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
未验证 提交 17ada1d2 编写于 作者: L lujun 提交者： GitHub
浏览文件
操作

Merge pull request #6 from PaddlePaddle/develop 

fetch dev
上级 d8291fa3 6f11f35a
展开全部 隐藏空白更改
内联 并排
Showing with 11077 addition and 3308 deletion
CMakeLists.txt
浏览文件 @ 17ada1d2
...@@ -25,12 +25,18 @@ message(STATUS "CXX compiler: ${CMAKE_CXX_COMPILER}, version: " ...@@ -25,12 +25,18 @@ message(STATUS "CXX compiler: ${CMAKE_CXX_COMPILER}, version: "
message(STATUS "C compiler: ${CMAKE_C_COMPILER}, version: " message(STATUS "C compiler: ${CMAKE_C_COMPILER}, version: "
"${CMAKE_C_COMPILER_ID} ${CMAKE_C_COMPILER_VERSION}") "${CMAKE_C_COMPILER_ID} ${CMAKE_C_COMPILER_VERSION}")
if(WIN32) if(WIN32)
set(CMAKE_SUPPRESS_REGENERATION ON)
set(CMAKE_STATIC_LIBRARY_PREFIX lib) set(CMAKE_STATIC_LIBRARY_PREFIX lib)
add_definitions("/DGOOGLE_GLOG_DLL_DECL=") add_definitions("/DGOOGLE_GLOG_DLL_DECL=")
set(CMAKE_C_FLAGS_DEBUG "${CMAKE_C_FLAGS_DEBUG} /bigobj /MTd") set(CMAKE_C_FLAGS_DEBUG "${CMAKE_C_FLAGS_DEBUG} /bigobj /MTd")
set(CMAKE_C_FLAGS_RELEASE "${CMAKE_C_FLAGS_RELEASE} /bigobj /MT") set(CMAKE_C_FLAGS_RELEASE "${CMAKE_C_FLAGS_RELEASE} /bigobj /MT")
set(CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} /bigobj /MTd") set(CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} /bigobj /MTd")
set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /bigobj /MT") set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /bigobj /MT")
add_compile_options(/wd4068 /wd4129 /wd4244 /wd4267 /wd4297 /wd4530 /wd4577 /wd4819 /wd4838)
set(PADDLE_LINK_FLAGS "/IGNORE:4006 /IGNORE:4098 /IGNORE:4217 /IGNORE:4221")
set(CMAKE_STATIC_LINKER_FLAGS "${CMAKE_STATIC_LINKER_FLAGS} ${PADDLE_LINK_FLAGS}")
set(CMAKE_SHARED_LINKER_FLAGS "${CMAKE_SHARED_LINKER_FLAGS} ${PADDLE_LINK_FLAGS}")
set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} ${PADDLE_LINK_FLAGS}")
endif(WIN32) endif(WIN32)
find_package(CUDA QUIET) find_package(CUDA QUIET)
...@@ -212,7 +218,7 @@ endif() ...@@ -212,7 +218,7 @@ endif()
if (WITH_JEMALLOC) if (WITH_JEMALLOC)
find_package(JeMalloc REQUIRED) find_package(JeMalloc REQUIRED)
include_directories(${JEMALLOC_INCLUDE_DIR}) include_directories(${JEMALLOC_INCLUDE_DIR})
add_definitions(-DWITH_JEMALLOC) add_definitions(-DPADDLE_WITH_JEMALLOC)
endif() endif()
include(generic) # simplify cmake module include(generic) # simplify cmake module
...@@ -276,9 +282,3 @@ add_subdirectory(paddle) ...@@ -276,9 +282,3 @@ add_subdirectory(paddle)
if(WITH_PYTHON) if(WITH_PYTHON)
add_subdirectory(python) add_subdirectory(python)
endif() endif()
if(WITH_DOC)
find_package(Sphinx REQUIRED)
find_python_module(recommonmark REQUIRED)
add_subdirectory(doc)
endif()
Dockerfile
浏览文件 @ 17ada1d2
...@@ -11,12 +11,10 @@ RUN /bin/bash -c 'if [[ -n ${UBUNTU_MIRROR} ]]; then sed -i 's#http://archive.ub ...@@ -11,12 +11,10 @@ RUN /bin/bash -c 'if [[ -n ${UBUNTU_MIRROR} ]]; then sed -i 's#http://archive.ub
# ENV variables # ENV variables
ARG WITH_GPU ARG WITH_GPU
ARG WITH_AVX ARG WITH_AVX
ARG WITH_DOC
ENV WOBOQ OFF ENV WOBOQ OFF
ENV WITH_GPU=${WITH_GPU:-ON} ENV WITH_GPU=${WITH_GPU:-ON}
ENV WITH_AVX=${WITH_AVX:-ON} ENV WITH_AVX=${WITH_AVX:-ON}
ENV WITH_DOC=${WITH_DOC:-OFF}
ENV HOME /root ENV HOME /root
# Add bash enhancements # Add bash enhancements
......
README.md
浏览文件 @ 17ada1d2
# PaddlePaddle # PaddlePaddle
English | [简体中文](./README_cn.md)
[![Build Status](https://travis-ci.org/PaddlePaddle/Paddle.svg?branch=develop)](https://travis-ci.org/PaddlePaddle/Paddle) [![Build Status](https://travis-ci.org/PaddlePaddle/Paddle.svg?branch=develop)](https://travis-ci.org/PaddlePaddle/Paddle)
[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://paddlepaddle.org/documentation/docs/en/1.2/getstarted/index_en.html) [![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://paddlepaddle.org/documentation/docs/en/1.2/getstarted/index_en.html)
...@@ -7,7 +8,6 @@ ...@@ -7,7 +8,6 @@
[![Release](https://img.shields.io/github/release/PaddlePaddle/Paddle.svg)](https://github.com/PaddlePaddle/Paddle/releases) [![Release](https://img.shields.io/github/release/PaddlePaddle/Paddle.svg)](https://github.com/PaddlePaddle/Paddle/releases)
[![License](https://img.shields.io/badge/license-Apache%202-blue.svg)](LICENSE) [![License](https://img.shields.io/badge/license-Apache%202-blue.svg)](LICENSE)
Welcome to the PaddlePaddle GitHub. Welcome to the PaddlePaddle GitHub.
PaddlePaddle (PArallel Distributed Deep LEarning) is an easy-to-use, PaddlePaddle (PArallel Distributed Deep LEarning) is an easy-to-use,
...@@ -18,16 +18,6 @@ learning to many products at Baidu. ...@@ -18,16 +18,6 @@ learning to many products at Baidu.
Our vision is to enable deep learning for everyone via PaddlePaddle. 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. 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) ### Latest PaddlePaddle Release: [Fluid 1.2.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.2)
### Install Latest Stable Release: ### Install Latest Stable Release:
``` ```
...@@ -43,23 +33,6 @@ pip install paddlepaddle-gpu==1.2.0.post85 ...@@ -43,23 +33,6 @@ pip install paddlepaddle-gpu==1.2.0.post85
# For installation on other platform, refer to http://paddlepaddle.org/ # 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 ## Features
- **Flexibility** - **Flexibility**
...@@ -100,38 +73,10 @@ pip install paddlepaddle-gpu==1.2.0.post85 ...@@ -100,38 +73,10 @@ pip install paddlepaddle-gpu==1.2.0.post85
Baidu and it has achieved a significant impact. We hope you can also explore 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. 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 ## Installation
It is recommended to read [this doc](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/install/index_cn.html) on our website. It is recommended to read [this doc](http://paddlepaddle.org/documentation/docs/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 ## Documentation
We provide [English](http://paddlepaddle.org/documentation/docs/en/1.2/getstarted/index_en.html) and We provide [English](http://paddlepaddle.org/documentation/docs/en/1.2/getstarted/index_en.html) and
...@@ -153,37 +98,9 @@ We provide [English](http://paddlepaddle.org/documentation/docs/en/1.2/getstarte ...@@ -153,37 +98,9 @@ We provide [English](http://paddlepaddle.org/documentation/docs/en/1.2/getstarte
We appreciate your contributions! 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 ## Ask Questions
You are welcome to submit questions and bug reports as [Github Issues](https://github.com/PaddlePaddle/Paddle/issues). 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 ## Copyright and License
PaddlePaddle is provided under the [Apache-2.0 license](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
README_cn.md 0 → 100644
浏览文件 @ 17ada1d2
# PaddlePaddle
[English](./README.md) | 简体中文
[![Build Status](https://travis-ci.org/PaddlePaddle/Paddle.svg?branch=develop)](https://travis-ci.org/PaddlePaddle/Paddle)
[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://paddlepaddle.org/documentation/docs/en/1.2/getstarted/index_en.html)
[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/index.html)
[![Release](https://img.shields.io/github/release/PaddlePaddle/Paddle.svg)](https://github.com/PaddlePaddle/Paddle/releases)
[![License](https://img.shields.io/badge/license-Apache%202-blue.svg)](LICENSE)
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)
### 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/
```
## 特性
- **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。
- **与产品相连**
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。
## 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)
## 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!
## 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
## 版权和许可证
PaddlePaddle由[Apache-2.0 license](LICENSE)提供
cmake/FindSphinx.cmake 已删除 100644 → 0
浏览文件 @ d8291fa3
# - This module looks for Sphinx
# Find the Sphinx documentation generator
#
# This modules defines
# SPHINX_EXECUTABLE
# SPHINX_FOUND
find_program(SPHINX_EXECUTABLE
NAMES sphinx-build
PATHS
/usr/bin
/usr/local/bin
/opt/local/bin
DOC "Sphinx documentation generator"
)
if( NOT SPHINX_EXECUTABLE )
set(_Python_VERSIONS
2.7 2.6 2.5 2.4 2.3 2.2 2.1 2.0 1.6 1.5
)
foreach( _version ${_Python_VERSIONS} )
set( _sphinx_NAMES sphinx-build-${_version} )
find_program( SPHINX_EXECUTABLE
NAMES ${_sphinx_NAMES}
PATHS
/usr/bin
/usr/local/bin
/opt/loca/bin
DOC "Sphinx documentation generator"
)
endforeach()
endif()
include(FindPackageHandleStandardArgs)
find_package_handle_standard_args(Sphinx DEFAULT_MSG
SPHINX_EXECUTABLE
)
option( SPHINX_HTML_OUTPUT "Build a single HTML with the whole content." ON )
option( SPHINX_DIRHTML_OUTPUT "Build HTML pages, but with a single directory per document." OFF )
option( SPHINX_HTMLHELP_OUTPUT "Build HTML pages with additional information for building a documentation collection in htmlhelp." OFF )
option( SPHINX_QTHELP_OUTPUT "Build HTML pages with additional information for building a documentation collection in qthelp." OFF )
option( SPHINX_DEVHELP_OUTPUT "Build HTML pages with additional information for building a documentation collection in devhelp." OFF )
option( SPHINX_EPUB_OUTPUT "Build HTML pages with additional information for building a documentation collection in epub." OFF )
option( SPHINX_LATEX_OUTPUT "Build LaTeX sources that can be compiled to a PDF document using pdflatex." OFF )
option( SPHINX_MAN_OUTPUT "Build manual pages in groff format for UNIX systems." OFF )
option( SPHINX_TEXT_OUTPUT "Build plain text files." OFF )
mark_as_advanced(
SPHINX_EXECUTABLE
SPHINX_HTML_OUTPUT
SPHINX_DIRHTML_OUTPUT
SPHINX_HTMLHELP_OUTPUT
SPHINX_QTHELP_OUTPUT
SPHINX_DEVHELP_OUTPUT
SPHINX_EPUB_OUTPUT
SPHINX_LATEX_OUTPUT
SPHINX_MAN_OUTPUT
SPHINX_TEXT_OUTPUT
)
function( Sphinx_add_target target_name builder conf cache source destination )
add_custom_target( ${target_name} ALL
COMMAND ${SPHINX_EXECUTABLE} -b ${builder}
-d ${cache}
-c ${conf}
${source}
${destination}
COMMENT "Generating sphinx documentation: ${builder}"
COMMAND cd ${destination} && ln -sf ./index_*.html index.html
)
set_property(
DIRECTORY APPEND PROPERTY
ADDITIONAL_MAKE_CLEAN_FILES
${destination}
)
endfunction()
# Target dependencies can be optionally listed at the end.
function( Sphinx_add_targets target_base_name conf source base_destination )
set( _dependencies )
foreach( arg IN LISTS ARGN )
set( _dependencies ${_dependencies} ${arg} )
endforeach()
if( ${SPHINX_HTML_OUTPUT} )
Sphinx_add_target( ${target_base_name}_html html ${conf} ${source} ${base_destination}/html )
add_dependencies( ${target_base_name}_html ${_dependencies} )
endif()
if( ${SPHINX_DIRHTML_OUTPUT} )
Sphinx_add_target( ${target_base_name}_dirhtml dirhtml ${conf} ${source} ${base_destination}/dirhtml )
add_dependencies( ${target_base_name}_dirhtml ${_dependencies} )
endif()
if( ${SPHINX_QTHELP_OUTPUT} )
Sphinx_add_target( ${target_base_name}_qthelp qthelp ${conf} ${source} ${base_destination}/qthelp )
add_dependencies( ${target_base_name}_qthelp ${_dependencies} )
endif()
if( ${SPHINX_DEVHELP_OUTPUT} )
Sphinx_add_target( ${target_base_name}_devhelp devhelp ${conf} ${source} ${base_destination}/devhelp )
add_dependencies( ${target_base_name}_devhelp ${_dependencies} )
endif()
if( ${SPHINX_EPUB_OUTPUT} )
Sphinx_add_target( ${target_base_name}_epub epub ${conf} ${source} ${base_destination}/epub )
add_dependencies( ${target_base_name}_epub ${_dependencies} )
endif()
if( ${SPHINX_LATEX_OUTPUT} )
Sphinx_add_target( ${target_base_name}_latex latex ${conf} ${source} ${base_destination}/latex )
add_dependencies( ${target_base_name}_latex ${_dependencies} )
endif()
if( ${SPHINX_MAN_OUTPUT} )
Sphinx_add_target( ${target_base_name}_man man ${conf} ${source} ${base_destination}/man )
add_dependencies( ${target_base_name}_man ${_dependencies} )
endif()
if( ${SPHINX_TEXT_OUTPUT} )
Sphinx_add_target( ${target_base_name}_text text ${conf} ${source} ${base_destination}/text )
add_dependencies( ${target_base_name}_text ${_dependencies} )
endif()
if( ${BUILD_TESTING} )
sphinx_add_target( ${target_base_name}_linkcheck linkcheck ${conf} ${source} ${base_destination}/linkcheck )
add_dependencies( ${target_base_name}_linkcheck ${_dependencies} )
endif()
endfunction()
cmake/configure.cmake
浏览文件 @ 17ada1d2
...@@ -152,7 +152,12 @@ endif() ...@@ -152,7 +152,12 @@ endif()
if (WITH_MKLML AND MKLML_IOMP_LIB) if (WITH_MKLML AND MKLML_IOMP_LIB)
message(STATUS "Enable Intel OpenMP with ${MKLML_IOMP_LIB}") message(STATUS "Enable Intel OpenMP with ${MKLML_IOMP_LIB}")
set(OPENMP_FLAGS "-fopenmp") if(WIN32)
# openmp not support well for now on windows
set(OPENMP_FLAGS "")
else(WIN32)
set(OPENMP_FLAGS "-fopenmp")
endif(WIN32)
set(CMAKE_C_CREATE_SHARED_LIBRARY_FORBIDDEN_FLAGS ${OPENMP_FLAGS}) set(CMAKE_C_CREATE_SHARED_LIBRARY_FORBIDDEN_FLAGS ${OPENMP_FLAGS})
set(CMAKE_CXX_CREATE_SHARED_LIBRARY_FORBIDDEN_FLAGS ${OPENMP_FLAGS}) set(CMAKE_CXX_CREATE_SHARED_LIBRARY_FORBIDDEN_FLAGS ${OPENMP_FLAGS})
set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${OPENMP_FLAGS}") set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${OPENMP_FLAGS}")
......
cmake/cuda.cmake
浏览文件 @ 17ada1d2
...@@ -203,25 +203,26 @@ list(APPEND CUDA_NVCC_FLAGS "-w") ...@@ -203,25 +203,26 @@ list(APPEND CUDA_NVCC_FLAGS "-w")
list(APPEND CUDA_NVCC_FLAGS "--expt-relaxed-constexpr") list(APPEND CUDA_NVCC_FLAGS "--expt-relaxed-constexpr")
if (NOT WIN32) if (NOT WIN32)
if(CMAKE_BUILD_TYPE STREQUAL "Debug") if(CMAKE_BUILD_TYPE STREQUAL "Debug")
list(APPEND CUDA_NVCC_FLAGS ${CMAKE_CXX_FLAGS_DEBUG}) list(APPEND CUDA_NVCC_FLAGS ${CMAKE_CXX_FLAGS_DEBUG})
elseif(CMAKE_BUILD_TYPE STREQUAL "Release") elseif(CMAKE_BUILD_TYPE STREQUAL "Release")
list(APPEND CUDA_NVCC_FLAGS ${CMAKE_CXX_FLAGS_RELEASE}) list(APPEND CUDA_NVCC_FLAGS ${CMAKE_CXX_FLAGS_RELEASE})
elseif(CMAKE_BUILD_TYPE STREQUAL "RelWithDebInfo") elseif(CMAKE_BUILD_TYPE STREQUAL "RelWithDebInfo")
list(APPEND CUDA_NVCC_FLAGS ${CMAKE_CXX_FLAGS_RELWITHDEBINFO}) list(APPEND CUDA_NVCC_FLAGS ${CMAKE_CXX_FLAGS_RELWITHDEBINFO})
elseif(CMAKE_BUILD_TYPE STREQUAL "MinSizeRel") elseif(CMAKE_BUILD_TYPE STREQUAL "MinSizeRel")
# nvcc 9 does not support -Os. Use Release flags instead # nvcc 9 does not support -Os. Use Release flags instead
list(APPEND CUDA_NVCC_FLAGS ${CMAKE_CXX_FLAGS_RELEASE}) list(APPEND CUDA_NVCC_FLAGS ${CMAKE_CXX_FLAGS_RELEASE})
endif() endif()
else(NOT WIN32) else(NOT WIN32)
list(APPEND CUDA_NVCC_FLAGS "--compiler-options;/bigobj") list(APPEND CUDA_NVCC_FLAGS "-Xcompiler \"/wd 4244 /wd 4267 /wd 4819\"")
if(CMAKE_BUILD_TYPE STREQUAL "Debug") list(APPEND CUDA_NVCC_FLAGS "--compiler-options;/bigobj")
list(APPEND CUDA_NVCC_FLAGS "-g -G") if(CMAKE_BUILD_TYPE STREQUAL "Debug")
# match the cl's _ITERATOR_DEBUG_LEVEL list(APPEND CUDA_NVCC_FLAGS "-g -G")
list(APPEND CUDA_NVCC_FLAGS "-D_DEBUG") # match the cl's _ITERATOR_DEBUG_LEVEL
elseif(CMAKE_BUILD_TYPE STREQUAL "Release") list(APPEND CUDA_NVCC_FLAGS "-D_DEBUG")
list(APPEND CUDA_NVCC_FLAGS "-O3 -DNDEBUG") elseif(CMAKE_BUILD_TYPE STREQUAL "Release")
else() list(APPEND CUDA_NVCC_FLAGS "-O3 -DNDEBUG")
else()
message(FATAL "Windows only support Release or Debug build now. Please set visual studio build type to Release/Debug, x64 build.") message(FATAL "Windows only support Release or Debug build now. Please set visual studio build type to Release/Debug, x64 build.")
endif() endif()
endif(NOT WIN32) endif(NOT WIN32)
......
cmake/external/glog.cmake
浏览文件 @ 17ada1d2
...@@ -20,8 +20,10 @@ SET(GLOG_INCLUDE_DIR "${GLOG_INSTALL_DIR}/include" CACHE PATH "glog include dire ...@@ -20,8 +20,10 @@ SET(GLOG_INCLUDE_DIR "${GLOG_INSTALL_DIR}/include" CACHE PATH "glog include dire
IF(WIN32) IF(WIN32)
SET(GLOG_LIBRARIES "${GLOG_INSTALL_DIR}/lib/libglog.lib" CACHE FILEPATH "glog library." FORCE) SET(GLOG_LIBRARIES "${GLOG_INSTALL_DIR}/lib/libglog.lib" CACHE FILEPATH "glog library." FORCE)
SET(GLOG_CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /wd4267 /wd4530")
ELSE(WIN32) ELSE(WIN32)
SET(GLOG_LIBRARIES "${GLOG_INSTALL_DIR}/lib/libglog.a" CACHE FILEPATH "glog library." FORCE) SET(GLOG_LIBRARIES "${GLOG_INSTALL_DIR}/lib/libglog.a" CACHE FILEPATH "glog library." FORCE)
SET(GLOG_CMAKE_CXX_FLAGS ${CMAKE_CXX_FLAGS})
ENDIF(WIN32) ENDIF(WIN32)
INCLUDE_DIRECTORIES(${GLOG_INCLUDE_DIR}) INCLUDE_DIRECTORIES(${GLOG_INCLUDE_DIR})
...@@ -39,7 +41,7 @@ ExternalProject_Add( ...@@ -39,7 +41,7 @@ ExternalProject_Add(
UPDATE_COMMAND "" UPDATE_COMMAND ""
CMAKE_ARGS -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER} CMAKE_ARGS -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
-DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
-DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS} -DCMAKE_CXX_FLAGS=${GLOG_CMAKE_CXX_FLAGS}
-DCMAKE_CXX_FLAGS_RELEASE=${CMAKE_CXX_FLAGS_RELEASE} -DCMAKE_CXX_FLAGS_RELEASE=${CMAKE_CXX_FLAGS_RELEASE}
-DCMAKE_CXX_FLAGS_DEBUG=${CMAKE_CXX_FLAGS_DEBUG} -DCMAKE_CXX_FLAGS_DEBUG=${CMAKE_CXX_FLAGS_DEBUG}
-DCMAKE_C_FLAGS=${CMAKE_C_FLAGS} -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}
......
cmake/external/mkldnn.cmake
浏览文件 @ 17ada1d2
...@@ -49,6 +49,8 @@ IF(NOT WIN32) ...@@ -49,6 +49,8 @@ IF(NOT WIN32)
SET(MKLDNN_FLAG "${MKLDNN_FLAG} -Wno-unused-result -Wno-unused-value") SET(MKLDNN_FLAG "${MKLDNN_FLAG} -Wno-unused-result -Wno-unused-value")
SET(MKLDNN_CFLAG "${CMAKE_C_FLAGS} ${MKLDNN_FLAG}") SET(MKLDNN_CFLAG "${CMAKE_C_FLAGS} ${MKLDNN_FLAG}")
SET(MKLDNN_CXXFLAG "${CMAKE_CXX_FLAGS} ${MKLDNN_FLAG}") SET(MKLDNN_CXXFLAG "${CMAKE_CXX_FLAGS} ${MKLDNN_FLAG}")
ELSE()
SET(MKLDNN_CXXFLAG "${CMAKE_CXX_FLAGS} /EHsc")
ENDIF(NOT WIN32) ENDIF(NOT WIN32)
ExternalProject_Add( ExternalProject_Add(
...@@ -61,7 +63,6 @@ ExternalProject_Add( ...@@ -61,7 +63,6 @@ ExternalProject_Add(
UPDATE_COMMAND "" UPDATE_COMMAND ""
CMAKE_ARGS -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER} CMAKE_ARGS -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
CMAKE_ARGS -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} CMAKE_ARGS -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
CMAKE_ARGS -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
CMAKE_ARGS -DCMAKE_CXX_FLAGS_RELEASE=${CMAKE_CXX_FLAGS_RELEASE} CMAKE_ARGS -DCMAKE_CXX_FLAGS_RELEASE=${CMAKE_CXX_FLAGS_RELEASE}
CMAKE_ARGS -DCMAKE_CXX_FLAGS_DEBUG=${CMAKE_CXX_FLAGS_DEBUG} CMAKE_ARGS -DCMAKE_CXX_FLAGS_DEBUG=${CMAKE_CXX_FLAGS_DEBUG}
CMAKE_ARGS -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS} CMAKE_ARGS -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}
......
cmake/external/snappy.cmake
浏览文件 @ 17ada1d2
...@@ -20,6 +20,12 @@ set(SNAPPY_SOURCES_DIR ${THIRD_PARTY_PATH}/snappy) ...@@ -20,6 +20,12 @@ set(SNAPPY_SOURCES_DIR ${THIRD_PARTY_PATH}/snappy)
set(SNAPPY_INSTALL_DIR ${THIRD_PARTY_PATH}/install/snappy) set(SNAPPY_INSTALL_DIR ${THIRD_PARTY_PATH}/install/snappy)
set(SNAPPY_INCLUDE_DIR "${SNAPPY_INSTALL_DIR}/include" CACHE PATH "snappy include directory." FORCE) set(SNAPPY_INCLUDE_DIR "${SNAPPY_INSTALL_DIR}/include" CACHE PATH "snappy include directory." FORCE)
if(WIN32)
SET(SNAPPY_CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /wd4244 /wd4267")
else()
SET(SNAPPY_CMAKE_CXX_FLAGS ${CMAKE_CXX_FLAGS})
endif()
ExternalProject_Add( ExternalProject_Add(
extern_snappy extern_snappy
GIT_REPOSITORY "https://github.com/google/snappy" GIT_REPOSITORY "https://github.com/google/snappy"
...@@ -31,7 +37,7 @@ ExternalProject_Add( ...@@ -31,7 +37,7 @@ ExternalProject_Add(
-DCMAKE_C_FLAGS=${CMAKE_C_FLAGS} -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}
-DCMAKE_C_FLAGS_DEBUG=${CMAKE_C_FLAGS_DEBUG} -DCMAKE_C_FLAGS_DEBUG=${CMAKE_C_FLAGS_DEBUG}
-DCMAKE_C_FLAGS_RELEASE=${CMAKE_C_FLAGS_RELEASE} -DCMAKE_C_FLAGS_RELEASE=${CMAKE_C_FLAGS_RELEASE}
-DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS} -DCMAKE_CXX_FLAGS=${SNAPPY_CMAKE_CXX_FLAGS}
-DCMAKE_CXX_FLAGS_RELEASE=${CMAKE_CXX_FLAGS_RELEASE} -DCMAKE_CXX_FLAGS_RELEASE=${CMAKE_CXX_FLAGS_RELEASE}
-DCMAKE_CXX_FLAGS_DEBUG=${CMAKE_CXX_FLAGS_DEBUG} -DCMAKE_CXX_FLAGS_DEBUG=${CMAKE_CXX_FLAGS_DEBUG}
-DCMAKE_INSTALL_PREFIX=${SNAPPY_INSTALL_DIR} -DCMAKE_INSTALL_PREFIX=${SNAPPY_INSTALL_DIR}
......
cmake/flags.cmake
浏览文件 @ 17ada1d2
...@@ -21,7 +21,7 @@ function(CheckCompilerCXX11Flag) ...@@ -21,7 +21,7 @@ function(CheckCompilerCXX11Flag)
if (${CMAKE_CXX_COMPILER_VERSION} VERSION_LESS 3.3) if (${CMAKE_CXX_COMPILER_VERSION} VERSION_LESS 3.3)
message(FATAL_ERROR "Unsupported Clang version. Clang >= 3.3 required.") message(FATAL_ERROR "Unsupported Clang version. Clang >= 3.3 required.")
endif() endif()
endif() endif()
endif() endif()
endfunction() endfunction()
...@@ -147,12 +147,7 @@ set(GPU_COMMON_FLAGS ...@@ -147,12 +147,7 @@ set(GPU_COMMON_FLAGS
-Wno-error=unused-function # Warnings in Numpy Header. -Wno-error=unused-function # Warnings in Numpy Header.
