Skip to content

P
Paddle

Crayon鑫 / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 1
Fork 0
P
Paddle
提交 893957f7 编写于 作者: T tensor-tang
浏览文件
操作

Merge remote-tracking branch 'ups/develop' into refine/jit

上级 6648995f c89a1fb2
展开全部 隐藏空白更改
内联 并排
Showing with 14809 addition and 3543 deletion
CMakeLists.txt
浏览文件 @ 893957f7
......@@ -54,7 +54,7 @@ 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" OFF)
option(WITH_PROFILER "Compile PaddlePaddle with GPU profiler and gperftools" 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)
......@@ -65,6 +65,7 @@ option(WITH_GOLANG "Compile PaddlePaddle with GOLANG" OFF)
option(GLIDE_INSTALL "Download and install go dependencies " ON)
option(USE_NNPACK "Compile PaddlePaddle with NNPACK library" OFF)
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)
......@@ -131,8 +132,6 @@ if (APPLE OR WIN32)
endif()
if (WIN32)
set(WITH_AVX OFF CACHE STRING
"Disable AVX when compiling for Windows" FORCE)
set(WITH_DSO OFF CACHE STRING
"Disable DSO when compiling for Windows" FORCE)
set(WITH_MKL OFF CACHE STRING
......@@ -217,6 +216,12 @@ include(cupti)
include(external/gzstream)
endif (NOT WIN32)
if(WITH_PSLIB)
include(external/libmct)
include(external/pslib_brpc)
include(external/pslib)
endif(WITH_PSLIB)
if(WITH_DISTRIBUTE)
if(WITH_GRPC)
include(external/grpc)
......@@ -254,6 +259,12 @@ elseif()
set(WITH_ANAKIN OFF CACHE STRING "Anakin is used in MKL only now." FORCE)
endif()
if (WITH_PROFILER)
find_package(Gperftools REQUIRED)
include_directories(${GPERFTOOLS_INCLUDE_DIR})
add_definitions(-DWITH_GPERFTOOLS)
endif()
include(generic) # simplify cmake module
include(package) # set paddle packages
include(ccache) # set ccache for compilation
......@@ -278,6 +289,12 @@ set(EXTERNAL_LIBS
${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)
......
README.md
浏览文件 @ 893957f7
......@@ -19,6 +19,15 @@ 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.
欢迎来到 PaddlePaddle GitHub
PaddlePaddle (PArallel Distributed Deep LEarning) 是一个简单易用、高效灵活、可扩展的深度学习平台，最初由百度科学家和工程师共同开发，目的是将深度学习技术应用到百度的众多产品中。
我们的愿景是让每个人都能通过PaddlePaddle接触深度学习
跟进PaddlePaddle最新特性请参考我们的[版本说明](https://github.com/PaddlePaddle/Paddle/releases)
### Latest PaddlePaddle Release: [Fluid 1.2.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.2)
### Install Latest Stable Release:
```
......@@ -34,6 +43,23 @@ pip install paddlepaddle-gpu==1.2.0.post85
# For installation on other platform, refer to http://paddlepaddle.org/
```
### PaddlePaddle最新版本: [Fluid 1.2.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.2)
### 安装最新稳定版本:
```
# Linux CPU
pip install paddlepaddle
# Linux GPU cuda9cudnn7
pip install paddlepaddle-gpu
# Linux GPU cuda8cudnn7
pip install paddlepaddle-gpu==1.2.0.post87
# Linux GPU cuda8cudnn5
pip install paddlepaddle-gpu==1.2.0.post85
# 其他平台上的安装指引请参考 http://paddlepaddle.org/
```
## Features
- **Flexibility**
......@@ -74,10 +100,38 @@ pip install paddlepaddle-gpu==1.2.0.post85
Baidu and it has achieved a significant impact. We hope you can also explore
the capability of PaddlePaddle to make an impact on your product.
## 特点
- **灵活性**
PaddlePaddle支持丰富的神经网络架构和优化算法。易于配置复杂模型，例如带有注意力机制或复杂记忆连接的神经网络机器翻译模型。
- **高效性**
为了高效使用异步计算资源，PaddlePaddle对框架的不同层进行优化，包括计算、存储、架构和通信。下面是一些样例：
- 通过SSE/AVX 内置函数、BLAS库(例如MKL、OpenBLAS、cuBLAS)或定制的CPU/GPU内核优化数学操作。
- 通过MKL-DNN库优化CNN网络
- 高度优化循环网络，无需执行 `padding` 操作即可处理 **变长** 序列
- 针对高维稀疏数据模型，优化了局部和分布式训练。
- **稳定性**
有了 PaddlePaddle，使得利用各种CPU/GPU和机器来加速训练变得简单。PaddlePaddle 通过优化通信可以实现巨大吞吐量和快速执行。
- **连接产品**
另外，PaddlePaddle 的设计也易于部署。在百度，PaddlePaddle 已经部署到含有巨大用户量的产品和服务上，包括广告点击率（CTR）预测、大规模图像分类、光学字符识别（OCR）、搜索排序，计算机病毒检测、推荐系统等等。PaddlePaddle广泛应用于百度产品中，产生了非常重要的影响。我们希望您也能探索 PaddlePaddle 的能力，为您的产品创造新的影响力和效果。
## 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.
## 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)
## Documentation
We provide [English](http://paddlepaddle.org/documentation/docs/en/1.2/getstarted/index_en.html) and
......@@ -99,10 +153,37 @@ We provide [English](http://paddlepaddle.org/documentation/docs/en/1.2/getstarte
We appreciate your contributions!
## 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
- [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)
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)
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)
欢迎您的贡献!
## Ask Questions
You are welcome to submit questions and bug reports as [Github Issues](https://github.com/PaddlePaddle/Paddle/issues).
## 答疑
欢迎您将问题和bug报告以[Github Issues](https://github.com/PaddlePaddle/Paddle/issues)的形式提交
## Copyright and License
PaddlePaddle is provided under the [Apache-2.0 license](LICENSE).
## 版权和许可证
PaddlePaddle由[Apache-2.0 license](LICENSE)提供
benchmark/fluid/fluid_benchmark.py
浏览文件 @ 893957f7
......@@ -81,9 +81,11 @@ def dist_transpile(trainer_id, args, train_prog, startup_prog):
# the role, should be either PSERVER or TRAINER
training_role = os.getenv("PADDLE_TRAINING_ROLE")
config = distribute_transpiler.DistributeTranspilerConfig()
config = fluid.DistributeTranspilerConfig()
config.slice_var_up = not args.no_split_var
config.min_block_size = 1048576
t = distribute_transpiler.DistributeTranspiler(config=config)
t.transpile(
trainer_id,
# NOTE: *MUST* use train_prog, for we are using with guard to
......
cmake/FindGperftools.cmake 0 → 100644
浏览文件 @ 893957f7
# Tries to find Gperftools.
#
# Usage of this module as follows:
#
# find_package(Gperftools)
#
# Variables used by this module, they can change the default behaviour and need
# to be set before calling find_package:
#
# Gperftools_ROOT_DIR Set this variable to the root installation of
# Gperftools if the module has problems finding
# the proper installation path.
#
# Variables defined by this module:
#
# GPERFTOOLS_FOUND System has Gperftools libs/headers
# GPERFTOOLS_LIBRARIES The Gperftools libraries (tcmalloc & profiler)
# GPERFTOOLS_INCLUDE_DIR The location of Gperftools headers
find_library(GPERFTOOLS_TCMALLOC
NAMES tcmalloc
HINTS ${Gperftools_ROOT_DIR}/lib)
find_library(GPERFTOOLS_PROFILER
NAMES profiler
HINTS ${Gperftools_ROOT_DIR}/lib)
find_library(GPERFTOOLS_TCMALLOC_AND_PROFILER
NAMES tcmalloc_and_profiler
HINTS ${Gperftools_ROOT_DIR}/lib)
find_path(GPERFTOOLS_INCLUDE_DIR
NAMES gperftools/heap-profiler.h
HINTS ${Gperftools_ROOT_DIR}/include)
set(GPERFTOOLS_LIBRARIES ${GPERFTOOLS_TCMALLOC_AND_PROFILER})
include(FindPackageHandleStandardArgs)
find_package_handle_standard_args(
Gperftools
DEFAULT_MSG
GPERFTOOLS_LIBRARIES
GPERFTOOLS_INCLUDE_DIR)
mark_as_advanced(
Gperftools_ROOT_DIR
GPERFTOOLS_TCMALLOC
GPERFTOOLS_PROFILER
GPERFTOOLS_TCMALLOC_AND_PROFILER
GPERFTOOLS_LIBRARIES
GPERFTOOLS_INCLUDE_DIR)
# create IMPORTED targets
if (Gperftools_FOUND AND NOT TARGET gperftools::tcmalloc)
add_library(gperftools::tcmalloc UNKNOWN IMPORTED)
set_target_properties(gperftools::tcmalloc PROPERTIES
IMPORTED_LOCATION ${GPERFTOOLS_TCMALLOC}
INTERFACE_INCLUDE_DIRECTORIES "${GPERFTOOLS_INCLUDE_DIR}")
add_library(gperftools::profiler UNKNOWN IMPORTED)
set_target_properties(gperftools::profiler PROPERTIES
IMPORTED_LOCATION ${GPERFTOOLS_PROFILER}
INTERFACE_INCLUDE_DIRECTORIES "${GPERFTOOLS_INCLUDE_DIR}")
endif()
cmake/configure.cmake
浏览文件 @ 893957f7
......@@ -84,8 +84,13 @@ if(NOT WITH_GOLANG)
add_definitions(-DPADDLE_WITHOUT_GOLANG)
endif(NOT WITH_GOLANG)
if(WITH_PSLIB)
add_definitions(-DPADDLE_WITH_PSLIB)
endif()
if(WITH_GPU)
add_definitions(-DPADDLE_WITH_CUDA)
add_definitions(-DEIGEN_USE_GPU)
FIND_PACKAGE(CUDA REQUIRED)
......
cmake/external/brpc.cmake
浏览文件 @ 893957f7
......@@ -14,14 +14,16 @@
INCLUDE(ExternalProject)
find_library(SSL_LIBRARY NAMES ssl)
find_package(OpenSSL REQUIRED)
message(STATUS "ssl:" ${OPENSSL_SSL_LIBRARY})
message(STATUS "crypto:" ${OPENSSL_CRYPTO_LIBRARY})
ADD_LIBRARY(ssl SHARED IMPORTED GLOBAL)
SET_PROPERTY(TARGET ssl PROPERTY IMPORTED_LOCATION ${SSL_LIBRARY})
SET_PROPERTY(TARGET ssl PROPERTY IMPORTED_LOCATION ${OPENSSL_SSL_LIBRARY})
find_library(CRYPTO_LIBRARY NAMES crypto)
ADD_LIBRARY(crypto SHARED IMPORTED GLOBAL)
SET_PROPERTY(TARGET crypto PROPERTY IMPORTED_LOCATION ${CRYPTO_LIBRARY})
SET_PROPERTY(TARGET crypto PROPERTY IMPORTED_LOCATION ${OPENSSL_CRYPTO_LIBRARY})
SET(BRPC_SOURCES_DIR ${THIRD_PARTY_PATH}/brpc)
SET(BRPC_INSTALL_DIR ${THIRD_PARTY_PATH}/install/brpc)
......@@ -31,14 +33,15 @@ SET(BRPC_LIBRARIES "${BRPC_INSTALL_DIR}/lib/libbrpc.a" CACHE FILEPATH "brpc libr
INCLUDE_DIRECTORIES(${BRPC_INCLUDE_DIR})
# Reference https://stackoverflow.com/questions/45414507/pass-a-list-of-prefix-paths-to-externalproject-add-in-cmake-args
set(prefix_path "${THIRD_PARTY_PATH}/install/gflags|${THIRD_PARTY_PATH}/install/leveldb|${THIRD_PARTY_PATH}/install/snappy|${THIRD_PARTY_PATH}/install/gtest|${THIRD_PARTY_PATH}/install/protobuf|${THIRD_PARTY_PATH}/install/zlib")
set(prefix_path "${THIRD_PARTY_PATH}/install/gflags|${THIRD_PARTY_PATH}/install/leveldb|${THIRD_PARTY_PATH}/install/snappy|${THIRD_PARTY_PATH}/install/gtest|${THIRD_PARTY_PATH}/install/protobuf|${THIRD_PARTY_PATH}/install/zlib|${THIRD_PARTY_PATH}/install/glog")
# If minimal .a is need, you can set WITH_DEBUG_SYMBOLS=OFF
ExternalProject_Add(
extern_brpc
${EXTERNAL_PROJECT_LOG_ARGS}
# TODO(gongwb): change to de newst repo when they changed.
GIT_REPOSITORY "https://github.com/gongweibao/brpc"
GIT_TAG "7dc04defad1fd4173aae170c3fcbde131b65155a"
GIT_TAG "e9b67ec1b7458f2af5fae76451afe1e27e01b4b4"
PREFIX ${BRPC_SOURCES_DIR}
UPDATE_COMMAND ""
CMAKE_ARGS -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
......@@ -50,7 +53,7 @@ ExternalProject_Add(
-DCMAKE_POSITION_INDEPENDENT_CODE=ON
-DCMAKE_BUILD_TYPE=${THIRD_PARTY_BUILD_TYPE}
-DCMAKE_PREFIX_PATH=${prefix_path}
-DBRPC_WITH_GLOG=ON
-DWITH_GLOG=ON
-DIOBUF_WITH_HUGE_BLOCK=ON
-DBRPC_WITH_RDMA=${WITH_BRPC_RDMA}
${EXTERNAL_OPTIONAL_ARGS}
......@@ -65,5 +68,6 @@ ADD_LIBRARY(brpc STATIC IMPORTED GLOBAL)
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/gtest.cmake
浏览文件 @ 893957f7
......@@ -12,8 +12,12 @@
# See the License for the specific language governing permissions and
# limitations under the License.
IF(WITH_TESTING)
ENABLE_TESTING()
#FIXME:(gongwb) Move brpc's gtest dependency.
IF(WITH_TESTING OR (WITH_DISTRIBUTE AND NOT WITH_GRPC))
IF(WITH_TESTING)
ENABLE_TESTING()
ENDIF(WITH_TESTING)
INCLUDE(ExternalProject)
SET(GTEST_SOURCES_DIR ${THIRD_PARTY_PATH}/gtest)
......@@ -76,4 +80,4 @@ IF(WITH_TESTING)
ADD_DEPENDENCIES(gtest_main extern_gtest)
LIST(APPEND external_project_dependencies gtest gtest_main)
ENDIF(WITH_TESTING)
ENDIF(WITH_TESTING OR (WITH_DISTRIBUTE AND NOT WITH_GRPC))
cmake/external/leveldb.cmake
浏览文件 @ 893957f7
......@@ -24,8 +24,8 @@ ExternalProject_Add(
extern_leveldb
${EXTERNAL_PROJECT_LOG_ARGS}
PREFIX ${LEVELDB_SOURCES_DIR}
URL "https://github.com/google/leveldb/archive/v1.18.tar.gz"
URL_MD5 "73770de34a2a5ab34498d2e05b2b7fa0"
GIT_REPOSITORY "https://github.com/google/leveldb"
GIT_TAG v1.18
CONFIGURE_COMMAND ""
BUILD_COMMAND CXXFLAGS=-fPIC make -j ${NUM_OF_PROCESSOR} libleveldb.a
INSTALL_COMMAND mkdir -p ${LEVELDB_INSTALL_DIR}/lib/
......
cmake/external/libmct.cmake 0 → 100644
浏览文件 @ 893957f7
# Copyright (c) 2017 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.
IF(NOT ${WITH_LIBMCT})
return()
ENDIF(NOT ${WITH_LIBMCT})
IF(WIN32 OR APPLE)
MESSAGE(WARNING
"Windows or Mac is not supported with LIBMCT in Paddle yet."
"Force WITH_LIBMCT=OFF")
SET(WITH_LIBMCT OFF CACHE STRING "Disable LIBMCT package in Windows and MacOS" FORCE)
return()
ENDIF()
INCLUDE(ExternalProject)
SET(LIBMCT_PROJECT "extern_libmct")
IF((NOT DEFINED LIBMCT_VER) OR (NOT DEFINED LIBMCT_URL))
MESSAGE(STATUS "use pre defined download url")
SET(LIBMCT_VER "0.1.0" CACHE STRING "" FORCE)
SET(LIBMCT_NAME "libmct" CACHE STRING "" FORCE)
SET(LIBMCT_URL "https://raw.githubusercontent.com/PaddlePaddle/Fleet/release/${LIBMCT_VER}/${LIBMCT_NAME}.tar.gz" CACHE STRING "" FORCE)
ENDIF()
MESSAGE(STATUS "LIBMCT_NAME: ${LIBMCT_NAME}, LIBMCT_URL: ${LIBMCT_URL}")
SET(LIBMCT_SOURCE_DIR "${THIRD_PARTY_PATH}/libmct")
SET(LIBMCT_DOWNLOAD_DIR "${LIBMCT_SOURCE_DIR}/src/${LIBMCT_PROJECT}")
SET(LIBMCT_DST_DIR "libmct")
SET(LIBMCT_INSTALL_ROOT "${THIRD_PARTY_PATH}/install")
SET(LIBMCT_INSTALL_DIR ${LIBMCT_INSTALL_ROOT}/${LIBMCT_DST_DIR})
SET(LIBMCT_ROOT ${LIBMCT_INSTALL_DIR})
SET(LIBMCT_INC_DIR ${LIBMCT_ROOT}/include)
SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${LIBMCT_ROOT}/lib")
INCLUDE_DIRECTORIES(${LIBMCT_INC_DIR})
FILE(WRITE ${LIBMCT_DOWNLOAD_DIR}/CMakeLists.txt
"PROJECT(LIBMCT)\n"
"cmake_minimum_required(VERSION 3.0)\n"
"install(DIRECTORY ${LIBMCT_NAME}/include ${LIBMCT_NAME}/lib \n"
" DESTINATION ${LIBMCT_DST_DIR})\n")
ExternalProject_Add(
${LIBMCT_PROJECT}
${EXTERNAL_PROJECT_LOG_ARGS}
PREFIX ${LIBMCT_SOURCE_DIR}
DOWNLOAD_DIR ${LIBMCT_DOWNLOAD_DIR}
DOWNLOAD_COMMAND wget --no-check-certificate ${LIBMCT_URL} -c -q -O ${LIBMCT_NAME}.tar.gz
&& tar zxvf ${LIBMCT_NAME}.tar.gz
DOWNLOAD_NO_PROGRESS 1
UPDATE_COMMAND ""
CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${LIBMCT_INSTALL_ROOT}
CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${LIBMCT_INSTALL_ROOT}
)
if (${CMAKE_VERSION} VERSION_LESS "3.3.0" OR NOT WIN32)
set(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/boost_dummy.c)
file(WRITE ${dummyfile} "const char *dummy = \"${dummyfile}\";")
add_library(libmct STATIC ${dummyfile})
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/pslib.cmake 0 → 100644
浏览文件 @ 893957f7
# Copyright (c) 2017 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.
IF(NOT ${WITH_PSLIB})
return()
ENDIF(NOT ${WITH_PSLIB})
IF(WIN32 OR APPLE)
MESSAGE(WARNING
"Windows or Mac is not supported with PSLIB in Paddle yet."
"Force WITH_PSLIB=OFF")
SET(WITH_PSLIB OFF CACHE STRING "Disable PSLIB package in Windows and MacOS" FORCE)
return()
ENDIF()
INCLUDE(ExternalProject)
SET(PSLIB_PROJECT "extern_pslib")
IF((NOT DEFINED PSLIB_VER) OR (NOT DEFINED PSLIB_URL))
MESSAGE(STATUS "use pre defined download url")
SET(PSLIB_VER "0.1.0" CACHE STRING "" FORCE)
SET(PSLIB_NAME "pslib" CACHE STRING "" FORCE)
SET(PSLIB_URL "https://raw.githubusercontent.com/PaddlePaddle/Fleet/release/${PSLIB_VER}/${PSLIB_NAME}.tar.gz" CACHE STRING "" FORCE)
ENDIF()
MESSAGE(STATUS "PSLIB_NAME: ${PSLIB_NAME}, PSLIB_URL: ${PSLIB_URL}")
SET(PSLIB_SOURCE_DIR "${THIRD_PARTY_PATH}/pslib")
SET(PSLIB_DOWNLOAD_DIR "${PSLIB_SOURCE_DIR}/src/${PSLIB_PROJECT}")
SET(PSLIB_DST_DIR "pslib")
SET(PSLIB_INSTALL_ROOT "${THIRD_PARTY_PATH}/install")
SET(PSLIB_INSTALL_DIR ${PSLIB_INSTALL_ROOT}/${PSLIB_DST_DIR})
SET(PSLIB_ROOT ${PSLIB_INSTALL_DIR})
SET(PSLIB_INC_DIR ${PSLIB_ROOT}/include)
SET(PSLIB_LIB_DIR ${PSLIB_ROOT}/lib)
SET(PSLIB_LIB ${PSLIB_LIB_DIR}/libps.so)
SET(PSLIB_IOMP_LIB ${PSLIB_LIB_DIR}/libiomp5.so) #todo what is this
SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${PSLIB_ROOT}/lib")
INCLUDE_DIRECTORIES(${PSLIB_INC_DIR})
FILE(WRITE ${PSLIB_DOWNLOAD_DIR}/CMakeLists.txt
"PROJECT(PSLIB)\n"
"cmake_minimum_required(VERSION 3.0)\n"
"install(DIRECTORY ${PSLIB_NAME}/include ${PSLIB_NAME}/lib \n"
" DESTINATION ${PSLIB_DST_DIR})\n")
ExternalProject_Add(
${PSLIB_PROJECT}
${EXTERNAL_PROJECT_LOG_ARGS}
PREFIX ${PSLIB_SOURCE_DIR}
DOWNLOAD_DIR ${PSLIB_DOWNLOAD_DIR}
DOWNLOAD_COMMAND wget --no-check-certificate ${PSLIB_URL} -c -q -O ${PSLIB_NAME}.tar.gz
&& tar zxvf ${PSLIB_NAME}.tar.gz
DOWNLOAD_NO_PROGRESS 1
UPDATE_COMMAND ""
CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${PSLIB_INSTALL_ROOT}
CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${PSLIB_INSTALL_ROOT}
)
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)
IF(WITH_C_API)
INSTALL(FILES ${PSLIB_LIB} ${PSLIB_IOMP_LIB} DESTINATION lib)
ENDIF()
cmake/external/pslib_brpc.cmake 0 → 100644
浏览文件 @ 893957f7
# Copyright (c) 2017 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.
IF(NOT ${WITH_PSLIB_BRPC})
return()
ENDIF(NOT ${WITH_PSLIB_BRPC})
IF(WIN32 OR APPLE)
MESSAGE(WARNING
"Windows or Mac is not supported with PSLIB_BRPC in Paddle yet."
