diff --git a/.github/ISSUE_TEMPLATE/---feature-request-.md b/.github/ISSUE_TEMPLATE/---feature-request-.md new file mode 100644 index 0000000000000..57708855dce4f --- /dev/null +++ b/.github/ISSUE_TEMPLATE/---feature-request-.md @@ -0,0 +1,27 @@ +--- +name: 建议(Feature request) +about: 您可以提出您的建议。 You could use this template for reporting a suggestion  issue. + +--- + +欢迎您对PaddlePaddle提出建议，非常感谢您对PaddlePaddle的贡献！ +在留下您的建议时，辛苦您同步提供如下信息： +- 版本、环境信息 +1）PaddlePaddle版本：请提供您的PaddlePaddle版本号，例如1.1 +2）CPU/GPU：您是否使用GPU进行训练，如是，请提供您的CUDA和cuDNN版本号 +3）系统环境：请您描述系统类型、版本，例如Mac OS 10.14 +- 复现信息：如为报错，请给出复现环境、复现步骤 +- 建议描述：请您详细描述，您认为需优化的功能 + +Thank you for contributing to PaddlePaddle. +Before submitting the issue, you could search issue in the github in case that there was a similar issue submitted or resolved before. +Please make sure that this is a feature request. +**System information** +-PaddlePaddle version （eg.1.1）or CommitID +-CPU: including CPUMKL/OpenBlas/MKLDNN version +-GPU: including CUDA/CUDNN version +-OS Platform (eg.Mac OS 10.14) +**To Reproduce** +Steps to reproduce the behavior +**Describe the feature and the current behavior/state.** +**Any Other info.** diff --git a/.github/ISSUE_TEMPLATE/---inference-issue-.md b/.github/ISSUE_TEMPLATE/---inference-issue-.md new file mode 100644 index 0000000000000..37bdc8889e272 --- /dev/null +++ b/.github/ISSUE_TEMPLATE/---inference-issue-.md @@ -0,0 +1,40 @@ +--- +name: 预测（Inference Issue） +about: 您可以提问预测中报错、应用等问题。 You could use this template for reporting an inference issue. + +--- + +为使您的问题得到快速解决，在建立Issue前，请您先通过如下方式搜索是否有相似问题:【搜索issue关键字】【使用labels筛选】【官方文档】 + +如果您没有查询到相似问题，为快速解决您的提问，建立issue时请提供如下细节信息： +- 标题：简洁、精准描述您的问题，例如“最新预测库的API文档在哪儿 ” +- 版本、环境信息： +    1）PaddlePaddle版本：请提供您的PaddlePaddle版本号（如1.1）或CommitID +    2）CPU：预测若用CPU，请提供CPU型号，MKL/OpenBlas/MKLDNN/等数学库使用情况 +    3）GPU：预测若用GPU，请提供GPU型号、CUDA和CUDNN版本号 +    4）系统环境：请您描述系统类型、版本（如Mac OS 10.14），Python版本 +-预测信息 +    1）C++预测：请您提供预测库安装包的版本信息，及其中的version.txt文件 +    2）CMake包含路径的完整命令 +    3）API信息（如调用请提供） +    4）预测库来源：官网下载/特殊环境（如BCLOUD编译） +- 复现信息：如为报错，请给出复现环境、复现步骤 +- 问题描述：请详细描述您的问题，同步贴出报错信息、日志/代码关键片段 + +Thank you for contributing to PaddlePaddle. +Before submitting the issue, you could search issue in the github in case that th +If there is no solution,please make sure that this is an inference issue including the following details : +**System information** +-PaddlePaddle version （eg.1.1）or CommitID +-CPU: including CPUMKL/OpenBlas/MKLDNN version +-GPU: including CUDA/CUDNN version +-OS Platform (eg.Mac OS 10.14) +-Python version +-Cmake orders +-C++version.txt +-API information +**To Reproduce** +Steps to reproduce the behavior +**Describe your current behavior** +**Code to reproduce the issue** +**Other info / logs** diff --git a/.github/ISSUE_TEMPLATE/---installation-issue-.md b/.github/ISSUE_TEMPLATE/---installation-issue-.md new file mode 100644 index 0000000000000..ce4ba58932467 --- /dev/null +++ b/.github/ISSUE_TEMPLATE/---installation-issue-.md @@ -0,0 +1,40 @@ +--- +name: 安装（Installation Issue） +about: 您可以提问安装、编译出现报错等问题。 You could use this template for reporting an installation +  issue. + +--- + +为使您的问题得到快速解决，在建立Issue前，请您先通过如下方式搜索是否有相似问题:【搜索issue关键字】【使用labels筛选】【官方文档】 + +建立issue时，为快速解决问题，请您根据使用情况给出如下信息： +- 标题：请包含关键词“安装错误”/“编译错误”，例如“Mac编译错误” +- 版本、环境信息： +    1）PaddlePaddle版本：请提供您的PaddlePaddle版本号（如1.1）或CommitID +    2）CPU：请提供CPU型号，MKL/OpenBlas/MKLDNN/等数学库的使用情况 +    3）GPU：请提供GPU型号，CUDA和CUDNN版本号 +    4）系统环境：请说明系统类型、版本（如Mac OS 10.14）、Python版本 +- 安装方式信息： +1）pip安装/docker安装 +2）本地编译：请提供cmake命令，编译命令 +3）docker编译：请提供docker镜像，编译命令            +  特殊环境请注明：如离线安装等 +- 复现信息：如为报错，请给出复现环境、复现步骤 +- 问题描述：请详细描述您的问题，同步贴出报错信息、日志/代码关键片段 + +Thank you for contributing to PaddlePaddle. +Before submitting the issue, you could search issue in Github in case that there was a similar issue submitted or resolved before. +If there is no solution,please make sure that this is an installation issue including the following details: +**System information** +-PaddlePaddle version （eg.1.1）or CommitID +-CPU: including CPUMKL/OpenBlas/MKLDNN version +-GPU: including CUDA/CUDNN version +-OS Platform (eg. Mac OS 10.14) +-Python version +- Install method: pip install/install with docker/build from source(without docker)/build within docker +- Other special cases that you think may be related to this problem, eg. offline install, special internet condition   +**To Reproduce** +Steps to reproduce the behavior +**Describe your current behavior** +**Code to reproduce the issue** +**Other info / logs** diff --git a/.github/ISSUE_TEMPLATE/---model-issue-.md b/.github/ISSUE_TEMPLATE/---model-issue-.md new file mode 100644 index 0000000000000..7cb52f37b9026 --- /dev/null +++ b/.github/ISSUE_TEMPLATE/---model-issue-.md @@ -0,0 +1,36 @@ +--- +name: 模型（Model Issue） +about: 您可以提问模型、算法、数据集方向的使用报错等问题。You could use this template for reporting a model/ + algorithm/dataset  issue. + +--- + +为使您的问题得到快速解决，在建立Issue前，请您先通过如下方式搜索是否有相似问题:【搜索issue关键字】【使用labels筛选】【官方文档】 + +建立issue时，为快速解决问题，请您根据使用情况给出如下信息： +- 标题：简洁、精准描述您的问题，例如“ssd 模型前置lstm报错  ” +- 版本、环境信息： +    1）PaddlePaddle版本：请提供PaddlePaddle版本号，例如1.1或CommitID +    2）CPU：请提供CPU型号，MKL/OpenBlas/MKLDNN/等数学库的使用情况 +    3）GPU：请提供GPU型号，CUDA和CUDNN版本号 +    4）系统环境：请说明系统类型、版本（例如Mac OS 10.14），Python版本 +- 模型信息 +    1）模型名称 2）使用数据集名称 3）使用算法名称 4）模型链接 +- 复现信息：如为报错，请给出复现环境、复现步骤 +- 问题描述：请详细描述您的问题，同步贴出报错信息、日志/代码关键片段 + +Thank you for contributing to PaddlePaddle. +Before submitting the issue, you could search issue in the github.Probably there was a similar issue submitted or resolved before. +If there is no solution,please make sure that this is a issue of models including the following details: +**System information** +-PaddlePaddle version （eg.1.1）or CommitID +-CPU: including CPUMKL/OpenBlas/MKLDNN version +-GPU: including CUDA/CUDNN version +-OS Platform (eg.Mac OS 10.14) +-Python version +-Name of Models&Dataset/details of operator +**To Reproduce** +Steps to reproduce the behavior +**Describe your current behavior** +**Code to reproduce the issue** +**Other info / logs** diff --git a/.github/ISSUE_TEMPLATE/---others-.md b/.github/ISSUE_TEMPLATE/---others-.md new file mode 100644 index 0000000000000..6a291153e43f5 --- /dev/null +++ b/.github/ISSUE_TEMPLATE/---others-.md @@ -0,0 +1,33 @@ +--- +name: 其他（Others） +about: 如上述分类未包含您的问题，可在此提出。 You could use this template for reporting other issues + +--- + +为使您的问题得到快速解决，在建立Issues前，请您先通过如下方式搜索是否有相似问题:【搜索issue关键字】【使用labels筛选】【官方文档】 + +如果您没有查询到相似问题，为快速解决您的提问，建立issue时请提供如下细节信息： +- 标题：简洁、精准概括您的问题 +- 版本、环境信息： +    1）PaddlePaddle版本：请提供您的PaddlePaddle版本号，例如1.1或CommitID +    2）CPU/GPU：如果您使用GPU训练，请提供GPU驱动版本、CUDA和cuDNN版本号 +    3）系统环境：请您描述系统类型、版本，例如Mac OS 10.14 +    4）Python版本号 +    5）显存信息 +- 复现信息：如为报错，请给出复现环境、复现步骤 +- 问题描述：请详细描述您的问题，同步贴出报错信息、日志/代码关键片段 + +Thank you for contributing to PaddlePaddle. +Before submitting the issue, you could search issue in the github in case that there was a similar issue submitted or resolved before. +If there is no solution,please provide us with the following details : +**System information** +-PaddlePaddle version （eg.1.1）or CommitID +-CPU: including CPUMKL/OpenBlas/MKLDNN version +-GPU: including CUDA/cuDNN version +-OS Platform and Distribution(eg.Mac OS 10.14) +-Python version +**To Reproduce** +Steps to reproduce the behavior +**Describe your current behavior** +**Code to reproduce the issue** +**Other info / logs** diff --git a/.github/ISSUE_TEMPLATE/---training-issue-.md b/.github/ISSUE_TEMPLATE/---training-issue-.md new file mode 100644 index 0000000000000..29e8383d97792 --- /dev/null +++ b/.github/ISSUE_TEMPLATE/---training-issue-.md @@ -0,0 +1,38 @@ +--- +name: 训练（Training issue） +about: 您可以提问训练中报错、应用、出core等问题。 You could use this template for reporting an training +  issue. + +--- + +为使您的问题得到快速解决，在建立Issues前，请您先通过如下方式搜索是否有相似问题:【搜索issue关键字】【使用labels筛选】【官方文档】 + +如果您没有查询到相似问题，为快速解决您的提问，建立issue时请提供如下细节信息： +- 标题：简洁、精准概括您的问题，例如“Insufficient Memory xxx" ” +- 版本、环境信息： +    1）PaddlePaddle版本：请提供您的PaddlePaddle版本号，例如1.1或CommitID +    2）CPU：预测若用CPU，请提供CPU型号，MKL/OpenBlas/MKLDNN/等数学库使用情况 +    3）GPU：预测若用GPU，请提供GPU型号、CUDA和CUDNN版本号 +    4）系统环境：请您描述系统类型、版本，例如Mac OS 10.14，Python版本 +- 训练信息 +    1）单机/多机，单卡/多卡 +    2）显存信息 +    3）Operator信息 +- 复现信息：如为报错，请给出复现环境、复现步骤 +- 问题描述：请详细描述您的问题，同步贴出报错信息、日志、可复现的代码片段 + +Thank you for contributing to PaddlePaddle. +Before submitting the issue, you could search issue in the github in case that there was a similar issue submitted or resolved before. +If there is no solution,please make sure that this is a training issue including the following details: +**System information** +-PaddlePaddle version （eg.1.1）or CommitID +-CPU: including CPUMKL/OpenBlas/MKLDNN version +-GPU: including CUDA/CUDNN version +-OS Platform (eg.Mac OS 10.14) +-Other imformation: Distriuted training/informantion of operator/ +Graphics card storage +**To Reproduce** +Steps to reproduce the behavior +**Describe your current behavior** +**Code to reproduce the issue** +**Other info / logs** diff --git a/CMakeLists.txt b/CMakeLists.txt index bc2ac2cd93969..66dcef0013efb 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -54,7 +54,7 @@ option(WITH_PYTHON "Compile PaddlePaddle with python interpreter" ON) option(WITH_DOUBLE "Compile PaddlePaddle with double precision" OFF) option(WITH_RDMA "Compile PaddlePaddle with RDMA support" OFF) option(WITH_TIMER "Compile PaddlePaddle with stats timer" OFF) -option(WITH_PROFILER "Compile PaddlePaddle with GPU profiler" OFF) +option(WITH_PROFILER "Compile PaddlePaddle with GPU profiler and gperftools" OFF) option(WITH_DOC "Compile PaddlePaddle with documentation" OFF) option(WITH_COVERAGE "Compile PaddlePaddle with code coverage" OFF) option(COVERALLS_UPLOAD "Package code coverage data to coveralls" OFF) @@ -65,6 +65,7 @@ option(WITH_GOLANG "Compile PaddlePaddle with GOLANG" OFF) option(GLIDE_INSTALL "Download and install go dependencies " ON) option(USE_NNPACK "Compile PaddlePaddle with NNPACK library" OFF) option(WITH_DISTRIBUTE "Compile with distributed support" OFF) +option(WITH_PSLIB "Compile with pslib support" OFF) option(USE_EIGEN_FOR_BLAS "Use matrix multiplication in Eigen" OFF) option(EIGEN_USE_THREADS "Compile with multi-threaded Eigen" OFF) option(WITH_ARM_FP16 "Use half precision support on armv8.2-a cpu" OFF) @@ -125,18 +126,12 @@ if(ANDROID OR IOS) add_definitions(-DPADDLE_MOBILE_INFERENCE) endif() -if (APPLE OR WIN32) +if (APPLE) set(WITH_MKL OFF CACHE STRING - "Disable MKL for building on mac and windows" FORCE) + "Disable MKL for building on mac" FORCE) endif() if (WIN32) - set(WITH_AVX OFF CACHE STRING - "Disable AVX when compiling for Windows" FORCE) - set(WITH_DSO OFF CACHE STRING - "Disable DSO when compiling for Windows" FORCE) - set(WITH_MKL OFF CACHE STRING - "Disable MKL when compiling for Windows" FORCE) set(WITH_DISTRIBUTE OFF CACHE STRING "Disable DISTRIBUTE when compiling for Windows" FORCE) set(WITH_C_API OFF CACHE STRING @@ -204,16 +199,25 @@ include(external/eigen) # download eigen3 include(external/pybind11) # download pybind11 include(external/cares) include(external/cub) +include(external/rocprim) include(external/xxhash) # download xxhash +include(external/dlpack) include(external/snappy) # download snappy include(external/snappystream) # download snappystream +include(external/warpctc) # download, build, install warpctc if (NOT WIN32) -# there is no official support of warpctc, nccl, cupti in windows -include(external/warpctc) # download, build, install warpctc +# there is no official support of nccl, cupti in windows include(cupti) +include(external/gzstream) endif (NOT WIN32) +if(WITH_PSLIB) + include(external/libmct) + include(external/pslib_brpc) + include(external/pslib) +endif(WITH_PSLIB) + if(WITH_DISTRIBUTE) if(WITH_GRPC) include(external/grpc) @@ -251,6 +255,12 @@ elseif() set(WITH_ANAKIN OFF CACHE STRING "Anakin is used in MKL only now." FORCE) endif() +if (WITH_PROFILER) + find_package(Gperftools REQUIRED) + include_directories(${GPERFTOOLS_INCLUDE_DIR}) + add_definitions(-DWITH_GPERFTOOLS) +endif() + include(generic) # simplify cmake module include(package) # set paddle packages include(ccache) # set ccache for compilation @@ -275,6 +285,12 @@ set(EXTERNAL_LIBS ${PYTHON_LIBRARIES} ) +if(WITH_PSLIB) + list(APPEND EXTERNAL_LIBS pslib) + list(APPEND EXTERNAL_LIBS pslib_brpc) + list(APPEND EXTERNAL_LIBS libmct) +endif(WITH_PSLIB) + if(WITH_AMD_GPU) find_package(HIP) include(hip) diff --git a/Dockerfile b/Dockerfile index c8b9eed6d60e5..84e1edbee91b0 100644 --- a/Dockerfile +++ b/Dockerfile @@ -22,6 +22,29 @@ ENV HOME /root # Add bash enhancements COPY ./paddle/scripts/docker/root/ /root/ +# Prepare packages for Python +RUN apt-get update && \ + apt-get install -y make build-essential libssl-dev zlib1g-dev libbz2-dev \ + libreadline-dev libsqlite3-dev wget curl llvm libncurses5-dev libncursesw5-dev \ + xz-utils tk-dev libffi-dev liblzma-dev + +# Install Python3.6 +RUN mkdir -p /root/python_build/ && wget -q https://www.sqlite.org/2018/sqlite-autoconf-3250300.tar.gz && \ + tar -zxf sqlite-autoconf-3250300.tar.gz && cd sqlite-autoconf-3250300 && \ + ./configure -prefix=/usr/local && make -j8 && make install && cd ../ && rm sqlite-autoconf-3250300.tar.gz && \ + wget -q https://www.python.org/ftp/python/3.6.0/Python-3.6.0.tgz && \ + tar -xzf Python-3.6.0.tgz && cd Python-3.6.0 && \ + CFLAGS="-Wformat" ./configure --prefix=/usr/local/ --enable-shared > /dev/null && \ + make -j8 > /dev/null && make altinstall > /dev/null + +# Install Python3.7 +RUN wget -q https://www.python.org/ftp/python/3.7.0/Python-3.7.0.tgz && \ + tar -xzf Python-3.7.0.tgz && cd Python-3.7.0 && \ + CFLAGS="-Wformat" ./configure --prefix=/usr/local/ --enable-shared > /dev/null && \ + make -j8 > /dev/null && make altinstall > /dev/null + +RUN rm -r /root/python_build + RUN apt-get update && \ apt-get install -y --allow-downgrades patchelf \ python3 python3-dev python3-pip \ @@ -74,6 +97,12 @@ RUN localedef -i en_US -f UTF-8 en_US.UTF-8 RUN pip3 install -U wheel && \ pip3 install -U docopt PyYAML sphinx==1.5.6 && \ pip3 install sphinx-rtd-theme==0.1.9 recommonmark && \ + pip3.6 install -U wheel && \ + pip3.6 install -U docopt PyYAML sphinx==1.5.6 && \ + pip3.6 install sphinx-rtd-theme==0.1.9 recommonmark && \ + pip3.7 install -U wheel && \ + pip3.7 install -U docopt PyYAML sphinx==1.5.6 && \ + pip3.7 install sphinx-rtd-theme==0.1.9 recommonmark && \ easy_install -U pip && \ pip install -U pip setuptools wheel && \ pip install -U docopt PyYAML sphinx==1.5.6 && \ @@ -82,22 +111,34 @@ RUN pip3 install -U wheel && \ RUN pip3 install 'pre-commit==1.10.4' 'ipython==5.3.0' && \ pip3 install 'ipykernel==4.6.0' 'jupyter==1.0.0' && \ pip3 install opencv-python && \ + pip3.6 install 'pre-commit==1.10.4' 'ipython==5.3.0' && \ + pip3.6 install 'ipykernel==4.6.0' 'jupyter==1.0.0' && \ + pip3.6 install opencv-python && \ + pip3.7 install 'pre-commit==1.10.4' 'ipython==5.3.0' && \ + pip3.7 install 'ipykernel==4.6.0' 'jupyter==1.0.0' && \ + pip3.7 install opencv-python && \ pip install 'pre-commit==1.10.4' 'ipython==5.3.0' && \ pip install 'ipykernel==4.6.0' 'jupyter==1.0.0' && \ pip install opencv-python #For docstring checker RUN pip3 install pylint pytest astroid isort +RUN pip3.6 install pylint pytest astroid isort +RUN pip3.7 install pylint pytest astroid isort RUN pip install pylint pytest astroid isort LinkChecker COPY ./python/requirements.txt /root/ RUN pip3 install -r /root/requirements.txt +RUN pip3.6 install -r /root/requirements.txt +RUN pip3.7 install -r /root/requirements.txt RUN pip install -r /root/requirements.txt # To fix https://github.com/PaddlePaddle/Paddle/issues/1954, we use # the solution in https://urllib3.readthedocs.io/en/latest/user-guide.html#ssl-py2 RUN apt-get install -y libssl-dev libffi-dev RUN pip3 install certifi urllib3[secure] +RUN pip3.6 install certifi urllib3[secure] +RUN pip3.7 install certifi urllib3[secure] RUN pip install certifi urllib3[secure] diff --git a/README.md b/README.md index 56d6c10c64278..32a302cc5431a 100644 --- a/README.md +++ b/README.md @@ -2,8 +2,8 @@ [![Build Status](https://travis-ci.org/PaddlePaddle/Paddle.svg?branch=develop)](https://travis-ci.org/PaddlePaddle/Paddle) -[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://paddlepaddle.org/documentation/docs/en/1.1/getstarted/index_en.html) -[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://paddlepaddle.org/documentation/docs/zh/1.1/beginners_guide/index.html) +[![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) @@ -19,7 +19,16 @@ Our vision is to enable deep learning for everyone via PaddlePaddle. Please refer to our [release announcement](https://github.com/PaddlePaddle/Paddle/releases) to track the latest feature of PaddlePaddle. -### Latest PaddlePaddle Release: [Fluid 1.1.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.1) +欢迎来到 PaddlePaddle GitHub + +PaddlePaddle (PArallel Distributed Deep LEarning) 是一个简单易用、高效灵活、可扩展的深度学习平台，最初由百度科学家和工程师共同开发，目的是将深度学习技术应用到百度的众多产品中。 + +我们的愿景是让每个人都能通过PaddlePaddle接触深度学习 + +跟进PaddlePaddle最新特性请参考我们的[版本说明](https://github.com/PaddlePaddle/Paddle/releases) + + +### Latest PaddlePaddle Release: [Fluid 1.2.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.2) ### Install Latest Stable Release: ``` # Linux CPU @@ -27,13 +36,30 @@ pip install paddlepaddle # Linux GPU cuda9cudnn7 pip install paddlepaddle-gpu # Linux GPU cuda8cudnn7 -pip install paddlepaddle-gpu==1.1.0.post87 +pip install paddlepaddle-gpu==1.2.0.post87 # Linux GPU cuda8cudnn5 -pip install paddlepaddle-gpu==1.1.0.post85 +pip install paddlepaddle-gpu==1.2.0.post85 # For installation on other platform, refer to http://paddlepaddle.org/ ``` + +### PaddlePaddle最新版本: [Fluid 1.2.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.2) +### 安装最新稳定版本: +``` +# Linux CPU +pip install paddlepaddle +# Linux GPU cuda9cudnn7 +pip install paddlepaddle-gpu +# Linux GPU cuda8cudnn7 +pip install paddlepaddle-gpu==1.2.0.post87 +# Linux GPU cuda8cudnn5 +pip install paddlepaddle-gpu==1.2.0.post85 + +# 其他平台上的安装指引请参考 http://paddlepaddle.org/ +``` + + ## Features - **Flexibility** @@ -74,35 +100,90 @@ pip install paddlepaddle-gpu==1.1.0.post85 Baidu and it has achieved a significant impact. We hope you can also explore the capability of PaddlePaddle to make an impact on your product. +## 特点 + +- **灵活性** + + PaddlePaddle支持丰富的神经网络架构和优化算法。易于配置复杂模型，例如带有注意力机制或复杂记忆连接的神经网络机器翻译模型。 + +- **高效性** + + 为了高效使用异步计算资源，PaddlePaddle对框架的不同层进行优化，包括计算、存储、架构和通信。下面是一些样例： + + - 通过SSE/AVX 内置函数、BLAS库(例如MKL、OpenBLAS、cuBLAS)或定制的CPU/GPU内核优化数学操作。 + - 通过MKL-DNN库优化CNN网络 + - 高度优化循环网络，无需执行 `padding` 操作即可处理 **变长** 序列 + - 针对高维稀疏数据模型，优化了局部和分布式训练。 + + +- **稳定性** + + 有了 PaddlePaddle，使得利用各种CPU/GPU和机器来加速训练变得简单。PaddlePaddle 通过优化通信可以实现巨大吞吐量和快速执行。 + +- **连接产品** + + 另外，PaddlePaddle 的设计也易于部署。在百度，PaddlePaddle 已经部署到含有巨大用户量的产品和服务上，包括广告点击率（CTR）预测、大规模图像分类、光学字符识别（OCR）、搜索排序，计算机病毒检测、推荐系统等等。PaddlePaddle广泛应用于百度产品中，产生了非常重要的影响。我们希望您也能探索 PaddlePaddle 的能力，为您的产品创造新的影响力和效果。 + ## Installation -It is recommended to read [this doc](http://paddlepaddle.org/documentation/docs/zh/1.1/beginners_guide/index.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 -We provide [English](http://paddlepaddle.org/documentation/docs/en/1.1/getstarted/index_en.html) and -[Chinese](http://paddlepaddle.org/documentation/docs/zh/1.1/beginners_guide/index.html) documentation. +We provide [English](http://paddlepaddle.org/documentation/docs/en/1.2/getstarted/index_en.html) and +[Chinese](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/index.html) documentation. - [Deep Learning 101](https://github.com/PaddlePaddle/book) You might want to start from this online interactive book that can run in a Jupyter Notebook. -- [Distributed Training](http://paddlepaddle.org/documentation/docs/zh/1.1/user_guides/howto/training/cluster_howto.html) +- [Distributed Training](http://paddlepaddle.org/documentation/docs/zh/1.2/user_guides/howto/training/cluster_howto.html) You can run distributed training jobs on MPI clusters. -- [Python API](http://paddlepaddle.org/documentation/api/zh/1.1/fluid.html) +- [Python API](http://paddlepaddle.org/documentation/docs/zh/1.2/api_cn/index_cn.html) Our new API enables much shorter programs. -- [How to Contribute](http://paddlepaddle.org/documentation/docs/zh/1.1/advanced_usage/development/contribute_to_paddle.html) +- [How to Contribute](http://paddlepaddle.org/documentation/docs/zh/1.2/advanced_usage/development/contribute_to_paddle/index_cn.html) We appreciate your contributions! +## 文档 + +我们提供[英文](http://paddlepaddle.org/documentation/docs/en/1.2/getstarted/index_en.html)和 +[中文](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/index.html) 文档 + +- [深度学习101](https://github.com/PaddlePaddle/book) + + 或许您想从这个在线交互式书籍开始，可以在Jupyter Notebook中运行 + +- [分布式训练](http://paddlepaddle.org/documentation/docs/zh/1.2/user_guides/howto/training/cluster_howto.html) + + 可以在MPI集群上运行分布式训练任务 + +- [Python API](http://paddlepaddle.org/documentation/docs/zh/1.2/api_cn/index_cn.html) + + 新的API支持代码更少更简洁的程序 + +- [贡献方式](http://paddlepaddle.org/documentation/docs/zh/1.2/advanced_usage/development/contribute_to_paddle/index_cn.html) + + 欢迎您的贡献! ## Ask Questions You are welcome to submit questions and bug reports as [Github Issues](https://github.com/PaddlePaddle/Paddle/issues). +## 答疑 + +欢迎您将问题和bug报告以[Github Issues](https://github.com/PaddlePaddle/Paddle/issues)的形式提交 + ## Copyright and License PaddlePaddle is provided under the [Apache-2.0 license](LICENSE). + +## 版权和许可证 +PaddlePaddle由[Apache-2.0 license](LICENSE)提供 diff --git a/benchmark/fluid/fluid_benchmark.py b/benchmark/fluid/fluid_benchmark.py index 5f3ce300acc44..10b633a4fc106 100644 --- a/benchmark/fluid/fluid_benchmark.py +++ b/benchmark/fluid/fluid_benchmark.py @@ -81,9 +81,11 @@ def dist_transpile(trainer_id, args, train_prog, startup_prog): # the role, should be either PSERVER or TRAINER training_role = os.getenv("PADDLE_TRAINING_ROLE") - config = distribute_transpiler.DistributeTranspilerConfig() + config = fluid.DistributeTranspilerConfig() config.slice_var_up = not args.no_split_var + config.min_block_size = 1048576 t = distribute_transpiler.DistributeTranspiler(config=config) + t.transpile( trainer_id, # NOTE: *MUST* use train_prog, for we are using with guard to diff --git a/cmake/FindGperftools.cmake b/cmake/FindGperftools.cmake new file mode 100644 index 0000000000000..928f573a4fb82 --- /dev/null +++ b/cmake/FindGperftools.cmake @@ -0,0 +1,63 @@ +# Tries to find Gperftools. +# +# Usage of this module as follows: +# +# find_package(Gperftools) +# +# Variables used by this module, they can change the default behaviour and need +# to be set before calling find_package: +# +# Gperftools_ROOT_DIR Set this variable to the root installation of +# Gperftools if the module has problems finding +# the proper installation path. +# +# Variables defined by this module: +# +# GPERFTOOLS_FOUND System has Gperftools libs/headers +# GPERFTOOLS_LIBRARIES The Gperftools libraries (tcmalloc & profiler) +# GPERFTOOLS_INCLUDE_DIR The location of Gperftools headers + +find_library(GPERFTOOLS_TCMALLOC + NAMES tcmalloc + HINTS ${Gperftools_ROOT_DIR}/lib) + +find_library(GPERFTOOLS_PROFILER + NAMES profiler + HINTS ${Gperftools_ROOT_DIR}/lib) + +find_library(GPERFTOOLS_TCMALLOC_AND_PROFILER + NAMES tcmalloc_and_profiler + HINTS ${Gperftools_ROOT_DIR}/lib) + +find_path(GPERFTOOLS_INCLUDE_DIR + NAMES gperftools/heap-profiler.h + HINTS ${Gperftools_ROOT_DIR}/include) + +set(GPERFTOOLS_LIBRARIES ${GPERFTOOLS_TCMALLOC_AND_PROFILER}) + +include(FindPackageHandleStandardArgs) +find_package_handle_standard_args( + Gperftools + DEFAULT_MSG + GPERFTOOLS_LIBRARIES + GPERFTOOLS_INCLUDE_DIR) + +mark_as_advanced( + Gperftools_ROOT_DIR + GPERFTOOLS_TCMALLOC + GPERFTOOLS_PROFILER + GPERFTOOLS_TCMALLOC_AND_PROFILER + GPERFTOOLS_LIBRARIES + GPERFTOOLS_INCLUDE_DIR) + +# create IMPORTED targets +if (Gperftools_FOUND AND NOT TARGET gperftools::tcmalloc) + add_library(gperftools::tcmalloc UNKNOWN IMPORTED) + set_target_properties(gperftools::tcmalloc PROPERTIES + IMPORTED_LOCATION ${GPERFTOOLS_TCMALLOC} + INTERFACE_INCLUDE_DIRECTORIES "${GPERFTOOLS_INCLUDE_DIR}") + add_library(gperftools::profiler UNKNOWN IMPORTED) + set_target_properties(gperftools::profiler PROPERTIES + IMPORTED_LOCATION ${GPERFTOOLS_PROFILER} + INTERFACE_INCLUDE_DIRECTORIES "${GPERFTOOLS_INCLUDE_DIR}") +endif() diff --git a/cmake/configure.cmake b/cmake/configure.cmake index 4e17ddee73958..4ee2fdcf2db6b 100644 --- a/cmake/configure.cmake +++ b/cmake/configure.cmake @@ -84,8 +84,13 @@ if(NOT WITH_GOLANG) add_definitions(-DPADDLE_WITHOUT_GOLANG) endif(NOT WITH_GOLANG) +if(WITH_PSLIB) + add_definitions(-DPADDLE_WITH_PSLIB) +endif() + if(WITH_GPU) add_definitions(-DPADDLE_WITH_CUDA) + add_definitions(-DEIGEN_USE_GPU) FIND_PACKAGE(CUDA REQUIRED) diff --git a/cmake/cuda.cmake b/cmake/cuda.cmake index 964d5fd45b350..414e92eb27f56 100644 --- a/cmake/cuda.cmake +++ b/cmake/cuda.cmake @@ -199,8 +199,11 @@ elseif(CMAKE_BUILD_TYPE STREQUAL "MinSizeRel") list(APPEND CUDA_NVCC_FLAGS ${CMAKE_CXX_FLAGS_RELEASE}) endif() else(NOT WIN32) +list(APPEND CUDA_NVCC_FLAGS "--compiler-options;/bigobj") if(CMAKE_BUILD_TYPE STREQUAL "Debug") - list(APPEND CUDA_NVCC_FLAGS "-g -G") + list(APPEND CUDA_NVCC_FLAGS "-g -G") + # match the cl's _ITERATOR_DEBUG_LEVEL + list(APPEND CUDA_NVCC_FLAGS "-D_DEBUG") elseif(CMAKE_BUILD_TYPE STREQUAL "Release") list(APPEND CUDA_NVCC_FLAGS "-O3 -DNDEBUG") else() diff --git a/cmake/cudnn.cmake b/cmake/cudnn.cmake index 09bec347dbd56..fb899e3d7cd42 100644 --- a/cmake/cudnn.cmake +++ b/cmake/cudnn.cmake @@ -44,9 +44,9 @@ if(WIN32) set(CUDNN_LIB_NAME "cudnn.lib" "cudnn64_7.dll") endif(WIN32) -if(Apple) +if(APPLE) set(CUDNN_LIB_NAME "libcudnn.dylib" "libcudnn.so") -endif(Apple) +endif(APPLE) find_library(CUDNN_LIBRARY NAMES ${CUDNN_LIB_NAME} # libcudnn_static.a PATHS ${CUDNN_CHECK_LIBRARY_DIRS} ${CUDNN_INCLUDE_DIR} ${__libpath_hist} diff --git a/cmake/external/brpc.cmake b/cmake/external/brpc.cmake index 30b227b6452ab..6b50cff7a66a3 100644 --- a/cmake/external/brpc.cmake +++ b/cmake/external/brpc.cmake @@ -14,14 +14,16 @@ INCLUDE(ExternalProject) -find_library(SSL_LIBRARY NAMES ssl) +find_package(OpenSSL REQUIRED) + +message(STATUS "ssl:" ${OPENSSL_SSL_LIBRARY}) +message(STATUS "crypto:" ${OPENSSL_CRYPTO_LIBRARY}) + ADD_LIBRARY(ssl SHARED IMPORTED GLOBAL) -SET_PROPERTY(TARGET ssl PROPERTY IMPORTED_LOCATION ${SSL_LIBRARY}) +SET_PROPERTY(TARGET ssl PROPERTY IMPORTED_LOCATION ${OPENSSL_SSL_LIBRARY}) -find_library(CRYPTO_LIBRARY NAMES crypto) ADD_LIBRARY(crypto SHARED IMPORTED GLOBAL) -SET_PROPERTY(TARGET crypto PROPERTY IMPORTED_LOCATION ${CRYPTO_LIBRARY}) - +SET_PROPERTY(TARGET crypto PROPERTY IMPORTED_LOCATION ${OPENSSL_CRYPTO_LIBRARY}) SET(BRPC_SOURCES_DIR ${THIRD_PARTY_PATH}/brpc) SET(BRPC_INSTALL_DIR ${THIRD_PARTY_PATH}/install/brpc) @@ -31,14 +33,15 @@ SET(BRPC_LIBRARIES "${BRPC_INSTALL_DIR}/lib/libbrpc.a" CACHE FILEPATH "brpc libr INCLUDE_DIRECTORIES(${BRPC_INCLUDE_DIR}) # Reference https://stackoverflow.com/questions/45414507/pass-a-list-of-prefix-paths-to-externalproject-add-in-cmake-args -set(prefix_path "${THIRD_PARTY_PATH}/install/gflags|${THIRD_PARTY_PATH}/install/leveldb|${THIRD_PARTY_PATH}/install/snappy|${THIRD_PARTY_PATH}/install/gtest|${THIRD_PARTY_PATH}/install/protobuf|${THIRD_PARTY_PATH}/install/zlib") +set(prefix_path "${THIRD_PARTY_PATH}/install/gflags|${THIRD_PARTY_PATH}/install/leveldb|${THIRD_PARTY_PATH}/install/snappy|${THIRD_PARTY_PATH}/install/gtest|${THIRD_PARTY_PATH}/install/protobuf|${THIRD_PARTY_PATH}/install/zlib|${THIRD_PARTY_PATH}/install/glog") # If minimal .a is need, you can set WITH_DEBUG_SYMBOLS=OFF ExternalProject_Add( extern_brpc ${EXTERNAL_PROJECT_LOG_ARGS} + # TODO(gongwb): change to de newst repo when they changed. GIT_REPOSITORY "https://github.com/gongweibao/brpc" - GIT_TAG "7dc04defad1fd4173aae170c3fcbde131b65155a" + GIT_TAG "e9b67ec1b7458f2af5fae76451afe1e27e01b4b4" PREFIX ${BRPC_SOURCES_DIR} UPDATE_COMMAND "" CMAKE_ARGS -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER} @@ -50,7 +53,7 @@ ExternalProject_Add( -DCMAKE_POSITION_INDEPENDENT_CODE=ON -DCMAKE_BUILD_TYPE=${THIRD_PARTY_BUILD_TYPE} -DCMAKE_PREFIX_PATH=${prefix_path} - -DBRPC_WITH_GLOG=ON + -DWITH_GLOG=ON -DIOBUF_WITH_HUGE_BLOCK=ON -DBRPC_WITH_RDMA=${WITH_BRPC_RDMA} ${EXTERNAL_OPTIONAL_ARGS} @@ -65,5 +68,6 @@ ADD_LIBRARY(brpc STATIC IMPORTED GLOBAL) SET_PROPERTY(TARGET brpc PROPERTY IMPORTED_LOCATION ${BRPC_LIBRARIES}) ADD_DEPENDENCIES(brpc extern_brpc) +add_definitions(-DBRPC_WITH_GLOG) LIST(APPEND external_project_dependencies brpc) diff --git a/cmake/external/dlpack.cmake b/cmake/external/dlpack.cmake new file mode 100644 index 0000000000000..94d8fcc668556 --- /dev/null +++ b/cmake/external/dlpack.cmake @@ -0,0 +1,31 @@ +include(ExternalProject) + +set(DLPACK_SOURCE_DIR ${THIRD_PARTY_PATH}/dlpack) +set(DLPACK_INCLUDE_DIR ${DLPACK_SOURCE_DIR}/src/extern_dlpack/include) + +include_directories(${DLPACK_INCLUDE_DIR}) + +ExternalProject_Add( + extern_dlpack + ${EXTERNAL_PROJECT_LOG_ARGS} + GIT_REPOSITORY "https://github.com/dmlc/dlpack.git" + GIT_TAG "v0.2" + PREFIX ${DLPACK_SOURCE_DIR} + UPDATE_COMMAND "" + CONFIGURE_COMMAND "" + BUILD_COMMAND "" + INSTALL_COMMAND "" + TEST_COMMAND "" +) + +if(${CMAKE_VERSION} VERSION_LESS "3.3.0") + set(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/dlpack_dummy.c) + file(WRITE ${dummyfile} "const char *dummy = \"${dummyfile}\";") + add_library(dlpack STATIC ${dummyfile}) +else() + add_library(dlpack INTERFACE) +endif() + +add_dependencies(dlpack extern_dlpack) + +LIST(APPEND externl_project_dependencies dlpack) diff --git a/cmake/external/eigen.cmake b/cmake/external/eigen.cmake index 573ad5e5f06a9..6aef97f21244e 100644 --- a/cmake/external/eigen.cmake +++ b/cmake/external/eigen.cmake @@ -17,7 +17,7 @@ if(WITH_AMD_GPU) extern_eigen3 ${EXTERNAL_PROJECT_LOG_ARGS} GIT_REPOSITORY "https://github.com/sabreshao/hipeigen.git" - GIT_TAG 0cba03ff9f8f9f70bbd92ac5857b031aa8fed6f9 + GIT_TAG 7cb2b6e5a4b4a1efe658abb215cd866c6fb2275e PREFIX ${EIGEN_SOURCE_DIR} UPDATE_COMMAND "" CONFIGURE_COMMAND "" diff --git a/cmake/external/gtest.cmake b/cmake/external/gtest.cmake index 4fe9c13fb7f2c..9be625b620287 100644 --- a/cmake/external/gtest.cmake +++ b/cmake/external/gtest.cmake @@ -12,8 +12,12 @@ # See the License for the specific language governing permissions and # limitations under the License. -IF(WITH_TESTING) - ENABLE_TESTING() +#FIXME:(gongwb) Move brpc's gtest dependency. +IF(WITH_TESTING OR (WITH_DISTRIBUTE AND NOT WITH_GRPC)) + IF(WITH_TESTING) + ENABLE_TESTING() + ENDIF(WITH_TESTING) + INCLUDE(ExternalProject) SET(GTEST_SOURCES_DIR ${THIRD_PARTY_PATH}/gtest) @@ -76,4 +80,4 @@ IF(WITH_TESTING) ADD_DEPENDENCIES(gtest_main extern_gtest) LIST(APPEND external_project_dependencies gtest gtest_main) -ENDIF(WITH_TESTING) +ENDIF(WITH_TESTING OR (WITH_DISTRIBUTE AND NOT WITH_GRPC)) diff --git a/cmake/external/gzstream.cmake b/cmake/external/gzstream.cmake new file mode 100644 index 0000000000000..3e36ef7ae205b --- /dev/null +++ b/cmake/external/gzstream.cmake @@ -0,0 +1,48 @@ +# 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. +# + +IF(MOBILE_INFERENCE) + return() +ENDIF() + +include (ExternalProject) + +# NOTE: gzstream is needed when linking with ctr reader. + +SET(GZSTREAM_SOURCES_DIR ${THIRD_PARTY_PATH}/gzstream) +SET(GZSTREAM_INSTALL_DIR ${THIRD_PARTY_PATH}/install/gzstream) +SET(GZSTREAM_INCLUDE_DIR "${GZSTREAM_INSTALL_DIR}/include/" CACHE PATH "gzstream include directory." FORCE) + +ExternalProject_Add( + extern_gzstream + DEPENDS zlib + GIT_REPOSITORY "https://github.com/jacquesqiao/gzstream.git" + GIT_TAG "" + PREFIX ${GZSTREAM_SOURCES_DIR} + UPDATE_COMMAND "" + CONFIGURE_COMMAND "" + BUILD_IN_SOURCE 1 + BUILD_COMMAND make EXTERN_CPPFLAGS="-I${THIRD_PARTY_PATH}/install/zlib/include" EXTERM_LDFLAGS="-L${THIRD_PARTY_PATH}/install/zlib/lib" -j8 + INSTALL_COMMAND mkdir -p ${GZSTREAM_INSTALL_DIR}/lib/ && mkdir -p ${GZSTREAM_INSTALL_DIR}/include/ + && cp ${GZSTREAM_SOURCES_DIR}/src/extern_gzstream/libgzstream.a ${GZSTREAM_INSTALL_DIR}/lib + && cp -r ${GZSTREAM_SOURCES_DIR}/src/extern_gzstream/gzstream.h ${GZSTREAM_INSTALL_DIR}/include +) + +ADD_LIBRARY(gzstream STATIC IMPORTED GLOBAL) +SET_PROPERTY(TARGET gzstream PROPERTY IMPORTED_LOCATION + "${GZSTREAM_INSTALL_DIR}/lib/libgzstream.a") + +include_directories(${GZSTREAM_INCLUDE_DIR}) +ADD_DEPENDENCIES(gzstream extern_gzstream zlib) diff --git a/cmake/external/leveldb.cmake b/cmake/external/leveldb.cmake index fb5091731da02..0df61b01ab64c 100644 --- a/cmake/external/leveldb.cmake +++ b/cmake/external/leveldb.cmake @@ -24,8 +24,8 @@ ExternalProject_Add( extern_leveldb ${EXTERNAL_PROJECT_LOG_ARGS} PREFIX ${LEVELDB_SOURCES_DIR} - URL "https://github.com/google/leveldb/archive/v1.18.tar.gz" - URL_MD5 "73770de34a2a5ab34498d2e05b2b7fa0" + GIT_REPOSITORY "https://github.com/google/leveldb" + GIT_TAG v1.18 CONFIGURE_COMMAND "" BUILD_COMMAND CXXFLAGS=-fPIC make -j ${NUM_OF_PROCESSOR} libleveldb.a INSTALL_COMMAND mkdir -p ${LEVELDB_INSTALL_DIR}/lib/ diff --git a/cmake/external/libmct.cmake b/cmake/external/libmct.cmake new file mode 100644 index 0000000000000..27cff8cfb6315 --- /dev/null +++ b/cmake/external/libmct.cmake @@ -0,0 +1,78 @@ +# Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +IF(NOT ${WITH_LIBMCT}) + return() +ENDIF(NOT ${WITH_LIBMCT}) + +IF(WIN32 OR APPLE) + MESSAGE(WARNING + "Windows or Mac is not supported with LIBMCT in Paddle yet." + "Force WITH_LIBMCT=OFF") + SET(WITH_LIBMCT OFF CACHE STRING "Disable LIBMCT package in Windows and MacOS" FORCE) + return() +ENDIF() + +INCLUDE(ExternalProject) + +SET(LIBMCT_PROJECT "extern_libmct") +IF((NOT DEFINED LIBMCT_VER) OR (NOT DEFINED LIBMCT_URL)) + MESSAGE(STATUS "use pre defined download url") + SET(LIBMCT_VER "0.1.0" CACHE STRING "" FORCE) + SET(LIBMCT_NAME "libmct" CACHE STRING "" FORCE) + SET(LIBMCT_URL "https://raw.githubusercontent.com/PaddlePaddle/Fleet/release/${LIBMCT_VER}/${LIBMCT_NAME}.tar.gz" CACHE STRING "" FORCE) +ENDIF() +MESSAGE(STATUS "LIBMCT_NAME: ${LIBMCT_NAME}, LIBMCT_URL: ${LIBMCT_URL}") +SET(LIBMCT_SOURCE_DIR "${THIRD_PARTY_PATH}/libmct") +SET(LIBMCT_DOWNLOAD_DIR "${LIBMCT_SOURCE_DIR}/src/${LIBMCT_PROJECT}") +SET(LIBMCT_DST_DIR "libmct") +SET(LIBMCT_INSTALL_ROOT "${THIRD_PARTY_PATH}/install") +SET(LIBMCT_INSTALL_DIR ${LIBMCT_INSTALL_ROOT}/${LIBMCT_DST_DIR}) +SET(LIBMCT_ROOT ${LIBMCT_INSTALL_DIR}) +SET(LIBMCT_INC_DIR ${LIBMCT_ROOT}/include) +SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${LIBMCT_ROOT}/lib") + +INCLUDE_DIRECTORIES(${LIBMCT_INC_DIR}) + +FILE(WRITE ${LIBMCT_DOWNLOAD_DIR}/CMakeLists.txt + "PROJECT(LIBMCT)\n" + "cmake_minimum_required(VERSION 3.0)\n" + "install(DIRECTORY ${LIBMCT_NAME}/include ${LIBMCT_NAME}/lib \n" + " DESTINATION ${LIBMCT_DST_DIR})\n") + +ExternalProject_Add( + ${LIBMCT_PROJECT} + ${EXTERNAL_PROJECT_LOG_ARGS} + PREFIX ${LIBMCT_SOURCE_DIR} + DOWNLOAD_DIR ${LIBMCT_DOWNLOAD_DIR} + DOWNLOAD_COMMAND wget --no-check-certificate ${LIBMCT_URL} -c -q -O ${LIBMCT_NAME}.tar.gz + && tar zxvf ${LIBMCT_NAME}.tar.gz + DOWNLOAD_NO_PROGRESS 1 + UPDATE_COMMAND "" + CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${LIBMCT_INSTALL_ROOT} + CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${LIBMCT_INSTALL_ROOT} +) + +if (${CMAKE_VERSION} VERSION_LESS "3.3.0" OR NOT WIN32) + set(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/boost_dummy.c) + file(WRITE ${dummyfile} "const char *dummy = \"${dummyfile}\";") + add_library(libmct STATIC ${dummyfile}) +else() + add_library(libmct INTERFACE) +endif() + +#ADD_LIBRARY(libmct SHARED IMPORTED GLOBAL) +ADD_DEPENDENCIES(libmct ${LIBMCT_PROJECT}) +LIST(APPEND external_project_dependencies libmct) + diff --git a/cmake/external/mkldnn.cmake b/cmake/external/mkldnn.cmake index 785148d4f9f44..c29375cd05897 100644 --- a/cmake/external/mkldnn.cmake +++ b/cmake/external/mkldnn.cmake @@ -23,15 +23,14 @@ SET(MKLDNN_SOURCES_DIR ${THIRD_PARTY_PATH}/mkldnn) SET(MKLDNN_INSTALL_DIR ${THIRD_PARTY_PATH}/install/mkldnn) SET(MKLDNN_INC_DIR "${MKLDNN_INSTALL_DIR}/include" CACHE PATH "mkldnn include directory." FORCE) -IF(WIN32 OR APPLE) +IF(APPLE) MESSAGE(WARNING - "Windows or Mac is not supported with MKLDNN in Paddle yet." + "Mac is not supported with MKLDNN in Paddle yet." "Force WITH_MKLDNN=OFF") - SET(WITH_MKLDNN OFF CACHE STRING "Disable MKLDNN in Windows and MacOS" FORCE) + SET(WITH_MKLDNN OFF CACHE STRING "Disable MKLDNN in MacOS" FORCE) return() ENDIF() -SET(MKLDNN_LIB "${MKLDNN_INSTALL_DIR}/lib/libmkldnn.so" CACHE FILEPATH "mkldnn library." FORCE) MESSAGE(STATUS "Set ${MKLDNN_INSTALL_DIR}/lib to runtime path") SET(CMAKE_INSTALL_RPATH_USE_LINK_PATH