Skip to content

P
Paddle

机器未来 / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 1
Fork 0
P
Paddle
未验证 提交 2412c27c 编写于 作者: S Sylwester Fraczek 提交者： GitHub
浏览文件
操作

Merge branch 'develop' into sfraczek/test-analyzer-mobilenet

上级 b5f617fa c7b3bfcd
展开全部 隐藏空白更改
内联 并排
Showing with 9601 addition and 6054 deletion
AUTHORS.md
浏览文件 @ 2412c27c
......@@ -43,6 +43,7 @@
| qingqing01 | Qing-Qing Dang |
| reyoung | Yang Yu |
| Superjom | Chun-Wei Yan |
| tensor-tang | Jian Tang |
| tianbingsz | Tian-Bing Xu |
| tpatejko | Tomasz Patejko |
| typhoonzero | Yi Wu |
......
CMakeLists.txt
浏览文件 @ 2412c27c
......@@ -41,6 +41,7 @@ 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)
......@@ -65,6 +66,8 @@ option(WITH_ARM_FP16 "Use half precision support on armv8.2-a cpu" 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" ${WITH_DISTRIBUTE})
option(WITH_BRPC_RDMA "Use brpc rdma as the rpc protocal" OFF)
option(ON_INFER "Turn on inference optimization." OFF)
......@@ -103,6 +106,8 @@ if(ANDROID OR IOS)
"Disable RDMA when cross-compiling for Android and iOS" FORCE)
set(WITH_MKL OFF CACHE STRING
"Disable MKL when cross-compiling for Android and iOS" FORCE)
set(WITH_NGRAPH OFF CACHE STRING
"Disable nGraph when cross-compiling for Android and iOS" FORCE)
set(WITH_GOLANG OFF CACHE STRING
"Disable golang when cross-compiling for Android and iOS" FORCE)
......@@ -171,6 +176,7 @@ include(external/protobuf) # download, build, install protobuf
include(external/python) # download, build, install python
include(external/openblas) # download, build, install openblas
include(external/mkldnn) # download, build, install mkldnn
include(external/ngraph) # download, build, install nGraph
include(external/swig) # download, build, install swig
include(external/boost) # download boost
include(external/any) # download libn::any
......
cmake/external/anakin.cmake
浏览文件 @ 2412c27c
......@@ -58,19 +58,21 @@ ExternalProject_Add(
-DPROTOBUF_ROOT=${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 ${ANAKIN_SABER_LIB})
add_dependencies(anakin_saber extern_anakin protobuf mklml)
add_dependencies(anakin_saber extern_anakin)
list(APPEND external_project_dependencies anakin_shared anakin_saber)
cmake/external/mkldnn.cmake
浏览文件 @ 2412c27c
......@@ -37,7 +37,6 @@ SET(CMAKE_INSTALL_RPATH_USE_LINK_PATH TRUE)
SET(CMAKE_INSTALL_RPATH "${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 @ 2412c27c
# 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/protobuf.cmake
浏览文件 @ 2412c27c
......@@ -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})
......
cmake/inference_lib.cmake
浏览文件 @ 2412c27c
......@@ -164,7 +164,7 @@ 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}
)
......@@ -202,10 +202,10 @@ copy(third_party DEPS fluid_lib_dist
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 ${FLUID_INFERENCE_INSTALL_DIR}/paddle/include
)
......
cmake/tensorrt.cmake
浏览文件 @ 2412c27c
......@@ -34,4 +34,5 @@ if(TENSORRT_FOUND)
"Current TensorRT version is v${TENSORRT_MAJOR_VERSION}. ")
include_directories(${TENSORRT_INCLUDE_DIR})
list(APPEND EXTERNAL_LIBS ${TENSORRT_LIBRARY})
add_definitions(-DPADDLE_WITH_TENSORRT)
endif()
paddle/fluid/API.spec
浏览文件 @ 2412c27c
......@@ -103,7 +103,7 @@ paddle.fluid.layers.beam_search ArgSpec(args=['pre_ids', 'pre_scores', 'ids', 's
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', 'numeric_stable_mode'], varargs=None, keywords=None, defaults=(False, -100, False))
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))
......@@ -118,9 +118,10 @@ 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,))
......@@ -174,13 +175,16 @@ 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)
......@@ -189,6 +193,7 @@ 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)
......@@ -198,6 +203,7 @@ 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,))
......@@ -268,6 +274,7 @@ 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/framework/CMakeLists.txt
浏览文件 @ 2412c27c
......@@ -136,6 +136,10 @@ 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)
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)
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)
......@@ -163,10 +167,10 @@ 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)
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)
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 @ 2412c27c
......@@ -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 @ 2412c27c
......@@ -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/broadcast_op_handle.cc
浏览文件 @ 2412c27c
......@@ -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 @ 2412c27c
......@@ -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/fast_threaded_ssa_graph_executor.cc
浏览文件 @ 2412c27c
......@@ -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;
}
......@@ -158,7 +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 @ 2412c27c
......@@ -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 @ 2412c27c
......@@ -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 @ 2412c27c
......@@ -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 @ 2412c27c
......@@ -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
浏览文件 @ 2412c27c
......@@ -16,6 +16,7 @@
#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 {
......@@ -35,17 +36,17 @@ static bool IsLockAndRecordEventFreeComputationOpHandle(
std::unique_ptr<ir::Graph> ModifyOpLockAndRecordEventPass::ApplyImpl(
std::unique_ptr<ir::Graph> ir_graph) const {
auto &all_ops = ir_graph->Get<GraphOps>(kGraphOps);
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.get());
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(10) << "Set is_lock_and_record_event_free be true in op "
<< compute_op->DebugString();
VLOG(100) << "Set is_lock_and_record_event_free be true in op "
<< compute_op->DebugString();
}
}
return ir_graph;
......
paddle/fluid/framework/details/multi_devices_graph_check_pass.cc
浏览文件 @ 2412c27c
......@@ -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 @ 2412c27c
......@@ -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 @ 2412c27c
......@@ -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 @ 2412c27c
......@@ -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 @ 2412c27c
......@@ -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
浏览文件 @ 2412c27c
......@@ -20,19 +20,16 @@ namespace paddle {
namespace framework {
namespace details {
OpGraphView::OpGraphView(
const std::vector<std::unique_ptr<OpHandleBase>> &ops) {
Build(ops);
}
OpGraphView::OpGraphView(const std::vector<OpHandleBase *> &ops) { Build(ops); }
void OpGraphView::Build(const std::vector<std::unique_ptr<OpHandleBase>> &ops) {
void OpGraphView::Build(const std::vector<OpHandleBase *> &ops) {
for (auto &op : ops) {
preceding_ops_[op.get()];
pending_ops_[op.get()];
preceding_ops_[op];
pending_ops_[op];
for (auto &var : op->Outputs()) {
for (auto &pending_op : var->PendingOps()) {
preceding_ops_[pending_op].insert(op.get());
pending_ops_[op.get()].insert(pending_op);
preceding_ops_[pending_op].insert(op);
pending_ops_[op].insert(pending_op);
}
}
}
......@@ -41,8 +38,6 @@ void OpGraphView::Build(const std::vector<std::unique_ptr<OpHandleBase>> &ops) {
"There are duplicate ops in graph.");
}
size_t OpGraphView::OpNumber() const { return preceding_ops_.size(); }
std::unordered_set<OpHandleBase *> OpGraphView::AllOps() const {
std::unordered_set<OpHandleBase *> ret;
for (auto &pair : preceding_ops_) {
......@@ -60,12 +55,6 @@ void OpGraphView::EnforceHasOp(OpHandleBase *op) const {
op == nullptr ? "nullptr" : op->DebugString());
}
const std::unordered_set<OpHandleBase *> &OpGraphView::PrecedingOps(
OpHandleBase *op) const {
EnforceHasOp(op);
return preceding_ops_.at(op);
}
const std::unordered_set<OpHandleBase *> &OpGraphView::PendingOps(
OpHandleBase *op) const {
EnforceHasOp(op);
......
paddle/fluid/framework/details/op_graph_view.h
浏览文件 @ 2412c27c
......@@ -26,21 +26,16 @@ namespace details {
class OpGraphView {
public:
explicit OpGraphView(const std::vector<std::unique_ptr<OpHandleBase>> &ops);
size_t OpNumber() const;
explicit OpGraphView(const std::vector<OpHandleBase *> &ops);
std::unordered_set<OpHandleBase *> AllOps() const;
const std::unordered_set<OpHandleBase *> &PrecedingOps(
OpHandleBase *op) const;
const std::unordered_set<OpHandleBase *> &PendingOps(OpHandleBase *op) const;
bool HasOp(OpHandleBase *op) const;
private:
void Build(const std::vector<std::unique_ptr<OpHandleBase>> &ops);
void Build(const std::vector<OpHandleBase *> &ops);
void EnforceHasOp(OpHandleBase *op) const;
std::unordered_map<OpHandleBase *, std::unordered_set<OpHandleBase *>>
......
paddle/fluid/framework/details/op_handle_base.h
浏览文件 @ 2412c27c
......@@ -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 @ 2412c27c
......@@ -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_pass.cc
浏览文件 @ 2412c27c
......@@ -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 {
......@@ -71,14 +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;
......@@ -121,9 +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;
......@@ -141,8 +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(
......@@ -151,21 +150,21 @@ std::unique_ptr<ir::Graph> ReferenceCountPass::ApplyImpl(
ref_cnt_node, next_compute_op->GetScope(), place, {var_name},
gcs[place.device].get(), cur_ref_cnts[place.device].get());
AddDependencyBetween(next_compute_op, ref_cnt_handle, graph.get());
compute_ref_cnt_map[next_compute_op].reset(ref_cnt_handle);
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);
......@@ -173,7 +172,7 @@ std::unique_ptr<ir::Graph> ReferenceCountPass::ApplyImpl(
ref_cnt_node, compute_op->GetScope(), place, in_var_names,
gcs[place.device].get(), cur_ref_cnts[place.device].get());
AddDependencyBetween(compute_op, ref_cnt_handle, graph.get());
compute_ref_cnt_map[compute_op].reset(ref_cnt_handle);
compute_ref_cnt_map[compute_op] = ref_cnt_handle;
}
for (auto &op : all_ops) {
......@@ -181,11 +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/scale_loss_grad_op_handle.cc
浏览文件 @ 2412c27c
......@@ -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
浏览文件 @ 2412c27c
......@@ -94,8 +94,8 @@ std::unique_ptr<ir::Graph> SequentialExecutionPass::ApplyImpl(
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(10) << "Add dependencies between " << op_node_list[i - 1]->Name()
<< " and " << op_node_list[i]->Name();
VLOG(100) << "Add dependencies between " << op_node_list[i - 1]->Name()
<< " and " << op_node_list[i]->Name();
}
return graph;
}
......
paddle/fluid/framework/details/ssa_graph_executor.cc
浏览文件 @ 2412c27c
......@@ -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 @ 2412c27c
......@@ -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 @ 2412c27c
......@@ -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());
}
}
}
......@@ -208,16 +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();
}
if (LIKELY(!strategy_.dry_run_)) {
op->Run(strategy_.use_cuda_);
}
VLOG(10) << op << " " << op->Name() << " Done ";
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 @ 2412c27c
......@@ -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 @ 2412c27c
......@@ -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 @ 2412c27c
......@@ -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 @ 2412c27c
......@@ -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>()) {
......@@ -81,6 +83,24 @@ 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() {
......@@ -141,21 +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;
}
}
}
......@@ -286,7 +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();
......@@ -309,7 +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();
......@@ -338,6 +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;
}
......@@ -359,6 +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) {
......@@ -398,8 +420,8 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
}
if (FLAGS_benchmark) {
VLOG(2) << "Memory used after operator " + op->Type() + " running: "
<< memory::memory_usage(place_);
VLOG(20) << "Memory used after operator " + op->Type() + " running: "
<< memory::memory_usage(place_);
}
}
......@@ -424,10 +446,10 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
}
if (FLAGS_benchmark) {
VLOG(2) << "-------------------------------------------------------";
VLOG(2) << "Memory used after deleting local scope: "
<< memory::memory_usage(place_);
VLOG(2) << "-------------------------------------------------------";
VLOG(20) << "-------------------------------------------------------";
VLOG(20) << "Memory used after deleting local scope: "
<< memory::memory_usage(place_);
VLOG(20) << "-------------------------------------------------------";
}
}
......@@ -471,7 +493,7 @@ 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()) {
......@@ -485,6 +507,5 @@ 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 @ 2412c27c
......@@ -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/ir/CMakeLists.txt
浏览文件 @ 2412c27c
......@@ -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)
......@@ -53,6 +55,7 @@ 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)
......
paddle/fluid/framework/ir/attention_lstm_fuse_pass.cc
浏览文件 @ 2412c27c
......@@ -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);
......@@ -208,7 +208,7 @@ 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(
......
paddle/fluid/framework/ir/conv_bias_mkldnn_fuse_pass.cc
浏览文件 @ 2412c27c
......@@ -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 @ 2412c27c
......@@ -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_relu_mkldnn_fuse_pass.cc
浏览文件 @ 2412c27c
......@@ -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/depthwise_conv_mkldnn_pass.cc
浏览文件 @ 2412c27c
......@@ -39,7 +39,7 @@ std::unique_ptr<ir::Graph> DepthwiseConvMKLDNNPass::ApplyImpl(
int found_depthwise_conv_mkldnn_count = 0;
auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
Graph* g) {
VLOG(3) << "handle DepthwiseConvMKLDNN fuse";
VLOG(30) << "handle DepthwiseConvMKLDNN fuse";
GET_NODE(depthwise_conv, (*pattern));
depthwise_conv->Op()->SetType("conv2d");
found_depthwise_conv_mkldnn_count++;
......
paddle/fluid/framework/ir/fc_fuse_pass.cc
浏览文件 @ 2412c27c
......@@ -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);
......
paddle/fluid/framework/ir/fuse_elewise_add_act_pass.cc
浏览文件 @ 2412c27c
......@@ -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 @ 2412c27c
......@@ -84,15 +84,13 @@ void CheckProgram(const ProgramDesc &program) {
Graph::Graph(const ProgramDesc &program) : program_(program) {
CheckProgram(program_);
// Make the nodes id start from 0.
Node::ResetId();
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;
......@@ -160,7 +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 @ 2412c27c
......@@ -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 @ 2412c27c
......@@ -33,8 +33,9 @@ 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 +104,9 @@ 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);
}
}
......@@ -163,10 +164,10 @@ size_t GraphNum(const Graph &graph) {
graph_nodes.emplace_back(g_nodes);
}
if (VLOG_IS_ON(10)) {
VLOG(10) << "graph_num: " << graph_nodes.size();
if (VLOG_IS_ON(100)) {
VLOG(100) << "graph_num: " << graph_nodes.size();
for (auto &g_n : graph_nodes) {
VLOG(10) << "graph_nodes: " << g_n.size();
VLOG(100) << "graph_nodes: " << g_n.size();
if (g_n.size() < 10) {
std::stringstream out;
for (auto &node : g_n) {
......@@ -180,7 +181,7 @@ size_t GraphNum(const Graph &graph) {
}
out << "]";
}
VLOG(10) << out.str();
VLOG(100) << out.str();
}
}
}
......
paddle/fluid/framework/ir/graph_helper.h
浏览文件 @ 2412c27c
......@@ -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.cc
浏览文件 @ 2412c27c
......@@ -12,6 +12,7 @@
// See the License for the specific language governing permissions and
// limitations under the License.