"Force WITH_PSLIB_BRPC=OFF")
SET(WITH_PSLIB_BRPC OFF CACHE STRING "Disable PSLIB_BRPC package in Windows and MacOS" FORCE)
return()
ENDIF()
INCLUDE(ExternalProject)
SET(PSLIB_BRPC_PROJECT "extern_pslib_brpc")
IF((NOT DEFINED PSLIB_BRPC_NAME) OR (NOT DEFINED PSLIB_BRPC_URL))
MESSAGE(STATUS "use pre defined download url")
SET(PSLIB_BRPC_VER "0.1.0" CACHE STRING "" FORCE)
SET(PSLIB_BRPC_NAME "pslib_brpc" CACHE STRING "" FORCE)
SET(PSLIB_BRPC_URL "https://raw.githubusercontent.com/PaddlePaddle/Fleet/release/${PSLIB_BRPC_VER}/${PSLIB_BRPC_NAME}.tar.gz" CACHE STRING "" FORCE)
ENDIF()
MESSAGE(STATUS "PSLIB_BRPC_NAME: ${PSLIB_BRPC_NAME}, PSLIB_BRPC_URL: ${PSLIB_BRPC_URL}")
SET(PSLIB_BRPC_SOURCE_DIR "${THIRD_PARTY_PATH}/pslib_brpc")
SET(PSLIB_BRPC_DOWNLOAD_DIR "${PSLIB_BRPC_SOURCE_DIR}/src/${PSLIB_BRPC_PROJECT}")
SET(PSLIB_BRPC_DST_DIR "pslib_brpc")
SET(PSLIB_BRPC_INSTALL_ROOT "${THIRD_PARTY_PATH}/install")
SET(PSLIB_BRPC_INSTALL_DIR ${PSLIB_BRPC_INSTALL_ROOT}/${PSLIB_BRPC_DST_DIR})
SET(PSLIB_BRPC_ROOT ${PSLIB_BRPC_INSTALL_DIR})
SET(PSLIB_BRPC_INC_DIR ${PSLIB_BRPC_ROOT}/include)
SET(PSLIB_BRPC_LIB_DIR ${PSLIB_BRPC_ROOT}/lib)
SET(PSLIB_BRPC_LIB ${PSLIB_BRPC_LIB_DIR}/libbrpc.a)
SET(PSLIB_BRPC_IOMP_LIB ${PSLIB_BRPC_LIB_DIR}/libiomp5.so) #todo what is this
SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${PSLIB_BRPC_ROOT}/lib")
INCLUDE_DIRECTORIES(${PSLIB_BRPC_INC_DIR})
FILE(WRITE ${PSLIB_BRPC_DOWNLOAD_DIR}/CMakeLists.txt
"PROJECT(PSLIB_BRPC)\n"
"cmake_minimum_required(VERSION 3.0)\n"
"install(DIRECTORY ${PSLIB_BRPC_NAME}/include ${PSLIB_BRPC_NAME}/lib \n"
" DESTINATION ${PSLIB_BRPC_DST_DIR})\n")
ExternalProject_Add(
${PSLIB_BRPC_PROJECT}
${EXTERNAL_PROJECT_LOG_ARGS}
PREFIX ${PSLIB_BRPC_SOURCE_DIR}
DOWNLOAD_DIR ${PSLIB_BRPC_DOWNLOAD_DIR}
DOWNLOAD_COMMAND wget --no-check-certificate ${PSLIB_BRPC_URL} -c -q -O ${PSLIB_BRPC_NAME}.tar.gz
&& tar zxvf ${PSLIB_BRPC_NAME}.tar.gz
DOWNLOAD_NO_PROGRESS 1
UPDATE_COMMAND ""
CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${PSLIB_BRPC_INSTALL_ROOT}
CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${PSLIB_BRPC_INSTALL_ROOT}
)
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)
IF(WITH_C_API)
INSTALL(FILES ${PSLIB_BRPC_LIB} ${PSLIB_BRPC_IOMP_LIB} DESTINATION lib)
ENDIF()
cmake/external/python.cmake
浏览文件 @ 893957f7
......@@ -18,8 +18,8 @@ ENDIF()
INCLUDE(python_module)
FIND_PACKAGE(PythonInterp ${PY_VERSION})
FIND_PACKAGE(PythonLibs ${PY_VERSION})
FIND_PACKAGE(PythonInterp ${PY_VERSION} REQUIRED)
FIND_PACKAGE(PythonLibs ${PY_VERSION} REQUIRED)
if(WIN32)
execute_process(COMMAND "${PYTHON_EXECUTABLE}" "-c"
......@@ -79,6 +79,5 @@ IF(PYTHONINTERP_FOUND)
"please use pip to upgrade protobuf. pip install -U protobuf")
ENDIF()
ENDIF(PYTHONINTERP_FOUND)
INCLUDE_DIRECTORIES(${PYTHON_INCLUDE_DIR})
INCLUDE_DIRECTORIES(${PYTHON_NUMPY_INCLUDE_DIR})
cmake/external/snappy.cmake
浏览文件 @ 893957f7
......@@ -24,12 +24,6 @@ set(SNAPPY_SOURCES_DIR ${THIRD_PARTY_PATH}/snappy)
set(SNAPPY_INSTALL_DIR ${THIRD_PARTY_PATH}/install/snappy)
set(SNAPPY_INCLUDE_DIR "${SNAPPY_INSTALL_DIR}/include" CACHE PATH "snappy include directory." FORCE)
if (WIN32)
set(SNAPPY_LIBRARIES "${SNAPPY_INSTALL_DIR}/lib/snappy.lib")
else(WIN32)
set(SNAPPY_LIBRARIES "${SNAPPY_INSTALL_DIR}/lib/libsnappy.a")
endif (WIN32)
ExternalProject_Add(
extern_snappy
GIT_REPOSITORY "https://github.com/google/snappy"
......@@ -56,6 +50,16 @@ ExternalProject_Add(
-DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON
-DCMAKE_BUILD_TYPE:STRING=${THIRD_PARTY_BUILD_TYPE}
)
IF(WIN32)
IF(NOT EXISTS "${SNAPPY_INSTALL_DIR}/lib/libsnappy.lib")
add_custom_command(TARGET extern_snappy POST_BUILD
COMMAND cmake -E copy ${SNAPPY_INSTALL_DIR}/lib/snappy.lib ${SNAPPY_INSTALL_DIR}/lib/libsnappy.lib
)
ENDIF()
set(SNAPPY_LIBRARIES "${SNAPPY_INSTALL_DIR}/lib/libsnappy.lib")
else(WIN32)
set(SNAPPY_LIBRARIES "${SNAPPY_INSTALL_DIR}/lib/libsnappy.a")
endif (WIN32)
add_library(snappy STATIC IMPORTED GLOBAL)
set_property(TARGET snappy PROPERTY IMPORTED_LOCATION ${SNAPPY_LIBRARIES})
......
cmake/external/xxhash.cmake
浏览文件 @ 893957f7
......@@ -56,7 +56,12 @@ else()
endif()
if (WIN32)
set(XXHASH_LIBRARIES "${XXHASH_INSTALL_DIR}/lib/xxhash.lib")
IF(NOT EXISTS "${XXHASH_INSTALL_DIR}/lib/libxxhash.lib")
add_custom_command(TARGET extern_xxhash POST_BUILD
COMMAND cmake -E copy ${XXHASH_INSTALL_DIR}/lib/xxhash.lib ${XXHASH_INSTALL_DIR}/lib/libxxhash.lib
)
ENDIF()
set(XXHASH_LIBRARIES "${XXHASH_INSTALL_DIR}/lib/libxxhash.lib")
else()
set(XXHASH_LIBRARIES "${XXHASH_INSTALL_DIR}/lib/libxxhash.a")
endif ()
......
cmake/external/zlib.cmake
浏览文件 @ 893957f7
......@@ -19,12 +19,6 @@ SET(ZLIB_INSTALL_DIR ${THIRD_PARTY_PATH}/install/zlib)
SET(ZLIB_ROOT ${ZLIB_INSTALL_DIR} CACHE FILEPATH "zlib root directory." FORCE)
SET(ZLIB_INCLUDE_DIR "${ZLIB_INSTALL_DIR}/include" CACHE PATH "zlib include directory." FORCE)
IF(WIN32)
SET(ZLIB_LIBRARIES "${ZLIB_INSTALL_DIR}/lib/zlibstatic.lib" CACHE FILEPATH "zlib library." FORCE)
ELSE(WIN32)
SET(ZLIB_LIBRARIES "${ZLIB_INSTALL_DIR}/lib/libz.a" CACHE FILEPATH "zlib library." FORCE)
ENDIF(WIN32)
INCLUDE_DIRECTORIES(${ZLIB_INCLUDE_DIR}) # For zlib code to include its own headers.
INCLUDE_DIRECTORIES(${THIRD_PARTY_PATH}/install) # For Paddle code to include zlib.h.
......@@ -49,6 +43,16 @@ ExternalProject_Add(
-DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON
-DCMAKE_BUILD_TYPE:STRING=${THIRD_PARTY_BUILD_TYPE}
)
IF(WIN32)
IF(NOT EXISTS "${ZLIB_INSTALL_DIR}/lib/libz.lib")
add_custom_command(TARGET extern_zlib POST_BUILD
COMMAND cmake -E copy ${ZLIB_INSTALL_DIR}/lib/zlibstatic.lib ${ZLIB_INSTALL_DIR}/lib/libz.lib
)
ENDIF()
SET(ZLIB_LIBRARIES "${ZLIB_INSTALL_DIR}/lib/libz.lib" CACHE FILEPATH "zlib library." FORCE)
ELSE(WIN32)
SET(ZLIB_LIBRARIES "${ZLIB_INSTALL_DIR}/lib/libz.a" CACHE FILEPATH "zlib library." FORCE)
ENDIF(WIN32)
ADD_LIBRARY(zlib STATIC IMPORTED GLOBAL)
SET_PROPERTY(TARGET zlib PROPERTY IMPORTED_LOCATION ${ZLIB_LIBRARIES})
......
cmake/generic.cmake
浏览文件 @ 893957f7
......@@ -110,6 +110,14 @@ function(find_fluid_modules TARGET_NAME)
endif()
endfunction(find_fluid_modules)
function(common_link TARGET_NAME)
if (WITH_PROFILER)
target_link_libraries(${TARGET_NAME} gperftools::profiler)
endif()
endfunction()
# find all third_party modules is used for paddle static library
# for reduce the dependency when building the inference libs.
set_property(GLOBAL PROPERTY FLUID_THIRD_PARTY)
......@@ -274,6 +282,7 @@ function(cc_library TARGET_NAME)
endif()
target_link_libraries(${TARGET_NAME} ${cc_library_DEPS})
add_dependencies(${TARGET_NAME} ${cc_library_DEPS})
common_link(${TARGET_NAME})
endif()
# cpplint code style
......@@ -340,6 +349,7 @@ function(cc_binary TARGET_NAME)
if(cc_binary_DEPS)
target_link_libraries(${TARGET_NAME} ${cc_binary_DEPS})
add_dependencies(${TARGET_NAME} ${cc_binary_DEPS})
common_link(${TARGET_NAME})
endif()
endfunction(cc_binary)
......@@ -362,6 +372,7 @@ function(cc_test TARGET_NAME)
target_link_libraries(${TARGET_NAME} ${win32_deps})
endif(WIN32)
add_dependencies(${TARGET_NAME} ${cc_test_DEPS} paddle_gtest_main lod_tensor memory gtest gflags glog)
common_link(${TARGET_NAME})
add_test(NAME ${TARGET_NAME}
COMMAND ${TARGET_NAME} ${cc_test_ARGS}
WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
......@@ -420,6 +431,7 @@ function(nv_binary TARGET_NAME)
if(nv_binary_DEPS)
target_link_libraries(${TARGET_NAME} ${nv_binary_DEPS})
add_dependencies(${TARGET_NAME} ${nv_binary_DEPS})
common_link(${TARGET_NAME})
endif()
endif()
endfunction(nv_binary)
......@@ -433,6 +445,7 @@ function(nv_test TARGET_NAME)
cuda_add_executable(${TARGET_NAME} ${nv_test_SRCS})
target_link_libraries(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main lod_tensor memory gtest gflags glog)
add_dependencies(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main lod_tensor memory gtest gflags glog)
common_link(${TARGET_NAME})
add_test(${TARGET_NAME} ${TARGET_NAME})
if (nv_test_SERIAL)
set_property(TEST ${TARGET_NAME} PROPERTY RUN_SERIAL 1)
......@@ -499,6 +512,7 @@ function(hip_binary TARGET_NAME)
if(hip_binary_DEPS)
target_link_libraries(${TARGET_NAME} ${hip_binary_DEPS})
add_dependencies(${TARGET_NAME} ${hip_binary_DEPS})
common_link(${TARGET_NAME})
endif()
endif()
endfunction(hip_binary)
......@@ -518,6 +532,7 @@ function(hip_test TARGET_NAME)
set_target_properties(${TARGET_NAME} PROPERTIES LINKER_LANGUAGE HIP)
target_link_libraries(${TARGET_NAME} ${hip_test_DEPS} paddle_gtest_main memory gtest gflags)
add_dependencies(${TARGET_NAME} ${hip_test_DEPS} paddle_gtest_main memory gtest gflags)
common_link(${TARGET_NAME})
add_test(${TARGET_NAME} ${TARGET_NAME})
endif()
endfunction(hip_test)
......@@ -560,6 +575,7 @@ function(go_library TARGET_NAME)
endif()
if(go_library_DEPS)
add_dependencies(${TARGET_NAME} ${go_library_DEPS})
common_link(${TARGET_NAME})
endif(go_library_DEPS)
# The "source file" of the library is `${dummyfile}` which never
......
cmake/inference_lib.cmake
浏览文件 @ 893957f7
......@@ -32,24 +32,35 @@ function(copy TARGET)
list(GET copy_lib_SRCS ${index} src)
list(GET copy_lib_DSTS ${index} dst)
if (WIN32)
# windows cmd shell will not expand wildcard automatically.
# below expand the files,libs and copy them by rules.
file(GLOB header_files ${src} "*.h")
file(GLOB static_lib_files ${src} "*.lib")
file(GLOB dll_lib_files ${src} "*.dll")
set(src_files ${header_files} ${static_lib_files} ${dll_lib_files})
if (NOT "${src_files}" STREQUAL "")
list(REMOVE_DUPLICATES src_files)
endif ()
add_custom_command(TARGET ${TARGET} PRE_BUILD
COMMAND ${CMAKE_COMMAND} -E make_directory "${dst}"
)
foreach (src_file ${src_files})
if(IS_DIRECTORY ${src})
get_filename_component(last_path ${src} NAME)
string(APPEND dst "/" ${last_path})
add_custom_command(TARGET ${TARGET} PRE_BUILD
COMMAND ${CMAKE_COMMAND} -E copy "${src_file}" "${dst}"
COMMENT "copying ${src_file} -> ${dst}")
endforeach ()
COMMAND ${CMAKE_COMMAND} -E make_directory "${dst}"
)
if(EXISTS ${src})
add_custom_command(TARGET ${TARGET} PRE_BUILD
COMMAND cmake -E copy_directory "${src}" "${dst}"
COMMENT "copying ${src} -> ${dst}")
else()
message(WARNING "${src} not exist!")
endif()
else()
# windows cmd shell will not expand wildcard automatically.
# below expand the files, and copy them by rules.
file(GLOB src_files ${src})
if (NOT "${src_files}" STREQUAL "")
list(REMOVE_DUPLICATES src_files)
endif ()
add_custom_command(TARGET ${TARGET} PRE_BUILD
COMMAND ${CMAKE_COMMAND} -E make_directory "${dst}"
)
foreach (src_file ${src_files})
add_custom_command(TARGET ${TARGET} PRE_BUILD
COMMAND ${CMAKE_COMMAND} -E copy "${src_file}" "${dst}"
COMMENT "copying ${src_file} -> ${dst}")
endforeach ()
endif()
else (WIN32) # not windows
add_custom_command(TARGET ${TARGET} PRE_BUILD
COMMAND mkdir -p "${dst}"
......@@ -95,7 +106,7 @@ copy(xxhash_lib
DEPS xxhash
)
if (NOT PROTOBUF_FOUND)
if (NOT PROTOBUF_FOUND OR WIN32)
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/protobuf")
copy(protobuf_lib
SRCS ${PROTOBUF_INCLUDE_DIR} ${PROTOBUF_LIBRARY}
......@@ -138,27 +149,25 @@ if (WITH_NGRAPH)
)
endif ()
if (NOT WIN32)
if (NOT MOBILE_INFERENCE AND NOT RPI)
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/snappy")
copy(snappy_lib
SRCS ${SNAPPY_INCLUDE_DIR} ${SNAPPY_LIBRARIES}
DSTS ${dst_dir} ${dst_dir}/lib
DEPS snappy)
if (NOT MOBILE_INFERENCE AND NOT RPI)
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/snappy")
copy(snappy_lib
SRCS ${SNAPPY_INCLUDE_DIR} ${SNAPPY_LIBRARIES}
DSTS ${dst_dir} ${dst_dir}/lib
DEPS snappy)
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/snappystream")
copy(snappystream_lib
SRCS ${SNAPPYSTREAM_INCLUDE_DIR} ${SNAPPYSTREAM_LIBRARIES}
DSTS ${dst_dir} ${dst_dir}/lib
DEPS snappystream)
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/snappystream")
copy(snappystream_lib
SRCS ${SNAPPYSTREAM_INCLUDE_DIR} ${SNAPPYSTREAM_LIBRARIES}
DSTS ${dst_dir} ${dst_dir}/lib
DEPS snappystream)
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/zlib")
copy(zlib_lib
SRCS ${ZLIB_INCLUDE_DIR} ${ZLIB_LIBRARIES}
DSTS ${dst_dir} ${dst_dir}/lib
DEPS zlib)
endif ()
endif (NOT WIN32)
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/zlib")
copy(zlib_lib
SRCS ${ZLIB_INCLUDE_DIR} ${ZLIB_LIBRARIES}
DSTS ${dst_dir} ${dst_dir}/lib
DEPS zlib)
endif ()
# paddle fluid module
set(src_dir "${PADDLE_SOURCE_DIR}/paddle/fluid")
......@@ -191,9 +200,21 @@ if (WITH_ANAKIN AND WITH_MKL)
list(APPEND inference_deps anakin_inference_lib)
endif ()
if (TENSORRT_FOUND)
copy(tensorrt_lib DEPS ${inference_deps}
SRCS ${TENSORRT_ROOT}/include/Nv*.h ${TENSORRT_ROOT}/lib/libnvinfer*
DSTS ${FLUID_INSTALL_DIR}/third_party/install/tensorrt/include ${FLUID_INSTALL_DIR}/third_party/install/tensorrt/lib)
endif ()
set(module "inference")
if(WIN32)
set(paddle_fluid_lib ${PADDLE_BINARY_DIR}/paddle/fluid/inference/${CMAKE_BUILD_TYPE}/libpaddle_fluid.*)
else(WIN32)
set(paddle_fluid_lib ${PADDLE_BINARY_DIR}/paddle/fluid/inference/libpaddle_fluid.*)
endif(WIN32)
copy(inference_lib DEPS ${inference_deps}
SRCS ${src_dir}/${module}/*.h ${PADDLE_BINARY_DIR}/paddle/fluid/inference/libpaddle_fluid.*
SRCS ${src_dir}/${module}/*.h ${paddle_fluid_lib}
${src_dir}/${module}/api/paddle_*.h
DSTS ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module}
)
......@@ -233,7 +254,7 @@ copy(third_party DEPS fluid_lib_dist
# only need libpaddle_fluid.so/a and paddle_*.h for inference-only library
copy(inference_api_lib DEPS fluid_lib_dist
SRCS ${FLUID_INSTALL_DIR}/paddle/fluid/inference/libpaddle_fluid.*
SRCS ${paddle_fluid_lib}
${FLUID_INSTALL_DIR}/paddle/fluid/inference/paddle_*.h
DSTS ${FLUID_INFERENCE_INSTALL_DIR}/paddle/lib ${FLUID_INFERENCE_INSTALL_DIR}/paddle/include
)
......