TRUE) SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${MKLDNN_INSTALL_DIR}/lib") @@ -44,22 +43,33 @@ IF(${CBLAS_PROVIDER} STREQUAL "MKLML") ELSE() MESSAGE(FATAL_ERROR "Should enable MKLML when build MKLDNN") ENDIF() -SET(MKLDNN_FLAG "-Wno-error=strict-overflow -Wno-error=unused-result -Wno-error=array-bounds") -SET(MKLDNN_FLAG "${MKLDNN_FLAG} -Wno-unused-result -Wno-unused-value") -SET(MKLDNN_CFLAG "${CMAKE_C_FLAGS} ${MKLDNN_FLAG}") -SET(MKLDNN_CXXFLAG "${CMAKE_CXX_FLAGS} ${MKLDNN_FLAG}") + +IF(NOT WIN32) + SET(MKLDNN_FLAG "-Wno-error=strict-overflow -Wno-error=unused-result -Wno-error=array-bounds") + SET(MKLDNN_FLAG "${MKLDNN_FLAG} -Wno-unused-result -Wno-unused-value") + SET(MKLDNN_CFLAG "${CMAKE_C_FLAGS} ${MKLDNN_FLAG}") + SET(MKLDNN_CXXFLAG "${CMAKE_CXX_FLAGS} ${MKLDNN_FLAG}") +ENDIF(NOT WIN32) + ExternalProject_Add( ${MKLDNN_PROJECT} ${EXTERNAL_PROJECT_LOG_ARGS} DEPENDS ${MKLDNN_DEPENDS} GIT_REPOSITORY "https://github.com/01org/mkl-dnn.git" - GIT_TAG "21fb5f2af1dd14e132af4f1b79160977ee487818" + GIT_TAG "830a10059a018cd2634d94195140cf2d8790a75a" PREFIX ${MKLDNN_SOURCES_DIR} UPDATE_COMMAND "" CMAKE_ARGS -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER} CMAKE_ARGS -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} + CMAKE_ARGS -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS} + CMAKE_ARGS -DCMAKE_CXX_FLAGS_RELEASE=${CMAKE_CXX_FLAGS_RELEASE} + CMAKE_ARGS -DCMAKE_CXX_FLAGS_DEBUG=${CMAKE_CXX_FLAGS_DEBUG} + CMAKE_ARGS -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS} + CMAKE_ARGS -DCMAKE_C_FLAGS_DEBUG=${CMAKE_C_FLAGS_DEBUG} + CMAKE_ARGS -DCMAKE_C_FLAGS_RELEASE=${CMAKE_C_FLAGS_RELEASE} CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${MKLDNN_INSTALL_DIR} CMAKE_ARGS -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE} + CMAKE_ARGS -DCMAKE_POSITION_INDEPENDENT_CODE=ON CMAKE_ARGS -DMKLROOT=${MKLML_ROOT} CMAKE_ARGS -DCMAKE_C_FLAGS=${MKLDNN_CFLAG} CMAKE_ARGS -DCMAKE_CXX_FLAGS=${MKLDNN_CXXFLAG} @@ -67,6 +77,11 @@ ExternalProject_Add( CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${MKLDNN_INSTALL_DIR} -DMKLROOT:PATH=${MKLML_ROOT} ) +if(WIN32) + SET(MKLDNN_LIB "${MKLDNN_INSTALL_DIR}/lib/mkldnn.lib" CACHE FILEPATH "mkldnn library." FORCE) +else(WIN32) + SET(MKLDNN_LIB "${MKLDNN_INSTALL_DIR}/lib/libmkldnn.so" CACHE FILEPATH "mkldnn library." FORCE) +endif(WIN32) ADD_LIBRARY(shared_mkldnn SHARED IMPORTED GLOBAL) SET_PROPERTY(TARGET shared_mkldnn PROPERTY IMPORTED_LOCATION ${MKLDNN_LIB}) @@ -85,10 +100,14 @@ ADD_DEPENDENCIES(mkldnn ${MKLDNN_PROJECT}) # copy the real so.0 lib to install dir # it can be directly contained in wheel or capi -SET(MKLDNN_SHARED_LIB ${MKLDNN_INSTALL_DIR}/libmkldnn.so.0) -ADD_CUSTOM_COMMAND(OUTPUT ${MKLDNN_SHARED_LIB} - COMMAND cp ${MKLDNN_LIB} ${MKLDNN_SHARED_LIB} - DEPENDS mkldnn) +if(WIN32) + SET(MKLDNN_SHARED_LIB ${MKLDNN_INSTALL_DIR}/lib/mkldnn.dll) +else(WIN32) + SET(MKLDNN_SHARED_LIB ${MKLDNN_INSTALL_DIR}/libmkldnn.so.0) + ADD_CUSTOM_COMMAND(OUTPUT ${MKLDNN_SHARED_LIB} + COMMAND ${CMAKE_COMMAND} -E copy ${MKLDNN_LIB} ${MKLDNN_SHARED_LIB} + DEPENDS mkldnn) +endif(WIN32) ADD_CUSTOM_TARGET(mkldnn_shared_lib ALL DEPENDS ${MKLDNN_SHARED_LIB}) IF(WITH_C_API) diff --git a/cmake/external/mklml.cmake b/cmake/external/mklml.cmake index dc5427acd45f5..d49839a89d788 100644 --- a/cmake/external/mklml.cmake +++ b/cmake/external/mklml.cmake @@ -16,56 +16,67 @@ IF(NOT ${WITH_MKLML}) return() ENDIF(NOT ${WITH_MKLML}) -IF(WIN32 OR APPLE) +IF(APPLE) MESSAGE(WARNING - "Windows or Mac is not supported with MKLML in Paddle yet." + "Mac is not supported with MKLML in Paddle yet." "Force WITH_MKLML=OFF") SET(WITH_MKLML OFF CACHE STRING "Disable MKLML package in Windows and MacOS" FORCE) return() ENDIF() INCLUDE(ExternalProject) - -SET(MKLML_PROJECT "extern_mklml") -IF((NOT DEFINED MKLML_VER) OR (NOT DEFINED MKLML_URL)) - MESSAGE(STATUS "use pre defined download url") - SET(MKLML_VER "mklml_lnx_2019.0.20180710" CACHE STRING "" FORCE) - SET(MKLML_URL "http://paddlepaddledeps.cdn.bcebos.com/${MKLML_VER}.tgz" CACHE STRING "" FORCE) -ENDIF() -MESSAGE(STATUS "MKLML_VER: ${MKLML_VER}, MKLML_URL: ${MKLML_URL}") -SET(MKLML_SOURCE_DIR "${THIRD_PARTY_PATH}/mklml") -SET(MKLML_DOWNLOAD_DIR "${MKLML_SOURCE_DIR}/src/${MKLML_PROJECT}") SET(MKLML_DST_DIR "mklml") SET(MKLML_INSTALL_ROOT "${THIRD_PARTY_PATH}/install") SET(MKLML_INSTALL_DIR ${MKLML_INSTALL_ROOT}/${MKLML_DST_DIR}) SET(MKLML_ROOT ${MKLML_INSTALL_DIR}) SET(MKLML_INC_DIR ${MKLML_ROOT}/include) SET(MKLML_LIB_DIR ${MKLML_ROOT}/lib) -SET(MKLML_LIB ${MKLML_LIB_DIR}/libmklml_intel.so) -SET(MKLML_IOMP_LIB ${MKLML_LIB_DIR}/libiomp5.so) +if(WIN32) + SET(MKLML_LIB ${MKLML_LIB_DIR}/mklml.lib) + SET(MKLML_IOMP_LIB ${MKLML_LIB_DIR}/libiomp5md.lib) + SET(MKLML_SHARED_LIB ${MKLML_LIB_DIR}/mklml.dll) + SET(MKLML_SHARED_IOMP_LIB ${MKLML_LIB_DIR}/libiomp5md.dll) +else() + SET(MKLML_LIB ${MKLML_LIB_DIR}/libmklml_intel.so) + SET(MKLML_IOMP_LIB ${MKLML_LIB_DIR}/libiomp5.so) + SET(MKLML_SHARED_LIB ${MKLML_LIB_DIR}/libmklml_intel.so) + SET(MKLML_SHARED_IOMP_LIB ${MKLML_LIB_DIR}/libiomp5.so) +endif() SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${MKLML_ROOT}/lib") -INCLUDE_DIRECTORIES(${MKLML_INC_DIR}) +IF((NOT DEFINED MKLML_VER) OR (NOT DEFINED MKLML_URL)) + MESSAGE(STATUS "use pre defined download url") + if(WIN32) + SET(MKLML_VER "mklml_win_2019.0.20180710" CACHE STRING "" FORCE) + SET(MKLML_URL "https://paddlepaddledeps.cdn.bcebos.com/${MKLML_VER}.zip" CACHE STRING "" FORCE) + else() + SET(MKLML_VER "mklml_lnx_2019.0.20180710" CACHE STRING "" FORCE) + SET(MKLML_URL "http://paddlepaddledeps.cdn.bcebos.com/${MKLML_VER}.tgz" CACHE STRING "" FORCE) + ENDIF() +endif() -FILE(WRITE ${MKLML_DOWNLOAD_DIR}/CMakeLists.txt - "PROJECT(MKLML)\n" - "cmake_minimum_required(VERSION 3.0)\n" - "install(DIRECTORY ${MKLML_VER}/include ${MKLML_VER}/lib \n" - " DESTINATION ${MKLML_DST_DIR})\n") +SET(MKLML_PROJECT "extern_mklml") +MESSAGE(STATUS "MKLML_VER: ${MKLML_VER}, MKLML_URL: ${MKLML_URL}") +SET(MKLML_SOURCE_DIR "${THIRD_PARTY_PATH}/mklml") +SET(MKLML_DOWNLOAD_DIR "${MKLML_SOURCE_DIR}/src/${MKLML_PROJECT}") ExternalProject_Add( ${MKLML_PROJECT} ${EXTERNAL_PROJECT_LOG_ARGS} - PREFIX ${MKLML_SOURCE_DIR} + PREFIX ${MKLML_SOURCE_DIR} + URL ${MKLML_URL} DOWNLOAD_DIR ${MKLML_DOWNLOAD_DIR} - DOWNLOAD_COMMAND wget --no-check-certificate ${MKLML_URL} -c -q -O ${MKLML_VER}.tgz - && tar zxf ${MKLML_VER}.tgz DOWNLOAD_NO_PROGRESS 1 - UPDATE_COMMAND "" - CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${MKLML_INSTALL_ROOT} - CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${MKLML_INSTALL_ROOT} + CONFIGURE_COMMAND "" + BUILD_COMMAND "" + UPDATE_COMMAND "" + INSTALL_COMMAND + ${CMAKE_COMMAND} -E copy_directory ${MKLML_DOWNLOAD_DIR}/include ${MKLML_INC_DIR} && + ${CMAKE_COMMAND} -E copy_directory ${MKLML_DOWNLOAD_DIR}/lib ${MKLML_LIB_DIR} ) +INCLUDE_DIRECTORIES(${MKLML_INC_DIR}) + ADD_LIBRARY(mklml SHARED IMPORTED GLOBAL) SET_PROPERTY(TARGET mklml PROPERTY IMPORTED_LOCATION ${MKLML_LIB}) ADD_DEPENDENCIES(mklml ${MKLML_PROJECT}) diff --git a/cmake/external/ngraph.cmake b/cmake/external/ngraph.cmake index 2e335579f32df..e66459fa3a150 100644 --- a/cmake/external/ngraph.cmake +++ b/cmake/external/ngraph.cmake @@ -32,6 +32,8 @@ IF(NOT ${WITH_NGRAPH}) return() ENDIF() +INCLUDE(GNUInstallDirs) + INCLUDE(ExternalProject) SET(NGRAPH_PROJECT "extern_ngraph") @@ -40,10 +42,14 @@ SET(NGRAPH_GIT_TAG "f9fd9d4cc318dc59dd4b68448e7fbb5f67a28bd0") SET(NGRAPH_SOURCES_DIR ${THIRD_PARTY_PATH}/ngraph) SET(NGRAPH_INSTALL_DIR ${THIRD_PARTY_PATH}/install/ngraph) SET(NGRAPH_INC_DIR ${NGRAPH_INSTALL_DIR}/include) +SET(NGRAPH_LIB_DIR ${NGRAPH_INSTALL_DIR}/${CMAKE_INSTALL_LIBDIR}) SET(NGRAPH_SHARED_LIB_NAME libngraph.so.${NGRAPH_VERSION}) SET(NGRAPH_CPU_LIB_NAME libcpu_backend.so) SET(NGRAPH_TBB_LIB_NAME libtbb.so.2) SET(NGRAPH_GIT_REPO "https://github.com/NervanaSystems/ngraph.git") +SET(NGRAPH_SHARED_LIB ${NGRAPH_LIB_DIR}/${NGRAPH_SHARED_LIB_NAME}) +SET(NGRAPH_CPU_LIB ${NGRAPH_LIB_DIR}/${NGRAPH_CPU_LIB_NAME}) +SET(NGRAPH_TBB_LIB ${NGRAPH_LIB_DIR}/${NGRAPH_TBB_LIB_NAME}) ExternalProject_Add( ${NGRAPH_PROJECT} @@ -63,18 +69,6 @@ ExternalProject_Add( CMAKE_ARGS -DMKLDNN_LIB_DIR=${MKLDNN_INSTALL_DIR}/lib ) -if(UNIX AND NOT APPLE) - include(GNUInstallDirs) - SET(NGRAPH_LIB_DIR ${NGRAPH_INSTALL_DIR}/${CMAKE_INSTALL_LIBDIR}) -else() - SET(NGRAPH_LIB_DIR ${NGRAPH_INSTALL_DIR}/lib) -endif() -MESSAGE(STATUS "nGraph lib will be installed at: ${NGRAPH_LIB_DIR}") - -SET(NGRAPH_SHARED_LIB ${NGRAPH_LIB_DIR}/${NGRAPH_SHARED_LIB_NAME}) -SET(NGRAPH_CPU_LIB ${NGRAPH_LIB_DIR}/${NGRAPH_CPU_LIB_NAME}) -SET(NGRAPH_TBB_LIB ${NGRAPH_LIB_DIR}/${NGRAPH_TBB_LIB_NAME}) - # Workaround for nGraph expecting mklml to be in mkldnn install directory. ExternalProject_Add_Step( ${NGRAPH_PROJECT} diff --git a/cmake/external/pslib.cmake b/cmake/external/pslib.cmake new file mode 100644 index 0000000000000..3b495d78e2c61 --- /dev/null +++ b/cmake/external/pslib.cmake @@ -0,0 +1,77 @@ +# Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +IF(NOT ${WITH_PSLIB}) + return() +ENDIF(NOT ${WITH_PSLIB}) + +IF(WIN32 OR APPLE) + MESSAGE(WARNING + "Windows or Mac is not supported with PSLIB in Paddle yet." + "Force WITH_PSLIB=OFF") + SET(WITH_PSLIB OFF CACHE STRING "Disable PSLIB package in Windows and MacOS" FORCE) + return() +ENDIF() + +INCLUDE(ExternalProject) + +SET(PSLIB_PROJECT "extern_pslib") +IF((NOT DEFINED PSLIB_VER) OR (NOT DEFINED PSLIB_URL)) + MESSAGE(STATUS "use pre defined download url") + SET(PSLIB_VER "0.1.0" CACHE STRING "" FORCE) + SET(PSLIB_NAME "pslib" CACHE STRING "" FORCE) + SET(PSLIB_URL "https://raw.githubusercontent.com/PaddlePaddle/Fleet/release/${PSLIB_VER}/${PSLIB_NAME}.tar.gz" CACHE STRING "" FORCE) +ENDIF() +MESSAGE(STATUS "PSLIB_NAME: ${PSLIB_NAME}, PSLIB_URL: ${PSLIB_URL}") +SET(PSLIB_SOURCE_DIR "${THIRD_PARTY_PATH}/pslib") +SET(PSLIB_DOWNLOAD_DIR "${PSLIB_SOURCE_DIR}/src/${PSLIB_PROJECT}") +SET(PSLIB_DST_DIR "pslib") +SET(PSLIB_INSTALL_ROOT "${THIRD_PARTY_PATH}/install") +SET(PSLIB_INSTALL_DIR ${PSLIB_INSTALL_ROOT}/${PSLIB_DST_DIR}) +SET(PSLIB_ROOT ${PSLIB_INSTALL_DIR}) +SET(PSLIB_INC_DIR ${PSLIB_ROOT}/include) +SET(PSLIB_LIB_DIR ${PSLIB_ROOT}/lib) +SET(PSLIB_LIB ${PSLIB_LIB_DIR}/libps.so) +SET(PSLIB_IOMP_LIB ${PSLIB_LIB_DIR}/libiomp5.so) #todo what is this +SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${PSLIB_ROOT}/lib") + +INCLUDE_DIRECTORIES(${PSLIB_INC_DIR}) + +FILE(WRITE ${PSLIB_DOWNLOAD_DIR}/CMakeLists.txt + "PROJECT(PSLIB)\n" + "cmake_minimum_required(VERSION 3.0)\n" + "install(DIRECTORY ${PSLIB_NAME}/include ${PSLIB_NAME}/lib \n" + " DESTINATION ${PSLIB_DST_DIR})\n") + +ExternalProject_Add( + ${PSLIB_PROJECT} + ${EXTERNAL_PROJECT_LOG_ARGS} + PREFIX ${PSLIB_SOURCE_DIR} + DOWNLOAD_DIR ${PSLIB_DOWNLOAD_DIR} + DOWNLOAD_COMMAND wget --no-check-certificate ${PSLIB_URL} -c -q -O ${PSLIB_NAME}.tar.gz + && tar zxvf ${PSLIB_NAME}.tar.gz + DOWNLOAD_NO_PROGRESS 1 + UPDATE_COMMAND "" + CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${PSLIB_INSTALL_ROOT} + CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${PSLIB_INSTALL_ROOT} +) + +ADD_LIBRARY(pslib SHARED IMPORTED GLOBAL) +SET_PROPERTY(TARGET pslib PROPERTY IMPORTED_LOCATION ${PSLIB_LIB}) +ADD_DEPENDENCIES(pslib ${PSLIB_PROJECT}) +LIST(APPEND external_project_dependencies pslib) + +IF(WITH_C_API) + INSTALL(FILES ${PSLIB_LIB} ${PSLIB_IOMP_LIB} DESTINATION lib) +ENDIF() diff --git a/cmake/external/pslib_brpc.cmake b/cmake/external/pslib_brpc.cmake new file mode 100644 index 0000000000000..7ff5a8aca1872 --- /dev/null +++ b/cmake/external/pslib_brpc.cmake @@ -0,0 +1,77 @@ +# Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +IF(NOT ${WITH_PSLIB_BRPC}) + return() +ENDIF(NOT ${WITH_PSLIB_BRPC}) + +IF(WIN32 OR APPLE) + MESSAGE(WARNING + "Windows or Mac is not supported with PSLIB_BRPC in Paddle yet." + "Force WITH_PSLIB_BRPC=OFF") + SET(WITH_PSLIB_BRPC OFF CACHE STRING "Disable PSLIB_BRPC package in Windows and MacOS" FORCE) + return() +ENDIF() + +INCLUDE(ExternalProject) + +SET(PSLIB_BRPC_PROJECT "extern_pslib_brpc") +IF((NOT DEFINED PSLIB_BRPC_NAME) OR (NOT DEFINED PSLIB_BRPC_URL)) + MESSAGE(STATUS "use pre defined download url") + SET(PSLIB_BRPC_VER "0.1.0" CACHE STRING "" FORCE) + SET(PSLIB_BRPC_NAME "pslib_brpc" CACHE STRING "" FORCE) + SET(PSLIB_BRPC_URL "https://raw.githubusercontent.com/PaddlePaddle/Fleet/release/${PSLIB_BRPC_VER}/${PSLIB_BRPC_NAME}.tar.gz" CACHE STRING "" FORCE) +ENDIF() +MESSAGE(STATUS "PSLIB_BRPC_NAME: ${PSLIB_BRPC_NAME}, PSLIB_BRPC_URL: ${PSLIB_BRPC_URL}") +SET(PSLIB_BRPC_SOURCE_DIR "${THIRD_PARTY_PATH}/pslib_brpc") +SET(PSLIB_BRPC_DOWNLOAD_DIR "${PSLIB_BRPC_SOURCE_DIR}/src/${PSLIB_BRPC_PROJECT}") +SET(PSLIB_BRPC_DST_DIR "pslib_brpc") +SET(PSLIB_BRPC_INSTALL_ROOT "${THIRD_PARTY_PATH}/install") +SET(PSLIB_BRPC_INSTALL_DIR ${PSLIB_BRPC_INSTALL_ROOT}/${PSLIB_BRPC_DST_DIR}) +SET(PSLIB_BRPC_ROOT ${PSLIB_BRPC_INSTALL_DIR}) +SET(PSLIB_BRPC_INC_DIR ${PSLIB_BRPC_ROOT}/include) +SET(PSLIB_BRPC_LIB_DIR ${PSLIB_BRPC_ROOT}/lib) +SET(PSLIB_BRPC_LIB ${PSLIB_BRPC_LIB_DIR}/libbrpc.a) +SET(PSLIB_BRPC_IOMP_LIB ${PSLIB_BRPC_LIB_DIR}/libiomp5.so) #todo what is this +SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${PSLIB_BRPC_ROOT}/lib") + +INCLUDE_DIRECTORIES(${PSLIB_BRPC_INC_DIR}) + +FILE(WRITE ${PSLIB_BRPC_DOWNLOAD_DIR}/CMakeLists.txt + "PROJECT(PSLIB_BRPC)\n" + "cmake_minimum_required(VERSION 3.0)\n" + "install(DIRECTORY ${PSLIB_BRPC_NAME}/include ${PSLIB_BRPC_NAME}/lib \n" + " DESTINATION ${PSLIB_BRPC_DST_DIR})\n") + +ExternalProject_Add( + ${PSLIB_BRPC_PROJECT} + ${EXTERNAL_PROJECT_LOG_ARGS} + PREFIX ${PSLIB_BRPC_SOURCE_DIR} + DOWNLOAD_DIR ${PSLIB_BRPC_DOWNLOAD_DIR} + DOWNLOAD_COMMAND wget --no-check-certificate ${PSLIB_BRPC_URL} -c -q -O ${PSLIB_BRPC_NAME}.tar.gz + && tar zxvf ${PSLIB_BRPC_NAME}.tar.gz + DOWNLOAD_NO_PROGRESS 1 + UPDATE_COMMAND "" + CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${PSLIB_BRPC_INSTALL_ROOT} + CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${PSLIB_BRPC_INSTALL_ROOT} +) + +ADD_LIBRARY(pslib_brpc SHARED IMPORTED GLOBAL) +SET_PROPERTY(TARGET pslib_brpc PROPERTY IMPORTED_LOCATION ${PSLIB_BRPC_LIB}) +ADD_DEPENDENCIES(pslib_brpc ${PSLIB_BRPC_PROJECT}) +LIST(APPEND external_project_dependencies pslib_brpc) + +IF(WITH_C_API) + INSTALL(FILES ${PSLIB_BRPC_LIB} ${PSLIB_BRPC_IOMP_LIB} DESTINATION lib) +ENDIF() diff --git a/cmake/external/pybind11.cmake b/cmake/external/pybind11.cmake index c885877a2bcd6..3a10ea945d3d1 100644 --- a/cmake/external/pybind11.cmake +++ b/cmake/external/pybind11.cmake @@ -26,7 +26,7 @@ ExternalProject_Add( extern_pybind ${EXTERNAL_PROJECT_LOG_ARGS} GIT_REPOSITORY "https://github.com/pybind/pybind11.git" - GIT_TAG "v2.1.1" + GIT_TAG "v2.2.4" PREFIX ${PYBIND_SOURCE_DIR} UPDATE_COMMAND "" CONFIGURE_COMMAND "" diff --git a/cmake/external/python.cmake b/cmake/external/python.cmake index a3599dd798c07..623c53f4f75bb 100644 --- a/cmake/external/python.cmake +++ b/cmake/external/python.cmake @@ -18,8 +18,8 @@ ENDIF() INCLUDE(python_module) -FIND_PACKAGE(PythonInterp ${PY_VERSION}) -FIND_PACKAGE(PythonLibs ${PY_VERSION}) +FIND_PACKAGE(PythonInterp ${PY_VERSION} REQUIRED) +FIND_PACKAGE(PythonLibs ${PY_VERSION} REQUIRED) if(WIN32) execute_process(COMMAND "${PYTHON_EXECUTABLE}" "-c" @@ -79,6 +79,5 @@ IF(PYTHONINTERP_FOUND) "please use pip to upgrade protobuf. pip install -U protobuf") ENDIF() ENDIF(PYTHONINTERP_FOUND) - INCLUDE_DIRECTORIES(${PYTHON_INCLUDE_DIR}) INCLUDE_DIRECTORIES(${PYTHON_NUMPY_INCLUDE_DIR}) diff --git a/cmake/external/rocprim.cmake b/cmake/external/rocprim.cmake new file mode 100644 index 0000000000000..914c064918905 --- /dev/null +++ b/cmake/external/rocprim.cmake @@ -0,0 +1,44 @@ +if (NOT WITH_AMD_GPU) + return() +endif() + +# rocprim is "ROCm Parallel Primitives" for short. +# It is a header-only library providing HIP and HC parallel primitives +# for developing performant GPU-accelerated code on AMD ROCm platform. + +if("x${HCC_HOME}" STREQUAL "x") + set(HCC_HOME "/opt/rocm/hcc") +endif() + +INCLUDE(ExternalProject) + +SET(ROCPRIM_SOURCE_DIR ${THIRD_PARTY_PATH}/rocprim) +SET(ROCPRIM_INSTALL_DIR ${THIRD_PARTY_PATH}/install/rocprim) +SET(ROCPRIM_INCLUDE_DIR ${ROCPRIM_INSTALL_DIR}/include) + +ExternalProject_Add( + extern_rocprim + GIT_REPOSITORY "https://github.com/ROCmSoftwarePlatform/rocPRIM.git" + GIT_TAG 5bd41b96ab8d8343330fb2c3e1b96775bde3b3fc + PREFIX ${ROCPRIM_SOURCE_DIR} + UPDATE_COMMAND "" + CMAKE_ARGS -DCMAKE_CXX_COMPILER=${HCC_HOME}/bin/hcc + CMAKE_ARGS -DONLY_INSTALL=ON + CMAKE_ARGS -DBUILD_TEST=OFF + CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${ROCPRIM_INSTALL_DIR} + + INSTALL_DIR ${ROCPRIM_INSTALL_DIR} + ${EXTERNAL_PROJECT_LOG_ARGS} +) + +INCLUDE_DIRECTORIES(${ROCPRIM_INCLUDE_DIR}) + +if (${CMAKE_VERSION} VERSION_LESS "3.3.0") + set(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/rocprim_dummy.c) + file(WRITE ${dummyfile} "const char *dummy_rocprim = \"${dummyfile}\";") + add_library(rocprim STATIC ${dummyfile}) +else() + add_library(rocprim INTERFACE) +endif() + +add_dependencies(rocprim extern_rocprim) diff --git a/cmake/external/snappy.cmake b/cmake/external/snappy.cmake index b30403d2d81ce..f9d4cd97400a6 100644 --- a/cmake/external/snappy.cmake +++ b/cmake/external/snappy.cmake @@ -24,12 +24,6 @@ set(SNAPPY_SOURCES_DIR ${THIRD_PARTY_PATH}/snappy) set(SNAPPY_INSTALL_DIR ${THIRD_PARTY_PATH}/install/snappy) set(SNAPPY_INCLUDE_DIR "${SNAPPY_INSTALL_DIR}/include" CACHE PATH "snappy include directory." FORCE) -if (WIN32) - set(SNAPPY_LIBRARIES "${SNAPPY_INSTALL_DIR}/lib/snappy.lib") -else(WIN32) - set(SNAPPY_LIBRARIES "${SNAPPY_INSTALL_DIR}/lib/libsnappy.a") -endif (WIN32) - ExternalProject_Add( extern_snappy GIT_REPOSITORY "https://github.com/google/snappy" @@ -56,6 +50,16 @@ ExternalProject_Add( -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON -DCMAKE_BUILD_TYPE:STRING=${THIRD_PARTY_BUILD_TYPE} ) +IF(WIN32) + IF(NOT EXISTS "${SNAPPY_INSTALL_DIR}/lib/libsnappy.lib") + add_custom_command(TARGET extern_snappy POST_BUILD + COMMAND cmake -E copy ${SNAPPY_INSTALL_DIR}/lib/snappy.lib ${SNAPPY_INSTALL_DIR}/lib/libsnappy.lib + ) + ENDIF() + set(SNAPPY_LIBRARIES "${SNAPPY_INSTALL_DIR}/lib/libsnappy.lib") +else(WIN32) + set(SNAPPY_LIBRARIES "${SNAPPY_INSTALL_DIR}/lib/libsnappy.a") +endif (WIN32) add_library(snappy STATIC IMPORTED GLOBAL) set_property(TARGET snappy PROPERTY IMPORTED_LOCATION ${SNAPPY_LIBRARIES}) diff --git a/cmake/external/warpctc.cmake b/cmake/external/warpctc.cmake index 07e1137e16afc..7b937c93febdf 100644 --- a/cmake/external/warpctc.cmake +++ b/cmake/external/warpctc.cmake @@ -26,25 +26,33 @@ SET(WARPCTC_INCLUDE_DIR "${WARPCTC_INSTALL_DIR}/include" # Used in unit test test_WarpCTCLayer SET(WARPCTC_LIB_DIR "${WARPCTC_INSTALL_DIR}/lib" CACHE PATH "Warp-ctc Library Directory" FORCE) -SET(WARPCTC_LIBRARIES "${WARPCTC_INSTALL_DIR}/lib/libwarpctc${CMAKE_SHARED_LIBRARY_SUFFIX}" - CACHE FILEPATH "Warp-ctc Library" FORCE) -IF(CMAKE_CXX_COMPILER_ID STREQUAL "Clang" OR CMAKE_CXX_COMPILER_ID STREQUAL "AppleClang" ) +IF(CMAKE_CXX_COMPILER_ID STREQUAL "Clang" OR CMAKE_CXX_COMPILER_ID STREQUAL "AppleClang" OR WIN32) SET(USE_OMP OFF) ELSE() SET(USE_OMP ON) ENDIF() +IF(WIN32) + SET(WARPCTC_REPOSITORY "https://github.com/wopeizl/warp-ctc.git") +ELSE() + SET(WARPCTC_REPOSITORY "https://github.com/dzhwinter/warp-ctc.git") +ENDIF() + ExternalProject_Add( extern_warpctc ${EXTERNAL_PROJECT_LOG_ARGS} - GIT_REPOSITORY "https://github.com/dzhwinter/warp-ctc.git" + GIT_REPOSITORY ${WARPCTC_REPOSITORY} PREFIX ${WARPCTC_SOURCES_DIR} UPDATE_COMMAND "" CMAKE_ARGS -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER} -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} - -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS} -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS} + -DCMAKE_C_FLAGS_DEBUG=${CMAKE_C_FLAGS_DEBUG} + -DCMAKE_C_FLAGS_RELEASE=${CMAKE_C_FLAGS_RELEASE} + -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS} + -DCMAKE_CXX_FLAGS_RELEASE=${CMAKE_CXX_FLAGS_RELEASE} + -DCMAKE_CXX_FLAGS_DEBUG=${CMAKE_CXX_FLAGS_DEBUG} -DCMAKE_INSTALL_PREFIX=${WARPCTC_INSTALL_DIR} -DWITH_GPU=${WITH_GPU} -DWITH_OMP=${USE_OMP} @@ -59,6 +67,18 @@ ExternalProject_Add( -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON -DCMAKE_INSTALL_PREFIX:PATH=${WARPCTC_INSTALL_DIR} ) +IF(WIN32) + IF(NOT EXISTS "${WARPCTC_INSTALL_DIR}/lib/warpctc${CMAKE_SHARED_LIBRARY_SUFFIX}") + add_custom_command(TARGET extern_warpctc POST_BUILD + COMMAND cmake -E copy ${WARPCTC_INSTALL_DIR}/bin/warpctc${CMAKE_SHARED_LIBRARY_SUFFIX} ${WARPCTC_INSTALL_DIR}/lib/warpctc${CMAKE_SHARED_LIBRARY_SUFFIX} + ) + ENDIF() + SET(WARPCTC_LIBRARIES "${WARPCTC_INSTALL_DIR}/lib/warpctc${CMAKE_SHARED_LIBRARY_SUFFIX}" + CACHE FILEPATH "Warp-ctc Library" FORCE) +else(WIN32) + SET(WARPCTC_LIBRARIES "${WARPCTC_INSTALL_DIR}/lib/libwarpctc${CMAKE_SHARED_LIBRARY_SUFFIX}" + CACHE FILEPATH "Warp-ctc Library" FORCE) +ENDIF(WIN32) MESSAGE(STATUS "warp-ctc library: ${WARPCTC_LIBRARIES}") INCLUDE_DIRECTORIES(${WARPCTC_INCLUDE_DIR}) # For warpctc code to include its headers. diff --git a/cmake/external/xxhash.cmake b/cmake/external/xxhash.cmake index 4c2d64f627401..c3e1212d8f835 100644 --- a/cmake/external/xxhash.cmake +++ b/cmake/external/xxhash.cmake @@ -56,7 +56,12 @@ else() endif() if (WIN32) - set(XXHASH_LIBRARIES "${XXHASH_INSTALL_DIR}/lib/xxhash.lib") + IF(NOT EXISTS "${XXHASH_INSTALL_DIR}/lib/libxxhash.lib") + add_custom_command(TARGET extern_xxhash POST_BUILD + COMMAND cmake -E copy ${XXHASH_INSTALL_DIR}/lib/xxhash.lib ${XXHASH_INSTALL_DIR}/lib/libxxhash.lib + ) + ENDIF() + set(XXHASH_LIBRARIES "${XXHASH_INSTALL_DIR}/lib/libxxhash.lib") else() set(XXHASH_LIBRARIES "${XXHASH_INSTALL_DIR}/lib/libxxhash.a") endif () diff --git a/cmake/external/zlib.cmake b/cmake/external/zlib.cmake index c3d73235453c8..d35073753725c 100644 --- a/cmake/external/zlib.cmake +++ b/cmake/external/zlib.cmake @@ -19,12 +19,6 @@ SET(ZLIB_INSTALL_DIR ${THIRD_PARTY_PATH}/install/zlib) SET(ZLIB_ROOT ${ZLIB_INSTALL_DIR} CACHE FILEPATH "zlib root directory." FORCE) SET(ZLIB_INCLUDE_DIR "${ZLIB_INSTALL_DIR}/include" CACHE PATH "zlib include directory." FORCE) -IF(WIN32) - SET(ZLIB_LIBRARIES "${ZLIB_INSTALL_DIR}/lib/zlibstatic.lib" CACHE FILEPATH "zlib library." FORCE) -ELSE(WIN32) - SET(ZLIB_LIBRARIES "${ZLIB_INSTALL_DIR}/lib/libz.a" CACHE FILEPATH "zlib library." FORCE) -ENDIF(WIN32) - INCLUDE_DIRECTORIES(${ZLIB_INCLUDE_DIR}) # For zlib code to include its own headers. INCLUDE_DIRECTORIES(${THIRD_PARTY_PATH}/install) # For Paddle code to include zlib.h. @@ -49,6 +43,16 @@ ExternalProject_Add( -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON -DCMAKE_BUILD_TYPE:STRING=${THIRD_PARTY_BUILD_TYPE} ) +IF(WIN32) + IF(NOT EXISTS "${ZLIB_INSTALL_DIR}/lib/libz.lib") + add_custom_command(TARGET extern_zlib POST_BUILD + COMMAND cmake -E copy ${ZLIB_INSTALL_DIR}/lib/zlibstatic.lib ${ZLIB_INSTALL_DIR}/lib/libz.lib + ) + ENDIF() + SET(ZLIB_LIBRARIES "${ZLIB_INSTALL_DIR}/lib/libz.lib" CACHE FILEPATH "zlib library." FORCE) +ELSE(WIN32) + SET(ZLIB_LIBRARIES "${ZLIB_INSTALL_DIR}/lib/libz.a" CACHE FILEPATH "zlib library." FORCE) +ENDIF(WIN32) ADD_LIBRARY(zlib STATIC IMPORTED GLOBAL) SET_PROPERTY(TARGET zlib PROPERTY IMPORTED_LOCATION ${ZLIB_LIBRARIES}) diff --git a/cmake/flags.cmake b/cmake/flags.cmake index 343e44ab4bc21..c4472040cef87 100644 --- a/cmake/flags.cmake +++ b/cmake/flags.cmake @@ -129,6 +129,9 @@ set(COMMON_FLAGS -Wno-error=parentheses-equality # Warnings in pybind11 -Wno-error=ignored-attributes # Warnings in Eigen, gcc 6.3 -Wno-error=terminate # Warning in PADDLE_ENFORCE + -Wno-error=int-in-bool-context # Warning in Eigen gcc 7.2 + -Wimplicit-fallthrough=0 # Warning in tinyformat.h + -Wno-error=maybe-uninitialized # Warning in boost gcc 7.2 ) set(GPU_COMMON_FLAGS diff --git a/cmake/generic.cmake b/cmake/generic.cmake index 111627a932afe..c6fe2e970d3e0 100644 --- a/cmake/generic.cmake +++ b/cmake/generic.cmake @@ -110,6 +110,14 @@ function(find_fluid_modules TARGET_NAME) endif() endfunction(find_fluid_modules) + +function(common_link TARGET_NAME) + if (WITH_PROFILER) + target_link_libraries(${TARGET_NAME} gperftools::profiler) + endif() +endfunction() + + # find all third_party modules is used for paddle static library # for reduce the dependency when building the inference libs. set_property(GLOBAL PROPERTY FLUID_THIRD_PARTY) @@ -259,7 +267,11 @@ function(cc_library TARGET_NAME) list(APPEND cc_library_DEPS dynload_mklml) endif() add_dependencies(${TARGET_NAME} mklml) - target_link_libraries(${TARGET_NAME} "-L${MKLML_LIB_DIR} -liomp5 -Wl,--as-needed") + if(WIN32) + target_link_libraries(${TARGET_NAME} ${MKLML_IOMP_LIB}) + else(WIN32) + target_link_libraries(${TARGET_NAME} "-L${MKLML_LIB_DIR} -liomp5 -Wl,--as-needed") + endif(WIN32) endif() # remove link to python, see notes at: # https://github.com/pybind/pybind11/blob/master/docs/compiling.rst#building-manually @@ -274,6 +286,7 @@ function(cc_library TARGET_NAME) endif() target_link_libraries(${TARGET_NAME} ${cc_library_DEPS}) add_dependencies(${TARGET_NAME} ${cc_library_DEPS}) + common_link(${TARGET_NAME}) endif() # cpplint code style @@ -340,6 +353,7 @@ function(cc_binary TARGET_NAME) if(cc_binary_DEPS) target_link_libraries(${TARGET_NAME} ${cc_binary_DEPS}) add_dependencies(${TARGET_NAME} ${cc_binary_DEPS}) + common_link(${TARGET_NAME}) endif() endfunction(cc_binary) @@ -349,12 +363,20 @@ function(cc_test TARGET_NAME) set(oneValueArgs "") set(multiValueArgs SRCS DEPS ARGS) cmake_parse_arguments(cc_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) + if(WIN32) + list(APPEND win32_deps shlwapi) + if("${cc_test_DEPS};" MATCHES "python;") + list(REMOVE_ITEM cc_test_DEPS python) + list(APPEND win32_deps ${PYTHON_LIBRARIES}) + endif() + endif(WIN32) add_executable(${TARGET_NAME} ${cc_test_SRCS}) target_link_libraries(${TARGET_NAME} ${cc_test_DEPS} paddle_gtest_main lod_tensor memory gtest gflags glog) if(WIN32) - target_link_libraries(${TARGET_NAME} shlwapi) + target_link_libraries(${TARGET_NAME} ${win32_deps}) endif(WIN32) add_dependencies(${TARGET_NAME} ${cc_test_DEPS} paddle_gtest_main lod_tensor memory gtest gflags glog) + common_link(${TARGET_NAME}) add_test(NAME ${TARGET_NAME} COMMAND ${TARGET_NAME} ${cc_test_ARGS} WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}) @@ -413,6 +435,7 @@ function(nv_binary TARGET_NAME) if(nv_binary_DEPS) target_link_libraries(${TARGET_NAME} ${nv_binary_DEPS}) add_dependencies(${TARGET_NAME} ${nv_binary_DEPS}) + common_link(${TARGET_NAME}) endif() endif() endfunction(nv_binary) @@ -426,6 +449,7 @@ function(nv_test TARGET_NAME) cuda_add_executable(${TARGET_NAME} ${nv_test_SRCS}) target_link_libraries(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main lod_tensor memory gtest gflags glog) add_dependencies(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main lod_tensor memory gtest gflags glog) + common_link(${TARGET_NAME}) add_test(${TARGET_NAME} ${TARGET_NAME}) if (nv_test_SERIAL) set_property(TEST ${TARGET_NAME} PROPERTY RUN_SERIAL 1) @@ -454,25 +478,29 @@ function(hip_library TARGET_NAME) else() add_library(${TARGET_NAME} STATIC ${_cmake_options} ${_generated_files} ${_sources}) set_target_properties(${TARGET_NAME} PROPERTIES LINKER_LANGUAGE CXX) - target_link_libraries(${TARGET_NAME} /opt/rocm/hip/lib/libhip_hcc.so /opt/rocm/hip/lib/libhip_device.a) - find_fluid_modules(${TARGET_NAME}) + target_link_libraries(${TARGET_NAME} /opt/rocm/hip/lib/libhip_hcc.so /opt/rocm/hip/lib/libhip_device.a /opt/rocm/rccl/lib/librccl.so /opt/rocm/hiprand/lib/libhiprand.so) + find_fluid_modules(${TARGET_NAME}) endif() - if (hip_library_DEPS) - add_dependencies(${TARGET_NAME} ${hip_library_DEPS}) - target_link_libraries(${TARGET_NAME} ${hip_library_DEPS}) + if("${hip_library_DEPS}" MATCHES "ARCHIVE_START") + # Support linking flags: --whole-archive (Linux) / -force_load (MacOS). + # WARNING: Please don't use ARCHIVE_START&ARCHIVE_END if TARGET_NAME will be linked by other libraries. + target_circle_link_libraries(${TARGET_NAME} ${hip_library_DEPS}) + list(REMOVE_ITEM hip_library_DEPS ARCHIVE_START ARCHIVE_END) + else() + target_link_libraries(${TARGET_NAME} ${hip_library_DEPS}) endif() # cpplint code style foreach(source_file ${hip_library_SRCS}) - string(REGEX REPLACE "\\.[^.]*$" "" source ${source_file}) - if(EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${source}.h) - list(APPEND hip_library_HEADERS ${CMAKE_CURRENT_SOURCE_DIR}/${source}.h) - endif() + string(REGEX REPLACE "\\.[^.]*$" "" source ${source_file}) + if(EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${source}.h) + list(APPEND hip_library_HEADERS ${CMAKE_CURRENT_SOURCE_DIR}/${source}.h) + endif() endforeach() else(hip_library_SRCS) if (hip_library_DEPS) - merge_static_libs(${TARGET_NAME} ${hip_library_DEPS}) + merge_static_libs(${TARGET_NAME} ${hip_library_DEPS}) else() - message(FATAL "Please specify source file or library in nv_library.") + message(FATAL "Please specify source file or library in nv_library.") endif() endif(hip_library_SRCS) endif() @@ -488,6 +516,7 @@ function(hip_binary TARGET_NAME) if(hip_binary_DEPS) target_link_libraries(${TARGET_NAME} ${hip_binary_DEPS}) add_dependencies(${TARGET_NAME} ${hip_binary_DEPS}) + common_link(${TARGET_NAME}) endif() endif() endfunction(hip_binary) @@ -507,6 +536,7 @@ function(hip_test TARGET_NAME) set_target_properties(${TARGET_NAME} PROPERTIES LINKER_LANGUAGE HIP) target_link_libraries(${TARGET_NAME} ${hip_test_DEPS} paddle_gtest_main memory gtest gflags) add_dependencies(${TARGET_NAME} ${hip_test_DEPS} paddle_gtest_main memory gtest gflags) + common_link(${TARGET_NAME}) add_test(${TARGET_NAME} ${TARGET_NAME}) endif() endfunction(hip_test) @@ -549,6 +579,7 @@ function(go_library TARGET_NAME) endif() if(go_library_DEPS) add_dependencies(${TARGET_NAME} ${go_library_DEPS}) + common_link(${TARGET_NAME}) endif(go_library_DEPS) # The "source file" of the library is `${dummyfile}` which never @@ -679,7 +710,7 @@ function(py_test TARGET_NAME) set(multiValueArgs SRCS DEPS ARGS ENVS) cmake_parse_arguments(py_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) add_test(NAME ${TARGET_NAME} - COMMAND env FLAGS_init_allocated_mem=true FLAGS_cudnn_deterministic=true + COMMAND ${CMAKE_COMMAND} -E env FLAGS_init_allocated_mem=true FLAGS_cudnn_deterministic=true FLAGS_cpu_deterministic=true PYTHONPATH=${PADDLE_BINARY_DIR}/python ${py_test_ENVS} ${PYTHON_EXECUTABLE} -u ${py_test_SRCS} ${py_test_ARGS} diff --git a/cmake/hip.cmake b/cmake/hip.cmake index bfe491bd6b760..4276bc5b08cd8 100644 --- a/cmake/hip.cmake +++ b/cmake/hip.cmake @@ -3,6 +3,8 @@ if(NOT WITH_AMD_GPU) endif() include_directories("/opt/rocm/include") +include_directories("/opt/rocm/hip/include") +include_directories("/opt/rocm/miopen/include") include_directories("/opt/rocm/hipblas/include") include_directories("/opt/rocm/hiprand/include") include_directories("/opt/rocm/rocrand/include") @@ -11,20 +13,40 @@ include_directories("/opt/rocm/thrust") list(APPEND EXTERNAL_LIBS "-L/opt/rocm/lib/ -lhip_hcc") -set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -fPIC -DPADDLE_WITH_HIP -std=c++14" ) +set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -fPIC -DPADDLE_WITH_HIP -std=c++11" ) if(WITH_DSO) set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DPADDLE_USE_DSO") endif(WITH_DSO) -if(WITH_DOUBLE) - set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DPADDLE_TYPE_DOUBLE") -endif(WITH_DOUBLE) - if(WITH_TESTING) set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DPADDLE_WITH_TESTING") endif(WITH_TESTING) +if(WITH_DISTRIBUTE) + set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DPADDLE_WITH_DISTRIBUTE") +endif(WITH_DISTRIBUTE) + +if(WITH_GRPC) + set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DPADDLE_WITH_GRPC") +endif(WITH_GRPC) + +if(NOT WITH_GOLANG) + set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DPADDLE_WITHOUT_GOLANG") +endif(NOT WITH_GOLANG) + +if(WITH_MKLDNN) + set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DPADDLE_WITH_MKLDNN") +endif(WITH_MKLDNN) + +set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DANY_IMPL_ANY_CAST_MOVEABLE") + +if(NOT WITH_RDMA) + set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DPADDLE_DISABLE_RDMA") +endif(NOT WITH_RDMA) + + + if(CMAKE_BUILD_TYPE STREQUAL "Debug") list(APPEND HIP_HCC_FLAGS ${CMAKE_CXX_FLAGS_DEBUG}) elseif(CMAKE_BUILD_TYPE STREQUAL "RelWithDebInfo") diff --git a/cmake/inference_lib.cmake b/cmake/inference_lib.cmake index 7355b67ab1020..48279bc809dde 100644 --- a/cmake/inference_lib.cmake +++ b/cmake/inference_lib.cmake @@ -32,24 +32,35 @@ function(copy TARGET) list(GET copy_lib_SRCS ${index} src) list(GET copy_lib_DSTS ${index} dst) if (WIN32) - # windows cmd shell will not expand wildcard automatically. - # below expand the files,libs and copy them by rules. - file(GLOB header_files ${src} "*.h") - file(GLOB static_lib_files ${src} "*.lib") - file(GLOB dll_lib_files ${src} "*.dll") - set(src_files ${header_files} ${static_lib_files} ${dll_lib_files}) - - if (NOT "${src_files}" STREQUAL "") - list(REMOVE_DUPLICATES src_files) - endif () - add_custom_command(TARGET ${TARGET} PRE_BUILD - COMMAND ${CMAKE_COMMAND} -E make_directory "${dst}" - ) - foreach (src_file ${src_files}) + if(IS_DIRECTORY ${src}) + get_filename_component(last_path ${src} NAME) + string(APPEND dst "/" ${last_path}) + add_custom_command(TARGET ${TARGET} PRE_BUILD + COMMAND ${CMAKE_COMMAND} -E make_directory "${dst}" + ) + if(EXISTS ${src}) + add_custom_command(TARGET ${TARGET} PRE_BUILD + COMMAND cmake -E copy_directory "${src}" "${dst}" + COMMENT "copying ${src} -> ${dst}") + else() + message(WARNING "${src} not exist!") + endif() + else() + # windows cmd shell will not expand wildcard automatically. + # below expand the files, and copy them by rules. + file(GLOB src_files ${src}) + if (NOT "${src_files}" STREQUAL "") + list(REMOVE_DUPLICATES src_files) + endif () add_custom_command(TARGET ${TARGET} PRE_BUILD - COMMAND ${CMAKE_COMMAND} -E copy "${src_file}" "${dst}" - COMMENT "copying ${src_file} -> ${dst}") - endforeach () + COMMAND ${CMAKE_COMMAND} -E make_directory "${dst}" + ) + foreach (src_file ${src_files}) + add_custom_command(TARGET ${TARGET} PRE_BUILD + COMMAND ${CMAKE_COMMAND} -E copy "${src_file}" "${dst}" + COMMENT "copying ${src_file} -> ${dst}") + endforeach () + endif() else (WIN32) # not windows add_custom_command(TARGET ${TARGET} PRE_BUILD COMMAND mkdir -p "${dst}" @@ -95,7 +106,7 @@ copy(xxhash_lib DEPS xxhash ) -if (NOT PROTOBUF_FOUND) +if (NOT PROTOBUF_FOUND OR WIN32) set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/protobuf") copy(protobuf_lib SRCS ${PROTOBUF_INCLUDE_DIR} ${PROTOBUF_LIBRARY} @@ -104,20 +115,20 @@ if (NOT PROTOBUF_FOUND) ) endif () -if (NOT CBLAS_FOUND) - set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/openblas") - copy(openblas_lib - SRCS ${CBLAS_INSTALL_DIR}/lib ${CBLAS_INSTALL_DIR}/include - DSTS ${dst_dir} ${dst_dir} - DEPS extern_openblas - ) -elseif (WITH_MKLML) +if (WITH_MKLML) set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/mklml") copy(mklml_lib SRCS ${MKLML_LIB} ${MKLML_IOMP_LIB} ${MKLML_INC_DIR} DSTS ${dst_dir}/lib ${dst_dir}/lib ${dst_dir} DEPS mklml ) +elseif (NOT CBLAS_FOUND OR WIN32) + set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/openblas") + copy(openblas_lib + SRCS ${CBLAS_INSTALL_DIR}/lib ${CBLAS_INSTALL_DIR}/include + DSTS ${dst_dir} ${dst_dir} + DEPS extern_openblas + ) endif () if (WITH_MKLDNN) @@ -129,27 +140,34 @@ if (WITH_MKLDNN) ) endif () -if (NOT WIN32) - if (NOT MOBILE_INFERENCE AND NOT RPI) - set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/snappy") - copy(snappy_lib - SRCS ${SNAPPY_INCLUDE_DIR} ${SNAPPY_LIBRARIES} - DSTS ${dst_dir} ${dst_dir}/lib - DEPS snappy) - - set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/snappystream") - copy(snappystream_lib - SRCS ${SNAPPYSTREAM_INCLUDE_DIR} ${SNAPPYSTREAM_LIBRARIES} - DSTS ${dst_dir} ${dst_dir}/lib - DEPS snappystream) - - set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/zlib") - copy(zlib_lib - SRCS ${ZLIB_INCLUDE_DIR} ${ZLIB_LIBRARIES} - DSTS ${dst_dir} ${dst_dir}/lib - DEPS zlib) - endif () -endif (NOT WIN32) +if (WITH_NGRAPH) + set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/ngraph") + copy(ngraph_lib + SRCS ${NGRAPH_INC_DIR} ${NGRAPH_LIB_DIR} + DSTS ${dst_dir} ${dst_dir} + DEPS ngraph + ) +endif () + +if (NOT MOBILE_INFERENCE AND NOT RPI) + set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/snappy") + copy(snappy_lib + SRCS ${SNAPPY_INCLUDE_DIR} ${SNAPPY_LIBRARIES} + DSTS ${dst_dir} ${dst_dir}/lib + DEPS snappy) + + set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/snappystream") + copy(snappystream_lib + SRCS ${SNAPPYSTREAM_INCLUDE_DIR} ${SNAPPYSTREAM_LIBRARIES} + DSTS ${dst_dir} ${dst_dir}/lib + DEPS snappystream) + + set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/zlib") + copy(zlib_lib + SRCS ${ZLIB_INCLUDE_DIR} ${ZLIB_LIBRARIES} + DSTS ${dst_dir} ${dst_dir}/lib + DEPS zlib) +endif () # paddle fluid module set(src_dir "${PADDLE_SOURCE_DIR}/paddle/fluid") @@ -182,12 +200,23 @@ if (WITH_ANAKIN AND WITH_MKL) list(APPEND inference_deps anakin_inference_lib) endif () +if (TENSORRT_FOUND) + copy(tensorrt_lib DEPS ${inference_deps} + SRCS ${TENSORRT_ROOT}/include/Nv*.h ${TENSORRT_ROOT}/lib/libnvinfer* + DSTS ${FLUID_INSTALL_DIR}/third_party/install/tensorrt/include ${FLUID_INSTALL_DIR}/third_party/install/tensorrt/lib) +endif () + + set(module "inference") +if(WIN32) + set(paddle_fluid_lib ${PADDLE_BINARY_DIR}/paddle/fluid/inference/${CMAKE_BUILD_TYPE}/libpaddle_fluid.