#include <algorithm>
#include <array>
#include <string>
#include <vector>
......@@ -91,19 +92,19 @@ void GraphPatternDetector::operator()(Graph *graph,
PrettyLogEndl(Style::detail(), "--- detect %d subgraphs", subgraphs.size());
int id = 0;
for (auto &g : subgraphs) {
VLOG(3) << "optimizing #" << id++ << " subgraph";
VLOG(30) << "optimizing #" << id++ << " subgraph";
handler(g, graph);
}
}
bool GraphPatternDetector::MarkPDNodesInGraph(const ir::Graph &graph) {
VLOG(3) << "mark pdnodes in graph";
VLOG(30) << "mark pdnodes in graph";
if (graph.Nodes().empty()) return false;
for (auto &node : GraphTraits::DFS(graph)) {
for (const auto &pdnode : pattern_.nodes()) {
if (pdnode->Tell(&node)) {
VLOG(4) << "pdnode " << pdnode->name() << " marked";
VLOG(40) << "pdnode " << pdnode->name() << " marked";
pdnodes2nodes_[pdnode.get()].insert(&node);
}
}
......@@ -111,7 +112,7 @@ bool GraphPatternDetector::MarkPDNodesInGraph(const ir::Graph &graph) {
// Check to early stop if some PDNode can't find matched Node.
for (auto &pdnode : pattern_.nodes()) {
if (!pdnodes2nodes_.count(pdnode.get())) {
VLOG(4) << pdnode->name() << " can't find matched Node, early stop";
VLOG(40) << pdnode->name() << " can't find matched Node, early stop";
// return false;
}
}
......@@ -120,7 +121,7 @@ bool GraphPatternDetector::MarkPDNodesInGraph(const ir::Graph &graph) {
GetMarkedNodes(const_cast<Graph *>(&graph)).insert(n);
}
}
VLOG(3) << pdnodes2nodes_.size() << " nodes marked";
VLOG(30) << pdnodes2nodes_.size() << " nodes marked";
return !pdnodes2nodes_.empty();
}
......@@ -166,10 +167,12 @@ struct HitGroup {
bool Match(Node *node, PDNode *pat) {
if (nodes_.count(node)) {
if (!roles.count(pat)) return false;
return roles[pat] == node;
if (roles.count(pat) && roles[pat] == node) return true;
return false;
} else {
if (roles.count(pat) && roles[pat] != node) return false;
return true;
}
return !roles.count(pat) || roles.at(pat) == node;
}
void Register(Node *node, PDNode *pat) {
......@@ -197,7 +200,6 @@ GraphPatternDetector::DetectPatterns() {
std::vector<GraphPatternDetector::subgraph_t> result;
std::vector<HitGroup> init_groups;
std::array<std::vector<HitGroup>, 2> bi_records;
// PADDLE_ENFORCE(!pattern_.edges().empty(), "At least one edge is needed");
auto *first_pnode = pattern_.edges().empty() ? pattern().nodes().front().get()
: pattern_.edges().front().first;
if (!pdnodes2nodes_.count(first_pnode)) return result;
......@@ -213,7 +215,7 @@ GraphPatternDetector::DetectPatterns() {
// Extend a PDNode to subgraphs by deducing the connection relations defined
// in edges of PDNodes.
for (const auto &edge : pattern_.edges()) {
VLOG(4) << "check " << edge.first->name() << " -> " << edge.second->name();
VLOG(40) << "check " << edge.first->name() << " -> " << edge.second->name();
// TODO(Superjomn) Fix bug here, the groups might be duplicate here.
// Each role has two PDNodes, which indicates two roles.
// Detect two Nodes that can match these two roles and they are connected.
......@@ -224,14 +226,15 @@ GraphPatternDetector::DetectPatterns() {
// source -> target
for (Node *source : pdnodes2nodes_[edge.first]) {
for (Node *target : pdnodes2nodes_[edge.second]) {
VLOG(8) << "check " << source->id() << " -- " << target->id();
VLOG(80) << "check " << source->id() << " -- " << target->id();
// TODO(Superjomn) add some prune strategies.
for (const auto &group : pre_groups) {
HitGroup new_group = group;
if (IsNodesLink(source, target) &&
new_group.Match(source, edge.first)) {
new_group.Register(source, edge.first);
if (new_group.Match(target, edge.second)) {
if (IsNodesLink(source, target)) {
HitGroup new_group = group;
bool flag = new_group.Match(source, edge.first) &&
new_group.Match(target, edge.second);
if (flag) {
new_group.Register(source, edge.first);
new_group.Register(target, edge.second);
cur_groups.push_back(new_group);
// TODO(Superjomn) need to unique
......@@ -240,12 +243,13 @@ GraphPatternDetector::DetectPatterns() {
}
}
}
VLOG(3) << "step " << step << " get records: " << cur_groups.size();
VLOG(30) << "step " << step << " get records: " << cur_groups.size();
for (auto &group : cur_groups) {
for (auto &item : group.roles) {
VLOG(4) << "node " << item.second->id() << " as " << item.first->name();
VLOG(40) << "node " << item.second->id() << " as "
<< item.first->name();
}
VLOG(4) << "=========================================================";
VLOG(40) << "=========================================================";
}
}
......@@ -259,14 +263,16 @@ GraphPatternDetector::DetectPatterns() {
return result;
}
bool GraphItemCMP(const std::pair<PDNode *, Node *> &a,
struct GraphItemLessThan {
bool operator()(const std::pair<PDNode *, Node *> &a,
const std::pair<PDNode *, Node *> &b) {
if (a.first != b.first) {
return a.first < b.first;
} else {
return a.second < b.second;
if (a.first != b.first) {
return a.first < b.first;
} else {
return a.second < b.second;
}
}
}
};
// TODO(Superjomn) enhance the function as it marks unique unique as duplicates
// see https://github.com/PaddlePaddle/Paddle/issues/13550
......@@ -280,7 +286,7 @@ void GraphPatternDetector::UniquePatterns(
for (auto &g : *subgraphs) {
// Sort the items in the sub-graph, and transform to a string key.
std::vector<std::pair<PDNode *, Node *>> sorted_keys(g.begin(), g.end());
std::sort(sorted_keys.begin(), sorted_keys.end(), GraphItemCMP);
std::sort(sorted_keys.begin(), sorted_keys.end(), GraphItemLessThan());
std::stringstream ss;
for (auto &item : sorted_keys) {
ss << item.first << ":" << item.second;
......
paddle/fluid/framework/ir/graph_pattern_detector.h
浏览文件 @ 2412c27c
......@@ -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 {
......
paddle/fluid/framework/ir/graph_to_program_pass.cc
浏览文件 @ 2412c27c
......@@ -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/graph_traits.cc
浏览文件 @ 2412c27c
......@@ -66,6 +66,76 @@ NodesDFSIterator &NodesDFSIterator::operator=(const NodesDFSIterator &other) {
}
Node *NodesDFSIterator::operator->() { return stack_.top(); }
inline bool CheckNodeIndegreeEquals(const Node &node, size_t n) {
return node.inputs.size() == n;
}
NodesTSIterator::NodesTSIterator(const std::vector<Node *> &source) {
PADDLE_ENFORCE(!source.empty(),
"Start points of topological sorting should not be empty!");
// CHECK all the inputs' in-degree is 0
for (auto *node : source) {
PADDLE_ENFORCE(CheckNodeIndegreeEquals(*node, 0));
}
std::unordered_set<Node *> visited;
std::unordered_set<Node *> to_visit{source.begin(), source.end()};
std::vector<Node *> inlink_visited;
while (!to_visit.empty()) {
std::vector<Node *> queue(to_visit.begin(), to_visit.end());
for (auto *p : queue) {
inlink_visited.clear();
std::copy_if(p->inputs.begin(), p->inputs.end(),
std::back_inserter(inlink_visited),
[&](Node *x) -> bool { return visited.count(x) != 0; });
if (inlink_visited.size() == p->inputs.size()) {
sorted_.push_back(p);
for (auto *_ : p->outputs) {
if (!visited.count(_)) {
to_visit.insert(_);
}
}
to_visit.erase(p);
visited.insert(p);
}
}
}
}
NodesTSIterator::NodesTSIterator(const NodesTSIterator &other)
: sorted_(other.sorted_), cursor_(other.cursor_) {}
Node &NodesTSIterator::operator*() {
PADDLE_ENFORCE_LT(cursor_, sorted_.size());
return *sorted_[cursor_];
}
NodesTSIterator &NodesTSIterator::operator++() {
if (++cursor_ >= sorted_.size()) {
sorted_.clear();
cursor_ = 0;
}
return *this;
}
NodesTSIterator &NodesTSIterator::operator=(const NodesTSIterator &other) {
cursor_ = other.cursor_;
sorted_ = other.sorted_;
return *this;
}
bool NodesTSIterator::operator==(const NodesTSIterator &other) {
return sorted_ == other.sorted_ && cursor_ == other.cursor_;
}
Node *NodesTSIterator::operator->() {
PADDLE_ENFORCE_LT(cursor_, sorted_.size());
return sorted_[cursor_];
}
} // namespace ir
} // namespace framework
} // namespace paddle
paddle/fluid/framework/ir/graph_traits.h
浏览文件 @ 2412c27c
......@@ -62,6 +62,32 @@ struct NodesDFSIterator
std::unordered_set<Node *> visited_;
};
// Topological sorting iterator on nodes.
struct NodesTSIterator
: public std::iterator<std::forward_iterator_tag, Node *> {
NodesTSIterator() = default;
NodesTSIterator(const std::vector<Node *> &source);
NodesTSIterator(NodesTSIterator &&other)
: sorted_(std::move(other.sorted_)), cursor_(other.cursor_) {
other.cursor_ = 0;
}
NodesTSIterator(const NodesTSIterator &other);
Node &operator*();
NodesTSIterator &operator++();
// TODO(Superjomn) current implementation just compare the first
// element, need to compare the graph and all the elements in the queue and
// set.
NodesTSIterator &operator=(const NodesTSIterator &other);
bool operator==(const NodesTSIterator &other);
bool operator!=(const NodesTSIterator &other) { return !(*this == other); }
Node *operator->();
private:
std::vector<Node *> sorted_;
size_t cursor_{0};
};
/*
* GraphTraits contains some graph traversal algorithms.
*
......@@ -76,6 +102,14 @@ struct GraphTraits {
NodesDFSIterator());
}
static iterator_range<NodesTSIterator> TS(const Graph &g) {
auto start_points = ExtractStartPoints(g);
PADDLE_ENFORCE(!start_points.empty());
NodesTSIterator x(start_points);
return iterator_range<NodesTSIterator>(NodesTSIterator(start_points),
NodesTSIterator());
}
private:
// The nodes those have no input will be treated as start points.
static std::vector<Node *> ExtractStartPoints(const Graph &g) {
......
paddle/fluid/framework/ir/graph_viz_pass.cc
浏览文件 @ 2412c27c
......@@ -41,7 +41,7 @@ std::string FormatName(const Node* node) {
std::unique_ptr<ir::Graph> GraphVizPass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
const std::string graph_viz_path = Get<std::string>(kGraphVizPath);
VLOG(3) << "draw IR graph viz to " << graph_viz_path;
VLOG(30) << "draw IR graph viz to " << graph_viz_path;
std::unique_ptr<std::ostream> fout(new std::ofstream(graph_viz_path));
PADDLE_ENFORCE(fout->good());
std::ostream& sout = *fout;
......
paddle/fluid/framework/ir/mkldnn_placement_pass.cc
浏览文件 @ 2412c27c
......@@ -20,7 +20,7 @@ namespace ir {
std::unique_ptr<ir::Graph> MKLDNNPlacementPass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
VLOG(3) << "Aplies MKL-DNN placement strategy.";
VLOG(30) << "Aplies MKL-DNN placement strategy.";
for (const Node* n : graph->Nodes()) {
if (n->IsOp() && n->Op()->HasAttr("use_mkldnn")) {
n->Op()->SetAttr("use_mkldnn", true);
......
paddle/fluid/framework/ir/multi_batch_merge_pass.cc
浏览文件 @ 2412c27c
......@@ -62,7 +62,7 @@ VarDesc UpdateGradVarDesc(
string::Sprintf("%s.repeat.%d", var_desc->Name(), repeat);
VarDesc repeated_var = CopyVarDesc(var_desc);
repeated_var.SetName(new_gname);
VLOG(3) << "update " << var_desc->Name() << " to repeat " << repeat;
VLOG(30) << "update " << var_desc->Name() << " to repeat " << repeat;
return repeated_var;
}
return *var_desc;
......@@ -78,7 +78,7 @@ std::unique_ptr<Graph> BatchMergePass::ApplyImpl(
std::vector<ir::Node*> nodes = TopologySortOperations(*graph);
auto origin_nodes = graph->ReleaseNodes();
VLOG(3) << "origin nodes count: " << origin_nodes.size();
VLOG(30) << "origin nodes count: " << origin_nodes.size();
ir::Graph& result = *graph;
// 1. record op nodes of different roles
......@@ -137,8 +137,8 @@ std::unique_ptr<Graph> BatchMergePass::ApplyImpl(
"%s.repeat.%d", repeated_op.Input("Variance")[0], i);
bn_vars_need_rename.insert(repeated_op.Input("Mean")[0]);
bn_vars_need_rename.insert(repeated_op.Input("Variance")[0]);
VLOG(3) << "renaming " << repeated_op.Input("Mean")[0] << " to "
<< new_mean_name;
VLOG(30) << "renaming " << repeated_op.Input("Mean")[0] << " to "
<< new_mean_name;
repeated_op.RenameInput(repeated_op.Input("Mean")[0], new_mean_name);
repeated_op.RenameInput(repeated_op.Input("Variance")[0], new_var_name);
repeated_op.RenameOutput(repeated_op.Output("MeanOut")[0],
......
paddle/fluid/framework/ir/node.cc
浏览文件 @ 2412c27c
......@@ -18,7 +18,6 @@ namespace paddle {
namespace framework {
namespace ir {
constexpr char Node::kControlDepVarName[];
int Node::count_ = 0;
std::unique_ptr<Node> CreateNodeForTest(const std::string& name,
Node::Type type) {
......
paddle/fluid/framework/ir/node.h
浏览文件 @ 2412c27c
......@@ -15,7 +15,10 @@ limitations under the License. */
#pragma once
#include <string>
#include <typeindex>
#include <typeinfo>
#include <vector>
#include "paddle/fluid/framework/op_desc.h"
#include "paddle/fluid/framework/var_desc.h"
#include "paddle/fluid/platform/macros.h"
......@@ -24,9 +27,33 @@ namespace paddle {
namespace framework {
namespace ir {
// Node should normally created by Graph::CreateXXXNode().
// Node should only created by Graph::CreateXXXNode().
// 1. Every Node should be part of a graph. No dangling Node exists.
// 2. Node only contains members necessary for building graph structure.
// It doesn't contain other unrelated members, such as device, etc.
//
// Sometimes, for specific usages, Node needs to have additional members,
// such as device_placement, version in order to be executed. It is suggested
// to use composition pattern.
//
// class RunnableOp {
// RunnableOp(ir::Node* n) : n_(n) { n_.WrappedBy(this); }
//
// int any_thing_;
// }
//
// RunnableOp is owned by the ir::Node that composes it. In other words.