paddle/fluid/API.spec
浏览文件 @ 893957f7
......@@ -37,8 +37,16 @@ paddle.fluid.DataFeedDesc.desc ArgSpec(args=['self'], varargs=None, keywords=Non
paddle.fluid.DataFeedDesc.set_batch_size ArgSpec(args=['self', 'batch_size'], varargs=None, keywords=None, defaults=None)
paddle.fluid.DataFeedDesc.set_dense_slots ArgSpec(args=['self', 'dense_slots_name'], varargs=None, keywords=None, defaults=None)
paddle.fluid.DataFeedDesc.set_use_slots ArgSpec(args=['self', 'use_slots_name'], varargs=None, keywords=None, defaults=None)
paddle.fluid.AsyncExecutor.__init__ ArgSpec(args=['self', 'place'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.AsyncExecutor.run ArgSpec(args=['self', 'program', 'data_feed', 'filelist', 'thread_num', 'fetch', 'debug'], varargs=None, keywords=None, defaults=(False,))
paddle.fluid.AsyncExecutor.__init__ ArgSpec(args=['self', 'place', 'run_mode'], varargs=None, keywords=None, defaults=(None, ''))
paddle.fluid.AsyncExecutor.config_distributed_nodes ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.AsyncExecutor.download_data ArgSpec(args=['self', 'afs_path', 'local_path', 'fs_default_name', 'ugi', 'file_cnt', 'hadoop_home', 'process_num'], varargs=None, keywords=None, defaults=('$HADOOP_HOME', 12))
paddle.fluid.AsyncExecutor.get_instance ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.AsyncExecutor.init_model ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.AsyncExecutor.init_server ArgSpec(args=['self', 'dist_desc'], varargs=None, keywords=None, defaults=None)
paddle.fluid.AsyncExecutor.init_worker ArgSpec(args=['self', 'dist_desc', 'startup_program'], varargs=None, keywords=None, defaults=None)
paddle.fluid.AsyncExecutor.run ArgSpec(args=['self', 'program', 'data_feed', 'filelist', 'thread_num', 'fetch', 'mode', 'debug'], varargs=None, keywords=None, defaults=('', False))
paddle.fluid.AsyncExecutor.save_model ArgSpec(args=['self', 'save_path'], varargs=None, keywords=None, defaults=None)
paddle.fluid.AsyncExecutor.stop ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.io.save_vars ArgSpec(args=['executor', 'dirname', 'main_program', 'vars', 'predicate', 'filename'], varargs=None, keywords=None, defaults=(None, None, None, None))
paddle.fluid.io.save_params ArgSpec(args=['executor', 'dirname', 'main_program', 'filename'], varargs=None, keywords=None, defaults=(None, None))
paddle.fluid.io.save_persistables ArgSpec(args=['executor', 'dirname', 'main_program', 'filename'], varargs=None, keywords=None, defaults=(None, None))
......@@ -77,6 +85,8 @@ paddle.fluid.layers.sequence_softmax ArgSpec(args=['input', 'use_cudnn', 'name']
paddle.fluid.layers.softmax ArgSpec(args=['input', 'use_cudnn', 'name'], varargs=None, keywords=None, defaults=(True, None))
paddle.fluid.layers.pool2d ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name', 'exclusive'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None, True))
paddle.fluid.layers.pool3d ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name', 'exclusive'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None, True))
paddle.fluid.layers.adaptive_pool2d ArgSpec(args=['input', 'pool_size', 'pool_type', 'require_index', 'name'], varargs=None, keywords=None, defaults=('max', False, None))
paddle.fluid.layers.adaptive_pool3d ArgSpec(args=['input', 'pool_size', 'pool_type', 'require_index', 'name'], varargs=None, keywords=None, defaults=('max', False, None))
paddle.fluid.layers.batch_norm ArgSpec(args=['input', 'act', 'is_test', 'momentum', 'epsilon', 'param_attr', 'bias_attr', 'data_layout', 'in_place', 'name', 'moving_mean_name', 'moving_variance_name', 'do_model_average_for_mean_and_var', 'fuse_with_relu', 'use_global_stats'], varargs=None, keywords=None, defaults=(None, False, 0.9, 1e-05, None, None, 'NCHW', False, None, None, None, False, False, False))
paddle.fluid.layers.beam_search_decode ArgSpec(args=['ids', 'scores', 'beam_size', 'end_id', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.conv2d_transpose ArgSpec(args=['input', 'num_filters', 'output_size', 'filter_size', 'padding', 'stride', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, None, 0, 1, 1, None, None, None, True, None, None))
......@@ -198,6 +208,8 @@ paddle.fluid.layers.bilinear_tensor_product ArgSpec(args=['x', 'y', 'size', 'act
paddle.fluid.layers.merge_selected_rows ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.get_tensor_from_selected_rows ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.lstm ArgSpec(args=['input', 'init_h', 'init_c', 'max_len', 'hidden_size', 'num_layers', 'dropout_prob', 'is_bidirec', 'is_test', 'name', 'default_initializer', 'seed'], varargs=None, keywords=None, defaults=(0.0, False, False, None, None, -1))
paddle.fluid.layers.psroi_pool ArgSpec(args=['input', 'rois', 'output_channels', 'spatial_scale', 'pooled_height', 'pooled_width', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.huber_loss ArgSpec(args=['input', 'label', 'delta'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.data ArgSpec(args=['name', 'shape', 'append_batch_size', 'dtype', 'lod_level', 'type', 'stop_gradient'], varargs=None, keywords=None, defaults=(True, 'float32', 0, VarType.LOD_TENSOR, True))
paddle.fluid.layers.open_files ArgSpec(args=['filenames', 'shapes', 'lod_levels', 'dtypes', 'thread_num', 'buffer_size', 'pass_num', 'is_test'], varargs=None, keywords=None, defaults=(None, None, 1, None))
paddle.fluid.layers.read_file ArgSpec(args=['reader'], varargs=None, keywords=None, defaults=None)
......@@ -364,7 +376,7 @@ paddle.fluid.optimizer.MomentumOptimizer.__init__ ArgSpec(args=['self', 'learnin
paddle.fluid.optimizer.MomentumOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None))
paddle.fluid.optimizer.AdagradOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'epsilon', 'regularization', 'name'], varargs=None, keywords=None, defaults=(1e-06, None, None))
paddle.fluid.optimizer.AdagradOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None))
paddle.fluid.optimizer.AdamOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'beta1', 'beta2', 'epsilon', 'regularization', 'name'], varargs=None, keywords=None, defaults=(0.001, 0.9, 0.999, 1e-08, None, None))
paddle.fluid.optimizer.AdamOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'beta1', 'beta2', 'epsilon', 'regularization', 'name', 'lazy_mode'], varargs=None, keywords=None, defaults=(0.001, 0.9, 0.999, 1e-08, None, None, False))
paddle.fluid.optimizer.AdamOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None))
paddle.fluid.optimizer.AdamaxOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'beta1', 'beta2', 'epsilon', 'regularization', 'name'], varargs=None, keywords=None, defaults=(0.001, 0.9, 0.999, 1e-08, None, None))
paddle.fluid.optimizer.AdamaxOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None))
......
paddle/fluid/framework/CMakeLists.txt
浏览文件 @ 893957f7
# windows treat symbolic file as a real file, which is different with unix
# We create a hidden file and compile it instead of origin source file.
#windows treat symbolic file as a real file, which is different with unix
#We create a hidden file and compile it instead of origin source file.
function(windows_symbolic TARGET)
set(oneValueArgs "")
set(multiValueArgs SRCS DEPS)
set(multiValueArgs SRCS PATH)
cmake_parse_arguments(windows_symbolic "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
set(final_path ${CMAKE_CURRENT_SOURCE_DIR}/${windows_symbolic_PATH})
foreach(src ${windows_symbolic_SRCS})
get_filename_component(src ${src} NAME_WE)
if (NOT EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${src}.cc OR NOT EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${src}.cu)
message(FATAL " ${src}.cc and ${src}.cu must exsits, and ${src}.cu must be symbolic file.")
endif()
# only copy the xx.cu to .xx.cu when the content are modified
#only copy the xx.cu to.xx.cu when the content are modified
set(copy_flag 1)
if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/.${src}.cu)
file(READ ${CMAKE_CURRENT_SOURCE_DIR}/${src}.cc SOURCE_STR)
......@@ -32,7 +33,7 @@ endfunction()
add_subdirectory(ir)
add_subdirectory(details)
# ddim lib
#ddim lib
proto_library(framework_proto SRCS framework.proto)
proto_library(async_executor_proto SRCS data_feed.proto)
......@@ -72,6 +73,8 @@ cc_library(lod_tensor SRCS lod_tensor.cc DEPS ddim place tensor framework_proto
cc_test(lod_tensor_test SRCS lod_tensor_test.cc DEPS lod_tensor memory)
nv_test(lod_tensor_gpu_test SRCS lod_tensor_test.cu DEPS lod_tensor)
cc_library(garbage_collector SRCS garbage_collector.cc DEPS device_context memory)
cc_library(reader SRCS reader.cc DEPS lod_tensor ddim)
cc_test(reader_test SRCS reader_test.cc DEPS reader)
......@@ -89,8 +92,8 @@ nv_test(data_device_transform_test SRCS data_device_transform_test.cu
if(WITH_GPU)
if (WIN32)
# windows treat symbolic file as a real file, which is different with unix
# We create a hidden file and compile it instead of origin source file.
#windows treat symbolic file as a real file, which is different with unix
#We create a hidden file and compile it instead of origin source file.
windows_symbolic(hidden_file SRCS data_type_transform.cu)
nv_library(data_type_transform SRCS .data_type_transform.cu DEPS tensor)
add_dependencies(data_type_transform hidden_file)
......@@ -129,17 +132,20 @@ cc_test(version_test SRCS version_test.cc DEPS version)
cc_library(proto_desc SRCS var_desc.cc op_desc.cc block_desc.cc program_desc.cc DEPS shape_inference op_info operator glog version)
if(NOT WIN32)
cc_library(ngraph_bridge SRCS ngraph_bridge.cc DEPS operator framework_proto ngraph)
cc_library(ngraph_operator SRCS ngraph_operator.cc DEPS ngraph_bridge operator op_info device_context tensor scope glog
shape_inference data_transform lod_tensor profiler)
endif(NOT WIN32)
if(WITH_NGRAPH)
if(NOT WIN32)
cc_library(ngraph_bridge SRCS ngraph_bridge.cc DEPS operator framework_proto ngraph)
cc_library(ngraph_operator SRCS ngraph_operator.cc DEPS ngraph_bridge operator op_info device_context tensor scope glog
shape_inference data_transform lod_tensor profiler ngraph)
endif(NOT WIN32)
endif(WITH_NGRAPH)
cc_library(op_registry SRCS op_registry.cc DEPS op_proto_maker op_info operator glog proto_desc)
nv_test(op_registry_test SRCS op_registry_test.cc DEPS op_registry)
py_proto_compile(framework_py_proto SRCS framework.proto data_feed.proto)
# Generate an empty __init__.py to make framework_py_proto as a valid python module.
#Generate an empty \
#__init__.py to make framework_py_proto as a valid python module.
add_custom_target(framework_py_proto_init ALL COMMAND ${CMAKE_COMMAND} -E touch __init__.py)
add_dependencies(framework_py_proto framework_py_proto_init)
if (NOT WIN32)
......@@ -165,24 +171,38 @@ cc_library(variable_helper SRCS variable_helper.cc DEPS lod_tensor)
cc_library(naive_executor SRCS naive_executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass variable_helper)
if(WITH_DISTRIBUTE)
cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method sendrecvop_grpc cares grpc++_unsecure grpc_unsecure gpr graph_to_program_pass variable_helper)
set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
set_source_files_properties(executor.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog
lod_rank_table feed_fetch_method sendrecvop_rpc ${GLOB_DISTRIBUTE_DEPS} graph_to_program_pass variable_helper)
set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
set_source_files_properties(executor.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
else()
if(NOT WIN32)
cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass ngraph_operator variable_helper)
else(NOT WIN32)
if(WITH_NGRAPH)
if(NOT WIN32)
cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass ngraph ngraph_operator variable_helper)
else(NOT WIN32)
cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass variable_helper)
endif(NOT WIN32)
else(WITH_NGRAPH)
cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass variable_helper)
endif(NOT WIN32)
endif(WITH_NGRAPH)
cc_test(test_naive_executor SRCS naive_executor_test.cc DEPS naive_executor elementwise_add_op)
endif()
target_link_libraries(executor garbage_collector)
cc_library(parallel_executor SRCS parallel_executor.cc DEPS
threaded_ssa_graph_executor scope_buffered_ssa_graph_executor
graph build_strategy
fast_threaded_ssa_graph_executor variable_helper)
cc_library(async_executor SRCS async_executor.cc data_feed.cc data_feed_factory.cc executor_thread_worker.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass async_executor_proto variable_helper)
if(WITH_PSLIB)
cc_library(async_executor SRCS async_executor.cc data_feed.cc data_feed_factory.cc executor_thread_worker.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass async_executor_proto variable_helper pslib_brpc pslib)
else()
cc_library(async_executor SRCS async_executor.cc data_feed.cc data_feed_factory.cc executor_thread_worker.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass async_executor_proto variable_helper)
endif(WITH_PSLIB)
cc_test(data_feed_test SRCS data_feed_test.cc DEPS async_executor)
cc_library(prune SRCS prune.cc DEPS framework_proto)
......
paddle/fluid/framework/async_executor.cc
浏览文件 @ 893957f7
......@@ -29,6 +29,9 @@ limitations under the License. */
#include "paddle/fluid/inference/io.h"
#include "paddle/fluid/platform/place.h"
#include "paddle/fluid/pybind/pybind.h"
#ifdef PADDLE_WITH_PSLIB
#include <pslib.h>
#endif
namespace paddle {
namespace framework {
......@@ -47,6 +50,11 @@ void AsyncExecutor::CreateThreads(
worker->SetDataFeed(reader);
worker->SetFetchVarNames(fetch_var_names);
worker->BindingDataFeedMemory();
#ifdef PADDLE_WITH_PSLIB
worker->SetPSlibPtr(_pslib_ptr);
worker->SetPullDenseThread(_pull_dense_thread);
worker->SetParamConfig(&_param_config);
#endif
}
void PrepareReaders(std::vector<std::shared_ptr<DataFeed>>& readers, // NOLINT
......@@ -60,12 +68,177 @@ void PrepareReaders(std::vector<std::shared_ptr<DataFeed>>& readers, // NOLINT
readers[0]->SetFileList(filelist);
}
#ifdef PADDLE_WITH_PSLIB
void AsyncExecutor::InitServer(const std::string& dist_desc, int index) {
_pslib_ptr = std::shared_ptr<paddle::distributed::PSlib>(
new paddle::distributed::PSlib());
_pslib_ptr->init_server(dist_desc, index);
InitParamConfig();
}
void AsyncExecutor::InitWorker(const std::string& dist_desc,
const std::vector<uint64_t>& host_sign_list,
int node_num, int index) {
_pslib_ptr = std::shared_ptr<paddle::distributed::PSlib>(
new paddle::distributed::PSlib());
_pslib_ptr->init_worker(
dist_desc, const_cast<uint64_t*>(host_sign_list.data()), node_num, index);
InitParamConfig();
}
uint64_t AsyncExecutor::StartServer() { return _pslib_ptr->run_server(); }
void AsyncExecutor::StopServer() { _pslib_ptr->stop_server(); }
void AsyncExecutor::GatherServers(const std::vector<uint64_t>& host_sign_list,
int node_num) {
_pslib_ptr->gather_servers(const_cast<uint64_t*>(host_sign_list.data()),
node_num);
}
void AsyncExecutor::InitParamConfig() {
for (int i = 0; i < _pslib_ptr->get_param()
->server_param()
.downpour_server_param()
.downpour_table_param_size();
++i) {
if (_pslib_ptr->get_param()
->server_param()
.downpour_server_param()
.downpour_table_param(i)
.table_class()
.find("SparseTable") != -1) {
_param_config.fea_dim = _pslib_ptr->get_param()
->server_param()
.downpour_server_param()
.downpour_table_param(i)
.accessor()
.fea_dim();
break;
}
}
_param_config.slot_dim = _param_config.fea_dim - 2;
_param_config.tmp_push_dense_wait_times = static_cast<int32_t>(
_pslib_ptr->get_param()->trainer_param().push_dense_per_batch());
_param_config.tmp_push_sparse_wait_times = static_cast<int32_t>(
_pslib_ptr->get_param()->trainer_param().push_sparse_per_batch());
for (auto t = 0u; t < _pslib_ptr->get_param()->trainer_param().skip_op_size();
++t) {
_param_config.skip_op.push_back(
_pslib_ptr->get_param()->trainer_param().skip_op(t));
}
for (auto t = 0u;
t < _pslib_ptr->get_param()->trainer_param().sparse_table_size(); ++t) {
auto& table = _pslib_ptr->get_param()->trainer_param().sparse_table(t);
std::vector<std::string> tmp_sparse_variable_name;
for (int i = 0u; i < table.slot_value_size(); ++i) {
tmp_sparse_variable_name.push_back(table.slot_value(i));
_param_config.slot_alias_to_table[table.slot_key(i)] = table.table_id();
}
std::vector<std::string> tmp_sparse_gradient_variable_name;
for (auto i = 0u; i < table.slot_gradient_size(); ++i) {
tmp_sparse_gradient_variable_name.push_back(table.slot_gradient(i));
}
_param_config.slot_input_vec[table.table_id()] =
std::move(tmp_sparse_variable_name);
_param_config.gradient_var[table.table_id()] =
std::move(tmp_sparse_gradient_variable_name);
_param_config.sparse_table_id.push_back(table.table_id());
}
for (auto t = 0u;
t < _pslib_ptr->get_param()->trainer_param().dense_table_size(); ++t) {
auto& table = _pslib_ptr->get_param()->trainer_param().dense_table(t);
std::vector<std::string> tmp_dense_variable_name;
for (int i = 0u; i < table.dense_variable_name_size(); ++i) {
tmp_dense_variable_name.push_back(table.dense_variable_name(i));
}
std::vector<std::string> tmp_dense_gradient_variable_name;
for (auto i = 0u; i < table.dense_gradient_variable_name_size(); ++i) {
tmp_dense_gradient_variable_name.push_back(
table.dense_gradient_variable_name(i));
}
_param_config.dense_variable_name[table.table_id()] =
std::move(tmp_dense_variable_name);
_param_config.dense_gradient_variable_name[table.table_id()] =
std::move(tmp_dense_gradient_variable_name);
_param_config.dense_table_id.push_back(table.table_id());
_param_config.dense_table_size.push_back(table.fea_dim());
}
}
void AsyncExecutor::InitModel() {
for (auto table_id : _param_config.dense_table_id) {
std::vector<paddle::ps::Region> regions;
for (auto& t : _param_config.dense_variable_name[table_id]) {
Variable* var = root_scope_->FindVar(t);
CHECK(var != nullptr) << "var[" << t << "] not found";
LoDTensor* tensor = var->GetMutable<LoDTensor>();
float* g = tensor->data<float>();
CHECK(g != nullptr) << "var[" << t << "] value not initialized";
float init_range = 0.2;
int rown = tensor->dims()[0];
init_range /= sqrt(rown);
std::normal_distribution<float> ndistr(0.0, 1.0);
for (auto i = 0u; i < tensor->numel(); ++i) {
g[i] = ndistr(local_random_engine()) * init_range;
}
paddle::ps::Region reg(g, tensor->numel());
regions.emplace_back(std::move(reg));
}
auto push_status = _pslib_ptr->_worker_ptr->push_dense_param(
regions.data(), regions.size(), table_id);
push_status.wait();
auto status = push_status.get();
if (status != 0) {
LOG(FATAL) << "push dense param failed, status[" << status << "]";
exit(-1);
}
}
}
void AsyncExecutor::SaveModel(const std::string& path) {
auto ret = _pslib_ptr->_worker_ptr->flush();
ret.wait();
ret = _pslib_ptr->_worker_ptr->save(path, 0);
ret.wait();
int32_t feasign_cnt = ret.get();
if (feasign_cnt == -1) { // (colourful-tree) TODO should be feasign_cnt < 0
LOG(FATAL) << "save model failed";
exit(-1);
}
}
void AsyncExecutor::PrepareDenseThread(const std::string& mode) {
if (mode == "mpi") {
DensePullThreadParam param;
param.ps_client = _pslib_ptr->_worker_ptr;
param.threshold = 1;
param.training_thread_num = actual_thread_num;
param.root_scope = root_scope_;
param.dense_params = &_param_config.dense_variable_name;
_pull_dense_thread =
std::shared_ptr<DensePullThread>(new DensePullThread(param));
_pull_dense_thread->start();
}
}
#endif
void AsyncExecutor::RunFromFile(const ProgramDesc& main_program,
const std::string& data_feed_desc_str,
const std::vector<std::string>& filelist,
const int thread_num,
const std::vector<std::string>& fetch_var_names,
const bool debug) {
const std::string& mode, const bool debug) {
std::vector<std::thread> threads;
auto& block = main_program.Block(0);
......@@ -82,7 +255,7 @@ void AsyncExecutor::RunFromFile(const ProgramDesc& main_program,
google::protobuf::TextFormat::ParseFromString(data_feed_desc_str,
&data_feed_desc);
int actual_thread_num = thread_num;
actual_thread_num = thread_num;
int file_cnt = filelist.size();
PADDLE_ENFORCE(file_cnt > 0, "File list cannot be empty");
......@@ -106,11 +279,21 @@ void AsyncExecutor::RunFromFile(const ProgramDesc& main_program,
// todo: should be factory method for creating datafeed
std::vector<std::shared_ptr<DataFeed>> readers;
PrepareReaders(readers, actual_thread_num, data_feed_desc, filelist);
#ifdef PADDLE_WITH_PSLIB
PrepareDenseThread(mode);
#endif
std::vector<std::shared_ptr<ExecutorThreadWorker>> workers;
workers.resize(actual_thread_num);
for (auto& worker : workers) {
#ifdef PADDLE_WITH_PSLIB
if (mode == "mpi") {
worker.reset(new AsyncExecutorThreadWorker);
} else {
worker.reset(new ExecutorThreadWorker);
}
#else
worker.reset(new ExecutorThreadWorker);
#endif
}
// prepare thread resource here
......@@ -128,7 +311,11 @@ void AsyncExecutor::RunFromFile(const ProgramDesc& main_program,
for (auto& th : threads) {
th.join();
}
#ifdef PADDLE_WITH_PSLIB
if (mode == "mpi") {
_pull_dense_thread->stop();
}
#endif
root_scope_->DropKids();
return;
......
