Skip to content

P
PaddleDetection

PaddlePaddle / PaddleDetection
12 个月 前同步成功

通知 692
Star 11112
Fork 2696
P
PaddleDetection

体验新版 GitCode，发现更多精彩内容 >>

提交 ce994190 编写于 作者: Q Qiao Longfei
浏览文件
操作

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into optimize-pyreader 

test=develop
上级 cc7f5514 1001f8e1
展开全部 隐藏空白更改
内联 并排
Showing with 20727 addition and 4952 deletion
.gitignore
浏览文件 @ ce994190
......@@ -28,3 +28,4 @@ third_party/
build_*
# clion workspace.
cmake-build-*
model_test
CMakeLists.txt
浏览文件 @ ce994190
......@@ -62,13 +62,12 @@ option(WITH_DISTRIBUTE "Compile with distributed support" OFF)
option(USE_EIGEN_FOR_BLAS "Use matrix multiplication in Eigen" OFF)
option(EIGEN_USE_THREADS "Compile with multi-threaded Eigen" OFF)
option(WITH_ARM_FP16 "Use half precision support on armv8.2-a cpu" OFF)
option(WITH_FAST_BUNDLE_TEST "Bundle tests that can be run in a single process together to reduce launch overhead" OFF)
option(WITH_CONTRIB "Compile the third-party contributation" OFF)
option(REPLACE_ENFORCE_GLOG "Replace PADDLE_ENFORCE with glog/CHECK for better debug." OFF)
option(WITH_ANAKIN "Compile with Anakin library" OFF)
option(WITH_GRPC "Use grpc as the default rpc framework" ${WITH_DISTRIBUTE})
option(WITH_BRPC_RDMA "Use brpc rdma as the rpc protocal" OFF)
option(WITH_INFERENCE "Compile fluid inference library" ON)
option(ON_INFER "Turn on inference optimization." OFF)
option(WITH_INFERENCE_API_TEST "Test fluid inference high-level api interface" OFF)
option(WITH_SYSTEM_BLAS "Use system blas library" OFF)
option(PY_VERSION "Compile PaddlePaddle with python3 support" ${PY_VERSION})
......@@ -127,6 +126,9 @@ set(THIRD_PARTY_PATH "${CMAKE_BINARY_DIR}/third_party" CACHE STRING
set(FLUID_INSTALL_DIR "${CMAKE_BINARY_DIR}/fluid_install_dir" CACHE STRING
"A path setting fluid shared and static libraries")
set(FLUID_INFERENCE_INSTALL_DIR "${CMAKE_BINARY_DIR}/fluid_inference_install_dir" CACHE STRING
"A path setting fluid inference shared and static libraries")
if (WITH_C_API AND WITH_PYTHON)
message(WARNING "It is suggest not embedded a python interpreter in Paddle "
"when using C-API. It will give an unpredictable behavior when using a "
......@@ -176,6 +178,7 @@ include(external/eigen) # download eigen3
include(external/pybind11) # download pybind11
include(external/cares)
include(external/cub)
include(external/xxhash) # download xxhash
if (NOT WIN32)
# there is no official support of snappystream, warpctc, nccl, cupti in windows
......@@ -298,3 +301,11 @@ if(WITH_DOC)
find_python_module(recommonmark REQUIRED)
add_subdirectory(doc)
endif()
if (ON_INFER)
message(STATUS "On inference mode, will take place some specific optimization.")
add_definitions(-DPADDLE_ON_INFERENCE)
else()
#TODO(luotao), combine this warning with `make inference_lib_dist` command.
message(WARNING "On inference mode, will take place some specific optimization. Turn on the ON_INFER flag when building inference_lib only.")
endif()
Dockerfile
浏览文件 @ ce994190
......@@ -75,14 +75,14 @@ RUN pip3 install -U wheel && \
pip3 install -U docopt PyYAML sphinx==1.5.6 && \
pip3 install sphinx-rtd-theme==0.1.9 recommonmark && \
easy_install -U pip && \
pip install -U wheel && \
pip install -U pip setuptools wheel && \
pip install -U docopt PyYAML sphinx==1.5.6 && \
pip install sphinx-rtd-theme==0.1.9 recommonmark
RUN pip3 install pre-commit 'ipython==5.3.0' && \
RUN pip3 install 'pre-commit==1.10.4' 'ipython==5.3.0' && \
pip3 install 'ipykernel==4.6.0' 'jupyter==1.0.0' && \
pip3 install opencv-python && \
pip install pre-commit 'ipython==5.3.0' && \
pip install 'pre-commit==1.10.4' 'ipython==5.3.0' && \
pip install 'ipykernel==4.6.0' 'jupyter==1.0.0' && \
pip install opencv-python
......
README.md
浏览文件 @ ce994190
......@@ -2,8 +2,8 @@
[![Build Status](https://travis-ci.org/PaddlePaddle/Paddle.svg?branch=develop)](https://travis-ci.org/PaddlePaddle/Paddle)
[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/index_en.html)
[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://www.paddlepaddle.org/docs/develop/documentation/zh/getstarted/index_cn.html)
[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://paddlepaddle.org/documentation/docs/en/1.1/getstarted/index_en.html)
[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://paddlepaddle.org/documentation/docs/zh/1.1/beginners_guide/index.html)
[![Release](https://img.shields.io/github/release/PaddlePaddle/Paddle.svg)](https://github.com/PaddlePaddle/Paddle/releases)
[![License](https://img.shields.io/badge/license-Apache%202-blue.svg)](LICENSE)
......@@ -19,7 +19,7 @@ Our vision is to enable deep learning for everyone via PaddlePaddle.
Please refer to our [release announcement](https://github.com/PaddlePaddle/Paddle/releases) to track the latest feature of PaddlePaddle.
### Latest PaddlePaddle Release: [Fluid 0.15.0](https://github.com/PaddlePaddle/Paddle/tree/v0.15.0)
### Latest PaddlePaddle Release: [Fluid 1.1.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.1)
### Install Latest Stable Release:
```
# Linux CPU
......@@ -27,9 +27,9 @@ pip install paddlepaddle
# Linux GPU cuda9cudnn7
pip install paddlepaddle-gpu
# Linux GPU cuda8cudnn7
pip install paddlepaddle-gpu==0.15.0.post87
pip install paddlepaddle-gpu==1.1.0.post87
# Linux GPU cuda8cudnn5
pip install paddlepaddle-gpu==0.15.0.post85
pip install paddlepaddle-gpu==1.1.0.post85
# For installation on other platform, refer to http://paddlepaddle.org/
```
......@@ -76,26 +76,26 @@ pip install paddlepaddle-gpu==0.15.0.post85
## Installation
It is recommended to read [this doc](http://paddlepaddle.org/documentation/docs/zh/0.15.0/new_docs/beginners_guide/install/install_doc.html) on our website.
It is recommended to read [this doc](http://paddlepaddle.org/documentation/docs/zh/1.1/beginners_guide/index.html) on our website.
## Documentation
We provide [English](http://paddlepaddle.org/documentation/docs/en/0.15.0/getstarted/index_en.html) and
[Chinese](http://paddlepaddle.org/documentation/docs/zh/0.15.0/new_docs/beginners_guide/index.html) documentation.
We provide [English](http://paddlepaddle.org/documentation/docs/en/1.1/getstarted/index_en.html) and
[Chinese](http://paddlepaddle.org/documentation/docs/zh/1.1/beginners_guide/index.html) documentation.
- [Deep Learning 101](https://github.com/PaddlePaddle/book)
You might want to start from this online interactive book that can run in a Jupyter Notebook.
- [Distributed Training](http://paddlepaddle.org/documentation/docs/zh/0.15.0/new_docs/user_guides/howto/training/cluster_howto.html)
- [Distributed Training](http://paddlepaddle.org/documentation/docs/zh/1.1/user_guides/howto/training/cluster_howto.html)
You can run distributed training jobs on MPI clusters.
- [Python API](http://paddlepaddle.org/documentation/api/zh/0.15.0/fluid.html)
- [Python API](http://paddlepaddle.org/documentation/api/zh/1.1/fluid.html)
Our new API enables much shorter programs.
- [How to Contribute](http://paddlepaddle.org/documentation/docs/zh/0.15.0/new_docs/advanced_usage/development/contribute_to_paddle.html)
- [How to Contribute](http://paddlepaddle.org/documentation/docs/zh/1.1/advanced_usage/development/contribute_to_paddle.html)
We appreciate your contributions!
......
benchmark/fluid/args.py
浏览文件 @ ce994190
......@@ -142,5 +142,10 @@ def parse_args():
choices=['reduce', 'all_reduce'],
default='all_reduce',
help='Specify the reduce strategy, can be reduce, all_reduce')
parser.add_argument(
'--fuse_broadcast_op',
action='store_true',
help='If set, would fuse multiple broadcast operators into one fused_broadcast operator.'
)
args = parser.parse_args()
return args
benchmark/fluid/fluid_benchmark.py
浏览文件 @ ce994190
......@@ -177,6 +177,7 @@ def train_parallel(train_args, test_args, args, train_prog, test_prog,
else:
build_strategy.reduce_strategy = fluid.BuildStrategy(
).ReduceStrategy.AllReduce
build_strategy.fuse_broadcast_op = args.fuse_broadcast_op
avg_loss = train_args[0]
......@@ -240,7 +241,6 @@ def train_parallel(train_args, test_args, args, train_prog, test_prog,
if args.use_fake_data or args.use_reader_op:
try:
fetch_ret = exe.run(fetch_list)
except fluid.core.EOFException as eof:
break
......
benchmark/fluid/run.sh 100644 → 100755
浏览文件 @ ce994190
文件模式从 100644 更改为 100755
cmake/configure.cmake
浏览文件 @ ce994190
......@@ -50,11 +50,7 @@ if(NOT WITH_PROFILER)
endif(NOT WITH_PROFILER)
if(NOT CMAKE_CROSSCOMPILING)
if(WITH_AVX AND AVX512F_FOUND)
set(SIMD_FLAG ${AVX512F_FLAG})
elseif(WITH_AVX AND AVX2_FOUND)
set(SIMD_FLAG ${AVX2_FLAG})
elseif(WITH_AVX AND AVX_FOUND)
if(WITH_AVX AND AVX_FOUND)
set(SIMD_FLAG ${AVX_FLAG})
elseif(SSE3_FOUND)
set(SIMD_FLAG ${SSE3_FLAG})
......
cmake/external/xxhash.cmake 0 → 100644
浏览文件 @ ce994190
INCLUDE(ExternalProject)
set(XXHASH_SOURCE_DIR ${THIRD_PARTY_PATH}/xxhash)
set(XXHASH_INSTALL_DIR ${THIRD_PARTY_PATH}/install/xxhash)
set(XXHASH_INCLUDE_DIR "${XXHASH_INSTALL_DIR}/include")
IF(WITH_STATIC_LIB)
SET(BUILD_CMD make lib)
ELSE()
IF(APPLE)
SET(BUILD_CMD sed -i \"\" "s/-Wstrict-prototypes -Wundef/-Wstrict-prototypes -Wundef -fPIC/g" ${XXHASH_SOURCE_DIR}/src/extern_xxhash/Makefile && make lib)
ELSE(APPLE)
SET(BUILD_CMD sed -i "s/-Wstrict-prototypes -Wundef/-Wstrict-prototypes -Wundef -fPIC/g" ${XXHASH_SOURCE_DIR}/src/extern_xxhash/Makefile && make lib)
ENDIF(APPLE)
ENDIF()
ExternalProject_Add(
extern_xxhash
${EXTERNAL_PROJECT_LOG_ARGS}
GIT_REPOSITORY "https://github.com/Cyan4973/xxHash"
GIT_TAG "v0.6.5"
PREFIX ${XXHASH_SOURCE_DIR}
DOWNLOAD_NAME "xxhash"
UPDATE_COMMAND ""
CONFIGURE_COMMAND ""
BUILD_IN_SOURCE 1
PATCH_COMMAND
BUILD_COMMAND ${BUILD_CMD}
INSTALL_COMMAND export PREFIX=${XXHASH_INSTALL_DIR}/ && make install
TEST_COMMAND ""
)
set(XXHASH_LIBRARIES "${XXHASH_INSTALL_DIR}/lib/libxxhash.a")
INCLUDE_DIRECTORIES(${XXHASH_INCLUDE_DIR})
add_library(xxhash STATIC IMPORTED GLOBAL)
set_property(TARGET xxhash PROPERTY IMPORTED_LOCATION ${XXHASH_LIBRARIES})
include_directories(${XXHASH_INCLUDE_DIR})
add_dependencies(xxhash extern_xxhash)
LIST(APPEND external_project_dependencies xxhash)
IF(WITH_C_API)
INSTALL(DIRECTORY ${XXHASH_INCLUDE_DIR} DESTINATION third_party/xxhash)
IF(ANDROID)
INSTALL(FILES ${XXHASH_LIBRARIES} DESTINATION third_party/xxhash/lib/${ANDROID_ABI})
ELSE()
INSTALL(FILES ${XXHASH_LIBRARIES} DESTINATION third_party/xxhash/lib)
ENDIF()
ENDIF()
cmake/generic.cmake
浏览文件 @ ce994190
......@@ -261,6 +261,13 @@ function(cc_library TARGET_NAME)
add_dependencies(${TARGET_NAME} mklml)
target_link_libraries(${TARGET_NAME} "-L${MKLML_LIB_DIR} -liomp5 -Wl,--as-needed")
endif()
# remove link to python, see notes at:
# https://github.com/pybind/pybind11/blob/master/docs/compiling.rst#building-manually
if("${cc_library_DEPS};" MATCHES "python;")
list(REMOVE_ITEM cc_library_DEPS python)
add_dependencies(${TARGET_NAME} python)
target_link_libraries(${TARGET_NAME} "-Wl,-undefined,dynamic_lookup")
endif()
target_link_libraries(${TARGET_NAME} ${cc_library_DEPS})
add_dependencies(${TARGET_NAME} ${cc_library_DEPS})
endif()
......@@ -311,6 +318,8 @@ function(cc_test TARGET_NAME)
set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_cpu_deterministic=true)
set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_init_allocated_mem=true)
set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_cudnn_deterministic=true)
# No unit test should exceed 10 minutes.
set_tests_properties(${TARGET_NAME} PROPERTIES TIMEOUT 600)
endif()
endfunction(cc_test)
......@@ -629,6 +638,8 @@ function(py_test TARGET_NAME)
PYTHONPATH=${PADDLE_BINARY_DIR}/python ${py_test_ENVS}
${PYTHON_EXECUTABLE} -u ${py_test_SRCS} ${py_test_ARGS}
WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
# No unit test should exceed 10 minutes.
set_tests_properties(${TARGET_NAME} PROPERTIES TIMEOUT 600)
endif()
endfunction()
......
cmake/inference_lib.cmake
浏览文件 @ ce994190
......@@ -31,7 +31,7 @@ function(copy TARGET)
foreach(index RANGE ${len})
list(GET copy_lib_SRCS ${index} src)
list(GET copy_lib_DSTS ${index} dst)
add_custom_command(TARGET ${TARGET} PRE_BUILD
add_custom_command(TARGET ${TARGET} PRE_BUILD
COMMAND mkdir -p "${dst}"
COMMAND cp -r "${src}" "${dst}"
COMMENT "copying ${src} -> ${dst}")
......@@ -67,6 +67,13 @@ copy(boost_lib
DEPS boost
)
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/xxhash")
copy(xxhash_lib
SRCS ${XXHASH_INCLUDE_DIR} ${XXHASH_LIBRARIES}
DSTS ${dst_dir} ${dst_dir}/lib
DEPS xxhash
)
if(NOT PROTOBUF_FOUND)
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/protobuf")
copy(protobuf_lib
......@@ -150,16 +157,16 @@ if (WITH_ANAKIN AND WITH_MKL)
SRCS
${PADDLE_BINARY_DIR}/paddle/fluid/inference/api/libinference_anakin_api* # compiled anakin api
${ANAKIN_INSTALL_DIR} # anakin release
DSTS ${dst_dir}/inference/anakin ${FLUID_INSTALL_DIR}/third_party/install/anakin)
DSTS ${FLUID_INSTALL_DIR}/third_party/install/anakin ${FLUID_INSTALL_DIR}/third_party/install/anakin)
list(APPEND inference_deps anakin_inference_lib)
endif()
set(module "inference")
copy(inference_lib DEPS ${inference_deps}
SRCS ${src_dir}/${module}/*.h ${PADDLE_BINARY_DIR}/paddle/fluid/inference/libpaddle_fluid.*
${src_dir}/${module}/api/paddle_inference_api.h ${src_dir}/${module}/api/demo_ci
${src_dir}/${module}/api/paddle_inference_api.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} ${dst_dir}/${module}
DSTS ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module}
)
set(module "platform")
......@@ -186,20 +193,40 @@ copy(cmake_cache
DSTS ${FLUID_INSTALL_DIR})
# This command generates a complete fluid library for both train and inference
add_custom_target(fluid_lib_dist DEPENDS ${fluid_lib_dist_dep})
add_custom_target(fluid_lib_dist DEPENDS ${fluid_lib_dist_dep})
# Following commands generate a inference-only fluid library
# third_party, version.txt and CMakeCache.txt are the same position with ${FLUID_INSTALL_DIR}
copy(third_party DEPS fluid_lib_dist
SRCS ${FLUID_INSTALL_DIR}/third_party ${FLUID_INSTALL_DIR}/CMakeCache.txt
DSTS ${FLUID_INFERENCE_INSTALL_DIR} ${FLUID_INFERENCE_INSTALL_DIR}
)
# only need libpaddle_fluid.so/a and paddle_inference_api.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
DSTS ${FLUID_INFERENCE_INSTALL_DIR}/paddle/lib ${FLUID_INFERENCE_INSTALL_DIR}/paddle/include
)
add_custom_target(inference_lib_dist DEPENDS third_party inference_api_lib)
# paddle fluid version
execute_process(
COMMAND ${GIT_EXECUTABLE} log --pretty=format:%H -1
WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}
OUTPUT_VARIABLE PADDLE_GIT_COMMIT)
set(version_file ${FLUID_INSTALL_DIR}/version.txt)
file(WRITE ${version_file}
"GIT COMMIT ID: ${PADDLE_GIT_COMMIT}\n"
"WITH_MKL: ${WITH_MKL}\n"
"WITH_GPU: ${WITH_GPU}\n")
if(WITH_GPU)
file(APPEND ${version_file}
"CUDA version: ${CUDA_VERSION}\n"
"CUDNN version: v${CUDNN_MAJOR_VERSION}\n")
endif()
function(version version_file)
execute_process(
COMMAND ${GIT_EXECUTABLE} log --pretty=format:%H -1
WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}
OUTPUT_VARIABLE PADDLE_GIT_COMMIT)
file(WRITE ${version_file}
"GIT COMMIT ID: ${PADDLE_GIT_COMMIT}\n"
"WITH_MKL: ${WITH_MKL}\n"
"WITH_MKLDNN: ${WITH_MKLDNN}\n"
"WITH_GPU: ${WITH_GPU}\n")
if(WITH_GPU)
file(APPEND ${version_file}
"CUDA version: ${CUDA_VERSION}\n"
"CUDNN version: v${CUDNN_MAJOR_VERSION}\n")
endif()
endfunction()
version(${FLUID_INSTALL_DIR}/version.txt)
version(${FLUID_INFERENCE_INSTALL_DIR}/version.txt)
cmake/simd.cmake
浏览文件 @ ce994190
......@@ -89,7 +89,9 @@ CHECK_CXX_SOURCE_RUNS("
#include <immintrin.h>
int main()
{
__m512i a = _mm512_undefined_epi32();
__m512i a = _mm512_set_epi32 (-1, 2, -3, 4, -1, 2, -3, 4,
13, -5, 6, -7, 9, 2, -6, 3);
__m512i result = _mm512_abs_epi32 (a);
return 0;
}" AVX512F_FOUND)
......