*) +else(WIN32) + set(paddle_fluid_lib ${PADDLE_BINARY_DIR}/paddle/fluid/inference/libpaddle_fluid.*) +endif(WIN32) copy(inference_lib DEPS ${inference_deps} - SRCS ${src_dir}/${module}/*.h ${PADDLE_BINARY_DIR}/paddle/fluid/inference/libpaddle_fluid.* + SRCS ${src_dir}/${module}/*.h ${paddle_fluid_lib} ${src_dir}/${module}/api/paddle_*.h - ${PADDLE_BINARY_DIR}/paddle/fluid/inference/api/paddle_inference_pass.h - DSTS ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module} + DSTS ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module} ) set(module "platform") @@ -225,7 +254,7 @@ copy(third_party DEPS fluid_lib_dist # only need libpaddle_fluid.so/a and paddle_*.h for inference-only library copy(inference_api_lib DEPS fluid_lib_dist - SRCS ${FLUID_INSTALL_DIR}/paddle/fluid/inference/libpaddle_fluid.* + SRCS ${paddle_fluid_lib} ${FLUID_INSTALL_DIR}/paddle/fluid/inference/paddle_*.h DSTS ${FLUID_INFERENCE_INSTALL_DIR}/paddle/lib ${FLUID_INFERENCE_INSTALL_DIR}/paddle/include ) diff --git a/cmake/operators.cmake b/cmake/operators.cmake index 17107e0698757..70d159b4f3549 100644 --- a/cmake/operators.cmake +++ b/cmake/operators.cmake @@ -84,7 +84,7 @@ function(op_library TARGET) endif() if (WIN32) # remove windows unsupported op, because windows has no nccl, no warpctc such ops. - foreach(windows_unsupport_op "nccl_op" "gen_nccl_id_op" "warpctc_op") + foreach(windows_unsupport_op "nccl_op" "gen_nccl_id_op") if ("${TARGET}" STREQUAL "${windows_unsupport_op}") return() endif() @@ -109,7 +109,8 @@ function(op_library TARGET) # Define operators that don't need pybind here. foreach(manual_pybind_op "compare_op" "logical_op" "nccl_op" -"tensor_array_read_write_op" "tensorrt_engine_op" "conv_fusion_op") +"tensor_array_read_write_op" "tensorrt_engine_op" "conv_fusion_op" +"fusion_transpose_flatten_concat_op") if ("${TARGET}" STREQUAL "${manual_pybind_op}") set(pybind_flag 1) endif() @@ -165,6 +166,8 @@ function(op_library TARGET) # Append first implemented MKLDNN activation operator if (${MKLDNN_FILE} STREQUAL "activation_mkldnn_op") file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(relu, MKLDNN);\n") + elseif(${MKLDNN_FILE} STREQUAL "conv_mkldnn_op") + file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL_WITH_CUSTOM_TYPE(conv2d, MKLDNN, FP32);\n") else() file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, MKLDNN);\n") endif() diff --git a/paddle/fluid/API.spec b/paddle/fluid/API.spec index 541c4db1fa091..b6974c6af2904 100644 --- a/paddle/fluid/API.spec +++ b/paddle/fluid/API.spec @@ -26,12 +26,27 @@ paddle.fluid.release_memory ArgSpec(args=['input_program', 'skip_opt_set'], vara paddle.fluid.DistributeTranspilerConfig.__init__ paddle.fluid.ParallelExecutor.__init__ ArgSpec(args=['self', 'use_cuda', 'loss_name', 'main_program', 'share_vars_from', 'exec_strategy', 'build_strategy', 'num_trainers', 'trainer_id', 'scope'], varargs=None, keywords=None, defaults=(None, None, None, None, None, 1, 0, None)) paddle.fluid.ParallelExecutor.run ArgSpec(args=['self', 'fetch_list', 'feed', 'feed_dict', 'return_numpy'], varargs=None, keywords=None, defaults=(None, None, True)) -paddle.fluid.ExecutionStrategy.__init__ __init__(self: paddle.fluid.core.ExecutionStrategy) -> None -paddle.fluid.BuildStrategy.GradientScaleStrategy.__init__ __init__(self: paddle.fluid.core.GradientScaleStrategy, arg0: int) -> None -paddle.fluid.BuildStrategy.ReduceStrategy.__init__ __init__(self: paddle.fluid.core.ReduceStrategy, arg0: int) -> None -paddle.fluid.BuildStrategy.__init__ __init__(self: paddle.fluid.core.BuildStrategy) -> None +paddle.fluid.ExecutionStrategy.__init__ __init__(self: paddle.fluid.core.ParallelExecutor.ExecutionStrategy) -> None +paddle.fluid.BuildStrategy.GradientScaleStrategy.__init__ __init__(self: paddle.fluid.core.ParallelExecutor.BuildStrategy.GradientScaleStrategy, arg0: int) -> None +paddle.fluid.BuildStrategy.ReduceStrategy.__init__ __init__(self: paddle.fluid.core.ParallelExecutor.BuildStrategy.ReduceStrategy, arg0: int) -> None +paddle.fluid.BuildStrategy.__init__ __init__(self: paddle.fluid.core.ParallelExecutor.BuildStrategy) -> None paddle.fluid.create_lod_tensor ArgSpec(args=['data', 'recursive_seq_lens', 'place'], varargs=None, keywords=None, defaults=None) paddle.fluid.create_random_int_lodtensor ArgSpec(args=['recursive_seq_lens', 'base_shape', 'place', 'low', 'high'], varargs=None, keywords=None, defaults=None) +paddle.fluid.DataFeedDesc.__init__ ArgSpec(args=['self', 'proto_file'], varargs=None, keywords=None, defaults=None) +paddle.fluid.DataFeedDesc.desc ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) +paddle.fluid.DataFeedDesc.set_batch_size ArgSpec(args=['self', 'batch_size'], varargs=None, keywords=None, defaults=None) +paddle.fluid.DataFeedDesc.set_dense_slots ArgSpec(args=['self', 'dense_slots_name'], varargs=None, keywords=None, defaults=None) +paddle.fluid.DataFeedDesc.set_use_slots ArgSpec(args=['self', 'use_slots_name'], varargs=None, keywords=None, defaults=None) +paddle.fluid.AsyncExecutor.__init__ ArgSpec(args=['self', 'place', 'run_mode'], varargs=None, keywords=None, defaults=(None, '')) +paddle.fluid.AsyncExecutor.config_distributed_nodes ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) +paddle.fluid.AsyncExecutor.download_data ArgSpec(args=['self', 'afs_path', 'local_path', 'fs_default_name', 'ugi', 'file_cnt', 'hadoop_home', 'process_num'], varargs=None, keywords=None, defaults=('$HADOOP_HOME', 12)) +paddle.fluid.AsyncExecutor.get_instance ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) +paddle.fluid.AsyncExecutor.init_model ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) +paddle.fluid.AsyncExecutor.init_server ArgSpec(args=['self', 'dist_desc'], varargs=None, keywords=None, defaults=None) +paddle.fluid.AsyncExecutor.init_worker ArgSpec(args=['self', 'dist_desc', 'startup_program'], varargs=None, keywords=None, defaults=None) +paddle.fluid.AsyncExecutor.run ArgSpec(args=['self', 'program', 'data_feed', 'filelist', 'thread_num', 'fetch', 'mode', 'debug'], varargs=None, keywords=None, defaults=('', False)) +paddle.fluid.AsyncExecutor.save_model ArgSpec(args=['self', 'save_path'], varargs=None, keywords=None, defaults=None) +paddle.fluid.AsyncExecutor.stop ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) paddle.fluid.io.save_vars ArgSpec(args=['executor', 'dirname', 'main_program', 'vars', 'predicate', 'filename'], varargs=None, keywords=None, defaults=(None, None, None, None)) paddle.fluid.io.save_params ArgSpec(args=['executor', 'dirname', 'main_program', 'filename'], varargs=None, keywords=None, defaults=(None, None)) paddle.fluid.io.save_persistables ArgSpec(args=['executor', 'dirname', 'main_program', 'filename'], varargs=None, keywords=None, defaults=(None, None)) @@ -59,6 +74,7 @@ paddle.fluid.layers.linear_chain_crf ArgSpec(args=['input', 'label', 'param_attr paddle.fluid.layers.crf_decoding ArgSpec(args=['input', 'param_attr', 'label'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.cos_sim ArgSpec(args=['X', 'Y'], varargs=None, keywords=None, defaults=None) paddle.fluid.layers.cross_entropy ArgSpec(args=['input', 'label', 'soft_label', 'ignore_index'], varargs=None, keywords=None, defaults=(False, -100)) +paddle.fluid.layers.bpr_loss ArgSpec(args=['input', 'label', 'name'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.square_error_cost ArgSpec(args=['input', 'label'], varargs=None, keywords=None, defaults=None) paddle.fluid.layers.chunk_eval ArgSpec(args=['input', 'label', 'chunk_scheme', 'num_chunk_types', 'excluded_chunk_types'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.sequence_conv ArgSpec(args=['input', 'num_filters', 'filter_size', 'filter_stride', 'padding', 'bias_attr', 'param_attr', 'act', 'name'], varargs=None, keywords=None, defaults=(3, 1, None, None, None, None, None)) @@ -69,7 +85,9 @@ paddle.fluid.layers.sequence_softmax ArgSpec(args=['input', 'use_cudnn', 'name'] paddle.fluid.layers.softmax ArgSpec(args=['input', 'use_cudnn', 'name'], varargs=None, keywords=None, defaults=(True, None)) paddle.fluid.layers.pool2d ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name', 'exclusive'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None, True)) paddle.fluid.layers.pool3d ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name', 'exclusive'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None, True)) -paddle.fluid.layers.batch_norm ArgSpec(args=['input', 'act', 'is_test', 'momentum', 'epsilon', 'param_attr', 'bias_attr', 'data_layout', 'in_place', 'name', 'moving_mean_name', 'moving_variance_name', 'do_model_average_for_mean_and_var', 'fuse_with_relu'], varargs=None, keywords=None, defaults=(None, False, 0.9, 1e-05, None, None, 'NCHW', False, None, None, None, False, False)) +paddle.fluid.layers.adaptive_pool2d ArgSpec(args=['input', 'pool_size', 'pool_type', 'require_index', 'name'], varargs=None, keywords=None, defaults=('max', False, None)) +paddle.fluid.layers.adaptive_pool3d ArgSpec(args=['input', 'pool_size', 'pool_type', 'require_index', 'name'], varargs=None, keywords=None, defaults=('max', False, None)) +paddle.fluid.layers.batch_norm ArgSpec(args=['input', 'act', 'is_test', 'momentum', 'epsilon', 'param_attr', 'bias_attr', 'data_layout', 'in_place', 'name', 'moving_mean_name', 'moving_variance_name', 'do_model_average_for_mean_and_var', 'fuse_with_relu', 'use_global_stats'], varargs=None, keywords=None, defaults=(None, False, 0.9, 1e-05, None, None, 'NCHW', False, None, None, None, False, False, False)) paddle.fluid.layers.beam_search_decode ArgSpec(args=['ids', 'scores', 'beam_size', 'end_id', 'name'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.conv2d_transpose ArgSpec(args=['input', 'num_filters', 'output_size', 'filter_size', 'padding', 'stride', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, None, 0, 1, 1, None, None, None, True, None, None)) paddle.fluid.layers.conv3d_transpose ArgSpec(args=['input', 'num_filters', 'output_size', 'filter_size', 'padding', 'stride', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, None, 0, 1, 1, None, None, None, True, None, None)) @@ -97,8 +115,8 @@ paddle.fluid.layers.warpctc ArgSpec(args=['input', 'label', 'blank', 'norm_by_ti paddle.fluid.layers.sequence_reshape ArgSpec(args=['input', 'new_dim'], varargs=None, keywords=None, defaults=None) paddle.fluid.layers.transpose ArgSpec(args=['x', 'perm', 'name'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.im2sequence ArgSpec(args=['input', 'filter_size', 'stride', 'padding', 'input_image_size', 'out_stride', 'name'], varargs=None, keywords=None, defaults=(1, 1, 0, None, 1, None)) -paddle.fluid.layers.nce ArgSpec(args=['input', 'label', 'num_total_classes', 'sample_weight', 'param_attr', 'bias_attr', 'num_neg_samples', 'name', 'sampler', 'custom_dist', 'seed'], varargs=None, keywords=None, defaults=(None, None, None, None, None, 'uniform', None, 0)) -paddle.fluid.layers.hsigmoid ArgSpec(args=['input', 'label', 'num_classes', 'param_attr', 'bias_attr', 'name'], varargs=None, keywords=None, defaults=(None, None, None)) +paddle.fluid.layers.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.beam_search ArgSpec(args=['pre_ids', 'pre_scores', 'ids', 'scores', 'beam_size', 'end_id', 'level', 'name'], varargs=None, keywords=None, defaults=(0, None)) paddle.fluid.layers.row_conv ArgSpec(args=['input', 'future_context_size', 'param_attr', 'act'], varargs=None, keywords=None, defaults=(None, None)) paddle.fluid.layers.multiplex ArgSpec(args=['inputs', 'index'], varargs=None, keywords=None, defaults=None) @@ -175,7 +193,7 @@ paddle.fluid.layers.clip ArgSpec(args=['x', 'min', 'max', 'name'], varargs=None, paddle.fluid.layers.clip_by_norm ArgSpec(args=['x', 'max_norm', 'name'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.mean ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.mul ArgSpec(args=['x', 'y', 'x_num_col_dims', 'y_num_col_dims', 'name'], varargs=None, keywords=None, defaults=(1, 1, None)) -paddle.fluid.layers.sigmoid_cross_entropy_with_logits ArgSpec(args=['x', 'label', 'name'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.layers.sigmoid_cross_entropy_with_logits ArgSpec(args=['x', 'label', 'ignore_index', 'name'], varargs=None, keywords=None, defaults=(-100, None)) paddle.fluid.layers.maxout ArgSpec(args=['x', 'groups', 'name'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.space_to_depth ArgSpec(args=['x', 'blocksize', 'name'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.affine_grid ArgSpec(args=['theta', 'out_shape', 'name'], varargs=None, keywords=None, defaults=(None,)) @@ -187,6 +205,12 @@ paddle.fluid.layers.grid_sampler ArgSpec(args=['x', 'grid', 'name'], varargs=Non paddle.fluid.layers.log_loss ArgSpec(args=['input', 'label', 'epsilon', 'name'], varargs=None, keywords=None, defaults=(0.0001, None)) paddle.fluid.layers.add_position_encoding ArgSpec(args=['input', 'alpha', 'beta', 'name'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.bilinear_tensor_product ArgSpec(args=['x', 'y', 'size', 'act', 'name', 'param_attr', 'bias_attr'], varargs=None, keywords=None, defaults=(None, None, None, None)) +paddle.fluid.layers.merge_selected_rows ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.layers.get_tensor_from_selected_rows ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.layers.lstm ArgSpec(args=['input', 'init_h', 'init_c', 'max_len', 'hidden_size', 'num_layers', 'dropout_prob', 'is_bidirec', 'is_test', 'name', 'default_initializer', 'seed'], varargs=None, keywords=None, defaults=(0.0, False, False, None, None, -1)) +paddle.fluid.layers.py_func ArgSpec(args=['func', 'x', 'out', 'backward_func', 'skip_vars_in_backward_input'], varargs=None, keywords=None, defaults=(None, None)) +paddle.fluid.layers.psroi_pool ArgSpec(args=['input', 'rois', 'output_channels', 'spatial_scale', 'pooled_height', 'pooled_width', 'name'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.layers.huber_loss ArgSpec(args=['input', 'label', 'delta'], varargs=None, keywords=None, defaults=None) paddle.fluid.layers.data ArgSpec(args=['name', 'shape', 'append_batch_size', 'dtype', 'lod_level', 'type', 'stop_gradient'], varargs=None, keywords=None, defaults=(True, 'float32', 0, VarType.LOD_TENSOR, True)) paddle.fluid.layers.open_files ArgSpec(args=['filenames', 'shapes', 'lod_levels', 'dtypes', 'thread_num', 'buffer_size', 'pass_num', 'is_test'], varargs=None, keywords=None, defaults=(None, None, 1, None)) paddle.fluid.layers.read_file ArgSpec(args=['reader'], varargs=None, keywords=None, defaults=None) @@ -276,7 +300,7 @@ paddle.fluid.layers.hard_shrink ArgSpec(args=['x', 'threshold'], varargs=None, k paddle.fluid.layers.cumsum ArgSpec(args=['x', 'axis', 'exclusive', 'reverse'], varargs=None, keywords=None, defaults=(None, None, None)) paddle.fluid.layers.thresholded_relu ArgSpec(args=['x', 'threshold'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.prior_box ArgSpec(args=['input', 'image', 'min_sizes', 'max_sizes', 'aspect_ratios', 'variance', 'flip', 'clip', 'steps', 'offset', 'name', 'min_max_aspect_ratios_order'], varargs=None, keywords=None, defaults=(None, [1.0], [0.1, 0.1, 0.2, 0.2], False, False, [0.0, 0.0], 0.5, None, False)) -paddle.fluid.layers.density_prior_box ArgSpec(args=['input', 'image', 'densities', 'fixed_sizes', 'fixed_ratios', 'variance', 'clip', 'steps', 'offset', 'name'], varargs=None, keywords=None, defaults=(None, None, None, [0.1, 0.1, 0.2, 0.2], False, [0.0, 0.0], 0.5, None)) +paddle.fluid.layers.density_prior_box ArgSpec(args=['input', 'image', 'densities', 'fixed_sizes', 'fixed_ratios', 'variance', 'clip', 'steps', 'offset', 'flatten_to_2d', 'name'], varargs=None, keywords=None, defaults=(None, None, None, [0.1, 0.1, 0.2, 0.2], False, [0.0, 0.0], 0.5, False, None)) paddle.fluid.layers.multi_box_head ArgSpec(args=['inputs', 'image', 'base_size', 'num_classes', 'aspect_ratios', 'min_ratio', 'max_ratio', 'min_sizes', 'max_sizes', 'steps', 'step_w', 'step_h', 'offset', 'variance', 'flip', 'clip', 'kernel_size', 'pad', 'stride', 'name', 'min_max_aspect_ratios_order'], varargs=None, keywords=None, defaults=(None, None, None, None, None, None, None, 0.5, [0.1, 0.1, 0.2, 0.2], True, False, 1, 0, 1, None, False)) paddle.fluid.layers.bipartite_match ArgSpec(args=['dist_matrix', 'match_type', 'dist_threshold', 'name'], varargs=None, keywords=None, defaults=(None, None, None)) paddle.fluid.layers.target_assign ArgSpec(args=['input', 'matched_indices', 'negative_indices', 'mismatch_value', 'name'], varargs=None, keywords=None, defaults=(None, None, None)) @@ -291,6 +315,7 @@ paddle.fluid.layers.generate_proposals ArgSpec(args=['scores', 'bbox_deltas', 'i 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.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.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.exponential_decay ArgSpec(args=['learning_rate', 'decay_steps', 'decay_rate', 'staircase'], varargs=None, keywords=None, defaults=(False,)) @@ -326,6 +351,39 @@ 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.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.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.add_strategy ArgSpec(args=['self', 'strategy'], varargs=None, keywords=None, defaults=None) +paddle.fluid.contrib.CompressPass.apply ArgSpec(args=['self', 'graph'], varargs=None, keywords=None, defaults=None) +paddle.fluid.contrib.ImitationGraph.__init__ ArgSpec(args=['self', 'program'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.contrib.ImitationGraph.all_parameters ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) +paddle.fluid.contrib.SensitivePruneStrategy.__init__ ArgSpec(args=['self', 'pruner', 'start_epoch', 'end_epoch', 'delta_rate', 'acc_loss_threshold', 'sensitivities'], varargs=None, keywords=None, defaults=(None, 0, 10, 0.2, 0.2, None)) +paddle.fluid.contrib.SensitivePruneStrategy.on_batch_begin ArgSpec(args=['self', 'context'], varargs=None, keywords=None, defaults=None) +paddle.fluid.contrib.SensitivePruneStrategy.on_batch_end ArgSpec(args=['self', 'context'], varargs=None, keywords=None, defaults=None) +paddle.fluid.contrib.SensitivePruneStrategy.on_compress_begin ArgSpec(args=['self', 'context'], varargs=None, keywords=None, defaults=None) +paddle.fluid.contrib.SensitivePruneStrategy.on_compress_end ArgSpec(args=['self', 'context'], varargs=None, keywords=None, defaults=None) +paddle.fluid.contrib.SensitivePruneStrategy.on_epoch_begin ArgSpec(args=['self', 'context'], varargs=None, keywords=None, defaults=None) +paddle.fluid.contrib.SensitivePruneStrategy.on_epoch_end ArgSpec(args=['self', 'context'], varargs=None, keywords=None, defaults=None) +paddle.fluid.contrib.MagnitudePruner.__init__ ArgSpec(args=['self', 'threshold'], varargs=None, keywords=None, defaults=None) +paddle.fluid.contrib.MagnitudePruner.prune ArgSpec(args=['self', 'param', 'threshold'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.contrib.RatioPruner.__init__ ArgSpec(args=['self', 'ratios'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.contrib.RatioPruner.prune ArgSpec(args=['self', 'param', 'ratio'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.contrib.load_persistables_for_increment ArgSpec(args=['dirname', 'executor', 'program', 'lookup_table_var', 'lookup_table_var_path'], varargs=None, keywords=None, defaults=None) +paddle.fluid.contrib.load_persistables_for_inference ArgSpec(args=['dirname', 'executor', 'program', 'lookup_table_var_name'], varargs=None, keywords=None, defaults=None) +paddle.fluid.contrib.convert_dist_to_sparse_program ArgSpec(args=['program'], varargs=None, keywords=None, defaults=None) +paddle.fluid.contrib.HDFSClient.__init__ ArgSpec(args=['self', 'hadoop_home', 'configs'], varargs=None, keywords=None, defaults=None) +paddle.fluid.contrib.HDFSClient.delete ArgSpec(args=['self', 'hdfs_path'], varargs=None, keywords=None, defaults=None) +paddle.fluid.contrib.HDFSClient.download ArgSpec(args=['self', 'hdfs_path', 'local_path', 'overwrite', 'unzip'], varargs=None, keywords=None, defaults=(False, False)) +paddle.fluid.contrib.HDFSClient.is_dir ArgSpec(args=['self', 'hdfs_path'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.contrib.HDFSClient.is_exist ArgSpec(args=['self', 'hdfs_path'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.contrib.HDFSClient.ls ArgSpec(args=['self', 'hdfs_path'], varargs=None, keywords=None, defaults=None) +paddle.fluid.contrib.HDFSClient.lsr ArgSpec(args=['self', 'hdfs_path', 'only_file', 'sort'], varargs=None, keywords=None, defaults=(True, True)) +paddle.fluid.contrib.HDFSClient.make_local_dirs ArgSpec(args=['local_path'], varargs=None, keywords=None, defaults=None) +paddle.fluid.contrib.HDFSClient.makedirs ArgSpec(args=['self', 'hdfs_path'], varargs=None, keywords=None, defaults=None) +paddle.fluid.contrib.HDFSClient.rename ArgSpec(args=['self', 'hdfs_src_path', 'hdfs_dst_path', 'overwrite'], varargs=None, keywords=None, defaults=(False,)) +paddle.fluid.contrib.HDFSClient.upload ArgSpec(args=['self', 'hdfs_path', 'local_path', 'overwrite', 'retry_times'], varargs=None, keywords=None, defaults=(False, 5)) +paddle.fluid.contrib.multi_download ArgSpec(args=['client', 'hdfs_path', 'local_path', 'trainer_id', 'trainers', 'multi_processes'], varargs=None, keywords=None, defaults=(5,)) +paddle.fluid.contrib.multi_upload ArgSpec(args=['client', 'hdfs_path', 'local_path', 'multi_processes', 'overwrite', 'sync'], varargs=None, keywords=None, defaults=(5, False, True)) paddle.fluid.transpiler.DistributeTranspiler.__init__ ArgSpec(args=['self', 'config'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.transpiler.DistributeTranspiler.get_pserver_program ArgSpec(args=['self', 'endpoint'], varargs=None, keywords=None, defaults=None) paddle.fluid.transpiler.DistributeTranspiler.get_pserver_programs ArgSpec(args=['self', 'endpoint'], varargs=None, keywords=None, defaults=None) @@ -342,7 +400,7 @@ paddle.fluid.transpiler.RoundRobin.dispatch ArgSpec(args=['self', 'varlist'], va paddle.fluid.transpiler.RoundRobin.reset ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) paddle.fluid.transpiler.DistributeTranspilerConfig.__init__ paddle.fluid.nets.simple_img_conv_pool ArgSpec(args=['input', 'num_filters', 'filter_size', 'pool_size', 'pool_stride', 'pool_padding', 'pool_type', 'global_pooling', 'conv_stride', 'conv_padding', 'conv_dilation', 'conv_groups', 'param_attr', 'bias_attr', 'act', 'use_cudnn'], varargs=None, keywords=None, defaults=(0, 'max', False, 1, 0, 1, 1, None, None, None, True)) -paddle.fluid.nets.sequence_conv_pool ArgSpec(args=['input', 'num_filters', 'filter_size', 'param_attr', 'act', 'pool_type'], varargs=None, keywords=None, defaults=(None, 'sigmoid', 'max')) +paddle.fluid.nets.sequence_conv_pool ArgSpec(args=['input', 'num_filters', 'filter_size', 'param_attr', 'act', 'pool_type', 'bias_attr'], varargs=None, keywords=None, defaults=(None, 'sigmoid', 'max', None)) paddle.fluid.nets.glu ArgSpec(args=['input', 'dim'], varargs=None, keywords=None, defaults=(-1,)) paddle.fluid.nets.scaled_dot_product_attention ArgSpec(args=['queries', 'keys', 'values', 'num_heads', 'dropout_rate'], varargs=None, keywords=None, defaults=(1, 0.0)) paddle.fluid.nets.img_conv_group ArgSpec(args=['input', 'conv_num_filter', 'pool_size', 'conv_padding', 'conv_filter_size', 'conv_act', 'param_attr', 'conv_with_batchnorm', 'conv_batchnorm_drop_rate', 'pool_stride', 'pool_type', 'use_cudnn'], varargs=None, keywords=None, defaults=(1, 3, None, None, False, 0.0, 1, 'max', True)) @@ -352,7 +410,7 @@ paddle.fluid.optimizer.MomentumOptimizer.__init__ ArgSpec(args=['self', 'learnin paddle.fluid.optimizer.MomentumOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)) paddle.fluid.optimizer.AdagradOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'epsilon', 'regularization', 'name'], varargs=None, keywords=None, defaults=(1e-06, None, None)) paddle.fluid.optimizer.AdagradOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)) -paddle.fluid.optimizer.AdamOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'beta1', 'beta2', 'epsilon', 'regularization', 'name'], varargs=None, keywords=None, defaults=(0.001, 0.9, 0.999, 1e-08, None, None)) +paddle.fluid.optimizer.AdamOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'beta1', 'beta2', 'epsilon', 'regularization', 'name', 'lazy_mode'], varargs=None, keywords=None, defaults=(0.001, 0.9, 0.999, 1e-08, None, None, False)) paddle.fluid.optimizer.AdamOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)) paddle.fluid.optimizer.AdamaxOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'beta1', 'beta2', 'epsilon', 'regularization', 'name'], varargs=None, keywords=None, defaults=(0.001, 0.9, 0.999, 1e-08, None, None)) paddle.fluid.optimizer.AdamaxOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)) @@ -411,3 +469,17 @@ paddle.fluid.Scope.drop_kids drop_kids(self: paddle.fluid.core.Scope) -> None paddle.fluid.Scope.find_var find_var(self: paddle.fluid.core.Scope, arg0: unicode) -> paddle.fluid.core.Variable paddle.fluid.Scope.new_scope new_scope(self: paddle.fluid.core.Scope) -> paddle.fluid.core.Scope paddle.fluid.Scope.var var(self: paddle.fluid.core.Scope, arg0: unicode) -> paddle.fluid.core.Variable +paddle.reader.map_readers ArgSpec(args=['func'], varargs='readers', keywords=None, defaults=None) +paddle.reader.buffered ArgSpec(args=['reader', 'size'], varargs=None, keywords=None, defaults=None) +paddle.reader.compose ArgSpec(args=[], varargs='readers', keywords='kwargs', defaults=None) +paddle.reader.chain ArgSpec(args=[], varargs='readers', keywords=None, defaults=None) +paddle.reader.shuffle ArgSpec(args=['reader', 'buf_size'], varargs=None, keywords=None, defaults=None) +paddle.reader.firstn ArgSpec(args=['reader', 'n'], varargs=None, keywords=None, defaults=None) +paddle.reader.xmap_readers ArgSpec(args=['mapper', 'reader', 'process_num', 'buffer_size', 'order'], varargs=None, keywords=None, defaults=(False,)) +paddle.reader.PipeReader.__init__ ArgSpec(args=['self', 'command', 'bufsize', 'file_type'], varargs=None, keywords=None, defaults=(8192, 'plain')) +paddle.reader.PipeReader.get_line ArgSpec(args=['self', 'cut_lines', 'line_break'], varargs=None, keywords=None, defaults=(True, '\n')) +paddle.reader.multiprocess_reader ArgSpec(args=['readers', 'use_pipe', 'queue_size'], varargs=None, keywords=None, defaults=(True, 1000)) +paddle.reader.Fake.__init__ ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) +paddle.reader.creator.np_array ArgSpec(args=['x'], varargs=None, keywords=None, defaults=None) +paddle.reader.creator.text_file ArgSpec(args=['path'], varargs=None, keywords=None, defaults=None) +paddle.reader.creator.recordio ArgSpec(args=['paths', 'buf_size'], varargs=None, keywords=None, defaults=(100,)) diff --git a/paddle/fluid/CMakeLists.txt b/paddle/fluid/CMakeLists.txt index 6b526f0103ad3..595454e90b9cd 100644 --- a/paddle/fluid/CMakeLists.txt +++ b/paddle/fluid/CMakeLists.txt @@ -1,6 +1,7 @@ add_subdirectory(memory) add_subdirectory(platform) add_subdirectory(framework) +add_subdirectory(imperative) add_subdirectory(operators) add_subdirectory(string) add_subdirectory(recordio) diff --git a/paddle/fluid/framework/CMakeLists.txt b/paddle/fluid/framework/CMakeLists.txt index 43e1bc6b2efec..412bc9cbe88b8 100644 --- a/paddle/fluid/framework/CMakeLists.txt +++ b/paddle/fluid/framework/CMakeLists.txt @@ -1,17 +1,18 @@ -# windows treat symbolic file as a real file, which is different with unix -# We create a hidden file and compile it instead of origin source file. +#windows treat symbolic file as a real file, which is different with unix +#We create a hidden file and compile it instead of origin source file. function(windows_symbolic TARGET) set(oneValueArgs "") - set(multiValueArgs SRCS DEPS) + set(multiValueArgs SRCS PATH) cmake_parse_arguments(windows_symbolic "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) + set(final_path ${CMAKE_CURRENT_SOURCE_DIR}/${windows_symbolic_PATH}) foreach(src ${windows_symbolic_SRCS}) get_filename_component(src ${src} NAME_WE) if (NOT EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${src}.cc OR NOT EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${src}.cu) message(FATAL " ${src}.cc and ${src}.cu must exsits, and ${src}.cu must be symbolic file.") endif() - # only copy the xx.cu to .xx.cu when the content are modified +#only copy the xx.cu to.xx.cu when the content are modified set(copy_flag 1) if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/.${src}.cu) file(READ ${CMAKE_CURRENT_SOURCE_DIR}/${src}.cc SOURCE_STR) @@ -32,8 +33,9 @@ endfunction() add_subdirectory(ir) add_subdirectory(details) -# ddim lib +#ddim lib proto_library(framework_proto SRCS framework.proto) +proto_library(async_executor_proto SRCS data_feed.proto) cc_library(ddim SRCS ddim.cc DEPS eigen3 boost) cc_test(ddim_test SRCS ddim_test.cc DEPS ddim) @@ -71,6 +73,8 @@ cc_library(lod_tensor SRCS lod_tensor.cc DEPS ddim place tensor framework_proto cc_test(lod_tensor_test SRCS lod_tensor_test.cc DEPS lod_tensor memory) nv_test(lod_tensor_gpu_test SRCS lod_tensor_test.cu DEPS lod_tensor) +cc_library(garbage_collector SRCS garbage_collector.cc DEPS device_context memory) + cc_library(reader SRCS reader.cc DEPS lod_tensor ddim) cc_test(reader_test SRCS reader_test.cc DEPS reader) @@ -88,8 +92,8 @@ nv_test(data_device_transform_test SRCS data_device_transform_test.cu if(WITH_GPU) if (WIN32) - # windows treat symbolic file as a real file, which is different with unix - # We create a hidden file and compile it instead of origin source file. +#windows treat symbolic file as a real file, which is different with unix +#We create a hidden file and compile it instead of origin source file. windows_symbolic(hidden_file SRCS data_type_transform.cu) nv_library(data_type_transform SRCS .data_type_transform.cu DEPS tensor) add_dependencies(data_type_transform hidden_file) @@ -116,8 +120,10 @@ cc_test(op_proto_maker_test SRCS op_proto_maker_test.cc DEPS op_proto_maker) cc_library(op_info SRCS op_info.cc DEPS attribute framework_proto) cc_library(shape_inference SRCS shape_inference.cc DEPS ddim attribute device_context) +cc_library(transfer_scope_cache SRCS transfer_scope_cache.cc DEPS scope framework_proto device_context) +cc_library(op_kernel_type SRCS op_kernel_type.cc DEPS device_context place) cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope glog - shape_inference data_transform lod_tensor profiler) + shape_inference data_transform lod_tensor profiler transfer_scope_cache op_kernel_type) cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry device_context) @@ -125,17 +131,21 @@ cc_library(version SRCS version.cc) cc_test(version_test SRCS version_test.cc DEPS version) cc_library(proto_desc SRCS var_desc.cc op_desc.cc block_desc.cc program_desc.cc DEPS shape_inference op_info operator glog version) -cc_library(ngraph_bridge SRCS ngraph_bridge.cc DEPS operator framework_proto) -if(NOT WIN32) -cc_library(ngraph_operator SRCS ngraph_operator.cc DEPS ngraph_bridge operator op_info device_context tensor scope glog - shape_inference data_transform lod_tensor profiler) -endif(NOT WIN32) + +if(WITH_NGRAPH) + if(NOT WIN32) + cc_library(ngraph_bridge SRCS ngraph_bridge.cc DEPS operator framework_proto ngraph) + cc_library(ngraph_operator SRCS ngraph_operator.cc DEPS ngraph_bridge operator op_info device_context tensor scope glog + shape_inference data_transform lod_tensor profiler ngraph) + endif(NOT WIN32) +endif(WITH_NGRAPH) cc_library(op_registry SRCS op_registry.cc DEPS op_proto_maker op_info operator glog proto_desc) nv_test(op_registry_test SRCS op_registry_test.cc DEPS op_registry) -py_proto_compile(framework_py_proto SRCS framework.proto) -# Generate an empty __init__.py to make framework_py_proto as a valid python module. +py_proto_compile(framework_py_proto SRCS framework.proto data_feed.proto) +#Generate an empty \ + #__init__.py to make framework_py_proto as a valid python module. add_custom_target(framework_py_proto_init ALL COMMAND ${CMAKE_COMMAND} -E touch __init__.py) add_dependencies(framework_py_proto framework_py_proto_init) if (NOT WIN32) @@ -156,27 +166,45 @@ endif(NOT WIN32) cc_library(lod_rank_table SRCS lod_rank_table.cc DEPS lod_tensor) cc_library(feed_fetch_method SRCS feed_fetch_method.cc DEPS lod_tensor scope glog) +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) +cc_library(naive_executor SRCS naive_executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass variable_helper) if(WITH_DISTRIBUTE) - cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method sendrecvop_grpc cares grpc++_unsecure grpc_unsecure gpr graph_to_program_pass) - set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor") - set_source_files_properties(executor.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) + cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog + lod_rank_table feed_fetch_method sendrecvop_rpc ${GLOB_DISTRIBUTE_DEPS} graph_to_program_pass variable_helper) + + set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor") + set_source_files_properties(executor.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) + else() - if(NOT WIN32) - cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass ngraph_operator) - else(NOT WIN32) - cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass) - endif(NOT WIN32) + if(WITH_NGRAPH) + if(NOT WIN32) + cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass ngraph ngraph_operator variable_helper) + else(NOT WIN32) + cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass variable_helper) + endif(NOT WIN32) + else(WITH_NGRAPH) + cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass variable_helper) + endif(WITH_NGRAPH) cc_test(test_naive_executor SRCS naive_executor_test.cc DEPS naive_executor elementwise_add_op) endif() +target_link_libraries(executor garbage_collector) + cc_library(parallel_executor SRCS parallel_executor.cc DEPS threaded_ssa_graph_executor scope_buffered_ssa_graph_executor graph build_strategy - fast_threaded_ssa_graph_executor) + fast_threaded_ssa_graph_executor variable_helper) +if(WITH_PSLIB) + cc_library(async_executor SRCS async_executor.cc data_feed.cc data_feed_factory.cc executor_thread_worker.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass async_executor_proto variable_helper pslib_brpc pslib) +else() + cc_library(async_executor SRCS async_executor.cc data_feed.cc data_feed_factory.cc executor_thread_worker.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass async_executor_proto variable_helper) +endif(WITH_PSLIB) + + +cc_test(data_feed_test SRCS data_feed_test.cc DEPS async_executor) cc_library(prune SRCS prune.cc DEPS framework_proto) 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 @@ -184,7 +212,7 @@ cc_test(var_type_inference_test SRCS var_type_inference_test.cc DEPS op_registry 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(op_kernel_type_test SRCS op_kernel_type_test.cc DEPS place device_context framework_proto) +cc_test(op_kernel_type_test SRCS op_kernel_type_test.cc DEPS place device_context framework_proto op_kernel_type) cc_test(cow_ptr_tests SRCS details/cow_ptr_test.cc) cc_test(tuple_test SRCS tuple_test.cc ) @@ -192,3 +220,6 @@ cc_test(tuple_test SRCS tuple_test.cc ) if (NOT WIN32) cc_test(rw_lock_test SRCS rw_lock_test.cc) endif (NOT WIN32) + +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) diff --git a/paddle/fluid/framework/async_executor.cc b/paddle/fluid/framework/async_executor.cc new file mode 100644 index 0000000000000..ee3c5e01f87ee --- /dev/null +++ b/paddle/fluid/framework/async_executor.cc @@ -0,0 +1,325 @@ +/* Copyright (c) 2016 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/async_executor.h" +#include "google/protobuf/io/zero_copy_stream_impl.h" +#include "google/protobuf/message.h" +#include "google/protobuf/text_format.h" + +#include "gflags/gflags.h" +#include "paddle/fluid/framework/data_feed_factory.h" +#include "paddle/fluid/framework/executor_thread_worker.h" +#include "paddle/fluid/framework/feed_fetch_method.h" +#include "paddle/fluid/framework/feed_fetch_type.h" +#include "paddle/fluid/framework/lod_rank_table.h" +#include "paddle/fluid/framework/lod_tensor_array.