-Wno-error=array-bounds # Warnings in Eigen::array -Wno-error=array-bounds # Warnings in Eigen::array
) )
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -m64")
else(NOT WIN32)
set(COMMON_FLAGS
"/w") #disable all warnings.
set(GPU_COMMON_FLAGS
"/w") #disable all warnings
endif(NOT WIN32) endif(NOT WIN32)
if (APPLE) if (APPLE)
...@@ -193,8 +188,7 @@ safe_set_static_flag() ...@@ -193,8 +188,7 @@ safe_set_static_flag()
CMAKE_CXX_FLAGS_MINSIZEREL CMAKE_CXX_FLAGS_RELWITHDEBINFO CMAKE_CXX_FLAGS_MINSIZEREL CMAKE_CXX_FLAGS_RELWITHDEBINFO
CMAKE_C_FLAGS CMAKE_C_FLAGS_DEBUG CMAKE_C_FLAGS_RELEASE CMAKE_C_FLAGS CMAKE_C_FLAGS_DEBUG CMAKE_C_FLAGS_RELEASE
CMAKE_C_FLAGS_MINSIZEREL CMAKE_C_FLAGS_RELWITHDEBINFO) CMAKE_C_FLAGS_MINSIZEREL CMAKE_C_FLAGS_RELWITHDEBINFO)
if(${flag_var} MATCHES "/W3") string(REGEX REPLACE "(^| )/W[0-9]( |$)" " " ${flag_var} "${${flag_var}}")
string(REGEX REPLACE "/W3" "/w" ${flag_var} "${${flag_var}}") set(flag_var "${flag_var} /w")
endif(${flag_var} MATCHES "/W3")
endforeach(flag_var) endforeach(flag_var)
endif(WIN32) endif(WIN32)
cmake/operators.cmake
浏览文件 @ 17ada1d2
...@@ -52,8 +52,8 @@ function(op_library TARGET) ...@@ -52,8 +52,8 @@ function(op_library TARGET)
endif() endif()
if(WITH_MKLDNN) if(WITH_MKLDNN)
string(REPLACE "_op" "_mkldnn_op" MKLDNN_FILE "${TARGET}") string(REPLACE "_op" "_mkldnn_op" MKLDNN_FILE "${TARGET}")
if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${MKLDNN_FILE}.cc) if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/mkldnn/${MKLDNN_FILE}.cc)
list(APPEND mkldnn_cc_srcs ${MKLDNN_FILE}.cc) list(APPEND mkldnn_cc_srcs mkldnn/${MKLDNN_FILE}.cc)
endif() endif()
endif() endif()
else() else()
......
cmake/version.cmake
浏览文件 @ 17ada1d2
...@@ -31,8 +31,23 @@ while ("${PADDLE_VERSION}" STREQUAL "") ...@@ -31,8 +31,23 @@ while ("${PADDLE_VERSION}" STREQUAL "")
set(tmp_version "${GIT_TAG_NAME}~1") set(tmp_version "${GIT_TAG_NAME}~1")
endif() endif()
else() else()
# otherwise, we always set PADDLE_VERSION to 0.0.0 to represent latest execute_process(
set(PADDLE_VERSION "0.0.0") COMMAND ${GIT_EXECUTABLE} describe --exact-match --tags ${tmp_version}
WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}
OUTPUT_VARIABLE GIT_EXACT_TAG_NAME
RESULT_VARIABLE GIT_EXACT_TAG_RESULT
ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)
if (NOT ${GIT_EXACT_TAG_NAME})
# Check if current branch is tag branch
if (${GIT_EXACT_TAG_NAME} MATCHES "v${TAG_VERSION_REGEX}")
string(REPLACE "v" "" PADDLE_VERSION ${GIT_EXACT_TAG_NAME})
else()
set(PADDLE_VERSION "0.0.0")
endif()
else()
# otherwise, we always set PADDLE_VERSION to 0.0.0 to represent latest
set(PADDLE_VERSION "0.0.0")
endif()
endif() endif()
else() else()
set(PADDLE_VERSION "0.0.0") set(PADDLE_VERSION "0.0.0")
......
paddle/fluid/API.spec
浏览文件 @ 17ada1d2
...@@ -8,13 +8,13 @@ paddle.fluid.Program.parse_from_string ArgSpec(args=['binary_str'], varargs=None ...@@ -8,13 +8,13 @@ paddle.fluid.Program.parse_from_string ArgSpec(args=['binary_str'], varargs=None
paddle.fluid.Program.to_string ArgSpec(args=['self', 'throw_on_error', 'with_details'], varargs=None, keywords=None, defaults=(False,)) paddle.fluid.Program.to_string ArgSpec(args=['self', 'throw_on_error', 'with_details'], varargs=None, keywords=None, defaults=(False,))
paddle.fluid.default_startup_program ArgSpec(args=[], varargs=None, keywords=None, defaults=None) paddle.fluid.default_startup_program ArgSpec(args=[], varargs=None, keywords=None, defaults=None)
paddle.fluid.default_main_program ArgSpec(args=[], varargs=None, keywords=None, defaults=None) paddle.fluid.default_main_program ArgSpec(args=[], varargs=None, keywords=None, defaults=None)
paddle.fluid.program_guard ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None) paddle.fluid.program_guard ArgSpec(args=['main_program', 'startup_program'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.name_scope ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None) paddle.fluid.name_scope ArgSpec(args=['prefix'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.Executor.__init__ ArgSpec(args=['self', 'place'], varargs=None, keywords=None, defaults=None) paddle.fluid.Executor.__init__ ArgSpec(args=['self', 'place'], varargs=None, keywords=None, defaults=None)
paddle.fluid.Executor.close ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) paddle.fluid.Executor.close ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.Executor.run ArgSpec(args=['self', 'program', 'feed', 'fetch_list', 'feed_var_name', 'fetch_var_name', 'scope', 'return_numpy', 'use_program_cache'], varargs=None, keywords=None, defaults=(None, None, None, 'feed', 'fetch', None, True, False)) paddle.fluid.Executor.run ArgSpec(args=['self', 'program', 'feed', 'fetch_list', 'feed_var_name', 'fetch_var_name', 'scope', 'return_numpy', 'use_program_cache'], varargs=None, keywords=None, defaults=(None, None, None, 'feed', 'fetch', None, True, False))
paddle.fluid.global_scope ArgSpec(args=[], varargs=None, keywords=None, defaults=None) paddle.fluid.global_scope ArgSpec(args=[], varargs=None, keywords=None, defaults=None)
paddle.fluid.scope_guard ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None) paddle.fluid.scope_guard ArgSpec(args=['scope'], varargs=None, keywords=None, defaults=None)
paddle.fluid.DistributeTranspiler.__init__ ArgSpec(args=['self', 'config'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.DistributeTranspiler.__init__ ArgSpec(args=['self', 'config'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.DistributeTranspiler.get_pserver_program ArgSpec(args=['self', 'endpoint'], varargs=None, keywords=None, defaults=None) paddle.fluid.DistributeTranspiler.get_pserver_program ArgSpec(args=['self', 'endpoint'], varargs=None, keywords=None, defaults=None)
paddle.fluid.DistributeTranspiler.get_pserver_programs ArgSpec(args=['self', 'endpoint'], varargs=None, keywords=None, defaults=None) paddle.fluid.DistributeTranspiler.get_pserver_programs ArgSpec(args=['self', 'endpoint'], varargs=None, keywords=None, defaults=None)
...@@ -66,7 +66,7 @@ paddle.fluid.initializer.XavierInitializer.__init__ ArgSpec(args=['self', 'unifo ...@@ -66,7 +66,7 @@ paddle.fluid.initializer.XavierInitializer.__init__ ArgSpec(args=['self', 'unifo
paddle.fluid.initializer.BilinearInitializer.__init__ ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) paddle.fluid.initializer.BilinearInitializer.__init__ ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.initializer.MSRAInitializer.__init__ ArgSpec(args=['self', 'uniform', 'fan_in', 'seed'], varargs=None, keywords=None, defaults=(True, None, 0)) paddle.fluid.initializer.MSRAInitializer.__init__ ArgSpec(args=['self', 'uniform', 'fan_in', 'seed'], varargs=None, keywords=None, defaults=(True, None, 0))
paddle.fluid.initializer.force_init_on_cpu ArgSpec(args=[], varargs=None, keywords=None, defaults=None) paddle.fluid.initializer.force_init_on_cpu ArgSpec(args=[], varargs=None, keywords=None, defaults=None)
paddle.fluid.initializer.init_on_cpu ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None) paddle.fluid.initializer.init_on_cpu ArgSpec(args=[], varargs=None, keywords=None, defaults=None)
paddle.fluid.initializer.NumpyArrayInitializer.__init__ ArgSpec(args=['self', 'value'], varargs=None, keywords=None, defaults=None) paddle.fluid.initializer.NumpyArrayInitializer.__init__ ArgSpec(args=['self', 'value'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.fc ArgSpec(args=['input', 'size', 'num_flatten_dims', 'param_attr', 'bias_attr', 'act', 'is_test', 'name'], varargs=None, keywords=None, defaults=(1, None, None, None, False, None)) paddle.fluid.layers.fc ArgSpec(args=['input', 'size', 'num_flatten_dims', 'param_attr', 'bias_attr', 'act', 'is_test', 'name'], varargs=None, keywords=None, defaults=(1, None, None, None, False, None))
paddle.fluid.layers.embedding ArgSpec(args=['input', 'size', 'is_sparse', 'is_distributed', 'padding_idx', 'param_attr', 'dtype'], varargs=None, keywords=None, defaults=(False, False, None, None, 'float32')) paddle.fluid.layers.embedding ArgSpec(args=['input', 'size', 'is_sparse', 'is_distributed', 'padding_idx', 'param_attr', 'dtype'], varargs=None, keywords=None, defaults=(False, False, None, None, 'float32'))
...@@ -122,7 +122,7 @@ paddle.fluid.layers.transpose ArgSpec(args=['x', 'perm', 'name'], varargs=None, ...@@ -122,7 +122,7 @@ paddle.fluid.layers.transpose ArgSpec(args=['x', 'perm', 'name'], varargs=None,
paddle.fluid.layers.im2sequence ArgSpec(args=['input', 'filter_size', 'stride', 'padding', 'input_image_size', 'out_stride', 'name'], varargs=None, keywords=None, defaults=(1, 1, 0, None, 1, None)) paddle.fluid.layers.im2sequence ArgSpec(args=['input', 'filter_size', 'stride', 'padding', 'input_image_size', 'out_stride', 'name'], varargs=None, keywords=None, defaults=(1, 1, 0, None, 1, None))
paddle.fluid.layers.nce ArgSpec(args=['input', 'label', 'num_total_classes', 'sample_weight', 'param_attr', 'bias_attr', 'num_neg_samples', 'name', 'sampler', 'custom_dist', 'seed', 'is_sparse'], varargs=None, keywords=None, defaults=(None, None, None, None, None, 'uniform', None, 0, False)) paddle.fluid.layers.nce ArgSpec(args=['input', 'label', 'num_total_classes', 'sample_weight', 'param_attr', 'bias_attr', 'num_neg_samples', 'name', 'sampler', 'custom_dist', 'seed', 'is_sparse'], varargs=None, keywords=None, defaults=(None, None, None, None, None, 'uniform', None, 0, False))
paddle.fluid.layers.hsigmoid ArgSpec(args=['input', 'label', 'num_classes', 'param_attr', 'bias_attr', 'name', 'path_table', 'path_code', 'is_custom', 'is_sparse'], varargs=None, keywords=None, defaults=(None, None, None, None, None, False, False)) paddle.fluid.layers.hsigmoid ArgSpec(args=['input', 'label', 'num_classes', 'param_attr', 'bias_attr', 'name', 'path_table', 'path_code', 'is_custom', 'is_sparse'], varargs=None, keywords=None, defaults=(None, None, None, None, None, False, False))
paddle.fluid.layers.beam_search ArgSpec(args=['pre_ids', 'pre_scores', 'ids', 'scores', 'beam_size', 'end_id', 'level', 'is_accumulated', 'name'], varargs=None, keywords=None, defaults=(0, True, None)) paddle.fluid.layers.beam_search ArgSpec(args=['pre_ids', 'pre_scores', 'ids', 'scores', 'beam_size', 'end_id', 'level', 'is_accumulated', 'name', 'return_parent_idx'], varargs=None, keywords=None, defaults=(0, True, None, False))
paddle.fluid.layers.row_conv ArgSpec(args=['input', 'future_context_size', 'param_attr', 'act'], varargs=None, keywords=None, defaults=(None, None)) paddle.fluid.layers.row_conv ArgSpec(args=['input', 'future_context_size', 'param_attr', 'act'], varargs=None, keywords=None, defaults=(None, None))
paddle.fluid.layers.multiplex ArgSpec(args=['inputs', 'index'], varargs=None, keywords=None, defaults=None) paddle.fluid.layers.multiplex ArgSpec(args=['inputs', 'index'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.layer_norm ArgSpec(args=['input', 'scale', 'shift', 'begin_norm_axis', 'epsilon', 'param_attr', 'bias_attr', 'act', 'name'], varargs=None, keywords=None, defaults=(True, True, 1, 1e-05, None, None, None, None)) paddle.fluid.layers.layer_norm ArgSpec(args=['input', 'scale', 'shift', 'begin_norm_axis', 'epsilon', 'param_attr', 'bias_attr', 'act', 'name'], varargs=None, keywords=None, defaults=(True, True, 1, 1e-05, None, None, None, None))
...@@ -142,10 +142,10 @@ paddle.fluid.layers.label_smooth ArgSpec(args=['label', 'prior_dist', 'epsilon', ...@@ -142,10 +142,10 @@ paddle.fluid.layers.label_smooth ArgSpec(args=['label', 'prior_dist', 'epsilon',
paddle.fluid.layers.roi_pool ArgSpec(args=['input', 'rois', 'pooled_height', 'pooled_width', 'spatial_scale'], varargs=None, keywords=None, defaults=(1, 1, 1.0)) paddle.fluid.layers.roi_pool ArgSpec(args=['input', 'rois', 'pooled_height', 'pooled_width', 'spatial_scale'], varargs=None, keywords=None, defaults=(1, 1, 1.0))
paddle.fluid.layers.roi_align ArgSpec(args=['input', 'rois', 'pooled_height', 'pooled_width', 'spatial_scale', 'sampling_ratio', 'name'], varargs=None, keywords=None, defaults=(1, 1, 1.0, -1, None)) paddle.fluid.layers.roi_align ArgSpec(args=['input', 'rois', 'pooled_height', 'pooled_width', 'spatial_scale', 'sampling_ratio', 'name'], varargs=None, keywords=None, defaults=(1, 1, 1.0, -1, None))
paddle.fluid.layers.dice_loss ArgSpec(args=['input', 'label', 'epsilon'], varargs=None, keywords=None, defaults=(1e-05,)) paddle.fluid.layers.dice_loss ArgSpec(args=['input', 'label', 'epsilon'], varargs=None, keywords=None, defaults=(1e-05,))
paddle.fluid.layers.image_resize ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'resample', 'actual_shape'], varargs=None, keywords=None, defaults=(None, None, None, 'BILINEAR', None)) paddle.fluid.layers.image_resize ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'resample', 'actual_shape', 'align_corners', 'align_mode'], varargs=None, keywords=None, defaults=(None, None, None, 'BILINEAR', None, True, 1))
paddle.fluid.layers.image_resize_short ArgSpec(args=['input', 'out_short_len', 'resample'], varargs=None, keywords=None, defaults=('BILINEAR',)) paddle.fluid.layers.image_resize_short ArgSpec(args=['input', 'out_short_len', 'resample'], varargs=None, keywords=None, defaults=('BILINEAR',))
paddle.fluid.layers.resize_bilinear ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'actual_shape'], varargs=None, keywords=None, defaults=(None, None, None, None)) paddle.fluid.layers.resize_bilinear ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'actual_shape', 'align_corners', 'align_mode'], varargs=None, keywords=None, defaults=(None, None, None, None, True, 1))
paddle.fluid.layers.resize_nearest ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'actual_shape'], varargs=None, keywords=None, defaults=(None, None, None, None)) paddle.fluid.layers.resize_nearest ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'actual_shape', 'align_corners'], varargs=None, keywords=None, defaults=(None, None, None, None, True))
paddle.fluid.layers.gather ArgSpec(args=['input', 'index'], varargs=None, keywords=None, defaults=None) paddle.fluid.layers.gather ArgSpec(args=['input', 'index'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.scatter ArgSpec(args=['input', 'index', 'updates', 'name'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.scatter ArgSpec(args=['input', 'index', 'updates', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.sequence_scatter ArgSpec(args=['input', 'index', 'updates', 'name'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.sequence_scatter ArgSpec(args=['input', 'index', 'updates', 'name'], varargs=None, keywords=None, defaults=(None,))
...@@ -229,7 +229,7 @@ paddle.fluid.layers.random_data_generator ArgSpec(args=['low', 'high', 'shapes', ...@@ -229,7 +229,7 @@ paddle.fluid.layers.random_data_generator ArgSpec(args=['low', 'high', 'shapes',
paddle.fluid.layers.py_reader ArgSpec(args=['capacity', 'shapes', 'dtypes', 'lod_levels', 'name', 'use_double_buffer'], varargs=None, keywords=None, defaults=(None, None, True)) paddle.fluid.layers.py_reader ArgSpec(args=['capacity', 'shapes', 'dtypes', 'lod_levels', 'name', 'use_double_buffer'], varargs=None, keywords=None, defaults=(None, None, True))
paddle.fluid.layers.create_py_reader_by_data ArgSpec(args=['capacity', 'feed_list', 'name', 'use_double_buffer'], varargs=None, keywords=None, defaults=(None, True)) paddle.fluid.layers.create_py_reader_by_data ArgSpec(args=['capacity', 'feed_list', 'name', 'use_double_buffer'], varargs=None, keywords=None, defaults=(None, True))
paddle.fluid.layers.Preprocessor.__init__ ArgSpec(args=['self', 'reader', 'name'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.Preprocessor.__init__ ArgSpec(args=['self', 'reader', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.Preprocessor.block ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None) paddle.fluid.layers.Preprocessor.block ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.Preprocessor.inputs ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) paddle.fluid.layers.Preprocessor.inputs ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.Preprocessor.outputs ArgSpec(args=['self'], varargs='outs', keywords=None, defaults=None) paddle.fluid.layers.Preprocessor.outputs ArgSpec(args=['self'], varargs='outs', keywords=None, defaults=None)
paddle.fluid.layers.load ArgSpec(args=['out', 'file_path', 'load_as_fp16'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.load ArgSpec(args=['out', 'file_path', 'load_as_fp16'], varargs=None, keywords=None, defaults=(None,))
...@@ -270,7 +270,7 @@ paddle.fluid.layers.IfElse.input ArgSpec(args=['self', 'x'], varargs=None, keywo ...@@ -270,7 +270,7 @@ paddle.fluid.layers.IfElse.input ArgSpec(args=['self', 'x'], varargs=None, keywo
paddle.fluid.layers.IfElse.output ArgSpec(args=['self'], varargs='outs', keywords=None, defaults=None) paddle.fluid.layers.IfElse.output ArgSpec(args=['self'], varargs='outs', keywords=None, defaults=None)
paddle.fluid.layers.IfElse.true_block ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) paddle.fluid.layers.IfElse.true_block ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.DynamicRNN.__init__ ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.DynamicRNN.__init__ ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.DynamicRNN.block ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None) paddle.fluid.layers.DynamicRNN.block ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.DynamicRNN.memory ArgSpec(args=['self', 'init', 'shape', 'value', 'need_reorder', 'dtype'], varargs=None, keywords=None, defaults=(None, None, 0.0, False, 'float32')) paddle.fluid.layers.DynamicRNN.memory ArgSpec(args=['self', 'init', 'shape', 'value', 'need_reorder', 'dtype'], varargs=None, keywords=None, defaults=(None, None, 0.0, False, 'float32'))
paddle.fluid.layers.DynamicRNN.output ArgSpec(args=['self'], varargs='outputs', keywords=None, defaults=None) paddle.fluid.layers.DynamicRNN.output ArgSpec(args=['self'], varargs='outputs', keywords=None, defaults=None)
paddle.fluid.layers.DynamicRNN.static_input ArgSpec(args=['self', 'x'], varargs=None, keywords=None, defaults=None) paddle.fluid.layers.DynamicRNN.static_input ArgSpec(args=['self', 'x'], varargs=None, keywords=None, defaults=None)
...@@ -322,9 +322,11 @@ paddle.fluid.layers.generate_proposal_labels ArgSpec(args=['rpn_rois', 'gt_class ...@@ -322,9 +322,11 @@ paddle.fluid.layers.generate_proposal_labels ArgSpec(args=['rpn_rois', 'gt_class
paddle.fluid.layers.generate_proposals ArgSpec(args=['scores', 'bbox_deltas', 'im_info', 'anchors', 'variances', 'pre_nms_top_n', 'post_nms_top_n', 'nms_thresh', 'min_size', 'eta', 'name'], varargs=None, keywords=None, defaults=(6000, 1000, 0.5, 0.1, 1.0, None)) paddle.fluid.layers.generate_proposals ArgSpec(args=['scores', 'bbox_deltas', 'im_info', 'anchors', 'variances', 'pre_nms_top_n', 'post_nms_top_n', 'nms_thresh', 'min_size', 'eta', 'name'], varargs=None, keywords=None, defaults=(6000, 1000, 0.5, 0.1, 1.0, None))
paddle.fluid.layers.generate_mask_labels ArgSpec(args=['im_info', 'gt_classes', 'is_crowd', 'gt_segms', 'rois', 'labels_int32', 'num_classes', 'resolution'], varargs=None, keywords=None, defaults=None) paddle.fluid.layers.generate_mask_labels ArgSpec(args=['im_info', 'gt_classes', 'is_crowd', 'gt_segms', 'rois', 'labels_int32', 'num_classes', 'resolution'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.iou_similarity ArgSpec(args=['x', 'y', 'name'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.iou_similarity ArgSpec(args=['x', 'y', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.box_coder ArgSpec(args=['prior_box', 'prior_box_var', 'target_box', 'code_type', 'box_normalized', 'name'], varargs=None, keywords=None, defaults=('encode_center_size', True, None)) paddle.fluid.layers.box_coder ArgSpec(args=['prior_box', 'prior_box_var', 'target_box', 'code_type', 'box_normalized', 'name', 'axis'], varargs=None, keywords=None, defaults=('encode_center_size', True, None, 0))
paddle.fluid.layers.polygon_box_transform ArgSpec(args=['input', 'name'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.polygon_box_transform ArgSpec(args=['input', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.yolov3_loss ArgSpec(args=['x', 'gtbox', 'gtlabel', 'anchors', 'class_num', 'ignore_thresh', 'loss_weight_xy', 'loss_weight_wh', 'loss_weight_conf_target', 'loss_weight_conf_notarget', 'loss_weight_class', 'name'], varargs=None, keywords=None, defaults=(None, None, None, None, None, None)) paddle.fluid.layers.yolov3_loss ArgSpec(args=['x', 'gtbox', 'gtlabel', 'anchors', 'anchor_mask', 'class_num', 'ignore_thresh', 'downsample_ratio', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.box_clip ArgSpec(args=['input', 'im_info', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.multiclass_nms ArgSpec(args=['bboxes', 'scores', 'score_threshold', 'nms_top_k', 'keep_top_k', 'nms_threshold', 'normalized', 'nms_eta', 'background_label', 'name'], varargs=None, keywords=None, defaults=(0.3, True, 1.0, 0, None))
paddle.fluid.layers.accuracy ArgSpec(args=['input', 'label', 'k', 'correct', 'total'], varargs=None, keywords=None, defaults=(1, None, None)) paddle.fluid.layers.accuracy ArgSpec(args=['input', 'label', 'k', 'correct', 'total'], varargs=None, keywords=None, defaults=(1, None, None))
paddle.fluid.layers.auc ArgSpec(args=['input', 'label', 'curve', 'num_thresholds', 'topk', 'slide_steps'], varargs=None, keywords=None, defaults=('ROC', 4095, 1, 1)) paddle.fluid.layers.auc ArgSpec(args=['input', 'label', 'curve', 'num_thresholds', 'topk', 'slide_steps'], varargs=None, keywords=None, defaults=('ROC', 4095, 1, 1))
paddle.fluid.layers.exponential_decay ArgSpec(args=['learning_rate', 'decay_steps', 'decay_rate', 'staircase'], varargs=None, keywords=None, defaults=(False,)) paddle.fluid.layers.exponential_decay ArgSpec(args=['learning_rate', 'decay_steps', 'decay_rate', 'staircase'], varargs=None, keywords=None, defaults=(False,))
...@@ -344,12 +346,12 @@ paddle.fluid.contrib.StateCell.set_state ArgSpec(args=['self', 'state_name', 'st ...@@ -344,12 +346,12 @@ paddle.fluid.contrib.StateCell.set_state ArgSpec(args=['self', 'state_name', 'st
paddle.fluid.contrib.StateCell.state_updater ArgSpec(args=['self', 'updater'], varargs=None, keywords=None, defaults=None) paddle.fluid.contrib.StateCell.state_updater ArgSpec(args=['self', 'updater'], varargs=None, keywords=None, defaults=None)
paddle.fluid.contrib.StateCell.update_states ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) paddle.fluid.contrib.StateCell.update_states ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.contrib.TrainingDecoder.__init__ ArgSpec(args=['self', 'state_cell', 'name'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.contrib.TrainingDecoder.__init__ ArgSpec(args=['self', 'state_cell', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.contrib.TrainingDecoder.block ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None) paddle.fluid.contrib.TrainingDecoder.block ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.contrib.TrainingDecoder.output ArgSpec(args=['self'], varargs='outputs', keywords=None, defaults=None) paddle.fluid.contrib.TrainingDecoder.output ArgSpec(args=['self'], varargs='outputs', keywords=None, defaults=None)
paddle.fluid.contrib.TrainingDecoder.static_input ArgSpec(args=['self', 'x'], varargs=None, keywords=None, defaults=None) paddle.fluid.contrib.TrainingDecoder.static_input ArgSpec(args=['self', 'x'], varargs=None, keywords=None, defaults=None)
paddle.fluid.contrib.TrainingDecoder.step_input ArgSpec(args=['self', 'x'], varargs=None, keywords=None, defaults=None) paddle.fluid.contrib.TrainingDecoder.step_input ArgSpec(args=['self', 'x'], varargs=None, keywords=None, defaults=None)
paddle.fluid.contrib.BeamSearchDecoder.__init__ ArgSpec(args=['self', 'state_cell', 'init_ids', 'init_scores', 'target_dict_dim', 'word_dim', 'input_var_dict', 'topk_size', 'sparse_emb', 'max_len', 'beam_size', 'end_id', 'name'], varargs=None, keywords=None, defaults=({}, 50, True, 100, 1, 1, None)) paddle.fluid.contrib.BeamSearchDecoder.__init__ ArgSpec(args=['self', 'state_cell', 'init_ids', 'init_scores', 'target_dict_dim', 'word_dim', 'input_var_dict', 'topk_size', 'sparse_emb', 'max_len', 'beam_size', 'end_id', 'name'], varargs=None, keywords=None, defaults=({}, 50, True, 100, 1, 1, None))
paddle.fluid.contrib.BeamSearchDecoder.block ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None) paddle.fluid.contrib.BeamSearchDecoder.block ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.contrib.BeamSearchDecoder.decode ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) paddle.fluid.contrib.BeamSearchDecoder.decode ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.contrib.BeamSearchDecoder.early_stop ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) paddle.fluid.contrib.BeamSearchDecoder.early_stop ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.contrib.BeamSearchDecoder.read_array ArgSpec(args=['self', 'init', 'is_ids', 'is_scores'], varargs=None, keywords=None, defaults=(False, False)) paddle.fluid.contrib.BeamSearchDecoder.read_array ArgSpec(args=['self', 'init', 'is_ids', 'is_scores'], varargs=None, keywords=None, defaults=(False, False))
...@@ -360,6 +362,9 @@ paddle.fluid.contrib.QuantizeTranspiler.__init__ ArgSpec(args=['self', 'weight_b ...@@ -360,6 +362,9 @@ paddle.fluid.contrib.QuantizeTranspiler.__init__ ArgSpec(args=['self', 'weight_b
paddle.fluid.contrib.QuantizeTranspiler.convert_to_int8 ArgSpec(args=['self', 'program', 'place', 'scope'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.contrib.QuantizeTranspiler.convert_to_int8 ArgSpec(args=['self', 'program', 'place', 'scope'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.contrib.QuantizeTranspiler.freeze_program ArgSpec(args=['self', 'program', 'place', 'fuse_bn', 'scope'], varargs=None, keywords=None, defaults=(False, None)) paddle.fluid.contrib.QuantizeTranspiler.freeze_program ArgSpec(args=['self', 'program', 'place', 'fuse_bn', 'scope'], varargs=None, keywords=None, defaults=(False, None))
paddle.fluid.contrib.QuantizeTranspiler.training_transpile ArgSpec(args=['self', 'program', 'startup_program'], varargs=None, keywords=None, defaults=(None, None)) paddle.fluid.contrib.QuantizeTranspiler.training_transpile ArgSpec(args=['self', 'program', 'startup_program'], varargs=None, keywords=None, defaults=(None, None))
paddle.fluid.contrib.Calibrator.__init__ ArgSpec(args=['self'], varargs='args', keywords='kwargs', defaults=None)
paddle.fluid.contrib.Calibrator.sample_data ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.contrib.Calibrator.save_int8_model ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.contrib.reader.ctr_reader.ctr_reader ArgSpec(args=['feed_dict', 'file_type', 'file_format', 'dense_slot_index', 'sparse_slot_index', 'capacity', 'thread_num', 'batch_size', 'file_list', 'slots', 'name'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.contrib.reader.ctr_reader.ctr_reader ArgSpec(args=['feed_dict', 'file_type', 'file_format', 'dense_slot_index', 'sparse_slot_index', 'capacity', 'thread_num', 'batch_size', 'file_list', 'slots', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.contrib.build_compressor ArgSpec(args=['place', 'data_reader', 'data_feeder', 'scope', 'metrics', 'epoch', 'config'], varargs=None, keywords=None, defaults=(None, None, None, None, None, None, None)) paddle.fluid.contrib.build_compressor ArgSpec(args=['place', 'data_reader', 'data_feeder', 'scope', 'metrics', 'epoch', 'config'], varargs=None, keywords=None, defaults=(None, None, None, None, None, None, None))
paddle.fluid.contrib.CompressPass.__init__ ArgSpec(args=['self', 'place', 'data_reader', 'data_feeder', 'scope', 'metrics', 'epoch', 'program_exe'], varargs=None, keywords=None, defaults=(None, None, None, None, None, None, None)) paddle.fluid.contrib.CompressPass.__init__ ArgSpec(args=['self', 'place', 'data_reader', 'data_feeder', 'scope', 'metrics', 'epoch', 'program_exe'], varargs=None, keywords=None, defaults=(None, None, None, None, None, None, None))
...@@ -451,7 +456,7 @@ paddle.fluid.optimizer.AdadeltaOptimizer.apply_gradients ArgSpec(args=['self', ' ...@@ -451,7 +456,7 @@ paddle.fluid.optimizer.AdadeltaOptimizer.apply_gradients ArgSpec(args=['self', '
paddle.fluid.optimizer.AdadeltaOptimizer.backward ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)) paddle.fluid.optimizer.AdadeltaOptimizer.backward ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None))
paddle.fluid.optimizer.AdadeltaOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)) paddle.fluid.optimizer.AdadeltaOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None))
paddle.fluid.optimizer.ModelAverage.__init__ ArgSpec(args=['self', 'average_window_rate', 'min_average_window', 'max_average_window', 'regularization', 'name'], varargs=None, keywords=None, defaults=(10000, 10000, None, None)) paddle.fluid.optimizer.ModelAverage.__init__ ArgSpec(args=['self', 'average_window_rate', 'min_average_window', 'max_average_window', 'regularization', 'name'], varargs=None, keywords=None, defaults=(10000, 10000, None, None))
paddle.fluid.optimizer.ModelAverage.apply ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None) paddle.fluid.optimizer.ModelAverage.apply ArgSpec(args=['self', 'executor', 'need_restore'], varargs=None, keywords=None, defaults=(True,))
paddle.fluid.optimizer.ModelAverage.apply_gradients ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None) paddle.fluid.optimizer.ModelAverage.apply_gradients ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None)
paddle.fluid.optimizer.ModelAverage.backward ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)) paddle.fluid.optimizer.ModelAverage.backward ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None))
paddle.fluid.optimizer.ModelAverage.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)) paddle.fluid.optimizer.ModelAverage.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None))
...@@ -486,14 +491,14 @@ paddle.fluid.clip.ErrorClipByValue.__init__ ArgSpec(args=['self', 'max', 'min'], ...@@ -486,14 +491,14 @@ paddle.fluid.clip.ErrorClipByValue.__init__ ArgSpec(args=['self', 'max', 'min'],
paddle.fluid.clip.GradientClipByValue.__init__ ArgSpec(args=['self', 'max', 'min'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.clip.GradientClipByValue.__init__ ArgSpec(args=['self', 'max', 'min'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.clip.GradientClipByNorm.__init__ ArgSpec(args=['self', 'clip_norm'], varargs=None, keywords=None, defaults=None) paddle.fluid.clip.GradientClipByNorm.__init__ ArgSpec(args=['self', 'clip_norm'], varargs=None, keywords=None, defaults=None)
paddle.fluid.clip.GradientClipByGlobalNorm.__init__ ArgSpec(args=['self', 'clip_norm', 'group_name'], varargs=None, keywords=None, defaults=('default_group',)) paddle.fluid.clip.GradientClipByGlobalNorm.__init__ ArgSpec(args=['self', 'clip_norm', 'group_name'], varargs=None, keywords=None, defaults=('default_group',))
paddle.fluid.profiler.cuda_profiler ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None) paddle.fluid.profiler.cuda_profiler ArgSpec(args=['output_file', 'output_mode', 'config'], varargs=None, keywords=None, defaults=(None, None))
paddle.fluid.profiler.reset_profiler ArgSpec(args=[], varargs=None, keywords=None, defaults=None) paddle.fluid.profiler.reset_profiler ArgSpec(args=[], varargs=None, keywords=None, defaults=None)
paddle.fluid.profiler.profiler ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None) paddle.fluid.profiler.profiler ArgSpec(args=['state', 'sorted_key', 'profile_path'], varargs=None, keywords=None, defaults=(None, '/tmp/profile'))
paddle.fluid.profiler.start_profiler ArgSpec(args=['state'], varargs=None, keywords=None, defaults=None) paddle.fluid.profiler.start_profiler ArgSpec(args=['state'], varargs=None, keywords=None, defaults=None)
paddle.fluid.profiler.stop_profiler ArgSpec(args=['sorted_key', 'profile_path'], varargs=None, keywords=None, defaults=(None, '/tmp/profile')) paddle.fluid.profiler.stop_profiler ArgSpec(args=['sorted_key', 'profile_path'], varargs=None, keywords=None, defaults=(None, '/tmp/profile'))
paddle.fluid.unique_name.generate ArgSpec(args=['key'], varargs=None, keywords=None, defaults=None) paddle.fluid.unique_name.generate ArgSpec(args=['key'], varargs=None, keywords=None, defaults=None)
paddle.fluid.unique_name.switch ArgSpec(args=['new_generator'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.unique_name.switch ArgSpec(args=['new_generator'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.unique_name.guard ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None) paddle.fluid.unique_name.guard ArgSpec(args=['new_generator'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.recordio_writer.convert_reader_to_recordio_file ArgSpec(args=['filename', 'reader_creator', 'feeder', 'compressor', 'max_num_records', 'feed_order'], varargs=None, keywords=None, defaults=(Compressor.Snappy, 1000, None)) paddle.fluid.recordio_writer.convert_reader_to_recordio_file ArgSpec(args=['filename', 'reader_creator', 'feeder', 'compressor', 'max_num_records', 'feed_order'], varargs=None, keywords=None, defaults=(Compressor.Snappy, 1000, None))
paddle.fluid.recordio_writer.convert_reader_to_recordio_files ArgSpec(args=['filename', 'batch_per_file', 'reader_creator', 'feeder', 'compressor', 'max_num_records', 'feed_order'], varargs=None, keywords=None, defaults=(Compressor.Snappy, 1000, None)) paddle.fluid.recordio_writer.convert_reader_to_recordio_files ArgSpec(args=['filename', 'batch_per_file', 'reader_creator', 'feeder', 'compressor', 'max_num_records', 'feed_order'], varargs=None, keywords=None, defaults=(Compressor.Snappy, 1000, None))
paddle.fluid.Scope Scope() -> paddle.fluid.core._Scope paddle.fluid.Scope Scope() -> paddle.fluid.core._Scope
......