paddle/CMakeLists.txt
浏览文件 @ ce994190
......@@ -24,6 +24,7 @@ if(NOT WITH_FLUID_ONLY)
endif()
add_subdirectory(testing)
set(PYTHON_TESTS_DIR ${PADDLE_BINARY_DIR}/python/paddle/fluid/tests CACHE INTERNAL "python tests directory")
if(NOT MOBILE_INFERENCE AND NOT RPI AND NOT WITH_C_API)
add_subdirectory(fluid)
endif()
paddle/fluid/API.spec
浏览文件 @ ce994190
......@@ -61,14 +61,14 @@ paddle.fluid.layers.cos_sim ArgSpec(args=['X', 'Y'], varargs=None, keywords=None
paddle.fluid.layers.cross_entropy ArgSpec(args=['input', 'label', 'soft_label', 'ignore_index'], varargs=None, keywords=None, defaults=(False, -100))
paddle.fluid.layers.square_error_cost ArgSpec(args=['input', 'label'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.chunk_eval ArgSpec(args=['input', 'label', 'chunk_scheme', 'num_chunk_types', 'excluded_chunk_types'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.sequence_conv ArgSpec(args=['input', 'num_filters', 'filter_size', 'filter_stride', 'padding', 'bias_attr', 'param_attr', 'act'], varargs=None, keywords=None, defaults=(3, 1, None, None, None, None))
paddle.fluid.layers.sequence_conv ArgSpec(args=['input', 'num_filters', 'filter_size', 'filter_stride', 'padding', 'bias_attr', 'param_attr', 'act', 'name'], varargs=None, keywords=None, defaults=(3, 1, None, None, None, None, None))
paddle.fluid.layers.conv2d ArgSpec(args=['input', 'num_filters', 'filter_size', 'stride', 'padding', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(1, 0, 1, None, None, None, True, None, None))
paddle.fluid.layers.conv3d ArgSpec(args=['input', 'num_filters', 'filter_size', 'stride', 'padding', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(1, 0, 1, None, None, None, True, None, None))
paddle.fluid.layers.sequence_pool ArgSpec(args=['input', 'pool_type'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.sequence_softmax ArgSpec(args=['input', 'param_attr', 'bias_attr', 'use_cudnn'], varargs=None, keywords=None, defaults=(None, None, False))
paddle.fluid.layers.softmax ArgSpec(args=['input', 'param_attr', 'bias_attr', 'use_cudnn', 'name'], varargs=None, keywords=None, defaults=(None, None, True, None))
paddle.fluid.layers.pool2d ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None))
paddle.fluid.layers.pool3d ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None))
paddle.fluid.layers.sequence_pool ArgSpec(args=['input', 'pool_type', 'is_test'], varargs=None, keywords=None, defaults=(False,))
paddle.fluid.layers.sequence_softmax ArgSpec(args=['input', 'use_cudnn', 'name'], varargs=None, keywords=None, defaults=(False, None))
paddle.fluid.layers.softmax ArgSpec(args=['input', 'use_cudnn', 'name'], varargs=None, keywords=None, defaults=(True, None))
paddle.fluid.layers.pool2d ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name', 'exclusive'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None, True))
paddle.fluid.layers.pool3d ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name', 'exclusive'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None, True))
paddle.fluid.layers.batch_norm ArgSpec(args=['input', 'act', 'is_test', 'momentum', 'epsilon', 'param_attr', 'bias_attr', 'data_layout', 'in_place', 'name', 'moving_mean_name', 'moving_variance_name', 'do_model_average_for_mean_and_var', 'fuse_with_relu'], varargs=None, keywords=None, defaults=(None, False, 0.9, 1e-05, None, None, 'NCHW', False, None, None, None, False, False))
paddle.fluid.layers.beam_search_decode ArgSpec(args=['ids', 'scores', 'beam_size', 'end_id', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.conv2d_transpose ArgSpec(args=['input', 'num_filters', 'output_size', 'filter_size', 'padding', 'stride', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, None, 0, 1, 1, None, None, None, True, None, None))
......@@ -86,7 +86,7 @@ paddle.fluid.layers.reduce_prod ArgSpec(args=['input', 'dim', 'keep_dim', 'name'
paddle.fluid.layers.sequence_first_step ArgSpec(args=['input'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.sequence_last_step ArgSpec(args=['input'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.sequence_slice ArgSpec(args=['input', 'offset', 'length', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.dropout ArgSpec(args=['x', 'dropout_prob', 'is_test', 'seed', 'name'], varargs=None, keywords=None, defaults=(False, None, None))
paddle.fluid.layers.dropout ArgSpec(args=['x', 'dropout_prob', 'is_test', 'seed', 'name', 'dropout_implementation'], varargs=None, keywords=None, defaults=(False, None, None, 'downgrade_in_infer'))
paddle.fluid.layers.split ArgSpec(args=['input', 'num_or_sections', 'dim', 'name'], varargs=None, keywords=None, defaults=(-1, None))
paddle.fluid.layers.ctc_greedy_decoder ArgSpec(args=['input', 'blank', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.edit_distance ArgSpec(args=['input', 'label', 'normalized', 'ignored_tokens'], varargs=None, keywords=None, defaults=(True, None))
......@@ -97,17 +97,17 @@ paddle.fluid.layers.warpctc ArgSpec(args=['input', 'label', 'blank', 'norm_by_ti
paddle.fluid.layers.sequence_reshape ArgSpec(args=['input', 'new_dim'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.transpose ArgSpec(args=['x', 'perm', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.im2sequence ArgSpec(args=['input', 'filter_size', 'stride', 'padding', 'input_image_size', 'out_stride', 'name'], varargs=None, keywords=None, defaults=(1, 1, 0, None, 1, None))
paddle.fluid.layers.nce ArgSpec(args=['input', 'label', 'num_total_classes', 'sample_weight', 'param_attr', 'bias_attr', 'num_neg_samples'], varargs=None, keywords=None, defaults=(None, None, None, None))
paddle.fluid.layers.hsigmoid ArgSpec(args=['input', 'label', 'num_classes', 'param_attr', 'bias_attr'], varargs=None, keywords=None, defaults=(None, None))
paddle.fluid.layers.nce ArgSpec(args=['input', 'label', 'num_total_classes', 'sample_weight', 'param_attr', 'bias_attr', 'num_neg_samples', 'name'], varargs=None, keywords=None, defaults=(None, None, None, None, None))
paddle.fluid.layers.hsigmoid ArgSpec(args=['input', 'label', 'num_classes', 'param_attr', 'bias_attr', 'name'], varargs=None, keywords=None, defaults=(None, None, None))
paddle.fluid.layers.beam_search ArgSpec(args=['pre_ids', 'pre_scores', 'ids', 'scores', 'beam_size', 'end_id', 'level', 'name'], varargs=None, keywords=None, defaults=(0, None))
paddle.fluid.layers.row_conv ArgSpec(args=['input', 'future_context_size', 'param_attr', 'act'], varargs=None, keywords=None, defaults=(None, None))
paddle.fluid.layers.multiplex ArgSpec(args=['inputs', 'index'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.layer_norm ArgSpec(args=['input', 'scale', 'shift', 'begin_norm_axis', 'epsilon', 'param_attr', 'bias_attr', 'act', 'name'], varargs=None, keywords=None, defaults=(True, True, 1, 1e-05, None, None, None, None))
paddle.fluid.layers.softmax_with_cross_entropy ArgSpec(args=['logits', 'label', 'soft_label', 'ignore_index'], varargs=None, keywords=None, defaults=(False, -100))
paddle.fluid.layers.softmax_with_cross_entropy ArgSpec(args=['logits', 'label', 'soft_label', 'ignore_index', 'numeric_stable_mode'], varargs=None, keywords=None, defaults=(False, -100, 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))
paddle.fluid.layers.reshape ArgSpec(args=['x', 'shape', 'actual_shape', 'act', 'inplace', 'name'], varargs=None, keywords=None, defaults=(None, None, True, None))
paddle.fluid.layers.reshape ArgSpec(args=['x', 'shape', 'actual_shape', 'act', 'inplace', 'name'], varargs=None, keywords=None, defaults=(None, None, False, None))
paddle.fluid.layers.squeeze ArgSpec(args=['input', 'axes', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.unsqueeze ArgSpec(args=['input', 'axes', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.lod_reset ArgSpec(args=['x', 'y', 'target_lod'], varargs=None, keywords=None, defaults=(None, None))
......@@ -116,6 +116,7 @@ paddle.fluid.layers.pad ArgSpec(args=['x', 'paddings', 'pad_value', 'name'], var
paddle.fluid.layers.pad_constant_like ArgSpec(args=['x', 'y', 'pad_value', 'name'], varargs=None, keywords=None, defaults=(0.0, None))
paddle.fluid.layers.label_smooth ArgSpec(args=['label', 'prior_dist', 'epsilon', 'dtype', 'name'], varargs=None, keywords=None, defaults=(None, 0.1, 'float32', None))
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_short ArgSpec(args=['input', 'out_short_len', 'resample'], varargs=None, keywords=None, defaults=('BILINEAR',))
......@@ -173,6 +174,13 @@ 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.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.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.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)
......@@ -352,6 +360,8 @@ paddle.fluid.optimizer.ModelAverage.__init__ ArgSpec(args=['self', 'average_wind
paddle.fluid.optimizer.ModelAverage.apply ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None)
paddle.fluid.optimizer.ModelAverage.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None))
paddle.fluid.optimizer.ModelAverage.restore ArgSpec(args=['self', 'executor'], varargs=None, keywords=None, defaults=None)
paddle.fluid.optimizer.LarsMomentumOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'momentum', 'lars_coeff', 'lars_weight_decay', 'regularization', 'name'], varargs=None, keywords=None, defaults=(0.001, 0.0005, None, None))
paddle.fluid.optimizer.LarsMomentumOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None))
paddle.fluid.backward.append_backward ArgSpec(args=['loss', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None))
paddle.fluid.regularizer.L1DecayRegularizer.__init__ ArgSpec(args=['self', 'regularization_coeff'], varargs=None, keywords=None, defaults=(0.0,))
paddle.fluid.regularizer.L2DecayRegularizer.__init__ ArgSpec(args=['self', 'regularization_coeff'], varargs=None, keywords=None, defaults=(0.0,))
......
paddle/fluid/CMakeLists.txt
浏览文件 @ ce994190
......@@ -9,8 +9,6 @@ add_subdirectory(pybind)
add_subdirectory(recordio)
endif(NOT WIN32)
if(WITH_INFERENCE)
# NOTE: please add subdirectory inference at last.
add_subdirectory(inference)
add_subdirectory(train)
endif()
# NOTE: please add subdirectory inference at last.
add_subdirectory(inference)
add_subdirectory(train)
paddle/fluid/framework/attribute.cc
浏览文件 @ ce994190
......@@ -64,6 +64,13 @@ Attribute GetAttrValue(const proto::OpDesc::Attr& attr_desc) {
case proto::AttrType::LONG: {
return attr_desc.l();
}
case proto::AttrType::LONGS: {
std::vector<int64_t> val(attr_desc.longs_size());
for (int i = 0; i < attr_desc.longs_size(); ++i) {
val[i] = attr_desc.longs(i);
}
return val;
}
default:
PADDLE_THROW("Unsupport attr type %d", attr_desc.type());
}
......
paddle/fluid/framework/attribute.h
浏览文件 @ ce994190
......@@ -26,6 +26,113 @@ limitations under the License. */
namespace paddle {
namespace framework {
template <typename T>
struct ExtractAttribute {
explicit ExtractAttribute(const std::string& attr_name)
: attr_name_(attr_name) {}
T* operator()(Attribute& attr) const {
T* attr_value = nullptr;
try {
attr_value = &boost::get<T>(attr);
} catch (boost::bad_get& bad_get) {
PADDLE_THROW("Cannot get attribute %s by type %s, its type is %s",
attr_name_, paddle::platform::demangle(typeid(T).name()),
paddle::platform::demangle(attr.type().name()));
}
return attr_value;
}
const std::string& attr_name_;
};
// special handle bool
// FIXME(yuyang18): Currently we cast bool into int in python binding. It is
// hard to change the logic there. In another way, we should correct handle
// if the user set `some_flag=1`.
//
// FIX ME anytime if there is a better solution.
template <>
struct ExtractAttribute<bool> {
explicit ExtractAttribute(const std::string& attr_name)
: attr_name_(attr_name) {}
bool* operator()(Attribute& attr) const {
if (attr.type() == typeid(int)) { // NOLINT
int val = boost::get<int>(attr);
attr = static_cast<bool>(val);
} else if (attr.type() == typeid(float)) { // NOLINT
float val = boost::get<float>(attr);
attr = static_cast<bool>(val);
}
bool* attr_value = nullptr;
try {
attr_value = &boost::get<bool>(attr);
} catch (boost::bad_get& bad_get) {
PADDLE_THROW("Cannot get attribute %s by type bool, its type is %s",
attr_name_, paddle::platform::demangle(attr.type().name()));
}
return attr_value;
}
const std::string& attr_name_;
};
template <>
struct ExtractAttribute<int64_t> {
explicit ExtractAttribute(const std::string& attr_name)
: attr_name_(attr_name) {}
int64_t* operator()(Attribute& attr) const {
if (attr.type() == typeid(int)) { // NOLINT
int val = boost::get<int>(attr);
attr = static_cast<int64_t>(val);
} else if (attr.type() == typeid(float)) { // NOLINT
int val = boost::get<float>(attr);
attr = static_cast<int64_t>(val);
}
int64_t* attr_value = nullptr;
try {
attr_value = &boost::get<int64_t>(attr);
} catch (boost::bad_get& bad_get) {
PADDLE_THROW("Cannot get attribute %s by type int64_t, its type is %s",
attr_name_, paddle::platform::demangle(attr.type().name()));
}
return attr_value;
}
const std::string& attr_name_;
};
template <>
struct ExtractAttribute<std::vector<int64_t>> {
explicit ExtractAttribute(const std::string& attr_name)
: attr_name_(attr_name) {}
std::vector<int64_t>* operator()(Attribute& attr) const {
if (attr.type() == typeid(std::vector<int>)) { // NOLINT
std::vector<int> val = boost::get<std::vector<int>>(attr);
std::vector<int64_t> vec(val.begin(), val.end());
attr = vec;
} else if (attr.type() == typeid(std::vector<float>)) { // NOLINT
std::vector<float> val = boost::get<std::vector<float>>(attr);
std::vector<int64_t> vec(val.begin(), val.end());
attr = vec;
}
std::vector<int64_t>* attr_value = nullptr;
try {
attr_value = &boost::get<std::vector<int64_t>>(attr);
} catch (boost::bad_get& bad_get) {
PADDLE_THROW("Cannot get attribute %s by type int64_t, its type is %s",
attr_name_, paddle::platform::demangle(attr.type().name()));
}
return attr_value;
}
const std::string& attr_name_;
};
template <typename T>
inline proto::AttrType AttrTypeID() {
Attribute tmp = T();
......@@ -42,7 +149,11 @@ class AttrReader {
inline const T& Get(const std::string& name) const {
PADDLE_ENFORCE(attrs_.count(name) != 0, "%s should be in AttributeMap",
name);
return boost::get<T>(attrs_.at(name));
Attribute& attr = const_cast<Attribute&>(attrs_.at(name));
ExtractAttribute<T> extract_attr(name);
T* attr_value = extract_attr(attr);
return *attr_value;
}
private:
......@@ -82,7 +193,7 @@ class DefaultValueSetter {
public:
explicit DefaultValueSetter(T default_value)
: default_value_(default_value) {}
void operator()(T& value) const { value = default_value_; }
void operator()(T& value) const { value = default_value_; } // NOLINT
private:
T default_value_;
......@@ -117,84 +228,6 @@ class EnumInContainer {
std::unordered_set<T> container_;
};
template <typename T>
struct ExtractAttribute {
explicit ExtractAttribute(const std::string& attr_name)
: attr_name_(attr_name) {}
T* operator()(Attribute& attr) const {
T* attr_value = nullptr;
try {
attr_value = &boost::get<T>(attr);
} catch (boost::bad_get& bad_get) {
PADDLE_THROW("Cannot get attribute %s by type %s, its type is %s",
attr_name_, paddle::platform::demangle(typeid(T).name()),
paddle::platform::demangle(attr.type().name()));
}
return attr_value;
}
const std::string& attr_name_;
};
// special handle bool
// FIXME(yuyang18): Currently we cast bool into int in python binding. It is
// hard to change the logic there. In another way, we should correct handle
// if the user set `some_flag=1`.
//
// FIX ME anytime if there is a better solution.
template <>
struct ExtractAttribute<bool> {
explicit ExtractAttribute(const std::string& attr_name)
: attr_name_(attr_name) {}
bool* operator()(Attribute& attr) const {
if (attr.type() == typeid(int)) { // NOLINT
int val = boost::get<int>(attr);
attr = static_cast<bool>(val);
} else if (attr.type() == typeid(float)) { // NOLINT
float val = boost::get<float>(attr);
attr = static_cast<bool>(val);
}
bool* attr_value = nullptr;
try {
attr_value = &boost::get<bool>(attr);
} catch (boost::bad_get& bad_get) {
PADDLE_THROW("Cannot get attribute %s by type bool, its type is %s",
attr_name_, paddle::platform::demangle(attr.type().name()));
}
return attr_value;
}
const std::string& attr_name_;
};
template <>
struct ExtractAttribute<int64_t> {
explicit ExtractAttribute(const std::string& attr_name)
: attr_name_(attr_name) {}
int64_t* operator()(Attribute& attr) const {
if (attr.type() == typeid(int)) { // NOLINT
int val = boost::get<int>(attr);
attr = static_cast<int64_t>(val);
} else if (attr.type() == typeid(float)) { // NOLINT
int val = boost::get<float>(attr);
attr = static_cast<int64_t>(val);
}
int64_t* attr_value = nullptr;
try {
attr_value = &boost::get<int64_t>(attr);
} catch (boost::bad_get& bad_get) {
PADDLE_THROW("Cannot get attribute %s by type int64_t, its type is %s",
attr_name_, paddle::platform::demangle(attr.type().name()));
}
return attr_value;
}
const std::string& attr_name_;
};
// check whether a certain attribute fit its limits
// an attribute can have more than one limits
template <typename T>
......@@ -235,7 +268,7 @@ class TypedAttrChecker {
return *this;
}
void operator()(AttributeMap& attr_map) const {
void operator()(AttributeMap& attr_map) const { // NOLINT
if (!attr_map.count(attr_name_)) {
// user do not set this attr
PADDLE_ENFORCE(!default_value_setter_.empty(),
......@@ -271,7 +304,7 @@ class OpAttrChecker {
return *(checker.target<TypedAttrChecker<T>>());
}
void Check(AttributeMap& attr_map) const {
void Check(AttributeMap& attr_map) const { // NOLINT
for (const auto& checker : attr_checkers_) {
checker(attr_map);
}
......
paddle/fluid/framework/details/CMakeLists.txt
浏览文件 @ ce994190
cc_library(var_handle SRCS var_handle.cc DEPS place framework_proto node)
cc_library(op_handle_base SRCS op_handle_base.cc DEPS var_handle device_context lod_tensor)
cc_library(op_graph_view SRCS op_graph_view.cc DEPS op_handle_base)
cc_library(scale_loss_grad_op_handle SRCS scale_loss_grad_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory)
cc_library(fetch_op_handle SRCS fetch_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory)
cc_library(computation_op_handle SRCS computation_op_handle.cc DEPS framework_proto scope place operator op_registry)
......@@ -16,32 +17,39 @@ if(WITH_GPU)
dynload_cuda variable_visitor)
nv_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope ddim dynload_cuda)
nv_library(broadcast_op_handle SRCS broadcast_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor dynload_cuda)
nv_library(fused_broadcast_op_handle SRCS fused_broadcast_op_handle.cc DEPS broadcast_op_handle)
else()
cc_library(all_reduce_op_handle SRCS all_reduce_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory
variable_visitor)
cc_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope ddim)
cc_library(broadcast_op_handle SRCS broadcast_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor)
cc_library(fused_broadcast_op_handle SRCS fused_broadcast_op_handle.cc DEPS broadcast_op_handle)
endif()
cc_library(data_balance_op_handle SRCS data_balance_op_handle.cc DEPS op_handle_base scope lod_tensor)
cc_library(gather_op_handle SRCS gather_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor)
cc_library(fuse_vars_op_handle SRCS fuse_vars_op_handle.cc DEPS op_handle_base scope)
if(WITH_GPU)
cc_library(modify_op_lock_and_record_event_pass SRCS modify_op_lock_and_record_event_pass.cc DEPS computation_op_handle op_graph_view multi_devices_helper)
if (WITH_GPU)
cc_library(reference_count_pass SRCS reference_count_pass.cc DEPS computation_op_handle scale_loss_grad_op_handle rpc_op_handle
all_reduce_op_handle reduce_op_handle broadcast_op_handle data_balance_op_handle graph graph_helper pass)
endif()
cc_library(sequential_execution_pass SRCS sequential_execution_pass.cc DEPS graph graph_helper pass)
cc_library(multi_devices_graph_pass SRCS multi_devices_graph_pass.cc DEPS multi_devices_helper computation_op_handle
scale_loss_grad_op_handle rpc_op_handle all_reduce_op_handle reduce_op_handle broadcast_op_handle data_balance_op_handle)
scale_loss_grad_op_handle rpc_op_handle all_reduce_op_handle reduce_op_handle broadcast_op_handle data_balance_op_handle fused_broadcast_op_handle)
if(WITH_GPU)
cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS graph framework_proto reference_count_pass)
else()
cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS graph framework_proto)
set(SSA_GRAPH_EXECUTOR_DEPS graph framework_proto sequential_execution_pass modify_op_lock_and_record_event_pass)
if (WITH_GPU)
list(APPEND SSA_GRAPH_EXECUTOR_DEPS reference_count_pass)
endif()
cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS ${SSA_GRAPH_EXECUTOR_DEPS})
cc_library(threaded_ssa_graph_executor SRCS threaded_ssa_graph_executor.cc DEPS fetch_op_handle ssa_graph_executor scope
simple_threadpool device_context)
......@@ -54,8 +62,9 @@ cc_library(scope_buffered_ssa_graph_executor SRCS scope_buffered_ssa_graph_execu
# device_context reduce_op_handle )
cc_library(fast_threaded_ssa_graph_executor SRCS fast_threaded_ssa_graph_executor.cc
DEPS fetch_op_handle ssa_graph_executor scope simple_threadpool device_context)
cc_test(fused_broadcast_op_test SRCS fused_broadcast_op_handle_test.cc DEPS fused_broadcast_op_handle)
cc_library(build_strategy SRCS build_strategy.cc DEPS
graph_viz_pass multi_devices_graph_pass
multi_devices_graph_print_pass multi_devices_graph_check_pass
fuse_elewise_add_act_pass)
fuse_elewise_add_act_pass multi_batch_merge_pass)
paddle/fluid/framework/details/all_reduce_op_handle.cc
浏览文件 @ ce994190
......@@ -34,7 +34,7 @@ AllReduceOpHandle::AllReduceOpHandle(ir::Node *node,
nccl_ctxs_(ctxs) {
if (nccl_ctxs_) {
for (auto &p : places_) {
this->dev_ctxes_[p] = nccl_ctxs_->DevCtx(p);
this->SetDeviceContext(p, nccl_ctxs_->DevCtx(p));
}
}
}
......@@ -46,7 +46,7 @@ AllReduceOpHandle::AllReduceOpHandle(ir::Node *node,
#endif
void AllReduceOpHandle::RunImpl() {
platform::RecordEvent record_event(Name(), dev_ctxes_.begin()->second);
platform::RecordEvent record_event(Name(), dev_ctxes_.cbegin()->second);
if (NoDummyInputSize() == 1) {
return; // No need to all reduce when GPU count = 1;
......@@ -127,7 +127,7 @@ void AllReduceOpHandle::RunImpl() {
*local_scopes_[i]->FindVar(kLocalExecScopeName)->Get<Scope *>();
auto &p = places_[i];
auto *var = scope.FindVar(out_var_handles[i]->name_);
auto *dev_ctx = dev_ctxes_[p];
auto *dev_ctx = dev_ctxes_.at(p);
RunAndRecordEvent(p, [&trg, var, dev_ctx, p] {
auto &tensor_gpu = *var->GetMutable<framework::LoDTensor>();
......
paddle/fluid/framework/details/broadcast_op_handle.cc
浏览文件 @ ce994190
......@@ -48,16 +48,27 @@ void BroadcastOpHandle::RunImpl() {
var_scopes.emplace_back(s->FindVar(kLocalExecScopeName)->Get<Scope *>());
}
BroadcastOneVar(*in_var_handle, out_var_handles, var_scopes);
}
void BroadcastOpHandle::BroadcastOneVar(
const VarHandle &in_var_handle,
const std::vector<VarHandle *> &out_var_handles,
const std::vector<const Scope *> &var_scopes) {
auto *in_var =
var_scopes.at(in_var_handle->scope_idx_)->FindVar(in_var_handle->name_);
var_scopes.at(in_var_handle.scope_idx_)->FindVar(in_var_handle.name_);
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!";
return;
}
InitOutputValue(*in_var_handle, out_var_handles);
InitOutputValue(in_var_handle, out_var_handles);
if (platform::is_cpu_place(in_tensor.place())) {
for (auto *out_var_handle : out_var_handles) {
if (out_var_handle->IsTheSameVar(*in_var_handle)) {
if (out_var_handle->IsTheSameVar(in_var_handle)) {
continue;
}
auto &out_p = out_var_handle->place_;
......@@ -114,12 +125,12 @@ void BroadcastOpHandle::RunImpl() {
}
}
if (!out_handle->IsTheSameVar(*in_var_handle)) {
auto out_var = var_scopes.at(in_var_handle->scope_idx_)
if (!out_handle->IsTheSameVar(in_var_handle)) {
auto out_var = var_scopes.at(in_var_handle.scope_idx_)
->FindVar(out_var_handles[0]->name_);
paddle::framework::TensorCopy(
in_tensor, in_var_handle->place_,
*(dev_ctxes_.at(in_var_handle->place_)),
in_tensor, in_var_handle.place_,
*(dev_ctxes_.at(in_var_handle.place_)),
&VariableVisitor::GetMutableTensor(out_var));
}
});
......
paddle/fluid/framework/details/broadcast_op_handle.h
浏览文件 @ ce994190
......@@ -44,7 +44,8 @@ struct BroadcastOpHandle : public OpHandleBase {
nccl_ctxs_(nccl_ctxs) {
if (nccl_ctxs_) {
for (auto &p_ctx : nccl_ctxs_->contexts_) {
dev_ctxes_[platform::CUDAPlace(p_ctx.first)] = p_ctx.second.ctx_.get();
this->SetDeviceContext(platform::CUDAPlace(p_ctx.first),
p_ctx.second.ctx_.get());
}
}
}
......@@ -61,7 +62,10 @@ struct BroadcastOpHandle : public OpHandleBase {
protected:
void RunImpl() override;
private:
void BroadcastOneVar(const VarHandle &in_var_handle,
const std::vector<VarHandle *> &out_var_handles,
const std::vector<const Scope *> &var_scopes);
std::vector<Scope *> local_scopes_;
std::vector<platform::Place> places_;
#ifdef PADDLE_WITH_CUDA
......
