Skip to content

P
Paddle

PaddlePaddle / Paddle
接近 2 年 前同步成功

通知 2323
Star 20933
Fork 5424
P
Paddle
未验证 提交 f7b55de9 编写于 作者: J Jiabin Yang 提交者： GitHub
浏览文件
操作

Merge branch 'develop' into enhance_hierachical_sigmod_op

上级 af9a3301 e68c1fcd
展开全部 隐藏空白更改
内联 并排
Showing with 26898 addition and 10722 deletion
.gitignore
浏览文件 @ f7b55de9
python/paddle/fluid/tests/unittests/reader_reset_test.recordio
paddle/operators/check_t.save
paddle/operators/check_tensor.ls
paddle/operators/tensor.save
python/paddle/v2/fluid/tests/book/image_classification_resnet.inference.model/
python/paddle/v2/fluid/tests/book/image_classification_vgg.inference.model/
python/paddle/v2/fluid/tests/book/label_semantic_roles.inference.model/
paddle/fluid/operators/distributed/send_recv.proto
*.DS_Store
*.vs
build/
......@@ -28,4 +30,5 @@ third_party/
build_*
# clion workspace.
cmake-build-*
paddle/fluid/operators/distributed/send_recv.proto
model_test
AUTHORS.md
浏览文件 @ f7b55de9
......@@ -42,7 +42,9 @@
| QiJune | Jun Qi |
| qingqing01 | Qing-Qing Dang |
| reyoung | Yang Yu |
| Sand3r- | Michal Gallus |
| Superjom | Chun-Wei Yan |
| tensor-tang | Jian Tang |
| tianbingsz | Tian-Bing Xu |
| tpatejko | Tomasz Patejko |
| typhoonzero | Yi Wu |
......
CMakeLists.txt
浏览文件 @ f7b55de9
......@@ -26,6 +26,11 @@ message(STATUS "C compiler: ${CMAKE_C_COMPILER}, version: "
"${CMAKE_C_COMPILER_ID} ${CMAKE_C_COMPILER_VERSION}")
if(WIN32)
set(CMAKE_STATIC_LIBRARY_PREFIX lib)
add_definitions("/DGOOGLE_GLOG_DLL_DECL=")
set(CMAKE_C_FLAGS_DEBUG "${CMAKE_C_FLAGS_DEBUG} /bigobj /MTd")
set(CMAKE_C_FLAGS_RELEASE "${CMAKE_C_FLAGS_RELEASE} /bigobj /MT")
set(CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} /bigobj /MTd")
set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /bigobj /MT")
endif(WIN32)
if(NOT CMAKE_CROSSCOMPILING)
......@@ -41,6 +46,7 @@ option(WITH_GPU "Compile PaddlePaddle with NVIDIA GPU" ${CUDA_F
option(WITH_AMD_GPU "Compile PaddlePaddle with AMD GPU" OFF)
option(WITH_AVX "Compile PaddlePaddle with AVX intrinsics" ${AVX_FOUND})
option(WITH_MKL "Compile PaddlePaddle with MKL support." ${AVX_FOUND})
option(WITH_NGRAPH "Compile PaddlePaddle with nGraph support." OFF)
option(WITH_DSO "Compile PaddlePaddle with dynamic linked CUDA" ON)
option(WITH_TESTING "Compile PaddlePaddle with unit testing" OFF)
option(WITH_SWIG_PY "Compile PaddlePaddle with inference api" ON)
......@@ -62,10 +68,11 @@ option(WITH_DISTRIBUTE "Compile with distributed support" OFF)
option(USE_EIGEN_FOR_BLAS "Use matrix multiplication in Eigen" OFF)
option(EIGEN_USE_THREADS "Compile with multi-threaded Eigen" OFF)
option(WITH_ARM_FP16 "Use half precision support on armv8.2-a cpu" OFF)
option(WITH_FAST_BUNDLE_TEST "Bundle tests that can be run in a single process together to reduce launch overhead" OFF)
option(WITH_CONTRIB "Compile the third-party contributation" OFF)
option(REPLACE_ENFORCE_GLOG "Replace PADDLE_ENFORCE with glog/CHECK for better debug." OFF)
option(WITH_ANAKIN "Compile with Anakin library" OFF)
option(ANAKIN_BUILD_FAT_BIN "Build anakin cuda fat-bin lib for all device plantform, ignored when WITH_ANAKIN=OFF" OFF)
option(ANAKIN_BUILD_CROSS_PLANTFORM "Build anakin lib for any nvidia device plantform. ignored when WITH_ANAKIN=OFF" ON)
option(WITH_GRPC "Use grpc as the default rpc framework" ${WITH_DISTRIBUTE})
option(WITH_BRPC_RDMA "Use brpc rdma as the rpc protocal" OFF)
option(ON_INFER "Turn on inference optimization." OFF)
......@@ -104,6 +111,8 @@ if(ANDROID OR IOS)
"Disable RDMA when cross-compiling for Android and iOS" FORCE)
set(WITH_MKL OFF CACHE STRING
"Disable MKL when cross-compiling for Android and iOS" FORCE)
set(WITH_NGRAPH OFF CACHE STRING
"Disable nGraph when cross-compiling for Android and iOS" FORCE)
set(WITH_GOLANG OFF CACHE STRING
"Disable golang when cross-compiling for Android and iOS" FORCE)
......@@ -172,6 +181,7 @@ include(external/protobuf) # download, build, install protobuf
include(external/python) # download, build, install python
include(external/openblas) # download, build, install openblas
include(external/mkldnn) # download, build, install mkldnn
include(external/ngraph) # download, build, install nGraph
include(external/swig) # download, build, install swig
include(external/boost) # download boost
include(external/any) # download libn::any
......@@ -292,6 +302,14 @@ set(PADDLE_PYTHON_BUILD_DIR "${CMAKE_CURRENT_BINARY_DIR}/python/build")
set(CMAKE_CXX_FLAGS_RELWITHDEBINFO "-O3 -g -DNDEBUG")
set(CMAKE_C_FLAGS_RELWITHDEBINFO "-O3 -g -DNDEBUG")
if (ON_INFER)
message(STATUS "On inference mode, will take place some specific optimization.")
add_definitions(-DPADDLE_ON_INFERENCE)
else()
#TODO(luotao), combine this warning with `make inference_lib_dist` command.
message(WARNING "On inference mode, will take place some specific optimization. Turn on the ON_INFER flag when building inference_lib only.")
endif()
add_subdirectory(paddle)
if(WITH_PYTHON)
add_subdirectory(python)
......@@ -302,11 +320,3 @@ if(WITH_DOC)
find_python_module(recommonmark REQUIRED)
add_subdirectory(doc)
endif()
if (ON_INFER)
message(STATUS "On inference mode, will take place some specific optimization.")
add_definitions(-DPADDLE_ON_INFERENCE)
else()
#TODO(luotao), combine this warning with `make inference_lib_dist` command.
message(WARNING "On inference mode, will take place some specific optimization. Turn on the ON_INFER flag when building inference_lib only.")
endif()
cmake/configure.cmake
浏览文件 @ f7b55de9
......@@ -50,11 +50,7 @@ if(NOT WITH_PROFILER)
endif(NOT WITH_PROFILER)
if(NOT CMAKE_CROSSCOMPILING)
if(WITH_AVX AND AVX512F_FOUND)
set(SIMD_FLAG ${AVX512F_FLAG})
elseif(WITH_AVX AND AVX2_FOUND)
set(SIMD_FLAG ${AVX2_FLAG})
elseif(WITH_AVX AND AVX_FOUND)
if(WITH_AVX AND AVX_FOUND)
set(SIMD_FLAG ${AVX_FLAG})
elseif(SSE3_FOUND)
set(SIMD_FLAG ${SSE3_FLAG})
......@@ -222,3 +218,7 @@ endif(WITH_GRPC)
if(WITH_BRPC_RDMA)
add_definitions(-DPADDLE_WITH_BRPC_RDMA)
endif(WITH_BRPC_RDMA)
if(ON_INFER)
add_definitions(-DPADDLE_ON_INFERENCE)
endif(ON_INFER)
cmake/cuda.cmake
浏览文件 @ f7b55de9
......@@ -157,6 +157,9 @@ list(APPEND EXTERNAL_LIBS ${CUDA_LIBRARIES} ${CUDA_rt_LIBRARY})
if(NOT WITH_DSO)
# TODO(panyx0718): CUPTI only allows DSO?
list(APPEND EXTERNAL_LIBS ${CUDNN_LIBRARY} ${CUPTI_LIBRARY} ${CUDA_CUBLAS_LIBRARIES} ${CUDA_curand_LIBRARY} ${NCCL_LIBRARY})
if(WIN32)
set_property(GLOBAL PROPERTY CUDA_MODULES ${CUDNN_LIBRARY} ${CUDA_CUBLAS_LIBRARIES} ${CUDA_curand_LIBRARY})
endif(WIN32)
endif(NOT WITH_DSO)
# setting nvcc arch flags
......@@ -196,10 +199,12 @@ elseif(CMAKE_BUILD_TYPE STREQUAL "MinSizeRel")
list(APPEND CUDA_NVCC_FLAGS ${CMAKE_CXX_FLAGS_RELEASE})
endif()
else(NOT WIN32)
if(CMAKE_BUILD_TYPE STREQUAL "Release")
if(CMAKE_BUILD_TYPE STREQUAL "Debug")
list(APPEND CUDA_NVCC_FLAGS "-g -G")
elseif(CMAKE_BUILD_TYPE STREQUAL "Release")
list(APPEND CUDA_NVCC_FLAGS "-O3 -DNDEBUG")
else()
message(FATAL "Windows only support Release build now. Please set visual studio build type to Release, x64 build.")
message(FATAL "Windows only support Release or Debug build now. Please set visual studio build type to Release/Debug, x64 build.")
endif()
endif(NOT WIN32)
......
cmake/cudnn.cmake
浏览文件 @ f7b55de9
......@@ -2,7 +2,12 @@ if(NOT WITH_GPU)
return()
endif()
set(CUDNN_ROOT "/usr" CACHE PATH "CUDNN ROOT")
if(WIN32)
set(CUDNN_ROOT ${CUDA_TOOLKIT_ROOT_DIR})
else(WIN32)
set(CUDNN_ROOT "/usr" CACHE PATH "CUDNN ROOT")
endif(WIN32)
find_path(CUDNN_INCLUDE_DIR cudnn.h
PATHS ${CUDNN_ROOT} ${CUDNN_ROOT}/include
$ENV{CUDNN_ROOT} $ENV{CUDNN_ROOT}/include ${CUDA_TOOLKIT_INCLUDE}
......
cmake/external/anakin.cmake
浏览文件 @ f7b55de9
......@@ -58,19 +58,21 @@ ExternalProject_Add(
-DPROTOBUF_ROOT=${THIRD_PARTY_PATH}/install/protobuf
-DMKLML_ROOT=${THIRD_PARTY_PATH}/install/mklml
-DENABLE_OP_TIMER=${ANAKIN_ENABLE_OP_TIMER}
-DBUILD_FAT_BIN=${ANAKIN_BUILD_FAT_BIN}
-DBUILD_CROSS_PLANTFORM=${ANAKIN_BUILD_CROSS_PLANTFORM}
${EXTERNAL_OPTIONAL_ARGS}
CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${ANAKIN_INSTALL_DIR}
)
message(STATUS "Anakin for inference is enabled")
message(STATUS "Anakin is set INCLUDE:${ANAKIN_INCLUDE} LIBRARY:${ANAKIN_LIBRARY}")
add_dependencies(extern_anakin protobuf mklml)
add_library(anakin_shared SHARED IMPORTED GLOBAL)
set_property(TARGET anakin_shared PROPERTY IMPORTED_LOCATION ${ANAKIN_SHARED_LIB})
add_dependencies(anakin_shared extern_anakin protobuf mklml)
add_dependencies(anakin_shared extern_anakin)
add_library(anakin_saber SHARED IMPORTED GLOBAL)
set_property(TARGET anakin_saber PROPERTY IMPORTED_LOCATION ${ANAKIN_SABER_LIB})
add_dependencies(anakin_saber extern_anakin protobuf mklml)
add_dependencies(anakin_saber extern_anakin)
list(APPEND external_project_dependencies anakin_shared anakin_saber)
cmake/external/boost.cmake
浏览文件 @ f7b55de9
......@@ -28,34 +28,28 @@ if((NOT DEFINED BOOST_TAR) OR (NOT DEFINED BOOST_URL))
set(BOOST_TAR "boost_1_41_0" CACHE STRING "" FORCE)
set(BOOST_URL "http://paddlepaddledeps.cdn.bcebos.com/${BOOST_TAR}.tar.gz" CACHE STRING "" FORCE)
endif()
IF (WIN32)
MESSAGE(WARNING, "In windows, boost can not be downloaded automaticlly, please build it manually and put it at " ${THIRD_PARTY_PATH}install/boost)
else()
MESSAGE(STATUS "BOOST_TAR: ${BOOST_TAR}, BOOST_URL: ${BOOST_URL}")
ENDIF(WIN32)
MESSAGE(STATUS "BOOST_TAR: ${BOOST_TAR}, BOOST_URL: ${BOOST_URL}")
set(BOOST_SOURCES_DIR ${THIRD_PARTY_PATH}/boost)
set(BOOST_DOWNLOAD_DIR "${BOOST_SOURCES_DIR}/src/${BOOST_PROJECT}")
set(BOOST_INCLUDE_DIR "${BOOST_DOWNLOAD_DIR}/${BOOST_TAR}" CACHE PATH "boost include directory." FORCE)
set_directory_properties(PROPERTIES CLEAN_NO_CUSTOM 1)
set(BOOST_INCLUDE_DIR "${BOOST_DOWNLOAD_DIR}" CACHE PATH "boost include directory." FORCE)
set_directory_properties(PROPERTIES CLEAN_NO_CUSTOM 1)
include_directories(${BOOST_INCLUDE_DIR})
if (NOT WIN32)
ExternalProject_Add(
${BOOST_PROJECT}
${EXTERNAL_PROJECT_LOG_ARGS}
DOWNLOAD_DIR ${BOOST_DOWNLOAD_DIR}
DOWNLOAD_COMMAND wget --no-check-certificate ${BOOST_URL} -c -q -O ${BOOST_TAR}.tar.gz
&& tar zxf ${BOOST_TAR}.tar.gz
URL ${BOOST_URL}
DOWNLOAD_NO_PROGRESS 1
PREFIX ${BOOST_SOURCES_DIR}
CONFIGURE_COMMAND ""
BUILD_COMMAND ""
INSTALL_COMMAND ""
UPDATE_COMMAND ""
)
endif(NOT WIN32)
)
if (${CMAKE_VERSION} VERSION_LESS "3.3.0" OR NOT WIN32)
set(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/boost_dummy.c)
......
cmake/external/gflags.cmake
浏览文件 @ f7b55de9
......@@ -35,7 +35,12 @@ ExternalProject_Add(
CMAKE_ARGS -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
-DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
-DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
-DCMAKE_CXX_FLAGS_RELEASE=${CMAKE_CXX_FLAGS_RELEASE}
-DCMAKE_CXX_FLAGS_DEBUG=${CMAKE_CXX_FLAGS_DEBUG}
-DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}
-DCMAKE_C_FLAGS_DEBUG=${CMAKE_C_FLAGS_DEBUG}
-DCMAKE_C_FLAGS_RELEASE=${CMAKE_C_FLAGS_RELEASE}
-DBUILD_STATIC_LIBS=ON
-DCMAKE_INSTALL_PREFIX=${GFLAGS_INSTALL_DIR}
-DCMAKE_POSITION_INDEPENDENT_CODE=ON
-DBUILD_TESTING=OFF
......@@ -48,8 +53,8 @@ ExternalProject_Add(
IF(WIN32)
IF(NOT EXISTS "${GFLAGS_INSTALL_DIR}/lib/libgflags.lib")
add_custom_command(TARGET extern_gflags POST_BUILD
COMMAND cmake -E rename ${GFLAGS_INSTALL_DIR}/lib/gflags_static.lib ${GFLAGS_INSTALL_DIR}/lib/libgflags.lib
)
COMMAND cmake -E copy ${GFLAGS_INSTALL_DIR}/lib/gflags_static.lib ${GFLAGS_INSTALL_DIR}/lib/libgflags.lib
)
ENDIF()
ENDIF(WIN32)
ADD_LIBRARY(gflags STATIC IMPORTED GLOBAL)
......
cmake/external/glog.cmake
浏览文件 @ f7b55de9
......@@ -46,7 +46,11 @@ ExternalProject_Add(
CMAKE_ARGS -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
-DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
-DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
-DCMAKE_CXX_FLAGS_RELEASE=${CMAKE_CXX_FLAGS_RELEASE}
-DCMAKE_CXX_FLAGS_DEBUG=${CMAKE_CXX_FLAGS_DEBUG}
-DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}
-DCMAKE_C_FLAGS_DEBUG=${CMAKE_C_FLAGS_DEBUG}
-DCMAKE_C_FLAGS_RELEASE=${CMAKE_C_FLAGS_RELEASE}
-DCMAKE_INSTALL_PREFIX=${GLOG_INSTALL_DIR}
-DCMAKE_INSTALL_LIBDIR=${GLOG_INSTALL_DIR}/lib
-DCMAKE_POSITION_INDEPENDENT_CODE=ON
......@@ -63,7 +67,7 @@ ExternalProject_Add(
IF(WIN32)
IF(NOT EXISTS "${GLOG_INSTALL_DIR}/lib/libglog.lib")
add_custom_command(TARGET extern_glog POST_BUILD
COMMAND cmake -E rename ${GLOG_INSTALL_DIR}/lib/glog.lib ${GLOG_INSTALL_DIR}/lib/libglog.lib
COMMAND cmake -E copy ${GLOG_INSTALL_DIR}/lib/glog.lib ${GLOG_INSTALL_DIR}/lib/libglog.lib
)
ENDIF()
ENDIF(WIN32)
......
cmake/external/mkldnn.cmake
浏览文件 @ f7b55de9
......@@ -37,7 +37,6 @@ SET(CMAKE_INSTALL_RPATH_USE_LINK_PATH TRUE)
SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${MKLDNN_INSTALL_DIR}/lib")
INCLUDE_DIRECTORIES(${MKLDNN_INC_DIR}) # For MKLDNN code to include internal headers.
INCLUDE_DIRECTORIES(${THIRD_PARTY_PATH}/install) # For Paddle code to include mkldnn.h
IF(${CBLAS_PROVIDER} STREQUAL "MKLML")
SET(MKLDNN_DEPENDS ${MKLML_PROJECT})
......@@ -45,7 +44,7 @@ IF(${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")
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}")
......@@ -54,7 +53,7 @@ ExternalProject_Add(
${EXTERNAL_PROJECT_LOG_ARGS}
DEPENDS ${MKLDNN_DEPENDS}
GIT_REPOSITORY "https://github.com/01org/mkl-dnn.git"
GIT_TAG "64e03a1939e0d526aa8e9f2e3f7dc0ad8d372944"
GIT_TAG "21fb5f2af1dd14e132af4f1b79160977ee487818"
PREFIX ${MKLDNN_SOURCES_DIR}
UPDATE_COMMAND ""
CMAKE_ARGS -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
......
cmake/external/ngraph.cmake 0 → 100644
浏览文件 @ f7b55de9
# 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.
add_library(ngraph INTERFACE)
IF(WIN32 OR APPLE)
MESSAGE(WARNING
"Windows or Mac is not supported with nGraph in Paddle yet."
"Force WITH_NGRAPH=OFF")
SET(WITH_NGRAPH OFF CACHE STRING "Disable nGraph in Windows and MacOS" FORCE)
ENDIF()
IF(${WITH_NGRAPH} AND NOT ${WITH_MKLDNN})
MESSAGE(WARNING
"nGraph needs mkl-dnn to be enabled."
"Force WITH_NGRAPH=OFF")
SET(WITH_NGRAPH OFF CACHE STRING "Disable nGraph if mkl-dnn is disabled" FORCE)
ENDIF()
IF(NOT ${WITH_NGRAPH})
return()
ENDIF()
INCLUDE(ExternalProject)
SET(NGRAPH_PROJECT "extern_ngraph")
SET(NGRAPH_VERSION "0.9")
SET(NGRAPH_GIT_TAG "f9fd9d4cc318dc59dd4b68448e7fbb5f67a28bd0")
SET(NGRAPH_SOURCES_DIR ${THIRD_PARTY_PATH}/ngraph)
SET(NGRAPH_INSTALL_DIR ${THIRD_PARTY_PATH}/install/ngraph)
SET(NGRAPH_INC_DIR ${NGRAPH_INSTALL_DIR}/include)
SET(NGRAPH_SHARED_LIB_NAME libngraph.so.${NGRAPH_VERSION})
SET(NGRAPH_CPU_LIB_NAME libcpu_backend.so)
SET(NGRAPH_TBB_LIB_NAME libtbb.so.2)
SET(NGRAPH_GIT_REPO "https://github.com/NervanaSystems/ngraph.git")
ExternalProject_Add(
${NGRAPH_PROJECT}
${EXTERNAL_PROJECT_LOG_ARGS}
DEPENDS ${MKLDNN_PROJECT} ${MKLML_PROJECT}
GIT_REPOSITORY ${NGRAPH_GIT_REPO}
GIT_TAG ${NGRAPH_GIT_TAG}
PREFIX ${NGRAPH_SOURCES_DIR}
UPDATE_COMMAND ""
CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${NGRAPH_INSTALL_DIR}
CMAKE_ARGS -DNGRAPH_UNIT_TEST_ENABLE=FALSE
CMAKE_ARGS -DNGRAPH_TOOLS_ENABLE=FALSE
CMAKE_ARGS -DNGRAPH_INTERPRETER_ENABLE=FALSE
CMAKE_ARGS -DNGRAPH_DEX_ONLY=TRUE
CMAKE_ARGS -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}
CMAKE_ARGS -DMKLDNN_INCLUDE_DIR=${MKLDNN_INC_DIR}
CMAKE_ARGS -DMKLDNN_LIB_DIR=${MKLDNN_INSTALL_DIR}/lib
)
if(UNIX AND NOT APPLE)
include(GNUInstallDirs)
SET(NGRAPH_LIB_DIR ${NGRAPH_INSTALL_DIR}/${CMAKE_INSTALL_LIBDIR})
else()
SET(NGRAPH_LIB_DIR ${NGRAPH_INSTALL_DIR}/lib)
endif()
MESSAGE(STATUS "nGraph lib will be installed at: ${NGRAPH_LIB_DIR}")
SET(NGRAPH_SHARED_LIB ${NGRAPH_LIB_DIR}/${NGRAPH_SHARED_LIB_NAME})
SET(NGRAPH_CPU_LIB ${NGRAPH_LIB_DIR}/${NGRAPH_CPU_LIB_NAME})
SET(NGRAPH_TBB_LIB ${NGRAPH_LIB_DIR}/${NGRAPH_TBB_LIB_NAME})
# Workaround for nGraph expecting mklml to be in mkldnn install directory.
ExternalProject_Add_Step(
${NGRAPH_PROJECT}
PrepareMKL
COMMAND ${CMAKE_COMMAND} -E create_symlink ${MKLML_LIB} ${MKLDNN_INSTALL_DIR}/lib/libmklml_intel.so
COMMAND ${CMAKE_COMMAND} -E create_symlink ${MKLML_IOMP_LIB} ${MKLDNN_INSTALL_DIR}/lib/libiomp5.so
DEPENDEES download
DEPENDERS configure
)
add_dependencies(ngraph ${NGRAPH_PROJECT})
target_compile_definitions(ngraph INTERFACE -DPADDLE_WITH_NGRAPH)
target_include_directories(ngraph INTERFACE ${NGRAPH_INC_DIR})
target_link_libraries(ngraph INTERFACE ${NGRAPH_SHARED_LIB})
LIST(APPEND external_project_dependencies ngraph)
cmake/external/openblas.cmake
浏览文件 @ f7b55de9
......@@ -17,12 +17,8 @@ IF(USE_EIGEN_FOR_BLAS)
ENDIF(USE_EIGEN_FOR_BLAS)
INCLUDE(cblas)
# IF(WIN32 AND NOT ${CBLAS_FOUND})
IF(NOT ${CBLAS_FOUND})
INCLUDE(ExternalProject)
SET(CBLAS_SOURCES_DIR ${THIRD_PARTY_PATH}/openblas)
......@@ -34,6 +30,7 @@ IF(NOT ${CBLAS_FOUND})
CACHE FILEPATH "openblas library." FORCE)
ADD_DEFINITIONS(-DPADDLE_USE_OPENBLAS)
IF (WIN32)
SET(CBLAS_FOUND true)
MESSAGE(WARNING, "In windows, openblas only support msvc build, please build it manually and put it at " ${CBLAS_INSTALL_DIR})
......
cmake/external/protobuf.cmake
浏览文件 @ f7b55de9
......@@ -30,66 +30,61 @@ UNSET_VAR(PROTOBUF_LITE_LIBRARY)
UNSET_VAR(PROTOBUF_LIBRARY)
UNSET_VAR(PROTOBUF_INCLUDE_DIR)
UNSET_VAR(Protobuf_PROTOC_EXECUTABLE)
function(protobuf_generate_python SRCS)
# shameless copy from https://github.com/Kitware/CMake/blob/master/Modules/FindProtobuf.cmake
if(NOT ARGN)
message(SEND_ERROR "Error: PROTOBUF_GENERATE_PYTHON() called without any proto files")
return()
endif()
if(NOT COMMAND protobuf_generate_python) # before cmake 3.4, protobuf_genrerate_python is not defined.
function(protobuf_generate_python SRCS)
# shameless copy from https://github.com/Kitware/CMake/blob/master/Modules/FindProtobuf.cmake
if(NOT ARGN)
message(SEND_ERROR "Error: PROTOBUF_GENERATE_PYTHON() called without any proto files")
return()
endif()
if(PROTOBUF_GENERATE_CPP_APPEND_PATH)
# Create an include path for each file specified
foreach(FIL ${ARGN})
get_filename_component(ABS_FIL ${FIL} ABSOLUTE)
get_filename_component(ABS_PATH ${ABS_FIL} PATH)
list(FIND _protobuf_include_path ${ABS_PATH} _contains_already)
if(${_contains_already} EQUAL -1)
list(APPEND _protobuf_include_path -I ${ABS_PATH})
endif()
endforeach()
else()
set(_protobuf_include_path -I ${CMAKE_CURRENT_SOURCE_DIR})
endif()
if(DEFINED PROTOBUF_IMPORT_DIRS AND NOT DEFINED Protobuf_IMPORT_DIRS)
set(Protobuf_IMPORT_DIRS "${PROTOBUF_IMPORT_DIRS}")
endif()
if(DEFINED Protobuf_IMPORT_DIRS)
foreach(DIR ${Protobuf_IMPORT_DIRS})
get_filename_component(ABS_PATH ${DIR} ABSOLUTE)
list(FIND _protobuf_include_path ${ABS_PATH} _contains_already)
if(${_contains_already} EQUAL -1)
list(APPEND _protobuf_include_path -I ${ABS_PATH})
endif()
endforeach()
endif()
set(${SRCS})
if(PROTOBUF_GENERATE_CPP_APPEND_PATH)
# Create an include path for each file specified
foreach(FIL ${ARGN})
get_filename_component(ABS_FIL ${FIL} ABSOLUTE)
get_filename_component(FIL_WE ${FIL} NAME_WE)
if(NOT PROTOBUF_GENERATE_CPP_APPEND_PATH)
get_filename_component(FIL_DIR ${FIL} DIRECTORY)
if(FIL_DIR)
set(FIL_WE "${FIL_DIR}/${FIL_WE}")
endif()
get_filename_component(ABS_PATH ${ABS_FIL} PATH)
list(FIND _protobuf_include_path ${ABS_PATH} _contains_already)
if(${_contains_already} EQUAL -1)
list(APPEND _protobuf_include_path -I ${ABS_PATH})
endif()
endforeach()
else()
set(_protobuf_include_path -I ${CMAKE_CURRENT_SOURCE_DIR})
endif()
if(DEFINED PROTOBUF_IMPORT_DIRS AND NOT DEFINED Protobuf_IMPORT_DIRS)
set(Protobuf_IMPORT_DIRS "${PROTOBUF_IMPORT_DIRS}")
endif()
list(APPEND ${SRCS} "${CMAKE_CURRENT_BINARY_DIR}/${FIL_WE}_pb2.py")
add_custom_command(
OUTPUT "${CMAKE_CURRENT_BINARY_DIR}/${FIL_WE}_pb2.py"
COMMAND ${Protobuf_PROTOC_EXECUTABLE} --python_out ${CMAKE_CURRENT_BINARY_DIR} ${_protobuf_include_path} ${ABS_FIL}
DEPENDS ${ABS_FIL} ${Protobuf_PROTOC_EXECUTABLE}
COMMENT "Running Python protocol buffer compiler on ${FIL}"
VERBATIM )
if(DEFINED Protobuf_IMPORT_DIRS)
foreach(DIR ${Protobuf_IMPORT_DIRS})
get_filename_component(ABS_PATH ${DIR} ABSOLUTE)
list(FIND _protobuf_include_path ${ABS_PATH} _contains_already)
if(${_contains_already} EQUAL -1)
list(APPEND _protobuf_include_path -I ${ABS_PATH})
endif()
endforeach()
endif()
set(${SRCS} ${${SRCS}} PARENT_SCOPE)
endfunction()
endif()
set(${SRCS})
foreach(FIL ${ARGN})
get_filename_component(ABS_FIL ${FIL} ABSOLUTE)
get_filename_component(FIL_WE ${FIL} NAME_WE)
if(NOT PROTOBUF_GENERATE_CPP_APPEND_PATH)
get_filename_component(FIL_DIR ${FIL} DIRECTORY)
if(FIL_DIR)
set(FIL_WE "${FIL_DIR}/${FIL_WE}")
endif()
endif()
list(APPEND ${SRCS} "${CMAKE_CURRENT_BINARY_DIR}/${FIL_WE}_pb2.py")
add_custom_command(
OUTPUT "${CMAKE_CURRENT_BINARY_DIR}/${FIL_WE}_pb2.py"
COMMAND ${PROTOBUF_PROTOC_EXECUTABLE} --python_out ${CMAKE_CURRENT_BINARY_DIR} ${_protobuf_include_path} ${ABS_FIL}
DEPENDS ${ABS_FIL} ${PROTOBUF_PROTOC_EXECUTABLE}
COMMENT "Running Python protocol buffer compiler on ${FIL}"
VERBATIM )
endforeach()
set(${SRCS} ${${SRCS}} PARENT_SCOPE)
endfunction()
# Print and set the protobuf library information,
# finish this cmake process and exit from this file.
......@@ -126,6 +121,7 @@ macro(PROMPT_PROTOBUF_LIB)
# FIND_Protobuf.cmake uses `Protobuf_PROTOC_EXECUTABLE`.
# make `protobuf_generate_cpp` happy.
SET(Protobuf_PROTOC_EXECUTABLE ${PROTOBUF_PROTOC_EXECUTABLE})
FOREACH(dep ${protobuf_DEPS})
ADD_DEPENDENCIES(protobuf ${dep})
ADD_DEPENDENCIES(protobuf_lite ${dep})
......@@ -144,7 +140,6 @@ endmacro()
set(PROTOBUF_ROOT "" CACHE PATH "Folder contains protobuf")
IF (WIN32)
SET(PROTOBUF_ROOT ${THIRD_PARTY_PATH}/install/protobuf)
MESSAGE(WARNING, "In windows, protobuf only support msvc build, please build it manually and put it at " ${PROTOBUF_ROOT})
ENDIF(WIN32)
if (NOT "${PROTOBUF_ROOT}" STREQUAL "")
......@@ -192,13 +187,20 @@ FUNCTION(build_protobuf TARGET_NAME BUILD_FOR_HOST)
SET(OPTIONAL_ARGS
"-DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}"
"-DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}"
"-DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}"
"-DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}"
"-DCMAKE_C_FLAGS_DEBUG=${CMAKE_C_FLAGS_DEBUG}"
"-DCMAKE_C_FLAGS_RELEASE=${CMAKE_C_FLAGS_RELEASE}"
"-DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}"
"-DCMAKE_CXX_FLAGS_RELEASE=${CMAKE_CXX_FLAGS_RELEASE}"
"-DCMAKE_CXX_FLAGS_DEBUG=${CMAKE_CXX_FLAGS_DEBUG}"
"-Dprotobuf_WITH_ZLIB=ON"
"-DZLIB_ROOT:FILEPATH=${ZLIB_ROOT}"
${EXTERNAL_OPTIONAL_ARGS})
SET(OPTIONAL_CACHE_ARGS "-DZLIB_ROOT:STRING=${ZLIB_ROOT}")
ENDIF()
IF(WIN32)
SET(OPTIONAL_ARGS ${OPTIONAL_ARGS} "-DCMAKE_GENERATOR_PLATFORM=x64")
ENDIF()
SET(PROTOBUF_REPO "https://github.com/google/protobuf.git")
SET(PROTOBUF_TAG "9f75c5aa851cd877fb0d93ccc31b8567a6706546")
......