paddle/fluid/framework/async_executor.h
浏览文件 @ 893957f7
......@@ -14,9 +14,11 @@ limitations under the License. */
#pragma once
#include <time.h>
#include <map>
#include <memory>
#include <mutex> // NOLINT
#include <mutex> // NOLINT
#include <random> // local_random_engine
#include <set>
#include <string>
#include <thread> // NOLINT
......@@ -30,6 +32,31 @@ limitations under the License. */
namespace paddle {
namespace framework {
inline double current_realtime() {
#if !defined(_WIN32)
struct timespec tp;
clock_gettime(CLOCK_REALTIME, &tp);
return tp.tv_sec + tp.tv_nsec * 1e-9;
#else
return 0.0;
#endif
}
inline std::default_random_engine& local_random_engine() {
struct engine_wrapper_t {
std::default_random_engine engine;
engine_wrapper_t() {
static std::atomic<uint64_t> x(0);
std::seed_seq sseq = {x++, x++, x++,
static_cast<uint64_t>(current_realtime() * 1000)};
engine.seed(sseq);
}
};
thread_local engine_wrapper_t r;
return r.engine;
}
class AsyncExecutor {
public:
AsyncExecutor(Scope* scope, const platform::Place& place);
......@@ -39,7 +66,19 @@ class AsyncExecutor {
const std::vector<std::string>& filelist,
const int thread_num,
const std::vector<std::string>& fetch_names,
const bool debug = false);
const std::string& mode, const bool debug = false);
#ifdef PADDLE_WITH_PSLIB
void InitServer(const std::string& dist_desc, int index);
void InitWorker(const std::string& dist_desc,
const std::vector<uint64_t>& host_sign_list, int node_num,
int index);
uint64_t StartServer();
void StopServer();
void GatherServers(const std::vector<uint64_t>& host_sign_list, int node_num);
void InitModel();
void SaveModel(const std::string& path);
void InitParamConfig();
#endif
private:
void CreateThreads(ExecutorThreadWorker* worker,
......@@ -48,10 +87,21 @@ class AsyncExecutor {
const std::vector<std::string>& fetch_var_names,
Scope* root_scope, const int thread_index,
const bool debug);
#ifdef PADDLE_WITH_PSLIB
void PrepareDenseThread(const std::string& mode);
#endif
public:
#ifdef PADDLE_WITH_PSLIB
std::shared_ptr<paddle::distributed::PSlib> _pslib_ptr;
std::shared_ptr<DensePullThread> _pull_dense_thread;
AsyncWorkerParamConfig _param_config;
#endif
Scope* root_scope_;
platform::Place place_;
private:
int actual_thread_num;
};
} // namespace framework
......
paddle/fluid/framework/data_feed.cc
浏览文件 @ 893957f7
......@@ -64,6 +64,7 @@ bool DataFeed::PickOneFile(std::string* filename) {
return false;
}
*filename = filelist_[file_idx_++];
LOG(ERROR) << "pick file:" << *filename;
return true;
}
......
paddle/fluid/framework/data_layout_transform.cc
浏览文件 @ 893957f7
......@@ -85,7 +85,7 @@ void TransDataLayout(const OpKernelType& kernel_type_for_var,
out->mutable_data(expected_kernel_type.place_, in.type());
framework::VisitDataType(
framework::ToDataType(in.type()),
in.type(),
CastDataLayout(pool.Get(expected_kernel_type.place_), axis, in, out));
out->set_layout(expected_kernel_type.data_layout_);
......@@ -101,7 +101,7 @@ void* GetDataFromTensor(const Tensor& tensor, mkldnn::memory::data_type type) {
case mkldnn::memory::data_type::f32:
return platform::to_void_cast(tensor.data<float>());
case mkldnn::memory::data_type::s8:
return platform::to_void_cast(tensor.data<char>());
return platform::to_void_cast(tensor.data<int8_t>());
case mkldnn::memory::data_type::u8:
return platform::to_void_cast(tensor.data<unsigned char>());
case mkldnn::memory::data_type::s16:
......@@ -144,7 +144,7 @@ void TransDataLayoutFromMKLDNN(const OpKernelType& kernel_type_for_var,
memory::data_type in_type = ToMKLDNNDataType(in.type());
PADDLE_ENFORCE(in_type != memory::data_type::data_undef,
"Input tensor type is not supported: ", in.type().name());
"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());
......
paddle/fluid/framework/data_layout_transform.h
浏览文件 @ 893957f7
......@@ -50,14 +50,14 @@ inline DataLayout ToPaddleLayout(const MKLDNNFormat& format) {
}
}
inline MKLDNNDataType ToMKLDNNDataType(const std::type_index type) {
static const std::map<std::type_index, MKLDNNDataType> dict{
{std::type_index(typeid(float)), MKLDNNDataType::f32}, // NOLINT
{std::type_index(typeid(char)), MKLDNNDataType::s8}, // NOLINT
{std::type_index(typeid(unsigned char)), MKLDNNDataType::u8},
{std::type_index(typeid(int16_t)), MKLDNNDataType::s16},
{std::type_index(typeid(int32_t)), MKLDNNDataType::s32}};
auto iter = dict.find(type);
inline MKLDNNDataType ToMKLDNNDataType(proto::VarType::Type type) {
static std::unordered_map<int, MKLDNNDataType> dict{
{DataTypeTrait<float>::DataType, MKLDNNDataType::f32},
{DataTypeTrait<int8_t>::DataType, MKLDNNDataType::s8},
{DataTypeTrait<uint8_t>::DataType, MKLDNNDataType::u8},
{DataTypeTrait<int16_t>::DataType, MKLDNNDataType::s16},
{DataTypeTrait<int32_t>::DataType, MKLDNNDataType::s32}};
auto iter = dict.find(static_cast<int>(type));
if (iter != dict.end()) return iter->second;
return MKLDNNDataType::data_undef;
}
......
paddle/fluid/framework/data_type.cc
浏览文件 @ 893957f7
......@@ -26,7 +26,7 @@ struct DataTypeMap {
std::unordered_map<std::type_index, proto::VarType::Type> cpp_to_proto_;
std::unordered_map<int, std::type_index> proto_to_cpp_;
std::unordered_map<int, std::string> proto_to_str_;
std::unordered_map<std::type_index, size_t> cpp_to_size_;
std::unordered_map<int, size_t> proto_to_size_;
};
static DataTypeMap* InitDataTypeMap();
......@@ -45,7 +45,7 @@ static inline void RegisterType(DataTypeMap* map,
map->proto_to_cpp_.emplace(static_cast<int>(proto_type), typeid(T));
map->cpp_to_proto_.emplace(typeid(T), proto_type);
map->proto_to_str_.emplace(static_cast<int>(proto_type), name);
map->cpp_to_size_.emplace(typeid(T), sizeof(T));
map->proto_to_size_.emplace(static_cast<int>(proto_type), sizeof(T));
}
static DataTypeMap* InitDataTypeMap() {
......@@ -54,17 +54,7 @@ static DataTypeMap* InitDataTypeMap() {
#define RegType(cc_type, proto_type) \
RegisterType<cc_type>(retv, proto_type, #cc_type)
// NOTE: Add your customize type here.
RegType(float16, proto::VarType::FP16);
RegType(float, proto::VarType::FP32);
RegType(double, proto::VarType::FP64);
RegType(int, proto::VarType::INT32);
RegType(int64_t, proto::VarType::INT64);
RegType(bool, proto::VarType::BOOL);
RegType(size_t, proto::VarType::SIZE_T);
RegType(int16_t, proto::VarType::INT16);
RegType(uint8_t, proto::VarType::UINT8);
RegType(int8_t, proto::VarType::INT8);
_ForEachDataType_(RegType);
#undef RegType
return retv;
......@@ -96,12 +86,12 @@ std::string DataTypeToString(const proto::VarType::Type type) {
static_cast<int>(type));
}
size_t SizeOfType(std::type_index type) {
auto it = gDataTypeMap().cpp_to_size_.find(type);
if (it != gDataTypeMap().cpp_to_size_.end()) {
size_t SizeOfType(proto::VarType::Type type) {
auto it = gDataTypeMap().proto_to_size_.find(static_cast<int>(type));
if (it != gDataTypeMap().proto_to_size_.end()) {
return it->second;
}
PADDLE_THROW("Not support %s as tensor type", type.name());
PADDLE_THROW("Not support %s as tensor type", DataTypeToString(type));
}
} // namespace framework
......
paddle/fluid/framework/data_type.h
浏览文件 @ 893957f7
......@@ -22,46 +22,59 @@ limitations under the License. */
namespace paddle {
namespace framework {
template <typename T>
struct DataTypeTrait {};
// Stub handle for void
template <>
struct DataTypeTrait<void> {
constexpr static auto DataType = proto::VarType::RAW;
};
#define _ForEachDataTypeHelper_(callback, cpp_type, proto_type) \
callback(cpp_type, ::paddle::framework::proto::VarType::proto_type);
#define _ForEachDataType_(callback) \
_ForEachDataTypeHelper_(callback, float, FP32); \
_ForEachDataTypeHelper_(callback, ::paddle::platform::float16, FP16); \
_ForEachDataTypeHelper_(callback, double, FP64); \
_ForEachDataTypeHelper_(callback, int, INT32); \
_ForEachDataTypeHelper_(callback, int64_t, INT64); \
_ForEachDataTypeHelper_(callback, bool, BOOL); \
_ForEachDataTypeHelper_(callback, uint8_t, UINT8); \
_ForEachDataTypeHelper_(callback, int16_t, INT16); \
_ForEachDataTypeHelper_(callback, int8_t, INT8)
#define DefineDataTypeTrait(cpp_type, proto_type) \
template <> \
struct DataTypeTrait<cpp_type> { \
constexpr static auto DataType = proto_type; \
}
_ForEachDataType_(DefineDataTypeTrait);
#undef DefineDataTypeTrait
extern proto::VarType::Type ToDataType(std::type_index type);
extern std::type_index ToTypeIndex(proto::VarType::Type type);
template <typename Visitor>
inline void VisitDataType(proto::VarType::Type type, Visitor visitor) {
switch (type) {
case proto::VarType::FP16:
visitor.template apply<platform::float16>();
break;
case proto::VarType::FP32:
visitor.template apply<float>();
break;
case proto::VarType::FP64:
visitor.template apply<double>();
break;
case proto::VarType::INT32:
visitor.template apply<int>();
break;
case proto::VarType::INT64:
visitor.template apply<int64_t>();
break;
case proto::VarType::BOOL:
visitor.template apply<bool>();
break;
case proto::VarType::UINT8:
visitor.template apply<uint8_t>();
break;
case proto::VarType::INT16:
visitor.template apply<int16_t>();
break;
case proto::VarType::INT8:
visitor.template apply<int8_t>();
break;
default:
PADDLE_THROW("Not supported %d", type);
}
#define VisitDataTypeCallback(cpp_type, proto_type) \
do { \
if (type == proto_type) { \
visitor.template apply<cpp_type>(); \
return; \
} \
} while (0)
_ForEachDataType_(VisitDataTypeCallback);
#undef VisitDataTypeCallback
PADDLE_THROW("Not supported %d", type);
}
extern std::string DataTypeToString(const proto::VarType::Type type);
extern size_t SizeOfType(std::type_index type);
extern size_t SizeOfType(proto::VarType::Type type);
inline std::ostream& operator<<(std::ostream& out,
const proto::VarType::Type& type) {
out << DataTypeToString(type);
......
paddle/fluid/framework/data_type_test.cc
浏览文件 @ 893957f7
......@@ -26,15 +26,15 @@ TEST(DataType, float16) {
Tensor tensor;
CPUPlace cpu;
tensor.mutable_data(cpu, f::ToTypeIndex(dtype));
tensor.mutable_data(cpu, dtype);
// test fp16 tensor
EXPECT_EQ(tensor.type(), std::type_index(typeid(float16)));
EXPECT_EQ(tensor.type(), f::ToDataType(typeid(float16)));
// test fp16 size
EXPECT_EQ(f::SizeOfType(f::ToTypeIndex(dtype)), 2u);
EXPECT_EQ(f::SizeOfType(dtype), 2u);
// test debug info
std::string type = "float16";
std::string type = "::paddle::platform::float16";
EXPECT_STREQ(f::DataTypeToString(dtype).c_str(), type.c_str());
}
paddle/fluid/framework/details/CMakeLists.txt
浏览文件 @ 893957f7
......@@ -12,12 +12,19 @@ cc_library(multi_devices_graph_check_pass SRCS multi_devices_graph_check_pass.cc
cc_library(variable_visitor SRCS variable_visitor.cc DEPS lod_tensor selected_rows)
if(WITH_DISTRIBUTE)
if(NOT WITH_GRPC)
set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
set_source_files_properties(reduce_op_handle.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
endif()
endif()
if(WITH_GPU)
nv_library(all_reduce_op_handle SRCS all_reduce_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory
dynload_cuda variable_visitor)
if(WITH_DISTRIBUTE)
nv_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope
ddim dynload_cuda selected_rows_functor sendrecvop_grpc)
ddim dynload_cuda selected_rows_functor sendrecvop_rpc)
else()
nv_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope
ddim dynload_cuda selected_rows_functor)
......@@ -30,7 +37,7 @@ else()
variable_visitor)
if(WITH_DISTRIBUTE)
cc_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope
ddim selected_rows_functor sendrecvop_grpc)
ddim selected_rows_functor sendrecvop_rpc)
else()
cc_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope
ddim selected_rows_functor)
......@@ -43,12 +50,14 @@ 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_pass SRCS analysis_var_pass.cc memory_reuse_types.cc DEPS graph graph_helper pass)
cc_library(modify_op_lock_and_record_event_pass SRCS modify_op_lock_and_record_event_pass.cc DEPS computation_op_handle op_graph_view multi_devices_helper)
if (WITH_GPU)
cc_library(reference_count_pass SRCS reference_count_pass.cc DEPS computation_op_handle scale_loss_grad_op_handle rpc_op_handle
all_reduce_op_handle reduce_op_handle broadcast_op_handle data_balance_op_handle graph graph_helper pass)
endif()
cc_library(memory_early_delete_pass SRCS memory_early_delete_pass.cc DEPS memory_optimize_pass computation_op_handle scale_loss_grad_op_handle rpc_op_handle
all_reduce_op_handle reduce_op_handle broadcast_op_handle data_balance_op_handle graph graph_helper pass)
cc_library(reference_count_pass_helper SRCS reference_count_pass_helper.cc DEPS garbage_collector computation_op_handle)
cc_library(eager_deletion_op_handle SRCS eager_deletion_op_handle.cc DEPS lod_tensor selected_rows reference_count_pass_helper)
cc_library(eager_deletion_pass SRCS eager_deletion_pass.cc DEPS computation_op_handle eager_deletion_op_handle graph graph_helper pass)
cc_library(reference_count_pass SRCS reference_count_pass.cc DEPS computation_op_handle graph graph_helper pass op_graph_view reference_count_pass_helper)
cc_library(sequential_execution_pass SRCS sequential_execution_pass.cc DEPS graph graph_helper pass)
cc_library(all_reduce_deps_pass SRCS all_reduce_deps_pass.cc DEPS graph graph_helper pass)
......@@ -56,10 +65,12 @@ cc_library(all_reduce_deps_pass SRCS all_reduce_deps_pass.cc DEPS graph graph_he
cc_library(multi_devices_graph_pass SRCS multi_devices_graph_pass.cc DEPS multi_devices_helper computation_op_handle
scale_loss_grad_op_handle rpc_op_handle all_reduce_op_handle reduce_op_handle broadcast_op_handle data_balance_op_handle fused_broadcast_op_handle)
set(SSA_GRAPH_EXECUTOR_DEPS graph framework_proto sequential_execution_pass modify_op_lock_and_record_event_pass all_reduce_deps_pass)
set(SSA_GRAPH_EXECUTOR_DEPS graph framework_proto sequential_execution_pass modify_op_lock_and_record_event_pass all_reduce_deps_pass reference_count_pass eager_deletion_pass memory_optimize_pass memory_early_delete_pass)
if (WITH_GPU)
list(APPEND SSA_GRAPH_EXECUTOR_DEPS reference_count_pass)
endif()
cc_test(memory_reuse_types_test SRCS memory_reuse_types_test.cc memory_reuse_types.cc DEPS framework_proto graph)
cc_test(analysis_var_pass_test SRCS analysis_var_pass_test.cc analysis_var_pass.cc memory_reuse_types.cc DEPS framework_proto graph graph_helper op_registry pass)
cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS ${SSA_GRAPH_EXECUTOR_DEPS})
......@@ -80,4 +91,5 @@ cc_test(fused_broadcast_op_test SRCS fused_broadcast_op_handle_test.cc DEPS fuse
cc_library(build_strategy SRCS build_strategy.cc DEPS
graph_viz_pass multi_devices_graph_pass
multi_devices_graph_print_pass multi_devices_graph_check_pass
fuse_elewise_add_act_pass multi_batch_merge_pass)
fuse_elewise_add_act_pass multi_batch_merge_pass
memory_optimize_pass)
paddle/fluid/framework/details/all_reduce_op_handle.cc
浏览文件 @ 893957f7
......@@ -127,7 +127,7 @@ void AllReduceOpHandle::RunImpl() {
// Reduce All Tensor to trg in CPU
ReduceLoDTensor func(lod_tensors, &trg);
VisitDataType(ToDataType(lod_tensors[0]->type()), func);
VisitDataType(lod_tensors[0]->type(), func);
for (size_t i = 1; i < local_scopes_.size(); ++i) {
auto &scope =
......
paddle/fluid/framework/details/analysis_var_pass.cc 0 → 100644
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/framework/details/analysis_var_pass.h 0 → 100644
浏览文件 @ 893957f7
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#pragma once
#include <algorithm>
#include <list>
#include <map>
#include <memory>
#include <set>
#include <string>
#include <unordered_map>
#include <utility>
#include <vector>
#include "paddle/fluid/framework/data_type.h"
#include "paddle/fluid/framework/details/memory_reuse_types.h"
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/pass.h"
namespace paddle {
namespace framework {
namespace details {
constexpr char kAllOpDescs[] = "all_op_descs";
std::vector<ir::Node*> SortOpLikeDescOrder(const ir::Graph& graph);
// sort op in bfs order
std::vector<ir::Node*> BFSSortGraphOps(const ir::Graph& graph);
class ControlFlowGraph;
class AnalysisVarPass : public ir::Pass {
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
private:
// fill the variable map(var_nodes) by version.
void InitSSAGraphNodes() const;
// update program descs
void RenameVarInGraphDesc(const std::string& var,
const std::string& cache_var, size_t idx) const;
// update ir nodes
void RenameVarInGraphNode(const std::string& var,
const std::string& cache_var, size_t idx,
ir::Graph* graph) const;
void SubGraphOptimize(OpDesc* op_desc) const;
// valid a tensor can be reuse or not
bool NodeCanReused(ir::Node* node) const;
// scan subblock and collect the output/input variables.
std::unordered_set<std::string> GetSubBlockVars(
const std::unordered_set<ir::Node*>&) const;
// check op has subblock or not
bool OpHasSubBlock(OpDesc* desc) const;
private:
// Reuse Node Pool, Owned.
mutable OrderedNodePairPool pool_;
// controlflow Graph
mutable std::unique_ptr<ControlFlowGraph> cfg_;
// skip set
mutable std::unordered_set<std::string> skip_set_;
// var nodes
mutable std::map<std::string, std::vector<ir::Node*>> var_nodes_;
};
class ControlFlowGraph {
public:
ControlFlowGraph() = default;
// For IR Graph in parallelexecutor
explicit ControlFlowGraph(const ir::Graph& graph);
void LiveVariableAnalysis();
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;
std::vector<ir::Node*>& Ops();
// for ssa-graph nodes
ir::Node* GetNodeFromVarName(const std::string& name, ir::Node* op) const;
private:
void BuildCFGGraph();
void ConnectNodes();
using NodeListMap = std::unordered_map<ir::Node*, std::set<ir::Node*>>;
using VarSetMap = std::map<ir::Node*, std::set<std::string>>;
// successors ops use the output variables.
NodeListMap successors_;
// predecessors ops generated input variables.
NodeListMap predecessors_;
// variables lived before run current op.
VarSetMap live_in_;
// variables lived after run current op.
VarSetMap live_out_;
VarSetMap uses_; // op inputs
VarSetMap defs_; // op outputs
std::vector<ir::Node*> ops_; // op sequence by topology sort
};
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/analysis_var_pass_test.cc 0 → 100644
浏览文件 @ 893957f7
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/framework/details/analysis_var_pass.h"
#include <algorithm>
#include <iostream>
#include <iterator>
#include "glog/logging.h"
#include "gtest/gtest.h"
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/graph_helper.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/framework/program_desc.h"
namespace paddle {
namespace framework {
class DummyOp : public OperatorBase {
public:
DummyOp(const std::string& type, const VariableNameMap& inputs,
const VariableNameMap& outputs, const AttributeMap& attrs)
: OperatorBase(type, inputs, outputs, attrs) {}
private:
void RunImpl(const Scope& scope,
const platform::Place& place) const override {}
};
class SumOpMaker : public OpProtoAndCheckerMaker {
public:
void Make() {
AddInput("X", "").AsDuplicable();
AddOutput("Out", "");
AddComment("");
}
};
class AssignOpMaker : public OpProtoAndCheckerMaker {
public:
void Make() {
AddInput("X", "").AsDuplicable();
AddOutput("Out", "");
AddComment("");
}
};
class DummyVarTypeInference : public VarTypeInference {
public:
void operator()(const OpDesc& op_desc, BlockDesc* block) const override {
auto& inputs = op_desc.Input("X");
auto type = block->Var(inputs.front())->GetType();
auto out_var_name = op_desc.Output("Out").front();
block->Var(out_var_name)->SetType(type);
}
};
} // namespace framework
} // namespace paddle
REGISTER_OPERATOR(sum, paddle::framework::DummyOp,
paddle::framework::SumOpMaker,
paddle::framework::DummyVarTypeInference);
REGISTER_OPERATOR(assign, paddle::framework::DummyOp,
paddle::framework::AssignOpMaker,
paddle::framework::DummyVarTypeInference);
REGISTER_OPERATOR(dummy, paddle::framework::DummyOp,
paddle::framework::SumOpMaker,
paddle::framework::DummyVarTypeInference);
/*
https://en.wikipedia.org/wiki/Live_variable_analysis
Create a customed classical dependency graph, left row is the instruction
number.