// ir::Node will be responsible for deleting RunnableOp, say, when ir::Node
// is deleted from the graph.
class Node {
public:
virtual ~Node() {
if (!wrapper_.empty()) {
VLOG(4) << "ir::Node deleting a wrapper node " << Name();
wrapper_deleter_();
}
}
enum class Type { kOperation, kVariable };
static constexpr char kControlDepVarName[] = "__control_var";
......@@ -44,6 +71,29 @@ class Node {
return op_desc_.get();
}
// Set the `wrapper` that wraps the Node. `wrapper` is owned by Node.
template <typename T>
void WrappedBy(T* wrapper) {
if (!wrapper_.empty()) {
wrapper_deleter_();
}
wrapper_ = wrapper;
wrapper_deleter_ = [wrapper]() { delete wrapper; };
wrapper_type_ = std::type_index(typeid(T));
}
// Return a reference to the `wrapper`.
template <typename T>
T& Wrapper() {
return *boost::any_cast<T*>(wrapper_);
}
// Test if the Node is wrapped by type T.
template <typename T>
bool IsWrappedBy() {
return std::type_index(typeid(T)) == wrapper_type_;
}
// Please don't use this API!
int id() const { return id_; }
......@@ -65,36 +115,34 @@ class Node {
int id_;
private:
// ID can only set by a Graph.
void SetId(int id) { id_ = id; }
friend class Graph;
friend std::unique_ptr<Node> CreateNodeForTest(const std::string& name,
Node::Type type);
explicit Node(const std::string& name, Type type)
: name_(name),
var_desc_(nullptr),
op_desc_(nullptr),
type_(type),
id_(count_++) {}
: name_(name), var_desc_(nullptr), op_desc_(nullptr), type_(type) {}
explicit Node(VarDesc* var_desc)
: name_(var_desc->Name()),
var_desc_(new VarDesc(*var_desc)),
op_desc_(nullptr),
type_(Type::kVariable),
id_(count_++) {}
type_(Type::kVariable) {}
explicit Node(OpDesc* op_desc)
: name_(op_desc->Type()),
var_desc_(nullptr),
op_desc_(new OpDesc(*op_desc, op_desc->Block())),
type_(Type::kOperation),
id_(count_++) {}
type_(Type::kOperation) {}
Node() = delete;
static int count_;
// Please don't use this API or make this public.
static void ResetId() { count_ = 0; }
boost::any wrapper_;
std::function<void(void)> wrapper_deleter_;
std::type_index wrapper_type_ = std::type_index(typeid(void));
DISABLE_COPY_AND_ASSIGN(Node);
};
......
paddle/fluid/inference/analysis/tensorrt_subgraph_pass.h → paddle/fluid/framework/ir/node_test.cc
浏览文件 @ 2412c27c
/* 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 <string>
#include "paddle/fluid/inference/analysis/analysis_pass.h"
#include "paddle/fluid/inference/analysis/node.h"
#include "paddle/fluid/inference/analysis/subgraph_splitter.h"
namespace paddle {
namespace inference {
namespace analysis {
/*
* Parse the graph and replace TensorRT supported nodes with SubGraphNode
*/
class TensorRTSubGraphPass : public DataFlowGraphPass {
public:
// Tell whether to transform a sub-graph into TensorRT.
using NodeInsideSubgraphTeller = SubGraphFuse::NodeInsideSubgraphTeller;
explicit TensorRTSubGraphPass(const NodeInsideSubgraphTeller& teller);
bool Initialize(Argument* argument) override {
argument_ = argument;
return true;
}
// This class get a sub-graph as input and determine whether to transform this
// sub-graph into TensorRT.
void Run(DataFlowGraph* graph) override;
bool Finalize() override { return true; }
std::string repr() const override { return "tensorrt-sub-graph"; }
std::string description() const override { return "tensorrt sub graph pass"; }
private:
NodeInsideSubgraphTeller node_inside_subgraph_teller_;
Argument* argument_;
};
} // namespace analysis
} // namespace inference
} // namespace paddle
/* 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 <string>
#include "gtest/gtest.h"
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/pass.h"
namespace paddle {
namespace framework {
namespace ir {
class RunnableOp {
public:
RunnableOp(Node* node, bool* alive) : node_(node), alive_(alive) {
node_->WrappedBy(this);
}
virtual ~RunnableOp() { *alive_ = false; }
private:
Node* node_;
bool* alive_;
};
class RunnableOp2 {
public:
RunnableOp2(Node* node, bool* alive) : node_(node), alive_(alive) {
node_->WrappedBy(this);
}
virtual ~RunnableOp2() { *alive_ = false; }
private:
Node* node_;
bool* alive_;
};
TEST(NodeTest, Basic) {
bool alive1 = true;
bool alive2 = true;
std::unique_ptr<Node> n1(CreateNodeForTest("n1", Node::Type::kVariable));
std::unique_ptr<Node> n2(CreateNodeForTest("n2", Node::Type::kVariable));
EXPECT_FALSE(n1->IsWrappedBy<RunnableOp>());
EXPECT_FALSE(n1->IsWrappedBy<RunnableOp2>());
EXPECT_FALSE(n2->IsWrappedBy<RunnableOp>());
EXPECT_FALSE(n2->IsWrappedBy<RunnableOp2>());
new RunnableOp(n1.get(), &alive1);
new RunnableOp2(n2.get(), &alive2);
EXPECT_TRUE(n1->IsWrappedBy<RunnableOp>());
EXPECT_FALSE(n1->IsWrappedBy<RunnableOp2>());
EXPECT_FALSE(n2->IsWrappedBy<RunnableOp>());
EXPECT_TRUE(n2->IsWrappedBy<RunnableOp2>());
EXPECT_TRUE(alive1);
EXPECT_TRUE(alive2);
n1.reset(nullptr);
n2.reset(nullptr);
EXPECT_FALSE(alive1);
EXPECT_FALSE(alive2);
}
} // namespace ir
} // namespace framework
} // namespace paddle
paddle/fluid/framework/ir/pass.h
浏览文件 @ 2412c27c
......@@ -76,7 +76,7 @@ class Pass {
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;
};
}
......@@ -93,6 +93,7 @@ class Pass {
protected:
virtual std::unique_ptr<Graph> ApplyImpl(std::unique_ptr<Graph> graph) const {
LOG(FATAL) << "Calling virtual Pass not implemented.";
return graph;
}
private:
......
paddle/fluid/framework/ir/seq_concat_fc_fuse_pass.cc
浏览文件 @ 2412c27c
......@@ -12,10 +12,13 @@
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/framework/ir/seq_concat_fc_fuse_pass.h"
#include <set>
#include <string>
#include "paddle/fluid/framework/ir/fuse_pass_base.h"
#include "paddle/fluid/framework/ir/graph_pattern_detector.h"
#include "paddle/fluid/framework/ir/graph_viz_pass.h"
#include "paddle/fluid/framework/ir/seq_concat_fc_fuse_pass.h"
#include "paddle/fluid/framework/lod_tensor.h"
namespace paddle {
......@@ -159,10 +162,7 @@ PDNode* BuildFCPattern(PDPattern* pattern, PDNode* fc_x) {
std::set<std::string> acts({"sigmoid", "tanh", "relu", "identity"});
PDNode* act = pattern->NewNode(
[=](Node* x) {
return x && x->IsOp() && acts.count(x->Op()->Type());
},
[=](Node* x) { return x && x->IsOp() && acts.count(x->Op()->Type()); },
"act");
PDNode* fc_out = pattern->NewNode(
......@@ -196,7 +196,7 @@ std::unique_ptr<ir::Graph> SeqConcatFcFusePass::ApplyImpl(
detector(graph.get(), [&](const GraphPatternDetector::subgraph_t& subgraph,
Graph* graph) {
VLOG(4) << "get one concat pattern";
VLOG(40) << "get one concat pattern";
// fc
GET_NODE(fc_w, detector.pattern());
GET_NODE(fc_bias, detector.pattern());
......
paddle/fluid/framework/ir/seqconv_eltadd_relu_fuse_pass.cc
浏览文件 @ 2412c27c
......@@ -60,7 +60,7 @@ int BuildFusion(Graph* graph, const std::string& name_scope, Scope* scope) {
auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
Graph* g) {
VLOG(4) << "handle SeqConv EltAdd Relu fuse";
VLOG(40) << "handle SeqConv EltAdd Relu fuse";
GET_IR_NODE_FROM_SUBGRAPH(seqconv, seqconv, fuse_pattern);
GET_IR_NODE_FROM_SUBGRAPH(seqconv_weight, seqconv_weight, fuse_pattern);
GET_IR_NODE_FROM_SUBGRAPH(seqconv_out, seqconv_out, fuse_pattern);
......
paddle/fluid/framework/lod_rank_table.cc
浏览文件 @ 2412c27c
......@@ -31,7 +31,7 @@ void LoDRankTable::Reset(const LoD& lod, size_t level) {
TableItem item;
item.index = i;
item.length = vec[i + 1] - vec[i];
VLOG(10) << "Add item to rank table " << item.index << " " << item.length;
VLOG(100) << "Add item to rank table " << item.index << " " << item.length;
items_.emplace_back(item);
}
// NOTE(yuyang18):
......
paddle/fluid/framework/mixed_vector_test.cc
浏览文件 @ 2412c27c
......@@ -51,7 +51,7 @@ TEST(mixed_vector, InitWithCount) {
TEST(mixed_vector, ForEach) {
vec<int> tmp;
for (auto& v : tmp) {
VLOG(3) << v;
VLOG(30) << v;
}
}
......
paddle/fluid/framework/naive_executor.cc
浏览文件 @ 2412c27c
......@@ -57,60 +57,58 @@ static void InitializeVariable(Variable *var, proto::VarType::Type var_type) {
}
}
void NaiveExecutor::Prepare(Scope *parent_scope,
const ProgramDesc &program_desc, int block_id,
bool with_feed_fetch_ops) {
if (!parent_scope) {
void NaiveExecutor::Prepare(Scope *scope, const ProgramDesc &program_desc,
int block_id, bool with_feed_fetch_ops) {
if (!scope) {
scope_ = new framework::Scope;
} else {
scope_ = &parent_scope->NewScope();
scope_ = scope;
}
CreateVariables(program_desc, scope_, block_id);
VLOG(3) << "NaiveExecutor init with scope " << scope;
CreateOps(program_desc, block_id, with_feed_fetch_ops);
}
void NaiveExecutor::Run() {
for (auto &op : ops_) {
VLOG(4) << "run " << op->Type();
VLOG(3) << std::this_thread::get_id() << " run " << op->Type()
<< " on scope " << scope_;
op->Run(*scope_, place_);
}
}
void NaiveExecutor::CreateVariables(const ProgramDesc &desc, Scope *scope,
int block_id) {
PADDLE_ENFORCE(scope);
void NaiveExecutor::CreateVariables(const ProgramDesc &desc, int block_id,
bool persistable, Scope *scope) {
PADDLE_ENFORCE_NOT_NULL(scope);
auto &global_block = desc.Block(block_id);
const Scope *ancestor_scope = scope;
while (ancestor_scope->parent()) {
ancestor_scope = ancestor_scope->parent();
const auto *anc = scope;
PADDLE_ENFORCE(anc->parent() != anc);
while (anc->parent()) {
anc = anc->parent();
}
if (ancestor_scope != scope) {
for (auto &var : global_block.AllVars()) {
if (var->Name() == framework::kEmptyVarName) {
continue;
}
// Create persistable vars in ancestor 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;
} else { // Create temporary variables in local scope.
auto *ptr = scope->Var(var->Name());
for (auto &var : global_block.AllVars()) {
if (var->Name() == framework::kEmptyVarName) {
continue;
}
if (persistable == var->Persistable()) {
if (persistable) {
if (!anc->FindVar(var->Name())) {
auto *ptr = const_cast<Scope *>(anc)->Var(var->Name());
VLOG(3) << scope << " Create persistable variable " << var->Name()
<< ", which pointer is " << ptr;
InitializeVariable(ptr, var->GetType());
}
} else {
auto *ptr = const_cast<Scope *>(scope)->Var(var->Name());
VLOG(3) << scope << " Create variable " << var->Name()
<< ", which pointer is " << ptr;
InitializeVariable(ptr, var->GetType());
VLOG(3) << "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;
}
}
}
......
paddle/fluid/framework/naive_executor.h
浏览文件 @ 2412c27c
......@@ -35,8 +35,14 @@ class NaiveExecutor {
// Create child scope.
// Create variables.
// @with_feed_fetch_ops: whether to work with the feed and fetch operators.
void Prepare(Scope* parent_scope, const ProgramDesc& program_desc,
int block_id, bool with_feed_fetch_ops);
void Prepare(Scope* scope, const ProgramDesc& program_desc, int block_id,
bool with_feed_fetch_ops);
// Create variables before head.
// Create parameters if persistable is ture, or create the temporary variables
// instead.
void CreateVariables(const ProgramDesc& desc, int block_id, bool persistable,
Scope* scope);
// Run all the operators.
void Run();
......@@ -49,8 +55,6 @@ class NaiveExecutor {
void CleanFeedFetchOps();
protected:
void CreateVariables(const ProgramDesc& desc, Scope* scope, int block_id);
void CreateOps(const ProgramDesc& desc, int block_id,
bool with_feed_fetch_ops);
......
paddle/fluid/framework/naive_executor_test.cc
浏览文件 @ 2412c27c
......@@ -39,7 +39,7 @@ TEST(NaiveExecutor, Basic) {
auto place = platform::CPUPlace();
NaiveExecutor exe(place);
exe.Prepare(nullptr, program, 0, false /*with feed fetch ops*/);
exe.Prepare(nullptr, program, 0, false);
auto* a_tensor = exe.FindTensor("a");
auto* b_tensor = exe.FindTensor("b");
auto* c_tensor = exe.FindTensor("c");
......
paddle/fluid/inference/analysis/graph_traits.cc → paddle/fluid/framework/ngraph_bridge.cc
浏览文件 @ 2412c27c
/* 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/inference/analysis/graph_traits.h"
/* 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. */
#ifdef PADDLE_WITH_NGRAPH
#include <algorithm>
#include <functional>
#include "paddle/fluid/framework/ngraph_bridge.h"
#include "ngraph/ngraph.hpp"
namespace paddle {
namespace framework {
std::map<std::string,
std::function<void(const std::shared_ptr<OperatorBase>&,
std::shared_ptr<std::unordered_map<
std::string, std::shared_ptr<ngraph::Node>>>)>>
NgraphBridge::NG_NODE_MAP = {};
void NgraphBridge::build_graph(const std::shared_ptr<OperatorBase>& op) {
auto& op_type = op->Type();
NG_NODE_MAP[op_type](op, ngb_node_map);
}
} // namespace framework
} // namespace paddle
#endif
paddle/fluid/inference/analysis/dfg_graphviz_draw_pass_tester.cc → paddle/fluid/framework/ngraph_bridge.h
浏览文件 @ 2412c27c
......@@ -4,7 +4,7 @@ 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
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,
......@@ -12,43 +12,47 @@ 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/inference/analysis/dfg_graphviz_draw_pass.h"
#pragma once
#include <gtest/gtest.h>
#include <fstream>
#ifdef PADDLE_WITH_NGRAPH
#include <algorithm>
#include <map>
#include <string>
#include "paddle/fluid/inference/analysis/ut_helper.h"
#include <unordered_map>
#include <vector>
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/platform/enforce.h"
#include "ngraph/ngraph.hpp"
namespace paddle {
namespace inference {
namespace analysis {
TEST(DFG_GraphvizDrawPass, dfg_graphviz_draw_pass_tester) {
Argument argument(FLAGS_inference_model_dir);
FluidToDataFlowGraphPass pass0;
ASSERT_TRUE(pass0.Initialize(&argument));
pass0.Run(argument.main_dfg.get());
// auto dfg = ProgramDescToDFG(*argument.origin_program_desc);
DFG_GraphvizDrawPass::Config config("./", "test");
DFG_GraphvizDrawPass pass(config);
pass.Initialize(&argument);
pass.Run(argument.main_dfg.get());
// test content
std::ifstream file("./0-graph_test.dot");
ASSERT_TRUE(file.is_open());
std::string line;
int no{0};
while (std::getline(file, line)) {
no++;
}
// DFG is sensitive to ProgramDesc, be careful to change the existing models.