cmake/external/python.cmake
浏览文件 @ f7b55de9
......@@ -21,6 +21,48 @@ INCLUDE(python_module)
FIND_PACKAGE(PythonInterp ${PY_VERSION})
FIND_PACKAGE(PythonLibs ${PY_VERSION})
if(WIN32)
execute_process(COMMAND "${PYTHON_EXECUTABLE}" "-c"
"from distutils import sysconfig as s;import sys;import struct;
print(sys.prefix);
print(s.get_config_var('LDVERSION') or s.get_config_var('VERSION'));
"
RESULT_VARIABLE _PYTHON_SUCCESS
OUTPUT_VARIABLE _PYTHON_VALUES
ERROR_VARIABLE _PYTHON_ERROR_VALUE)
if(NOT _PYTHON_SUCCESS MATCHES 0)
set(PYTHONLIBS_FOUND FALSE)
return()
endif()
# Convert the process output into a list
string(REGEX REPLACE ";" "\\\\;" _PYTHON_VALUES ${_PYTHON_VALUES})
string(REGEX REPLACE "\n" ";" _PYTHON_VALUES ${_PYTHON_VALUES})
list(GET _PYTHON_VALUES 0 PYTHON_PREFIX)
list(GET _PYTHON_VALUES 1 PYTHON_LIBRARY_SUFFIX)
# Make sure all directory separators are '/'
string(REGEX REPLACE "\\\\" "/" PYTHON_PREFIX ${PYTHON_PREFIX})
set(PYTHON_LIBRARY
"${PYTHON_PREFIX}/libs/Python${PYTHON_LIBRARY_SUFFIX}.lib")
# when run in a venv, PYTHON_PREFIX points to it. But the libraries remain in the
# original python installation. They may be found relative to PYTHON_INCLUDE_DIR.
if(NOT EXISTS "${PYTHON_LIBRARY}")
get_filename_component(_PYTHON_ROOT ${PYTHON_INCLUDE_DIR} DIRECTORY)
set(PYTHON_LIBRARY
"${_PYTHON_ROOT}/libs/Python${PYTHON_LIBRARY_SUFFIX}.lib")
endif()
# raise an error if the python libs are still not found.
if(NOT EXISTS "${PYTHON_LIBRARY}")
message(FATAL_ERROR "Python libraries not found")
endif()
SET(PYTHON_LIBRARIES "${PYTHON_LIBRARY}")
endif(WIN32)
# Fixme: Maybe find a static library. Get SHARED/STATIC by FIND_PACKAGE.
ADD_LIBRARY(python SHARED IMPORTED GLOBAL)
SET_PROPERTY(TARGET python PROPERTY IMPORTED_LOCATION ${PYTHON_LIBRARIES})
......
cmake/external/xxhash.cmake
浏览文件 @ f7b55de9
......@@ -14,23 +14,52 @@ ELSE()
ENDIF(APPLE)
ENDIF()
ExternalProject_Add(
extern_xxhash
${EXTERNAL_PROJECT_LOG_ARGS}
GIT_REPOSITORY "https://github.com/Cyan4973/xxHash"
GIT_TAG "v0.6.5"
PREFIX ${XXHASH_SOURCE_DIR}
DOWNLOAD_NAME "xxhash"
UPDATE_COMMAND ""
CONFIGURE_COMMAND ""
BUILD_IN_SOURCE 1
PATCH_COMMAND
BUILD_COMMAND ${BUILD_CMD}
INSTALL_COMMAND export PREFIX=${XXHASH_INSTALL_DIR}/ && make install
TEST_COMMAND ""
)
if(WIN32)
ExternalProject_Add(
extern_xxhash
${EXTERNAL_PROJECT_LOG_ARGS}
GIT_REPOSITORY "https://github.com/Cyan4973/xxHash"
GIT_TAG "v0.6.5"
PREFIX ${XXHASH_SOURCE_DIR}
DOWNLOAD_NAME "xxhash"
UPDATE_COMMAND ""
BUILD_IN_SOURCE 1
PATCH_COMMAND
CONFIGURE_COMMAND
${CMAKE_COMMAND} ${XXHASH_SOURCE_DIR}/src/extern_xxhash/cmake_unofficial
-DCMAKE_INSTALL_PREFIX:PATH=${XXHASH_INSTALL_DIR}
-DCMAKE_BUILD_TYPE:STRING=${CMAKE_BUILD_TYPE}
-DCMAKE_VERBOSE_MAKEFILE:BOOL=OFF
-DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON
-DBUILD_XXHSUM=OFF
-DCMAKE_GENERATOR_PLATFORM=x64
-DBUILD_SHARED_LIBS=OFF
${OPTIONAL_CACHE_ARGS}
TEST_COMMAND ""
)
else()
ExternalProject_Add(
extern_xxhash
${EXTERNAL_PROJECT_LOG_ARGS}
GIT_REPOSITORY "https://github.com/Cyan4973/xxHash"
GIT_TAG "v0.6.5"
PREFIX ${XXHASH_SOURCE_DIR}
DOWNLOAD_NAME "xxhash"
UPDATE_COMMAND ""
CONFIGURE_COMMAND ""
BUILD_IN_SOURCE 1
PATCH_COMMAND
BUILD_COMMAND ${BUILD_CMD}
INSTALL_COMMAND export PREFIX=${XXHASH_INSTALL_DIR}/ && make install
TEST_COMMAND ""
)
endif()
set(XXHASH_LIBRARIES "${XXHASH_INSTALL_DIR}/lib/libxxhash.a")
if (WIN32)
set(XXHASH_LIBRARIES "${XXHASH_INSTALL_DIR}/lib/xxhash.lib")
else()
set(XXHASH_LIBRARIES "${XXHASH_INSTALL_DIR}/lib/libxxhash.a")
endif ()
INCLUDE_DIRECTORIES(${XXHASH_INCLUDE_DIR})
add_library(xxhash STATIC IMPORTED GLOBAL)
......
cmake/generic.cmake
浏览文件 @ f7b55de9
......@@ -266,7 +266,11 @@ function(cc_library TARGET_NAME)
if("${cc_library_DEPS};" MATCHES "python;")
list(REMOVE_ITEM cc_library_DEPS python)
add_dependencies(${TARGET_NAME} python)
target_link_libraries(${TARGET_NAME} "-Wl,-undefined,dynamic_lookup")
if(WIN32)
target_link_libraries(${TARGET_NAME} ${PYTHON_LIBRARIES})
else()
target_link_libraries(${TARGET_NAME} "-Wl,-undefined,dynamic_lookup")
endif(WIN32)
endif()
target_link_libraries(${TARGET_NAME} ${cc_library_DEPS})
add_dependencies(${TARGET_NAME} ${cc_library_DEPS})
......@@ -288,6 +292,45 @@ function(cc_library TARGET_NAME)
endif(cc_library_SRCS)
endfunction(cc_library)
# The link operation under windows may exceeds the maximum characters limit, simply break the link command
# into multiple link opeartion can fix that, say
# original:
# lib /out:target.lib a.lib b.lib c.lib d.lib
# after:
# 1. lib /out:dummy_lib_1.lib a.lib b.lib
# 2. lib /out:dummy_lib_2.lib c.lib d.lib
# 1. lib /out:target.lib dummy_lib_1.lib dummy_lib_2.lib
function(sep_library TARGET_NAME)
set(options STATIC static SHARED shared)
set(oneValueArgs "")
set(multiValueArgs SRCS DEPS)
cmake_parse_arguments(sep_library "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
set(dummy_index 1)
set(dummy_offset 1)
# the dummy target would be consisted of limit size libraries
set(dummy_limit 50)
list(LENGTH sep_library_DEPS sep_all_len)
foreach(v ${sep_library_DEPS})
list(APPEND dummy_list ${v})
list(LENGTH dummy_list listlen )
if ((${listlen} GREATER ${dummy_limit}) OR (${dummy_offset} EQUAL ${sep_all_len}))
message("create dummy library ${TARGET_NAME}_dummy_lib_${dummy_index} for ${TARGET_NAME}")
cc_library(${TARGET_NAME}_dummy_lib_${dummy_index} STATIC DEPS ${dummy_list})
foreach(i ${dummy_list})
list(REMOVE_AT dummy_list 0)
endforeach()
list(APPEND ${TARGET_NAME}_dummy_list ${TARGET_NAME}_dummy_lib_${dummy_index})
MATH(EXPR dummy_index "${dummy_index}+1")
endif()
MATH(EXPR dummy_offset "${dummy_offset}+1")
endforeach()
if(${sep_library_SHARED})
cc_library(${TARGET_NAME} SHARED SRCS ${sep_library_SRCS} DEPS ${${TARGET_NAME}_dummy_list})
else(${sep_library_SHARED})
cc_library(${TARGET_NAME} STATIC SRCS ${sep_library_SRCS} DEPS ${${TARGET_NAME}_dummy_list})
endif(${sep_library_SHARED})
endfunction(sep_library)
function(cc_binary TARGET_NAME)
set(options "")
set(oneValueArgs "")
......
cmake/inference_lib.cmake
浏览文件 @ f7b55de9
......@@ -22,175 +22,196 @@ function(copy TARGET)
list(LENGTH copy_lib_SRCS copy_lib_SRCS_len)
list(LENGTH copy_lib_DSTS copy_lib_DSTS_len)
if(NOT ${copy_lib_SRCS_len} EQUAL ${copy_lib_DSTS_len})
if (NOT ${copy_lib_SRCS_len} EQUAL ${copy_lib_DSTS_len})
message(FATAL_ERROR "${TARGET} source numbers are not equal to destination numbers")
endif()
endif ()
math(EXPR len "${copy_lib_SRCS_len} - 1")
add_custom_target(${TARGET} DEPENDS ${copy_lib_DEPS})
foreach(index RANGE ${len})
foreach (index RANGE ${len})
list(GET copy_lib_SRCS ${index} src)
list(GET copy_lib_DSTS ${index} dst)
add_custom_command(TARGET ${TARGET} PRE_BUILD
COMMAND mkdir -p "${dst}"
COMMAND cp -r "${src}" "${dst}"
COMMENT "copying ${src} -> ${dst}")
endforeach()
if (WIN32)
# windows cmd shell will not expand wildcard automatically.
# below expand the files,libs and copy them by rules.
file(GLOB header_files ${src} "*.h")
file(GLOB static_lib_files ${src} "*.lib")
file(GLOB dll_lib_files ${src} "*.dll")
set(src_files ${header_files} ${static_lib_files} ${dll_lib_files})
if (NOT "${src_files}" STREQUAL "")
list(REMOVE_DUPLICATES src_files)
endif ()
add_custom_command(TARGET ${TARGET} PRE_BUILD
COMMAND ${CMAKE_COMMAND} -E make_directory "${dst}"
)
foreach (src_file ${src_files})
add_custom_command(TARGET ${TARGET} PRE_BUILD
COMMAND ${CMAKE_COMMAND} -E copy "${src_file}" "${dst}"
COMMENT "copying ${src_file} -> ${dst}")
endforeach ()
else (WIN32) # not windows
add_custom_command(TARGET ${TARGET} PRE_BUILD
COMMAND mkdir -p "${dst}"
COMMAND cp -r "${src}" "${dst}"
COMMENT "copying ${src} -> ${dst}")
endif (WIN32) # not windows
endforeach ()
endfunction()
# third party
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/eigen3")
copy(eigen3_lib
SRCS ${EIGEN_INCLUDE_DIR}/Eigen/Core ${EIGEN_INCLUDE_DIR}/Eigen/src ${EIGEN_INCLUDE_DIR}/unsupported/Eigen
DSTS ${dst_dir}/Eigen ${dst_dir}/Eigen ${dst_dir}/unsupported
DEPS eigen3
)
SRCS ${EIGEN_INCLUDE_DIR}/Eigen/Core ${EIGEN_INCLUDE_DIR}/Eigen/src ${EIGEN_INCLUDE_DIR}/unsupported/Eigen
DSTS ${dst_dir}/Eigen ${dst_dir}/Eigen ${dst_dir}/unsupported
DEPS eigen3
)
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/gflags")
copy(gflags_lib
SRCS ${GFLAGS_INCLUDE_DIR} ${GFLAGS_LIBRARIES}
DSTS ${dst_dir} ${dst_dir}/lib
DEPS gflags
)
SRCS ${GFLAGS_INCLUDE_DIR} ${GFLAGS_LIBRARIES}
DSTS ${dst_dir} ${dst_dir}/lib
DEPS gflags
)
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/glog")
copy(glog_lib
SRCS ${GLOG_INCLUDE_DIR} ${GLOG_LIBRARIES}
DSTS ${dst_dir} ${dst_dir}/lib
DEPS glog
)
SRCS ${GLOG_INCLUDE_DIR} ${GLOG_LIBRARIES}
DSTS ${dst_dir} ${dst_dir}/lib
DEPS glog
)
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/boost/")
copy(boost_lib
SRCS ${BOOST_INCLUDE_DIR}/boost
DSTS ${dst_dir}
DEPS boost
)
SRCS ${BOOST_INCLUDE_DIR}/boost
DSTS ${dst_dir}
DEPS boost
)
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/xxhash")
copy(xxhash_lib
SRCS ${XXHASH_INCLUDE_DIR} ${XXHASH_LIBRARIES}
DSTS ${dst_dir} ${dst_dir}/lib
DEPS xxhash
)
SRCS ${XXHASH_INCLUDE_DIR} ${XXHASH_LIBRARIES}
DSTS ${dst_dir} ${dst_dir}/lib
DEPS xxhash
)
if(NOT PROTOBUF_FOUND)
if (NOT PROTOBUF_FOUND)
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/protobuf")
copy(protobuf_lib
SRCS ${PROTOBUF_INCLUDE_DIR} ${PROTOBUF_LIBRARY}
DSTS ${dst_dir} ${dst_dir}/lib
DEPS extern_protobuf
)
endif()
SRCS ${PROTOBUF_INCLUDE_DIR} ${PROTOBUF_LIBRARY}
DSTS ${dst_dir} ${dst_dir}/lib
DEPS extern_protobuf
)
endif ()
if(NOT CBLAS_FOUND)
if (NOT CBLAS_FOUND)
set(dst_dir "${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
)
SRCS ${CBLAS_INSTALL_DIR}/lib ${CBLAS_INSTALL_DIR}/include
DSTS ${dst_dir} ${dst_dir}
DEPS extern_openblas
)
elseif (WITH_MKLML)
set(dst_dir "${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 ${dst_dir}
DEPS mklml
)
endif()
if(WITH_MKLDNN)
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/mkldnn")
copy(mkldnn_lib
SRCS ${MKLDNN_INC_DIR} ${MKLDNN_SHARED_LIB}
DSTS ${dst_dir} ${dst_dir}/lib
DEPS mkldnn
)
endif()
SRCS ${MKLML_LIB} ${MKLML_IOMP_LIB} ${MKLML_INC_DIR}
DSTS ${dst_dir}/lib ${dst_dir}/lib ${dst_dir}
DEPS mklml
)
endif ()
if (WITH_MKLDNN)
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/mkldnn")
copy(mkldnn_lib
SRCS ${MKLDNN_INC_DIR} ${MKLDNN_SHARED_LIB}
DSTS ${dst_dir} ${dst_dir}/lib
DEPS mkldnn
)
endif ()
if (NOT WIN32)
if(NOT MOBILE_INFERENCE AND NOT RPI)
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/snappy")
copy(snappy_lib
SRCS ${SNAPPY_INCLUDE_DIR} ${SNAPPY_LIBRARIES}
DSTS ${dst_dir} ${dst_dir}/lib
DEPS snappy)
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/snappystream")
copy(snappystream_lib
SRCS ${SNAPPYSTREAM_INCLUDE_DIR} ${SNAPPYSTREAM_LIBRARIES}
DSTS ${dst_dir} ${dst_dir}/lib
DEPS snappystream)
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/zlib")
copy(zlib_lib
SRCS ${ZLIB_INCLUDE_DIR} ${ZLIB_LIBRARIES}
DSTS ${dst_dir} ${dst_dir}/lib
DEPS zlib)
endif()
endif(NOT WIN32)
if (NOT MOBILE_INFERENCE AND NOT RPI)
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/snappy")
copy(snappy_lib
SRCS ${SNAPPY_INCLUDE_DIR} ${SNAPPY_LIBRARIES}
DSTS ${dst_dir} ${dst_dir}/lib
DEPS snappy)
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/snappystream")
copy(snappystream_lib
SRCS ${SNAPPYSTREAM_INCLUDE_DIR} ${SNAPPYSTREAM_LIBRARIES}
DSTS ${dst_dir} ${dst_dir}/lib
DEPS snappystream)
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/zlib")
copy(zlib_lib
SRCS ${ZLIB_INCLUDE_DIR} ${ZLIB_LIBRARIES}
DSTS ${dst_dir} ${dst_dir}/lib
DEPS zlib)
endif ()
endif (NOT WIN32)
# paddle fluid module
set(src_dir "${PADDLE_SOURCE_DIR}/paddle/fluid")
set(dst_dir "${FLUID_INSTALL_DIR}/paddle/fluid")
set(module "framework")
if (NOT WIN32)
set(framework_lib_deps framework_py_proto)
endif(NOT WIN32)
set(framework_lib_deps framework_py_proto)
endif (NOT WIN32)
copy(framework_lib DEPS ${framework_lib_deps}
SRCS ${src_dir}/${module}/*.h ${src_dir}/${module}/details/*.h ${PADDLE_BINARY_DIR}/paddle/fluid/framework/framework.pb.h
${src_dir}/${module}/ir/*.h
DSTS ${dst_dir}/${module} ${dst_dir}/${module}/details ${dst_dir}/${module} ${dst_dir}/${module}/ir
)
SRCS ${src_dir}/${module}/*.h ${src_dir}/${module}/details/*.h ${PADDLE_BINARY_DIR}/paddle/fluid/framework/framework.pb.h
${src_dir}/${module}/ir/*.h
DSTS ${dst_dir}/${module} ${dst_dir}/${module}/details ${dst_dir}/${module} ${dst_dir}/${module}/ir
)
set(module "memory")
copy(memory_lib
SRCS ${src_dir}/${module}/*.h ${src_dir}/${module}/detail/*.h
DSTS ${dst_dir}/${module} ${dst_dir}/${module}/detail
)
SRCS ${src_dir}/${module}/*.h ${src_dir}/${module}/detail/*.h ${src_dir}/${module}/allocation/*.h
DSTS ${dst_dir}/${module} ${dst_dir}/${module}/detail ${dst_dir}/${module}/allocation
)
set(inference_deps paddle_fluid_shared paddle_fluid)
set(module "inference/api")
if (WITH_ANAKIN AND WITH_MKL)
copy(anakin_inference_lib DEPS paddle_inference_api inference_anakin_api
SRCS
${PADDLE_BINARY_DIR}/paddle/fluid/inference/api/libinference_anakin_api* # compiled anakin api
${ANAKIN_INSTALL_DIR} # anakin release
DSTS ${FLUID_INSTALL_DIR}/third_party/install/anakin ${FLUID_INSTALL_DIR}/third_party/install/anakin)
list(APPEND inference_deps anakin_inference_lib)
endif()
SRCS
${PADDLE_BINARY_DIR}/paddle/fluid/inference/api/libinference_anakin_api* # compiled anakin api
${ANAKIN_INSTALL_DIR} # anakin release
DSTS ${FLUID_INSTALL_DIR}/third_party/install/anakin ${FLUID_INSTALL_DIR}/third_party/install/anakin)
list(APPEND inference_deps anakin_inference_lib)
endif ()
set(module "inference")
copy(inference_lib DEPS ${inference_deps}
SRCS ${src_dir}/${module}/*.h ${PADDLE_BINARY_DIR}/paddle/fluid/inference/libpaddle_fluid.*
${src_dir}/${module}/api/paddle_inference_api.h
${src_dir}/${module}/api/paddle_*.h
${PADDLE_BINARY_DIR}/paddle/fluid/inference/api/paddle_inference_pass.h
DSTS ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module}
)
)
set(module "platform")
copy(platform_lib DEPS profiler_py_proto
SRCS ${src_dir}/${module}/*.h ${src_dir}/${module}/dynload/*.h ${src_dir}/${module}/details/*.h
DSTS ${dst_dir}/${module} ${dst_dir}/${module}/dynload ${dst_dir}/${module}/details
)
SRCS ${src_dir}/${module}/*.h ${src_dir}/${module}/dynload/*.h ${src_dir}/${module}/details/*.h
DSTS ${dst_dir}/${module} ${dst_dir}/${module}/dynload ${dst_dir}/${module}/details
)
set(module "string")
copy(string_lib
SRCS ${src_dir}/${module}/*.h ${src_dir}/${module}/tinyformat/*.h
DSTS ${dst_dir}/${module} ${dst_dir}/${module}/tinyformat
)
SRCS ${src_dir}/${module}/*.h ${src_dir}/${module}/tinyformat/*.h
DSTS ${dst_dir}/${module} ${dst_dir}/${module}/tinyformat
)
set(module "pybind")
copy(pybind_lib
SRCS ${CMAKE_CURRENT_BINARY_DIR}/paddle/fluid/${module}/pybind.h
DSTS ${dst_dir}/${module}
)
SRCS ${CMAKE_CURRENT_BINARY_DIR}/paddle/fluid/${module}/pybind.h
DSTS ${dst_dir}/${module}
)
# CMakeCache Info
copy(cmake_cache
SRCS ${CMAKE_CURRENT_BINARY_DIR}/CMakeCache.txt
DSTS ${FLUID_INSTALL_DIR})
SRCS ${CMAKE_CURRENT_BINARY_DIR}/CMakeCache.txt
DSTS ${FLUID_INSTALL_DIR})
# This command generates a complete fluid library for both train and inference
add_custom_target(fluid_lib_dist DEPENDS ${fluid_lib_dist_dep})
......@@ -198,14 +219,14 @@ add_custom_target(fluid_lib_dist DEPENDS ${fluid_lib_dist_dep})
# Following commands generate a inference-only fluid library
# third_party, version.txt and CMakeCache.txt are the same position with ${FLUID_INSTALL_DIR}
copy(third_party DEPS fluid_lib_dist
SRCS ${FLUID_INSTALL_DIR}/third_party ${FLUID_INSTALL_DIR}/CMakeCache.txt
DSTS ${FLUID_INFERENCE_INSTALL_DIR} ${FLUID_INFERENCE_INSTALL_DIR}
)
SRCS ${FLUID_INSTALL_DIR}/third_party ${FLUID_INSTALL_DIR}/CMakeCache.txt
DSTS ${FLUID_INFERENCE_INSTALL_DIR} ${FLUID_INFERENCE_INSTALL_DIR}
)
# only need libpaddle_fluid.so/a and paddle_inference_api.h for inference-only library
# only need libpaddle_fluid.so/a and paddle_*.h for inference-only library
copy(inference_api_lib DEPS fluid_lib_dist
SRCS ${FLUID_INSTALL_DIR}/paddle/fluid/inference/libpaddle_fluid.*
${FLUID_INSTALL_DIR}/paddle/fluid/inference/paddle_inference_api.h
${FLUID_INSTALL_DIR}/paddle/fluid/inference/paddle_*.h
DSTS ${FLUID_INFERENCE_INSTALL_DIR}/paddle/lib ${FLUID_INFERENCE_INSTALL_DIR}/paddle/include
)
......@@ -213,20 +234,20 @@ add_custom_target(inference_lib_dist DEPENDS third_party inference_api_lib)
# paddle fluid version
function(version version_file)
execute_process(
COMMAND ${GIT_EXECUTABLE} log --pretty=format:%H -1
WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}
OUTPUT_VARIABLE PADDLE_GIT_COMMIT)
file(WRITE ${version_file}
"GIT COMMIT ID: ${PADDLE_GIT_COMMIT}\n"
"WITH_MKL: ${WITH_MKL}\n"
"WITH_MKLDNN: ${WITH_MKLDNN}\n"
"WITH_GPU: ${WITH_GPU}\n")
if(WITH_GPU)
file(APPEND ${version_file}
"CUDA version: ${CUDA_VERSION}\n"
"CUDNN version: v${CUDNN_MAJOR_VERSION}\n")
endif()
execute_process(
COMMAND ${GIT_EXECUTABLE} log --pretty=format:%H -1
WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}
OUTPUT_VARIABLE PADDLE_GIT_COMMIT)
file(WRITE ${version_file}
"GIT COMMIT ID: ${PADDLE_GIT_COMMIT}\n"
"WITH_MKL: ${WITH_MKL}\n"
"WITH_MKLDNN: ${WITH_MKLDNN}\n"
"WITH_GPU: ${WITH_GPU}\n")
if (WITH_GPU)
file(APPEND ${version_file}
"CUDA version: ${CUDA_VERSION}\n"
"CUDNN version: v${CUDNN_MAJOR_VERSION}\n")
endif ()
endfunction()
version(${FLUID_INSTALL_DIR}/version.txt)
version(${FLUID_INFERENCE_INSTALL_DIR}/version.txt)
cmake/operators.cmake 0 → 100644
浏览文件 @ f7b55de9
set(PART_CUDA_KERNEL_FILES)
function(op_library TARGET)
# op_library is a function to create op library. The interface is same as
# cc_library. But it handle split GPU/CPU code and link some common library
# for ops.
set(cc_srcs)
set(cu_srcs)
set(hip_cu_srcs)
set(miopen_hip_cc_srcs)
set(cu_cc_srcs)
set(cudnn_cu_cc_srcs)
set(CUDNN_FILE)
set(mkldnn_cc_srcs)
set(MKLDNN_FILE)
set(op_common_deps operator op_registry math_function)
set(options "")
set(oneValueArgs "")
set(multiValueArgs SRCS DEPS)
set(pybind_flag 0)
cmake_parse_arguments(op_library "${options}" "${oneValueArgs}"
"${multiValueArgs}" ${ARGN})
list(LENGTH op_library_SRCS op_library_SRCS_len)
if (${op_library_SRCS_len} EQUAL 0)
if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.cc)
list(APPEND cc_srcs ${TARGET}.cc)
endif()
if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.cu.cc)
list(APPEND cu_cc_srcs ${TARGET}.cu.cc)
endif()
if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.cu)
list(APPEND cu_srcs ${TARGET}.cu)
endif()
if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.part.cu)
set(PART_CUDA_KERNEL_FILES ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.part.cu
${PART_CUDA_KERNEL_FILES} PARENT_SCOPE)
list(APPEND cu_srcs ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.part.cu)
endif()
if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.hip.cu)
list(APPEND hip_cu_srcs ${TARGET}.hip.cu)
endif()
string(REPLACE "_op" "_cudnn_op" CUDNN_FILE "${TARGET}")
if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${CUDNN_FILE}.cu.cc)
list(APPEND cudnn_cu_cc_srcs ${CUDNN_FILE}.cu.cc)
endif()
if(WITH_AMD_GPU)
string(REPLACE "_op" "_miopen_op" MIOPEN_FILE "${TARGET}")
if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${MIOPEN_FILE}.hip.cc)
list(APPEND miopen_hip_cc_srcs ${MIOPEN_FILE}.hip.cc)
endif()
endif()
if(WITH_MKLDNN)
string(REPLACE "_op" "_mkldnn_op" MKLDNN_FILE "${TARGET}")
if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${MKLDNN_FILE}.cc)
list(APPEND mkldnn_cc_srcs ${MKLDNN_FILE}.cc)
endif()
endif()
else()
foreach(src ${op_library_SRCS})
if (${src} MATCHES ".*\\.hip.cu$")
list(APPEND hip_cu_srcs ${src})
elseif (${src} MATCHES ".*\\.cu$")
list(APPEND cu_srcs ${src})
elseif(${src} MATCHES ".*_cudnn_op.cu.cc$")
list(APPEND cudnn_cu_cc_srcs ${src})
elseif(WITH_AMD_GPU AND ${src} MATCHES ".*_miopen_op.hip.cc$")
list(APPEND miopen_hip_cc_srcs ${src})
elseif(WITH_MKLDNN AND ${src} MATCHES ".*_mkldnn_op.cc$")
list(APPEND mkldnn_cc_srcs ${src})
elseif(${src} MATCHES ".*\\.cu.cc$")
list(APPEND cu_cc_srcs ${src})
elseif(${src} MATCHES ".*\\.cc$")
list(APPEND cc_srcs ${src})
else()
message(FATAL_ERROR "${TARGET} Source file ${src} should only be .cc or .cu")
endif()
endforeach()
endif()
list(LENGTH cc_srcs cc_srcs_len)
if (${cc_srcs_len} EQUAL 0)
message(FATAL_ERROR "The op library ${TARGET} should contains at least one .cc file")
endif()
if (WIN32)
# remove windows unsupported op, because windows has no nccl, no warpctc such ops.
foreach(windows_unsupport_op "nccl_op" "gen_nccl_id_op" "warpctc_op" "hierarchical_sigmoid_op"
"crf_decoding_op" "select_op" "lstmp_op" "gru_op" "fusion_gru_op" "lstm_op" "fusion_lstm_op" "cumsum_op"
"fusion_seqconv_eltadd_relu_op" "channel_send_op" "channel_create_op" "channel_close_op" "channel_recv_op")
if ("${TARGET}" STREQUAL "${windows_unsupport_op}")
return()
endif()
endforeach()
endif(WIN32)
set(OP_LIBRARY ${TARGET} ${OP_LIBRARY} CACHE INTERNAL "op libs")
list(LENGTH op_library_DEPS op_library_DEPS_len)
if (${op_library_DEPS_len} GREATER 0)
set(DEPS_OPS ${TARGET} ${DEPS_OPS} PARENT_SCOPE)
endif()
if (WITH_GPU)
nv_library(${TARGET} SRCS ${cc_srcs} ${cu_cc_srcs} ${cudnn_cu_cc_srcs} ${mkldnn_cc_srcs} ${cu_srcs} DEPS ${op_library_DEPS}
${op_common_deps})
elseif (WITH_AMD_GPU)
hip_library(${TARGET} SRCS ${cc_srcs} ${hip_cu_srcs} ${miopen_hip_cc_srcs} ${mkldnn_cc_srcs} DEPS ${op_library_DEPS}
${op_common_deps})
else()
cc_library(${TARGET} SRCS ${cc_srcs} ${mkldnn_cc_srcs} DEPS ${op_library_DEPS}
${op_common_deps})
endif()
# Define operators that don't need pybind here.
foreach(manual_pybind_op "compare_op" "logical_op" "nccl_op"
"tensor_array_read_write_op" "tensorrt_engine_op" "conv_fusion_op")
if ("${TARGET}" STREQUAL "${manual_pybind_op}")
set(pybind_flag 1)
endif()
endforeach()