1. a = 1
2. b = a
3. c = a
4. d = b + c
5. e = d
a--------+
| |
b c
| |
d--------+
|
e
Then analysis these variable's liveness range
*/
namespace paddle {
namespace framework {
namespace details {
static inline bool IsSameDesc(OpDesc* op1, OpDesc* op2) {
return op1->Type() == op2->Type() && op1->Inputs() == op2->Inputs() &&
op1->Outputs() == op2->Outputs();
}
inline static ProgramDesc FillProgramDesc() {
ProgramDesc prog;
prog.MutableBlock(0)->Var("a")->SetType(proto::VarType::LOD_TENSOR);
prog.MutableBlock(0)->Var("b")->SetType(proto::VarType::LOD_TENSOR);
prog.MutableBlock(0)->Var("c")->SetType(proto::VarType::LOD_TENSOR);
prog.MutableBlock(0)->Var("d")->SetType(proto::VarType::LOD_TENSOR);
prog.MutableBlock(0)->Var("e")->SetType(proto::VarType::LOD_TENSOR);
{
auto* op = prog.MutableBlock(0)->AppendOp();
op->SetType("assign");
op->SetInput("X", {"a"});
op->SetOutput("Out", {"b"});
}
{
auto* op = prog.MutableBlock(0)->AppendOp();
op->SetType("assign");
op->SetInput("X", {"a"});
op->SetOutput("Out", {"c"});
}
{
auto* op = prog.MutableBlock(0)->AppendOp();
op->SetType("sum");
op->SetInput("X", {"b", "c"});
op->SetOutput("Out", {"d"});
}
{
auto* op = prog.MutableBlock(0)->AppendOp();
op->SetType("assign");
op->SetInput("X", {"d"});
op->SetOutput("Out", {"e"});
}
return prog;
}
template <typename Container>
inline static std::string DebugString(const Container& c) {
std::stringstream ss;
for (auto& item : c) {
ss << item << " ";
}
return ss.str();
}
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();
// test assign op
ASSERT_TRUE((std::set<std::string>{"a"} == cfg.LiveIn(cfg.Ops()[0])));
ASSERT_TRUE((std::set<std::string>{"a", "b"} == cfg.LiveOut(cfg.Ops()[0])));
// test assign op
ASSERT_TRUE((std::set<std::string>{"a", "b"} == cfg.LiveIn(cfg.Ops()[1])));
ASSERT_TRUE((std::set<std::string>{"b", "c"} == cfg.LiveOut(cfg.Ops()[1])));
// test sum op
ASSERT_TRUE((std::set<std::string>{"b", "c"} == cfg.LiveIn(cfg.Ops()[2])));
ASSERT_TRUE((std::set<std::string>{"d"} == cfg.LiveOut(cfg.Ops()[2])));
// test assign op
ASSERT_TRUE((std::set<std::string>{"d"} == cfg.LiveIn(cfg.Ops()[3])));
ASSERT_TRUE((std::set<std::string>{} == cfg.LiveOut(cfg.Ops()[3])));
}
// 1. normal test
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();
for (size_t i = 0; i < nodes.size(); ++i) {
auto node = nodes[i];
auto op_desc = op_descs[i];
ASSERT_TRUE(IsSameDesc(node->Op(), op_desc));
}
}
// 2. remove some op_desc
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;
for (auto node : nodes) {
if (node->IsOp() && node->outputs.back()->Name() == "e") {
found_node = node;
break;
}
}
PADDLE_ENFORCE(found_node != nullptr);
for (auto it = op_descs.begin(); it != op_descs.end();) {
if (IsSameDesc(*it, found_node->Op())) {
it = op_descs.erase(it);
} else {
++it;
}
}
auto find_node_in_graph = [&](std::string s) {
ir::Node* ret = nullptr;
for (auto n : graph.Nodes()) {
if (n->Name() == s) {
ret = n;
break;
}
}
PADDLE_ENFORCE(ret != nullptr);
return ret;
};
ir::Node* e = find_node_in_graph("e");
ir::Node* d = find_node_in_graph("d");
std::remove(d->outputs.begin(), d->outputs.end(), found_node);
graph.RemoveNode(found_node);
graph.RemoveNode(e);
// other node keeps the same order
auto remain_nodes = SortOpLikeDescOrder(graph);
for (size_t i = 0; i < remain_nodes.size(); ++i) {
auto node = remain_nodes[i];
auto op_desc = op_descs[i];
ASSERT_TRUE(IsSameDesc(node->Op(), op_desc));
}
}
// 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) {
ir::Node* ret = nullptr;
for (auto n : graph.Nodes()) {
if (n->Name() == s) {
ret = n;
break;
}
}
PADDLE_ENFORCE(ret != nullptr);
return ret;
};
// 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();
auto op = prog.MutableBlock(0)->AppendOp();
prog.MutableBlock(0)->Var("d1")->SetType(proto::VarType::LOD_TENSOR);
op->SetType("sum");
op->SetInput("X", {"b", "c"});
op->SetOutput("Out", {"d1"});
ir::Node* node = graph.CreateOpNode(op);
ir::Node* d1 = graph.CreateVarNode(prog.MutableBlock(0)->Var("d1"));
ir::Node* b = find_node_in_graph("b");
ir::Node* c = find_node_in_graph("c");
node->outputs.emplace_back(d1);
node->inputs.emplace_back(b);
node->inputs.emplace_back(c);
d1->inputs.emplace_back(node);
b->outputs.emplace_back(node);
c->outputs.emplace_back(node);
op_descs.insert(op_descs.begin() + 4, op);
auto nodes = SortOpLikeDescOrder(graph);
for (size_t i = 0; i < nodes.size(); ++i) {
auto node = nodes[i];
auto op_desc = op_descs[i];
ASSERT_TRUE(IsSameDesc(node->Op(), op_desc));
}
}
// 4. add and delete some op_desc
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;
for (auto n : graph.Nodes()) {
if (n->Name() == s) {
ret = n;
break;
}
}
PADDLE_ENFORCE(ret != nullptr);
return ret;
};
// remove sum node
auto op_descs = prog.Block(0).AllOps();
ir::Node* found_node = nullptr;
auto nodes = graph.Nodes();
for (auto node : nodes) {
if (node->Name() == "sum") {
found_node = node;
break;
}
}
PADDLE_ENFORCE(found_node != nullptr);
for (auto it = op_descs.begin(); it != op_descs.end();) {
if (IsSameDesc(*it, found_node->Op())) {
it = op_descs.erase(it);
} else {
++it;
}
}
{
ir::Node* d = find_node_in_graph("d");
ir::Node* c = find_node_in_graph("c");
ir::Node* e = find_node_in_graph("e");
std::remove(d->outputs.begin(), d->outputs.end(), found_node);
std::remove(c->outputs.begin(), c->outputs.end(), found_node);
ir::Node* pending_op = found_node->outputs[0]->outputs[0];
graph.RemoveNode(e);
graph.RemoveNode(pending_op);
graph.RemoveNode(found_node);
}
// add node
auto op = prog.MutableBlock(0)->AppendOp();
prog.MutableBlock(0)->Var("d1")->SetType(proto::VarType::LOD_TENSOR);
op->SetType("sum");
op->SetInput("X", {"b", "c"});
op->SetOutput("Out", {"d1"});
{
ir::Node* node = graph.CreateOpNode(op);
ir::Node* d1 = graph.CreateVarNode(prog.MutableBlock(0)->Var("d1"));
ir::Node* b = find_node_in_graph("b");
ir::Node* c = find_node_in_graph("c");
node->outputs.emplace_back(d1);
node->inputs.emplace_back(b);
node->inputs.emplace_back(c);
b->outputs.emplace_back(node);
c->outputs.emplace_back(node);
}
op_descs.insert(op_descs.begin() + 2, op);
// check the order
auto mynodes = SortOpLikeDescOrder(graph);
for (size_t i = 0; i < mynodes.size(); ++i) {
auto node = mynodes[i];
auto op_desc = op_descs[i];
ASSERT_TRUE(IsSameDesc(node->Op(), op_desc));
}
}
// 5. add and replace some op_desc inplace.
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
auto find_node_in_graph = [&](std::string s) {
ir::Node* ret = nullptr;
for (auto n : graph.Nodes()) {
if (n->Name() == s) {
ret = n;
break;
}
}
PADDLE_ENFORCE(ret != nullptr);
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);
op->SetType("sum");
op->SetInput("X", {"b", "c"});
op->SetOutput("Out", {"d1"});
{
ir::Node* node = graph.CreateOpNode(op);
ir::Node* d1 = graph.CreateVarNode(prog.MutableBlock(0)->Var("d1"));
ir::Node* b = find_node_in_graph("b");
ir::Node* c = find_node_in_graph("c");
node->outputs.emplace_back(d1);
node->inputs.emplace_back(b);
node->inputs.emplace_back(c);
d1->inputs.emplace_back(node);
b->outputs.emplace_back(node);
c->outputs.emplace_back(node);
}
op_descs.emplace_back(op);
// replace op_desc inplace
auto nodes = graph.Nodes();
ir::Node* found_node = nullptr;
for (auto node : nodes) {
if (node->IsOp() && node->Op() && node->Name() == "assign") {
if (node->outputs.size() == 1 && node->outputs[0]->Name() == "e") {
found_node = node;
break;
}
}
}
{
ir::Node* d = find_node_in_graph("d");
ir::Node* e = find_node_in_graph("e");
std::remove(d->outputs.begin(), d->outputs.end(), found_node);
std::remove(e->inputs.begin(), e->inputs.end(), found_node);
graph.RemoveNode(found_node);
}
op_descs.erase(op_descs.begin() + 3);
auto replace_op = prog.MutableBlock(0)->AppendOp();
replace_op->SetType("sum");
replace_op->SetInput("X", {"d", "d1"});
replace_op->SetOutput("Out", {"e"});
{
ir::Node* sum2 = graph.CreateOpNode(replace_op);
ir::Node* e = find_node_in_graph("e");
ir::Node* d = find_node_in_graph("d");
ir::Node* d1 = find_node_in_graph("d1");
sum2->inputs.emplace_back(d);
sum2->inputs.emplace_back(d1);
sum2->outputs.emplace_back(e);
e->inputs.emplace_back(sum2);
d->outputs.emplace_back(sum2);
d1->outputs.emplace_back(sum2);
}
op_descs.emplace_back(replace_op);
// compare op order
auto graph_nodes = SortOpLikeDescOrder(graph);
for (size_t i = 0; i < graph_nodes.size(); ++i) {
auto node = graph_nodes[i];
auto op_desc = op_descs[i];
ASSERT_TRUE(IsSameDesc(node->Op(), op_desc));
}
}
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/build_strategy.cc
浏览文件 @ 893957f7
......@@ -14,11 +14,16 @@ limitations under the License. */
#include "paddle/fluid/framework/details/build_strategy.h"
#include <glog/logging.h>
#include <memory>
#include "paddle/fluid/framework/details/memory_reuse_types.h"
#include "paddle/fluid/framework/details/multi_devices_graph_check_pass.h"
#include "paddle/fluid/framework/details/multi_devices_graph_print_pass.h"
#include "paddle/fluid/framework/details/reduce_op_handle.h"
#include "paddle/fluid/framework/details/sequential_execution_pass.h"
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/graph_helper.h"
#include "paddle/fluid/framework/ir/graph_viz_pass.h"
namespace paddle {
......@@ -69,6 +74,14 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
}
VLOG(1) << "CollectiveContext:" << context->String();
// NOTE(dzh): memory optimize should be a runtime pass.
// However, after multi_devices_pass, VarHandle, OpHandle is
// the de-fact IR, any reuse on Graph is meaningless.
// A side-effect of that, memory optimize cannot forsee the fetched vars
// , so fetchlist should be set persistable before call the Run interface.
if (strategy.memory_optimize_) {
auto analysis_var_pass = AppendPass("analysis_var_pass");
}
// Convert graph to run on multi-devices.
auto multi_devices_pass = AppendPass("multi_devices_pass");
multi_devices_pass->SetNotOwned<const BuildStrategy>("strategy",
......@@ -79,8 +92,11 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
// Add a graph print pass to record a graph with device info.
if (!strategy_.debug_graphviz_path_.empty()) {
auto multi_devices_print_pass = AppendPass("multi_devices_print_pass");
multi_devices_print_pass->SetNotOwned<const std::string>(
"debug_graphviz_path", &strategy_.debug_graphviz_path_);
const std::string graph_path =
string::Sprintf("%s%s", strategy_.debug_graphviz_path_.c_str(),
"_multi_devices_graph");
multi_devices_print_pass->Set<std::string>(kGraphvizPath,
new std::string(graph_path));
multi_devices_print_pass->Set<details::GraphvizSSAGraphPrinter>(
"graph_printer", new details::GraphvizSSAGraphPrinter);
}
......@@ -127,7 +143,6 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
CreatePassesFromStrategy(false);
std::unique_ptr<ir::Graph> graph(new ir::Graph(main_program));
for (std::shared_ptr<ir::Pass> &pass : pass_builder_->AllPasses()) {
if (pass->Type() == "multi_devices_pass") {
pass->Erase("places");
......@@ -145,6 +160,17 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
pass->Erase("nccl_ctxs");
pass->SetNotOwned<platform::NCCLContextMap>("nccl_ctxs", nctx);
#endif
} else if (pass->Type() == "analysis_var_pass") {
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
graph->Set<GraphNodePool>(kGraphNodePool,
new GraphNodePool); // 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_;
......@@ -166,6 +192,7 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
}
return graph;
}
} // namespace details
} // namespace framework
} // namespace paddle
......@@ -176,6 +203,7 @@ USE_PASS(multi_batch_merge_pass);
USE_PASS(multi_devices_pass);
USE_PASS(multi_devices_check_pass);
USE_PASS(multi_devices_print_pass);
USE_PASS(analysis_var_pass);
USE_PASS(sequential_execution_pass);
USE_PASS(all_reduce_deps_pass);
USE_PASS(modify_op_lock_and_record_event_pass);
paddle/fluid/framework/details/build_strategy.h
浏览文件 @ 893957f7
......@@ -60,8 +60,15 @@ struct BuildStrategy {
kCustomized = 2,
};
enum class OptimizeStrategy {
// To be Implemented,bruteforce, recursive compute unused var names.
kBruteForce = 0,
kControlFlowGraph = 1, // use cfg_graph algorithm, faster speed.
};
ReduceStrategy reduce_{ReduceStrategy::kAllReduce};
GradientScaleStrategy gradient_scale_{GradientScaleStrategy::kCoeffNumDevice};
OptimizeStrategy strategy_{OptimizeStrategy::kControlFlowGraph};
std::string debug_graphviz_path_{""};
......@@ -69,6 +76,10 @@ struct BuildStrategy {
bool enable_data_balance_{false};
bool memory_optimize_{false};
bool memory_early_delete_{false};
bool enable_sequential_execution_{false};
bool fuse_broadcast_op_{false};
......
paddle/fluid/framework/details/computation_op_handle.cc
浏览文件 @ 893957f7
......@@ -20,11 +20,13 @@ namespace paddle {
namespace framework {
namespace details {
ComputationOpHandle::ComputationOpHandle(ir::Node *node, Scope *scope,
platform::Place place)
platform::Place place,
size_t scope_idx)
: OpHandleBase(node),
op_(framework::OpRegistry::CreateOp(*node->Op())),
scope_(scope),
place_(place) {}
place_(place),
scope_idx_(scope_idx) {}
void ComputationOpHandle::RunImpl() {
WaitInputVarGenerated(place_);
......
paddle/fluid/framework/details/computation_op_handle.h
浏览文件 @ 893957f7
......@@ -28,7 +28,8 @@ namespace framework {
namespace details {
struct ComputationOpHandle : public OpHandleBase {
public:
ComputationOpHandle(ir::Node *node, Scope *scope, platform::Place place);
ComputationOpHandle(ir::Node *node, Scope *scope, platform::Place place,
size_t scope_idx);
std::string Name() const override;
......@@ -38,6 +39,8 @@ struct ComputationOpHandle : public OpHandleBase {
void SetLockAndRecordEventFree(bool b) { is_lock_and_record_event_free_ = b; }
size_t GetScopeIdx() const { return scope_idx_; }
protected:
void RunImpl() override;
......@@ -47,6 +50,7 @@ struct ComputationOpHandle : public OpHandleBase {
std::unique_ptr<OperatorBase> op_;
Scope *scope_;
platform::Place place_;
size_t scope_idx_;
bool is_lock_and_record_event_free_{false};
};
} // namespace details
......
paddle/fluid/framework/details/eager_deletion_op_handle.cc 0 → 100644
浏览文件 @ 893957f7
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/framework/details/eager_deletion_op_handle.h"
#include "paddle/fluid/framework/lod_tensor_array.h"
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/framework/selected_rows.h"
#ifdef PADDLE_WITH_CUDA
#include "paddle/fluid/platform/cuda_device_guard.h"
#endif
namespace paddle {
namespace framework {
namespace details {
EagerDeletionOpHandle::EagerDeletionOpHandle(
ir::Node *node, const Scope *scope, const platform::Place &place,
const std::unordered_set<std::string> &var_names, GarbageCollector *gc,
AtomicReferenceCountMap *ref_cnts)
: OpHandleBase(node),
scope_(scope),
var_names_(var_names),
gc_(gc),
ref_cnts_(ref_cnts) {
#ifdef PADDLE_WITH_CUDA
if (platform::is_gpu_place(place)) {
dev_ctx_ = reinterpret_cast<platform::CUDADeviceContext *>(
platform::DeviceContextPool::Instance().Get(place));
if (dynamic_cast<StreamGarbageCollector *>(gc_)) {
platform::CUDADeviceGuard guard(
boost::get<platform::CUDAPlace>(place).device);
PADDLE_ENFORCE(cudaEventCreateWithFlags(&event_, cudaEventDisableTiming));
PADDLE_ENFORCE_NOT_NULL(event_);
}
}
#endif
}
EagerDeletionOpHandle::~EagerDeletionOpHandle() {
#ifdef PADDLE_WITH_CUDA
if (event_) {
auto gpu_place = boost::get<platform::CUDAPlace>(dev_ctx_->GetPlace());
platform::CUDADeviceGuard guard(gpu_place.device);
PADDLE_ENFORCE(cudaEventDestroy(event_));
}
#endif
}
std::string EagerDeletionOpHandle::Name() const { return "eager_deletion"; }
void EagerDeletionOpHandle::RunImpl() {
auto *exec_scope = scope_->FindVar(kLocalExecScopeName)->Get<Scope *>();
std::deque<std::shared_ptr<memory::Allocation>> garbages;
for (auto &name : var_names_) {
auto it = ref_cnts_->find(name);
// Var not found, not reference count has not decreased to 0
if (it == ref_cnts_->end() || it->second.fetch_sub(1) != 1) {
continue;
}
auto *var = exec_scope->FindVar(name);
if (var == nullptr) {
continue;
}
VLOG(2) << "Erase variable " << name;
if (var->IsType<LoDTensor>()) {
garbages.emplace_back(var->GetMutable<LoDTensor>()->MoveMemoryHolder());
} else if (var->IsType<SelectedRows>()) {
garbages.emplace_back(
var->GetMutable<SelectedRows>()->mutable_value()->MoveMemoryHolder());
} else if (var->IsType<LoDTensorArray>()) {
auto *tensor_arr = var->GetMutable<LoDTensorArray>();
for (auto &t : *tensor_arr) {
garbages.emplace_back(t.MoveMemoryHolder());
}
} else {
PADDLE_THROW("Type %s of %s is not supported eager deletion",
var->Type().name(), name);
}
}
if (!garbages.empty()) {
ClearGarbages(&garbages);
}
}
void EagerDeletionOpHandle::ClearGarbages(
std::deque<std::shared_ptr<memory::Allocation>> *garbages) {
#ifdef PADDLE_WITH_CUDA
if (event_) {
auto compute_stream = dev_ctx_->stream();
auto callback_stream =
reinterpret_cast<StreamGarbageCollector *>(gc_)->stream();
auto callback_func = [=]() {
PADDLE_ENFORCE(cudaEventRecord(event_, compute_stream));
PADDLE_ENFORCE(cudaStreamWaitEvent(callback_stream, event_, 0));
};
gc_->Add(std::move(*garbages), callback_func);
} else {
#endif
gc_->Add(std::move(*garbages));
#ifdef PADDLE_WITH_CUDA
}
#endif
}
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/eager_deletion_op_handle.h 0 → 100644
浏览文件 @ 893957f7
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#pragma once
#include <deque>
#include <string>
#include "paddle/fluid/framework/details/op_handle_base.h"
#include "paddle/fluid/framework/details/reference_count_pass_helper.h"
namespace paddle {
namespace framework {
class Scope;
namespace details {
class EagerDeletionOpHandle : public OpHandleBase {
public:
EagerDeletionOpHandle(ir::Node *node, const Scope *scope,
const platform::Place &place,
const std::unordered_set<std::string> &var_names,
GarbageCollector *gc,
AtomicReferenceCountMap *ref_cnts);
~EagerDeletionOpHandle();
std::string Name() const override;
protected:
void RunImpl() override;
private:
void ClearGarbages(std::deque<std::shared_ptr<memory::Allocation>> *garbages);
const Scope *scope_;
std::unordered_set<std::string> var_names_;
GarbageCollector *gc_; // not own
AtomicReferenceCountMap *ref_cnts_; // not own
#ifdef PADDLE_WITH_CUDA
platform::CUDADeviceContext *dev_ctx_{nullptr};
cudaEvent_t event_{nullptr};
#endif
};
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/eager_deletion_pass.cc 0 → 100644
浏览文件 @ 893957f7
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include <queue>
#include <string>
#include <vector>
#include "paddle/fluid/framework/details/computation_op_handle.h"
#include "paddle/fluid/framework/details/eager_deletion_op_handle.h"
#include "paddle/fluid/framework/details/eager_deletion_pass.h"
#include "paddle/fluid/framework/details/multi_devices_helper.h"
#include "paddle/fluid/framework/ir/graph_helper.h"
namespace paddle {
namespace framework {
namespace details {
std::unique_ptr<ir::Graph> EagerDeletionPass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
auto &ref_cnts =
Get<std::vector<AtomicReferenceCountMap>>(kRuntimeReferenceCount);
PADDLE_ENFORCE(ref_cnts.empty(),
"kRuntimeReferenceCount should be initialized here!");
const auto &vars = graph->Get<GraphVars>(kGraphVars);
ref_cnts.resize(vars.size());
const auto &last_live_ops =
Get<std::vector<LastLiveOpsOfVars>>(kLastLiveOpsOfVars);
const auto &gcs = Get<GarbageCollectorMap>(kGarbageCollector);
const auto &places = Get<std::vector<platform::Place>>(kAllPlaces);
// a reverse map of last_live_ops
// i.e., last op --> variable names which can be deleted.
