Skip to content

P
Paddle

BaiXuePrincess / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 1
Fork 0
P
Paddle
提交 d497bd90 编写于 作者: J jerrywgz
浏览文件
操作

resolve conflict, test=develop

上级 a9c0968b 02c106c7
展开全部 隐藏空白更改
内联 并排
Showing with 12126 addition and 10822 deletion
.pre-commit-config.yaml
浏览文件 @ d497bd90
......@@ -42,12 +42,6 @@ repos:
entry: bash ./tools/codestyle/pylint_pre_commit.hook
language: system
files: \.(py)$
- repo: https://github.com/PaddlePaddle/pre-commit-golang
sha: 8337620115c25ff8333f1b1a493bd031049bd7c0
hooks:
- id: go-fmt
types:
- go
- repo: local
hooks:
- id: copyright_checker
......
AUTHORS.md
浏览文件 @ d497bd90
......@@ -44,6 +44,7 @@
| qingqing01 | Qing-Qing Dang |
| reyoung | Yang Yu |
| Sand3r- | Michal Gallus |
| sfraczek | Sylwester Fraczek |
| Superjom | Chun-Wei Yan |
| tensor-tang | Jian Tang |
| tianbingsz | Tian-Bing Xu |
......@@ -54,6 +55,7 @@
| wangyang59 | Yang Wang |
| wangzhen-nlp | Zhen Wang |
| wen-bo-yang | Wen-Bo Yang |
| wojtuss | Wojciech Uss |
| wwhu | Wei-Wei Hu |
| xinghai-sun | Xing-Hai Sun |
| Xreki | Yi-Qun Liu |
......
CMakeLists.txt
浏览文件 @ d497bd90
......@@ -54,23 +54,12 @@ option(WITH_NGRAPH "Compile PaddlePaddle with nGraph support." OFF)
option(WITH_DSO "Compile PaddlePaddle with dynamic linked CUDA" ON)
option(WITH_TESTING "Compile PaddlePaddle with unit testing" OFF)
option(WITH_PYTHON "Compile PaddlePaddle with python interpreter" ON)
option(WITH_DOUBLE "Compile PaddlePaddle with double precision" OFF)
option(WITH_RDMA "Compile PaddlePaddle with RDMA support" OFF)
option(WITH_TIMER "Compile PaddlePaddle with stats timer" OFF)
option(WITH_PROFILER "Compile PaddlePaddle with GPU profiler and gperftools" OFF)
option(WITH_JEMALLOC "Compile PaddlePaddle with jemalloc" OFF)
option(WITH_DOC "Compile PaddlePaddle with documentation" OFF)
option(WITH_COVERAGE "Compile PaddlePaddle with code coverage" OFF)
option(COVERALLS_UPLOAD "Package code coverage data to coveralls" OFF)
option(ON_TRAVIS "Exclude special unit test on Travis CI" OFF)
option(WITH_FLUID_ONLY "Compile PaddlePaddle fluid only" OFF)
option(WITH_GOLANG "Compile PaddlePaddle with GOLANG" OFF)
option(GLIDE_INSTALL "Download and install go dependencies " ON)
option(WITH_DISTRIBUTE "Compile with distributed support" OFF)
option(WITH_PSLIB "Compile with pslib 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_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)
......@@ -105,8 +94,6 @@ endif()
if (WIN32)
set(WITH_DISTRIBUTE OFF CACHE STRING
"Disable DISTRIBUTE when compiling for Windows" FORCE)
set(WITH_FLUID_ONLY ON CACHE STRING
"Enable FLUID_ONLY when compiling for Windows" FORCE)
endif()
set(THIRD_PARTY_PATH "${CMAKE_BINARY_DIR}/third_party" CACHE STRING
......@@ -148,7 +135,6 @@ include(external/openblas) # download, build, install openblas
include(external/mkldnn) # download, build, install mkldnn
include(external/ngraph) # download, build, install nGraph
include(external/boost) # download boost
include(external/any) # download libn::any
include(external/eigen) # download eigen3
include(external/pybind11) # download pybind11
include(external/cares)
......@@ -225,7 +211,6 @@ include(generic) # simplify cmake module
include(package) # set paddle packages
include(ccache) # set ccache for compilation
include(util) # set unittest and link libs
include(rdma) # set rdma libraries
include(version) # set PADDLE_VERSION
include(coveralls) # set code coverage
include(inference_lib) # add paddle fluid inference libraries
......@@ -233,38 +218,11 @@ include(inference_lib) # add paddle fluid inference libraries
include_directories("${PADDLE_SOURCE_DIR}")
set(EXTERNAL_LIBS
gflags
glog
${CBLAS_LIBRARIES}
protobuf
zlib
${PYTHON_LIBRARIES}
)
if(WITH_PSLIB)
list(APPEND EXTERNAL_LIBS pslib)
list(APPEND EXTERNAL_LIBS pslib_brpc)
list(APPEND EXTERNAL_LIBS libmct)
endif(WITH_PSLIB)
if(WITH_AMD_GPU)
find_package(HIP)
include(hip)
endif(WITH_AMD_GPU)
if(WITH_MKLML)
list(APPEND EXTERNAL_LIBS ${MKLML_IOMP_LIB})
endif()
if(WITH_LIBXSMM)
list(APPEND EXTERNAL_LIBS ${LIBXSMM_LIBS})
endif()
if(WITH_MKLDNN)
list(APPEND EXTERNAL_LIBS ${MKLDNN_LIB})
endif()
set(PADDLE_PYTHON_BUILD_DIR "${CMAKE_CURRENT_BINARY_DIR}/python/build")
set(CMAKE_CXX_FLAGS_RELWITHDEBINFO "-O3 -g -DNDEBUG")
......
README.md
浏览文件 @ d497bd90
......@@ -3,8 +3,8 @@
English | [简体中文](./README_cn.md)
[![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://paddlepaddle.org/documentation/docs/en/1.2/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/1.2/beginners_guide/index.html)
[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://paddlepaddle.org/documentation/docs/en/1.3/beginners_guide/index_en.html)
[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://paddlepaddle.org/documentation/docs/zh/1.3/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)
......@@ -18,7 +18,7 @@ learning to many products at Baidu.
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 1.2.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.2)
### Latest PaddlePaddle Release: [Fluid 1.3.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.3)
### Install Latest Stable Release:
```
# Linux CPU
......@@ -26,9 +26,9 @@ pip install paddlepaddle
# Linux GPU cuda9cudnn7
pip install paddlepaddle-gpu
# Linux GPU cuda8cudnn7
pip install paddlepaddle-gpu==1.2.0.post87
pip install paddlepaddle-gpu==1.3.0.post87
# Linux GPU cuda8cudnn5
pip install paddlepaddle-gpu==1.2.0.post85
pip install paddlepaddle-gpu==1.3.0.post85
# For installation on other platform, refer to http://paddlepaddle.org/
```
......@@ -75,26 +75,26 @@ pip install paddlepaddle-gpu==1.2.0.post85
## Installation
It is recommended to read [this doc](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/install/index_cn.html) on our website.
It is recommended to read [this doc](http://paddlepaddle.org/documentation/docs/en/1.3/beginners_guide/index_en.html) on our website.
## Documentation
We provide [English](http://paddlepaddle.org/documentation/docs/en/1.2/getstarted/index_en.html) and
[Chinese](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/index.html) documentation.
We provide [English](http://paddlepaddle.org/documentation/docs/en/1.3/beginners_guide/index_en.html) and
[Chinese](http://paddlepaddle.org/documentation/docs/zh/1.3/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/1.2/user_guides/howto/training/cluster_howto.html)
- [Distributed Training](http://paddlepaddle.org/documentation/docs/en/1.3/user_guides/howto/training/multi_node_en.html)
You can run distributed training jobs on MPI clusters.
- [Python API](http://paddlepaddle.org/documentation/docs/zh/1.2/api_cn/index_cn.html)
- [Python API](http://paddlepaddle.org/documentation/docs/en/1.3/api/index_en.html)
Our new API enables much shorter programs.
- [How to Contribute](http://paddlepaddle.org/documentation/docs/zh/1.2/advanced_usage/development/contribute_to_paddle/index_cn.html)
- [How to Contribute](http://paddlepaddle.org/documentation/docs/en/1.3/advanced_usage/development/contribute_to_paddle/index_en.html)
We appreciate your contributions!
......
README_cn.md
浏览文件 @ d497bd90
......@@ -3,8 +3,8 @@
[English](./README.md) | 简体中文
[![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://paddlepaddle.org/documentation/docs/en/1.2/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/1.2/beginners_guide/index.html)
[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://paddlepaddle.org/documentation/docs/en/1.3/beginners_guide/index_en.html)
[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://paddlepaddle.org/documentation/docs/zh/1.3/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)
......@@ -16,7 +16,7 @@ PaddlePaddle (PArallel Distributed Deep LEarning) 是一个简单易用、高效
跟进PaddlePaddle最新特性请参考我们的[版本说明](https://github.com/PaddlePaddle/Paddle/releases)
### PaddlePaddle最新版本: [Fluid 1.2.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.2)
### PaddlePaddle最新版本: [Fluid 1.3.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.3)
### 安装最新稳定版本:
```
# Linux CPU
......@@ -24,9 +24,9 @@ pip install paddlepaddle
# Linux GPU cuda9cudnn7
pip install paddlepaddle-gpu
# Linux GPU cuda8cudnn7
pip install paddlepaddle-gpu==1.2.0.post87
pip install paddlepaddle-gpu==1.3.0.post87
# Linux GPU cuda8cudnn5
pip install paddlepaddle-gpu==1.2.0.post85
pip install paddlepaddle-gpu==1.3.0.post85
# 其他平台上的安装指引请参考 http://paddlepaddle.org/
```
......@@ -57,26 +57,26 @@ pip install paddlepaddle-gpu==1.2.0.post85
## 安装
推荐阅读官网上的[安装说明](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/install/index_cn.html)
推荐阅读官网上的[安装说明](http://paddlepaddle.org/documentation/docs/zh/1.3/beginners_guide/install/index_cn.html)
## 文档
我们提供[英文](http://paddlepaddle.org/documentation/docs/en/1.2/getstarted/index_en.html)和
[中文](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/index.html) 文档
我们提供[英文](http://paddlepaddle.org/documentation/docs/en/1.3/beginners_guide/index_en.html)和
[中文](http://paddlepaddle.org/documentation/docs/zh/1.3/beginners_guide/index.html) 文档
- [深度学习101](https://github.com/PaddlePaddle/book)
或许您想从这个在线交互式书籍开始，可以在Jupyter Notebook中运行
- [分布式训练](http://paddlepaddle.org/documentation/docs/zh/1.2/user_guides/howto/training/cluster_howto.html)
- [分布式训练](http://paddlepaddle.org/documentation/docs/zh/1.3/user_guides/howto/training/multi_node.html)
可以在MPI集群上运行分布式训练任务
- [Python API](http://paddlepaddle.org/documentation/docs/zh/1.2/api_cn/index_cn.html)
- [Python API](http://paddlepaddle.org/documentation/docs/zh/1.3/api_cn/index_cn.html)
新的API支持代码更少更简洁的程序
- [贡献方式](http://paddlepaddle.org/documentation/docs/zh/1.2/advanced_usage/development/contribute_to_paddle/index_cn.html)
- [贡献方式](http://paddlepaddle.org/documentation/docs/zh/1.3/advanced_usage/development/contribute_to_paddle/index_cn.html)
欢迎您的贡献!
......
cmake/configure.cmake
浏览文件 @ d497bd90
......@@ -20,31 +20,10 @@ if(WITH_DSO)
add_definitions(-DPADDLE_USE_DSO)
endif(WITH_DSO)
if(WITH_DOUBLE)
add_definitions(-DPADDLE_TYPE_DOUBLE)
endif(WITH_DOUBLE)
if(WITH_ARM_FP16)
add_definitions(-DPADDLE_ARM_FP16)
add_definitions("-march=armv8.2-a+fp16+simd")
endif(WITH_ARM_FP16)
if(WITH_TESTING)
add_definitions(-DPADDLE_WITH_TESTING)
endif(WITH_TESTING)
if(NOT WITH_TIMER)
add_definitions(-DPADDLE_DISABLE_TIMER)
endif(NOT WITH_TIMER)
if(USE_EIGEN_FOR_BLAS)
add_definitions(-DPADDLE_USE_EIGEN_FOR_BLAS)
endif(USE_EIGEN_FOR_BLAS)
if(EIGEN_USE_THREADS)
add_definitions(-DEIGEN_USE_THREADS)
endif(EIGEN_USE_THREADS)
if(NOT WITH_PROFILER)
add_definitions(-DPADDLE_DISABLE_PROFILER)
endif(NOT WITH_PROFILER)
......@@ -78,10 +57,6 @@ if(WIN32)
endif(NOT MSVC)
endif(WIN32)
if(NOT WITH_GOLANG)
add_definitions(-DPADDLE_WITHOUT_GOLANG)
endif(NOT WITH_GOLANG)
if(WITH_PSLIB)
add_definitions(-DPADDLE_WITH_PSLIB)
endif()
......@@ -171,55 +146,6 @@ if(WITH_DISTRIBUTE)
add_definitions(-DPADDLE_WITH_DISTRIBUTE)
endif()
if(WITH_GOLANG)
# we need to symlink Paddle directory into GOPATH. If we
# don't do it and we have code that depends on Paddle, go
# get ./... will download a new Paddle repo from Github,
# without the changes in our current Paddle repo that we
# want to build.
set(GOPATH "${CMAKE_CURRENT_BINARY_DIR}/go")
file(MAKE_DIRECTORY ${GOPATH})
set(PADDLE_IN_GOPATH "${GOPATH}/src/github.com/PaddlePaddle/Paddle")
file(MAKE_DIRECTORY "${PADDLE_IN_GOPATH}")
set(PADDLE_GO_PATH "${CMAKE_SOURCE_DIR}/go")
add_custom_target(go_path)
add_custom_command(TARGET go_path
# Symlink Paddle directory into GOPATH
COMMAND mkdir -p ${PADDLE_IN_GOPATH}
COMMAND rm -rf ${PADDLE_IN_GOPATH}
COMMAND ln -sf ${CMAKE_SOURCE_DIR} ${PADDLE_IN_GOPATH}
# Automatically get all dependencies specified in the source code
# We can't run `go get -d ./...` for every target, because
# multiple `go get` can not run concurrently, but make need to be
# able to run with multiple jobs.
WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
)
if (GLIDE_INSTALL)
if(EXISTS $ENV{GOPATH}/bin/glide)
set(GLIDE "$ENV{GOPATH}/bin/glide")
else()
message(FATAL_ERROR "no glide executeble found: $ENV{GOPATH}/bin/glide")
endif()
# this command will only run when the file it depends is missing
# or has changed, or the output is missing.
add_custom_command(OUTPUT ${CMAKE_BINARY_DIR}/glide
COMMAND env GOPATH=${GOPATH} ${GLIDE} install
COMMAND touch ${CMAKE_BINARY_DIR}/glide
DEPENDS ${PADDLE_SOURCE_DIR}/go/glide.lock
WORKING_DIRECTORY "${PADDLE_IN_GOPATH}/go"
)
# depends on the custom command which outputs
# ${CMAKE_BINARY_DIR}/glide, the custom command does not need to
# run every time this target is built.
add_custom_target(go_vendor DEPENDS ${CMAKE_BINARY_DIR}/glide go_path)
endif()
endif(WITH_GOLANG)
if(WITH_GRPC)
add_definitions(-DPADDLE_WITH_GRPC)
endif(WITH_GRPC)
......
cmake/cuda.cmake
浏览文件 @ d497bd90
......@@ -168,10 +168,7 @@ elseif (${CUDA_VERSION} LESS 11.0) # CUDA 10.x
endif()
include_directories(${CUDA_INCLUDE_DIRS})
list(APPEND EXTERNAL_LIBS ${CUDA_LIBRARIES} ${CUDA_rt_LIBRARY})
if(NOT WITH_DSO)
# TODO(panyx0718): CUPTI only allows DSO?
list(APPEND EXTERNAL_LIBS ${CUDNN_LIBRARY} ${CUPTI_LIBRARY} ${CUDA_CUBLAS_LIBRARIES} ${CUDA_curand_LIBRARY} ${NCCL_LIBRARY})
if(WIN32)
set_property(GLOBAL PROPERTY CUDA_MODULES ${CUDNN_LIBRARY} ${CUDA_CUBLAS_LIBRARIES} ${CUDA_curand_LIBRARY})
endif(WIN32)
......