paddle/fluid/framework/CMakeLists.txt
浏览文件 @ 17ada1d2
#windows treat symbolic file as a real file, which is different with unix #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. #We create a hidden file and compile it instead of origin source file.
function(windows_symbolic TARGET) function(windows_symbolic TARGET)
...@@ -129,7 +128,7 @@ cc_test(version_test SRCS version_test.cc DEPS version) ...@@ -129,7 +128,7 @@ cc_test(version_test SRCS version_test.cc DEPS version)
cc_library(proto_desc SRCS var_desc.cc op_desc.cc block_desc.cc program_desc.cc DEPS shape_inference op_info operator glog version) cc_library(proto_desc SRCS var_desc.cc op_desc.cc block_desc.cc program_desc.cc DEPS shape_inference op_info operator glog version)
cc_library(op_registry SRCS op_registry.cc DEPS op_proto_maker op_info operator glog proto_desc) cc_library(op_registry SRCS op_registry.cc DEPS op_proto_maker op_info operator glog proto_desc memory_optimize_helper)
nv_test(op_registry_test SRCS op_registry_test.cc DEPS op_registry) 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) py_proto_compile(framework_py_proto SRCS framework.proto data_feed.proto)
...@@ -159,18 +158,19 @@ cc_library(variable_helper SRCS variable_helper.cc DEPS lod_tensor) ...@@ -159,18 +158,19 @@ 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) 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) if(WITH_NGRAPH)
cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog set(NGRAPH_EXE_DEPS ngraph_engine)
lod_rank_table feed_fetch_method sendrecvop_rpc ${GLOB_DISTRIBUTE_DEPS} graph_to_program_pass variable_helper) else()
set(NGRAPH_EXE_DEPS)
endif()
set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor") if(WITH_DISTRIBUTE)
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 ${NGRAPH_EXE_DEPS})
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() else()
if (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 ${NGRAPH_EXE_DEPS})
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 ngraph_engine)
else ()
cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass variable_helper)
endif()
cc_test(test_naive_executor SRCS naive_executor_test.cc DEPS naive_executor elementwise_add_op) cc_test(test_naive_executor SRCS naive_executor_test.cc DEPS naive_executor elementwise_add_op)
endif() endif()
...@@ -193,6 +193,7 @@ cc_library(prune SRCS prune.cc DEPS framework_proto) ...@@ -193,6 +193,7 @@ cc_library(prune SRCS prune.cc DEPS framework_proto)
cc_test(prune_test SRCS prune_test.cc DEPS op_info prune recurrent_op device_context) cc_test(prune_test SRCS prune_test.cc DEPS op_info prune recurrent_op device_context)
cc_test(var_type_inference_test SRCS var_type_inference_test.cc DEPS op_registry cc_test(var_type_inference_test SRCS var_type_inference_test.cc DEPS op_registry
proto_desc) proto_desc)
cc_test(inplace_op_inference_test SRCS inplace_op_inference_test.cc DEPS op_registry proto_desc op_info memory_optimize_helper)
cc_library(selected_rows SRCS selected_rows.cc DEPS tensor) cc_library(selected_rows SRCS selected_rows.cc DEPS tensor)
cc_test(selected_rows_test SRCS selected_rows_test.cc DEPS selected_rows) cc_test(selected_rows_test SRCS selected_rows_test.cc DEPS selected_rows)
...@@ -207,3 +208,24 @@ endif (NOT WIN32) ...@@ -207,3 +208,24 @@ endif (NOT WIN32)
cc_library(dlpack_tensor SRCS dlpack_tensor.cc DEPS tensor dlpack) cc_library(dlpack_tensor SRCS dlpack_tensor.cc DEPS tensor dlpack)
cc_test(dlpack_tensor_test SRCS dlpack_tensor_test.cc DEPS dlpack_tensor glog) cc_test(dlpack_tensor_test SRCS dlpack_tensor_test.cc DEPS dlpack_tensor glog)
# Get the current working branch
execute_process(
COMMAND git rev-parse --abbrev-ref HEAD
WORKING_DIRECTORY ${CMAKE_SOURCE_DIR}
OUTPUT_VARIABLE PADDLE_BRANCH
OUTPUT_STRIP_TRAILING_WHITESPACE
)
# Get the latest abbreviated commit hash of the working branch
execute_process(
COMMAND git log -1 --format=%h
WORKING_DIRECTORY ${CMAKE_SOURCE_DIR}
OUTPUT_VARIABLE PADDLE_COMMIT
OUTPUT_STRIP_TRAILING_WHITESPACE
)
message(STATUS "commit: ${PADDLE_COMMIT}")
message(STATUS "branch: ${PADDLE_BRANCH}")
configure_file(commit.h.in commit.h)
paddle/fluid/framework/async_executor.cc
浏览文件 @ 17ada1d2
...@@ -244,6 +244,7 @@ void AsyncExecutor::RunFromFile(const ProgramDesc& main_program, ...@@ -244,6 +244,7 @@ void AsyncExecutor::RunFromFile(const ProgramDesc& main_program,
auto& block = main_program.Block(0); auto& block = main_program.Block(0);
for (auto var_name : fetch_var_names) { for (auto var_name : fetch_var_names) {
auto var_desc = block.FindVar(var_name); auto var_desc = block.FindVar(var_name);
PADDLE_ENFORCE_NOT_NULL(var_desc, "%s is not found.", var_name);
auto shapes = var_desc->GetShape(); auto shapes = var_desc->GetShape();
PADDLE_ENFORCE(shapes[shapes.size() - 1] == 1, PADDLE_ENFORCE(shapes[shapes.size() - 1] == 1,
"var %s: Fetched var has wrong shape, " "var %s: Fetched var has wrong shape, "
......
paddle/fluid/framework/commit.h.in 0 → 100644
浏览文件 @ 17ada1d2
#pragma once
#include <string>
namespace paddle {
namespace framework {
static std::string paddle_commit() {
return "@PADDLE_COMMIT@";
}
static std::string paddle_compile_branch() {
return "@PADDLE_BRANCH@";
}
static std::string paddle_version() {
return "@PADDLE_VERSION@";
}
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/CMakeLists.txt
浏览文件 @ 17ada1d2
...@@ -50,10 +50,10 @@ cc_library(data_balance_op_handle SRCS data_balance_op_handle.cc DEPS op_handle_ ...@@ -50,10 +50,10 @@ 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(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(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(memory_optimize_helper SRCS memory_optimize_helper.cc DEPS graph graph_helper)
cc_library(memory_optimize_pass SRCS memory_optimize_pass.cc DEPS memory_optimize_helper pass)
cc_library(inplace_op_pass SRCS inplace_op_pass.cc DEPS memory_optimize_pass op_info)
cc_library(modify_op_lock_and_record_event_pass SRCS modify_op_lock_and_record_event_pass.cc DEPS computation_op_handle op_graph_view multi_devices_helper) 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)
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(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_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(eager_deletion_pass SRCS eager_deletion_pass.cc DEPS computation_op_handle eager_deletion_op_handle graph graph_helper pass)
...@@ -65,13 +65,11 @@ cc_library(all_reduce_deps_pass SRCS all_reduce_deps_pass.cc DEPS graph graph_he ...@@ -65,13 +65,11 @@ 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 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) 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 reference_count_pass eager_deletion_pass memory_optimize_pass memory_early_delete_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 inplace_op_pass)
if (WITH_GPU) if (WITH_GPU)
list(APPEND SSA_GRAPH_EXECUTOR_DEPS reference_count_pass) list(APPEND SSA_GRAPH_EXECUTOR_DEPS reference_count_pass)
endif() endif()
cc_test(memory_reuse_types_test SRCS memory_reuse_types_test.cc memory_reuse_types.cc DEPS framework_proto graph) cc_test(memory_optimize_helper_test SRCS memory_optimize_helper_test.cc memory_optimize_helper.cc DEPS framework_proto graph graph_helper op_registry)
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}) cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS ${SSA_GRAPH_EXECUTOR_DEPS})
cc_library(threaded_ssa_graph_executor SRCS threaded_ssa_graph_executor.cc DEPS fetch_op_handle ssa_graph_executor scope cc_library(threaded_ssa_graph_executor SRCS threaded_ssa_graph_executor.cc DEPS fetch_op_handle ssa_graph_executor scope
......
paddle/fluid/framework/details/build_strategy.cc
浏览文件 @ 17ada1d2
...@@ -17,7 +17,7 @@ limitations under the License. */ ...@@ -17,7 +17,7 @@ limitations under the License. */
#include <glog/logging.h> #include <glog/logging.h>
#include <memory> #include <memory>
#include "paddle/fluid/framework/details/memory_reuse_types.h" #include "paddle/fluid/framework/details/memory_optimize_helper.h"
#include "paddle/fluid/framework/details/multi_devices_graph_pass.h" #include "paddle/fluid/framework/details/multi_devices_graph_pass.h"
#include "paddle/fluid/framework/details/multi_devices_graph_print_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/reduce_op_handle.h"
...@@ -47,6 +47,22 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder { ...@@ -47,6 +47,22 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
AppendPass("sequential_execution_pass"); AppendPass("sequential_execution_pass");
} }
// Add op fusion.
if (strategy.fuse_relu_depthwise_conv_) {
AppendPass("fuse_relu_depthwise_conv_pass");
}
// NOTE(dzhwinter): A note for automatical inplace.
// 1. modify program desc passes should put
// before inplace pass.
// 2. manually configured inplace should put
// before inplace_pass
// Add automatically inplace.
if (strategy_.enable_inplace_) {
AppendPass("inplace_pass");
}
// Add a graph viz pass to record a graph. // Add a graph viz pass to record a graph.
if (!strategy_.debug_graphviz_path_.empty()) { if (!strategy_.debug_graphviz_path_.empty()) {
auto viz_pass = AppendPass("graph_viz_pass"); auto viz_pass = AppendPass("graph_viz_pass");
...@@ -55,10 +71,6 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder { ...@@ -55,10 +71,6 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
viz_pass->Set<std::string>("graph_viz_path", new std::string(graph_path)); viz_pass->Set<std::string>("graph_viz_path", new std::string(graph_path));
} }
// Add op fusion.
if (strategy.fuse_relu_depthwise_conv_) {
AppendPass("fuse_relu_depthwise_conv_pass");
}
if (strategy.fuse_elewise_add_act_ops_) { if (strategy.fuse_elewise_add_act_ops_) {
auto fuse_elewise_add_act_pass = AppendPass("fuse_elewise_add_act_pass"); auto fuse_elewise_add_act_pass = AppendPass("fuse_elewise_add_act_pass");
// Add a graph viz pass to record a graph. // Add a graph viz pass to record a graph.
...@@ -88,7 +100,7 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder { ...@@ -88,7 +100,7 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
// A side-effect of that, memory optimize cannot forsee the fetched vars // A side-effect of that, memory optimize cannot forsee the fetched vars
// , so fetchlist should be set persistable before call the Run interface. // , so fetchlist should be set persistable before call the Run interface.
if (strategy.memory_optimize_) { if (strategy.memory_optimize_) {
auto analysis_var_pass = AppendPass("analysis_var_pass"); auto memory_optimize_pass = AppendPass("memory_optimize_pass");
} }
AppendMultiDevPass(strategy); AppendMultiDevPass(strategy);
...@@ -186,14 +198,14 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply( ...@@ -186,14 +198,14 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
pass->Erase("nccl_ctxs"); pass->Erase("nccl_ctxs");
pass->SetNotOwned<platform::NCCLContextMap>("nccl_ctxs", nctx); pass->SetNotOwned<platform::NCCLContextMap>("nccl_ctxs", nctx);
#endif #endif
} else if (pass->Type() == "memory_optimize_pass") {
} else if (pass->Type() == "analysis_var_pass") { if (graph->Has(kAllOpDescs)) {
graph->Erase(kAllOpDescs);
}
const std::vector<OpDesc *> *all_op_descs = const std::vector<OpDesc *> *all_op_descs =
new std::vector<OpDesc *>(main_program.Block(0).AllOps()); new std::vector<OpDesc *>(main_program.Block(0).AllOps());
graph->Set<const std::vector<OpDesc *>>(kAllOpDescs, graph->Set<const std::vector<OpDesc *>>(kAllOpDescs,
all_op_descs); // take ownership all_op_descs); // take ownership
graph->Set<GraphNodePool>(kGraphNodePool,
new GraphNodePool); // take ownership
pass->Erase(kAllOpDescs); pass->Erase(kAllOpDescs);
pass->SetNotOwned<const std::vector<OpDesc *>>(kAllOpDescs, all_op_descs); pass->SetNotOwned<const std::vector<OpDesc *>>(kAllOpDescs, all_op_descs);
...@@ -214,6 +226,13 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply( ...@@ -214,6 +226,13 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
pass->Set<const std::vector<OpDesc *>>( pass->Set<const std::vector<OpDesc *>>(
kAllOpDescs, kAllOpDescs,
new std::vector<OpDesc *>(main_program.Block(0).AllOps())); new std::vector<OpDesc *>(main_program.Block(0).AllOps()));
} else if (pass->Type() == "inplace_pass") {
if (graph->Has(kAllOpDescs)) {
graph->Erase(kAllOpDescs);
}
graph->Set<const std::vector<OpDesc *>>(
kAllOpDescs,
new std::vector<OpDesc *>(main_program.Block(0).AllOps()));
} else if (pass->Type() == "fuse_relu_depthwise_conv_pass") { } else if (pass->Type() == "fuse_relu_depthwise_conv_pass") {
if (!use_cuda) { if (!use_cuda) {
LOG(WARNING) << "fuse_relu_depthwise_conv_pass is only supported on " LOG(WARNING) << "fuse_relu_depthwise_conv_pass is only supported on "
...@@ -239,9 +258,10 @@ USE_PASS(allreduce_mode_multi_devices_pass); ...@@ -239,9 +258,10 @@ USE_PASS(allreduce_mode_multi_devices_pass);
USE_PASS(dist_multi_devices_pass); USE_PASS(dist_multi_devices_pass);
USE_PASS(multi_devices_check_pass); USE_PASS(multi_devices_check_pass);
USE_PASS(multi_devices_print_pass); USE_PASS(multi_devices_print_pass);
USE_PASS(analysis_var_pass); USE_PASS(memory_optimize_pass);
USE_PASS(sequential_execution_pass); USE_PASS(sequential_execution_pass);
USE_PASS(all_reduce_deps_pass); USE_PASS(all_reduce_deps_pass);
USE_PASS(modify_op_lock_and_record_event_pass); USE_PASS(modify_op_lock_and_record_event_pass);
USE_PASS(inplace_pass);
USE_PASS(lock_free_optimize_pass); USE_PASS(lock_free_optimize_pass);
USE_PASS(graph_to_program_pass); USE_PASS(graph_to_program_pass);
paddle/fluid/framework/details/build_strategy.h
浏览文件 @ 17ada1d2
...@@ -77,8 +77,10 @@ struct BuildStrategy { ...@@ -77,8 +77,10 @@ struct BuildStrategy {
bool fuse_relu_depthwise_conv_{false}; bool fuse_relu_depthwise_conv_{false};
bool memory_optimize_{false}; bool memory_optimize_{false};
// TODO(dzhwinter):
bool memory_early_delete_{false}; // make enable_inplace, memory_optimize_
// memory_early_delete_ true by default
bool enable_inplace_{false};
bool enable_sequential_execution_{false}; bool enable_sequential_execution_{false};
......
paddle/fluid/framework/details/computation_op_handle.h
浏览文件 @ 17ada1d2
...@@ -26,7 +26,7 @@ ...@@ -26,7 +26,7 @@
namespace paddle { namespace paddle {
namespace framework { namespace framework {
namespace details { namespace details {
struct ComputationOpHandle : public OpHandleBase { class ComputationOpHandle : public OpHandleBase {
public: public:
ComputationOpHandle(ir::Node *node, Scope *scope, platform::Place place, ComputationOpHandle(ir::Node *node, Scope *scope, platform::Place place,
size_t scope_idx); size_t scope_idx);
......
paddle/fluid/framework/details/fused_broadcast_op_handle_test.cc
浏览文件 @ 17ada1d2
...@@ -34,8 +34,8 @@ struct TestFusedBroadcastOpHandle : TestBroadcastOpHandle { ...@@ -34,8 +34,8 @@ struct TestFusedBroadcastOpHandle : TestBroadcastOpHandle {
->Var(details::kLocalExecScopeName) ->Var(details::kLocalExecScopeName)
->GetMutable<Scope*>() = &local_scope; ->GetMutable<Scope*>() = &local_scope;
for (size_t j = 0; j < input_scope_idxes.size(); ++j) { for (size_t j = 0; j < input_scope_idxes.size(); ++j) {
local_scope.Var("out_var" + j); local_scope.Var("out_var" + std::to_string(j));
if (i == j) local_scope.Var("in_var" + j); if (i == j) local_scope.Var("in_var" + std::to_string(j));
} }
param_scopes_.emplace_back(&local_scope); param_scopes_.emplace_back(&local_scope);
} }
...@@ -62,20 +62,21 @@ struct TestFusedBroadcastOpHandle : TestBroadcastOpHandle { ...@@ -62,20 +62,21 @@ struct TestFusedBroadcastOpHandle : TestBroadcastOpHandle {
for (size_t i = 0; i < input_scope_idxes.size(); ++i) { for (size_t i = 0; i < input_scope_idxes.size(); ++i) {
// add input var handle // add input var handle
nodes_.emplace_back( nodes_.emplace_back(ir::CreateNodeForTest("in_node" + std::to_string(i),
ir::CreateNodeForTest("in_node" + i, ir::Node::Type::kVariable)); ir::Node::Type::kVariable));
VarHandle* in_var_handle = VarHandle* in_var_handle = new VarHandle(
new VarHandle(nodes_.back().get(), 1, input_scope_idxes[i], nodes_.back().get(), 1, input_scope_idxes[i],
"in_var" + i, place_list_[input_scope_idxes[i]]); "in_var" + std::to_string(i), place_list_[input_scope_idxes[i]]);
vars_.emplace_back(in_var_handle); vars_.emplace_back(in_var_handle);
op_handle_->AddInput(in_var_handle); op_handle_->AddInput(in_var_handle);
// add output var handle // add output var handle
for (size_t j = 0; j < place_list_.size(); ++j) { for (size_t j = 0; j < place_list_.size(); ++j) {
nodes_.emplace_back( nodes_.emplace_back(ir::CreateNodeForTest(
ir::CreateNodeForTest("out_node" + i, ir::Node::Type::kVariable)); "out_node" + std::to_string(i), ir::Node::Type::kVariable));
VarHandle* out_var_handle = new VarHandle( VarHandle* out_var_handle =
nodes_.back().get(), 2, j, "out_var" + i, place_list_[j]); new VarHandle(nodes_.back().get(), 2, j,
"out_var" + std::to_string(i), place_list_[j]);
vars_.emplace_back(out_var_handle); vars_.emplace_back(out_var_handle);
op_handle_->AddOutput(out_var_handle); op_handle_->AddOutput(out_var_handle);
} }
...@@ -86,7 +87,7 @@ struct TestFusedBroadcastOpHandle : TestBroadcastOpHandle { ...@@ -86,7 +87,7 @@ struct TestFusedBroadcastOpHandle : TestBroadcastOpHandle {
std::vector<std::vector<float>> send_vec; std::vector<std::vector<float>> send_vec;
f::LoD lod{{0, 10, 20}}; f::LoD lod{{0, 10, 20}};
for (size_t i = 0; i < input_scope_idxes.size(); ++i) { for (size_t i = 0; i < input_scope_idxes.size(); ++i) {
const std::string varname("in_var" + i); const std::string varname("in_var" + std::to_string(i));
float val_scalar = static_cast<float>(i); float val_scalar = static_cast<float>(i);
send_vec.push_back( send_vec.push_back(
InitLoDTensor(varname, input_scope_idxes[i], lod, val_scalar)); InitLoDTensor(varname, input_scope_idxes[i], lod, val_scalar));
...@@ -96,7 +97,7 @@ struct TestFusedBroadcastOpHandle : TestBroadcastOpHandle { ...@@ -96,7 +97,7 @@ struct TestFusedBroadcastOpHandle : TestBroadcastOpHandle {
WaitAll(); WaitAll();
for (size_t i = 0; i < input_scope_idxes.size(); ++i) { for (size_t i = 0; i < input_scope_idxes.size(); ++i) {
const std::string& varname("out_var" + i); const std::string& varname("out_var" + std::to_string(i));
for (size_t j = 0; j < place_list_.size(); ++j) { for (size_t j = 0; j < place_list_.size(); ++j) {
LoDTensorEqual(varname, send_vec[i], lod, param_scopes_[j]); LoDTensorEqual(varname, send_vec[i], lod, param_scopes_[j]);
} }
...@@ -109,7 +110,7 @@ struct TestFusedBroadcastOpHandle : TestBroadcastOpHandle { ...@@ -109,7 +110,7 @@ struct TestFusedBroadcastOpHandle : TestBroadcastOpHandle {
2, 4, 6, 3, 1, 1, 1, 1, 3, 7}; 2, 4, 6, 3, 1, 1, 1, 1, 3, 7};
int height = static_cast<int>(kDims[0] * 2); int height = static_cast<int>(kDims[0] * 2);
for (size_t i = 0; i < input_scope_idxes.size(); ++i) { for (size_t i = 0; i < input_scope_idxes.size(); ++i) {
const std::string varname("in_var" + i); const std::string varname("in_var" + std::to_string(i));
float val_scalar = static_cast<float>(i); float val_scalar = static_cast<float>(i);
send_vector.push_back(InitSelectedRows(varname, input_scope_idxes[i], send_vector.push_back(InitSelectedRows(varname, input_scope_idxes[i],
rows, height, val_scalar)); rows, height, val_scalar));
...@@ -119,7 +120,7 @@ struct TestFusedBroadcastOpHandle : TestBroadcastOpHandle { ...@@ -119,7 +120,7 @@ struct TestFusedBroadcastOpHandle : TestBroadcastOpHandle {
WaitAll(); WaitAll();
for (size_t i = 0; i < input_scope_idxes.size(); ++i) { for (size_t i = 0; i < input_scope_idxes.size(); ++i) {
const std::string& varname("out_var" + i); const std::string& varname("out_var" + std::to_string(i));
for (size_t j = 0; j < place_list_.size(); ++j) { for (size_t j = 0; j < place_list_.size(); ++j) {
SelectedRowsEqual(varname, input_scope_idxes[i], send_vector[i], rows, SelectedRowsEqual(varname, input_scope_idxes[i], send_vector[i], rows,
height); height);
......
paddle/fluid/framework/details/graph_test_base.h 0 → 100644
浏览文件 @ 17ada1d2
// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#pragma once
#include <algorithm>
#include <iostream>
#include <iterator>
#include <string>
#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/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 SplitOpMaker : public OpProtoAndCheckerMaker {
public:
void Make() {
AddInput("X", "");
AddOutput("Out", "").AsDuplicable();
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
paddle/fluid/framework/details/inplace_op_pass.cc 0 → 100644
浏览文件 @ 17ada1d2
// 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/inplace_op_pass.h"
#include <algorithm>
#include <deque>
#include <iterator>
#include <stack>
#include <string>
#include <unordered_map>
#include <unordered_set>
#include <vector>
#include "paddle/fluid/framework/details/memory_optimize_pass.h"
#include "paddle/fluid/framework/ir/graph_helper.h"
#include "paddle/fluid/framework/op_info.h"
// NOTE(dzhwinter): inplace means one op output variable reuse the input space.