std::unordered_map<ComputationOpHandle *, std::unordered_set<std::string>>
op_vars_map;
for (auto &var_ops_map : last_live_ops) {
for (auto &var_ops_pair : var_ops_map) {
const std::string &var_name = var_ops_pair.first;
for (auto *op : var_ops_pair.second) {
op_vars_map[op].insert(var_name);
}
}
}
for (auto &pair : op_vars_map) {
auto *op = pair.first;
auto &var_names = pair.second;
auto *eager_deletion_node =
graph->CreateEmptyNode("eager_deletion", ir::Node::Type::kOperation);
auto *eager_deletion_op = new EagerDeletionOpHandle(
eager_deletion_node, op->GetScope(), op->GetPlace(), var_names,
gcs.at(places[op->GetScopeIdx()]).get(),
&(ref_cnts[op->GetScopeIdx()]));
auto it = std::find_if(
op->Outputs().begin(), op->Outputs().end(), [](VarHandleBase *var) {
return dynamic_cast<DummyVarHandle *>(var) != nullptr;
});
if (it != op->Outputs().end()) {
eager_deletion_op->AddInput(*it);
} else {
auto *dep_var = new DummyVarHandle(graph->CreateControlDepVar());
graph->Get<GraphDepVars>(kGraphDepVars).emplace(dep_var);
op->AddOutput(dep_var);
eager_deletion_op->AddInput(dep_var);
}
auto *dummy_leaf = new DummyVarHandle(graph->CreateControlDepVar());
graph->Get<GraphDepVars>(kGraphDepVars).emplace(dummy_leaf);
eager_deletion_op->AddOutput(dummy_leaf);
}
VLOG(10) << "Create " << op_vars_map.size() << " EagerDeletionOpHandle(s)";
return graph;
}
} // namespace details
} // namespace framework
} // namespace paddle
REGISTER_PASS(eager_deletion_pass,
paddle::framework::details::EagerDeletionPass)
.RequirePassAttr(paddle::framework::details::kRuntimeReferenceCount)
.RequirePassAttr(paddle::framework::details::kLastLiveOpsOfVars)
.RequirePassAttr(paddle::framework::details::kAllPlaces)
.RequirePassAttr(paddle::framework::details::kGarbageCollector);
paddle/fluid/framework/details/eager_deletion_pass.h 0 → 100644
浏览文件 @ 893957f7
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#pragma once
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/pass.h"
namespace paddle {
namespace framework {
namespace details {
class EagerDeletionPass : public ir::Pass {
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
};
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/reference_count_op_handle.h → paddle/fluid/framework/details/early_delete_op_handle.h
浏览文件 @ 893957f7
......@@ -13,14 +13,13 @@
// limitations under the License.
#pragma once
#include <atomic>
#include <string>
#include <unordered_map>
#include <vector>
#include "paddle/fluid/framework/details/computation_op_handle.h"
#include "paddle/fluid/framework/details/op_handle_base.h"
#include "paddle/fluid/framework/details/var_handle.h"
#include "paddle/fluid/framework/garbage_collector.h"
#include "paddle/fluid/framework/lod_tensor_array.h"
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/framework/selected_rows.h"
#include "paddle/fluid/framework/tensor.h"
......@@ -29,85 +28,85 @@ namespace paddle {
namespace framework {
namespace details {
using ReferenceCountMap = std::unordered_map<std::string, int>;
using AtomicReferenceCountMap =
std::unordered_map<std::string, std::atomic<int>>;
using DeviceReferenceCountMap =
std::unordered_map<int, std::unique_ptr<ReferenceCountMap>>;
using AtomicDeviceReferenceCountMap =
std::unordered_map<int, std::unique_ptr<AtomicReferenceCountMap>>;
using DeviceGarbageCollectorMap =
std::unordered_map<int,
std::unique_ptr<GarbageCollector<framework::Tensor>>>;
class ReferenceCountOpHandle : public OpHandleBase {
class EarlyDeleteOpHandle : public OpHandleBase {
public:
ReferenceCountOpHandle(ir::Node *node, const Scope *scope,
const platform::CUDAPlace &place,
const std::vector<std::string> &var_names,
GarbageCollector<Tensor> *gc,
AtomicReferenceCountMap *ref_cnts)
: OpHandleBase(node), scope_(scope), gc_(gc), ref_cnts_(ref_cnts) {
dev_ctx_ = static_cast<platform::CUDADeviceContext *>(
platform::DeviceContextPool::Instance().Get(place));
EarlyDeleteOpHandle(ir::Node* node, const Scope* scope,
const platform::Place& place,
const std::vector<std::string>& names,
GarbageCollector* gc)
: OpHandleBase(node),
scope_(scope),
place_(place),
names_(names),
gc_(gc) {
#ifdef PADDLE_WITH_CUDA
if (IsStreamGarabageCollector()) {
platform::SetDeviceId(place.device);
auto gpu_place = boost::get<platform::CUDAPlace>(place);
PADDLE_ENFORCE(cudaSetDevice(gpu_place.device));
PADDLE_ENFORCE(cudaEventCreateWithFlags(&event_, cudaEventDisableTiming));
}
for (auto &name : var_names) AddVar(name);
#endif
}
~ReferenceCountOpHandle() {
~EarlyDeleteOpHandle() {
#ifdef PADDLE_WITH_CUDA
if (IsStreamGarabageCollector()) {
auto gpu_place = boost::get<platform::CUDAPlace>(dev_ctx_->GetPlace());
platform::SetDeviceId(gpu_place.device);
PADDLE_ENFORCE(cudaSetDevice(gpu_place.device));
PADDLE_ENFORCE(cudaEventDestroy(event_));
}
#endif
}
std::string Name() const override { return "reference_count"; }
void AddVar(const std::string &name) {
auto it = var_names_.find(name);
if (it != var_names_.end())
++(it->second);
else
var_names_[name] = 1;
}
std::string Name() const override { return "early_delete"; }
protected:
void RunImpl() override {
auto *exec_scope = scope_->FindVar(kLocalExecScopeName)->Get<Scope *>();
std::vector<Tensor *> tensors;
for (auto &pair : var_names_) {
auto &name = pair.first;
auto it = ref_cnts_->find(name);
if (it == ref_cnts_->end()) continue;
auto *var = exec_scope->FindVar(name);
if (var == nullptr) continue;
std::vector<std::shared_ptr<memory::Allocation>> tensors;
auto* local_scope = scope_->FindVar(kLocalExecScopeName)->Get<Scope*>();
for (auto& var_name : names_) {
auto* var = local_scope->FindVar(var_name);
PADDLE_ENFORCE(var != nullptr,
string::Sprintf("Local Scope not has var %s", var_name));
if (var->IsType<LoDTensor>()) {
if (it->second.fetch_sub(pair.second) <= pair.second) {
tensors.emplace_back(var->GetMutable<LoDTensor>());
}
tensors.emplace_back(var->GetMutable<LoDTensor>()->MoveMemoryHolder());
} else if (var->IsType<SelectedRows>()) {
if (it->second.fetch_sub(pair.second) <= pair.second) {
tensors.emplace_back(
var->GetMutable<SelectedRows>()->mutable_value());
tensors.emplace_back(var->GetMutable<SelectedRows>()
->mutable_value()
->MoveMemoryHolder());
} else if (var->IsType<LoDTensorArray>()) {
LoDTensorArray* tensor_array = var->GetMutable<LoDTensorArray>();
for (auto& tensor : *tensor_array) {
tensors.emplace_back(tensor.MoveMemoryHolder());
}
}
}
if (!tensors.empty()) {
ClearTensors(tensors);
}
}
private:
void ClearTensors(const std::vector<Tensor *> &tensors) {
auto *gc = dynamic_cast<StreamGarbageCollector<Tensor> *>(gc_);
void ClearTensors(
const std::vector<std::shared_ptr<memory::Allocation>>& tensors) {
if (platform::is_cpu_place(place_)) {
ClearCPUTensors(tensors);
} else {
ClearGPUTensors(tensors);
}
}
void ClearCPUTensors(
const std::vector<std::shared_ptr<memory::Allocation>>& tensors) {
auto* gc = dynamic_cast<CPUGarbageCollector*>(gc_);
if (gc != nullptr) {
gc->Add(tensors);
}
}
void ClearGPUTensors(
const std::vector<std::shared_ptr<memory::Allocation>>& tensors) {
#ifdef PADDLE_WITH_CUDA
auto* gc = dynamic_cast<StreamGarbageCollector*>(gc_);
if (gc != nullptr) {
auto compute_stream = dev_ctx_->stream();
auto callback_stream = gc->stream();
......@@ -122,15 +121,18 @@ class ReferenceCountOpHandle : public OpHandleBase {
}
bool IsStreamGarabageCollector() const {
return dynamic_cast<const StreamGarbageCollector<Tensor> *>(gc_) != nullptr;
return dynamic_cast<const StreamGarbageCollector*>(gc_) != nullptr;
#endif
}
const Scope *scope_;
platform::CUDADeviceContext *dev_ctx_;
std::unordered_map<std::string, int> var_names_;
GarbageCollector<Tensor> *gc_; // not own
AtomicReferenceCountMap *ref_cnts_; // not own
const Scope* scope_;
const platform::Place place_;
std::vector<std::string> names_;
GarbageCollector* gc_;
#ifdef PADDLE_WITH_CUDA
platform::CUDADeviceContext* dev_ctx_;
cudaEvent_t event_;
#endif
};
} // namespace details
......
paddle/fluid/framework/details/fuse_vars_op_handle.h
浏览文件 @ 893957f7
......@@ -33,7 +33,7 @@ struct FuseVarsOpHandle : public OpHandleBase {
FuseVarsOpHandle(ir::Node *node, Scope *local_scope,
const platform::Place &place,
const std::unordered_map<std::string, int64_t> &inputs_numel,
const std::type_index &var_type)
const proto::VarType::Type var_type)
: OpHandleBase(node),
local_scope_(local_scope),
place_(place),
......@@ -57,7 +57,7 @@ struct FuseVarsOpHandle : public OpHandleBase {
Scope *local_scope_;
const platform::Place place_;
const std::unordered_map<std::string, int64_t> inputs_numel_;
const std::type_index type_;
const proto::VarType::Type type_;
int64_t total_numel_;
};
} // namespace details
......
paddle/fluid/framework/details/memory_early_delete_pass.cc 0 → 100644
浏览文件 @ 893957f7
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/framework/details/memory_early_delete_pass.h"
#include <queue>
#include <string>
#include <vector>
#include "paddle/fluid/framework/details/memory_reuse_types.h"
#include "paddle/fluid/framework/details/multi_devices_helper.h"
#include "paddle/fluid/framework/details/reference_count_pass_helper.h"
#include "paddle/fluid/framework/ir/graph_helper.h"
namespace paddle {
namespace framework {
namespace details {
static ComputationOpHandle* FindNextComputationOpHandle(VarHandle* var_in) {
std::queue<VarHandleBase*> queue;
queue.push(var_in);
do {
auto* var = queue.front();
queue.pop();
for (auto* op : var->PendingOps()) {
auto* compute_op = dynamic_cast<ComputationOpHandle*>(op);
if (compute_op != nullptr && compute_op->GetPlace() == var_in->place_) {
return compute_op;
}
for (auto* out_var : op->Outputs()) {
queue.push(out_var);
}
}
} while (!queue.empty());
return nullptr;
}
std::unique_ptr<ir::Graph> MemoryEarlyDeletePass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
auto& graph_pool = Get<GraphNodePool>(kGraphNodePool);
auto& gcs = Get<GarbageCollectorMap>(kGarbageCollector);
std::unordered_map<std::string, std::unordered_set<OpDesc*>> unlived_vars;
unlived_vars.reserve(graph_pool.size());
for (auto& pair : graph_pool) {
unlived_vars.insert(std::make_pair(pair.first, pair.second));
}
auto compare_and_insert_early_delete_op = [&](
OpHandleBase* op, const std::vector<VarHandleBase*>& vars) {
if (unlived_vars.empty()) return;
// unlived vars can be deleted after the last used op has finished.
auto* compute_op = dynamic_cast<ComputationOpHandle*>(op);
const auto& places = Get<std::vector<platform::Place>>(kAllPlaces);
for (auto& var : vars) {
auto* var_handle = dynamic_cast<VarHandle*>(var);
auto var_name = var->Node()->Name();
auto& var_place = var_handle->place_;
if (unlived_vars.count(var_name) == 0) continue;
if (!unlived_vars[var_name].empty()) {
if (compute_op != nullptr &&
unlived_vars[var_name].count(compute_op->Node()->Op()) != 0) {
unlived_vars[var_name].erase(compute_op->Node()->Op());
}
continue;
}
if (var_handle == nullptr || !var_handle->Node()->IsVar() ||
var_handle->Node()->IsCtrlVar())
continue;
// shameless copyed from reference count pass.
if (compute_op == nullptr) {
// use next computation op scope
compute_op = FindNextComputationOpHandle(var_handle);
}
auto* early_delete_node =
graph->CreateEmptyNode("early_delete", ir::Node::Type::kOperation);
GarbageCollector* gc = gcs.at(places[compute_op->GetScopeIdx()]).get();
auto* early_delete_handle = new EarlyDeleteOpHandle(
early_delete_node, compute_op->GetScope(), var_place, {var_name}, gc);
if (compute_op->Outputs().empty()) {
auto* dep_var = new DummyVarHandle(graph->CreateControlDepVar());
compute_op->AddOutput(dep_var);
graph->Get<GraphDepVars>(kGraphDepVars).emplace(dep_var);
}
early_delete_handle->AddInput(compute_op->Outputs().front());
VLOG(5) << "Add early delete op " << var_name << " to Operator"
<< compute_op->Name();
}
};
auto all_ops = ir::FilterByNodeWrapper<OpHandleBase>(*graph);
for (auto& op : all_ops) {
compare_and_insert_early_delete_op(op, op->Inputs());
compare_and_insert_early_delete_op(op, op->Outputs());
}
return graph;
}
} // namespace details
} // namespace framework
} // namespace paddle
REGISTER_PASS(memory_early_delete_pass,
paddle::framework::details::MemoryEarlyDeletePass)
.RequireGraphAttr(paddle::framework::details::kGraphNodePool)
.RequireGraphAttr(paddle::framework::details::kGarbageCollector);
paddle/fluid/framework/details/memory_early_delete_pass.h 0 → 100644
浏览文件 @ 893957f7
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#pragma once
#include "paddle/fluid/framework/details/early_delete_op_handle.h"
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/pass.h"
namespace paddle {
namespace framework {
namespace details {
class MemoryEarlyDeletePass : public ir::Pass {
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
};
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/memory_reuse_types.cc 0 → 100644
浏览文件 @ 893957f7
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/framework/details/memory_reuse_types.h"
#include <iostream>
#include <sstream>
#include <string>
namespace paddle {
namespace framework {
namespace details {
size_t NodeSizeInBytes(ir::Node* n) {
auto* desc = FindVarDescInBlock(n);
auto shape = desc->GetShape();
size_t type_size = SizeOfType(desc->GetDataType());
int size = 1;
for (auto& s : shape) {
size *= s;
}
return type_size * std::abs(size);
}
std::string DebugStringImpl(VarDesc* var) {
std::stringstream ss;
ss << var->Name();
ss << "[";
try {
auto shape = var->GetShape();
for (size_t i = 0; i < shape.size(); ++i) {
if (i != shape.size() - 1) {
ss << shape[i] << ",";
} else {
ss << shape[i];
}
}
ss << "]";
} catch (...) {
ss << "Var has no VarDesc !!! Name:" << var->Name();
}
return ss.str();
}
std::string DebugString(ir::Node* var) {
return DebugStringImpl(FindVarDescInBlock(var));
}
// return DebugString(var->Var()); }
// NOTE(dzh): based ir node, if a large node has been reused
// by a small size node, then next time it appear in pool, it will
// have the small size. Find the original node shap from blockdesc.
VarDesc* FindVarDescInBlock(ir::Node* n) {
PADDLE_ENFORCE(n->IsVar() && !n->IsCtrlVar() && n->inputs.size() == 1);
BlockDesc* block = n->inputs[0]->Op()->Block();
PADDLE_ENFORCE(block->HasVar(n->Name()),
string::Sprintf("Block do not has var %s", n->Name()));
return block->FindVar(n->Name());
}
struct NodeComparator {
bool operator()(ir::Node* lhs, ir::Node* rhs) const {
auto* lhs_desc = FindVarDescInBlock(lhs);
auto* rhs_desc = FindVarDescInBlock(rhs);
auto lhs_shape = lhs_desc->GetShape();
auto rhs_shape = rhs_desc->GetShape();
if ((lhs_shape[0] == -1 && rhs_shape[0] == -1) ||
(lhs_shape[0] != -1 && rhs_shape[0] != -1)) {
return NodeSizeInBytes(lhs) <= NodeSizeInBytes(rhs);
} else {
return false;
}
}
};
void OrderedNodePairPool::Insert(ir::Node* var, ir::Node* op) {
PADDLE_ENFORCE(var->IsVar() && !var->IsCtrlVar());
PADDLE_ENFORCE(op->IsOp());
if (mark_table_.count(var->Name()) != 0) {
mark_table_[var->Name()]->second.insert(op);
return;
}
auto* var_desc = FindVarDescInBlock(var);
auto var_shape = var_desc->GetShape();
int batch_size = static_cast<int>(var_shape[0]);
NodeComparator compare_node;
Iter it = nodes_.begin();
while (it != nodes_.end()) {
auto* cache_desc = FindVarDescInBlock(it->first);
int cache_batch_size = cache_desc->GetShape()[0];
if ((cache_batch_size == -1 && batch_size == -1) ||
(cache_batch_size != -1 && batch_size != -1)) {
if (compare_node(it->first, var)) {
++it;
} else {
break;
}
} else if (cache_batch_size == -1 && batch_size != -1) {
++it;
} else if (cache_batch_size != -1 && batch_size == -1) {
break;
}
}
it =
nodes_.insert(it, std::make_pair(var, std::unordered_set<ir::Node*>{op}));
mark_table_[var->Name()] = it;
}
int OrderedNodePairPool::GetIndex(ir::Node* var) {
return std::distance(nodes_.begin(), mark_table_[var->Name()]);
}
ir::Node* OrderedNodePairPool::NodeMatch(ir::Node* var) const {
ir::Node* found_node = nullptr;
NodeComparator compare_node;
for (auto it = nodes_.begin(); it != nodes_.end(); ++it) {
if (compare_node(var, it->first)) {
found_node = it->first;
break;
}
}
return found_node;
}
void OrderedNodePairPool::Erase(ir::Node* var) {
PADDLE_ENFORCE(mark_table_.count(var->Name()));
nodes_.erase(mark_table_[var->Name()]);
mark_table_.erase(var->Name());
}
std::string OrderedNodePairPool::ToString() const {
std::stringstream ss;
for (auto it = nodes_.begin(); it != nodes_.end(); ++it) {
ss << DebugString(it->first) << " ";
}
return ss.str();
}
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/memory_reuse_types.h 0 → 100644
浏览文件 @ 893957f7
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#pragma once
#include <algorithm>
#include <iostream>
#include <iterator>
#include <list>
#include <string>
#include <utility>
#include <vector>
#include "paddle/fluid/framework/data_type.h"
#include "paddle/fluid/framework/ir/graph.h"
namespace paddle {
namespace framework {
namespace details {
constexpr char kFetchedVars[] = "fetched_vars";
constexpr char kGraphNodePool[] = "graph_node_pool";
// NOTE(dzh): Variable and the operators use the var.
// for early delete pass.
// Because analysis var pass build base on ir::Node, which maybe released
// or modified between passes, so we use OpDesc* to mark ops.
using GraphNodePool = std::vector<
std::pair<std::string /*var node*/, std::unordered_set<OpDesc*> /* ops */>>;
// NOTE(dzh): by default, it sort node in ascend order(by node bytes size).
// in fluid, -1 means the batch_size is determined in runtime.
// the node batch_size equal -1 always ranking in the front than the node not.
// For example,
// node0[-1, 1] node1[-1, 1, 1], node2[1,1], node3[1,1024], ..