ASSERT_EQ(no, 83);
}
} // namespace analysis
} // namespace inference
namespace framework {
class NgraphBridge {
public:
static std::map<
std::string,
std::function<void(const std::shared_ptr<OperatorBase>&,
std::shared_ptr<std::unordered_map<
std::string, std::shared_ptr<ngraph::Node>>>)>>
NG_NODE_MAP;
explicit NgraphBridge(
std::shared_ptr<
std::unordered_map<std::string, std::shared_ptr<ngraph::Node>>>
var_node_map)
: ngb_node_map(var_node_map) {}
void build_graph(const std::shared_ptr<OperatorBase>& op);
private:
std::shared_ptr<
std::unordered_map<std::string, std::shared_ptr<ngraph::Node>>>
ngb_node_map;
};
} // namespace framework
} // namespace paddle
#endif
paddle/fluid/framework/ngraph_operator.cc 0 → 100644
浏览文件 @ 2412c27c
/* 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. */
#ifdef PADDLE_WITH_NGRAPH
#include <glog/logging.h>
#include <algorithm>
#include <map>
#include "paddle/fluid/framework/feed_fetch_type.h"
#include "paddle/fluid/framework/ngraph_operator.h"
#include "paddle/fluid/framework/shape_inference.h"
#include "paddle/fluid/framework/var_desc.h"
#include "paddle/fluid/framework/var_type.h"
namespace paddle {
namespace framework {
static std::map<proto::VarType::Type, ngraph::element::Type> pd2ng_type_map = {
{proto::VarType::FP32, ngraph::element::f32},
{proto::VarType::FP64, ngraph::element::f64},
{proto::VarType::INT32, ngraph::element::i32},
{proto::VarType::INT64, ngraph::element::i64},
{proto::VarType::BOOL, ngraph::element::boolean},
};
typedef enum { /* nGraph support state on ops */
FULL_TRAIN, /* Support full ops for train */
PARTIAL_TRAIN, /* Support partial ops for train */
FULL_TEST, /* Support full list of ops for test */
PARTIAL_TEST /* Support partial list of ops for test */
} op_state;
class NgraphOperator {
public:
explicit NgraphOperator(const Scope& scope, const platform::Place& place,
const std::vector<std::shared_ptr<OperatorBase>>& ops,
const std::unordered_map<
std::string, ngraph::element::Type>& var_type_map,
const std::unordered_set<std::string>& persist,
const std::unordered_set<std::string>& fetches,
const std::unordered_set<std::string>& post_op_inputs,
op_state ng_op_state)
: scope_(scope),
place_(place),
fused_ops_(ops),
var_type_map_(var_type_map),
persistables_(persist),
fetches_(fetches),
post_op_inputs_(post_op_inputs),
ng_op_state_(ng_op_state) {}
void Run(const Scope& scope, const platform::Place& place) const;
private:
static std::unordered_map<std::string, std::shared_ptr<ngraph::Function>>
func_cache;
const Scope& scope_;
const platform::Place& place_;
std::vector<std::shared_ptr<OperatorBase>> fused_ops_;
std::unordered_map<std::string, ngraph::element::Type> var_type_map_;
std::unordered_set<std::string> persistables_;
std::unordered_set<std::string> fetches_;
std::unordered_set<std::string> post_op_inputs_;
op_state ng_op_state_;
};
std::vector<std::vector<std::vector<std::unique_ptr<OperatorBase>>::iterator>>
FusedOperator::FusedOpIntervals(
std::vector<std::unique_ptr<paddle::framework::OperatorBase>>* ops) {
std::vector<std::vector<std::vector<std::unique_ptr<OperatorBase>>::iterator>>
intervals;
if (ops->empty()) {
return intervals;
}
size_t size = ops->size();
size_t left = 0;
while (left < size && ops.at(left)->Type() != kFeedOpType) {
++left;
}
if (left == size) {
return intervals;
}
while (left < size && ops->at(left)->Type() == kFeedOpType) {
++left;
}
size_t right = left;
while (right < size && ops->at(right)->Type() != kFetchOpType) {
++right;
}
if (right == size) {
return intervals;
}
if (left >= right) return intervals;
// (left, right - 1) represents indices between feed and fetch
size_t pivot = left;
while (pivot < right) {
auto op_type = ops->at(pivot)->Type();
if (paddle::framework::NgraphBridge::NG_NODE_MAP.find(op_type) ==
paddle::framework::NgraphBridge::NG_NODE_MAP.end()) {
++pivot;
} else {
size_t start = pivot, end = start;
while (pivot < right &&
(paddle::framework::NgraphBridge::NG_NODE_MAP.find(
ops.at(pivot)->Type()) !=
paddle::framework::NgraphBridge::NG_NODE_MAP.end())) {
++pivot;
++end;
}
std::vector<std::vector<std::unique_ptr<OperatorBase>>::iterator>
interval = {ops->begin() + start, ops->begin() + end};
intervals.push_back(interval);
}
} // end while
return intervals;
}
FusedOperator::FusedOperator(
const ProgramDesc& prog, size_t block_id,
std::vector<std::unique_ptr<OperatorBase>>::iterator start,
std::vector<std::unique_ptr<OperatorBase>>::iterator end,
const std::string& type, const VariableNameMap& inputs,
const VariableNameMap& outputs, const AttributeMap& attrs)
: OperatorBase(type, inputs, outputs, attrs), pdesc(prog), block(block_id) {
for (std::vector<std::unique_ptr<OperatorBase>>::iterator it = start;
it != end; ++it) {
fused_ops_.push_back(std::move(*it));
}
for (std::vector<std::unique_ptr<OperatorBase>>::iterator it = end;
(*it)->Type() != kFetchOpType; ++it) {
for (auto& var_name_item : (*it)->Inputs()) {
for (auto& var_name : var_name_item.second) {
post_op_inputs_.insert(var_name);
}
}
}
if ((*(start - 1))->Type() == kFeedOpType && (*end)->Type() == kFetchOpType) {
is_complete = true;
}
Process();
}
void FusedOperator::Process() {
auto& bdesc = pdesc_.Block(block_);
for (auto& var : bdesc.AllVars()) {
if (!(var->GetType() == proto::VarType::SELECTED_ROWS ||
var->GetType() == proto::VarType::LOD_TENSOR ||
var->GetType() == proto::VarType::LOD_TENSOR_ARRAY)) {
continue;
}
auto var_name = var->Name();
if (var->Name() == framework::kEmptyVarName) {
continue;
}
if (var_name != "fetch" && var_name != "feed") {
auto pd_type = var->GetDataType();
if (pd2ng_type_map.find(pd_type) == pd2ng_type_map.end()) {
PADDLE_THROW("Data type of var %s not found in pd2ng_type_map",
var_name);
}
var_type_map_[var_name] = pd2ng_type_map[pd_type];
}
if (var->Persistable()) {
persistables_.insert(var->Name());
}
}
for (auto* op : bdesc.AllOps()) {
if (op->Type() == kFetchOpType) {
std::string fetch_target_name = op->Input("X")[0];
fetches_.insert(fetch_target_name);
}
}
}
void FusedOperator::RunImpl(const Scope& scope,
const platform::Place& place) const {
op_state ng_op_state = PARTIAL_TEST;
auto& bdesc = pdesc_.Block(block_);
for (auto* op : bdesc.AllOps()) {
if (op->Type().find("_grad") != std::string::npos) {
ng_op_state = PARTIAL_TRAIN;
break;
}
}
if (is_full) {
ng_op_state = ng_op_state == PARTIAL_TEST ? FULL_TEST : FULL_TRAIN;
}
NgraphOperator ngraph_op(scope, place, fused_ops_, var_type_map_,
persistables_, fetches_, post_op_inputs_,
ng_op_state);
ngraph_op.Run(scope, place);
}
} // namespace framework
} // namespace paddle
#endif
paddle/fluid/framework/ngraph_operator.h 0 → 100644
浏览文件 @ 2412c27c
/* 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
#ifdef PADDLE_WITH_NGRAPH
#include <algorithm>
#include <atomic>
#include <string>
#include <unordered_map>
#include <vector>
#include "paddle/fluid/framework/attribute.h"
#include "paddle/fluid/framework/framework.pb.h"
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/ngraph_bridge.h"
#include "paddle/fluid/framework/op_info.h"
#include "paddle/fluid/framework/op_kernel_type.h"
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/framework/program_desc.h"
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/framework/tensor.h"
#include "paddle/fluid/platform/variant.h"
#include "ngraph/ngraph.hpp"
namespace paddle {
namespace framework {
class FusedOperator : public OperatorBase {
public:
static std::vector<
std::vector<std::vector<std::unique_ptr<OperatorBase>>::iterator>>
FusedOpIntervals(
std::vector<std::unique_ptr<paddle::framework::OperatorBase>>* ops);
explicit FusedOperator(
const ProgramDesc& prog, size_t block_id,
std::vector<std::unique_ptr<OperatorBase>>::iterator start,
std::vector<std::unique_ptr<OperatorBase>>::iterator end,
const std::string& type = "fused_op", const VariableNameMap& inputs = {},
const VariableNameMap& outputs = {}, const AttributeMap& attrs = {});
void RunImpl(const Scope& scope, const platform::Place& place) const final;
private:
const ProgramDesc pdesc_;
size_t block_;
std::vector<std::shared_ptr<OperatorBase>> fused_ops_;
std::unordered_map<std::string, ngraph::element::Type> var_type_map_;
std::unordered_set<std::string> persistables_;
std::unordered_set<std::string> fetches_;
std::unordered_set<std::string> post_op_inputs_;
bool is_full_ = false;
void Process();
};
} // namespace framework
} // namespace paddle
#endif
paddle/fluid/framework/op_desc.cc
浏览文件 @ 2412c27c
......@@ -82,7 +82,7 @@ class CompileTimeInferShapeContext : public InferShapeContext {
auto *in_var = block_.FindVarRecursive(Inputs(in)[i]);
auto *out_var = block_.FindVarRecursive(Outputs(out)[j]);
if (in_var->GetType() != proto::VarType::LOD_TENSOR) {
VLOG(3) << "input " << in << " is not LodTensor";
VLOG(30) << "input " << in << " is not LodTensor";
return;
}
out_var->SetLoDLevel(in_var->GetLoDLevel());
......@@ -241,32 +241,32 @@ void OpDesc::SetAttr(const std::string &name, const Attribute &v) {
const proto::OpProto::Attr &attr = GetProtoAttr(name);
switch (attr.type()) {
case proto::AttrType::BOOLEANS: {
VLOG(11) << "SetAttr: " << Type() << ", " << name
<< " from INTS to BOOLEANS";
VLOG(110) << "SetAttr: " << Type() << ", " << name
<< " from INTS to BOOLEANS";
this->attrs_[name] = std::vector<bool>();
break;
}
case proto::AttrType::INTS: {
VLOG(11) << "SetAttr: " << Type() << ", " << name
<< " from INTS to INTS";
VLOG(110) << "SetAttr: " << Type() << ", " << name
<< " from INTS to INTS";
this->attrs_[name] = std::vector<int>();
break;
}
case proto::AttrType::FLOATS: {
VLOG(11) << "SetAttr: " << Type() << ", " << name
<< " from INTS to FLOATS";
VLOG(110) << "SetAttr: " << Type() << ", " << name
<< " from INTS to FLOATS";
this->attrs_[name] = std::vector<float>();
break;
}
case proto::AttrType::STRINGS: {
VLOG(11) << "SetAttr: " << Type() << ", " << name
<< " from INTS to STRINGS";
VLOG(110) << "SetAttr: " << Type() << ", " << name
<< " from INTS to STRINGS";
this->attrs_[name] = std::vector<std::string>();
break;
}
case proto::AttrType::BLOCKS: {
VLOG(11) << "SetAttr: " << Type() << ", " << name
<< " from INTS to BLOCKS";
VLOG(110) << "SetAttr: " << Type() << ", " << name
<< " from INTS to BLOCKS";
this->SetBlocksAttr(name, std::vector<BlockDesc *>());
return;
}
......@@ -499,13 +499,13 @@ void OpDesc::CheckAttrs() {
}
void OpDesc::InferShape(const BlockDesc &block) const {
VLOG(3) << "CompileTime infer shape on " << Type();
VLOG(30) << "CompileTime infer shape on " << Type();
InitInferShapeFuncs();
auto &infer_shape = OpInfoMap::Instance().Get(this->Type()).infer_shape_;
PADDLE_ENFORCE(static_cast<bool>(infer_shape),
"%s's infer_shape has not been registered", this->Type());
CompileTimeInferShapeContext ctx(*this, block);
if (VLOG_IS_ON(10)) {
if (VLOG_IS_ON(100)) {
std::ostringstream sout;
auto inames = this->InputArgumentNames();
sout << " From [";
......@@ -516,7 +516,7 @@ void OpDesc::InferShape(const BlockDesc &block) const {
std::copy(onames.begin(), onames.end(),
std::ostream_iterator<std::string>(sout, ", "));
sout << "]";
VLOG(10) << sout.str();
VLOG(100) << sout.str();
}
infer_shape(&ctx);
}
......@@ -607,7 +607,7 @@ DDim CompileTimeInferShapeContext::GetDim(const std::string &name) const {
auto shape = var->GetShape();
res = shape.empty() ? make_ddim({0UL}) : make_ddim(shape);
} catch (...) {
VLOG(5) << "GetDim of variable " << name << " error";
VLOG(50) << "GetDim of variable " << name << " error";
std::rethrow_exception(std::current_exception());
}
return res;
......@@ -624,7 +624,7 @@ std::vector<DDim> CompileTimeInferShapeContext::GetRepeatedDims(
res.push_back(s.empty() ? make_ddim({0UL}) : make_ddim(s));
}
} catch (...) {
VLOG(5) << "GetRepeatedDim of variable " << name << " error.";
VLOG(50) << "GetRepeatedDim of variable " << name << " error.";
std::rethrow_exception(std::current_exception());
}
return res;
......
paddle/fluid/framework/op_registry.cc
浏览文件 @ 2412c27c
......@@ -46,9 +46,9 @@ static VariableNameMap ConvertOpDescVarsToVarNameMap(
std::unique_ptr<OperatorBase> OpRegistry::CreateOp(
const proto::OpDesc& op_desc) {
VLOG(1) << "CreateOp directly from OpDesc is deprecated. It should only be"
"used in unit tests. Use CreateOp(const OpDesc& op_desc) "
"instead.";
VLOG(10) << "CreateOp directly from OpDesc is deprecated. It should only be"
"used in unit tests. Use CreateOp(const OpDesc& op_desc) "
"instead.";
VariableNameMap inputs = ConvertOpDescVarsToVarNameMap(op_desc.inputs());
VariableNameMap outputs = ConvertOpDescVarsToVarNameMap(op_desc.outputs());
AttributeMap attrs;
......
