(
place,
data_reader=train_reader,
data_feeder=train_feeder,
scope=fluid.global_scope(),
metrics=train_metrics,
epoch=1,
config=config)
comp_pass.apply(graph)
if __name__ == "__main__":
model = LinearModel()
model.train()
python/paddle/fluid/contrib/slim/graph/__init__.py 0 → 100644
浏览文件 @ f01c9668
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed 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.
from . import executor
from .executor import *
from . import graph
from .graph import *
from . import graph_pass
from .graph_pass import *
__all__ = executor.__all__
__all__ += graph.__all__
__all__ += graph_pass.__all__
python/paddle/fluid/contrib/slim/graph/executor.py 0 → 100644
浏览文件 @ f01c9668
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed 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.
import abc
from abc import abstractmethod
from .... import executor
from .graph import IRGraph, ImitationGraph
__all__ = ['get_executor']
class GraphExecutor(object):
__metaclass__ = abc.ABCMeta
def __init__(self, place):
self.place = place
@abstractmethod
def run(self, graph, feches=None, feed=None):
pass
class IRGraphExecutor(GraphExecutor):
def run(self, grah, fetches, feed=None):
pass
class ImitationGraphExecutor(GraphExecutor):
def __init__(self, place):
super(ImitationGraphExecutor, self).__init__(place)
self.exe = executor.Executor(place)
def run(self, graph, scope=None, fetches=None, feed=None):
assert isinstance(graph, ImitationGraph)
fetch_list = None
if fetches:
fetch_list = [
graph.program.global_block().var(name) for name in fetches
]
results = self.exe.run(graph.program,
scope=scope,
fetch_list=fetch_list,
feed=feed)
return results
def get_executor(graph, place):
if isinstance(graph, ImitationGraph):
return ImitationGraphExecutor(place)
if isinstance(graph, IRGraph):
return IRGraphExecutor(place)
python/paddle/fluid/contrib/slim/graph/graph.py 0 → 100644
浏览文件 @ f01c9668
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed 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.
from ....framework import Program
__all__ = ['Graph', 'ImitationGraph', 'IRGraph']
class Graph(object):
"""
Base class for all graph.
"""
def __init__(self):
pass
def all_parameters(self):
"""
Return all the parameters in current graph.
"""
pass
class ImitationGraph(Graph):
def __init__(self, program=None):
super(ImitationGraph, self).__init__()
self.program = Program() if program is None else program
def all_parameters(self):
return self.program.global_block().all_parameters()
class IRGraph(Graph):
pass
python/paddle/fluid/contrib/slim/graph/graph_pass.py 0 → 100644
浏览文件 @ f01c9668
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed 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.
__all__ = ['GraphPass', 'PruneParameterPass']
class GraphPass(object):
"""
Base class for all graph pass.
"""
def __init__(self):
pass
def apply(self, graph):
pass
class PruneParameterPass(GraphPass):
"""
Generate a graph for pruning parameters from target graph.
"""
def __init__(self, pruned_params, thresholds):
super(PruneParameterPass, self).__init__()
self.pruned_params = pruned_params
self.thresholds = thresholds
self.default_threshold = thresholds['*']
def apply(self, graph):
pass
python/paddle/fluid/contrib/slim/prune/__init__.py 0 → 100644
浏览文件 @ f01c9668
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed 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.
from . import pruner
from .pruner import *
from . import prune_strategy
from .prune_strategy import *
__all__ = pruner.__all__
__all__ += prune_strategy.__all__
python/paddle/fluid/contrib/slim/prune/prune_strategy.py 0 → 100644
浏览文件 @ f01c9668
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed 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.
from ..core.strategy import Strategy
from ....framework import Program, program_guard
from .... import layers
import numpy as np
__all__ = ['SensitivePruneStrategy', 'PruneStrategy']
class SensitivePruneStrategy(Strategy):
def __init__(self,
pruner=None,
start_epoch=0,
end_epoch=10,
delta_rate=0.20,
acc_loss_threshold=0.2,
sensitivities=None):
super(SensitivePruneStrategy, self).__init__(start_epoch, end_epoch)
self.pruner = pruner
self.delta_rate = delta_rate
self.acc_loss_threshold = acc_loss_threshold
self.sensitivities = sensitivities
class PruneStrategy(Strategy):
"""
The strategy that pruning weights by threshold or ratio iteratively.
"""
def __init__(self,
pruner,
mini_batch_pruning_frequency=1,
start_epoch=0,
end_epoch=10):
super(PruneStrategy, self).__init__(start_epoch, end_epoch)
self.pruner = pruner
self.mini_batch_pruning_frequency = mini_batch_pruning_frequency
def _triger(self, context):
return (context.batch_id % self.mini_batch_pruning_frequency == 0 and
self.start_epoch <= context.epoch_id < self.end_epoch)
def on_batch_end(self, context):
if self._triger(context):
prune_program = Program()
with program_guard(prune_program):
for param in context.graph.all_parameters():
prune_program.global_block().clone_variable(param)
p = prune_program.global_block().var(param.name)
zeros_mask = self.pruner.prune(p)
pruned_param = p * zeros_mask
layers.assign(input=pruned_param, output=param)
context.program_exe.run(prune_program, scope=context.scope)
python/paddle/fluid/contrib/slim/prune/pruner.py 0 → 100644
浏览文件 @ f01c9668
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed 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.
import numpy as np
from .... import layers
__all__ = ['Pruner', 'MagnitudePruner', 'RatioPruner']
class Pruner(object):
"""
Base class of all pruners.