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/framework/reader.h" +#include "paddle/fluid/inference/io.h" +#include "paddle/fluid/platform/place.h" +#include "paddle/fluid/pybind/pybind.h" +#ifdef PADDLE_WITH_PSLIB +#include +#endif + +namespace paddle { +namespace framework { +AsyncExecutor::AsyncExecutor(Scope* scope, const platform::Place& place) + : root_scope_(scope), place_(place) {} + +void AsyncExecutor::CreateThreads( + ExecutorThreadWorker* worker, const ProgramDesc& main_program, + const std::shared_ptr& reader, + const std::vector& fetch_var_names, Scope* root_scope, + const int thread_index, const bool debug) { + worker->SetThreadId(thread_index); + worker->SetDebug(debug); + worker->SetRootScope(root_scope); + worker->CreateThreadResource(main_program, place_); + worker->SetDataFeed(reader); + worker->SetFetchVarNames(fetch_var_names); + worker->BindingDataFeedMemory(); +#ifdef PADDLE_WITH_PSLIB + worker->SetPSlibPtr(_pslib_ptr); + worker->SetPullDenseThread(_pull_dense_thread); + worker->SetParamConfig(&_param_config); +#endif +} + +void PrepareReaders(std::vector>& readers, // NOLINT + const int thread_num, const DataFeedDesc& data_feed_desc, + const std::vector& filelist) { + readers.resize(thread_num); + for (size_t i = 0; i < readers.size(); ++i) { + readers[i] = DataFeedFactory::CreateDataFeed(data_feed_desc.name()); + readers[i]->Init(data_feed_desc); // set batch_size and queue_size here + } + readers[0]->SetFileList(filelist); +} + +#ifdef PADDLE_WITH_PSLIB +void AsyncExecutor::InitServer(const std::string& dist_desc, int index) { + _pslib_ptr = std::shared_ptr( + new paddle::distributed::PSlib()); + _pslib_ptr->init_server(dist_desc, index); + InitParamConfig(); +} + +void AsyncExecutor::InitWorker(const std::string& dist_desc, + const std::vector& host_sign_list, + int node_num, int index) { + _pslib_ptr = std::shared_ptr( + new paddle::distributed::PSlib()); + _pslib_ptr->init_worker( + dist_desc, const_cast(host_sign_list.data()), node_num, index); + + InitParamConfig(); +} + +uint64_t AsyncExecutor::StartServer() { return _pslib_ptr->run_server(); } + +void AsyncExecutor::StopServer() { _pslib_ptr->stop_server(); } + +void AsyncExecutor::GatherServers(const std::vector& host_sign_list, + int node_num) { + _pslib_ptr->gather_servers(const_cast(host_sign_list.data()), + node_num); +} + +void AsyncExecutor::InitParamConfig() { + for (int i = 0; i < _pslib_ptr->get_param() + ->server_param() + .downpour_server_param() + .downpour_table_param_size(); + ++i) { + if (_pslib_ptr->get_param() + ->server_param() + .downpour_server_param() + .downpour_table_param(i) + .table_class() + .find("SparseTable") != -1) { + _param_config.fea_dim = _pslib_ptr->get_param() + ->server_param() + .downpour_server_param() + .downpour_table_param(i) + .accessor() + .fea_dim(); + break; + } + } + _param_config.slot_dim = _param_config.fea_dim - 2; + _param_config.tmp_push_dense_wait_times = static_cast( + _pslib_ptr->get_param()->trainer_param().push_dense_per_batch()); + _param_config.tmp_push_sparse_wait_times = static_cast( + _pslib_ptr->get_param()->trainer_param().push_sparse_per_batch()); + + for (auto t = 0u; t < _pslib_ptr->get_param()->trainer_param().skip_op_size(); + ++t) { + _param_config.skip_op.push_back( + _pslib_ptr->get_param()->trainer_param().skip_op(t)); + } + + for (auto t = 0u; + t < _pslib_ptr->get_param()->trainer_param().sparse_table_size(); ++t) { + auto& table = _pslib_ptr->get_param()->trainer_param().sparse_table(t); + std::vector tmp_sparse_variable_name; + for (int i = 0u; i < table.slot_value_size(); ++i) { + tmp_sparse_variable_name.push_back(table.slot_value(i)); + _param_config.slot_alias_to_table[table.slot_key(i)] = table.table_id(); + } + std::vector tmp_sparse_gradient_variable_name; + for (auto i = 0u; i < table.slot_gradient_size(); ++i) { + tmp_sparse_gradient_variable_name.push_back(table.slot_gradient(i)); + } + _param_config.slot_input_vec[table.table_id()] = + std::move(tmp_sparse_variable_name); + _param_config.gradient_var[table.table_id()] = + std::move(tmp_sparse_gradient_variable_name); + _param_config.sparse_table_id.push_back(table.table_id()); + } + + for (auto t = 0u; + t < _pslib_ptr->get_param()->trainer_param().dense_table_size(); ++t) { + auto& table = _pslib_ptr->get_param()->trainer_param().dense_table(t); + std::vector tmp_dense_variable_name; + for (int i = 0u; i < table.dense_variable_name_size(); ++i) { + tmp_dense_variable_name.push_back(table.dense_variable_name(i)); + } + std::vector tmp_dense_gradient_variable_name; + for (auto i = 0u; i < table.dense_gradient_variable_name_size(); ++i) { + tmp_dense_gradient_variable_name.push_back( + table.dense_gradient_variable_name(i)); + } + _param_config.dense_variable_name[table.table_id()] = + std::move(tmp_dense_variable_name); + _param_config.dense_gradient_variable_name[table.table_id()] = + std::move(tmp_dense_gradient_variable_name); + _param_config.dense_table_id.push_back(table.table_id()); + _param_config.dense_table_size.push_back(table.fea_dim()); + } +} + +void AsyncExecutor::InitModel() { + for (auto table_id : _param_config.dense_table_id) { + std::vector regions; + for (auto& t : _param_config.dense_variable_name[table_id]) { + Variable* var = root_scope_->FindVar(t); + CHECK(var != nullptr) << "var[" << t << "] not found"; + LoDTensor* tensor = var->GetMutable(); + + float* g = tensor->data(); + CHECK(g != nullptr) << "var[" << t << "] value not initialized"; + + float init_range = 0.2; + int rown = tensor->dims()[0]; + init_range /= sqrt(rown); + + std::normal_distribution ndistr(0.0, 1.0); + for (auto i = 0u; i < tensor->numel(); ++i) { + g[i] = ndistr(local_random_engine()) * init_range; + } + + paddle::ps::Region reg(g, tensor->numel()); + regions.emplace_back(std::move(reg)); + } + + auto push_status = _pslib_ptr->_worker_ptr->push_dense_param( + regions.data(), regions.size(), table_id); + push_status.wait(); + auto status = push_status.get(); + if (status != 0) { + LOG(FATAL) << "push dense param failed, status[" << status << "]"; + exit(-1); + } + } +} + +void AsyncExecutor::SaveModel(const std::string& path) { + auto ret = _pslib_ptr->_worker_ptr->flush(); + ret.wait(); + ret = _pslib_ptr->_worker_ptr->save(path, 0); + ret.wait(); + int32_t feasign_cnt = ret.get(); + if (feasign_cnt == -1) { // (colourful-tree) TODO should be feasign_cnt < 0 + LOG(FATAL) << "save model failed"; + exit(-1); + } +} + +void AsyncExecutor::PrepareDenseThread(const std::string& mode) { + if (mode == "mpi") { + DensePullThreadParam param; + param.ps_client = _pslib_ptr->_worker_ptr; + param.threshold = 1; + param.training_thread_num = actual_thread_num; + param.root_scope = root_scope_; + param.dense_params = &_param_config.dense_variable_name; + + _pull_dense_thread = + std::shared_ptr(new DensePullThread(param)); + _pull_dense_thread->start(); + } +} +#endif + +void AsyncExecutor::RunFromFile(const ProgramDesc& main_program, + const std::string& data_feed_desc_str, + const std::vector& filelist, + const int thread_num, + const std::vector& fetch_var_names, + const std::string& mode, const bool debug) { + std::vector threads; + + auto& block = main_program.Block(0); + for (auto var_name : fetch_var_names) { + auto var_desc = block.FindVar(var_name); + auto shapes = var_desc->GetShape(); + PADDLE_ENFORCE(shapes[shapes.size() - 1] == 1, + "var %s: Fetched var has wrong shape, " + "only variables with the last dimension size 1 supported", + var_name); + } + + DataFeedDesc data_feed_desc; + google::protobuf::TextFormat::ParseFromString(data_feed_desc_str, + &data_feed_desc); + + actual_thread_num = thread_num; + int file_cnt = filelist.size(); + PADDLE_ENFORCE(file_cnt > 0, "File list cannot be empty"); + + if (actual_thread_num > file_cnt) { + VLOG(1) << "Thread num = " << thread_num << ", file num = " << file_cnt + << ". Changing thread_num = " << file_cnt; + actual_thread_num = file_cnt; + } + + /* + readerDesc: protobuf description for reader initlization + argument: class_name, batch_size, use_slot, queue_size, buffer_size, + padding_index + + reader: + 1) each thread has a reader, reader will read input data and + put it into input queue + 2) each reader has a Next() iterface, that can fetch an instance + from the input queue + */ + // todo: should be factory method for creating datafeed + std::vector> readers; + PrepareReaders(readers, actual_thread_num, data_feed_desc, filelist); +#ifdef PADDLE_WITH_PSLIB + PrepareDenseThread(mode); +#endif + std::vector> workers; + workers.resize(actual_thread_num); + for (auto& worker : workers) { +#ifdef PADDLE_WITH_PSLIB + if (mode == "mpi") { + worker.reset(new AsyncExecutorThreadWorker); + } else { + worker.reset(new ExecutorThreadWorker); + } +#else + worker.reset(new ExecutorThreadWorker); +#endif + } + + // prepare thread resource here + for (int thidx = 0; thidx < actual_thread_num; ++thidx) { + CreateThreads(workers[thidx].get(), main_program, readers[thidx], + fetch_var_names, root_scope_, thidx, debug); + } + + // start executing ops in multiple threads + for (int thidx = 0; thidx < actual_thread_num; ++thidx) { + threads.push_back( + std::thread(&ExecutorThreadWorker::TrainFiles, workers[thidx].get())); + } + + for (auto& th : threads) { + th.join(); + } +#ifdef PADDLE_WITH_PSLIB + if (mode == "mpi") { + _pull_dense_thread->stop(); + } +#endif + root_scope_->DropKids(); + + return; +} + +} // einit_modelnd namespace framework +} // end namespace paddle diff --git a/paddle/fluid/framework/async_executor.h b/paddle/fluid/framework/async_executor.h new file mode 100644 index 0000000000000..95c8472b2f3b6 --- /dev/null +++ b/paddle/fluid/framework/async_executor.h @@ -0,0 +1,108 @@ +/* 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 +#include +#include +#include // NOLINT +#include // local_random_engine +#include +#include +#include // NOLINT +#include +#include +#include "paddle/fluid/framework/data_feed.pb.h" +#include "paddle/fluid/framework/executor.h" +#include "paddle/fluid/framework/executor_thread_worker.h" +#include "paddle/fluid/framework/program_desc.h" +#include "paddle/fluid/framework/scope.h" + +namespace paddle { +namespace framework { + +inline double current_realtime() { +#if !defined(_WIN32) + struct timespec tp; + clock_gettime(CLOCK_REALTIME, &tp); + return tp.tv_sec + tp.tv_nsec * 1e-9; +#else + return 0.0; +#endif +} + +inline std::default_random_engine& local_random_engine() { + struct engine_wrapper_t { + std::default_random_engine engine; + engine_wrapper_t() { + static std::atomic x(0); + std::seed_seq sseq = {x++, x++, x++, + static_cast(current_realtime() * 1000)}; + engine.seed(sseq); + } + }; + thread_local engine_wrapper_t r; + return r.engine; +} + +class AsyncExecutor { + public: + AsyncExecutor(Scope* scope, const platform::Place& place); + virtual ~AsyncExecutor() {} + void RunFromFile(const ProgramDesc& main_program, + const std::string& data_feed_desc_str, + const std::vector& filelist, + const int thread_num, + const std::vector& fetch_names, + const std::string& mode, const bool debug = false); +#ifdef PADDLE_WITH_PSLIB + void InitServer(const std::string& dist_desc, int index); + void InitWorker(const std::string& dist_desc, + const std::vector& host_sign_list, int node_num, + int index); + uint64_t StartServer(); + void StopServer(); + void GatherServers(const std::vector& host_sign_list, int node_num); + void InitModel(); + void SaveModel(const std::string& path); + void InitParamConfig(); +#endif + + private: + void CreateThreads(ExecutorThreadWorker* worker, + const ProgramDesc& main_program, + const std::shared_ptr& reader, + const std::vector& fetch_var_names, + Scope* root_scope, const int thread_index, + const bool debug); +#ifdef PADDLE_WITH_PSLIB + void PrepareDenseThread(const std::string& mode); +#endif + + public: +#ifdef PADDLE_WITH_PSLIB + std::shared_ptr _pslib_ptr; + std::shared_ptr _pull_dense_thread; + AsyncWorkerParamConfig _param_config; +#endif + Scope* root_scope_; + platform::Place place_; + + private: + int actual_thread_num; +}; + +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/data_device_transform.cc b/paddle/fluid/framework/data_device_transform.cc index 57ff061fe5e61..fee6ba4004705 100644 --- a/paddle/fluid/framework/data_device_transform.cc +++ b/paddle/fluid/framework/data_device_transform.cc @@ -18,8 +18,8 @@ namespace framework { void TransDataDevice(const Tensor &in, const platform::Place &dst_place, Tensor *out) { - VLOG(30) << "DeviceTransform in, src_place " << in.place() - << " dst_place: " << dst_place; + VLOG(3) << "DeviceTransform in, src_place " << in.place() + << " dst_place: " << dst_place; PADDLE_ENFORCE_NE( in.place().which(), dst_place.which(), diff --git a/paddle/fluid/framework/data_device_transform_test.cu b/paddle/fluid/framework/data_device_transform_test.cu index 2d2323edc3a66..c9ec5e7a7b37b 100644 --- a/paddle/fluid/framework/data_device_transform_test.cu +++ b/paddle/fluid/framework/data_device_transform_test.cu @@ -49,10 +49,10 @@ class TestOpWithKernel : public OperatorWithKernel { OpKernelType GetExpectedKernelType( const ExecutionContext& ctx) const override { if (Attr("use_gpu")) { - VLOG(30) << "force use gpu kernel"; + VLOG(3) << "force use gpu kernel"; return OpKernelType(proto::VarType::FP32, platform::CUDAPlace(0)); } else { - VLOG(30) << "use default kernel"; + VLOG(3) << "use default kernel"; return OpKernelType(proto::VarType::FP32, ctx.Input("input")->place()); } @@ -148,7 +148,7 @@ TEST(Operator, CPUtoGPU) { // get output auto* output2 = scope.Var("OUT2"); gpu_op->Run(scope, cuda_place); - VLOG(30) << "after gpu_op run"; + VLOG(3) << "after gpu_op run"; // auto* output2_ptr = output2->Get().data(); paddle::platform::DeviceContextPool& pool = diff --git a/paddle/fluid/framework/data_feed.cc b/paddle/fluid/framework/data_feed.cc new file mode 100644 index 0000000000000..41155cfb7714b --- /dev/null +++ b/paddle/fluid/framework/data_feed.cc @@ -0,0 +1,374 @@ +/* Copyright (c) 2016 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 "google/protobuf/io/zero_copy_stream_impl.h" +#include "google/protobuf/message.h" +#include "google/protobuf/text_format.h" + +#include "gflags/gflags.h" +#include "paddle/fluid/framework/data_feed.h" +#include "paddle/fluid/framework/feed_fetch_method.h" +#include "paddle/fluid/framework/feed_fetch_type.h" + +namespace paddle { +namespace framework { + +std::vector DataFeed::filelist_; +size_t DataFeed::file_idx_; +std::mutex DataFeed::mutex_for_pick_file_; +bool DataFeed::finish_set_filelist_; + +void DataFeed::AddFeedVar(Variable* var, const std::string& name) { + CheckInit(); + for (size_t i = 0; i < use_slots_.size(); ++i) { + if (name == use_slots_[i]) { + feed_vec_[i] = var->GetMutable(); + } + } +} + +bool DataFeed::SetFileList(const std::vector& files) { + std::unique_lock lock(mutex_for_pick_file_); + CheckInit(); + if (finish_set_filelist_) { + VLOG(3) << "info: you have set the filelist."; + return false; + } + PADDLE_ENFORCE(files.size(), "You have set an empty filelist."); + filelist_.assign(files.begin(), files.end()); + file_idx_ = 0; + + finish_set_filelist_ = true; + return true; +} + +void DataFeed::SetBatchSize(int batch_size) { + PADDLE_ENFORCE(batch_size > 0, "Illegal batch size: %d.", batch_size); + default_batch_size_ = batch_size; +} + +bool DataFeed::PickOneFile(std::string* filename) { + std::unique_lock lock(mutex_for_pick_file_); + if (file_idx_ == filelist_.size()) { + return false; + } + *filename = filelist_[file_idx_++]; + LOG(ERROR) << "pick file:" << *filename; + return true; +} + +void DataFeed::CheckInit() { + PADDLE_ENFORCE(finish_init_, "Initialization did not succeed."); +} + +void DataFeed::CheckSetFileList() { + PADDLE_ENFORCE(finish_set_filelist_, "Set filelist did not succeed."); +} + +void DataFeed::CheckStart() { + PADDLE_ENFORCE(finish_start_, "Datafeed has not started running yet."); +} + +template +void PrivateQueueDataFeed::SetQueueSize(int queue_size) { + PADDLE_ENFORCE(queue_size > 0, "Illegal queue size: %d.", queue_size); + queue_size_ = queue_size; + queue_ = std::unique_ptr>( + new paddle::operators::reader::BlockingQueue(queue_size_)); +} + +template +bool PrivateQueueDataFeed::Start() { + CheckSetFileList(); + read_thread_ = std::thread(&PrivateQueueDataFeed::ReadThread, this); + read_thread_.detach(); + + finish_start_ = true; + return true; +} + +template +void PrivateQueueDataFeed::ReadThread() { + std::string filename; + while (PickOneFile(&filename)) { + file_.open(filename.c_str()); // is_text_feed + PADDLE_ENFORCE(file_.good(), "Open file<%s> fail.", filename.c_str()); + T instance; + while (ParseOneInstance(&instance)) { + queue_->Send(instance); + } + file_.close(); + } + queue_->Close(); +} + +template +int PrivateQueueDataFeed::Next() { + CheckStart(); + int index = 0; + T instance; + T ins_vec; + while (index < default_batch_size_) { + if (!queue_->Receive(&instance)) { + break; + } + AddInstanceToInsVec(&ins_vec, instance, index++); + } + batch_size_ = index; + if (batch_size_ != 0) { + PutToFeedVec(ins_vec); + } + return batch_size_; +} + +#ifdef _WIN32 +template class PrivateQueueDataFeed>; +#endif + +void MultiSlotDataFeed::Init( + const paddle::framework::DataFeedDesc& data_feed_desc) { + finish_init_ = false; + finish_set_filelist_ = false; + finish_start_ = false; + + PADDLE_ENFORCE(data_feed_desc.has_multi_slot_desc(), + "Multi_slot_desc has not been set."); + paddle::framework::MultiSlotDesc multi_slot_desc = + data_feed_desc.multi_slot_desc(); + SetBatchSize(data_feed_desc.batch_size()); + SetQueueSize(data_feed_desc.batch_size()); + size_t all_slot_num = multi_slot_desc.slots_size(); + all_slots_.resize(all_slot_num); + all_slots_type_.resize(all_slot_num); + use_slots_index_.resize(all_slot_num); + use_slots_.clear(); + use_slots_is_dense_.clear(); + for (size_t i = 0; i < all_slot_num; ++i) { + const auto& slot = multi_slot_desc.slots(i); + all_slots_[i] = slot.name(); + all_slots_type_[i] = slot.type(); + use_slots_index_[i] = slot.is_used() ? use_slots_.size() : -1; + if (slot.is_used()) { + use_slots_.push_back(all_slots_[i]); + use_slots_is_dense_.push_back(slot.is_dense()); + } + } + feed_vec_.resize(use_slots_.size()); + finish_init_ = true; +} + +bool MultiSlotDataFeed::CheckFile(const char* filename) { + CheckInit(); // get info of slots + std::ifstream fin(filename); + if (!fin.good()) { + VLOG(1) << "error: open file<" << filename << "> fail"; + return false; + } + std::string line; + int instance_cout = 0; + std::string all_slots_alias = ""; + for (const auto& alias : all_slots_) { + all_slots_alias += alias + " "; + } + std::string use_slots_alias = ""; + for (const auto& alias : use_slots_) { + use_slots_alias += alias + " "; + } + VLOG(3) << "total slots num: " << all_slots_.size(); + VLOG(3) << "total slots alias: " << all_slots_alias; + VLOG(3) << "used slots num: " << use_slots_.size(); + VLOG(3) << "used slots alias: " << use_slots_alias; + while (getline(fin, line)) { + ++instance_cout; + const char* str = line.c_str(); + char* endptr = const_cast(str); + int len = line.length(); + for (size_t i = 0; i < all_slots_.size(); ++i) { + int num = strtol(endptr, &endptr, 10); + if (num < 0) { + VLOG(0) << "error: the number of ids is a negative number: " << num; + VLOG(0) << "please check line<" << instance_cout << "> in file<" + << filename << ">"; + return false; + } else if (num == 0) { + VLOG(0) + << "error: the number of ids can not be zero, you need " + "padding it in data generator; or if there is something wrong" + " with the data, please check if the data contains unresolvable " + "characters."; + VLOG(0) << "please check line<" << instance_cout << "> in file<" + << filename << ">"; + return false; + } else if (errno == ERANGE || num > INT_MAX) { + VLOG(0) << "error: the number of ids greater than INT_MAX"; + VLOG(0) << "please check line<" << instance_cout << "> in file<" + << filename << ">"; + return false; + } + if (all_slots_type_[i] == "float") { + for (int i = 0; i < num; ++i) { + strtof(endptr, &endptr); + if (errno == ERANGE) { + VLOG(0) << "error: the value is out of the range of " + "representable values for float"; + VLOG(0) << "please check line<" << instance_cout << "> in file<" + << filename << ">"; + return false; + } + if (i + 1 != num && endptr - str == len) { + VLOG(0) << "error: there is a wrong with the number of ids."; + VLOG(0) << "please check line<" << instance_cout << "> in file<" + << filename << ">"; + return false; + } + } + } else if (all_slots_type_[i] == "uint64") { + for (int i = 0; i < num; ++i) { + strtoull(endptr, &endptr, 10); + if (errno == ERANGE) { + VLOG(0) << "error: the value is out of the range of " + "representable values for uint64_t"; + VLOG(0) << "please check line<" << instance_cout << "> in file<" + << filename << ">"; + return false; + } + if (i + 1 != num && endptr - str == len) { + VLOG(0) << "error: there is a wrong with the number of ids."; + VLOG(0) << "please check line<" << instance_cout << "> in file<" + << filename << ">"; + return false; + } + } + } else { + VLOG(0) << "error: this type<" << all_slots_type_[i] + << "> is not supported"; + return false; + } + } + // It may be added '\t' character to the end of the output of reduce + // task when processes data by Hadoop(when the output of the reduce + // task of Hadoop has only one field, it will add a '\t' at the end + // of the line by default, and you can use this option to avoid it: + // `-D mapred.textoutputformat.ignoreseparator=true`), which does + // not affect the correctness of the data. Therefore, it should be + // judged that the data is not normal when the end of each line of + // data contains characters which are not spaces. + while (endptr - str != len) { + if (!isspace(*(endptr++))) { + VLOG(0) + << "error: there is some extra characters at the end of the line."; + VLOG(0) << "please check line<" << instance_cout << "> in file<" + << filename << ">"; + return false; + } + } + } + VLOG(3) << "instances cout: " << instance_cout; + VLOG(3) << "The file format is correct"; + return true; +} + +bool MultiSlotDataFeed::ParseOneInstance(std::vector* instance) { + std::string line; + if (getline(file_, line)) { + int use_slots_num = use_slots_.size(); + instance->resize(use_slots_num); + // parse line + const char* str = line.c_str(); + char* endptr = const_cast(str); + int pos = 0; + for (size_t i = 0; i < use_slots_index_.size(); ++i) { + int idx = use_slots_index_[i]; + int num = strtol(&str[pos], &endptr, 10); + PADDLE_ENFORCE( + num, + "The number of ids can not be zero, you need padding " + "it in data generator; or if there is something wrong with " + "the data, please check if the data contains unresolvable " + "characters.\nplease check this error line: %s", + str); + + if (idx != -1) { + (*instance)[idx].Init(all_slots_type_[i]); + if ((*instance)[idx].GetType()[0] == 'f') { // float + for (int j = 0; j < num; ++j) { + float feasign = strtof(endptr, &endptr); + (*instance)[idx].AddValue(feasign); + } + } else if ((*instance)[idx].GetType()[0] == 'u') { // uint64 + for (int j = 0; j < num; ++j) { + uint64_t feasign = (uint64_t)strtoull(endptr, &endptr, 10); + (*instance)[idx].AddValue(feasign); + } + } + pos = endptr - str; + } else { + for (int j = 0; j <= num; ++j) { + pos = line.find_first_of(' ', pos + 1); + } + } + } + } else { + return false; + } + return true; +} + +void MultiSlotDataFeed::AddInstanceToInsVec( + std::vector* ins_vec, + const std::vector& instance, int index) { + if (index == 0) { + ins_vec->resize(instance.size()); + for (size_t i = 0; i < instance.size(); ++i) { + (*ins_vec)[i].Init(instance[i].GetType()); + (*ins_vec)[i].InitOffset(); + } + } + + for (size_t i = 0; i < instance.size(); ++i) { + (*ins_vec)[i].AddIns(instance[i]); + } +} + +void MultiSlotDataFeed::PutToFeedVec( + const std::vector& ins_vec) { + for (size_t i = 0; i < use_slots_.size(); ++i) { + const auto& type = ins_vec[i].GetType(); + const auto& offset = ins_vec[i].GetOffset(); + int total_instance = static_cast(offset.back()); + + if (type[0] == 'f') { // float + const auto& feasign = ins_vec[i].GetFloatData(); + float* tensor_ptr = feed_vec_[i]->mutable_data( + {total_instance, 1}, platform::CPUPlace()); + memcpy(tensor_ptr, &feasign[0], total_instance * sizeof(float)); + } else if (type[0] == 'u') { // uint64 + // no uint64_t type in paddlepaddle + const auto& feasign = ins_vec[i].GetUint64Data(); + int64_t* tensor_ptr = feed_vec_[i]->mutable_data( + {total_instance, 1}, platform::CPUPlace()); + memcpy(tensor_ptr, &feasign[0], total_instance * sizeof(int64_t)); + } + + LoD data_lod{offset}; + feed_vec_[i]->set_lod(data_lod); + if (use_slots_is_dense_[i]) { + int dim = total_instance / batch_size_; + feed_vec_[i]->Resize({batch_size_, dim}); + } + } +} + +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/data_feed.h b/paddle/fluid/framework/data_feed.h new file mode 100644 index 0000000000000..7cc6919703680 --- /dev/null +++ b/paddle/fluid/framework/data_feed.h @@ -0,0 +1,240 @@ +/* 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 +#include +#include // NOLINT +#include +#include // NOLINT +#include + +#include "paddle/fluid/framework/data_feed.pb.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/reader.h" +#include "paddle/fluid/framework/variable.h" +#include "paddle/fluid/operators/reader/blocking_queue.h" + +namespace paddle { +namespace framework { + +// DataFeed is the base virtual class for all ohther DataFeeds. +// It is used to read files and parse the data for subsequent trainer. +// Example: +// DataFeed* reader = +// paddle::framework::DataFeedFactory::CreateDataFeed(data_feed_name); +// reader->Init(data_feed_desc); // data_feed_desc is a protobuf object +// reader->SetFileList(filelist); +// const std::vector & use_slot_alias = +// reader->GetUseSlotAlias(); +// for (auto name: use_slot_alias){ // for binding memory +// reader->AddFeedVar(scope->Var(name), name); +// } +// reader->Start(); +// while (reader->Next()) { +// // trainer do something +// } +class DataFeed { + public: + DataFeed() {} + virtual ~DataFeed() {} + virtual void Init(const paddle::framework::DataFeedDesc& data_feed_desc) = 0; + virtual bool CheckFile(const char* filename) { + PADDLE_THROW("This function(CheckFile) is not implemented."); + } + // Set filelist for DataFeed. + // Pay attention that it must init all readers before call this function. + // Otherwise, Init() function will init finish_set_filelist_ flag. + virtual bool SetFileList(const std::vector& files); + virtual bool Start() = 0; + // The trainer calls the Next() function, and the DataFeed will load a new + // batch to the feed_vec. The return value of this function is the batch + // size of the current batch. + virtual int Next() = 0; + // Get all slots' alias which defined in protofile + virtual const std::vector& GetAllSlotAlias() { + return all_slots_; + } + // Get used slots' alias which defined in protofile + virtual const std::vector& GetUseSlotAlias() { + return use_slots_; + } + // This function is used for binding feed_vec memory + virtual void AddFeedVar(Variable* var, const std::string& name); + + protected: + // The following three functions are used to check if it is executed in this + // order: + // Init() -> SetFileList() -> Start() -> Next() + virtual void CheckInit(); + virtual void CheckSetFileList(); + virtual void CheckStart(); + virtual void SetBatchSize( + int batch); // batch size will be set in Init() function + // This function is used to pick one file from the global filelist(thread + // safe). + virtual bool PickOneFile(std::string* filename); + + static std::vector filelist_; + static size_t file_idx_; + static std::mutex mutex_for_pick_file_; + + // the alias of used slots, and its order is determined by + // data_feed_desc(proto object) + std::vector use_slots_; + std::vector use_slots_is_dense_; + + // the alias of all slots, and its order is determined by data_feed_desc(proto + // object) + std::vector all_slots_; + std::vector all_slots_type_; + std::vector + use_slots_index_; // -1: not used; >=0: the index of use_slots_ + + // The data read by DataFeed will be stored here + std::vector feed_vec_; + + // the batch size defined by user + int default_batch_size_; + // current batch size + int batch_size_; + + bool finish_init_; + static bool finish_set_filelist_; + bool finish_start_; +}; + +// PrivateQueueDataFeed is the base virtual class for ohther DataFeeds. +// It use a read-thread to read file and parse data to a private-queue +// (thread level), and get data from this queue when trainer call Next(). +template +class PrivateQueueDataFeed : public DataFeed { + public: + PrivateQueueDataFeed() {} + virtual ~PrivateQueueDataFeed() {} + virtual void Init(const paddle::framework::DataFeedDesc& data_feed_desc) = 0; + virtual bool Start(); + virtual int Next(); + + protected: + // The thread implementation function for reading file and parse. + virtual void ReadThread(); + // This function is used to set private-queue size, and the most + // efficient when the queue size is close to the batch size. + virtual void SetQueueSize(int queue_size); + // The reading and parsing method called in the ReadThread. + virtual bool ParseOneInstance(T* instance) = 0; + // This function is used to put instance to vec_ins + virtual void AddInstanceToInsVec(T* vec_ins, const T& instance, + int index) = 0; + // This function is used to put ins_vec to feed_vec + virtual void PutToFeedVec(const T& ins_vec) = 0; + + // The thread for read files + std::thread read_thread_; + // using ifstream one line and one line parse is faster + // than using fread one buffer and one buffer parse. + // for a 601M real data: + // ifstream one line and one line parse: 6034 ms + // fread one buffer and one buffer parse: 7097 ms + std::ifstream file_; + size_t queue_size_; + // The queue for store parsed data + std::unique_ptr> queue_; +}; + +// This class define the data type of instance(ins_vec) in MultiSlotDataFeed +class MultiSlotType { + public: + MultiSlotType() {} + ~MultiSlotType() {} + void Init(const std::string& type) { + CheckType(type); + if (type_[0] == 'f') { + float_feasign_.clear(); + } else if (type_[0] == 'u') { + uint64_feasign_.clear(); + } + type_ = type; + } + void InitOffset() { + offset_.resize(1); + // LoDTensor' lod is counted from 0, the size of lod + // is one size larger than the size of data. + offset_[0] = 0; + } + const std::vector& GetOffset() const { return offset_; } + void AddValue(const float v) { + CheckFloat(); + float_feasign_.push_back(v); + } + void AddValue(const uint64_t v) { + CheckUint64(); + uint64_feasign_.push_back(v); + } + void AddIns(const MultiSlotType& ins) { + if (ins.GetType()[0] == 'f') { // float + CheckFloat(); + auto& vec = ins.GetFloatData(); + offset_.push_back(offset_.back() + vec.size()); + float_feasign_.insert(float_feasign_.end(), vec.begin(), vec.end()); + } else if (ins.GetType()[0] == 'u') { // uint64 + CheckUint64(); + auto& vec = ins.GetUint64Data(); + offset_.push_back(offset_.back() + vec.size()); + uint64_feasign_.insert(uint64_feasign_.end(), vec.begin(), vec.end()); + } + } + const std::vector& GetFloatData() const { return float_feasign_; } + const std::vector& GetUint64Data() const { return uint64_feasign_; } + const std::string& GetType() const { return type_; } + + private: + void CheckType(const std::string& type) const { + PADDLE_ENFORCE((type == "uint64") || (type == "float"), + "There is no this type<%s>.", type); + } + void CheckFloat() const { + PADDLE_ENFORCE(type_[0] == 'f', "Add %s value to float slot.", type_); + } + void CheckUint64() const { + PADDLE_ENFORCE(type_[0] == 'u', "Add %s value to uint64 slot.", type_); + } + std::vector float_feasign_; + std::vector uint64_feasign_; + std::string type_; + std::vector offset_; +}; + +// This DataFeed is used to feed multi-slot type data. +// The format of multi-slot type data: +// [n feasign_0 feasign_1 ... feasign_n]* +class MultiSlotDataFeed + : public PrivateQueueDataFeed> { + public: + MultiSlotDataFeed() {} + virtual ~MultiSlotDataFeed() {} + virtual void Init(const paddle::framework::DataFeedDesc& data_feed_desc); + virtual bool CheckFile(const char* filename); + + protected: + virtual void AddInstanceToInsVec(std::vector* vec_ins, + const std::vector& instance, + int index); + virtual bool ParseOneInstance(std::vector* instance); + virtual void PutToFeedVec(const std::vector& ins_vec); +}; +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/data_feed.proto b/paddle/fluid/framework/data_feed.proto new file mode 100644 index 0000000000000..489fec08d86cc --- /dev/null +++ b/paddle/fluid/framework/data_feed.proto @@ -0,0 +1,30 @@ +/* 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. */ +syntax = "proto2"; +package paddle.framework; + +message Slot { + required string name = 1; + required string type = 2; + optional bool is_dense = 3 [ default = false ]; + optional bool is_used = 4 [ default = false ]; +} + +message MultiSlotDesc { repeated Slot slots = 1; } + +message DataFeedDesc { + optional string name = 1; + optional int32 batch_size = 2 [ default = 32 ]; + optional MultiSlotDesc multi_slot_desc = 3; +} diff --git a/paddle/fluid/framework/data_feed_factory.cc b/paddle/fluid/framework/data_feed_factory.cc new file mode 100644 index 0000000000000..72148b9f7d343 --- /dev/null +++ b/paddle/fluid/framework/data_feed_factory.cc @@ -0,0 +1,64 @@ +/* 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/data_feed_factory.h" +#include +#include +#include + +#include "paddle/fluid/framework/data_feed.h" + +namespace paddle { +namespace framework { +typedef std::shared_ptr (*Createdata_feedFunction)(); +typedef std::unordered_map data_feedMap; +data_feedMap g_data_feed_map; + +#define REGISTER_DATAFEED_CLASS(data_feed_class) \ + namespace { \ + std::shared_ptr Creator_##data_feed_class() { \ + return std::shared_ptr(new data_feed_class); \ + } \ + class __Registerer_##data_feed_class { \ + public: \ + __Registerer_##data_feed_class() { \ + g_data_feed_map[#data_feed_class] = &Creator_##data_feed_class; \ + } \ + }; \ + __Registerer_##data_feed_class g_registerer_##data_feed_class; \ + } // namespace + +std::string DataFeedFactory::DataFeedTypeList() { + std::string data_feed_types; + for (auto iter = g_data_feed_map.begin(); iter != g_data_feed_map.end(); + ++iter) { + if (iter != g_data_feed_map.begin()) { + data_feed_types += ", "; + } + data_feed_types += iter->first; + } + return data_feed_types; +} + +std::shared_ptr DataFeedFactory::CreateDataFeed( + std::string data_feed_class) { + if (g_data_feed_map.count(data_feed_class) < 1) { + exit(-1); + } + return g_data_feed_map[data_feed_class](); +} + +REGISTER_DATAFEED_CLASS(MultiSlotDataFeed); +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/operators/math/jit_kernel.cc b/paddle/fluid/framework/data_feed_factory.h similarity index 52% rename from paddle/fluid/operators/math/jit_kernel.cc rename to paddle/fluid/framework/data_feed_factory.h index 68b708b345334..13678edb0b8d0 100644 --- a/paddle/fluid/operators/math/jit_kernel.cc +++ b/paddle/fluid/framework/data_feed_factory.h @@ -4,7 +4,7 @@ Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at -http://www.apache.org/licenses/LICENSE-2.0 + http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, @@ -12,30 +12,18 @@ 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/operators/math/jit_kernel.h" -#include +#pragma once + +#include #include +#include "paddle/fluid/framework/data_feed.h" namespace paddle { -namespace operators { -namespace math { -namespace jitkernel { - -namespace jit = platform::jit; - -KernelPool& KernelPool::Instance() { - static thread_local KernelPool g_jit_kernels; - return g_jit_kernels; -} - -std::shared_ptr KernelPool::Get(const std::string& key) const { - if (kers_.find(key) == kers_.end()) { - return nullptr; - } - return kers_.at(key); -} - -} // namespace jitkernel -} // namespace math -} // namespace operators +namespace framework { +class DataFeedFactory { + public: + static std::string DataFeedTypeList(); + static std::shared_ptr CreateDataFeed(std::string data_feed_class); +}; +} // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/data_feed_test.cc b/paddle/fluid/framework/data_feed_test.cc new file mode 100644 index 0000000000000..b3e9698715923 --- /dev/null +++ b/paddle/fluid/framework/data_feed_test.cc @@ -0,0 +1,330 @@ +// 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/data_feed.h" +#include +#include // NOLINT +#include +#include +#include +#include // NOLINT +#include +#include // NOLINT +#include +#include +#include "google/protobuf/io/zero_copy_stream_impl.h" +#include "google/protobuf/text_format.h" +#include "gtest/gtest.h" +#include "paddle/fluid/framework/data_feed_factory.