cmake/external/anakin.cmake
浏览文件 @ d497bd90
......@@ -74,5 +74,3 @@ add_dependencies(anakin_shared extern_anakin)
add_library(anakin_saber SHARED IMPORTED GLOBAL)
set_property(TARGET anakin_saber PROPERTY IMPORTED_LOCATION ${ANAKIN_SABER_LIB})
add_dependencies(anakin_saber extern_anakin)
list(APPEND external_project_dependencies anakin_shared anakin_saber)
cmake/external/any.cmake 已删除 100644 → 0
浏览文件 @ a9c0968b
INCLUDE(ExternalProject)
SET(ANY_SOURCE_DIR ${THIRD_PARTY_PATH}/any)
INCLUDE_DIRECTORIES(${ANY_SOURCE_DIR}/src/extern_lib_any)
ExternalProject_Add(
extern_lib_any
${EXTERNAL_PROJECT_LOG_ARGS}
GIT_REPOSITORY "https://github.com/PaddlePaddle/any.git"
GIT_TAG "15595d8324be9e8a9a80d9ae442fdd12bd66df5d"
PREFIX ${ANY_SOURCE_DIR}
UPDATE_COMMAND ""
CONFIGURE_COMMAND ""
BUILD_COMMAND ""
INSTALL_COMMAND ""
TEST_COMMAND ""
)
if (${CMAKE_VERSION} VERSION_LESS "3.3.0")
set(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/lib_any_dummy.c)
file(WRITE ${dummyfile} "const char * dummy_any = \"${dummyfile}\";")
add_library(lib_any STATIC ${dummyfile})
else()
add_library(lib_any INTERFACE)
endif()
add_dependencies(lib_any extern_lib_any)
add_definitions(-DANY_IMPL_ANY_CAST_MOVEABLE)
LIST(APPEND external_project_dependencies lib_any)
cmake/external/boost.cmake
浏览文件 @ d497bd90
......@@ -57,5 +57,4 @@ else()
endif()
add_dependencies(boost ${BOOST_PROJECT})
list(APPEND external_project_dependencies boost)
set(Boost_INCLUDE_DIR ${BOOST_INCLUDE_DIR})
cmake/external/brpc.cmake
浏览文件 @ d497bd90
......@@ -69,5 +69,3 @@ SET_PROPERTY(TARGET brpc PROPERTY IMPORTED_LOCATION ${BRPC_LIBRARIES})
ADD_DEPENDENCIES(brpc extern_brpc)
add_definitions(-DBRPC_WITH_GLOG)
LIST(APPEND external_project_dependencies brpc)
cmake/external/cub.cmake
浏览文件 @ d497bd90
......@@ -31,5 +31,3 @@ else()
endif()
add_dependencies(cub extern_cub)
LIST(APPEND external_project_dependencies cub)
cmake/external/dlpack.cmake
浏览文件 @ d497bd90
......@@ -27,5 +27,3 @@ else()
endif()
add_dependencies(dlpack extern_dlpack)
LIST(APPEND external_project_dependencies dlpack)
cmake/external/eigen.cmake
浏览文件 @ d497bd90
......@@ -52,5 +52,3 @@ else()
endif()
add_dependencies(eigen3 extern_eigen3)
LIST(APPEND external_project_dependencies eigen3)
cmake/external/gflags.cmake
浏览文件 @ d497bd90
......@@ -61,8 +61,6 @@ ADD_LIBRARY(gflags STATIC IMPORTED GLOBAL)
SET_PROPERTY(TARGET gflags PROPERTY IMPORTED_LOCATION ${GFLAGS_LIBRARIES})
ADD_DEPENDENCIES(gflags extern_gflags)
LIST(APPEND external_project_dependencies gflags)
# On Windows (including MinGW), the Shlwapi library is used by gflags if available.
if (WIN32)
include(CheckIncludeFileCXX)
......
cmake/external/glog.cmake
浏览文件 @ d497bd90
......@@ -72,5 +72,3 @@ ADD_LIBRARY(glog STATIC IMPORTED GLOBAL)
SET_PROPERTY(TARGET glog PROPERTY IMPORTED_LOCATION ${GLOG_LIBRARIES})
ADD_DEPENDENCIES(glog extern_glog gflags)
LINK_LIBRARIES(glog gflags)
LIST(APPEND external_project_dependencies glog)
cmake/external/gtest.cmake
浏览文件 @ d497bd90
......@@ -79,5 +79,4 @@ IF(WITH_TESTING OR (WITH_DISTRIBUTE AND NOT WITH_GRPC))
SET_PROPERTY(TARGET gtest_main PROPERTY IMPORTED_LOCATION ${GTEST_MAIN_LIBRARIES})
ADD_DEPENDENCIES(gtest_main extern_gtest)
LIST(APPEND external_project_dependencies gtest gtest_main)
ENDIF(WITH_TESTING OR (WITH_DISTRIBUTE AND NOT WITH_GRPC))
cmake/external/leveldb.cmake
浏览文件 @ d497bd90
......@@ -39,6 +39,3 @@ ADD_DEPENDENCIES(extern_leveldb snappy)
ADD_LIBRARY(leveldb STATIC IMPORTED GLOBAL)
SET_PROPERTY(TARGET leveldb PROPERTY IMPORTED_LOCATION ${LEVELDB_LIBRARIES})
ADD_DEPENDENCIES(leveldb extern_leveldb)
LIST(APPEND external_project_dependencies leveldb)
cmake/external/libmct.cmake
浏览文件 @ d497bd90
......@@ -72,7 +72,4 @@ else()
add_library(libmct INTERFACE)
endif()
#ADD_LIBRARY(libmct SHARED IMPORTED GLOBAL)
ADD_DEPENDENCIES(libmct ${LIBMCT_PROJECT})
LIST(APPEND external_project_dependencies libmct)
cmake/external/libxsmm.cmake
浏览文件 @ d497bd90
......@@ -53,5 +53,3 @@ MESSAGE(STATUS "Libxsmm library: ${LIBXSMM_LIBS}")
include_directories(${LIBXSMM_INCLUDE_DIR})
ADD_DEFINITIONS(-DPADDLE_WITH_LIBXSMM)
ADD_DEPENDENCIES(libxsmm extern_libxsmm)
LIST(APPEND external_project_dependencies libxsmm)
cmake/external/mkldnn.cmake
浏览文件 @ d497bd90
......@@ -89,7 +89,6 @@ SET_PROPERTY(TARGET shared_mkldnn PROPERTY IMPORTED_LOCATION ${MKLDNN_LIB})
ADD_DEPENDENCIES(shared_mkldnn ${MKLDNN_PROJECT})
MESSAGE(STATUS "MKLDNN library: ${MKLDNN_LIB}")
add_definitions(-DPADDLE_WITH_MKLDNN)
LIST(APPEND external_project_dependencies shared_mkldnn)
# generate a static dummy target to track mkldnn dependencies
# for cc_library(xxx SRCS xxx.c DEPS mkldnn)
......
cmake/external/mklml.cmake
浏览文件 @ d497bd90
......@@ -39,8 +39,10 @@ IF(WIN32)
SET(MKLML_IOMP_LIB ${MKLML_LIB_DIR}/libiomp5md.lib)
SET(MKLML_SHARED_LIB ${MKLML_LIB_DIR}/mklml.dll)
SET(MKLML_SHARED_IOMP_LIB ${MKLML_LIB_DIR}/libiomp5md.dll)
ELSE()
SET(MKLML_VER "mklml_lnx_${TIME_VERSION}" CACHE STRING "" FORCE)
ELSE()
#TODO(intel-huying):
# Now enable Erf function in mklml library temporarily, it will be updated as offical version later.
SET(MKLML_VER "Glibc225_vsErf_mklml_lnx_${TIME_VERSION}" CACHE STRING "" FORCE)
SET(MKLML_URL "http://paddlepaddledeps.cdn.bcebos.com/${MKLML_VER}.tgz" CACHE STRING "" FORCE)
SET(MKLML_LIB ${MKLML_LIB_DIR}/libmklml_intel.so)
SET(MKLML_IOMP_LIB ${MKLML_LIB_DIR}/libiomp5.so)
......@@ -73,4 +75,3 @@ INCLUDE_DIRECTORIES(${MKLML_INC_DIR})
ADD_LIBRARY(mklml SHARED IMPORTED GLOBAL)
SET_PROPERTY(TARGET mklml PROPERTY IMPORTED_LOCATION ${MKLML_LIB})
ADD_DEPENDENCIES(mklml ${MKLML_PROJECT})
LIST(APPEND external_project_dependencies mklml)
cmake/external/ngraph.cmake
浏览文件 @ d497bd90
......@@ -37,7 +37,7 @@ INCLUDE(GNUInstallDirs)
INCLUDE(ExternalProject)
SET(NGRAPH_PROJECT "extern_ngraph")
SET(NGRAPH_GIT_TAG "20bd8bbc79ae3a81c57313846a2be7313e5d1dab")
SET(NGRAPH_GIT_TAG "a444f7a959b7d87f2c117c9b57a4c387759e481e")
SET(NGRAPH_SOURCES_DIR ${THIRD_PARTY_PATH}/ngraph)
SET(NGRAPH_INSTALL_DIR ${THIRD_PARTY_PATH}/install/ngraph)
SET(NGRAPH_INC_DIR ${NGRAPH_INSTALL_DIR}/include)
......@@ -69,7 +69,7 @@ ExternalProject_Add(
CMAKE_ARGS -DNGRAPH_DEX_ONLY=TRUE
CMAKE_ARGS -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}
CMAKE_ARGS -DMKLDNN_INCLUDE_DIR=${MKLDNN_INC_DIR}
CMAKE_ARGS -DMKLDNN_LIB_DIR=${MKLDNN_INSTALL_DIR}/lib
CMAKE_ARGS -DMKLDNN_LIB_DIR=${MKLDNN_INSTALL_DIR}/${CMAKE_INSTALL_LIBDIR}
CMAKE_ARGS -DMKLML_LIB_DIR=${MKLML_INSTALL_DIR}/lib
)
......@@ -77,4 +77,3 @@ add_dependencies(ngraph ${NGRAPH_PROJECT})
target_compile_definitions(ngraph INTERFACE -DPADDLE_WITH_NGRAPH)
target_include_directories(ngraph INTERFACE ${NGRAPH_INC_DIR})
target_link_libraries(ngraph INTERFACE ${NGRAPH_SHARED_LIB})
LIST(APPEND external_project_dependencies ngraph)
cmake/external/openblas.cmake
浏览文件 @ d497bd90
......@@ -11,11 +11,6 @@
# 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.
IF(USE_EIGEN_FOR_BLAS)
return()
ENDIF(USE_EIGEN_FOR_BLAS)
INCLUDE(cblas)
IF(NOT ${CBLAS_FOUND})
......@@ -91,7 +86,6 @@ ENDIF()
IF(NOT ${CBLAS_FOUND})
ADD_DEPENDENCIES(cblas extern_openblas)
LIST(APPEND external_project_dependencies cblas)
ELSE()
IF("${CBLAS_PROVIDER}" STREQUAL "MKLML")
ADD_DEPENDENCIES(cblas mklml)
......
cmake/external/protobuf.cmake
浏览文件 @ d497bd90
......@@ -129,7 +129,6 @@ macro(PROMPT_PROTOBUF_LIB)
ADD_DEPENDENCIES(protoc ${dep})
ENDFOREACH()
LIST(APPEND external_project_dependencies protobuf)
RETURN()
endmacro()
macro(SET_PROTOBUF_VERSION)
......@@ -231,7 +230,7 @@ FUNCTION(build_protobuf TARGET_NAME BUILD_FOR_HOST)
)
ENDFUNCTION()
SET(PROTOBUF_VERSION 3.1)
SET(PROTOBUF_VERSION 3.1.0)
IF(NOT PROTOBUF_FOUND)
build_protobuf(extern_protobuf FALSE)
......
cmake/external/pslib.cmake
浏览文件 @ d497bd90
......@@ -70,4 +70,3 @@ ExternalProject_Add(
ADD_LIBRARY(pslib SHARED IMPORTED GLOBAL)
SET_PROPERTY(TARGET pslib PROPERTY IMPORTED_LOCATION ${PSLIB_LIB})
ADD_DEPENDENCIES(pslib ${PSLIB_PROJECT})
LIST(APPEND external_project_dependencies pslib)
cmake/external/pslib_brpc.cmake
浏览文件 @ d497bd90
......@@ -70,4 +70,3 @@ ExternalProject_Add(
ADD_LIBRARY(pslib_brpc SHARED IMPORTED GLOBAL)
SET_PROPERTY(TARGET pslib_brpc PROPERTY IMPORTED_LOCATION ${PSLIB_BRPC_LIB})
ADD_DEPENDENCIES(pslib_brpc ${PSLIB_BRPC_PROJECT})
LIST(APPEND external_project_dependencies pslib_brpc)
cmake/external/threadpool.cmake
浏览文件 @ d497bd90
......@@ -26,5 +26,3 @@ else()
endif()
add_dependencies(simple_threadpool extern_threadpool)
LIST(APPEND external_project_dependencies simple_threadpool)
cmake/external/warpctc.cmake
浏览文件 @ d497bd90
......@@ -83,5 +83,3 @@ INCLUDE_DIRECTORIES(${THIRD_PARTY_PATH}/install) # For Paddle code to include wa
ADD_LIBRARY(warpctc SHARED IMPORTED GLOBAL)
SET_PROPERTY(TARGET warpctc PROPERTY IMPORTED_LOCATION ${WARPCTC_LIBRARIES})
ADD_DEPENDENCIES(warpctc extern_warpctc)
LIST(APPEND external_project_dependencies warpctc)
cmake/external/xbyak.cmake
浏览文件 @ d497bd90
......@@ -55,4 +55,3 @@ else()
endif()
add_dependencies(xbyak ${XBYAK_PROJECT})
list(APPEND external_project_dependencies xbyak)
cmake/external/xxhash.cmake
浏览文件 @ d497bd90
......@@ -71,5 +71,3 @@ 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)
cmake/external/zlib.cmake
浏览文件 @ d497bd90
......@@ -57,5 +57,3 @@ ENDIF(WIN32)
ADD_LIBRARY(zlib STATIC IMPORTED GLOBAL)
SET_PROPERTY(TARGET zlib PROPERTY IMPORTED_LOCATION ${ZLIB_LIBRARIES})
ADD_DEPENDENCIES(zlib extern_zlib)
LIST(APPEND external_project_dependencies zlib)
cmake/hip.cmake
浏览文件 @ d497bd90
......@@ -11,8 +11,6 @@ include_directories("/opt/rocm/rocrand/include")
include_directories("/opt/rocm/rccl/include")
include_directories("/opt/rocm/thrust")
list(APPEND EXTERNAL_LIBS "-L/opt/rocm/lib/ -lhip_hcc")
set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -fPIC -DPADDLE_WITH_HIP -std=c++11" )
if(WITH_DSO)
......@@ -31,22 +29,12 @@ if(WITH_GRPC)
set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DPADDLE_WITH_GRPC")
endif(WITH_GRPC)
if(NOT WITH_GOLANG)
set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DPADDLE_WITHOUT_GOLANG")
endif(NOT WITH_GOLANG)
if(WITH_MKLDNN)
set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DPADDLE_WITH_MKLDNN")
endif(WITH_MKLDNN)
set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DANY_IMPL_ANY_CAST_MOVEABLE")
if(NOT WITH_RDMA)
set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DPADDLE_DISABLE_RDMA")
endif(NOT WITH_RDMA)
if(CMAKE_BUILD_TYPE STREQUAL "Debug")
list(APPEND HIP_HCC_FLAGS ${CMAKE_CXX_FLAGS_DEBUG})
elseif(CMAKE_BUILD_TYPE STREQUAL "RelWithDebInfo")
......
cmake/operators.cmake
浏览文件 @ d497bd90
......@@ -153,7 +153,11 @@ function(op_library TARGET)
# pybind USE_OP_DEVICE_KERNEL for CUDNN
list(LENGTH cudnn_cu_cc_srcs cudnn_cu_cc_srcs_len)
if (WITH_GPU AND ${cudnn_cu_cc_srcs_len} GREATER 0)
if(${TARGET} STREQUAL "activation")
file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(relu, CUDNN);\n")
else()
file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, CUDNN);\n")
endif()
endif()
# pybind USE_OP_DEVICE_KERNEL for MIOPEN
......@@ -168,6 +172,9 @@ function(op_library TARGET)
file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(relu, MKLDNN);\n")
elseif(${MKLDNN_FILE} STREQUAL "conv_mkldnn_op")
file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL_WITH_CUSTOM_TYPE(conv2d, MKLDNN, FP32);\n")
file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL_WITH_CUSTOM_TYPE(conv2d, MKLDNN, S8);\n")
file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL_WITH_CUSTOM_TYPE(conv2d, MKLDNN, U8);\n")
else()
file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, MKLDNN);\n")
endif()
......
cmake/rdma.cmake 已删除 100644 → 0
浏览文件 @ a9c0968b
# user should download rdma first from subversion repository
# execute following instruction to download svn mannally
# svn co https://svn.baidu.com/sys/ip/trunk/rdma/sockrdmav1 rdma/
# svn co https://svn.baidu.com/sys/ip/trunk/rdma/thirdparty rdma/
# we use static output in svn repositories to avoid implict bugs from not standard runtime env.
if(WITH_RDMA)
set(RDMA_ROOT $ENV{RDMA_ROOT} CACHE PATH "Folder contains RDMA sock library and thirdparty library")
function(generate_rdma_links)
#redirect to current DIR to isolate the pollution from system runtime environment
#it can benifits unified control for different gcc environment.