// By our design, one operator only can read its input(const Variable),
// write its output(non-const Variable). If one operator is inplaced, means
// user have chance to write the space before reading happens.
// Especially when some optimize code writing style is applied.
//
//
// /* wrong case in operator */
// /*In this case, a larger allocation is allocated, input content is lost*/
// const Tensor* in = ctx.Input<Tensor>("In")
// Tensor* out = ctx.Output<Tensor>("Out");
// auto* out_ptr = out->mutable_data<T>(ctx.GetPlace());
// out_ptr[0] = 0; // input contect is overwrited.
// NOTE(dzhwinter):
// Only for backward compacity and stable. if enable_inplace_whitelist is turn
// on.
// only the ops in whitelist will be use inplace strategy.
// if not, all the op will be inplaced if it registered with InplaceClass
DEFINE_bool(
enable_inplace_whitelist, false,
"If this option turns on, only these op in whitelist can be inplaced."
"If it turns off, all of the running op can be candidate of inplaced op."
"Such as scale, elementwise_add"
"By default, it's turned on");
DECLARE_string(memory_optimize_debug);
// clang-format off
const std::string kInplacedOpWhiteList[] = { // NOLINT
"sigmoid",
"exp",
"relu",
"tanh",
"sqrt",
"ceil",
"floor",
"reciprocal",
"relu6",
"soft_relu",
"hard_sigmoid",
"batch_norm",
"batch_norm_grad",
"sum",
"sum_grad",
"scale",
"reshape",
"elementwise_add",
"elementwise_add_grad",
};
// clang-format on
namespace paddle {
namespace framework {
namespace details {
static inline ir::Node* GetNextCascadeInplacedVar(ir::Node* var) {
// if next op is inplaced, then return the output var
// otherwise return nullptr
PADDLE_ENFORCE(var && var->IsVar() && !var->IsCtrlVar());
ir::Node* inplaced_var = nullptr;
for (auto* next_op : var->outputs) {
for (auto* output : next_op->outputs) {
if (output->IsVar() && !output->IsCtrlVar() &&
output->Name() == var->Name()) {
inplaced_var = output;
}
}
}
return inplaced_var;
}
static inline ir::Node* GetPrevCascadeInplacedVar(ir::Node* var) {
PADDLE_ENFORCE(var && var->IsVar() && !var->IsCtrlVar());
if (var->inputs.empty()) return nullptr;
auto* prev_op = var->inputs.at(0);
auto input_it = std::find_if(prev_op->inputs.begin(), prev_op->inputs.end(),
[&](ir::Node* node) {
if (node->IsVar() && !node->IsCtrlVar() &&
node->Name() == var->Name()) {
return true;
} else {
return false;
}
});
return input_it == prev_op->inputs.end() ? nullptr : *input_it;
}
InplacePass::InplacePass() : Pass() {
if (FLAGS_enable_inplace_whitelist) {
for (auto& s : kInplacedOpWhiteList) {
whitelist_.emplace(s);
}
}
}
void InplacePass::InitSSAGraphNodes() const {
std::unordered_map<std::string, std::unordered_set<ir::Node*>> all_vars;
for (auto* op : view_.AllOps()) {
for (auto* node : op->inputs) {
if (!node->IsVar() || node->IsCtrlVar()) continue;
if (all_vars[node->Name()].count(node) == 0) {
all_vars[node->Name()].emplace(node);
var_nodes_[node->Name()].emplace_back(node);
}
}
for (auto* node : op->outputs) {
if (!node->IsVar() || node->IsCtrlVar()) continue;
if (all_vars[node->Name()].count(node) == 0) {
all_vars[node->Name()].emplace(node);
var_nodes_[node->Name()].emplace_back(node);
}
}
}
}
std::unique_ptr<ir::Graph> InplacePass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
var_nodes_.clear();
view_.Build(graph.get());
InitSSAGraphNodes();
for (auto* op : view_.AllOps()) {
if (FLAGS_enable_inplace_whitelist && !whitelist_.count(op->Name()))
continue;
TryInplaceOpInputOutput(op, graph.get());
}
graph->ResolveHazard(var_nodes_);
return graph;
}
void InplacePass::InplaceModifyDesc(const std::string& var,
const std::string& cache_var,
const size_t& idx) const {
for (size_t i = idx; i < view_.AllOps().size(); ++i) {
ir::Node* op = view_.AllOps()[i];
PADDLE_ENFORCE(op->IsOp() && op->Op());
auto* op_desc = op->Op();
op_desc->RenameInput(var, cache_var);
op_desc->RenameOutput(var, cache_var);
if (op_desc->Block()->HasVar(var)) op_desc->Block()->RemoveVar(var);
op_desc->Flush();
}
}
const NodeSwapQueue InplacePass::TryInplaceModifyVar(
const std::string& var, const std::string& cache_var, const size_t& idx,
ir::Graph* graph) const {
PADDLE_ENFORCE(var_nodes_[var].size() >= 1 &&
var_nodes_[var].at(0)->Var() != nullptr);
std::unique_ptr<VarDesc> var_desc(new VarDesc(*var_nodes_[var].at(0)->Var()));
var_desc->SetName(cache_var);
NodeSwapQueue swap_nodes;
for (size_t i = idx; i < view_.AllOps().size(); ++i) {
auto* op = view_.AllOps()[i];
// redirect the input to the latest version of cache_var
for (auto* node : op->inputs) {
if (node->Name() == var) {
ir::Node* cache_node = graph->CreateVarNode(var_desc.get());
// swap node to cache_node
cache_node->outputs.insert(cache_node->outputs.end(),
node->outputs.begin(), node->outputs.end());
PADDLE_ENFORCE(node->inputs.size() == 1 && node->inputs[0]->IsOp());
auto* prev_op = node->inputs[0];
std::replace(prev_op->outputs.begin(), prev_op->outputs.end(), node,
cache_node);
cache_node->inputs.emplace_back(prev_op);
for (auto* next_op : node->outputs) {
std::replace(next_op->inputs.begin(), next_op->inputs.end(), node,
cache_node);
}
swap_nodes.emplace_back(std::make_pair(node, cache_node));
}
}
// if we need to rename the output,
// always create a newer version of cache_var
for (auto* node : op->outputs) {
if (node->Name() == var) {
ir::Node* cache_node = graph->CreateVarNode(var_desc.get());
// swap node to cache node
cache_node->outputs.insert(cache_node->outputs.end(),
node->outputs.begin(), node->outputs.end());
cache_node->inputs.emplace_back(op);
std::replace(op->outputs.begin(), op->outputs.end(), node, cache_node);
for (auto* next_op : node->outputs) {
std::replace(next_op->inputs.begin(), next_op->inputs.end(), node,
cache_node);
}
swap_nodes.emplace_back(std::make_pair(node, cache_node));
}
}
}
return swap_nodes;
}
void InplacePass::CommitModify(const NodeSwapQueue& swap_nodes,
ir::Graph* graph) const {
for (auto& pair : swap_nodes) {
auto *node = pair.first, *cache_node = pair.second;
const std::string var = node->Name(), cache_var = cache_node->Name();
var_nodes_[cache_var].emplace_back(cache_node);
graph->RemoveNode(node);
auto& nodes = var_nodes_.at(var);
// release unused var in graph. Because python side memory optimize
// may reused the var in same name, so we only clear the var node
// after current inplaced index.
nodes.erase(std::remove(nodes.begin(), nodes.end(), node), nodes.end());
}
}
void InplacePass::WithdrawModify(const NodeSwapQueue& nodes,
ir::Graph* graph) const {
for (auto& pair : nodes) {
auto *node = pair.first, *cache_node = pair.second;
const std::string var = node->Name(), cache_var = cache_node->Name();
auto* prev_op = node->inputs[0];
std::replace(prev_op->outputs.begin(), prev_op->outputs.end(), cache_node,
node);
for (auto* next_op : node->outputs) {
std::replace(next_op->inputs.begin(), next_op->inputs.end(), cache_node,
node);
}
graph->RemoveNode(cache_node);
}
}
void InplacePass::TryInplaceOpInputOutput(ir::Node* op,
ir::Graph* graph) const {
VLOG(4) << "Try to inplace op " << op->Name();
PADDLE_ENFORCE(op->Op() != nullptr && op->Op()->Block() != nullptr,
"op_desc is nullptr");
// some pre-requirments need to meet if the op want to inplaced.
auto* op_desc = op->Op();
auto& infer_inplace =
OpInfoMap::Instance().Get(op_desc->Type()).infer_inplace_;
// 1. infer_inplace_ is registered.
if (!static_cast<bool>(infer_inplace)) return;
PADDLE_ENFORCE(static_cast<bool>(infer_inplace),
"%s's infer_inplace has not been registered", op_desc->Type());
auto* block = op_desc->Block();
auto in_to_outs = infer_inplace(*op_desc, block);
auto& all_ops = view_.AllOps();
auto cursor = std::find(all_ops.begin(), all_ops.end(), op);
size_t idx = std::distance(all_ops.begin(), cursor);
for (auto& pair : in_to_outs) {
auto& in_var_name = pair.first;
auto& out_var_name = pair.second;
auto* in_node = view_.GetNodeByName(in_var_name, op->inputs);
auto* out_node = view_.GetNodeByName(out_var_name, op->outputs);
// 2. there is no external pending op on the input node
if (view_.PendingOpsOnVar(in_node).size() > 1) {
VLOG(4) << string::Sprintf(
"Skiped pair %s => %s. %s input has external dependency."
"inplace such pair will overwrite the memory.",
out_var_name, in_var_name, op->Name());
continue;
}
// 3. if output has been memory optimize by python(fluid.memory_optmize()).
// this candidate can not be inplaced. Will be deprecated in the future.
if (view_.InSkipSet(out_node->Name())) {
VLOG(4) << string::Sprintf(
"Skiped %s => %s reused previous memory block in python memory "
"optmize,"
"it inplace may generate a circle",
out_var_name, in_var_name, op->Name());
continue;
}
// Debug Interface. Which would be skipped by the pass.
if (out_node->Name() == FLAGS_memory_optimize_debug) {
VLOG(3) << "Skiped var by force. FLAGS_memory_optimize_debug="
<< out_node->Name();
continue;
}
// NOTE(dzhwinter):
// two stage commit of inplaced process. if after inplace happens generate a
// circle,
// then withdraw the changes. Otherwise, safely add the node.
auto swap_nodes =
TryInplaceModifyVar(out_var_name, in_var_name, idx, graph);
if (!ir::HasCircle(*graph)) {
VLOG(3) << string::Sprintf("!!! %s, %s => %s inplaced", op->Name(),
out_var_name, in_var_name);
InplaceModifyDesc(out_var_name, in_var_name, idx);
CommitModify(swap_nodes, graph);
} else {
VLOG(3) << string::Sprintf(
"Skiped pair %s => %s, inplace will generate a circle. withdraw %s",
out_var_name, in_var_name, op->Name());
WithdrawModify(swap_nodes, graph);
}
}
}
ir::Node* GraphView::GetNodeByName(const std::string& name,
const std::vector<ir::Node*>& nodes) const {
// nodes should be op->inputs/outputs
// node in same node do have different name.
std::unordered_set<std::string> nodes_in_op;
bool has_dup_node =
std::all_of(nodes.begin(), nodes.end(), [&nodes_in_op](ir::Node* node) {
if (!node->IsVar() || node->IsCtrlVar() || node->Var() == nullptr) {
if (nodes_in_op.count(node->Name())) return true;
nodes_in_op.emplace(node->Name());
}
return false;
});
PADDLE_ENFORCE(has_dup_node == false, "nodes has same name!");
ir::Node* node = nullptr;
for (auto* it : nodes) {
if (!it->IsVar() || it->IsCtrlVar() || it->Var() == nullptr) continue;
if (it->Name() == name) {
node = it;
break;
}
}
PADDLE_ENFORCE(node != nullptr,
string::Sprintf("Not found var %s in nodes!", name));
return node;
}
std::vector<ir::Node*> GraphView::PendingOpsOnVar(ir::Node* node) {
// get the pending ops depends on same var node.
// because node also maybe a inplaced variable, so need to backtrack all the
// previous inplaced vars.
std::vector<ir::Node*> pending_ops;
ir::Node* p = node;
while (p != nullptr) {
pending_ops.insert(pending_ops.end(), p->outputs.begin(), p->outputs.end());
p = GetPrevCascadeInplacedVar(p);
}
return pending_ops;
}
void GraphView::Build(ir::Graph* g) {
// track the var nodes in correct order.
// Because we insert some new created node. Which may have data race between
// nodes.
// resolve data harzards depends on the var nodes in right order.
ops_ = SortOpLikeDescOrder(*g);
// 1. track the nodes which reused previous node in Python memory optimize.
// these node can not be inplaced, otherwise may generate a circle in graph.
std::unordered_set<std::string> all_vars;
for (auto& node : g->Nodes()) {
if (node->IsVar()) continue;
for (auto& out : node->outputs) {
if (out->IsCtrlVar() || out->Var() == nullptr) continue;
if (all_vars.count(out->Name())) {
dup_nodes_.emplace(out->Name());
} else {
all_vars.emplace(out->Name());
}
}
}
// 2. track the nodes which used by parameter server.
// these node can not be inplaced, otherwise trainer
// pserver can not find each other name.
auto update_skip_set = [&](ir::Node* node) {
for (auto& in : node->inputs) {
if (in->IsVar() && in->Var() != nullptr) dup_nodes_.emplace(in->Name());
}
for (auto& out : node->outputs) {
if (out->IsVar() && out->Var() != nullptr)
dup_nodes_.emplace(out->Name());
}
};
for (auto& node : g->Nodes()) {
if (!node->IsOp()) continue;
if (node->Name() == "send") update_skip_set(node);
if (node->Name() == "recv") update_skip_set(node);
if (node->Name() == "prefetch") update_skip_set(node);
}
}
const std::vector<ir::Node*>& GraphView::AllOps() { return ops_; }
bool GraphView::InSkipSet(const std::string& var) const {
return dup_nodes_.count(var);
}
} // namespace details
} // namespace framework
} // namespace paddle
REGISTER_PASS(inplace_pass, paddle::framework::details::InplacePass);
paddle/fluid/framework/details/inplace_op_pass.h 0 → 100644
浏览文件 @ 17ada1d2
// 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 abtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#pragma once
#include <map>
#include <string>
#include <unordered_map>
#include <unordered_set>
#include <utility>
#include <vector>
#include "paddle/fluid/framework/details/memory_optimize_helper.h"
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/pass.h"
namespace paddle {
namespace framework {
namespace details {
class GraphView {
public:
GraphView() = default;
void Build(ir::Graph* g);
const std::vector<ir::Node*>& AllOps();
ir::Node* GetNodeByName(const std::string& name,
const std::vector<ir::Node*>& nodes) const;
std::vector<ir::Node*> PendingOpsOnVar(ir::Node* var);
// Will Deperated in the future.
// NOTE(dzhwinter) :
// 1. Python memory optimize will reuse
// memory based var name, so different op output may
// have the same variable name. enable inplace on such node
// will generate a circle in ssa graph.
// 2. DistributeTranspiler will use unique name to
// map the parameter and gradient, must be skipped.
bool InSkipSet(const std::string& var) const;
private:
std::vector<ir::Node*> ops_;
std::unordered_set<std::string> dup_nodes_; // mem opt affect nodes
std::map<ir::Node*, std::unordered_set<ir::Node*>> adj_list_;
};
// swap pairs in sequence
typedef std::vector<std::pair<ir::Node*, ir::Node*>> NodeSwapQueue;
class InplacePass : public ir::Pass {
public:
InplacePass();
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
void InitSSAGraphNodes() const;
private:
const NodeSwapQueue TryInplaceModifyVar(const std::string& var,
const std::string& cache_var,
const size_t& idx,
ir::Graph* graph) const;
void CommitModify(const NodeSwapQueue&, ir::Graph* graph) const;
void WithdrawModify(const NodeSwapQueue& nodes, ir::Graph* graph) const;
void InplaceModifyDesc(const std::string& in_var, const std::string& out_var,
const size_t& idx) const;
void TryInplaceOpInputOutput(ir::Node* op, ir::Graph* graph) const;
mutable std::map<std::string, std::vector<ir::Node*>> var_nodes_;
mutable std::unordered_set<std::string> whitelist_;
mutable GraphView view_;
};
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/memory_early_delete_pass.cc 已删除 100644 → 0
浏览文件 @ d8291fa3
// 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_reuse_types.h → paddle/fluid/framework/details/memory_optimize_helper.h
浏览文件 @ 17ada1d2
...@@ -17,6 +17,8 @@ ...@@ -17,6 +17,8 @@
#include <iostream> #include <iostream>
#include <iterator> #include <iterator>
#include <list> #include <list>
#include <map>
#include <set>
#include <string> #include <string>
#include <utility> #include <utility>
#include <vector> #include <vector>
...@@ -27,37 +29,41 @@ namespace paddle { ...@@ -27,37 +29,41 @@ namespace paddle {
namespace framework { namespace framework {
namespace details { namespace details {
constexpr char kFetchedVars[] = "fetched_vars"; constexpr char kAllOpDescs[] = "all_op_descs";
constexpr char kGraphNodePool[] = "graph_node_pool";
// NOTE(dzh): Variable and the operators use the var. std::vector<ir::Node*> SortOpLikeDescOrder(const ir::Graph& graph);
// 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). // NOTE(dzh): A ordered set for node reuse in memory optimize.
// in fluid, -1 means the batch_size is determined in runtime. // the orderedset sort node in ascend order(by node bytes size).
// the node batch_size equal -1 always ranking in the front than the node not. // in fluid, -1 means the batch_size, which is determined in runtime.
// So the reuse happens between nodes who's batch_size both are -1
// simultaneously or not.
//
// sort rule:
// rule 0 : smaller node ranking in front.
// rule 1 : batch_size equal -1 ranking in the front than the node not.
//
// For example, // For example,
// node0[-1, 1] node1[-1, 1, 1], node2[1,1], node3[1,1024], .. // 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); class OrderedSet {
public:
// nodes with same name exists in pool.
using NodeVector = std::vector<ir::Node*>;
using Iter = typename std::list<NodeVector>::iterator;
using ConstIter = typename std::list<NodeVector>::const_iterator;
void Insert(ir::Node* var);
void Erase(ir::Node* var); void Erase(ir::Node* var);
bool Has(ir::Node* var) const;
bool Has(ir::Node* var) { return mark_table_.count(var->Name()); } void Clear() {
mark_table_.clear();
ir::Node* NodeMatch(ir::Node* var) const; nodes_.clear();
}
// find the bestfit shape node block with var.
ir::Node* FindBestFitNode(ir::Node* var) const;
// map store non-const iterator, can not promise const // map store non-const iterator, can not promise const
int GetIndex(ir::Node* var); int GetNodeIndexInPool(ir::Node* var);
// pool all node to string // pool all node to string
std::string ToString() const; std::string ToString() const;
...@@ -65,23 +71,112 @@ class OrderedNodePairPool { ...@@ -65,23 +71,112 @@ class OrderedNodePairPool {
Iter end() { return nodes_.end(); } Iter end() { return nodes_.end(); }
ConstIter begin() const { return nodes_.begin(); } ConstIter begin() const { return nodes_.begin(); }
ConstIter end() const { return nodes_.end(); } ConstIter end() const { return nodes_.end(); }
size_t size() const { return nodes_.size(); } size_t size() const { return nodes_.size(); }
private: private:
// for searching. // for searching.
std::unordered_map<std::string, Iter> mark_table_; std::unordered_map<std::string, Iter> mark_table_;
// node swap pairs. var -> ops dep var // node pool
std::list<NodePair> nodes_; std::list<NodeVector> nodes_;
}; };
class ControlFlowGraph {
public:
ControlFlowGraph() = default;
// IR Graph
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* GetNodeByName(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
};
// valid a tensor can be reuse or not
bool NodeCanReused(ir::Node* node);
// valid a tensor can be reuse or not.
bool NodeCanReused(const VarDesc& node);
// check op has subblock or not
bool OpHasSubBlock(OpDesc* desc);
// node memory size in bytes
size_t NodeSize(ir::Node* n);
// node memory size in bytes // node memory size in bytes
size_t NodeSizeInBytes(ir::Node* n); size_t NodeSize(const VarDesc&);
std::string DebugString(ir::Node* var); std::string DebugString(ir::Node* var);
// std::string DebugString(VarDesc* var); // NOTE(dzhwinter)
// after node reuse, the replaced node shape is
// different with its VarDesc. So need to find the
// correct VarDesc in Block.
VarDesc* FindVarDescInBlock(ir::Node* n); VarDesc* FindVarDescInBlock(ir::Node* n);
static inline bool IsSameDesc(OpDesc* op1, OpDesc* op2) {
return op1->Type() == op2->Type() && op1->Inputs() == op2->Inputs() &&
op1->Outputs() == op2->Outputs();
}
template <typename Container, typename Callback>
class FilterVariableImpl {
public:
void operator()(const Container& nodes, Callback callback) {
for (auto* node : nodes) {
callback(node);
}
}
};
// filter var node for op->inputs/outputs
template <typename Callback>
class FilterVariableImpl<std::vector<ir::Node*>, Callback> {
public:
void operator()(const std::vector<ir::Node*>& nodes, Callback callback) {
for (auto* var : nodes) {
if (var->IsVar() && !var->IsCtrlVar()) {
callback(var);
}
}
}
};
template <typename Container, typename Callback>
void FilterVariables(const Container& nodes, Callback callback) {
FilterVariableImpl<Container, Callback>()(nodes, callback);
}
} // namespace details } // namespace details
} // namespace framework } // namespace framework
} // namespace paddle } // namespace paddle
paddle/fluid/framework/details/analysis_var_pass_test.cc → paddle/fluid/framework/details/memory_optimize_helper_test.cc
浏览文件 @ 17ada1d2
...@@ -12,12 +12,18 @@ ...@@ -12,12 +12,18 @@
// See the License for the specific language governing permissions and // See the License for the specific language governing permissions and
// limitations under the License. // limitations under the License.
#include "paddle/fluid/framework/details/analysis_var_pass.h" #include "paddle/fluid/framework/details/memory_optimize_helper.h"
#include <algorithm> #include <algorithm>
#include <iostream> #include <iostream>
#include <iterator> #include <iterator>
#include <memory>
#include <sstream>
#include <string>
#include <utility>
#include <vector>
#include "glog/logging.h" #include "glog/logging.h"
#include "gtest/gtest.h" #include "gtest/gtest.h"
#include "paddle/fluid/framework/details/graph_test_base.h"
#include "paddle/fluid/framework/ir/graph.h" #include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/graph_helper.h" #include "paddle/fluid/framework/ir/graph_helper.h"
#include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/op_registry.h"
...@@ -26,46 +32,82 @@ ...@@ -26,46 +32,82 @@
namespace paddle { namespace paddle {
namespace framework { namespace framework {
namespace details {
class DummyOp : public OperatorBase { TEST(OrderedSet, Normal) {
public: OrderedSet pool;
DummyOp(const std::string& type, const VariableNameMap& inputs, std::vector<std::unique_ptr<ir::Node>> nodes;
const VariableNameMap& outputs, const AttributeMap& attrs)
: OperatorBase(type, inputs, outputs, attrs) {} // clang-format off
std::vector<std::vector<int64_t>> shapes = {{-1, 10},
private: {-1, 20},
void RunImpl(const Scope& scope, {1, 2},
const platform::Place& place) const override {} {5, 2},
}; {10, 20},
{-1, 2, 5},
class SumOpMaker : public OpProtoAndCheckerMaker { {-1, 1, 5},
public: {-1, 1}};
void Make() { // clang-format on
AddInput("X", "").AsDuplicable(); const int COUNT = shapes.size();
AddOutput("Out", ""); ProgramDesc prog;
AddComment(""); 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));
} }
};
// Insert
class AssignOpMaker : public OpProtoAndCheckerMaker { for (auto& node : nodes) {
public: pool.Insert(node.get());
void Make() {
AddInput("X", "").AsDuplicable();
AddOutput("Out", "");
AddComment("");
} }
};
// Has/size
class DummyVarTypeInference : public VarTypeInference { ASSERT_EQ(pool.size(), shapes.size());
public: for (auto& node : nodes) {
void operator()(const OpDesc& op_desc, BlockDesc* block) const override { ASSERT_TRUE(pool.Has(node.get()));
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);
} }
};
// 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.GetNodeIndexInPool(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.FindBestFitNode(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.FindBestFitNode(node1.get());
ASSERT_EQ(pool.GetNodeIndexInPool(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.FindBestFitNode(node1.get());
ASSERT_EQ(pool.GetNodeIndexInPool(cache), 5); // match 4:[5,2]
}
}
} // namespace details
} // namespace framework } // namespace framework
} // namespace paddle } // namespace paddle
...@@ -102,11 +144,6 @@ namespace paddle { ...@@ -102,11 +144,6 @@ namespace paddle {
namespace framework { namespace framework {
namespace details { 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() { inline static ProgramDesc FillProgramDesc() {
ProgramDesc prog; ProgramDesc prog;
prog.MutableBlock(0)->Var("a")->SetType(proto::VarType::LOD_TENSOR); prog.MutableBlock(0)->Var("a")->SetType(proto::VarType::LOD_TENSOR);
...@@ -141,15 +178,6 @@ inline static ProgramDesc FillProgramDesc() { ...@@ -141,15 +178,6 @@ inline static ProgramDesc FillProgramDesc() {
return prog; 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) { TEST(CFGGraph, IRGraph) {
// prepare ir graph // prepare ir graph
auto prog = FillProgramDesc(); auto prog = FillProgramDesc();
......
paddle/fluid/framework/details/memory_optimize_pass.cc 0 → 100644
浏览文件 @ 17ada1d2
// 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_optimize_pass.h"
#include <algorithm>
#include <atomic>
#include <deque>
#include <fstream>
#include <iostream>
#include <iterator>
#include <memory>
#include <queue>
#include <sstream>
#include <string>
#include <type_traits>
#include <vector>
#include "gflags/gflags.h"
#include "paddle/fluid/framework/data_type.h"
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/graph_helper.h"
DEFINE_bool(enable_subgraph_optimize, false,
"SubGraph also reuse global graph variables, it will reduce the "
"memory occupation"
"but a higher risk of memory reuse error. default disabled.");
DEFINE_string(memory_optimize_debug, "",
"debug the operator output variable when do the variable reuse."
"memory reuse pass."
"only for debug, default disabled.");
namespace paddle {
namespace framework {
namespace details {
std::unique_ptr<ir::Graph> MemoryOptimizePass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
auto nodes = graph->Nodes();
CollectSkipVarsSet(nodes);
cfg_.reset(new details::ControlFlowGraph(*graph));
cfg_->LiveVariableAnalysis();
InitSSAGraphNodes();
int reuse_id = 0;
for (size_t idx = 0; idx < cfg_->Ops().size(); ++idx) {
auto& op = cfg_->Ops()[idx];
auto* op_desc = op->Op();
// some op in graph has no op desc
if (op_desc == nullptr) continue;
if (OpHasSubBlock(op_desc)) {
if (FLAGS_enable_subgraph_optimize) {
SubGraphOptimize(op_desc);
} else {
VLOG(3) << op->Name()
<< " has subblock, but disable subgraph optimize. skipped.";
continue;
}
}
for (auto& var : op->outputs) {
if (!NodeCanReused(var) || cfg_->Use(op).count(var->Name()) == 0 ||
skip_set_.count(var->Name()))
continue;
ir::Node* cache = pool_.FindBestFitNode(var);
if (var->Name() == FLAGS_memory_optimize_debug) {
VLOG(3) << "start match var " << DebugString(var) << " of op "
<< op->Name();
VLOG(3) << pool_.ToString();
VLOG(3) << "matched in pool : "
<< ((cache == nullptr) ? "False" : "True");
}
if (cache == nullptr) continue;
if (var->Name() == cache->Name()) {
VLOG(3) << "The same cache variable is cascade reused." << var->Name()
<< " is re-filled to the pool after"
<< "the reused op is finished. Current op can not "
<< "replace it again. Skip this candidate.";
continue;
int node_idx_in_pool = pool_.GetNodeIndexInPool(cache);
VLOG(3) << string::Sprintf(
"!!! %s, %s => %s, cache idx %d, pool size %d",
std::to_string(reuse_id++), DebugString(var), DebugString(cache),
node_idx_in_pool, static_cast<int>(pool_.size()));
// update CFG Graph on the fly.
// reused var maybe re-fill into the pool
cfg_->RenameVarInCFGGraph(var->Name(), cache->Name(), idx);
// NOTE(dzhwinter): we need to both update the ProgramDesc
// and IR Graph. because op_desc/var_desc is used in CreateOp,
// CreateVar when running happens. But IR Graph
// define the dependence relationship between nodes.
RenameVarInGraphDesc(var->Name(), cache->Name(), idx);
RenameVarInGraphNode(var->Name(), cache->Name(), idx, graph.get());
pool_.Erase(cache);
}
// fill the pool
std::unordered_set<std::string> unlived_vars;
for (auto var : cfg_->LiveIn(op)) {
if (cfg_->LiveOut(op).count(var) == 0) {
unlived_vars.emplace(var);
}
}
for (auto var : unlived_vars) {
ir::Node* var_node = cfg_->GetNodeByName(var, op);
if (NodeCanReused(var_node) && !pool_.Has(var_node)) {
pool_.Insert(var_node);
}
}
}
}
graph->ResolveHazard(var_nodes_);
return graph;
}
void MemoryOptimizePass::SubGraphOptimize(OpDesc* op_desc) const {
// conditional block, while op and their grad op
auto* sub_block_desc =
AttrReader(op_desc->GetAttrMap()).Get<BlockDesc*>("sub_block");
// create a mirror block to construct an IR Graph.