"""
def __init__(self):
pass
def prune(self, param):
pass
class MagnitudePruner(Pruner):
"""
Pruner used to pruning a parameter by threshold.
"""
def __init__(self, threshold):
self.threshold = threshold
def prune(self, param, threshold=None):
if threshold is None:
thres = layers.fill_constant(
shape=[1], dtype='float32', value=self.threshold)
else:
thres = threshold
zeros_mask = layers.less_than(x=param, y=thres)
return zeros_mask
class RatioPruner(Pruner):
"""
Pruner used to pruning a parameter by ratio.
"""
def __init__(self, ratios=None):
"""
Args:
ratios: dict with pair (paramer_name, pruned_ratio).
"""
self.ratios = ratios
def prune(self, param, ratio=None):
"""
Args:
ratio: `ratio=40%` means pruning (1 - 40%) weights to zero.
"""
if ratio is None:
rat = self.ratios[
param.name] if param.name in self.ratios else self.ratios['*']
else:
rat = ratio
if rat < 1.0:
k = max(int(rat * np.prod(param.shape)), 1)
param_vec = layers.reshape(x=param, shape=[1, -1])
param_topk, _ = layers.topk(param_vec, k=k)
threshold = layers.slice(
param_topk, axes=[1], starts=[-1], ends=[k])
threshold = layers.reshape(x=threshold, shape=[1])
zeros_mask = layers.less_than(x=param, y=threshold)
else:
zeros_mask = layers.ones(param.shape)
return zeros_mask
python/paddle/fluid/contrib/slim/unitest/__init__.py 0 → 100644
浏览文件 @ f01c9668
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed 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.
python/paddle/fluid/contrib/slim/unitest/configs/config.yaml 0 → 100644
浏览文件 @ f01c9668
version: 1.0
include: ["./unitest/configs/pruners.yaml", "./unitest/configs/pruners_0.yaml"]
pruners:
pruner_1:
class: 'RatioPruner'
ratios:
'conv1_1.w': 0.3
'conv1_2.w': 0.4
'*': 0.9
group_dims:
'*': [1, 2, 3]
criterions:
'*': 'l1-norm'
strategies:
strategy_1:
class: 'SensitivePruneStrategy'
pruner: 'pruner_2'
start_epoch: 0
end_epoch: 10
delta_rate: 0.20
acc_loss_threshold: 0.2
sensitivities:
'conv1_1.w': 0.4
compress_pass:
class: 'CompressPass'
epoch: 100
strategies:
- strategy_1
python/paddle/fluid/contrib/slim/unitest/configs/pruners.yaml 0 → 100644
浏览文件 @ f01c9668
version: 1.0
pruners:
pruner_2:
class: 'RatioPruner'
ratios:
'conv1_1.w': 0.5
'conv1_2.w': 0.2
'*': 0.7
group_dims:
'*': [1, 2, 3]
criterions:
'*': 'l1-norm'
python/paddle/fluid/contrib/slim/unitest/configs/pruners_0.yaml 0 → 100644
浏览文件 @ f01c9668
version: 1.0
pruners:
pruner_3:
class: 'RatioPruner'
ratios:
'conv1_1.w': 0.5
'conv1_2.w': 0.2
'*': 0.7
group_dims:
'*': [1, 2, 3]
criterions:
'*': 'l1-norm'
python/paddle/fluid/contrib/slim/unitest/test_factory.py 0 → 100644
浏览文件 @ f01c9668
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed 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.
from paddle.fluid.contrib.slim import ConfigFactory
import unittest
class TestFactory(unittest.TestCase):
def test_parse(self):
factory = ConfigFactory('./unitest/configs/config.yaml')
pruner = factory.instance('pruner_1')
self.assertEquals(pruner.ratios['conv1_1.w'], 0.3)
pruner = factory.instance('pruner_2')
self.assertEquals(pruner.ratios['*'], 0.7)
strategy = factory.instance('strategy_1')
pruner = strategy.pruner
self.assertEquals(pruner.ratios['*'], 0.7)
compress_pass = factory.get_compress_pass()
self.assertEquals(compress_pass.epoch, 100)
strategy = compress_pass.strategies[0]
self.assertEquals(strategy.delta_rate, 0.2)
if __name__ == '__main__':
unittest.main()
python/paddle/fluid/net_drawer.py
浏览文件 @ f01c9668
......@@ -26,7 +26,7 @@ logger = logging.getLogger(__name__)
logger.setLevel(logging.INFO)
try:
from .graphviz import Digraph
from .graphviz import Graph
except ImportError:
logger.info(
'Cannot import graphviz, which is required for drawing a network. This '
......@@ -112,7 +112,7 @@ def draw_graph(startup_program, main_program, **kwargs):
filename = kwargs.get("filename")
if filename == None:
filename = str(graph_id) + ".gv"
g = Digraph(
g = Graph(
name=str(graph_id),
filename=filename,
graph_attr=GRAPH_STYLE,
......