#e.g, by default gcc48 did not refer /usr/lib64 which could contain low version
#runtime libraries that will crash process while loading it. That redirect trick
#can fix it.
execute_process(
COMMAND mkdir -p librdma
COMMAND ln -s -f /usr/lib64/libibverbs.so.1.0.0 librdma/libibverbs.so.1
COMMAND ln -s -f /usr/lib64/libibverbs.so.1.0.0 librdma/libibverbs.so
COMMAND ln -s -f /usr/lib64/librdmacm.so.1.0.0 librdma/librdmacm.so.1
COMMAND ln -s -f /usr/lib64/librdmacm.so.1.0.0 librdma/librdmacm.so
COMMAND ln -s -f /lib64/libnl.so.1.1.4 librdma/libnl.so.1
COMMAND ln -s -f /lib64/libnl.so.1.1.4 librdma/libnl.so
WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}
)
endfunction(generate_rdma_links)
#check and set headers
find_path(RDMA_INC_SXISOCK sxi_sock.h PATHS ${RDMA_ROOT}/sockrdmav1/output/include)
find_path(RDMA_INC_XIO libxio.h PATHS ${RDMA_ROOT}/thirdparty/output/accelio)
find_path(RDMA_INC_EVENT event2 PATHS ${RDMA_ROOT}/thirdparty/output/libevent)
find_path(RDMA_INC_NUMA numa.h PATHS ${RDMA_ROOT}/thirdparty/output/libnuma)
#check and set libs
find_library(RDMA_LIB_SXISOCK NAMES sxisock PATHS ${RDMA_ROOT}/sockrdmav1/output)
find_library(RDMA_LIB_XIO NAMES xio PATHS ${RDMA_ROOT}/thirdparty/output/accelio)
find_library(RDMA_LIB_EVENT NAMES event PATHS ${RDMA_ROOT}/thirdparty/output/libevent)
find_library(RDMA_LIB_EVENT_CORE NAMES event_core PATHS ${RDMA_ROOT}/thirdparty/output/libevent)
find_library(RDMA_LIB_EVENT_EXTRA NAMES event_extra PATHS ${RDMA_ROOT}/thirdparty/output/libevent)
find_library(RDMA_LIB_EVENT_PTHREADS NAMES event_pthreads PATHS ${RDMA_ROOT}/thirdparty/output/libevent)
find_library(RDMA_LIB_NUMA NAMES numa PATHS ${RDMA_ROOT}/thirdparty/output/libnuma)
if(
RDMA_INC_SXISOCK AND
RDMA_INC_XIO AND
RDMA_INC_EVENT AND
RDMA_INC_NUMA AND
RDMA_LIB_SXISOCK AND
RDMA_LIB_XIO AND
RDMA_LIB_EVENT AND
RDMA_LIB_EVENT_CORE AND
RDMA_LIB_EVENT_EXTRA AND
RDMA_LIB_EVENT_PTHREADS AND
RDMA_LIB_NUMA
)
set(RDMA_INC_DIR
${RDMA_INC_SXISOCK}
${RDMA_INC_XIO}
${RDMA_INC_EVENT}
${RDMA_INC_NUMA})
set(RDMA_LIBS
${RDMA_LIB_SXISOCK}
${RDMA_LIB_XIO}
${RDMA_LIB_EVENT}
${RDMA_LIB_EVENT_CORE}
${RDMA_LIB_EVENT_EXTRA}
${RDMA_LIB_EVENT_PTHREADS}
${RDMA_LIB_NUMA}
)
set(RDMA_LD_FLAGS "-L./librdma -libverbs -lrdmacm -Xlinker -rpath ./librdma")
include_directories("${RDMA_INC_DIR}")
else()
#if this module is not called, RDMA_INC_DIR RDMA_LIBS will be null, so top module always refer this variable
message(FATAL_ERROR, "RDMA libraries are not found, try to set RDMA_ROOT or check all related libraries.")
endif()
else(WITH_RDMA)
set(RDMA_LIBS "")
set(RDMA_LD_FLAGS "")
add_definitions(-DPADDLE_DISABLE_RDMA)
endif(WITH_RDMA)
cmake/tensorrt.cmake
浏览文件 @ d497bd90
......@@ -33,6 +33,5 @@ if(TENSORRT_FOUND)
message(STATUS "Current TensorRT header is ${TENSORRT_INCLUDE_DIR}/NvInfer.h. "
"Current TensorRT version is v${TENSORRT_MAJOR_VERSION}. ")
include_directories(${TENSORRT_INCLUDE_DIR})
list(APPEND EXTERNAL_LIBS ${TENSORRT_LIBRARY})
add_definitions(-DPADDLE_WITH_TENSORRT)
endif()
paddle/contrib/float16/run_float16_demo.sh
浏览文件 @ d497bd90
......@@ -14,9 +14,7 @@ cmake .. -DWITH_AVX=OFF \
-DWITH_MKL=OFF \
-DWITH_GPU=ON \
-DWITH_TESTING=ON \
-DWITH_TIMER=ON \
-DWITH_PROFILER=ON \
-DWITH_FLUID_ONLY=ON
make -j `nproc`
pip install -U "$WHEEL_PATH/$(ls $WHEEL_PATH)"
......
paddle/fluid/API.spec
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/framework/block_desc.cc
浏览文件 @ d497bd90
......@@ -13,7 +13,11 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/framework/block_desc.h"
#include <queue>
#include <unordered_set>
#include <utility>
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/framework/program_desc.h"
......@@ -155,6 +159,16 @@ void BlockDesc::RemoveOp(size_t s, size_t e) {
ops_.erase(ops_.begin() + s, ops_.begin() + e);
}
void BlockDesc::RemoveOpInternal(const OpDesc *op_desc) {
// TODO(minqiyang): make this faster
for (auto it = ops_.begin(); it != ops_.end(); ++it) {
if (it->get() == op_desc) {
ops_.erase(it);
break;
}
}
}
std::vector<OpDesc *> BlockDesc::AllOps() const {
std::vector<OpDesc *> res;
for (const auto &op : ops_) {
......
paddle/fluid/framework/block_desc.h
浏览文件 @ d497bd90
......@@ -93,6 +93,8 @@ class BlockDesc {
*/
void RemoveOp(size_t s, size_t e);
void RemoveOpInternal(const OpDesc *op_desc);
void RemoveVar(const std::string &name) { vars_.erase(name); }
std::vector<OpDesc *> AllOps() const;
......
paddle/fluid/framework/data_layout_transform.cc
浏览文件 @ d497bd90
......@@ -134,11 +134,6 @@ void TransDataLayoutFromMKLDNN(const OpKernelType& kernel_type_for_var,
out_layout =
out_layout == DataLayout::kAnyLayout ? DataLayout::kNCHW : out_layout;
auto& pool = platform::DeviceContextPool::Instance();
auto* dev_ctx = dynamic_cast<platform::MKLDNNDeviceContext*>(
pool.Get(expected_kernel_type.place_));
auto& cpu_engine = dev_ctx->GetEngine();
std::vector<int> in_tz = paddle::framework::vectorize2int(in.dims());
std::vector<int> out_tz = in_tz;
......@@ -147,29 +142,25 @@ void TransDataLayoutFromMKLDNN(const OpKernelType& kernel_type_for_var,
"Input tensor type is not supported: %s", in.type());
memory::data_type out_type = in_type;
auto in_format = platform::MKLDNNFormatForSize(in_tz.size(), in.format());
auto out_format =
platform::MKLDNNFormatForSize(in_tz.size(), ToMKLDNNFormat(out_layout));
// output tensor has the same dims as input. Reorder don't change dims
out->Resize(in.dims());
if (in_format != out_format) {
// tempory mem pd fr out , to make reorder
auto out_mem_pd = paddle::platform::create_prim_desc_from_dims(
paddle::framework::vectorize2int(out->dims()),
mkldnn::memory::format::blocked, out_type);
if (in.get_mkldnn_prim_desc() != out_mem_pd) {
void* in_data = GetDataFromTensor(in, in_type);
auto out_data = out->mutable_data(expected_kernel_type.place_, in.type());
auto in_memory =
memory({{{in_tz}, in_type, in_format}, cpu_engine}, in_data);
auto out_memory =
memory({{{out_tz}, out_type, out_format}, cpu_engine}, out_data);
auto in_memory = memory(in.get_mkldnn_prim_desc(), in_data);
auto out_memory = memory(out_mem_pd, out_data);
platform::Reorder(in_memory, out_memory);
} else {
out->ShareDataWith(in);
}
out->set_layout(out_layout);
// reset format since the out tensor will be feed to non-MKLDNN OPkernel
out->set_format(memory::format::format_undef);
#endif
}
......
paddle/fluid/framework/data_transform.cc
浏览文件 @ d497bd90
......@@ -51,13 +51,31 @@ void TransformData(const OpKernelType &expected_kernel_type,
#ifdef PADDLE_WITH_MKLDNN
// Case1 - transform from Non-MKLDNN OPKernel to MKLDNN OPKernel
// Just set layout/format. No real transform occur
auto out_format = platform::MKLDNNFormatForSize(in.dims().size(),
ToMKLDNNFormat(lin));
out.ShareDataWith(input_tensor);
out.set_layout(DataLayout::kMKLDNN);
out.set_format(out_format);
// TODO(jczaja): Remove that once all mkldnn ops
// are modified to work with mkldnn_blocked
auto mkldnn_fmt = [&](int rank) {
switch (rank) {
case 5:
return mkldnn::memory::format::ncdhw;
case 4:
return mkldnn::memory::format::nchw;
case 3:
return mkldnn::memory::format::ncw;
case 2:
return mkldnn::memory::format::nc;
case 1:
return mkldnn::memory::format::x;
default:
return mkldnn::memory::format::blocked;
}
};
auto out_mem_pd = paddle::platform::create_prim_desc_from_dims(
paddle::framework::vectorize2int(out.dims()),
mkldnn_fmt(out.dims().size()));
out.set_mkldnn_prim_desc(out_mem_pd);
#endif
} else {
// Case2 - transfrom from MKLDNN OPKernel to Non-MKLDNN OPKernel
......
paddle/fluid/framework/details/CMakeLists.txt
浏览文件 @ d497bd90
......@@ -50,7 +50,12 @@ cc_library(data_balance_op_handle SRCS data_balance_op_handle.cc DEPS op_handle_
cc_library(gather_op_handle SRCS gather_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor)
cc_library(fuse_vars_op_handle SRCS fuse_vars_op_handle.cc DEPS op_handle_base scope)
cc_library(memory_optimize_helper SRCS memory_optimize_helper.cc DEPS graph graph_helper)
if(WITH_GPU)
cc_library(memory_optimize_helper SRCS memory_optimize_helper.cc DEPS graph graph_helper gpu_info)
else()
cc_library(memory_optimize_helper SRCS memory_optimize_helper.cc DEPS graph graph_helper cpu_info)
endif()
cc_library(memory_optimize_pass SRCS memory_optimize_pass.cc DEPS memory_optimize_helper pass)
cc_library(inplace_op_pass SRCS inplace_op_pass.cc DEPS memory_optimize_pass op_info)
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)
......
paddle/fluid/framework/details/all_reduce_deps_pass.cc
浏览文件 @ d497bd90
......@@ -30,8 +30,6 @@ namespace paddle {
namespace framework {
namespace details {
static constexpr char kAllOpDescs[] = "all_op_descs";
VarHandle* GetValidInput(const OpHandleBase* a) {
for (auto p : a->Inputs()) {
VarHandle* b = dynamic_cast<VarHandle*>(p);
......@@ -52,7 +50,7 @@ std::unique_ptr<ir::Graph> AllReduceDepsPass::ApplyImpl(
std::unordered_map<std::string, int> vars;
// TODO(gongwb): use graph topology sort to find the order of operators.
// Note that must assert topology sort is stable
auto& ops = Get<const std::vector<OpDesc*>>(kAllOpDescs);
auto& ops = graph->Get<const std::vector<OpDesc*>>(kStaleProgramOpDescs);
for (auto* op_desc : ops) {
auto outputs = op_desc->Outputs();
for (auto& o_it : outputs) {
......@@ -122,4 +120,4 @@ std::unique_ptr<ir::Graph> AllReduceDepsPass::ApplyImpl(
REGISTER_PASS(all_reduce_deps_pass,
paddle::framework::details::AllReduceDepsPass)
.RequirePassAttr(paddle::framework::details::kAllOpDescs);
.RequireGraphAttr(paddle::framework::details::kStaleProgramOpDescs);
paddle/fluid/framework/details/all_reduce_op_handle.cc
浏览文件 @ d497bd90
......@@ -53,7 +53,7 @@ AllReduceOpHandle::AllReduceOpHandle(ir::Node *node,
#endif
void AllReduceOpHandle::RunImpl() {
platform::RecordEvent record_event(Name(), dev_ctxes_.cbegin()->second);
platform::RecordEvent record_event(Name());
WaitInputVarGenerated();
auto in_var_handles = DynamicCast<VarHandle>(this->Inputs());
......
paddle/fluid/framework/details/broadcast_op_handle.cc
浏览文件 @ d497bd90
......@@ -22,7 +22,7 @@ namespace framework {
namespace details {
void BroadcastOpHandle::RunImpl() {
platform::RecordEvent record_event(Name(), dev_ctxes_.begin()->second);
platform::RecordEvent record_event(Name());
if (places_.size() == 1) return;
......@@ -30,7 +30,7 @@ void BroadcastOpHandle::RunImpl() {
VarHandle *in_var_handle;
{
auto in_var_handles = DynamicCast<VarHandle>(inputs_);
PADDLE_ENFORCE_EQ(in_var_handles.size(), 1,
PADDLE_ENFORCE_EQ(in_var_handles.size(), 1UL,
"The number of input should be one.");
in_var_handle = in_var_handles[0];
}
......
paddle/fluid/framework/details/build_strategy.cc
浏览文件 @ d497bd90
......@@ -34,9 +34,11 @@ namespace details {
static inline bool SeqOnlyAllReduceOps(const BuildStrategy &strategy) {
// Should fix the allreduce op order if scheduling
// them in multiple threads or processes to avoid hang.
// NOTE: ParallelGraph would execute this pass on each graph, so
// don't need to append it here.
return (!strategy.enable_sequential_execution_ &&
strategy.num_trainers_ > 1) ||
strategy.enable_parallel_graph_;
strategy.num_trainers_ > 1) &&
!strategy.enable_parallel_graph_;
}
class ParallelExecutorPassBuilder : public ir::PassBuilder {
......@@ -133,12 +135,15 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
void AppendMultiDevPass(const BuildStrategy &strategy) {
ir::Pass *multi_devices_pass;
if (strategy_.is_distribution_) {
VLOG(3) << "multi device parameter server mode";
multi_devices_pass = AppendPass("dist_multi_devices_pass").get();
} else {
if (strategy.reduce_ == BuildStrategy::ReduceStrategy::kAllReduce) {
VLOG(3) << "multi devices collective mode with allreduce";
multi_devices_pass =
AppendPass("allreduce_mode_multi_devices_pass").get();
} else if (strategy.reduce_ == BuildStrategy::ReduceStrategy::kReduce) {
VLOG(3) << "multi deivces collective mode with reduce";
multi_devices_pass = AppendPass("reduce_mode_multi_devices_pass").get();
} else {
PADDLE_THROW("Unknown reduce strategy.");
......@@ -169,7 +174,8 @@ bool BuildStrategy::IsMultiDevPass(const std::string &pass_name) const {
}
std::unique_ptr<ir::Graph> BuildStrategy::Apply(
const ProgramDesc &main_program, const std::vector<platform::Place> &places,
std::unique_ptr<ir::Graph> graph,
const std::vector<platform::Place> &places,
const std::string &loss_var_name, const std::vector<Scope *> &local_scopes,
const size_t &nranks,
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
......@@ -180,7 +186,6 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
// Create a default one if not finalized by user.