# The registration of USE_OP, please refer to paddle/fluid/framework/op_registry.h.
# Note that it's enough to just adding one operator to pybind in a *_op.cc file.
# And for detail pybind information, please see generated paddle/pybind/pybind.h.
file(READ ${TARGET}.cc TARGET_CONTENT)
string(REGEX MATCH "REGISTER_OPERATOR\\(.*REGISTER_OPERATOR\\(" multi_register "${TARGET_CONTENT}")
string(REGEX MATCH "REGISTER_OPERATOR\\([a-z0-9_]*," one_register "${multi_register}")
if (one_register STREQUAL "")
string(REPLACE "_op" "" TARGET "${TARGET}")
else ()
string(REPLACE "REGISTER_OPERATOR(" "" TARGET "${one_register}")
string(REPLACE "," "" TARGET "${TARGET}")
endif()
# pybind USE_NO_KERNEL_OP
# HACK: if REGISTER_OP_CPU_KERNEL presents the operator must have kernel
string(REGEX MATCH "REGISTER_OP_CPU_KERNEL" regex_result "${TARGET_CONTENT}")
string(REPLACE "_op" "" TARGET "${TARGET}")
if (${pybind_flag} EQUAL 0 AND regex_result STREQUAL "")
file(APPEND ${pybind_file} "USE_NO_KERNEL_OP(${TARGET});\n")
set(pybind_flag 1)
endif()
# pybind USE_CPU_ONLY_OP
list(LENGTH cu_srcs cu_srcs_len)
list(LENGTH cu_cc_srcs cu_cc_srcs_len)
list(LENGTH mkldnn_cc_srcs mkldnn_cc_srcs_len)
list(LENGTH hip_cu_srcs hip_cu_srcs_len)
list(LENGTH miopen_hip_cc_srcs miopen_hip_cc_srcs_len)
if (${pybind_flag} EQUAL 0 AND ${mkldnn_cc_srcs_len} EQUAL 0 AND ${cu_srcs_len} EQUAL 0 AND ${cu_cc_srcs_len} EQUAL 0 AND
${hip_cu_srcs_len} EQUAL 0 AND ${miopen_hip_cc_srcs_len} EQUAL 0)
file(APPEND ${pybind_file} "USE_CPU_ONLY_OP(${TARGET});\n")
set(pybind_flag 1)
endif()
# pybind USE_OP_DEVICE_KERNEL for CUDNN
list(LENGTH cudnn_cu_cc_srcs cudnn_cu_cc_srcs_len)
if (WITH_GPU AND ${cudnn_cu_cc_srcs_len} GREATER 0)
file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, CUDNN);\n")
endif()
# pybind USE_OP_DEVICE_KERNEL for MIOPEN
if (WITH_AMD_GPU AND ${miopen_hip_cc_srcs_len} GREATER 0)
file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, MIOPEN);\n")
endif()
# pybind USE_OP_DEVICE_KERNEL for MKLDNN
if (WITH_MKLDNN AND ${mkldnn_cc_srcs_len} GREATER 0)
# Append first implemented MKLDNN activation operator
if (${MKLDNN_FILE} STREQUAL "activation_mkldnn_op")
file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(relu, MKLDNN);\n")
else()
file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, MKLDNN);\n")
endif()
endif()
# pybind USE_OP
if (${pybind_flag} EQUAL 0)
# NOTE(*): activation use macro to regist the kernels, set use_op manually.
if(${TARGET} STREQUAL "activation")
file(APPEND ${pybind_file} "USE_OP(relu);\n")
elseif(${TARGET} STREQUAL "fake_dequantize")
file(APPEND ${pybind_file} "USE_OP(fake_dequantize_max_abs);\n")
elseif(${TARGET} STREQUAL "fake_quantize")
file(APPEND ${pybind_file} "USE_OP(fake_quantize_abs_max);\n")
elseif(${TARGET} STREQUAL "tensorrt_engine_op")
message(STATUS "Pybind skips [tensorrt_engine_op], for this OP is only used in inference")
elseif(${TARGET} STREQUAL "fc")
# HACK: fc only have mkldnn and cpu, which would mismatch the cpu only condition
file(APPEND ${pybind_file} "USE_CPU_ONLY_OP(${TARGET});\n")
else()
file(APPEND ${pybind_file} "USE_OP(${TARGET});\n")
endif()
endif()
endfunction()
function(register_operators)
set(options "")
set(oneValueArgs "")
set(multiValueArgs EXCLUDES DEPS)
cmake_parse_arguments(register_operators "${options}" "${oneValueArgs}"
"${multiValueArgs}" ${ARGN})
file(GLOB OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "*_op.cc")
string(REPLACE "_mkldnn" "" OPS "${OPS}")
string(REPLACE ".cc" "" OPS "${OPS}")
list(REMOVE_DUPLICATES OPS)
list(LENGTH register_operators_DEPS register_operators_DEPS_len)
foreach(src ${OPS})
list(FIND register_operators_EXCLUDES ${src} _index)
if (${_index} EQUAL -1)
if (${register_operators_DEPS_len} GREATER 0)
op_library(${src} DEPS ${register_operators_DEPS})
else()
op_library(${src})
endif()
endif()
endforeach()
endfunction()
cmake/simd.cmake
浏览文件 @ f7b55de9
......@@ -89,7 +89,9 @@ CHECK_CXX_SOURCE_RUNS("
#include <immintrin.h>
int main()
{
__m512i a = _mm512_undefined_epi32();
__m512i a = _mm512_set_epi32 (-1, 2, -3, 4, -1, 2, -3, 4,
13, -5, 6, -7, 9, 2, -6, 3);
__m512i result = _mm512_abs_epi32 (a);
return 0;
}" AVX512F_FOUND)
......
cmake/tensorrt.cmake
浏览文件 @ f7b55de9
......@@ -34,4 +34,5 @@ if(TENSORRT_FOUND)
"Current TensorRT version is v${TENSORRT_MAJOR_VERSION}. ")
include_directories(${TENSORRT_INCLUDE_DIR})
list(APPEND EXTERNAL_LIBS ${TENSORRT_LIBRARY})
add_definitions(-DPADDLE_WITH_TENSORRT)
endif()
doc/v2/dev/contribute_to_paddle_en.md
浏览文件 @ f7b55de9
../../../CONTRIBUTING.md
\ No newline at end of file
../../../CONTRIBUTING.md
paddle/fluid/API.spec
浏览文件 @ f7b55de9
......@@ -67,8 +67,8 @@ paddle.fluid.layers.conv3d ArgSpec(args=['input', 'num_filters', 'filter_size',
paddle.fluid.layers.sequence_pool ArgSpec(args=['input', 'pool_type', 'is_test'], varargs=None, keywords=None, defaults=(False,))
paddle.fluid.layers.sequence_softmax ArgSpec(args=['input', 'use_cudnn', 'name'], varargs=None, keywords=None, defaults=(False, None))
paddle.fluid.layers.softmax ArgSpec(args=['input', 'use_cudnn', 'name'], varargs=None, keywords=None, defaults=(True, None))
paddle.fluid.layers.pool2d ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None))
paddle.fluid.layers.pool3d ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None))
paddle.fluid.layers.pool2d ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name', 'exclusive'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None, True))
paddle.fluid.layers.pool3d ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name', 'exclusive'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None, True))
paddle.fluid.layers.batch_norm ArgSpec(args=['input', 'act', 'is_test', 'momentum', 'epsilon', 'param_attr', 'bias_attr', 'data_layout', 'in_place', 'name', 'moving_mean_name', 'moving_variance_name', 'do_model_average_for_mean_and_var', 'fuse_with_relu'], varargs=None, keywords=None, defaults=(None, False, 0.9, 1e-05, None, None, 'NCHW', False, None, None, None, False, False))
paddle.fluid.layers.beam_search_decode ArgSpec(args=['ids', 'scores', 'beam_size', 'end_id', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.conv2d_transpose ArgSpec(args=['input', 'num_filters', 'output_size', 'filter_size', 'padding', 'stride', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, None, 0, 1, 1, None, None, None, True, None, None))
......@@ -93,17 +93,17 @@ paddle.fluid.layers.edit_distance ArgSpec(args=['input', 'label', 'normalized',
paddle.fluid.layers.l2_normalize ArgSpec(args=['x', 'axis', 'epsilon', 'name'], varargs=None, keywords=None, defaults=(1e-12, None))
paddle.fluid.layers.matmul ArgSpec(args=['x', 'y', 'transpose_x', 'transpose_y', 'alpha', 'name'], varargs=None, keywords=None, defaults=(False, False, 1.0, None))
paddle.fluid.layers.topk ArgSpec(args=['input', 'k', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.warpctc ArgSpec(args=['input', 'label', 'blank', 'norm_by_times'], varargs=None, keywords=None, defaults=(0, False))
paddle.fluid.layers.warpctc ArgSpec(args=['input', 'label', 'blank', 'norm_by_times', 'use_cudnn'], varargs=None, keywords=None, defaults=(0, False, False))
paddle.fluid.layers.sequence_reshape ArgSpec(args=['input', 'new_dim'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.transpose ArgSpec(args=['x', 'perm', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.im2sequence ArgSpec(args=['input', 'filter_size', 'stride', 'padding', 'input_image_size', 'out_stride', 'name'], varargs=None, keywords=None, defaults=(1, 1, 0, None, 1, None))
paddle.fluid.layers.nce ArgSpec(args=['input', 'label', 'num_total_classes', 'sample_weight', 'param_attr', 'bias_attr', 'num_neg_samples', 'name'], varargs=None, keywords=None, defaults=(None, None, None, None, None))
paddle.fluid.layers.hsigmoid ArgSpec(args=['input', 'label', 'num_classes', 'non_leaf_num', 'ptable', 'pcode', 'param_attr', 'bias_attr', 'name', 'is_costum'], varargs=None, keywords=None, defaults=(None, None, None, None, None, None, None, False))
paddle.fluid.layers.nce ArgSpec(args=['input', 'label', 'num_total_classes', 'sample_weight', 'param_attr', 'bias_attr', 'num_neg_samples', 'name', 'sampler', 'custom_dist', 'seed'], varargs=None, keywords=None, defaults=(None, None, None, None, None, 'uniform', None, 0))
paddle.fluid.layers.hsigmoid ArgSpec(args=['input', 'label', 'num_classes', 'non_leaf_num', 'ptable', 'pcode', 'param_attr', 'bias_attr', 'name', 'is_costum', 'is_sparse'], varargs=None, keywords=None, defaults=(None, None, None, None, None, None, None, False, False))
paddle.fluid.layers.beam_search ArgSpec(args=['pre_ids', 'pre_scores', 'ids', 'scores', 'beam_size', 'end_id', 'level', 'name'], varargs=None, keywords=None, defaults=(0, None))
paddle.fluid.layers.row_conv ArgSpec(args=['input', 'future_context_size', 'param_attr', 'act'], varargs=None, keywords=None, defaults=(None, None))
paddle.fluid.layers.multiplex ArgSpec(args=['inputs', 'index'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.layer_norm ArgSpec(args=['input', 'scale', 'shift', 'begin_norm_axis', 'epsilon', 'param_attr', 'bias_attr', 'act', 'name'], varargs=None, keywords=None, defaults=(True, True, 1, 1e-05, None, None, None, None))
paddle.fluid.layers.softmax_with_cross_entropy ArgSpec(args=['logits', 'label', 'soft_label', 'ignore_index'], varargs=None, keywords=None, defaults=(False, -100))
paddle.fluid.layers.softmax_with_cross_entropy ArgSpec(args=['logits', 'label', 'soft_label', 'ignore_index', 'numeric_stable_mode', 'return_softmax'], varargs=None, keywords=None, defaults=(False, -100, False, False))
paddle.fluid.layers.smooth_l1 ArgSpec(args=['x', 'y', 'inside_weight', 'outside_weight', 'sigma'], varargs=None, keywords=None, defaults=(None, None, None))
paddle.fluid.layers.one_hot ArgSpec(args=['input', 'depth'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.autoincreased_step_counter ArgSpec(args=['counter_name', 'begin', 'step'], varargs=None, keywords=None, defaults=(None, 1, 1))
......@@ -118,15 +118,17 @@ paddle.fluid.layers.label_smooth ArgSpec(args=['label', 'prior_dist', 'epsilon',
paddle.fluid.layers.roi_pool ArgSpec(args=['input', 'rois', 'pooled_height', 'pooled_width', 'spatial_scale'], varargs=None, keywords=None, defaults=(1, 1, 1.0))
paddle.fluid.layers.roi_align ArgSpec(args=['input', 'rois', 'pooled_height', 'pooled_width', 'spatial_scale', 'sampling_ratio', 'name'], varargs=None, keywords=None, defaults=(1, 1, 1.0, -1, None))
paddle.fluid.layers.dice_loss ArgSpec(args=['input', 'label', 'epsilon'], varargs=None, keywords=None, defaults=(1e-05,))
paddle.fluid.layers.image_resize ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'resample'], varargs=None, keywords=None, defaults=(None, None, None, 'BILINEAR'))
paddle.fluid.layers.image_resize ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'resample', 'actual_shape'], varargs=None, keywords=None, defaults=(None, None, None, 'BILINEAR', None))
paddle.fluid.layers.image_resize_short ArgSpec(args=['input', 'out_short_len', 'resample'], varargs=None, keywords=None, defaults=('BILINEAR',))
paddle.fluid.layers.resize_bilinear ArgSpec(args=['input', 'out_shape', 'scale', 'name'], varargs=None, keywords=None, defaults=(None, None, None))
paddle.fluid.layers.resize_bilinear ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'actual_shape'], varargs=None, keywords=None, defaults=(None, None, None, None))
paddle.fluid.layers.resize_nearest ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'actual_shape'], varargs=None, keywords=None, defaults=(None, None, None, None))
paddle.fluid.layers.gather ArgSpec(args=['input', 'index'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.scatter ArgSpec(args=['input', 'index', 'updates', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.sequence_scatter ArgSpec(args=['input', 'index', 'updates', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.random_crop ArgSpec(args=['x', 'shape', 'seed'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.mean_iou ArgSpec(args=['input', 'label', 'num_classes'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.relu ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.selu ArgSpec(args=['x', 'scale', 'alpha', 'name'], varargs=None, keywords=None, defaults=(None, None, None))
paddle.fluid.layers.log ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.crop ArgSpec(args=['x', 'shape', 'offsets', 'name'], varargs=None, keywords=None, defaults=(None, None, None))
paddle.fluid.layers.rank_loss ArgSpec(args=['label', 'left', 'right', 'name'], varargs=None, keywords=None, defaults=(None,))
......@@ -174,11 +176,16 @@ paddle.fluid.layers.mean ArgSpec(args=['x', 'name'], varargs=None, keywords=None
paddle.fluid.layers.mul ArgSpec(args=['x', 'y', 'x_num_col_dims', 'y_num_col_dims', 'name'], varargs=None, keywords=None, defaults=(1, 1, None))
paddle.fluid.layers.sigmoid_cross_entropy_with_logits ArgSpec(args=['x', 'label', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.maxout ArgSpec(args=['x', 'groups', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.space_to_depth ArgSpec(args=['x', 'blocksize', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.affine_grid ArgSpec(args=['theta', 'out_shape', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.sequence_reverse ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.affine_channel ArgSpec(args=['x', 'scale', 'bias', 'data_layout', 'name'], varargs=None, keywords=None, defaults=(None, None, 'NCHW', None))
paddle.fluid.layers.similarity_focus ArgSpec(args=['input', 'axis', 'indexes', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.hash ArgSpec(args=['input', 'hash_size', 'num_hash', 'name'], varargs=None, keywords=None, defaults=(1, None))
paddle.fluid.layers.grid_sampler ArgSpec(args=['x', 'grid', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.log_loss ArgSpec(args=['input', 'label', 'epsilon', 'name'], varargs=None, keywords=None, defaults=(0.0001, None))
paddle.fluid.layers.add_position_encoding ArgSpec(args=['input', 'alpha', 'beta', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.bilinear_tensor_product ArgSpec(args=['x', 'y', 'size', 'act', 'name', 'param_attr', 'bias_attr'], varargs=None, keywords=None, defaults=(None, None, None, None))
paddle.fluid.layers.data ArgSpec(args=['name', 'shape', 'append_batch_size', 'dtype', 'lod_level', 'type', 'stop_gradient'], varargs=None, keywords=None, defaults=(True, 'float32', 0, VarType.LOD_TENSOR, True))
paddle.fluid.layers.open_files ArgSpec(args=['filenames', 'shapes', 'lod_levels', 'dtypes', 'thread_num', 'buffer_size', 'pass_num', 'is_test'], varargs=None, keywords=None, defaults=(None, None, 1, None))
paddle.fluid.layers.read_file ArgSpec(args=['reader'], varargs=None, keywords=None, defaults=None)
......@@ -187,6 +194,7 @@ paddle.fluid.layers.batch ArgSpec(args=['reader', 'batch_size'], varargs=None, k
paddle.fluid.layers.double_buffer ArgSpec(args=['reader', 'place', 'name'], varargs=None, keywords=None, defaults=(None, None))
paddle.fluid.layers.random_data_generator ArgSpec(args=['low', 'high', 'shapes', 'lod_levels', 'for_parallel'], varargs=None, keywords=None, defaults=(True,))
paddle.fluid.layers.py_reader ArgSpec(args=['capacity', 'shapes', 'dtypes', 'lod_levels', 'name', 'use_double_buffer'], varargs=None, keywords=None, defaults=(None, None, True))
paddle.fluid.layers.create_py_reader_by_data ArgSpec(args=['capacity', 'feed_list', 'name', 'use_double_buffer'], varargs=None, keywords=None, defaults=(None, True))
paddle.fluid.layers.Preprocessor.__init__ ArgSpec(args=['self', 'reader', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.Preprocessor.block ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None)
paddle.fluid.layers.Preprocessor.inputs ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
......@@ -196,6 +204,7 @@ paddle.fluid.layers.create_tensor ArgSpec(args=['dtype', 'name', 'persistable'],
paddle.fluid.layers.create_parameter ArgSpec(args=['shape', 'dtype', 'name', 'attr', 'is_bias', 'default_initializer'], varargs=None, keywords=None, defaults=(None, None, False, None))
paddle.fluid.layers.create_global_var ArgSpec(args=['shape', 'value', 'dtype', 'persistable', 'force_cpu', 'name'], varargs=None, keywords=None, defaults=(False, False, None))
paddle.fluid.layers.cast ArgSpec(args=['x', 'dtype'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.tensor_array_to_tensor ArgSpec(args=['input', 'axis', 'name'], varargs=None, keywords=None, defaults=(1, None))
paddle.fluid.layers.concat ArgSpec(args=['input', 'axis', 'name'], varargs=None, keywords=None, defaults=(0, None))
paddle.fluid.layers.sums ArgSpec(args=['input', 'out'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.assign ArgSpec(args=['input', 'output'], varargs=None, keywords=None, defaults=(None,))
......@@ -266,6 +275,7 @@ paddle.fluid.layers.hard_shrink ArgSpec(args=['x', 'threshold'], varargs=None, k
paddle.fluid.layers.cumsum ArgSpec(args=['x', 'axis', 'exclusive', 'reverse'], varargs=None, keywords=None, defaults=(None, None, None))
paddle.fluid.layers.thresholded_relu ArgSpec(args=['x', 'threshold'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.prior_box ArgSpec(args=['input', 'image', 'min_sizes', 'max_sizes', 'aspect_ratios', 'variance', 'flip', 'clip', 'steps', 'offset', 'name', 'min_max_aspect_ratios_order'], varargs=None, keywords=None, defaults=(None, [1.0], [0.1, 0.1, 0.2, 0.2], False, False, [0.0, 0.0], 0.5, None, False))
paddle.fluid.layers.density_prior_box ArgSpec(args=['input', 'image', 'densities', 'fixed_sizes', 'fixed_ratios', 'variance', 'clip', 'steps', 'offset', 'name'], varargs=None, keywords=None, defaults=(None, None, None, [0.1, 0.1, 0.2, 0.2], False, [0.0, 0.0], 0.5, None))
paddle.fluid.layers.multi_box_head ArgSpec(args=['inputs', 'image', 'base_size', 'num_classes', 'aspect_ratios', 'min_ratio', 'max_ratio', 'min_sizes', 'max_sizes', 'steps', 'step_w', 'step_h', 'offset', 'variance', 'flip', 'clip', 'kernel_size', 'pad', 'stride', 'name', 'min_max_aspect_ratios_order'], varargs=None, keywords=None, defaults=(None, None, None, None, None, None, None, 0.5, [0.1, 0.1, 0.2, 0.2], True, False, 1, 0, 1, None, False))
paddle.fluid.layers.bipartite_match ArgSpec(args=['dist_matrix', 'match_type', 'dist_threshold', 'name'], varargs=None, keywords=None, defaults=(None, None, None))
paddle.fluid.layers.target_assign ArgSpec(args=['input', 'matched_indices', 'negative_indices', 'mismatch_value', 'name'], varargs=None, keywords=None, defaults=(None, None, None))
......
paddle/fluid/CMakeLists.txt
浏览文件 @ f7b55de9
......@@ -4,11 +4,12 @@ add_subdirectory(framework)
add_subdirectory(operators)
add_subdirectory(string)
if (NOT WIN32)
add_subdirectory(pybind)
if (NOT WIN32)
add_subdirectory(recordio)
endif(NOT WIN32)
# NOTE: please add subdirectory inference at last.
add_subdirectory(inference)
add_subdirectory(train)
paddle/fluid/framework/CMakeLists.txt
浏览文件 @ f7b55de9
......@@ -136,20 +136,32 @@ cc_library(version SRCS version.cc)
cc_test(version_test SRCS version_test.cc DEPS version)
cc_library(proto_desc SRCS var_desc.cc op_desc.cc block_desc.cc program_desc.cc DEPS shape_inference op_info operator glog version)
cc_library(ngraph_bridge SRCS ngraph_bridge.cc DEPS operator framework_proto)
if(NOT WIN32)
cc_library(ngraph_operator SRCS ngraph_operator.cc DEPS ngraph_bridge operator op_info device_context tensor scope glog
shape_inference data_transform lod_tensor profiler)
endif(NOT WIN32)
cc_library(op_registry SRCS op_registry.cc DEPS op_proto_maker op_info operator glog proto_desc)
nv_test(op_registry_test SRCS op_registry_test.cc DEPS op_registry)
if (NOT WIN32)
py_proto_compile(framework_py_proto SRCS framework.proto)
# Generate an empty __init__.py to make framework_py_proto as a valid python module.
add_custom_target(framework_py_proto_init ALL COMMAND ${CMAKE_COMMAND} -E touch __init__.py)
add_dependencies(framework_py_proto framework_py_proto_init)
add_custom_command(TARGET framework_py_proto POST_BUILD
COMMAND ${CMAKE_COMMAND} -E make_directory ${PADDLE_BINARY_DIR}/python/paddle/fluid/proto
COMMAND cp *.py ${PADDLE_BINARY_DIR}/python/paddle/fluid/proto/
COMMENT "Copy generated python proto into directory paddle/fluid/proto."
WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
if (NOT WIN32)
add_custom_command(TARGET framework_py_proto POST_BUILD
COMMAND ${CMAKE_COMMAND} -E make_directory ${PADDLE_BINARY_DIR}/python/paddle/fluid/proto
COMMAND cp *.py ${PADDLE_BINARY_DIR}/python/paddle/fluid/proto/
COMMENT "Copy generated python proto into directory paddle/fluid/proto."
WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
else(NOT WIN32)
string(REPLACE "/" "\\" proto_dstpath "${PADDLE_BINARY_DIR}/python/paddle/fluid/proto/")
add_custom_command(TARGET framework_py_proto POST_BUILD
COMMAND ${CMAKE_COMMAND} -E make_directory ${PADDLE_BINARY_DIR}/python/paddle/fluid/proto
COMMAND copy /Y *.py ${proto_dstpath}
COMMENT "Copy generated python proto into directory paddle/fluid/proto."
WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
endif(NOT WIN32)
cc_library(lod_rank_table SRCS lod_rank_table.cc DEPS lod_tensor)
......@@ -163,10 +175,14 @@ if(WITH_DISTRIBUTE)
set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
set_source_files_properties(executor.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
else()
cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass)
if(NOT WIN32)
cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass ngraph_operator)
else(NOT WIN32)
cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass)
endif(NOT WIN32)
cc_test(test_naive_executor SRCS naive_executor_test.cc DEPS naive_executor elementwise_add_op)
endif()
if (NOT WIN32)
cc_library(parallel_executor SRCS parallel_executor.cc DEPS
threaded_ssa_graph_executor scope_buffered_ssa_graph_executor
......
paddle/fluid/framework/data_device_transform.cc
浏览文件 @ f7b55de9
......@@ -18,8 +18,8 @@ namespace framework {
void TransDataDevice(const Tensor &in, const platform::Place &dst_place,
Tensor *out) {
VLOG(3) << "DeviceTransform in, src_place " << in.place()
<< " dst_place: " << dst_place;
VLOG(30) << "DeviceTransform in, src_place " << in.place()
<< " dst_place: " << dst_place;
PADDLE_ENFORCE_NE(
in.place().which(), dst_place.which(),
......
paddle/fluid/framework/data_device_transform_test.cu
浏览文件 @ f7b55de9
......@@ -17,7 +17,7 @@ limitations under the License. */
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/op_info.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/elementwise_op_function.h"
#include "paddle/fluid/operators/elementwise/elementwise_op_function.h"
#include "paddle/fluid/operators/math/math_function.h"
#include "paddle/fluid/platform/device_context.h"
#include "paddle/fluid/platform/init.h"
......@@ -49,10 +49,10 @@ class TestOpWithKernel : public OperatorWithKernel {
OpKernelType GetExpectedKernelType(
const ExecutionContext& ctx) const override {
if (Attr<bool>("use_gpu")) {
VLOG(3) << "force use gpu kernel";
VLOG(30) << "force use gpu kernel";
return OpKernelType(proto::VarType::FP32, platform::CUDAPlace(0));
} else {
VLOG(3) << "use default kernel";
VLOG(30) << "use default kernel";
return OpKernelType(proto::VarType::FP32,
ctx.Input<Tensor>("input")->place());
}
......@@ -148,7 +148,7 @@ TEST(Operator, CPUtoGPU) {
// get output
auto* output2 = scope.Var("OUT2");
gpu_op->Run(scope, cuda_place);
VLOG(3) << "after gpu_op run";
VLOG(30) << "after gpu_op run";
// auto* output2_ptr = output2->Get<LoDTensor>().data<float>();
paddle::platform::DeviceContextPool& pool =
......
paddle/fluid/framework/details/CMakeLists.txt
浏览文件 @ f7b55de9
cc_library(var_handle SRCS var_handle.cc DEPS place framework_proto node)
cc_library(op_handle_base SRCS op_handle_base.cc DEPS var_handle device_context lod_tensor)
cc_library(op_graph_view SRCS op_graph_view.cc DEPS op_handle_base)
cc_library(scale_loss_grad_op_handle SRCS scale_loss_grad_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory)
cc_library(fetch_op_handle SRCS fetch_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory)
cc_library(computation_op_handle SRCS computation_op_handle.cc DEPS framework_proto scope place operator op_registry)
......@@ -30,20 +31,25 @@ cc_library(data_balance_op_handle SRCS data_balance_op_handle.cc DEPS op_handle_
cc_library(gather_op_handle SRCS gather_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor)
cc_library(fuse_vars_op_handle SRCS fuse_vars_op_handle.cc DEPS op_handle_base scope)
if(WITH_GPU)
cc_library(modify_op_lock_and_record_event_pass SRCS modify_op_lock_and_record_event_pass.cc DEPS computation_op_handle op_graph_view multi_devices_helper)
if (WITH_GPU)
cc_library(reference_count_pass SRCS reference_count_pass.cc DEPS computation_op_handle scale_loss_grad_op_handle rpc_op_handle
all_reduce_op_handle reduce_op_handle broadcast_op_handle data_balance_op_handle graph graph_helper pass)
endif()
cc_library(sequential_execution_pass SRCS sequential_execution_pass.cc DEPS graph graph_helper pass)
cc_library(multi_devices_graph_pass SRCS multi_devices_graph_pass.cc DEPS multi_devices_helper computation_op_handle
scale_loss_grad_op_handle rpc_op_handle all_reduce_op_handle reduce_op_handle broadcast_op_handle data_balance_op_handle fused_broadcast_op_handle)
if(WITH_GPU)
cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS graph framework_proto reference_count_pass)
else()
cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS graph framework_proto)
set(SSA_GRAPH_EXECUTOR_DEPS graph framework_proto sequential_execution_pass modify_op_lock_and_record_event_pass)
if (WITH_GPU)
list(APPEND SSA_GRAPH_EXECUTOR_DEPS reference_count_pass)
endif()
cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS ${SSA_GRAPH_EXECUTOR_DEPS})
cc_library(threaded_ssa_graph_executor SRCS threaded_ssa_graph_executor.cc DEPS fetch_op_handle ssa_graph_executor scope
simple_threadpool device_context)
......
paddle/fluid/framework/details/broadcast_op_handle.cc
浏览文件 @ f7b55de9
......@@ -60,7 +60,7 @@ void BroadcastOpHandle::BroadcastOneVar(
PADDLE_ENFORCE_NOT_NULL(in_var);
Tensor &in_tensor = VariableVisitor::GetMutableTensor(in_var);
if (UNLIKELY(!in_tensor.IsInitialized())) {
VLOG(3) << "in var " << in_var_handle.name_ << "not inited, return!";
VLOG(30) << "in var " << in_var_handle.name_ << "not inited, return!";
return;
}
......
paddle/fluid/framework/details/broadcast_op_handle_test.h
浏览文件 @ f7b55de9
......@@ -37,8 +37,9 @@ struct TestBroadcastOpHandle {
std::vector<Scope*> local_scopes_;
std::vector<Scope*> param_scopes_;
Scope g_scope_;
std::unique_ptr<OpHandleBase> op_handle_;
std::vector<std::unique_ptr<VarHandleBase>> vars_;
OpHandleBase* op_handle_;
std::vector<VarHandleBase*> vars_;
std::vector<std::unique_ptr<ir::Node>> nodes_;
std::vector<p::Place> place_list_;
bool use_gpu_;
#ifdef PADDLE_WITH_CUDA
......@@ -90,6 +91,7 @@ struct TestBroadcastOpHandle {
}
void InitBroadcastOp(size_t input_scope_idx) {
nodes_.clear();
for (size_t j = 0; j < place_list_.size(); ++j) {
local_scopes_.push_back(&(g_scope_.NewScope()));
Scope& local_scope = local_scopes_.back()->NewScope();
......@@ -101,39 +103,39 @@ struct TestBroadcastOpHandle {
}
param_scopes_[input_scope_idx]->Var("input");
std::unique_ptr<ir::Node> n =
ir::CreateNodeForTest("node0", ir::Node::Type::kOperation);
nodes_.emplace_back(
ir::CreateNodeForTest("node0", ir::Node::Type::kOperation));
if (use_gpu_) {
#ifdef PADDLE_WITH_CUDA
op_handle_.reset(new BroadcastOpHandle(n.get(), local_scopes_,
place_list_, nccl_ctxs_.get()));
op_handle_ = new BroadcastOpHandle(nodes_.back().get(), local_scopes_,
place_list_, nccl_ctxs_.get());
#else
PADDLE_THROW("CUDA is not support.");
#endif
} else {
#ifdef PADDLE_WITH_CUDA
op_handle_.reset(new BroadcastOpHandle(n.get(), local_scopes_,
place_list_, nccl_ctxs_.get()));
op_handle_ = new BroadcastOpHandle(nodes_.back().get(), local_scopes_,
place_list_, nccl_ctxs_.get());
#else
op_handle_.reset(
new BroadcastOpHandle(n.get(), local_scopes_, place_list_));
op_handle_ = new BroadcastOpHandle(nodes_.back().get(), local_scopes_,
place_list_);
#endif
}
std::unique_ptr<ir::Node> v =
ir::CreateNodeForTest("node1", ir::Node::Type::kVariable);
auto* in_var_handle = new VarHandle(v.get(), 1, input_scope_idx, "input",
place_list_[input_scope_idx]);
nodes_.emplace_back(
ir::CreateNodeForTest("node1", ir::Node::Type::kVariable));
auto* in_var_handle = new VarHandle(nodes_.back().get(), 1, input_scope_idx,
"input", place_list_[input_scope_idx]);
vars_.emplace_back(in_var_handle);
op_handle_->AddInput(in_var_handle);
// add dummy var
std::unique_ptr<ir::Node> v2 =
ir::CreateNodeForTest("node2", ir::Node::Type::kVariable);
vars_.emplace_back(new DummyVarHandle(v2.get()));
nodes_.emplace_back(
ir::CreateNodeForTest("node2", ir::Node::Type::kVariable));
vars_.emplace_back(new DummyVarHandle(nodes_.back().get()));
DummyVarHandle* dummy_var_handle =
static_cast<DummyVarHandle*>(vars_.back().get());
static_cast<DummyVarHandle*>(vars_.back());
dummy_var_handle->ClearGeneratedOp();
op_handle_->AddInput(dummy_var_handle);
......@@ -141,20 +143,20 @@ struct TestBroadcastOpHandle {
if (!use_gpu_) {
op_handle_->SetDeviceContext(place_list_[j], ctxs_[j].get());
}
std::unique_ptr<ir::Node> v3 =
ir::CreateNodeForTest("node3", ir::Node::Type::kVariable);
nodes_.emplace_back(
ir::CreateNodeForTest("node3", ir::Node::Type::kVariable));
VarHandle* out_var_handle =
new VarHandle(v3.get(), 2, j, "out", place_list_[j]);
new VarHandle(nodes_.back().get(), 2, j, "out", place_list_[j]);
vars_.emplace_back(out_var_handle);
op_handle_->AddOutput(out_var_handle);
}
// add dummy var
std::unique_ptr<ir::Node> v4 =
ir::CreateNodeForTest("node4", ir::Node::Type::kVariable);
vars_.emplace_back(new DummyVarHandle(v4.get()));
nodes_.emplace_back(
ir::CreateNodeForTest("node4", ir::Node::Type::kVariable));
vars_.emplace_back(new DummyVarHandle(nodes_.back().get()));
DummyVarHandle* out_dummy_var_handle =
static_cast<DummyVarHandle*>(vars_.back().get());
static_cast<DummyVarHandle*>(vars_.back());
out_dummy_var_handle->ClearGeneratedOp();
op_handle_->AddOutput(out_dummy_var_handle);
}
......
paddle/fluid/framework/details/build_strategy.cc
浏览文件 @ f7b55de9
......@@ -16,6 +16,7 @@ limitations under the License. */
#include "paddle/fluid/framework/details/multi_devices_graph_check_pass.h"
#include "paddle/fluid/framework/details/multi_devices_graph_print_pass.h"
#include "paddle/fluid/framework/details/sequential_execution_pass.h"
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/graph_viz_pass.h"
......@@ -27,6 +28,10 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
public:
explicit ParallelExecutorPassBuilder(const BuildStrategy &strategy)
: ir::PassBuilder(), strategy_(strategy) {
if (strategy_.enable_sequential_execution_) {
AppendPass("sequential_execution_pass");
}
// Add a graph viz pass to record a graph.
if (!strategy_.debug_graphviz_path_.empty()) {
auto viz_pass = AppendPass("graph_viz_pass");
......@@ -64,15 +69,25 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
// Verify that the graph is correct for multi-device executor.