python/paddle/fluid/tests/unittests/test_transpose_mkldnn_op.py
浏览文件 @ f01c9668
......@@ -23,16 +23,6 @@ class TestTransposeMKLDNN(TestTransposeOp):
def init_op_type(self):
self.op_type = "transpose2"
self.use_mkldnn = True
self.is_test = True
return
def test_check_grad(self):
return
def test_check_grad_no_input(self):
return
def test_check_grad_no_filter(self):
return
......
python/paddle/fluid/tests/unittests/test_transpose_op.py
浏览文件 @ f01c9668
......@@ -27,7 +27,6 @@ class TestTransposeOp(OpTest):
self.attrs = {
'axis': list(self.axis),
'use_mkldnn': self.use_mkldnn,
'is_test': self.is_test,
}
self.outputs = {
'XShape': np.random.random(self.shape).astype("float32"),
......@@ -37,7 +36,6 @@ class TestTransposeOp(OpTest):
def init_op_type(self):
self.op_type = "transpose2"
self.use_mkldnn = False
self.is_test = False
def test_check_output(self):
self.check_output(no_check_set=['XShape'])
......
python/requirements.txt
浏览文件 @ f01c9668
......@@ -9,3 +9,5 @@ Pillow
nltk>=3.2.2
graphviz
six
funcsigs
pyyaml
python/setup.py.in
浏览文件 @ f01c9668
......@@ -109,6 +109,10 @@ packages=['paddle',
'paddle.fluid.contrib',
'paddle.fluid.contrib.decoder',
'paddle.fluid.contrib.quantize',
'paddle.fluid.contrib.slim',
'paddle.fluid.contrib.slim.core',
'paddle.fluid.contrib.slim.graph',
'paddle.fluid.contrib.slim.prune',
'paddle.fluid.contrib.utils',
'paddle.fluid.transpiler',
'paddle.fluid.transpiler.details']
......@@ -140,8 +144,6 @@ if '${WITH_FLUID_ONLY}'== 'OFF':
'${PADDLE_BINARY_DIR}/paddle/scripts/paddle']
package_data={'paddle.fluid': ['core' + (ext_name if os.name != 'nt' else '.pyd')]}
if os.name == 'nt':
package_data['paddle.fluid'] += ['openblas' + ext_name]
if '${WITH_FLUID_ONLY}'== 'OFF':
package_data['paddle.v2.master']=['libpaddle_master' + ext_name]
......@@ -166,11 +168,17 @@ package_data['paddle.libs']=[('libwarpctc' if os.name != 'nt' else 'warpctc') +
shutil.copy('${WARPCTC_LIBRARIES}', libs_path)
if '${WITH_MKL}' == 'ON':
shutil.copy('${MKLML_LIB}', libs_path)
shutil.copy('${MKLML_IOMP_LIB}', libs_path)
package_data['paddle.libs']+=['libmklml_intel' + ext_name,'libiomp5' + ext_name]
shutil.copy('${MKLML_SHARED_LIB}', libs_path)
shutil.copy('${MKLML_SHARED_IOMP_LIB}', libs_path)
package_data['paddle.libs']+=[('libmklml_intel' if os.name != 'nt' else 'mklml') + ext_name, ('libiomp5' if os.name != 'nt' else 'libiomp5md') + ext_name]
else:
if os.name == 'nt':
# copy the openblas.dll
shutil.copy(os.path.dirname('${CBLAS_LIBRARIES}') + '/openblas' + ext_name, libs_path)
package_data['paddle.libs'] += ['openblas' + ext_name]
if '${WITH_MKLDNN}' == 'ON':
if '${CMAKE_BUILD_TYPE}' == 'Release':
if '${CMAKE_BUILD_TYPE}' == 'Release' and os.name != 'nt':
# only change rpath in Release mode.
# TODO(typhoonzero): use install_name_tool to patch mkl libs once
# we can support mkl on mac.
......@@ -181,7 +189,7 @@ if '${WITH_MKLDNN}' == 'ON':
command = "patchelf --set-rpath '$ORIGIN/' ${MKLDNN_SHARED_LIB}"
if os.system(command) != 0:
raise Exception("patch libmkldnn.so failed, command: %s" % command)
package_data['paddle.libs']+=['libmkldnn.so.0']
package_data['paddle.libs']+=['libmkldnn.so.0' if os.name != 'nt' else ('mkldnn' + ext_name)]
shutil.copy('${MKLDNN_SHARED_LIB}', libs_path)
if '${WITH_NGRAPH}' == 'ON':
# only change rpath in Release mode,
......
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或