CreatePassesFromStrategy(false);
std::unique_ptr<ir::Graph> graph(new ir::Graph(main_program));
for (std::shared_ptr<ir::Pass> &pass : pass_builder_->AllPasses()) {
if (IsMultiDevPass(pass->Type())) {
pass->Erase(kPlaces);
......@@ -198,41 +203,12 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
pass->Erase("nccl_ctxs");
pass->SetNotOwned<platform::NCCLContextMap>("nccl_ctxs", nctx);
#endif
} else if (pass->Type() == "memory_optimize_pass") {
if (graph->Has(kAllOpDescs)) {
graph->Erase(kAllOpDescs);
}
const std::vector<OpDesc *> *all_op_descs =
new std::vector<OpDesc *>(main_program.Block(0).AllOps());
graph->Set<const std::vector<OpDesc *>>(kAllOpDescs,
all_op_descs); // take ownership
pass->Erase(kAllOpDescs);
pass->SetNotOwned<const std::vector<OpDesc *>>(kAllOpDescs, all_op_descs);
} else if (pass->Type() == "sequential_execution_pass") {
LOG(INFO) << "set enable_sequential_execution:"
<< enable_sequential_execution_;
pass->Erase(kAllOpDescs);
pass->Set<const std::vector<OpDesc *>>(
kAllOpDescs,
new std::vector<OpDesc *>(main_program.Block(0).AllOps()));
} else if (pass->Type() == "all_reduce_deps_pass") {
LOG(INFO) << "SeqOnlyAllReduceOps:" << SeqOnlyAllReduceOps(*this)
<< ", num_trainers:" << num_trainers_;
pass->Erase(kAllOpDescs);
pass->Set<const std::vector<OpDesc *>>(
kAllOpDescs,
new std::vector<OpDesc *>(main_program.Block(0).AllOps()));
} else if (pass->Type() == "inplace_pass") {
if (graph->Has(kAllOpDescs)) {
graph->Erase(kAllOpDescs);
}
graph->Set<const std::vector<OpDesc *>>(
kAllOpDescs,
new std::vector<OpDesc *>(main_program.Block(0).AllOps()));
} else if (pass->Type() == "fuse_relu_depthwise_conv_pass") {
if (!use_cuda) {
LOG(WARNING) << "fuse_relu_depthwise_conv_pass is only supported on "
......@@ -240,7 +216,9 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
continue;
}
}
VLOG(3) << "Start Apply Pass " << pass->Type();
graph = pass->Apply(std::move(graph));
VLOG(3) << "Finish Apply Pass " << pass->Type();
}
return graph;
}
......
paddle/fluid/framework/details/build_strategy.h
浏览文件 @ d497bd90
......@@ -114,7 +114,7 @@ struct BuildStrategy {
// Apply the passes built by the pass_builder_. The passes will be
// applied to the Program and output an ir::Graph.
std::unique_ptr<ir::Graph> Apply(const ProgramDesc &main_program,
std::unique_ptr<ir::Graph> Apply(std::unique_ptr<ir::Graph> graph,
const std::vector<platform::Place> &places,
const std::string &loss_var_name,
const std::vector<Scope *> &local_scopes,
......
paddle/fluid/framework/details/data_balance_op_handle.cc
浏览文件 @ d497bd90
......@@ -86,7 +86,7 @@ std::vector<std::array<int, 3>> DataBalanceOpHandle::GetBalancePlan(
}
void DataBalanceOpHandle::RunImpl() {
PADDLE_ENFORCE_GT(places_.size(), 1,
PADDLE_ENFORCE_GT(places_.size(), 1UL,
"Data balance can only be enabled when the number of "
"places to run larger than 1.");
auto in_var_handles = DynamicCast<VarHandle>(this->Inputs());
......
paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.cc
浏览文件 @ d497bd90
......@@ -12,7 +12,9 @@
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h"
#include <memory>
#include <string>
#include <unordered_map>
#include <vector>
#include "paddle/fluid/framework/details/fetch_op_handle.h"
#include "paddle/fluid/framework/details/multi_devices_helper.h"
......@@ -24,12 +26,11 @@ namespace details {
FastThreadedSSAGraphExecutor::FastThreadedSSAGraphExecutor(
const ExecutionStrategy &strategy, const std::vector<Scope *> &local_scopes,
const std::vector<platform::Place> &places,
std::unique_ptr<ir::Graph> &&graph)
const std::vector<platform::Place> &places, ir::Graph *graph)
: strategy_(strategy),
local_scopes_(local_scopes),
places_(places),
graph_(std::move(graph)),
graph_(graph),
pool_(strategy.num_threads_),
prepare_pool_(1), // add one more thread for generate op_deps
fetch_ctxs_(places) {
......@@ -56,7 +57,7 @@ FeedFetchList FastThreadedSSAGraphExecutor::Run(
std::vector<FetchOpHandle *> fetch_ops;
for (auto &fetch_var_name : fetch_tensors) {
for (auto &var_map : graph_->Get<details::GraphVars>("vars")) {
for (auto &var_map : graph_->Get<details::GraphVars>(details::kGraphVars)) {
auto it = var_map.find(fetch_var_name);
if (it != var_map.end()) {
fetched_vars[fetch_var_name].push_back(*it->second.rbegin());
......@@ -110,14 +111,14 @@ FeedFetchList FastThreadedSSAGraphExecutor::Run(
}
}
if (exception_.IsCaught()) {
ClearFetchOp(graph_.get(), &fetch_ops);
ClearFetchOp(graph_, &fetch_ops);
exception_.ReThrow();
}
}
num_complete += num_comp;
}
// Wait FetchOps.
ClearFetchOp(graph_.get(), &fetch_ops);
ClearFetchOp(graph_, &fetch_ops);
return fetches;
}
......
paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h
浏览文件 @ d497bd90
......@@ -32,7 +32,7 @@ class FastThreadedSSAGraphExecutor : public SSAGraphExecutor {
FastThreadedSSAGraphExecutor(const ExecutionStrategy &strategy,
const std::vector<Scope *> &local_scopes,
const std::vector<platform::Place> &places,
std::unique_ptr<ir::Graph> &&graph);
ir::Graph *graph);
FeedFetchList Run(const std::vector<std::string> &fetch_tensors) override;
const ir::Graph &Graph() const override;
......@@ -40,7 +40,7 @@ class FastThreadedSSAGraphExecutor : public SSAGraphExecutor {
ExecutionStrategy strategy_;
std::vector<Scope *> local_scopes_;
std::vector<platform::Place> places_;
std::unique_ptr<ir::Graph> graph_;
ir::Graph *graph_;
std::unordered_map<OpHandleBase *, int> op_deps_;
std::vector<OpHandleBase *> bootstrap_ops_;
......
paddle/fluid/framework/details/fuse_vars_op_handle.cc
浏览文件 @ d497bd90
......@@ -23,7 +23,7 @@ void FuseVarsOpHandle::RunImpl() {
auto in_var_handles = DynamicCast<VarHandle>(this->Inputs());
auto out_var_handles = DynamicCast<VarHandle>(this->Outputs());
PADDLE_ENFORCE_EQ(in_var_handles.size(), 0);
PADDLE_ENFORCE_EQ(in_var_handles.size(), 0UL);
PADDLE_ENFORCE_EQ(out_var_handles.size() - 1, inputs_numel_.size(), "");
auto scope = local_scope_->FindVar(kLocalExecScopeName)->Get<Scope *>();
......
paddle/fluid/framework/details/fused_broadcast_op_handle.cc
浏览文件 @ d497bd90
......@@ -22,7 +22,7 @@ namespace framework {
namespace details {
void FusedBroadcastOpHandle::RunImpl() {
platform::RecordEvent record_event(Name(), dev_ctxes_.begin()->second);
platform::RecordEvent record_event(Name());
if (places_.size() == 1UL) return;
......
paddle/fluid/framework/details/inplace_op_pass.cc
浏览文件 @ d497bd90
......@@ -49,7 +49,7 @@ DEFINE_bool(
"If this option turns on, only these op in whitelist can be inplaced."
"If it turns off, all of the running op can be candidate of inplaced op."
"Such as scale, elementwise_add"
"By default, it's turned on");
"By default, it's turned off");
DECLARE_string(memory_optimize_debug);
......
paddle/fluid/framework/details/memory_optimize_helper.cc
浏览文件 @ d497bd90
......@@ -13,13 +13,19 @@
// limitations under the License.
#include "paddle/fluid/framework/details/memory_optimize_helper.h"
#include <algorithm>
#include <deque>
#include <functional>
#include <iostream>
#include <iterator>
#include <numeric>
#include <sstream>
#include <string>
#include "paddle/fluid/framework/var_desc.h"
#include "paddle/fluid/platform/cpu_info.h"
#ifdef PADDLE_WITH_CUDA
#include "paddle/fluid/platform/gpu_info.h"
#endif // PADDLE_WITH_CUDA
namespace paddle {
namespace framework {
......@@ -27,10 +33,10 @@ namespace details {
using paddle::framework::VarDesc;
std::vector<ir::Node*> SortOpLikeDescOrder(const ir::Graph& graph) {
PADDLE_ENFORCE(graph.Has(kAllOpDescs),
"Graph has no attribute of kAllOpDescs.");
PADDLE_ENFORCE(graph.Has(kStaleProgramOpDescs),
"Graph has no attribute of kStaleProgramOpDescs.");
// 1. get op desc order
auto& op_descs = graph.Get<const std::vector<OpDesc*>>(kAllOpDescs);
auto& op_descs = graph.Get<const std::vector<OpDesc*>>(kStaleProgramOpDescs);
// 2. topology sort order
auto nodes = graph.Nodes();
......@@ -123,7 +129,13 @@ size_t NodeSize(const VarDesc& node) {
}
size_t NodeSize(ir::Node* n) {
auto* desc = FindVarDescInBlock(n);
VarDesc* desc = nullptr;
// some op do not have block pointer
if (n->inputs[0]->Op() != nullptr) {
desc = FindVarDescInBlock(n);
} else {
desc = n->Var();
}
return NodeSize(*desc);
}
......@@ -166,6 +178,11 @@ struct NodeComparator {
bool operator()(ir::Node* lhs, ir::Node* rhs) const {
auto* lhs_desc = FindVarDescInBlock(lhs);
auto* rhs_desc = FindVarDescInBlock(rhs);
// match data type
if (lhs_desc->GetDataType() != rhs_desc->GetDataType()) {
return false;
}
// match shape
auto lhs_shape = lhs_desc->GetShape();
auto rhs_shape = rhs_desc->GetShape();
if ((lhs_shape[0] == -1 && rhs_shape[0] == -1) ||
......@@ -230,6 +247,27 @@ ir::Node* OrderedSet::FindBestFitNode(ir::Node* var) const {
return found_node;
}
ir::Node* OrderedSet::FindNextBestFitNode(ir::Node* var, ir::Node* prev) const {
ir::Node* found_node = nullptr;
NodeComparator functor;
auto it =
std::find_if(nodes_.begin(), nodes_.end(), [&](const NodeVector& v) {
if (v.front() == prev)
return true;
else
return false;
});
PADDLE_ENFORCE(it != nodes_.end(), "Not found previous in node list!");
for (it = std::next(it); it != nodes_.end(); ++it) {
auto& candidate = it->front();
if (functor(var, candidate)) {
found_node = candidate;
break;
}
}
return found_node;
}
bool OrderedSet::Has(ir::Node* var) const {
if (mark_table_.count(var->Name())) {
auto& node_in_samename = mark_table_.at(var->Name());
......@@ -241,10 +279,15 @@ bool OrderedSet::Has(ir::Node* var) const {
return false;
}
void OrderedSet::Erase(const std::string& var) {
PADDLE_ENFORCE(mark_table_.count(var));
nodes_.erase(mark_table_[var]);
mark_table_.erase(var);
}
void OrderedSet::Erase(ir::Node* var) {
PADDLE_ENFORCE(mark_table_.count(var->Name()));
nodes_.erase(mark_table_[var->Name()]);
mark_table_.erase(var->Name());
PADDLE_ENFORCE(var != nullptr);
Erase(var->Name());
}
std::string OrderedSet::ToString() const {
......@@ -274,14 +317,35 @@ bool NodeCanReused(ir::Node* node) {
return flag;
}
int MinChunkSize() {
int size{0};
#ifdef PADDLE_WITH_CUDA
size = platform::GpuMinChunkSize();
#else
size = platform::CpuMinChunkSize();
#endif // PADDLE_WITH_CUDA
return size;
}
bool NodeCanReused(const VarDesc& node) {
auto type = node.GetType();
// only these types holds bulk of gpu memory
if (!(type == proto::VarType::LOD_TENSOR ||
type == proto::VarType::SELECTED_ROWS ||
type == proto::VarType::LOD_TENSOR_ARRAY)) {
return false;
}
if (node.Persistable() || node.GetShape().empty()) {
// persistable variable is parameter
if (node.Persistable()) {
return false;
}
// shape < min_chunk_size is meaningless.
// further more, fetched loss always has size = 1
// which should not be reused.
auto shape = node.GetShape();
int size = std::abs(
std::accumulate(shape.begin(), shape.end(), 1, std::multiplies<int>()));
if (shape.empty() || size < MinChunkSize()) {
return false;
}
// vars can be @EMPTY@, @LR_DECAY_REUSE_ID@. For example, while_grad
......@@ -397,11 +461,21 @@ void ControlFlowGraph::LiveVariableAnalysis() {
}
}
}
for (auto* op : ops_) {
unlived_vars_[op] = std::set<std::string>();
for (auto& var : this->LiveIn(op)) {
if (!this->LiveOut(op).count(var)) {
unlived_vars_[op].insert(var);
}
}
}
}
void ControlFlowGraph::RenameVarInCFGGraph(const std::string& old_node,
const std::string& new_node,
int begin_idx) {
std::vector<bool> need_update(ops_.size(), false);
// update graph from begin idx to the end
for (size_t i = begin_idx; i != ops_.size(); ++i) {
auto* op = ops_[i];
......@@ -416,15 +490,27 @@ void ControlFlowGraph::RenameVarInCFGGraph(const std::string& old_node,
if (live_in_[op].find(old_node) != live_in_[op].end()) {
live_in_[op].erase(old_node);
live_in_[op].insert(new_node);
need_update[i] = true;
}
if (live_out_[op].find(old_node) != live_out_[op].end()) {
live_out_[op].erase(old_node);
live_out_[op].insert(new_node);
need_update[i] = true;
}
}
for (size_t i = begin_idx; i < ops_.size(); ++i) {
if (!need_update[i]) continue;
auto* op = ops_[i];
for (auto& var : this->LiveIn(op)) {
if (!this->LiveOut(op).count(var)) {
unlived_vars_[op].insert(var);
}
}
}
}
const std::set<std::string> ControlFlowGraph::LiveIn(ir::Node* op) const {
const std::set<std::string>& ControlFlowGraph::LiveIn(ir::Node* op) const {
auto it = live_in_.find(op);
PADDLE_ENFORCE(
it != live_in_.end(),
......@@ -432,7 +518,7 @@ const std::set<std::string> ControlFlowGraph::LiveIn(ir::Node* op) const {
return it->second;
}
const std::set<std::string> ControlFlowGraph::LiveOut(ir::Node* op) const {
const std::set<std::string>& ControlFlowGraph::LiveOut(ir::Node* op) const {
auto it = live_out_.find(op);
PADDLE_ENFORCE(
it != live_out_.end(),
......@@ -440,15 +526,24 @@ const std::set<std::string> ControlFlowGraph::LiveOut(ir::Node* op) const {
return it->second;
}
const std::set<std::string> ControlFlowGraph::Use(ir::Node* op) const {
const std::set<std::string>& ControlFlowGraph::Use(ir::Node* op) const {
auto it = uses_.find(op);
PADDLE_ENFORCE(
it != uses_.end(),
string::Sprintf("Expect %s in live_out, but Not Found.", op->Name()));
string::Sprintf("Expect %s in use, but Not Found.", op->Name()));
return it->second;
}
const std::set<std::string>& ControlFlowGraph::Unlived(ir::Node* op) const {
auto it = unlived_vars_.find(op);
PADDLE_ENFORCE(
it != unlived_vars_.end(),
string::Sprintf("Expect %s in unlived_set, but Not Found.", op->Name()));
return it->second;
return it->second;
}
const std::vector<ir::Node*> ControlFlowGraph::Ops() const { return ops_; }
const std::vector<ir::Node*>& ControlFlowGraph::Ops() const { return ops_; }
std::vector<ir::Node*>& ControlFlowGraph::Ops() { return ops_; }
......@@ -461,7 +556,9 @@ ir::Node* ControlFlowGraph::GetNodeByName(const std::string& name,
for (auto* node : ops_) {
if (node == op) break;
for (auto& output : node->outputs) {
if (output->Name() == name) {
PADDLE_ENFORCE((output != nullptr && output->IsVar()),
"Output is empty!");
if (output->Var() && output->Name() == name) {
found_node = output;
}
}
......
paddle/fluid/framework/details/memory_optimize_helper.h
浏览文件 @ d497bd90
......@@ -29,8 +29,6 @@ namespace paddle {
namespace framework {
namespace details {
constexpr char kAllOpDescs[] = "all_op_descs";
std::vector<ir::Node*> SortOpLikeDescOrder(const ir::Graph& graph);
// NOTE(dzh): A ordered set for node reuse in memory optimize.