AppendPass("multi_devices_check_pass");
if (strategy_.remove_unnecessary_lock_) {
AppendPass("modify_op_lock_and_record_event_pass");
}
}
private:
BuildStrategy strategy_;
};
std::shared_ptr<ir::PassBuilder> BuildStrategy::CreatePassesFromStrategy()
const {
std::shared_ptr<ir::PassBuilder> BuildStrategy::CreatePassesFromStrategy(
bool finalize_strategy) const {
if (is_finalized_) {
return pass_builder_;
}
pass_builder_.reset(new ParallelExecutorPassBuilder(*this));
if (finalize_strategy) {
is_finalized_ = true;
}
return pass_builder_;
}
......@@ -86,10 +101,8 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
#else
const bool use_cuda) const {
#endif
// Create a default one if not initialized by user.
if (!pass_builder_) {
CreatePassesFromStrategy();
}
// Create a default one if not finalized by user.
CreatePassesFromStrategy(false);
std::unique_ptr<ir::Graph> graph(new ir::Graph(main_program));
......@@ -110,6 +123,11 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
pass->Erase("nccl_ctxs");
pass->SetNotOwned<platform::NCCLContextMap>("nccl_ctxs", nctx);
#endif
} else if (pass->Type() == "sequential_execution_pass") {
pass->Erase(kAllOpDescs);
pass->Set<const std::vector<OpDesc *>>(
kAllOpDescs,
new std::vector<OpDesc *>(main_program.Block(0).AllOps()));
}
graph = pass->Apply(std::move(graph));
}
......@@ -125,3 +143,5 @@ USE_PASS(multi_batch_merge_pass);
USE_PASS(multi_devices_pass);
USE_PASS(multi_devices_check_pass);
USE_PASS(multi_devices_print_pass);
USE_PASS(sequential_execution_pass);
USE_PASS(modify_op_lock_and_record_event_pass);
paddle/fluid/framework/details/build_strategy.h
浏览文件 @ f7b55de9
......@@ -69,14 +69,26 @@ struct BuildStrategy {
bool enable_data_balance_{false};
bool enable_sequential_execution_{false};
bool fuse_broadcast_op_{false};
bool remove_unnecessary_lock_{false};
// NOTE:
// Before you add new options, think if it's a general strategy that works
// with other strategy. If not, the strategy should be created through
// CreatePassesFromStrategy and the pass can be managed separately.
// User normally doesn't need to call this API.
// The PassBuilder allows for more customized insert, remove of passes
// from python side.
// A new PassBuilder is created based on configs defined above and
// passes are owned by the PassBuilder.
std::shared_ptr<ir::PassBuilder> CreatePassesFromStrategy() const;
std::shared_ptr<ir::PassBuilder> CreatePassesFromStrategy(
bool finalize_strategy) const;
bool IsFinalized() const { return is_finalized_; }
// Apply the passes built by the pass_builder_. The passes will be
// applied to the Program and output an ir::Graph.
......@@ -93,6 +105,7 @@ struct BuildStrategy {
#endif
private:
mutable bool is_finalized_ = false;
mutable std::shared_ptr<ir::PassBuilder> pass_builder_;
};
......
paddle/fluid/framework/details/computation_op_handle.cc
浏览文件 @ f7b55de9
......@@ -29,9 +29,15 @@ ComputationOpHandle::ComputationOpHandle(ir::Node *node, Scope *scope,
void ComputationOpHandle::RunImpl() {
WaitInputVarGenerated(place_);
this->RunAndRecordEvent([this] {
auto run_func = [this]() {
op_->Run(*scope_->FindVar(kLocalExecScopeName)->Get<Scope *>(), place_);
});
};
if (is_lock_and_record_event_free_) {
run_func();
} else {
this->RunAndRecordEvent(run_func);
}
}
bool ComputationOpHandle::NeedWait(VarHandleBase *in_var) {
......
paddle/fluid/framework/details/computation_op_handle.h
浏览文件 @ f7b55de9
......@@ -36,6 +36,8 @@ struct ComputationOpHandle : public OpHandleBase {
const platform::Place &GetPlace() const { return place_; }
void SetLockAndRecordEventFree(bool b) { is_lock_and_record_event_free_ = b; }
protected:
void RunImpl() override;
......@@ -45,6 +47,7 @@ struct ComputationOpHandle : public OpHandleBase {
std::unique_ptr<OperatorBase> op_;
Scope *scope_;
platform::Place place_;
bool is_lock_and_record_event_free_{false};
};
} // namespace details
} // namespace framework
......
paddle/fluid/framework/details/exception_holder.h
浏览文件 @ f7b55de9
......@@ -30,6 +30,8 @@ class ExceptionHolder {
Catch(exp);
} catch (platform::EnforceNotMet exp) {
Catch(exp);
} catch (std::exception& ex) {
LOG(FATAL) << "std::exception caught, " << ex.what();
} catch (...) {
LOG(FATAL) << "Unknown exception caught";
}
......
paddle/fluid/framework/details/execution_strategy.h
浏览文件 @ f7b55de9
......@@ -13,6 +13,7 @@
// limitations under the License.
#pragma once
#include <cstddef> // for size_t
namespace paddle {
namespace framework {
......@@ -26,6 +27,7 @@ struct ExecutionStrategy {
bool allow_op_delay_{false};
size_t num_iteration_per_drop_scope_{100};
ExecutorType type_{kDefault};
bool dry_run_{false};
};
} // namespace details
......
paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.cc
浏览文件 @ f7b55de9
......@@ -16,6 +16,7 @@
#include <vector>
#include "paddle/fluid/framework/details/fetch_op_handle.h"
#include "paddle/fluid/framework/details/multi_devices_helper.h"
#include "paddle/fluid/framework/ir/graph_helper.h"
namespace paddle {
namespace framework {
......@@ -29,16 +30,14 @@ FastThreadedSSAGraphExecutor::FastThreadedSSAGraphExecutor(
local_scopes_(local_scopes),
places_(places),
graph_(std::move(graph)),
pool_(strategy.num_threads_ +
1), // add one more thread for generate op_deps
pool_(strategy.num_threads_),
prepare_pool_(1), // add one more thread for generate op_deps
fetch_ctxs_(places) {
auto &ops = graph_->Get<details::GraphOps>("ops");
for (auto &op : ops) {
for (auto &op : ir::FilterByNodeWrapper<OpHandleBase>(*graph_)) {
int dep = static_cast<int>(op->NotReadyInputSize());
op_deps_.emplace(op.get(), dep);
op_deps_.emplace(op, dep);
if (dep == 0) {
bootstrap_ops_.emplace_back(op.get());
bootstrap_ops_.emplace_back(op);
}
}
......@@ -54,13 +53,13 @@ FeedFetchList FastThreadedSSAGraphExecutor::Run(
paddle::framework::FeedFetchList fetches;
fetches.resize(fetch_tensors.size());
std::unordered_map<std::string, std::vector<VarHandleBase *>> fetched_vars;
std::vector<std::unique_ptr<FetchOpHandle>> fetch_ops;
std::vector<FetchOpHandle *> fetch_ops;
for (auto &fetch_var_name : fetch_tensors) {
for (auto &var_map : graph_->Get<details::GraphVars>("vars")) {
auto it = var_map.find(fetch_var_name);
if (it != var_map.end()) {
fetched_vars[fetch_var_name].push_back(it->second.rbegin()->get());
fetched_vars[fetch_var_name].push_back(*it->second.rbegin());
}
}
}
......@@ -110,7 +109,10 @@ FeedFetchList FastThreadedSSAGraphExecutor::Run(
complete_q->Pop();
}
}
exception_.ReThrow();
if (exception_.IsCaught()) {
ClearFetchOp(graph_.get(), &fetch_ops);
exception_.ReThrow();
}
}
num_complete += num_comp;
}
......@@ -128,7 +130,9 @@ void FastThreadedSSAGraphExecutor::RunOpAsync(
size_t complete = 0;
while (op_to_run != nullptr) {
try {
op_to_run->Run(strategy_.use_cuda_);
if (LIKELY(!strategy_.dry_run_)) {
op_to_run->Run(strategy_.use_cuda_);
}
++complete;
} catch (...) {
exception_.Catch(std::current_exception());
......@@ -156,7 +160,7 @@ void FastThreadedSSAGraphExecutor::RunOpAsync(
});
}
void FastThreadedSSAGraphExecutor::PrepareAtomicOpDeps() {
atomic_op_deps_ = pool_.enqueue([&] {
atomic_op_deps_ = prepare_pool_.enqueue([&] {
auto *op_deps = new std::unordered_map<OpHandleBase *, std::atomic<int>>;
for (auto &pair : op_deps_) {
(*op_deps)[pair.first] = pair.second;
......
paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h
浏览文件 @ f7b55de9
......@@ -46,6 +46,7 @@ class FastThreadedSSAGraphExecutor : public SSAGraphExecutor {
std::vector<OpHandleBase *> bootstrap_ops_;
::ThreadPool pool_;
::ThreadPool prepare_pool_;
platform::DeviceContextPool fetch_ctxs_;
std::atomic<int> remaining_;
......
paddle/fluid/framework/details/fetch_op_handle.cc
浏览文件 @ f7b55de9
......@@ -28,11 +28,7 @@ FetchOpHandle::FetchOpHandle(ir::Node *node, FeedFetchList *data, size_t offset,
offset_(offset),
local_scopes_(local_scopes) {}
FetchOpHandle::~FetchOpHandle() {
for (auto *input_var : inputs_) {
input_var->RemoveOutput(this, this->Node());
}
}
FetchOpHandle::~FetchOpHandle() {}
void FetchOpHandle::RecordWaitEventOnCtx(platform::DeviceContext *waited_ctx) {
PADDLE_THROW("Nobody should wait FetchOp. Unexpceted Error");
......
paddle/fluid/framework/details/fused_broadcast_op_handle_test.cc
浏览文件 @ f7b55de9
......@@ -22,8 +22,10 @@ namespace details {
struct TestFusedBroadcastOpHandle : TestBroadcastOpHandle {
std::vector<std::string> out_varnames_;
std::vector<std::unique_ptr<ir::Node>> nodes_;
void InitFusedBroadcastOp(std::vector<size_t> input_scope_idxes) {
nodes_.clear();
// initialize scope and var
for (size_t i = 0; i < place_list_.size(); ++i) {
local_scopes_.push_back(&(g_scope_.NewScope()));
......@@ -39,41 +41,41 @@ struct TestFusedBroadcastOpHandle : TestBroadcastOpHandle {
}
// create op handle node
std::unique_ptr<ir::Node> n =
ir::CreateNodeForTest("fused_broadcast", ir::Node::Type::kOperation);
nodes_.emplace_back(
ir::CreateNodeForTest("fused_broadcast", ir::Node::Type::kOperation));
if (use_gpu_) {
#ifdef PADDLE_WITH_CUDA
op_handle_.reset(new FusedBroadcastOpHandle(
n.get(), local_scopes_, place_list_, nccl_ctxs_.get()));
op_handle_ = new FusedBroadcastOpHandle(
nodes_.back().get(), local_scopes_, place_list_, nccl_ctxs_.get());
#else
PADDLE_THROW("CUDA is not supported.");
#endif
} else {
#ifdef PADDLE_WITH_CUDA
op_handle_.reset(new FusedBroadcastOpHandle(
n.get(), local_scopes_, place_list_, nccl_ctxs_.get()));
op_handle_ = new FusedBroadcastOpHandle(
nodes_.back().get(), local_scopes_, place_list_, nccl_ctxs_.get());
#else
op_handle_.reset(
new FusedBroadcastOpHandle(n.get(), local_scopes_, place_list_));
op_handle_ = new FusedBroadcastOpHandle(nodes_.back().get(),
local_scopes_, place_list_);
#endif
}
for (size_t i = 0; i < input_scope_idxes.size(); ++i) {
// add input var handle
std::unique_ptr<ir::Node> in_node =
ir::CreateNodeForTest("in_node" + i, ir::Node::Type::kVariable);
nodes_.emplace_back(
ir::CreateNodeForTest("in_node" + i, ir::Node::Type::kVariable));
VarHandle* in_var_handle =
new VarHandle(in_node.get(), 1, input_scope_idxes[i], "in_var" + i,
place_list_[input_scope_idxes[i]]);
new VarHandle(nodes_.back().get(), 1, input_scope_idxes[i],
"in_var" + i, place_list_[input_scope_idxes[i]]);
vars_.emplace_back(in_var_handle);
op_handle_->AddInput(in_var_handle);
// add output var handle
for (size_t j = 0; j < place_list_.size(); ++j) {
std::unique_ptr<ir::Node> out_node =
ir::CreateNodeForTest("out_node" + i, ir::Node::Type::kVariable);
VarHandle* out_var_handle =
new VarHandle(out_node.get(), 2, j, "out_var" + i, place_list_[j]);
nodes_.emplace_back(
ir::CreateNodeForTest("out_node" + i, ir::Node::Type::kVariable));
VarHandle* out_var_handle = new VarHandle(
nodes_.back().get(), 2, j, "out_var" + i, place_list_[j]);
vars_.emplace_back(out_var_handle);
op_handle_->AddOutput(out_var_handle);
}
......
paddle/fluid/framework/details/gather_op_handle_test.cc
浏览文件 @ f7b55de9
......@@ -31,9 +31,10 @@ struct TestGatherOpHandle {
std::vector<Scope*> local_scopes_;
std::vector<Scope*> param_scopes_;
Scope g_scope_;
std::unique_ptr<OpHandleBase> op_handle_;
std::vector<std::unique_ptr<VarHandleBase>> vars_;
OpHandleBase* op_handle_;
std::vector<VarHandleBase*> vars_;
std::vector<p::Place> gpu_list_;
std::vector<std::unique_ptr<ir::Node>> nodes_;
void WaitAll() {
for (size_t j = 0; j < ctxs_.size(); ++j) {
......@@ -70,7 +71,7 @@ struct TestGatherOpHandle {
}
void InitGatherOp(size_t input_scope_idx) {
std::vector<std::unique_ptr<ir::Node>> nodes;
nodes_.clear();
for (size_t j = 0; j < gpu_list_.size(); ++j) {
local_scopes_.push_back(&(g_scope_.NewScope()));
Scope& local_scope = local_scopes_.back()->NewScope();
......@@ -82,44 +83,45 @@ struct TestGatherOpHandle {
}
param_scopes_[input_scope_idx]->Var("out");
nodes.emplace_back(
nodes_.emplace_back(
ir::CreateNodeForTest("node", ir::Node::Type::kOperation).release());
op_handle_.reset(
new GatherOpHandle(nodes.back().get(), local_scopes_, gpu_list_));
op_handle_ =
new GatherOpHandle(nodes_.back().get(), local_scopes_, gpu_list_);
// add input
for (size_t j = 0; j < gpu_list_.size(); ++j) {
op_handle_->SetDeviceContext(gpu_list_[j], ctxs_[j].get());
nodes.emplace_back(
nodes_.emplace_back(
ir::CreateNodeForTest("node1", ir::Node::Type::kVariable).release());
auto* in_var_handle =
new VarHandle(nodes.back().get(), 1, j, "input", gpu_list_[j]);
new VarHandle(nodes_.back().get(), 1, j, "input", gpu_list_[j]);
vars_.emplace_back(in_var_handle);
op_handle_->AddInput(in_var_handle);
}
// add dummy var
nodes.emplace_back(
nodes_.emplace_back(
ir::CreateNodeForTest("node2", ir::Node::Type::kVariable).release());
vars_.emplace_back(new DummyVarHandle(nodes.back().get()));
vars_.emplace_back(new DummyVarHandle(nodes_.back().get()));
DummyVarHandle* in_dummy_var_handle =
static_cast<DummyVarHandle*>(vars_.back().get());
static_cast<DummyVarHandle*>(vars_.back());
in_dummy_var_handle->ClearGeneratedOp();
op_handle_->AddInput(in_dummy_var_handle);
// add output
nodes.emplace_back(
nodes_.emplace_back(
ir::CreateNodeForTest("node3", ir::Node::Type::kVariable).release());
auto* out_var_handle = new VarHandle(nodes.back().get(), 2, input_scope_idx,
"out", gpu_list_[input_scope_idx]);
auto* out_var_handle =
new VarHandle(nodes_.back().get(), 2, input_scope_idx, "out",
gpu_list_[input_scope_idx]);
vars_.emplace_back(out_var_handle);
op_handle_->AddOutput(out_var_handle);
// add dummy var
nodes.emplace_back(
nodes_.emplace_back(
ir::CreateNodeForTest("node4", ir::Node::Type::kVariable).release());
vars_.emplace_back(new DummyVarHandle(nodes.back().get()));
vars_.emplace_back(new DummyVarHandle(nodes_.back().get()));
DummyVarHandle* dummy_var_handle =
static_cast<DummyVarHandle*>(vars_.back().get());
static_cast<DummyVarHandle*>(vars_.back());
op_handle_->AddOutput(dummy_var_handle);
}
......
paddle/fluid/framework/details/modify_op_lock_and_record_event_pass.cc 0 → 100644
浏览文件 @ f7b55de9
// 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/framework/details/modify_op_lock_and_record_event_pass.h"
#include "paddle/fluid/framework/details/computation_op_handle.h"
#include "paddle/fluid/framework/details/multi_devices_helper.h"
#include "paddle/fluid/framework/details/op_graph_view.h"
#include "paddle/fluid/framework/ir/graph_helper.h"
namespace paddle {
namespace framework {
namespace details {
static bool IsLockAndRecordEventFreeComputationOpHandle(
ComputationOpHandle *op, const OpGraphView &graph_view) {
if (!platform::is_gpu_place(op->GetPlace())) return false;
for (auto &pending_op : graph_view.PendingOps(op)) {
auto *tmp = dynamic_cast<ComputationOpHandle *>(pending_op);
if (tmp == nullptr || !(tmp->GetPlace() == op->GetPlace())) {
return false;
}
}
return true;
}
std::unique_ptr<ir::Graph> ModifyOpLockAndRecordEventPass::ApplyImpl(
std::unique_ptr<ir::Graph> ir_graph) const {
auto all_ops = ir::FilterByNodeWrapper<OpHandleBase>(*ir_graph);
OpGraphView graph_view(all_ops);
for (auto &op : all_ops) {
auto *compute_op = dynamic_cast<ComputationOpHandle *>(op);
if (compute_op == nullptr) continue;
bool is_lock_and_record_event_free =
IsLockAndRecordEventFreeComputationOpHandle(compute_op, graph_view);
compute_op->SetLockAndRecordEventFree(is_lock_and_record_event_free);
if (is_lock_and_record_event_free) {
VLOG(100) << "Set is_lock_and_record_event_free be true in op "
<< compute_op->DebugString();
}
}
return ir_graph;
}
} // namespace details
} // namespace framework
} // namespace paddle
REGISTER_PASS(modify_op_lock_and_record_event_pass,
paddle::framework::details::ModifyOpLockAndRecordEventPass);
paddle/fluid/framework/details/modify_op_lock_and_record_event_pass.h 0 → 100644
浏览文件 @ f7b55de9
// 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/ir/graph.h"
#include "paddle/fluid/framework/ir/pass.h"
namespace paddle {
namespace framework {
namespace details {
class ModifyOpLockAndRecordEventPass : public ir::Pass {
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
};
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/multi_devices_graph_check_pass.cc
浏览文件 @ f7b55de9
......@@ -15,6 +15,7 @@
#include "paddle/fluid/framework/details/multi_devices_graph_check_pass.h"
#include <string>
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/graph_helper.h"
namespace paddle {
namespace framework {
......@@ -36,20 +37,20 @@ bool SSAGraghBuilderWithChecker::IsValidGraph(const ir::Graph *graph) const {
for (auto &var_map : graph->Get<GraphVars>(kGraphVars)) {
for (auto &name_pair : var_map) {
for (auto &version_pair : name_pair.second) {
insert_pending_var(version_pair.get());
insert_pending_var(version_pair);
}
}
}
for (auto &var : graph->Get<GraphDepVars>(kGraphDepVars)) {
insert_pending_var(var.get());
insert_pending_var(var);
}
for (auto &op : graph->Get<GraphOps>(kGraphOps)) {
for (OpHandleBase *op : ir::FilterByNodeWrapper<OpHandleBase>(*graph)) {
if (op->Inputs().empty()) {
ready_ops.insert(op.get());
ready_ops.insert(op);
} else {
pending_ops.insert({op.get(), op.get()->NoDupInputSize()});
pending_ops.insert({op, op->NoDupInputSize()});
}
}
......@@ -89,6 +90,4 @@ bool SSAGraghBuilderWithChecker::IsValidGraph(const ir::Graph *graph) const {
REGISTER_PASS(multi_devices_check_pass,
paddle::framework::details::SSAGraghBuilderWithChecker)
.RequireGraphAttr(paddle::framework::details::kGraphVars)
.RequireGraphAttr(paddle::framework::details::kGraphDepVars)
.RequireGraphAttr(paddle::framework::details::kGraphOps)
.RequireGraphAttr(paddle::framework::details::kShardedVarDevice);
.RequireGraphAttr(paddle::framework::details::kGraphDepVars);
paddle/fluid/framework/details/multi_devices_graph_pass.cc
浏览文件 @ f7b55de9
......@@ -34,7 +34,14 @@
namespace paddle {
namespace framework {
namespace details {
namespace {
// TODO(panyx0718): Clean this up as well.
// all operators. NOTE that even we use a vector here, the operators is
// unordered.
typedef std::vector<OpHandleBase *> GraphOps;
const char kGraphOps[] = "ops";
void PolishGraphToSupportDataHazards(ir::Graph *graph) {
for (auto &var_map : graph->Get<GraphVars>(kGraphVars)) {
for (auto &name_pair : var_map) {
......@@ -92,7 +99,7 @@ VarHandle *CreateOrGetLatestVarHandle(ir::Graph *graph, ir::Node *node,
}
var_holder.emplace_back(var);
} else {
var = var_holder.rbegin()->get();
var = *var_holder.rbegin();
}
return var;
}
......@@ -154,7 +161,7 @@ void MultiDevSSAGraphBuilder::CreateOpHandleIOs(ir::Graph *result,
ir::Node *node,
size_t place_id) const {
auto p = places_[place_id];
auto *op_handle = result->Get<GraphOps>(kGraphOps).back().get();
auto *op_handle = result->Get<GraphOps>(kGraphOps).back();
op_handle->SetDeviceContext(p,
platform::DeviceContextPool::Instance().Get(p));
......@@ -303,7 +310,6 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilder::ApplyImpl(
result.Set(kGraphVars, new GraphVars(places_.size()));
result.Set(kGraphDepVars, new GraphDepVars);
result.Set(kGraphOps, new GraphOps);
result.Set(kShardedVarDevice, new ShardedVarDevice);
// find send/recv vars so that we can place the distributed training
// related op in the place 0
......@@ -317,11 +323,13 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilder::ApplyImpl(
bool is_forwarding = true;
bool is_dist_train = false;
std::unordered_map<std::string, int> sharded_var_device;
for (ir::Node *node : sorted_ops) {
if (boost::get<int>(
node->Op()->GetAttr(OpProtoAndCheckerMaker::OpRoleAttrName())) ==
static_cast<int>(OpRole::kRPC)) {
int op_dev_id = CreateRPCOp(&result, node);
int op_dev_id = CreateRPCOp(&result, node, &sharded_var_device);
PADDLE_ENFORCE(op_dev_id != -1,
"Can not schedule the RPC operator to the right place.");
if (node->Op()->Type() == "recv") {
......@@ -337,7 +345,7 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilder::ApplyImpl(
} else if (boost::get<int>(node->Op()->GetAttr(
OpProtoAndCheckerMaker::OpRoleAttrName())) ==
static_cast<int>(OpRole::kDist)) {
int op_dev_id = CreateDistTrainOp(&result, node);
int op_dev_id = CreateDistTrainOp(&result, node, &sharded_var_device);
if (node->Op()->Type() == "concat") {
auto origin_param_name = node->Op()->OutputArgumentNames()[0];
bcast_var_name_set[op_dev_id].emplace(origin_param_name);
......@@ -356,12 +364,11 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilder::ApplyImpl(
// the block.
is_forwarding = false;
} else {
int op_dev_id = GetOpDeviceID(result, node);
int op_dev_id = GetOpDeviceID(result, node, sharded_var_device);
if (op_dev_id != -1) { // This op only runs on one specific device.
CreateComputationalOp(&result, node, op_dev_id);
for (ir::Node *n : node->outputs) {
graph->Get<ShardedVarDevice>(kShardedVarDevice)
.emplace(n->Name(), op_dev_id);
sharded_var_device.emplace(n->Name(), op_dev_id);
}
} else {
// This op runs on all devices, and its output may have parameter's
......@@ -392,14 +399,13 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilder::ApplyImpl(
for (size_t i = 0; i < backward_vars.size(); i += 2) {
auto &p_name = backward_vars[i];
auto &g_name = backward_vars[i + 1];
VLOG(10) << "Bcast " << g_name << " for parameter " << p_name;
VLOG(100) << "Bcast " << g_name << " for parameter " << p_name;
switch (strategy_.reduce_) {
case BuildStrategy::ReduceStrategy::kReduce:
cur_device_id = GetAppropriateDeviceID({g_name});
CreateReduceOp(&result, g_name, cur_device_id);
graph->Get<ShardedVarDevice>(kShardedVarDevice)
.emplace(g_name, cur_device_id);
sharded_var_device.emplace(g_name, cur_device_id);
if (!is_dist_train) {
bcast_var_name_set[cur_device_id].emplace(p_name);
}
......@@ -458,7 +464,7 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilder::ApplyImpl(
* Only variables should be the leaves of graph.