paddle/fluid/framework/operator.cc
浏览文件 @ 2412c27c
......@@ -140,7 +140,7 @@ static LoD GetLoD(const Scope& scope, const std::string& name) {
}
void OperatorBase::Run(const Scope& scope, const platform::Place& place) {
VLOG(4) << place << " " << DebugStringEx(&scope);
VLOG(40) << place << " " << DebugStringEx(&scope);
if (platform::is_gpu_place(place)) {
#ifndef PADDLE_WITH_CUDA
PADDLE_THROW("Cannot run operator on place %s", place);
......@@ -160,7 +160,7 @@ void OperatorBase::Run(const Scope& scope, const platform::Place& place) {
} else {
RunImpl(scope, place);
}
VLOG(3) << place << " " << DebugStringEx(&scope);
VLOG(30) << place << " " << DebugStringEx(&scope);
}
bool OperatorBase::HasInputs(const std::string& name) const {
......@@ -259,6 +259,8 @@ std::string OperatorBase::DebugStringEx(const Scope* scope) const {
if (row_size >= 0) {
ss << "[row_size=" << row_size << "]";
}
std::string dtype = GetDtype(*scope, output.second[i]);
ss << ":" << dtype;
ss << "[" << GetDims(*scope, var_name, true) << "]";
ss << "(" << GetLoD(*scope, var_name) << ")";
}
......@@ -358,7 +360,7 @@ static bool VarIsTensor(const Variable& var) {
return var.IsType<LoDTensor>() || var.IsType<SelectedRows>();
}
const Tensor* GetTensorFromVar(const Variable& var) {
const Tensor* GetLoDTensorOrSelectedRowsValueFromVar(const Variable& var) {
if (var.IsType<LoDTensor>()) {
return static_cast<const Tensor*>(&(var.Get<LoDTensor>()));
} else if (var.IsType<SelectedRows>()) {
......@@ -369,7 +371,7 @@ const Tensor* GetTensorFromVar(const Variable& var) {
}
}
static Tensor* GetMutableTensorFromVar(Variable* var) {
Tensor* GetMutableLoDTensorOrSelectedRowsValueFromVar(Variable* var) {
if (var->IsType<LoDTensor>()) {
return var->GetMutable<LoDTensor>();
} else if (var->IsType<SelectedRows>()) {
......@@ -414,8 +416,7 @@ bool ExecutionContext::HasOutput(const std::string& name) const {
template <>
const Tensor* ExecutionContext::Input<Tensor>(const std::string& name) const {
auto* var = InputVar(name);
return var == nullptr ? nullptr : GetTensorFromVar(*var);
return Input<LoDTensor>(name);
}
template <>
......@@ -425,17 +426,21 @@ const std::vector<const Tensor*> ExecutionContext::MultiInput<Tensor>(
std::vector<const Tensor*> res;
res.reserve(names.size());
std::transform(names.begin(), names.end(), std::back_inserter(res),
[&](const std::string& sub_name) {
[&](const std::string& sub_name) -> const Tensor* {
auto var = scope_.FindVar(sub_name);
return var == nullptr ? nullptr : GetTensorFromVar(*var);
if (var == nullptr) return nullptr;
PADDLE_ENFORCE(
var->IsType<LoDTensor>(),
"%s should be LoDTensor, but the received type is %s",
sub_name, var->Type().name());
return &(var->Get<LoDTensor>());
});
return res;
}
template <>
Tensor* ExecutionContext::Output<Tensor>(const std::string& name) const {
auto var = OutputVar(name);
return var == nullptr ? nullptr : GetMutableTensorFromVar(var);
return Output<LoDTensor>(name);
}
template <>
......@@ -445,10 +450,14 @@ std::vector<Tensor*> ExecutionContext::MultiOutput<Tensor>(
std::vector<Tensor*> res;
res.reserve(names.size());
std::transform(names.begin(), names.end(), std::back_inserter(res),
[&](const std::string& sub_name) {
[&](const std::string& sub_name) -> Tensor* {
auto var = scope_.FindVar(sub_name);
return var == nullptr ? nullptr
: GetMutableTensorFromVar(var);
if (var == nullptr) return nullptr;
PADDLE_ENFORCE(
var->IsType<LoDTensor>(),
"%s should be LoDTensor, but the received type is %s",
sub_name, var->Type().name());
return var->GetMutable<LoDTensor>();
});
return res;
}
......@@ -708,14 +717,14 @@ void OperatorWithKernel::RunImpl(const Scope& scope,
auto expected_kernel_key =
this->GetExpectedKernelType(ExecutionContext(*this, scope, *dev_ctx));
VLOG(3) << "expected_kernel_key:" << expected_kernel_key;
VLOG(30) << "expected_kernel_key:" << expected_kernel_key;
auto kernel_iter = kernels.find(expected_kernel_key);
#ifdef PADDLE_WITH_MKLDNN
// workaround for missing MKLDNN kernel when FLAGS_use_mkldnn env var is set
if (kernel_iter == kernels.end() &&
expected_kernel_key.library_type_ == LibraryType::kMKLDNN) {
VLOG(3) << "missing MKLDNN kernel: fallbacking to PLAIN one";
VLOG(30) << "missing MKLDNN kernel: fallbacking to PLAIN one";
expected_kernel_key.library_type_ = LibraryType::kPlain;
expected_kernel_key.data_layout_ = DataLayout::kAnyLayout;
kernel_iter = kernels.find(expected_kernel_key);
......@@ -767,12 +776,14 @@ void OperatorWithKernel::TransferInplaceVarsBack(
const Scope& scope, const std::vector<std::string>& inplace_vars,
const Scope& transfer_scope) const {
for (auto& var_name : inplace_vars) {
VLOG(3) << "share inplace var " + var_name + " back to it's original scope";
auto* original_tensor = GetMutableTensorFromVar(scope.FindVar(var_name));
VLOG(30) << "share inplace var " + var_name +
" back to it's original scope";
auto* original_tensor =
GetMutableLoDTensorOrSelectedRowsValueFromVar(scope.FindVar(var_name));
auto* var = transfer_scope.FindVar(var_name);
PADDLE_ENFORCE(var != nullptr, "The var[%s] should not be nullptr",
var_name);
auto* transformed_tensor = GetTensorFromVar(*var);
auto* transformed_tensor = GetLoDTensorOrSelectedRowsValueFromVar(*var);
original_tensor->ShareDataWith(*transformed_tensor);
}
}
......@@ -789,7 +800,7 @@ Scope* OperatorWithKernel::TryTransferData(
continue;
}
auto* tensor_in = GetTensorFromVar(*var);
auto* tensor_in = GetLoDTensorOrSelectedRowsValueFromVar(*var);
if (!tensor_in->IsInitialized()) {
continue;
}
......@@ -807,8 +818,8 @@ Scope* OperatorWithKernel::TryTransferData(
transfered_inplace_vars->emplace_back(var_name);
}
VLOG(3) << "Transform Variable " << var_name << " from "
<< kernel_type_for_var << " to " << expected_kernel_key;
VLOG(30) << "Transform Variable " << var_name << " from "
<< kernel_type_for_var << " to " << expected_kernel_key;
if (new_scope == nullptr) {
new_scope = &scope.NewScope();
......
paddle/fluid/framework/operator.h
浏览文件 @ 2412c27c
......@@ -54,6 +54,9 @@ constexpr char kGradVarSuffix[] = "@GRAD";
/// Variables with this suffix are supposed to be filled up with zeros.
constexpr char kZeroVarSuffix[] = "@ZERO";
/// Variables with this suffix are the new Gradient.
constexpr char kNewGradSuffix[] = "@NEWGRAD@";
// define some kernel priority
/* Define multiple kernel type fallback order*/
extern std::vector<std::tuple<platform::Place, LibraryType>> kKernelPriority;
......@@ -63,7 +66,8 @@ inline std::string GradVarName(const std::string& var_name) {
}
proto::VarType::Type GetDataTypeOfVar(const Variable* var);
const Tensor* GetTensorFromVar(const Variable& var);
const Tensor* GetLoDTensorOrSelectedRowsValueFromVar(const Variable& var);
Tensor* GetMutableLoDTensorOrSelectedRowsValueFromVar(Variable* var);
class OperatorBase;
class ExecutionContext;
......@@ -224,7 +228,7 @@ class ExecutionContext {
std::vector<const T*> res;
res.reserve(names.size());
std::transform(names.begin(), names.end(), std::back_inserter(res),
[&](const std::string& sub_name) {
[&](const std::string& sub_name) -> const T* {
auto var = scope_.FindVar(sub_name);
return var == nullptr ? nullptr : &var->Get<T>();
});
......@@ -237,7 +241,7 @@ class ExecutionContext {
std::vector<T*> res;
res.reserve(names.size());
std::transform(names.begin(), names.end(), std::back_inserter(res),
[&](const std::string& sub_name) {
[&](const std::string& sub_name) -> T* {
auto var = scope_.FindVar(sub_name);
return var == nullptr ? nullptr : var->GetMutable<T>();
});
......
paddle/fluid/framework/parallel_executor.cc
浏览文件 @ 2412c27c
......@@ -199,7 +199,7 @@ void ParallelExecutor::BCastParamsToDevices(
auto &main_tensor = main_var->Get<LoDTensor>();
if (!main_tensor.IsInitialized()) {
VLOG(3) << "one in var not inited, return!";
VLOG(30) << "one in var not inited, return!";
continue;
}
auto &dims = main_tensor.dims();
......
paddle/fluid/framework/scope.cc
浏览文件 @ 2412c27c
......@@ -15,7 +15,9 @@ limitations under the License. */
#include "paddle/fluid/framework/scope.h"
#include <memory> // for unique_ptr
#include <queue>
#include <set>
#include <unordered_set>
#include "glog/logging.h"
#include "paddle/fluid/framework/threadpool.h"
#include "paddle/fluid/string/printf.h"
......@@ -36,6 +38,16 @@ DEFINE_double(
"Memory size threshold (GB) when the garbage collector clear tensors."
"Disabled when this value is less than 0");
// When in inference scenario, the scopes will not be written by two threads in
// a mean time, but a scope may be read by multiple threads concurrently, and
// the mutex will cause serious performance issue.
// So the mutex is disabled when `ON_INFER`.
#ifdef ON_INFER
#define SCOPE_LOCK_GUARD
#else
#define SCOPE_LOCK_GUARD std::lock_guard<std::mutex> lock(mutex_);
#endif
namespace paddle {
namespace framework {
......@@ -49,18 +61,18 @@ int64_t GetEagerDeletionThreshold() {
Scope::~Scope() { DropKids(); }
Scope& Scope::NewScope() const {
std::lock_guard<std::mutex> lock(mutex_);
SCOPE_LOCK_GUARD
kids_.push_back(new Scope(this));
return *kids_.back();
}
Variable* Scope::Var(const std::string& name) {
std::lock_guard<std::mutex> lock(mutex_);
SCOPE_LOCK_GUARD
return VarInternal(name);
}
Variable* Scope::Var(std::string* name) {
std::lock_guard<std::mutex> lock(mutex_);
SCOPE_LOCK_GUARD
auto new_name = string::Sprintf("%p.%d", this, vars_.size());
if (name != nullptr) {
*name = new_name;
......@@ -69,34 +81,34 @@ Variable* Scope::Var(std::string* name) {
}
Variable* Scope::FindVar(const std::string& name) const {
std::lock_guard<std::mutex> lock(mutex_);
SCOPE_LOCK_GUARD
return FindVarInternal(name);
}
Variable* Scope::FindLocalVar(const std::string& name) const {
std::lock_guard<std::mutex> lock(mutex_);
SCOPE_LOCK_GUARD
return FindVarLocally(name);
}
const Scope* Scope::FindScope(const Variable* var) const {
std::lock_guard<std::mutex> lock(mutex_);
SCOPE_LOCK_GUARD
return FindScopeInternal(var);
}
void Scope::DropKids() {
std::lock_guard<std::mutex> lock(mutex_);
SCOPE_LOCK_GUARD
for (Scope* s : kids_) delete s;
kids_.clear();
}
bool Scope::HasKid(const Scope* scope) const {
std::lock_guard<std::mutex> lock(mutex_);
SCOPE_LOCK_GUARD
auto it = std::find(this->kids_.begin(), this->kids_.end(), scope);
return it != this->kids_.end();
}
std::vector<std::string> Scope::LocalVarNames() const {
std::lock_guard<std::mutex> lock(mutex_);
SCOPE_LOCK_GUARD
std::vector<std::string> known_vars;
known_vars.reserve(this->vars_.size());
for (auto& p : vars_) {
......@@ -106,9 +118,10 @@ std::vector<std::string> Scope::LocalVarNames() const {
}
void Scope::DeleteScope(Scope* scope) const {
std::lock_guard<std::mutex> lock(mutex_);
SCOPE_LOCK_GUARD
auto it = std::find(this->kids_.begin(), this->kids_.end(), scope);
PADDLE_ENFORCE(it != this->kids_.end(), "Cannot find %p as kid scope", scope);
PADDLE_ENFORCE(it != this->kids_.end(), "%p Cannot find %p as kid scope",
this, scope);
this->kids_.erase(it);
// When making memory benchmark on Fluid, we have to delete scope sync.
if (FLAGS_benchmark || FLAGS_eager_delete_scope) {
......@@ -119,7 +132,7 @@ void Scope::DeleteScope(Scope* scope) const {
}
void Scope::EraseVars(const std::vector<std::string>& var_names) {
std::lock_guard<std::mutex> lock(mutex_);
SCOPE_LOCK_GUARD
std::set<std::string> var_set(var_names.begin(), var_names.end());
for (auto it = vars_.begin(); it != vars_.end();) {
if (var_set.find(it->first) != var_set.end()) {
......@@ -132,12 +145,12 @@ void Scope::EraseVars(const std::vector<std::string>& var_names) {
void Scope::Rename(const std::string& origin_name,
const std::string& new_name) const {
std::lock_guard<std::mutex> lock(mutex_);
SCOPE_LOCK_GUARD
RenameInternal(origin_name, new_name);
}
std::string Scope::Rename(const std::string& origin_name) const {
std::lock_guard<std::mutex> lock(mutex_);
SCOPE_LOCK_GUARD
auto new_name = string::Sprintf("%p.%d", this, vars_.size());
RenameInternal(origin_name, new_name);
return new_name;
......@@ -149,7 +162,7 @@ Variable* Scope::VarInternal(const std::string& name) {
v = new Variable();
vars_[name].reset(v);
VLOG(3) << "Create variable " << name;
VLOG(30) << "Create variable " << name;
v->name_ = &(vars_.find(name)->first);
return v;
}
......@@ -189,5 +202,46 @@ Variable* Scope::FindVarLocally(const std::string& name) const {
return nullptr;
}
std::string GenScopeTreeDebugInfo(Scope* root) {
std::stringstream os;
if (!root) return "";
// level traversal
std::queue<Scope*> queue;
queue.push(root);
std::vector<Scope*> scopes;
while (!queue.empty()) {
auto* end = queue.back();
Scope* q = nullptr;
while (q != end) {
q = queue.front();
queue.pop();
os << q << " ";
scopes.push_back(q);
for (auto* c : q->kids()) {
queue.push(c);
}
}
// end of a level
os << "\n------------------------------------------\n";
}
os << "\nDetails:\n\n";
for (Scope* q : scopes) {
os << "====\n";
os << q << ":\n";
for (auto& var : q->LocalVarNames()) {
os << " - " << var << "\n";
}
}
return os.str();
}
} // namespace framework
} // namespace paddle
paddle/fluid/framework/scope.h
浏览文件 @ 2412c27c
......@@ -78,11 +78,11 @@ class Scope {
/// Drop all kids scopes belonged to this scope.
void DropKids();
std::list<Scope*>& kids() const { return kids_; }
/// Find if a scope exists in the kid scopes
bool HasKid(const Scope* scope) const;
const std::list<Scope*>& kids() const { return kids_; }
// enumerate all the variables current contains.
std::vector<std::string> LocalVarNames() const;
......@@ -118,12 +118,17 @@ class Scope {
// Scope in `kids_` are owned by this class.
mutable std::list<Scope*> kids_;
Scope const* parent_{nullptr};
const Scope* parent_{nullptr};
DISABLE_COPY_AND_ASSIGN(Scope);
private:
mutable std::mutex mutex_;
};
// Generate some debug string about the inherience structure of scope, quite
// naive.