......@@ -55,6 +53,7 @@ class OrderedSet {
void Insert(ir::Node* var);
void Erase(ir::Node* var);
void Erase(const std::string& var);
bool Has(ir::Node* var) const;
void Clear() {
mark_table_.clear();
......@@ -62,6 +61,7 @@ class OrderedSet {
}
// find the bestfit shape node block with var.
ir::Node* FindBestFitNode(ir::Node* var) const;
ir::Node* FindNextBestFitNode(ir::Node* var, ir::Node* prev) const;
// map store non-const iterator, can not promise const
int GetNodeIndexInPool(ir::Node* var);
// pool all node to string
......@@ -92,10 +92,11 @@ class ControlFlowGraph {
void RenameVarInCFGGraph(const std::string& old_node,
const std::string& new_node, int begin_idx);
const std::set<std::string> LiveIn(ir::Node* op) const;
const std::set<std::string> LiveOut(ir::Node* op) const;
const std::set<std::string> Use(ir::Node* op) const;
const std::vector<ir::Node*> Ops() const;
const std::set<std::string>& LiveIn(ir::Node* op) const;
const std::set<std::string>& LiveOut(ir::Node* op) const;
const std::set<std::string>& Use(ir::Node* op) const;
const std::set<std::string>& Unlived(ir::Node* op) const;
const std::vector<ir::Node*>& Ops() const;
std::vector<ir::Node*>& Ops();
// for ssa-graph nodes
......@@ -117,6 +118,7 @@ class ControlFlowGraph {
VarSetMap live_out_;
VarSetMap uses_; // op inputs
VarSetMap defs_; // op outputs
std::unordered_map<ir::Node*, std::set<std::string>> unlived_vars_;
std::vector<ir::Node*> ops_; // op sequence by topology sort
};
......
paddle/fluid/framework/details/memory_optimize_helper_test.cc
浏览文件 @ d497bd90
......@@ -107,6 +107,52 @@ TEST(OrderedSet, Normal) {
ASSERT_EQ(pool.GetNodeIndexInPool(cache), 5); // match 4:[5,2]
}
}
TEST(OrderedSet, FindBestFitNode) {
OrderedSet pool;
std::vector<std::unique_ptr<ir::Node>> nodes;
ProgramDesc prog;
BlockDesc* block_desc = prog.MutableBlock(0);
auto* op_desc = block_desc->AppendOp();
op_desc->SetType("dummy");
std::unique_ptr<ir::Node> op = ir::CreateNodeForTest(op_desc);
{
auto desc = block_desc->Var("a");
desc->SetShape({128, 128});
std::unique_ptr<ir::Node> node = ir::CreateNodeForTest(desc);
node->inputs.emplace_back(op.get());
nodes.emplace_back(std::move(node));
}
{
auto desc = block_desc->Var("b");
desc->SetShape({128, 129});
std::unique_ptr<ir::Node> node = ir::CreateNodeForTest(desc);
node->inputs.emplace_back(op.get());
nodes.emplace_back(std::move(node));
}
{
auto desc = block_desc->Var("c");
desc->SetShape({128, 128});
std::unique_ptr<ir::Node> node = ir::CreateNodeForTest(desc);
node->inputs.emplace_back(op.get());
nodes.emplace_back(std::move(node));
}
for (auto& node : nodes) {
pool.Insert(node.get());
}
// FindNextBestFitNode
auto* n = nodes[0].get();
auto* cache = pool.FindBestFitNode(n);
PADDLE_ENFORCE(cache->Name() == "a");
cache = pool.FindNextBestFitNode(n, cache);
PADDLE_ENFORCE(cache->Name() == "c");
cache = pool.FindNextBestFitNode(n, cache);
PADDLE_ENFORCE(cache->Name() == "b");
}
} // namespace details
} // namespace framework
} // namespace paddle
......@@ -182,9 +228,6 @@ TEST(CFGGraph, IRGraph) {
// prepare ir graph
auto prog = FillProgramDesc();
ir::Graph graph(prog);
const std::vector<OpDesc*>* all_op_descs =
new std::vector<OpDesc*>(prog.Block(0).AllOps());
graph.Set(details::kAllOpDescs, all_op_descs); // take ownership
ControlFlowGraph cfg(graph);
cfg.LiveVariableAnalysis();
......@@ -210,9 +253,6 @@ TEST(CFGGraph, IRGraph) {
TEST(SortOpLikeDescOrder, NormalTest) {
auto prog = FillProgramDesc();
ir::Graph graph(prog);
const std::vector<OpDesc*>* all_op_descs =
new std::vector<OpDesc*>(prog.Block(0).AllOps());
graph.Set(details::kAllOpDescs, all_op_descs); // take ownership
auto nodes = SortOpLikeDescOrder(graph);
auto op_descs = prog.Block(0).AllOps();
......@@ -227,9 +267,6 @@ TEST(SortOpLikeDescOrder, NormalTest) {
TEST(SortOpLikeDescOrder, RemoveOpDesc) {
auto prog = FillProgramDesc();
ir::Graph graph(prog);
const std::vector<OpDesc*>* all_op_descs =
new std::vector<OpDesc*>(prog.Block(0).AllOps());
graph.Set(details::kAllOpDescs, all_op_descs); // take ownership
auto nodes = graph.Nodes();
auto op_descs = prog.Block(0).AllOps();
ir::Node* found_node = nullptr;
......@@ -278,8 +315,6 @@ TEST(SortOpLikeDescOrder, RemoveOpDesc) {
// 3. add some op_desc
TEST(SortOpLikeDescOrder, AddOpDesc) {
auto prog = FillProgramDesc();
const std::vector<OpDesc*>* all_op_descs =
new std::vector<OpDesc*>(prog.Block(0).AllOps());
ir::Graph graph(prog);
auto find_node_in_graph = [&](std::string s) {
......@@ -296,9 +331,7 @@ TEST(SortOpLikeDescOrder, AddOpDesc) {
// cached desc different with real one
// mimic the intermidiete pass modify the programdesc.
graph.Set(details::kAllOpDescs, all_op_descs); // take ownership
auto op_descs = prog.Block(0).AllOps();
std::vector<OpDesc*> op_descs = graph.OriginProgram().Block(0).AllOps();
auto op = prog.MutableBlock(0)->AppendOp();
prog.MutableBlock(0)->Var("d1")->SetType(proto::VarType::LOD_TENSOR);
......@@ -330,9 +363,6 @@ TEST(SortOpLikeDescOrder, AddOpDesc) {
TEST(SortOpLikeDescOrder, AddAndDeleteOpDesc) {
auto prog = FillProgramDesc();
ir::Graph graph(prog);
const std::vector<OpDesc*>* all_op_descs =
new std::vector<OpDesc*>(prog.Block(0).AllOps());
graph.Set(details::kAllOpDescs, all_op_descs); // take ownership
auto find_node_in_graph = [&](std::string s) {
ir::Node* ret = nullptr;
......@@ -346,8 +376,9 @@ TEST(SortOpLikeDescOrder, AddAndDeleteOpDesc) {
return ret;
};
std::vector<OpDesc*> op_descs = graph.OriginProgram().Block(0).AllOps();
// remove sum node
auto op_descs = prog.Block(0).AllOps();
ir::Node* found_node = nullptr;
auto nodes = graph.Nodes();
for (auto node : nodes) {
......@@ -408,9 +439,7 @@ TEST(SortOpLikeDescOrder, AddAndDeleteOpDesc) {
TEST(SortOpLikeDescOrder, AddAndReplaceOpDescInplace) {
auto prog = FillProgramDesc();
ir::Graph graph(prog);
const std::vector<OpDesc*>* all_op_descs =
new std::vector<OpDesc*>(prog.Block(0).AllOps());
graph.Set(details::kAllOpDescs, all_op_descs); // take ownership
std::vector<OpDesc*> op_descs = graph.OriginProgram().Block(0).AllOps();
auto find_node_in_graph = [&](std::string s) {
ir::Node* ret = nullptr;
......@@ -424,7 +453,6 @@ TEST(SortOpLikeDescOrder, AddAndReplaceOpDescInplace) {
return ret;
};
auto op_descs = prog.Block(0).AllOps();
// add node
auto op = prog.MutableBlock(0)->AppendOp();
prog.MutableBlock(0)->Var("d1")->SetType(proto::VarType::LOD_TENSOR);
......
paddle/fluid/framework/details/memory_optimize_pass.cc
浏览文件 @ d497bd90
......@@ -69,58 +69,60 @@ std::unique_ptr<ir::Graph> MemoryOptimizePass::ApplyImpl(
}
for (auto& var : op->outputs) {
if (!NodeCanReused(var) || cfg_->Use(op).count(var->Name()) == 0 ||
skip_set_.count(var->Name()))
if (var->IsVar() && !var->IsCtrlVar() && skip_set_.count(var->Name())) {
VLOG(3) << "Skip set contains variable of " << var->Name()
<< "disable reuse on it. skipped";
continue;
ir::Node* cache = pool_.FindBestFitNode(var);
if (var->Name() == FLAGS_memory_optimize_debug) {
VLOG(3) << "start match var " << DebugString(var) << " of op "
<< op->Name();
VLOG(3) << pool_.ToString();
VLOG(3) << "matched in pool : "
<< ((cache == nullptr) ? "False" : "True");
}
if (NodeCanReused(var) && cfg_->Use(op).count(var->Name()) == 0) {
ir::Node* cache = pool_.FindBestFitNode(var);
while (cache != nullptr && var->Name() == cache->Name()) {
VLOG(3) << "The same cache variable is cascade reused. "
<< cache->Name() << " is re-filled to the pool after "
<< "the reused op is finished. Current op can not "
<< "replace it again. Skip this candidate.";
cache = pool_.FindNextBestFitNode(var, cache);
}
if (var->Name() == FLAGS_memory_optimize_debug) {
VLOG(3) << "start match var " << DebugString(var) << " of op "
<< op->Name();
VLOG(3) << pool_.ToString();
VLOG(3) << "matched in pool : "
<< ((cache == nullptr) ? "False" : "True");
}
if (cache == nullptr) continue;
if (var->Name() == cache->Name()) {
VLOG(3) << "The same cache variable is cascade reused." << var->Name()
<< " is re-filled to the pool after"
<< "the reused op is finished. Current op can not "
<< "replace it again. Skip this candidate.";
continue;
int node_idx_in_pool = pool_.GetNodeIndexInPool(cache);
VLOG(3) << string::Sprintf(
"!!! %s, %s => %s, cache idx %d, pool size %d",
std::to_string(reuse_id++), DebugString(var), DebugString(cache),
node_idx_in_pool, static_cast<int>(pool_.size()));
// update CFG Graph on the fly.
// reused var maybe re-fill into the pool
cfg_->RenameVarInCFGGraph(var->Name(), cache->Name(), idx);
// NOTE(dzhwinter): we need to both update the ProgramDesc
// and IR Graph. because op_desc/var_desc is used in CreateOp,
// CreateVar when running happens. But IR Graph
// define the dependence relationship between nodes.
RenameVarInGraphDesc(var->Name(), cache->Name(), idx);
RenameVarInGraphNode(var->Name(), cache->Name(), idx, graph.get());
pool_.Erase(cache);
}
if (cache != nullptr) {
int node_idx_in_pool = pool_.GetNodeIndexInPool(cache);
VLOG(3) << string::Sprintf(
"!!! %s, %s => %s, cache idx %d, pool size %d",
std::to_string(reuse_id++), DebugString(var), DebugString(cache),
node_idx_in_pool, static_cast<int>(pool_.size()));
// NOTE(dzhwinter): update the ProgramDesc/IR Graph
// and the CFG Graph on the fly.
//
// IR Graph define the dependence relationship between nodes.
//
// ProgramDesc defines the input/output vars. Its used in
// CreateOp, CreateVar when running happens.
//
// CFG Graph store the liveness information, when reuse happens
// we also need to update the variable liveness.
const std::string var_name = var->Name();
const std::string cache_name = cache->Name();
// fill the pool
std::unordered_set<std::string> unlived_vars;
for (auto var : cfg_->LiveIn(op)) {
if (cfg_->LiveOut(op).count(var) == 0) {
unlived_vars.emplace(var);
cfg_->RenameVarInCFGGraph(var_name, cache_name, idx);
RenameVarInGraphDesc(var_name, cache_name, idx);
RenameVarInGraphNode(var_name, cache_name, idx, graph.get());
pool_.Erase(cache_name);
}
}
for (auto var : unlived_vars) {
ir::Node* var_node = cfg_->GetNodeByName(var, op);
if (NodeCanReused(var_node) && !pool_.Has(var_node)) {
pool_.Insert(var_node);
}
}
// fill the pool
for (auto& var : cfg_->Unlived(op)) {
ir::Node* var_node = cfg_->GetNodeByName(var, op);
if (var_node == nullptr || var_node->IsCtrlVar()) continue;
if (NodeCanReused(var_node) && !pool_.Has(var_node)) {
pool_.Insert(var_node);
}
}
}
......@@ -190,7 +192,8 @@ void MemoryOptimizePass::SubGraphOptimize(OpDesc* op_desc) const {
// effect. Because it is a single op in graph. No need to
// update the ir nodes.
sub_op_desc->Rename(var->Name(), cache->Name());
if (sub_op_desc->Block()->HasVar(var->Name())) {
if (sub_op_desc->Block() != nullptr &&
sub_op_desc->Block()->HasVar(var->Name())) {
sub_op_desc->Block()->RemoveVar(var->Name());
}
}
......@@ -231,7 +234,13 @@ void MemoryOptimizePass::RenameVarInGraphDesc(const std::string& var,
auto* op_desc = op->Op();
op_desc->RenameInput(var, cache_var);
op_desc->RenameOutput(var, cache_var);
if (op_desc->Block()->HasVar(var)) op_desc->Block()->RemoveVar(var);
if (op_desc->Block() != nullptr) {
op_desc->Block()->RemoveVar(var);
} else {
LOG(WARNING) << "op " << op->Name() << " not know its block."
<< "Is the op_desc created without block pointer? "
<< "Can not find " << var << " in Block(0)";
}
op_desc->Flush();
}
}
......@@ -273,8 +282,7 @@ void MemoryOptimizePass::RenameVarInGraphNode(const std::string& var,
// redirect the input to the latest version of cache_var
for (auto* node : op->inputs) {
if (node->Name() == var) {
ir::Node* cache_node = graph->CreateVarNode(var_desc.get());
var_nodes_[cache_var].emplace_back(cache_node);
ir::Node* cache_node = var_nodes_[cache_var].back();
// swap node to cache_node
cache_node->outputs.insert(cache_node->outputs.end(),
......@@ -283,11 +291,15 @@ void MemoryOptimizePass::RenameVarInGraphNode(const std::string& var,
auto* prev_op = node->inputs[0];
std::replace(prev_op->outputs.begin(), prev_op->outputs.end(), node,
cache_node);
cache_node->inputs.emplace_back(prev_op);
for (auto* next_op : node->outputs) {
std::replace(next_op->inputs.begin(), next_op->inputs.end(), node,
cache_node);
}
// erase unused node
auto& nodes = var_nodes_.at(var);
nodes.erase(std::remove(nodes.begin(), nodes.end(), node), nodes.end());
graph->RemoveNode(node);
}
}
......@@ -307,15 +319,14 @@ void MemoryOptimizePass::RenameVarInGraphNode(const std::string& var,
std::replace(next_op->inputs.begin(), next_op->inputs.end(), node,
cache_node);
}
// erase unused node
auto& nodes = var_nodes_.at(var);
nodes.erase(std::remove(nodes.begin(), nodes.end(), node), nodes.end());
graph->RemoveNode(node);
}
}
}
// release node of unused var in graph
for (auto* node : var_nodes_[var]) {
graph->RemoveNode(node);
}
var_nodes_.at(var).clear();
}
} // namespace details
......@@ -324,4 +335,4 @@ void MemoryOptimizePass::RenameVarInGraphNode(const std::string& var,
REGISTER_PASS(memory_optimize_pass,
paddle::framework::details::MemoryOptimizePass)
.RequireGraphAttr(paddle::framework::details::kAllOpDescs);
.RequireGraphAttr(paddle::framework::details::kStaleProgramOpDescs);
paddle/fluid/framework/details/multi_devices_graph_pass.cc
浏览文件 @ d497bd90
......@@ -392,20 +392,32 @@ void MultiDevSSAGraphBuilderBase::CreateComputationalOp(ir::Graph *result,
void MultiDevSSAGraphBuilderBase::CreateAllReduceOp(
ir::Graph *result, const std::string &og) const {
OpHandleBase *op_handle = nullptr;
auto append_allreduce_op = [&](
const std::vector<Scope *> &scopes,
const std::vector<platform::Place> &places) -> OpHandleBase * {
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
result->Get<GraphOps>(kGraphOps).emplace_back(new AllReduceOpHandle(
result->CreateEmptyNode("allreduce", ir::Node::Type::kOperation),
local_scopes_, places_, nccl_ctxs_));
result->Get<GraphOps>(kGraphOps).emplace_back(new AllReduceOpHandle(
result->CreateEmptyNode("allreduce", ir::Node::Type::kOperation),
scopes, places, nccl_ctxs_));
#else
result->Get<GraphOps>(kGraphOps).emplace_back(new AllReduceOpHandle(
result->CreateEmptyNode("allreduce", ir::Node::Type::kOperation),
local_scopes_, places_));
result->Get<GraphOps>(kGraphOps).emplace_back(new AllReduceOpHandle(
result->CreateEmptyNode("allreduce", ir::Node::Type::kOperation),
scopes, places));
#endif
auto *op_handle = result->Get<GraphOps>(kGraphOps).back();
return result->Get<GraphOps>(kGraphOps).back();
};
if (!strategy_.enable_parallel_graph_)
op_handle = append_allreduce_op(local_scopes_, places_);
for (size_t i = 0; i < places_.size(); ++i) {
auto &p = places_[i];
SetCommunicationContext(op_handle, p);
if (strategy_.enable_parallel_graph_) {
op_handle = append_allreduce_op({local_scopes_[i]}, {places_[i]});
}
SetCommunicationContext(op_handle, places_[i]);
auto &vars = result->Get<GraphVars>(kGraphVars)[i][og];
PADDLE_ENFORCE(!vars.empty());
auto &prev_grad = vars.back();
......@@ -413,7 +425,7 @@ void MultiDevSSAGraphBuilderBase::CreateAllReduceOp(
auto var =
new VarHandle(result->CreateEmptyNode(og, ir::Node::Type::kVariable),
vars.size(), i, og, p);
vars.size(), i, og, places_[i]);
vars.emplace_back(var);
op_handle->AddOutput(var);
}
......@@ -925,9 +937,21 @@ void DistSSAGraphBuilder::InsertCollectiveOp(ir::Graph *result,
}
void DistSSAGraphBuilder::InsertPostprocessOps(ir::Graph *result) const {
if (need_broadcast_var_ ||
(UseGPU() &&
strategy_.reduce_ == BuildStrategy::ReduceStrategy::kReduce)) {
// broad cast received parameters when training in parameter server mode.
if (need_broadcast_var_) {
// There are 4 conditions:
// 1. GPU && Reduce: Reduce gradient then broadcast gradient to other GPUS.