ProgramDesc prog;
auto* copy_block = prog.MutableBlock(0);
for (auto* op : sub_block_desc->AllOps()) {
auto* copy_op = copy_block->AppendOp();
copy_op->CopyFrom(*op);
copy_op->Flush();
}
for (auto* var : sub_block_desc->AllVars()) {
auto* copy_var = copy_block->Var(var->Name());
copy_var->SetDataType(var->GetDataType());
// only lod tensor can be reused. So ignore the multiple dims case.
copy_var->SetType(var->GetType());
copy_var->SetShape(var->GetShape());
copy_var->SetPersistable(var->Persistable());
}
ir::Graph sub_graph(prog);
std::unordered_set<ir::Node*> sub_graph_all_ops;
FilterVariables(sub_graph.Nodes(), [&](ir::Node* var) {
// sub_graph_all_ops.emplace(var);
if (var->IsVar() && !var->IsCtrlVar()) {
sub_graph_all_ops.emplace(var);
}
});
int sub_reuse_id = 0;
// subgraph nodes is unordered, reuse need to follow the desc order.
// find the right op node through the descs
for (auto* sub_op_desc : sub_block_desc->AllOps()) {
ir::Node* sub_op = nullptr;
for (auto* node : sub_graph_all_ops) {
if (node->Op() == sub_op_desc) {
sub_op = node;
break;
}
}
PADDLE_ENFORCE(sub_op != nullptr);
for (auto* var : sub_op->outputs) {
if (NodeCanReused(var)) {
ir::Node* cache = pool_.FindBestFitNode(var);
if (cache != nullptr) {
if (var->Var()->GetDataType() != cache->Var()->GetDataType()) {
continue;
}
int node_idx_in_pool = pool_.GetNodeIndexInPool(cache);
VLOG(3) << string::Sprintf(
"!!! %s, %s => %s, cache idx %d, pool size %d",
std::to_string(sub_reuse_id++), DebugString(var),
DebugString(cache), node_idx_in_pool,
static_cast<int>(pool_.size()));
// NOTE(dzh): subblock is not in IR graph. Modify the block_desc
// immediately to make the subblock variable reuse strategy take
// effect. Because it is a single op in graph. No need to
// update the ir nodes.
sub_op_desc->Rename(var->Name(), cache->Name());
if (sub_op_desc->Block()->HasVar(var->Name())) {
sub_op_desc->Block()->RemoveVar(var->Name());
}
}
}
}
}
}
void MemoryOptimizePass::CollectSkipVarsSet(
const std::unordered_set<ir::Node*>& nodes) const {
auto update_skip_set = [&](OpDesc* op_desc) {
auto inputs = op_desc->InputArgumentNames();
auto outputs = op_desc->OutputArgumentNames();
skip_set_.insert(inputs.begin(), inputs.end());
skip_set_.insert(outputs.begin(), outputs.end());
};
for (auto& op : nodes) {
if (!op->IsOp() || op->Op() == nullptr) continue;
auto* op_desc = op->Op();
// NOTE(dzhwinter):
// current block can not reuse next level block vars.
if (OpHasSubBlock(op_desc)) update_skip_set(op_desc);
// NOTE(dzhwinter):
// distributed ops input/output name need to
// keep same bettwen trainer/pserver
if (op_desc->Type() == "send") update_skip_set(op_desc);
if (op_desc->Type() == "recv") update_skip_set(op_desc);
if (op_desc->Type() == "prefetch") update_skip_set(op_desc);
}
}
void MemoryOptimizePass::RenameVarInGraphDesc(const std::string& var,
const std::string& cache_var,
size_t idx) const {
for (size_t i = idx; i < cfg_->Ops().size(); ++i) {
auto* op = cfg_->Ops()[i];
PADDLE_ENFORCE(op->IsOp() && op->Op());
auto* op_desc = op->Op();
op_desc->RenameInput(var, cache_var);
op_desc->RenameOutput(var, cache_var);
if (op_desc->Block()->HasVar(var)) op_desc->Block()->RemoveVar(var);
op_desc->Flush();
}
}
void MemoryOptimizePass::InitSSAGraphNodes() const {
std::unordered_map<std::string, std::unordered_set<ir::Node*>> all_vars;
if (var_nodes_.empty()) {
for (auto* op : cfg_->Ops()) {
for (auto* node : op->inputs) {
if (all_vars[node->Name()].count(node) == 0) {
all_vars[node->Name()].emplace(node);
var_nodes_[node->Name()].emplace_back(node);
}
}
for (auto* node : op->outputs) {
if (all_vars[node->Name()].count(node) == 0) {
all_vars[node->Name()].emplace(node);
var_nodes_[node->Name()].emplace_back(node);
}
}
}
}
}
void MemoryOptimizePass::RenameVarInGraphNode(const std::string& var,
const std::string& cache_var,
size_t idx,
ir::Graph* graph) const {
// if replace happens, we need to create a newer version cache_var
// but use the same dims/data_type with var.
PADDLE_ENFORCE(var_nodes_[var].size() >= 1 &&
var_nodes_[var].at(0)->Var() != nullptr);
std::unique_ptr<VarDesc> var_desc(new VarDesc(*var_nodes_[var].at(0)->Var()));
var_desc->SetName(cache_var);
for (size_t i = idx; i < cfg_->Ops().size(); ++i) {
auto* op = cfg_->Ops()[i];
// redirect the input to the latest version of cache_var
for (auto* node : op->inputs) {
if (node->Name() == var) {
ir::Node* cache_node = graph->CreateVarNode(var_desc.get());
var_nodes_[cache_var].emplace_back(cache_node);
// swap node to cache_node
cache_node->outputs.insert(cache_node->outputs.end(),
node->outputs.begin(), node->outputs.end());
PADDLE_ENFORCE(node->inputs.size() == 1 && node->inputs[0]->IsOp());
auto* prev_op = node->inputs[0];
std::replace(prev_op->outputs.begin(), prev_op->outputs.end(), node,
cache_node);
cache_node->inputs.emplace_back(prev_op);
for (auto* next_op : node->outputs) {
std::replace(next_op->inputs.begin(), next_op->inputs.end(), node,
cache_node);
}
}
}
// if we need to rename the output,
// always create a newer version of cache_var
for (auto* node : op->outputs) {
if (node->Name() == var) {
ir::Node* cache_node = graph->CreateVarNode(var_desc.get());
var_nodes_[cache_var].emplace_back(cache_node);
// swap node to cache node
cache_node->outputs.insert(cache_node->outputs.end(),
node->outputs.begin(), node->outputs.end());
cache_node->inputs.emplace_back(op);
std::replace(op->outputs.begin(), op->outputs.end(), node, cache_node);
for (auto* next_op : node->outputs) {
std::replace(next_op->inputs.begin(), next_op->inputs.end(), node,
cache_node);
}
}
}
}
// release node of unused var in graph
for (auto* node : var_nodes_[var]) {
graph->RemoveNode(node);
}
var_nodes_.at(var).clear();
}
} // namespace details
} // namespace framework
} // namespace paddle
REGISTER_PASS(memory_optimize_pass,
paddle::framework::details::MemoryOptimizePass)
.RequireGraphAttr(paddle::framework::details::kAllOpDescs);
paddle/fluid/framework/details/analysis_var_pass.h → paddle/fluid/framework/details/memory_optimize_pass.h
浏览文件 @ 17ada1d2
...@@ -25,29 +25,22 @@ ...@@ -25,29 +25,22 @@
#include <vector> #include <vector>
#include "paddle/fluid/framework/data_type.h" #include "paddle/fluid/framework/data_type.h"
#include "paddle/fluid/framework/details/memory_reuse_types.h" #include "paddle/fluid/framework/details/memory_optimize_helper.h"
#include "paddle/fluid/framework/ir/graph.h" #include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/pass.h" #include "paddle/fluid/framework/ir/pass.h"
namespace paddle { namespace paddle {
namespace framework { namespace framework {
namespace details { namespace details {
constexpr char kAllOpDescs[] = "all_op_descs";
std::vector<ir::Node*> SortOpLikeDescOrder(const ir::Graph& graph); class MemoryOptimizePass : public ir::Pass {
// sort op in bfs order
std::vector<ir::Node*> BFSSortGraphOps(const ir::Graph& graph);
class ControlFlowGraph;
class AnalysisVarPass : public ir::Pass {
protected: protected:
std::unique_ptr<ir::Graph> ApplyImpl( std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override; std::unique_ptr<ir::Graph> graph) const override;
private:
// fill the variable map(var_nodes) by version. // fill the variable map(var_nodes) by version.
void InitSSAGraphNodes() const; void InitSSAGraphNodes() const;
private:
// update program descs // update program descs
void RenameVarInGraphDesc(const std::string& var, void RenameVarInGraphDesc(const std::string& var,
const std::string& cache_var, size_t idx) const; const std::string& cache_var, size_t idx) const;
...@@ -57,17 +50,14 @@ class AnalysisVarPass : public ir::Pass { ...@@ -57,17 +50,14 @@ class AnalysisVarPass : public ir::Pass {
ir::Graph* graph) const; ir::Graph* graph) const;
void SubGraphOptimize(OpDesc* op_desc) const; void SubGraphOptimize(OpDesc* op_desc) const;
// valid a tensor can be reuse or not // 1. scan op with subblock and collect the output/input vars.
bool NodeCanReused(ir::Node* node) const; // while, while_grad, conditional_block
// scan subblock and collect the output/input variables. // 2. scan distributed ops and collect the output/input vars
std::unordered_set<std::string> GetSubBlockVars( void CollectSkipVarsSet(const std::unordered_set<ir::Node*>&) const;
const std::unordered_set<ir::Node*>&) const;
// check op has subblock or not
bool OpHasSubBlock(OpDesc* desc) const;
private: private:
// Reuse Node Pool, Owned. // Reuse Node Pool, Owned.
mutable OrderedNodePairPool pool_; mutable OrderedSet pool_;
// controlflow Graph // controlflow Graph
mutable std::unique_ptr<ControlFlowGraph> cfg_; mutable std::unique_ptr<ControlFlowGraph> cfg_;
// skip set // skip set
...@@ -76,45 +66,6 @@ class AnalysisVarPass : public ir::Pass { ...@@ -76,45 +66,6 @@ class AnalysisVarPass : public ir::Pass {
mutable std::map<std::string, std::vector<ir::Node*>> 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 details
} // namespace framework } // namespace framework
} // namespace paddle } // namespace paddle
paddle/fluid/framework/details/memory_reuse_types.cc 已删除 100644 → 0
浏览文件 @ d8291fa3
// 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_test.cc 已删除 100644 → 0
浏览文件 @ d8291fa3
// 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/op_registry.h
浏览文件 @ 17ada1d2
...@@ -18,6 +18,7 @@ limitations under the License. */ ...@@ -18,6 +18,7 @@ limitations under the License. */
#include <tuple> #include <tuple>
#include <vector> #include <vector>
#include "paddle/fluid/framework/grad_op_desc_maker.h" #include "paddle/fluid/framework/grad_op_desc_maker.h"
#include "paddle/fluid/framework/inplace_op_inference.h"
#include "paddle/fluid/framework/op_info.h" #include "paddle/fluid/framework/op_info.h"
#include "paddle/fluid/framework/op_proto_maker.h" #include "paddle/fluid/framework/op_proto_maker.h"
#include "paddle/fluid/framework/operator.h" #include "paddle/fluid/framework/operator.h"
...@@ -32,7 +33,8 @@ enum OpInfoFillType { ...@@ -32,7 +33,8 @@ enum OpInfoFillType {
kOpProtoAndCheckerMaker = 1, kOpProtoAndCheckerMaker = 1,
kGradOpDescMaker = 2, kGradOpDescMaker = 2,
kVarTypeInference = 3, kVarTypeInference = 3,
kShapeInference = 4 kShapeInference = 4,
kInplaceOpInference = 5
}; };
template <typename T> template <typename T>
...@@ -48,8 +50,11 @@ struct OpInfoFillTypeID { ...@@ -48,8 +50,11 @@ struct OpInfoFillTypeID {
? kVarTypeInference ? kVarTypeInference
: (std::is_base_of<InferShapeBase, T>::value : (std::is_base_of<InferShapeBase, T>::value
? kShapeInference ? kShapeInference
: static_cast<OpInfoFillType>( : (std::is_base_of<
-1))))); InplaceOpInference, T>::value
? kInplaceOpInference
: static_cast<OpInfoFillType>(
-1))))));
} }
}; };
...@@ -139,6 +144,16 @@ struct OpInfoFiller<T, kShapeInference> { ...@@ -139,6 +144,16 @@ struct OpInfoFiller<T, kShapeInference> {
} }
}; };
template <typename T>
struct OpInfoFiller<T, kInplaceOpInference> {
void operator()(const char* op_type, OpInfo* info) const {
info->infer_inplace_ = [](const OpDesc& op_desc, BlockDesc* block) {
T infer;
return infer(op_desc, block);
};
}
};
} // namespace details } // namespace details
} // namespace framework } // namespace framework
......
paddle/fluid/framework/details/parallel_ssa_graph_executor.cc
浏览文件 @ 17ada1d2
...@@ -65,7 +65,7 @@ FeedFetchList ParallelSSAGraphExecutor::Run( ...@@ -65,7 +65,7 @@ FeedFetchList ParallelSSAGraphExecutor::Run(
if (pool_) { if (pool_) {
run_futures.emplace_back(pool_->enqueue(std::move(call))); run_futures.emplace_back(pool_->enqueue(std::move(call)));
} else { } else {
fetch_data.emplace_back(std::move(call())); fetch_data.emplace_back(call());
} }
} }
...@@ -74,7 +74,7 @@ FeedFetchList ParallelSSAGraphExecutor::Run( ...@@ -74,7 +74,7 @@ FeedFetchList ParallelSSAGraphExecutor::Run(
if (exception_holder_.IsCaught()) { if (exception_holder_.IsCaught()) {
f.wait(); f.wait();
} else { } else {
fetch_data.emplace_back(std::move(f.get())); fetch_data.emplace_back(f.get());
} }
} }
} }
......
paddle/fluid/framework/details/sequential_execution_pass.cc
浏览文件 @ 17ada1d2
...@@ -17,6 +17,7 @@ ...@@ -17,6 +17,7 @@
#include <unordered_map> #include <unordered_map>
#include <unordered_set> #include <unordered_set>
#include <vector> #include <vector>
#include "paddle/fluid/framework/details/memory_optimize_helper.h"
#include "paddle/fluid/framework/op_proto_maker.h" #include "paddle/fluid/framework/op_proto_maker.h"
namespace paddle { namespace paddle {
......
paddle/fluid/framework/details/sequential_execution_pass.h
浏览文件 @ 17ada1d2
...@@ -21,8 +21,6 @@ namespace paddle { ...@@ -21,8 +21,6 @@ namespace paddle {
namespace framework { namespace framework {
namespace details { namespace details {
constexpr char kAllOpDescs[] = "all_op_descs";
class SequentialExecutionPass : public ir::Pass { class SequentialExecutionPass : public ir::Pass {
protected: protected:
std::unique_ptr<ir::Graph> ApplyImpl( std::unique_ptr<ir::Graph> ApplyImpl(
......
paddle/fluid/framework/feed_fetch_method.cc
浏览文件 @ 17ada1d2
...@@ -44,6 +44,7 @@ LoDTensor& GetFetchVariable(const Scope& scope, const std::string& var_name, ...@@ -44,6 +44,7 @@ LoDTensor& GetFetchVariable(const Scope& scope, const std::string& var_name,
// Since we want to fetch LodTensor from a variable, the variable must // Since we want to fetch LodTensor from a variable, the variable must
// be created alreadly. // be created alreadly.
Variable* g_fetch_value = scope.FindVar(var_name); Variable* g_fetch_value = scope.FindVar(var_name);
PADDLE_ENFORCE_NOT_NULL(g_fetch_value, "%s is not found.", var_name);
PADDLE_ENFORCE(g_fetch_value->IsType<FeedFetchList>(), PADDLE_ENFORCE(g_fetch_value->IsType<FeedFetchList>(),
"Only %s can be invoked by GetFetchVariable", "Only %s can be invoked by GetFetchVariable",
typeid(FeedFetchList).name()); typeid(FeedFetchList).name());
......
paddle/fluid/framework/inplace_op_inference.h 0 → 100644
浏览文件 @ 17ada1d2
// 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 <functional>
#include <numeric>
#include <string>
#include <unordered_map>
#include "glog/logging.h"
#include "paddle/fluid/framework/block_desc.h"
#include "paddle/fluid/framework/details/memory_optimize_helper.h"
#include "paddle/fluid/framework/op_desc.h"
#include "paddle/fluid/framework/type_defs.h"
namespace paddle {
namespace framework {
/*
Inplace Inference for create In->Out pairs for inplaced operator.
If we specify a pair of corresponding names. For example, X->Out.
then Out will inplaced use X's memory. The base class will do
legality validation for both variables.
*/
class InplaceOpInference {
public:
virtual ~InplaceOpInference() {}
virtual std::unordered_map<std::string, std::string> operator()(
const OpDesc& op_desc, BlockDesc* block) const = 0;
};
class InplaceInToOut : public InplaceOpInference {
public:
std::unordered_map<std::string, std::string> operator()(
const OpDesc& op_desc, BlockDesc* block) const {
std::unordered_map<std::string, std::string> ret;
auto in_out_var_names_pair = this->Apply(op_desc, block);
for (auto& pair : in_out_var_names_pair) {
PADDLE_ENFORCE(!op_desc.Input(pair.first).empty(),
string::Sprintf("op %s do not have input of %s!",
op_desc.Type(), pair.first));
PADDLE_ENFORCE(!op_desc.Output(pair.second).empty(),
string::Sprintf("op %s do not have output of %s!",
op_desc.Type(), pair.second));
auto& in_name = op_desc.Input(pair.first).at(0);
auto& out_name = op_desc.Output(pair.second).at(0);
auto in = block->FindRecursiveOrCreateVar(in_name);
auto out = block->FindRecursiveOrCreateVar(out_name);
if (TryInplaceInputOutput(in, out)) ret.insert({in_name, out_name});
}
return ret;
}
protected:
virtual std::unordered_map<std::string, std::string> Apply(
const OpDesc& op_desc, BlockDesc* block) const = 0;
bool TryInplaceInputOutput(const VarDesc& in, const VarDesc& out) const {
return in.Name() != out.Name() && details::NodeCanReused(in) &&
details::NodeCanReused(out) &&
details::NodeSize(out) <= details::NodeSize(in);
}
};
/*
Inplace In and Out for operator only have an Input and an Output.
For example, activation op.
*/
class SingleOpInplaceInToOut : public InplaceInToOut {
protected:
std::unordered_map<std::string, std::string> Apply(
const OpDesc& op_desc, BlockDesc* block) const override {
PADDLE_ENFORCE(!op_desc.InputNames().empty(),
"Op inputs must not be empty");
PADDLE_ENFORCE(!op_desc.OutputNames().empty(),
"Op outputs must not be empty");
auto x_name = op_desc.InputNames().at(0);
auto out_name = op_desc.OutputNames().at(0);
return std::unordered_map<std::string, std::string>{{x_name, out_name}};
}
};
/*
Gradient op. Inplace output use it's Input.
For example, Input@Grad->Input reuse strategy.
*/
class GradOpInplaceInToOut : public InplaceInToOut {
protected:
std::unordered_map<std::string, std::string> Apply(
const OpDesc& op_desc, BlockDesc* block) const override {
std::unordered_map<std::string, std::string> ret;
std::unordered_set<std::string> output_names(op_desc.OutputNames().begin(),
op_desc.OutputNames().end());
for (auto& input_name : op_desc.InputNames()) {
if (output_names.count(GradVarName(input_name))) {
ret.insert({input_name, GradVarName(input_name)});
}
}
return ret;
}
};
} // namespace framework
} // namespace paddle
paddle/fluid/framework/inplace_op_inference_test.cc 0 → 100644
浏览文件 @ 17ada1d2
/* 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 <iterator>
#include <string>
#include "gtest/gtest.h"
#include "paddle/fluid/framework/op_info.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/framework/program_desc.h"
#include "paddle/fluid/framework/var_type_inference.h"
namespace paddle {
namespace framework {
class NOP : public OperatorBase {
public:
NOP(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 SingleOpMaker : public OpProtoAndCheckerMaker {
public:
void Make() {
AddInput("X", "").AsDuplicable();
AddOutput("Out", "");
AddComment("");
}
};
class SingleGradOpMaker : public framework::SingleGradOpDescMaker {
public:
using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
protected:
std::unique_ptr<framework::OpDesc> Apply() const override {
auto* op = new framework::OpDesc();
op->SetType("single_op_grad");
op->SetInput("Out", OutputGrad("Out"));
op->SetOutput(framework::GradVarName("X"), InputGrad("X"));
return std::unique_ptr<OpDesc>(op);
}
};
class SingleOpShapeInference : public framework::InferShapeBase {
public:
void operator()(framework::InferShapeContext* ctx) const override {
ctx->HasInput("X");
ctx->HasOutput("Out");
ctx->SetOutputDim("Out", ctx->GetInputDim("X"));
}
};
class SingleGradOpShapeInference : public framework::InferShapeBase {
public:
void operator()(framework::InferShapeContext* ctx) const override {
ctx->HasInput(framework::GradVarName("Out"));
ctx->HasOutput(framework::GradVarName("X"));
ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("Out"));
}
};
class MultiOutOpMaker : public OpProtoAndCheckerMaker {
public:
void Make() {
AddInput("X", "").AsDuplicable();
AddInput("Y", "").AsDuplicable();
AddInput("Z", "").AsDuplicable();
AddOutput("Out", "");
AddOutput("YOut", "");
AddOutput("ZOut", "");
AddOutput("NotReuseOut", "");
AddComment("");
}
};
class MultiOutShapeInference : public framework::InferShapeBase {
public:
void operator()(framework::InferShapeContext* ctx) const override {
ctx->ShareDim("X", "Out");
ctx->ShareDim("Y", "YOut");
ctx->ShareDim("Z", "ZOut");
}
};
class MultiGradOpMaker : public framework::SingleGradOpDescMaker {
public:
using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
protected:
std::unique_ptr<framework::OpDesc> Apply() const override {
auto* op = new framework::OpDesc();
op->SetType("multi_out_grad");
op->SetInput("X", Input("X"));
op->SetOutput(framework::GradVarName("Y"), OutputGrad("YOut"));
op->SetOutput(framework::GradVarName("X"), OutputGrad("Out"));
op->SetOutput(framework::GradVarName("Z"), OutputGrad("ZOut"));
return std::unique_ptr<framework::OpDesc>(op);
}
};
class MultiOutGradShapeInference : public framework::InferShapeBase {
public:
void operator()(framework::InferShapeContext* ctx) const override {
ctx->SetOutputDim(framework::GradVarName("Y"),
ctx->GetInputDim(framework::GradVarName("YOut")));
ctx->SetOutputDim(framework::GradVarName("X"),
ctx->GetInputDim(framework::GradVarName("Out")));
ctx->SetOutputDim(framework::GradVarName("Z"),
ctx->GetInputDim(framework::GradVarName("ZOut")));
}
};
class MultiOutInplaceInToOut : public framework::InplaceInToOut {
public:
using framework::InplaceInToOut::InplaceInToOut;
protected:
std::unordered_map<std::string, std::string> Apply(
const OpDesc& op_desc, BlockDesc* block) const override {
return std::unordered_map<std::string, std::string>{
{"X", "Out"}, {"Y", "YOut"}, {"Z", "ZOut"},
};
}
};
class MultiOutGradInplaceInToOut : public framework::InplaceInToOut {
public:
using framework::InplaceInToOut::InplaceInToOut;
protected:
std::unordered_map<std::string, std::string> Apply(
const OpDesc& op_desc, BlockDesc* block) const override {
return std::unordered_map<std::string, std::string>{
{framework::GradVarName("YOut"), framework::GradVarName("Y")},
{framework::GradVarName("Out"), framework::GradVarName("X")},
{framework::GradVarName("ZOut"), framework::GradVarName("Z")},
};
}
};
} // namespace framework
} // namespace paddle
namespace f = paddle::framework;
REGISTER_OPERATOR(single_op, f::NOP, f::SingleOpMaker, f::SingleGradOpMaker,
f::SingleOpInplaceInToOut, f::SingleOpShapeInference);
REGISTER_OPERATOR(single_op_grad, f::NOP, f::SingleOpInplaceInToOut,
f::SingleGradOpShapeInference);
REGISTER_OPERATOR(multi_out_op, f::NOP, f::MultiOutOpMaker, f::MultiGradOpMaker,
f::MultiOutInplaceInToOut, f::MultiOutShapeInference);
REGISTER_OPERATOR(multi_out_grad, f::NOP, f::MultiOutGradInplaceInToOut,
f::MultiOutGradShapeInference);
namespace paddle {
namespace framework {
TEST(InferInplace, SingleOpInplaceInToOut) {
ProgramDesc prog;
auto* op = prog.MutableBlock(0)->AppendOp();
op->SetType("single_op");
op->SetInput("X", {"test2_a", "test2_b", "test2_c"});
op->SetOutput("Out", {"test2_out"});
prog.MutableBlock(0)->Var("test2_a")->SetType(proto::VarType::LOD_TENSOR);
prog.MutableBlock(0)->Var("test2_a")->SetShape({32, 64});
prog.MutableBlock(0)->Var("test2_b")->SetType(proto::VarType::LOD_TENSOR);
prog.MutableBlock(0)->Var("test2_c")->SetType(proto::VarType::LOD_TENSOR);
prog.MutableBlock(0)->Var("test2_out");
prog.MutableBlock(0)->Var("test2_out")->SetShape({32, 16});
auto& infer_inplace = OpInfoMap::Instance().Get(op->Type()).infer_inplace_;
auto in_to_outs = infer_inplace(*op, op->Block());
EXPECT_EQ(in_to_outs.size(), 1ul);
auto it = in_to_outs.begin();
EXPECT_EQ(it->first, "test2_a");
EXPECT_EQ(it->second, "test2_out");
}
TEST(InferInplace, SingleGradOpInplaceInToOut) {
ProgramDesc prog;
auto* op = prog.MutableBlock(0)->AppendOp();
op->SetType("single_op_grad");
op->SetInput(GradVarName("Out"), {"test2_out"});
op->SetOutput(GradVarName("X"), {"test2_a", "test2_b", "test2_c"});
prog.MutableBlock(0)->Var("test2_a")->SetType(proto::VarType::LOD_TENSOR);
prog.MutableBlock(0)->Var("test2_a")->SetShape({32, 16});
prog.MutableBlock(0)->Var("test2_b")->SetType(proto::VarType::LOD_TENSOR);
prog.MutableBlock(0)->Var("test2_c")->SetType(proto::VarType::LOD_TENSOR);
prog.MutableBlock(0)->Var("test2_out");
prog.MutableBlock(0)->Var("test2_out")->SetShape({32, 16});
auto& infer_inplace = OpInfoMap::Instance().Get(op->Type()).infer_inplace_;
auto in_to_outs = infer_inplace(*op, op->Block());
EXPECT_EQ(in_to_outs.size(), 1ul);
auto it = in_to_outs.begin();
EXPECT_EQ(it->first, "test2_out");
EXPECT_EQ(it->second, "test2_a");
}
TEST(InferInplace, MultiOutInplaceInToOut) {
ProgramDesc prog;
auto* op = prog.MutableBlock(0)->AppendOp();
op->SetType("multi_out_op");
op->SetInput("X", {"a0", "a1"});
op->SetInput("Y", {"b0"});
op->SetInput("Z", {"c0", "c1"});
op->SetOutput("Out", {"o0"});
op->SetOutput("YOut", {"y0"});
op->SetOutput("ZOut", {"z0"});
prog.MutableBlock(0)->Var("a0")->SetType(proto::VarType::LOD_TENSOR);
prog.MutableBlock(0)->Var("b0")->SetType(proto::VarType::LOD_TENSOR);
prog.MutableBlock(0)->Var("c0")->SetType(proto::VarType::LOD_TENSOR);
prog.MutableBlock(0)->Var("c1")->SetType(proto::VarType::LOD_TENSOR);
prog.MutableBlock(0)->Var("o0");
prog.MutableBlock(0)->Var("y0");
prog.MutableBlock(0)->Var("z0");
prog.MutableBlock(0)->Var("a0")->SetShape({32, 16});
prog.MutableBlock(0)->Var("b0")->SetShape({32, 16});
prog.MutableBlock(0)->Var("c0")->SetShape({32, 16});
prog.MutableBlock(0)->Var("o0")->SetShape({32, 16});
prog.MutableBlock(0)->Var("y0")->SetShape({32, 16});
prog.MutableBlock(0)->Var("z0")->SetShape({32, 16});
auto& infer_inplace = OpInfoMap::Instance().Get(op->Type()).infer_inplace_;
auto in_to_outs = infer_inplace(*op, op->Block());
EXPECT_EQ(in_to_outs.size(), 3ul);
std::unordered_map<std::string, std::string> expects = {
{"a0", "o0"}, {"b0", "y0"}, {"c0", "z0"},
};
EXPECT_TRUE(expects == in_to_outs);
}
TEST(InferInplace, MultiGradInplaceInToOut) {
ProgramDesc prog;
auto* op = prog.MutableBlock(0)->AppendOp();
op->SetType("multi_out_grad");
op->SetInput(GradVarName("Out"), {"o0"});
op->SetInput(GradVarName("YOut"), {"y0"});
op->SetInput(GradVarName("ZOut"), {"z0"});
op->SetOutput(GradVarName("X"), {"a0", "a1"});
op->SetOutput(GradVarName("Y"), {"b0"});
op->SetOutput(GradVarName("Z"), {"c0", "c1"});
prog.MutableBlock(0)->Var("a0")->SetType(proto::VarType::LOD_TENSOR);
prog.MutableBlock(0)->Var("b0")->SetType(proto::VarType::LOD_TENSOR);
prog.MutableBlock(0)->Var("c0")->SetType(proto::VarType::LOD_TENSOR);
prog.MutableBlock(0)->Var("c1")->SetType(proto::VarType::LOD_TENSOR);
prog.MutableBlock(0)->Var("o0");
prog.MutableBlock(0)->Var("y0");
prog.MutableBlock(0)->Var("z0");
prog.MutableBlock(0)->Var("a0")->SetShape({32, 16});
prog.MutableBlock(0)->Var("b0")->SetShape({32, 16});
prog.MutableBlock(0)->Var("c0")->SetShape({32, 16});
prog.MutableBlock(0)->Var("o0")->SetShape({32, 16});
prog.MutableBlock(0)->Var("y0")->SetShape({32, 16});
prog.MutableBlock(0)->Var("z0")->SetShape({32, 16});
auto& infer_inplace = OpInfoMap::Instance().Get(op->Type()).infer_inplace_;
auto in_to_outs = infer_inplace(*op, op->Block());
EXPECT_EQ(in_to_outs.size(), 3ul);
std::unordered_map<std::string, std::string> expects = {
{"o0", "a0"}, {"y0", "b0"}, {"z0", "c0"},
};
EXPECT_TRUE(expects == in_to_outs);
}
} // namespace framework
} // namespace paddle
paddle/fluid/framework/ir/CMakeLists.txt
浏览文件 @ 17ada1d2
...@@ -10,8 +10,22 @@ function(pass_library TARGET DEST) ...@@ -10,8 +10,22 @@ function(pass_library TARGET DEST)
set(options "") set(options "")
set(oneValueArgs "") set(oneValueArgs "")
set(multiValueArgs SRCS DEPS) set(multiValueArgs SRCS DEPS)
set(targetPrefix "")
# Get optional argument
set(extraMacroArgs ${ARGN})
list(LENGTH extraMacroArgs numExtraMacroArgs)
if(numExtraMacroArgs GREATER 0)
list(GET extraMacroArgs 0 targetPrefix)
endif()
cmake_parse_arguments(op_library "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) cmake_parse_arguments(op_library "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
cc_library(${TARGET} SRCS ${TARGET}.cc DEPS graph_pattern_detector pass fuse_pass_base ${op_library_DEPS}) if(targetPrefix)
cc_library(${TARGET} SRCS ${targetPrefix}/${TARGET}.cc DEPS graph_pattern_detector pass fuse_pass_base ${op_library_DEPS})
else()
cc_library(${TARGET} SRCS ${TARGET}.cc DEPS graph_pattern_detector pass fuse_pass_base ${op_library_DEPS})
endif()
# add more DEST here, such as train, dist and collect USE_PASS into a file automatically. # add more DEST here, such as train, dist and collect USE_PASS into a file automatically.