// O(1) insert, delete
class OrderedNodePairPool {
public:
using NodePair = std::pair<ir::Node*, std::unordered_set<ir::Node*>>;
using Iter = typename std::list<NodePair>::iterator;
using ConstIter = typename std::list<NodePair>::const_iterator;
void Insert(ir::Node* var, ir::Node* op);
void Erase(ir::Node* var);
bool Has(ir::Node* var) { return mark_table_.count(var->Name()); }
ir::Node* NodeMatch(ir::Node* var) const;
// map store non-const iterator, can not promise const
int GetIndex(ir::Node* var);
// pool all node to string
std::string ToString() const;
Iter begin() { return nodes_.begin(); }
Iter end() { return nodes_.end(); }
ConstIter begin() const { return nodes_.begin(); }
ConstIter end() const { return nodes_.end(); }
size_t size() const { return nodes_.size(); }
private:
// for searching.
std::unordered_map<std::string, Iter> mark_table_;
// node swap pairs. var -> ops dep var
std::list<NodePair> nodes_;
};
// node memory size in bytes
size_t NodeSizeInBytes(ir::Node* n);
std::string DebugString(ir::Node* var);
// std::string DebugString(VarDesc* var);
VarDesc* FindVarDescInBlock(ir::Node* n);
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/memory_reuse_types_test.cc 0 → 100644
浏览文件 @ 893957f7
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/framework/details/memory_reuse_types.h"
#include <algorithm>
#include <iostream>
#include <memory>
#include <sstream>
#include <string>
#include <utility>
#include <vector>
#include "glog/logging.h"
#include "gtest/gtest.h"
namespace paddle {
namespace framework {
namespace details {
TEST(OrderedNodePairPool, Normal) {
OrderedNodePairPool pool;
std::vector<std::unique_ptr<ir::Node>> nodes;
// clang-format off
std::vector<std::vector<int64_t>> shapes = {{-1, 10},
{-1, 20},
{1, 2},
{5, 2},
{10, 20},
{-1, 2, 5},
{-1, 1, 5},
{-1, 1}};
// clang-format on
const int COUNT = shapes.size();
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);
for (int i = 0; i < COUNT; ++i) {
auto desc = block_desc->Var(std::to_string(i));
desc->SetShape(shapes[i]);
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(), op.get());
}
// assert its order and interface.
std::cout << pool.ToString() << std::endl;
pool.Erase(nodes.front().get());
std::cout << pool.ToString() << std::endl;
ASSERT_EQ(pool.size(), static_cast<size_t>(COUNT - 1));
ASSERT_EQ(pool.GetIndex(nodes.back().get()), 0);
{
auto v1 = block_desc->Var("11");
v1->SetShape({-1, 256, 56, 56});
std::unique_ptr<ir::Node> node1 = ir::CreateNodeForTest(v1);
node1->inputs.emplace_back(op.get());
auto* cache = pool.NodeMatch(node1.get());
ASSERT_EQ(cache, nullptr);
}
{
auto v2 = block_desc->Var("12");
v2->SetShape({-1, 2, 5});
std::unique_ptr<ir::Node> node1 = ir::CreateNodeForTest(v2);
node1->inputs.emplace_back(op.get());
auto* cache = pool.NodeMatch(node1.get());
ASSERT_EQ(pool.GetIndex(cache), 2); // match 6:[-1,2,5]
}
{
auto v3 = block_desc->Var("13");
v3->SetShape({2, 5});
std::unique_ptr<ir::Node> node1 = ir::CreateNodeForTest(v3);
node1->inputs.emplace_back(op.get());
auto* cache = pool.NodeMatch(node1.get());
ASSERT_EQ(pool.GetIndex(cache), 5); // match 4:[5,2]
}
}
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/multi_devices_graph_pass.cc
浏览文件 @ 893957f7
......@@ -565,7 +565,7 @@ void MultiDevSSAGraphBuilder::CreateComputationalOp(ir::Graph *result,
int dev_id) const {
result->Get<GraphOps>(kGraphOps).emplace_back(
new ComputationOpHandle(result->CreateOpNode(node->Op()),
local_scopes_[dev_id], places_[dev_id]));
local_scopes_[dev_id], places_[dev_id], dev_id));
CreateOpHandleIOs(result, node, dev_id);
}
......@@ -688,8 +688,8 @@ void MultiDevSSAGraphBuilder::CreateComputationalOps(ir::Graph *result,
for (size_t scope_idx = 0; scope_idx < num_places; ++scope_idx) {
auto p = places_[scope_idx];
auto s = local_scopes_[scope_idx];
result->Get<GraphOps>(kGraphOps).emplace_back(
new ComputationOpHandle(result->CreateOpNode(node->Op()), s, p));
result->Get<GraphOps>(kGraphOps).emplace_back(new ComputationOpHandle(
result->CreateOpNode(node->Op()), s, p, scope_idx));
CreateOpHandleIOs(result, node, scope_idx);
}
}
......
paddle/fluid/framework/details/multi_devices_graph_print_pass.cc
浏览文件 @ 893957f7
......@@ -85,4 +85,5 @@ void GraphvizSSAGraphPrinter::Print(const ir::Graph &graph,
} // namespace paddle
REGISTER_PASS(multi_devices_print_pass,
paddle::framework::details::SSAGraghBuilderWithPrinter);
paddle::framework::details::SSAGraghBuilderWithPrinter)
.RequirePassAttr(paddle::framework::details::kGraphvizPath);
paddle/fluid/framework/details/multi_devices_graph_print_pass.h
浏览文件 @ 893957f7
......@@ -14,6 +14,7 @@
#pragma once
#include <glog/logging.h>
#include <fstream>
#include <iosfwd>
#include <ostream>
......@@ -24,6 +25,8 @@ namespace paddle {
namespace framework {
namespace details {
constexpr char kGraphvizPath[] = "debug_graphviz_path";
class SSAGraphPrinter {
public:
virtual ~SSAGraphPrinter() {}
......@@ -40,7 +43,7 @@ class SSAGraghBuilderWithPrinter : public ir::Pass {
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override {
std::unique_ptr<std::ostream> fout(
new std::ofstream(Get<const std::string>("debug_graphviz_path")));
new std::ofstream(Get<std::string>(kGraphvizPath)));
PADDLE_ENFORCE(fout->good());
Get<GraphvizSSAGraphPrinter>("graph_printer").Print(*graph, *fout);
return graph;
......
paddle/fluid/framework/details/op_graph_view.cc
浏览文件 @ 893957f7
......@@ -23,6 +23,8 @@ namespace details {
OpGraphView::OpGraphView(const std::vector<OpHandleBase *> &ops) { Build(ops); }
void OpGraphView::Build(const std::vector<OpHandleBase *> &ops) {
preceding_ops_.clear();
pending_ops_.clear();
for (auto &op : ops) {
preceding_ops_[op];
pending_ops_[op];
......@@ -40,6 +42,7 @@ void OpGraphView::Build(const std::vector<OpHandleBase *> &ops) {
std::unordered_set<OpHandleBase *> OpGraphView::AllOps() const {
std::unordered_set<OpHandleBase *> ret;
ret.reserve(preceding_ops_.size());
for (auto &pair : preceding_ops_) {
ret.insert(pair.first);
}
......
paddle/fluid/framework/details/op_graph_view.h
浏览文件 @ 893957f7
......@@ -14,7 +14,7 @@
#pragma once
#include <memory>
#include <queue>
#include <unordered_map>
#include <unordered_set>
#include <vector>
......@@ -34,6 +34,11 @@ class OpGraphView {
bool HasOp(OpHandleBase *op) const;
// Use a visitor to visit all pending ops of op
// Stop when callback returns false
template <typename Callback>
bool VisitAllPendingOps(OpHandleBase *op, Callback &&callback) const;
private:
void Build(const std::vector<OpHandleBase *> &ops);
void EnforceHasOp(OpHandleBase *op) const;
......@@ -44,6 +49,28 @@ class OpGraphView {
pending_ops_;
};
template <typename Callback>
bool OpGraphView::VisitAllPendingOps(OpHandleBase *op,
Callback &&callback) const {
EnforceHasOp(op);
std::unordered_set<OpHandleBase *> visited;
std::queue<OpHandleBase *> q;
q.push(op);
do {
op = q.front();
q.pop();
for (auto &pending_op : pending_ops_.at(op)) {
if (visited.count(pending_op) == 0) {
visited.insert(pending_op);
if (!callback(pending_op)) {
return false;
}
}
}
} while (!q.empty());
return true;
}
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/op_handle_base.h
浏览文件 @ 893957f7
......@@ -25,7 +25,7 @@ namespace paddle {
namespace framework {
namespace details {
constexpr char kLocalExecScopeName[] = "@LCOAL_SCOPE@";
constexpr char kLocalExecScopeName[] = "@LOCAL_SCOPE@";
// Wraps ir::Node and provide helper utilities.
// It's responsible for populating necessary fields of ir::Node.
......
paddle/fluid/framework/details/reduce_op_handle.cc
浏览文件 @ 893957f7
......@@ -218,18 +218,18 @@ void ReduceOpHandle::RunImpl() {
}
#if defined PADDLE_WITH_CUDA && defined PADDLE_WITH_DISTRIBUTE
if (framework::IsType<const float>(in_selected_rows[0]->value().type())) {
if (in_selected_rows[0]->value().type() ==
framework::proto::VarType::FP32) {
GatherSelectedRows<platform::CUDADeviceContext, float>(
in_selected_rows, in_places, dev_ctxes_, out_var_handle, t_out_p,
out_var->GetMutable<framework::SelectedRows>());
} else if (framework::IsType<const double>(
in_selected_rows[0]->value().type())) {
} else if (in_selected_rows[0]->value().type() ==
framework::proto::VarType::FP64) {
GatherSelectedRows<platform::CUDADeviceContext, double>(
in_selected_rows, in_places, dev_ctxes_, out_var_handle, t_out_p,
out_var->GetMutable<framework::SelectedRows>());
} else {
PADDLE_ENFORCE(false,
"only support double or float when gahter SelectedRows");
PADDLE_THROW("only support double or float when gather SelectedRows");
}
#endif
});
......@@ -246,7 +246,7 @@ void ReduceOpHandle::RunImpl() {
if (!FLAGS_cpu_deterministic) {
ReduceLoDTensor func(lod_tensors,
out_var->GetMutable<framework::LoDTensor>());
VisitDataType(ToDataType(lod_tensors[0]->type()), func);
VisitDataType(lod_tensors[0]->type(), func);
} else {
// We sum lod_tensors to reduce_sum_trg which is in local_scopes_0
// here, but it doesn't mean reduce_sum_trg must be in local_scopes_0.
......@@ -256,7 +256,7 @@ void ReduceOpHandle::RunImpl() {
->FindVar(out_var_handle->name_)
->GetMutable<framework::LoDTensor>();
ReduceLoDTensor func(lod_tensors, &reduce_sum_trg);
VisitDataType(ToDataType(lod_tensors[0]->type()), func);
VisitDataType(lod_tensors[0]->type(), func);
auto trg = out_var->GetMutable<framework::LoDTensor>();
if (reduce_sum_trg.data<void>() != trg->data<void>()) {
......
paddle/fluid/framework/details/reference_count_pass.cc
浏览文件 @ 893957f7
......@@ -14,187 +14,240 @@
#include <queue>
#include <string>
#include <type_traits>
#include <vector>
#include "paddle/fluid/framework/details/computation_op_handle.h"
#include "paddle/fluid/framework/details/eager_deletion_op_handle.h"
#include "paddle/fluid/framework/details/multi_devices_helper.h"
#include "paddle/fluid/framework/details/op_graph_view.h"
#include "paddle/fluid/framework/details/reference_count_pass.h"
#include "paddle/fluid/framework/details/reference_count_pass_helper.h"
#include "paddle/fluid/framework/ir/graph_helper.h"
namespace paddle {
namespace framework {
namespace details {
static ComputationOpHandle *FindNextComputationOpHandle(VarHandle *var_in) {
std::queue<VarHandleBase *> queue;
queue.push(var_in);
do {
auto *var = queue.front();
queue.pop();
for (auto *op : var->PendingOps()) {
auto *compute_op = dynamic_cast<ComputationOpHandle *>(op);
if (compute_op != nullptr && compute_op->GetPlace() == var_in->place_) {
return compute_op;
// A functor to shrink/remove operators who depend on other operators in a set
class ShrinkDepsOpFunctor {
private:
enum RelationShip { kSame = 0, kNoDeps = 1, kBefore = 2, kAfter = 3 };
public:
explicit ShrinkDepsOpFunctor(const std::vector<OpHandleBase *> &all_ops)
: graph_(all_ops) {}
template <typename OpSet>
OpSet operator()(const OpSet &op_set) const {
using KeyType = typename OpSet::key_type;
static_assert(
std::is_base_of<OpHandleBase,
typename std::remove_pointer<KeyType>::type>::value,
"Key type of OpSet must be OpHandleBase, or derived of OpHandleBase");
if (op_set.size() <= 1) return op_set;
std::vector<OpHandleBase *> ops(op_set.begin(), op_set.end());
OpSet ret;
auto rels = GetRelations(ops);
auto not_before = [](RelationShip r) { return r != kBefore; };
for (size_t i = 0; i < rels.size(); ++i) {
if (std::all_of(rels[i].begin(), rels[i].end(), not_before)) {
ret.emplace(static_cast<KeyType>(ops[i]));
}
for (auto *out_var : op->Outputs()) {
queue.push(out_var);
}
return ret;
}
private:
std::vector<std::vector<RelationShip>> GetRelations(
const std::vector<OpHandleBase *> &ops) const {
std::unordered_map<OpHandleBase *, size_t> op_to_idx;
for (size_t i = 0; i < ops.size(); ++i) {
PADDLE_ENFORCE(graph_.HasOp(ops[i]), "Op does not exist in graph");
op_to_idx[ops[i]] = i;
}
PADDLE_ENFORCE(op_to_idx.size() == ops.size(), "Duplicate ops");
std::vector<std::vector<RelationShip>> ret(ops.size());
for (auto &e : ret) {
e.assign(ops.size(), kSame);
}
size_t found_num = ops.size();
size_t total_num = ops.size() * ops.size();
auto visitor = [&](OpHandleBase *op, size_t i) {
auto it = op_to_idx.find(op);
if (it != op_to_idx.end()) {
size_t j = it->second;
if (i != j && ret[i][j] == kSame) {
ret[i][j] = kBefore;
ret[j][i] = kAfter;
found_num += 2;
if (found_num == total_num) {
return false;
}
}
}
return true;
};
for (size_t i = 0; i < ops.size(); ++i) {
auto sub_visitor = [&, i](OpHandleBase *op) { return visitor(op, i); };
if (!graph_.VisitAllPendingOps(ops[i], sub_visitor)) {
break;
}
}
for (size_t i = 0; i < ops.size(); ++i) {
for (size_t j = i + 1; j < ops.size(); ++j) {
if (ret[i][j] != kSame) continue;
ret[i][j] = kNoDeps;
ret[j][i] = kNoDeps;
}
}
return ret;
}
const OpGraphView graph_;
};
/**
* Find the nearest downstream computation op handle. If the op is a
* computation op, just return itself.
*/
static ComputationOpHandle *FindNextComputationOpHandleOrReturnItself(
OpHandleBase *op, size_t scope_idx) {
std::queue<OpHandleBase *> q;
std::unordered_set<OpHandleBase *> visited;
q.push(op);
do {
auto *op = q.front();
q.pop();
auto *compute_op = dynamic_cast<ComputationOpHandle *>(op);
if (compute_op != nullptr && compute_op->GetScopeIdx() == scope_idx) {
return compute_op;
}
for (auto *out_var : op->Outputs()) {
for (auto *pending_op : out_var->PendingOps()) {
if (visited.count(pending_op)) continue;
visited.insert(pending_op);
}
}
} while (!queue.empty());
} while (!q.empty());
return nullptr;
}
static void AddDependencyBetween(OpHandleBase *in, OpHandleBase *out,
ir::Graph *graph) {
auto it = std::find_if(
in->Outputs().begin(), in->Outputs().end(), [](VarHandleBase *var) {
return dynamic_cast<DummyVarHandle *>(var) != nullptr;
});
if (it != in->Outputs().end()) {
out->AddInput(*it);
} else {
auto *dep_var = new DummyVarHandle(graph->CreateControlDepVar());
graph->Get<GraphDepVars>(kGraphDepVars).emplace(dep_var);
in->AddOutput(dep_var);
out->AddInput(dep_var);
static std::unordered_set<ComputationOpHandle *>
ExtractComputationOpFromLastLivedVar(VarHandle *var, size_t scope_idx,
const ShrinkDepsOpFunctor &shrink_func,
bool *ok) {
// stage one. Get last op for variable.
std::unordered_set<OpHandleBase *> candidates;
{
if (var->PendingOps().empty() && var->GeneratedOp()) {
// No operator depends on this variable. So the last operator is the op
// who generates this variable.
candidates.emplace(var->GeneratedOp());
} else {
candidates = var->PendingOps();
}
// No pending ops or generated op is nullptr
if (candidates.empty()) {
*ok = false;
return {};
}
}
// stage two. Try to cast them to computation op.
// return (*ok=false) when failed.
//
// The reason why we cannot make any types of op handle to be the last lived
// op is:
// some op handle may operate on many DeviceContext, however, our garbage
// collector can only wait one DeviceContext for now. So currently, we wait
// the nearest compute op.
std::unordered_set<ComputationOpHandle *> computation_op;
{
for (auto *op : candidates) {
auto *compute_op =
FindNextComputationOpHandleOrReturnItself(op, scope_idx);
if (compute_op == nullptr) {
*ok = false;
return {};
}
computation_op.emplace(compute_op);
}
}
// stage three. Try to shrink computation op if they depend on each other.
// Get the smallest set of the most ops.
*ok = true;
return shrink_func(computation_op);
}
static VarDesc *TryGetLatestVarDesc(const std::vector<VarHandle *> &vars) {
VarDesc *var_desc = nullptr;
std::find_if(vars.rbegin(), vars.rend(), [&](VarHandle *var_handle) -> bool {
var_desc = var_handle->Node()->Var();
return var_desc != nullptr;
});
return var_desc;
}
std::unique_ptr<ir::Graph> ReferenceCountPass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
auto &ref_cnts = Get<DeviceReferenceCountMap>(kGlobalReferenceCount);
auto &cur_ref_cnts = Get<AtomicDeviceReferenceCountMap>(kCurReferenceCount);
auto &gcs = Get<DeviceGarbageCollectorMap>(kGarbageCollector);
// It is not easy to find the right reference counts of varaibles in graph
// Step 1: Find all variables in computation ops
// Step 2: Find all variables in non-computation ops which refers to variables
// in computation ops
std::unordered_set<std::string> names;
std::unordered_map<OpHandleBase *, ReferenceCountOpHandle *>
compute_ref_cnt_map;
auto get_ref_cnts_from_compute_op = [&](
OpHandleBase *op, const std::vector<VarHandleBase *> &vars) {
std::vector<std::string> var_names_in_op;
auto *compute_op = dynamic_cast<ComputationOpHandle *>(op);
if (compute_op == nullptr ||
!platform::is_gpu_place(compute_op->GetPlace()))
return var_names_in_op;
auto place = boost::get<platform::CUDAPlace>(compute_op->GetPlace());
for (VarHandleBase *var_handle_base : vars) {
auto *var_handle = dynamic_cast<VarHandle *>(var_handle_base);
if (var_handle == nullptr || !var_handle->Node()->IsVar()) continue;
if (!platform::is_gpu_place(var_handle->place_) ||
boost::get<platform::CUDAPlace>(var_handle->place_) != place)
continue;
auto &ref_cnts = Get<std::vector<ReferenceCountMap>>(kGlobalReferenceCount);
auto &last_live_ops_of_vars =
Get<std::vector<LastLiveOpsOfVars>>(kLastLiveOpsOfVars);
PADDLE_ENFORCE(last_live_ops_of_vars.empty() && ref_cnts.empty(),
"Last Live Ops and Reference Counts of vars should be "
"initialized at here.");
VarDesc *var_desc = var_handle->Node()->Var();
auto var_name = var_handle->Node()->Name();
const auto &vars = graph->Get<GraphVars>(kGraphVars);
// This is weird but there is really some variables without var_desc
// in computation_op
if (var_desc == nullptr) {
var_desc = compute_op->Node()->Op()->Block()->FindVar(var_name);
if (var_desc == nullptr) continue;
last_live_ops_of_vars.resize(vars.size());
ref_cnts.resize(vars.size());
ShrinkDepsOpFunctor shrink_func(
ir::FilterByNodeWrapper<OpHandleBase>(*graph));
for (size_t i = 0; i < vars.size(); ++i) {
for (auto &name_var_pair : vars[i]) {
// Whether this variable can be reused or deleted? If not, we do not
// compute reference counts and dependencies.