paddle/fluid/framework/details/broadcast_op_handle_test.cc
浏览文件 @ ce994190
......@@ -12,232 +12,12 @@
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/framework/details/broadcast_op_handle.h"
#include "gtest/gtest.h"
#include "paddle/fluid/platform/device_context.h"
#include "paddle/fluid/framework/details/broadcast_op_handle_test.h"
namespace paddle {
namespace framework {
namespace details {
namespace f = paddle::framework;
namespace p = paddle::platform;
// test data amount
const f::DDim kDims = {20, 20};
struct TestBroadcastOpHandle {
std::vector<std::unique_ptr<p::DeviceContext>> ctxs_;
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_;
std::vector<p::Place> gpu_list_;
bool use_gpu_;
#ifdef PADDLE_WITH_CUDA
std::unique_ptr<platform::NCCLContextMap> nccl_ctxs_;
#endif
void WaitAll() {
for (size_t j = 0; j < ctxs_.size(); ++j) {
ctxs_[j]->Wait();
}
#ifdef PADDLE_WITH_CUDA
if (nccl_ctxs_) {
nccl_ctxs_->WaitAll();
}
#endif
}
void InitCtxOnGpu(bool use_gpu) {
use_gpu_ = use_gpu;
if (use_gpu_) {
#ifdef PADDLE_WITH_CUDA
int count = p::GetCUDADeviceCount();
if (count <= 1) {
LOG(WARNING) << "Cannot test multi-gpu Broadcast, because the CUDA "
"device count is "
<< count;
exit(0);
}
for (int i = 0; i < count; ++i) {
auto p = p::CUDAPlace(i);
gpu_list_.push_back(p);
ctxs_.emplace_back(new p::CUDADeviceContext(p));
}
nccl_ctxs_.reset(new platform::NCCLContextMap(gpu_list_));
#else
PADDLE_THROW("CUDA is not support.");
#endif
} else {
int count = 8;
for (int i = 0; i < count; ++i) {
auto p = p::CPUPlace();
gpu_list_.push_back(p);
ctxs_.emplace_back(new p::CPUDeviceContext(p));
}
#ifdef PADDLE_WITH_CUDA
nccl_ctxs_.reset(nullptr);
#endif
}
}
void InitBroadcastOp(size_t input_scope_idx) {
for (size_t j = 0; j < gpu_list_.size(); ++j) {
local_scopes_.push_back(&(g_scope_.NewScope()));
Scope& local_scope = local_scopes_.back()->NewScope();
*local_scopes_.back()
->Var(details::kLocalExecScopeName)
->GetMutable<Scope*>() = &local_scope;
local_scope.Var("out");
param_scopes_.emplace_back(&local_scope);
}
param_scopes_[input_scope_idx]->Var("input");
std::unique_ptr<ir::Node> n =
ir::CreateNodeForTest("node0", ir::Node::Type::kOperation);
if (use_gpu_) {
#ifdef PADDLE_WITH_CUDA
op_handle_.reset(new BroadcastOpHandle(n.get(), local_scopes_, gpu_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_, gpu_list_,
nccl_ctxs_.get()));
#else
op_handle_.reset(
new BroadcastOpHandle(n.get(), local_scopes_, gpu_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",
gpu_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()));
DummyVarHandle* dummy_var_handle =
static_cast<DummyVarHandle*>(vars_.back().get());
dummy_var_handle->ClearGeneratedOp();
op_handle_->AddInput(dummy_var_handle);
for (size_t j = 0; j < gpu_list_.size(); ++j) {
if (!use_gpu_) {
op_handle_->SetDeviceContext(gpu_list_[j], ctxs_[j].get());
}
std::unique_ptr<ir::Node> v3 =
ir::CreateNodeForTest("node3", ir::Node::Type::kVariable);
VarHandle* out_var_handle =
new VarHandle(v3.get(), 2, j, "out", gpu_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()));
DummyVarHandle* out_dummy_var_handle =
static_cast<DummyVarHandle*>(vars_.back().get());
out_dummy_var_handle->ClearGeneratedOp();
op_handle_->AddOutput(out_dummy_var_handle);
}
void TestBroadcastLodTensor(size_t input_scope_idx) {
auto in_var = param_scopes_[input_scope_idx]->FindVar("input");
PADDLE_ENFORCE_NOT_NULL(in_var);
auto in_lod_tensor = in_var->GetMutable<f::LoDTensor>();
in_lod_tensor->mutable_data<float>(kDims, gpu_list_[input_scope_idx]);
std::vector<float> send_vector(static_cast<size_t>(f::product(kDims)));
for (size_t k = 0; k < send_vector.size(); ++k) {
send_vector[k] = k;
}
f::LoD lod{{0, 10, 20}};
paddle::framework::TensorFromVector<float>(
send_vector, *(ctxs_[input_scope_idx]), in_lod_tensor);
in_lod_tensor->set_lod(lod);
in_lod_tensor->Resize(kDims);
op_handle_->Run(false);
WaitAll();
p::CPUPlace cpu_place;
for (size_t j = 0; j < gpu_list_.size(); ++j) {
auto out_var = param_scopes_[j]->FindVar("out");
PADDLE_ENFORCE_NOT_NULL(out_var);
auto out_tensor = out_var->Get<f::LoDTensor>();
PADDLE_ENFORCE_EQ(out_tensor.lod(), lod, "lod is not equal.");
f::Tensor result_tensor;
f::TensorCopySync(out_tensor, cpu_place, &result_tensor);
float* ct = result_tensor.mutable_data<float>(cpu_place);
for (int64_t i = 0; i < f::product(kDims); ++i) {
ASSERT_NEAR(ct[i], send_vector[i], 1e-5);
}
}
}
void TestBroadcastSelectedRows(size_t input_scope_idx) {
auto in_var = param_scopes_[input_scope_idx]->FindVar("input");
PADDLE_ENFORCE_NOT_NULL(in_var);
auto in_selected_rows = in_var->GetMutable<f::SelectedRows>();
auto value = in_selected_rows->mutable_value();
value->mutable_data<float>(kDims, gpu_list_[input_scope_idx]);
int height = static_cast<int>(kDims[0]) * 2;
std::vector<int64_t> rows{0, 1, 2, 3, 3, 0, 14, 7, 3, 1,
2, 4, 6, 3, 1, 1, 1, 1, 3, 7};
in_selected_rows->set_height(height);
in_selected_rows->set_rows(rows);
std::vector<float> send_vector(static_cast<size_t>(f::product(kDims)));
for (size_t k = 0; k < send_vector.size(); ++k) {
send_vector[k] = k;
}
paddle::framework::TensorFromVector<float>(
send_vector, *(ctxs_[input_scope_idx]), value);
op_handle_->Run(false);
WaitAll();
p::CPUPlace cpu_place;
for (size_t j = 0; j < gpu_list_.size(); ++j) {
auto out_var = param_scopes_[j]->FindVar("out");
PADDLE_ENFORCE_NOT_NULL(out_var);
auto& out_select_rows = out_var->Get<f::SelectedRows>();
auto rt = out_select_rows.value();
PADDLE_ENFORCE_EQ(out_select_rows.height(), height,
"height is not equal.");
for (size_t k = 0; k < out_select_rows.rows().size(); ++k) {
PADDLE_ENFORCE_EQ(out_select_rows.rows()[k], rows[k]);
}
f::Tensor result_tensor;
f::TensorCopySync(rt, cpu_place, &result_tensor);
float* ct = result_tensor.data<float>();
for (int64_t i = 0; i < f::product(kDims); ++i) {
ASSERT_NEAR(ct[i], send_vector[i], 1e-5);
}
}
}
};
TEST(BroadcastTester, TestCPUBroadcastTestLodTensor) {
TestBroadcastOpHandle test_op;
size_t input_scope_idx = 0;
......
paddle/fluid/framework/details/broadcast_op_handle_test.h 0 → 100644
浏览文件 @ ce994190
// 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 <vector>
#include "gtest/gtest.h"
#include "paddle/fluid/framework/details/broadcast_op_handle.h"
#include "paddle/fluid/platform/device_context.h"
namespace paddle {
namespace framework {
namespace details {
namespace f = paddle::framework;
namespace p = paddle::platform;
// test data amount
const f::DDim kDims = {20, 20};
struct TestBroadcastOpHandle {
std::vector<std::unique_ptr<p::DeviceContext>> ctxs_;
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_;
std::vector<p::Place> place_list_;
bool use_gpu_;
#ifdef PADDLE_WITH_CUDA
std::unique_ptr<platform::NCCLContextMap> nccl_ctxs_;
#endif
void WaitAll() {
for (size_t j = 0; j < ctxs_.size(); ++j) {
ctxs_[j]->Wait();
}
#ifdef PADDLE_WITH_CUDA
if (nccl_ctxs_) {
nccl_ctxs_->WaitAll();
}
#endif
}
void InitCtxOnGpu(bool use_gpu) {
use_gpu_ = use_gpu;
if (use_gpu_) {
#ifdef PADDLE_WITH_CUDA
int count = p::GetCUDADeviceCount();
if (count <= 1) {
LOG(WARNING) << "Cannot test multi-gpu Broadcast, because the CUDA "
"device count is "
<< count;
exit(0);
}
for (int i = 0; i < count; ++i) {
auto p = p::CUDAPlace(i);
place_list_.push_back(p);
ctxs_.emplace_back(new p::CUDADeviceContext(p));
}
nccl_ctxs_.reset(new platform::NCCLContextMap(place_list_));
#else
PADDLE_THROW("CUDA is not support.");
#endif
} else {
int count = 8;
for (int i = 0; i < count; ++i) {
auto p = p::CPUPlace();
place_list_.push_back(p);
ctxs_.emplace_back(new p::CPUDeviceContext(p));
}
#ifdef PADDLE_WITH_CUDA
nccl_ctxs_.reset(nullptr);
#endif
}
}
void InitBroadcastOp(size_t input_scope_idx) {
for (size_t j = 0; j < place_list_.size(); ++j) {
local_scopes_.push_back(&(g_scope_.NewScope()));
Scope& local_scope = local_scopes_.back()->NewScope();
*local_scopes_.back()
->Var(details::kLocalExecScopeName)
->GetMutable<Scope*>() = &local_scope;
local_scope.Var("out");
param_scopes_.emplace_back(&local_scope);
}
param_scopes_[input_scope_idx]->Var("input");
std::unique_ptr<ir::Node> n =
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()));
#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()));
#else
op_handle_.reset(
new BroadcastOpHandle(n.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]);
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()));
DummyVarHandle* dummy_var_handle =
static_cast<DummyVarHandle*>(vars_.back().get());
dummy_var_handle->ClearGeneratedOp();
op_handle_->AddInput(dummy_var_handle);
for (size_t j = 0; j < place_list_.size(); ++j) {
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);
VarHandle* out_var_handle =
new VarHandle(v3.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()));
DummyVarHandle* out_dummy_var_handle =
static_cast<DummyVarHandle*>(vars_.back().get());
out_dummy_var_handle->ClearGeneratedOp();
op_handle_->AddOutput(out_dummy_var_handle);
}
std::vector<float> InitLoDTensor(const std::string& varname,
size_t input_scope_idx, const f::LoD& lod,
float val_scalar = 0.0) {
auto var = param_scopes_[input_scope_idx]->FindVar(varname);
PADDLE_ENFORCE_NOT_NULL(var);
auto lod_tensor = var->GetMutable<f::LoDTensor>();
std::vector<float> send_vector(static_cast<size_t>(f::product(kDims)));
for (size_t k = 0; k < send_vector.size(); ++k) {
send_vector[k] = k + val_scalar;
}
paddle::framework::TensorFromVector<float>(
send_vector, *(ctxs_[input_scope_idx]), lod_tensor);
lod_tensor->set_lod(lod);
lod_tensor->Resize(kDims);
return send_vector;
}
std::vector<float> InitSelectedRows(const std::string& varname,
size_t input_scope_idx,
const std::vector<int64_t>& rows,
int height, float value_scalar = 0.0) {
std::vector<float> send_vector(static_cast<size_t>(f::product(kDims)));
for (size_t k = 0; k < send_vector.size(); ++k) {
send_vector[k] = k + value_scalar;
}
auto var = param_scopes_[input_scope_idx]->FindVar(varname);
PADDLE_ENFORCE_NOT_NULL(var);
auto selected_rows = var->GetMutable<f::SelectedRows>();
auto value = selected_rows->mutable_value();
value->mutable_data<float>(kDims, place_list_[input_scope_idx]);
selected_rows->set_height(height);
selected_rows->set_rows(rows);
paddle::framework::TensorFromVector<float>(
send_vector, *(ctxs_[input_scope_idx]), value);
return send_vector;
}
void SelectedRowsEqual(const std::string& varname, int input_scope_idx,
const std::vector<float>& send_vector,
const std::vector<int64_t>& rows, int height) {
auto var = param_scopes_[input_scope_idx]->FindVar(varname);
PADDLE_ENFORCE_NOT_NULL(var);
auto& selected_rows = var->Get<f::SelectedRows>();
auto rt = selected_rows.value();
PADDLE_ENFORCE_EQ(selected_rows.height(), height, "height is not equal.");
for (size_t k = 0; k < selected_rows.rows().size(); ++k) {
PADDLE_ENFORCE_EQ(selected_rows.rows()[k], rows[k]);
}
p::CPUPlace cpu_place;
f::Tensor result_tensor;
f::TensorCopySync(rt, cpu_place, &result_tensor);
float* ct = result_tensor.data<float>();
for (int64_t i = 0; i < f::product(kDims); ++i) {
ASSERT_NEAR(ct[i], send_vector[i], 1e-5);
}
}
void LoDTensorEqual(const std::string& varname,
const std::vector<float>& send_vec, const f::LoD& lod,
framework::Scope* scope) {
p::CPUPlace cpu_place;
auto var = scope->FindVar(varname);
PADDLE_ENFORCE_NOT_NULL(var);
auto tensor = var->Get<f::LoDTensor>();
PADDLE_ENFORCE_EQ(tensor.lod(), lod, "lod is not equal.");
f::Tensor result_tensor;
f::TensorCopySync(tensor, cpu_place, &result_tensor);
float* ct = result_tensor.mutable_data<float>(cpu_place);
for (int64_t k = 0; k < f::product(kDims); ++k) {
ASSERT_NEAR(ct[k], send_vec[k], 1e-5);
}
}
void TestBroadcastLodTensor(size_t input_scope_idx) {
f::LoD lod{{0, 10, 20}};
auto send_vector = InitLoDTensor("input", input_scope_idx, lod);
op_handle_->Run(false);
WaitAll();
for (size_t j = 0; j < place_list_.size(); ++j) {
LoDTensorEqual("out", send_vector, lod, param_scopes_[j]);
}
}
void TestBroadcastSelectedRows(size_t input_scope_idx) {
std::vector<int64_t> rows{0, 1, 2, 3, 3, 0, 14, 7, 3, 1,
2, 4, 6, 3, 1, 1, 1, 1, 3, 7};
int height = static_cast<int>(kDims[0] * 2);
auto send_vector = InitSelectedRows("input", input_scope_idx, rows, height);
op_handle_->Run(false);
WaitAll();
for (size_t j = 0; j < place_list_.size(); ++j) {
SelectedRowsEqual("out", input_scope_idx, send_vector, rows, height);
}
}
};
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/build_strategy.cc
浏览文件 @ ce994190
......@@ -16,6 +16,7 @@ limitations under the License. */
#include "paddle/fluid/framework/details/multi_devices_graph_check_pass.h"
#include "paddle/fluid/framework/details/multi_devices_graph_print_pass.h"
#include "paddle/fluid/framework/details/sequential_execution_pass.h"
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/graph_viz_pass.h"
......@@ -27,6 +28,10 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
public:
explicit ParallelExecutorPassBuilder(const BuildStrategy &strategy)
: ir::PassBuilder(), strategy_(strategy) {
if (strategy_.enable_sequential_execution_) {
AppendPass("sequential_execution_pass");
}
// Add a graph viz pass to record a graph.
if (!strategy_.debug_graphviz_path_.empty()) {
auto viz_pass = AppendPass("graph_viz_pass");
......@@ -64,6 +69,10 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
// Verify that the graph is correct for multi-device executor.
AppendPass("multi_devices_check_pass");
if (strategy_.remove_unnecessary_lock_) {
AppendPass("modify_op_lock_and_record_event_pass");
}
}
private:
......@@ -110,6 +119,11 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
pass->Erase("nccl_ctxs");
pass->SetNotOwned<platform::NCCLContextMap>("nccl_ctxs", nctx);
#endif
} else if (pass->Type() == "sequential_execution_pass") {
pass->Erase(kAllOpDescs);
pass->Set<const std::vector<OpDesc *>>(
kAllOpDescs,
new std::vector<OpDesc *>(main_program.Block(0).AllOps()));
}
graph = pass->Apply(std::move(graph));
}
......@@ -121,6 +135,9 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
USE_PASS(fuse_elewise_add_act_pass);
USE_PASS(graph_viz_pass);
USE_PASS(multi_batch_merge_pass);
USE_PASS(multi_devices_pass);
USE_PASS(multi_devices_check_pass);
USE_PASS(multi_devices_print_pass);
USE_PASS(sequential_execution_pass);
USE_PASS(modify_op_lock_and_record_event_pass);
paddle/fluid/framework/details/build_strategy.h
浏览文件 @ ce994190
......@@ -69,6 +69,12 @@ struct BuildStrategy {
bool enable_data_balance_{false};
bool enable_sequential_execution_{false};
bool fuse_broadcast_op_{false};
bool remove_unnecessary_lock_{false};
// User normally doesn't need to call this API.
// The PassBuilder allows for more customized insert, remove of passes
// from python side.
......
paddle/fluid/framework/details/computation_op_handle.cc
浏览文件 @ ce994190
......@@ -29,15 +29,21 @@ ComputationOpHandle::ComputationOpHandle(ir::Node *node, Scope *scope,
void ComputationOpHandle::RunImpl() {
WaitInputVarGenerated(place_);
this->RunAndRecordEvent([this] {
auto run_func = [this]() {
op_->Run(*scope_->FindVar(kLocalExecScopeName)->Get<Scope *>(), place_);
});
};
if (is_lock_and_record_event_free_) {
run_func();
} else {
this->RunAndRecordEvent(run_func);
}
}
bool ComputationOpHandle::NeedWait(VarHandleBase *in_var) {
bool need_wait =
in_var && in_var->GeneratedOp() &&
in_var->GeneratedOp()->DeviceContext(place_) != dev_ctxes_[place_];
in_var->GeneratedOp()->DeviceContext(place_) != dev_ctxes_.at(place_);
return need_wait;
}
......
paddle/fluid/framework/details/computation_op_handle.h
浏览文件 @ ce994190
......@@ -36,6 +36,8 @@ struct ComputationOpHandle : public OpHandleBase {
const platform::Place &GetPlace() const { return place_; }
void SetLockAndRecordEventFree(bool b) { is_lock_and_record_event_free_ = b; }
protected:
void RunImpl() override;
......@@ -45,6 +47,7 @@ struct ComputationOpHandle : public OpHandleBase {
std::unique_ptr<OperatorBase> op_;
Scope *scope_;
platform::Place place_;
bool is_lock_and_record_event_free_{false};
};
} // namespace details
} // namespace framework
......
paddle/fluid/framework/details/data_balance_op_handle.cc
浏览文件 @ ce994190
......@@ -28,7 +28,7 @@ DataBalanceOpHandle::DataBalanceOpHandle(
: OpHandleBase(node), local_scopes_(local_scopes), places_(places) {
if (ctxs) {
for (auto &p : places_) {
this->dev_ctxes_[p] = ctxs->DevCtx(p);
this->SetDeviceContext(p, ctxs->DevCtx(p));
}
}
}
......@@ -89,8 +89,8 @@ void DataBalanceOpHandle::RunImpl() {
PADDLE_ENFORCE_GT(places_.size(), 1,
"Data balance can only be enabled when the number of "
"places to run larger than 1.");
auto in_var_handles = DynamicCast<VarHandle>(inputs_);
auto out_var_handles = DynamicCast<VarHandle>(outputs_);
auto in_var_handles = DynamicCast<VarHandle>(this->Inputs());
auto out_var_handles = DynamicCast<VarHandle>(this->Outputs());
PADDLE_ENFORCE(in_var_handles.size() % places_.size() == 0);
PADDLE_ENFORCE_EQ(
in_var_handles.size(), out_var_handles.size(),
......
paddle/fluid/framework/details/execution_strategy.h
浏览文件 @ ce994190
......@@ -13,6 +13,7 @@
// limitations under the License.
#pragma once
#include <cstddef> // for size_t
namespace paddle {
namespace framework {
......@@ -26,6 +27,7 @@ struct ExecutionStrategy {
bool allow_op_delay_{false};
size_t num_iteration_per_drop_scope_{100};
ExecutorType type_{kDefault};
bool dry_run_{false};
};
} // namespace details
......
paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.cc
浏览文件 @ ce994190
......@@ -92,13 +92,13 @@ FeedFetchList FastThreadedSSAGraphExecutor::Run(
size_t num_complete = 0;
remaining_ = 0;
BlockingQueue<size_t> complete_q;
auto complete_q = std::make_shared<BlockingQueue<size_t>>();
for (auto op : bootstrap_ops_) {
RunOpAsync(op_deps.get(), op, &complete_q);
RunOpAsync(op_deps.get(), op, complete_q);
}
while (num_complete != op_deps->size()) {
size_t num_comp = complete_q.Pop();
size_t num_comp = complete_q->Pop();
if (num_comp == -1UL) {
int remaining = 0;
while (true) {
......@@ -107,7 +107,7 @@ FeedFetchList FastThreadedSSAGraphExecutor::Run(
break;
}
for (int i = 0; i < remaining; ++i) {
complete_q.Pop();
complete_q->Pop();
}
}
exception_.ReThrow();
......@@ -120,14 +120,17 @@ FeedFetchList FastThreadedSSAGraphExecutor::Run(
}
void FastThreadedSSAGraphExecutor::RunOpAsync(
std::unordered_map<OpHandleBase *, std::atomic<int>> *op_deps,
OpHandleBase *op, BlockingQueue<size_t> *complete_q) {
OpHandleBase *op,
const std::shared_ptr<BlockingQueue<size_t>> &complete_q) {
++remaining_;
this->pool_.enqueue([=] {
OpHandleBase *op_to_run = op;
size_t complete = 0;
while (op_to_run != nullptr) {
try {
op_to_run->Run(strategy_.use_cuda_);
if (LIKELY(!strategy_.dry_run_)) {
op_to_run->Run(strategy_.use_cuda_);
}
++complete;
} catch (...) {
exception_.Catch(std::current_exception());
......@@ -144,7 +147,7 @@ void FastThreadedSSAGraphExecutor::RunOpAsync(
if (op_to_run == nullptr) {
op_to_run = pending_op;
} else {
this->RunOpAsync(op_deps, pending_op, complete_q);
RunOpAsync(op_deps, pending_op, complete_q);
}
}
}
......@@ -156,8 +159,7 @@ void FastThreadedSSAGraphExecutor::RunOpAsync(
}
void FastThreadedSSAGraphExecutor::PrepareAtomicOpDeps() {
atomic_op_deps_ = pool_.enqueue([&] {
std::unordered_map<OpHandleBase *, std::atomic<int>> *op_deps =
new std::unordered_map<OpHandleBase *, std::atomic<int>>;
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 @ ce994190
......@@ -50,7 +50,8 @@ class FastThreadedSSAGraphExecutor : public SSAGraphExecutor {
std::atomic<int> remaining_;
void RunOpAsync(std::unordered_map<OpHandleBase *, std::atomic<int>> *op_deps,
OpHandleBase *op, BlockingQueue<size_t> *complete_q);
OpHandleBase *op,
const std::shared_ptr<BlockingQueue<size_t>> &complete_q);
void PrepareAtomicOpDeps();
......