if (${DEST} STREQUAL "base" OR ${DEST} STREQUAL "inference") if (${DEST} STREQUAL "base" OR ${DEST} STREQUAL "inference")
message(STATUS "add pass ${TARGET} ${DEST}") message(STATUS "add pass ${TARGET} ${DEST}")
...@@ -51,6 +65,7 @@ pass_library(conv_elementwise_add2_act_fuse_pass inference) ...@@ -51,6 +65,7 @@ pass_library(conv_elementwise_add2_act_fuse_pass inference)
pass_library(conv_elementwise_add_fuse_pass inference) pass_library(conv_elementwise_add_fuse_pass inference)
pass_library(conv_affine_channel_fuse_pass inference) pass_library(conv_affine_channel_fuse_pass inference)
pass_library(transpose_flatten_concat_fuse_pass inference) pass_library(transpose_flatten_concat_fuse_pass inference)
pass_library(identity_scale_op_clean_pass base)
# There may be many transpose-flatten structures in a model, and the output of # There may be many transpose-flatten structures in a model, and the output of
# these structures will be used as inputs to the concat Op. This pattern will # these structures will be used as inputs to the concat Op. This pattern will
...@@ -62,11 +77,11 @@ foreach (index RANGE 3 6) ...@@ -62,11 +77,11 @@ foreach (index RANGE 3 6)
endforeach() endforeach()
if(WITH_MKLDNN) if(WITH_MKLDNN)
pass_library(mkldnn_placement_pass base) pass_library(mkldnn_placement_pass base mkldnn)
pass_library(depthwise_conv_mkldnn_pass base) pass_library(depthwise_conv_mkldnn_pass base mkldnn)
pass_library(conv_bias_mkldnn_fuse_pass inference) pass_library(conv_bias_mkldnn_fuse_pass inference mkldnn)
pass_library(conv_relu_mkldnn_fuse_pass inference) pass_library(conv_relu_mkldnn_fuse_pass inference mkldnn)
pass_library(conv_elementwise_add_mkldnn_fuse_pass inference) pass_library(conv_elementwise_add_mkldnn_fuse_pass inference mkldnn)
endif() endif()
cc_library(fuse_elewise_add_act_pass SRCS fuse_elewise_add_act_pass.cc DEPS pass graph_pattern_detector ) cc_library(fuse_elewise_add_act_pass SRCS fuse_elewise_add_act_pass.cc DEPS pass graph_pattern_detector )
...@@ -86,7 +101,7 @@ cc_test(test_fc_fuse_pass SRCS fc_fuse_pass_tester.cc DEPS fc_fuse_pass framewor ...@@ -86,7 +101,7 @@ cc_test(test_fc_fuse_pass SRCS fc_fuse_pass_tester.cc DEPS fc_fuse_pass framewor
cc_test(test_seqpool_concat_fuse_pass SRCS seqpool_concat_fuse_pass_tester.cc DEPS seqpool_concat_fuse_pass framework_proto) cc_test(test_seqpool_concat_fuse_pass SRCS seqpool_concat_fuse_pass_tester.cc DEPS seqpool_concat_fuse_pass framework_proto)
cc_test(test_is_test_pass SRCS is_test_pass_tester.cc DEPS is_test_pass) cc_test(test_is_test_pass SRCS is_test_pass_tester.cc DEPS is_test_pass)
if (WITH_MKLDNN) if (WITH_MKLDNN)
cc_test(test_depthwise_conv_mkldnn_pass SRCS depthwise_conv_mkldnn_pass_tester.cc DEPS depthwise_conv_mkldnn_pass) cc_test(test_depthwise_conv_mkldnn_pass SRCS mkldnn/depthwise_conv_mkldnn_pass_tester.cc DEPS depthwise_conv_mkldnn_pass)
cc_test(test_conv_relu_mkldnn_fuse_pass SRCS conv_relu_mkldnn_fuse_pass_tester.cc DEPS conv_relu_mkldnn_fuse_pass) cc_test(test_conv_relu_mkldnn_fuse_pass SRCS mkldnn/conv_relu_mkldnn_fuse_pass_tester.cc DEPS conv_relu_mkldnn_fuse_pass)
cc_test(test_conv_elementwise_add_mkldnn_fuse_pass SRCS conv_elementwise_add_mkldnn_fuse_pass_tester.cc DEPS conv_elementwise_add_mkldnn_fuse_pass) cc_test(test_conv_elementwise_add_mkldnn_fuse_pass SRCS mkldnn/conv_elementwise_add_mkldnn_fuse_pass_tester.cc DEPS conv_elementwise_add_mkldnn_fuse_pass)
endif () endif ()
paddle/fluid/framework/ir/graph.cc
浏览文件 @ 17ada1d2
...@@ -76,7 +76,7 @@ std::map<std::string, std::vector<ir::Node *>> Graph::InitFromProgram( ...@@ -76,7 +76,7 @@ std::map<std::string, std::vector<ir::Node *>> Graph::InitFromProgram(
var->inputs.push_back(node); var->inputs.push_back(node);
} }
} }
return std::move(var_nodes); return var_nodes;
} }
void Graph::ResolveHazard( void Graph::ResolveHazard(
......
paddle/fluid/framework/ir/graph.h
浏览文件 @ 17ada1d2
...@@ -141,7 +141,8 @@ class Graph { ...@@ -141,7 +141,8 @@ class Graph {
ir::Node *CreateControlDepVar() { ir::Node *CreateControlDepVar() {
// TODO(panyx0718): control var name should be really unique. // TODO(panyx0718): control var name should be really unique.
const std::string name = string::Sprintf( const std::string name = string::Sprintf(
"%s@%llu", ir::Node::kControlDepVarName, node_set_.size()); "%s@%llu", static_cast<const char *>(ir::Node::kControlDepVarName),
num_node_created_);
auto *x = AddNode(new ir::Node(name, ir::Node::Type::kVariable)); auto *x = AddNode(new ir::Node(name, ir::Node::Type::kVariable));
x->SetId(num_node_created_++); x->SetId(num_node_created_++);
return x; return x;
......
paddle/fluid/framework/ir/graph_helper.cc
浏览文件 @ 17ada1d2
...@@ -52,16 +52,29 @@ bool HasCircleHelper( ...@@ -52,16 +52,29 @@ bool HasCircleHelper(
ir::Node *node, ir::Node *node,
const std::map<ir::Node *, std::unordered_set<ir::Node *>> &adj_list, const std::map<ir::Node *, std::unordered_set<ir::Node *>> &adj_list,
std::unordered_set<ir::Node *> *visited, std::unordered_set<ir::Node *> *visited,
std::unordered_set<ir::Node *> *in_trace) { std::unordered_set<ir::Node *> *in_trace,
std::vector<std::vector<ir::Node *>> *circles) {
if (visited->find(node) == visited->end()) { if (visited->find(node) == visited->end()) {
visited->insert(node); visited->insert(node);
in_trace->insert(node); in_trace->insert(node);
for (ir::Node *in : adj_list.at(node)) { for (ir::Node *in : adj_list.at(node)) {
if (visited->find(in) == visited->end() && if (visited->find(in) == visited->end() &&
HasCircleHelper(in, adj_list, visited, in_trace)) { HasCircleHelper(in, adj_list, visited, in_trace, circles)) {
return true; return true;
} else if (in_trace->find(in) != in_trace->end()) { } else if (in_trace->find(in) != in_trace->end()) {
if (circles != nullptr) {
std::vector<ir::Node *> circle;
circle.emplace_back(in);
ir::Node *p = in;
for (auto &adj : adj_list.at(p)) {
if (in_trace->count(adj)) {
circle.emplace_back(adj);
p = adj;
}
}
circles->emplace_back(circle);
}
return true; return true;
} }
} }
...@@ -71,11 +84,12 @@ bool HasCircleHelper( ...@@ -71,11 +84,12 @@ bool HasCircleHelper(
} }
bool HasCircleInternal( bool HasCircleInternal(
const std::map<ir::Node *, std::unordered_set<ir::Node *>> &adj_list) { const std::map<ir::Node *, std::unordered_set<ir::Node *>> &adj_list,
std::vector<std::vector<ir::Node *>> *circles) {
std::unordered_set<ir::Node *> visited; std::unordered_set<ir::Node *> visited;
std::unordered_set<ir::Node *> in_trace; std::unordered_set<ir::Node *> in_trace;
for (auto &adj : adj_list) { for (auto &adj : adj_list) {
if (HasCircleHelper(adj.first, adj_list, &visited, &in_trace)) { if (HasCircleHelper(adj.first, adj_list, &visited, &in_trace, circles)) {
return true; return true;
} }
} }
...@@ -84,13 +98,18 @@ bool HasCircleInternal( ...@@ -84,13 +98,18 @@ bool HasCircleInternal(
} // namespace } // namespace
bool HasCircle(const Graph &graph) { bool HasCircle(const Graph &graph) {
return HasCircleInternal(BuildOperationAdjList(graph)); return HasCircleInternal(BuildOperationAdjList(graph), nullptr);
}
bool FindCircleSubGraph(const Graph &graph,
std::vector<std::vector<ir::Node *>> *circles) {
return HasCircleInternal(BuildOperationAdjList(graph), circles);
} }
std::vector<ir::Node *> TopologySortOperations(const Graph &graph) { std::vector<ir::Node *> TopologySortOperations(const Graph &graph) {
std::map<ir::Node *, std::unordered_set<ir::Node *>> adj_list = std::map<ir::Node *, std::unordered_set<ir::Node *>> adj_list =
BuildOperationAdjList(graph); BuildOperationAdjList(graph);
PADDLE_ENFORCE(!HasCircleInternal(adj_list)); PADDLE_ENFORCE(!HasCircleInternal(adj_list, nullptr));
std::unordered_set<ir::Node *> visited; std::unordered_set<ir::Node *> visited;
std::vector<ir::Node *> ret; std::vector<ir::Node *> ret;
for (auto adj : adj_list) { for (auto adj : adj_list) {
......
paddle/fluid/framework/ir/graph_helper.h
浏览文件 @ 17ada1d2
...@@ -28,6 +28,11 @@ namespace ir { ...@@ -28,6 +28,11 @@ namespace ir {
// Test if the graph contains circle. // Test if the graph contains circle.
bool HasCircle(const Graph &graph); bool HasCircle(const Graph &graph);
// Find All Circles for debugging,
// store all subgraph in circles.
bool FindCircleSubGraph(const Graph &graph,
std::vector<std::vector<ir::Node *>> *circles);
size_t GraphNum(const Graph &graph); size_t GraphNum(const Graph &graph);
// Topology Sort the operations in the graph from inputs to outputs. // Topology Sort the operations in the graph from inputs to outputs.
......
paddle/fluid/framework/ir/graph_helper_test.cc
浏览文件 @ 17ada1d2
...@@ -195,6 +195,17 @@ void BuildTwoGraphs(Graph* g) { ...@@ -195,6 +195,17 @@ void BuildTwoGraphs(Graph* g) {
// v4->outputs.push_back(o5); // v4->outputs.push_back(o5);
} }
TEST(GraphHelperTest, Circles) {
ProgramDesc prog;
Graph g(prog);
BuildCircleGraph(&g);
std::vector<std::vector<ir::Node*>> circles;
ASSERT_TRUE(FindCircleSubGraph(g, &circles));
ASSERT_EQ(circles.size(), 1UL);
}
TEST(GraphHelperTest, GraphNum) { TEST(GraphHelperTest, GraphNum) {
ProgramDesc prog; ProgramDesc prog;
......
paddle/fluid/framework/ir/graph_pattern_detector.cc
浏览文件 @ 17ada1d2
...@@ -117,11 +117,6 @@ bool GraphPatternDetector::MarkPDNodesInGraph(const ir::Graph &graph) { ...@@ -117,11 +117,6 @@ bool GraphPatternDetector::MarkPDNodesInGraph(const ir::Graph &graph) {
// return false; // return false;
} }
} }
for (auto &item : pdnodes2nodes_) {
for (auto &n : item.second) {
GetMarkedNodes(const_cast<Graph *>(&graph)).insert(n);
}
}
VLOG(3) << pdnodes2nodes_.size() << " nodes marked"; VLOG(3) << pdnodes2nodes_.size() << " nodes marked";
return !pdnodes2nodes_.empty(); return !pdnodes2nodes_.empty();
......
paddle/fluid/framework/ir/identity_scale_op_clean_pass.cc 0 → 100644
浏览文件 @ 17ada1d2
// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/framework/ir/identity_scale_op_clean_pass.h"
#include <string>
#include "paddle/fluid/framework/ir/graph_pattern_detector.h"
namespace paddle {
namespace framework {
namespace ir {
std::unique_ptr<ir::Graph> IdentityScaleOpCleanPass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
FusePassBase::Init("identity_scale_op_clean", graph.get());
// pre_op -> scale_in -> scale_op -> scale_out
// ->
// pre_op -> scale_out
GraphPatternDetector detector;
auto pre_op = detector.mutable_pattern()->NewNode("pre_op")->assert_is_op();
auto scale_in = detector.mutable_pattern()
->NewNode("scale_in")
->assert_is_op_input("scale")
->AsIntermediate();
auto scale_op = detector.mutable_pattern()
->NewNode("scale_fuse")
->assert_is_op("scale")
->assert_op_attr<float>("scale", 1.)
->assert_op_attr<float>("bias", 0.);
auto scale_out = detector.mutable_pattern()
->NewNode("scale_out")
->assert_is_op_output("scale");
pre_op->LinksTo({scale_in});
scale_op->LinksFrom({scale_in}).LinksTo({scale_out});
GraphPatternDetector::handle_t handler = [&](
const GraphPatternDetector::subgraph_t& subgraph, Graph* graph) {
Node* scale_op_var = subgraph.at(scale_op);
Node* scale_in_var = subgraph.at(scale_in);
Node* scale_out_var = subgraph.at(scale_out);
Node* pre_op_var = subgraph.at(pre_op);
// Link pre_op directly to scale_out
const std::string scale_in_name = scale_in_var->Name();
const std::string scale_out_name = scale_out_var->Name();
// Remove links in graph
GraphSafeRemoveNodes(graph, {scale_in_var, scale_op_var});
// Modify proto message
auto* pre_op_desc = pre_op_var->Op();
for (auto& parameter : *pre_op_desc->Proto()->mutable_outputs()) {
auto* arguments = parameter.mutable_arguments();
auto it = std::find(arguments->begin(), arguments->end(), scale_in_name);
PADDLE_ENFORCE(it != arguments->end());
*it = scale_out_name;
}
IR_NODE_LINK_TO(pre_op_var, scale_out_var);
};
detector(graph.get(), handler);
return graph;
}
} // namespace ir
} // namespace framework
} // namespace paddle
REGISTER_PASS(identity_scale_op_clean_pass,
paddle::framework::ir::IdentityScaleOpCleanPass);
paddle/fluid/framework/details/memory_early_delete_pass.h → paddle/fluid/framework/ir/identity_scale_op_clean_pass.h
浏览文件 @ 17ada1d2
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. // Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
// //
// Licensed under the Apache License, Version 2.0 (the "License"); // Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License. // you may not use this file except in compliance with the License.
...@@ -13,20 +13,21 @@ ...@@ -13,20 +13,21 @@
// limitations under the License. // limitations under the License.
#pragma once #pragma once
#include "paddle/fluid/framework/details/early_delete_op_handle.h"
#include "paddle/fluid/framework/ir/graph.h" #include "paddle/fluid/framework/ir/fuse_pass_base.h"
#include "paddle/fluid/framework/ir/pass.h"
namespace paddle { namespace paddle {
namespace framework { namespace framework {
namespace details { namespace ir {
class MemoryEarlyDeletePass : public ir::Pass { class IdentityScaleOpCleanPass : public FusePassBase {
protected: protected:
std::unique_ptr<ir::Graph> ApplyImpl( std::unique_ptr<ir::Graph> ApplyImpl(std::unique_ptr<ir::Graph> graph) const;
std::unique_ptr<ir::Graph> graph) const override;
private:
virtual ~IdentityScaleOpCleanPass() = default;
}; };
} // namespace details } // namespace ir
} // namespace framework } // namespace framework
} // namespace paddle } // namespace paddle
paddle/fluid/framework/ir/infer_clean_graph_pass.cc
浏览文件 @ 17ada1d2
...@@ -37,6 +37,7 @@ class InferCleanGraphPass : public FusePassBase { ...@@ -37,6 +37,7 @@ class InferCleanGraphPass : public FusePassBase {
std::unordered_set<const Node*> invalid_nodes; std::unordered_set<const Node*> invalid_nodes;
int valid_op = 0; int valid_op = 0;
for (auto* node : graph->Nodes()) { for (auto* node : graph->Nodes()) {
PADDLE_ENFORCE_NOT_NULL(node);
if (is_valid_node(node)) { if (is_valid_node(node)) {
invalid_nodes.insert(node); invalid_nodes.insert(node);
} else if (node->IsOp()) { } else if (node->IsOp()) {
......
paddle/fluid/framework/ir/conv_bias_mkldnn_fuse_pass.cc → paddle/fluid/framework/ir/mkldnn/conv_bias_mkldnn_fuse_pass.cc
浏览文件 @ 17ada1d2
...@@ -12,7 +12,7 @@ ...@@ -12,7 +12,7 @@
// See the License for the specific language governing permissions and // See the License for the specific language governing permissions and
// limitations under the License. // limitations under the License.
#include "paddle/fluid/framework/ir/conv_bias_mkldnn_fuse_pass.h" #include "paddle/fluid/framework/ir/mkldnn/conv_bias_mkldnn_fuse_pass.h"
#include <functional> #include <functional>
#include <string> #include <string>
#include <vector> #include <vector>
......
paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.cc → paddle/fluid/framework/ir/mkldnn/conv_elementwise_add_mkldnn_fuse_pass.cc
浏览文件 @ 17ada1d2
...@@ -12,7 +12,7 @@ ...@@ -12,7 +12,7 @@
// See the License for the specific language governing permissions and // See the License for the specific language governing permissions and
// limitations under the License. // limitations under the License.
#include "paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.h" #include "paddle/fluid/framework/ir/mkldnn/conv_elementwise_add_mkldnn_fuse_pass.h"
#include <functional> #include <functional>
#include <list> #include <list>
#include <map> #include <map>
......
paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass_tester.cc → paddle/fluid/framework/ir/mkldnn/conv_elementwise_add_mkldnn_fuse_pass_tester.cc
浏览文件 @ 17ada1d2
...@@ -15,8 +15,8 @@ ...@@ -15,8 +15,8 @@
#include <gtest/gtest.h> #include <gtest/gtest.h>
#include <string> #include <string>
#include "paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.h"
#include "paddle/fluid/framework/ir/graph_traits.h" #include "paddle/fluid/framework/ir/graph_traits.h"
#include "paddle/fluid/framework/ir/mkldnn/conv_elementwise_add_mkldnn_fuse_pass.h"
namespace paddle { namespace paddle {
namespace framework { namespace framework {
......
paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.cc → paddle/fluid/framework/ir/mkldnn/conv_relu_mkldnn_fuse_pass.cc
浏览文件 @ 17ada1d2
...@@ -12,7 +12,7 @@ ...@@ -12,7 +12,7 @@
// See the License for the specific language governing permissions and // See the License for the specific language governing permissions and
// limitations under the License. // limitations under the License.
#include "paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.h" #include "paddle/fluid/framework/ir/mkldnn/conv_relu_mkldnn_fuse_pass.h"
#include <string> #include <string>
#include <vector> #include <vector>
#include "paddle/fluid/platform/enforce.h" #include "paddle/fluid/platform/enforce.h"
......
paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass_tester.cc → paddle/fluid/framework/ir/mkldnn/conv_relu_mkldnn_fuse_pass_tester.cc
浏览文件 @ 17ada1d2
...@@ -12,7 +12,7 @@ ...@@ -12,7 +12,7 @@
// See the License for the specific language governing permissions and // See the License for the specific language governing permissions and
// limitations under the License. // limitations under the License.
#include "paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.h" #include "paddle/fluid/framework/ir/mkldnn/conv_relu_mkldnn_fuse_pass.h"
#include <gtest/gtest.h> #include <gtest/gtest.h>
#include "paddle/fluid/framework/op_proto_maker.h" #include "paddle/fluid/framework/op_proto_maker.h"
......
paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass.cc → paddle/fluid/framework/ir/mkldnn/depthwise_conv_mkldnn_pass.cc
浏览文件 @ 17ada1d2
...@@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. ...@@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and See the License for the specific language governing permissions and
limitations under the License. */ limitations under the License. */
#include "paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass.h" #include "paddle/fluid/framework/ir/mkldnn/depthwise_conv_mkldnn_pass.h"
#include "paddle/fluid/framework/ir/graph_pattern_detector.h" #include "paddle/fluid/framework/ir/graph_pattern_detector.h"
namespace paddle { namespace paddle {
......
paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass_tester.cc → paddle/fluid/framework/ir/mkldnn/depthwise_conv_mkldnn_pass_tester.cc
浏览文件 @ 17ada1d2
...@@ -12,7 +12,7 @@ ...@@ -12,7 +12,7 @@
// See the License for the specific language governing permissions and // See the License for the specific language governing permissions and
// limitations under the License. // limitations under the License.
#include "paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass.h" #include "paddle/fluid/framework/ir/mkldnn/depthwise_conv_mkldnn_pass.h"
#include <gtest/gtest.h> #include <gtest/gtest.h>
......
paddle/fluid/framework/ir/mkldnn_placement_pass.cc → paddle/fluid/framework/ir/mkldnn/mkldnn_placement_pass.cc
浏览文件 @ 17ada1d2
...@@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. ...@@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and See the License for the specific language governing permissions and
limitations under the License. */ limitations under the License. */
#include "paddle/fluid/framework/ir/mkldnn_placement_pass.h" #include "paddle/fluid/framework/ir/mkldnn/mkldnn_placement_pass.h"
#include <string> #include <string>
namespace paddle { namespace paddle {
......
paddle/fluid/framework/ir/seqpool_concat_fuse_pass_tester.cc
浏览文件 @ 17ada1d2
...@@ -164,7 +164,7 @@ ProgramDesc BuildProgramDesc(int num_inputs_of_concat) { ...@@ -164,7 +164,7 @@ ProgramDesc BuildProgramDesc(int num_inputs_of_concat) {
}; };
std::vector<std::string> concat_inputs; std::vector<std::string> concat_inputs;
for (int i = 0; i < num_inputs_of_concat; ++i) { for (int i = 0; i < num_inputs_of_concat; ++i) {
std::string prefix = "seqpool_op_" + i; std::string prefix = "seqpool_op_" + std::to_string(i);
new_var(prefix + "in"); new_var(prefix + "in");
new_var(prefix + "out"); new_var(prefix + "out");
new_var(prefix + "out_unused"); new_var(prefix + "out_unused");
......
paddle/fluid/framework/op_info.h
浏览文件 @ 17ada1d2
...@@ -38,6 +38,7 @@ struct OpInfo { ...@@ -38,6 +38,7 @@ struct OpInfo {
OpAttrChecker* checker_{nullptr}; OpAttrChecker* checker_{nullptr};
InferVarTypeFN infer_var_type_; InferVarTypeFN infer_var_type_;
InferShapeFN infer_shape_; InferShapeFN infer_shape_;
InferInplaceOpFN infer_inplace_;
bool HasOpProtoAndChecker() const { bool HasOpProtoAndChecker() const {
return proto_ != nullptr && checker_ != nullptr; return proto_ != nullptr && checker_ != nullptr;
......