std::string GenScopeTreeDebugInfo(Scope*);
} // namespace framework
} // namespace paddle
paddle/fluid/framework/selected_rows.cc
浏览文件 @ 2412c27c
......@@ -63,6 +63,26 @@ struct TensorCopyVisitor {
int64_t size_;
};
struct TensorFillVisitor {
TensorFillVisitor(framework::Tensor* dst, int64_t dst_offset, int64_t size,
float value)
: dst_(dst), dst_offset_(dst_offset), size_(size) {}
template <typename T>
void apply() const {
// TODO(qiao): support other place
platform::CPUPlace cpu;
auto* tensor_data = dst_->mutable_data<T>(cpu);
auto* start = tensor_data + dst_offset_;
auto* end = start + size_;
std::fill(start, end, static_cast<T>(0.0));
}
framework::Tensor* dst_;
int64_t dst_offset_;
int64_t size_;
};
void SerializeToStream(std::ostream& os, const SelectedRows& selected_rows,
const platform::DeviceContext& dev_ctx) {
{ // the 1st field, uint32_t version
......@@ -120,7 +140,17 @@ bool SelectedRows::HasKey(int64_t key) const {
: true;
}
int64_t SelectedRows::AutoGrownIndex(int64_t key, bool auto_grown) {
int64_t SelectedRows::AutoGrownIndex(int64_t key, bool auto_grown,
bool is_test) {
if (is_test) {
auto iter = id_to_index_.find(key);
if (iter == id_to_index_.end()) {
return -1;
} else {
return iter->second;
}
}
rwlock_->RDLock();
auto iter = id_to_index_.find(key);
if (iter == id_to_index_.end()) {
......@@ -172,22 +202,30 @@ void SelectedRows::SyncIndex() {
}
void SelectedRows::Get(const framework::Tensor& ids, framework::Tensor* value,
bool auto_grown) {
bool auto_grown, bool is_test) {
PADDLE_ENFORCE(value->IsInitialized(),
"The value tensor should be initialized.");
if (ids.numel() == 0) {
VLOG(3) << "keys is empty, please check data!";
VLOG(30) << "keys is empty, please check data!";
} else {
int64_t value_width = value_->numel() / value_->dims()[0];
PADDLE_ENFORCE_EQ(value_width, value->numel() / value->dims()[0],
"output tensor should have the same shape with table "
"except the dims[0].");
for (int i = 0; i < ids.numel(); ++i) {
int64_t index = AutoGrownIndex(ids.data<int64_t>()[i], auto_grown);
framework::VisitDataType(
framework::ToDataType(value_->type()),
TensorCopyVisitor(value, i * value_width, *value_.get(),
index * value_width, value_width));
auto id = ids.data<int64_t>()[i];
int64_t index = AutoGrownIndex(id, auto_grown, is_test);
if (index < 0) {
VLOG(5) << "id " << id << " not in the table, return 0";
framework::VisitDataType(
framework::ToDataType(value_->type()),
TensorFillVisitor(value, i * value_width, value_width, 0.0));
} else {
framework::VisitDataType(
framework::ToDataType(value_->type()),
TensorCopyVisitor(value, i * value_width, *value_.get(),
index * value_width, value_width));
}
}
}
}
......
paddle/fluid/framework/selected_rows.h
浏览文件 @ 2412c27c
......@@ -105,7 +105,7 @@ class SelectedRows {
* the value
*/
void Get(const framework::Tensor& ids, framework::Tensor* value,
bool auto_grown = false);
bool auto_grown = false, bool is_test = false);
/*
* @brief Get the index of the key from id_to_index_ map. If the key not
......@@ -118,7 +118,7 @@ class SelectedRows {
*
* @return index of the key.
*/
int64_t AutoGrownIndex(int64_t key, bool auto_grown);
int64_t AutoGrownIndex(int64_t key, bool auto_grown, bool is_test = false);
void SyncIndex();
......
paddle/fluid/framework/selected_rows_test.cc
浏览文件 @ 2412c27c
......@@ -84,10 +84,14 @@ TEST(SelectedRows, SparseTable) {
data[i * embedding_width + j] = static_cast<float>(i);
}
}
ASSERT_EQ(table.AutoGrownIndex(10, true), 0);
ASSERT_EQ(table.AutoGrownIndex(8, true), 1);
ASSERT_EQ(table.AutoGrownIndex(8, true), 1);
ASSERT_EQ(table.AutoGrownIndex(6, true), 2);
ASSERT_EQ(table.AutoGrownIndex(10, true, false), 0);
ASSERT_EQ(table.AutoGrownIndex(8, true, false), 1);
ASSERT_EQ(table.AutoGrownIndex(8, true, false), 1);
ASSERT_EQ(table.AutoGrownIndex(6, true, false), 2);
for (int64_t i = 11; i < 20; i++) {
ASSERT_EQ(table.AutoGrownIndex(i, true, true), -1);
ASSERT_TRUE(!table.HasKey(i));
}
ASSERT_TRUE(table.HasKey(10));
ASSERT_TRUE(table.HasKey(8));
ASSERT_TRUE(table.HasKey(6));
......
paddle/fluid/framework/tensor_util.cc
浏览文件 @ 2412c27c
......@@ -22,8 +22,8 @@ namespace framework {
void TensorCopy(const Tensor& src, const platform::Place& dst_place,
const platform::DeviceContext& ctx, Tensor* dst) {
VLOG(3) << "TensorCopy " << src.dims() << " from " << src.place() << " to "
<< dst_place;
VLOG(30) << "TensorCopy " << src.dims() << " from " << src.place() << " to "
<< dst_place;
src.check_memory_size();
dst->Resize(src.dims());
......@@ -37,8 +37,8 @@ void TensorCopy(const Tensor& src, const platform::Place& dst_place,
if (platform::is_cpu_place(src_place) && platform::is_cpu_place(dst_place)) {
if (src_ptr == dst_ptr) {
VLOG(3) << "Skip copy the same data async from " << src_place << " to "
<< dst_place;
VLOG(30) << "Skip copy the same data async from " << src_place << " to "
<< dst_place;
return;
}
memory::Copy(boost::get<platform::CPUPlace>(dst_place), dst_ptr,
......@@ -77,8 +77,8 @@ void TensorCopy(const Tensor& src, const platform::Place& dst_place,
reinterpret_cast<const platform::CUDADeviceContext&>(ctx).stream();
if (platform::is_same_place(src_place, dst_place)) {
if (src_ptr == dst_ptr) {
VLOG(3) << "Skip copy the same data async from " << src_place << " to "
<< dst_place;
VLOG(30) << "Skip copy the same data async from " << src_place << " to "
<< dst_place;
return;
}
memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size,
......@@ -114,8 +114,8 @@ void TensorCopy(const Tensor& src, const platform::Place& dst_place,
void TensorCopySync(const Tensor& src, const platform::Place& dst_place,
Tensor* dst) {
VLOG(3) << "TensorCopySync " << src.dims() << " from " << src.place()
<< " to " << dst_place;
VLOG(30) << "TensorCopySync " << src.dims() << " from " << src.place()
<< " to " << dst_place;
src.check_memory_size();
dst->Resize(src.dims());
dst->set_layout(src.layout());
......@@ -125,8 +125,8 @@ void TensorCopySync(const Tensor& src, const platform::Place& dst_place,
auto size = src.numel() * SizeOfType(src.type());
if (platform::is_cpu_place(src_place) && platform::is_cpu_place(dst_place)) {
if (src_ptr == dst_ptr) {
VLOG(3) << "Skip copy the same data from " << src_place << " to "
<< dst_place;
VLOG(30) << "Skip copy the same data from " << src_place << " to "
<< dst_place;
return;
}
memory::Copy(boost::get<platform::CPUPlace>(dst_place), dst_ptr,
......@@ -146,8 +146,8 @@ void TensorCopySync(const Tensor& src, const platform::Place& dst_place,
} else if (platform::is_gpu_place(src_place) &&
platform::is_gpu_place(dst_place)) {
if (src_ptr == dst_ptr && platform::is_same_place(src_place, dst_place)) {
VLOG(3) << "Skip copy the same data from " << src_place << " to "
<< dst_place;
VLOG(30) << "Skip copy the same data from " << src_place << " to "
<< dst_place;
return;
}
auto src_gpu_place = boost::get<platform::CUDAPlace>(src_place);
......
paddle/fluid/framework/threadpool.cc
浏览文件 @ 2412c27c
......@@ -39,7 +39,7 @@ void ThreadPool::Init() {
int num_threads = std::thread::hardware_concurrency();
if (FLAGS_dist_threadpool_size > 0) {
num_threads = FLAGS_dist_threadpool_size;
VLOG(1) << "set dist_threadpool_size to " << num_threads;
VLOG(10) << "set dist_threadpool_size to " << num_threads;
}
PADDLE_ENFORCE_GT(num_threads, 0);
threadpool_.reset(new ThreadPool(num_threads));
......
paddle/fluid/framework/var_desc.cc
浏览文件 @ 2412c27c
......@@ -61,10 +61,10 @@ size_t VarDesc::GetTensorDescNum() const {
void VarDesc::SetShapes(
const std::vector<std::vector<int64_t>> &multiple_dims) {
if (multiple_dims.size() != GetTensorDescNum()) {
VLOG(3) << "WARNING: The number of given shapes(" << multiple_dims.size()
<< ") doesn't match the existing tensor number("
<< GetTensorDescNum()
<< "). The Reader is going to be reinitialized.";
VLOG(30) << "WARNING: The number of given shapes(" << multiple_dims.size()
<< ") doesn't match the existing tensor number("
<< GetTensorDescNum()
<< "). The Reader is going to be reinitialized.";
SetTensorDescNum(multiple_dims.size());
}
std::vector<proto::VarType::TensorDesc *> tensors = mutable_tensor_descs();
......@@ -94,11 +94,11 @@ void VarDesc::SetDataType(proto::VarType::Type data_type) {
void VarDesc::SetDataTypes(
const std::vector<proto::VarType::Type> &multiple_data_type) {
if (multiple_data_type.size() != GetTensorDescNum()) {
VLOG(3) << "WARNING: The number of given data types("
<< multiple_data_type.size()
<< ") doesn't match the existing tensor number("
<< GetTensorDescNum()
<< "). The Reader is going to be reinitialized.";
VLOG(30) << "WARNING: The number of given data types("
<< multiple_data_type.size()
<< ") doesn't match the existing tensor number("
<< GetTensorDescNum()
<< "). The Reader is going to be reinitialized.";
SetTensorDescNum(multiple_data_type.size());
}
std::vector<proto::VarType::TensorDesc *> tensor_descs =
......@@ -139,11 +139,11 @@ void VarDesc::SetLoDLevel(int32_t lod_level) {
void VarDesc::SetLoDLevels(const std::vector<int32_t> &multiple_lod_level) {
if (multiple_lod_level.size() != GetTensorDescNum()) {
VLOG(3) << "WARNING: The number of given lod_levels("
<< multiple_lod_level.size()
<< ") doesn't match the existing tensor number("
<< GetTensorDescNum()
<< "). The Reader is going to be reinitialized.";
VLOG(30) << "WARNING: The number of given lod_levels("
<< multiple_lod_level.size()
<< ") doesn't match the existing tensor number("
<< GetTensorDescNum()
<< "). The Reader is going to be reinitialized.";
SetTensorDescNum(multiple_lod_level.size());
}
switch (desc_.type().type()) {
......
paddle/fluid/framework/var_type_inference.h
浏览文件 @ 2412c27c
......@@ -13,6 +13,9 @@ See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include <string>
#include "paddle/fluid/framework/block_desc.h"
#include "paddle/fluid/framework/op_desc.h"
#include "paddle/fluid/framework/type_defs.h"
namespace paddle {
......@@ -24,5 +27,27 @@ class VarTypeInference {
virtual void operator()(const OpDesc& op_desc, BlockDesc* block) const = 0;
};
class PassInDtypeAndVarTypeToOutput : public framework::VarTypeInference {
public:
void operator()(const framework::OpDesc& op_desc,
framework::BlockDesc* block) const final {
auto in_out_var_names = this->GetInputOutputWithSameType();
for (auto& i_o_n : in_out_var_names) {
auto& x_name = op_desc.Input(i_o_n.first).at(0);
auto& out_name = op_desc.Output(i_o_n.second).at(0);
auto& x = block->FindRecursiveOrCreateVar(x_name);
auto& out = block->FindRecursiveOrCreateVar(out_name);
out.SetType(x.GetType());
out.SetDataType(x.GetDataType());
}
}
protected:
virtual std::unordered_map<std::string, std::string>
GetInputOutputWithSameType() const = 0;
};
} // namespace framework
} // namespace paddle
paddle/fluid/inference/CMakeLists.txt
浏览文件 @ 2412c27c
......@@ -27,13 +27,9 @@ set(SHARED_INFERENCE_SRCS
io.cc ${CMAKE_CURRENT_SOURCE_DIR}/api/api.cc ${CMAKE_CURRENT_SOURCE_DIR}/api/api_impl.cc
${CMAKE_CURRENT_SOURCE_DIR}/api/analysis_predictor.cc
${CMAKE_CURRENT_SOURCE_DIR}/api/details/zero_copy_tensor.cc)
if (WITH_GPU AND TENSORRT_FOUND)
set(STATIC_INFERENCE_APIS ${STATIC_INFERENCE_APIS} paddle_inference_tensorrt_subgraph_engine)
set(SHARED_INFERENCE_SRCS ${SHARED_INFERENCE_SRCS} ${CMAKE_CURRENT_SOURCE_DIR}/api/api_tensorrt_subgraph_engine.cc)
endif()
# Create static library
cc_library(paddle_fluid DEPS ${fluid_modules} ${STATIC_INFERENCE_APIS} zero_copy_tensor reset_tensor_array)
cc_library(paddle_fluid DEPS ${fluid_modules} ${STATIC_INFERENCE_APIS} zero_copy_tensor reset_tensor_array analysis_config paddle_pass_builder)
if(NOT APPLE)
# TODO(liuyiqu: Temporarily disable the link flag because it is not support on Mac.
......@@ -43,7 +39,7 @@ endif()
# Create shared library
cc_library(paddle_fluid_shared SHARED SRCS ${SHARED_INFERENCE_SRCS}
DEPS ${fluid_modules} paddle_fluid_api reset_tensor_array)
DEPS ${fluid_modules} paddle_fluid_api reset_tensor_array analysis_config paddle_pass_builder)
set_target_properties(paddle_fluid_shared PROPERTIES OUTPUT_NAME paddle_fluid)
if(NOT APPLE)
......