VarDesc *var_desc = TryGetLatestVarDesc(name_var_pair.second);
if (var_desc == nullptr || var_desc->Persistable()) {
continue;
}
if (var_desc->Persistable()) continue;
auto var_type = var_desc->Proto()->type().type();
if (var_type != proto::VarType::LOD_TENSOR &&
var_type != proto::VarType::SELECTED_ROWS) {
var_type != proto::VarType::SELECTED_ROWS &&
var_type != proto::VarType::LOD_TENSOR_ARRAY) {
// Var type cannot be deleted
continue;
}
// compute op only runs in one device
if (ref_cnts[place.device]->count(var_name))
++(*ref_cnts[place.device])[var_name];
else
(*ref_cnts[place.device])[var_name] = 1;
bool ok;
auto result = ExtractComputationOpFromLastLivedVar(
name_var_pair.second.back(), i, shrink_func, &ok);
names.insert(var_name);
var_names_in_op.push_back(var_name);
}
return var_names_in_op;
};
auto update_ref_cnts_from_non_compute_op = [&](
OpHandleBase *op, const std::vector<VarHandleBase *> &vars) {
if (dynamic_cast<ComputationOpHandle *>(op) != nullptr) return;
for (VarHandleBase *var_handle_base : vars) {
auto *var_handle = dynamic_cast<VarHandle *>(var_handle_base);
if (var_handle == nullptr || !var_handle->Node()->IsVar()) continue;
auto var_name = var_handle->Node()->Name();
auto var_place = var_handle->place_;
if (!platform::is_gpu_place(var_place)) continue;
auto place = boost::get<platform::CUDAPlace>(var_place);
if (names.count(var_name) == 0) continue;
if (ref_cnts.count(place.device) &&
ref_cnts[place.device]->count(var_name)) {
++(*ref_cnts[place.device])[var_name];
auto *next_compute_op = FindNextComputationOpHandle(var_handle);
if (next_compute_op != nullptr) {
if (compute_ref_cnt_map.count(next_compute_op)) {
compute_ref_cnt_map[next_compute_op]->AddVar(var_name);
VLOG(5) << "Add reference count of " << var_name << " to Operator "
<< next_compute_op->Name();
} else {
// Create new reference_count_op_handle
ir::Node *ref_cnt_node = graph->CreateEmptyNode(
"reference_count", ir::Node::Type::kOperation);
auto *ref_cnt_handle = new ReferenceCountOpHandle(
ref_cnt_node, next_compute_op->GetScope(), place, {var_name},
gcs[place.device].get(), cur_ref_cnts[place.device].get());
AddDependencyBetween(next_compute_op, ref_cnt_handle, graph.get());
compute_ref_cnt_map[next_compute_op] = ref_cnt_handle;
}
}
if (ok) {
auto &var_name = name_var_pair.first;
PADDLE_ENFORCE(!result.empty(), "Last living ops of %s cannot be empty",
var_name);
ref_cnts[i].emplace(var_name, result.size());
last_live_ops_of_vars[i].emplace(var_name, std::move(result));
}
}
};
auto all_ops = ir::FilterByNodeWrapper<OpHandleBase>(*graph);
for (auto &op : all_ops) {
auto in_var_names = get_ref_cnts_from_compute_op(op, op->Inputs());
auto out_var_names = get_ref_cnts_from_compute_op(op, op->Outputs());
if (in_var_names.empty() && out_var_names.empty()) continue;
in_var_names.insert(in_var_names.end(), out_var_names.begin(),
out_var_names.end());
auto *compute_op = dynamic_cast<ComputationOpHandle *>(op);
auto place = boost::get<platform::CUDAPlace>(compute_op->GetPlace());
ir::Node *ref_cnt_node =
graph->CreateEmptyNode("reference_count", ir::Node::Type::kOperation);
auto *ref_cnt_handle = new ReferenceCountOpHandle(
ref_cnt_node, compute_op->GetScope(), place, in_var_names,
gcs[place.device].get(), cur_ref_cnts[place.device].get());
AddDependencyBetween(compute_op, ref_cnt_handle, graph.get());
compute_ref_cnt_map[compute_op] = ref_cnt_handle;
}
for (auto &op : all_ops) {
update_ref_cnts_from_non_compute_op(op, op->Inputs());
update_ref_cnts_from_non_compute_op(op, op->Outputs());
}
std::vector<OpHandleBase *> new_all_ops;
new_all_ops.reserve(compute_ref_cnt_map.size() + all_ops.size());
for (auto &op : all_ops) {
new_all_ops.emplace_back(std::move(op));
auto it = compute_ref_cnt_map.find(new_all_ops.back());
if (it != compute_ref_cnt_map.end()) {
// Add LeafNode to ReferenceCountOpHandle
auto *dummy_leaf = new DummyVarHandle(graph->CreateControlDepVar());
graph->Get<GraphDepVars>(kGraphDepVars).emplace(dummy_leaf);
it->second->AddOutput(dummy_leaf);
new_all_ops.emplace_back(std::move(it->second));
}
}
all_ops.swap(new_all_ops);
return graph;
}
......@@ -205,5 +258,4 @@ std::unique_ptr<ir::Graph> ReferenceCountPass::ApplyImpl(
REGISTER_PASS(reference_count_pass,
paddle::framework::details::ReferenceCountPass)
.RequirePassAttr(paddle::framework::details::kGlobalReferenceCount)
.RequirePassAttr(paddle::framework::details::kCurReferenceCount)
.RequirePassAttr(paddle::framework::details::kGarbageCollector);
.RequirePassAttr(paddle::framework::details::kLastLiveOpsOfVars);
paddle/fluid/framework/details/reference_count_pass.h
浏览文件 @ 893957f7
......@@ -14,7 +14,6 @@
#pragma once
#include "paddle/fluid/framework/details/reference_count_op_handle.h"
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/pass.h"
......@@ -22,10 +21,6 @@ namespace paddle {
namespace framework {
namespace details {
constexpr char kGlobalReferenceCount[] = "reference_count";
constexpr char kCurReferenceCount[] = "current_reference_count";
constexpr char kGarbageCollector[] = "garbage_collector";
class ReferenceCountPass : public ir::Pass {
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
......
paddle/fluid/framework/details/reference_count_pass_helper.cc 0 → 100644
浏览文件 @ 893957f7
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/framework/details/reference_count_pass_helper.h"
namespace paddle {
namespace framework {
namespace details {} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/reference_count_pass_helper.h 0 → 100644
浏览文件 @ 893957f7
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#pragma once
#include <atomic>
#include <map>
#include <string>
#include <unordered_map>
#include <unordered_set>
#include <vector>
#include "paddle/fluid/framework/garbage_collector.h"
namespace paddle {
namespace framework {
namespace details {
class ComputationOpHandle;
using ReferenceCountMap = std::unordered_map<std::string, size_t>;
using AtomicReferenceCountMap =
std::unordered_map<std::string, std::atomic<size_t>>;
using GarbageCollectorMap =
std::map<platform::Place, std::unique_ptr<GarbageCollector>>;
const char kGlobalReferenceCount[] = "global_reference_count";
const char kRuntimeReferenceCount[] = "runtime_reference_count";
const char kGarbageCollector[] = "garbage_collector";
const char kAllPlaces[] = "all_places";
using LastLiveOpsOfVars =
std::unordered_map<std::string, std::unordered_set<ComputationOpHandle*>>;
const char kLastLiveOpsOfVars[] = "last_live_ops_of_var";
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.cc
浏览文件 @ 893957f7
......@@ -18,9 +18,6 @@
#include <vector>
#include "paddle/fluid/framework/variable_helper.h"
#include "paddle/fluid/platform/profiler.h"
#ifdef PADDLE_WITH_CUDA
#include "paddle/fluid/framework/details/reference_count_op_handle.h"
#endif
namespace paddle {
namespace framework {
......@@ -69,27 +66,12 @@ FeedFetchList ScopeBufferedSSAGraphExecutor::Run(
platform::RecordEvent e("ScopeBufferedSSAGraphExecutorAfterRun", nullptr);
drop_scope_counter_ += 1;
#ifdef PADDLE_WITH_CUDA
const std::string gc_name = "garbage_collector";
DeviceGarbageCollectorMap *gc =
Graph().Has(gc_name) ? &(Graph().Get<DeviceGarbageCollectorMap>(gc_name))
: nullptr;
#endif
if (!fetch_tensors.empty() ||
drop_scope_counter_ == strategy_.num_iteration_per_drop_scope_) {
drop_scope_counter_ = 0;
// Wait All computational streams
for (auto p : places_) {
platform::DeviceContextPool::Instance().Get(p)->Wait();
#ifdef PADDLE_WITH_CUDA
if (gc != nullptr && platform::is_gpu_place(p)) {
auto gpu_place = boost::get<platform::CUDAPlace>(p);
auto &gc_at_place = gc->at(gpu_place.device);
gc_at_place->Wait();
gc_at_place->Reset();
}
#endif
}
for (auto &scope : local_scopes_) {
auto &local_scope =
......
paddle/fluid/framework/dlpack_tensor.cc
浏览文件 @ 893957f7
......@@ -13,7 +13,7 @@
// limitations under the License.
#include "paddle/fluid/framework/dlpack_tensor.h"
#include "paddle/fluid/framework/data_type.h"
namespace paddle {
namespace framework {
......@@ -36,26 +36,23 @@ static ::DLDataType GetDLDataTypeCode() {
return dtype;
}
static DLDataType GetDLDataTypeFromTypeIndex(const std::type_index &type) {
#define REG_DL_DATA_TYPE(type) \
{ std::type_index(typeid(type)), GetDLDataTypeCode<type>() }
static const std::unordered_map<std::type_index, ::DLDataType>
type_to_dtype_map({
REG_DL_DATA_TYPE(platform::float16), // NOLINT
REG_DL_DATA_TYPE(float), // NOLINT
REG_DL_DATA_TYPE(double), // NOLINT
REG_DL_DATA_TYPE(int), // NOLINT
REG_DL_DATA_TYPE(int64_t), // NOLINT
REG_DL_DATA_TYPE(bool), // NOLINT
REG_DL_DATA_TYPE(size_t), // NOLINT
REG_DL_DATA_TYPE(int16_t), // NOLINT
REG_DL_DATA_TYPE(uint8_t), // NOLINT
REG_DL_DATA_TYPE(int8_t) // NOLINT
});
static std::unordered_map<int, ::DLDataType> CreateDLDataTypeMap() {
static std::unordered_map<int, ::DLDataType> result;
#define REG_DL_DATA_TYPE(cpp_type, proto_type) \
result[static_cast<int>(proto_type)] = GetDLDataTypeCode<cpp_type>()
_ForEachDataType_(REG_DL_DATA_TYPE);
#undef REG_DL_DATA_TYPE
return result;
}
static DLDataType GetDLDataTypeFromTypeIndex(proto::VarType::Type type) {
static auto type_to_dtype_map = CreateDLDataTypeMap();
static auto type_to_dtype_map_end_it = type_to_dtype_map.end();
auto it = type_to_dtype_map.find(type);
PADDLE_ENFORCE(it != type_to_dtype_map_end_it, "Unsupported data type %s",
type.name());
auto it = type_to_dtype_map.find(static_cast<int>(type));
PADDLE_ENFORCE(it != type_to_dtype_map_end_it, "Unsupported data type %d",
type);
return it->second;
#undef REG_DL_DATA_TYPE
}
......
paddle/fluid/framework/dlpack_tensor_test.cc
浏览文件 @ 893957f7
......@@ -91,23 +91,11 @@ void TestMainLoop() {
}
}
}
TEST(dlpack, test_all) {
#define TestCallback(cpp_type, proto_type) TestMainLoop<cpp_type>()
#define PADDLE_DLPACK_TEST(type) \
TEST(dlpack, test_##type) { TestMainLoop<type>(); }
using float16 = platform::float16;
PADDLE_DLPACK_TEST(float16);
PADDLE_DLPACK_TEST(float);
PADDLE_DLPACK_TEST(double);
PADDLE_DLPACK_TEST(int);
PADDLE_DLPACK_TEST(int64_t);
PADDLE_DLPACK_TEST(bool);
PADDLE_DLPACK_TEST(size_t);
PADDLE_DLPACK_TEST(int16_t);
PADDLE_DLPACK_TEST(uint8_t);
PADDLE_DLPACK_TEST(int8_t);
#undef PADDLE_DLPACK_TEST
_ForEachDataType_(TestCallback);
}
} // namespace framework
} // namespace paddle
paddle/fluid/framework/executor.cc
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/framework/executor.h
浏览文件 @ 893957f7
......@@ -27,52 +27,21 @@ limitations under the License. */
namespace paddle {
namespace framework {
template <typename T>
std::unordered_map<std::string, T> GetNonPersistableReferenceCount(
const ProgramDesc& prog, size_t block_id) {
auto& block = prog.Block(block_id);
std::unordered_map<std::string, T> ref_cnts;
auto update_ref_cnts = [&](OpDesc* op_desc, const VariableNameMap& name_map) {
for (auto& name_pair : name_map) {
for (auto& name : name_pair.second) {
auto* var_desc = block.FindVar(name);
if (var_desc == nullptr || var_desc->Persistable()) continue;
auto type = var_desc->Proto()->type().type();
if (type != proto::VarType::LOD_TENSOR &&
type != proto::VarType::SELECTED_ROWS) {
continue;
}
auto it = ref_cnts.find(name);
if (it != ref_cnts.end()) {
++it->second;
} else {
ref_cnts[name] = 1;
}
}
}
};
for (auto op_desc : block.AllOps()) {
update_ref_cnts(op_desc, op_desc->Inputs());
update_ref_cnts(op_desc, op_desc->Outputs());
}
return ref_cnts;
}
struct ExecutorPrepareContext {
ExecutorPrepareContext(const framework::ProgramDesc& prog, size_t block_id);
ExecutorPrepareContext(const framework::ProgramDesc& prog, size_t block_id,
const std::vector<std::string>& skip_ref_cnt_vars =
std::vector<std::string>());
~ExecutorPrepareContext();
void ResetReferenceCount() { cur_ref_cnts_ = ref_cnts_; }
void ResetReferenceCount() { runtime_ref_cnts_ = global_ref_cnts_; }
const framework::ProgramDesc& prog_;
size_t block_id_;
std::vector<std::unique_ptr<OperatorBase>> ops_;
std::unordered_map<std::string, int> ref_cnts_;
std::unordered_map<std::string, int> cur_ref_cnts_;
std::unordered_map<std::string, size_t> global_ref_cnts_;
std::unordered_map<std::string, size_t> runtime_ref_cnts_;
};
class Executor {
......@@ -108,10 +77,14 @@ class Executor {
const std::string& fetch_holder_name = "fetch");
static std::unique_ptr<ExecutorPrepareContext> Prepare(
const ProgramDesc& program, int block_id);
const ProgramDesc& program, int block_id,
const std::vector<std::string>& skip_ref_cnt_vars =
std::vector<std::string>());
static std::vector<std::shared_ptr<ExecutorPrepareContext>> Prepare(
const ProgramDesc& program, const std::vector<int>& block_ids);
const ProgramDesc& program, const std::vector<int>& block_ids,
const std::vector<std::vector<std::string>>& skip_ref_cnt_vars =
std::vector<std::vector<std::string>>());
void CreateVariables(const ProgramDesc& pdesc, Scope* scope, int block_id);
......
paddle/fluid/framework/garbage_collector.cc 0 → 100644
浏览文件 @ 893957f7
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include <algorithm>
#ifdef PADDLE_WITH_CUDA
#include "paddle/fluid/platform/cuda_device_guard.h"
#endif
#include "paddle/fluid/framework/garbage_collector.h"
namespace paddle {
namespace framework {
GarbageCollector::GarbageCollector(const platform::Place &place,
size_t max_memory_size)
: max_memory_size_((std::max)(max_memory_size, static_cast<size_t>(1))) {
garbages_.reset(new GarbageQueue());
dev_ctx_ = platform::DeviceContextPool::Instance().Get(place);
}
CPUGarbageCollector::CPUGarbageCollector(const platform::CPUPlace &place,
size_t max_memory_size)
: GarbageCollector(place, max_memory_size) {}
void CPUGarbageCollector::ClearCallback(const std::function<void()> &callback) {
callback();
}
#ifdef PADDLE_WITH_CUDA
UnsafeFastGPUGarbageCollector::UnsafeFastGPUGarbageCollector(
const platform::CUDAPlace &place, size_t max_memory_size)
: GarbageCollector(place, max_memory_size) {}
void UnsafeFastGPUGarbageCollector::ClearCallback(
const std::function<void()> &callback) {
callback();
}
DefaultStreamGarbageCollector::DefaultStreamGarbageCollector(
const platform::CUDAPlace &place, size_t max_memory_size)
: GarbageCollector(place, max_memory_size) {}
void DefaultStreamGarbageCollector::Wait() const {
static_cast<platform::CUDADeviceContext *>(this->dev_ctx_)
->WaitStreamCallback();
}
void DefaultStreamGarbageCollector::ClearCallback(
const std::function<void()> &callback) {
static_cast<platform::CUDADeviceContext *>(this->dev_ctx_)
->AddStreamCallback(callback);
}
StreamGarbageCollector::StreamGarbageCollector(const platform::CUDAPlace &place,
size_t max_memory_size)
: GarbageCollector(place, max_memory_size) {
platform::CUDADeviceGuard guard(place.device);
PADDLE_ENFORCE(cudaStreamCreate(&stream_));
callback_manager_.reset(new platform::StreamCallbackManager(stream_));
}
StreamGarbageCollector::~StreamGarbageCollector() {
auto place = boost::get<platform::CUDAPlace>(this->dev_ctx_->GetPlace());
platform::CUDADeviceGuard guard(place.device);
PADDLE_ENFORCE(cudaStreamSynchronize(stream_));
PADDLE_ENFORCE(cudaStreamDestroy(stream_));
}
cudaStream_t StreamGarbageCollector::stream() const { return stream_; }
void StreamGarbageCollector::Wait() const { callback_manager_->Wait(); }
void StreamGarbageCollector::ClearCallback(
const std::function<void()> &callback) {
callback_manager_->AddCallback(callback);
}
#endif
} // namespace framework
} // namespace paddle
paddle/fluid/framework/ir/CMakeLists.txt
浏览文件 @ 893957f7
......@@ -42,6 +42,9 @@ pass_library(multi_batch_merge_pass base)
pass_library(conv_bn_fuse_pass inference)
pass_library(seqconv_eltadd_relu_fuse_pass inference)
pass_library(is_test_pass base)
pass_library(conv_elementwise_add_act_fuse_pass inference)
pass_library(conv_elementwise_add2_act_fuse_pass inference)
pass_library(conv_elementwise_add_fuse_pass inference)
if(WITH_MKLDNN)
pass_library(mkldnn_placement_pass base)
pass_library(depthwise_conv_mkldnn_pass base)
......
paddle/fluid/framework/ir/conv_elementwise_add2_act_fuse_pass.h 0 → 100644
浏览文件 @ 893957f7
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#pragma once
#include "paddle/fluid/framework/ir/fuse_pass_base.h"
#include "paddle/fluid/framework/ir/graph_pattern_detector.h"
namespace paddle {
namespace framework {
namespace ir {
class ConvElementwiseAdd2ActFusePass : public FusePassBase {
public:
virtual ~ConvElementwiseAdd2ActFusePass() {}
protected:
std::unique_ptr<ir::Graph> ApplyImpl(std::unique_ptr<ir::Graph> graph) const;
};
} // namespace ir
} // namespace framework
} // namespace paddle
paddle/fluid/framework/ir/graph.cc
浏览文件 @ 893957f7
......@@ -162,7 +162,10 @@ void Graph::ResolveHazard(
(*it_new)->inputs.empty() ? nullptr : (*it_new)->inputs[0];
const auto &read_ops = (*it_old)->outputs;
PADDLE_ENFORCE(write_op, "The write_op should not be empty.");
PADDLE_ENFORCE(
write_op,
string::Sprintf("The write_op of var %s should not be empty.",
(*it_new)->Name()));
// Add write after write dependence
ir::Node *upstream_op =
......
paddle/fluid/framework/ir/graph.h
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/framework/ir/graph_helper.h
浏览文件 @ 893957f7
......@@ -24,6 +24,7 @@ limitations under the License. */
namespace paddle {
namespace framework {
namespace ir {
// Test if the graph contains circle.
bool HasCircle(const Graph &graph);
......
paddle/fluid/framework/ir/node.cc
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/framework/ir/node.h
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/framework/ir/pass.h
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/framework/operator.cc
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/framework/scope.cc
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/framework/scope.h
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/framework/tensor.cc
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/framework/tensor.h
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/inference/io.cc
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/operators/concat_mkldnn_op.cc 0 → 100644
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/operators/concat_op.cc
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/operators/conv_op.cc
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/operators/crop_op.cc
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/operators/crop_op.cu
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/operators/distributed/brpc_rdma_pool.cc 0 → 100644
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/operators/distributed/brpc_rdma_pool.h 0 → 100644
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/operators/distributed/brpc_serde_test.cc 0 → 100644
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/operators/expand_op.cu
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/operators/fc_op.cc
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/operators/fill_op.cc
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/operators/gather_op.cc
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/operators/load_op.cc
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/operators/lrn_op.cc
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/operators/lstm_op.cc
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/operators/lstmp_op.cc
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/operators/math/blas.h
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/operators/math/detail/avx_mathfun.h 0 → 100644
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/operators/mean_op.cc
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/operators/mean_op.cu
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/operators/nce_op.cc
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/operators/pad2d_op.cc
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/operators/pad_op.cu
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/operators/pool_op.cc
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/operators/pool_op.h
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/operators/prelu_op.cc
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/operators/print_op.cc
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/operators/psroi_pool_op.cc 0 → 100644
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/operators/psroi_pool_op.cu 0 → 100644
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/operators/psroi_pool_op.h 0 → 100644
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/operators/save_op.cc
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/operators/slice_op.cc
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/operators/spp_op.h
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/operators/sum_op.cc
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/operators/sum_op.cu
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/operators/unpool_op.cc
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/platform/cpu_info.h
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/platform/enforce.h
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/platform/float16.h
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/platform/nccl_helper.h
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/platform/stream_callback_manager.cc 0 → 100644
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/pybind/CMakeLists.txt
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/pybind/pybind.cc
浏览文件 @ 893957f7
此差异已折叠。
paddle/fluid/pybind/tensor_py.h
浏览文件 @ 893957f7
此差异已折叠。
paddle/scripts/paddle_build.sh
浏览文件 @ 893957f7
此差异已折叠。
python/paddle/fluid/__init__.py
浏览文件 @ 893957f7
此差异已折叠。
python/paddle/fluid/average.py
浏览文件 @ 893957f7
此差异已折叠。
python/paddle/fluid/backward.py
浏览文件 @ 893957f7
此差异已折叠。
python/paddle/fluid/distributed/__init__.py 0 → 100644
浏览文件 @ 893957f7
此差异已折叠。
python/paddle/fluid/distributed/downpour.py 0 → 100644
浏览文件 @ 893957f7
此差异已折叠。
python/paddle/fluid/distributed/helper.py 0 → 100644
浏览文件 @ 893957f7
此差异已折叠。
python/paddle/fluid/distributed/node.py 0 → 100644
浏览文件 @ 893957f7
此差异已折叠。
python/paddle/fluid/distributed/ps_instance.py 0 → 100644
浏览文件 @ 893957f7
此差异已折叠。
python/paddle/fluid/distributed/ps_pb2.py 0 → 100644
浏览文件 @ 893957f7
此差异已折叠。
python/paddle/fluid/framework.py
浏览文件 @ 893957f7
此差异已折叠。
python/paddle/fluid/layers/nn.py
浏览文件 @ 893957f7
此差异已折叠。
python/paddle/fluid/optimizer.py
浏览文件 @ 893957f7
此差异已折叠。
python/paddle/fluid/tests/unittests/ngraph/__init__.py 0 → 100644
浏览文件 @ 893957f7
此差异已折叠。
python/setup.py.in
浏览文件 @ 893957f7
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