paddle/fluid/framework/details/fused_broadcast_op_handle.cc 0 → 100644
浏览文件 @ ce994190
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/framework/details/fused_broadcast_op_handle.h"
#include "paddle/fluid/framework/details/container_cast.h"
#include "paddle/fluid/framework/details/variable_visitor.h"
#include "paddle/fluid/platform/profiler.h"
namespace paddle {
namespace framework {
namespace details {
void FusedBroadcastOpHandle::RunImpl() {
platform::RecordEvent record_event(Name(), dev_ctxes_.begin()->second);
if (places_.size() == 1UL) return;
auto in_var_handles = DynamicCast<VarHandle>(inputs_);
auto out_var_handles = DynamicCast<VarHandle>(outputs_);
WaitInputVarGenerated();
std::vector<const Scope *> var_scopes;
for (auto *s : local_scopes_) {
var_scopes.emplace_back(s->FindVar(kLocalExecScopeName)->Get<Scope *>());
}
size_t place_num = places_.size();
PADDLE_ENFORCE_EQ(in_var_handles.size() * place_num, out_var_handles.size());
for (size_t i = 0; i < in_var_handles.size(); ++i) {
BroadcastOneVar(
*in_var_handles[i],
std::vector<VarHandle *>(out_var_handles.begin() + i * place_num,
out_var_handles.begin() + (i + 1) * place_num),
var_scopes);
}
}
std::string FusedBroadcastOpHandle::Name() const { return "fused_broadcast"; }
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/fused_broadcast_op_handle.h 0 → 100644
浏览文件 @ ce994190
// 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 <map>
#include <string>
#include <vector>
#include "paddle/fluid/framework/details/broadcast_op_handle.h"
#include "paddle/fluid/framework/details/multi_devices_helper.h"
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/framework/selected_rows.h"
#include "paddle/fluid/platform/device_context.h"
#ifdef PADDLE_WITH_CUDA
#include "paddle/fluid/platform/nccl_helper.h"
#endif
namespace paddle {
namespace framework {
namespace details {
struct FusedBroadcastOpHandle : public BroadcastOpHandle {
public:
#ifdef PADDLE_WITH_CUDA
FusedBroadcastOpHandle(ir::Node *node,
const std::vector<Scope *> local_scopes,
const std::vector<platform::Place> &places,
const platform::NCCLContextMap *nccl_ctx)
: BroadcastOpHandle(node, local_scopes, places, nccl_ctx) {}
#else
FusedBroadcastOpHandle(ir::Node* node, const std::vector<Scope*> local_scopes,
const std::vector<platform::Place>& places)
: BroadcastOpHandle(node, local_scopes, places) {}
#endif
std::string Name() const override;
protected:
void RunImpl() override;
};
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/fused_broadcast_op_handle_test.cc 0 → 100644
浏览文件 @ ce994190
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/framework/details/fused_broadcast_op_handle.h"
#include "gtest/gtest.h"
#include "paddle/fluid/framework/details/broadcast_op_handle_test.h"
namespace paddle {
namespace framework {
namespace details {
struct TestFusedBroadcastOpHandle : TestBroadcastOpHandle {
std::vector<std::string> out_varnames_;
void InitFusedBroadcastOp(std::vector<size_t> input_scope_idxes) {
// initialize scope and var
for (size_t i = 0; i < place_list_.size(); ++i) {
local_scopes_.push_back(&(g_scope_.NewScope()));
Scope& local_scope = local_scopes_.back()->NewScope();
*local_scopes_.back()
->Var(details::kLocalExecScopeName)
->GetMutable<Scope*>() = &local_scope;
for (size_t j = 0; j < input_scope_idxes.size(); ++j) {
local_scope.Var("out_var" + j);
if (i == j) local_scope.Var("in_var" + j);
}
param_scopes_.emplace_back(&local_scope);
}
// create op handle node
std::unique_ptr<ir::Node> n =
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()));
#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()));
#else
op_handle_.reset(
new FusedBroadcastOpHandle(n.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);
VarHandle* in_var_handle =
new VarHandle(in_node.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]);
vars_.emplace_back(out_var_handle);
op_handle_->AddOutput(out_var_handle);
}
}
}
void TestFusedBroadcastLoDTensor(std::vector<size_t> input_scope_idxes) {
std::vector<std::vector<float>> send_vec;
f::LoD lod{{0, 10, 20}};
for (size_t i = 0; i < input_scope_idxes.size(); ++i) {
const std::string varname("in_var" + i);
float val_scalar = static_cast<float>(i);
send_vec.push_back(
InitLoDTensor(varname, input_scope_idxes[i], lod, val_scalar));
}
op_handle_->Run(false);
WaitAll();
for (size_t i = 0; i < input_scope_idxes.size(); ++i) {
const std::string& varname("out_var" + i);
for (size_t j = 0; j < place_list_.size(); ++j) {
LoDTensorEqual(varname, send_vec[i], lod, param_scopes_[j]);
}
}
}
void TestFusedBroadcastSelectedRows(std::vector<size_t> input_scope_idxes) {
std::vector<std::vector<float>> send_vector;
std::vector<int64_t> rows{0, 1, 2, 3, 3, 0, 14, 7, 3, 1,
2, 4, 6, 3, 1, 1, 1, 1, 3, 7};
int height = static_cast<int>(kDims[0] * 2);
for (size_t i = 0; i < input_scope_idxes.size(); ++i) {
const std::string varname("in_var" + i);
float val_scalar = static_cast<float>(i);
send_vector.push_back(InitSelectedRows(varname, input_scope_idxes[i],
rows, height, val_scalar));
}
op_handle_->Run(false);
WaitAll();
for (size_t i = 0; i < input_scope_idxes.size(); ++i) {
const std::string& varname("out_var" + i);
for (size_t j = 0; j < place_list_.size(); ++j) {
SelectedRowsEqual(varname, input_scope_idxes[i], send_vector[i], rows,
height);
}
}
}
};
TEST(FusedBroadcastTester, CPULodTensor) {
TestFusedBroadcastOpHandle test_op;
std::vector<size_t> input_scope_idxes = {0, 1};
test_op.InitCtxOnGpu(false);
test_op.InitFusedBroadcastOp(input_scope_idxes);
test_op.TestFusedBroadcastLoDTensor(input_scope_idxes);
}
TEST(FusedBroadcastTester, CPUSelectedRows) {
TestFusedBroadcastOpHandle test_op;
std::vector<size_t> input_scope_idxes = {0, 1};
test_op.InitCtxOnGpu(false);
test_op.InitFusedBroadcastOp(input_scope_idxes);
test_op.TestFusedBroadcastSelectedRows(input_scope_idxes);
}
#ifdef PADDLE_WITH_CUDA
TEST(FusedBroadcastTester, GPULodTensor) {
TestFusedBroadcastOpHandle test_op;
std::vector<size_t> input_scope_idxes = {0, 1};
test_op.InitCtxOnGpu(true);
test_op.InitFusedBroadcastOp(input_scope_idxes);
test_op.TestFusedBroadcastLoDTensor(input_scope_idxes);
}
TEST(FusedBroadcastTester, GPUSelectedRows) {
TestFusedBroadcastOpHandle test_op;
std::vector<size_t> input_scope_idxes = {0, 1};
test_op.InitCtxOnGpu(true);
test_op.InitFusedBroadcastOp(input_scope_idxes);
test_op.TestFusedBroadcastSelectedRows(input_scope_idxes);
}
#endif
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/gather_op_handle.cc
浏览文件 @ ce994190
......@@ -36,7 +36,7 @@ void GatherOpHandle::RunImpl() {
VarHandle *out_var_handle;
{
auto out_var_handles = DynamicCast<VarHandle>(outputs_);
auto out_var_handles = DynamicCast<VarHandle>(this->Outputs());
PADDLE_ENFORCE_EQ(out_var_handles.size(), 1,
"The number of output should be one.");
out_var_handle = out_var_handles.front();
......@@ -99,7 +99,7 @@ void GatherOpHandle::RunImpl() {
Tensor *out_tensor = out_value->mutable_value();
// copy
auto dev_ctx = dev_ctxes_[out_var_handle->place_];
auto dev_ctx = dev_ctxes_.at(out_var_handle->place_);
RunAndRecordEvent(out_var_handle->place_, [in_tensors, out_tensor, &dev_ctx,
t_out_p] {
int s = 0, e = 0;
......
paddle/fluid/framework/details/modify_op_lock_and_record_event_pass.cc 0 → 100644
浏览文件 @ ce994190
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/framework/details/modify_op_lock_and_record_event_pass.h"
#include "paddle/fluid/framework/details/computation_op_handle.h"
#include "paddle/fluid/framework/details/multi_devices_helper.h"
#include "paddle/fluid/framework/details/op_graph_view.h"
namespace paddle {
namespace framework {
namespace details {
static bool IsLockAndRecordEventFreeComputationOpHandle(
ComputationOpHandle *op, const OpGraphView &graph_view) {
if (!platform::is_gpu_place(op->GetPlace())) return false;
for (auto &pending_op : graph_view.PendingOps(op)) {
auto *tmp = dynamic_cast<ComputationOpHandle *>(pending_op);
if (tmp == nullptr || !(tmp->GetPlace() == op->GetPlace())) {
return false;
}
}
return true;
}
std::unique_ptr<ir::Graph> ModifyOpLockAndRecordEventPass::ApplyImpl(
std::unique_ptr<ir::Graph> ir_graph) const {
auto &all_ops = ir_graph->Get<GraphOps>(kGraphOps);
OpGraphView graph_view(all_ops);
for (auto &op : all_ops) {
auto *compute_op = dynamic_cast<ComputationOpHandle *>(op.get());
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();
}
}
return ir_graph;
}
} // namespace details
} // namespace framework
} // namespace paddle
REGISTER_PASS(modify_op_lock_and_record_event_pass,
paddle::framework::details::ModifyOpLockAndRecordEventPass);
paddle/fluid/framework/details/modify_op_lock_and_record_event_pass.h 0 → 100644
浏览文件 @ ce994190
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#pragma once
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/pass.h"
namespace paddle {
namespace framework {
namespace details {
class ModifyOpLockAndRecordEventPass : public ir::Pass {
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
};
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/multi_devices_graph_pass.cc
浏览文件 @ ce994190
......@@ -21,6 +21,7 @@
#include "paddle/fluid/framework/details/broadcast_op_handle.h"
#include "paddle/fluid/framework/details/computation_op_handle.h"
#include "paddle/fluid/framework/details/data_balance_op_handle.h"
#include "paddle/fluid/framework/details/fused_broadcast_op_handle.h"
#include "paddle/fluid/framework/details/multi_devices_graph_pass.h"
#include "paddle/fluid/framework/details/reduce_op_handle.h"
#include "paddle/fluid/framework/details/rpc_op_handle.h"
......@@ -252,9 +253,9 @@ std::vector<ir::Node *> SortOpsAndDelayOptimizeOp(const ir::Graph &graph) {
std::vector<ir::Node *> sorted_ret;
for (size_t i = 0; i < ret.size(); ++i) {
if (i < last_backward) {
if (boost::get<int>(ret[i]->Op()->GetAttr(
OpProtoAndCheckerMaker::OpRoleAttrName())) ==
static_cast<int>(OpRole::kOptimize)) {
if (static_cast<bool>(boost::get<int>(ret[i]->Op()->GetAttr(
OpProtoAndCheckerMaker::OpRoleAttrName())) &
static_cast<int>(OpRole::kOptimize))) {
optimize_ops.push_back(ret[i]);
} else {
sorted_ret.push_back(ret[i]);
......@@ -347,7 +348,7 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilder::ApplyImpl(
BuildStrategy::GradientScaleStrategy::kCustomized) {
// TODO(paddle-dev): Why is there no input for this op_handle?
auto loss_grad_name = node->Op()->OutputArgumentNames()[0];
CreateScaleLossGradOp(&result, loss_grad_name);
CreateScaleLossGradOp(&result, loss_grad_name, node->outputs[0]);
}
// This assumes the backward generating code will ensure IsScaleLossOp
// is true only for the op that scale the final scalar loss.
......@@ -436,10 +437,14 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilder::ApplyImpl(
if ((use_gpu &&
strategy_.reduce_ == BuildStrategy::ReduceStrategy::kReduce) ||
is_dist_train) {
for (size_t dev_id = 0; dev_id < bcast_var_name_set.size(); ++dev_id) {
auto &to_bcast_set = bcast_var_name_set[dev_id];
for (auto &bcast_name : to_bcast_set) {
CreateBroadcastOp(&result, bcast_name, dev_id);
if (strategy_.fuse_broadcast_op_) {
CreateFusedBroadcastOp(&result, bcast_var_name_set);
} else {
for (size_t dev_id = 0; dev_id < bcast_var_name_set.size(); ++dev_id) {
auto &to_bcast_set = bcast_var_name_set[dev_id];
for (auto &bcast_name : to_bcast_set) {
CreateBroadcastOp(&result, bcast_name, dev_id);
}
}
}
}
......@@ -508,6 +513,44 @@ void MultiDevSSAGraphBuilder::CreateBroadcastOp(ir::Graph *result,
}
}
void MultiDevSSAGraphBuilder::CreateFusedBroadcastOp(
ir::Graph *result,
const std::vector<std::unordered_set<std::string>> &bcast_varnames) const {
#ifdef PADDLE_WITH_CUDA
auto *op_handle = new FusedBroadcastOpHandle(
result->CreateEmptyNode("fused_broadcast", ir::Node::Type::kOperation),
local_scopes_, places_, nccl_ctxs_);
#else
auto *op_handle = new FusedBroadcastOpHandle(
result->CreateEmptyNode("fused_broadcast", ir::Node::Type::kOperation),
local_scopes_, places_);
#endif
result->Get<GraphOps>(kGraphOps).emplace_back(op_handle);
for (size_t i = 0; i < places_.size(); ++i) {
auto &p = places_[i];
SetCommunicationContext(op_handle, p);
}
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();
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];
auto &vars =
result->Get<GraphVars>(kGraphVars).at(out_dev_id).at(p_name);
auto *out_var = new VarHandle(
result->CreateEmptyNode(p_name, ir::Node::Type::kVariable),
vars.size(), out_dev_id, p_name, p);
vars.emplace_back(out_var);
op_handle->AddOutput(out_var);
}
}
}
}
void MultiDevSSAGraphBuilder::CreateComputationalOp(ir::Graph *result,
ir::Node *node,
int dev_id) const {
......@@ -602,7 +645,8 @@ int MultiDevSSAGraphBuilder::GetVarDeviceID(const ir::Graph &graph,
}
void MultiDevSSAGraphBuilder::CreateScaleLossGradOp(
ir::Graph *result, const std::string &loss_grad_name) const {
ir::Graph *result, const std::string &loss_grad_name,
ir::Node *out_var_node) const {
for (size_t i = 0; i < places_.size(); ++i) {
// Insert ScaleCost OpHandle
auto *dev_ctx = platform::DeviceContextPool::Instance().Get(places_[i]);
......@@ -617,10 +661,8 @@ void MultiDevSSAGraphBuilder::CreateScaleLossGradOp(
// loss->pending_ops_.emplace_back(op_handle);
// op_handle->inputs_.emplace_back(loss);
CreateOpOutput(
result, op_handle,
result->CreateEmptyNode(loss_grad_name, ir::Node::Type::kVariable),
places_[i], i);
CreateOpOutput(result, op_handle,
result->CreateVarNode(out_var_node->Var()), places_[i], i);
}
}
......@@ -680,7 +722,8 @@ int MultiDevSSAGraphBuilder::CreateDistTrainOp(ir::Graph *result,
}
if (node->Op()->Type() == "split_byref" ||
node->Op()->Type() == "split_selected_rows") {
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]);
if (strategy_.reduce_ == BuildStrategy::ReduceStrategy::kAllReduce) {
......
paddle/fluid/framework/details/multi_devices_graph_pass.h
浏览文件 @ ce994190
......@@ -61,7 +61,8 @@ class MultiDevSSAGraphBuilder : public ir::Pass {
size_t num_places) const;
void CreateScaleLossGradOp(ir::Graph *result,
const std::string &loss_grad_name) const;
const std::string &loss_grad_name,
ir::Node *out_var_node) const;
VarHandle *CreateReduceOp(ir::Graph *result, const std::string &og,
int dst_dev_id) const;
......@@ -78,6 +79,10 @@ class MultiDevSSAGraphBuilder : public ir::Pass {
void CreateBroadcastOp(ir::Graph *result, const std::string &p_name,
size_t src_dev_id) const;
void CreateFusedBroadcastOp(
ir::Graph *result,
const std::vector<std::unordered_set<std::string>> &bcast_varnames) const;
bool IsSparseGradient(const std::string &og) const;
size_t GetAppropriateDeviceID(
......
paddle/fluid/framework/details/op_graph_view.cc 0 → 100644
浏览文件 @ ce994190
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/framework/details/op_graph_view.h"
#include <queue>
#include <utility>
namespace paddle {
namespace framework {
namespace details {
OpGraphView::OpGraphView(
const std::vector<std::unique_ptr<OpHandleBase>> &ops) {
Build(ops);
}
void OpGraphView::Build(const std::vector<std::unique_ptr<OpHandleBase>> &ops) {
for (auto &op : ops) {
preceding_ops_[op.get()];
pending_ops_[op.get()];
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);
}
}
}
PADDLE_ENFORCE(
preceding_ops_.size() == ops.size() && pending_ops_.size() == ops.size(),
"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_) {
ret.insert(pair.first);
}
return ret;
}
bool OpGraphView::HasOp(OpHandleBase *op) const {
return preceding_ops_.count(op) != 0;
}
void OpGraphView::EnforceHasOp(OpHandleBase *op) const {
PADDLE_ENFORCE(HasOp(op), "Cannot find op %s in OpGraphView",
op == nullptr ? "nullptr" : op->DebugString());
}
const std::unordered_set<OpHandleBase *> &OpGraphView::PrecedingOps(
OpHandleBase *op) const {
EnforceHasOp(op);
return preceding_ops_.at(op);
}
const std::unordered_set<OpHandleBase *> &OpGraphView::PendingOps(
OpHandleBase *op) const {
EnforceHasOp(op);
return pending_ops_.at(op);
}
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/op_graph_view.h 0 → 100644
浏览文件 @ ce994190
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#pragma once
#include <memory>
#include <unordered_map>
#include <unordered_set>
#include <vector>
#include "paddle/fluid/framework/details/op_handle_base.h"
namespace paddle {
namespace framework {
namespace details {
class OpGraphView {
public:
explicit OpGraphView(const std::vector<std::unique_ptr<OpHandleBase>> &ops);
size_t OpNumber() const;
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 EnforceHasOp(OpHandleBase *op) const;
std::unordered_map<OpHandleBase *, std::unordered_set<OpHandleBase *>>
preceding_ops_;
std::unordered_map<OpHandleBase *, std::unordered_set<OpHandleBase *>>
pending_ops_;
};
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/op_handle_base.cc
浏览文件 @ ce994190
......@@ -103,7 +103,7 @@ void OpHandleBase::WaitInputVarGenerated() {
void OpHandleBase::WaitInputVarGenerated(const platform::Place &place) {
for (auto *in : inputs_) {
if (NeedWait(in)) {
in->GeneratedOp()->RecordWaitEventOnCtx(dev_ctxes_[place]);
in->GeneratedOp()->RecordWaitEventOnCtx(dev_ctxes_.at(place));
}
}
}
......
paddle/fluid/framework/details/reduce_op_handle.cc
浏览文件 @ ce994190
......@@ -27,7 +27,7 @@ namespace framework {
namespace details {
void ReduceOpHandle::RunImpl() {
platform::RecordEvent record_event(Name(), dev_ctxes_.begin()->second);
platform::RecordEvent record_event(Name(), dev_ctxes_.cbegin()->second);
if (places_.size() == 1) return;
// the input and output may have dummy var.
......
paddle/fluid/framework/details/reduce_op_handle.h
浏览文件 @ ce994190
......@@ -46,7 +46,8 @@ struct ReduceOpHandle : public OpHandleBase {
nccl_ctxs_(nccl_ctxs) {
if (nccl_ctxs_) {
for (auto &p_ctx : nccl_ctxs_->contexts_) {
dev_ctxes_[platform::CUDAPlace(p_ctx.first)] = p_ctx.second.ctx_.get();
this->SetDeviceContext(platform::CUDAPlace(p_ctx.first),
p_ctx.second.ctx_.get());
}
}
}
......
paddle/fluid/framework/details/reference_count_op_handle.h
浏览文件 @ ce994190
......@@ -51,7 +51,7 @@ class ReferenceCountOpHandle : public OpHandleBase {
dev_ctx_ = static_cast<platform::CUDADeviceContext *>(
platform::DeviceContextPool::Instance().Get(place));
if (IsStreamGarabageCollector()) {
PADDLE_ENFORCE(cudaSetDevice(place.device));
platform::SetDeviceId(place.device);
PADDLE_ENFORCE(cudaEventCreateWithFlags(&event_, cudaEventDisableTiming));
}
......@@ -61,7 +61,7 @@ class ReferenceCountOpHandle : public OpHandleBase {
~ReferenceCountOpHandle() {
if (IsStreamGarabageCollector()) {
auto gpu_place = boost::get<platform::CUDAPlace>(dev_ctx_->GetPlace());
PADDLE_ENFORCE(cudaSetDevice(gpu_place.device));
platform::SetDeviceId(gpu_place.device);
PADDLE_ENFORCE(cudaEventDestroy(event_));
}
}
......
paddle/fluid/framework/details/reference_count_pass.cc
浏览文件 @ ce994190
......@@ -43,6 +43,23 @@ static ComputationOpHandle *FindNextComputationOpHandle(VarHandle *var_in) {
return nullptr;
}
static void AddDependencyBetween(OpHandleBase *in, OpHandleBase *out,
ir::Graph *graph) {
auto it = std::find_if(
in->Outputs().begin(), in->Outputs().end(), [](VarHandleBase *var) {
return dynamic_cast<DummyVarHandle *>(var) != nullptr;
});
if (it != in->Outputs().end()) {
out->AddInput(*it);
} else {
auto *dep_var = new DummyVarHandle(graph->CreateControlDepVar());
graph->Get<GraphDepVars>(kGraphDepVars).emplace(dep_var);
in->AddOutput(dep_var);
out->AddInput(dep_var);
}
}
std::unique_ptr<ir::Graph> ReferenceCountPass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
auto &ref_cnts = Get<DeviceReferenceCountMap>(kGlobalReferenceCount);
......@@ -133,12 +150,7 @@ std::unique_ptr<ir::Graph> ReferenceCountPass::ApplyImpl(
auto *ref_cnt_handle = new ReferenceCountOpHandle(
ref_cnt_node, next_compute_op->GetScope(), place, {var_name},
gcs[place.device].get(), cur_ref_cnts[place.device].get());
if (next_compute_op->Outputs().empty()) {
auto *dep_var = new DummyVarHandle(graph->CreateControlDepVar());
next_compute_op->AddOutput(dep_var);
graph->Get<GraphDepVars>(kGraphDepVars).emplace(dep_var);
}
ref_cnt_handle->AddInput(next_compute_op->Outputs().front());
AddDependencyBetween(next_compute_op, ref_cnt_handle, graph.get());
compute_ref_cnt_map[next_compute_op].reset(ref_cnt_handle);
}
}
......@@ -160,12 +172,7 @@ std::unique_ptr<ir::Graph> ReferenceCountPass::ApplyImpl(
auto *ref_cnt_handle = new ReferenceCountOpHandle(
ref_cnt_node, compute_op->GetScope(), place, in_var_names,
gcs[place.device].get(), cur_ref_cnts[place.device].get());
if (compute_op->Outputs().empty()) {
auto *dep_var = new DummyVarHandle(graph->CreateControlDepVar());
compute_op->AddOutput(dep_var);
graph->Get<GraphDepVars>(kGraphDepVars).emplace(dep_var);
}
ref_cnt_handle->AddInput(compute_op->Outputs().front());
AddDependencyBetween(compute_op, ref_cnt_handle, graph.get());
compute_ref_cnt_map[compute_op].reset(ref_cnt_handle);
}
......
paddle/fluid/framework/details/rpc_op_handle.cc
浏览文件 @ ce994190
......@@ -29,22 +29,19 @@ RPCOpHandle::RPCOpHandle(ir::Node *node, const framework::OpDesc &op_desc,
place_(place) {}
void RPCOpHandle::RunImpl() {
// TODO(wuyi): need further analysis whether wait VarDummyHandle.
// Wait input done
for (auto *in : inputs_) {
auto &p = static_cast<VarHandle *>(in)->place_;
// FIXME(Yancey1989): need a better solution instead of use DebugString()
if (ir::IsControlDepVar(*in->Node())) { // HACK
if (ir::IsControlDepVar(*in->Node())) {
continue;
}
if (in->GeneratedOp()) {
in->GeneratedOp()->RecordWaitEventOnCtx(dev_ctxes_[p]);
in->GeneratedOp()->RecordWaitEventOnCtx(dev_ctxes_.at(p));
}
}
auto &tmp_scope = local_scope_->FindVar(kLocalExecScopeName)->Get<Scope *>();
// FIXME(wuyi): can not use RunAndRecordEvent here, for it will cause dead
// lock.
op_->Run(*tmp_scope, place_);
this->RunAndRecordEvent([this] {
op_->Run(*local_scope_->FindVar(kLocalExecScopeName)->Get<Scope *>(),
place_);
});
}
std::string RPCOpHandle::Name() const { return name_; }
......
paddle/fluid/framework/details/scale_loss_grad_op_handle.cc
浏览文件 @ ce994190
......@@ -27,7 +27,7 @@ ScaleLossGradOpHandle::ScaleLossGradOpHandle(ir::Node *node, size_t num_dev,
coeff_(static_cast<float>(1.0 / num_dev)),
scope_(scope),
place_(place) {
dev_ctxes_[place_] = dev_ctx;
this->SetDeviceContext(place_, dev_ctx);
}
ScaleLossGradOpHandle::~ScaleLossGradOpHandle() {}
......@@ -46,9 +46,9 @@ void ScaleLossGradOpHandle::RunImpl() {
} else {
#ifdef PADDLE_WITH_CUDA
this->RunAndRecordEvent([&] {
auto stream =
static_cast<platform::CUDADeviceContext *>(this->dev_ctxes_[place_])
->stream();
auto stream = static_cast<platform::CUDADeviceContext *>(
this->dev_ctxes_.at(place_))
->stream();
memory::Copy(boost::get<platform::CUDAPlace>(place_), tmp,
platform::CPUPlace(), &coeff_, sizeof(float), stream);
VLOG(10) << place_ << "RUN Scale loss grad op";
......
paddle/fluid/framework/details/sequential_execution_pass.cc 0 → 100644
浏览文件 @ ce994190
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/framework/details/sequential_execution_pass.h"
#include <string>
#include <unordered_map>
#include <unordered_set>
#include <vector>
#include "paddle/fluid/framework/op_proto_maker.h"
namespace paddle {
namespace framework {
namespace details {
static bool IsSameOpDesc(OpDesc *op1, OpDesc *op2) {
return op1->Type() == op2->Type() && op1->Inputs() == op2->Inputs() &&
op1->Outputs() == op2->Outputs();
}
std::unique_ptr<ir::Graph> SequentialExecutionPass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
// FIXME(zjl): Insert dependencies between some distributed ops may cause
// the multi_devices_graph_pass fails. So we skip these ops here.