paddle/fluid/inference/analysis/CMakeLists.txt
浏览文件 @ 2412c27c
cc_library(ir_pass_manager SRCS ir_pass_manager.cc DEPS graph pass)
set(analysis_deps
framework_proto proto_desc ir_pass_manager graph pass paddle_fluid_api executor pretty_log)
unset(analysis_deps CACHE)
set(analysis_deps # analysis_deps can be extended accross the project
framework_proto proto_desc graph pass paddle_fluid_api executor pretty_log
ir_pass_manager
CACHE INTERNAL "")
cc_library(analysis SRCS pass_manager.cc node.cc data_flow_graph.cc graph_traits.cc subgraph_splitter.cc
add_subdirectory(ir_passes)
add_subdirectory(passes)
cc_library(ir_pass_manager SRCS ir_pass_manager.cc DEPS graph pass ${INFER_IR_PASSES})
cc_library(argument SRCS argument.cc DEPS scope proto_desc)
cc_library(analysis_pass SRCS analysis_pass.cc DEPS proto_desc)
cc_library(analysis SRCS
analyzer.cc
helper.cc
# passes
analysis_pass.cc
fluid_to_data_flow_graph_pass.cc
data_flow_graph_to_fluid_pass.cc
dfg_graphviz_draw_pass.cc
tensorrt_subgraph_pass.cc
tensorrt_subgraph_node_mark_pass.cc
fluid_to_ir_pass.cc
model_store_pass.cc
DEPS ${analysis_deps})
analysis_pass
DEPS ${analysis_deps}
)
cc_test(test_node SRCS node_tester.cc DEPS analysis)
cc_test(test_dot SRCS dot_tester.cc DEPS analysis)
cc_binary(inference_analyzer SRCS analyzer_main.cc DEPS analysis paddle_fluid)
function(inference_analysis_test TARGET)
if(WITH_TESTING)
......@@ -34,13 +35,3 @@ function(inference_analysis_test TARGET)
endfunction(inference_analysis_test)
inference_analysis_test(test_analyzer SRCS analyzer_tester.cc EXTRA_DEPS paddle_inference_api)
inference_analysis_test(test_data_flow_graph SRCS data_flow_graph_tester.cc)
inference_analysis_test(test_data_flow_graph_to_fluid_pass SRCS data_flow_graph_to_fluid_pass_tester.cc)
inference_analysis_test(test_fluid_to_ir_pass SRCS fluid_to_ir_pass_tester.cc)
inference_analysis_test(test_fluid_to_data_flow_graph_pass SRCS fluid_to_data_flow_graph_pass_tester.cc)
inference_analysis_test(test_subgraph_splitter SRCS subgraph_splitter_tester.cc)
inference_analysis_test(test_dfg_graphviz_draw_pass SRCS dfg_graphviz_draw_pass_tester.cc)
inference_analysis_test(test_tensorrt_subgraph_pass SRCS tensorrt_subgraph_pass_tester.cc)
inference_analysis_test(test_pass_manager SRCS pass_manager_tester.cc)
inference_analysis_test(test_tensorrt_subgraph_node_mark_pass SRCS tensorrt_subgraph_node_mark_pass_tester.cc)
inference_analysis_test(test_model_store_pass SRCS model_store_pass_tester.cc)
paddle/fluid/inference/analysis/analysis_pass.h
浏览文件 @ 2412c27c
......@@ -19,42 +19,36 @@ limitations under the License. */
#include <string>
#include "paddle/fluid/framework/framework.pb.h"
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/inference/analysis/argument.h"
#include "paddle/fluid/inference/analysis/data_flow_graph.h"
#include "paddle/fluid/inference/analysis/helper.h"
#include "paddle/fluid/inference/analysis/node.h"
namespace paddle {
namespace inference {
namespace analysis {
/*
* AnalysisPass is a pass used to control the IR passes.
*/
class AnalysisPass {
public:
AnalysisPass() = default;
virtual ~AnalysisPass() = default;
// Mutable Pass.
virtual bool Initialize(Argument *argument) { return false; }
// Readonly Pass.
virtual bool Initialize(const Argument &argument) { return false; }
// Virtual method overriden by subclasses to do any necessary clean up after
// all passes have run.
virtual bool Finalize() { return false; }
// Create a debugger Pass that draw the DFG by graphviz toolkit.
virtual AnalysisPass *CreateGraphvizDebugerPass() const { return nullptr; }
// Run on a single DataFlowGraph.
virtual void Run(DataFlowGraph *x) = 0;
// Run on a single Graph.
void Run(Argument* argument) { RunImpl(argument); }
// Human-readable short representation.
virtual std::string repr() const = 0;
// Human-readable long description.
virtual std::string description() const { return "No DOC"; }
};
// GraphPass processes on any GraphType.
class DataFlowGraphPass : public AnalysisPass {};
protected:
// User should implement these.
virtual void RunImpl(Argument* argument) = 0;
Argument* argument_{nullptr};
};
} // namespace analysis
} // namespace inference
......
paddle/fluid/inference/analysis/analyzer.cc
浏览文件 @ 2412c27c
......@@ -15,135 +15,23 @@
#include "paddle/fluid/inference/analysis/analyzer.h"
#include <string>
#include <vector>
#include "paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.h"
#include "paddle/fluid/inference/analysis/dfg_graphviz_draw_pass.h"
#include "paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.h"
#include "paddle/fluid/inference/analysis/fluid_to_ir_pass.h"
#include "paddle/fluid/inference/analysis/model_store_pass.h"
#include "paddle/fluid/inference/analysis/pass_manager.h"
#include "paddle/fluid/inference/analysis/tensorrt_subgraph_node_mark_pass.h"
#include "paddle/fluid/inference/analysis/tensorrt_subgraph_pass.h"
DEFINE_bool(IA_enable_tensorrt_subgraph_engine, false,
"Enable subgraph to TensorRT engine for acceleration");
DEFINE_bool(IA_enable_ir, false, "Turn on IR support");
DEFINE_string(IA_graphviz_log_root, "./",
"Graphviz debuger for data flow graphs.");
DEFINE_string(IA_output_storage_path, "", "optimized model output path");
#include "paddle/fluid/inference/analysis/passes/ir_analysis_compose_pass.h"
#include "paddle/fluid/inference/analysis/passes/passes.h"
namespace paddle {
namespace inference {
namespace analysis {
class DfgPassManagerImpl final : public DfgPassManager {
public:
DfgPassManagerImpl() {
// TODO(Superjomn) set the key with pass reprs.
if (!FLAGS_IA_enable_ir) {
AddPass("fluid-to-data-flow-graph", new FluidToDataFlowGraphPass);
} else {
AddPass("fluid-to-ir-pass", new FluidToIrPass);
}
TryAddTensorRtPass();
AddPass("data-flow-graph-to-fluid", new DataFlowGraphToFluidPass);
if (!FLAGS_IA_output_storage_path.empty()) {
AddPass("model-store-pass", new ModelStorePass);
}
}
Analyzer::Analyzer() {}
std::string repr() const override { return "dfg-pass-manager"; }
std::string description() const override { return "DFG pass manager."; }
void Analyzer::Run(Argument *argument) { RunIrAnalysis(argument); }
private:
void AddPass(const std::string& name, AnalysisPass* pass) {
VLOG(3) << "Adding pass " << name;
Register(name, pass);
AddGraphvizDebugerPass(pass);
}
void Analyzer::RunIrAnalysis(Argument *argument) {
std::vector<std::string> passes({"ir_analysis_compose_pass"});
void TryAddTensorRtPass() {
if (FLAGS_IA_enable_tensorrt_subgraph_engine) {
auto trt_teller = [&](const Node* node) {
std::unordered_set<std::string> teller_set(
{"mul", "conv2d", "pool2d", "relu", "softmax", "sigmoid",
"depthwise_conv2d", "batch_norm", "concat", "tanh", "pad",
"elementwise_add", "dropout"});
if (!node->IsFunction()) return false;
const auto* func = static_cast<const Function*>(node);
if (teller_set.count(func->func_type())) {
return true;
} else {
return false;
}
};
AddPass("tensorrt-subgraph-marker",
new TensorRTSubgraphNodeMarkPass(trt_teller));
AddPass("tensorrt-subgraph", new TensorRTSubGraphPass(trt_teller));
}
}
// Add the graphviz debuger pass if the parent pass has one.
void AddGraphvizDebugerPass(AnalysisPass* pass) {
auto* debuger_pass = pass->CreateGraphvizDebugerPass();
if (debuger_pass) {
Register(debuger_pass->repr(), debuger_pass);
}
for (auto &pass : passes) {
PassRegistry::Global().Retreive(pass)->Run(argument);
}
};
Analyzer::Analyzer() { Register("manager1", new DfgPassManagerImpl); }
void Analyzer::Run(Argument* argument) {
std::vector<std::string> passes;
#ifdef PADDLE_WITH_MKLDNN
if (use_mkldnn_) {
VLOG(3) << "Adding MKL-DNN placement pass";
passes.push_back("mkldnn_placement_pass");
}
#endif
// infer_clean_graph_pass should be the first default pass
// after mkldnn_placement_pass.
passes.push_back("infer_clean_graph_pass");
for (auto& pass : ir_passes_) {
if (!disabled_ir_passes_.count(pass)) {
passes.push_back(pass);
passes.push_back("graph_viz_pass"); // add graphviz for debug.
}
}
passes.push_back("graph_viz_pass");
argument->Set(kFluidToIrPassesAttr, new std::vector<std::string>(passes));
for (auto& x : data_) {
PADDLE_ENFORCE(x->Initialize(argument));
x->RunAll();
PADDLE_ENFORCE(x->Finalize());
}
}
Analyzer& Analyzer::IncludeAllIrPasses() {
ir_passes_ = all_ir_passes_;
return *this;
}
Analyzer& Analyzer::DisableIrPasses(const std::vector<std::string>& passes) {
disabled_ir_passes_.insert(passes.begin(), passes.end());
return *this;
}
Analyzer& Analyzer::IncludeIrPasses(const std::vector<std::string>& passes) {
ir_passes_ = passes;
return *this;
}
Analyzer& Analyzer::SetUseMkldnn(bool use_mkldnn) {
use_mkldnn_ = use_mkldnn;
return *this;
}
} // namespace analysis
......
paddle/fluid/inference/analysis/analyzer.h
浏览文件 @ 2412c27c
......@@ -40,59 +40,21 @@ limitations under the License. */
#include <vector>
#include "paddle/fluid/inference/analysis/analysis_pass.h"
#include "paddle/fluid/inference/analysis/flags.h"
#include "paddle/fluid/inference/analysis/pass_manager.h"
namespace paddle {
namespace inference {
namespace analysis {
class Analyzer : public OrderedRegistry<PassManager> {
class Analyzer final {
public:
// Register all the pass-managers.
Analyzer();
void Run(Argument* argument);
Analyzer& DisableIrPasses(const std::vector<std::string>& passes);
Analyzer& IncludeIrPasses(const std::vector<std::string>& passes);
Analyzer& IncludeAllIrPasses();
Analyzer& SetUseMkldnn(bool use_mkldnn);
DISABLE_COPY_AND_ASSIGN(Analyzer);
private:
// All avaiable IR passes.
// The bigger fuse comes first, so that the small operators prefer to be
// merged in a larger fuse op. The small fusion will not break the pattern of
// larger fusion.
const std::vector<std::string> all_ir_passes_{{
// Manual update the passes here.
#ifdef PADDLE_WITH_MKLDNN
// This pass should run before any other convolution fuse.
"depthwise_conv_mkldnn_pass", //
#endif
"attention_lstm_fuse_pass", //
"seqconv_eltadd_relu_fuse_pass", //
"embedding_fc_lstm_fuse_pass", //
"fc_lstm_fuse_pass", //
"mul_lstm_fuse_pass", //
"fc_gru_fuse_pass", //
"mul_gru_fuse_pass", //
"seq_concat_fc_fuse_pass", //
"fc_fuse_pass", //
"conv_bn_fuse_pass", //
"conv_eltwiseadd_bn_fuse_pass", //
#ifdef PADDLE_WITH_MKLDNN
"conv_bias_mkldnn_fuse_pass", //
"conv_relu_mkldnn_fuse_pass", //
"conv_elementwise_add_mkldnn_fuse_pass", //
#endif
}};
std::unordered_set<std::string> disabled_ir_passes_;
// Ir passes to run
std::vector<std::string> ir_passes_;
bool use_mkldnn_;
protected:
void RunIrAnalysis(Argument* argument);
};
} // namespace analysis
......
paddle/fluid/inference/analysis/analyzer_tester.cc
浏览文件 @ 2412c27c
......@@ -27,21 +27,21 @@ namespace analysis {
using namespace framework; // NOLINT
TEST(Analyzer, analysis_without_tensorrt) {
FLAGS_IA_enable_tensorrt_subgraph_engine = false;
Argument argument;
argument.fluid_model_dir.reset(new std::string(FLAGS_inference_model_dir));
argument.SetModelDir(FLAGS_inference_model_dir);
argument.SetIrAnalysisPasses({"infer_clean_graph_pass"});
Analyzer analyser;
analyser.Run(&argument);
}
TEST(Analyzer, analysis_with_tensorrt) {
FLAGS_IA_enable_tensorrt_subgraph_engine = true;
Argument argument;
argument.Set<int>("minimum_subgraph_size", new int(0));
argument.Set<int>("max_batch_size", new int(3));
argument.Set<int>("workspace_size", new int(1 << 20));
argument.Set<std::string>("precision_mode", new std::string("FP32"));
argument.fluid_model_dir.reset(new std::string(FLAGS_inference_model_dir));
argument.SetTensorRtMaxBatchSize(3);
argument.SetTensorRtWorkspaceSize(1 << 20);
argument.SetModelDir(FLAGS_inference_model_dir);
argument.SetIrAnalysisPasses({"infer_clean_graph_pass"});
Analyzer analyser;
analyser.Run(&argument);
}
......
paddle/fluid/inference/analysis/argument.h
浏览文件 @ 2412c27c
......@@ -24,13 +24,16 @@
#pragma once
#include <string>
#include <vector>
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/program_desc.h"
#include "paddle/fluid/inference/analysis/data_flow_graph.h"
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/platform/variant.h"
namespace paddle {
namespace inference {
namespace analysis {
using framework::ir::Graph;
/*
* The argument definition of both Pass and PassManagers.
......@@ -39,75 +42,99 @@ namespace analysis {
*/
struct Argument {
Argument() = default;
explicit Argument(const std::string& fluid_model_dir)
: fluid_model_dir(new std::string(fluid_model_dir)) {}
// The directory of the trained model.
std::unique_ptr<std::string> fluid_model_dir;
// The path of `__model__` and `param`, this is used when the file name of
// model and param is changed.
std::unique_ptr<std::string> fluid_model_program_path;
std::unique_ptr<std::string> fluid_model_param_path;
// The graph that process by the Passes or PassManagers.
std::unique_ptr<DataFlowGraph> main_dfg;
// The original program desc.
std::unique_ptr<framework::proto::ProgramDesc> origin_program_desc;
// The processed program desc.
std::unique_ptr<framework::proto::ProgramDesc> transformed_program_desc;
// The output storage path of ModelStorePass.
std::unique_ptr<std::string> model_output_store_path;
// Support for any other attributes.
template <typename T>
void Set(const std::string& key, T* data) {
PADDLE_ENFORCE_NOT_NULL(data);
PADDLE_ENFORCE(!attrs_.count(key), "Duplicate set Argument's attr [%s]",
key);
attrs_[key] = data;
attr_deleters_[key] = [data, key]() {
VLOG(3) << "xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx";
VLOG(3) << "argument delete attr: " << key;
delete data;
};
}
bool Has(const std::string& name) const { return attrs_.count(name); }
template <typename T>
T* Release(const std::string& key) {
PADDLE_ENFORCE(attrs_.count(key));
auto* res = boost::any_cast<T*>(attrs_.at(key));
attrs_.erase(key);
attr_deleters_.erase(key);
return res;
}
template <typename T>
T& Get(const std::string& key) {
PADDLE_ENFORCE(Has(key));
return *boost::any_cast<T*>(attrs_.at(key));
}
~Argument() {
for (auto& item : attr_deleters_) {
item.second();
}
}
explicit Argument(const std::string& model_dir) { SetModelDir(model_dir); }
using unique_ptr_t = std::unique_ptr<void, std::function<void(void*)>>;
using fusion_statis_t = std::unordered_map<std::string, int>;
bool Has(const std::string& key) const { return valid_fields_.count(key); }
#define DECL_ARGUMENT_FIELD(field__, Field, type__) \
public: \
type__& field__() { \
PADDLE_ENFORCE(Has(#field__)); \
return field__##_; \
} \
void Set##Field(const type__& x) { \
field__##_ = x; \
valid_fields_.insert(#field__); \
} \
DECL_ARGUMENT_FIELD_VALID(field__); \
type__* field__##_ptr() { return &field__##_; } \
\
private: \
type__ field__##_;
#define DECL_ARGUMENT_FIELD_VALID(field__) \
bool field__##_valid() { return Has(#field__); }
#define DECL_ARGUMENT_UNIQUE_FIELD(field__, Field, type__) \
public: \
type__& field__() { \
PADDLE_ENFORCE_NOT_NULL(field__##_); \
PADDLE_ENFORCE(Has(#field__)); \
return *static_cast<type__*>(field__##_.get()); \
} \
void Set##Field(type__* x) { \
field__##_ = \
unique_ptr_t(x, [](void* x) { delete static_cast<type__*>(x); }); \
valid_fields_.insert(#field__); \
} \
void Set##Field##NotOwned(type__* x) { \
valid_fields_.insert(#field__); \
field__##_ = unique_ptr_t(x, [](void* x) {}); \
} \
DECL_ARGUMENT_FIELD_VALID(field__); \
type__* field__##_ptr() { \
PADDLE_ENFORCE(Has(#field__)); \
return static_cast<type__*>(field__##_.get()); \
} \
type__* Release##Field() { \
PADDLE_ENFORCE(Has(#field__)); \
valid_fields_.erase(#field__); \
return static_cast<type__*>(field__##_.release()); \
} \
\
private: \
unique_ptr_t field__##_;
// Model path
DECL_ARGUMENT_FIELD(model_dir, ModelDir, std::string);
// Model specified with program and parameters files.