// Need to broadcast received parameters to other GPU.
// 2. GPU && AllReduce: AllReduce all graident to each GPU. Need to
// broadcast received parameters to other GPU.
// 3. CPU && AllReduce: AllReduce all gradient to each thread. Need to
// broadcast received parameters to other scope.
// 4. CPU && Reduce: because all parameters share the same memory, did not
// broadcast received parameters.
if (!UseGPU() &&
strategy_.reduce_ == BuildStrategy::ReduceStrategy::kReduce) {
return;
}
if (strategy_.fuse_broadcast_op_) {
CreateFusedBroadcastOp(result, bcast_var_name_set_);
} else {
......
paddle/fluid/framework/details/multi_devices_helper.h
浏览文件 @ d497bd90
......@@ -36,13 +36,14 @@ namespace details {
// map from variable name to variables. The variables, who have the same name,
// will have a differsent version. The offset in the
// `std::vector<VarHandle*>` is the version of varaibles.
typedef std::vector<std::unordered_map<std::string, std::vector<VarHandle*>>>
typedef std::vector<std::unordered_map<std::string, std::vector<VarHandle *>>>
GraphVars;
const char kGraphVars[] = "vars";
// aux variables to represent dependency. Useful to resolve data hazard.
typedef std::unordered_set<VarHandleBase*> GraphDepVars;
typedef std::unordered_set<VarHandleBase *> GraphDepVars;
const char kGraphDepVars[] = "dep_vars";
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/op_handle_base.h
浏览文件 @ d497bd90
......@@ -70,6 +70,9 @@ class OpHandleBase {
auto it = dev_ctxes_.find(place);
return it != dev_ctxes_.end() ? it->second : nullptr;
}
const std::map<platform::Place, platform::DeviceContext *> &DeviceContext() {
return dev_ctxes_;
}
void SetDeviceContext(platform::Place place, platform::DeviceContext *ctx_) {
dev_ctxes_[place] = ctx_;
......
paddle/fluid/framework/details/parallel_ssa_graph_executor.cc
浏览文件 @ d497bd90
......@@ -13,22 +13,85 @@
// limitations under the License.
#include "paddle/fluid/framework/details/parallel_ssa_graph_executor.h"
#include "paddle/fluid/framework/ir/graph_helper.h"
namespace paddle {
namespace framework {
namespace details {
std::vector<std::unique_ptr<ir::Graph>>
ParallelSSAGraphExecutor::SeparateMultiDevicesGraph(ir::Graph *graph) {
std::vector<std::unique_ptr<ir::Graph>> graphs;
graphs.reserve(places_.size());
for (size_t i = 0; i < places_.size(); ++i) {
ProgramDesc empty;
graphs.emplace_back(std::unique_ptr<ir::Graph>(new ir::Graph(empty)));
auto &g = graphs.back();
g->Set(kGraphVars, new GraphVars(1UL));
g->Set(kGraphDepVars, new GraphDepVars);
}
auto op_handles = ir::FilterByNodeWrapper<OpHandleBase>(*graph);
for (auto &op : op_handles) {
auto &dev_ctx = op->DeviceContext();
auto &p = dev_ctx.begin()->first;
int dev_id = boost::get<platform::CUDAPlace>(p).device;
auto &dev_dummys = graphs[dev_id]->Get<GraphDepVars>(kGraphDepVars);
graphs[dev_id]->AddNode(graph->RemoveNode(op->Node()).release());
for (auto &var : op->Inputs()) {
auto dummy_ptr = dynamic_cast<DummyVarHandle *>(var);
if (dummy_ptr) {
dev_dummys.insert(var);
if (graph->Nodes().count(var->Node()))
graphs[dev_id]->AddNode(graph->RemoveNode(var->Node()).release());
}
}
for (auto &var : op->Outputs()) {
auto dummy_ptr = dynamic_cast<DummyVarHandle *>(var);
if (dummy_ptr) {
dev_dummys.insert(var);
if (graph->Nodes().count(var->Node()))
graphs[dev_id]->AddNode(graph->RemoveNode(var->Node()).release());
}
}
}
for (size_t dev_id = 0; dev_id < places_.size(); ++dev_id) {
auto &dev_vars = graphs[dev_id]->Get<GraphVars>(kGraphVars)[0];
auto &origin_vars = graph->Get<GraphVars>(kGraphVars)[dev_id];
for (auto &name_pair : origin_vars) {
dev_vars.emplace(name_pair.first, name_pair.second);
for (auto &version_pair : name_pair.second) {
if (graph->Nodes().count(version_pair->Node())) {
graphs[dev_id]->AddNode(
graph->RemoveNode(version_pair->Node()).release());
}
}
}
}
return graphs;
}
ParallelSSAGraphExecutor::ParallelSSAGraphExecutor(
const ExecutionStrategy &strategy, const std::vector<Scope *> &local_scopes,
const std::vector<platform::Place> &places,
std::vector<std::unique_ptr<ir::Graph>> &&graphs)
const std::vector<platform::Place> &places, ir::Graph *graph)
: strategy_(std::move(strategy)),
local_scopes_(std::move(local_scopes)),
pool_(places.size() >= 2 ? new ::ThreadPool(places.size()) : nullptr),
places_(std::move(places)),
graphs_(std::move(graphs)) {
// TODO(Yancey1989): Copying graphs is not safely since it deleted the
// attrs.
graphs_(SeparateMultiDevicesGraph(graph)) {
PADDLE_ENFORCE_EQ(places_.size(), local_scopes_.size());
auto seq_allreduce_pass =
ir::PassRegistry::Instance().Get("all_reduce_deps_pass");
for (size_t i = 0; i < graphs_.size(); ++i) {
graphs_[i] = seq_allreduce_pass->Apply(std::move(graphs_[i]));
}
// set the correct size of thread pool to each device.
strategy_.num_threads_ = strategy_.num_threads_ < places_.size()
? 1UL
......@@ -37,7 +100,7 @@ ParallelSSAGraphExecutor::ParallelSSAGraphExecutor(
<< " to run the operators of the graph on each device.";
for (size_t i = 0; i < places.size(); ++i) {
executors_.emplace_back(new details::ThreadedSSAGraphExecutor(
strategy_, {local_scopes_[i]}, {places_[i]}, std::move(graphs_[i])));
strategy_, local_scopes_, {places_[i]}, graphs_.at(i).get()));
}
}
......
paddle/fluid/framework/details/parallel_ssa_graph_executor.h
浏览文件 @ d497bd90
......@@ -18,7 +18,9 @@
#include <vector>
#include "ThreadPool.h"
#include "paddle/fluid/framework/details/multi_devices_helper.h"
#include "paddle/fluid/framework/details/threaded_ssa_graph_executor.h"
#include "paddle/fluid/framework/ir/graph.h"
namespace paddle {
namespace framework {
......@@ -29,13 +31,17 @@ class ParallelSSAGraphExecutor : public SSAGraphExecutor {
ParallelSSAGraphExecutor(const ExecutionStrategy &strategy,
const std::vector<Scope *> &local_scopes,
const std::vector<platform::Place> &places,
std::vector<std::unique_ptr<ir::Graph>> &&graphs);
ir::Graph *graph);
~ParallelSSAGraphExecutor() final = default;
const ir::Graph &Graph() const override { return *graphs_[0]; }
FeedFetchList Run(const std::vector<std::string> &fetch_tensors) override;
private:
std::vector<std::unique_ptr<ir::Graph>> SeparateMultiDevicesGraph(
ir::Graph *graph);
ExecutionStrategy strategy_;
std::vector<Scope *> local_scopes_;
std::unique_ptr<::ThreadPool> pool_{nullptr};
......
paddle/fluid/framework/details/reduce_op_handle.cc
浏览文件 @ d497bd90
......@@ -139,7 +139,7 @@ void ReduceOpHandle::GatherSelectedRows(
#endif
void ReduceOpHandle::RunImpl() {
platform::RecordEvent record_event(Name(), dev_ctxes_.cbegin()->second);
platform::RecordEvent record_event(Name());
if (places_.size() == 1) return;
// the input and output may have dummy var.
......@@ -153,7 +153,7 @@ void ReduceOpHandle::RunImpl() {
{
auto out_var_handles = DynamicCast<VarHandle>(outputs_);
PADDLE_ENFORCE_EQ(out_var_handles.size(), 1,
PADDLE_ENFORCE_EQ(out_var_handles.size(), 1UL,
"The number of output should be one.");
out_var_handle = out_var_handles.front();
}
......
paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.cc
浏览文件 @ d497bd90
......@@ -63,7 +63,7 @@ FeedFetchList ScopeBufferedSSAGraphExecutor::Run(
eptr = std::current_exception();
}
platform::RecordEvent e("ScopeBufferedSSAGraphExecutorAfterRun", nullptr);
platform::RecordEvent e("ScopeBufferedSSAGraphExecutorAfterRun");
++drop_scope_counter_;
bool stream_end = false;
......
paddle/fluid/framework/details/sequential_execution_pass.cc
浏览文件 @ d497bd90
......@@ -40,7 +40,7 @@ std::unique_ptr<ir::Graph> SequentialExecutionPass::ApplyImpl(
static std::unordered_set<std::string> skip_dist_ops{
"send", "recv", "send_barrier", "fetch_barrier"};
auto &ops = Get<const std::vector<OpDesc *>>(kAllOpDescs);
auto &ops = graph->Get<const std::vector<OpDesc *>>(kStaleProgramOpDescs);
std::vector<ir::Node *> op_node_list;
op_node_list.reserve(ops.size());
......@@ -107,4 +107,4 @@ std::unique_ptr<ir::Graph> SequentialExecutionPass::ApplyImpl(
REGISTER_PASS(sequential_execution_pass,
paddle::framework::details::SequentialExecutionPass)
.RequirePassAttr(paddle::framework::details::kAllOpDescs);
.RequireGraphAttr(paddle::framework::details::kStaleProgramOpDescs);
paddle/fluid/framework/details/threaded_ssa_graph_executor.cc
浏览文件 @ d497bd90
......@@ -23,9 +23,8 @@ namespace framework {
namespace details {
ThreadedSSAGraphExecutor::ThreadedSSAGraphExecutor(
const ExecutionStrategy &strategy, const std::vector<Scope *> &local_scopes,
const std::vector<platform::Place> &places,
std::unique_ptr<ir::Graph> &&graph)
: graph_(std::move(graph)),
const std::vector<platform::Place> &places, ir::Graph *graph)
: graph_(graph),
pool_(strategy.num_threads_ >= 2 ? new ::ThreadPool(strategy.num_threads_)
: nullptr),
local_scopes_(local_scopes),
......@@ -37,7 +36,7 @@ ThreadedSSAGraphExecutor::ThreadedSSAGraphExecutor(
FeedFetchList ThreadedSSAGraphExecutor::Run(
const std::vector<std::string> &fetch_tensors) {
std::unique_ptr<platform::RecordEvent> event(
new platform::RecordEvent("ThreadedSSAGraphExecutorPrepare", nullptr));
new platform::RecordEvent("ThreadedSSAGraphExecutorPrepare"));
std::unordered_map<OpHandleBase *, size_t> pending_ops;
std::unordered_set<VarHandleBase *> pending_vars;
auto ready_vars = std::make_shared<BlockingQueue<VarHandleBase *>>();
......@@ -110,7 +109,7 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
for (auto &run_op_future : run_op_futures_) {
run_op_future.wait();
}
ClearFetchOp(graph_.get(), &fetch_ops);
ClearFetchOp(graph_, &fetch_ops);
exception_holder_.ReThrow();
} else {
continue;
......@@ -135,7 +134,7 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
}
PADDLE_ENFORCE(ready_ops.empty());
// Wait FetchOps.
ClearFetchOp(graph_.get(), &fetch_ops);
ClearFetchOp(graph_, &fetch_ops);
return fetch_data;
}
......@@ -219,7 +218,7 @@ void ThreadedSSAGraphExecutor::RunOp(
VLOG(10) << op << " " << op->Name() << " Done ";
running_ops_--;
ready_var_q->Extend(op->Outputs());
VLOG(10) << op << " " << op->Name() << "Signal posted";
VLOG(10) << op << " " << op->Name() << " Signal posted";
} catch (...) {
exception_holder_.Catch(std::current_exception());
}
......
paddle/fluid/framework/details/threaded_ssa_graph_executor.h
浏览文件 @ d497bd90
......@@ -41,7 +41,7 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor {
ThreadedSSAGraphExecutor(const ExecutionStrategy &strategy,
const std::vector<Scope *> &local_scopes,
const std::vector<platform::Place> &places,
std::unique_ptr<ir::Graph> &&graph);
ir::Graph *graph);
const ir::Graph &Graph() const override { return *graph_; }
// Run a SSAGraph by a thread pool
......@@ -55,7 +55,7 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor {
details::OpHandleBase *op);
private:
std::unique_ptr<ir::Graph> graph_;
ir::Graph *graph_;
std::unique_ptr<::ThreadPool> pool_;
std::vector<Scope *> local_scopes_;
std::vector<platform::Place> places_;
......
paddle/fluid/framework/executor.cc
浏览文件 @ d497bd90
......@@ -20,6 +20,7 @@ limitations under the License. */
#include "paddle/fluid/framework/lod_tensor_array.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/reader.h"
#include "paddle/fluid/framework/threadpool.h"
#include "paddle/fluid/framework/transfer_scope_cache.h"
#include "paddle/fluid/framework/variable_helper.h"
#include "paddle/fluid/operators/distributed/distributed.h"
......
paddle/fluid/framework/inplace_op_inference_test.cc
浏览文件 @ d497bd90
......@@ -179,11 +179,11 @@ TEST(InferInplace, SingleOpInplaceInToOut) {
op->SetOutput("Out", {"test2_out"});
prog.MutableBlock(0)->Var("test2_a")->SetType(proto::VarType::LOD_TENSOR);
prog.MutableBlock(0)->Var("test2_a")->SetShape({32, 64});
prog.MutableBlock(0)->Var("test2_a")->SetShape({32, 64, 128, 128});
prog.MutableBlock(0)->Var("test2_b")->SetType(proto::VarType::LOD_TENSOR);
prog.MutableBlock(0)->Var("test2_c")->SetType(proto::VarType::LOD_TENSOR);
prog.MutableBlock(0)->Var("test2_out");
prog.MutableBlock(0)->Var("test2_out")->SetShape({32, 16});
prog.MutableBlock(0)->Var("test2_out")->SetShape({32, 16, 128, 128});
auto& infer_inplace = OpInfoMap::Instance().Get(op->Type()).infer_inplace_;
auto in_to_outs = infer_inplace(*op, op->Block());
......@@ -201,11 +201,11 @@ TEST(InferInplace, SingleGradOpInplaceInToOut) {
op->SetOutput(GradVarName("X"), {"test2_a", "test2_b", "test2_c"});
prog.MutableBlock(0)->Var("test2_a")->SetType(proto::VarType::LOD_TENSOR);
prog.MutableBlock(0)->Var("test2_a")->SetShape({32, 16});
prog.MutableBlock(0)->Var("test2_a")->SetShape({32, 16, 1024, 1024});
prog.MutableBlock(0)->Var("test2_b")->SetType(proto::VarType::LOD_TENSOR);
prog.MutableBlock(0)->Var("test2_c")->SetType(proto::VarType::LOD_TENSOR);
prog.MutableBlock(0)->Var("test2_out");
prog.MutableBlock(0)->Var("test2_out")->SetShape({32, 16});
prog.MutableBlock(0)->Var("test2_out")->SetShape({32, 16, 1024, 1024});
auto& infer_inplace = OpInfoMap::Instance().Get(op->Type()).infer_inplace_;
auto in_to_outs = infer_inplace(*op, op->Block());
......@@ -233,12 +233,12 @@ TEST(InferInplace, MultiOutInplaceInToOut) {
prog.MutableBlock(0)->Var("o0");
prog.MutableBlock(0)->Var("y0");
prog.MutableBlock(0)->Var("z0");
prog.MutableBlock(0)->Var("a0")->SetShape({32, 16});
prog.MutableBlock(0)->Var("b0")->SetShape({32, 16});
prog.MutableBlock(0)->Var("c0")->SetShape({32, 16});
prog.MutableBlock(0)->Var("o0")->SetShape({32, 16});
prog.MutableBlock(0)->Var("y0")->SetShape({32, 16});
prog.MutableBlock(0)->Var("z0")->SetShape({32, 16});
prog.MutableBlock(0)->Var("a0")->SetShape({32, 16, 1024, 1024});
prog.MutableBlock(0)->Var("b0")->SetShape({32, 16, 1024, 1024});
prog.MutableBlock(0)->Var("c0")->SetShape({32, 16, 1024, 1024});
prog.MutableBlock(0)->Var("o0")->SetShape({32, 16, 1024, 1024});
prog.MutableBlock(0)->Var("y0")->SetShape({32, 16, 1024, 1024});
prog.MutableBlock(0)->Var("z0")->SetShape({32, 16, 1024, 1024});
auto& infer_inplace = OpInfoMap::Instance().Get(op->Type()).infer_inplace_;
auto in_to_outs = infer_inplace(*op, op->Block());
......@@ -267,12 +267,12 @@ TEST(InferInplace, MultiGradInplaceInToOut) {
prog.MutableBlock(0)->Var("o0");
prog.MutableBlock(0)->Var("y0");
prog.MutableBlock(0)->Var("z0");
prog.MutableBlock(0)->Var("a0")->SetShape({32, 16});
prog.MutableBlock(0)->Var("b0")->SetShape({32, 16});
prog.MutableBlock(0)->Var("c0")->SetShape({32, 16});
prog.MutableBlock(0)->Var("o0")->SetShape({32, 16});
prog.MutableBlock(0)->Var("y0")->SetShape({32, 16});
prog.MutableBlock(0)->Var("z0")->SetShape({32, 16});
prog.MutableBlock(0)->Var("a0")->SetShape({32, 16, 1024, 1024});
prog.MutableBlock(0)->Var("b0")->SetShape({32, 16, 1024, 1024});
prog.MutableBlock(0)->Var("c0")->SetShape({32, 16, 1024, 1024});
prog.MutableBlock(0)->Var("o0")->SetShape({32, 16, 1024, 1024});
prog.MutableBlock(0)->Var("y0")->SetShape({32, 16, 1024, 1024});
prog.MutableBlock(0)->Var("z0")->SetShape({32, 16, 1024, 1024});
auto& infer_inplace = OpInfoMap::Instance().Get(op->Type()).infer_inplace_;
auto in_to_outs = infer_inplace(*op, op->Block());
......