// Indeed, maybe we should not insert dependencies between these ops
// casually, which may cause deadlock easily.
// We should add more skipped distributed ops when found errors in
// multi_devices_graph_pass
static std::unordered_set<std::string> skip_dist_ops{
"send", "recv", "send_barrier", "fetch_barrier"};
auto &ops = Get<const std::vector<OpDesc *>>(kAllOpDescs);
std::vector<ir::Node *> op_node_list;
op_node_list.reserve(ops.size());
std::unordered_map<ir::Node *, size_t> op_deps;
std::unordered_map<ir::Node *, std::unordered_set<ir::Node *>> pending_ops;
std::unordered_set<ir::Node *> ready_ops;
for (ir::Node *node : graph->Nodes()) {
if (!node->IsOp()) continue;
std::unordered_set<ir::Node *> preceding_ops;
for (auto *in : node->inputs) {
PADDLE_ENFORCE(in->IsVar(),
"Preceding Node of Op Nodes must be Var Node");
if (in->inputs.empty()) continue;
PADDLE_ENFORCE(in->inputs.size() == 1 && in->inputs[0]->IsOp(),
"Preceding Op Node of Var Node must be unique");
preceding_ops.insert(in->inputs[0]);
pending_ops[in->inputs[0]].insert(node);
}
op_deps[node] = preceding_ops.size();
if (preceding_ops.empty()) {
ready_ops.insert(node);
}
}
for (auto *op_desc : ops) {
ir::Node *found_node = nullptr;
for (auto *node : ready_ops) {
if (IsSameOpDesc(op_desc, node->Op())) {
PADDLE_ENFORCE(found_node == nullptr,
"Found multiple op_desc in graph: %s", op_desc->Type());
found_node = node;
}
}
PADDLE_ENFORCE_NOT_NULL(found_node, "Cannot find op_desc in graph: %s",
op_desc->Type());
for (auto *pending_op : pending_ops[found_node]) {
if (--op_deps.at(pending_op) == 0) {
ready_ops.insert(pending_op);
}
}
ready_ops.erase(found_node);
if (skip_dist_ops.count(op_desc->Type()) == 0) {
op_node_list.push_back(found_node);
}
}
for (size_t i = 1; i < op_node_list.size(); ++i) {
auto *dep_var = graph->CreateControlDepVar();
op_node_list[i]->inputs.push_back(dep_var);
op_node_list[i - 1]->outputs.push_back(dep_var);
dep_var->outputs.push_back(op_node_list[i]);
dep_var->inputs.push_back(op_node_list[i - 1]);
VLOG(10) << "Add dependencies between " << op_node_list[i - 1]->Name()
<< " and " << op_node_list[i]->Name();
}
return graph;
}
} // namespace details
} // namespace framework
} // namespace paddle
REGISTER_PASS(sequential_execution_pass,
paddle::framework::details::SequentialExecutionPass)
.RequirePassAttr(paddle::framework::details::kAllOpDescs);
paddle/fluid/framework/details/sequential_execution_pass.h 0 → 100644
浏览文件 @ ce994190
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#pragma once
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/pass.h"
namespace paddle {
namespace framework {
namespace details {
constexpr char kAllOpDescs[] = "all_op_descs";
class SequentialExecutionPass : public ir::Pass {
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
};
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/threaded_ssa_graph_executor.cc
浏览文件 @ ce994190
......@@ -39,7 +39,7 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
new platform::RecordEvent("ThreadedSSAGraphExecutorPrepare", nullptr));
std::unordered_map<OpHandleBase *, size_t> pending_ops;
std::unordered_set<VarHandleBase *> pending_vars;
BlockingQueue<VarHandleBase *> ready_vars;
auto ready_vars = std::make_shared<BlockingQueue<VarHandleBase *>>();
std::unordered_set<OpHandleBase *> ready_ops;
// For ops (e.g. nccl_all_reduce) that need to coordinate multiple
// streams from multiple GPUs, it's faster to buffer them and schedule
......@@ -51,12 +51,12 @@ 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, version_pair.get());
InsertPendingVar(&pending_vars, ready_vars.get(), version_pair.get());
}
}
}
for (auto &var : graph_->Get<details::GraphDepVars>(details::kGraphDepVars)) {
InsertPendingVar(&pending_vars, &ready_vars, var.get());
InsertPendingVar(&pending_vars, ready_vars.get(), var.get());
}
for (auto &op : graph_->Get<details::GraphOps>(details::kGraphOps)) {
......@@ -73,12 +73,12 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
FeedFetchList fetch_data(fetch_tensors.size());
InsertFetchOps(fetch_tensors, &fetch_ops, &fetch_dependencies, &pending_ops,
&pending_vars, &ready_vars, &fetch_data);
&pending_vars, ready_vars.get(), &fetch_data);
auto run_all_ops = [&](std::unordered_set<OpHandleBase *> &set) {
for (auto *op : set) {
running_ops_++;
RunOp(&ready_vars, op);
RunOp(ready_vars, op);
}
set.clear();
};
......@@ -87,7 +87,6 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
run_op_futures_.clear();
exception_holder_.Clear();
event.reset(nullptr);
// Step 3. Execution
while (!pending_vars.empty()) {
// 1. Run All Ready ops
......@@ -103,7 +102,7 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
// 2. Find ready variable
bool timeout;
auto cur_ready_vars = ready_vars.PopAll(1, &timeout);
auto cur_ready_vars = ready_vars->PopAll(1, &timeout);
if (timeout) {
if (exception_holder_.IsCaught()) {
......@@ -133,7 +132,6 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
}
}
PADDLE_ENFORCE(ready_ops.empty());
// Wait FetchOps.
ClearFetchOp(graph_.get(), &fetch_ops);
......@@ -206,13 +204,16 @@ void ThreadedSSAGraphExecutor::InsertPendingVar(
}
void ThreadedSSAGraphExecutor::RunOp(
BlockingQueue<VarHandleBase *> *ready_var_q, details::OpHandleBase *op) {
const std::shared_ptr<BlockingQueue<VarHandleBase *>> &ready_var_q,
details::OpHandleBase *op) {
auto op_run = [ready_var_q, op, this] {
try {
if (VLOG_IS_ON(10)) {
VLOG(10) << op << " " << op->Name() << " : " << op->DebugString();
}
op->Run(strategy_.use_cuda_);
if (LIKELY(!strategy_.dry_run_)) {
op->Run(strategy_.use_cuda_);
}
VLOG(10) << op << " " << op->Name() << " Done ";
running_ops_--;
ready_var_q->Extend(op->Outputs());
......
paddle/fluid/framework/details/threaded_ssa_graph_executor.h
浏览文件 @ ce994190
......@@ -48,10 +48,10 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor {
// Use topological sort algorithm
FeedFetchList Run(const std::vector<std::string> &fetch_tensors) override;
~ThreadedSSAGraphExecutor() {}
~ThreadedSSAGraphExecutor() final = default;
private:
void RunOp(BlockingQueue<VarHandleBase *> *ready_var_q,
void RunOp(const std::shared_ptr<BlockingQueue<VarHandleBase *>> &ready_var_q,
details::OpHandleBase *op);
private:
......
paddle/fluid/framework/details/var_handle.h
浏览文件 @ ce994190
......@@ -49,6 +49,8 @@ struct VarHandleBase {
void AddOutput(OpHandleBase* out, ir::Node* node) {
if (pending_ops_.find(out) == pending_ops_.end()) {
PADDLE_ENFORCE(out != nullptr, "The output of %s should not be nullptr",
this->Node()->Name());
pending_ops_.insert(out);
node_->outputs.push_back(node);
}
......
paddle/fluid/framework/executor.cc
浏览文件 @ ce994190
......@@ -85,8 +85,10 @@ Executor::Executor(const platform::Place& place) : place_(place) {}
void Executor::Close() {
#ifdef PADDLE_WITH_DISTRIBUTE
// TODO(typhoonzero): complete message will need to use real trainer_id,
// except 0.
::paddle::operators::distributed::RPCClient::GetInstance<
::paddle::operators::distributed::GRPCClient>()
::paddle::operators::distributed::GRPCClient>(0)
->SendComplete();
#endif
}
......
paddle/fluid/framework/framework.proto
浏览文件 @ ce994190
......@@ -35,6 +35,7 @@ enum AttrType {
BLOCK = 8;
LONG = 9;
BLOCKS = 10;
LONGS = 11;
}
// OpDesc describes an instance of a C++ framework::OperatorBase
......@@ -55,6 +56,7 @@ message OpDesc {
optional int32 block_idx = 12;
optional int64 l = 13;
repeated int32 blocks_idx = 14;
repeated int64 longs = 15;
};
message Var {
......@@ -80,7 +82,6 @@ message OpProto {
optional bool duplicable = 3 [ default = false ];
optional bool intermediate = 4 [ default = false ];
optional bool dispensable = 5 [ default = false ];
optional string reuse = 6;
}
// AttrProto describes the C++ type Attribute.
......
paddle/fluid/framework/ir/CMakeLists.txt
浏览文件 @ ce994190
......@@ -10,7 +10,7 @@ function(pass_library TARGET DEST)
set(oneValueArgs "")
set(multiValueArgs SRCS DEPS)
cmake_parse_arguments(op_library "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
cc_library(${TARGET} SRCS ${TARGET}.cc DEPS graph_pattern_detector pass ${op_library_DEPS})
cc_library(${TARGET} SRCS ${TARGET}.cc DEPS graph_pattern_detector pass fuse_pass_base ${op_library_DEPS})
# add more DEST here, such as train, dist and collect USE_PASS into a file automatically.
if (${DEST} STREQUAL "base" OR ${DEST} STREQUAL "inference")
message(STATUS "add pass ${TARGET} ${DEST}")
......@@ -25,20 +25,27 @@ cc_library(graph_helper SRCS graph_helper.cc DEPS graph)
cc_library(pass SRCS pass.cc DEPS graph node graph_helper)
cc_library(graph_traits SRCS graph_traits.cc DEPS graph)
cc_library(graph_pattern_detector SRCS graph_pattern_detector.cc DEPS graph graph_helper graph_traits)
cc_library(fuse_pass_base SRCS fuse_pass_base.cc DEPS pass)
pass_library(graph_to_program_pass base)
pass_library(graph_viz_pass base)
pass_library(fc_fuse_pass inference)
if (WITH_MKLDNN)
pass_library(conv_relu_mkldnn_fuse_pass inference)
endif ()
pass_library(attention_lstm_fuse_pass inference)
pass_library(infer_clean_graph_pass inference)
pass_library(fc_lstm_fuse_pass inference)
pass_library(embedding_fc_lstm_fuse_pass inference)
pass_library(fc_gru_fuse_pass inference)
pass_library(seq_concat_fc_fuse_pass inference)
pass_library(multi_batch_merge_pass base)
pass_library(conv_bn_fuse_pass inference)
pass_library(seqconv_eltadd_relu_fuse_pass inference)
if(WITH_MKLDNN)
pass_library(mkldnn_placement_pass base)
pass_library(depthwise_conv_mkldnn_pass base)
pass_library(conv_bias_mkldnn_fuse_pass inference)
pass_library(conv_relu_mkldnn_fuse_pass inference)
pass_library(conv_elementwise_add_mkldnn_fuse_pass inference)
endif()
cc_library(fuse_elewise_add_act_pass SRCS fuse_elewise_add_act_pass.cc DEPS pass graph_pattern_detector )
......@@ -53,5 +60,7 @@ cc_test(graph_to_program_pass_test SRCS graph_to_program_pass_test.cc DEPS graph
cc_test(test_graph_pattern_detector SRCS graph_pattern_detector_tester.cc DEPS graph_pattern_detector)
cc_test(test_fc_fuse_pass SRCS fc_fuse_pass_tester.cc DEPS fc_fuse_pass framework_proto)
if (WITH_MKLDNN)
cc_test(test_depthwise_conv_mkldnn_pass SRCS depthwise_conv_mkldnn_pass_tester.cc DEPS depthwise_conv_mkldnn_pass)
cc_test(test_conv_relu_mkldnn_fuse_pass SRCS conv_relu_mkldnn_fuse_pass_tester.cc DEPS conv_relu_mkldnn_fuse_pass)
cc_test(test_conv_elementwise_add_mkldnn_fuse_pass SRCS conv_elementwise_add_mkldnn_fuse_pass_tester.cc DEPS conv_elementwise_add_mkldnn_fuse_pass)
endif ()
paddle/fluid/framework/ir/attention_lstm_fuse_pass.cc
浏览文件 @ ce994190
......@@ -262,7 +262,7 @@ std::unique_ptr<ir::Graph> AttentionLSTMFusePass::ApplyImpl(
std::unordered_set<std::string> specified_vars({"data_lod_attention",
"cell_init", "hidden_init",
"data", "week", "minute"});
int count = 0;
size_t count = 0;
for (auto* node : graph->Nodes()) {
if (node->IsVar() && specified_vars.count(node->Name())) {
++count;
......
paddle/fluid/framework/ir/conv_bias_mkldnn_fuse_pass.cc 0 → 100644
浏览文件 @ ce994190
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/framework/ir/conv_bias_mkldnn_fuse_pass.h"
#include <functional>
#include <string>
#include <vector>
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/platform/enforce.h"
namespace paddle {
namespace framework {
namespace ir {
template <typename BinaryOperation>
LoDTensor tensor_apply_eltwise(const LoDTensor& vec_a, const LoDTensor& vec_b,
BinaryOperation f) {
PADDLE_ENFORCE_EQ(vec_a.dims(), vec_b.dims());
LoDTensor vec_y;
vec_y.Resize(vec_a.dims());
const float* a = vec_a.data<float>();
const float* b = vec_b.data<float>();
float* y = vec_y.mutable_data<float>(platform::CPUPlace());
for (int i = 0; i < vec_a.numel(); i++) {
y[i] = f(a[i], b[i]);
}
return vec_y;
}
std::unique_ptr<ir::Graph> ConvBiasFusePass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
PADDLE_ENFORCE(graph.get());
FusePassBase::Init(name_scope_, graph.get());
auto* scope = param_scope();
PADDLE_ENFORCE(scope);
GraphPatternDetector gpd;
auto* conv_input =
gpd.mutable_pattern()
->NewNode(patterns::PDNodeName(name_scope_, "conv_input"))
->AsInput()
->assert_is_op_input("conv2d", "Input");
patterns::ConvBias conv_bias_pattern(gpd.mutable_pattern(), name_scope_);
conv_bias_pattern(conv_input);
int found_conv_bias_count = 0;
auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
Graph* g) {
VLOG(4) << "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
GET_IR_NODE_FROM_SUBGRAPH(conv, conv, conv_bias_pattern); // CONV op
// bias
GET_IR_NODE_FROM_SUBGRAPH(eltwise_bias, eltwise_bias, conv_bias_pattern);
// output
GET_IR_NODE_FROM_SUBGRAPH(eltwise_out, eltwise_out, conv_bias_pattern);
// elementwise_add op
GET_IR_NODE_FROM_SUBGRAPH(eltwise, eltwise, conv_bias_pattern);
PADDLE_ENFORCE(subgraph.count(conv_input));
// 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";
return;
}
auto* eltwise_bias_tensor =
scope->FindVar(eltwise_bias->Name())->GetMutable<LoDTensor>();
auto input_names = conv->Op()->InputNames();
bool has_bias = std::find(input_names.begin(), input_names.end(), "Bias") !=
input_names.end();
if (has_bias && conv->Op()->Input("Bias").size() > 0) {
auto conv_bias_names = conv->Op()->Input("Bias");
// add eltwise bias to existing conv bias
PADDLE_ENFORCE_EQ(conv_bias_names.size(), 1);
auto* conv_bias_var = scope->FindVar(conv_bias_names[0]);
auto* conv_bias_tensor = conv_bias_var->GetMutable<LoDTensor>();
PADDLE_ENFORCE_EQ(conv_bias_tensor->dims(), eltwise_bias_tensor->dims());
*conv_bias_tensor = tensor_apply_eltwise(
*conv_bias_tensor, *eltwise_bias_tensor, std::plus<float>());
conv->Op()->SetOutput("Output",
std::vector<std::string>({eltwise_out->Name()}));
GraphSafeRemoveNodes(graph.get(), {eltwise, conv_out});
IR_NODE_LINK_TO(conv, eltwise_out);
} else {
// take eltwise bias as conv bias
OpDesc desc;
desc.SetInput(
"Input", std::vector<std::string>({subgraph.at(conv_input)->Name()}));
desc.SetInput("Filter", std::vector<std::string>({conv_weight->Name()}));
desc.SetInput("Bias", std::vector<std::string>({eltwise_bias->Name()}));
desc.SetOutput("Output", std::vector<std::string>({eltwise_out->Name()}));
desc.SetType("conv2d");
for (auto& attr : conv->Op()->GetAttrMap()) {
desc.SetAttr(attr.first, attr.second);
}
auto conv_bias_node = g->CreateOpNode(&desc);
IR_NODE_LINK_TO(subgraph.at(conv_input), conv_bias_node);
IR_NODE_LINK_TO(conv_weight, conv_bias_node);
IR_NODE_LINK_TO(eltwise_bias, conv_bias_node);
IR_NODE_LINK_TO(conv_bias_node, eltwise_out);
GraphSafeRemoveNodes(graph.get(), {conv, eltwise, conv_out});
}
found_conv_bias_count++;
};
gpd(graph.get(), handler);
AddStatis(found_conv_bias_count);
return graph;
}
} // namespace ir
} // namespace framework
} // namespace paddle
REGISTER_PASS(conv_bias_mkldnn_fuse_pass,
paddle::framework::ir::ConvBiasFusePass);
paddle/fluid/framework/ir/conv_bias_mkldnn_fuse_pass.h 0 → 100644
浏览文件 @ ce994190
// 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/framework/ir/fuse_pass_base.h"
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/graph_pattern_detector.h"
#include "paddle/fluid/framework/ir/pass.h"
namespace paddle {
namespace framework {
namespace ir {
/*
* Fuse the Conv and Elementwise_add to a ConvBiasOp.
*/
class ConvBiasFusePass : public FusePassBase {
public:
virtual ~ConvBiasFusePass() {}
protected:
std::unique_ptr<ir::Graph> ApplyImpl(std::unique_ptr<ir::Graph> graph) const;
const std::string name_scope_{"conv_bias_mkldnn_fuse"};
};
} // namespace ir
} // namespace framework
} // namespace paddle
paddle/fluid/framework/ir/conv_bn_fuse_pass.cc
浏览文件 @ ce994190
......@@ -126,12 +126,21 @@ std::unique_ptr<ir::Graph> ConvBNFusePass::ApplyImpl(
// conv, batch_norm,
// conv_weight, conv_out,
// bn_scale, bn_bias, bn_mean, bn_variance,
// bn_out, bn_mean_out, bn_variance_out, bn_saved_mean, bn_saved_variance
// bn_out, bn_mean_out, bn_variance_out, bn_saved_mean,
// bn_saved_variance
GET_CONV_BN_NODES(conv_bn_pattern);
// 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";
return;
}
// Create eltwise_y (conv bias) variable
VarDesc eltwise_y_in_desc(
patterns::PDNodeName(name_scope_, "eltwise_y_in"));
eltwise_y_in_desc.SetPersistable(true);
auto* eltwise_y_in_node = g->CreateVarNode(&eltwise_y_in_desc);
auto* eltwise_y_in_tensor =
scope->Var(eltwise_y_in_node->Name())->GetMutable<LoDTensor>();
......@@ -151,27 +160,59 @@ std::unique_ptr<ir::Graph> ConvBNFusePass::ApplyImpl(
*bn_mean, *bn_variance, eltwise_y_in_tensor,
epsilon);
// Create an elementwise add node
OpDesc desc;
desc.SetInput("X", std::vector<std::string>({conv_out->Name()}));
desc.SetInput("Y", std::vector<std::string>({eltwise_y_in_node->Name()}));
desc.SetOutput("Out", std::vector<std::string>({bn_out->Name()}));
desc.SetType("elementwise_add");
desc.SetAttr("axis", 1);
bool a = boost::get<bool>(conv->Op()->GetAttr("use_mkldnn"));
desc.SetAttr("use_mkldnn", a);
auto eltwise_op = g->CreateOpNode(&desc); // OpDesc will be copied.
GraphSafeRemoveNodes(graph.get(), {bn_scale, bn_bias, bn_mean, bn_variance,
batch_norm, bn_mean_out, bn_variance_out,
bn_saved_mean, bn_saved_variance});
PADDLE_ENFORCE(subgraph.count(conv_input));
IR_NODE_LINK_TO(conv_out, eltwise_op);
IR_NODE_LINK_TO(eltwise_y_in_node, eltwise_op);
IR_NODE_LINK_TO(eltwise_op, bn_out);
found_conv_bn_count++;
// with MKL-DNN fuse conv+bn into conv with bias
// without MKL-DNN fuse conv+bn into conv+elementwise_add
if (fuse_option == FUSE_MKLDNN) {
auto input_names = conv->Op()->InputNames();
bool has_bias = std::find(input_names.begin(), input_names.end(),
"Bias") != input_names.end();
if (has_bias && conv->Op()->Input("Bias").size() > 0) {
// reuse existing conv bias node
auto conv_bias_names = conv->Op()->Input("Bias");
PADDLE_ENFORCE_EQ(conv_bias_names.size(), 1);
auto* conv_bias_var = scope->FindVar(conv_bias_names[0]);
auto* conv_bias_tensor = conv_bias_var->GetMutable<LoDTensor>();
PADDLE_ENFORCE_EQ(conv_bias_tensor->dims(),
eltwise_y_in_tensor->dims());
auto eigen_conv_bias = EigenVector<float>::From(*conv_bias_tensor);
eigen_conv_bias += EigenVector<float>::From(*eltwise_y_in_tensor);
} else {
// add new conv_bias node
conv->Op()->SetInput(
"Bias", std::vector<std::string>({eltwise_y_in_node->Name()}));
IR_NODE_LINK_TO(eltwise_y_in_node, conv);
}
conv->Op()->SetOutput("Output",
std::vector<std::string>({bn_out->Name()}));
GraphSafeRemoveNodes(
graph.get(),
{conv_out, bn_scale, bn_bias, bn_mean, bn_variance, batch_norm,
bn_mean_out, bn_variance_out, bn_saved_mean, bn_saved_variance});
IR_NODE_LINK_TO(conv, bn_out);
found_conv_bn_count++;
} else { // fuse_option == FUSE_NATIVE
// create an elementwise add node.
OpDesc desc;
desc.SetInput("X", std::vector<std::string>({conv_out->Name()}));
desc.SetInput("Y", std::vector<std::string>({eltwise_y_in_node->Name()}));
desc.SetOutput("Out", std::vector<std::string>({bn_out->Name()}));
desc.SetType("elementwise_add");
desc.SetAttr("axis", 1);
auto eltwise_op = g->CreateOpNode(&desc); // OpDesc will be copied.