DECL_ARGUMENT_FIELD(model_program_path, ModelProgramPath, std::string);
DECL_ARGUMENT_FIELD(model_params_path, ModelParamsPath, std::string);
// The overall graph to work on.
DECL_ARGUMENT_UNIQUE_FIELD(main_graph, MainGraph, framework::ir::Graph);
// The overall Scope to work on.
DECL_ARGUMENT_UNIQUE_FIELD(scope, Scope, framework::Scope);
DECL_ARGUMENT_UNIQUE_FIELD(main_program, MainProgram, framework::ProgramDesc);
// The ir passes to perform in analysis phase.
DECL_ARGUMENT_FIELD(ir_analysis_passes, IrAnalysisPasses,
std::vector<std::string>);
DECL_ARGUMENT_FIELD(use_gpu, UseGPU, bool);
DECL_ARGUMENT_FIELD(use_tensorrt, UseTensorRT, bool);
DECL_ARGUMENT_FIELD(tensorrt_node_teller, TensorRtNodeTeller,
std::function<bool(const framework::ir::Node*)>);
DECL_ARGUMENT_FIELD(tensorrt_max_batch_size, TensorRtMaxBatchSize, int);
DECL_ARGUMENT_FIELD(tensorrt_workspace_size, TensorRtWorkspaceSize, int);
// The program transformed by IR analysis phase.
DECL_ARGUMENT_UNIQUE_FIELD(ir_analyzed_program, IrAnalyzedProgram,
framework::proto::ProgramDesc);
DECL_ARGUMENT_FIELD(fusion_statis, FusionStatis, fusion_statis_t);
private:
std::unordered_map<std::string, boost::any> attrs_;
std::unordered_map<std::string, std::function<void()>> attr_deleters_;
std::unordered_set<std::string> valid_fields_;
};
#define UNLIKELY(condition) __builtin_expect(static_cast<bool>(condition), 0)
#define ANALYSIS_ARGUMENT_CHECK_FIELD(field__) \
if (UNLIKELY(!(field__))) { \
LOG(ERROR) << "field " << #field__ << " should be set."; \
return false; \
}
#define ARGUMENT_CHECK_FIELD(argument__, fieldname__) \
PADDLE_ENFORCE(argument__->Has(#fieldname__), \
"the argument field [%s] should be set", #fieldname__);
} // namespace analysis
} // namespace inference
......
paddle/fluid/inference/analysis/data_flow_graph.h 已删除 100644 → 0
浏览文件 @ b5f617fa
/* 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. */
/*
* Data flow graph is an pass that build the basic graph. It contains a graph
* and the iterators that enable the iteration over the graph.
*/
#pragma once
#include <deque>
#include <stack>
#include <string>
#include <unordered_set>
#include <utility>
#include <vector>
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/inference/analysis/graph_traits.h"
#include "paddle/fluid/inference/analysis/node.h"
#include "paddle/fluid/platform/enforce.h"
namespace paddle {
namespace inference {
namespace analysis {
/*
* DataFlowGraph - A container of Value and Function Nodes.
*
* This is the base graph for any other type of graphs, such as SSA or CFG.
*/
struct DataFlowGraph {
NodeMap nodes;
// inputs and outputs are deduced from the graph.
// Used to interact with IR.
const framework::ir::Graph *ir_graph{nullptr};
// Extract inputs and outputs of the graph.
void Build();
void Build(const framework::proto::ProgramDesc &prog);
// Build a graph from ir::Graph.
void Build(const framework::ir::Graph &graph);
// Get an attribute.
AnyAttr &Attr(const std::string &key) { return attrs_[key]; }
// Output a DOT graph file for debug.
std::string DotString() const;
std::string HumanReadableInfo(bool show_values = true,
bool show_functions = true) const;
const std::vector<Node *> &inputs() const {
PADDLE_ENFORCE(!inputs_.empty(),
"No inputs are deduced, need to Build() first.");
return inputs_;
}
const std::vector<Node *> &outputs() const {
PADDLE_ENFORCE(!outputs_.empty(),
"No outputs are deduced, need to Build() first.");
return outputs_;
}
private:
mutable std::vector<Node *> inputs_;
mutable std::vector<Node *> outputs_;
std::unordered_map<std::string, AnyAttr> attrs_;
// Remove duplicate edges and so on.
void Clean();
};
/*
* An graph trait help to traverse the graph using BFS.
* The BFS start from a graph's inputs, the graph should be fully-connected, so
* that the iterator can reach the end.
*/
template <>
struct GraphTraits<DataFlowGraph> {
// BFS iterator on nodes.
struct NodesBFSIterator
: public std::iterator<std::forward_iterator_tag, Node *> {
NodesBFSIterator() = default;
explicit NodesBFSIterator(const std::vector<Node *> &source);
NodesBFSIterator(NodesBFSIterator &&other) noexcept;
// NOTE Heavy to use.
NodesBFSIterator(const NodesBFSIterator &other);
Node &operator*();
NodesBFSIterator &operator++();
Node *operator->();
// TODO(Superjomn) current implementation just compare the first
// element, need to compare the graph and all the elements in the queue and
// set.
NodesBFSIterator &operator=(const NodesBFSIterator &other);
bool operator==(const NodesBFSIterator &other);
bool operator!=(const NodesBFSIterator &other) { return !(*this == other); }
private:
std::deque<Node *> queue_;
std::unordered_set<Node *> visited_;
};
// DFS iterator on nodes.
struct NodesDFSIterator
: public std::iterator<std::forward_iterator_tag, Node *> {
NodesDFSIterator() = default;
NodesDFSIterator(const std::vector<Node *> &source);
NodesDFSIterator(NodesDFSIterator &&other) noexcept;
NodesDFSIterator(const NodesDFSIterator &other);
Node &operator*();
NodesDFSIterator &operator++();
// TODO(Superjomn) current implementation just compare the first
// element, need to compare the graph and all the elements in the queue and
// set.
NodesDFSIterator &operator=(const NodesDFSIterator &other);
bool operator==(const NodesDFSIterator &other);
bool operator!=(const NodesDFSIterator &other) { return !(*this == other); }
Node *operator->();
private:
std::stack<Node *> stack_;
std::unordered_set<Node *> visited_;
};
// Topological sorting iterator on nodes.
struct NodesTSIterator
: public std::iterator<std::forward_iterator_tag, Node *> {
NodesTSIterator() = default;
NodesTSIterator(const std::vector<Node *> &source);
NodesTSIterator(NodesTSIterator &&other)
: sorted_(std::move(other.sorted_)), cursor_(other.cursor_) {
other.cursor_ = 0;
}
NodesTSIterator(const NodesTSIterator &other);
Node &operator*();
NodesTSIterator &operator++();
// TODO(Superjomn) current implementation just compare the first
// element, need to compare the graph and all the elements in the queue and
// set.
NodesTSIterator &operator=(const NodesTSIterator &other);
bool operator==(const NodesTSIterator &other);
bool operator!=(const NodesTSIterator &other) { return !(*this == other); }
Node *operator->();
private:
std::vector<Node *> sorted_;
size_t cursor_{0};
};
explicit GraphTraits(const DataFlowGraph &graph) : graph_(graph) {}
// default use BFS to visit the nodes.
iterator_range<NodesBFSIterator> nodes() {
return iterator_range<NodesBFSIterator>(nodes_bfs_begin(), nodes_bfs_end());
}
iterator_range<NodesBFSIterator> nodes_in_BFS() {
return iterator_range<NodesBFSIterator>(nodes_bfs_begin(), nodes_bfs_end());
}
iterator_range<NodesDFSIterator> nodes_in_DFS() {
return iterator_range<NodesDFSIterator>(nodes_dfs_begin(), nodes_dfs_end());
}
iterator_range<NodesTSIterator> nodes_in_TS() {
return iterator_range<NodesTSIterator>(nodes_ts_begin(), nodes_ts_end());
}
private:
NodesBFSIterator nodes_bfs_begin() {
return NodesBFSIterator(graph_.inputs());
}
NodesBFSIterator nodes_bfs_end() { return NodesBFSIterator(); }
NodesDFSIterator nodes_dfs_begin() {
return NodesDFSIterator(graph_.inputs());
}
NodesDFSIterator nodes_dfs_end() { return NodesDFSIterator(); }
NodesTSIterator nodes_ts_begin() { return NodesTSIterator(graph_.inputs()); }
NodesTSIterator nodes_ts_end() { return NodesTSIterator(); }
private:
const DataFlowGraph &graph_;
};
// Extract the inputs and outputs of a graph. The inputs and outputs of a
// sub-graph is the inputs nodes and output nodes that doesn't inside the
// sub-graph.
std::pair<std::vector<Node *>, std::vector<Node *>>
ExtractInputAndOutputOfSubGraph(std::vector<Node *> &graph); // NOLINT
void FilterRedundantOutputOfSubGraph(DataFlowGraph *graph);
} // namespace analysis
} // namespace inference
} // namespace paddle
paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.h 已删除 100644 → 0
浏览文件 @ b5f617fa
/* 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. */
/*
* This file implements the transformation from fluid ProgramDesc to data flow
* graph.
*/
#pragma once
#include <string>
#include "paddle/fluid/framework/program_desc.h"
#include "paddle/fluid/inference/analysis/analysis_pass.h"
#include "paddle/fluid/inference/analysis/data_flow_graph.h"
namespace paddle {
namespace inference {
namespace analysis {
class DataFlowGraphToFluidPass final : public DataFlowGraphPass {
public:
DataFlowGraphToFluidPass() = default;
bool Initialize(Argument *argument) override;
bool Finalize() override;
void Run(DataFlowGraph *graph) override;
std::string repr() const override { return "DFG to fluid"; }
std::string description() const override {
return "Transform a DFG to a Fluid ProgramDesc";
}
AnalysisPass *CreateGraphvizDebugerPass() const override;
protected:
// Add a Fluid Op into the ProgramDesc.
void AddFluidOp(Node *node);
// Add a EngineOp into the ProgramDesc.
void AddEngineOp(Node *node);
private:
framework::proto::ProgramDesc *desc_;
Argument *argument_;
};
} // namespace analysis
} // namespace inference
} // namespace paddle
paddle/fluid/inference/analysis/node.cc 已删除 100644 → 0
浏览文件 @ b5f617fa
此差异已折叠。
paddle/fluid/inference/analysis/node.h 已删除 100644 → 0
浏览文件 @ b5f617fa
此差异已折叠。
paddle/fluid/inference/analysis/passes/CMakeLists.txt 0 → 100644
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/inference/api/analysis_config.cc 0 → 100644
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/inference/api/api.cc
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/inference/api/helper.h
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/inference/api/paddle_analysis_config.h 0 → 100644
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/inference/api/paddle_api.h 0 → 100644
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/inference/api/paddle_pass_builder.h 0 → 100644
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/inference/io.cc
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/memory/malloc.cc
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/operators/adagrad_op.h
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/operators/adam_op.h
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/operators/auc_op.cc
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/operators/bilinear_interp_op.h 已删除 100644 → 0
浏览文件 @ b5f617fa
此差异已折叠。
paddle/fluid/operators/concat_op.cc
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/operators/conv_cudnn_op_cache.h 0 → 100644
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/operators/conv_op.cc
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/operators/detection/density_prior_box_op.h 0 → 100644
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/operators/extract_rows_op.cc 已删除 100644 → 0
浏览文件 @ b5f617fa
此差异已折叠。
paddle/fluid/operators/feed_op.cc
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/operators/fetch_op.cc
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/operators/gather.cu.h
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/operators/gather.h
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/operators/gather_op.cc
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/operators/interpolate_op.h 0 → 100644
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/operators/load_op.cc
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/operators/lrn_op.cc
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/operators/lrn_op.h
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/operators/math/blas.h
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/operators/mean_op.cc
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/operators/mul_op.cc
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/operators/nce_op.cc
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/operators/pool_op.cc
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/operators/recv_op.cc
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/operators/reduce_max_op.part.cu 0 → 100644
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/operators/reduce_mean_op.part.cu 0 → 100644
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/operators/reduce_min_op.part.cu 0 → 100644
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/operators/reduce_prod_op.part.cu 0 → 100644
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/operators/reduce_sum_op.part.cu 0 → 100644
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/operators/save_op.cc
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/operators/scale_op.h
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/operators/scatter.cu.h
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/operators/scatter.h
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/operators/send_op.cc
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/operators/sgd_op.h
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/operators/similarity_focus_op.cc 0 → 100644
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/operators/similarity_focus_op.h 0 → 100644
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/operators/space_to_depth_op.cc 0 → 100644
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/operators/space_to_depth_op.cu 0 → 100644
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/operators/space_to_depth_op.h 0 → 100644
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/operators/stack_op.cc
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/operators/stack_op.cu
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/operators/sum_op.cc
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/operators/tensor_array_to_tensor_op.cc 0 → 100644
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/operators/while_op.cc
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/platform/gpu_info.cc
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/platform/init.cc
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/platform/nccl_helper.h
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/pybind/const_value.cc
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/pybind/protobuf.cc
浏览文件 @ 2412c27c
此差异已折叠。
paddle/fluid/string/printf.h
浏览文件 @ 2412c27c
此差异已折叠。
paddle/scripts/paddle_build.sh
浏览文件 @ 2412c27c
此差异已折叠。
paddle/testing/TestUtil.cpp
浏览文件 @ 2412c27c
此差异已折叠。
python/paddle/fluid/__init__.py
浏览文件 @ 2412c27c
此差异已折叠。
python/paddle/fluid/layers/io.py
浏览文件 @ 2412c27c
此差异已折叠。
python/paddle/fluid/layers/nn.py
浏览文件 @ 2412c27c
此差异已折叠。
python/paddle/fluid/op.py
浏览文件 @ 2412c27c
此差异已折叠。
python/paddle/fluid/optimizer.py
浏览文件 @ 2412c27c
此差异已折叠。
python/paddle/fluid/regularizer.py
浏览文件 @ 2412c27c
此差异已折叠。
python/setup.py.in
浏览文件 @ 2412c27c
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