paddle/fluid/framework/ir/CMakeLists.txt
浏览文件 @ d497bd90
......@@ -102,6 +102,8 @@ cc_test(test_seqpool_concat_fuse_pass SRCS seqpool_concat_fuse_pass_tester.cc DE
cc_test(test_is_test_pass SRCS is_test_pass_tester.cc DEPS is_test_pass)
if (WITH_MKLDNN)
cc_test(test_depthwise_conv_mkldnn_pass SRCS mkldnn/depthwise_conv_mkldnn_pass_tester.cc DEPS depthwise_conv_mkldnn_pass)
cc_test(test_conv_bias_mkldnn_fuse_pass SRCS mkldnn/conv_bias_mkldnn_fuse_pass_tester.cc DEPS conv_bias_mkldnn_fuse_pass naive_executor)
cc_test(test_conv_relu_mkldnn_fuse_pass SRCS mkldnn/conv_relu_mkldnn_fuse_pass_tester.cc DEPS conv_relu_mkldnn_fuse_pass)
cc_test(test_conv_elementwise_add_mkldnn_fuse_pass SRCS mkldnn/conv_elementwise_add_mkldnn_fuse_pass_tester.cc DEPS conv_elementwise_add_mkldnn_fuse_pass)
cc_test(test_mkldnn_placement_pass SRCS mkldnn/mkldnn_placement_pass_tester.cc DEPS mkldnn_placement_pass)
endif ()
paddle/fluid/framework/ir/attention_lstm_fuse_pass.h
浏览文件 @ d497bd90
......@@ -22,7 +22,8 @@ namespace ir {
class AttentionLSTMFusePass : public FusePassBase {
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_affine_channel_fuse_pass.h
浏览文件 @ d497bd90
......@@ -31,7 +31,8 @@ class ConvAffineChannelFusePass : public FusePassBase {
virtual ~ConvAffineChannelFusePass() {}
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;
const std::string name_scope_{"conv_affine_channel_fuse"};
};
......@@ -40,7 +41,8 @@ class ConvEltwiseAddAffineChannelFusePass : public FusePassBase {
virtual ~ConvEltwiseAddAffineChannelFusePass() {}
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;
const std::string name_scope_{"conv_eltwiseadd_affine_channel_fuse"};
};
......
paddle/fluid/framework/ir/conv_bn_fuse_pass.cc
浏览文件 @ d497bd90
......@@ -169,7 +169,7 @@ std::unique_ptr<ir::Graph> ConvBNFusePass::ApplyImpl(
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);
PADDLE_ENFORCE_EQ(conv_bias_names.size(), 1UL);
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(),
......
paddle/fluid/framework/ir/conv_bn_fuse_pass.h
浏览文件 @ d497bd90
......@@ -31,7 +31,8 @@ class ConvBNFusePass : public FusePassBase {
virtual ~ConvBNFusePass() {}
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;
const std::string name_scope_{"conv_bn_fuse"};
};
......@@ -40,7 +41,8 @@ class ConvEltwiseAddBNFusePass : public FusePassBase {
virtual ~ConvEltwiseAddBNFusePass() {}
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;
const std::string name_scope_{"conv_eltwiseadd_bn_fuse"};
};
......
paddle/fluid/framework/ir/conv_elementwise_add2_act_fuse_pass.h
浏览文件 @ d497bd90
......@@ -25,7 +25,8 @@ class ConvElementwiseAdd2ActFusePass : public FusePassBase {
virtual ~ConvElementwiseAdd2ActFusePass() {}
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_elementwise_add_act_fuse_pass.h
浏览文件 @ d497bd90
......@@ -25,7 +25,8 @@ class ConvElementwiseAddActFusePass : public FusePassBase {
virtual ~ConvElementwiseAddActFusePass() {}
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_elementwise_add_fuse_pass.h
浏览文件 @ d497bd90
......@@ -25,7 +25,8 @@ class ConvElementwiseAddFusePass : public FusePassBase {
virtual ~ConvElementwiseAddFusePass() {}
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/embedding_fc_lstm_fuse_pass.h
浏览文件 @ d497bd90
......@@ -14,6 +14,8 @@
#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"
......@@ -30,7 +32,8 @@ class EmbeddingFCLSTMFusePass : public FusePassBase {
virtual ~EmbeddingFCLSTMFusePass() {}
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;
const std::string name_scope_{"embedding_fc_lstm_fuse"};
};
......
paddle/fluid/framework/ir/fc_fuse_pass.h
浏览文件 @ d497bd90
......@@ -12,6 +12,8 @@
// 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/graph_pattern_detector.h"
......@@ -29,7 +31,8 @@ class FCFusePass : public FusePassBase {
virtual ~FCFusePass() {}
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/fc_gru_fuse_pass.h
浏览文件 @ d497bd90
......@@ -30,7 +30,8 @@ class FCGRUFusePass : public FusePassBase {
virtual ~FCGRUFusePass() {}
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;
const std::string name_scope_{"fc_gru_fuse"};
};
......@@ -41,7 +42,8 @@ class MulGRUFusePass : public FusePassBase {
virtual ~MulGRUFusePass() {}
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;
const std::string name_scope_{"fc_nobias_gru_fuse"};
};
......
paddle/fluid/framework/ir/fc_lstm_fuse_pass.h
浏览文件 @ d497bd90
......@@ -14,6 +14,8 @@
#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"
......@@ -30,7 +32,8 @@ class FCLstmFusePass : public FusePassBase {
virtual ~FCLstmFusePass() {}
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;
const std::string name_scope_{"fc_lstm_fuse"};
};
......@@ -40,7 +43,8 @@ class MulLstmFusePass : public FusePassBase {
virtual ~MulLstmFusePass() {}
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;
const std::string name_scope_{"fc_nobias_lstm_fuse"};
};
......
paddle/fluid/framework/ir/fuse_elewise_add_act_pass.h
浏览文件 @ d497bd90
......@@ -32,7 +32,8 @@ class FuseElewiseAddActPass : public FusePassBase {
virtual ~FuseElewiseAddActPass() {}
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;
std::unique_ptr<ir::Graph> FuseElewiseAddAct(
std::unique_ptr<ir::Graph> graph,
......
paddle/fluid/framework/ir/fuse_relu_depthwise_conv_pass.cc
浏览文件 @ d497bd90
......@@ -111,7 +111,7 @@ std::unique_ptr<ir::Graph> FuseReluDepthwiseConvPass::FuseReluDepthwiseConv(
xg_var = subgraph.at(xg)->Var();
}
PADDLE_ENFORCE_EQ(layer_op->Input("Input").size(), 1);
PADDLE_ENFORCE_EQ(layer_op->Input("Input").size(), 1UL);
PADDLE_ENFORCE_EQ(layer_op->Input("Input")[0], y_var->Name());
layer_op->SetInput("Input", {x_var->Name()});
subgraph.at(layer)->inputs.push_back(subgraph.at(x));
......@@ -119,13 +119,13 @@ std::unique_ptr<ir::Graph> FuseReluDepthwiseConvPass::FuseReluDepthwiseConv(
VLOG(4) << "replace " << y_var->Name() << " -> " << x_var->Name();
if (!only_forward) {
PADDLE_ENFORCE_EQ(layer_g_op->Input("Input").size(), 1);
PADDLE_ENFORCE_EQ(layer_g_op->Input("Input").size(), 1UL);
PADDLE_ENFORCE_EQ(layer_g_op->Input("Input")[0], y_var->Name());
layer_g_op->SetInput("Input", {x_var->Name()});
subgraph.at(layer_g)->inputs.push_back(subgraph.at(x));
subgraph.at(x)->outputs.push_back(subgraph.at(layer_g));
PADDLE_ENFORCE_EQ(layer_g_op->Output(GradVarName("Input")).size(), 1);
PADDLE_ENFORCE_EQ(layer_g_op->Output(GradVarName("Input")).size(), 1UL);
PADDLE_ENFORCE_EQ(layer_g_op->Output(GradVarName("Input"))[0],
yg_var->Name());
layer_g_op->SetOutput(GradVarName("Input"), {xg_var->Name()});
......
paddle/fluid/framework/ir/fuse_relu_depthwise_conv_pass.h
浏览文件 @ d497bd90
......@@ -32,7 +32,8 @@ class FuseReluDepthwiseConvPass : public FusePassBase {
virtual ~FuseReluDepthwiseConvPass() {}
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;
std::unique_ptr<ir::Graph> FuseReluDepthwiseConv(
std::unique_ptr<ir::Graph> graph, bool only_forward) const;
};
......
paddle/fluid/framework/ir/graph.cc
浏览文件 @ d497bd90
......@@ -76,6 +76,9 @@ std::map<std::string, std::vector<ir::Node *>> Graph::InitFromProgram(
var->inputs.push_back(node);
}
}
Set<const std::vector<OpDesc *>>(
details::kStaleProgramOpDescs,
new std::vector<OpDesc *>(program.Block(0).AllOps()));
return var_nodes;
}
......
paddle/fluid/framework/ir/graph.h
浏览文件 @ d497bd90
......@@ -26,6 +26,14 @@ limitations under the License. */
namespace paddle {
namespace framework {
namespace details {
// This attr is not recommended, because the graph should not dependence
// the program once it is built.
constexpr char kStaleProgramOpDescs[] = "stale_program_op_descs";
} // namespace details
namespace ir {
/*
......@@ -168,10 +176,13 @@ class Graph {
return ret;
}
void RemoveNode(ir::Node *node) {
std::unique_ptr<ir::Node> RemoveNode(ir::Node *node) {
PADDLE_ENFORCE(node_set_.find(node) != node_set_.end());
node_set_.erase(node);
std::unique_ptr<ir::Node> ret;
ret.reset(nodes_.at(node).release());
nodes_.erase(node);
node_set_.erase(node);
return ret;
}
// NOTE low performance, but simple and secure.
......@@ -184,12 +195,11 @@ class Graph {
return nullptr;
}
void ResolveHazard(
const std::map<std::string, std::vector<ir::Node *>> &var_nodes);
private:
std::map<std::string, std::vector<ir::Node *>> InitFromProgram(
const ProgramDesc &program);
// Returns reference to the original program.
// WARN: After a series of passes, the current graph can be quite
// different from OriginProgram. Caller shouldn't assume much from
// the returned OriginProgram.
const ProgramDesc &OriginProgram() const { return program_; }
// This method takes ownership of `node`.
ir::Node *AddNode(ir::Node *node) {
......@@ -199,6 +209,13 @@ class Graph {
return node;
}
void ResolveHazard(
const std::map<std::string, std::vector<ir::Node *>> &var_nodes);
private:
std::map<std::string, std::vector<ir::Node *>> InitFromProgram(
const ProgramDesc &program);
// NOTE: program_ shouldn't be exposed to user.
const ProgramDesc program_;
std::map<std::string, boost::any> attrs_;
......
paddle/fluid/framework/ir/graph_pattern_detector.cc
浏览文件 @ d497bd90
......@@ -38,7 +38,7 @@ size_t PDPattern::id_ = 0UL;
PDNode *PDPattern::NewNode(const std::string &name) {
if (!name.empty()) {
PADDLE_ENFORCE_EQ(node_map_.count(name), 0,
PADDLE_ENFORCE_EQ(node_map_.count(name), 0UL,
"PDNode's name should be unique, get duplicate [%s]",
name);
}
......@@ -51,7 +51,7 @@ PDNode *PDPattern::NewNode(const std::string &name) {
PDNode *PDPattern::NewNode(PDNode::teller_t &&teller, const std::string &name) {
if (!name.empty()) {
PADDLE_ENFORCE_EQ(node_map_.count(name), 0,
PADDLE_ENFORCE_EQ(node_map_.count(name), 0UL,
"PDNode's name should be unique, get duplicate [%s]",
name);
}
......
paddle/fluid/framework/ir/identity_scale_op_clean_pass.h
浏览文件 @ d497bd90
......@@ -22,7 +22,8 @@ namespace ir {
class IdentityScaleOpCleanPass : public FusePassBase {
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;
private:
virtual ~IdentityScaleOpCleanPass() = default;
......
paddle/fluid/framework/ir/lock_free_optimize_pass.h
浏览文件 @ d497bd90
......@@ -60,7 +60,8 @@ class LockFreeOptimizePass : public Pass {
virtual ~LockFreeOptimizePass() {}
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;
private:
// Create a new sgd node via current optimizer node
......
paddle/fluid/framework/ir/mkldnn/conv_bias_mkldnn_fuse_pass.h
浏览文件 @ d497bd90
......@@ -29,7 +29,8 @@ class ConvBiasFusePass : public FusePassBase {
virtual bool is_conv3d() const { return false; }
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;
const std::string name_scope_{"conv_bias_mkldnn_fuse"};
};
/*
......