GraphSafeRemoveNodes(
graph.get(),
{bn_scale, bn_bias, bn_mean, bn_variance, batch_norm, bn_mean_out,
bn_variance_out, bn_saved_mean, bn_saved_variance});
IR_NODE_LINK_TO(conv_out, eltwise_op);
IR_NODE_LINK_TO(eltwise_y_in_node, eltwise_op);
IR_NODE_LINK_TO(eltwise_op, bn_out);
found_conv_bn_count++;
}
};
gpd(graph.get(), handler);
......@@ -237,7 +278,6 @@ std::unique_ptr<ir::Graph> ConvEltwiseAddBNFusePass::ApplyImpl(
{bn_scale, bn_bias, bn_mean, bn_variance, batch_norm, bn_mean_out,
bn_variance_out, bn_saved_mean, bn_saved_variance, eltwise_out});
PADDLE_ENFORCE(subgraph.count(conv_input));
IR_NODE_LINK_TO(eltwise, bn_out);
found_conv_bn_count++;
......
paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.cc 0 → 100644
浏览文件 @ ce994190
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.h"
#include <functional>
#include <utility>
#include "paddle/fluid/framework/ir/graph_traits.h"
namespace paddle {
namespace framework {
namespace ir {
namespace {
// The function keeps the graph consistent by replacing
// a node 'from' in the set of inputs nodes
// of the visited node by a node 'to'.
void CorrectGraphEdges(Graph* graph, Node* from, Node* to) {
for (auto& node : GraphTraits::DFS(*graph)) {
auto from_in_inputs =
std::find(std::begin(node.inputs), std::end(node.inputs), from);
if (from_in_inputs != std::end(node.inputs)) {
IR_NODE_LINK_TO(to, (&node));
auto inputs = node.Op()->Inputs();
using input_type = VariableNameMap::value_type;
std::for_each(std::begin(inputs), std::end(inputs),
[from, to, &node](const input_type& i) -> void {
auto param_names = i.second;
auto pi = std::find(std::begin(param_names),
std::end(param_names), from->Name());
if (pi != std::end(param_names)) {
node.Op()->SetInput(i.first, {to->Name()});
}
});
}
}
}
} // namespace
using graph_ptr = std::unique_ptr<ir::Graph>;
graph_ptr ConvElementwiseAddMKLDNNFusePass::ApplyImpl(graph_ptr graph) const {
FusePassBase::Init(name_scope_, graph.get());
GraphPatternDetector gpd;
auto pattern = gpd.mutable_pattern();
patterns::Conv conv_pattern{pattern, name_scope_};
auto conv_output = conv_pattern();
patterns::ElementwiseAdd elementwise_add_pattern{pattern, name_scope_};
elementwise_add_pattern(conv_output);
conv_output->AsIntermediate();
auto conv_op_has_bias = [](const Node& conv_op) -> std::pair<bool, Node*> {
auto bias_input_names = conv_op.Op()->Inputs();
auto bias_it = bias_input_names.find("Bias");
if (bias_it != std::end(bias_input_names)) {
bool has_bias = !bias_it->second.empty();
if (has_bias) {
auto conv_bias_names = bias_it->second;
auto conv_bias_names_it =
std::find_if(std::begin(conv_op.inputs), std::end(conv_op.inputs),
[&conv_bias_names](Node* n) -> bool {
return n->Name() == conv_bias_names[0];
});
return std::make_pair(has_bias, *conv_bias_names_it);
}
}
return std::make_pair(false, nullptr);
};
auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
Graph* g) {
GET_IR_NODE_FROM_SUBGRAPH(conv_op, conv_op, conv_pattern);
GET_IR_NODE_FROM_SUBGRAPH(conv_input, conv_input, conv_pattern);
GET_IR_NODE_FROM_SUBGRAPH(conv_filter, conv_filter, conv_pattern);
GET_IR_NODE_FROM_SUBGRAPH(conv_output, conv_output, conv_pattern);
GET_IR_NODE_FROM_SUBGRAPH(elementwise_add_op, elementwise_add_op,
elementwise_add_pattern);
GET_IR_NODE_FROM_SUBGRAPH(elementwise_add_x, elementwise_add_x,
elementwise_add_pattern);
GET_IR_NODE_FROM_SUBGRAPH(elementwise_add_out, elementwise_add_out,
elementwise_add_pattern);
if (FindFuseOption(*conv_op, *elementwise_add_op) != FUSE_MKLDNN) return;
OpDesc op_desc;
op_desc.SetType("conv2d");
op_desc.SetInput("Input", {conv_input->Name()});
op_desc.SetInput("Filter", {conv_filter->Name()});
op_desc.SetInput("ResidualData", {elementwise_add_x->Name()});
op_desc.SetOutput("Output", {conv_output->Name()});
bool has_bias;
Node* conv_bias;
std::tie(has_bias, conv_bias) = conv_op_has_bias(*conv_op);
if (has_bias) {
op_desc.SetInput("Bias", {conv_bias->Name()});
}
for (const auto& attr : conv_op->Op()->GetAttrMap()) {
op_desc.SetAttr(attr.first, attr.second);
}
op_desc.SetAttr("fuse_residual_connection", true);
auto fused_conv_op = g->CreateOpNode(&op_desc);
IR_NODE_LINK_TO(conv_input, fused_conv_op);
IR_NODE_LINK_TO(conv_filter, fused_conv_op);
IR_NODE_LINK_TO(elementwise_add_x, fused_conv_op);
IR_NODE_LINK_TO(fused_conv_op, conv_output);
if (has_bias) {
IR_NODE_LINK_TO(conv_bias, fused_conv_op);
}
CorrectGraphEdges(g, elementwise_add_out, conv_output);
GraphSafeRemoveNodes(g, {elementwise_add_out, conv_op, elementwise_add_op});
};
gpd(graph.get(), handler);
return graph;
}
} // namespace ir
} // namespace framework
} // namespace paddle
REGISTER_PASS(conv_elementwise_add_mkldnn_fuse_pass,
paddle::framework::ir::ConvElementwiseAddMKLDNNFusePass);
paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.h 0 → 100644
浏览文件 @ ce994190
// 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/framework/ir/fuse_pass_base.h"
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/graph_pattern_detector.h"
namespace paddle {
namespace framework {
namespace ir {
class ConvElementwiseAddMKLDNNFusePass : public FusePassBase {
public:
virtual ~ConvElementwiseAddMKLDNNFusePass() {}
protected:
std::unique_ptr<ir::Graph> ApplyImpl(std::unique_ptr<ir::Graph> graph) const;
const std::string name_scope_{"residual_connections_fuse_pass"};
};
} // namespace ir
} // namespace framework
} // namespace paddle
paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass_tester.cc 0 → 100644
浏览文件 @ ce994190
// 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 <gtest/gtest.h>
#include <string>
#include "paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.h"
#include "paddle/fluid/framework/ir/graph_traits.h"
namespace paddle {
namespace framework {
namespace ir {
namespace {
constexpr int nodes_removed = 3;
constexpr int nodes_added = 1;
void SetOp(ProgramDesc* prog, const std::string& type,
const std::vector<std::pair<std::string, std::string>>& inputs,
const std::pair<std::string, std::string>& output) {
auto op = prog->MutableBlock(0)->AppendOp();
op->SetType(type);
op->SetAttr("use_mkldnn", true);
for (const auto& input : inputs) {
op->SetInput(input.first, {input.second});
}
op->SetOutput(output.first, {output.second});
}
struct IsReachable {
using func = std::function<bool(const std::string&, const std::string&)>;
auto operator()(const std::unique_ptr<ir::Graph>& graph) -> func {
auto find_node = [](const std::unique_ptr<ir::Graph>& graph,
const std::string& name) -> Node* {
for (auto& node : GraphTraits::DFS(*graph)) {
if (name == node.Name()) {
return &node;
}
}
return nullptr;
};
return [&](std::string from, const std::string to) -> bool {
if (from == to) return true;
std::map<std::string, bool> visited;
for (auto& node : GraphTraits::DFS(*graph)) {
visited[node.Name()] = false;
}
visited[from] = true;
std::list<std::string> queue;
queue.push_back(from);
while (!queue.empty()) {
auto cur = find_node(graph, queue.front());
queue.pop_front();
if (cur == nullptr) return false;
for (auto n : cur->outputs) {
if (n->Name() == to) return true;
if (!visited[n->Name()]) {
visited[n->Name()] = true;
queue.push_back(n->Name());
}
}
}
return false;
};
}
};
void AssertOpsCount(const std::unique_ptr<ir::Graph>& graph) {
int conv_count = 0;
int elementwise_add_count = 0;
for (auto* node : graph->Nodes()) {
if (node->IsOp() && node->Op()->Type() == "conv2d") {
++conv_count;
}
if (node->IsOp() && node->Op()->Type() == "elementwise_add") {
++elementwise_add_count;
}
}
EXPECT_EQ(conv_count, 1);
EXPECT_EQ(elementwise_add_count, 0);
}
ProgramDesc BuildProgramDesc(const std::vector<std::string>& transient_vars,
const std::vector<std::string>& persistent_vars) {
ProgramDesc prog;
auto add_var_to_prog = [&prog](const std::string& var_name) -> VarDesc* {
auto var = prog.MutableBlock(0)->Var(var_name);
var->SetType(proto::VarType::LOD_TENSOR);
return var;
};
for (const auto& v : transient_vars) {
add_var_to_prog(v);
}
for (const auto& v : persistent_vars) {
auto var = add_var_to_prog(v);
var->SetPersistable(true);
}
return prog;
}
} // namespace
TEST(ConvElementwiseAddMKLDNNFusePass, ConvolutionWithElementwiseAddRelu) {
auto prog =
BuildProgramDesc({"a", "b", "c", "d", "e", "f"}, {"bias", "weights"});
SetOp(&prog, "conv2d",
{{"Input", "a"}, {"Bias", "bias"}, {"Filter", "weights"}},
{"Output", "b"});
SetOp(&prog, "elementwise_add", {{"X", "b"}, {"Y", "c"}}, {"Out", "d"});
SetOp(&prog, "relu", {{"X", "d"}}, {"Out", "e"});
std::unique_ptr<ir::Graph> graph(new ir::Graph(prog));
IsReachable is_reachable;
EXPECT_TRUE(is_reachable(graph)("a", "relu"));
auto pass =
PassRegistry::Instance().Get("conv_elementwise_add_mkldnn_fuse_pass");
int original_nodes_num = graph->Nodes().size();
graph = pass->Apply(std::move(graph));
int current_nodes_num = graph->Nodes().size();
EXPECT_TRUE(is_reachable(graph)("a", "relu"));
EXPECT_EQ(original_nodes_num - nodes_removed + nodes_added,
current_nodes_num);
AssertOpsCount(graph);
}
TEST(ConvElementwiseAddMKLDNNFusePass,
ConvolutionWithElementwiseAddReluNoBias) {
auto prog = BuildProgramDesc({"a", "b", "c", "d", "e"}, {"weights"});
SetOp(&prog, "conv2d", {{"Input", "a"}, {"Filter", "weights"}},
{"Output", "b"});
SetOp(&prog, "elementwise_add", {{"X", "b"}, {"Y", "c"}}, {"Out", "d"});
SetOp(&prog, "relu", {{"X", "d"}}, {"Out", "e"});
std::unique_ptr<ir::Graph> graph(new ir::Graph(prog));
IsReachable is_reachable;
EXPECT_TRUE(is_reachable(graph)("a", "relu"));
auto pass =
PassRegistry::Instance().Get("conv_elementwise_add_mkldnn_fuse_pass");
int original_nodes_num = graph->Nodes().size();
graph = pass->Apply(std::move(graph));
int current_nodes_num = graph->Nodes().size();
EXPECT_TRUE(is_reachable(graph)("a", "relu"));
EXPECT_EQ(original_nodes_num - nodes_removed + nodes_added,
current_nodes_num);
AssertOpsCount(graph);
}
TEST(ConvElementwiseAddMKLDNNFusePass, ConvolutionElementwiseAdd) {
auto prog = BuildProgramDesc({"a", "b", "c", "d"}, {"bias", "weights"});
SetOp(&prog, "conv2d",
{{"Input", "a"}, {"Bias", "bias"}, {"Filter", "weights"}},
{"Output", "b"});
SetOp(&prog, "elementwise_add", {{"X", "b"}, {"Y", "c"}}, {"Out", "d"});
std::unique_ptr<ir::Graph> graph(new ir::Graph(prog));
IsReachable is_reachable;
EXPECT_TRUE(is_reachable(graph)("a", "d"));
auto pass =
PassRegistry::Instance().Get("conv_elementwise_add_mkldnn_fuse_pass");
int original_nodes_num = graph->Nodes().size();
graph = pass->Apply(std::move(graph));
int current_nodes_num = graph->Nodes().size();
EXPECT_FALSE(is_reachable(graph)("a", "d"));
EXPECT_EQ(original_nodes_num - nodes_removed + nodes_added,
current_nodes_num);
AssertOpsCount(graph);
}
TEST(ConvElementwiseAddMKLDNNFusePass, SigmoidConvolutionAddElementwiseRelu) {
auto prog =
BuildProgramDesc({"a", "b", "c", "d", "e", "f"}, {"bias", "weights"});
SetOp(&prog, "sigmoid", {{"X", "a"}}, {"Out", "b"});
SetOp(&prog, "conv2d",
{{"Input", "b"}, {"Bias", "bias"}, {"Filter", "weights"}},
{"Output", "c"});
SetOp(&prog, "elementwise_add", {{"X", "c"}, {"Y", "d"}}, {"Out", "e"});
SetOp(&prog, "relu", {{"X", "e"}}, {"Out", "f"});
std::unique_ptr<ir::Graph> graph(new ir::Graph(prog));
IsReachable is_reachable;
EXPECT_TRUE(is_reachable(graph)("a", "f"));
auto pass =
PassRegistry::Instance().Get("conv_elementwise_add_mkldnn_fuse_pass");
int original_nodes_num = graph->Nodes().size();
graph = pass->Apply(std::move(graph));
int current_nodes_num = graph->Nodes().size();
EXPECT_TRUE(is_reachable(graph)("a", "f"));
EXPECT_EQ(original_nodes_num - nodes_removed + nodes_added,
current_nodes_num);
AssertOpsCount(graph);
}
} // namespace ir
} // namespace framework
} // namespace paddle
USE_PASS(conv_elementwise_add_mkldnn_fuse_pass);
paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.cc
浏览文件 @ ce994190
......@@ -46,6 +46,12 @@ std::unique_ptr<ir::Graph> ConvReLUFusePass::ApplyImpl(
GET_IR_NODE_FROM_SUBGRAPH(relu_out, relu_out, conv_relu_pattern); // Out
GET_IR_NODE_FROM_SUBGRAPH(relu, relu, conv_relu_pattern); // ReLU op
FuseOptions fuse_option = FindFuseOption(*conv, *relu);
if (fuse_option == DO_NOT_FUSE) {
VLOG(3) << "do not perform conv+relu fuse";
return;
}
// Transform Conv node into ConvReLU node.
OpDesc* desc = conv->Op();
desc->SetOutput("Output", std::vector<std::string>({relu_out->Name()}));
......
paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.h
浏览文件 @ ce994190
......@@ -31,7 +31,8 @@ class ConvReLUFusePass : public FusePassBase {
virtual ~ConvReLUFusePass() {}
protected:
std::unique_ptr<ir::Graph> ApplyImpl(std::unique_ptr<ir::Graph> graph) const;
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
};
} // namespace ir
......
paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass_tester.cc
浏览文件 @ ce994190
......@@ -15,25 +15,30 @@
#include "paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.h"
#include <gtest/gtest.h>
#include "paddle/fluid/framework/op_proto_maker.h"
namespace paddle {
namespace framework {
namespace ir {
void SetOp(ProgramDesc* prog, const std::string& type,
void SetOp(ProgramDesc* prog, const std::string& type, const std::string& name,
const std::vector<std::string>& inputs,
const std::vector<std::string>& outputs) {
const std::vector<std::string>& outputs, bool use_mkldnn = false) {
auto* op = prog->MutableBlock(0)->AppendOp();
op->SetType(type);
if (type == "conv2d") {
op->SetAttr("use_mkldnn", true);
op->SetAttr("use_mkldnn", use_mkldnn);
op->SetAttr("name", name);
op->SetInput("Input", {inputs[0]});
op->SetInput("Filter", {inputs[1]});
op->SetInput("Bias", {inputs[2]});
} else if (type == "relu") {
op->SetAttr("use_mkldnn", use_mkldnn);
op->SetInput("X", inputs);
}
op->SetOutput("Out", outputs);
op->SetAttr(OpProtoAndCheckerMaker::OpRoleAttrName(),
static_cast<int>(OpRole::kForward));
}
// a->OP0->b
......@@ -43,7 +48,8 @@ void SetOp(ProgramDesc* prog, const std::string& type,
ProgramDesc BuildProgramDesc() {
ProgramDesc prog;
for (auto& v :
std::vector<std::string>({"a", "b", "c", "weights", "bias", "f", "g"})) {
std::vector<std::string>({"a", "b", "c", "weights", "bias", "f", "g",
"h", "weights2", "bias2", "k", "l"})) {
auto* var = prog.MutableBlock(0)->Var(v);
var->SetType(proto::VarType::SELECTED_ROWS);
if (v == "weights" || v == "bias") {
......@@ -51,14 +57,24 @@ ProgramDesc BuildProgramDesc() {
}
}
SetOp(&prog, "OP0", std::vector<std::string>({"a"}),
SetOp(&prog, "OP0", "op0", std::vector<std::string>({"a"}),
std::vector<std::string>({"b"}));
SetOp(&prog, "OP1", std::vector<std::string>({"b"}),
SetOp(&prog, "OP1", "op1", std::vector<std::string>({"b"}),
std::vector<std::string>({"c"}));
SetOp(&prog, "conv2d", std::vector<std::string>({"c", "weights", "bias"}),
std::vector<std::string>({"f"}));
SetOp(&prog, "relu", std::vector<std::string>({"f"}),
std::vector<std::string>({"g"}));
// conv+relu, both with MKL-DNN
SetOp(&prog, "conv2d", "conv1",
std::vector<std::string>({"c", "weights", "bias"}),
std::vector<std::string>({"f"}), true);
SetOp(&prog, "relu", "relu1", std::vector<std::string>({"f"}),
std::vector<std::string>({"g"}), true);
SetOp(&prog, "OP3", "op3", std::vector<std::string>({"g"}),
std::vector<std::string>({"h"}));
// conv+relu, only one with MKL-DNN
SetOp(&prog, "conv2d", "conv2",
std::vector<std::string>({"h", "weights2", "bias2"}),
std::vector<std::string>({"k"}), true);
SetOp(&prog, "relu", "relu2", std::vector<std::string>({"k"}),
std::vector<std::string>({"l"}));
return prog;
}
......@@ -88,10 +104,16 @@ TEST(ConvReLUFusePass, basic) {
auto* op = node->Op();
ASSERT_TRUE(op->HasAttr("use_mkldnn"));
EXPECT_TRUE(boost::get<bool>(op->GetAttr("use_mkldnn")));
ASSERT_TRUE(op->HasAttr("fuse_relu"));
bool fuse_relu = boost::get<bool>(op->GetAttr("fuse_relu"));
if (fuse_relu) {
++conv_relu_count;
// check if only "conv1" convolution is fused
auto op_name = boost::get<std::string>(op->GetAttr("name"));
if (op_name == "conv1") {
ASSERT_TRUE(op->HasAttr("fuse_relu"));
bool fuse_relu = boost::get<bool>(op->GetAttr("fuse_relu"));
if (fuse_relu) {
++conv_relu_count;
}
} else if (op_name == "conv2") {
ASSERT_FALSE(op->HasAttr("fuse_relu"));
}
}
}
......
paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass.cc 0 → 100644
浏览文件 @ ce994190
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass.h"
#include "paddle/fluid/framework/ir/graph_pattern_detector.h"
namespace paddle {
namespace framework {
namespace ir {
#define GET_NODE(id, pattern) \
PADDLE_ENFORCE(subgraph.count(pattern.RetrieveNode(#id)), \
"pattern has no Node called %s", #id); \
auto* id = subgraph.at(pattern.RetrieveNode(#id)); \
PADDLE_ENFORCE_NOT_NULL(id, "subgraph has no node %s", #id);
std::unique_ptr<ir::Graph> DepthwiseConvMKLDNNPass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
PADDLE_ENFORCE(graph.get());
FusePassBase::Init("depthwise_conv_mkldnn_pass", graph.get());
GraphPatternDetector gpd;
auto* pattern = gpd.mutable_pattern();
pattern->NewNode("depthwise_conv")
->assert_is_op("depthwise_conv2d")
->assert_op_attr("use_mkldnn", true);
int found_depthwise_conv_mkldnn_count = 0;
auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
Graph* g) {
VLOG(3) << "handle DepthwiseConvMKLDNN fuse";
GET_NODE(depthwise_conv, (*pattern));
depthwise_conv->Op()->SetType("conv2d");
found_depthwise_conv_mkldnn_count++;
};
gpd(graph.get(), handler);
AddStatis(found_depthwise_conv_mkldnn_count);
return graph;
}
} // namespace ir
} // namespace framework
} // namespace paddle
REGISTER_PASS(depthwise_conv_mkldnn_pass,
paddle::framework::ir::DepthwiseConvMKLDNNPass);
paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass.h 0 → 100644
浏览文件 @ ce994190
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include "paddle/fluid/framework/ir/fuse_pass_base.h"
namespace paddle {
namespace framework {
namespace ir {
class DepthwiseConvMKLDNNPass : public FusePassBase {
public:
virtual ~DepthwiseConvMKLDNNPass() {}
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
};
} // namespace ir
} // namespace framework
} // namespace paddle
paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass_tester.cc 0 → 100644
浏览文件 @ ce994190
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass.h"
#include <gtest/gtest.h>
namespace paddle {
namespace framework {
namespace ir {
void SetOp(ProgramDesc* prog, const std::string& type, const std::string& name,
const std::vector<std::string>& inputs,
const std::vector<std::string>& outputs, bool use_mkldnn = false) {
auto* op = prog->MutableBlock(0)->AppendOp();
op->SetType(type);
op->SetAttr("use_mkldnn", use_mkldnn);
op->SetAttr("name", name);
op->SetInput("Input", {inputs[0]});
op->SetInput("Filter", {inputs[1]});
op->SetInput("Bias", {inputs[2]});
op->SetOutput("Out", outputs);
}
// (a, weights, bias)->depthwise conv mkldnn->b
// (b, weights2, bias2)->depthwise conv no mkldnn->c
// (c, weights3, bias3)->conv mkldnn->d
// (d, weights3, bias3)->conv no mkldnn->e
ProgramDesc BuildProgramDesc() {
ProgramDesc prog;
for (auto& v : std::vector<std::string>(
{"a", "b", "c", "d", "e", "weights", "bias", "weights2", "bias2",
"weights3", "bias3", "weights4", "bias4"})) {
auto* var = prog.MutableBlock(0)->Var(v);
var->SetType(proto::VarType::SELECTED_ROWS);
if (v == "weights" || v == "bias" || v == "weights2" || v == "bias2" ||
v == "weights3" || v == "bias3" || v == "weights4" || v == "bias4") {
var->SetPersistable(true);
}
}
// depthwise conv with MKL-DNN
SetOp(&prog, "depthwise_conv2d", "conv1",
std::vector<std::string>({"a", "weights", "bias"}),
std::vector<std::string>({"b"}), true);
// depthwise conv without MKL-DNN
SetOp(&prog, "depthwise_conv2d", "conv2",
std::vector<std::string>({"b", "weights2", "bias2"}),
std::vector<std::string>({"c"}), false);
// conv with MKL-DNN
SetOp(&prog, "conv2d", "conv3",
std::vector<std::string>({"c", "weights3", "bias3"}),
std::vector<std::string>({"d"}), true);
// conv without MKL-dNN
SetOp(&prog, "conv2d", "conv4",
std::vector<std::string>({"d", "weights4", "bias4"}),
std::vector<std::string>({"e"}), false);
return prog;
}
TEST(DepthwiseConvMKLDNNPass, basic) {
auto prog = BuildProgramDesc();
std::unique_ptr<ir::Graph> graph(new ir::Graph(prog));
auto pass = PassRegistry::Instance().Get("depthwise_conv_mkldnn_pass");
struct counters {
int mkldnn_depthwise_conv_nodes;
int other_depthwise_conv_nodes;
int mkldnn_conv_nodes;
int other_conv_nodes;
};
counters before{1, 1, 1, 1};
graph = pass->Apply(std::move(graph));
// initialize counters before loop
counters after{0, 0, 0, 0};
for (auto* node : graph->Nodes()) {
if (node->IsOp()) {
auto* op = node->Op();
if (op->Type() == "conv2d") {
if (boost::get<bool>(op->GetAttr("use_mkldnn")))
after.mkldnn_conv_nodes++;
else
after.other_conv_nodes++;
} else if (op->Type() == "depthwise_conv2d") {
if (boost::get<bool>(op->GetAttr("use_mkldnn")))
after.mkldnn_depthwise_conv_nodes++;
else
after.other_depthwise_conv_nodes++;
}
}
}
EXPECT_EQ(after.other_depthwise_conv_nodes,
before.other_depthwise_conv_nodes);
EXPECT_EQ(after.other_conv_nodes, before.other_conv_nodes);
EXPECT_EQ(after.mkldnn_depthwise_conv_nodes,
before.mkldnn_depthwise_conv_nodes - 1);
EXPECT_EQ(after.mkldnn_conv_nodes, before.mkldnn_conv_nodes + 1);
}
} // namespace ir
} // namespace framework
} // namespace paddle
USE_PASS(depthwise_conv_mkldnn_pass);
paddle/fluid/framework/ir/fc_fuse_pass_tester.cc
浏览文件 @ ce994190
......@@ -15,6 +15,7 @@
#include "paddle/fluid/framework/ir/fc_fuse_pass.h"
#include <gtest/gtest.h>
#include "paddle/fluid/framework/op_proto_maker.h"
namespace paddle {
namespace framework {
......@@ -32,6 +33,8 @@ void SetOp(ProgramDesc* prog, const std::string& type,
op->SetInput("X", inputs);
}
op->SetOutput("Out", outputs);
op->SetAttr(OpProtoAndCheckerMaker::OpRoleAttrName(),
static_cast<int>(OpRole::kForward));
}
// a->OP0->b
......