*/
AddOutputToLeafOps(&result);
PADDLE_ENFORCE(!ir::HasCircle(result));
result.Erase<GraphOps>(kGraphOps);
return graph;
}
......@@ -498,7 +504,7 @@ void MultiDevSSAGraphBuilder::CreateBroadcastOp(ir::Graph *result,
result->Get<GraphOps>(kGraphOps).emplace_back(op_handle);
auto *in =
result->Get<GraphVars>(kGraphVars).at(src_dev_id).at(p_name).back().get();
result->Get<GraphVars>(kGraphVars).at(src_dev_id).at(p_name).back();
op_handle->AddInput(in);
for (size_t i = 0; i < places_.size(); ++i) {
......@@ -535,7 +541,7 @@ void MultiDevSSAGraphBuilder::CreateFusedBroadcastOp(
for (size_t dev_id = 0; dev_id < bcast_varnames.size(); ++dev_id) {
for (auto &p_name : bcast_varnames[dev_id]) {
auto *in =
result->Get<GraphVars>(kGraphVars).at(dev_id).at(p_name).back().get();
result->Get<GraphVars>(kGraphVars).at(dev_id).at(p_name).back();
op_handle->AddInput(in);
for (size_t out_dev_id = 0; out_dev_id < places_.size(); ++out_dev_id) {
auto &p = places_[out_dev_id];
......@@ -571,7 +577,7 @@ void MultiDevSSAGraphBuilder::InsertAllReduceOp(ir::Graph *result,
result->CreateEmptyNode("allreduce", ir::Node::Type::kOperation),
local_scopes_, places_));
#endif
auto *op_handle = result->Get<GraphOps>(kGraphOps).back().get();
auto *op_handle = result->Get<GraphOps>(kGraphOps).back();
for (size_t i = 0; i < places_.size(); ++i) {
auto &p = places_[i];
......@@ -579,7 +585,7 @@ void MultiDevSSAGraphBuilder::InsertAllReduceOp(ir::Graph *result,
auto &vars = result->Get<GraphVars>(kGraphVars)[i][og];
PADDLE_ENFORCE(!vars.empty());
auto &prev_grad = vars.back();
op_handle->AddInput(prev_grad.get());
op_handle->AddInput(prev_grad);
auto var =
new VarHandle(result->CreateEmptyNode(og, ir::Node::Type::kVariable),
......@@ -600,14 +606,14 @@ void MultiDevSSAGraphBuilder::InsertDataBalanceOp(
result->CreateEmptyNode("data_balance", ir::Node::Type::kOperation),
local_scopes_, places_));
#endif
auto *op_handle = result->Get<GraphOps>(kGraphOps).back().get();
auto *op_handle = result->Get<GraphOps>(kGraphOps).back();
for (size_t i = 0; i < places_.size(); ++i) {
auto &p = places_[i];
SetCommunicationContext(op_handle, p);
for (const std::string &d_name : datas) {
auto &vars = result->Get<GraphVars>(kGraphVars)[i][d_name];
PADDLE_ENFORCE(!vars.empty());
op_handle->AddInput(vars.back().get());
op_handle->AddInput(vars.back());
auto var = new VarHandle(
result->CreateEmptyNode(d_name, ir::Node::Type::kVariable),
vars.size(), i, d_name, p);
......@@ -617,8 +623,9 @@ void MultiDevSSAGraphBuilder::InsertDataBalanceOp(
}
}
int MultiDevSSAGraphBuilder::GetOpDeviceID(const ir::Graph &graph,
ir::Node *node) const {
int MultiDevSSAGraphBuilder::GetOpDeviceID(
const ir::Graph &graph, ir::Node *node,
const std::unordered_map<std::string, int> &sharded_var_device) const {
if (strategy_.reduce_ != BuildStrategy::ReduceStrategy::kReduce) {
return -1;
}
......@@ -631,16 +638,22 @@ int MultiDevSSAGraphBuilder::GetOpDeviceID(const ir::Graph &graph,
node->Op()->GetAttr(OpProtoAndCheckerMaker::OpRoleVarAttrName()));
PADDLE_ENFORCE_EQ(param_grad.size(), 2U);
int dev_id = GetVarDeviceID(graph, param_grad[1]);
int dev_id = GetVarDeviceID(graph, param_grad[1], sharded_var_device);
PADDLE_ENFORCE_NE(dev_id, -1, "dev_id should not be -1.[%s, %s, %s]",
node->Op()->Type(), param_grad[0], param_grad[1]);
return dev_id;
}
int MultiDevSSAGraphBuilder::GetVarDeviceID(const ir::Graph &graph,
const std::string &varname) const {
auto &sharded_var_device = graph.Get<ShardedVarDevice>(kShardedVarDevice);
int MultiDevSSAGraphBuilder::GetVarDeviceID(
const ir::Graph &graph, const std::string &varname,
const std::unordered_map<std::string, int> &sharded_var_device) const {
auto got = sharded_var_device.find(varname);
if (got == sharded_var_device.end()) {
auto pos = varname.find(framework::kNewGradSuffix);
if (pos != std::string::npos) {
got = sharded_var_device.find(varname.substr(0, pos));
}
}
return got == sharded_var_device.end() ? -1 : got->second;
}
......@@ -690,7 +703,7 @@ VarHandle *MultiDevSSAGraphBuilder::CreateReduceOp(ir::Graph *result,
result->CreateEmptyNode("reduce", ir::Node::Type::kOperation),
local_scopes_, places_));
#endif
auto *op_handle = result->Get<GraphOps>(kGraphOps).back().get();
auto *op_handle = result->Get<GraphOps>(kGraphOps).back();
for (size_t i = 0; i < places_.size(); ++i) {
auto &p = places_[i];
......@@ -698,7 +711,7 @@ VarHandle *MultiDevSSAGraphBuilder::CreateReduceOp(ir::Graph *result,
auto &vars = result->Get<GraphVars>(kGraphVars)[i][og];
PADDLE_ENFORCE(!vars.empty());
auto &prev_grad = vars.back();
op_handle->AddInput(prev_grad.get());
op_handle->AddInput(prev_grad);
}
auto &vars = result->Get<GraphVars>(kGraphVars)[dst_dev_id][og];
auto var =
......@@ -709,8 +722,9 @@ VarHandle *MultiDevSSAGraphBuilder::CreateReduceOp(ir::Graph *result,
return var;
}
int MultiDevSSAGraphBuilder::CreateDistTrainOp(ir::Graph *result,
ir::Node *node) const {
int MultiDevSSAGraphBuilder::CreateDistTrainOp(
ir::Graph *result, ir::Node *node,
std::unordered_map<std::string, int> *sharded_var_device) const {
int op_dev_id = -1;
std::vector<std::string> input_var_names;
std::vector<std::string> output_var_names;
......@@ -725,23 +739,22 @@ int MultiDevSSAGraphBuilder::CreateDistTrainOp(ir::Graph *result,
node->Op()->Type() == "split_selected_rows" ||
node->Op()->Type() == "split_ids") {
// TODO(paddle-dev): getting the first var is not safe.
op_dev_id = GetVarDeviceID(*result, input_var_names[0]);
op_dev_id =
GetVarDeviceID(*result, input_var_names[0], *sharded_var_device);
if (strategy_.reduce_ == BuildStrategy::ReduceStrategy::kAllReduce) {
op_dev_id = GetAppropriateDeviceID(input_var_names);
for (auto &varname : input_var_names) {
result->Get<ShardedVarDevice>(kShardedVarDevice)
.emplace(varname, op_dev_id);
sharded_var_device->emplace(varname, op_dev_id);
}
}
for (auto &varname : output_var_names) {
result->Get<ShardedVarDevice>(kShardedVarDevice)
.emplace(varname, op_dev_id);
sharded_var_device->emplace(varname, op_dev_id);
}
} else if (node->Op()->Type() == "concat") {
op_dev_id = GetVarDeviceID(*result, input_var_names[0]);
op_dev_id =
GetVarDeviceID(*result, input_var_names[0], *sharded_var_device);
for (auto &varname : output_var_names) {
result->Get<ShardedVarDevice>(kShardedVarDevice)
.emplace(varname, op_dev_id);
sharded_var_device->emplace(varname, op_dev_id);
}
} else {
LOG(ERROR) << "got unexpected dist op: " << node->Op()->Type();
......@@ -759,14 +772,14 @@ int MultiDevSSAGraphBuilder::CreateDistTrainOp(ir::Graph *result,
}
void SetOpInputsAllPlaces(ir::Graph *result, ir::Node *node, int num_places) {
auto *op_handle = result->Get<GraphOps>(kGraphOps).back().get();
auto *op_handle = result->Get<GraphOps>(kGraphOps).back();
for (ir::Node *input : node->inputs) {
VarHandle *var = nullptr;
for (int place_offset = 0; place_offset < num_places; ++place_offset) {
auto &var_holders = result->Get<GraphVars>(kGraphVars)[place_offset];
auto &var_holder = var_holders[input->Name()];
if (!var_holder.empty()) {
var = var_holder.rbegin()->get();
var = *var_holder.rbegin();
op_handle->AddInput(var);
}
}
......@@ -774,12 +787,14 @@ void SetOpInputsAllPlaces(ir::Graph *result, ir::Node *node, int num_places) {
}
// Create RPC related op handles that connects its in ops and out ops.
int MultiDevSSAGraphBuilder::CreateRPCOp(ir::Graph *result,
ir::Node *node) const {
int MultiDevSSAGraphBuilder::CreateRPCOp(
ir::Graph *result, ir::Node *node,
std::unordered_map<std::string, int> *sharded_var_device) const {
int op_dev_id = -1;
if (node->Op()->Type() == "send") {
// TODO(paddle-dev): getting the first var is not safe.
op_dev_id = GetVarDeviceID(*result, node->inputs[0]->Name());
op_dev_id =
GetVarDeviceID(*result, node->inputs[0]->Name(), *sharded_var_device);
PADDLE_ENFORCE(!ir::IsControlDepVar(*node->inputs[0]),
"This hack no longer holds, please fix.");
// the variable name which contains .block means it was splited by
......@@ -794,14 +809,12 @@ int MultiDevSSAGraphBuilder::CreateRPCOp(ir::Graph *result,
node->Op()->GetAttr(OpProtoAndCheckerMaker::OpRoleVarAttrName()));
PADDLE_ENFORCE_EQ(send_param_grad.size(), 2U);
op_dev_id = GetAppropriateDeviceID({send_param_grad[1]});
VLOG(10) << "send grad " << input_var_names[0] << " origin "
<< send_param_grad[1] << " place: " << op_dev_id;
VLOG(100) << "send grad " << input_var_names[0] << " origin "
<< send_param_grad[1] << " place: " << op_dev_id;
for (auto &varname : input_var_names) {
result->Get<ShardedVarDevice>(kShardedVarDevice)
.emplace(varname, op_dev_id);
sharded_var_device->emplace(varname, op_dev_id);
}
result->Get<ShardedVarDevice>(kShardedVarDevice)
.emplace(send_param_grad[1], op_dev_id);
sharded_var_device->emplace(send_param_grad[1], op_dev_id);
}
} else if (node->Op()->Type() == "recv") {
std::vector<std::string> output_var_names;
......@@ -811,16 +824,16 @@ int MultiDevSSAGraphBuilder::CreateRPCOp(ir::Graph *result,
auto recv_param_grad = boost::get<std::vector<std::string>>(
node->Op()->GetAttr(OpProtoAndCheckerMaker::OpRoleVarAttrName()));
if (recv_param_grad.size() == 2U) {
op_dev_id = GetVarDeviceID(*result, recv_param_grad[1]);
VLOG(10) << "recv param " << recv_param_grad[0]
<< " get grad place: " << recv_param_grad[1]
<< " place: " << op_dev_id;
op_dev_id =
GetVarDeviceID(*result, recv_param_grad[1], *sharded_var_device);
VLOG(100) << "recv param " << recv_param_grad[0]
<< " get grad place: " << recv_param_grad[1]
<< " place: " << op_dev_id;
} else {
op_dev_id = GetAppropriateDeviceID(output_var_names);
}
for (auto &varname : output_var_names) {
result->Get<ShardedVarDevice>(kShardedVarDevice)
.emplace(varname, op_dev_id);
sharded_var_device->emplace(varname, op_dev_id);
}
} else {
// send_barrier, fetch_barrier will run on place 0;
......@@ -839,7 +852,7 @@ int MultiDevSSAGraphBuilder::CreateRPCOp(ir::Graph *result,
// send_barrier, recv, fetch_barrier's inputs are deps var, get them from
// all places
auto p = places_[op_dev_id];
auto *op_handle = result->Get<GraphOps>(kGraphOps).back().get();
auto *op_handle = result->Get<GraphOps>(kGraphOps).back();
op_handle->SetDeviceContext(p,
platform::DeviceContextPool::Instance().Get(p));
......@@ -847,7 +860,8 @@ int MultiDevSSAGraphBuilder::CreateRPCOp(ir::Graph *result,
for (ir::Node *output : node->outputs) {
int outvar_dev_id = op_dev_id;
if (node->Op()->Type() == "fetch_barrier") {
outvar_dev_id = GetVarDeviceID(*result, output->Name());
outvar_dev_id =
GetVarDeviceID(*result, output->Name(), *sharded_var_device);
PADDLE_ENFORCE_NE(outvar_dev_id, -1);
}
p = places_[outvar_dev_id];
......
paddle/fluid/framework/details/multi_devices_graph_pass.h
浏览文件 @ f7b55de9
......@@ -44,12 +44,18 @@ class MultiDevSSAGraphBuilder : public ir::Pass {
mutable platform::NCCLContextMap *nccl_ctxs_;
#endif
int GetVarDeviceID(const ir::Graph &graph, const std::string &varname) const;
int GetVarDeviceID(
const ir::Graph &graph, const std::string &varname,
const std::unordered_map<std::string, int> &sharded_var_device) const;
bool IsScaleLossOp(ir::Node *node) const;
int CreateRPCOp(ir::Graph *result, ir::Node *node) const;
int CreateDistTrainOp(ir::Graph *result, ir::Node *node) const;
int CreateRPCOp(
ir::Graph *result, ir::Node *node,
std::unordered_map<std::string, int> *sharded_var_device) const;
int CreateDistTrainOp(
ir::Graph *result, ir::Node *node,
std::unordered_map<std::string, int> *sharded_var_device) const;
std::vector<std::string> FindDistTrainSendVars(
const std::vector<ir::Node *> &nodes) const;
......@@ -69,7 +75,9 @@ class MultiDevSSAGraphBuilder : public ir::Pass {
void CreateComputationalOp(ir::Graph *result, ir::Node *node,
int dev_id) const;
int GetOpDeviceID(const ir::Graph &graph, ir::Node *node) const;
int GetOpDeviceID(
const ir::Graph &graph, ir::Node *node,
const std::unordered_map<std::string, int> &sharded_var_device) const;
void InsertAllReduceOp(ir::Graph *result, const std::string &og) const;
......
paddle/fluid/framework/details/multi_devices_graph_print_pass.cc
浏览文件 @ f7b55de9
......@@ -15,6 +15,7 @@
#include "paddle/fluid/framework/details/multi_devices_graph_print_pass.h"
#include <string>
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/graph_helper.h"
namespace paddle {
namespace framework {
......@@ -62,7 +63,7 @@ void GraphvizSSAGraphPrinter::Print(const ir::Graph &graph,
});
size_t op_id = 0;
for (auto &op : graph.Get<GraphOps>(kGraphOps)) {
for (auto &op : ir::FilterByNodeWrapper<OpHandleBase>(graph)) {
std::string op_name = "op_" + std::to_string(op_id++);
sout << op_name << " [label=\"" << op->Name() << "\", shape=rect]"
<< std::endl;
......
paddle/fluid/framework/details/multi_devices_helper.h
浏览文件 @ f7b55de9
......@@ -35,23 +35,14 @@ namespace details {
// The outside vector is the device vector. Each element of this vector is a
// map from variable name to variables. The variables, who have the same name,
// will have a differsent version. The offset in the
// `std::vector<std::unique_ptr<VarHandle>>` is the version of varaibles.
typedef std::vector<
std::unordered_map<std::string, std::vector<std::unique_ptr<VarHandle>>>>
// `std::vector<VarHandle*>` is the version of varaibles.
typedef std::vector<std::unordered_map<std::string, std::vector<VarHandle*>>>
GraphVars;
const char kGraphVars[] = "vars";
// aux variables to represent dependency. Useful to resolve data hazard.
typedef std::unordered_set<std::unique_ptr<VarHandleBase>> GraphDepVars;
typedef std::unordered_set<VarHandleBase*> GraphDepVars;
const char kGraphDepVars[] = "dep_vars";
// all operators. NOTE that even we use a vector here, the operators is
// unordered.
typedef std::vector<std::unique_ptr<OpHandleBase>> GraphOps;
const char kGraphOps[] = "ops";
typedef std::unordered_map<std::string, int> ShardedVarDevice;
const char kShardedVarDevice[] = "sharded_var_device";
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/op_graph_view.cc 0 → 100644
浏览文件 @ f7b55de9
// 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/framework/details/op_graph_view.h"
#include <queue>
#include <utility>
namespace paddle {
namespace framework {
namespace details {
OpGraphView::OpGraphView(const std::vector<OpHandleBase *> &ops) { Build(ops); }
void OpGraphView::Build(const std::vector<OpHandleBase *> &ops) {
for (auto &op : ops) {
preceding_ops_[op];
pending_ops_[op];
for (auto &var : op->Outputs()) {
for (auto &pending_op : var->PendingOps()) {
preceding_ops_[pending_op].insert(op);
pending_ops_[op].insert(pending_op);
}
}
}
PADDLE_ENFORCE(
preceding_ops_.size() == ops.size() && pending_ops_.size() == ops.size(),
"There are duplicate ops in graph.");
}
std::unordered_set<OpHandleBase *> OpGraphView::AllOps() const {
std::unordered_set<OpHandleBase *> ret;
for (auto &pair : preceding_ops_) {
ret.insert(pair.first);
}
return ret;
}
bool OpGraphView::HasOp(OpHandleBase *op) const {
return preceding_ops_.count(op) != 0;
}
void OpGraphView::EnforceHasOp(OpHandleBase *op) const {
PADDLE_ENFORCE(HasOp(op), "Cannot find op %s in OpGraphView",
op == nullptr ? "nullptr" : op->DebugString());
}
const std::unordered_set<OpHandleBase *> &OpGraphView::PendingOps(
OpHandleBase *op) const {
EnforceHasOp(op);
return pending_ops_.at(op);
}
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/op_graph_view.h 0 → 100644
浏览文件 @ f7b55de9
// 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 <memory>
#include <unordered_map>
#include <unordered_set>
#include <vector>
#include "paddle/fluid/framework/details/op_handle_base.h"
namespace paddle {
namespace framework {
namespace details {
class OpGraphView {
public:
explicit OpGraphView(const std::vector<OpHandleBase *> &ops);
std::unordered_set<OpHandleBase *> AllOps() const;
const std::unordered_set<OpHandleBase *> &PendingOps(OpHandleBase *op) const;
bool HasOp(OpHandleBase *op) const;
private:
void Build(const std::vector<OpHandleBase *> &ops);
void EnforceHasOp(OpHandleBase *op) const;
std::unordered_map<OpHandleBase *, std::unordered_set<OpHandleBase *>>
preceding_ops_;
std::unordered_map<OpHandleBase *, std::unordered_set<OpHandleBase *>>
pending_ops_;
};
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/op_handle_base.h
浏览文件 @ f7b55de9
......@@ -31,7 +31,10 @@ constexpr char kLocalExecScopeName[] = "@LCOAL_SCOPE@";
// It's responsible for populating necessary fields of ir::Node.
class OpHandleBase {
public:
explicit OpHandleBase(ir::Node *node) : node_(node) {}
// Owned by `node`. No need to be deleted explicitly.
explicit OpHandleBase(ir::Node *node) : node_(node) {
node_->WrappedBy(this);
}
virtual ~OpHandleBase();
......
paddle/fluid/framework/details/reduce_op_handle_test.cc
浏览文件 @ f7b55de9
......@@ -30,8 +30,8 @@ struct TestReduceOpHandle {
Scope g_scope_;
std::vector<Scope *> local_scopes_;
std::vector<Scope *> param_scopes_;
std::unique_ptr<OpHandleBase> op_handle_;
std::vector<std::unique_ptr<VarHandleBase>> vars_;
OpHandleBase *op_handle_;
std::vector<VarHandleBase *> vars_;
std::vector<p::Place> gpu_list_;
std::vector<std::unique_ptr<p::DeviceContext>> ctxs_;
......
paddle/fluid/framework/details/reference_count_op_handle.h
浏览文件 @ f7b55de9
......@@ -51,7 +51,7 @@ class ReferenceCountOpHandle : public OpHandleBase {
dev_ctx_ = static_cast<platform::CUDADeviceContext *>(
platform::DeviceContextPool::Instance().Get(place));
if (IsStreamGarabageCollector()) {
PADDLE_ENFORCE(cudaSetDevice(place.device));
platform::SetDeviceId(place.device);
PADDLE_ENFORCE(cudaEventCreateWithFlags(&event_, cudaEventDisableTiming));
}
......@@ -61,7 +61,7 @@ class ReferenceCountOpHandle : public OpHandleBase {
~ReferenceCountOpHandle() {
if (IsStreamGarabageCollector()) {
auto gpu_place = boost::get<platform::CUDAPlace>(dev_ctx_->GetPlace());
PADDLE_ENFORCE(cudaSetDevice(gpu_place.device));
platform::SetDeviceId(gpu_place.device);
PADDLE_ENFORCE(cudaEventDestroy(event_));
}
}
......
paddle/fluid/framework/details/reference_count_pass.cc
浏览文件 @ f7b55de9
......@@ -19,6 +19,7 @@
#include "paddle/fluid/framework/details/computation_op_handle.h"
#include "paddle/fluid/framework/details/multi_devices_helper.h"
#include "paddle/fluid/framework/details/reference_count_pass.h"
#include "paddle/fluid/framework/ir/graph_helper.h"
namespace paddle {
namespace framework {
......@@ -43,6 +44,23 @@ static ComputationOpHandle *FindNextComputationOpHandle(VarHandle *var_in) {
return nullptr;
}
static void AddDependencyBetween(OpHandleBase *in, OpHandleBase *out,
ir::Graph *graph) {
auto it = std::find_if(
in->Outputs().begin(), in->Outputs().end(), [](VarHandleBase *var) {
return dynamic_cast<DummyVarHandle *>(var) != nullptr;
});
if (it != in->Outputs().end()) {
out->AddInput(*it);
} else {
auto *dep_var = new DummyVarHandle(graph->CreateControlDepVar());
graph->Get<GraphDepVars>(kGraphDepVars).emplace(dep_var);
in->AddOutput(dep_var);
out->AddInput(dep_var);
}
}
std::unique_ptr<ir::Graph> ReferenceCountPass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
auto &ref_cnts = Get<DeviceReferenceCountMap>(kGlobalReferenceCount);
......@@ -54,14 +72,13 @@ std::unique_ptr<ir::Graph> ReferenceCountPass::ApplyImpl(
// Step 2: Find all variables in non-computation ops which refers to variables
// in computation ops
std::unordered_set<std::string> names;
std::unordered_map<OpHandleBase *, std::unique_ptr<ReferenceCountOpHandle>>
std::unordered_map<OpHandleBase *, ReferenceCountOpHandle *>
compute_ref_cnt_map;
auto get_ref_cnts_from_compute_op = [&](
const std::unique_ptr<OpHandleBase> &op,
const std::vector<VarHandleBase *> &vars) {
OpHandleBase *op, const std::vector<VarHandleBase *> &vars) {
std::vector<std::string> var_names_in_op;
auto *compute_op = dynamic_cast<ComputationOpHandle *>(op.get());
auto *compute_op = dynamic_cast<ComputationOpHandle *>(op);
if (compute_op == nullptr ||
!platform::is_gpu_place(compute_op->GetPlace()))
return var_names_in_op;
......@@ -104,9 +121,8 @@ std::unique_ptr<ir::Graph> ReferenceCountPass::ApplyImpl(
};
auto update_ref_cnts_from_non_compute_op = [&](
const std::unique_ptr<OpHandleBase> &op,
const std::vector<VarHandleBase *> &vars) {
if (dynamic_cast<ComputationOpHandle *>(op.get()) != nullptr) return;
OpHandleBase *op, const std::vector<VarHandleBase *> &vars) {
if (dynamic_cast<ComputationOpHandle *>(op) != nullptr) return;
for (VarHandleBase *var_handle_base : vars) {
auto *var_handle = dynamic_cast<VarHandle *>(var_handle_base);
if (var_handle == nullptr || !var_handle->Node()->IsVar()) continue;
......@@ -124,8 +140,8 @@ std::unique_ptr<ir::Graph> ReferenceCountPass::ApplyImpl(
if (next_compute_op != nullptr) {
if (compute_ref_cnt_map.count(next_compute_op)) {
compute_ref_cnt_map[next_compute_op]->AddVar(var_name);
VLOG(5) << "Add reference count of " << var_name << " to Operator "
<< next_compute_op->Name();
VLOG(50) << "Add reference count of " << var_name << " to Operator "
<< next_compute_op->Name();
} else {
// Create new reference_count_op_handle
ir::Node *ref_cnt_node = graph->CreateEmptyNode(
......@@ -133,40 +149,30 @@ std::unique_ptr<ir::Graph> ReferenceCountPass::ApplyImpl(
auto *ref_cnt_handle = new ReferenceCountOpHandle(
ref_cnt_node, next_compute_op->GetScope(), place, {var_name},
gcs[place.device].get(), cur_ref_cnts[place.device].get());
if (next_compute_op->Outputs().empty()) {
auto *dep_var = new DummyVarHandle(graph->CreateControlDepVar());
next_compute_op->AddOutput(dep_var);
graph->Get<GraphDepVars>(kGraphDepVars).emplace(dep_var);
}
ref_cnt_handle->AddInput(next_compute_op->Outputs().front());
compute_ref_cnt_map[next_compute_op].reset(ref_cnt_handle);
AddDependencyBetween(next_compute_op, ref_cnt_handle, graph.get());
compute_ref_cnt_map[next_compute_op] = ref_cnt_handle;
}
}
}
}
};
auto &all_ops = graph->Get<GraphOps>(kGraphOps);
auto all_ops = ir::FilterByNodeWrapper<OpHandleBase>(*graph);
for (auto &op : all_ops) {
auto in_var_names = get_ref_cnts_from_compute_op(op, op->Inputs());
auto out_var_names = get_ref_cnts_from_compute_op(op, op->Outputs());
if (in_var_names.empty() && out_var_names.empty()) continue;
in_var_names.insert(in_var_names.end(), out_var_names.begin(),
out_var_names.end());
auto *compute_op = dynamic_cast<ComputationOpHandle *>(op.get());
auto *compute_op = dynamic_cast<ComputationOpHandle *>(op);
auto place = boost::get<platform::CUDAPlace>(compute_op->GetPlace());
ir::Node *ref_cnt_node =
graph->CreateEmptyNode("reference_count", ir::Node::Type::kOperation);
auto *ref_cnt_handle = new ReferenceCountOpHandle(
ref_cnt_node, compute_op->GetScope(), place, in_var_names,
gcs[place.device].get(), cur_ref_cnts[place.device].get());
if (compute_op->Outputs().empty()) {
auto *dep_var = new DummyVarHandle(graph->CreateControlDepVar());
compute_op->AddOutput(dep_var);
graph->Get<GraphDepVars>(kGraphDepVars).emplace(dep_var);
}
ref_cnt_handle->AddInput(compute_op->Outputs().front());
compute_ref_cnt_map[compute_op].reset(ref_cnt_handle);
AddDependencyBetween(compute_op, ref_cnt_handle, graph.get());
compute_ref_cnt_map[compute_op] = ref_cnt_handle;
}
for (auto &op : all_ops) {
......@@ -174,11 +180,11 @@ std::unique_ptr<ir::Graph> ReferenceCountPass::ApplyImpl(
update_ref_cnts_from_non_compute_op(op, op->Outputs());
}
std::vector<std::unique_ptr<OpHandleBase>> new_all_ops;
std::vector<OpHandleBase *> new_all_ops;
new_all_ops.reserve(compute_ref_cnt_map.size() + all_ops.size());
for (auto &op : all_ops) {
new_all_ops.emplace_back(std::move(op));
auto it = compute_ref_cnt_map.find(new_all_ops.back().get());
auto it = compute_ref_cnt_map.find(new_all_ops.back());
if (it != compute_ref_cnt_map.end()) {
// Add LeafNode to ReferenceCountOpHandle
auto *dummy_leaf = new DummyVarHandle(graph->CreateControlDepVar());
......
paddle/fluid/framework/details/rpc_op_handle.cc
浏览文件 @ f7b55de9
......@@ -29,22 +29,19 @@ RPCOpHandle::RPCOpHandle(ir::Node *node, const framework::OpDesc &op_desc,
place_(place) {}
void RPCOpHandle::RunImpl() {
// TODO(wuyi): need further analysis whether wait VarDummyHandle.
// Wait input done
for (auto *in : inputs_) {
auto &p = static_cast<VarHandle *>(in)->place_;
// FIXME(Yancey1989): need a better solution instead of use DebugString()
if (ir::IsControlDepVar(*in->Node())) { // HACK
if (ir::IsControlDepVar(*in->Node())) {
continue;
}
if (in->GeneratedOp()) {
in->GeneratedOp()->RecordWaitEventOnCtx(dev_ctxes_.at(p));
}
}
auto &tmp_scope = local_scope_->FindVar(kLocalExecScopeName)->Get<Scope *>();
// FIXME(wuyi): can not use RunAndRecordEvent here, for it will cause dead
// lock.
op_->Run(*tmp_scope, place_);
this->RunAndRecordEvent([this] {
op_->Run(*local_scope_->FindVar(kLocalExecScopeName)->Get<Scope *>(),
place_);
});
}
std::string RPCOpHandle::Name() const { return name_; }
......
paddle/fluid/framework/details/scale_loss_grad_op_handle.cc
浏览文件 @ f7b55de9
......@@ -51,7 +51,7 @@ void ScaleLossGradOpHandle::RunImpl() {
->stream();
memory::Copy(boost::get<platform::CUDAPlace>(place_), tmp,
platform::CPUPlace(), &coeff_, sizeof(float), stream);
VLOG(10) << place_ << "RUN Scale loss grad op";
VLOG(100) << place_ << "RUN Scale loss grad op";
});
#endif
}
......
paddle/fluid/framework/details/sequential_execution_pass.cc 0 → 100644
浏览文件 @ f7b55de9
// 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/framework/details/sequential_execution_pass.h"
#include <string>
#include <unordered_map>
#include <unordered_set>
#include <vector>
#include "paddle/fluid/framework/op_proto_maker.h"
namespace paddle {
namespace framework {
namespace details {
static bool IsSameOpDesc(OpDesc *op1, OpDesc *op2) {
return op1->Type() == op2->Type() && op1->Inputs() == op2->Inputs() &&
op1->Outputs() == op2->Outputs();
}
std::unique_ptr<ir::Graph> SequentialExecutionPass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
// FIXME(zjl): Insert dependencies between some distributed ops may cause
// the multi_devices_graph_pass fails. So we skip these ops here.
// Indeed, maybe we should not insert dependencies between these ops
// casually, which may cause deadlock easily.