paddle/fluid/framework/ir/mkldnn/conv_bias_mkldnn_fuse_pass_tester.cc 0 → 100644
浏览文件 @ d497bd90
// Copyright (c) 2019 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/conv_bias_mkldnn_fuse_pass.h"
#include "paddle/fluid/framework/naive_executor.h"
#include "paddle/fluid/platform/place.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, const std::string& name,
const std::vector<std::string>& inputs,
const std::vector<std::string>& outputs) {
auto* op = prog->MutableBlock(0)->AppendOp();
op->SetType(type);
if (type == "conv2d") {
op->SetAttr("use_mkldnn", true);
op->SetAttr("name", name);
op->SetInput("Input", {inputs[0]});
op->SetInput("Filter", {inputs[1]});
if (inputs.size() > 2)
op->SetInput("Bias", {inputs[2]});
else
op->SetInput("Bias", {});
} else if (type == "elementwise_add") {
op->SetAttr("use_mkldnn", true);
op->SetInput("X", {inputs[0]});
op->SetInput("Y", {inputs[1]});
}
op->SetOutput("Out", outputs);
op->SetAttr(OpProtoAndCheckerMaker::OpRoleAttrName(),
static_cast<int>(OpRole::kForward));
}
// (c, weights)->conv->f
// (f)->elementwise_add->g
ProgramDesc BuildProgramDesc(bool convWithExistingBias) {
ProgramDesc prog;
std::vector<std::string> nodes{"c", "weights", "f", "eltwise_bias", "g"};
if (convWithExistingBias) nodes.push_back("conv_bias");
for (auto& v : nodes) {
auto* var = prog.MutableBlock(0)->Var(v);
var->SetType(proto::VarType::LOD_TENSOR);
if (v == "weights" || v == "conv_bias" || v == "eltwise_bias") {
var->SetPersistable(true);
}
}
// conv+bias, both with MKL-DNN
if (convWithExistingBias) {
SetOp(&prog, "conv2d", "conv",
std::vector<std::string>({"c", "weights", "conv_bias"}),
std::vector<std::string>({"f"}));
} else {
SetOp(&prog, "conv2d", "conv", std::vector<std::string>({"c", "weights"}),
std::vector<std::string>({"f"}));
}
SetOp(&prog, "elementwise_add", "eltwise",
std::vector<std::string>({"f", "eltwise_bias"}),
std::vector<std::string>({"g"}));
return prog;
}
void InitTensorHolder(Scope* scope, const paddle::platform::Place& place,
const char* var_name) {
auto x = scope->Var(var_name);
auto tensor = x->GetMutable<LoDTensor>();
tensor->mutable_data(place, proto::VarType::FP32,
::paddle::memory::Allocator::kDefault, 1);
}
void MainTest(bool convWithExistingBias) {
auto prog = BuildProgramDesc(convWithExistingBias);
std::unique_ptr<ir::Graph> graph(new ir::Graph(prog));
auto place = paddle::platform::CPUPlace();
NaiveExecutor exe{place};
Scope scope;
// Init scope, as it is used in pass
exe.CreateVariables(prog, 0, true, &scope);
if (convWithExistingBias) {
InitTensorHolder(&scope, place, "conv_bias");
InitTensorHolder(&scope, place, "eltwise_bias");
}
graph->Set(kParamScopeAttr, new framework::Scope*(&scope));
auto pass = PassRegistry::Instance().Get("conv_bias_mkldnn_fuse_pass");
int original_nodes_num = graph->Nodes().size();
graph = pass->Apply(std::move(graph));
int current_nodes_num = graph->Nodes().size();
// Remove 3 Nodes: Conv, Bias, conv_out
// Add 1 Node: ConvBias
EXPECT_EQ(original_nodes_num - 2, current_nodes_num);
// Assert conv_bias op in newly generated graph
int conv_bias_count = 0;
for (auto* node : graph->Nodes()) {
if (node->IsOp() && node->Op()->Type() == "conv2d") {
auto* op = node->Op();
ASSERT_TRUE(op->HasAttr("use_mkldnn"));
EXPECT_TRUE(boost::get<bool>(op->GetAttr("use_mkldnn")));
// check if "conv" convolution is fused
auto op_name = boost::get<std::string>(op->GetAttr("name"));
if (op_name == "conv") {
auto input_names = op->InputNames();
ASSERT_TRUE(std::find(input_names.begin(), input_names.end(), "Bias") !=
input_names.end());
auto bias = boost::get<std::vector<std::string>>(op->Input("Bias"));
if (bias.size()) {
++conv_bias_count;
}
}
}
}
EXPECT_EQ(conv_bias_count, 1);
}
TEST(ConvBiasFusePass, bias_free_conv) { MainTest(false); }
TEST(ConvBiasFusePass, conv_with_existing_bias) { MainTest(true); }
TEST(ConvBiasFusePass, conv3d) {
Conv3DBiasFusePass pass;
ASSERT_TRUE(pass.is_conv3d());
}
} // namespace ir
} // namespace framework
} // namespace paddle
USE_PASS(conv_bias_mkldnn_fuse_pass);
paddle/fluid/framework/ir/mkldnn/conv_elementwise_add_mkldnn_fuse_pass_tester.cc
浏览文件 @ d497bd90
......@@ -44,10 +44,14 @@ struct TestIsReachable {
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* {
auto hash = [](const Node* node) -> std::string {
return node->Name() + std::to_string(node->id());
};
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()) {
if (name == hash(&node)) {
return &node;
}
}
......@@ -55,13 +59,17 @@ struct TestIsReachable {
return nullptr;
};
return [&](std::string from, const std::string to) -> bool {
// update the from and to strings to hashed equivs in loop from graph traits
return [&](std::string from, 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;
auto hashed = hash(&node);
if (node.Name() == from) from = hashed;
if (node.Name() == to) to = hashed;
visited[hashed] = false;
}
visited[from] = true;
......@@ -72,15 +80,15 @@ struct TestIsReachable {
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;
auto hashed_name = hash(n);
if (hashed_name == to) return true;
if (!visited[n->Name()]) {
visited[n->Name()] = true;
queue.push_back(n->Name());
if (!visited[hashed_name]) {
visited[hashed_name] = true;
queue.push_back(hashed_name);
}
}
}
......@@ -166,6 +174,28 @@ TEST(ConvElementwiseAddMKLDNNFusePass, ConvolutionAsYWithElementwiseAddRelu) {
RunPassAndAssert(&prog, "a", "relu", 1);
}
TEST(ConvElementwiseAddMKLDNNFusePass,
ConvolutionProjectionAsYWithElementwiseAddRelu) {
auto prog = BuildProgramDesc({"a", "b", "c", "d", "e", "f"},
{"bias", "weights", "bias2", "weights2"});
SetOp(&prog, "sigmoid", {{"X", "a"}}, {"Out", "b"});
// right branch
SetOp(&prog, "conv2d",
{{"Input", "b"}, {"Bias", "bias"}, {"Filter", "weights"}},
{"Output", "c"});
// left branch
SetOp(&prog, "conv2d",
{{"Input", "a"}, {"Bias", "bias2"}, {"Filter", "weights2"}},
{"Output", "f"});
SetOp(&prog, "elementwise_add", {{"X", "f"}, {"Y", "c"}}, {"Out", "d"});
SetOp(&prog, "relu", {{"X", "d"}}, {"Out", "e"});
RunPassAndAssert(&prog, "a", "relu", 2);
}
TEST(ConvElementwiseAddMKLDNNFusePass,
ConvolutionAsYWithElementwiseAddReluNoBias) {
auto prog = BuildProgramDesc({"a", "b", "c", "d", "e"}, {"weights"});
......
paddle/fluid/framework/ir/mkldnn/mkldnn_placement_pass.cc
浏览文件 @ d497bd90
......@@ -21,7 +21,7 @@ namespace ir {
std::unique_ptr<ir::Graph> MKLDNNPlacementPass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
VLOG(3) << "Aplies MKL-DNN placement strategy.";
VLOG(3) << "Applies MKL-DNN placement strategy.";
const auto& op_types_list =
Get<std::unordered_set<std::string>>("mkldnn_enabled_op_types");
for (const Node* n : graph->Nodes()) {
......
paddle/fluid/framework/ir/mkldnn/mkldnn_placement_pass_tester.cc 0 → 100644
浏览文件 @ d497bd90
// Copyright (c) 2019 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/mkldnn_placement_pass.h"
#include <gtest/gtest.h>
#include <boost/logic/tribool.hpp>
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, boost::tribool use_mkldnn) {
auto* op = prog->MutableBlock(0)->AppendOp();
op->SetType(type);
if (!boost::indeterminate(use_mkldnn)) op->SetAttr("use_mkldnn", use_mkldnn);
if (type == "conv2d") {
op->SetAttr("name", name);
op->SetInput("Input", {inputs[0]});
op->SetInput("Filter", {inputs[1]});
op->SetInput("Bias", {inputs[2]});
} else if (type == "relu") {
op->SetInput("X", inputs);
} else if (type == "concat") {
op->SetAttr("axis", 1);
op->SetInput("X", {inputs[0], inputs[1]});
} else if (type == "pool2d") {
op->SetInput("X", {inputs[0]});
} else {
FAIL() << "Unexpected operator type.";
}
op->SetOutput("Out", {outputs[0]});
}
// operator use_mkldnn
// ---------------------------------------
// (a,b)->concat->c none
// (c,weights,bias)->conv->f none
// f->relu->g false
// g->pool->h false
// (h,weights2,bias2)->conv->k true
// k->relu->l true
ProgramDesc BuildProgramDesc() {
ProgramDesc prog;
for (auto& v :
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") {
var->SetPersistable(true);
}
}
SetOp(&prog, "concat", "concat1", std::vector<std::string>({"a", "b"}),
std::vector<std::string>({"c"}), boost::indeterminate);
SetOp(&prog, "conv2d", "conv1",
std::vector<std::string>({"c", "weights", "bias"}),
std::vector<std::string>({"f"}), boost::indeterminate);
SetOp(&prog, "relu", "relu1", std::vector<std::string>({"f"}),
std::vector<std::string>({"g"}), false);
SetOp(&prog, "pool2d", "pool1", std::vector<std::string>({"g"}),
std::vector<std::string>({"h"}), false);
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"}), true);
return prog;
}
void MainTest(std::initializer_list<std::string> mkldnn_enabled_op_types,
unsigned expected_use_mkldnn_true_count) {
auto prog = BuildProgramDesc();
std::unique_ptr<ir::Graph> graph(new ir::Graph(prog));
auto pass = PassRegistry::Instance().Get("mkldnn_placement_pass");
pass->Set("mkldnn_enabled_op_types",
new std::unordered_set<std::string>(mkldnn_enabled_op_types));
graph = pass->Apply(std::move(graph));
unsigned use_mkldnn_true_count = 0;
for (auto* node : graph->Nodes()) {
if (node->IsOp()) {
auto* op = node->Op();
if (op->HasAttr("use_mkldnn") &&
boost::get<bool>(op->GetAttr("use_mkldnn"))) {
++use_mkldnn_true_count;
}
}
}
EXPECT_EQ(use_mkldnn_true_count, expected_use_mkldnn_true_count);
}
TEST(MKLDNNPlacementPass, enable_conv_relu) {
// 1 conv (1 conv is always true) + 2 relu (1 relu is always true) + 0 pool
MainTest({"conv2d", "relu"}, 3);
}
TEST(MKLDNNPlacementPass, enable_relu_pool) {
// 1 conv (1 conv is always true) + 2 relu (1 relu is always true) + 1 pool
MainTest({"relu", "pool2d"}, 4);
}
TEST(MKLDNNPlacementPass, enable_all) {
// 1 conv (1 conv is always true) + 2 relu (1 relu is always true) + 1 pool
MainTest({}, 4);
}
} // namespace ir
} // namespace framework
} // namespace paddle
USE_PASS(mkldnn_placement_pass);
paddle/fluid/framework/ir/repeated_fc_relu_fuse_pass.h
浏览文件 @ d497bd90
......@@ -31,7 +31,8 @@ class RepeatedFCReluFusePass : public FusePassBase {
virtual ~RepeatedFCReluFusePass() {}
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;
const std::string name_scope_{"repeated_fc_relu_fuse"};
};
......
paddle/fluid/framework/ir/seq_concat_fc_fuse_pass.h
浏览文件 @ d497bd90
......@@ -12,6 +12,8 @@
// 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"
......@@ -25,7 +27,8 @@ class SeqConcatFcFusePass : public FusePassBase {
virtual ~SeqConcatFcFusePass() {}
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/seqconv_eltadd_relu_fuse_pass.h
浏览文件 @ d497bd90
......@@ -28,7 +28,8 @@ class SeqConvEltAddReluFusePass : public FusePassBase {
virtual ~SeqConvEltAddReluFusePass() {}
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;
const std::string name_scope_{"seqconv_eltadd_relu_fuse"};
};
......
paddle/fluid/framework/ir/seqpool_concat_fuse_pass.h
浏览文件 @ d497bd90
......@@ -42,7 +42,8 @@ class SeqPoolConcatFusePass : public FusePassBase {
virtual ~SeqPoolConcatFusePass() {}
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;
const std::string name_scope_{"seqpool_concat_fuse"};
};
......
paddle/fluid/framework/ir/squared_mat_sub_fuse_pass.h
浏览文件 @ d497bd90
......@@ -31,7 +31,8 @@ class SquaredMatSubFusePass : public FusePassBase {
virtual ~SquaredMatSubFusePass() {}
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;
const std::string name_scope_{"squared_mat_sub_fuse"};
};
......
paddle/fluid/framework/operator.cc
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/framework/operator.h
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/framework/operator_kernel_configs.h 0 → 100644
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/framework/tensor.h
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/imperative/layer.cc
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/imperative/layer.h
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/imperative/tracer.cc
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/imperative/tracer.h
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/inference/anakin/CMakeLists.txt 0 → 100644
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/inference/anakin/convert/CMakeLists.txt 0 → 100644
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/inference/anakin/convert/fc.cc 0 → 100644
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/inference/anakin/convert/fc.h 0 → 100644
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/inference/anakin/convert/op_converter.h 0 → 100644
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/inference/anakin/convert/registrar.cc 0 → 100644
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/inference/anakin/convert/registrar.h 0 → 100644
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/inference/anakin/convert/test_fc_op.cc 0 → 100644
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/inference/anakin/convert/ut_helper.h 0 → 100644
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/inference/anakin/engine.cc 0 → 100644
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/inference/anakin/engine.h 0 → 100644
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/inference/anakin/test_anakin_engine.cc 0 → 100644
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/inference/api/api.cc
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/operators/activation_cudnn.cu.cc 0 → 100644
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/operators/activation_cudnn_op.cu.cc 0 → 100644
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/operators/alloc_continuous_space_op.cc 0 → 100644
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/operators/benchmark/CMakeLists.txt 0 → 100644
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/operators/benchmark/op_tester.cc 0 → 100644
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/operators/benchmark/op_tester.h 0 → 100644
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/operators/benchmark/op_tester_config.cc 0 → 100644
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/operators/benchmark/op_tester_config.h 0 → 100644
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/operators/conv_op.cc
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/operators/fc_op.cc
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/operators/hash_op.cc
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/operators/hash_op.h
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/operators/is_empty_op.cu.cc 0 → 100644
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/operators/jit/gen/embseqpool.cc 0 → 100644
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/operators/jit/gen/embseqpool.h 0 → 100644
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/operators/jit/gen/sgd.cc 0 → 100644
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/operators/jit/gen/sgd.h 0 → 100644
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/operators/jit/helper.h
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/operators/jit/test.cc
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/operators/lstm_op.h
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/operators/lstmp_op.cc
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/operators/lstmp_op.h
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/operators/math/blas.h
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/operators/math/sample_prob.cu 0 → 100644
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/operators/math/sample_prob.h 0 → 100644
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/operators/ngraph/ops/CMakeLists.txt 0 → 100644
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/operators/ngraph/ops/adam_op.h 0 → 100644
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/operators/ngraph/ops/concat_op.h 0 → 100644
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/operators/ngraph/ops/cross_entropy_op.h 0 → 100644
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/operators/ngraph/ops/momentum_op.h 0 → 100644
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/operators/ngraph/ops/op_bridge.h 0 → 100644
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/operators/pool_op.cc
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/operators/sample_logits_op.cc 0 → 100644
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/operators/sample_logits_op.cu 0 → 100644
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/operators/sample_logits_op.h 0 → 100644
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/operators/shape_op.cc
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/platform/cudnn_desc.h 0 → 100644
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/platform/cudnn_desc_test.cc 0 → 100644
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/platform/enforce.h
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/platform/event.h 0 → 100644
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/platform/init.cc
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/platform/mkldnn_utils.h 0 → 100644
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/platform/profiler.cc
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/platform/profiler.cu 0 → 100644
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/platform/profiler.h
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/pybind/imperative.cc
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/pybind/imperative.h
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/pybind/ir.cc
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/pybind/pybind.cc
浏览文件 @ d497bd90
此差异已折叠。
paddle/fluid/train/demo/README.md
浏览文件 @ d497bd90
此差异已折叠。
paddle/scripts/README.md
浏览文件 @ d497bd90
此差异已折叠。
paddle/scripts/cpplint.py 已删除 100644 → 0
浏览文件 @ a9c0968b
此差异已折叠。
paddle/scripts/fast_install.sh
浏览文件 @ d497bd90
此差异已折叠。
paddle/scripts/paddle_build.sh
浏览文件 @ d497bd90
此差异已折叠。
paddle/scripts/submit_local.sh.in
浏览文件 @ d497bd90
此差异已折叠。
python/CMakeLists.txt
浏览文件 @ d497bd90
此差异已折叠。
python/paddle/fluid/__init__.py
浏览文件 @ d497bd90
此差异已折叠。
python/paddle/fluid/compiler.py
浏览文件 @ d497bd90
此差异已折叠。
python/paddle/fluid/executor.py
浏览文件 @ d497bd90
此差异已折叠。
python/paddle/fluid/framework.py
浏览文件 @ d497bd90
此差异已折叠。
python/paddle/fluid/imperative/tracer.py 0 → 100644
浏览文件 @ d497bd90
此差异已折叠。
python/paddle/fluid/initializer.py
浏览文件 @ d497bd90
此差异已折叠。
python/paddle/fluid/io.py
浏览文件 @ d497bd90
此差异已折叠。
python/paddle/fluid/layer_helper.py
浏览文件 @ d497bd90
此差异已折叠。
python/paddle/fluid/layers/io.py
浏览文件 @ d497bd90
此差异已折叠。
python/paddle/fluid/layers/nn.py
浏览文件 @ d497bd90
此差异已折叠。
python/paddle/fluid/layers/ops.py
浏览文件 @ d497bd90
此差异已折叠。
python/paddle/fluid/optimizer.py
浏览文件 @ d497bd90
此差异已折叠。
python/paddle/utils/plot.py
浏览文件 @ d497bd90
此差异已折叠。
python/requirements.txt
浏览文件 @ d497bd90
此差异已折叠。
tools/check_doc_approval.py 已删除 100644 → 0
浏览文件 @ a9c0968b
此差异已折叠。
tools/diff_api.py
浏览文件 @ d497bd90
此差异已折叠。
tools/manylinux1/Dockerfile.x64
浏览文件 @ d497bd90
此差异已折叠。
tools/manylinux1/build_all.sh
浏览文件 @ d497bd90
此差异已折叠。
tools/print_signatures.py
浏览文件 @ d497bd90
此差异已折叠。
tools/timeline.py
浏览文件 @ d497bd90
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