paddle/fluid/framework/operator.cc
浏览文件 @ 17ada1d2
...@@ -188,14 +188,14 @@ void OperatorBase::Run(const Scope& scope, const platform::Place& place) { ...@@ -188,14 +188,14 @@ void OperatorBase::Run(const Scope& scope, const platform::Place& place) {
VLOG(3) << place << " " << DebugStringEx(&scope); VLOG(3) << place << " " << DebugStringEx(&scope);
} catch (platform::EnforceNotMet exception) { } catch (platform::EnforceNotMet exception) {
if (Attrs().count("sub_block") != 0) { if (Attrs().count("sub_block") != 0) {
throw exception; throw;
} }
auto& callstack = Attr<std::vector<std::string>>( auto& callstack = Attr<std::vector<std::string>>(
OpProtoAndCheckerMaker::OpCreationCallstackAttrName()); OpProtoAndCheckerMaker::OpCreationCallstackAttrName());
if (callstack.empty()) { if (callstack.empty()) {
throw exception; throw;
} }
std::ostringstream sout; std::ostringstream sout;
sout << "Invoke operator " << Type() << " error.\n"; sout << "Invoke operator " << Type() << " error.\n";
...@@ -206,7 +206,7 @@ void OperatorBase::Run(const Scope& scope, const platform::Place& place) { ...@@ -206,7 +206,7 @@ void OperatorBase::Run(const Scope& scope, const platform::Place& place) {
sout << "C++ Callstacks: \n"; sout << "C++ Callstacks: \n";
sout << exception.err_str_; sout << exception.err_str_;
exception.err_str_ = sout.str(); exception.err_str_ = sout.str();
throw exception; throw;
} catch (...) { } catch (...) {
std::rethrow_exception(std::current_exception()); std::rethrow_exception(std::current_exception());
} }
...@@ -555,18 +555,17 @@ Tensor* ExecutionContext::LegacyOutput<Tensor>(const std::string& name) const { ...@@ -555,18 +555,17 @@ Tensor* ExecutionContext::LegacyOutput<Tensor>(const std::string& name) const {
template <> template <>
std::vector<Tensor*> ExecutionContext::MultiOutput<Tensor>( std::vector<Tensor*> ExecutionContext::MultiOutput<Tensor>(
const std::string& name) const { const std::string& name) const {
auto names = op().Outputs(name); auto it = ctx_.outputs.find(name);
if (it == ctx_.outputs.end()) {
return {};
}
const std::vector<Variable*>& vars = it->second;
std::vector<Tensor*> res; std::vector<Tensor*> res;
res.reserve(names.size()); res.reserve(vars.size());
std::transform(names.begin(), names.end(), std::back_inserter(res), std::transform(vars.begin(), vars.end(), std::back_inserter(res),
[&](const std::string& sub_name) -> Tensor* { [&](Variable* var) -> Tensor* {
auto var = scope_.FindVar(sub_name); return var == nullptr ? nullptr
if (var == nullptr) return nullptr; : var->GetMutable<LoDTensor>();
PADDLE_ENFORCE(
var->IsType<LoDTensor>(),
"%s should be LoDTensor, but the received type is %s",
sub_name, ToTypeName(var->Type()));
return var->GetMutable<LoDTensor>();
}); });
return res; return res;
} }
...@@ -590,7 +589,7 @@ class RuntimeInferShapeContext : public InferShapeContext { ...@@ -590,7 +589,7 @@ class RuntimeInferShapeContext : public InferShapeContext {
public: public:
RuntimeInferShapeContext(const OperatorBase& op, const Scope& scope, RuntimeInferShapeContext(const OperatorBase& op, const Scope& scope,
const RuntimeContext& ctx) const RuntimeContext& ctx)
: op_(op), scope_(scope), ctx_(ctx) {} : op_(op), ctx_(ctx) {}
bool HasInput(const std::string& name) const override { bool HasInput(const std::string& name) const override {
// has only one input // has only one input
...@@ -882,7 +881,6 @@ class RuntimeInferShapeContext : public InferShapeContext { ...@@ -882,7 +881,6 @@ class RuntimeInferShapeContext : public InferShapeContext {
} }
const OperatorBase& op_; const OperatorBase& op_;
const Scope& scope_;
const RuntimeContext& ctx_; const RuntimeContext& ctx_;
}; };
...@@ -991,11 +989,14 @@ void OperatorWithKernel::TransferInplaceVarsBack( ...@@ -991,11 +989,14 @@ void OperatorWithKernel::TransferInplaceVarsBack(
const Scope& transfer_scope) const { const Scope& transfer_scope) const {
for (auto& var_name : inplace_vars) { for (auto& var_name : inplace_vars) {
VLOG(3) << "share inplace var " + var_name + " back to it's original scope"; VLOG(3) << "share inplace var " + var_name + " back to it's original scope";
auto* origin_var = scope.FindVar(var_name);
PADDLE_ENFORCE_NOT_NULL(origin_var, "The var[%s] should not be nullptr.",
var_name);
auto* original_tensor = auto* original_tensor =
GetMutableLoDTensorOrSelectedRowsValueFromVar(scope.FindVar(var_name)); GetMutableLoDTensorOrSelectedRowsValueFromVar(origin_var);
auto* var = transfer_scope.FindVar(var_name); auto* var = transfer_scope.FindVar(var_name);
PADDLE_ENFORCE(var != nullptr, "The var[%s] should not be nullptr", PADDLE_ENFORCE_NOT_NULL(var, "The var[%s] should not be nullptr.",
var_name); var_name);
auto* transformed_tensor = GetLoDTensorOrSelectedRowsValueFromVar(*var); auto* transformed_tensor = GetLoDTensorOrSelectedRowsValueFromVar(*var);
original_tensor->ShareDataWith(*transformed_tensor); original_tensor->ShareDataWith(*transformed_tensor);
} }
......
paddle/fluid/framework/operator.h
浏览文件 @ 17ada1d2
...@@ -222,12 +222,7 @@ class ExecutionContext { ...@@ -222,12 +222,7 @@ class ExecutionContext {
if (it == ctx_.inputs.end()) { if (it == ctx_.inputs.end()) {
return {}; return {};
} }
std::vector<const Variable*> res; return {it->second.begin(), it->second.end()};
res.reserve(it->second.size());
std::transform(it->second.begin(), it->second.end(),
std::back_inserter(res),
[this](Variable* var) { return var; });
return res;
} }
std::vector<Variable*> MultiOutputVar(const std::string& name) const { std::vector<Variable*> MultiOutputVar(const std::string& name) const {
......
paddle/fluid/framework/parallel_executor.cc
浏览文件 @ 17ada1d2
...@@ -171,14 +171,6 @@ std::unique_ptr<ir::Graph> ParallelExecutorPrivate::PrepareGCAndRefCnts( ...@@ -171,14 +171,6 @@ std::unique_ptr<ir::Graph> ParallelExecutorPrivate::PrepareGCAndRefCnts(
eager_deletion_pass->SetNotOwned(details::kAllPlaces, &places_); eager_deletion_pass->SetNotOwned(details::kAllPlaces, &places_);
graph = eager_deletion_pass->Apply(std::move(graph)); graph = eager_deletion_pass->Apply(std::move(graph));
VLOG(10) << "EagerDeletionPass Applied"; VLOG(10) << "EagerDeletionPass Applied";
if (build_strategy_.memory_early_delete_) {
auto early_delete_pass =
ir::PassRegistry::Instance().Get("memory_early_delete_pass");
early_delete_pass->SetNotOwned(details::kGarbageCollector, &gcs_);
graph = early_delete_pass->Apply(std::move(graph));
}
VLOG(10) << "MemoryEarlyDeletePass Applied.";
} }
return graph; return graph;
...@@ -288,6 +280,8 @@ ParallelExecutor::ParallelExecutor( ...@@ -288,6 +280,8 @@ ParallelExecutor::ParallelExecutor(
graphs.push_back(std::move(graph)); graphs.push_back(std::move(graph));
#endif #endif
auto max_memory_size = GetEagerDeletionThreshold(); auto max_memory_size = GetEagerDeletionThreshold();
VLOG(10) << "Eager Deletion Threshold "
<< static_cast<float>(max_memory_size) / (1 << 30);
if (max_memory_size >= 0) { if (max_memory_size >= 0) {
for (size_t i = 0; i < graphs.size(); ++i) { for (size_t i = 0; i < graphs.size(); ++i) {
graphs[i] = member_->PrepareGCAndRefCnts( graphs[i] = member_->PrepareGCAndRefCnts(
...@@ -506,6 +500,5 @@ ParallelExecutor::~ParallelExecutor() { ...@@ -506,6 +500,5 @@ ParallelExecutor::~ParallelExecutor() {
} // namespace framework } // namespace framework
} // namespace paddle } // namespace paddle
USE_PASS(memory_early_delete_pass);
USE_PASS(reference_count_pass); USE_PASS(reference_count_pass);
USE_PASS(eager_deletion_pass); USE_PASS(eager_deletion_pass);
paddle/fluid/framework/scope.cc
浏览文件 @ 17ada1d2
...@@ -22,11 +22,7 @@ limitations under the License. */ ...@@ -22,11 +22,7 @@ limitations under the License. */
#include "paddle/fluid/framework/threadpool.h" #include "paddle/fluid/framework/threadpool.h"
#include "paddle/fluid/string/printf.h" #include "paddle/fluid/string/printf.h"
DEFINE_bool(benchmark, false, DECLARE_bool(benchmark);
"Doing memory benchmark. It will make deleting scope synchronized, "
"and add some memory usage logs."
"Default cuda is asynchronous device, set to True will"
"force op run in synchronous mode.");
DEFINE_bool( DEFINE_bool(
eager_delete_scope, true, eager_delete_scope, true,
......
paddle/fluid/framework/type_defs.h
浏览文件 @ 17ada1d2
...@@ -57,5 +57,8 @@ using InferVarTypeFN = ...@@ -57,5 +57,8 @@ using InferVarTypeFN =
using InferShapeFN = std::function<void(InferShapeContext*)>; using InferShapeFN = std::function<void(InferShapeContext*)>;
using InplacePair = std::unordered_map<std::string, std::string>;
using InferInplaceOpFN = std::function<InplacePair(const OpDesc&, BlockDesc*)>;
} // namespace framework } // namespace framework
} // namespace paddle } // namespace paddle
paddle/fluid/imperative/CMakeLists.txt
浏览文件 @ 17ada1d2
if(WITH_PYTHON) if(WITH_PYTHON)
cc_library(layer SRCS layer.cc DEPS proto_desc operator device_context blas) cc_library(layer SRCS layer.cc DEPS proto_desc operator device_context blas pybind)
cc_library(tracer SRCS tracer.cc DEPS proto_desc device_context) cc_library(tracer SRCS tracer.cc DEPS proto_desc device_context pybind)
cc_library(engine SRCS engine.cc) cc_library(engine SRCS engine.cc)
endif() endif()
paddle/fluid/imperative/layer.cc
浏览文件 @ 17ada1d2
...@@ -156,6 +156,8 @@ class Autograd { ...@@ -156,6 +156,8 @@ class Autograd {
for (auto it : candidate->pre_ops_) { for (auto it : candidate->pre_ops_) {
for (OpBase* pre_op : it.second) { for (OpBase* pre_op : it.second) {
if (!pre_op) continue; if (!pre_op) continue;
VLOG(5) << "op dep " << candidate->op_desc_->Type() << " <---- "
<< it.first << " <---- " << pre_op->op_desc_->Type();
if (visited.find(pre_op) == visited.end()) { if (visited.find(pre_op) == visited.end()) {
visited.insert(pre_op); visited.insert(pre_op);
queue.push_back(pre_op); queue.push_back(pre_op);
......
paddle/fluid/imperative/layer.h
浏览文件 @ 17ada1d2
...@@ -28,6 +28,7 @@ ...@@ -28,6 +28,7 @@
#include "paddle/fluid/framework/var_desc.h" #include "paddle/fluid/framework/var_desc.h"
#include "paddle/fluid/platform/enforce.h" #include "paddle/fluid/platform/enforce.h"
#include "paddle/fluid/platform/device_context.h" #include "paddle/fluid/platform/device_context.h"
#include "paddle/fluid/operators/math/math_function.h"
#include "paddle/fluid/imperative/type_defs.h" #include "paddle/fluid/imperative/type_defs.h"
...@@ -140,16 +141,24 @@ class VarBase { ...@@ -140,16 +141,24 @@ class VarBase {
void RunBackward(); void RunBackward();
void TrackPreOp(OpBase* pre_op, const std::string& pre_op_out_name, void TrackPreOp(OpBase* pre_op, const std::string& pre_op_out_name,
int pre_op_out_idx, bool stop_gradient) { int pre_op_out_idx, bool pre_op_stop_gradient) {
pre_op_ = pre_op; pre_op_ = pre_op;
pre_op_out_name_ = pre_op_out_name; pre_op_out_name_ = pre_op_out_name;
pre_op_out_idx_ = pre_op_out_idx; pre_op_out_idx_ = pre_op_out_idx;
stop_gradient_ = stop_gradient; if (pre_op_stop_gradient) {
stop_gradient_ = pre_op_stop_gradient;
}
} }
void ClearGradient() { void ClearGradient() {
delete grads_; VLOG(1) << "clear gradient of " << var_desc_->Name();
grads_ = new VarBase(true); if (grads_ && grads_->var_ && grads_->var_->IsInitialized()) {
auto grads_t = grads_->var_->GetMutable<framework::LoDTensor>();
operators::math::set_constant(
*(platform::DeviceContextPool::Instance().Get(
grads_->var_->Get<framework::LoDTensor>().place())),
grads_t, 0.0);
}
} }
framework::LoDTensor& GradValue(); framework::LoDTensor& GradValue();
......
paddle/fluid/imperative/tracer.cc
浏览文件 @ 17ada1d2
...@@ -31,6 +31,7 @@ void CreateGradOp(const framework::OpDesc& op_desc, ...@@ -31,6 +31,7 @@ void CreateGradOp(const framework::OpDesc& op_desc,
framework::OpInfoMap::Instance() framework::OpInfoMap::Instance()
.Get(op_desc.Type()) .Get(op_desc.Type())
.GradOpMaker()(op_desc, no_grad_set, grad_to_var, grad_sub_block); .GradOpMaker()(op_desc, no_grad_set, grad_to_var, grad_sub_block);
for (auto& desc : descs) { for (auto& desc : descs) {
grad_op_descs->emplace_back(desc.release()); grad_op_descs->emplace_back(desc.release());
} }
...@@ -84,11 +85,12 @@ void Tracer::Trace(OpBase* op, const VarBasePtrMap& inputs, ...@@ -84,11 +85,12 @@ void Tracer::Trace(OpBase* op, const VarBasePtrMap& inputs,
op->input_vars_ = inputs; op->input_vars_ = inputs;
for (auto it : op->input_vars_) { for (auto it : op->input_vars_) {
auto& invars = invars_map[it.first]; auto& invars = invars_map[it.first];
invars.reserve(it.second.size());
for (VarBase* inp : it.second) { for (VarBase* inp : it.second) {
PADDLE_ENFORCE_NOT_NULL(inp->var_, "op %s input %s nullptr", PADDLE_ENFORCE_NOT_NULL(inp->var_, "op %s input %s nullptr",
op->op_desc_->Type(), inp->var_desc_->Name()); op->op_desc_->Type(), inp->var_desc_->Name());
invars.push_back(inp->var_); invars.emplace_back(inp->var_);
vars[inp->var_desc_->Name()] = inp; vars[inp->var_desc_->Name()] = inp;
if (inp->PreOp()) { if (inp->PreOp()) {
op->pre_ops_[it.first].push_back(inp->PreOp()); op->pre_ops_[it.first].push_back(inp->PreOp());
...@@ -105,9 +107,10 @@ void Tracer::Trace(OpBase* op, const VarBasePtrMap& inputs, ...@@ -105,9 +107,10 @@ void Tracer::Trace(OpBase* op, const VarBasePtrMap& inputs,
for (auto it : op->output_vars_) { for (auto it : op->output_vars_) {
auto& outvars = outvars_map[it.first]; auto& outvars = outvars_map[it.first];
const std::vector<VarBase*>& outputs = it.second; const std::vector<VarBase*>& outputs = it.second;
outvars.reserve(outputs.size());
for (size_t i = 0; i < outputs.size(); ++i) { for (size_t i = 0; i < outputs.size(); ++i) {
VarBase* out = outputs[i]; VarBase* out = outputs[i];
outvars.push_back(out->var_); outvars.emplace_back(out->var_);
vars[out->var_desc_->Name()] = out; vars[out->var_desc_->Name()] = out;
framework::VarDesc* var_desc = block->FindVar(out->var_desc_->Name()); framework::VarDesc* var_desc = block->FindVar(out->var_desc_->Name());
......
paddle/fluid/inference/CMakeLists.txt
浏览文件 @ 17ada1d2
...@@ -58,12 +58,13 @@ if(WIN32) ...@@ -58,12 +58,13 @@ if(WIN32)
sep_library(paddle_fluid_shared SHARED SRCS ${SHARED_INFERENCE_SRCS} sep_library(paddle_fluid_shared SHARED SRCS ${SHARED_INFERENCE_SRCS}
DEPS ${fluid_modules} paddle_fluid_api reset_tensor_array DEPS ${fluid_modules} paddle_fluid_api reset_tensor_array
analysis_config paddle_pass_builder) analysis_config paddle_pass_builder)
target_link_libraries(paddle_fluid_shared shlwapi)
else(WIN32) else(WIN32)
cc_library(paddle_fluid_shared SHARED SRCS ${SHARED_INFERENCE_SRCS} cc_library(paddle_fluid_shared SHARED SRCS ${SHARED_INFERENCE_SRCS}
DEPS ${fluid_modules} paddle_fluid_api reset_tensor_array DEPS ${fluid_modules} paddle_fluid_api reset_tensor_array
analysis_config paddle_pass_builder) analysis_config paddle_pass_builder)
endif() endif()
get_property(os_dependency_modules GLOBAL PROPERTY OS_DEPENDENCY_MODULES)
target_link_libraries(paddle_fluid_shared ${os_dependency_modules})
set_target_properties(paddle_fluid_shared PROPERTIES OUTPUT_NAME paddle_fluid) set_target_properties(paddle_fluid_shared PROPERTIES OUTPUT_NAME paddle_fluid)
if(NOT APPLE AND NOT WIN32) if(NOT APPLE AND NOT WIN32)
......
paddle/fluid/inference/analysis/argument.h
浏览文件 @ 17ada1d2
...@@ -132,7 +132,7 @@ struct Argument { ...@@ -132,7 +132,7 @@ struct Argument {
DECL_ARGUMENT_FIELD(tensorrt_workspace_size, TensorRtWorkspaceSize, int); DECL_ARGUMENT_FIELD(tensorrt_workspace_size, TensorRtWorkspaceSize, int);
DECL_ARGUMENT_FIELD(tensorrt_min_subgraph_size, TensorRtMinSubgraphSize, int); DECL_ARGUMENT_FIELD(tensorrt_min_subgraph_size, TensorRtMinSubgraphSize, int);
DECL_ARGUMENT_FIELD(tensorrt_precision_mode, TensorRtPrecisionMode, DECL_ARGUMENT_FIELD(tensorrt_precision_mode, TensorRtPrecisionMode,
contrib::AnalysisConfig::Precision); AnalysisConfig::Precision);
// Memory optimized related. // Memory optimized related.
DECL_ARGUMENT_FIELD(enable_memory_optim, EnableMemoryOptim, bool); DECL_ARGUMENT_FIELD(enable_memory_optim, EnableMemoryOptim, bool);
......
paddle/fluid/inference/analysis/helper.h
浏览文件 @ 17ada1d2
...@@ -32,7 +32,7 @@ limitations under the License. */ ...@@ -32,7 +32,7 @@ limitations under the License. */
#ifdef _WIN32 #ifdef _WIN32
#include <direct.h> #include <direct.h>
#include <io.h> #include <io.h>
#define GCC_ATTRIBUTE(attr__) ; #define GCC_ATTRIBUTE(attr__)
#define MKDIR(path) _mkdir(path) #define MKDIR(path) _mkdir(path)
#else #else
#include <unistd.h> #include <unistd.h>
......
paddle/fluid/inference/analysis/ir_pass_manager.cc
浏览文件 @ 17ada1d2
...@@ -71,7 +71,7 @@ void IRPassManager::CreatePasses(Argument *argument, ...@@ -71,7 +71,7 @@ void IRPassManager::CreatePasses(Argument *argument,
new framework::ProgramDesc *(&argument->main_program())); new framework::ProgramDesc *(&argument->main_program()));
bool enable_int8 = argument->tensorrt_precision_mode() == bool enable_int8 = argument->tensorrt_precision_mode() ==
contrib::AnalysisConfig::Precision::kInt8; AnalysisConfig::Precision::kInt8;
pass->Set("enable_int8", new bool(enable_int8)); pass->Set("enable_int8", new bool(enable_int8));
std::string model_opt_cache_dir = std::string model_opt_cache_dir =
...@@ -83,7 +83,6 @@ void IRPassManager::CreatePasses(Argument *argument, ...@@ -83,7 +83,6 @@ void IRPassManager::CreatePasses(Argument *argument,
new std::string(GetOrCreateModelOptCacheDir(model_opt_cache_dir))); new std::string(GetOrCreateModelOptCacheDir(model_opt_cache_dir)));
} }
// graph_ = pass->Apply(std::move(graph_));
pre_pass = pass_name; pre_pass = pass_name;
passes_.emplace_back(std::move(pass)); passes_.emplace_back(std::move(pass));
...@@ -97,11 +96,12 @@ std::unique_ptr<Graph> IRPassManager::Apply(std::unique_ptr<Graph> graph) { ...@@ -97,11 +96,12 @@ std::unique_ptr<Graph> IRPassManager::Apply(std::unique_ptr<Graph> graph) {
PADDLE_ENFORCE(graph.get()); PADDLE_ENFORCE(graph.get());
// Apply all the passes // Apply all the passes
for (const auto &pass : passes_) { for (const auto &pass : passes_) {
if (pass->Type() == "graph_viz_pass") continue; if (pass->Type() != "graph_viz_pass") {
PrettyLogEndl(Style::H2(), "--- Running IR pass [%s]", pass->Type()); PrettyLogEndl(Style::H2(), "--- Running IR pass [%s]", pass->Type());
}
graph = pass->Apply(std::move(graph)); graph = pass->Apply(std::move(graph));
} }
return std::move(graph); return graph;
} }
framework::proto::ProgramDesc IRPassManager::AcquireProgram( framework::proto::ProgramDesc IRPassManager::AcquireProgram(
......
paddle/fluid/inference/analysis/ir_passes/CMakeLists.txt
浏览文件 @ 17ada1d2
cc_library(subgraph_detector SRCS subgraph_detector.cc DEPS proto_desc) cc_library(subgraph_detector SRCS subgraph_detector.cc DEPS proto_desc)
if(WITH_TESTING)
add_dependencies(subgraph_detector gtest)
endif()
if (WITH_GPU AND TENSORRT_FOUND) if (WITH_GPU AND TENSORRT_FOUND)
cc_library(tensorrt_subgraph_pass SRCS tensorrt_subgraph_pass.cc DEPS subgraph_detector tensorrt_op_teller) cc_library(tensorrt_subgraph_pass SRCS tensorrt_subgraph_pass.cc DEPS subgraph_detector tensorrt_op_teller)
......
paddle/fluid/inference/analysis/passes/memory_optimize_pass.cc
浏览文件 @ 17ada1d2
...@@ -18,6 +18,7 @@ ...@@ -18,6 +18,7 @@
#include <limits> #include <limits>
#include <map> #include <map>
#include <string> #include <string>
#include <type_traits>
#include <utility> #include <utility>
#include <vector> #include <vector>
#include "paddle/fluid/framework/ir/graph_helper.h" #include "paddle/fluid/framework/ir/graph_helper.h"
...@@ -168,7 +169,11 @@ bool FindSuitableTensorToReuse( ...@@ -168,7 +169,11 @@ bool FindSuitableTensorToReuse(
if (!cluster->count(candidate)) continue; if (!cluster->count(candidate)) continue;
size_t space = space_table.at(candidate); size_t space = space_table.at(candidate);
size_t space_diff = std::abs<size_t>(space - space_required); PADDLE_ENFORCE(
space <= std::numeric_limits<std::make_signed<size_t>::type>::max(),
"space overload");
size_t space_diff =
std::abs((std::make_signed<size_t>::type)space - space_required);
if (space_diff < best_fit.second) { if (space_diff < best_fit.second) {
best_fit.first = candidate; best_fit.first = candidate;
best_fit.second = space_diff; best_fit.second = space_diff;
......
paddle/fluid/inference/analysis/passes/memory_optimize_pass.h
浏览文件 @ 17ada1d2
...@@ -13,7 +13,9 @@ ...@@ -13,7 +13,9 @@
// limitations under the License. // limitations under the License.
#pragma once #pragma once
#include <string>
#include <utility>
#include <vector>
#include "paddle/fluid/inference/analysis/analysis_pass.h" #include "paddle/fluid/inference/analysis/analysis_pass.h"
#include "paddle/fluid/platform/port.h" #include "paddle/fluid/platform/port.h"
......
paddle/fluid/inference/api/CMakeLists.txt
浏览文件 @ 17ada1d2
...@@ -52,8 +52,8 @@ cc_test(test_analysis_predictor SRCS analysis_predictor_tester.cc DEPS analysis_ ...@@ -52,8 +52,8 @@ cc_test(test_analysis_predictor SRCS analysis_predictor_tester.cc DEPS analysis_
if (WITH_ANAKIN AND WITH_MKL) # only needed in CI if (WITH_ANAKIN AND WITH_MKL) # only needed in CI
# compile the libinference_anakin_api.a and anakin.so. # compile the libinference_anakin_api.a and anakin.so.
cc_library(inference_anakin_api SRCS api.cc api_anakin_engine.cc DEPS anakin_shared anakin_saber mklml zero_copy_tensor_dummy) cc_library(inference_anakin_api SRCS api.cc api_anakin_engine.cc DEPS anakin_shared anakin_saber mklml zero_copy_tensor_dummy device_context)
cc_library(inference_anakin_api_shared SHARED SRCS api.cc api_anakin_engine.cc DEPS anakin_shared anakin_saber zero_copy_tensor_dummy) cc_library(inference_anakin_api_shared SHARED SRCS api.cc api_anakin_engine.cc DEPS anakin_shared anakin_saber zero_copy_tensor_dummy device_context)
function(anakin_target target_name) function(anakin_target target_name)
target_compile_options(${target_name} BEFORE PUBLIC ${ANAKIN_COMPILE_EXTRA_FLAGS}) target_compile_options(${target_name} BEFORE PUBLIC ${ANAKIN_COMPILE_EXTRA_FLAGS})
endfunction() endfunction()
......
paddle/fluid/inference/api/analysis_config.cc
浏览文件 @ 17ada1d2
...@@ -22,7 +22,7 @@ ...@@ -22,7 +22,7 @@
namespace paddle { namespace paddle {
PassStrategy *contrib::AnalysisConfig::pass_builder() const { PassStrategy *AnalysisConfig::pass_builder() const {
if (!pass_builder_.get()) { if (!pass_builder_.get()) {
if (use_gpu_) { if (use_gpu_) {
LOG(INFO) << "Create GPU IR passes"; LOG(INFO) << "Create GPU IR passes";
...@@ -42,27 +42,27 @@ PassStrategy *contrib::AnalysisConfig::pass_builder() const { ...@@ -42,27 +42,27 @@ PassStrategy *contrib::AnalysisConfig::pass_builder() const {
return pass_builder_.get(); return pass_builder_.get();
} }
contrib::AnalysisConfig::AnalysisConfig(const std::string &model_dir) { AnalysisConfig::AnalysisConfig(const std::string &model_dir) {
model_dir_ = model_dir; model_dir_ = model_dir;
Update(); Update();
} }
contrib::AnalysisConfig::AnalysisConfig(const std::string &prog_file, AnalysisConfig::AnalysisConfig(const std::string &prog_file,
const std::string &params_file) { const std::string &params_file) {
prog_file_ = prog_file; prog_file_ = prog_file;
params_file_ = params_file; params_file_ = params_file;
Update(); Update();
} }
void contrib::AnalysisConfig::SetModel(const std::string &prog_file_path, void AnalysisConfig::SetModel(const std::string &prog_file_path,
const std::string &params_file_path) { const std::string &params_file_path) {
prog_file_ = prog_file_path; prog_file_ = prog_file_path;
params_file_ = params_file_path; params_file_ = params_file_path;
Update(); Update();
} }
void contrib::AnalysisConfig::EnableUseGpu(uint64_t memory_pool_init_size_mb, void AnalysisConfig::EnableUseGpu(uint64_t memory_pool_init_size_mb,
int device_id) { int device_id) {
#ifdef PADDLE_WITH_CUDA #ifdef PADDLE_WITH_CUDA
use_gpu_ = true; use_gpu_ = true;
memory_pool_init_size_mb_ = memory_pool_init_size_mb; memory_pool_init_size_mb_ = memory_pool_init_size_mb;
...@@ -74,13 +74,13 @@ void contrib::AnalysisConfig::EnableUseGpu(uint64_t memory_pool_init_size_mb, ...@@ -74,13 +74,13 @@ void contrib::AnalysisConfig::EnableUseGpu(uint64_t memory_pool_init_size_mb,
Update(); Update();
} }
void contrib::AnalysisConfig::DisableGpu() { void AnalysisConfig::DisableGpu() {
use_gpu_ = false; use_gpu_ = false;
Update(); Update();
} }
contrib::AnalysisConfig::AnalysisConfig(const contrib::AnalysisConfig &other) { AnalysisConfig::AnalysisConfig(const AnalysisConfig &other) {
#define CP_MEMBER(member__) member__ = other.member__; #define CP_MEMBER(member__) member__ = other.member__;
// Model related. // Model related.
...@@ -130,7 +130,7 @@ contrib::AnalysisConfig::AnalysisConfig(const contrib::AnalysisConfig &other) { ...@@ -130,7 +130,7 @@ contrib::AnalysisConfig::AnalysisConfig(const contrib::AnalysisConfig &other) {
Update(); Update();
} }
void contrib::AnalysisConfig::EnableMKLDNN() { void AnalysisConfig::EnableMKLDNN() {
#ifdef PADDLE_WITH_MKLDNN #ifdef PADDLE_WITH_MKLDNN
pass_builder()->EnableMKLDNN(); pass_builder()->EnableMKLDNN();
use_mkldnn_ = true; use_mkldnn_ = true;
...@@ -142,9 +142,9 @@ void contrib::AnalysisConfig::EnableMKLDNN() { ...@@ -142,9 +142,9 @@ void contrib::AnalysisConfig::EnableMKLDNN() {
Update(); Update();
} }
void contrib::AnalysisConfig::EnableTensorRtEngine( void AnalysisConfig::EnableTensorRtEngine(
int workspace_size, int max_batch_size, int min_subgraph_size, int workspace_size, int max_batch_size, int min_subgraph_size,
contrib::AnalysisConfig::Precision precision_mode) { AnalysisConfig::Precision precision_mode) {
#ifdef PADDLE_WITH_CUDA #ifdef PADDLE_WITH_CUDA
if (!use_gpu()) { if (!use_gpu()) {
LOG(ERROR) << "To use TensorRT engine, please call EnableGpu() first"; LOG(ERROR) << "To use TensorRT engine, please call EnableGpu() first";
...@@ -165,7 +165,7 @@ void contrib::AnalysisConfig::EnableTensorRtEngine( ...@@ -165,7 +165,7 @@ void contrib::AnalysisConfig::EnableTensorRtEngine(
} }
// TODO(Superjomn) refactor this, buggy. // TODO(Superjomn) refactor this, buggy.
void contrib::AnalysisConfig::Update() { void AnalysisConfig::Update() {
auto info = SerializeInfoCache(); auto info = SerializeInfoCache();
if (info == serialized_info_cache_) return; if (info == serialized_info_cache_) return;
...@@ -225,7 +225,7 @@ void contrib::AnalysisConfig::Update() { ...@@ -225,7 +225,7 @@ void contrib::AnalysisConfig::Update() {
} }
} }
std::string contrib::AnalysisConfig::SerializeInfoCache() { std::string AnalysisConfig::SerializeInfoCache() {
std::stringstream ss; std::stringstream ss;
ss << model_dir_; ss << model_dir_;
ss << prog_file_; ss << prog_file_;
...@@ -260,14 +260,14 @@ std::string contrib::AnalysisConfig::SerializeInfoCache() { ...@@ -260,14 +260,14 @@ std::string contrib::AnalysisConfig::SerializeInfoCache() {
return ss.str(); return ss.str();
} }
void contrib::AnalysisConfig::SetCpuMathLibraryNumThreads( void AnalysisConfig::SetCpuMathLibraryNumThreads(
int cpu_math_library_num_threads) { int cpu_math_library_num_threads) {
cpu_math_library_num_threads_ = cpu_math_library_num_threads; cpu_math_library_num_threads_ = cpu_math_library_num_threads;
Update(); Update();
} }
float contrib::AnalysisConfig::fraction_of_gpu_memory_for_pool() const { float AnalysisConfig::fraction_of_gpu_memory_for_pool() const {
#ifdef PADDLE_WITH_CUDA #ifdef PADDLE_WITH_CUDA
// Get the GPU memory details and calculate the fraction of memory for the // Get the GPU memory details and calculate the fraction of memory for the
// GPU memory pool. // GPU memory pool.