// We should add more skipped distributed ops when found errors in
// multi_devices_graph_pass
static std::unordered_set<std::string> skip_dist_ops{
"send", "recv", "send_barrier", "fetch_barrier"};
auto &ops = Get<const std::vector<OpDesc *>>(kAllOpDescs);
std::vector<ir::Node *> op_node_list;
op_node_list.reserve(ops.size());
std::unordered_map<ir::Node *, size_t> op_deps;
std::unordered_map<ir::Node *, std::unordered_set<ir::Node *>> pending_ops;
std::unordered_set<ir::Node *> ready_ops;
for (ir::Node *node : graph->Nodes()) {
if (!node->IsOp()) continue;
std::unordered_set<ir::Node *> preceding_ops;
for (auto *in : node->inputs) {
PADDLE_ENFORCE(in->IsVar(),
"Preceding Node of Op Nodes must be Var Node");
if (in->inputs.empty()) continue;
PADDLE_ENFORCE(in->inputs.size() == 1 && in->inputs[0]->IsOp(),
"Preceding Op Node of Var Node must be unique");
preceding_ops.insert(in->inputs[0]);
pending_ops[in->inputs[0]].insert(node);
}
op_deps[node] = preceding_ops.size();
if (preceding_ops.empty()) {
ready_ops.insert(node);
}
}
for (auto *op_desc : ops) {
ir::Node *found_node = nullptr;
for (auto *node : ready_ops) {
if (IsSameOpDesc(op_desc, node->Op())) {
PADDLE_ENFORCE(found_node == nullptr,
"Found multiple op_desc in graph: %s", op_desc->Type());
found_node = node;
}
}
PADDLE_ENFORCE_NOT_NULL(found_node, "Cannot find op_desc in graph: %s",
op_desc->Type());
for (auto *pending_op : pending_ops[found_node]) {
if (--op_deps.at(pending_op) == 0) {
ready_ops.insert(pending_op);
}
}
ready_ops.erase(found_node);
if (skip_dist_ops.count(op_desc->Type()) == 0) {
op_node_list.push_back(found_node);
}
}
for (size_t i = 1; i < op_node_list.size(); ++i) {
auto *dep_var = graph->CreateControlDepVar();
op_node_list[i]->inputs.push_back(dep_var);
op_node_list[i - 1]->outputs.push_back(dep_var);
dep_var->outputs.push_back(op_node_list[i]);
dep_var->inputs.push_back(op_node_list[i - 1]);
VLOG(100) << "Add dependencies between " << op_node_list[i - 1]->Name()
<< " and " << op_node_list[i]->Name();
}
return graph;
}
} // namespace details
} // namespace framework
} // namespace paddle
REGISTER_PASS(sequential_execution_pass,
paddle::framework::details::SequentialExecutionPass)
.RequirePassAttr(paddle::framework::details::kAllOpDescs);
paddle/fluid/framework/details/sequential_execution_pass.h 0 → 100644
浏览文件 @ f7b55de9
// 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/ir/graph.h"
#include "paddle/fluid/framework/ir/pass.h"
namespace paddle {
namespace framework {
namespace details {
constexpr char kAllOpDescs[] = "all_op_descs";
class SequentialExecutionPass : public ir::Pass {
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
};
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/ssa_graph_executor.cc
浏览文件 @ f7b55de9
......@@ -19,14 +19,16 @@ namespace framework {
namespace details {
SSAGraphExecutor::~SSAGraphExecutor() {}
void ClearFetchOp(ir::Graph* graph,
std::vector<std::unique_ptr<FetchOpHandle>>* fetch_ops) {
void ClearFetchOp(ir::Graph* graph, std::vector<FetchOpHandle*>* fetch_ops) {
if (fetch_ops->empty()) return;
for (auto& op : *fetch_ops) {
for (auto& out_var : op->Node()->outputs) {
graph->RemoveNode(out_var);
}
for (auto& in_var : op->Inputs()) {
in_var->RemoveOutput(op, op->Node());
}
graph->RemoveNode(op->Node());
}
fetch_ops->clear();
......
paddle/fluid/framework/details/ssa_graph_executor.h
浏览文件 @ f7b55de9
......@@ -38,8 +38,7 @@ class SSAGraphExecutor {
virtual FeedFetchList Run(const std::vector<std::string>& fetch_tensors) = 0;
};
void ClearFetchOp(ir::Graph* graph,
std::vector<std::unique_ptr<FetchOpHandle>>* fetch_ops);
void ClearFetchOp(ir::Graph* graph, std::vector<FetchOpHandle*>* fetch_ops);
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/threaded_ssa_graph_executor.cc
浏览文件 @ f7b55de9
......@@ -15,6 +15,7 @@
#include "paddle/fluid/framework/details/threaded_ssa_graph_executor.h"
#include "paddle/fluid/framework/details/multi_devices_helper.h"
#include "paddle/fluid/framework/ir/graph_helper.h"
#include "paddle/fluid/platform/profiler.h"
namespace paddle {
......@@ -51,25 +52,25 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
for (auto &var_map : graph_->Get<details::GraphVars>(details::kGraphVars)) {
for (auto &name_pair : var_map) {
for (auto &version_pair : name_pair.second) {
InsertPendingVar(&pending_vars, ready_vars.get(), version_pair.get());
InsertPendingVar(&pending_vars, ready_vars.get(), version_pair);
}
}
}
for (auto &var : graph_->Get<details::GraphDepVars>(details::kGraphDepVars)) {
InsertPendingVar(&pending_vars, ready_vars.get(), var.get());
InsertPendingVar(&pending_vars, ready_vars.get(), var);
}
for (auto &op : graph_->Get<details::GraphOps>(details::kGraphOps)) {
for (auto &op : ir::FilterByNodeWrapper<OpHandleBase>(*graph_)) {
if (op->Inputs().empty()) { // Special case, Op has no input.
ready_ops.insert(op.get());
ready_ops.insert(op);
} else {
InsertPendingOp(&pending_ops, op.get());
InsertPendingOp(&pending_ops, op);
}
}
// Step 2. Insert FetchOps
std::vector<std::unique_ptr<FetchOpHandle>> fetch_ops;
std::unordered_set<std::unique_ptr<VarHandleBase>> fetch_dependencies;
std::vector<FetchOpHandle *> fetch_ops;
std::unordered_set<VarHandleBase *> fetch_dependencies;
FeedFetchList fetch_data(fetch_tensors.size());
InsertFetchOps(fetch_tensors, &fetch_ops, &fetch_dependencies, &pending_ops,
......@@ -109,6 +110,7 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
for (auto &run_op_future : run_op_futures_) {
run_op_future.wait();
}
ClearFetchOp(graph_.get(), &fetch_ops);
exception_holder_.ReThrow();
} else {
continue;
......@@ -140,8 +142,8 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
void ThreadedSSAGraphExecutor::InsertFetchOps(
const std::vector<std::string> &fetch_tensors,
std::vector<std::unique_ptr<FetchOpHandle>> *fetch_ops,
std::unordered_set<std::unique_ptr<VarHandleBase>> *fetch_dependencies,
std::vector<FetchOpHandle *> *fetch_ops,
std::unordered_set<VarHandleBase *> *fetch_dependencies,
std::unordered_map<OpHandleBase *, size_t> *pending_ops,
std::unordered_set<VarHandleBase *> *pending_vars,
BlockingQueue<VarHandleBase *> *ready_vars, FeedFetchList *fetch_data) {
......@@ -151,7 +153,7 @@ void ThreadedSSAGraphExecutor::InsertFetchOps(
for (auto &var_map : graph_->Get<details::GraphVars>(details::kGraphVars)) {
auto it = var_map.find(fetch_var_name);
if (it != var_map.end()) {
fetched_vars[fetch_var_name].push_back(it->second.rbegin()->get());
fetched_vars[fetch_var_name].push_back(*it->second.rbegin());
}
}
}
......@@ -208,14 +210,16 @@ void ThreadedSSAGraphExecutor::RunOp(
details::OpHandleBase *op) {
auto op_run = [ready_var_q, op, this] {
try {
if (VLOG_IS_ON(10)) {
VLOG(10) << op << " " << op->Name() << " : " << op->DebugString();
if (VLOG_IS_ON(100)) {
VLOG(100) << op << " " << op->Name() << " : " << op->DebugString();
}
op->Run(strategy_.use_cuda_);
VLOG(10) << op << " " << op->Name() << " Done ";
if (LIKELY(!strategy_.dry_run_)) {
op->Run(strategy_.use_cuda_);
}
VLOG(100) << op << " " << op->Name() << " Done ";
running_ops_--;
ready_var_q->Extend(op->Outputs());
VLOG(10) << op << " " << op->Name() << "Signal posted";
VLOG(100) << op << " " << op->Name() << "Signal posted";
} catch (...) {
exception_holder_.Catch(std::current_exception());
}
......
paddle/fluid/framework/details/threaded_ssa_graph_executor.h
浏览文件 @ f7b55de9
......@@ -48,7 +48,7 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor {
// Use topological sort algorithm
FeedFetchList Run(const std::vector<std::string> &fetch_tensors) override;
~ThreadedSSAGraphExecutor() {}
~ThreadedSSAGraphExecutor() final = default;
private:
void RunOp(const std::shared_ptr<BlockingQueue<VarHandleBase *>> &ready_var_q,
......@@ -70,13 +70,13 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor {
BlockingQueue<VarHandleBase *> *ready_vars,
VarHandleBase *var) const;
void InsertFetchOps(
const std::vector<std::string> &fetch_tensors,
std::vector<std::unique_ptr<FetchOpHandle>> *fetch_ops,
std::unordered_set<std::unique_ptr<VarHandleBase>> *fetch_dependencies,
std::unordered_map<OpHandleBase *, size_t> *pending_ops,
std::unordered_set<VarHandleBase *> *pending_vars,
BlockingQueue<VarHandleBase *> *ready_vars, FeedFetchList *fetch_data);
void InsertFetchOps(const std::vector<std::string> &fetch_tensors,
std::vector<FetchOpHandle *> *fetch_ops,
std::unordered_set<VarHandleBase *> *fetch_dependencies,
std::unordered_map<OpHandleBase *, size_t> *pending_ops,
std::unordered_set<VarHandleBase *> *pending_vars,
BlockingQueue<VarHandleBase *> *ready_vars,
FeedFetchList *fetch_data);
private:
ExecutionStrategy strategy_;
......
paddle/fluid/framework/details/var_handle.cc
浏览文件 @ f7b55de9
......@@ -20,6 +20,8 @@ namespace details {
VarHandleBase::~VarHandleBase() {}
VarHandle::~VarHandle() { VLOG(4) << "deleting var handle " << DebugString(); }
std::string VarHandle::DebugString() const {
std::stringstream ss;
ss << name_ << ":" << place_;
......@@ -27,6 +29,10 @@ std::string VarHandle::DebugString() const {
}
std::string DummyVarHandle::DebugString() const { return node_->Name(); }
DummyVarHandle::~DummyVarHandle() {
VLOG(4) << "deleting dummy var handle " << DebugString();
}
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/var_handle.h
浏览文件 @ f7b55de9
......@@ -35,7 +35,10 @@ class OpHandleBase;
// A variable can only be generated by a single operator. i.e.
// This is a single assignment graph.
struct VarHandleBase {
explicit VarHandleBase(ir::Node* node) : node_(node) {}
// Owned by `node`. No need to be deleted explicitly.
explicit VarHandleBase(ir::Node* node) : node_(node) {
node_->WrappedBy(this);
}
virtual ~VarHandleBase();
......@@ -94,6 +97,8 @@ struct VarHandleBase {
struct VarHandle : public VarHandleBase {
explicit VarHandle(ir::Node* node) : VarHandleBase(node) {}
virtual ~VarHandle();
std::string DebugString() const override;
VarHandle(ir::Node* node, size_t version, size_t scope_index,
......@@ -121,6 +126,8 @@ struct VarHandle : public VarHandleBase {
struct DummyVarHandle : public VarHandleBase {
explicit DummyVarHandle(ir::Node* node) : VarHandleBase(node) {}
virtual ~DummyVarHandle();
std::string DebugString() const override;
};
......
paddle/fluid/framework/executor.cc
浏览文件 @ f7b55de9
......@@ -17,6 +17,7 @@ limitations under the License. */
#include "paddle/fluid/framework/feed_fetch_method.h"
#include "paddle/fluid/framework/lod_rank_table.h"
#include "paddle/fluid/framework/lod_tensor_array.h"
#include "paddle/fluid/framework/ngraph_operator.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/reader.h"
#include "paddle/fluid/operators/detail/macros.h"
......@@ -25,6 +26,7 @@ limitations under the License. */
DECLARE_bool(benchmark);
DEFINE_bool(use_mkldnn, false, "Use MKLDNN to run");
DEFINE_bool(use_ngraph, false, "Use NGRAPH to run");
namespace paddle {
namespace framework {
......@@ -43,7 +45,7 @@ ExecutorPrepareContext::ExecutorPrepareContext(
}
ExecutorPrepareContext::~ExecutorPrepareContext() {
VLOG(5) << "destroy ExecutorPrepareContext";
VLOG(50) << "destroy ExecutorPrepareContext";
}
template <typename RefCntMap>
......@@ -60,7 +62,7 @@ static void DeleteUnusedTensors(const Scope& scope, const OperatorBase* op,
if ((it->second)-- == 1) {
auto* var = scope.FindVar(name);
if (var != nullptr) {
VLOG(10) << "Erase tensor \'" << name << "\'";
VLOG(100) << "Erase tensor \'" << name << "\'";
if (var->IsType<LoDTensor>()) {
erase_tensors.insert(var->GetMutable<LoDTensor>());
} else if (var->IsType<SelectedRows>()) {
......@@ -81,12 +83,32 @@ static void DeleteUnusedTensors(const Scope& scope, const OperatorBase* op,
}
}
static void EnableFusedOp(ExecutorPrepareContext* ctx) {
#ifdef PADDLE_WITH_NGRAPH
VLOG(3) << "use_ngraph=True";
auto intervals = FusedOperator::FusedOpIntervals(&ctx->ops_);
for (auto& interval : intervals) {
auto* fused_op = new FusedOperator(ctx->prog_, ctx->block_id_,
interval.at(0), interval.at(1));
*interval[0] = std::unique_ptr<OperatorBase>(fused_op);
}
for (auto it = intervals.rbegin(); it != intervals.rend(); ++it) {
ctx->ops_.erase(it->at(0) + 1, it->at(1));
}
#else
LOG(WARNING)
<< "'NGRAPH' is not supported, Please re-compile with WITH_NGRAPH option";
#endif
}
Executor::Executor(const platform::Place& place) : place_(place) {}
void Executor::Close() {
#ifdef PADDLE_WITH_DISTRIBUTE
// TODO(typhoonzero): complete message will need to use real trainer_id,
// except 0.
::paddle::operators::distributed::RPCClient::GetInstance<
::paddle::operators::distributed::GRPCClient>()
::paddle::operators::distributed::GRPCClient>(0)
->SendComplete();
#endif
}
......@@ -139,21 +161,21 @@ void Executor::CreateVariables(const ProgramDesc& pdesc, Scope* scope,
if (var->Persistable()) {
auto* ptr = const_cast<Scope*>(ancestor_scope)->Var(var->Name());
InitializeVariable(ptr, var->GetType());
VLOG(3) << "Create Variable " << var->Name()
<< " global, which pointer is " << ptr;
VLOG(30) << "Create Variable " << var->Name()
<< " global, which pointer is " << ptr;
} else {
auto* ptr = scope->Var(var->Name());
InitializeVariable(ptr, var->GetType());
VLOG(3) << "Create Variable " << var->Name()
<< " locally, which pointer is " << ptr;
VLOG(30) << "Create Variable " << var->Name()
<< " locally, which pointer is " << ptr;
}
}
} else {
for (auto& var : global_block.AllVars()) {
auto* ptr = scope->Var(var->Name());
InitializeVariable(ptr, var->GetType());
VLOG(3) << "Create variable " << var->Name() << ", which pointer is "
<< ptr;
VLOG(30) << "Create variable " << var->Name() << ", which pointer is "
<< ptr;
}
}
}
......@@ -284,7 +306,7 @@ void Executor::Run(const ProgramDesc& program, Scope* scope,
int i = 0;
for (auto& feed_target : (*feed_targets)) {
std::string var_name = feed_target.first;
VLOG(3) << "feed target's name: " << var_name;
VLOG(30) << "feed target's name: " << var_name;
// prepend feed op
auto* op = global_block->PrependOp();
......@@ -307,7 +329,7 @@ void Executor::Run(const ProgramDesc& program, Scope* scope,
int i = 0;
for (auto& fetch_target : (*fetch_targets)) {
std::string var_name = fetch_target.first;
VLOG(3) << "fetch target's name: " << var_name;
VLOG(30) << "fetch target's name: " << var_name;
// append fetch op
auto* op = global_block->AppendOp();
......@@ -336,6 +358,7 @@ std::unique_ptr<ExecutorPrepareContext> Executor::Prepare(
for (auto& op_desc : block.AllOps()) {
ctx->ops_.push_back(OpRegistry::CreateOp(*op_desc));
}
if (FLAGS_use_ngraph) EnableFusedOp(ctx.get());
return ctx;
}
......@@ -357,6 +380,7 @@ std::vector<std::shared_ptr<ExecutorPrepareContext>> Executor::Prepare(
void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
bool create_local_scope, bool create_vars,
bool keep_kids) {
PADDLE_ENFORCE_NOT_NULL(scope);
Scope* local_scope = scope;
if (create_vars) {
if (create_local_scope) {
......@@ -394,11 +418,6 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
DeleteUnusedTensors(*local_scope, op.get(), gc.get(),
&(ctx->cur_ref_cnts_));
}
if (FLAGS_benchmark) {
VLOG(2) << "Memory used after operator " + op->Type() + " running: "
<< memory::memory_usage(place_);
}
}
if (gc != nullptr) {
......@@ -420,13 +439,6 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
scope->DropKids();
}
}
if (FLAGS_benchmark) {
VLOG(2) << "-------------------------------------------------------";
VLOG(2) << "Memory used after deleting local scope: "
<< memory::memory_usage(place_);
VLOG(2) << "-------------------------------------------------------";
}
}
void Executor::RunPreparedContext(
......@@ -469,7 +481,7 @@ void Executor::RunPreparedContext(
void Executor::EnableMKLDNN(const ProgramDesc& program) {
#ifdef PADDLE_WITH_MKLDNN
VLOG(3) << "use_mkldnn=True";
VLOG(30) << "use_mkldnn=True";
for (size_t bid = 0; bid < program.Size(); ++bid) {
auto* block = const_cast<ProgramDesc&>(program).MutableBlock(bid);
for (auto* op : block->AllOps()) {
......@@ -483,6 +495,5 @@ void Executor::EnableMKLDNN(const ProgramDesc& program) {
<< "'MKLDNN' is not supported, Please re-compile with WITH_MKLDNN option";
#endif
}
} // namespace framework
} // namespace paddle
paddle/fluid/framework/feed_fetch_method.cc
浏览文件 @ f7b55de9
......@@ -25,7 +25,7 @@ void SetFeedVariable(Scope* scope, const LoDTensor& input,
const std::string& var_name, size_t index) {
// If var_name Variable is not found in GlobalScope, a new variable will
// be created.
VLOG(3) << "SetFeedVariable name=" << var_name << " index=" << index;
VLOG(30) << "SetFeedVariable name=" << var_name << " index=" << index;
Variable* g_feed_value = scope->Var(var_name);
auto& feed_inputs = *(g_feed_value->GetMutable<FeedFetchList>());
if (index >= feed_inputs.size()) {
......@@ -47,8 +47,8 @@ LoDTensor& GetFetchVariable(const Scope& scope, const std::string& var_name,
typeid(FeedFetchList).name());
auto& fetch_outputs = *g_fetch_value->GetMutable<FeedFetchList>();
auto& tensor = fetch_outputs[index];
VLOG(3) << "Fetch " << var_name << " with index " << index
<< " shape= " << tensor.dims();
VLOG(30) << "Fetch " << var_name << " with index " << index
<< " shape= " << tensor.dims();
PADDLE_ENFORCE_LT(index, fetch_outputs.size());
return tensor;
}
......
paddle/fluid/framework/garbage_collector.h
浏览文件 @ f7b55de9
......@@ -29,7 +29,7 @@ template <typename T>
class GarbageCollector {
public:
GarbageCollector(const platform::Place &place, size_t max_memory_size)
: max_memory_size_(std::max(max_memory_size, static_cast<size_t>(1))) {
: max_memory_size_((std::max)(max_memory_size, static_cast<size_t>(1))) {
garbages_.reset(new std::deque<T *>());
dev_ctx_ = platform::DeviceContextPool::Instance().Get(place);
}
......
paddle/fluid/framework/ir/CMakeLists.txt
浏览文件 @ f7b55de9
......@@ -5,6 +5,7 @@ file(APPEND ${pass_file} "\#include \"paddle/fluid/framework/ir/pass.h\"\n")
# Usage: pass_library(target inference) will append to paddle_inference_pass.h
unset(INFER_IR_PASSES CACHE) # clear the global variable
function(pass_library TARGET DEST)
set(options "")
set(oneValueArgs "")
......@@ -15,10 +16,11 @@ function(pass_library TARGET DEST)
if (${DEST} STREQUAL "base" OR ${DEST} STREQUAL "inference")
message(STATUS "add pass ${TARGET} ${DEST}")
file(APPEND ${pass_file} "USE_PASS(${TARGET});\n")
set(PASS_LIBRARY ${TARGET} ${PASS_LIBRARY} PARENT_SCOPE)
set(INFER_IR_PASSES ${INFER_IR_PASSES} ${TARGET} CACHE INTERNAL "")
endif()
endfunction()
cc_library(node SRCS node.cc DEPS proto_desc)
cc_library(graph SRCS graph.cc DEPS node pretty_log)
cc_library(graph_helper SRCS graph_helper.cc DEPS graph)
......@@ -39,8 +41,10 @@ pass_library(seq_concat_fc_fuse_pass inference)
pass_library(multi_batch_merge_pass base)
pass_library(conv_bn_fuse_pass inference)
pass_library(seqconv_eltadd_relu_fuse_pass inference)
pass_library(is_test_pass base)
if(WITH_MKLDNN)
pass_library(mkldnn_placement_pass base)
pass_library(depthwise_conv_mkldnn_pass base)
pass_library(conv_bias_mkldnn_fuse_pass inference)
pass_library(conv_relu_mkldnn_fuse_pass inference)
pass_library(conv_elementwise_add_mkldnn_fuse_pass inference)
......@@ -52,13 +56,16 @@ set(GLOB_PASS_LIB ${PASS_LIBRARY} CACHE INTERNAL "Global PASS library")
cc_library(pass_builder SRCS pass_builder.cc DEPS pass)
cc_test(node_test SRCS node_test.cc DEPS node)
cc_test(pass_test SRCS pass_test.cc DEPS graph pass graph_helper)
cc_test(graph_test SRCS graph_test.cc DEPS graph graph_helper op_registry)
cc_test(graph_helper_test SRCS graph_helper_test.cc DEPS graph graph_helper op_registry)
cc_test(graph_to_program_pass_test SRCS graph_to_program_pass_test.cc DEPS graph_to_program_pass)
cc_test(test_graph_pattern_detector SRCS graph_pattern_detector_tester.cc DEPS graph_pattern_detector)
cc_test(test_fc_fuse_pass SRCS fc_fuse_pass_tester.cc DEPS fc_fuse_pass framework_proto)
cc_test(test_is_test_pass SRCS is_test_pass_tester.cc DEPS is_test_pass)
if (WITH_MKLDNN)
cc_test(test_depthwise_conv_mkldnn_pass SRCS depthwise_conv_mkldnn_pass_tester.cc DEPS depthwise_conv_mkldnn_pass)
cc_test(test_conv_relu_mkldnn_fuse_pass SRCS conv_relu_mkldnn_fuse_pass_tester.cc DEPS conv_relu_mkldnn_fuse_pass)
cc_test(test_conv_elementwise_add_mkldnn_fuse_pass SRCS conv_elementwise_add_mkldnn_fuse_pass_tester.cc DEPS conv_elementwise_add_mkldnn_fuse_pass)
endif ()
paddle/fluid/framework/ir/attention_lstm_fuse_pass.cc
浏览文件 @ f7b55de9
......@@ -91,10 +91,10 @@ void FindWhileOp(Graph* graph) {
#undef OP_SET_IN
#undef OP_SET_OUT
auto* X = graph->RetriveNode(34);
auto* LSTMOUT = graph->RetriveNode(81);
auto* cell_init = graph->RetriveNode(6);
auto* hidden_init = graph->RetriveNode(8);
auto* X = graph->RetrieveNode(34);
auto* LSTMOUT = graph->RetrieveNode(81);
auto* cell_init = graph->RetrieveNode(6);
auto* hidden_init = graph->RetrieveNode(8);
auto* lstm_op = graph->CreateOpNode(&op_desc);
PrepareParameters(graph, param);
......@@ -147,19 +147,19 @@ void PrepareParameters(Graph* graph, const Param& param) {
scope->Var(param.LSTMX)->GetMutable<LoDTensor>();
scope->Var(param.LSTMOUT)->GetMutable<LoDTensor>();
#define GATE_W(name__) \
auto* W_##name__##_w0 = scope->FindVar(#name__ ".w_0"); \
auto* W_##name__##_w1 = scope->FindVar(#name__ ".w_1"); \
auto* W_##name__##_b0 = scope->FindVar(#name__ ".b_0"); \
CHECK_P3(W_##name__##_w0, W_##name__##_w1, W_##name__##_b0); \
VLOG(4) << #name__ "_w0" \
<< " shape: " << W_##name__##_w0->Get<LoDTensor>().dims(); \
VLOG(4) << #name__ "_w1" \
<< " shape: " << W_##name__##_w1->Get<LoDTensor>().dims(); \
VLOG(4) << #name__ "_b0" \
<< " shape: " << W_##name__##_b0->Get<LoDTensor>().dims(); \
auto& W_##name__##_w0_t = W_##name__##_w0->Get<LoDTensor>(); \
auto& W_##name__##_w1_t = W_##name__##_w1->Get<LoDTensor>(); \
#define GATE_W(name__) \
auto* W_##name__##_w0 = scope->FindVar(#name__ ".w_0"); \
auto* W_##name__##_w1 = scope->FindVar(#name__ ".w_1"); \
auto* W_##name__##_b0 = scope->FindVar(#name__ ".b_0"); \
CHECK_P3(W_##name__##_w0, W_##name__##_w1, W_##name__##_b0); \
VLOG(40) << #name__ "_w0" \
<< " shape: " << W_##name__##_w0->Get<LoDTensor>().dims(); \
VLOG(40) << #name__ "_w1" \
<< " shape: " << W_##name__##_w1->Get<LoDTensor>().dims(); \
VLOG(40) << #name__ "_b0" \
<< " shape: " << W_##name__##_b0->Get<LoDTensor>().dims(); \
auto& W_##name__##_w0_t = W_##name__##_w0->Get<LoDTensor>(); \
auto& W_##name__##_w1_t = W_##name__##_w1->Get<LoDTensor>(); \
auto& W_##name__##_b0_t = W_##name__##_b0->Get<LoDTensor>();
GATE_W(forget);
......@@ -208,15 +208,15 @@ void PrepareLSTMWeight(const LoDTensor& W_forget_w0,
int D = W_forget_w0.dims()[0];
int M = W_forget_w1.dims()[0];
out->Resize(make_ddim({D + M, 4 * D}));
VLOG(3) << "LSTMWeight resized to " << out->dims();
VLOG(30) << "LSTMWeight resized to " << out->dims();
float* out_data = out->mutable_data<float>(platform::CPUPlace());
std::array<const float*, 4> tensors(
{{W_forget_w0.data<float>(), W_input_w0.data<float>(),
W_output_w0.data<float>(), W_cell_w0.data<float>()}});
std::array<const float*, 4> tensors1(
{{W_forget_w1.data<float>(), W_input_w1.data<float>(),
W_output_w1.data<float>(), W_cell_w1.data<float>()}});
std::array<const float*, 4> tensors{
W_forget_w0.data<float>(), W_input_w0.data<float>(),
W_output_w0.data<float>(), W_cell_w0.data<float>()};
std::array<const float*, 4> tensors1{
W_forget_w1.data<float>(), W_input_w1.data<float>(),
W_output_w1.data<float>(), W_cell_w1.data<float>()};
for (int row = 0; row < D; row++) {
for (int col = 0; col < 4; col++) {
......@@ -238,9 +238,9 @@ void PrepareLSTMWeight(const LoDTensor& W_forget_w0,
void PrepareLSTMBias(const LoDTensor& B_forget, const LoDTensor& B_input,
const LoDTensor& B_output, const LoDTensor& B_cell,
LoDTensor* out) {
std::array<const float*, 4> tensors(
{{B_forget.data<float>(), B_input.data<float>(), B_output.data<float>(),
B_cell.data<float>()}});
std::array<const float*, 4> tensors{
B_forget.data<float>(), B_input.data<float>(), B_output.data<float>(),
B_cell.data<float>()};
PADDLE_ENFORCE_EQ(B_forget.dims().size(), 1);
int D = B_forget.dims()[0];
......
paddle/fluid/framework/ir/conv_bias_mkldnn_fuse_pass.cc
浏览文件 @ f7b55de9
......@@ -57,7 +57,7 @@ std::unique_ptr<ir::Graph> ConvBiasFusePass::ApplyImpl(
int found_conv_bias_count = 0;
auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
Graph* g) {
VLOG(4) << "handle ConvBias fuse";
VLOG(40) << "handle ConvBias fuse";
GET_IR_NODE_FROM_SUBGRAPH(conv_weight, conv_weight,
conv_bias_pattern); // Filter
GET_IR_NODE_FROM_SUBGRAPH(conv_out, conv_out, conv_bias_pattern); // tmp
......@@ -74,7 +74,7 @@ std::unique_ptr<ir::Graph> ConvBiasFusePass::ApplyImpl(
// check if fuse can be done and if MKL-DNN should be used
FuseOptions fuse_option = FindFuseOption(*conv, *eltwise);
if (fuse_option == DO_NOT_FUSE || fuse_option == FUSE_NATIVE) {
VLOG(3) << "do not perform conv+bias fuse";
VLOG(30) << "do not perform conv+bias fuse";
return;
}
......
paddle/fluid/framework/ir/conv_bn_fuse_pass.cc
浏览文件 @ f7b55de9
......@@ -121,7 +121,7 @@ std::unique_ptr<ir::Graph> ConvBNFusePass::ApplyImpl(
int found_conv_bn_count = 0;
auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
Graph* g) {
VLOG(4) << "handle ConvBN fuse";
VLOG(40) << "handle ConvBN fuse";
// conv, batch_norm,
// conv_weight, conv_out,
......@@ -133,7 +133,7 @@ std::unique_ptr<ir::Graph> ConvBNFusePass::ApplyImpl(
// check if fuse can be done and if MKL-DNN should be used
FuseOptions fuse_option = FindFuseOption(*conv, *batch_norm);
if (fuse_option == DO_NOT_FUSE) {
VLOG(3) << "do not perform conv+bn fuse";
VLOG(30) << "do not perform conv+bn fuse";
return;
}
......@@ -241,7 +241,7 @@ std::unique_ptr<ir::Graph> ConvEltwiseAddBNFusePass::ApplyImpl(
int found_conv_bn_count = 0;
auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
Graph* g) {
VLOG(4) << "handle ConvBN fuse";
VLOG(40) << "handle ConvBN fuse";
// conv, batch_norm,
// conv_weight, conv_out,
......
paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.cc
浏览文件 @ f7b55de9
......@@ -14,14 +14,15 @@
#include "paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.h"
#include <functional>
#include <utility>
#include <list>
#include <map>
#include <tuple>
#include "paddle/fluid/framework/ir/graph_traits.h"
namespace paddle {
namespace framework {
namespace ir {
namespace {
// The function keeps the graph consistent by replacing
// a node 'from' in the set of inputs nodes
......@@ -51,99 +52,382 @@ void CorrectGraphEdges(Graph* graph, Node* from, Node* to) {
}
}
}
} // namespace
using graph_ptr = std::unique_ptr<ir::Graph>;
graph_ptr ConvElementwiseAddMKLDNNFusePass::ApplyImpl(graph_ptr graph) const {
FusePassBase::Init(name_scope_, graph.get());
bool IsReachable(ir::Graph* graph, Node* from, Node* to) {
auto find_node = [](ir::Graph* graph, const Node* node) -> Node* {
for (auto n : graph->Nodes()) {
if (n == node) {
return n;
}
}
GraphPatternDetector gpd;
auto pattern = gpd.mutable_pattern();
return nullptr;
};
patterns::Conv conv_pattern{pattern, name_scope_};
auto conv_output = conv_pattern();
if (from == to) {
return true;
}
patterns::ElementwiseAdd elementwise_add_pattern{pattern, name_scope_};
elementwise_add_pattern(conv_output);
std::map<Node*, bool> visited;
conv_output->AsIntermediate();
for (auto& node : GraphTraits::DFS(*graph)) {
visited[&node] = false;
}
auto conv_op_has_bias = [](const Node& conv_op) -> std::pair<bool, Node*> {
auto bias_input_names = conv_op.Op()->Inputs();
auto bias_it = bias_input_names.find("Bias");
if (bias_it != std::end(bias_input_names)) {
bool has_bias = !bias_it->second.empty();
if (has_bias) {
auto conv_bias_names = bias_it->second;
auto conv_bias_names_it =
std::find_if(std::begin(conv_op.inputs), std::end(conv_op.inputs),
[&conv_bias_names](Node* n) -> bool {
return n->Name() == conv_bias_names[0];
});
return std::make_pair(has_bias, *conv_bias_names_it);
}
}
visited[from] = true;
return std::make_pair(false, nullptr);
};
std::list<Node*> queue;
queue.push_back(from);
auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
Graph* g) {
GET_IR_NODE_FROM_SUBGRAPH(conv_op, conv_op, conv_pattern);
GET_IR_NODE_FROM_SUBGRAPH(conv_input, conv_input, conv_pattern);
GET_IR_NODE_FROM_SUBGRAPH(conv_filter, conv_filter, conv_pattern);
GET_IR_NODE_FROM_SUBGRAPH(conv_output, conv_output, conv_pattern);
GET_IR_NODE_FROM_SUBGRAPH(elementwise_add_op, elementwise_add_op,
elementwise_add_pattern);
GET_IR_NODE_FROM_SUBGRAPH(elementwise_add_x, elementwise_add_x,
elementwise_add_pattern);
GET_IR_NODE_FROM_SUBGRAPH(elementwise_add_out, elementwise_add_out,
elementwise_add_pattern);
while (!queue.empty()) {
auto cur = find_node(graph, queue.front());
queue.pop_front();
if (FindFuseOption(*conv_op, *elementwise_add_op) != FUSE_MKLDNN) return;
if (!cur) return false;
OpDesc op_desc;
op_desc.SetType("conv2d");
for (auto n : cur->outputs) {
if (n == to) {
return true;
}
op_desc.SetInput("Input", {conv_input->Name()});
op_desc.SetInput("Filter", {conv_filter->Name()});
op_desc.SetInput("ResidualData", {elementwise_add_x->Name()});
op_desc.SetOutput("Output", {conv_output->Name()});
if (!visited[n]) {
visited[n] = true;
queue.push_back(n);
}
}
}
return false;
}
bool has_bias;
Node* conv_bias;
boost::optional<Node*> HasBias(const Node& op, const std::string& bias_name) {
auto bias_input_names = op.Op()->Inputs();
auto bias_it = bias_input_names.find(bias_name);
std::tie(has_bias, conv_bias) = conv_op_has_bias(*conv_op);
if (bias_it != std::end(bias_input_names)) {
bool has_bias = !bias_it->second.empty();
if (has_bias) {
op_desc.SetInput("Bias", {conv_bias->Name()});
auto bias_names = bias_it->second;
auto bias_names_it =
std::find_if(std::begin(op.inputs), std::end(op.inputs),
[&bias_names](Node* n) -> bool {
return n->Name() == bias_names[0];
});
return *bias_names_it;
}
}
for (const auto& attr : conv_op->Op()->GetAttrMap()) {
op_desc.SetAttr(attr.first, attr.second);
}
return boost::none;
}
op_desc.SetAttr("fuse_residual_connection", true);
ResidualConnectionMKLDNNFusePass::IdentityFuseHandle::IdentityFuseHandle(
const ResidualConnectionMKLDNNFusePass::CanFuseFunc& can_fuse_func,
const ResidualConnectionMKLDNNFusePass::IdentityConvFunc&
get_node_from_conv_op,
const ResidualConnectionMKLDNNFusePass::IdentityElementwiseAddFunc&
get_node_from_elementwise_add_op)
: fusion_stats{std::make_shared<int>(0)},
can_fuse_func{can_fuse_func},
get_node_from_conv_op{get_node_from_conv_op},
get_node_from_elementwise_add_op{get_node_from_elementwise_add_op} {}
void ResidualConnectionMKLDNNFusePass::IdentityFuseHandle::operator()(
const GraphPatternDetector::subgraph_t& subgraph, Graph* graph) {
Node* conv_op;
Node* conv_input;
Node* conv_filter;
Node* conv_output;
Node* elementwise_add_op;
Node* elementwise_add_identity;
Node* elementwise_add_out;
std::tie(conv_op, conv_input, conv_filter, conv_output) =
get_node_from_conv_op(subgraph);
std::tie(elementwise_add_op, elementwise_add_identity, elementwise_add_out) =
get_node_from_elementwise_add_op(subgraph);
if (!can_fuse_func(conv_op, elementwise_add_op)) return;
if (!IsReachable(graph, elementwise_add_identity, conv_output)) return;
OpDesc op_desc;
op_desc.SetType("conv2d");
op_desc.SetInput("Input", {conv_input->Name()});
op_desc.SetInput("Filter", {conv_filter->Name()});
op_desc.SetInput("ResidualData", {elementwise_add_identity->Name()});
op_desc.SetOutput("Output", {conv_output->Name()});
auto conv_bias = HasBias(*conv_op, "Bias");
if (conv_bias) {
op_desc.SetInput("Bias", {(*conv_bias)->Name()});
}
auto fused_conv_op = g->CreateOpNode(&op_desc);
for (const auto& attr : conv_op->Op()->GetAttrMap()) {
op_desc.SetAttr(attr.first, attr.second);
}
IR_NODE_LINK_TO(conv_input, fused_conv_op);
IR_NODE_LINK_TO(conv_filter, fused_conv_op);
IR_NODE_LINK_TO(elementwise_add_x, fused_conv_op);
IR_NODE_LINK_TO(fused_conv_op, conv_output);
op_desc.SetAttr("fuse_residual_connection", true);
if (has_bias) {
IR_NODE_LINK_TO(conv_bias, fused_conv_op);
}
auto fused_conv_op = graph->CreateOpNode(&op_desc);
CorrectGraphEdges(g, elementwise_add_out, conv_output);
GraphSafeRemoveNodes(g, {elementwise_add_out, conv_op, elementwise_add_op});
};
IR_NODE_LINK_TO(conv_input, fused_conv_op);
IR_NODE_LINK_TO(conv_filter, fused_conv_op);
IR_NODE_LINK_TO(elementwise_add_identity, fused_conv_op);
IR_NODE_LINK_TO(fused_conv_op, conv_output);
gpd(graph.get(), handler);
if (conv_bias) {
IR_NODE_LINK_TO((*conv_bias), fused_conv_op);
}
CorrectGraphEdges(graph, elementwise_add_out, conv_output);
GraphSafeRemoveNodes(graph,
{elementwise_add_out, conv_op, elementwise_add_op});
(*fusion_stats)++;
}
ResidualConnectionMKLDNNFusePass::ProjectionFuseHandle::ProjectionFuseHandle(
const ResidualConnectionMKLDNNFusePass::CanFuseFunc& can_fuse_func,
const ResidualConnectionMKLDNNFusePass::ProjectionConvFunc&
get_node_from_conv_x_op,
const ResidualConnectionMKLDNNFusePass::ProjectionConvFunc&
get_node_from_conv_y_op,
const ResidualConnectionMKLDNNFusePass::ProjectionElementwiseAddFunc&
get_node_from_elementwise_add_op)
: fusion_stats{std::make_shared<int>(0)},
can_fuse_func{can_fuse_func},
get_node_from_conv_x_op{get_node_from_conv_x_op},
get_node_from_conv_y_op{get_node_from_conv_y_op},
get_node_from_elementwise_add_op{get_node_from_elementwise_add_op} {}
void ResidualConnectionMKLDNNFusePass::ProjectionFuseHandle::operator()(
const GraphPatternDetector::subgraph_t& subgraph, Graph* graph) {
Node* conv_x_op;
Node* conv_x_input;
Node* conv_x_filter;
Node* conv_x_output;
Node* conv_y_op;
Node* conv_y_input;
Node* conv_y_filter;
Node* conv_y_output;
Node* elementwise_add_op;
Node* elementwise_add_out;
std::tie(conv_x_op, conv_x_input, conv_x_filter, conv_x_output) =
get_node_from_conv_x_op(subgraph);
std::tie(conv_y_op, conv_y_input, conv_y_filter, conv_y_output) =
get_node_from_conv_y_op(subgraph);
std::tie(elementwise_add_op, elementwise_add_out) =
get_node_from_elementwise_add_op(subgraph);
if (!can_fuse_func(conv_x_op, elementwise_add_op)) return;
if (!can_fuse_func(conv_y_op, elementwise_add_op)) return;
Node* projection_node;
Node* residual_conv_op;
Node* residual_conv_input;
Node* residual_conv_filter;
Node* residual_conv_output;
if (IsReachable(graph, conv_x_input, conv_y_output)) {
projection_node = conv_x_output;
residual_conv_op = conv_y_op;
residual_conv_input = conv_y_input;
residual_conv_filter = conv_y_filter;
residual_conv_output = conv_y_output;
} else if (IsReachable(graph, conv_y_input, conv_x_output)) {
projection_node = conv_y_output;
residual_conv_op = conv_x_op;
residual_conv_input = conv_x_input;
residual_conv_filter = conv_x_filter;
residual_conv_output = conv_x_output;
} else {
return;
}
OpDesc op_desc;
op_desc.SetType("conv2d");
op_desc.SetInput("Input", {residual_conv_input->Name()});
op_desc.SetInput("Filter", {residual_conv_filter->Name()});
op_desc.SetInput("ResidualData", {projection_node->Name()});
op_desc.SetOutput("Output", {residual_conv_output->Name()});
auto residual_conv_bias = HasBias(*residual_conv_op, "Bias");
if (residual_conv_bias) {
op_desc.SetInput("Bias", {(*residual_conv_bias)->Name()});
}
for (const auto& attr : residual_conv_op->Op()->GetAttrMap()) {
op_desc.SetAttr(attr.first, attr.second);
}
op_desc.SetAttr("fuse_residual_connection", true);
auto fused_conv_op = graph->CreateOpNode(&op_desc);
IR_NODE_LINK_TO(residual_conv_input, fused_conv_op);
IR_NODE_LINK_TO(residual_conv_filter, fused_conv_op);
IR_NODE_LINK_TO(projection_node, fused_conv_op);
IR_NODE_LINK_TO(fused_conv_op, residual_conv_output);
if (residual_conv_bias) {
IR_NODE_LINK_TO((*residual_conv_bias), fused_conv_op);
}
CorrectGraphEdges(graph, elementwise_add_out, residual_conv_output);
GraphSafeRemoveNodes(
graph, {elementwise_add_out, residual_conv_op, elementwise_add_op});
(*fusion_stats)++;
}
std::tuple<Node*, Node*, Node*, Node*>
ResidualConnectionMKLDNNFusePass::GetNodesFromConv(
const patterns::Conv& conv_pattern,
const GraphPatternDetector::subgraph_t& subgraph) const {
GET_IR_NODE_FROM_SUBGRAPH(conv_op, conv_op, conv_pattern);
GET_IR_NODE_FROM_SUBGRAPH(conv_input, conv_input, conv_pattern);
GET_IR_NODE_FROM_SUBGRAPH(conv_filter, conv_filter, conv_pattern);
GET_IR_NODE_FROM_SUBGRAPH(conv_output, conv_output, conv_pattern);
return std::make_tuple(conv_op, conv_input, conv_filter, conv_output);
}
GraphWithStats ResidualConnectionMKLDNNFusePass::FuseConvAsX(
const std::string& name_scope,
const GraphWithStats& graph_with_stats) const {
ir::Graph* graph;
int stats;
std::tie(graph, stats) = graph_with_stats;
GraphPatternDetector gpd;
auto pattern = gpd.mutable_pattern();
patterns::Conv conv_pattern{pattern, name_scope};
auto conv_output = conv_pattern();
patterns::ElementwiseAdd elementwise_add_pattern{pattern, name_scope};
elementwise_add_pattern(
conv_output,
pattern->NewNode(elementwise_add_pattern.elementwise_add_y_repr()));
conv_output->AsIntermediate();
auto get_node_from_elementwise_add = [&elementwise_add_pattern](
const GraphPatternDetector::subgraph_t& subgraph)
-> std::tuple<Node*, Node*, Node*> {
GET_IR_NODE_FROM_SUBGRAPH(elementwise_add_op, elementwise_add_op,
elementwise_add_pattern);
GET_IR_NODE_FROM_SUBGRAPH(elementwise_add_y, elementwise_add_y,
elementwise_add_pattern);
GET_IR_NODE_FROM_SUBGRAPH(elementwise_add_out, elementwise_add_out,
elementwise_add_pattern);
return std::make_tuple(elementwise_add_op, elementwise_add_y,
elementwise_add_out);
};
return ExecuteHandleOnGraph<IdentityFuseHandle>(
&gpd, graph_with_stats,
[this, &conv_pattern](const GraphPatternDetector::subgraph_t& subgraph) {
return GetNodesFromConv(conv_pattern, subgraph);
},
get_node_from_elementwise_add);
}
GraphWithStats ResidualConnectionMKLDNNFusePass::FuseConvAsY(
const std::string& name_scope,
const GraphWithStats& graph_with_stats) const {
GraphPatternDetector gpd;
auto pattern = gpd.mutable_pattern();
patterns::Conv conv_pattern{pattern, name_scope};
auto conv_output = conv_pattern();
patterns::ElementwiseAdd elementwise_add_pattern{pattern, name_scope};
elementwise_add_pattern(
pattern->NewNode(elementwise_add_pattern.elementwise_add_x_repr()),
conv_output);
conv_output->AsIntermediate();
auto get_node_from_elementwise_add = [&elementwise_add_pattern](
const GraphPatternDetector::subgraph_t& subgraph)
-> std::tuple<Node*, Node*, Node*> {
GET_IR_NODE_FROM_SUBGRAPH(elementwise_add_op, elementwise_add_op,
elementwise_add_pattern);
GET_IR_NODE_FROM_SUBGRAPH(elementwise_add_x, elementwise_add_x,
elementwise_add_pattern);
GET_IR_NODE_FROM_SUBGRAPH(elementwise_add_out, elementwise_add_out,
elementwise_add_pattern);
return std::make_tuple(elementwise_add_op, elementwise_add_x,
elementwise_add_out);
};
return ExecuteHandleOnGraph<IdentityFuseHandle>(
&gpd, graph_with_stats,
[this, &conv_pattern](const GraphPatternDetector::subgraph_t& subgraph) {
return GetNodesFromConv(conv_pattern, subgraph);
},
get_node_from_elementwise_add);
}
GraphWithStats ResidualConnectionMKLDNNFusePass::FuseProjectionConv(
const std::string& name_scope,
const GraphWithStats& graph_with_stats) const {
GraphPatternDetector gpd;
auto pattern = gpd.mutable_pattern();
patterns::Conv conv_x_pattern{pattern, name_scope};
auto conv_x_output = conv_x_pattern();
patterns::Conv conv_y_pattern{pattern, name_scope};
auto conv_y_output = conv_y_pattern();
patterns::ElementwiseAdd elementwise_add_pattern{pattern, name_scope};
elementwise_add_pattern(conv_x_output, conv_y_output);
conv_x_output->AsIntermediate();
conv_y_output->AsIntermediate();
auto get_node_from_elementwise_add = [&elementwise_add_pattern](
const GraphPatternDetector::subgraph_t& subgraph)
-> std::tuple<Node*, Node*> {
GET_IR_NODE_FROM_SUBGRAPH(elementwise_add_op, elementwise_add_op,
elementwise_add_pattern);
GET_IR_NODE_FROM_SUBGRAPH(elementwise_add_out, elementwise_add_out,
elementwise_add_pattern);
return std::make_tuple(elementwise_add_op, elementwise_add_out);
};
return ExecuteHandleOnGraph<ProjectionFuseHandle>(
&gpd, graph_with_stats,
[this,
&conv_x_pattern](const GraphPatternDetector::subgraph_t& subgraph) {
return GetNodesFromConv(conv_x_pattern, subgraph);
},
[this,
&conv_y_pattern](const GraphPatternDetector::subgraph_t& subgraph) {
return GetNodesFromConv(conv_y_pattern, subgraph);
},
get_node_from_elementwise_add);
}
graph_ptr ResidualConnectionMKLDNNFusePass::ApplyImpl(graph_ptr graph) const {
FusePassBase::Init(name_scope_, graph.get());
auto fused_graph_with_stats = FuseConvAsY(
name_scope_,
FuseConvAsX(
name_scope_,
FuseProjectionConv(name_scope_, std::make_pair(graph.get(), 0))));
std::cout << "Fused graph " << fused_graph_with_stats.second << std::endl;
AddStatis(fused_graph_with_stats.second);
return graph;
}
} // namespace ir
......@@ -151,4 +435,4 @@ graph_ptr ConvElementwiseAddMKLDNNFusePass::ApplyImpl(graph_ptr graph) const {
} // namespace paddle
REGISTER_PASS(conv_elementwise_add_mkldnn_fuse_pass,
paddle::framework::ir::ConvElementwiseAddMKLDNNFusePass);
paddle::framework::ir::ResidualConnectionMKLDNNFusePass);
paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.h
浏览文件 @ f7b55de9
......@@ -15,24 +15,119 @@
#pragma once
#include <string>
#include <tuple>
#include <utility>
#include "paddle/fluid/framework/ir/fuse_pass_base.h"
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/graph_pattern_detector.h"
#include <boost/optional.hpp>
namespace paddle {
namespace framework {
namespace ir {
class ConvElementwiseAddMKLDNNFusePass : public FusePassBase {
using graph_ptr = std::unique_ptr<ir::Graph>;
using GraphWithStats = std::pair<ir::Graph*, int>;
void CorrectGraphEdges(Graph* graph, Node* from, Node* to);
bool IsReachable(ir::Graph* graph, Node* from, Node* to);
boost::optional<Node*> HasBias(const Node& op, const std::string& bias_name);
class ResidualConnectionMKLDNNFusePass : public FusePassBase {
private:
GraphWithStats FuseConvAsX(const std::string& name_scope,
const GraphWithStats& graph_with_stats) const;
GraphWithStats FuseConvAsY(const std::string& name_scope,
const GraphWithStats& graph_with_stats) const;
GraphWithStats FuseProjectionConv(
const std::string& name_scope,
const GraphWithStats& graph_with_stats) const;
template <typename RetType>
using GetNodeFunc =
std::function<RetType(const GraphPatternDetector::subgraph_t& subgraph)>;
using IdentityConvFunc = GetNodeFunc<std::tuple<Node*, Node*, Node*, Node*>>;
using IdentityElementwiseAddFunc =
GetNodeFunc<std::tuple<Node*, Node*, Node*>>;
using ProjectionConvFunc = IdentityConvFunc;
using ProjectionElementwiseAddFunc = GetNodeFunc<std::tuple<Node*, Node*>>;
using CanFuseFunc = std::function<bool(Node*, Node*)>;
std::tuple<Node*, Node*, Node*, Node*> GetNodesFromConv(
const patterns::Conv& conv_pattern,
const GraphPatternDetector::subgraph_t& subgraph) const;
std::tuple<Node*, Node*, Node*, Node*> GetNodesFromProjectionConv(
const patterns::Conv& conv_pattern,
const GraphPatternDetector::subgraph_t& subgraph) const;
template <typename HandleType, typename... OpFuncs>
GraphWithStats ExecuteHandleOnGraph(GraphPatternDetector* gpd,
const GraphWithStats& graph_with_stats,
OpFuncs&&... op_funcs) const {
ir::Graph* graph;
int stats;
std::tie(graph, stats) = graph_with_stats;
auto can_fuse = [this](Node* op1, Node* op2) -> bool {
return this->FindFuseOption(*op1, *op2) == FUSE_MKLDNN;
};
auto fuse_handle = HandleType{can_fuse, std::forward<OpFuncs>(op_funcs)...};
(*gpd)(graph, fuse_handle);
return std::make_pair(graph, stats + fuse_handle.get_stats());
}
struct IdentityFuseHandle {
IdentityFuseHandle(
const CanFuseFunc& can_fuse_func,
const IdentityConvFunc& get_node_from_conv_op,
const IdentityElementwiseAddFunc& get_node_from_elementwise_add_op);
void operator()(const GraphPatternDetector::subgraph_t& subgraph,
Graph* graph);
int get_stats() const { return *fusion_stats; }
private:
std::shared_ptr<int> fusion_stats;
CanFuseFunc can_fuse_func;
IdentityConvFunc get_node_from_conv_op;
IdentityElementwiseAddFunc get_node_from_elementwise_add_op;
};
struct ProjectionFuseHandle {
ProjectionFuseHandle(
const CanFuseFunc& can_fuse_func,
const ProjectionConvFunc& get_node_from_conv_x_op,
const ProjectionConvFunc& get_node_from_conv_y_op,
const ProjectionElementwiseAddFunc& get_node_from_elementwise_add_op);
void operator()(const GraphPatternDetector::subgraph_t& subgraph,
Graph* graph);
int get_stats() const { return *fusion_stats; }
private:
std::shared_ptr<int> fusion_stats;
CanFuseFunc can_fuse_func;
ProjectionConvFunc get_node_from_conv_x_op;
ProjectionConvFunc get_node_from_conv_y_op;
ProjectionElementwiseAddFunc get_node_from_elementwise_add_op;
};
public:
virtual ~ConvElementwiseAddMKLDNNFusePass() {}
virtual ~ResidualConnectionMKLDNNFusePass() {}
protected:
std::unique_ptr<ir::Graph> ApplyImpl(std::unique_ptr<ir::Graph> graph) const;
std::unique_ptr<ir::Graph> ApplyImpl(graph_ptr graph) const;
const std::string name_scope_{"residual_connections_fuse_pass"};
const std::string name_scope_{"residual_connection_fuse_pass"};
};
} // namespace ir
} // namespace framework
} // namespace paddle
paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass_tester.cc
浏览文件 @ f7b55de9
......@@ -40,7 +40,7 @@ void SetOp(ProgramDesc* prog, const std::string& type,
op->SetOutput(output.first, {output.second});
}
struct IsReachable {
struct TestIsReachable {
using func = std::function<bool(const std::string&, const std::string&)>;
auto operator()(const std::unique_ptr<ir::Graph>& graph) -> func {
......@@ -89,7 +89,9 @@ struct IsReachable {
}
};
void AssertOpsCount(const std::unique_ptr<ir::Graph>& graph) {
void AssertOpsCount(const std::unique_ptr<ir::Graph>& graph,
int expected_conv_count,
int expected_elementwise_add_count = 0) {
int conv_count = 0;
int elementwise_add_count = 0;
......@@ -101,8 +103,8 @@ void AssertOpsCount(const std::unique_ptr<ir::Graph>& graph) {
++elementwise_add_count;
}
}
EXPECT_EQ(conv_count, 1);
EXPECT_EQ(elementwise_add_count, 0);
EXPECT_EQ(conv_count, expected_conv_count);
EXPECT_EQ(elementwise_add_count, expected_elementwise_add_count);
}
ProgramDesc BuildProgramDesc(const std::vector<std::string>& transient_vars,
......@@ -127,22 +129,13 @@ ProgramDesc BuildProgramDesc(const std::vector<std::string>& transient_vars,
return prog;
}
} // namespace
TEST(ConvElementwiseAddMKLDNNFusePass, ConvolutionWithElementwiseAddRelu) {
auto prog =
BuildProgramDesc({"a", "b", "c", "d", "e", "f"}, {"bias", "weights"});
SetOp(&prog, "conv2d",
{{"Input", "a"}, {"Bias", "bias"}, {"Filter", "weights"}},
{"Output", "b"});
SetOp(&prog, "elementwise_add", {{"X", "b"}, {"Y", "c"}}, {"Out", "d"});
SetOp(&prog, "relu", {{"X", "d"}}, {"Out", "e"});
std::unique_ptr<ir::Graph> graph(new ir::Graph(prog));
void RunPassAndAssert(ProgramDesc* prog, const std::string& from,
const std::string& to, int expected_conv_num) {
std::unique_ptr<ir::Graph> graph(new ir::Graph(*prog));
IsReachable is_reachable;
EXPECT_TRUE(is_reachable(graph)("a", "relu"));
TestIsReachable is_reachable;
EXPECT_TRUE(is_reachable(graph)(from, to));
auto pass =
PassRegistry::Instance().Get("conv_elementwise_add_mkldnn_fuse_pass");
......@@ -150,82 +143,87 @@ TEST(ConvElementwiseAddMKLDNNFusePass, ConvolutionWithElementwiseAddRelu) {
graph = pass->Apply(std::move(graph));
int current_nodes_num = graph->Nodes().size();
EXPECT_TRUE(is_reachable(graph)("a", "relu"));
EXPECT_TRUE(is_reachable(graph)(from, to));
EXPECT_EQ(original_nodes_num - nodes_removed + nodes_added,
current_nodes_num);
AssertOpsCount(graph);
AssertOpsCount(graph, expected_conv_num);
}
} // namespace
TEST(ConvElementwiseAddMKLDNNFusePass,
ConvolutionWithElementwiseAddReluNoBias) {
auto prog = BuildProgramDesc({"a", "b", "c", "d", "e"}, {"weights"});
SetOp(&prog, "conv2d", {{"Input", "a"}, {"Filter", "weights"}},
{"Output", "b"});
SetOp(&prog, "elementwise_add", {{"X", "b"}, {"Y", "c"}}, {"Out", "d"});
SetOp(&prog, "relu", {{"X", "d"}}, {"Out", "e"});
std::unique_ptr<ir::Graph> graph(new ir::Graph(prog));
TEST(ConvElementwiseAddMKLDNNFusePass, ConvolutionAsYWithElementwiseAddRelu) {
auto prog = BuildProgramDesc({"a", "b", "c", "d", "e"}, {"bias", "weights"});
IsReachable is_reachable;
SetOp(&prog, "sigmoid", {{"X", "a"}}, {"Out", "b"});
SetOp(&prog, "conv2d",
{{"Input", "b"}, {"Bias", "bias"}, {"Filter", "weights"}},
{"Output", "c"});
EXPECT_TRUE(is_reachable(graph)("a", "relu"));
SetOp(&prog, "elementwise_add", {{"X", "a"}, {"Y", "c"}}, {"Out", "d"});
SetOp(&prog, "relu", {{"X", "d"}}, {"Out", "e"});
auto pass =
PassRegistry::Instance().Get("conv_elementwise_add_mkldnn_fuse_pass");
int original_nodes_num = graph->Nodes().size();
graph = pass->Apply(std::move(graph));
int current_nodes_num = graph->Nodes().size();
RunPassAndAssert(&prog, "a", "relu", 1);
}
EXPECT_TRUE(is_reachable(graph)("a", "relu"));
TEST(ConvElementwiseAddMKLDNNFusePass,
ConvolutionAsYWithElementwiseAddReluNoBias) {
auto prog = BuildProgramDesc({"a", "b", "c", "d", "e"}, {"weights"});
EXPECT_EQ(original_nodes_num - nodes_removed + nodes_added,
current_nodes_num);
SetOp(&prog, "sigmoid", {{"X", "a"}}, {"Out", "b"});
SetOp(&prog, "conv2d", {{"Input", "b"}, {"Filter", "weights"}},
{"Output", "c"});
SetOp(&prog, "elementwise_add", {{"X", "a"}, {"Y", "c"}}, {"Out", "d"});
SetOp(&prog, "relu", {{"X", "d"}}, {"Out", "e"});
AssertOpsCount(graph);
RunPassAndAssert(&prog, "a", "relu", 1);
}
TEST(ConvElementwiseAddMKLDNNFusePass, ConvolutionElementwiseAdd) {
auto prog = BuildProgramDesc({"a", "b", "c", "d"}, {"bias", "weights"});
TEST(ConvElementwiseAddMKLDNNFusePass, ConvolutionAsXWithElementwiseAddRelu) {
auto prog = BuildProgramDesc({"a", "b", "c", "d", "e"}, {"bias", "weights"});
SetOp(&prog, "sigmoid", {{"X", "a"}}, {"Out", "b"});
SetOp(&prog, "conv2d",
{{"Input", "a"}, {"Bias", "bias"}, {"Filter", "weights"}},
{"Output", "b"});
SetOp(&prog, "elementwise_add", {{"X", "b"}, {"Y", "c"}}, {"Out", "d"});
{{"Input", "b"}, {"Bias", "bias"}, {"Filter", "weights"}},
{"Output", "c"});
std::unique_ptr<ir::Graph> graph(new ir::Graph(prog));
SetOp(&prog, "elementwise_add", {{"X", "c"}, {"Y", "a"}}, {"Out", "d"});
SetOp(&prog, "relu", {{"X", "d"}}, {"Out", "e"});
IsReachable is_reachable;
EXPECT_TRUE(is_reachable(graph)("a", "d"));
RunPassAndAssert(&prog, "a", "relu", 1);
}
auto pass =
PassRegistry::Instance().Get("conv_elementwise_add_mkldnn_fuse_pass");
int original_nodes_num = graph->Nodes().size();
graph = pass->Apply(std::move(graph));
int current_nodes_num = graph->Nodes().size();
TEST(ConvElementwiseAddMKLDNNFusePass,
ConvolutionAsXWithElementwiseAddReluNoBias) {
auto prog = BuildProgramDesc({"a", "b", "c", "d", "e"}, {"weights"});
EXPECT_FALSE(is_reachable(graph)("a", "d"));
SetOp(&prog, "sigmoid", {{"X", "a"}}, {"Out", "b"});
SetOp(&prog, "conv2d", {{"Input", "b"}, {"Filter", "weights"}},
{"Output", "c"});
SetOp(&prog, "elementwise_add", {{"X", "c"}, {"Y", "a"}}, {"Out", "d"});
SetOp(&prog, "relu", {{"X", "d"}}, {"Out", "e"});
EXPECT_EQ(original_nodes_num - nodes_removed + nodes_added,
current_nodes_num);
AssertOpsCount(graph);
RunPassAndAssert(&prog, "a", "relu", 1);
}
TEST(ConvElementwiseAddMKLDNNFusePass, SigmoidConvolutionAddElementwiseRelu) {
TEST(ConvElementwiseAddMKLDNNFusePass, NoFusion) {
auto prog =
BuildProgramDesc({"a", "b", "c", "d", "e", "f"}, {"bias", "weights"});
BuildProgramDesc({"a", "b", "c", "d", "e", "f", "g"}, {"weights"});
SetOp(&prog, "sigmoid", {{"X", "a"}}, {"Out", "b"});
SetOp(&prog, "conv2d",
{{"Input", "b"}, {"Bias", "bias"}, {"Filter", "weights"}},
SetOp(&prog, "conv2d", {{"Input", "b"}, {"Filter", "weights"}},
{"Output", "c"});
SetOp(&prog, "elementwise_add", {{"X", "c"}, {"Y", "d"}}, {"Out", "e"});
SetOp(&prog, "relu", {{"X", "e"}}, {"Out", "f"});
std::unique_ptr<ir::Graph> graph(new ir::Graph(prog));
SetOp(&prog, "conv2d", {{"Input", "d"}, {"Filter", "weights"}},
{"Output", "e"});
IsReachable is_reachable;
SetOp(&prog, "elementwise_add", {{"X", "c"}, {"Y", "e"}}, {"Out", "f"});
SetOp(&prog, "relu", {{"X", "f"}}, {"Out", "g"});
EXPECT_TRUE(is_reachable(graph)("a", "f"));
std::unique_ptr<ir::Graph> graph(new ir::Graph(prog));
TestIsReachable is_reachable;
EXPECT_TRUE(is_reachable(graph)("a", "g"));
auto pass =
PassRegistry::Instance().Get("conv_elementwise_add_mkldnn_fuse_pass");
......@@ -233,11 +231,10 @@ TEST(ConvElementwiseAddMKLDNNFusePass, SigmoidConvolutionAddElementwiseRelu) {
graph = pass->Apply(std::move(graph));
int current_nodes_num = graph->Nodes().size();
EXPECT_TRUE(is_reachable(graph)("a", "f"));
EXPECT_TRUE(is_reachable(graph)("a", "g"));
EXPECT_EQ(original_nodes_num, current_nodes_num);
EXPECT_EQ(original_nodes_num - nodes_removed + nodes_added,
current_nodes_num);
AssertOpsCount(graph);
AssertOpsCount(graph, 2, 1);
}
} // namespace ir
......
paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.cc
浏览文件 @ f7b55de9
......@@ -38,7 +38,7 @@ std::unique_ptr<ir::Graph> ConvReLUFusePass::ApplyImpl(
int found_conv_relu_count = 0;
auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
Graph* g) {
VLOG(4) << "handle ConvReLU fuse";
VLOG(40) << "handle ConvReLU fuse";
GET_IR_NODE_FROM_SUBGRAPH(conv_weight, conv_weight,
conv_relu_pattern); // Filter
GET_IR_NODE_FROM_SUBGRAPH(conv_out, conv_out, conv_relu_pattern); // tmp
......@@ -48,7 +48,7 @@ std::unique_ptr<ir::Graph> ConvReLUFusePass::ApplyImpl(
FuseOptions fuse_option = FindFuseOption(*conv, *relu);
if (fuse_option == DO_NOT_FUSE) {
VLOG(3) << "do not perform conv+relu fuse";
VLOG(30) << "do not perform conv+relu fuse";
return;
}
......
paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.h
浏览文件 @ f7b55de9
......@@ -31,7 +31,8 @@ class ConvReLUFusePass : public FusePassBase {
virtual ~ConvReLUFusePass() {}
protected:
std::unique_ptr<ir::Graph> ApplyImpl(std::unique_ptr<ir::Graph> graph) const;
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
};
} // namespace ir
......
paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass_tester.cc
浏览文件 @ f7b55de9
......@@ -15,6 +15,7 @@
#include "paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.h"
#include <gtest/gtest.h>
#include "paddle/fluid/framework/op_proto_maker.h"
namespace paddle {
namespace framework {
......@@ -36,6 +37,8 @@ void SetOp(ProgramDesc* prog, const std::string& type, const std::string& name,
op->SetInput("X", inputs);
}
op->SetOutput("Out", outputs);
op->SetAttr(OpProtoAndCheckerMaker::OpRoleAttrName(),
static_cast<int>(OpRole::kForward));
}
// a->OP0->b
......
paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass.cc 0 → 100644
浏览文件 @ f7b55de9
/* 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/framework/ir/depthwise_conv_mkldnn_pass.h"
#include "paddle/fluid/framework/ir/graph_pattern_detector.h"
namespace paddle {
namespace framework {
namespace ir {
#define GET_NODE(id, pattern) \
PADDLE_ENFORCE(subgraph.count(pattern.RetrieveNode(#id)), \
"pattern has no Node called %s", #id); \
auto* id = subgraph.at(pattern.RetrieveNode(#id)); \
PADDLE_ENFORCE_NOT_NULL(id, "subgraph has no node %s", #id);
std::unique_ptr<ir::Graph> DepthwiseConvMKLDNNPass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
PADDLE_ENFORCE(graph.get());
FusePassBase::Init("depthwise_conv_mkldnn_pass", graph.get());
GraphPatternDetector gpd;
auto* pattern = gpd.mutable_pattern();
pattern->NewNode("depthwise_conv")
->assert_is_op("depthwise_conv2d")
->assert_op_attr("use_mkldnn", true);
int found_depthwise_conv_mkldnn_count = 0;
auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
Graph* g) {
VLOG(30) << "handle DepthwiseConvMKLDNN fuse";
GET_NODE(depthwise_conv, (*pattern));
depthwise_conv->Op()->SetType("conv2d");
found_depthwise_conv_mkldnn_count++;
};
gpd(graph.get(), handler);
AddStatis(found_depthwise_conv_mkldnn_count);
return graph;
}
} // namespace ir
} // namespace framework
} // namespace paddle
REGISTER_PASS(depthwise_conv_mkldnn_pass,
paddle::framework::ir::DepthwiseConvMKLDNNPass);
paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass.h 0 → 100644
浏览文件 @ f7b55de9
/* 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/ir/fuse_pass_base.h"
namespace paddle {
namespace framework {
namespace ir {
class DepthwiseConvMKLDNNPass : public FusePassBase {
public:
virtual ~DepthwiseConvMKLDNNPass() {}
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
};
} // namespace ir
} // namespace framework
} // namespace paddle
paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass_tester.cc 0 → 100644
浏览文件 @ f7b55de9
// 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/framework/ir/depthwise_conv_mkldnn_pass.h"
#include <gtest/gtest.h>
namespace paddle {
namespace framework {
namespace ir {
void SetOp(ProgramDesc* prog, const std::string& type, const std::string& name,
const std::vector<std::string>& inputs,
const std::vector<std::string>& outputs, bool use_mkldnn = false) {
auto* op = prog->MutableBlock(0)->AppendOp();
op->SetType(type);
op->SetAttr("use_mkldnn", use_mkldnn);
op->SetAttr("name", name);
op->SetInput("Input", {inputs[0]});
op->SetInput("Filter", {inputs[1]});
op->SetInput("Bias", {inputs[2]});
op->SetOutput("Out", outputs);
}
// (a, weights, bias)->depthwise conv mkldnn->b
// (b, weights2, bias2)->depthwise conv no mkldnn->c
// (c, weights3, bias3)->conv mkldnn->d
// (d, weights3, bias3)->conv no mkldnn->e
ProgramDesc BuildProgramDesc() {
ProgramDesc prog;
for (auto& v : std::vector<std::string>(
{"a", "b", "c", "d", "e", "weights", "bias", "weights2", "bias2",
"weights3", "bias3", "weights4", "bias4"})) {
auto* var = prog.MutableBlock(0)->Var(v);
var->SetType(proto::VarType::SELECTED_ROWS);
if (v == "weights" || v == "bias" || v == "weights2" || v == "bias2" ||
v == "weights3" || v == "bias3" || v == "weights4" || v == "bias4") {
var->SetPersistable(true);
}
}
// depthwise conv with MKL-DNN
SetOp(&prog, "depthwise_conv2d", "conv1",
std::vector<std::string>({"a", "weights", "bias"}),
std::vector<std::string>({"b"}), true);
// depthwise conv without MKL-DNN
SetOp(&prog, "depthwise_conv2d", "conv2",
std::vector<std::string>({"b", "weights2", "bias2"}),
std::vector<std::string>({"c"}), false);
// conv with MKL-DNN
SetOp(&prog, "conv2d", "conv3",
std::vector<std::string>({"c", "weights3", "bias3"}),
std::vector<std::string>({"d"}), true);
// conv without MKL-dNN
SetOp(&prog, "conv2d", "conv4",
std::vector<std::string>({"d", "weights4", "bias4"}),
std::vector<std::string>({"e"}), false);
return prog;
}
TEST(DepthwiseConvMKLDNNPass, basic) {
auto prog = BuildProgramDesc();
std::unique_ptr<ir::Graph> graph(new ir::Graph(prog));
auto pass = PassRegistry::Instance().Get("depthwise_conv_mkldnn_pass");
struct counters {
int mkldnn_depthwise_conv_nodes;
int other_depthwise_conv_nodes;
int mkldnn_conv_nodes;
int other_conv_nodes;
};
counters before{1, 1, 1, 1};
graph = pass->Apply(std::move(graph));
// initialize counters before loop
counters after{0, 0, 0, 0};
for (auto* node : graph->Nodes()) {
if (node->IsOp()) {
auto* op = node->Op();
if (op->Type() == "conv2d") {
if (boost::get<bool>(op->GetAttr("use_mkldnn")))
after.mkldnn_conv_nodes++;
else
after.other_conv_nodes++;
} else if (op->Type() == "depthwise_conv2d") {
if (boost::get<bool>(op->GetAttr("use_mkldnn")))
after.mkldnn_depthwise_conv_nodes++;
else
after.other_depthwise_conv_nodes++;
}
}
}
EXPECT_EQ(after.other_depthwise_conv_nodes,
before.other_depthwise_conv_nodes);
EXPECT_EQ(after.other_conv_nodes, before.other_conv_nodes);
EXPECT_EQ(after.mkldnn_depthwise_conv_nodes,
before.mkldnn_depthwise_conv_nodes - 1);
EXPECT_EQ(after.mkldnn_conv_nodes, before.mkldnn_conv_nodes + 1);
}
} // namespace ir
} // namespace framework
} // namespace paddle
USE_PASS(depthwise_conv_mkldnn_pass);
paddle/fluid/framework/ir/fc_fuse_pass.cc
浏览文件 @ f7b55de9
......@@ -39,7 +39,7 @@ std::unique_ptr<ir::Graph> FCFusePass::ApplyImpl(
int found_fc_count = 0;
auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
Graph* g) {
VLOG(4) << "handle FC fuse";
VLOG(40) << "handle FC fuse";
GET_IR_NODE_FROM_SUBGRAPH(w, w, fc_pattern);
GET_IR_NODE_FROM_SUBGRAPH(fc_bias, bias, fc_pattern);
GET_IR_NODE_FROM_SUBGRAPH(fc_out, Out, fc_pattern);
......@@ -57,6 +57,7 @@ std::unique_ptr<ir::Graph> FCFusePass::ApplyImpl(
desc.SetInput("W", std::vector<std::string>({fc_Y_in}));
desc.SetInput("Bias", std::vector<std::string>({fc_bias_in}));
desc.SetOutput("Out", std::vector<std::string>({fc_out_out}));
desc.SetAttr("in_num_col_dims", mul->Op()->GetAttr("x_num_col_dims"));
desc.SetType("fc");
auto fc_node = g->CreateOpNode(&desc); // OpDesc will be copied.
GraphSafeRemoveNodes(graph.get(), {mul, elementwise_add, mul_out});
......
paddle/fluid/framework/ir/fc_fuse_pass_tester.cc
浏览文件 @ f7b55de9
......@@ -15,6 +15,7 @@
#include "paddle/fluid/framework/ir/fc_fuse_pass.h"
#include <gtest/gtest.h>
#include "paddle/fluid/framework/op_proto_maker.h"
namespace paddle {
namespace framework {
......@@ -28,10 +29,13 @@ void SetOp(ProgramDesc* prog, const std::string& type,
if (type == "mul") {
op->SetInput("X", {inputs[0]});
op->SetInput("Y", {inputs[1]});
op->SetAttr("x_num_col_dims", {1});
} else if (type == "elementwise_add") {
op->SetInput("X", inputs);
}
op->SetOutput("Out", outputs);
op->SetAttr(OpProtoAndCheckerMaker::OpRoleAttrName(),
static_cast<int>(OpRole::kForward));
}
// a->OP0->b
......
paddle/fluid/framework/ir/fuse_elewise_add_act_pass.cc
浏览文件 @ f7b55de9
......@@ -61,7 +61,7 @@ std::unique_ptr<ir::Graph> FuseElewiseAddActPass::FuseElewiseAddAct(
auto handler = [&](const GraphPatternDetector::subgraph_t &subgraph,
Graph *g) {
VLOG(4) << "handle FuseElewiseAddAct fuse";
VLOG(40) << "handle FuseElewiseAddAct fuse";
GET_IR_NODE_FROM_SUBGRAPH(ele_y, ele_y, elewise_add_act_pattern);
GET_IR_NODE_FROM_SUBGRAPH(ele_out, elewise_add_out,
elewise_add_act_pattern);
......@@ -77,10 +77,10 @@ std::unique_ptr<ir::Graph> FuseElewiseAddActPass::FuseElewiseAddAct(
Node *elewise_add_act_node = CreateFuseElewiseAddActNode(
g, act, ele_add, ele_x_n, ele_y_n, ele_out_n, act_out_n);
VLOG(4) << "\n\t " << ele_x_n << " and " << ele_y_n << " -> "
<< ele_add->Name() << " -> " << ele_out_n << "\n"
<< "\t " << ele_out_n << " -> " << act->Name() << " -> "
<< act_out_n;
VLOG(40) << "\n\t " << ele_x_n << " and " << ele_y_n << " -> "
<< ele_add->Name() << " -> " << ele_out_n << "\n"
<< "\t " << ele_out_n << " -> " << act->Name() << " -> "
<< act_out_n;
ReLinkNodes(g, ele_out, ele_add, act, elewise_add_act_node);
found_elewise_add_act_count++;
......@@ -113,7 +113,7 @@ std::unique_ptr<ir::Graph> FuseElewiseAddActPass::FuseActElewiseAdd(
auto handler = [&](const GraphPatternDetector::subgraph_t &subgraph,
Graph *g) {
VLOG(4) << "handle FuseElewiseAddAct fuse";
VLOG(40) << "handle FuseElewiseAddAct fuse";
GET_IR_NODE_FROM_SUBGRAPH(act_out, act_out, act_elewise_add_pattern);
GET_IR_NODE_FROM_SUBGRAPH(ele_x, ele_x, act_elewise_add_pattern);
GET_IR_NODE_FROM_SUBGRAPH(ele_out, elewise_add_out,
......@@ -129,9 +129,9 @@ std::unique_ptr<ir::Graph> FuseElewiseAddActPass::FuseActElewiseAdd(
Node *elewise_add_act_node = CreateFuseElewiseAddActNode(
g, ele_add, act, elewise_add_x_n, act_i_n, act_o_n, elewise_add_out_n);
VLOG(4) << "\n\t " << act_i_n << " -> " << act->Name() << " -> " << act_o_n
<< "\n\t " << act_o_n << " and " << elewise_add_x_n << " -> "
<< ele_add->Name() << " -> " << elewise_add_out_n;
VLOG(40) << "\n\t " << act_i_n << " -> " << act->Name() << " -> " << act_o_n
<< "\n\t " << act_o_n << " and " << elewise_add_x_n << " -> "
<< ele_add->Name() << " -> " << elewise_add_out_n;
ReLinkNodes(g, act_out, act, ele_add, elewise_add_act_node);
found_elewise_add_act_count++;
......@@ -165,7 +165,7 @@ std::unique_ptr<ir::Graph> FuseElewiseAddActPass::FuseElewiseAddActInplaceGrad(
auto handler = [&](const GraphPatternDetector::subgraph_t &subgraph,
Graph *g) {
VLOG(4) << "handle FuseElewiseAddActGrad1 fuse";
VLOG(40) << "handle FuseElewiseAddActGrad1 fuse";
GET_IR_NODE_FROM_SUBGRAPH(act_out, act_out, elewise_add_act_grad_pattern);
GET_IR_NODE_FROM_SUBGRAPH(act_grad, act_grad, elewise_add_act_grad_pattern);
GET_IR_NODE_FROM_SUBGRAPH(d_itermediate_out, d_itermediate_out,
......@@ -208,10 +208,10 @@ std::unique_ptr<ir::Graph> FuseElewiseAddActPass::FuseElewiseAddActInplaceGrad(
auto fused_node = g->CreateOpNode(&desc);
VLOG(4) << "\n\t " << d_act_out_n << " and " << act_out_n << " -> "
<< act_grad->Name() << " -> " << d_itermediate_out_n << "\n\t "
<< d_itermediate_out_n << " and " << act_out_n << " -> "
<< ele_add_grad->Name() << " -> " << d_itermediate_out_n;
VLOG(40) << "\n\t " << d_act_out_n << " and " << act_out_n << " -> "
<< act_grad->Name() << " -> " << d_itermediate_out_n << "\n\t "
<< d_itermediate_out_n << " and " << act_out_n << " -> "
<< ele_add_grad->Name() << " -> " << d_itermediate_out_n;
ReLinkNodes(g, d_itermediate_out, act_grad, ele_add_grad, fused_node);
found_elewise_add_act_count++;
......
paddle/fluid/framework/ir/graph.cc
浏览文件 @ f7b55de9
......@@ -23,17 +23,74 @@ limitations under the License. */
namespace paddle {
namespace framework {
namespace ir {
namespace {
void CheckProgram(const ProgramDesc &program) {
#define _INT(role) static_cast<int>(role)
std::map<int, bool> visit;
for (OpDesc *op : program.Block(0).AllOps()) {
// For backward compatibility, some program doesn't have role added.
if (!op->HasAttr(OpProtoAndCheckerMaker::OpRoleAttrName())) continue;
int role_id =
boost::get<int>(op->GetAttr(OpProtoAndCheckerMaker::OpRoleAttrName()));
visit[role_id] = true;
switch (role_id) {
case _INT(OpRole::kForward):
if (visit.find(_INT(OpRole::kBackward)) != visit.end()) {
LOG(ERROR)
<< "Cannot add backward operator before forward operator %s."
<< op->Type();
}
break;
case _INT(OpRole::kBackward):
case _INT(OpRole::kBackward) | _INT(OpRole::kLoss):
PADDLE_ENFORCE(
visit.find(_INT(OpRole::kOptimize)) == visit.end(),
"Cannot add backward operator %s after optimize operator.",
op->Type());
break;
case _INT(OpRole::kForward) | _INT(OpRole::kLoss):
PADDLE_ENFORCE(visit.find(_INT(OpRole::kBackward) |
_INT(OpRole::kLoss)) == visit.end(),
"Cannot add backward|loss operator before "
"forward|loss operator %s.",
op->Type());
PADDLE_ENFORCE(
visit.find(_INT(OpRole::kOptimize)) == visit.end(),
"Cannot add forward|loss operator %s after optimize operator.",
op->Type());
break;
case _INT(OpRole::kOptimize):
case _INT(OpRole::kOptimize) | _INT(OpRole::kLRSched):
PADDLE_ENFORCE(visit.find(_INT(OpRole::kBackward)) != visit.end(),
"Optimize operators %s must follow backward operator.",
op->Type());
break;
case _INT(OpRole::kLRSched):
case _INT(OpRole::kDist):
case _INT(OpRole::kRPC):
case _INT(OpRole::kNotSpecified):
break;
default:
LOG(FATAL) << "Unknown operator role. Don't add new role because "
"you don't know what you are doing.";
}
}
#undef _INT
}
} // namespace
Graph::Graph(const ProgramDesc &program) : program_(program) {
// Make the nodes id start from 0.
Node::ResetId();
CheckProgram(program_);
auto var_nodes = InitFromProgram(program_);
ResolveHazard(var_nodes);
}
std::map<std::string, std::vector<ir::Node *>> Graph::InitFromProgram(
const ProgramDesc &program) {
VLOG(3) << "block in program:" << program_.Size();
VLOG(30) << "block in program:" << program_.Size();
std::unordered_map<std::string, VarDesc *> all_vars;
// var nodes for each var name, will have multiple versions in SSA
std::map<std::string, std::vector<ir::Node *>> var_nodes;
......@@ -101,7 +158,7 @@ void Graph::ResolveHazard(
auto it_old = versions.rbegin();
++it_old;
for (; it_old != versions.rend(); it_new = it_old, ++it_old) {
VLOG(3) << "deal with var: " << (*it_new)->Name();
VLOG(30) << "deal with var: " << (*it_new)->Name();
ir::Node *write_op =
(*it_new)->inputs.empty() ? nullptr : (*it_new)->inputs[0];
const auto &read_ops = (*it_old)->outputs;
......
paddle/fluid/framework/ir/graph.h
浏览文件 @ f7b55de9
......@@ -89,7 +89,7 @@ class Graph {
attr_name);
attrs_[attr_name] = attr;
attr_dels_[attr_name] = [attr, attr_name]() {
VLOG(3) << "deleting " << attr_name;
VLOG(30) << "deleting " << attr_name;
delete attr;
};
}
......@@ -102,18 +102,31 @@ class Graph {
attr_dels_[attr_name] = []() {};
}
template <typename AttrType>
void Erase(const std::string &attr_name) {
PADDLE_ENFORCE(attrs_.count(attr_name) != 0, "%s not set in the graph",
attr_name);
attr_dels_[attr_name]();
attrs_.erase(attr_name);
attr_dels_.erase(attr_name);
}
const std::unordered_set<ir::Node *> &Nodes() const { return node_set_; }
// Create a normal variable with non-null VarDesc.
ir::Node *CreateVarNode(VarDesc *var_desc) {
PADDLE_ENFORCE(var_desc);
return AddNode(new ir::Node(var_desc));
auto *x = AddNode(new ir::Node(var_desc));
x->SetId(num_node_created_++);
return x;
}
// Create a normal runnable operator with OpDesc.
ir::Node *CreateOpNode(OpDesc *op_desc) {
PADDLE_ENFORCE(op_desc);
return AddNode(new ir::Node(op_desc));
auto *x = AddNode(new ir::Node(op_desc));
x->SetId(num_node_created_++);
return x;
}
// Create a control dependency var that connects 2 operations. The
......@@ -123,13 +136,17 @@ class Graph {
// TODO(panyx0718): control var name should be really unique.
const std::string name = string::Sprintf(
"%s@%llu", ir::Node::kControlDepVarName, node_set_.size());
return AddNode(new ir::Node(name, ir::Node::Type::kVariable));
auto *x = AddNode(new ir::Node(name, ir::Node::Type::kVariable));
x->SetId(num_node_created_++);
return x;
}
// A more free style way of creating a graph node. Mostly use for test
// or "copy" from another node. Avoid using it if possible.
ir::Node *CreateEmptyNode(const std::string &name, ir::Node::Type type) {
return AddNode(new ir::Node(name, type));
auto *x = AddNode(new ir::Node(name, type));
x->SetId(num_node_created_++);
return x;
}
// Clear all node information of the graph and return the ownership of the
......@@ -151,7 +168,7 @@ class Graph {
}
// NOTE low performance, but simple and secure.
Node *RetriveNode(int id) {
Node *RetrieveNode(int id) {
for (auto &node : nodes_) {
if (node.second->id() == id) {
return node.second.get();
......@@ -160,6 +177,7 @@ class Graph {
return nullptr;
}
const ProgramDesc &program() const { return program_; }
std::map<std::string, std::vector<ir::Node *>> InitFromProgram(
const ProgramDesc &program);
......@@ -181,6 +199,7 @@ class Graph {
std::map<std::string, std::function<void(void)>> attr_dels_;
std::map<ir::Node *, std::unique_ptr<ir::Node>> nodes_;
std::unordered_set<ir::Node *> node_set_;
size_t num_node_created_{0}; // help to generate a unique node id.
};
bool IsControlDepVar(const ir::Node &var);
......
paddle/fluid/framework/ir/graph_helper.cc
浏览文件 @ f7b55de9
......@@ -15,8 +15,15 @@ limitations under the License. */
#include "paddle/fluid/framework/ir/graph_helper.h"
#include <algorithm>
#include <deque>
#include <fstream>
#include <iosfwd>
#include <ostream>
#include <unordered_set>
DEFINE_string(print_sub_graph_dir, "",
"FLAGS_print_sub_graph_dir is used "
"to print the nodes of sub_graphs.");
namespace paddle {
namespace framework {
namespace ir {
......@@ -33,8 +40,9 @@ void SortHelper(
}
}
VLOG(3) << "topology sort insert: " << node->Name()
<< reinterpret_cast<void *>(node) << " input " << node->inputs.size();
VLOG(30) << "topology sort insert: " << node->Name()
<< reinterpret_cast<void *>(node) << " input "
<< node->inputs.size();
ret->push_back(node);
}
......@@ -103,9 +111,9 @@ std::map<ir::Node *, std::unordered_set<ir::Node *>> BuildOperationAdjList(
for (auto &var : n->inputs) {
for (auto &adj_n : var->inputs) {
PADDLE_ENFORCE(adj_n->NodeType() == ir::Node::Type::kOperation);
VLOG(4) << "adj " << adj_n->Name() << reinterpret_cast<void *>(adj_n)
<< " -> " << n->Name() << reinterpret_cast<void *>(n)
<< " via " << var->Name() << reinterpret_cast<void *>(var);
VLOG(40) << "adj " << adj_n->Name() << reinterpret_cast<void *>(adj_n)
<< " -> " << n->Name() << reinterpret_cast<void *>(n)
<< " via " << var->Name() << reinterpret_cast<void *>(var);
adj_list[n].insert(adj_n);
}
}
......@@ -163,12 +171,15 @@ size_t GraphNum(const Graph &graph) {
graph_nodes.emplace_back(g_nodes);
}
if (VLOG_IS_ON(10)) {
VLOG(10) << "graph_num: " << graph_nodes.size();
for (auto &g_n : graph_nodes) {
VLOG(10) << "graph_nodes: " << g_n.size();
if (g_n.size() < 10) {
std::stringstream out;
if (FLAGS_print_sub_graph_dir.size()) {
if (graph_nodes.size() > 1) {
std::stringstream out;
for (auto &g_n : graph_nodes) {
out << "graph_nodes: " << g_n.size() << "\n";
}
out << "\n\n";
for (auto &g_n : graph_nodes) {
out << "graph_nodes: " << g_n.size();
for (auto &node : g_n) {
out << "\nNode: " << node->Name() << " in [";
for (auto &n : node->inputs) {
......@@ -180,8 +191,12 @@ size_t GraphNum(const Graph &graph) {
}
out << "]";
}
VLOG(10) << out.str();
out << "\n\n\n";
}
std::unique_ptr<std::ostream> fout(
new std::ofstream(FLAGS_print_sub_graph_dir));
PADDLE_ENFORCE(fout->good());
*fout << out.str();
}
}
......
paddle/fluid/framework/ir/graph_helper.h
浏览文件 @ f7b55de9
......@@ -37,6 +37,15 @@ std::vector<ir::Node *> TopologySortOperations(const Graph &graph);
std::map<ir::Node *, std::unordered_set<ir::Node *>> BuildOperationAdjList(
const Graph &graph);
template <typename T>
std::vector<T *> FilterByNodeWrapper(const Graph &graph) {
std::vector<T *> ret;
for (ir::Node *n : graph.Nodes()) {
if (n->IsWrappedBy<T>()) ret.push_back(&n->Wrapper<T>());
}
return ret;
}
} // namespace ir
} // namespace framework
} // namespace paddle
paddle/fluid/framework/ir/graph_pattern_detector.h
浏览文件 @ f7b55de9
......@@ -310,8 +310,8 @@ void GraphSafeRemoveNodes(Graph* graph,
const std::unordered_set<const Node*>& nodes);
// Some pre-defined patterns those can be reused in multiple passes.
// The related Fluid Layer or Op should be one pattern here for better reusage
// accross different fusion.
// The related Fluid Layer or Op should be one pattern here for better re-usage
// across different fusion.
namespace patterns {
struct KeyCounter {
......@@ -664,7 +664,7 @@ struct ElementwiseAdd : public PatternBase {
ElementwiseAdd(PDPattern* pattern, const std::string& name_scope)
: PatternBase(pattern, name_scope, "elementwise_add") {}
PDNode* operator()(PDNode* x_var);
PDNode* operator()(PDNode* x_var, PDNode* y_var);
PATTERN_DECL_NODE(elementwise_add_op);
PATTERN_DECL_NODE(elementwise_add_x);
......
paddle/fluid/framework/ir/graph_to_program_pass.cc
浏览文件 @ f7b55de9
......@@ -35,10 +35,11 @@ std::unique_ptr<Graph> GraphToProgramPass::ApplyImpl(
new proto::ProgramDesc(*program.Proto()));
auto block = program_pb->mutable_blocks(kRootBlockIndex);
block->set_idx(kRootBlockIndex);
block->clear_vars();
std::unordered_set<std::string> visited_vars;
for (ir::Node* n : graph->Nodes()) {
if (n->NodeType() == ir::Node::Type::kVariable) {
if (n->IsVar()) {
if (n->Var() && visited_vars.count(n->Var()->Name()) == 0) {
visited_vars.insert(n->Var()->Name());
block->add_vars()->MergeFrom(*n->Var()->Proto());
......
paddle/fluid/framework/ir/is_test_pass.cc 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/framework/ir/is_test_pass_tester.cc 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/framework/ir/node.cc
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/framework/ir/node.h
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/framework/ir/node_test.cc 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/framework/ir/pass.h
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/framework/lod_tensor.h
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/framework/ngraph_bridge.cc 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/framework/ngraph_bridge.h 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/framework/ngraph_operator.cc 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/framework/ngraph_operator.h 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/framework/op_desc.cc
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/framework/operator.cc
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/framework/operator.h
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/framework/scope.cc
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/framework/scope.h
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/framework/tensor.cc
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/framework/tensor.h
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/framework/threadpool.h
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/framework/var_desc.cc
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/inference/analysis/node.cc 已删除 100644 → 0
浏览文件 @ af9a3301
此差异已折叠。
paddle/fluid/inference/analysis/node.h 已删除 100644 → 0
浏览文件 @ af9a3301
此差异已折叠。
paddle/fluid/inference/analysis/passes/CMakeLists.txt 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/inference/api/analysis_config.cc 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/inference/api/api.cc
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/inference/api/helper.h
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/inference/api/paddle_anakin_config.h 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/inference/api/paddle_analysis_config.h 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/inference/api/paddle_api.h 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/inference/api/paddle_pass_builder.cc 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/inference/api/paddle_pass_builder.h 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/inference/io.cc
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/inference/tensorrt/convert/prelu_op.cc 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/inference/tensorrt/convert/split_op.cc 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/inference/tensorrt/plugin/CMakeLists.txt 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/inference/tensorrt/plugin/serialize.h 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/inference/tensorrt/plugin/trt_plugin.cc 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/inference/tensorrt/plugin/trt_plugin.h 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/inference/tests/api/analyzer_dam_tester.cc 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/inference/tests/api/config_printer.h 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/inference/tests/test.cmake 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/memory/CMakeLists.txt
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/memory/allocation/CMakeLists.txt 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/memory/allocation/aligned_allocator.h 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/memory/allocation/allocator.cc 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/memory/allocation/allocator.h 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/memory/allocation/allocator_facade.cc 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/memory/allocation/allocator_facade_test.cc 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/memory/allocation/allocator_strategy.cc 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/memory/allocation/allocator_strategy.h 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/memory/allocation/best_fit_allocator.cc 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/memory/allocation/best_fit_allocator.h 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/memory/allocation/buffered_allocator.cc 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/memory/allocation/buffered_allocator.h 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/memory/allocation/conditional_allocator.h 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/memory/allocation/cpu_allocator.cc 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/memory/allocation/cpu_allocator.h 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/memory/allocation/cuda_allocator.cc 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/memory/allocation/cuda_allocator.h 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/memory/allocation/legacy_allocator.cc 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/memory/allocation/legacy_allocator.h 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/memory/allocation/locked_allocator.cc 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/memory/allocation/locked_allocator.h 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/memory/allocation/pinned_allocator.cc 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/memory/allocation/pinned_allocator.h 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/memory/allocation/retry_allocator.cc 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/memory/allocation/retry_allocator.h 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/memory/allocation/retry_allocator_test.cc 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/memory/allocation/zero_size_allocator.cc 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/memory/allocation/zero_size_allocator.h 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/memory/malloc.cc
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/memory/malloc.h
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/memory/malloc_test.cc 已删除 100644 → 0
浏览文件 @ af9a3301
此差异已折叠。
paddle/fluid/memory/memcpy.cc
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/affine_grid_cudnn_op.cu.cc 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/affine_grid_op.cc 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/affine_grid_op.h 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/bilinear_interp_op.h 已删除 100644 → 0
浏览文件 @ af9a3301
此差异已折叠。
paddle/fluid/operators/concat_op.cc
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/controlflow/CMakeLists.txt 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/conv_cudnn_op_cache.h 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/conv_fusion_op.cc 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/conv_fusion_op.cu.cc 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/conv_op.cc
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/conv_op.h
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/csp/CMakeLists.txt 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/detection/density_prior_box_op.h 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/distributed_ops/CMakeLists.txt 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/elementwise/CMakeLists.txt 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/expand_op.cc
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/extract_rows_op.cc 已删除 100644 → 0
浏览文件 @ af9a3301
此差异已折叠。
paddle/fluid/operators/fc_op.cc
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/fused/CMakeLists.txt 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/gather.cu.h
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/gather.h
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/gather_op.cc
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/grid_sampler_cudnn_op.cu.cc 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/grid_sampler_op.cc 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/grid_sampler_op.h 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/hash_op.cc
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/interpolate_op.h 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/load_op.cc
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/lrn_op.cc
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/lrn_op.h
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/math/blas.h
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/math/jit_code.cc 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/math/jit_code.h 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/math/jit_gen.cc 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/math/jit_gen.h 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/math/jit_kernel_layer_norm.cc 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/mean_op.cc
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/mean_op.cu
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/mean_op.h
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/metrics/CMakeLists.txt 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/mul_op.cc
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/mul_op.cu.cc
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/nce_op.cc
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/nce_op.h
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/optimizers/CMakeLists.txt 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/pool_op.cc
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/pool_op.h
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/prelu_op.h
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/reduce_ops/CMakeLists.txt 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/reduce_ops/reduce_max_op.part.cu 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/reduce_ops/reduce_min_op.part.cu 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/reduce_ops/reduce_sum_op.part.cu 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/save_op.cc
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/scale_op.cu
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/scale_op.h
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/scatter.cu.h
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/scatter.h
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/selu_op.cc 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/selu_op.cu 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/selu_op.h 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/sequence_ops/CMakeLists.txt 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/sign_op.cc
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/sign_op.cu
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/similarity_focus_op.cc 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/similarity_focus_op.h 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/softmax_op.h
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/space_to_depth_op.cc 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/space_to_depth_op.cu 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/space_to_depth_op.h 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/spp_op.h
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/stack_op.cc
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/stack_op.cu
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/stack_op.h
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/sum_op.cc
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/sum_op.cu
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/sum_op.h
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/tensor_array_to_tensor_op.cc 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/tensorrt/CMakeLists.txt 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/unpool_op.cc
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/operators/warpctc_cudnn_op.cu.cc 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/platform/cpu_info.cc
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/platform/cpu_info.h
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/platform/cuda_device_guard.cc 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/platform/cuda_device_guard.h 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/platform/gpu_info.cc
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/platform/init.cc
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/platform/lock_guard_ptr.h 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/platform/nccl_helper.h
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/platform/place.h
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/platform/port.h
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/platform/profiler.cc
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/platform/variant.h
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/pybind/CMakeLists.txt
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/pybind/const_value.cc
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/pybind/protobuf.cc
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/pybind/pybind.cc
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/pybind/tensor_py.h
浏览文件 @ f7b55de9
此差异已折叠。
paddle/fluid/string/printf.h
浏览文件 @ f7b55de9
此差异已折叠。
paddle/scripts/paddle_build.sh
浏览文件 @ f7b55de9
此差异已折叠。
paddle/testing/TestUtil.cpp
浏览文件 @ f7b55de9
此差异已折叠。
paddle/testing/paddle_gtest_main.cc
浏览文件 @ f7b55de9
此差异已折叠。
python/CMakeLists.txt
浏览文件 @ f7b55de9
此差异已折叠。
python/paddle/dataset/wmt16.py
浏览文件 @ f7b55de9
此差异已折叠。
python/paddle/fluid/__init__.py
浏览文件 @ f7b55de9
此差异已折叠。
python/paddle/fluid/distribute_lookup_table.py 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
python/paddle/fluid/io.py
浏览文件 @ f7b55de9
此差异已折叠。
python/paddle/fluid/layers/io.py
浏览文件 @ f7b55de9
此差异已折叠。
python/paddle/fluid/layers/nn.py
浏览文件 @ f7b55de9
此差异已折叠。
python/paddle/fluid/layers/ops.py
浏览文件 @ f7b55de9
此差异已折叠。
python/paddle/fluid/op.py
浏览文件 @ f7b55de9
此差异已折叠。
python/paddle/fluid/optimizer.py
浏览文件 @ f7b55de9
此差异已折叠。
python/paddle/fluid/regularizer.py
浏览文件 @ f7b55de9
此差异已折叠。
python/paddle/fluid/tests/unittests/dist_save_load.py 0 → 100644
浏览文件 @ f7b55de9
此差异已折叠。
python/paddle/v2/dataset/wmt16.py
浏览文件 @ f7b55de9
此差异已折叠。
python/requirements.txt
浏览文件 @ f7b55de9
此差异已折叠。
python/setup.py.in
浏览文件 @ f7b55de9
此差异已折叠。
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