paddle/fluid/framework/ir/fuse_pass_base.cc 0 → 100644
浏览文件 @ ce994190
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/framework/ir/fuse_pass_base.h"
namespace paddle {
namespace framework {
namespace ir {
void FusePassBase::Init(const std::string& repr, Graph* graph) const {
repr_ = repr;
graph_ = graph;
}
Scope* FusePassBase::param_scope() const {
PADDLE_ENFORCE(graph_->Has(kParamScopeAttr));
return graph_->Get<framework::Scope*>(kParamScopeAttr);
}
void FusePassBase::AddStatis(int count_of_fused) const {
PADDLE_ENFORCE(graph_);
PADDLE_ENFORCE(!repr_.empty());
if (!graph_->Has(kFuseStatisAttr)) {
graph_->Set(kFuseStatisAttr, new std::unordered_map<std::string, int>);
}
auto& info =
graph_->Get<std::unordered_map<std::string, int>>(kFuseStatisAttr);
info[repr_] = count_of_fused;
}
FuseOptions FusePassBase::FindFuseOption(const Node& node1,
const Node& node2) const {
#ifdef PADDLE_WITH_MKLDNN
bool node1_mkldnn = node1.Op()->HasAttr("use_mkldnn") &&
boost::get<bool>(node1.Op()->GetAttr("use_mkldnn"));
bool node2_mkldnn = node2.Op()->HasAttr("use_mkldnn") &&
boost::get<bool>(node2.Op()->GetAttr("use_mkldnn"));
if (node1_mkldnn && node2_mkldnn)
return FUSE_MKLDNN;
else if (!node1_mkldnn && !node2_mkldnn)
return FUSE_NATIVE;
else
return DO_NOT_FUSE;
#else
return FUSE_NATIVE;
#endif
};
} // namespace ir
} // namespace framework
} // namespace paddle
paddle/fluid/framework/ir/fuse_pass_base.h
浏览文件 @ ce994190
......@@ -25,32 +25,24 @@ namespace ir {
static const char kParamScopeAttr[] = "__param_scope__";
static const char kFuseStatisAttr[] = "__fuse_statis__";
enum FuseOptions {
DO_NOT_FUSE, // fusing will not be done
FUSE_NATIVE, // fusing will be done without MKL-DNN
FUSE_MKLDNN // fusing will be done with MKL-DNN
};
class FusePassBase : public Pass {
public:
void Init(const std::string& repr, Graph* graph) const {
repr_ = repr;
graph_ = graph;
}
Scope* param_scope() const {
PADDLE_ENFORCE(graph_->Has(kParamScopeAttr));
return graph_->Get<framework::Scope*>(kParamScopeAttr);
}
void AddStatis(int count_of_fused) const {
PADDLE_ENFORCE(graph_);
PADDLE_ENFORCE(!repr_.empty());
if (!graph_->Has(kFuseStatisAttr)) {
graph_->Set(kFuseStatisAttr, new std::unordered_map<std::string, int>);
}
auto& info =
graph_->Get<std::unordered_map<std::string, int>>(kFuseStatisAttr);
info[repr_] = count_of_fused;
}
void Init(const std::string& repr, Graph* graph) const;
Scope* param_scope() const;
void AddStatis(int count_of_fused) const;
virtual ~FusePassBase() {}
protected:
virtual FuseOptions FindFuseOption(const Node& node1,
const Node& node2) const;
mutable Graph* graph_;
mutable std::string repr_;
};
......
paddle/fluid/framework/ir/graph.cc
浏览文件 @ ce994190
......@@ -23,80 +23,83 @@ limitations under the License. */
namespace paddle {
namespace framework {
namespace ir {
std::vector<std::string> FindDistTrainSendVars(
const std::vector<ir::Node *> &nodes) {
std::vector<std::string> send_vars;
// since parameters are all in block 0,
// it's enough to only scan send ops in block 0
for (auto &node : nodes) {
auto op_vars = node->Op()->InputArgumentNames();
send_vars.reserve(send_vars.size() +
std::distance(op_vars.begin(), op_vars.end()));
send_vars.insert(send_vars.end(), op_vars.begin(), op_vars.end());
}
return send_vars;
}
std::vector<std::string> FindDistTrainRecvVars(
const std::vector<ir::Node *> &nodes) {
std::vector<std::string> recv_vars;
for (auto &node : nodes) {
auto op_vars = node->Op()->OutputArgumentNames();
recv_vars.reserve(recv_vars.size() +
std::distance(op_vars.begin(), op_vars.end()));
recv_vars.insert(recv_vars.end(), op_vars.begin(), op_vars.end());
}
return recv_vars;
}
bool IsDistTrainOp(ir::Node *node, const std::vector<std::string> &send_vars,
const std::vector<std::string> &recv_vars) {
if (send_vars.size() == 0 || recv_vars.size() == 0) {
return false;
}
/**
* Check any of opvars contains `.block` and in sendvars
*/
auto checker = [](const std::vector<std::string> &opvars,
const std::vector<std::string> &rpc_vars) -> bool {
for (auto &var : opvars) {
// a variable name with the suffix `.block` means it's a splited
// variable by (DistributeTranspiler)
// [python/paddle/fluid/transpiler/distribute_transpiler.py]
if (var.find(".block") != std::string::npos &&
std::find(rpc_vars.begin(), rpc_vars.end(), var) != rpc_vars.end()) {
return true;
}
namespace {
void CheckProgram(const ProgramDesc &program) {
#define _INT(role) static_cast<int>(role)
std::map<int, bool> visit;
for (OpDesc *op : program.Block(0).AllOps()) {
// For backward compatibility, some program doesn't have role added.
if (!op->HasAttr(OpProtoAndCheckerMaker::OpRoleAttrName())) continue;
int role_id =
boost::get<int>(op->GetAttr(OpProtoAndCheckerMaker::OpRoleAttrName()));
visit[role_id] = true;
switch (role_id) {
case _INT(OpRole::kForward):
if (visit.find(_INT(OpRole::kBackward)) != visit.end()) {
LOG(ERROR)
<< "Cannot add backward operator before forward operator %s."
<< op->Type();
}
break;
case _INT(OpRole::kBackward):
case _INT(OpRole::kBackward) | _INT(OpRole::kLoss):
PADDLE_ENFORCE(
visit.find(_INT(OpRole::kOptimize)) == visit.end(),
"Cannot add backward operator %s after optimize operator.",
op->Type());
break;
case _INT(OpRole::kForward) | _INT(OpRole::kLoss):
PADDLE_ENFORCE(visit.find(_INT(OpRole::kBackward) |
_INT(OpRole::kLoss)) == visit.end(),
"Cannot add backward|loss operator before "
"forward|loss operator %s.",
op->Type());
PADDLE_ENFORCE(
visit.find(_INT(OpRole::kOptimize)) == visit.end(),
"Cannot add forward|loss operator %s after optimize operator.",
op->Type());
break;
case _INT(OpRole::kOptimize):
case _INT(OpRole::kOptimize) | _INT(OpRole::kLRSched):
PADDLE_ENFORCE(visit.find(_INT(OpRole::kBackward)) != visit.end(),
"Optimize operators %s must follow backward operator.",
op->Type());
break;
case _INT(OpRole::kLRSched):
case _INT(OpRole::kDist):
case _INT(OpRole::kRPC):
case _INT(OpRole::kNotSpecified):
break;
default:
LOG(FATAL) << "Unknown operator role. Don't add new role because "
"you don't know what you are doing.";
}
return false;
};
std::vector<std::string> input_var_names;
std::vector<std::string> output_var_names;
for (ir::Node *input : node->inputs) {
input_var_names.push_back(input->Name());
}
for (ir::Node *output : node->outputs) {
output_var_names.push_back(output->Name());
}
return checker(output_var_names, send_vars) ||
checker(input_var_names, recv_vars);
#undef _INT
}
} // namespace
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();
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;
for (auto *var : program.Block(0).AllVars()) {
all_vars.emplace(var->Name(), var);
}
std::map<std::string, std::vector<ir::Node *>> var_nodes;
for (auto *op : program.Block(0).AllOps()) {
ir::Node *node = CreateOpNode(op);
// For input args, reuse the same var name if it was created before.
......@@ -134,7 +137,11 @@ Graph::Graph(const ProgramDesc &program) : program_(program) {
var->inputs.push_back(node);
}
}
return std::move(var_nodes);
}
void Graph::ResolveHazard(
const std::map<std::string, std::vector<ir::Node *>> &var_nodes) {
/**
* We should handle write after read(WAR) and write after write(WAW) here.
* Because some of the operators of the program can be executed parallelly.
......@@ -153,6 +160,7 @@ Graph::Graph(const ProgramDesc &program) : program_(program) {
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();
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 @ ce994190
......@@ -160,6 +160,12 @@ class Graph {
return nullptr;
}
std::map<std::string, std::vector<ir::Node *>> InitFromProgram(
const ProgramDesc &program);
void ResolveHazard(
const std::map<std::string, std::vector<ir::Node *>> &var_nodes);
private:
// This method takes ownership of `node`.
ir::Node *AddNode(ir::Node *node) {
......
paddle/fluid/framework/ir/graph_helper.cc
浏览文件 @ ce994190
......@@ -120,19 +120,25 @@ size_t GraphNum(const Graph &graph) {
std::deque<ir::Node *> q_nodes;
std::vector<std::unordered_set<ir::Node *>> graph_nodes;
std::unordered_set<ir::Node *> g_nodes;
// q_set used to record records in the queue.
std::unordered_set<ir::Node *> q_set;
size_t graph_count = 0;
auto traverse_nodes = [&visited_nodes,
&q_nodes](const std::vector<ir::Node *> &nodes) {
std::copy_if(
nodes.begin(), nodes.end(), std::back_inserter(q_nodes),
[&visited_nodes](Node *node) { return !visited_nodes.count(node); });
auto traverse_nodes = [&visited_nodes, &q_nodes,
&q_set](const std::vector<ir::Node *> &nodes) {
for (auto n : nodes) {
if (visited_nodes.count(n) == 0 && q_set.count(n) == 0) {
q_nodes.push_back(n);
q_set.insert(n);
}
}
};
while (visited_nodes.size() != nodes.size()) {
if (!q_nodes.empty()) {
auto cur_node = q_nodes.front();
q_nodes.pop_front();
q_set.erase(cur_node);
visited_nodes.insert(cur_node);
g_nodes.insert(cur_node);
traverse_nodes(cur_node->inputs);
......@@ -146,6 +152,7 @@ size_t GraphNum(const Graph &graph) {
for (auto &n : nodes) {
if (visited_nodes.count(n) == 0) {
q_nodes.push_back(n);
q_set.insert(n);
break;
}
}
......
paddle/fluid/framework/ir/graph_helper_test.cc
浏览文件 @ ce994190
......@@ -200,15 +200,15 @@ TEST(GraphHelperTest, GraphNum) {
Graph g(prog);
BuildZeroGraph(&g);
ASSERT_EQ(GraphNum(g), 0);
ASSERT_EQ(GraphNum(g), 0UL);
Graph g2(prog);
BuildOneGraph(&g2);
ASSERT_EQ(GraphNum(g2), 1);
ASSERT_EQ(GraphNum(g2), 1UL);
Graph g3(prog);
BuildTwoGraphs(&g3);
ASSERT_EQ(GraphNum(g3), 2);
ASSERT_EQ(GraphNum(g3), 2UL);
}
} // namespace ir
......
paddle/fluid/framework/ir/graph_pattern_detector.cc
浏览文件 @ ce994190
......@@ -259,18 +259,33 @@ GraphPatternDetector::DetectPatterns() {
return result;
}
bool GraphItemCMP(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;
}
}
// TODO(Superjomn) enhance the function as it marks unique unique as duplicates
// see https://github.com/PaddlePaddle/Paddle/issues/13550
void GraphPatternDetector::UniquePatterns(
std::vector<GraphPatternDetector::subgraph_t> *subgraphs) {
if (subgraphs->empty()) return;
std::vector<GraphPatternDetector::subgraph_t> result;
std::unordered_set<size_t> set;
std::hash<std::string> hasher;
for (auto &g : *subgraphs) {
size_t key = 0;
for (auto &item : g) {
key ^= std::hash<void *>{}(item.first);
key ^= std::hash<void *>{}(item.second);
// 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::stringstream ss;
for (auto &item : sorted_keys) {
ss << item.first << ":" << item.second;
}
auto key = hasher(ss.str());
if (!set.count(key)) {
result.emplace_back(g);
set.insert(key);
......@@ -759,6 +774,51 @@ PDNode *patterns::ConvReLU::operator()(
return relu_out_var;
}
PDNode *patterns::SeqConvEltAddRelu::operator()(
paddle::framework::ir::PDNode *seqconv_input) {
// Create Operators
seqconv_input->assert_is_op_input("sequence_conv", "X");
auto *seqconv_op = pattern->NewNode(seqconv_repr())
->assert_is_op("sequence_conv")
->assert_op_attr<bool>("paddingTrainable", false)
->assert_op_attr<int>("contextStride", 1);
auto *eltadd_op =
pattern->NewNode(eltadd_repr())->assert_is_op("elementwise_add");
auto *relu_op = pattern->NewNode(relu_repr())->assert_is_op("relu");
// Create variables
// Filter
auto *seqconv_weight_var =
pattern->NewNode(seqconv_weight_repr())
->AsInput()
->assert_is_persistable_var()
->assert_is_op_input("sequence_conv", "Filter");
// Bias
auto *eltadd_bias_var = pattern->NewNode(eltadd_bias_repr())
->AsInput()
->assert_is_op_input("elementwise_add");
// intermediate variable, will be removed in the IR after fuse.
auto *seqconv_out_var = pattern->NewNode(seqconv_out_repr())
->AsIntermediate()
->assert_is_only_output_of_op("sequence_conv")
->assert_is_op_input("elementwise_add");
auto *eltadd_out_var = pattern->NewNode(eltadd_out_repr())
->AsIntermediate()
->assert_is_only_output_of_op("elementwise_add")
->assert_is_only_input_of_op("relu");
// output
auto *relu_out_var = pattern->NewNode(relu_out_repr())
->AsOutput()
->assert_is_op_output("relu");
seqconv_op->LinksFrom({seqconv_input, seqconv_weight_var})
.LinksTo({seqconv_out_var});
eltadd_op->LinksFrom({seqconv_out_var, eltadd_bias_var})
.LinksTo({eltadd_out_var});
relu_op->LinksFrom({eltadd_out_var}).LinksTo({relu_out_var});
return relu_out_var;
}
PDNode *patterns::FC::operator()(paddle::framework::ir::PDNode *x,
bool with_bias) {
// Create shared nodes.
......@@ -964,6 +1024,79 @@ PDNode *patterns::ElewiseAddActInplaceGrad::operator()(
return ele_add_grad;
}
PDNode *patterns::ConvBias::operator()(
paddle::framework::ir::PDNode *conv_input) {
// Create Operators
conv_input->assert_is_op_input("conv2d", "Input");
auto *conv_op = pattern->NewNode(conv_repr())->assert_is_op("conv2d");
auto *eltiwse_op =
pattern->NewNode(eltwise_repr())->assert_is_op("elementwise_add");
// Create variables
// Filter
auto *conv_weight_var = pattern->NewNode(conv_weight_repr())
->AsInput()
->assert_is_persistable_var()
->assert_is_op_input("conv2d", "Filter");
// intermediate variable, will be removed in the IR after fuse.
auto *conv_out_var = pattern->NewNode(conv_out_repr())
->AsIntermediate()
->assert_is_only_output_of_op("conv2d")
->assert_is_op_input("elementwise_add");
// Bias stored in elementwise_add
auto *eltwise_bias_var = pattern->NewNode(eltwise_bias_repr())
->AsInput()
->assert_is_persistable_var()
->assert_is_op_input("elementwise_add", "Y");
// output
auto *eltwise_out_var = pattern->NewNode(eltwise_out_repr())
->AsOutput()
->assert_is_op_output("elementwise_add");
conv_op->LinksFrom({conv_input, conv_weight_var}).LinksTo({conv_out_var});
eltiwse_op->LinksFrom({conv_out_var, eltwise_bias_var})
.LinksTo({eltwise_out_var});
return eltwise_out_var;
}
PDNode *patterns::Conv::operator()() {
auto conv_op = pattern->NewNode(conv_op_repr())->assert_is_op("conv2d");
auto input_var = pattern->NewNode(conv_input_repr())
->AsInput()
->assert_is_op_input("conv2d", "Input");
auto filter_var = pattern->NewNode(conv_filter_repr())
->AsInput()
->assert_is_op_input("conv2d", "Filter");
auto output_var = pattern->NewNode(conv_output_repr())
->AsOutput()
->assert_is_op_output("conv2d", "Output");
conv_op->LinksFrom({input_var, filter_var});
conv_op->LinksTo({output_var});
return output_var;
}
PDNode *patterns::ElementwiseAdd::operator()(PDNode *x_var) {
auto elementwise_add_op = pattern->NewNode(elementwise_add_op_repr())
->assert_is_op("elementwise_add");
x_var->assert_is_op_input("elementwise_add", "X");
auto y_var = pattern->NewNode(elementwise_add_x_repr())
->AsInput()
->assert_is_op_input("elementwise_add", "Y");
auto out_var = pattern->NewNode(elementwise_add_out_repr())
->AsOutput()
->assert_is_op_output("elementwise_add", "Out");
elementwise_add_op->LinksFrom({x_var, y_var});
elementwise_add_op->LinksTo({out_var});
return out_var;
}
} // namespace ir
} // namespace framework
} // namespace paddle
paddle/fluid/framework/ir/graph_pattern_detector.h
浏览文件 @ ce994190
......@@ -128,6 +128,15 @@ struct PDNode {
const std::unordered_set<std::string>& op_types,
const std::string& argument, int nth);
template <typename T>
PDNode* assert_op_attr(const std::string& attr_name, const T& attr) {
asserts_.emplace_back([=](Node* x) {
return x && x->IsOp() && x->Op()->HasAttr(attr_name) &&
boost::get<T>(x->Op()->GetAttr(attr_name)) == attr;
});
return this;
}
private:
PDNode(PDPattern* pattern, const std::string& name = "",
Type type = Type::kVar)
......@@ -434,6 +443,31 @@ struct ConvReLU : public PatternBase {
PATTERN_DECL_NODE(relu_out);
};
// SEQCONV with Elementwise_Add ReLU
// op: seqconv + elementwise_add + relu
// named nodes:
// seqconv_input, seqconv_weight,
// seqconv_out, seqconv,
// elementwise_add_bias, elementwise_add_out, elementwise_add
// relu_out, relu
struct SeqConvEltAddRelu : public PatternBase {
SeqConvEltAddRelu(PDPattern* pattern, const std::string& name_scope)
: PatternBase(pattern, name_scope, "seqconv_eltadd_relu") {}
PDNode* operator()(PDNode* seqconv_input);
// declare operator node's name
PATTERN_DECL_NODE(seqconv);
PATTERN_DECL_NODE(eltadd);
PATTERN_DECL_NODE(relu);
// declare variable node's name
PATTERN_DECL_NODE(seqconv_weight);
PATTERN_DECL_NODE(seqconv_out);
PATTERN_DECL_NODE(eltadd_bias);
PATTERN_DECL_NODE(eltadd_out);
PATTERN_DECL_NODE(relu_out);
};
// FC with bias
// op: mul + elementwise_add
// named nodes:
......@@ -578,6 +612,65 @@ struct ElewiseAddActInplaceGrad : public PatternBase {
PATTERN_DECL_NODE(d_ele_y);
PATTERN_DECL_NODE(ele_y);
};
// Conv with Elementwise_add as bias
// op: conv + elementwise_add
// named nodes:
// conv_input, conv_weight,
// conv_out, conv,
// eltwise_bias, eltwise_out,
// elementwise_add
struct ConvBias : public PatternBase {
ConvBias(PDPattern* pattern, const std::string& name_scope)
: PatternBase(pattern, name_scope, "conv_bias") {}
PDNode* operator()(PDNode* conv_input);
// declare operator node's name
PATTERN_DECL_NODE(conv);
PATTERN_DECL_NODE(eltwise);
// declare variable node's name
PATTERN_DECL_NODE(conv_weight);
PATTERN_DECL_NODE(conv_out);
PATTERN_DECL_NODE(eltwise_bias);
PATTERN_DECL_NODE(eltwise_out);
};
// Convolution op
// Forward pass for convolution.
// conv_input, conv_bias and conv_filter are inputs.
// conv_output is a result of the operator.
// residual_data is data used by skip connection.
// If residual connection fusion is on, the formula is:
// conv_output = conv_op(conv_filter, conv_input, conv_bias)
// + conv_residual_data
// If the fusion is off, conv_residual_data is not added.
struct Conv : public PatternBase {
Conv(PDPattern* pattern, const std::string& name_scope)
: PatternBase(pattern, name_scope, "convolution") {}
PDNode* operator()();
PATTERN_DECL_NODE(conv_op);
PATTERN_DECL_NODE(conv_input);
PATTERN_DECL_NODE(conv_filter);
PATTERN_DECL_NODE(conv_residual_data);
PATTERN_DECL_NODE(conv_output);
};
// ElementwiseAdd used in residual connections.
// y_var is used and convolution output.
// The operator is removed, when residual
// connection fusion is on.
struct ElementwiseAdd : public PatternBase {
ElementwiseAdd(PDPattern* pattern, const std::string& name_scope)
: PatternBase(pattern, name_scope, "elementwise_add") {}
PDNode* operator()(PDNode* x_var);
PATTERN_DECL_NODE(elementwise_add_op);
PATTERN_DECL_NODE(elementwise_add_x);
PATTERN_DECL_NODE(elementwise_add_y);
PATTERN_DECL_NODE(elementwise_add_out);
};
} // namespace patterns
// Link two ir::Nodes from each other.