...@@ -282,8 +282,8 @@ float contrib::AnalysisConfig::fraction_of_gpu_memory_for_pool() const { ...@@ -282,8 +282,8 @@ float contrib::AnalysisConfig::fraction_of_gpu_memory_for_pool() const {
#endif #endif
} }
void contrib::AnalysisConfig::EnableMemoryOptim( void AnalysisConfig::EnableMemoryOptim(bool static_optim,
bool static_optim, bool force_update_static_cache) { bool force_update_static_cache) {
enable_memory_optim_ = true; enable_memory_optim_ = true;
static_memory_optim_ = static_optim; static_memory_optim_ = static_optim;
static_memory_optim_force_update_ = force_update_static_cache; static_memory_optim_force_update_ = force_update_static_cache;
...@@ -291,14 +291,14 @@ void contrib::AnalysisConfig::EnableMemoryOptim( ...@@ -291,14 +291,14 @@ void contrib::AnalysisConfig::EnableMemoryOptim(
Update(); Update();
} }
bool contrib::AnalysisConfig::enable_memory_optim() const { bool AnalysisConfig::enable_memory_optim() const {
return enable_memory_optim_; return enable_memory_optim_;
} }
void contrib::AnalysisConfig::SetModelBuffer(const char *prog_buffer, void AnalysisConfig::SetModelBuffer(const char *prog_buffer,
size_t prog_buffer_size, size_t prog_buffer_size,
const char *param_buffer, const char *param_buffer,
size_t param_buffer_size) { size_t param_buffer_size) {
prog_file_ = std::string(prog_buffer, prog_buffer + prog_buffer_size); prog_file_ = std::string(prog_buffer, prog_buffer + prog_buffer_size);
params_file_ = std::string(param_buffer, param_buffer + param_buffer_size); params_file_ = std::string(param_buffer, param_buffer + param_buffer_size);
model_from_memory_ = true; model_from_memory_ = true;
...@@ -306,7 +306,7 @@ void contrib::AnalysisConfig::SetModelBuffer(const char *prog_buffer, ...@@ -306,7 +306,7 @@ void contrib::AnalysisConfig::SetModelBuffer(const char *prog_buffer,
Update(); Update();
} }
NativeConfig contrib::AnalysisConfig::ToNativeConfig() const { NativeConfig AnalysisConfig::ToNativeConfig() const {
NativeConfig config; NativeConfig config;
config.model_dir = model_dir_; config.model_dir = model_dir_;
config.prog_file = prog_file_; config.prog_file = prog_file_;
...@@ -318,4 +318,9 @@ NativeConfig contrib::AnalysisConfig::ToNativeConfig() const { ...@@ -318,4 +318,9 @@ NativeConfig contrib::AnalysisConfig::ToNativeConfig() const {
return config; return config;
} }
void AnalysisConfig::SwitchIrDebug(int x) {
ir_debug_ = x;
Update();
}
} // namespace paddle } // namespace paddle
paddle/fluid/inference/api/analysis_predictor.cc
浏览文件 @ 17ada1d2
...@@ -47,7 +47,6 @@ DECLARE_bool(profile); ...@@ -47,7 +47,6 @@ DECLARE_bool(profile);
namespace paddle { namespace paddle {
using contrib::AnalysisConfig;
using inference::Singleton; using inference::Singleton;
#if PADDLE_WITH_TENSORRT #if PADDLE_WITH_TENSORRT
using inference::tensorrt::TRTInt8Calibrator; using inference::tensorrt::TRTInt8Calibrator;
...@@ -59,7 +58,8 @@ namespace { ...@@ -59,7 +58,8 @@ namespace {
bool IsPersistable(const framework::VarDesc *var) { bool IsPersistable(const framework::VarDesc *var) {
if (var->Persistable() && if (var->Persistable() &&
var->GetType() != framework::proto::VarType::FEED_MINIBATCH && var->GetType() != framework::proto::VarType::FEED_MINIBATCH &&
var->GetType() != framework::proto::VarType::FETCH_LIST) { var->GetType() != framework::proto::VarType::FETCH_LIST &&
var->GetType() != framework::proto::VarType::RAW) {
return true; return true;
} }
return false; return false;
...@@ -123,6 +123,15 @@ bool AnalysisPredictor::PrepareProgram( ...@@ -123,6 +123,15 @@ bool AnalysisPredictor::PrepareProgram(
if (!program) { if (!program) {
if (!LoadProgramDesc()) return false; if (!LoadProgramDesc()) return false;
// If not cloned, the parameters should be loaded.
// If config_.ir_optim() is True, parameters is loaded in
// OptimizeInferenceProgram(), but other persistable variables
// (like RAW type var) are not created in scope.
// If config_.ir_optim() is False, parameters is loaded in LoadParameters(),
// still need to create other persistable variables.
// So in both case, create persistable variables at first.
executor_->CreateVariables(*inference_program_, 0, true, sub_scope_);
// Optimize the program, and load parameters and modify them in the // Optimize the program, and load parameters and modify them in the
// scope_. // scope_.
// This will change the scope_ address. // This will change the scope_ address.
...@@ -130,15 +139,6 @@ bool AnalysisPredictor::PrepareProgram( ...@@ -130,15 +139,6 @@ bool AnalysisPredictor::PrepareProgram(
status_ir_optim_enabled_ = true; status_ir_optim_enabled_ = true;
OptimizeInferenceProgram(); OptimizeInferenceProgram();
} else { } else {
// If the parent_scope is passed, we assert that the persistable variables
// are already created, so just create the no persistable variables.
// If not cloned, the parameters should be loaded
// OptimizeInferenceProgram.
// So in both cases, just the local variables are needed to load, not the
// parematers.
executor_->CreateVariables(*inference_program_, 0, true, sub_scope_);
// Load parameters // Load parameters
LOG(INFO) << "load parameters "; LOG(INFO) << "load parameters ";
LoadParameters(); LoadParameters();
...@@ -376,7 +376,7 @@ void AnalysisPredictor::OptimizeInferenceProgram() { ...@@ -376,7 +376,7 @@ void AnalysisPredictor::OptimizeInferenceProgram() {
} }
argument_.SetIrAnalysisPasses(passes); argument_.SetIrAnalysisPasses(passes);
argument_.SetAnalysisPasses(config_.pass_builder()->AnalysisPasses()); argument_.SetAnalysisPasses(config_.pass_builder()->AnalysisPasses());
argument_.SetScopeNotOwned(const_cast<framework::Scope *>(scope_.get())); argument_.SetScopeNotOwned(scope_.get());
Analyzer().Run(&argument_); Analyzer().Run(&argument_);
PADDLE_ENFORCE(argument_.scope_valid()); PADDLE_ENFORCE(argument_.scope_valid());
...@@ -421,7 +421,7 @@ std::unique_ptr<PaddlePredictor> CreatePaddlePredictor< ...@@ -421,7 +421,7 @@ std::unique_ptr<PaddlePredictor> CreatePaddlePredictor<
if (!dynamic_cast<AnalysisPredictor *>(predictor.get())->Init(nullptr)) { if (!dynamic_cast<AnalysisPredictor *>(predictor.get())->Init(nullptr)) {
return nullptr; return nullptr;
} }
return std::move(predictor); return predictor;
} }
void AnalysisPredictor::PrepareFeedFetch() { void AnalysisPredictor::PrepareFeedFetch() {
...@@ -726,11 +726,15 @@ bool AnalysisPredictor::need_collect_var_shapes_for_memory_optim() { ...@@ -726,11 +726,15 @@ bool AnalysisPredictor::need_collect_var_shapes_for_memory_optim() {
return need; return need;
} }
std::string AnalysisPredictor::GetSeriazlizedProgram() const {
return inference_program_->Proto()->SerializeAsString();
}
template <> template <>
std::unique_ptr<PaddlePredictor> CreatePaddlePredictor<contrib::AnalysisConfig>( std::unique_ptr<PaddlePredictor> CreatePaddlePredictor<AnalysisConfig>(
const contrib::AnalysisConfig &config) { const AnalysisConfig &config) {
return CreatePaddlePredictor<contrib::AnalysisConfig, return CreatePaddlePredictor<AnalysisConfig, PaddleEngineKind::kAnalysis>(
PaddleEngineKind::kAnalysis>(config); config);
} }
} // namespace paddle } // namespace paddle
......
paddle/fluid/inference/api/analysis_predictor.h
浏览文件 @ 17ada1d2
...@@ -33,7 +33,6 @@ using inference::analysis::Argument; ...@@ -33,7 +33,6 @@ using inference::analysis::Argument;
using inference::analysis::Analyzer; using inference::analysis::Analyzer;
using framework::proto::ProgramDesc; using framework::proto::ProgramDesc;
using framework::NaiveExecutor; using framework::NaiveExecutor;
using contrib::AnalysisConfig;
/** \brief This predictor is based on the original native predictor with IR and /** \brief This predictor is based on the original native predictor with IR and
* Analysis support. * Analysis support.
...@@ -75,6 +74,8 @@ class AnalysisPredictor : public PaddlePredictor { ...@@ -75,6 +74,8 @@ class AnalysisPredictor : public PaddlePredictor {
void SetMkldnnThreadID(int tid); void SetMkldnnThreadID(int tid);
std::string GetSeriazlizedProgram() const override;
protected: protected:
// For memory optimization. // For memory optimization.
bool need_collect_var_shapes_for_memory_optim(); bool need_collect_var_shapes_for_memory_optim();
...@@ -121,7 +122,7 @@ class AnalysisPredictor : public PaddlePredictor { ...@@ -121,7 +122,7 @@ class AnalysisPredictor : public PaddlePredictor {
#endif #endif
private: private:
contrib::AnalysisConfig config_; AnalysisConfig config_;
Argument argument_; Argument argument_;
std::unique_ptr<NaiveExecutor> executor_; std::unique_ptr<NaiveExecutor> executor_;
platform::Place place_; platform::Place place_;
......
paddle/fluid/inference/api/analysis_predictor_tester.cc
浏览文件 @ 17ada1d2
...@@ -24,7 +24,6 @@ ...@@ -24,7 +24,6 @@
DEFINE_string(dirname, "", "dirname to tests."); DEFINE_string(dirname, "", "dirname to tests.");
namespace paddle { namespace paddle {
using contrib::AnalysisConfig;
TEST(AnalysisPredictor, analysis_off) { TEST(AnalysisPredictor, analysis_off) {
AnalysisConfig config; AnalysisConfig config;
...@@ -197,7 +196,7 @@ TEST(AnalysisPredictor, memory_optim) { ...@@ -197,7 +196,7 @@ TEST(AnalysisPredictor, memory_optim) {
AnalysisConfig config(FLAGS_dirname); AnalysisConfig config(FLAGS_dirname);
config.DisableGpu(); config.DisableGpu();
config.EnableMemoryOptim(true); config.EnableMemoryOptim(true);
config.pass_builder()->TurnOnDebug(); config.SwitchIrDebug();
auto native_predictor = auto native_predictor =
CreatePaddlePredictor<NativeConfig>(config.ToNativeConfig()); CreatePaddlePredictor<NativeConfig>(config.ToNativeConfig());
...@@ -215,6 +214,8 @@ TEST(AnalysisPredictor, memory_optim) { ...@@ -215,6 +214,8 @@ TEST(AnalysisPredictor, memory_optim) {
{ {
// The first predictor help to cache the memory optimize strategy. // The first predictor help to cache the memory optimize strategy.
auto predictor = CreatePaddlePredictor<AnalysisConfig>(config); auto predictor = CreatePaddlePredictor<AnalysisConfig>(config);
LOG(INFO) << "serialized program: " << predictor->GetSeriazlizedProgram();
ASSERT_FALSE(predictor->GetSeriazlizedProgram().empty());
// Run several times to check the parameters are not reused by mistake. // Run several times to check the parameters are not reused by mistake.
for (int i = 0; i < 5; i++) { for (int i = 0; i < 5; i++) {
......
paddle/fluid/inference/api/api.cc
浏览文件 @ 17ada1d2
...@@ -12,6 +12,8 @@ ...@@ -12,6 +12,8 @@
// See the License for the specific language governing permissions and // See the License for the specific language governing permissions and
// limitations under the License. // limitations under the License.
#include <sstream>
#include "paddle/fluid/framework/commit.h"
#include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/scope.h" #include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/inference/api/paddle_inference_api.h" #include "paddle/fluid/inference/api/paddle_inference_api.h"
...@@ -97,4 +99,12 @@ void PaddleBuf::Free() { ...@@ -97,4 +99,12 @@ void PaddleBuf::Free() {
} }
} }
std::string get_version() {
std::stringstream ss;
ss << "version: " << framework::paddle_version() << "\n";
ss << "commit: " << framework::paddle_commit() << "\n";
ss << "branch: " << framework::paddle_compile_branch() << "\n";
return ss.str();
}
} // namespace paddle } // namespace paddle
paddle/fluid/inference/api/api_impl_tester.cc
浏览文件 @ 17ada1d2
...@@ -295,7 +295,7 @@ TEST(inference_api_native, image_classification_gpu) { ...@@ -295,7 +295,7 @@ TEST(inference_api_native, image_classification_gpu) {
#endif #endif
TEST(PassBuilder, Delete) { TEST(PassBuilder, Delete) {
contrib::AnalysisConfig config; AnalysisConfig config;
config.DisableGpu(); config.DisableGpu();
config.pass_builder()->DeletePass("attention_lstm_fuse_pass"); config.pass_builder()->DeletePass("attention_lstm_fuse_pass");
const auto& passes = config.pass_builder()->AllPasses(); const auto& passes = config.pass_builder()->AllPasses();
......
paddle/fluid/inference/api/api_tester.cc
浏览文件 @ 17ada1d2
...@@ -61,4 +61,10 @@ TEST(paddle_inference_api, demo) { ...@@ -61,4 +61,10 @@ TEST(paddle_inference_api, demo) {
predictor->Run({}, &outputs); predictor->Run({}, &outputs);
} }
TEST(paddle_inference_api, get_version) {
LOG(INFO) << "paddle version:\n" << get_version();
auto version = get_version();
ASSERT_FALSE(version.empty());
}
} // namespace paddle } // namespace paddle
paddle/fluid/inference/api/demo_ci/trt_mobilenet_demo.cc
浏览文件 @ 17ada1d2
...@@ -36,7 +36,7 @@ namespace demo { ...@@ -36,7 +36,7 @@ namespace demo {
*/ */
void Main() { void Main() {
std::unique_ptr<PaddlePredictor> predictor; std::unique_ptr<PaddlePredictor> predictor;
paddle::contrib::AnalysisConfig config; paddle::AnalysisConfig config;
config.EnableUseGpu(100, 0); config.EnableUseGpu(100, 0);
config.SetModel(FLAGS_modeldir + "/__model__", config.SetModel(FLAGS_modeldir + "/__model__",
FLAGS_modeldir + "/__params__"); FLAGS_modeldir + "/__params__");
......
paddle/fluid/inference/api/demo_ci/vis_demo.cc
浏览文件 @ 17ada1d2
...@@ -34,7 +34,6 @@ DEFINE_bool(use_gpu, false, "Whether use gpu."); ...@@ -34,7 +34,6 @@ DEFINE_bool(use_gpu, false, "Whether use gpu.");
namespace paddle { namespace paddle {
namespace demo { namespace demo {
using contrib::AnalysisConfig;
/* /*
* Use the native and analysis fluid engine to inference the demo. * Use the native and analysis fluid engine to inference the demo.
*/ */
......
paddle/fluid/inference/api/paddle_analysis_config.h
浏览文件 @ 17ada1d2
...@@ -29,11 +29,6 @@ ...@@ -29,11 +29,6 @@
namespace paddle { namespace paddle {
class AnalysisPredictor; class AnalysisPredictor;
// ==
//
// -----------------------------------------------------------------------------------
// NOTE: The following APIs are not mature yet, we are still working on them.
namespace contrib {
// NOTE WIP, not stable yet. // NOTE WIP, not stable yet.
struct AnalysisConfig { struct AnalysisConfig {
...@@ -145,9 +140,12 @@ struct AnalysisConfig { ...@@ -145,9 +140,12 @@ struct AnalysisConfig {
*/ */
bool tensorrt_engine_enabled() const { return use_tensorrt_; } bool tensorrt_engine_enabled() const { return use_tensorrt_; }
/** Control whther to debug IR graph analysis phase. /** \brief Control whether to debug IR graph analysis phase.
*
* This will generate DOT files for visualizing the computation graph after
* each analysis pass applied.
*/ */
void SwitchIrDebug(int x = true) { ir_debug_ = x; } void SwitchIrDebug(int x = true);
/** Turn on MKLDNN. /** Turn on MKLDNN.
*/ */
...@@ -260,5 +258,4 @@ struct AnalysisConfig { ...@@ -260,5 +258,4 @@ struct AnalysisConfig {
mutable std::unique_ptr<PassStrategy> pass_builder_; mutable std::unique_ptr<PassStrategy> pass_builder_;
}; };
} // namespace contrib
} // namespace paddle } // namespace paddle
paddle/fluid/inference/api/paddle_pass_builder.h
浏览文件 @ 17ada1d2
...@@ -117,6 +117,7 @@ class CpuPassStrategy : public PassStrategy { ...@@ -117,6 +117,7 @@ class CpuPassStrategy : public PassStrategy {
"conv_bn_fuse_pass", // "conv_bn_fuse_pass", //
"conv_eltwiseadd_bn_fuse_pass", // "conv_eltwiseadd_bn_fuse_pass", //
"is_test_pass", // "is_test_pass", //
"identity_scale_op_clean_pass", //
}); });
use_gpu_ = false; use_gpu_ = false;
} }
...@@ -155,6 +156,7 @@ class GpuPassStrategy : public PassStrategy { ...@@ -155,6 +156,7 @@ class GpuPassStrategy : public PassStrategy {
GpuPassStrategy() : PassStrategy({}) { GpuPassStrategy() : PassStrategy({}) {
passes_.assign({ passes_.assign({
"infer_clean_graph_pass", // "infer_clean_graph_pass", //
"identity_scale_op_clean_pass", //
"conv_affine_channel_fuse_pass", // "conv_affine_channel_fuse_pass", //
"conv_eltwiseadd_affine_channel_fuse_pass", // "conv_eltwiseadd_affine_channel_fuse_pass", //
"conv_bn_fuse_pass", // "conv_bn_fuse_pass", //
......
paddle/fluid/inference/tensorrt/trt_int8_calibrator.h
浏览文件 @ 17ada1d2
...@@ -13,16 +13,16 @@ ...@@ -13,16 +13,16 @@
// limitations under the License. // limitations under the License.
#pragma once #pragma once
#include <NvInfer.h>
#include <cuda_runtime_api.h>
#include <atomic> #include <atomic>
#include <memory> #include <memory>
#include <mutex> #include <mutex> // NOLINT
#include <string> #include <string>
#include <unordered_map> #include <unordered_map>
#include <utility> #include <utility>
#include <vector> #include <vector>
#include <NvInfer.h>
#include <cuda_runtime_api.h>
#include "paddle/fluid/framework/tensor.h" #include "paddle/fluid/framework/tensor.h"
#include "paddle/fluid/inference/tensorrt/engine.h" #include "paddle/fluid/inference/tensorrt/engine.h"
#include "paddle/fluid/platform/place.h" #include "paddle/fluid/platform/place.h"
......
paddle/fluid/inference/tests/api/CMakeLists.txt
浏览文件 @ 17ada1d2
...@@ -128,6 +128,11 @@ inference_analysis_api_test_with_fake_data(test_analyzer_resnet50 ...@@ -128,6 +128,11 @@ inference_analysis_api_test_with_fake_data(test_analyzer_resnet50
inference_analysis_api_test_with_fake_data(test_analyzer_mobilenet_depthwise_conv inference_analysis_api_test_with_fake_data(test_analyzer_mobilenet_depthwise_conv
"${INFERENCE_DEMO_INSTALL_DIR}/mobilenet_depthwise_conv" analyzer_resnet50_tester.cc "mobilenet_model.tar.gz" SERIAL) "${INFERENCE_DEMO_INSTALL_DIR}/mobilenet_depthwise_conv" analyzer_resnet50_tester.cc "mobilenet_model.tar.gz" SERIAL)
# bert, max_len=20, embedding_dim=128
set(BERT_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/bert_emb128")
download_model_and_data(${BERT_INSTALL_DIR} "bert_emb128_model.tar.gz" "bert_data_len20.txt.tar.gz")
inference_analysis_api_test(test_analyzer_bert ${BERT_INSTALL_DIR} analyzer_bert_tester.cc SERIAL)
# anakin # anakin
if (WITH_ANAKIN AND WITH_MKL) # only needed in CI if (WITH_ANAKIN AND WITH_MKL) # only needed in CI
# anakin rnn1 # anakin rnn1
......
paddle/fluid/inference/tests/api/analyzer_lac_tester.cc
浏览文件 @ 17ada1d2
...@@ -18,8 +18,6 @@ namespace paddle { ...@@ -18,8 +18,6 @@ namespace paddle {
namespace inference { namespace inference {
namespace analysis { namespace analysis {
using contrib::AnalysisConfig;
struct DataRecord { struct DataRecord {
std::vector<int64_t> data; std::vector<int64_t> data;
std::vector<size_t> lod; std::vector<size_t> lod;
......
paddle/fluid/operators/conv_op.cc
浏览文件 @ 17ada1d2
此差异已折叠。
paddle/fluid/operators/detection/box_clip_op.cc 0 → 100644
浏览文件 @ 17ada1d2
此差异已折叠。
paddle/fluid/operators/detection/box_clip_op.cu 0 → 100644
浏览文件 @ 17ada1d2
此差异已折叠。
paddle/fluid/operators/detection/box_clip_op.h 0 → 100644
浏览文件 @ 17ada1d2
此差异已折叠。
paddle/fluid/operators/detection/yolov3_loss_op.h 0 → 100644
浏览文件 @ 17ada1d2
此差异已折叠。
paddle/fluid/operators/expand_op.cc
浏览文件 @ 17ada1d2
此差异已折叠。
paddle/fluid/operators/expand_op.cu
浏览文件 @ 17ada1d2
此差异已折叠。
paddle/fluid/operators/gather_op.cu
浏览文件 @ 17ada1d2
此差异已折叠。
paddle/fluid/operators/gather_op.h
浏览文件 @ 17ada1d2
此差异已折叠。
paddle/fluid/operators/jit/gen/hopv.cc 0 → 100644
浏览文件 @ 17ada1d2
此差异已折叠。
paddle/fluid/operators/jit/gen/hopv.h 0 → 100644
浏览文件 @ 17ada1d2
此差异已折叠。
paddle/fluid/operators/jit/gen/matmul.cc 0 → 100644
浏览文件 @ 17ada1d2
此差异已折叠。
paddle/fluid/operators/jit/gen/matmul.h 0 → 100644
浏览文件 @ 17ada1d2
此差异已折叠。
paddle/fluid/operators/jit/helper.h
浏览文件 @ 17ada1d2
此差异已折叠。
paddle/fluid/operators/jit/test.cc
浏览文件 @ 17ada1d2
此差异已折叠。
paddle/fluid/operators/ngraph/ops/accuracy_op.h 0 → 100644
浏览文件 @ 17ada1d2
此差异已折叠。
paddle/fluid/operators/ngraph/ops/activation_op.h 0 → 100644
浏览文件 @ 17ada1d2
此差异已折叠。
paddle/fluid/operators/ngraph/ops/batch_norm_op.h 0 → 100644
浏览文件 @ 17ada1d2
此差异已折叠。
paddle/fluid/operators/ngraph/ops/conv2d_op.h 0 → 100644
浏览文件 @ 17ada1d2
此差异已折叠。
paddle/fluid/operators/ngraph/ops/pool2d_op.h 0 → 100644
浏览文件 @ 17ada1d2
此差异已折叠。
paddle/fluid/operators/ngraph/ops/sum_op.h 0 → 100644
浏览文件 @ 17ada1d2
此差异已折叠。
paddle/fluid/operators/norm_op.h
浏览文件 @ 17ada1d2
此差异已折叠。
paddle/fluid/operators/pool_op.cc
浏览文件 @ 17ada1d2
此差异已折叠。
paddle/fluid/operators/scale_op.cc
浏览文件 @ 17ada1d2
此差异已折叠。
paddle/fluid/operators/slice_op.cc
浏览文件 @ 17ada1d2
此差异已折叠。
paddle/fluid/operators/stack_op.cu
浏览文件 @ 17ada1d2
此差异已折叠。
paddle/fluid/operators/yolov3_loss_op.h 已删除 100644 → 0
浏览文件 @ d8291fa3
此差异已折叠。
paddle/fluid/platform/enforce.h
浏览文件 @ 17ada1d2
此差异已折叠。
paddle/fluid/platform/nccl_helper.h
浏览文件 @ 17ada1d2
此差异已折叠。
paddle/fluid/platform/place.cc
浏览文件 @ 17ada1d2
此差异已折叠。
paddle/fluid/pybind/CMakeLists.txt
浏览文件 @ 17ada1d2
此差异已折叠。
paddle/fluid/pybind/pybind.cc
浏览文件 @ 17ada1d2
此差异已折叠。
paddle/fluid/string/printf.h
浏览文件 @ 17ada1d2
此差异已折叠。
paddle/scripts/fast_install.sh 0 → 100644
浏览文件 @ 17ada1d2
此差异已折叠。
paddle/scripts/paddle_build.sh
浏览文件 @ 17ada1d2
此差异已折叠。
python/CMakeLists.txt
浏览文件 @ 17ada1d2
此差异已折叠。
python/paddle/__init__.py
浏览文件 @ 17ada1d2
此差异已折叠。
python/paddle/distributed/__init__.py 0 → 100644
浏览文件 @ 17ada1d2
此差异已折叠。
python/paddle/fluid/__init__.py
浏览文件 @ 17ada1d2
此差异已折叠。
python/paddle/fluid/compiler.py
浏览文件 @ 17ada1d2
此差异已折叠。
python/paddle/fluid/contrib/int8_inference/README.md 0 → 100644
浏览文件 @ 17ada1d2
此差异已折叠。
python/paddle/fluid/executor.py
浏览文件 @ 17ada1d2
此差异已折叠。
python/paddle/fluid/framework.py
浏览文件 @ 17ada1d2
此差异已折叠。
python/paddle/fluid/initializer.py
浏览文件 @ 17ada1d2
此差异已折叠。
python/paddle/fluid/io.py
浏览文件 @ 17ada1d2
此差异已折叠。
python/paddle/fluid/layer_helper.py
浏览文件 @ 17ada1d2
此差异已折叠。
python/paddle/fluid/layers/io.py
浏览文件 @ 17ada1d2
此差异已折叠。
python/paddle/fluid/layers/nn.py
浏览文件 @ 17ada1d2
此差异已折叠。
python/paddle/fluid/layers/ops.py
浏览文件 @ 17ada1d2
此差异已折叠。
python/paddle/fluid/optimizer.py
浏览文件 @ 17ada1d2
此差异已折叠。
python/paddle/fluid/profiler.py
浏览文件 @ 17ada1d2
此差异已折叠。
python/paddle/fluid/tests/unittests/mkldnn/__init__.py 0 → 100644
浏览文件 @ 17ada1d2
此差异已折叠。
python/paddle/fluid/tests/unittests/test_box_clip_op.py 0 → 100644
浏览文件 @ 17ada1d2
此差异已折叠。
python/paddle/fluid/unique_name.py
浏览文件 @ 17ada1d2
此差异已折叠。
python/paddle/fluid/wrapped_decorator.py 0 → 100644
浏览文件 @ 17ada1d2
此差异已折叠。
python/requirements.txt
浏览文件 @ 17ada1d2
此差异已折叠。
python/setup.py.in
浏览文件 @ 17ada1d2
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