......
paddle/fluid/framework/ir/graph_test.cc
浏览文件 @ ce994190
......@@ -124,7 +124,7 @@ TEST(GraphTest, Basic) {
ASSERT_EQ(n->outputs.size(), 0UL);
}
}
ASSERT_EQ(nodes.size(), 5);
ASSERT_EQ(nodes.size(), 5UL);
}
TEST(GraphTest, WriteAfterRead) {
......
paddle/fluid/framework/ir/mkldnn_placement_pass.cc 0 → 100644
浏览文件 @ ce994190
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/framework/ir/mkldnn_placement_pass.h"
namespace paddle {
namespace framework {
namespace ir {
std::unique_ptr<ir::Graph> MKLDNNPlacementPass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
VLOG(3) << "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);
}
}
return graph;
}
} // namespace ir
} // namespace framework
} // namespace paddle
REGISTER_PASS(mkldnn_placement_pass,
paddle::framework::ir::MKLDNNPlacementPass);
paddle/fluid/framework/ir/mkldnn_placement_pass.h 0 → 100644
浏览文件 @ ce994190
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include "paddle/fluid/framework/ir/pass.h"
namespace paddle {
namespace framework {
namespace ir {
class MKLDNNPlacementPass : public Pass {
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
};
} // namespace ir
} // namespace framework
} // namespace paddle
paddle/fluid/framework/ir/multi_batch_merge_pass.cc 0 → 100644
浏览文件 @ ce994190
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/framework/ir/multi_batch_merge_pass.h"
#include <map>
#include <string>
#include <vector>
#include "paddle/fluid/framework/ir/graph_helper.h"
#include "paddle/fluid/framework/op_proto_maker.h"
namespace paddle {
namespace framework {
namespace ir {
static const char kNumRepeats[] = "num_repeats";
typedef std::unordered_map<std::string, std::vector<ir::Node*>> SSAVarList;
ir::Node* SameNameVar(std::unordered_set<ir::Node*> all, ir::Node* target) {
for (auto n : all) {
if (target->IsVar() && target->Name() == n->Name()) {
return n;
}
}
return nullptr;
}
VarDesc CopyVarDesc(VarDesc* var_desc) {
VarDesc repeated_var(var_desc->Name());
// copy other variable attributes
if (var_desc->GetType() != proto::VarType::READER) {
repeated_var.SetType(var_desc->GetType());
repeated_var.SetShape(var_desc->GetShape());
repeated_var.SetDataType(var_desc->GetDataType());
repeated_var.SetLoDLevel(var_desc->GetLoDLevel());
repeated_var.SetPersistable(var_desc->Persistable());
} else {
// TODO(typhoonzero): copy reader var
}
return repeated_var;
}
VarDesc UpdateGradVarDesc(
VarDesc* var_desc, int repeat,
const std::unordered_set<std::string>& grad_names,
const std::unordered_set<std::string>& bn_vars_need_rename) {
if (grad_names.find(var_desc->Name()) != grad_names.end() ||
bn_vars_need_rename.find(var_desc->Name()) != bn_vars_need_rename.end()) {
std::string new_gname =
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;
return repeated_var;
}
return *var_desc;
}
std::unique_ptr<Graph> BatchMergePass::ApplyImpl(
std::unique_ptr<Graph> graph) const {
int num_repeats = Get<const int>(kNumRepeats);
std::vector<Node*> forward_backward_ops;
std::vector<Node*> optimize_ops;
std::vector<Node*> lr_ops; // ops other than forward/backward/optimize
std::unordered_set<std::string> grad_names;
std::vector<ir::Node*> nodes = TopologySortOperations(*graph);
auto origin_nodes = graph->ReleaseNodes();
VLOG(3) << "origin nodes count: " << origin_nodes.size();
ir::Graph& result = *graph;
// 1. record op nodes of different roles
for (auto node : nodes) {
if (node->IsVar()) continue;
int op_role = boost::get<int>(node->Op()->GetAttr(
framework::OpProtoAndCheckerMaker::OpRoleAttrName()));
if ((op_role == static_cast<int>(framework::OpRole::kForward)) ||
(op_role & static_cast<int>(framework::OpRole::kBackward)) ||
(op_role & static_cast<int>(framework::OpRole::kLoss))) {
forward_backward_ops.push_back(node);
} else if ((op_role & static_cast<int>(framework::OpRole::kOptimize)) ||
(op_role & static_cast<int>(framework::OpRole::kDist)) ||
(op_role & static_cast<int>(framework::OpRole::kRPC))) {
optimize_ops.push_back(node);
auto op_role_var = node->Op()->GetNullableAttr(
OpProtoAndCheckerMaker::OpRoleVarAttrName());
auto op_role_vars = boost::get<std::vector<std::string>>(op_role_var);
for (size_t i = 0; i < op_role_vars.size(); i += 2) {
grad_names.insert(op_role_vars[i + 1]);
}
} else if (op_role & static_cast<int>(framework::OpRole::kLRSched)) {
lr_ops.push_back(node);
} else { // NOLINT
PADDLE_THROW("Invalid op_role: %d", static_cast<int>(op_role));
}
}
// 2. copy forward backward
ir::Node* prev_repeat_last_op_node = nullptr;
// record origin_grad -> repeated grad list map.
std::map<ir::Node*, std::vector<ir::Node*>> grad_repeated_map;
std::map<std::string, std::vector<ir::Node*>> created;
std::unordered_set<std::string> bn_vars_need_rename;
for (int i = 0; i < num_repeats; ++i) {
std::unordered_set<ir::Node*> copied;
for (size_t node_idx = 0; node_idx < forward_backward_ops.size();
++node_idx) {
auto node = forward_backward_ops[node_idx];
OpDesc repeated_op(*(node->Op()), node->Op()->Block());
// 3. rename grad outputs to current repeat.
for (auto outname : repeated_op.OutputArgumentNames()) {
if (grad_names.find(outname) != grad_names.end()) {
std::string new_gname = string::Sprintf("%s.repeat.%d", outname, i);
repeated_op.RenameOutput(outname, new_gname);
}
}
// 3.5 let batch_norm ops use independent vars, note batch_norm_grad do
// not need this update
if (node->Name() == "batch_norm") {
// NOTE: assume bn op created by layers use save var as output mean and
// variance
std::string new_mean_name =
string::Sprintf("%s.repeat.%d", repeated_op.Input("Mean")[0], i);
std::string new_var_name = string::Sprintf(
"%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;
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],
new_mean_name);
repeated_op.RenameOutput(repeated_op.Output("VarianceOut")[0],
new_var_name);
}
// 3.9 do copy
auto repeated_node = result.CreateOpNode(&repeated_op);
copied.insert(node);
// 4. add deps between repeats
if (node_idx == forward_backward_ops.size() - 1) {
prev_repeat_last_op_node = repeated_node;
}
if (node_idx == 0 && prev_repeat_last_op_node) {
auto* depvar = result.CreateControlDepVar();
prev_repeat_last_op_node->outputs.push_back(depvar);
depvar->inputs.push_back(prev_repeat_last_op_node);
repeated_node->inputs.push_back(depvar);
depvar->outputs.push_back(repeated_node);
}
for (auto in_node : node->inputs) {
if (in_node->IsCtrlVar()) {
continue;
}
ir::Node* var = nullptr;
auto updated_var = UpdateGradVarDesc(in_node->Var(), i, grad_names,
bn_vars_need_rename);
// should be initialized by startup, how to initilize tensor in the
// scope?
if (node->Name() == "batch_norm" &&
bn_vars_need_rename.find(in_node->Name()) !=
bn_vars_need_rename.end()) {
// Create bn mean/variance for each repeat
var = result.CreateVarNode(&updated_var);
created[updated_var.Name()].push_back(var);
copied.insert(in_node);
repeated_node->inputs.push_back(var);
var->outputs.push_back(repeated_node);
continue;
}
// for other ops
if (in_node->inputs.empty() && i > 0) {
// do not copy head vars (inputs, params) in repeats > 0
var = created.at(in_node->Name()).back();
} else {
if (copied.find(in_node) == copied.end()) {
var = result.CreateVarNode(&updated_var);
if (grad_names.find(in_node->Var()->Name()) != grad_names.end()) {
grad_repeated_map[in_node].push_back(var);
}
copied.insert(in_node);
created[updated_var.Name()].push_back(var);
} else {
var = created.at(updated_var.Name()).back();
}
}
repeated_node->inputs.push_back(var);
var->outputs.push_back(repeated_node);
}
for (auto out_node : node->outputs) {
if (out_node->IsCtrlVar()) {
continue;
}
ir::Node* var = nullptr;
auto updated_var = UpdateGradVarDesc(out_node->Var(), i, grad_names,
bn_vars_need_rename);
if (copied.find(out_node) == copied.end()) {
var = result.CreateVarNode(&updated_var);
if (grad_names.find(out_node->Var()->Name()) != grad_names.end()) {
grad_repeated_map[out_node].push_back(var);
}
copied.insert(out_node);
created[updated_var.Name()].push_back(var);
} else {
var = created.at(updated_var.Name()).back();
}
repeated_node->outputs.push_back(var);
var->inputs.push_back(repeated_node);
}
}
}
// 5. create GRAD merge op node
for (auto kv : grad_repeated_map) {
OpDesc sum_op;
sum_op.SetType("sum");
std::vector<std::string> repeated_grad_names;
for (auto r : kv.second) {
repeated_grad_names.push_back(r->Var()->Name());
}
sum_op.SetInput("X", repeated_grad_names);
sum_op.SetOutput("Out", {kv.first->Var()->Name()});
sum_op.SetAttr(OpProtoAndCheckerMaker::OpRoleAttrName(),
static_cast<int>(OpRole::kBackward));
auto sum_op_node = result.CreateOpNode(&sum_op);
for (auto r : kv.second) {
sum_op_node->inputs.push_back(r);
r->outputs.push_back(sum_op_node);
}
auto sum_out_var_node = result.CreateVarNode(kv.first->Var());
sum_op_node->outputs.push_back(sum_out_var_node);
sum_out_var_node->inputs.push_back(sum_op_node);
created[sum_out_var_node->Name()].push_back(sum_out_var_node);
OpDesc scale_op;
scale_op.SetType("scale");
scale_op.SetInput("X", {sum_out_var_node->Var()->Name()});
// NOTE: inplace scale.
scale_op.SetOutput("Out", {sum_out_var_node->Var()->Name()});
scale_op.SetAttr("scale", static_cast<float>(1.0f / num_repeats));
scale_op.SetAttr(OpProtoAndCheckerMaker::OpRoleAttrName(),
static_cast<int>(OpRole::kBackward));
auto scale_op_node = result.CreateOpNode(&scale_op);
scale_op_node->inputs.push_back(sum_out_var_node);
sum_out_var_node->outputs.push_back(scale_op_node);
auto scale_out_var_node = result.CreateVarNode(sum_out_var_node->Var());
scale_op_node->outputs.push_back(scale_out_var_node);
scale_out_var_node->inputs.push_back(scale_op_node);
created[scale_out_var_node->Name()].push_back(scale_out_var_node);
}
// 6. add optimize ops
{
auto copy_node = [&result, &created](ir::Node* node) {
auto op_node = result.CreateOpNode(node->Op());
// copy op ins/outs
// NOTE: for send/recv ops, the OpDesc uses ctrldepvar to describe
// dependencies, so create those depvars if OpDesc have in/outs.
for (auto in_node : node->inputs) {
if (in_node->IsCtrlVar() && !in_node->Var()) {
continue;
}
ir::Node* var = nullptr;
if (created.find(in_node->Name()) == created.end()) {
var = result.CreateVarNode(in_node->Var());
created[in_node->Name()].push_back(var);
} else {
var = created.at(in_node->Name()).back();
}
op_node->inputs.push_back(var);
var->outputs.push_back(op_node);
}
for (auto out_node : node->outputs) {
if (out_node->IsCtrlVar() && !out_node->Var()) {
continue;
}
auto var = result.CreateVarNode(out_node->Var());
created[out_node->Name()].push_back(var);
op_node->outputs.push_back(var);
var->inputs.push_back(op_node);
}
};
for (auto node : lr_ops) {
copy_node(node);
}
for (auto node : optimize_ops) {
copy_node(node);
}
}
result.ResolveHazard(created);
return graph;
}
} // namespace ir
} // namespace framework
} // namespace paddle
REGISTER_PASS(multi_batch_merge_pass, paddle::framework::ir::BatchMergePass)
.RequirePassAttr(paddle::framework::ir::kNumRepeats);
paddle/fluid/framework/ir/multi_batch_merge_pass.h 0 → 100644
浏览文件 @ ce994190
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#pragma once
#include "paddle/fluid/framework/ir/fuse_pass_base.h"
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/pass.h"
namespace paddle {
namespace framework {
namespace ir {
// BatchMergePass is used to copy forward and backward ops for several
// times to run several batches to simulate large batch size training
// as if we have more than 1 GPUs.
// User can define how many batches to run, gradients will be merged
// through those repeats, and then do optimization using merged gradients.
// This pass is extremely useful when doing large batch-size distributed
// sync training, we can simulate even large batch size as if we have more
// GPUs.
class BatchMergePass : public Pass {
public:
virtual ~BatchMergePass() {}
protected:
std::unique_ptr<Graph> ApplyImpl(std::unique_ptr<Graph> graph) const override;
};
} // namespace ir
} // namespace framework
} // namespace paddle
paddle/fluid/framework/ir/node.h
浏览文件 @ ce994190
......@@ -44,6 +44,7 @@ class Node {
return op_desc_.get();
}
// Please don't use this API!
int id() const { return id_; }
bool IsOp() const { return type_ == Type::kOperation; }
......@@ -92,6 +93,7 @@ class Node {
Node() = delete;
static int count_;
// Please don't use this API or make this public.
static void ResetId() { count_ = 0; }
DISABLE_COPY_AND_ASSIGN(Node);
};
......
paddle/fluid/framework/ir/seqconv_eltadd_relu_fuse_pass.cc 0 → 100644
浏览文件 @ ce994190
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/framework/ir/seqconv_eltadd_relu_fuse_pass.h"
#include <string>
#include "paddle/fluid/framework/lod_tensor.h"
namespace paddle {
namespace framework {
namespace ir {
int BuildFusion(Graph* graph, const std::string& name_scope, Scope* scope) {
GraphPatternDetector gpd;
auto* pattern = gpd.mutable_pattern();
PDNode* x = pattern->NewNode(patterns::PDNodeName(name_scope, "X"))
->assert_is_op_input("sequence_conv")
->assert_var_not_persistable();
patterns::SeqConvEltAddRelu fuse_pattern(pattern, name_scope);
fuse_pattern(x);
// Create New OpDesc
auto fuse_creator = [&](Node* seqconv, Node* input, Node* seqconv_weight,
Node* eltadd_bias, Node* relu_out) {
OpDesc op_desc;
op_desc.SetType("fusion_seqconv_eltadd_relu");
op_desc.SetInput("X", {input->Name()});
op_desc.SetInput("Filter", {seqconv_weight->Name()});
op_desc.SetInput("Bias", {eltadd_bias->Name()});
op_desc.SetAttr("contextLength", seqconv->Op()->GetAttr("contextLength"));
op_desc.SetAttr("contextStart", seqconv->Op()->GetAttr("contextStart"));
op_desc.SetAttr("contextStride", seqconv->Op()->GetAttr("contextStride"));
PADDLE_ENFORCE(graph->Has(kParamScopeAttr));
auto* scope = graph->Get<Scope*>(kParamScopeAttr);
const std::string ColMat = patterns::UniqueKey("SeqConvColMat");
op_desc.SetOutput("ColMat", {ColMat});
op_desc.SetOutput("Out", {relu_out->Name()});
scope->Var(ColMat)->GetMutable<LoDTensor>();
auto* op = graph->CreateOpNode(&op_desc);
IR_NODE_LINK_TO(input, op);
IR_NODE_LINK_TO(seqconv_weight, op);
IR_NODE_LINK_TO(eltadd_bias, op);
IR_NODE_LINK_TO(op, relu_out);
return op;
};
int fusion_count{0};
auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
Graph* g) {
VLOG(4) << "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);
GET_IR_NODE_FROM_SUBGRAPH(eltadd, eltadd, fuse_pattern);
GET_IR_NODE_FROM_SUBGRAPH(eltadd_bias, eltadd_bias, fuse_pattern);
GET_IR_NODE_FROM_SUBGRAPH(eltadd_out, eltadd_out, fuse_pattern);
GET_IR_NODE_FROM_SUBGRAPH(relu, relu, fuse_pattern);
GET_IR_NODE_FROM_SUBGRAPH(relu_out, relu_out, fuse_pattern);
fuse_creator(seqconv, subgraph.at(x), seqconv_weight, eltadd_bias,
relu_out);
std::unordered_set<const Node*> marked_nodes(
{seqconv, seqconv_out, eltadd, eltadd_out, relu});
GraphSafeRemoveNodes(graph, marked_nodes);
++fusion_count;
};
gpd(graph, handler);
return fusion_count;
}
std::unique_ptr<ir::Graph> SeqConvEltAddReluFusePass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
FusePassBase::Init(name_scope_, graph.get());
int fusion_count = BuildFusion(graph.get(), name_scope_, param_scope());
AddStatis(fusion_count);
return graph;
}
} // namespace ir
} // namespace framework
} // namespace paddle
REGISTER_PASS(seqconv_eltadd_relu_fuse_pass,
paddle::framework::ir::SeqConvEltAddReluFusePass);
paddle/fluid/framework/ir/seqconv_eltadd_relu_fuse_pass.h 0 → 100644
浏览文件 @ ce994190
// 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/framework/ir/fuse_pass_base.h"
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/graph_pattern_detector.h"
namespace paddle {
namespace framework {
namespace ir {
class SeqConvEltAddReluFusePass : public FusePassBase {
public:
virtual ~SeqConvEltAddReluFusePass() {}
protected:
std::unique_ptr<ir::Graph> ApplyImpl(std::unique_ptr<ir::Graph> graph) const;
const std::string name_scope_{"seqconv_eltadd_relu_fuse"};
};
} // namespace ir
} // namespace framework
} // namespace paddle
paddle/fluid/framework/lod_tensor.cc
浏览文件 @ ce994190
......@@ -418,7 +418,7 @@ void LoDTensor::MergeLoDTensor(
PADDLE_ENFORCE_EQ(new_lod.size(), lod.size());
for (size_t j = 0; j < lod.size(); ++j) {
auto &sub_lod = new_lod[j];
auto &offset = sub_lod.back();
size_t offset = sub_lod.back();
for (size_t k = 1; k < lod[j].size(); ++k) {
sub_lod.push_back(lod[j][k] + offset);
}
......
paddle/fluid/framework/lod_tensor_array.h
浏览文件 @ ce994190
......@@ -18,6 +18,8 @@ limitations under the License. */
namespace paddle {
namespace framework {
using LoDTensorArray = std::vector<LoDTensor>;
}
} // namespace framework
} // namespace paddle
paddle/fluid/framework/mixed_vector.h
浏览文件 @ ce994190
......@@ -542,6 +542,33 @@ class CPUVector : public std::vector<T, std::allocator<T>> {
this->reserve(this->size() + size_t(end - begin));
this->insert(this->end(), begin, end);
}
const T *CUDAData(platform::Place place) const {
PADDLE_THROW(
"Vector::CUDAData() method is not supported in CPU-only version");
}
T *CUDAMutableData(platform::Place place) {
PADDLE_THROW(
"Vector::CUDAMutableData() method is not supported in CPU-only "
"version");
}
const T *Data(platform::Place place) const {
PADDLE_ENFORCE(
platform::is_cpu_place(place),
"Vector::Data() method is not supported when not in CPUPlace");
return this->data();
}
T *MutableData(platform::Place place) {
PADDLE_ENFORCE(
platform::is_cpu_place(place),
"Vector::MutableData() method is not supported when not in CPUPlace");
return this->data();
}
const void *Handle() const { return static_cast<const void *>(this); }
};
template <typename T>
......
paddle/fluid/framework/naive_executor.cc
浏览文件 @ ce994190
......@@ -146,22 +146,5 @@ void NaiveExecutor::CleanFeedFetchOps() {
ops_.swap(ops);
}
void NaiveExecutor::EnableMKLDNN(const ProgramDesc &program) {
#ifdef PADDLE_WITH_MKLDNN
VLOG(3) << "use_mkldnn=True";
for (size_t block_id = 0; block_id < program.Size(); ++block_id) {
auto *block = const_cast<ProgramDesc &>(program).MutableBlock(block_id);
for (auto *op : block->AllOps()) {
if (op->HasAttr("use_mkldnn")) {
op->SetAttr("use_mkldnn", true);
}
}
}
#else
LOG(WARNING)
<< "'MKLDNN' is not supported, Please re-compile with WITH_MKLDNN option";
#endif
}
} // namespace framework
} // namespace paddle
paddle/fluid/framework/naive_executor.h
浏览文件 @ ce994190
......@@ -48,8 +48,6 @@ class NaiveExecutor {
void CleanFeedFetchOps();
void EnableMKLDNN(const ProgramDesc& program);
protected:
void CreateVariables(const ProgramDesc& desc, Scope* scope, int block_id);
......
paddle/fluid/framework/op_desc.cc
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/framework/op_desc.h
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/framework/operator.cc
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/framework/operator.h
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/framework/scope.h
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/framework/threadpool.h
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/framework/type_defs.h
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/inference/api/details/reset_tensor_array.h 0 → 100644
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/inference/api/helper.h
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/inference/tests/api/analyzer_dam_tester.cc 0 → 100644
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/inference/tests/test.cmake 0 → 100644
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/adagrad_op.h
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/adam_op.cc
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/add_position_encoding_op.cc 0 → 100644
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/add_position_encoding_op.h 0 → 100644
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/affine_channel_op.cc 0 → 100644
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/affine_channel_op.cu 0 → 100644
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/affine_grid_cudnn_op.cu.cc 0 → 100644
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/affine_grid_op.cc 0 → 100644
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/affine_grid_op.h 0 → 100644
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/concat_op.h
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/conv_op.cc
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/detection/gpc.cc 0 → 100644
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/detection/gpc.h 0 → 100644
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/detection/poly_util.cc 0 → 100644
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/detection/poly_util.h 0 → 100644
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/dropout_op.h
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/fake_init_op.cc 0 → 100644
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/fusion_seqconv_eltadd_relu_op.cc 0 → 100644
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/fusion_seqconv_eltadd_relu_op.h 0 → 100644
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/gather.h
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/gather_op.cc
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/gather_op.cu
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/grid_sampler_cudnn_op.cu.cc 0 → 100644
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/grid_sampler_op.cc 0 → 100644
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/grid_sampler_op.h 0 → 100644
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/hash_op.cc 0 → 100644
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/hash_op.h 0 → 100644
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/lars_momentum_op.cc 0 → 100644
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/lars_momentum_op.cu 0 → 100644
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/lars_momentum_op.h 0 → 100644
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/math/jit_code.cc 0 → 100644
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/math/jit_code.h 0 → 100644
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/math/jit_gen.cc 0 → 100644
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/math/jit_gen.h 0 → 100644
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/math/jit_kernel_crf_decode.cc 0 → 100644
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/math/sequence_pooling_test.cc 0 → 100644
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/mean_op.cc
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/mean_op.cu
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/mean_op.h
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/mul_op.cu.cc
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/pool_op.cc
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/pool_op.h
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/read_op.cc
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/recv_op.cc
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/ref_by_trainer_id_op.cc 0 → 100644
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/ref_by_trainer_id_op.cu.cc 0 → 100644
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/ref_by_trainer_id_op.h 0 → 100644
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/rmsprop_op.h
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/roi_align_op.cc 0 → 100644
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/roi_align_op.cu 0 → 100644
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/roi_align_op.h 0 → 100644
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/scale_op.cu
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/send_op.cc
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/sequence_reverse_op.cc 0 → 100644
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/sequence_reverse_op.cu 0 → 100644
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/sequence_reverse_op.h 0 → 100644
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/sgd_op.cc
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/sign_op.cc
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/sign_op.cu
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/split_op.cc
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/split_op.h
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/spp_op.h
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/sum_op.cc
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/sum_op.cu
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/sum_op.h
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/top_k_op.cc
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/top_k_op.cu
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/operators/top_k_op.h
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/platform/init.cc
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/platform/profiler.cc
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/platform/profiler.h
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/pybind/protobuf.cc
浏览文件 @ ce994190
此差异已折叠。
paddle/fluid/pybind/pybind.cc
浏览文件 @ ce994190
此差异已折叠。
paddle/scripts/paddle_build.sh
浏览文件 @ ce994190
此差异已折叠。
python/paddle/dataset/wmt16.py
浏览文件 @ ce994190
此差异已折叠。
python/paddle/fluid/__init__.py
浏览文件 @ ce994190
此差异已折叠。
python/paddle/fluid/clip.py
浏览文件 @ ce994190
此差异已折叠。
python/paddle/fluid/evaluator.py
浏览文件 @ ce994190
此差异已折叠。
python/paddle/fluid/framework.py
浏览文件 @ ce994190
此差异已折叠。
python/paddle/fluid/io.py
浏览文件 @ ce994190
此差异已折叠。
python/paddle/fluid/layer_helper.py
浏览文件 @ ce994190
此差异已折叠。
python/paddle/fluid/layers/io.py
浏览文件 @ ce994190
此差异已折叠。
python/paddle/fluid/layers/nn.py
浏览文件 @ ce994190
此差异已折叠。
python/paddle/fluid/metrics.py
浏览文件 @ ce994190
此差异已折叠。
python/paddle/fluid/nets.py
浏览文件 @ ce994190
此差异已折叠。
python/paddle/fluid/op.py
浏览文件 @ ce994190
此差异已折叠。
python/paddle/fluid/optimizer.py
浏览文件 @ ce994190
此差异已折叠。
python/paddle/fluid/regularizer.py
浏览文件 @ ce994190
此差异已折叠。
python/paddle/fluid/tests/unittests/dist_mnist_lars.py 0 → 100644
浏览文件 @ ce994190
此差异已折叠。
python/paddle/fluid/tests/unittests/dist_save_load.py 0 → 100644
浏览文件 @ ce994190
此差异已折叠。
python/paddle/fluid/tests/unittests/test_hash_op.py 0 → 100644
浏览文件 @ ce994190
此差异已折叠。
python/paddle/fluid/tests/unittests/test_metrics.py 0 → 100644
浏览文件 @ ce994190
此差异已折叠。
python/paddle/utils/__init__.py
浏览文件 @ ce994190
此差异已折叠。
python/paddle/utils/plot.py 0 → 100644
浏览文件 @ ce994190
此差异已折叠。
python/setup.py.in
浏览文件 @ ce994190
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