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/scope.h" + +paddle::framework::DataFeedDesc load_datafeed_param_from_file( + const char* filename) { + paddle::framework::DataFeedDesc data_feed_desc; + int file_descriptor = open(filename, O_RDONLY); + PADDLE_ENFORCE(file_descriptor != -1, "Can not open %s.", filename); + google::protobuf::io::FileInputStream fileInput(file_descriptor); + google::protobuf::TextFormat::Parse(&fileInput, &data_feed_desc); + close(file_descriptor); + return data_feed_desc; +} + +const std::vector load_filelist_from_file(const char* filename) { + std::vector filelist; + std::ifstream fin(filename); + PADDLE_ENFORCE(fin.good(), "Can not open %s.", filename); + std::string line; + while (getline(fin, line)) { + filelist.push_back(line); + } + fin.close(); + return filelist; +} + +void GenerateFileForTest(const char* protofile, const char* filelist) { + std::ofstream w_protofile(protofile); + w_protofile << "name: \"MultiSlotDataFeed\"\n" + "batch_size: 2\n" + "multi_slot_desc {\n" + " slots {\n" + " name: \"uint64_sparse_slot\"\n" + " type: \"uint64\"\n" + " is_dense: false\n" + " is_used: true\n" + " }\n" + " slots {\n" + " name: \"float_sparse_slot\"\n" + " type: \"float\"\n" + " is_dense: false\n" + " is_used: true\n" + " }\n" + " slots {\n" + " name: \"uint64_dense_slot\"\n" + " type: \"uint64\"\n" + " is_dense: true\n" + " is_used: true\n" + " }\n" + " slots {\n" + " name: \"float_dense_slot\"\n" + " type: \"float\"\n" + " is_dense: true\n" + " is_used: true\n" + " }\n" + " slots {\n" + " name: \"not_used_slot\"\n" + " type: \"uint64\"\n" + " is_dense: false\n" + " is_used: false\n" + " }\n" + "}"; + w_protofile.close(); + std::ofstream w_filelist(filelist); + int total_file = 4; + for (int i = 0; i < total_file; ++i) { + std::string filename = "TestMultiSlotDataFeed.data." + std::to_string(i); + w_filelist << filename; + if (i + 1 != total_file) { + w_filelist << std::endl; + } + std::ofstream w_datafile(filename.c_str()); + w_datafile << "3 3978 620 82 1 1926.08 1 1926 1 6.02 1 1996\n" + "2 1300 2983353 1 985.211 1 8 1 0.618 1 12\n" + "1 19260827 2 3.14 2.718 1 27 1 2.236 1 28\n"; + w_datafile.close(); + } + w_filelist.close(); +} + +class MultiTypeSet { + public: + MultiTypeSet() { + uint64_set_.clear(); + float_set_.clear(); + } + ~MultiTypeSet() {} + void AddValue(uint64_t v) { uint64_set_.insert(v); } + void AddValue(float v) { float_set_.insert(v); } + const std::set& GetUint64Set() const { return uint64_set_; } + const std::set& GetFloatSet() const { return float_set_; } + + private: + std::set uint64_set_; + std::set float_set_; +}; + +void GetElemSetFromReader(std::vector* reader_elem_set, + const paddle::framework::DataFeedDesc& data_feed_desc, + const std::vector& filelist, + const int thread_num) { + int used_slot_num = 0; + for (auto i = 0; i < data_feed_desc.multi_slot_desc().slots_size(); ++i) { + if (data_feed_desc.multi_slot_desc().slots(i).is_used()) { + ++used_slot_num; + } + } + reader_elem_set->resize(used_slot_num); + std::vector threads; + std::vector> readers; + readers.resize(thread_num); + for (int i = 0; i < thread_num; ++i) { + readers[i] = paddle::framework::DataFeedFactory::CreateDataFeed( + data_feed_desc.name()); + readers[i]->Init(data_feed_desc); + } + readers[0]->SetFileList(filelist); + std::mutex mu; + for (int idx = 0; idx < thread_num; ++idx) { + threads.emplace_back(std::thread([&, idx] { + std::unique_ptr scope( + new paddle::framework::Scope()); + const auto& multi_slot_desc = data_feed_desc.multi_slot_desc(); + std::map + lodtensor_targets; + for (int i = 0; i < multi_slot_desc.slots_size(); ++i) { + const auto& slot = multi_slot_desc.slots(i); + if (slot.is_used()) { + const auto& name = slot.name(); + readers[idx]->AddFeedVar(scope->Var(name), name); + lodtensor_targets[name] = + &scope->FindVar(name)->Get(); + } + } + readers[idx]->Start(); + while (readers[idx]->Next()) { + int index = 0; + for (int k = 0; k < multi_slot_desc.slots_size(); ++k) { + const auto& slot = multi_slot_desc.slots(k); + if (!slot.is_used()) { + continue; + } + const paddle::framework::LoDTensor* tens = + lodtensor_targets[slot.name()]; + if (slot.is_dense()) { // dense branch + if (slot.type() == "uint64") { + const int64_t* data = tens->data(); + int batch_size = tens->dims()[0]; + int dim = tens->dims()[1]; + for (int i = 0; i < batch_size; ++i) { + for (int j = 0; j < dim; ++j) { + std::lock_guard lock(mu); + (*reader_elem_set)[index].AddValue( + (uint64_t)data[i * dim + j]); + } + } + } else if (slot.type() == "float") { + const float* data = tens->data(); + int batch_size = tens->dims()[0]; + int dim = tens->dims()[1]; + for (int i = 0; i < batch_size; ++i) { + for (int j = 0; j < dim; ++j) { + std::lock_guard lock(mu); + (*reader_elem_set)[index].AddValue(data[i * dim + j]); + } + } + } else { + PADDLE_THROW("Error type in proto file."); + } + } else { // sparse branch + if (slot.type() == "uint64") { + const int64_t* data = tens->data(); + for (size_t i = 0; i < tens->NumElements(); ++i) { + std::pair element = tens->lod_element(0, i); + for (size_t j = element.first; j < element.second; ++j) { + std::lock_guard lock(mu); + (*reader_elem_set)[index].AddValue((uint64_t)data[j]); + } + } + } else if (slot.type() == "float") { + const float* data = tens->data(); + for (size_t i = 0; i < tens->NumElements(); ++i) { + std::pair element = tens->lod_element(0, i); + for (size_t j = element.first; j < element.second; ++j) { + std::lock_guard lock(mu); + (*reader_elem_set)[index].AddValue(data[j]); + } + } + } else { + PADDLE_THROW("Error type in proto file."); + } + } // end sparse branch + ++index; + } // end slots loop + } // end while Next() + })); // end anonymous function + } + for (auto& th : threads) { + th.join(); + } +} + +void CheckIsUnorderedSame(const std::vector& s1, + const std::vector& s2) { + EXPECT_EQ(s1.size(), s2.size()); + for (size_t i = 0; i < s1.size(); ++i) { + // check for uint64 + const std::set& uint64_s1 = s1[i].GetUint64Set(); + const std::set& uint64_s2 = s2[i].GetUint64Set(); + EXPECT_EQ(uint64_s1.size(), uint64_s2.size()); + auto uint64_it1 = uint64_s1.begin(); + auto uint64_it2 = uint64_s2.begin(); + while (uint64_it1 != uint64_s1.end()) { + EXPECT_EQ(*uint64_it1, *uint64_it2); + ++uint64_it1; + ++uint64_it2; + } + // check for float + const std::set& float_s1 = s1[i].GetFloatSet(); + const std::set& float_s2 = s2[i].GetFloatSet(); + EXPECT_EQ(float_s1.size(), float_s2.size()); + auto float_it1 = float_s1.begin(); + auto float_it2 = float_s2.begin(); + while (float_it1 != float_s1.end()) { + EXPECT_EQ(*float_it1, *float_it2); + ++float_it1; + ++float_it2; + } + } +} + +void GetElemSetFromFile(std::vector* file_elem_set, + const paddle::framework::DataFeedDesc& data_feed_desc, + const std::vector& filelist) { + int used_slot_num = 0; + for (auto i = 0; i < data_feed_desc.multi_slot_desc().slots_size(); ++i) { + if (data_feed_desc.multi_slot_desc().slots(i).is_used()) { + ++used_slot_num; + } + } + file_elem_set->resize(used_slot_num); + for (const auto& file : filelist) { + std::ifstream fin(file.c_str()); + PADDLE_ENFORCE(fin.good(), "Can not open %s.", file.c_str()); + while (1) { + bool end_flag = false; + int index = 0; + for (auto i = 0; i < data_feed_desc.multi_slot_desc().slots_size(); ++i) { + int num; + if (fin >> num) { + auto slot = data_feed_desc.multi_slot_desc().slots(i); + auto type = slot.type(); + if (type == "uint64") { + while (num--) { + uint64_t feasign; + fin >> feasign; + if (slot.is_used()) { + (*file_elem_set)[index].AddValue(feasign); + } + } + } else if (type == "float") { + while (num--) { + float feasign; + fin >> feasign; + if (slot.is_used()) { + (*file_elem_set)[index].AddValue(feasign); + } + } + } else { + PADDLE_THROW("Error type in proto file."); + } + if (slot.is_used()) { + ++index; + } + } else { + end_flag = true; + break; + } + } + if (end_flag) { + break; + } + } + fin.close(); + } +} + +TEST(DataFeed, MultiSlotUnitTest) { + const char* protofile = "data_feed_desc.prototxt"; + const char* filelist_name = "filelist.txt"; + GenerateFileForTest(protofile, filelist_name); + const std::vector filelist = + load_filelist_from_file(filelist_name); + paddle::framework::DataFeedDesc data_feed_desc = + load_datafeed_param_from_file(protofile); + std::vector reader_elem_set; + std::vector file_elem_set; + GetElemSetFromReader(&reader_elem_set, data_feed_desc, filelist, 4); + GetElemSetFromFile(&file_elem_set, data_feed_desc, filelist); + CheckIsUnorderedSame(reader_elem_set, file_elem_set); +} diff --git a/paddle/fluid/framework/data_layout_transform.cc b/paddle/fluid/framework/data_layout_transform.cc index c9e3a8ac1d1e5..72c50518af08b 100644 --- a/paddle/fluid/framework/data_layout_transform.cc +++ b/paddle/fluid/framework/data_layout_transform.cc @@ -85,7 +85,7 @@ void TransDataLayout(const OpKernelType& kernel_type_for_var, out->mutable_data(expected_kernel_type.place_, in.type()); framework::VisitDataType( - framework::ToDataType(in.type()), + in.type(), CastDataLayout(pool.Get(expected_kernel_type.place_), axis, in, out)); out->set_layout(expected_kernel_type.data_layout_); @@ -101,7 +101,7 @@ void* GetDataFromTensor(const Tensor& tensor, mkldnn::memory::data_type type) { case mkldnn::memory::data_type::f32: return platform::to_void_cast(tensor.data()); case mkldnn::memory::data_type::s8: - return platform::to_void_cast(tensor.data()); + return platform::to_void_cast(tensor.data()); case mkldnn::memory::data_type::u8: return platform::to_void_cast(tensor.data()); case mkldnn::memory::data_type::s16: @@ -144,26 +144,29 @@ void TransDataLayoutFromMKLDNN(const OpKernelType& kernel_type_for_var, memory::data_type in_type = ToMKLDNNDataType(in.type()); PADDLE_ENFORCE(in_type != memory::data_type::data_undef, - "Input tensor type is not supported: ", in.type().name()); + "Input tensor type is not supported: %s", in.type()); memory::data_type out_type = in_type; auto in_format = platform::MKLDNNFormatForSize(in_tz.size(), in.format()); auto out_format = platform::MKLDNNFormatForSize(in_tz.size(), ToMKLDNNFormat(out_layout)); - void* in_data = GetDataFromTensor(in, in_type); - // output tensor has the same dims as input. Reorder don't change dims out->Resize(in.dims()); - auto out_data = out->mutable_data(expected_kernel_type.place_, in.type()); - - auto in_memory = memory({{{in_tz}, in_type, in_format}, cpu_engine}, in_data); - auto out_memory = - memory({{{out_tz}, out_type, out_format}, cpu_engine}, out_data); + if (in_format != out_format) { + void* in_data = GetDataFromTensor(in, in_type); + auto out_data = out->mutable_data(expected_kernel_type.place_, in.type()); - platform::Reorder(in_memory, out_memory); + auto in_memory = + memory({{{in_tz}, in_type, in_format}, cpu_engine}, in_data); + auto out_memory = + memory({{{out_tz}, out_type, out_format}, cpu_engine}, out_data); + platform::Reorder(in_memory, out_memory); + } else { + out->ShareDataWith(in); + } out->set_layout(out_layout); // reset format since the out tensor will be feed to non-MKLDNN OPkernel out->set_format(memory::format::format_undef); diff --git a/paddle/fluid/framework/data_layout_transform.h b/paddle/fluid/framework/data_layout_transform.h index 90bb206ec6b69..2479de4fd4680 100644 --- a/paddle/fluid/framework/data_layout_transform.h +++ b/paddle/fluid/framework/data_layout_transform.h @@ -50,14 +50,14 @@ inline DataLayout ToPaddleLayout(const MKLDNNFormat& format) { } } -inline MKLDNNDataType ToMKLDNNDataType(const std::type_index type) { - static const std::map dict{ - {std::type_index(typeid(float)), MKLDNNDataType::f32}, // NOLINT - {std::type_index(typeid(char)), MKLDNNDataType::s8}, // NOLINT - {std::type_index(typeid(unsigned char)), MKLDNNDataType::u8}, - {std::type_index(typeid(int16_t)), MKLDNNDataType::s16}, - {std::type_index(typeid(int32_t)), MKLDNNDataType::s32}}; - auto iter = dict.find(type); +inline MKLDNNDataType ToMKLDNNDataType(proto::VarType::Type type) { + static std::unordered_map dict{ + {DataTypeTrait::DataType, MKLDNNDataType::f32}, + {DataTypeTrait::DataType, MKLDNNDataType::s8}, + {DataTypeTrait::DataType, MKLDNNDataType::u8}, + {DataTypeTrait::DataType, MKLDNNDataType::s16}, + {DataTypeTrait::DataType, MKLDNNDataType::s32}}; + auto iter = dict.find(static_cast(type)); if (iter != dict.end()) return iter->second; return MKLDNNDataType::data_undef; } diff --git a/paddle/fluid/framework/data_type.cc b/paddle/fluid/framework/data_type.cc index 28f3da88fa180..a0248cf3c7569 100644 --- a/paddle/fluid/framework/data_type.cc +++ b/paddle/fluid/framework/data_type.cc @@ -26,7 +26,7 @@ struct DataTypeMap { std::unordered_map cpp_to_proto_; std::unordered_map proto_to_cpp_; std::unordered_map proto_to_str_; - std::unordered_map cpp_to_size_; + std::unordered_map proto_to_size_; }; static DataTypeMap* InitDataTypeMap(); @@ -45,7 +45,7 @@ static inline void RegisterType(DataTypeMap* map, map->proto_to_cpp_.emplace(static_cast(proto_type), typeid(T)); map->cpp_to_proto_.emplace(typeid(T), proto_type); map->proto_to_str_.emplace(static_cast(proto_type), name); - map->cpp_to_size_.emplace(typeid(T), sizeof(T)); + map->proto_to_size_.emplace(static_cast(proto_type), sizeof(T)); } static DataTypeMap* InitDataTypeMap() { @@ -54,17 +54,7 @@ static DataTypeMap* InitDataTypeMap() { #define RegType(cc_type, proto_type) \ RegisterType(retv, proto_type, #cc_type) - // NOTE: Add your customize type here. - RegType(float16, proto::VarType::FP16); - RegType(float, proto::VarType::FP32); - RegType(double, proto::VarType::FP64); - RegType(int, proto::VarType::INT32); - RegType(int64_t, proto::VarType::INT64); - RegType(bool, proto::VarType::BOOL); - RegType(size_t, proto::VarType::SIZE_T); - RegType(int16_t, proto::VarType::INT16); - RegType(uint8_t, proto::VarType::UINT8); - RegType(int8_t, proto::VarType::INT8); + _ForEachDataType_(RegType); #undef RegType return retv; @@ -96,12 +86,12 @@ std::string DataTypeToString(const proto::VarType::Type type) { static_cast(type)); } -size_t SizeOfType(std::type_index type) { - auto it = gDataTypeMap().cpp_to_size_.find(type); - if (it != gDataTypeMap().cpp_to_size_.end()) { +size_t SizeOfType(proto::VarType::Type type) { + auto it = gDataTypeMap().proto_to_size_.find(static_cast(type)); + if (it != gDataTypeMap().proto_to_size_.end()) { return it->second; } - PADDLE_THROW("Not support %s as tensor type", type.name()); + PADDLE_THROW("Not support %s as tensor type", DataTypeToString(type)); } } // namespace framework diff --git a/paddle/fluid/framework/data_type.h b/paddle/fluid/framework/data_type.h index d5be43b33edab..76df78ea5e17c 100644 --- a/paddle/fluid/framework/data_type.h +++ b/paddle/fluid/framework/data_type.h @@ -22,46 +22,59 @@ limitations under the License. */ namespace paddle { namespace framework { +template +struct DataTypeTrait {}; + +// Stub handle for void +template <> +struct DataTypeTrait { + constexpr static auto DataType = proto::VarType::RAW; +}; + +#define _ForEachDataTypeHelper_(callback, cpp_type, proto_type) \ + callback(cpp_type, ::paddle::framework::proto::VarType::proto_type); + +#define _ForEachDataType_(callback) \ + _ForEachDataTypeHelper_(callback, float, FP32); \ + _ForEachDataTypeHelper_(callback, ::paddle::platform::float16, FP16); \ + _ForEachDataTypeHelper_(callback, double, FP64); \ + _ForEachDataTypeHelper_(callback, int, INT32); \ + _ForEachDataTypeHelper_(callback, int64_t, INT64); \ + _ForEachDataTypeHelper_(callback, bool, BOOL); \ + _ForEachDataTypeHelper_(callback, uint8_t, UINT8); \ + _ForEachDataTypeHelper_(callback, int16_t, INT16); \ + _ForEachDataTypeHelper_(callback, int8_t, INT8) + +#define DefineDataTypeTrait(cpp_type, proto_type) \ + template <> \ + struct DataTypeTrait { \ + constexpr static auto DataType = proto_type; \ + } + +_ForEachDataType_(DefineDataTypeTrait); + +#undef DefineDataTypeTrait + extern proto::VarType::Type ToDataType(std::type_index type); extern std::type_index ToTypeIndex(proto::VarType::Type type); template inline void VisitDataType(proto::VarType::Type type, Visitor visitor) { - switch (type) { - case proto::VarType::FP16: - visitor.template apply(); - break; - case proto::VarType::FP32: - visitor.template apply(); - break; - case proto::VarType::FP64: - visitor.template apply(); - break; - case proto::VarType::INT32: - visitor.template apply(); - break; - case proto::VarType::INT64: - visitor.template apply(); - break; - case proto::VarType::BOOL: - visitor.template apply(); - break; - case proto::VarType::UINT8: - visitor.template apply(); - break; - case proto::VarType::INT16: - visitor.template apply(); - break; - case proto::VarType::INT8: - visitor.template apply(); - break; - default: - PADDLE_THROW("Not supported %d", type); - } +#define VisitDataTypeCallback(cpp_type, proto_type) \ + do { \ + if (type == proto_type) { \ + visitor.template apply(); \ + return; \ + } \ + } while (0) + + _ForEachDataType_(VisitDataTypeCallback); +#undef VisitDataTypeCallback + PADDLE_THROW("Not supported %d", type); } extern std::string DataTypeToString(const proto::VarType::Type type); -extern size_t SizeOfType(std::type_index type); +extern size_t SizeOfType(proto::VarType::Type type); inline std::ostream& operator<<(std::ostream& out, const proto::VarType::Type& type) { out << DataTypeToString(type); diff --git a/paddle/fluid/framework/data_type_test.cc b/paddle/fluid/framework/data_type_test.cc index 54c41c55ba63c..2a380201f297f 100644 --- a/paddle/fluid/framework/data_type_test.cc +++ b/paddle/fluid/framework/data_type_test.cc @@ -26,15 +26,15 @@ TEST(DataType, float16) { Tensor tensor; CPUPlace cpu; - tensor.mutable_data(cpu, f::ToTypeIndex(dtype)); + tensor.mutable_data(cpu, dtype); // test fp16 tensor - EXPECT_EQ(tensor.type(), std::type_index(typeid(float16))); + EXPECT_EQ(tensor.type(), f::ToDataType(typeid(float16))); // test fp16 size - EXPECT_EQ(f::SizeOfType(f::ToTypeIndex(dtype)), 2u); + EXPECT_EQ(f::SizeOfType(dtype), 2u); // test debug info - std::string type = "float16"; + std::string type = "::paddle::platform::float16"; EXPECT_STREQ(f::DataTypeToString(dtype).c_str(), type.c_str()); } diff --git a/paddle/fluid/framework/details/CMakeLists.txt b/paddle/fluid/framework/details/CMakeLists.txt index d6b5ad4570c1d..63a68ba3a5c28 100644 --- a/paddle/fluid/framework/details/CMakeLists.txt +++ b/paddle/fluid/framework/details/CMakeLists.txt @@ -12,17 +12,36 @@ cc_library(multi_devices_graph_check_pass SRCS multi_devices_graph_check_pass.cc cc_library(variable_visitor SRCS variable_visitor.cc DEPS lod_tensor selected_rows) +if(WITH_DISTRIBUTE) + if(NOT WITH_GRPC) + set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor") + set_source_files_properties(reduce_op_handle.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) + endif() +endif() + if(WITH_GPU) nv_library(all_reduce_op_handle SRCS all_reduce_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory dynload_cuda variable_visitor) - nv_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope ddim dynload_cuda) + if(WITH_DISTRIBUTE) + nv_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope + ddim dynload_cuda selected_rows_functor sendrecvop_rpc) + else() + nv_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope + ddim dynload_cuda selected_rows_functor) + endif() nv_library(broadcast_op_handle SRCS broadcast_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor dynload_cuda) nv_library(fused_broadcast_op_handle SRCS fused_broadcast_op_handle.cc DEPS broadcast_op_handle) else() cc_library(all_reduce_op_handle SRCS all_reduce_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory variable_visitor) - cc_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope ddim) + if(WITH_DISTRIBUTE) + cc_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope + ddim selected_rows_functor sendrecvop_rpc) + else() + cc_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope + ddim selected_rows_functor) + endif() cc_library(broadcast_op_handle SRCS broadcast_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor) cc_library(fused_broadcast_op_handle SRCS fused_broadcast_op_handle.cc DEPS broadcast_op_handle) endif() @@ -31,22 +50,27 @@ cc_library(data_balance_op_handle SRCS data_balance_op_handle.cc DEPS op_handle_ cc_library(gather_op_handle SRCS gather_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor) cc_library(fuse_vars_op_handle SRCS fuse_vars_op_handle.cc DEPS op_handle_base scope) +cc_library(memory_optimize_pass SRCS analysis_var_pass.cc memory_reuse_types.cc DEPS graph graph_helper pass) cc_library(modify_op_lock_and_record_event_pass SRCS modify_op_lock_and_record_event_pass.cc DEPS computation_op_handle op_graph_view multi_devices_helper) - -if (WITH_GPU) - cc_library(reference_count_pass SRCS reference_count_pass.cc DEPS computation_op_handle scale_loss_grad_op_handle rpc_op_handle - all_reduce_op_handle reduce_op_handle broadcast_op_handle data_balance_op_handle graph graph_helper pass) -endif() +cc_library(memory_early_delete_pass SRCS memory_early_delete_pass.cc DEPS memory_optimize_pass computation_op_handle scale_loss_grad_op_handle rpc_op_handle + all_reduce_op_handle reduce_op_handle broadcast_op_handle data_balance_op_handle graph graph_helper pass) +cc_library(reference_count_pass_helper SRCS reference_count_pass_helper.cc DEPS garbage_collector computation_op_handle) +cc_library(eager_deletion_op_handle SRCS eager_deletion_op_handle.cc DEPS lod_tensor selected_rows reference_count_pass_helper) +cc_library(eager_deletion_pass SRCS eager_deletion_pass.cc DEPS computation_op_handle eager_deletion_op_handle graph graph_helper pass) +cc_library(reference_count_pass SRCS reference_count_pass.cc DEPS computation_op_handle graph graph_helper pass op_graph_view reference_count_pass_helper) cc_library(sequential_execution_pass SRCS sequential_execution_pass.cc DEPS graph graph_helper pass) +cc_library(all_reduce_deps_pass SRCS all_reduce_deps_pass.cc DEPS graph graph_helper pass) cc_library(multi_devices_graph_pass SRCS multi_devices_graph_pass.cc DEPS multi_devices_helper computation_op_handle scale_loss_grad_op_handle rpc_op_handle all_reduce_op_handle reduce_op_handle broadcast_op_handle data_balance_op_handle fused_broadcast_op_handle) -set(SSA_GRAPH_EXECUTOR_DEPS graph framework_proto sequential_execution_pass modify_op_lock_and_record_event_pass) +set(SSA_GRAPH_EXECUTOR_DEPS graph framework_proto sequential_execution_pass modify_op_lock_and_record_event_pass all_reduce_deps_pass reference_count_pass eager_deletion_pass memory_optimize_pass memory_early_delete_pass) if (WITH_GPU) list(APPEND SSA_GRAPH_EXECUTOR_DEPS reference_count_pass) endif() +cc_test(memory_reuse_types_test SRCS memory_reuse_types_test.cc memory_reuse_types.cc DEPS framework_proto graph) +cc_test(analysis_var_pass_test SRCS analysis_var_pass_test.cc analysis_var_pass.cc memory_reuse_types.cc DEPS framework_proto graph graph_helper op_registry pass) cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS ${SSA_GRAPH_EXECUTOR_DEPS}) @@ -67,4 +91,5 @@ cc_test(fused_broadcast_op_test SRCS fused_broadcast_op_handle_test.cc DEPS fuse cc_library(build_strategy SRCS build_strategy.cc DEPS graph_viz_pass multi_devices_graph_pass multi_devices_graph_print_pass multi_devices_graph_check_pass - fuse_elewise_add_act_pass multi_batch_merge_pass) + fuse_elewise_add_act_pass multi_batch_merge_pass + memory_optimize_pass) diff --git a/paddle/fluid/framework/details/all_reduce_deps_pass.cc b/paddle/fluid/framework/details/all_reduce_deps_pass.cc new file mode 100644 index 0000000000000..fe21e21bcfc42 --- /dev/null +++ b/paddle/fluid/framework/details/all_reduce_deps_pass.cc @@ -0,0 +1,125 @@ +// 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 +#include +#include +#include +#include + +#include "paddle/fluid/framework/details/all_reduce_deps_pass.h" +#include "paddle/fluid/framework/details/all_reduce_op_handle.h" +#include "paddle/fluid/framework/details/multi_devices_helper.h" +#include "paddle/fluid/framework/details/op_graph_view.h" +#include "paddle/fluid/framework/details/var_handle.h" +#include "paddle/fluid/framework/ir/graph_helper.h" +#include "paddle/fluid/framework/op_proto_maker.h" + +namespace paddle { +namespace framework { +namespace details { + +static constexpr char kAllOpDescs[] = "all_op_descs"; + +VarHandle* GetValidInput(const OpHandleBase* a) { + for (auto p : a->Inputs()) { + VarHandle* b = dynamic_cast(p); + if (b) { + return b; + } + } + + return nullptr; +} + +std::unique_ptr AllReduceDepsPass::ApplyImpl( + std::unique_ptr graph) const { + auto graph_ops = ir::FilterByNodeWrapper(*graph); + + // get vars order + int order = 0; + std::unordered_map vars; + // TODO(gongwb): use graph topology sort to find the order of operators. + // Note that must assert topology sort is stable + auto& ops = Get>(kAllOpDescs); + for (auto* op_desc : ops) { + auto outputs = op_desc->Outputs(); + for (auto& o_it : outputs) { + for (auto& v : o_it.second) { // values + vars[v] = order; + } + } + order++; + } + + std::vector dist_ops; + // get allreduce ops. + for (auto& op : graph_ops) { + // FIXME(gongwb):add broad cast. + if (op->Name() == "all_reduce" || op->Name() == "reduce") { + dist_ops.push_back(op); + } + } + + VLOG(10) << "dist_ops size:" << dist_ops.size() << std::endl; + + std::sort(dist_ops.begin(), dist_ops.end(), [&](OpHandleBase* op1, + OpHandleBase* op2) { + VarHandle* i0 = dynamic_cast(GetValidInput(op1)); + VarHandle* i1 = dynamic_cast(GetValidInput(op2)); + + PADDLE_ENFORCE(i0 != nullptr && i1 != nullptr, "%s convert to %s error", + op1->DebugString(), op2->DebugString()); + + auto l_it = vars.find(i0->name_); + auto r_it = vars.find(i1->name_); + + if (l_it->second < r_it->second) return true; + + if (l_it->second == r_it->second) { + return i0->name_ < i1->name_; + } + + return false; + }); + + // add dependency. + auto& sorted_ops = dist_ops; + for (size_t i = 1; i < sorted_ops.size(); ++i) { + auto* dep_var = new DummyVarHandle(graph->CreateControlDepVar()); + + auto* pre_op = sorted_ops[i - 1]; + auto* op = sorted_ops[i]; + + pre_op->AddOutput(dep_var); + op->AddInput(dep_var); + graph->Get(kGraphDepVars).emplace(dep_var); + + VLOG(10) << "add all_reduce sequential dependencies between " << pre_op + << " and " << op; + + VLOG(10) << "pre_op:" << pre_op->DebugString() + << ", op:" << op->DebugString(); + } + + return graph; +} + +} // namespace details +} // namespace framework +} // namespace paddle + +REGISTER_PASS(all_reduce_deps_pass, + paddle::framework::details::AllReduceDepsPass) + .RequirePassAttr(paddle::framework::details::kAllOpDescs); diff --git a/paddle/fluid/framework/details/all_reduce_deps_pass.h b/paddle/fluid/framework/details/all_reduce_deps_pass.h new file mode 100644 index 0000000000000..e8b91089816c7 --- /dev/null +++ b/paddle/fluid/framework/details/all_reduce_deps_pass.h @@ -0,0 +1,33 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include "paddle/fluid/framework/ir/graph.h" +#include "paddle/fluid/framework/ir/pass.h" + +namespace paddle { +namespace framework { +namespace details { + +// TODO(gongwb): overlap allreduce with backward computation. +class AllReduceDepsPass : public ir::Pass { + protected: + std::unique_ptr ApplyImpl( + std::unique_ptr graph) const override; +}; + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/all_reduce_op_handle.cc b/paddle/fluid/framework/details/all_reduce_op_handle.cc index b8690156763e4..9eaff1f560147 100644 --- a/paddle/fluid/framework/details/all_reduce_op_handle.cc +++ b/paddle/fluid/framework/details/all_reduce_op_handle.cc @@ -23,7 +23,7 @@ namespace paddle { namespace framework { namespace details { -#ifdef PADDLE_WITH_CUDA +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) AllReduceOpHandle::AllReduceOpHandle(ir::Node *node, const std::vector &local_scopes, const std::vector &places, @@ -48,7 +48,14 @@ AllReduceOpHandle::AllReduceOpHandle(ir::Node *node, void AllReduceOpHandle::RunImpl() { platform::RecordEvent record_event(Name(), dev_ctxes_.cbegin()->second); +// FIXME(typhoonzero): If scope0(global scope) have NCCL_ID_VAR, +// this is a distributed or inter-process call, find a better way. +#ifdef PADDLE_WITH_CUDA + if (NoDummyInputSize() == 1 && + local_scopes_[0]->FindLocalVar(NCCL_ID_VARNAME) == nullptr) { +#else if (NoDummyInputSize() == 1) { +#endif return; // No need to all reduce when GPU count = 1; } else { // Wait input done @@ -74,7 +81,7 @@ void AllReduceOpHandle::RunImpl() { } if (platform::is_gpu_place(lod_tensors[0]->place())) { -#ifdef PADDLE_WITH_CUDA +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) PADDLE_ENFORCE(nccl_ctxs_, "nccl_ctxs should not be nullptr."); int dtype = -1; size_t numel = 0; @@ -120,7 +127,7 @@ void AllReduceOpHandle::RunImpl() { // Reduce All Tensor to trg in CPU ReduceLoDTensor func(lod_tensors, &trg); - VisitDataType(ToDataType(lod_tensors[0]->type()), func); + VisitDataType(lod_tensors[0]->type(), func); for (size_t i = 1; i < local_scopes_.size(); ++i) { auto &scope = diff --git a/paddle/fluid/framework/details/all_reduce_op_handle.h b/paddle/fluid/framework/details/all_reduce_op_handle.h index f6ef3a1367b91..b449796fcaee7 100644 --- a/paddle/fluid/framework/details/all_reduce_op_handle.h +++ b/paddle/fluid/framework/details/all_reduce_op_handle.h @@ -20,7 +20,7 @@ #include "paddle/fluid/framework/details/op_handle_base.h" #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/scope.h" -#ifdef PADDLE_WITH_CUDA +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) #include "paddle/fluid/platform/nccl_helper.h" #endif @@ -29,7 +29,7 @@ namespace framework { namespace details { struct AllReduceOpHandle : public OpHandleBase { -#ifdef PADDLE_WITH_CUDA +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) AllReduceOpHandle(ir::Node *node, const std::vector &local_scopes, const std::vector &places, const platform::NCCLContextMap *ctxs); @@ -49,7 +49,7 @@ struct AllReduceOpHandle : public OpHandleBase { private: std::vector local_scopes_; std::vector places_; -#ifdef PADDLE_WITH_CUDA +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) const platform::NCCLContextMap *nccl_ctxs_; #endif }; diff --git a/paddle/fluid/framework/details/analysis_var_pass.cc b/paddle/fluid/framework/details/analysis_var_pass.cc new file mode 100644 index 0000000000000..223b9da3cfba3 --- /dev/null +++ b/paddle/fluid/framework/details/analysis_var_pass.cc @@ -0,0 +1,656 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/framework/details/analysis_var_pass.h" +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#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 { + +static inline bool IsSameDesc(OpDesc* op1, OpDesc* op2) { + return op1->Type() == op2->Type() && op1->Inputs() == op2->Inputs() && + op1->Outputs() == op2->Outputs(); +} + +template +class FilterVariableImpl { + public: + void operator()(const Container& nodes, Callback callback) { + for (auto* node : nodes) { + callback(node); + } + } +}; + +// filter var node for op->inputs/outputs +template +class FilterVariableImpl, Callback> { + public: + void operator()(const std::vector& nodes, Callback callback) { + for (auto* var : nodes) { + if (var->IsVar() && !var->IsCtrlVar()) { + callback(var); + } + } + } +}; + +template +void FilterVariables(const Container& nodes, Callback callback) { + FilterVariableImpl()(nodes, callback); +} + +std::unique_ptr AnalysisVarPass::ApplyImpl( + std::unique_ptr graph) const { + auto nodes = graph->Nodes(); + auto subblock_vars = GetSubBlockVars(nodes); + skip_set_.insert(subblock_vars.begin(), subblock_vars.end()); + + 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) { + ir::Node* cache = pool_.NodeMatch(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) { + 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_.GetIndex(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(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 + for (auto var : cfg_->LiveIn(op)) { + if (cfg_->LiveOut(op).count(var) == 0) { + ir::Node* var_node = cfg_->GetNodeFromVarName(var, op); + if (var_node == nullptr) continue; + if (NodeCanReused(var_node) && !pool_.Has(var_node)) { + pool_.Insert(var_node, op); + } + } + } + } + graph->ResolveHazard(var_nodes_); + + // For early delete pass. use GraphNodePool load the unlived vars. + // 1. find all deps op for each unlived var in memory pool. + for (auto& op : graph->Nodes()) { + for (auto& var : op->inputs) { + if (pool_.Has(var)) { + pool_.Insert(var, op); + } + } + } + // 2. convert ir node based memory pool to graph node + // because Node* maybe released bettwen passes. + auto& graph_pool = graph->Get(kGraphNodePool); + for (auto it = pool_.begin(); it != pool_.end(); ++it) { + std::unordered_set descs; + for (auto& op : it->second) { + PADDLE_ENFORCE(op->IsOp()); + descs.insert(op->Op()); + } + graph_pool.push_back(std::make_pair(it->first->Name(), descs)); + } + + return graph; +} + +void AnalysisVarPass::SubGraphOptimize(OpDesc* op_desc) const { + // conditional block, while op and their grad op + auto* sub_block_desc = + AttrReader(op_desc->GetAttrMap()).Get("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 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_.NodeMatch(var); + if (cache != nullptr) { + if (var->Var()->GetDataType() != cache->Var()->GetDataType()) { + continue; + } + int node_idx_in_pool = pool_.GetIndex(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(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()); + } + } + } + } + } +} + +std::unordered_set AnalysisVarPass::GetSubBlockVars( + const std::unordered_set& nodes) const { + std::unordered_set vars; + for (auto& op : nodes) { + if (!op->IsOp() || op->Op() == nullptr) continue; + auto* op_desc = op->Op(); + if (OpHasSubBlock(op_desc)) { + auto inputs = op_desc->InputArgumentNames(); + auto outputs = op_desc->OutputArgumentNames(); + vars.insert(inputs.begin(), inputs.end()); + vars.insert(outputs.begin(), outputs.end()); + } + } + return vars; +} + +void AnalysisVarPass::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 AnalysisVarPass::InitSSAGraphNodes() const { + std::unordered_map> 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 AnalysisVarPass::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 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(); +} + +bool AnalysisVarPass::NodeCanReused(ir::Node* node) const { + if (!node->IsVar() || node->IsCtrlVar()) return false; + auto* desc = node->Var(); + auto type = desc->GetType(); + if (desc->Persistable() || type != proto::VarType::LOD_TENSOR || + desc->GetShape().empty()) { + return false; + } + // vars can be @EMPTY@, @LR_DECAY_REUSE_ID@. For example, while_grad + std::string name = node->Name(); + if (!name.empty() && name[0] == '@' && name[name.size() - 1] == '@') + return false; + if (skip_set_.count(name)) return false; + for (auto* op : node->inputs) { + if (op->Op()->HasAttr("force_cpu")) { + // op output force generated in cpu, can not be reused. + return framework::AttrReader(op->Op()->GetAttrMap()) + .Get("force_cpu") == 0; + } + } + return true; +} + +bool AnalysisVarPass::OpHasSubBlock(OpDesc* desc) const { + const AttributeMap& attrs = desc->GetAttrMap(); + for (auto& attr : attrs) { + if (attr.second.type() == typeid(BlockDesc*) || // NOLINT + attr.second.type() == typeid(std::vector)) // NOLINT + return true; + } + return false; +} + +std::vector SortOpLikeDescOrder(const ir::Graph& graph) { + PADDLE_ENFORCE(graph.Has(kAllOpDescs), + "Graph has no attribute of kAllOpDescs."); + // 1. get op desc order + auto& op_descs = graph.Get>(kAllOpDescs); + + // 2. topology sort order + auto nodes = graph.Nodes(); + std::deque ops; + FilterVariables(nodes, [&](ir::Node* op) { + if (op->IsOp() && op->Op() != nullptr) { + ops.emplace_back(op); + } + }); + std::unordered_map op_deps; + std::list ready_ops; + std::unordered_map> pending_ops; + + for (auto* op : ops) { + std::unordered_set preceding_op; + for (auto* in : op->inputs) { + if (in->inputs.empty()) continue; + PADDLE_ENFORCE(in->inputs.size() == 1 && in->inputs[0]->IsOp()); + preceding_op.emplace(in->inputs[0]); + pending_ops[in->inputs[0]].emplace(op); + } + op_deps[op] = preceding_op.size(); + if (preceding_op.empty()) { + ready_ops.emplace_back(op); + } + } + + // 3. generated op list based desc order and the topology order + std::vector ret; + std::list op_descs_list(op_descs.begin(), op_descs.end()); + + auto update_by_found_node = [&](ir::Node* found_node) { + for (auto* pending_op : pending_ops[found_node]) { + if (--op_deps[pending_op] == 0) { + ready_ops.emplace_back(pending_op); + } + } + ready_ops.remove(found_node); + ret.emplace_back(found_node); + }; + + while (!ready_ops.empty()) { + bool all_of_ready_op_unmatched = true; + for (auto it = op_descs_list.begin(); it != op_descs_list.end();) { + auto op_desc = *it; + ir::Node* found_node = nullptr; + for (auto* op : ready_ops) { + if (IsSameDesc(op->Op(), op_desc)) { + found_node = op; + break; + } + } + + // 3.1 op desc deleted by other pass + if (found_node == nullptr) { + ++it; + continue; + } else { + all_of_ready_op_unmatched = false; + it = op_descs_list.erase(it); + } + update_by_found_node(found_node); + } + + // 3.2 op descs are added by other pass + // preceding op non empty means some new op descs are + // created, but not contained in return node list. + // these new op desc may depend on each other. + std::list prev_ready_ops(ready_ops); + if (all_of_ready_op_unmatched) { + for (auto op : prev_ready_ops) { + update_by_found_node(op); + } + } + } + + PADDLE_ENFORCE(std::all_of( + op_deps.begin(), op_deps.end(), + [&](const std::pair& p) { return p.second == 0; })); + + return ret; +} + +ControlFlowGraph::ControlFlowGraph(const ir::Graph& graph) { + ops_ = SortOpLikeDescOrder(graph); + ConnectNodes(); +} + +void ControlFlowGraph::BuildCFGGraph() { + // FIXME(dzh): same effect with ConnectNodes, but use the control + // link to build dependency graph, it goes wrong in transformer. + for (ir::Node* op : ops_) { + for (auto& input_var : op->inputs) { + if (!input_var->inputs.empty()) { + PADDLE_ENFORCE( + input_var->inputs.size() == 1 && input_var->inputs[0]->IsOp(), + "Preceding Op Node of Var Node must be unique"); + auto* pred_op = input_var->inputs[0]; + if (pred_op->Op() != nullptr) { + predecessors_[op].insert(pred_op); + successors_[pred_op].insert(op); + } + } + if (input_var->IsVar() && !input_var->IsCtrlVar()) { + uses_[op].insert(input_var->Name()); + } + } + for (auto& output_var : op->outputs) { + // output var may be used by many op + for (auto* succ_op : output_var->outputs) { + if (succ_op->Op() != nullptr) { + successors_[op].insert(succ_op); + predecessors_[succ_op].insert(op); + } + } + if (output_var->IsVar() && !output_var->IsCtrlVar()) { + defs_[op].insert(output_var->Name()); + } + } + } +} + +void ControlFlowGraph::ConnectNodes() { + for (size_t i = 0; i < ops_.size(); ++i) { + auto& op = ops_[i]; + try { + auto& next_op = ops_.at(i + 1); + successors_[op].insert(next_op); + predecessors_[next_op].insert(op); + } catch (...) { + // do nothing + } + + FilterVariables(op->inputs, + [&](ir::Node* var) { uses_[op].emplace(var->Name()); }); + + FilterVariables(op->outputs, + [&](ir::Node* var) { defs_[op].emplace(var->Name()); }); + } +} + +void ControlFlowGraph::LiveVariableAnalysis() { + // NOTE(dzh): variable liveless analysis (a.k.a reversed_ops algorithm) + // compute the liveness of for each variable though reversed_ops algorithm. + // It iterates the operators from end to begin, compute the live in/live out + // variable set for each op, then the diff between in/out will be used for + // the variable reuse. For detail refer to + // http://www.cs.cornell.edu/courses/cs4120/2013fa/lectures/lec26-fa13.pdf + std::list work_list(ops_.rbegin(), ops_.rend()); + while (!work_list.empty()) { + ir::Node* op = work_list.front(); + work_list.pop_front(); + // get the live_in calculated before. Empty if first. + auto prev_live_in = std::move(live_in_[op]); + for (auto& s : successors_[op]) { + for (auto& var : live_in_[s]) { + live_out_[op].insert(var); + } + } + for (auto& var : uses_[op]) { + live_in_[op].insert(var); + } + for (auto& var : live_out_[op]) { + live_in_[op].insert(var); + } + for (auto& var : defs_[op]) { + live_in_[op].erase(var); + } + + // If the live_in is not changed, then the liveness analysis of + // predecessors is completed. + // + // Otherwise, recalculate the predecessors liveness + if (live_in_[op] != prev_live_in) { + for (auto& pre : predecessors_[op]) { + work_list.push_back(pre); + } + } + } +} + +void ControlFlowGraph::RenameVarInCFGGraph(const std::string& old_node, + const std::string& new_node, + int begin_idx) { + // update graph from begin idx to the end + for (size_t i = begin_idx; i != ops_.size(); ++i) { + auto* op = ops_[i]; + if (uses_[op].find(old_node) != uses_[op].end()) { + uses_[op].erase(old_node); + uses_[op].insert(new_node); + } + if (defs_[op].find(old_node) != defs_[op].end()) { + defs_[op].erase(old_node); + defs_[op].insert(new_node); + } + if (live_in_[op].find(old_node) != live_in_[op].end()) { + live_in_[op].erase(old_node); + live_in_[op].insert(new_node); + } + if (live_out_[op].find(old_node) != live_out_[op].end()) { + live_out_[op].erase(old_node); + live_out_[op].insert(new_node); + } + } +} + +const std::set ControlFlowGraph::LiveIn(ir::Node* op) const { + auto it = live_in_.find(op); + PADDLE_ENFORCE( + it != live_in_.end(), + string::Sprintf("Expect %s in live_in, but Not Found.", op->Name())); + return it->second; +} + +const std::set ControlFlowGraph::LiveOut(ir::Node* op) const { + auto it = live_out_.find(op); + PADDLE_ENFORCE( + it != live_out_.end(), + string::Sprintf("Expect %s in live_out, but Not Found.", op->Name())); + return it->second; +} + +const std::set ControlFlowGraph::Use(ir::Node* op) const { + auto it = uses_.find(op); + PADDLE_ENFORCE( + it != uses_.end(), + string::Sprintf("Expect %s in live_out, but Not Found.", op->Name())); + return it->second; +} + +const std::vector ControlFlowGraph::Ops() const { return ops_; } + +std::vector& ControlFlowGraph::Ops() { return ops_; } + +ir::Node* ControlFlowGraph::GetNodeFromVarName(const std::string& name, + ir::Node* op) const { + // in ssa-graph, different version nodes have same name, + // this function get the latest version var before target op + // It may return nullptr, such as data node. + ir::Node* found_node = nullptr; + for (auto* node : ops_) { + if (node == op) break; + for (auto& output : node->outputs) { + if (output->Name() == name) { + found_node = output; + } + } + } + return found_node; +} + +} // namespace details +} // namespace framework +} // namespace paddle + +REGISTER_PASS(analysis_var_pass, paddle::framework::details::AnalysisVarPass) + .RequireGraphAttr(paddle::framework::details::kGraphNodePool) + .RequireGraphAttr(paddle::framework::details::kAllOpDescs); diff --git a/paddle/fluid/framework/details/analysis_var_pass.h b/paddle/fluid/framework/details/analysis_var_pass.h new file mode 100644 index 0000000000000..144204beafb34 --- /dev/null +++ b/paddle/fluid/framework/details/analysis_var_pass.h @@ -0,0 +1,120 @@ +// 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 +#include +#include +#include +#include +#include +#include +#include +#include + +#include "paddle/fluid/framework/data_type.h" +#include "paddle/fluid/framework/details/memory_reuse_types.h" +#include "paddle/fluid/framework/ir/graph.h" +#include "paddle/fluid/framework/ir/pass.h" + +namespace paddle { +namespace framework { +namespace details { +constexpr char kAllOpDescs[] = "all_op_descs"; + +std::vector SortOpLikeDescOrder(const ir::Graph& graph); +// sort op in bfs order +std::vector BFSSortGraphOps(const ir::Graph& graph); + +class ControlFlowGraph; + +class AnalysisVarPass : public ir::Pass { + protected: + std::unique_ptr ApplyImpl( + std::unique_ptr graph) const override; + + private: + // fill the variable map(var_nodes) by version. + void InitSSAGraphNodes() const; + // update program descs + void RenameVarInGraphDesc(const std::string& var, + const std::string& cache_var, size_t idx) const; + // update ir nodes + void RenameVarInGraphNode(const std::string& var, + const std::string& cache_var, size_t idx, + ir::Graph* graph) const; + + void SubGraphOptimize(OpDesc* op_desc) const; + // valid a tensor can be reuse or not + bool NodeCanReused(ir::Node* node) const; + // scan subblock and collect the output/input variables. + std::unordered_set GetSubBlockVars( + const std::unordered_set&) const; + // check op has subblock or not + bool OpHasSubBlock(OpDesc* desc) const; + + private: + // Reuse Node Pool, Owned. + mutable OrderedNodePairPool pool_; + // controlflow Graph + mutable std::unique_ptr cfg_; + // skip set + mutable std::unordered_set skip_set_; + // var nodes + mutable std::map> 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 LiveIn(ir::Node* op) const; + const std::set LiveOut(ir::Node* op) const; + const std::set Use(ir::Node* op) const; + const std::vector Ops() const; + std::vector& 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>; + using VarSetMap = std::map>; + // 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 ops_; // op sequence by topology sort +}; + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/analysis_var_pass_test.cc b/paddle/fluid/framework/details/analysis_var_pass_test.cc new file mode 100644 index 0000000000000..9bc4fd33f7058 --- /dev/null +++ b/paddle/fluid/framework/details/analysis_var_pass_test.cc @@ -0,0 +1,470 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/framework/details/analysis_var_pass.h" +#include +#include +#include +#include "glog/logging.h" +#include "gtest/gtest.h" +#include "paddle/fluid/framework/ir/graph.h" +#include "paddle/fluid/framework/ir/graph_helper.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/framework/operator.h" +#include "paddle/fluid/framework/program_desc.h" + +namespace paddle { +namespace framework { + +class DummyOp : public OperatorBase { + public: + DummyOp(const std::string& type, const VariableNameMap& inputs, + const VariableNameMap& outputs, const AttributeMap& attrs) + : OperatorBase(type, inputs, outputs, attrs) {} + + private: + void RunImpl(const Scope& scope, + const platform::Place& place) const override {} +}; + +class SumOpMaker : public OpProtoAndCheckerMaker { + public: + void Make() { + AddInput("X", "").AsDuplicable(); + AddOutput("Out", ""); + AddComment(""); + } +}; + +class AssignOpMaker : public OpProtoAndCheckerMaker { + public: + void Make() { + AddInput("X", "").AsDuplicable(); + AddOutput("Out", ""); + AddComment(""); + } +}; + +class DummyVarTypeInference : public VarTypeInference { + public: + void operator()(const OpDesc& op_desc, BlockDesc* block) const override { + auto& inputs = op_desc.Input("X"); + auto type = block->Var(inputs.front())->GetType(); + auto out_var_name = op_desc.Output("Out").front(); + block->Var(out_var_name)->SetType(type); + } +}; + +} // namespace framework +} // namespace paddle + +REGISTER_OPERATOR(sum, paddle::framework::DummyOp, + paddle::framework::SumOpMaker, + paddle::framework::DummyVarTypeInference); +REGISTER_OPERATOR(assign, paddle::framework::DummyOp, + paddle::framework::AssignOpMaker, + paddle::framework::DummyVarTypeInference); +REGISTER_OPERATOR(dummy, paddle::framework::DummyOp, + paddle::framework::SumOpMaker, + paddle::framework::DummyVarTypeInference); +/* + https://en.wikipedia.org/wiki/Live_variable_analysis + Create a customed classical dependency graph, left row is the instruction + number. + 1. a = 1 + 2. b = a + 3. c = a + 4. d = b + c + 5. e = d + + a--------+ + | | + b c + | | + d--------+ + | + e + Then analysis these variable's liveness range + */ + +namespace paddle { +namespace framework { +namespace details { + +static inline bool IsSameDesc(OpDesc* op1, OpDesc* op2) { + return op1->Type() == op2->Type() && op1->Inputs() == op2->Inputs() && + op1->Outputs() == op2->Outputs(); +} + +inline static ProgramDesc FillProgramDesc() { + ProgramDesc prog; + prog.MutableBlock(0)->Var("a")->SetType(proto::VarType::LOD_TENSOR); + prog.MutableBlock(0)->Var("b")->SetType(proto::VarType::LOD_TENSOR); + prog.MutableBlock(0)->Var("c")->SetType(proto::VarType::LOD_TENSOR); + prog.MutableBlock(0)->Var("d")->SetType(proto::VarType::LOD_TENSOR); + prog.MutableBlock(0)->Var("e")->SetType(proto::VarType::LOD_TENSOR); + { + auto* op = prog.MutableBlock(0)->AppendOp(); + op->SetType("assign"); + op->SetInput("X", {"a"}); + op->SetOutput("Out", {"b"}); + } + { + auto* op = prog.MutableBlock(0)->AppendOp(); + op->SetType("assign"); + op->SetInput("X", {"a"}); + op->SetOutput("Out", {"c"}); + } + { + auto* op = prog.MutableBlock(0)->AppendOp(); + op->SetType("sum"); + op->SetInput("X", {"b", "c"}); + op->SetOutput("Out", {"d"}); + } + { + auto* op = prog.MutableBlock(0)->AppendOp(); + op->SetType("assign"); + op->SetInput("X", {"d"}); + op->SetOutput("Out", {"e"}); + } + return prog; +} + +template +inline static std::string DebugString(const Container& c) { + std::stringstream ss; + for (auto& item : c) { + ss << item << " "; + } + return ss.str(); +} + +TEST(CFGGraph, IRGraph) { + // prepare ir graph + auto prog = FillProgramDesc(); + ir::Graph graph(prog); + const std::vector* all_op_descs = + new std::vector(prog.Block(0).AllOps()); + graph.Set(details::kAllOpDescs, all_op_descs); // take ownership + + ControlFlowGraph cfg(graph); + cfg.LiveVariableAnalysis(); + + // test assign op + ASSERT_TRUE((std::set{"a"} == cfg.LiveIn(cfg.Ops()[0]))); + ASSERT_TRUE((std::set{"a", "b"} == cfg.LiveOut(cfg.Ops()[0]))); + + // test assign op + ASSERT_TRUE((std::set{"a", "b"} == cfg.LiveIn(cfg.Ops()[1]))); + ASSERT_TRUE((std::set{"b", "c"} == cfg.LiveOut(cfg.Ops()[1]))); + + // test sum op + ASSERT_TRUE((std::set{"b", "c"} == cfg.LiveIn(cfg.Ops()[2]))); + ASSERT_TRUE((std::set{"d"} == cfg.LiveOut(cfg.Ops()[2]))); + + // test assign op + ASSERT_TRUE((std::set{"d"} == cfg.LiveIn(cfg.Ops()[3]))); + ASSERT_TRUE((std::set{} == cfg.LiveOut(cfg.Ops()[3]))); +} + +// 1. normal test +TEST(SortOpLikeDescOrder, NormalTest) { + auto prog = FillProgramDesc(); + ir::Graph graph(prog); + const std::vector* all_op_descs = + new std::vector(prog.Block(0).AllOps()); + graph.Set(details::kAllOpDescs, all_op_descs); // take ownership + + auto nodes = SortOpLikeDescOrder(graph); + auto op_descs = prog.Block(0).AllOps(); + for (size_t i = 0; i < nodes.size(); ++i) { + auto node = nodes[i]; + auto op_desc = op_descs[i]; + ASSERT_TRUE(IsSameDesc(node->Op(), op_desc)); + } +} + +// 2. remove some op_desc +TEST(SortOpLikeDescOrder, RemoveOpDesc) { + auto prog = FillProgramDesc(); + ir::Graph graph(prog); + const std::vector* all_op_descs = + new std::vector(prog.Block(0).AllOps()); + graph.Set(details::kAllOpDescs, all_op_descs); // take ownership + auto nodes = graph.Nodes(); + auto op_descs = prog.Block(0).AllOps(); + ir::Node* found_node = nullptr; + for (auto node : nodes) { + if (node->IsOp() && node->outputs.back()->Name() == "e") { + found_node = node; + break; + } + } + PADDLE_ENFORCE(found_node != nullptr); + for (auto it = op_descs.begin(); it != op_descs.end();) { + if (IsSameDesc(*it, found_node->Op())) { + it = op_descs.erase(it); + } else { + ++it; + } + } + + auto find_node_in_graph = [&](std::string s) { + ir::Node* ret = nullptr; + for (auto n : graph.Nodes()) { + if (n->Name() == s) { + ret = n; + break; + } + } + PADDLE_ENFORCE(ret != nullptr); + return ret; + }; + + ir::Node* e = find_node_in_graph("e"); + ir::Node* d = find_node_in_graph("d"); + std::remove(d->outputs.begin(), d->outputs.end(), found_node); + graph.RemoveNode(found_node); + graph.RemoveNode(e); + + // other node keeps the same order + auto remain_nodes = SortOpLikeDescOrder(graph); + for (size_t i = 0; i < remain_nodes.size(); ++i) { + auto node = remain_nodes[i]; + auto op_desc = op_descs[i]; + ASSERT_TRUE(IsSameDesc(node->Op(), op_desc)); + } +} + +// 3. add some op_desc +TEST(SortOpLikeDescOrder, AddOpDesc) { + auto prog = FillProgramDesc(); + const std::vector* all_op_descs = + new std::vector(prog.Block(0).AllOps()); + ir::Graph graph(prog); + + auto find_node_in_graph = [&](std::string s) { + ir::Node* ret = nullptr; + for (auto n : graph.Nodes()) { + if (n->Name() == s) { + ret = n; + break; + } + } + PADDLE_ENFORCE(ret != nullptr); + return ret; + }; + + // cached desc different with real one + // mimic the intermidiete pass modify the programdesc. + graph.Set(details::kAllOpDescs, all_op_descs); // take ownership + + auto op_descs = prog.Block(0).AllOps(); + + auto op = prog.MutableBlock(0)->AppendOp(); + prog.MutableBlock(0)->Var("d1")->SetType(proto::VarType::LOD_TENSOR); + op->SetType("sum"); + op->SetInput("X", {"b", "c"}); + op->SetOutput("Out", {"d1"}); + ir::Node* node = graph.CreateOpNode(op); + ir::Node* d1 = graph.CreateVarNode(prog.MutableBlock(0)->Var("d1")); + ir::Node* b = find_node_in_graph("b"); + ir::Node* c = find_node_in_graph("c"); + node->outputs.emplace_back(d1); + node->inputs.emplace_back(b); + node->inputs.emplace_back(c); + d1->inputs.emplace_back(node); + b->outputs.emplace_back(node); + c->outputs.emplace_back(node); + op_descs.insert(op_descs.begin() + 4, op); + + auto nodes = SortOpLikeDescOrder(graph); + + for (size_t i = 0; i < nodes.size(); ++i) { + auto node = nodes[i]; + auto op_desc = op_descs[i]; + ASSERT_TRUE(IsSameDesc(node->Op(), op_desc)); + } +} + +// 4. add and delete some op_desc +TEST(SortOpLikeDescOrder, AddAndDeleteOpDesc) { + auto prog = FillProgramDesc(); + ir::Graph graph(prog); + const std::vector* all_op_descs = + new std::vector(prog.Block(0).AllOps()); + graph.Set(details::kAllOpDescs, all_op_descs); // take ownership + + auto find_node_in_graph = [&](std::string s) { + ir::Node* ret = nullptr; + for (auto n : graph.Nodes()) { + if (n->Name() == s) { + ret = n; + break; + } + } + PADDLE_ENFORCE(ret != nullptr); + return ret; + }; + + // remove sum node + auto op_descs = prog.Block(0).AllOps(); + ir::Node* found_node = nullptr; + auto nodes = graph.Nodes(); + for (auto node : nodes) { + if (node->Name() == "sum") { + found_node = node; + break; + } + } + PADDLE_ENFORCE(found_node != nullptr); + for (auto it = op_descs.begin(); it != op_descs.end();) { + if (IsSameDesc(*it, found_node->Op())) { + it = op_descs.erase(it); + } else { + ++it; + } + } + { + ir::Node* d = find_node_in_graph("d"); + ir::Node* c = find_node_in_graph("c"); + ir::Node* e = find_node_in_graph("e"); + std::remove(d->outputs.begin(), d->outputs.end(), found_node); + std::remove(c->outputs.begin(), c->outputs.end(), found_node); + ir::Node* pending_op = found_node->outputs[0]->outputs[0]; + graph.RemoveNode(e); + graph.RemoveNode(pending_op); + graph.RemoveNode(found_node); + } + + // add node + auto op = prog.MutableBlock(0)->AppendOp(); + prog.MutableBlock(0)->Var("d1")->SetType(proto::VarType::LOD_TENSOR); + op->SetType("sum"); + op->SetInput("X", {"b", "c"}); + op->SetOutput("Out", {"d1"}); + { + ir::Node* node = graph.CreateOpNode(op); + ir::Node* d1 = graph.CreateVarNode(prog.MutableBlock(0)->Var("d1")); + ir::Node* b = find_node_in_graph("b"); + ir::Node* c = find_node_in_graph("c"); + node->outputs.emplace_back(d1); + node->inputs.emplace_back(b); + node->inputs.emplace_back(c); + b->outputs.emplace_back(node); + c->outputs.emplace_back(node); + } + op_descs.insert(op_descs.begin() + 2, op); + + // check the order + auto mynodes = SortOpLikeDescOrder(graph); + for (size_t i = 0; i < mynodes.size(); ++i) { + auto node = mynodes[i]; + auto op_desc = op_descs[i]; + ASSERT_TRUE(IsSameDesc(node->Op(), op_desc)); + } +} + +// 5. add and replace some op_desc inplace. +TEST(SortOpLikeDescOrder, AddAndReplaceOpDescInplace) { + auto prog = FillProgramDesc(); + ir::Graph graph(prog); + const std::vector* all_op_descs = + new std::vector(prog.Block(0).AllOps()); + graph.Set(details::kAllOpDescs, all_op_descs); // take ownership + + auto find_node_in_graph = [&](std::string s) { + ir::Node* ret = nullptr; + for (auto n : graph.Nodes()) { + if (n->Name() == s) { + ret = n; + break; + } + } + PADDLE_ENFORCE(ret != nullptr); + return ret; + }; + + auto op_descs = prog.Block(0).AllOps(); + // add node + auto op = prog.MutableBlock(0)->AppendOp(); + prog.MutableBlock(0)->Var("d1")->SetType(proto::VarType::LOD_TENSOR); + op->SetType("sum"); + op->SetInput("X", {"b", "c"}); + op->SetOutput("Out", {"d1"}); + { + ir::Node* node = graph.CreateOpNode(op); + ir::Node* d1 = graph.CreateVarNode(prog.MutableBlock(0)->Var("d1")); + ir::Node* b = find_node_in_graph("b"); + ir::Node* c = find_node_in_graph("c"); + node->outputs.emplace_back(d1); + node->inputs.emplace_back(b); + node->inputs.emplace_back(c); + d1->inputs.emplace_back(node); + b->outputs.emplace_back(node); + c->outputs.emplace_back(node); + } + + op_descs.emplace_back(op); + + // replace op_desc inplace + auto nodes = graph.Nodes(); + ir::Node* found_node = nullptr; + for (auto node : nodes) { + if (node->IsOp() && node->Op() && node->Name() == "assign") { + if (node->outputs.size() == 1 && node->outputs[0]->Name() == "e") { + found_node = node; + break; + } + } + } + { + ir::Node* d = find_node_in_graph("d"); + ir::Node* e = find_node_in_graph("e"); + std::remove(d->outputs.begin(), d->outputs.end(), found_node); + std::remove(e->inputs.begin(), e->inputs.end(), found_node); + graph.RemoveNode(found_node); + } + op_descs.erase(op_descs.begin() + 3); + + auto replace_op = prog.MutableBlock(0)->AppendOp(); + replace_op->SetType("sum"); + replace_op->SetInput("X", {"d", "d1"}); + replace_op->SetOutput("Out", {"e"}); + { + ir::Node* sum2 = graph.CreateOpNode(replace_op); + ir::Node* e = find_node_in_graph("e"); + ir::Node* d = find_node_in_graph("d"); + ir::Node* d1 = find_node_in_graph("d1"); + sum2->inputs.emplace_back(d); + sum2->inputs.emplace_back(d1); + sum2->outputs.emplace_back(e); + e->inputs.emplace_back(sum2); + d->outputs.emplace_back(sum2); + d1->outputs.emplace_back(sum2); + } + + op_descs.emplace_back(replace_op); + // compare op order + auto graph_nodes = SortOpLikeDescOrder(graph); + for (size_t i = 0; i < graph_nodes.size(); ++i) { + auto node = graph_nodes[i]; + auto op_desc = op_descs[i]; + ASSERT_TRUE(IsSameDesc(node->Op(), op_desc)); + } +} + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/broadcast_op_handle.cc b/paddle/fluid/framework/details/broadcast_op_handle.cc index 8e5e542765938..cf280c29ff8c7 100644 --- a/paddle/fluid/framework/details/broadcast_op_handle.cc +++ b/paddle/fluid/framework/details/broadcast_op_handle.cc @@ -60,7 +60,7 @@ void BroadcastOpHandle::BroadcastOneVar( PADDLE_ENFORCE_NOT_NULL(in_var); Tensor &in_tensor = VariableVisitor::GetMutableTensor(in_var); if (UNLIKELY(!in_tensor.IsInitialized())) { - VLOG(30) << "in var " << in_var_handle.name_ << "not inited, return!"; + VLOG(3) << "in var " << in_var_handle.name_ << "not inited, return!"; return; } @@ -82,7 +82,7 @@ void BroadcastOpHandle::BroadcastOneVar( }); } } else { -#ifdef PADDLE_WITH_CUDA +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) VarHandle *out_handle = nullptr; int root_id = boost::get(in_tensor.place()).device; std::vector> broadcast_calls; diff --git a/paddle/fluid/framework/details/broadcast_op_handle.h b/paddle/fluid/framework/details/broadcast_op_handle.h index 72180fac86425..0c75e05f86163 100644 --- a/paddle/fluid/framework/details/broadcast_op_handle.h +++ b/paddle/fluid/framework/details/broadcast_op_handle.h @@ -24,7 +24,7 @@ #include "paddle/fluid/framework/selected_rows.h" #include "paddle/fluid/platform/device_context.h" -#ifdef PADDLE_WITH_CUDA +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) #include "paddle/fluid/platform/nccl_helper.h" #endif @@ -34,7 +34,7 @@ namespace details { struct BroadcastOpHandle : public OpHandleBase { public: -#ifdef PADDLE_WITH_CUDA +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) BroadcastOpHandle(ir::Node *node, const std::vector &local_scopes, const std::vector &places, const platform::NCCLContextMap *nccl_ctxs) @@ -68,7 +68,7 @@ struct BroadcastOpHandle : public OpHandleBase { std::vector local_scopes_; std::vector places_; -#ifdef PADDLE_WITH_CUDA +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) const platform::NCCLContextMap *nccl_ctxs_; #endif diff --git a/paddle/fluid/framework/details/broadcast_op_handle_test.h b/paddle/fluid/framework/details/broadcast_op_handle_test.h index 4305eb65733a7..df3b3cc9ca012 100644 --- a/paddle/fluid/framework/details/broadcast_op_handle_test.h +++ b/paddle/fluid/framework/details/broadcast_op_handle_test.h @@ -42,7 +42,7 @@ struct TestBroadcastOpHandle { std::vector> nodes_; std::vector place_list_; bool use_gpu_; -#ifdef PADDLE_WITH_CUDA +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) std::unique_ptr nccl_ctxs_; #endif @@ -50,7 +50,7 @@ struct TestBroadcastOpHandle { for (size_t j = 0; j < ctxs_.size(); ++j) { ctxs_[j]->Wait(); } -#ifdef PADDLE_WITH_CUDA +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) if (nccl_ctxs_) { nccl_ctxs_->WaitAll(); } @@ -60,7 +60,7 @@ struct TestBroadcastOpHandle { void InitCtxOnGpu(bool use_gpu) { use_gpu_ = use_gpu; if (use_gpu_) { -#ifdef PADDLE_WITH_CUDA +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) int count = p::GetCUDADeviceCount(); if (count <= 1) { LOG(WARNING) << "Cannot test multi-gpu Broadcast, because the CUDA " @@ -84,7 +84,7 @@ struct TestBroadcastOpHandle { place_list_.push_back(p); ctxs_.emplace_back(new p::CPUDeviceContext(p)); } -#ifdef PADDLE_WITH_CUDA +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) nccl_ctxs_.reset(nullptr); #endif } @@ -106,14 +106,14 @@ struct TestBroadcastOpHandle { nodes_.emplace_back( ir::CreateNodeForTest("node0", ir::Node::Type::kOperation)); if (use_gpu_) { -#ifdef PADDLE_WITH_CUDA +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) op_handle_ = new BroadcastOpHandle(nodes_.back().get(), local_scopes_, place_list_, nccl_ctxs_.get()); #else PADDLE_THROW("CUDA is not support."); #endif } else { -#ifdef PADDLE_WITH_CUDA +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) op_handle_ = new BroadcastOpHandle(nodes_.back().get(), local_scopes_, place_list_, nccl_ctxs_.get()); #else diff --git a/paddle/fluid/framework/details/build_strategy.cc b/paddle/fluid/framework/details/build_strategy.cc index 37202f869508c..389366a8a98c5 100644 --- a/paddle/fluid/framework/details/build_strategy.cc +++ b/paddle/fluid/framework/details/build_strategy.cc @@ -14,16 +14,26 @@ limitations under the License. */ #include "paddle/fluid/framework/details/build_strategy.h" +#include +#include + +#include "paddle/fluid/framework/details/memory_reuse_types.h" #include "paddle/fluid/framework/details/multi_devices_graph_check_pass.h" #include "paddle/fluid/framework/details/multi_devices_graph_print_pass.h" +#include "paddle/fluid/framework/details/reduce_op_handle.h" #include "paddle/fluid/framework/details/sequential_execution_pass.h" #include "paddle/fluid/framework/ir/graph.h" +#include "paddle/fluid/framework/ir/graph_helper.h" #include "paddle/fluid/framework/ir/graph_viz_pass.h" namespace paddle { namespace framework { namespace details { +static inline bool SeqOnlyAllReduceOps(const BuildStrategy &strategy) { + return (!strategy.enable_sequential_execution_ && strategy.num_trainers_ > 1); +} + class ParallelExecutorPassBuilder : public ir::PassBuilder { public: explicit ParallelExecutorPassBuilder(const BuildStrategy &strategy) @@ -53,16 +63,40 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder { } } + CollectiveContext *context = CollectiveContext::GetInstance(); + context->endpoints_ = strategy_.trainers_endpoints_; + context->trainer_id_ = strategy_.trainer_id_; + PADDLE_ENFORCE(strategy_.trainer_id_ >= 0, "trainer_id_ >= 0"); + if (strategy_.trainer_id_ > 0) { + PADDLE_ENFORCE((unsigned)(strategy_.trainer_id_) < + strategy_.trainers_endpoints_.size(), + "trainer_id_ < endpoints_ size"); + } + VLOG(1) << "CollectiveContext:" << context->String(); + + // NOTE(dzh): memory optimize should be a runtime pass. + // However, after multi_devices_pass, VarHandle, OpHandle is + // the de-fact IR, any reuse on Graph is meaningless. + // A side-effect of that, memory optimize cannot forsee the fetched vars + // , so fetchlist should be set persistable before call the Run interface. + if (strategy.memory_optimize_) { + auto analysis_var_pass = AppendPass("analysis_var_pass"); + } // Convert graph to run on multi-devices. auto multi_devices_pass = AppendPass("multi_devices_pass"); multi_devices_pass->SetNotOwned("strategy", &strategy_); + multi_devices_pass->Set("num_trainers", + new int(strategy_.num_trainers_)); // Add a graph print pass to record a graph with device info. if (!strategy_.debug_graphviz_path_.empty()) { auto multi_devices_print_pass = AppendPass("multi_devices_print_pass"); - multi_devices_print_pass->SetNotOwned( - "debug_graphviz_path", &strategy_.debug_graphviz_path_); + const std::string graph_path = + string::Sprintf("%s%s", strategy_.debug_graphviz_path_.c_str(), + "_multi_devices_graph"); + multi_devices_print_pass->Set(kGraphvizPath, + new std::string(graph_path)); multi_devices_print_pass->Set( "graph_printer", new details::GraphvizSSAGraphPrinter); } @@ -70,6 +104,10 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder { // Verify that the graph is correct for multi-device executor. AppendPass("multi_devices_check_pass"); + if (SeqOnlyAllReduceOps(strategy)) { + AppendPass("all_reduce_deps_pass"); + } + if (strategy_.remove_unnecessary_lock_) { AppendPass("modify_op_lock_and_record_event_pass"); } @@ -93,10 +131,8 @@ std::shared_ptr BuildStrategy::CreatePassesFromStrategy( std::unique_ptr BuildStrategy::Apply( const ProgramDesc &main_program, const std::vector &places, - const std::string &loss_var_name, - const std::unordered_set ¶m_names, - const std::vector &local_scopes, -#ifdef PADDLE_WITH_CUDA + const std::string &loss_var_name, const std::vector &local_scopes, +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) const bool use_cuda, platform::NCCLContextMap *nccl_ctxs) const { #else const bool use_cuda) const { @@ -105,25 +141,43 @@ std::unique_ptr BuildStrategy::Apply( CreatePassesFromStrategy(false); std::unique_ptr graph(new ir::Graph(main_program)); - for (std::shared_ptr &pass : pass_builder_->AllPasses()) { if (pass->Type() == "multi_devices_pass") { pass->Erase("places"); pass->SetNotOwned>("places", &places); pass->Erase("loss_var_name"); pass->SetNotOwned("loss_var_name", &loss_var_name); - pass->Erase("params"); - pass->SetNotOwned>("params", - ¶m_names); pass->Erase("local_scopes"); pass->SetNotOwned>("local_scopes", &local_scopes); -#ifdef PADDLE_WITH_CUDA +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) platform::NCCLContextMap *nctx = use_cuda ? nccl_ctxs : nullptr; pass->Erase("nccl_ctxs"); pass->SetNotOwned("nccl_ctxs", nctx); #endif + } else if (pass->Type() == "analysis_var_pass") { + const std::vector *all_op_descs = + new std::vector(main_program.Block(0).AllOps()); + graph->Set>(kAllOpDescs, + all_op_descs); // take ownership + graph->Set(kGraphNodePool, + new GraphNodePool); // take ownership + + pass->Erase(kAllOpDescs); + pass->SetNotOwned>(kAllOpDescs, all_op_descs); + } else if (pass->Type() == "sequential_execution_pass") { + LOG(INFO) << "set enable_sequential_execution:" + << enable_sequential_execution_; + + pass->Erase(kAllOpDescs); + pass->Set>( + kAllOpDescs, + new std::vector(main_program.Block(0).AllOps())); + } else if (pass->Type() == "all_reduce_deps_pass") { + LOG(INFO) << "SeqOnlyAllReduceOps:" << SeqOnlyAllReduceOps(*this) + << ", num_trainers:" << num_trainers_; + pass->Erase(kAllOpDescs); pass->Set>( kAllOpDescs, @@ -133,6 +187,7 @@ std::unique_ptr BuildStrategy::Apply( } return graph; } + } // namespace details } // namespace framework } // namespace paddle @@ -143,5 +198,7 @@ USE_PASS(multi_batch_merge_pass); USE_PASS(multi_devices_pass); USE_PASS(multi_devices_check_pass); USE_PASS(multi_devices_print_pass); +USE_PASS(analysis_var_pass); USE_PASS(sequential_execution_pass); +USE_PASS(all_reduce_deps_pass); USE_PASS(modify_op_lock_and_record_event_pass); diff --git a/paddle/fluid/framework/details/build_strategy.h b/paddle/fluid/framework/details/build_strategy.h index fc2641dbd4827..11db184cb4efe 100644 --- a/paddle/fluid/framework/details/build_strategy.h +++ b/paddle/fluid/framework/details/build_strategy.h @@ -23,7 +23,7 @@ #include "paddle/fluid/platform/device_context.h" #include "paddle/fluid/platform/enforce.h" -#ifdef PADDLE_WITH_CUDA +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) #include "paddle/fluid/platform/nccl_helper.h" #endif @@ -60,8 +60,15 @@ struct BuildStrategy { kCustomized = 2, }; + enum class OptimizeStrategy { + // To be Implemented,bruteforce, recursive compute unused var names. + kBruteForce = 0, + kControlFlowGraph = 1, // use cfg_graph algorithm, faster speed. + }; + ReduceStrategy reduce_{ReduceStrategy::kAllReduce}; GradientScaleStrategy gradient_scale_{GradientScaleStrategy::kCoeffNumDevice}; + OptimizeStrategy strategy_{OptimizeStrategy::kControlFlowGraph}; std::string debug_graphviz_path_{""}; @@ -69,10 +76,17 @@ struct BuildStrategy { bool enable_data_balance_{false}; + bool memory_optimize_{false}; + + bool memory_early_delete_{false}; + bool enable_sequential_execution_{false}; bool fuse_broadcast_op_{false}; + int num_trainers_{1}; + int trainer_id_{0}; + std::vector trainers_endpoints_; bool remove_unnecessary_lock_{false}; // NOTE: @@ -92,16 +106,15 @@ struct BuildStrategy { // Apply the passes built by the pass_builder_. The passes will be // applied to the Program and output an ir::Graph. - std::unique_ptr Apply( - const ProgramDesc &main_program, - const std::vector &places, - const std::string &loss_var_name, - const std::unordered_set ¶m_names, - const std::vector &local_scopes, -#ifdef PADDLE_WITH_CUDA - const bool use_cuda, platform::NCCLContextMap *nccl_ctxs) const; + std::unique_ptr Apply(const ProgramDesc &main_program, + const std::vector &places, + const std::string &loss_var_name, + const std::vector &local_scopes, +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) + const bool use_cuda, + platform::NCCLContextMap *nccl_ctxs) const; #else - const bool use_cuda) const; + const bool use_cuda) const; #endif private: diff --git a/paddle/fluid/framework/details/computation_op_handle.cc b/paddle/fluid/framework/details/computation_op_handle.cc index 7ad1e40c600c6..7beb8c8de9fc4 100644 --- a/paddle/fluid/framework/details/computation_op_handle.cc +++ b/paddle/fluid/framework/details/computation_op_handle.cc @@ -20,11 +20,13 @@ namespace paddle { namespace framework { namespace details { ComputationOpHandle::ComputationOpHandle(ir::Node *node, Scope *scope, - platform::Place place) + platform::Place place, + size_t scope_idx) : OpHandleBase(node), op_(framework::OpRegistry::CreateOp(*node->Op())), scope_(scope), - place_(place) {} + place_(place), + scope_idx_(scope_idx) {} void ComputationOpHandle::RunImpl() { WaitInputVarGenerated(place_); diff --git a/paddle/fluid/framework/details/computation_op_handle.h b/paddle/fluid/framework/details/computation_op_handle.h index 662a91d6b4dfc..601ae4f8c6de1 100644 --- a/paddle/fluid/framework/details/computation_op_handle.h +++ b/paddle/fluid/framework/details/computation_op_handle.h @@ -28,7 +28,8 @@ namespace framework { namespace details { struct ComputationOpHandle : public OpHandleBase { public: - ComputationOpHandle(ir::Node *node, Scope *scope, platform::Place place); + ComputationOpHandle(ir::Node *node, Scope *scope, platform::Place place, + size_t scope_idx); std::string Name() const override; @@ -38,6 +39,8 @@ struct ComputationOpHandle : public OpHandleBase { void SetLockAndRecordEventFree(bool b) { is_lock_and_record_event_free_ = b; } + size_t GetScopeIdx() const { return scope_idx_; } + protected: void RunImpl() override; @@ -47,6 +50,7 @@ struct ComputationOpHandle : public OpHandleBase { std::unique_ptr op_; Scope *scope_; platform::Place place_; + size_t scope_idx_; bool is_lock_and_record_event_free_{false}; }; } // namespace details diff --git a/paddle/fluid/framework/details/data_balance_op_handle.cc b/paddle/fluid/framework/details/data_balance_op_handle.cc index 0b772f9b63e2c..cc562c7b102ce 100644 --- a/paddle/fluid/framework/details/data_balance_op_handle.cc +++ b/paddle/fluid/framework/details/data_balance_op_handle.cc @@ -20,7 +20,7 @@ namespace paddle { namespace framework { namespace details { -#ifdef PADDLE_WITH_CUDA +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) DataBalanceOpHandle::DataBalanceOpHandle( ir::Node *node, const std::vector &local_scopes, const std::vector &places, diff --git a/paddle/fluid/framework/details/data_balance_op_handle.h b/paddle/fluid/framework/details/data_balance_op_handle.h index 0462fb6ec713e..2db18a1a7203f 100644 --- a/paddle/fluid/framework/details/data_balance_op_handle.h +++ b/paddle/fluid/framework/details/data_balance_op_handle.h @@ -19,7 +19,7 @@ #include "paddle/fluid/framework/details/op_handle_base.h" #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/scope.h" -#ifdef PADDLE_WITH_CUDA +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) #include "paddle/fluid/platform/nccl_helper.h" #endif @@ -29,7 +29,7 @@ namespace details { struct DataBalanceOpHandle : public OpHandleBase { public: -#ifdef PADDLE_WITH_CUDA +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) DataBalanceOpHandle(ir::Node *node, const std::vector &local_scopes, const std::vector &places, const platform::NCCLContextMap *ctxs); diff --git a/paddle/fluid/framework/details/eager_deletion_op_handle.cc b/paddle/fluid/framework/details/eager_deletion_op_handle.cc new file mode 100644 index 0000000000000..abacb11e3b018 --- /dev/null +++ b/paddle/fluid/framework/details/eager_deletion_op_handle.cc @@ -0,0 +1,122 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/framework/details/eager_deletion_op_handle.h" +#include "paddle/fluid/framework/lod_tensor_array.h" +#include "paddle/fluid/framework/scope.h" +#include "paddle/fluid/framework/selected_rows.h" +#ifdef PADDLE_WITH_CUDA +#include "paddle/fluid/platform/cuda_device_guard.h" +#endif + +namespace paddle { +namespace framework { +namespace details { + +EagerDeletionOpHandle::EagerDeletionOpHandle( + ir::Node *node, const Scope *scope, const platform::Place &place, + const std::unordered_set &var_names, GarbageCollector *gc, + AtomicReferenceCountMap *ref_cnts) + : OpHandleBase(node), + scope_(scope), + var_names_(var_names), + gc_(gc), + ref_cnts_(ref_cnts) { +#ifdef PADDLE_WITH_CUDA + if (platform::is_gpu_place(place)) { + dev_ctx_ = reinterpret_cast( + platform::DeviceContextPool::Instance().Get(place)); + if (dynamic_cast(gc_)) { + platform::CUDADeviceGuard guard( + boost::get(place).device); + PADDLE_ENFORCE(cudaEventCreateWithFlags(&event_, cudaEventDisableTiming)); + PADDLE_ENFORCE_NOT_NULL(event_); + } + } +#endif +} + +EagerDeletionOpHandle::~EagerDeletionOpHandle() { +#ifdef PADDLE_WITH_CUDA + if (event_) { + auto gpu_place = boost::get(dev_ctx_->GetPlace()); + platform::CUDADeviceGuard guard(gpu_place.device); + PADDLE_ENFORCE(cudaEventDestroy(event_)); + } +#endif +} + +std::string EagerDeletionOpHandle::Name() const { return "eager_deletion"; } + +void EagerDeletionOpHandle::RunImpl() { + auto *exec_scope = scope_->FindVar(kLocalExecScopeName)->Get(); + std::deque> garbages; + for (auto &name : var_names_) { + auto it = ref_cnts_->find(name); + // Var not found, not reference count has not decreased to 0 + if (it == ref_cnts_->end() || it->second.fetch_sub(1) != 1) { + continue; + } + + auto *var = exec_scope->FindVar(name); + if (var == nullptr) { + continue; + } + + VLOG(2) << "Erase variable " << name; + + if (var->IsType()) { + garbages.emplace_back(var->GetMutable()->MoveMemoryHolder()); + } else if (var->IsType()) { + garbages.emplace_back( + var->GetMutable()->mutable_value()->MoveMemoryHolder()); + } else if (var->IsType()) { + auto *tensor_arr = var->GetMutable(); + for (auto &t : *tensor_arr) { + garbages.emplace_back(t.MoveMemoryHolder()); + } + } else { + PADDLE_THROW("Type %s of %s is not supported eager deletion", + var->Type().name(), name); + } + } + + if (!garbages.empty()) { + ClearGarbages(&garbages); + } +} + +void EagerDeletionOpHandle::ClearGarbages( + std::deque> *garbages) { +#ifdef PADDLE_WITH_CUDA + if (event_) { + auto compute_stream = dev_ctx_->stream(); + auto callback_stream = + reinterpret_cast(gc_)->stream(); + auto callback_func = [=]() { + PADDLE_ENFORCE(cudaEventRecord(event_, compute_stream)); + PADDLE_ENFORCE(cudaStreamWaitEvent(callback_stream, event_, 0)); + }; + gc_->Add(std::move(*garbages), callback_func); + } else { +#endif + gc_->Add(std::move(*garbages)); +#ifdef PADDLE_WITH_CUDA + } +#endif +} + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/eager_deletion_op_handle.h b/paddle/fluid/framework/details/eager_deletion_op_handle.h new file mode 100644 index 0000000000000..64867afad5b70 --- /dev/null +++ b/paddle/fluid/framework/details/eager_deletion_op_handle.h @@ -0,0 +1,58 @@ +// 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 +#include +#include "paddle/fluid/framework/details/op_handle_base.h" +#include "paddle/fluid/framework/details/reference_count_pass_helper.h" + +namespace paddle { +namespace framework { +class Scope; + +namespace details { + +class EagerDeletionOpHandle : public OpHandleBase { + public: + EagerDeletionOpHandle(ir::Node *node, const Scope *scope, + const platform::Place &place, + const std::unordered_set &var_names, + GarbageCollector *gc, + AtomicReferenceCountMap *ref_cnts); + + ~EagerDeletionOpHandle(); + + std::string Name() const override; + + protected: + void RunImpl() override; + + private: + void ClearGarbages(std::deque> *garbages); + + const Scope *scope_; + std::unordered_set var_names_; + GarbageCollector *gc_; // not own + AtomicReferenceCountMap *ref_cnts_; // not own +#ifdef PADDLE_WITH_CUDA + platform::CUDADeviceContext *dev_ctx_{nullptr}; + cudaEvent_t event_{nullptr}; +#endif +}; + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/eager_deletion_pass.cc b/paddle/fluid/framework/details/eager_deletion_pass.cc new file mode 100644 index 0000000000000..4e42d0b4972d5 --- /dev/null +++ b/paddle/fluid/framework/details/eager_deletion_pass.cc @@ -0,0 +1,101 @@ +// 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 +#include +#include + +#include "paddle/fluid/framework/details/computation_op_handle.h" +#include "paddle/fluid/framework/details/eager_deletion_op_handle.h" +#include "paddle/fluid/framework/details/eager_deletion_pass.h" +#include "paddle/fluid/framework/details/multi_devices_helper.h" +#include "paddle/fluid/framework/ir/graph_helper.h" + +namespace paddle { +namespace framework { +namespace details { + +std::unique_ptr EagerDeletionPass::ApplyImpl( + std::unique_ptr graph) const { + auto &ref_cnts = + Get>(kRuntimeReferenceCount); + PADDLE_ENFORCE(ref_cnts.empty(), + "kRuntimeReferenceCount should be initialized here!"); + + const auto &vars = graph->Get(kGraphVars); + ref_cnts.resize(vars.size()); + + const auto &last_live_ops = + Get>(kLastLiveOpsOfVars); + const auto &gcs = Get(kGarbageCollector); + const auto &places = Get>(kAllPlaces); + + // a reverse map of last_live_ops + // i.e., last op --> variable names which can be deleted. + std::unordered_map> + op_vars_map; + + for (auto &var_ops_map : last_live_ops) { + for (auto &var_ops_pair : var_ops_map) { + const std::string &var_name = var_ops_pair.first; + for (auto *op : var_ops_pair.second) { + op_vars_map[op].insert(var_name); + } + } + } + + for (auto &pair : op_vars_map) { + auto *op = pair.first; + auto &var_names = pair.second; + + auto *eager_deletion_node = + graph->CreateEmptyNode("eager_deletion", ir::Node::Type::kOperation); + auto *eager_deletion_op = new EagerDeletionOpHandle( + eager_deletion_node, op->GetScope(), op->GetPlace(), var_names, + gcs.at(places[op->GetScopeIdx()]).get(), + &(ref_cnts[op->GetScopeIdx()])); + + auto it = std::find_if( + op->Outputs().begin(), op->Outputs().end(), [](VarHandleBase *var) { + return dynamic_cast(var) != nullptr; + }); + + if (it != op->Outputs().end()) { + eager_deletion_op->AddInput(*it); + } else { + auto *dep_var = new DummyVarHandle(graph->CreateControlDepVar()); + graph->Get(kGraphDepVars).emplace(dep_var); + op->AddOutput(dep_var); + eager_deletion_op->AddInput(dep_var); + } + + auto *dummy_leaf = new DummyVarHandle(graph->CreateControlDepVar()); + graph->Get(kGraphDepVars).emplace(dummy_leaf); + eager_deletion_op->AddOutput(dummy_leaf); + } + + VLOG(10) << "Create " << op_vars_map.size() << " EagerDeletionOpHandle(s)"; + return graph; +} + +} // namespace details +} // namespace framework +} // namespace paddle + +REGISTER_PASS(eager_deletion_pass, + paddle::framework::details::EagerDeletionPass) + .RequirePassAttr(paddle::framework::details::kRuntimeReferenceCount) + .RequirePassAttr(paddle::framework::details::kLastLiveOpsOfVars) + .RequirePassAttr(paddle::framework::details::kAllPlaces) + .RequirePassAttr(paddle::framework::details::kGarbageCollector); diff --git a/paddle/fluid/framework/details/eager_deletion_pass.h b/paddle/fluid/framework/details/eager_deletion_pass.h new file mode 100644 index 0000000000000..d7a7a9709d970 --- /dev/null +++ b/paddle/fluid/framework/details/eager_deletion_pass.h @@ -0,0 +1,32 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include "paddle/fluid/framework/ir/graph.h" +#include "paddle/fluid/framework/ir/pass.h" + +namespace paddle { +namespace framework { +namespace details { + +class EagerDeletionPass : public ir::Pass { + protected: + std::unique_ptr ApplyImpl( + std::unique_ptr graph) const override; +}; + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/early_delete_op_handle.h b/paddle/fluid/framework/details/early_delete_op_handle.h new file mode 100644 index 0000000000000..c8382d34b790b --- /dev/null +++ b/paddle/fluid/framework/details/early_delete_op_handle.h @@ -0,0 +1,140 @@ +// 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 +#include +#include "paddle/fluid/framework/details/computation_op_handle.h" +#include "paddle/fluid/framework/details/op_handle_base.h" +#include "paddle/fluid/framework/details/var_handle.h" +#include "paddle/fluid/framework/garbage_collector.h" +#include "paddle/fluid/framework/lod_tensor_array.h" +#include "paddle/fluid/framework/scope.h" +#include "paddle/fluid/framework/selected_rows.h" +#include "paddle/fluid/framework/tensor.h" + +namespace paddle { +namespace framework { +namespace details { + +class EarlyDeleteOpHandle : public OpHandleBase { + public: + EarlyDeleteOpHandle(ir::Node* node, const Scope* scope, + const platform::Place& place, + const std::vector& names, + GarbageCollector* gc) + : OpHandleBase(node), + scope_(scope), + place_(place), + names_(names), + gc_(gc) { +#ifdef PADDLE_WITH_CUDA + if (IsStreamGarabageCollector()) { + auto gpu_place = boost::get(place); + PADDLE_ENFORCE(cudaSetDevice(gpu_place.device)); + PADDLE_ENFORCE(cudaEventCreateWithFlags(&event_, cudaEventDisableTiming)); + } +#endif + } + ~EarlyDeleteOpHandle() { +#ifdef PADDLE_WITH_CUDA + if (IsStreamGarabageCollector()) { + auto gpu_place = boost::get(dev_ctx_->GetPlace()); + PADDLE_ENFORCE(cudaSetDevice(gpu_place.device)); + PADDLE_ENFORCE(cudaEventDestroy(event_)); + } +#endif + } + + std::string Name() const override { return "early_delete"; } + + protected: + void RunImpl() override { + std::vector> tensors; + auto* local_scope = scope_->FindVar(kLocalExecScopeName)->Get(); + for (auto& var_name : names_) { + auto* var = local_scope->FindVar(var_name); + PADDLE_ENFORCE(var != nullptr, + string::Sprintf("Local Scope not has var %s", var_name)); + if (var->IsType()) { + tensors.emplace_back(var->GetMutable()->MoveMemoryHolder()); + } else if (var->IsType()) { + tensors.emplace_back(var->GetMutable() + ->mutable_value() + ->MoveMemoryHolder()); + } else if (var->IsType()) { + LoDTensorArray* tensor_array = var->GetMutable(); + for (auto& tensor : *tensor_array) { + tensors.emplace_back(tensor.MoveMemoryHolder()); + } + } + } + if (!tensors.empty()) { + ClearTensors(tensors); + } + } + + private: + void ClearTensors( + const std::vector>& tensors) { + if (platform::is_cpu_place(place_)) { + ClearCPUTensors(tensors); + } else { + ClearGPUTensors(tensors); + } + } + + void ClearCPUTensors( + const std::vector>& tensors) { + auto* gc = dynamic_cast(gc_); + if (gc != nullptr) { + gc->Add(tensors); + } + } + + void ClearGPUTensors( + const std::vector>& tensors) { +#ifdef PADDLE_WITH_CUDA + auto* gc = dynamic_cast(gc_); + if (gc != nullptr) { + auto compute_stream = dev_ctx_->stream(); + auto callback_stream = gc->stream(); + auto callback_func = [=]() { + PADDLE_ENFORCE(cudaEventRecord(event_, compute_stream)); + PADDLE_ENFORCE(cudaStreamWaitEvent(callback_stream, event_, 0)); + }; + gc_->Add(tensors, callback_func); + } else { + gc_->Add(tensors); + } + } + + bool IsStreamGarabageCollector() const { + return dynamic_cast(gc_) != nullptr; +#endif + } + + const Scope* scope_; + const platform::Place place_; + std::vector names_; + GarbageCollector* gc_; +#ifdef PADDLE_WITH_CUDA + platform::CUDADeviceContext* dev_ctx_; + cudaEvent_t event_; +#endif +}; + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/fuse_vars_op_handle.h b/paddle/fluid/framework/details/fuse_vars_op_handle.h index 3f360c510a4fd..b40b01df36479 100644 --- a/paddle/fluid/framework/details/fuse_vars_op_handle.h +++ b/paddle/fluid/framework/details/fuse_vars_op_handle.h @@ -33,7 +33,7 @@ struct FuseVarsOpHandle : public OpHandleBase { FuseVarsOpHandle(ir::Node *node, Scope *local_scope, const platform::Place &place, const std::unordered_map &inputs_numel, - const std::type_index &var_type) + const proto::VarType::Type var_type) : OpHandleBase(node), local_scope_(local_scope), place_(place), @@ -57,7 +57,7 @@ struct FuseVarsOpHandle : public OpHandleBase { Scope *local_scope_; const platform::Place place_; const std::unordered_map inputs_numel_; - const std::type_index type_; + const proto::VarType::Type type_; int64_t total_numel_; }; } // namespace details diff --git a/paddle/fluid/framework/details/fused_broadcast_op_handle.h b/paddle/fluid/framework/details/fused_broadcast_op_handle.h index e37259526a5f6..e43d545c9c0d0 100644 --- a/paddle/fluid/framework/details/fused_broadcast_op_handle.h +++ b/paddle/fluid/framework/details/fused_broadcast_op_handle.h @@ -25,7 +25,7 @@ #include "paddle/fluid/framework/selected_rows.h" #include "paddle/fluid/platform/device_context.h" -#ifdef PADDLE_WITH_CUDA +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) #include "paddle/fluid/platform/nccl_helper.h" #endif @@ -35,7 +35,7 @@ namespace details { struct FusedBroadcastOpHandle : public BroadcastOpHandle { public: -#ifdef PADDLE_WITH_CUDA +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) FusedBroadcastOpHandle(ir::Node *node, const std::vector local_scopes, const std::vector &places, diff --git a/paddle/fluid/framework/details/fused_broadcast_op_handle_test.cc b/paddle/fluid/framework/details/fused_broadcast_op_handle_test.cc index 541993c74332c..be0d941c4f9c2 100644 --- a/paddle/fluid/framework/details/fused_broadcast_op_handle_test.cc +++ b/paddle/fluid/framework/details/fused_broadcast_op_handle_test.cc @@ -44,14 +44,14 @@ struct TestFusedBroadcastOpHandle : TestBroadcastOpHandle { nodes_.emplace_back( ir::CreateNodeForTest("fused_broadcast", ir::Node::Type::kOperation)); if (use_gpu_) { -#ifdef PADDLE_WITH_CUDA +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) op_handle_ = new FusedBroadcastOpHandle( nodes_.back().get(), local_scopes_, place_list_, nccl_ctxs_.get()); #else PADDLE_THROW("CUDA is not supported."); #endif } else { -#ifdef PADDLE_WITH_CUDA +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) op_handle_ = new FusedBroadcastOpHandle( nodes_.back().get(), local_scopes_, place_list_, nccl_ctxs_.get()); #else diff --git a/paddle/fluid/framework/details/memory_early_delete_pass.cc b/paddle/fluid/framework/details/memory_early_delete_pass.cc new file mode 100644 index 0000000000000..06a2451c136e3 --- /dev/null +++ b/paddle/fluid/framework/details/memory_early_delete_pass.cc @@ -0,0 +1,117 @@ +// 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 +#include +#include +#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 queue; + queue.push(var_in); + do { + auto* var = queue.front(); + queue.pop(); + for (auto* op : var->PendingOps()) { + auto* compute_op = dynamic_cast(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 MemoryEarlyDeletePass::ApplyImpl( + std::unique_ptr graph) const { + auto& graph_pool = Get(kGraphNodePool); + auto& gcs = Get(kGarbageCollector); + + std::unordered_map> 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& vars) { + if (unlived_vars.empty()) return; + // unlived vars can be deleted after the last used op has finished. + auto* compute_op = dynamic_cast(op); + const auto& places = Get>(kAllPlaces); + for (auto& var : vars) { + auto* var_handle = dynamic_cast(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(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(*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); diff --git a/paddle/fluid/framework/details/memory_early_delete_pass.h b/paddle/fluid/framework/details/memory_early_delete_pass.h new file mode 100644 index 0000000000000..8215aa1b2baa2 --- /dev/null +++ b/paddle/fluid/framework/details/memory_early_delete_pass.h @@ -0,0 +1,32 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once +#include "paddle/fluid/framework/details/early_delete_op_handle.h" +#include "paddle/fluid/framework/ir/graph.h" +#include "paddle/fluid/framework/ir/pass.h" + +namespace paddle { +namespace framework { +namespace details { + +class MemoryEarlyDeletePass : public ir::Pass { + protected: + std::unique_ptr ApplyImpl( + std::unique_ptr graph) const override; +}; + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/memory_reuse_types.cc b/paddle/fluid/framework/details/memory_reuse_types.cc new file mode 100644 index 0000000000000..2b9ff518b9adc --- /dev/null +++ b/paddle/fluid/framework/details/memory_reuse_types.cc @@ -0,0 +1,155 @@ +// 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 +#include +#include + +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(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{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 diff --git a/paddle/fluid/framework/details/memory_reuse_types.h b/paddle/fluid/framework/details/memory_reuse_types.h new file mode 100644 index 0000000000000..9a9c1d948e869 --- /dev/null +++ b/paddle/fluid/framework/details/memory_reuse_types.h @@ -0,0 +1,87 @@ +// 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 +#include +#include +#include +#include +#include +#include +#include "paddle/fluid/framework/data_type.h" +#include "paddle/fluid/framework/ir/graph.h" + +namespace paddle { +namespace framework { +namespace details { + +constexpr char kFetchedVars[] = "fetched_vars"; +constexpr char kGraphNodePool[] = "graph_node_pool"; + +// NOTE(dzh): Variable and the operators use the var. +// for early delete pass. +// Because analysis var pass build base on ir::Node, which maybe released +// or modified between passes, so we use OpDesc* to mark ops. +using GraphNodePool = std::vector< + std::pair /* ops */>>; + +// NOTE(dzh): by default, it sort node in ascend order(by node bytes size). +// in fluid, -1 means the batch_size is determined in runtime. +// the node batch_size equal -1 always ranking in the front than the node not. +// For example, +// node0[-1, 1] node1[-1, 1, 1], node2[1,1], node3[1,1024], .. +// O(1) insert, delete +class OrderedNodePairPool { + public: + using NodePair = std::pair>; + using Iter = typename std::list::iterator; + using ConstIter = typename std::list::const_iterator; + + void Insert(ir::Node* var, ir::Node* op); + + void Erase(ir::Node* var); + + bool Has(ir::Node* var) { return mark_table_.count(var->Name()); } + + ir::Node* NodeMatch(ir::Node* var) const; + // map store non-const iterator, can not promise const + int GetIndex(ir::Node* var); + // pool all node to string + std::string ToString() const; + + Iter begin() { return nodes_.begin(); } + Iter end() { return nodes_.end(); } + ConstIter begin() const { return nodes_.begin(); } + ConstIter end() const { return nodes_.end(); } + size_t size() const { return nodes_.size(); } + + private: + // for searching. + std::unordered_map mark_table_; + // node swap pairs. var -> ops dep var + std::list nodes_; +}; + +// node memory size in bytes +size_t NodeSizeInBytes(ir::Node* n); + +std::string DebugString(ir::Node* var); + +// std::string DebugString(VarDesc* var); +VarDesc* FindVarDescInBlock(ir::Node* n); + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/memory_reuse_types_test.cc b/paddle/fluid/framework/details/memory_reuse_types_test.cc new file mode 100644 index 0000000000000..d2fabf5ce068e --- /dev/null +++ b/paddle/fluid/framework/details/memory_reuse_types_test.cc @@ -0,0 +1,99 @@ +// 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 +#include +#include +#include +#include +#include +#include +#include "glog/logging.h" +#include "gtest/gtest.h" + +namespace paddle { +namespace framework { +namespace details { + +TEST(OrderedNodePairPool, Normal) { + OrderedNodePairPool pool; + std::vector> nodes; + + // clang-format off + std::vector> 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 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 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(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 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 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 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 diff --git a/paddle/fluid/framework/details/modify_op_lock_and_record_event_pass.cc b/paddle/fluid/framework/details/modify_op_lock_and_record_event_pass.cc index bf3f3637b551a..67aad9f94f088 100644 --- a/paddle/fluid/framework/details/modify_op_lock_and_record_event_pass.cc +++ b/paddle/fluid/framework/details/modify_op_lock_and_record_event_pass.cc @@ -45,8 +45,8 @@ std::unique_ptr ModifyOpLockAndRecordEventPass::ApplyImpl( IsLockAndRecordEventFreeComputationOpHandle(compute_op, graph_view); compute_op->SetLockAndRecordEventFree(is_lock_and_record_event_free); if (is_lock_and_record_event_free) { - VLOG(100) << "Set is_lock_and_record_event_free be true in op " - << compute_op->DebugString(); + VLOG(10) << "Set is_lock_and_record_event_free be true in op " + << compute_op->DebugString(); } } return ir_graph; diff --git a/paddle/fluid/framework/details/multi_devices_graph_pass.cc b/paddle/fluid/framework/details/multi_devices_graph_pass.cc index 8c98b781301e8..036cef1daaae4 100644 --- a/paddle/fluid/framework/details/multi_devices_graph_pass.cc +++ b/paddle/fluid/framework/details/multi_devices_graph_pass.cc @@ -130,9 +130,9 @@ void AddOutputToLeafOps(ir::Graph *graph) { static const char kLossVarName[] = "loss_var_name"; static const char kPlaces[] = "places"; -static const char kParams[] = "params"; static const char kLocalScopes[] = "local_scopes"; static const char kStrategy[] = "strategy"; +static const char kNumTrainers[] = "num_trainers"; void MultiDevSSAGraphBuilder::Init() const { all_vars_.clear(); @@ -142,13 +142,10 @@ void MultiDevSSAGraphBuilder::Init() const { places_ = Get>(kPlaces); local_scopes_ = Get>(kLocalScopes); strategy_ = Get(kStrategy); -#ifdef PADDLE_WITH_CUDA +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) nccl_ctxs_ = &Get("nccl_ctxs"); #endif - for (auto &p : Get>(kParams)) { - grad_names_.insert(GradVarName(p)); - } balance_vars_.resize(places_.size(), 0); if (strategy_.enable_data_balance_ && places_.size() == 1) { LOG(WARNING) << "It is no need to enable data balance when there is only " @@ -299,6 +296,8 @@ std::unique_ptr MultiDevSSAGraphBuilder::ApplyImpl( auto nodes = graph->ReleaseNodes(); ir::Graph &result = *graph; + int num_trainers = Get(kNumTrainers); + for (auto &node : nodes) { if (node->IsVar() && node->Var()) { all_vars_.emplace(node->Name(), node->Var()); @@ -383,7 +382,7 @@ std::unique_ptr MultiDevSSAGraphBuilder::ApplyImpl( CreateComputationalOps(&result, node, places_.size()); } - if (!is_forwarding && places_.size() > 1) { + if (!is_forwarding && (places_.size() > 1 || num_trainers > 1)) { // Currently, we assume that once gradient is generated, it can be // broadcast, and each gradient is only broadcast once. if (static_cast(boost::get(node->Op()->GetAttr( @@ -399,7 +398,7 @@ std::unique_ptr MultiDevSSAGraphBuilder::ApplyImpl( for (size_t i = 0; i < backward_vars.size(); i += 2) { auto &p_name = backward_vars[i]; auto &g_name = backward_vars[i + 1]; - VLOG(100) << "Bcast " << g_name << " for parameter " << p_name; + VLOG(10) << "Bcast " << g_name << " for parameter " << p_name; switch (strategy_.reduce_) { case BuildStrategy::ReduceStrategy::kReduce: @@ -431,7 +430,7 @@ std::unique_ptr MultiDevSSAGraphBuilder::ApplyImpl( } } bool use_gpu = false; -#ifdef PADDLE_WITH_CUDA +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) use_gpu = nccl_ctxs_ != nullptr; #endif @@ -478,7 +477,7 @@ bool MultiDevSSAGraphBuilder::IsSparseGradient(const std::string &og) const { void MultiDevSSAGraphBuilder::SetCommunicationContext( OpHandleBase *op_handle, const platform::Place &p) const { -#ifdef PADDLE_WITH_CUDA +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) if (nccl_ctxs_ == nullptr) { op_handle->SetDeviceContext(p, platform::DeviceContextPool::Instance().Get(p)); @@ -492,7 +491,7 @@ void MultiDevSSAGraphBuilder::SetCommunicationContext( void MultiDevSSAGraphBuilder::CreateBroadcastOp(ir::Graph *result, const std::string &p_name, size_t src_dev_id) const { -#ifdef PADDLE_WITH_CUDA +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) auto *op_handle = new BroadcastOpHandle( result->CreateEmptyNode("broadcast", ir::Node::Type::kOperation), local_scopes_, places_, nccl_ctxs_); @@ -522,7 +521,7 @@ void MultiDevSSAGraphBuilder::CreateBroadcastOp(ir::Graph *result, void MultiDevSSAGraphBuilder::CreateFusedBroadcastOp( ir::Graph *result, const std::vector> &bcast_varnames) const { -#ifdef PADDLE_WITH_CUDA +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) auto *op_handle = new FusedBroadcastOpHandle( result->CreateEmptyNode("fused_broadcast", ir::Node::Type::kOperation), local_scopes_, places_, nccl_ctxs_); @@ -562,13 +561,13 @@ void MultiDevSSAGraphBuilder::CreateComputationalOp(ir::Graph *result, int dev_id) const { result->Get(kGraphOps).emplace_back( new ComputationOpHandle(result->CreateOpNode(node->Op()), - local_scopes_[dev_id], places_[dev_id])); + local_scopes_[dev_id], places_[dev_id], dev_id)); CreateOpHandleIOs(result, node, dev_id); } void MultiDevSSAGraphBuilder::InsertAllReduceOp(ir::Graph *result, const std::string &og) const { -#ifdef PADDLE_WITH_CUDA +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) result->Get(kGraphOps).emplace_back(new AllReduceOpHandle( result->CreateEmptyNode("allreduce", ir::Node::Type::kOperation), local_scopes_, places_, nccl_ctxs_)); @@ -597,7 +596,7 @@ void MultiDevSSAGraphBuilder::InsertAllReduceOp(ir::Graph *result, void MultiDevSSAGraphBuilder::InsertDataBalanceOp( ir::Graph *result, const std::vector &datas) const { -#ifdef PADDLE_WITH_CUDA +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) result->Get(kGraphOps).emplace_back(new DataBalanceOpHandle( result->CreateEmptyNode("data_balance", ir::Node::Type::kOperation), local_scopes_, places_, nccl_ctxs_)); @@ -685,8 +684,8 @@ void MultiDevSSAGraphBuilder::CreateComputationalOps(ir::Graph *result, for (size_t scope_idx = 0; scope_idx < num_places; ++scope_idx) { auto p = places_[scope_idx]; auto s = local_scopes_[scope_idx]; - result->Get(kGraphOps).emplace_back( - new ComputationOpHandle(result->CreateOpNode(node->Op()), s, p)); + result->Get(kGraphOps).emplace_back(new ComputationOpHandle( + result->CreateOpNode(node->Op()), s, p, scope_idx)); CreateOpHandleIOs(result, node, scope_idx); } } @@ -694,7 +693,7 @@ void MultiDevSSAGraphBuilder::CreateComputationalOps(ir::Graph *result, VarHandle *MultiDevSSAGraphBuilder::CreateReduceOp(ir::Graph *result, const std::string &og, int dst_dev_id) const { -#ifdef PADDLE_WITH_CUDA +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) result->Get(kGraphOps).emplace_back(new ReduceOpHandle( result->CreateEmptyNode("reduce", ir::Node::Type::kOperation), local_scopes_, places_, nccl_ctxs_)); @@ -809,8 +808,8 @@ int MultiDevSSAGraphBuilder::CreateRPCOp( node->Op()->GetAttr(OpProtoAndCheckerMaker::OpRoleVarAttrName())); PADDLE_ENFORCE_EQ(send_param_grad.size(), 2U); op_dev_id = GetAppropriateDeviceID({send_param_grad[1]}); - VLOG(100) << "send grad " << input_var_names[0] << " origin " - << send_param_grad[1] << " place: " << op_dev_id; + VLOG(10) << "send grad " << input_var_names[0] << " origin " + << send_param_grad[1] << " place: " << op_dev_id; for (auto &varname : input_var_names) { sharded_var_device->emplace(varname, op_dev_id); } @@ -826,9 +825,9 @@ int MultiDevSSAGraphBuilder::CreateRPCOp( if (recv_param_grad.size() == 2U) { op_dev_id = GetVarDeviceID(*result, recv_param_grad[1], *sharded_var_device); - VLOG(100) << "recv param " << recv_param_grad[0] - << " get grad place: " << recv_param_grad[1] - << " place: " << op_dev_id; + VLOG(10) << "recv param " << recv_param_grad[0] + << " get grad place: " << recv_param_grad[1] + << " place: " << op_dev_id; } else { op_dev_id = GetAppropriateDeviceID(output_var_names); } @@ -862,7 +861,7 @@ int MultiDevSSAGraphBuilder::CreateRPCOp( if (node->Op()->Type() == "fetch_barrier") { outvar_dev_id = GetVarDeviceID(*result, output->Name(), *sharded_var_device); - PADDLE_ENFORCE_NE(outvar_dev_id, -1); + PADDLE_ENFORCE_NE(outvar_dev_id, -1, "output name %s", output->Name()); } p = places_[outvar_dev_id]; ir::Node *new_node = nullptr; @@ -893,6 +892,6 @@ REGISTER_PASS(multi_devices_pass, paddle::framework::details::MultiDevSSAGraphBuilder) .RequirePassAttr(paddle::framework::details::kLossVarName) .RequirePassAttr(paddle::framework::details::kPlaces) - .RequirePassAttr(paddle::framework::details::kParams) .RequirePassAttr(paddle::framework::details::kLocalScopes) - .RequirePassAttr(paddle::framework::details::kStrategy); + .RequirePassAttr(paddle::framework::details::kStrategy) + .RequirePassAttr(paddle::framework::details::kNumTrainers); diff --git a/paddle/fluid/framework/details/multi_devices_graph_pass.h b/paddle/fluid/framework/details/multi_devices_graph_pass.h index f3ec2d2941524..0556232aa4754 100644 --- a/paddle/fluid/framework/details/multi_devices_graph_pass.h +++ b/paddle/fluid/framework/details/multi_devices_graph_pass.h @@ -40,7 +40,7 @@ class MultiDevSSAGraphBuilder : public ir::Pass { size_t device_id) const; void Init() const; -#ifdef PADDLE_WITH_CUDA +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) mutable platform::NCCLContextMap *nccl_ctxs_; #endif @@ -102,7 +102,6 @@ class MultiDevSSAGraphBuilder : public ir::Pass { mutable std::string loss_var_name_; mutable std::vector places_; mutable std::vector local_scopes_; - mutable std::unordered_set grad_names_; mutable BuildStrategy strategy_; mutable std::unordered_map all_vars_; diff --git a/paddle/fluid/framework/details/multi_devices_graph_print_pass.cc b/paddle/fluid/framework/details/multi_devices_graph_print_pass.cc index 8f92f0948d7d3..c203073845375 100644 --- a/paddle/fluid/framework/details/multi_devices_graph_print_pass.cc +++ b/paddle/fluid/framework/details/multi_devices_graph_print_pass.cc @@ -85,4 +85,5 @@ void GraphvizSSAGraphPrinter::Print(const ir::Graph &graph, } // namespace paddle REGISTER_PASS(multi_devices_print_pass, - paddle::framework::details::SSAGraghBuilderWithPrinter); + paddle::framework::details::SSAGraghBuilderWithPrinter) + .RequirePassAttr(paddle::framework::details::kGraphvizPath); diff --git a/paddle/fluid/framework/details/multi_devices_graph_print_pass.h b/paddle/fluid/framework/details/multi_devices_graph_print_pass.h index c00685fa1629c..b06c87a5c185c 100644 --- a/paddle/fluid/framework/details/multi_devices_graph_print_pass.h +++ b/paddle/fluid/framework/details/multi_devices_graph_print_pass.h @@ -14,6 +14,7 @@ #pragma once +#include #include #include #include @@ -24,6 +25,8 @@ namespace paddle { namespace framework { namespace details { +constexpr char kGraphvizPath[] = "debug_graphviz_path"; + class SSAGraphPrinter { public: virtual ~SSAGraphPrinter() {} @@ -40,7 +43,7 @@ class SSAGraghBuilderWithPrinter : public ir::Pass { std::unique_ptr ApplyImpl( std::unique_ptr graph) const override { std::unique_ptr fout( - new std::ofstream(Get("debug_graphviz_path"))); + new std::ofstream(Get(kGraphvizPath))); PADDLE_ENFORCE(fout->good()); Get("graph_printer").Print(*graph, *fout); return graph; diff --git a/paddle/fluid/framework/details/op_graph_view.cc b/paddle/fluid/framework/details/op_graph_view.cc index 4838c4198ff35..d3865c2c2919c 100644 --- a/paddle/fluid/framework/details/op_graph_view.cc +++ b/paddle/fluid/framework/details/op_graph_view.cc @@ -23,6 +23,8 @@ namespace details { OpGraphView::OpGraphView(const std::vector &ops) { Build(ops); } void OpGraphView::Build(const std::vector &ops) { + preceding_ops_.clear(); + pending_ops_.clear(); for (auto &op : ops) { preceding_ops_[op]; pending_ops_[op]; @@ -40,6 +42,7 @@ void OpGraphView::Build(const std::vector &ops) { std::unordered_set OpGraphView::AllOps() const { std::unordered_set ret; + ret.reserve(preceding_ops_.size()); for (auto &pair : preceding_ops_) { ret.insert(pair.first); } diff --git a/paddle/fluid/framework/details/op_graph_view.h b/paddle/fluid/framework/details/op_graph_view.h index afb3e8e59461e..77aa02eba56ac 100644 --- a/paddle/fluid/framework/details/op_graph_view.h +++ b/paddle/fluid/framework/details/op_graph_view.h @@ -14,7 +14,7 @@ #pragma once -#include +#include #include #include #include @@ -34,6 +34,11 @@ class OpGraphView { bool HasOp(OpHandleBase *op) const; + // Use a visitor to visit all pending ops of op + // Stop when callback returns false + template + bool VisitAllPendingOps(OpHandleBase *op, Callback &&callback) const; + private: void Build(const std::vector &ops); void EnforceHasOp(OpHandleBase *op) const; @@ -44,6 +49,28 @@ class OpGraphView { pending_ops_; }; +template +bool OpGraphView::VisitAllPendingOps(OpHandleBase *op, + Callback &&callback) const { + EnforceHasOp(op); + std::unordered_set visited; + std::queue q; + q.push(op); + do { + op = q.front(); + q.pop(); + for (auto &pending_op : pending_ops_.at(op)) { + if (visited.count(pending_op) == 0) { + visited.insert(pending_op); + if (!callback(pending_op)) { + return false; + } + } + } + } while (!q.empty()); + return true; +} + } // namespace details } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/details/op_handle_base.h b/paddle/fluid/framework/details/op_handle_base.h index ba12ca3c61c05..b1a82e8771b92 100644 --- a/paddle/fluid/framework/details/op_handle_base.h +++ b/paddle/fluid/framework/details/op_handle_base.h @@ -25,7 +25,7 @@ namespace paddle { namespace framework { namespace details { -constexpr char kLocalExecScopeName[] = "@LCOAL_SCOPE@"; +constexpr char kLocalExecScopeName[] = "@LOCAL_SCOPE@"; // Wraps ir::Node and provide helper utilities. // It's responsible for populating necessary fields of ir::Node. diff --git a/paddle/fluid/framework/details/reduce_and_gather.h b/paddle/fluid/framework/details/reduce_and_gather.h index bd6153c0c736f..2e5256fbd49a3 100644 --- a/paddle/fluid/framework/details/reduce_and_gather.h +++ b/paddle/fluid/framework/details/reduce_and_gather.h @@ -53,7 +53,7 @@ struct ReduceLoDTensor { } }; -inline void GatherSelectedRows( +inline void GatherLocalSelectedRows( const std::vector &src_selecte_rows_, const std::vector &in_places, const std::map &dev_ctxes, diff --git a/paddle/fluid/framework/details/reduce_op_handle.cc b/paddle/fluid/framework/details/reduce_op_handle.cc index 4503123eac810..7a5f7de57ef20 100644 --- a/paddle/fluid/framework/details/reduce_op_handle.cc +++ b/paddle/fluid/framework/details/reduce_op_handle.cc @@ -16,6 +16,12 @@ #include "paddle/fluid/framework/details/container_cast.h" #include "paddle/fluid/framework/details/reduce_and_gather.h" #include "paddle/fluid/framework/details/variable_visitor.h" +#if defined PADDLE_WITH_CUDA && defined PADDLE_WITH_DISTRIBUTE +#include "paddle/fluid/operators/distributed/collective_client.h" +#include "paddle/fluid/operators/distributed/collective_server.h" +#include "paddle/fluid/operators/distributed/request_handler.h" +#endif +#include "paddle/fluid/operators/math/selected_rows_functor.h" #include "paddle/fluid/platform/profiler.h" DEFINE_bool( @@ -26,6 +32,112 @@ namespace paddle { namespace framework { namespace details { +std::once_flag CollectiveContext::init_flag_; +std::unique_ptr CollectiveContext::context_; + +static inline std::string GetRemoteVarName(const std::string &var_name, + int trainer_id) { + return string::Sprintf("%s_merged_tmp@trainer_%d", var_name, trainer_id); +} + +void ReduceOpHandle::Wait( + const std::map &dev_ctxes) { + // TODO(gongwb): use event wait? + for (auto &dev_ctx : dev_ctxes) { + dev_ctx.second->Wait(); + } +} + +#if defined PADDLE_WITH_CUDA && defined PADDLE_WITH_DISTRIBUTE +template +void ReduceOpHandle::GatherSelectedRows( + const std::vector &src_selected_rows, + const std::vector &in_places, + const std::map &dev_ctxes, + VarHandle *out_var_handle, const platform::Place &out_place, + SelectedRows *dst_selected_rows) { + const CollectiveContext &collective_context = + *CollectiveContext::GetInstance(); + + // 1. gather local selected rows, merge them + std::string gathered_var_name = out_var_handle->name_ + "_gathered_tmp"; + auto scope = local_scopes_.at(out_var_handle->scope_idx_); + auto gathered_var_mid = scope->Var(gathered_var_name); + auto gathered_select_rows = + gathered_var_mid->GetMutable(); + GatherLocalSelectedRows(src_selected_rows, in_places, dev_ctxes, out_place, + gathered_select_rows); + // FIXME(gongwb): remove this Wait. + Wait(dev_ctxes); + + // merge them + auto merged_dev_ctx = dynamic_cast(dev_ctxes.at(out_place)); + std::string merged_var_name = + GetRemoteVarName(out_var_handle->name_, collective_context.trainer_id_); + auto merged_select_rows = + scope->Var(merged_var_name)->GetMutable(); + operators::math::scatter::MergeAdd merge_func; + merge_func(*merged_dev_ctx, *gathered_select_rows, merged_select_rows); + + // 2. start collective server if it doesn't exist + operators::distributed::CollectiveServer *server = + operators::distributed::CollectiveServer::GetInstance( + collective_context.endpoints_[collective_context.trainer_id_], + collective_context.endpoints_.size() - 1); + + auto rpc_server = server->GetRPCServer(); + rpc_server->RegisterVar(merged_var_name, + operators::distributed::kRequestGetMonomerVariable, + scope, merged_dev_ctx); + + // 3. gather them from all remote nodes. + std::vector remote; + operators::distributed::CollectiveClient *client = + operators::distributed::CollectiveClient::GetInstance(); + + std::vector vars; + for (unsigned int i = 0; i < collective_context.endpoints_.size(); i++) { + if (i == (unsigned)collective_context.trainer_id_) continue; + + operators::distributed::RemoteVar var; + var.trainer_id_ = i; + var.var_name_ = GetRemoteVarName(out_var_handle->name_, i); + var.ep_ = collective_context.endpoints_[i]; + + vars.push_back(var); + VLOG(4) << "gather from:" << var.String(); + } + + // erase gathered vars + merged_dev_ctx->Wait(); + scope->EraseVars(std::vector{gathered_var_name}); + + PADDLE_ENFORCE(client->Gather(vars, &remote, *merged_dev_ctx, scope)); + PADDLE_ENFORCE(remote.size() == vars.size()); + + // 4. merged local selected rows. + std::vector all; + all.resize(collective_context.endpoints_.size()); + for (auto v : vars) { + all[v.trainer_id_] = + scope->FindVar(v.var_name_)->GetMutable(); + } + all[collective_context.trainer_id_] = merged_select_rows; + + merge_func(*merged_dev_ctx, all, dst_selected_rows); + + rpc_server->WaitVarBarrier(merged_var_name); + rpc_server->ClearVar(merged_var_name); + + // 5. clear mid vars + std::vector tmp_vars{merged_var_name}; + for (auto r : vars) { + tmp_vars.push_back(r.var_name_); + } + scope->EraseVars(tmp_vars); +} +#endif + void ReduceOpHandle::RunImpl() { platform::RecordEvent record_event(Name(), dev_ctxes_.cbegin()->second); @@ -90,8 +202,36 @@ void ReduceOpHandle::RunImpl() { this->RunAndRecordEvent([&] { std::vector in_selected_rows = GetInputValues(in_var_handles, var_scopes); - GatherSelectedRows(in_selected_rows, in_places, dev_ctxes_, t_out_p, - out_var->GetMutable()); + + const CollectiveContext &collective_context = + *CollectiveContext::GetInstance(); + VLOG(10) << "GatherSelectedRows CollectiveContext:" + << collective_context.String(); + + // TODO(gongwb): add cpu support + if (collective_context.endpoints_.size() <= 1 || + is_cpu_place(in_places[0]) || is_cpu_place(t_out_p)) { + GatherLocalSelectedRows(in_selected_rows, in_places, dev_ctxes_, + t_out_p, + out_var->GetMutable()); + return; + } + +#if defined PADDLE_WITH_CUDA && defined PADDLE_WITH_DISTRIBUTE + if (in_selected_rows[0]->value().type() == + framework::proto::VarType::FP32) { + GatherSelectedRows( + in_selected_rows, in_places, dev_ctxes_, out_var_handle, t_out_p, + out_var->GetMutable()); + } else if (in_selected_rows[0]->value().type() == + framework::proto::VarType::FP64) { + GatherSelectedRows( + in_selected_rows, in_places, dev_ctxes_, out_var_handle, t_out_p, + out_var->GetMutable()); + } else { + PADDLE_THROW("only support double or float when gather SelectedRows"); + } +#endif }); } else { std::vector lod_tensors = @@ -106,7 +246,7 @@ void ReduceOpHandle::RunImpl() { if (!FLAGS_cpu_deterministic) { ReduceLoDTensor func(lod_tensors, out_var->GetMutable()); - VisitDataType(ToDataType(lod_tensors[0]->type()), func); + VisitDataType(lod_tensors[0]->type(), func); } else { // We sum lod_tensors to reduce_sum_trg which is in local_scopes_0 // here, but it doesn't mean reduce_sum_trg must be in local_scopes_0. @@ -116,7 +256,7 @@ void ReduceOpHandle::RunImpl() { ->FindVar(out_var_handle->name_) ->GetMutable(); ReduceLoDTensor func(lod_tensors, &reduce_sum_trg); - VisitDataType(ToDataType(lod_tensors[0]->type()), func); + VisitDataType(lod_tensors[0]->type(), func); auto trg = out_var->GetMutable(); if (reduce_sum_trg.data() != trg->data()) { @@ -125,7 +265,7 @@ void ReduceOpHandle::RunImpl() { } }); } else if (paddle::platform::is_gpu_place(lod_tensors[0]->place())) { -#ifdef PADDLE_WITH_CUDA +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) auto pre_in = pre_in_var->Get(); VariableVisitor::ShareDimsAndLoD(*pre_in_var, out_var); VariableVisitor::GetMutableTensor(out_var).mutable_data( diff --git a/paddle/fluid/framework/details/reduce_op_handle.h b/paddle/fluid/framework/details/reduce_op_handle.h index 999828ae457ba..5491f00f45e9d 100644 --- a/paddle/fluid/framework/details/reduce_op_handle.h +++ b/paddle/fluid/framework/details/reduce_op_handle.h @@ -23,19 +23,45 @@ #include "paddle/fluid/framework/scope.h" #include "paddle/fluid/framework/selected_rows.h" #include "paddle/fluid/platform/device_context.h" -#ifdef PADDLE_WITH_CUDA +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) #include "paddle/fluid/platform/nccl_helper.h" #endif namespace paddle { namespace framework { namespace details { +struct CollectiveContext { + std::vector endpoints_; + int trainer_id_{0}; + + std::string String() const { + std::stringstream ss; + ss << "endpoints_:"; + for (auto e : endpoints_) { + ss << e << ","; + } + + ss << "trainer_id_:" << trainer_id_; + + return ss.str(); + } + + static CollectiveContext *GetInstance() { + std::call_once(init_flag_, + [&]() { context_.reset(new CollectiveContext()); }); + return context_.get(); + } + + private: + static std::once_flag init_flag_; + static std::unique_ptr context_; +}; struct ReduceOpHandle : public OpHandleBase { std::vector local_scopes_; std::vector places_; -#ifdef PADDLE_WITH_CUDA +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) const platform::NCCLContextMap *nccl_ctxs_; ReduceOpHandle(ir::Node *node, const std::vector &local_scopes, const std::vector &places, @@ -64,6 +90,19 @@ struct ReduceOpHandle : public OpHandleBase { protected: void RunImpl() override; +#if defined PADDLE_WITH_CUDA && defined PADDLE_WITH_DISTRIBUTE + template + void GatherSelectedRows( + const std::vector &src_selecte_rows_, + const std::vector &in_places, + const std::map &dev_ctxes, + VarHandle *out_var_handle, const platform::Place &out_place, + SelectedRows *dst_selecte_rows); +#endif + + void Wait( + const std::map &dev_ctxes); + template std::vector GetInputValues( const std::vector &in_var_handles, diff --git a/paddle/fluid/framework/details/reduce_op_handle_test.cc b/paddle/fluid/framework/details/reduce_op_handle_test.cc index 72299c0bfa916..6cee4770e6435 100644 --- a/paddle/fluid/framework/details/reduce_op_handle_test.cc +++ b/paddle/fluid/framework/details/reduce_op_handle_test.cc @@ -35,7 +35,7 @@ struct TestReduceOpHandle { std::vector gpu_list_; std::vector> ctxs_; -#ifdef PADDLE_WITH_CUDA +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) std::unique_ptr nccl_ctxs_; #endif @@ -43,7 +43,7 @@ struct TestReduceOpHandle { for (size_t j = 0; j < ctxs_.size(); ++j) { ctxs_[j]->Wait(); } -#ifdef PADDLE_WITH_CUDA +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) if (nccl_ctxs_) { nccl_ctxs_->WaitAll(); } @@ -53,7 +53,7 @@ struct TestReduceOpHandle { void InitCtxOnGpu(bool use_gpu) { use_gpu_ = use_gpu; if (use_gpu) { -#ifdef PADDLE_WITH_CUDA +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) int count = p::GetCUDADeviceCount(); if (count <= 1) { LOG(WARNING) << "Cannot test multi-gpu Broadcast, because the CUDA " @@ -77,7 +77,7 @@ struct TestReduceOpHandle { gpu_list_.push_back(p); ctxs_.emplace_back(new p::CPUDeviceContext(p)); } -#ifdef PADDLE_WITH_CUDA +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) nccl_ctxs_.reset(nullptr); #endif } @@ -99,14 +99,14 @@ struct TestReduceOpHandle { nodes.emplace_back(new ir::Node("node")); if (use_gpu_) { -#ifdef PADDLE_WITH_CUDA +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) op_handle_.reset(new ReduceOpHandle(nodes.back().get(), local_scopes_, gpu_list_, nccl_ctxs_.get())); #else PADDLE_THROW("CUDA is not support."); #endif } else { -#ifdef PADDLE_WITH_CUDA +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) op_handle_.reset(new ReduceOpHandle(nodes.back().get(), local_scopes_, gpu_list_, nccl_ctxs_.get())); #else diff --git a/paddle/fluid/framework/details/reference_count_op_handle.h b/paddle/fluid/framework/details/reference_count_op_handle.h deleted file mode 100644 index cc4ccfbdfc720..0000000000000 --- a/paddle/fluid/framework/details/reference_count_op_handle.h +++ /dev/null @@ -1,138 +0,0 @@ -// 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 -#include -#include -#include - -#include "paddle/fluid/framework/details/op_handle_base.h" -#include "paddle/fluid/framework/garbage_collector.h" -#include "paddle/fluid/framework/scope.h" -#include "paddle/fluid/framework/selected_rows.h" -#include "paddle/fluid/framework/tensor.h" - -namespace paddle { -namespace framework { -namespace details { - -using ReferenceCountMap = std::unordered_map; -using AtomicReferenceCountMap = - std::unordered_map>; -using DeviceReferenceCountMap = - std::unordered_map>; -using AtomicDeviceReferenceCountMap = - std::unordered_map>; -using DeviceGarbageCollectorMap = - std::unordered_map>>; - -class ReferenceCountOpHandle : public OpHandleBase { - public: - ReferenceCountOpHandle(ir::Node *node, const Scope *scope, - const platform::CUDAPlace &place, - const std::vector &var_names, - GarbageCollector *gc, - AtomicReferenceCountMap *ref_cnts) - : OpHandleBase(node), scope_(scope), gc_(gc), ref_cnts_(ref_cnts) { - dev_ctx_ = static_cast( - platform::DeviceContextPool::Instance().Get(place)); - if (IsStreamGarabageCollector()) { - platform::SetDeviceId(place.device); - PADDLE_ENFORCE(cudaEventCreateWithFlags(&event_, cudaEventDisableTiming)); - } - - for (auto &name : var_names) AddVar(name); - } - - ~ReferenceCountOpHandle() { - if (IsStreamGarabageCollector()) { - auto gpu_place = boost::get(dev_ctx_->GetPlace()); - platform::SetDeviceId(gpu_place.device); - PADDLE_ENFORCE(cudaEventDestroy(event_)); - } - } - - std::string Name() const override { return "reference_count"; } - - void AddVar(const std::string &name) { - auto it = var_names_.find(name); - if (it != var_names_.end()) - ++(it->second); - else - var_names_[name] = 1; - } - - protected: - void RunImpl() override { - auto *exec_scope = scope_->FindVar(kLocalExecScopeName)->Get(); - std::vector tensors; - for (auto &pair : var_names_) { - auto &name = pair.first; - auto it = ref_cnts_->find(name); - if (it == ref_cnts_->end()) continue; - - auto *var = exec_scope->FindVar(name); - if (var == nullptr) continue; - - if (var->IsType()) { - if (it->second.fetch_sub(pair.second) <= pair.second) { - tensors.emplace_back(var->GetMutable()); - } - } else if (var->IsType()) { - if (it->second.fetch_sub(pair.second) <= pair.second) { - tensors.emplace_back( - var->GetMutable()->mutable_value()); - } - } - } - - if (!tensors.empty()) { - ClearTensors(tensors); - } - } - - private: - void ClearTensors(const std::vector &tensors) { - auto *gc = dynamic_cast *>(gc_); - if (gc != nullptr) { - auto compute_stream = dev_ctx_->stream(); - auto callback_stream = gc->stream(); - auto callback_func = [=]() { - PADDLE_ENFORCE(cudaEventRecord(event_, compute_stream)); - PADDLE_ENFORCE(cudaStreamWaitEvent(callback_stream, event_, 0)); - }; - gc_->Add(tensors, callback_func); - } else { - gc_->Add(tensors); - } - } - - bool IsStreamGarabageCollector() const { - return dynamic_cast *>(gc_) != nullptr; - } - - const Scope *scope_; - platform::CUDADeviceContext *dev_ctx_; - std::unordered_map var_names_; - GarbageCollector *gc_; // not own - AtomicReferenceCountMap *ref_cnts_; // not own - cudaEvent_t event_; -}; - -} // namespace details -} // namespace framework -} // namespace paddle diff --git a/paddle/fluid/framework/details/reference_count_pass.cc b/paddle/fluid/framework/details/reference_count_pass.cc index 28443cc886e4c..13a042d8e6ed7 100644 --- a/paddle/fluid/framework/details/reference_count_pass.cc +++ b/paddle/fluid/framework/details/reference_count_pass.cc @@ -14,187 +14,240 @@ #include #include +#include #include #include "paddle/fluid/framework/details/computation_op_handle.h" +#include "paddle/fluid/framework/details/eager_deletion_op_handle.h" #include "paddle/fluid/framework/details/multi_devices_helper.h" +#include "paddle/fluid/framework/details/op_graph_view.h" #include "paddle/fluid/framework/details/reference_count_pass.h" +#include "paddle/fluid/framework/details/reference_count_pass_helper.h" #include "paddle/fluid/framework/ir/graph_helper.h" namespace paddle { namespace framework { namespace details { -static ComputationOpHandle *FindNextComputationOpHandle(VarHandle *var_in) { - std::queue queue; - queue.push(var_in); - do { - auto *var = queue.front(); - queue.pop(); - for (auto *op : var->PendingOps()) { - auto *compute_op = dynamic_cast(op); - if (compute_op != nullptr && compute_op->GetPlace() == var_in->place_) { - return compute_op; +// A functor to shrink/remove operators who depend on other operators in a set +class ShrinkDepsOpFunctor { + private: + enum RelationShip { kSame = 0, kNoDeps = 1, kBefore = 2, kAfter = 3 }; + + public: + explicit ShrinkDepsOpFunctor(const std::vector &all_ops) + : graph_(all_ops) {} + + template + OpSet operator()(const OpSet &op_set) const { + using KeyType = typename OpSet::key_type; + static_assert( + std::is_base_of::type>::value, + "Key type of OpSet must be OpHandleBase, or derived of OpHandleBase"); + + if (op_set.size() <= 1) return op_set; + std::vector ops(op_set.begin(), op_set.end()); + OpSet ret; + auto rels = GetRelations(ops); + auto not_before = [](RelationShip r) { return r != kBefore; }; + for (size_t i = 0; i < rels.size(); ++i) { + if (std::all_of(rels[i].begin(), rels[i].end(), not_before)) { + ret.emplace(static_cast(ops[i])); } - for (auto *out_var : op->Outputs()) { - queue.push(out_var); + } + return ret; + } + + private: + std::vector> GetRelations( + const std::vector &ops) const { + std::unordered_map op_to_idx; + for (size_t i = 0; i < ops.size(); ++i) { + PADDLE_ENFORCE(graph_.HasOp(ops[i]), "Op does not exist in graph"); + op_to_idx[ops[i]] = i; + } + + PADDLE_ENFORCE(op_to_idx.size() == ops.size(), "Duplicate ops"); + + std::vector> ret(ops.size()); + for (auto &e : ret) { + e.assign(ops.size(), kSame); + } + + size_t found_num = ops.size(); + size_t total_num = ops.size() * ops.size(); + auto visitor = [&](OpHandleBase *op, size_t i) { + auto it = op_to_idx.find(op); + if (it != op_to_idx.end()) { + size_t j = it->second; + if (i != j && ret[i][j] == kSame) { + ret[i][j] = kBefore; + ret[j][i] = kAfter; + found_num += 2; + if (found_num == total_num) { + return false; + } + } + } + return true; + }; + + for (size_t i = 0; i < ops.size(); ++i) { + auto sub_visitor = [&, i](OpHandleBase *op) { return visitor(op, i); }; + if (!graph_.VisitAllPendingOps(ops[i], sub_visitor)) { + break; + } + } + + for (size_t i = 0; i < ops.size(); ++i) { + for (size_t j = i + 1; j < ops.size(); ++j) { + if (ret[i][j] != kSame) continue; + ret[i][j] = kNoDeps; + ret[j][i] = kNoDeps; + } + } + + return ret; + } + + const OpGraphView graph_; +}; + +/** + * Find the nearest downstream computation op handle. If the op is a + * computation op, just return itself. + */ +static ComputationOpHandle *FindNextComputationOpHandleOrReturnItself( + OpHandleBase *op, size_t scope_idx) { + std::queue q; + std::unordered_set visited; + q.push(op); + do { + auto *op = q.front(); + q.pop(); + auto *compute_op = dynamic_cast(op); + if (compute_op != nullptr && compute_op->GetScopeIdx() == scope_idx) { + return compute_op; + } + for (auto *out_var : op->Outputs()) { + for (auto *pending_op : out_var->PendingOps()) { + if (visited.count(pending_op)) continue; + visited.insert(pending_op); } } - } while (!queue.empty()); + } while (!q.empty()); return nullptr; } -static void AddDependencyBetween(OpHandleBase *in, OpHandleBase *out, - ir::Graph *graph) { - auto it = std::find_if( - in->Outputs().begin(), in->Outputs().end(), [](VarHandleBase *var) { - return dynamic_cast(var) != nullptr; - }); - - if (it != in->Outputs().end()) { - out->AddInput(*it); - } else { - auto *dep_var = new DummyVarHandle(graph->CreateControlDepVar()); - graph->Get(kGraphDepVars).emplace(dep_var); - in->AddOutput(dep_var); - out->AddInput(dep_var); +static std::unordered_set +ExtractComputationOpFromLastLivedVar(VarHandle *var, size_t scope_idx, + const ShrinkDepsOpFunctor &shrink_func, + bool *ok) { + // stage one. Get last op for variable. + std::unordered_set candidates; + { + if (var->PendingOps().empty() && var->GeneratedOp()) { + // No operator depends on this variable. So the last operator is the op + // who generates this variable. + candidates.emplace(var->GeneratedOp()); + } else { + candidates = var->PendingOps(); + } + + // No pending ops or generated op is nullptr + if (candidates.empty()) { + *ok = false; + return {}; + } + } + + // stage two. Try to cast them to computation op. + // return (*ok=false) when failed. + // + // The reason why we cannot make any types of op handle to be the last lived + // op is: + // some op handle may operate on many DeviceContext, however, our garbage + // collector can only wait one DeviceContext for now. So currently, we wait + // the nearest compute op. + std::unordered_set computation_op; + { + for (auto *op : candidates) { + auto *compute_op = + FindNextComputationOpHandleOrReturnItself(op, scope_idx); + if (compute_op == nullptr) { + *ok = false; + return {}; + } + computation_op.emplace(compute_op); + } } + + // stage three. Try to shrink computation op if they depend on each other. + // Get the smallest set of the most ops. + *ok = true; + return shrink_func(computation_op); +} + +static VarDesc *TryGetLatestVarDesc(const std::vector &vars) { + VarDesc *var_desc = nullptr; + std::find_if(vars.rbegin(), vars.rend(), [&](VarHandle *var_handle) -> bool { + var_desc = var_handle->Node()->Var(); + return var_desc != nullptr; + }); + return var_desc; } std::unique_ptr ReferenceCountPass::ApplyImpl( std::unique_ptr graph) const { - auto &ref_cnts = Get(kGlobalReferenceCount); - auto &cur_ref_cnts = Get(kCurReferenceCount); - auto &gcs = Get(kGarbageCollector); - - // It is not easy to find the right reference counts of varaibles in graph - // Step 1: Find all variables in computation ops - // Step 2: Find all variables in non-computation ops which refers to variables - // in computation ops - std::unordered_set names; - std::unordered_map - compute_ref_cnt_map; - - auto get_ref_cnts_from_compute_op = [&]( - OpHandleBase *op, const std::vector &vars) { - std::vector var_names_in_op; - auto *compute_op = dynamic_cast(op); - if (compute_op == nullptr || - !platform::is_gpu_place(compute_op->GetPlace())) - return var_names_in_op; - auto place = boost::get(compute_op->GetPlace()); - for (VarHandleBase *var_handle_base : vars) { - auto *var_handle = dynamic_cast(var_handle_base); - if (var_handle == nullptr || !var_handle->Node()->IsVar()) continue; - - if (!platform::is_gpu_place(var_handle->place_) || - boost::get(var_handle->place_) != place) - continue; + auto &ref_cnts = Get>(kGlobalReferenceCount); + auto &last_live_ops_of_vars = + Get>(kLastLiveOpsOfVars); + + PADDLE_ENFORCE(last_live_ops_of_vars.empty() && ref_cnts.empty(), + "Last Live Ops and Reference Counts of vars should be " + "initialized at here."); - VarDesc *var_desc = var_handle->Node()->Var(); - auto var_name = var_handle->Node()->Name(); + const auto &vars = graph->Get(kGraphVars); - // This is weird but there is really some variables without var_desc - // in computation_op - if (var_desc == nullptr) { - var_desc = compute_op->Node()->Op()->Block()->FindVar(var_name); - if (var_desc == nullptr) continue; + last_live_ops_of_vars.resize(vars.size()); + ref_cnts.resize(vars.size()); + + ShrinkDepsOpFunctor shrink_func( + ir::FilterByNodeWrapper(*graph)); + + for (size_t i = 0; i < vars.size(); ++i) { + for (auto &name_var_pair : vars[i]) { + // Whether this variable can be reused or deleted? If not, we do not + // compute reference counts and dependencies. + VarDesc *var_desc = TryGetLatestVarDesc(name_var_pair.second); + + if (var_desc == nullptr || var_desc->Persistable()) { + continue; } - if (var_desc->Persistable()) continue; auto var_type = var_desc->Proto()->type().type(); if (var_type != proto::VarType::LOD_TENSOR && - var_type != proto::VarType::SELECTED_ROWS) { + var_type != proto::VarType::SELECTED_ROWS && + var_type != proto::VarType::LOD_TENSOR_ARRAY) { + // Var type cannot be deleted continue; } - // compute op only runs in one device - if (ref_cnts[place.device]->count(var_name)) - ++(*ref_cnts[place.device])[var_name]; - else - (*ref_cnts[place.device])[var_name] = 1; + bool ok; + auto result = ExtractComputationOpFromLastLivedVar( + name_var_pair.second.back(), i, shrink_func, &ok); - names.insert(var_name); - var_names_in_op.push_back(var_name); - } - return var_names_in_op; - }; - - auto update_ref_cnts_from_non_compute_op = [&]( - OpHandleBase *op, const std::vector &vars) { - if (dynamic_cast(op) != nullptr) return; - for (VarHandleBase *var_handle_base : vars) { - auto *var_handle = dynamic_cast(var_handle_base); - if (var_handle == nullptr || !var_handle->Node()->IsVar()) continue; - - auto var_name = var_handle->Node()->Name(); - auto var_place = var_handle->place_; - if (!platform::is_gpu_place(var_place)) continue; - auto place = boost::get(var_place); - if (names.count(var_name) == 0) continue; - if (ref_cnts.count(place.device) && - ref_cnts[place.device]->count(var_name)) { - ++(*ref_cnts[place.device])[var_name]; - - auto *next_compute_op = FindNextComputationOpHandle(var_handle); - if (next_compute_op != nullptr) { - if (compute_ref_cnt_map.count(next_compute_op)) { - compute_ref_cnt_map[next_compute_op]->AddVar(var_name); - VLOG(50) << "Add reference count of " << var_name << " to Operator " - << next_compute_op->Name(); - } else { - // Create new reference_count_op_handle - ir::Node *ref_cnt_node = graph->CreateEmptyNode( - "reference_count", ir::Node::Type::kOperation); - auto *ref_cnt_handle = new ReferenceCountOpHandle( - ref_cnt_node, next_compute_op->GetScope(), place, {var_name}, - gcs[place.device].get(), cur_ref_cnts[place.device].get()); - AddDependencyBetween(next_compute_op, ref_cnt_handle, graph.get()); - compute_ref_cnt_map[next_compute_op] = ref_cnt_handle; - } - } + if (ok) { + auto &var_name = name_var_pair.first; + PADDLE_ENFORCE(!result.empty(), "Last living ops of %s cannot be empty", + var_name); + ref_cnts[i].emplace(var_name, result.size()); + last_live_ops_of_vars[i].emplace(var_name, std::move(result)); } } - }; - - auto all_ops = ir::FilterByNodeWrapper(*graph); - for (auto &op : all_ops) { - auto in_var_names = get_ref_cnts_from_compute_op(op, op->Inputs()); - auto out_var_names = get_ref_cnts_from_compute_op(op, op->Outputs()); - if (in_var_names.empty() && out_var_names.empty()) continue; - in_var_names.insert(in_var_names.end(), out_var_names.begin(), - out_var_names.end()); - auto *compute_op = dynamic_cast(op); - auto place = boost::get(compute_op->GetPlace()); - ir::Node *ref_cnt_node = - graph->CreateEmptyNode("reference_count", ir::Node::Type::kOperation); - auto *ref_cnt_handle = new ReferenceCountOpHandle( - ref_cnt_node, compute_op->GetScope(), place, in_var_names, - gcs[place.device].get(), cur_ref_cnts[place.device].get()); - AddDependencyBetween(compute_op, ref_cnt_handle, graph.get()); - compute_ref_cnt_map[compute_op] = ref_cnt_handle; - } - - for (auto &op : all_ops) { - update_ref_cnts_from_non_compute_op(op, op->Inputs()); - update_ref_cnts_from_non_compute_op(op, op->Outputs()); - } - - std::vector new_all_ops; - new_all_ops.reserve(compute_ref_cnt_map.size() + all_ops.size()); - for (auto &op : all_ops) { - new_all_ops.emplace_back(std::move(op)); - auto it = compute_ref_cnt_map.find(new_all_ops.back()); - if (it != compute_ref_cnt_map.end()) { - // Add LeafNode to ReferenceCountOpHandle - auto *dummy_leaf = new DummyVarHandle(graph->CreateControlDepVar()); - graph->Get(kGraphDepVars).emplace(dummy_leaf); - it->second->AddOutput(dummy_leaf); - new_all_ops.emplace_back(std::move(it->second)); - } } - all_ops.swap(new_all_ops); return graph; } @@ -205,5 +258,4 @@ std::unique_ptr ReferenceCountPass::ApplyImpl( REGISTER_PASS(reference_count_pass, paddle::framework::details::ReferenceCountPass) .RequirePassAttr(paddle::framework::details::kGlobalReferenceCount) - .RequirePassAttr(paddle::framework::details::kCurReferenceCount) - .RequirePassAttr(paddle::framework::details::kGarbageCollector); + .RequirePassAttr(paddle::framework::details::kLastLiveOpsOfVars); diff --git a/paddle/fluid/framework/details/reference_count_pass.h b/paddle/fluid/framework/details/reference_count_pass.h index 7081280b0600b..bcbef027354ef 100644 --- a/paddle/fluid/framework/details/reference_count_pass.h +++ b/paddle/fluid/framework/details/reference_count_pass.h @@ -14,7 +14,6 @@ #pragma once -#include "paddle/fluid/framework/details/reference_count_op_handle.h" #include "paddle/fluid/framework/ir/graph.h" #include "paddle/fluid/framework/ir/pass.h" @@ -22,10 +21,6 @@ namespace paddle { namespace framework { namespace details { -constexpr char kGlobalReferenceCount[] = "reference_count"; -constexpr char kCurReferenceCount[] = "current_reference_count"; -constexpr char kGarbageCollector[] = "garbage_collector"; - class ReferenceCountPass : public ir::Pass { protected: std::unique_ptr ApplyImpl( diff --git a/paddle/fluid/framework/details/reference_count_pass_helper.cc b/paddle/fluid/framework/details/reference_count_pass_helper.cc new file mode 100644 index 0000000000000..89bd08c2d041d --- /dev/null +++ b/paddle/fluid/framework/details/reference_count_pass_helper.cc @@ -0,0 +1,21 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/framework/details/reference_count_pass_helper.h" + +namespace paddle { +namespace framework { +namespace details {} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/reference_count_pass_helper.h b/paddle/fluid/framework/details/reference_count_pass_helper.h new file mode 100644 index 0000000000000..1c083dbf001b0 --- /dev/null +++ b/paddle/fluid/framework/details/reference_count_pass_helper.h @@ -0,0 +1,51 @@ +// 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 +#include +#include +#include +#include +#include + +#include "paddle/fluid/framework/garbage_collector.h" + +namespace paddle { +namespace framework { +namespace details { + +class ComputationOpHandle; + +using ReferenceCountMap = std::unordered_map; + +using AtomicReferenceCountMap = + std::unordered_map>; + +using GarbageCollectorMap = + std::map>; + +const char kGlobalReferenceCount[] = "global_reference_count"; +const char kRuntimeReferenceCount[] = "runtime_reference_count"; +const char kGarbageCollector[] = "garbage_collector"; +const char kAllPlaces[] = "all_places"; + +using LastLiveOpsOfVars = + std::unordered_map>; +const char kLastLiveOpsOfVars[] = "last_live_ops_of_var"; + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/scale_loss_grad_op_handle.cc b/paddle/fluid/framework/details/scale_loss_grad_op_handle.cc index 6ab6cb2332b0a..ef1626599795a 100644 --- a/paddle/fluid/framework/details/scale_loss_grad_op_handle.cc +++ b/paddle/fluid/framework/details/scale_loss_grad_op_handle.cc @@ -51,7 +51,7 @@ void ScaleLossGradOpHandle::RunImpl() { ->stream(); memory::Copy(boost::get(place_), tmp, platform::CPUPlace(), &coeff_, sizeof(float), stream); - VLOG(100) << place_ << "RUN Scale loss grad op"; + VLOG(10) << place_ << "RUN Scale loss grad op"; }); #endif } diff --git a/paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.cc b/paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.cc index e5b1eaa7318ae..57f6fc66c57e2 100644 --- a/paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.cc +++ b/paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.cc @@ -16,11 +16,8 @@ #include #include #include -#include "paddle/fluid/framework/executor.h" +#include "paddle/fluid/framework/variable_helper.h" #include "paddle/fluid/platform/profiler.h" -#ifdef PADDLE_WITH_CUDA -#include "paddle/fluid/framework/details/reference_count_op_handle.h" -#endif namespace paddle { namespace framework { @@ -69,27 +66,12 @@ FeedFetchList ScopeBufferedSSAGraphExecutor::Run( platform::RecordEvent e("ScopeBufferedSSAGraphExecutorAfterRun", nullptr); drop_scope_counter_ += 1; -#ifdef PADDLE_WITH_CUDA - const std::string gc_name = "garbage_collector"; - DeviceGarbageCollectorMap *gc = - Graph().Has(gc_name) ? &(Graph().Get(gc_name)) - : nullptr; -#endif - if (!fetch_tensors.empty() || drop_scope_counter_ == strategy_.num_iteration_per_drop_scope_) { drop_scope_counter_ = 0; // Wait All computational streams for (auto p : places_) { platform::DeviceContextPool::Instance().Get(p)->Wait(); -#ifdef PADDLE_WITH_CUDA - if (gc != nullptr && platform::is_gpu_place(p)) { - auto gpu_place = boost::get(p); - auto &gc_at_place = gc->at(gpu_place.device); - gc_at_place->Wait(); - gc_at_place->Reset(); - } -#endif } for (auto &scope : local_scopes_) { auto &local_scope = diff --git a/paddle/fluid/framework/details/sequential_execution_pass.cc b/paddle/fluid/framework/details/sequential_execution_pass.cc index f78a47bb78e6f..cc2c8bfef9f9f 100644 --- a/paddle/fluid/framework/details/sequential_execution_pass.cc +++ b/paddle/fluid/framework/details/sequential_execution_pass.cc @@ -94,8 +94,8 @@ std::unique_ptr SequentialExecutionPass::ApplyImpl( op_node_list[i - 1]->outputs.push_back(dep_var); dep_var->outputs.push_back(op_node_list[i]); dep_var->inputs.push_back(op_node_list[i - 1]); - VLOG(100) << "Add dependencies between " << op_node_list[i - 1]->Name() - << " and " << op_node_list[i]->Name(); + VLOG(10) << "Add dependencies between " << op_node_list[i - 1]->Name() + << " and " << op_node_list[i]->Name(); } return graph; } diff --git a/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc b/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc index f781f02a07659..677a2937945b0 100644 --- a/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc +++ b/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc @@ -210,16 +210,16 @@ void ThreadedSSAGraphExecutor::RunOp( details::OpHandleBase *op) { auto op_run = [ready_var_q, op, this] { try { - if (VLOG_IS_ON(100)) { - VLOG(100) << op << " " << op->Name() << " : " << op->DebugString(); + if (VLOG_IS_ON(10)) { + VLOG(10) << op << " " << op->Name() << " : " << op->DebugString(); } if (LIKELY(!strategy_.dry_run_)) { op->Run(strategy_.use_cuda_); } - VLOG(100) << op << " " << op->Name() << " Done "; + VLOG(10) << op << " " << op->Name() << " Done "; running_ops_--; ready_var_q->Extend(op->Outputs()); - VLOG(100) << op << " " << op->Name() << "Signal posted"; + VLOG(10) << op << " " << op->Name() << "Signal posted"; } catch (...) { exception_holder_.Catch(std::current_exception()); } diff --git a/paddle/fluid/framework/dlpack_tensor.cc b/paddle/fluid/framework/dlpack_tensor.cc new file mode 100644 index 0000000000000..eaef093ed3b6e --- /dev/null +++ b/paddle/fluid/framework/dlpack_tensor.cc @@ -0,0 +1,124 @@ +// 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/dlpack_tensor.h" +#include "paddle/fluid/framework/data_type.h" +namespace paddle { +namespace framework { + +namespace internal { +template +static ::DLDataType GetDLDataTypeCode() { + ::DLDataType dtype; + if (std::is_same::value || + std::is_floating_point::value) { + dtype.code = kDLFloat; + } else if (std::is_unsigned::value) { + dtype.code = kDLUInt; + } else if (std::is_integral::value) { + dtype.code = kDLInt; + } else { + PADDLE_THROW("Unsupported data type %s", typeid(T).name()); + } + dtype.bits = 8 * sizeof(T); + dtype.lanes = 1; + return dtype; +} + +static std::unordered_map CreateDLDataTypeMap() { + static std::unordered_map result; + +#define REG_DL_DATA_TYPE(cpp_type, proto_type) \ + result[static_cast(proto_type)] = GetDLDataTypeCode() + + _ForEachDataType_(REG_DL_DATA_TYPE); +#undef REG_DL_DATA_TYPE + return result; +} + +static DLDataType GetDLDataTypeFromTypeIndex(proto::VarType::Type type) { + static auto type_to_dtype_map = CreateDLDataTypeMap(); + static auto type_to_dtype_map_end_it = type_to_dtype_map.end(); + auto it = type_to_dtype_map.find(static_cast(type)); + PADDLE_ENFORCE(it != type_to_dtype_map_end_it, "Unsupported data type %d", + type); + return it->second; +#undef REG_DL_DATA_TYPE +} + +struct DLContextVisitor : public boost::static_visitor<::DLContext> { + inline ::DLContext operator()(const platform::CPUPlace &place) const { + DLContext ctx; + ctx.device_type = kDLCPU; + ctx.device_id = 0; + return ctx; + } + + inline ::DLContext operator()(const platform::CUDAPlace &place) const { +#ifdef PADDLE_WITH_CUDA + DLContext ctx; + ctx.device_type = kDLGPU; + ctx.device_id = place.device; + return ctx; +#else + PADDLE_THROW("platform::CUDAPlace is not supported in CPU only version"); +#endif + } + + inline ::DLContext operator()(const platform::CUDAPinnedPlace &place) const { +#ifdef PADDLE_WITH_CUDA + DLContext ctx; + ctx.device_type = kDLCPUPinned; + ctx.device_id = 0; + return ctx; +#else + PADDLE_THROW( + "platform::CUDAPinnedPlace is not supported in CPU only version"); +#endif + } +}; +} // namespace internal + +DLPackTensor::DLPackTensor(const Tensor &tensor, LaneType lanes) { + // init data, data buffer + t_.data = const_cast(tensor.data()); + + // init ctx, DLContext type with device_type and device_id + auto place = tensor.place(); + t_.ctx = boost::apply_visitor(internal::DLContextVisitor(), place); + + // init dtype + t_.dtype = internal::GetDLDataTypeFromTypeIndex(tensor.type()); + t_.dtype.lanes = lanes; + + // init ndim, tensor rank + auto &dims = tensor.dims(); + using DimType = decltype(t_.ndim); // int + t_.ndim = static_cast(dims.size()); + + // init shape, tensor dims + t_.shape = shape_; + for (DimType i = 0; i < t_.ndim; ++i) { + t_.shape[i] = dims[i]; + } + + // init strides, nullptr means the tensor is compact + t_.strides = nullptr; + + // init byte_offset + t_.byte_offset = 0; +} + +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/dlpack_tensor.h b/paddle/fluid/framework/dlpack_tensor.h new file mode 100644 index 0000000000000..0c52bce1ef6af --- /dev/null +++ b/paddle/fluid/framework/dlpack_tensor.h @@ -0,0 +1,45 @@ +// 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 +#include "paddle/fluid/framework/tensor.h" + +namespace paddle { +namespace framework { + +class DLPackTensor { + public: + using LaneType = decltype(::DLTensor::dtype.lanes); // uint16_t + using ShapeType = + std::remove_reference::type; // int64_t + + // lanes is only used in CPU to enable vectorization + explicit DLPackTensor(const Tensor& tensor, LaneType lanes = 1); + + inline operator const ::DLTensor&() const { return t_; } + + inline operator ::DLTensor&() { return t_; } + + private: + ::DLTensor t_; + + // The shape in DLTensor is defined as int64_t* + // Add this member to make TVMTensor init without heap allocation + ShapeType shape_[9]; +}; + +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/dlpack_tensor_test.cc b/paddle/fluid/framework/dlpack_tensor_test.cc new file mode 100644 index 0000000000000..c0a8e1bcdfa3a --- /dev/null +++ b/paddle/fluid/framework/dlpack_tensor_test.cc @@ -0,0 +1,101 @@ +// 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/dlpack_tensor.h" +#include +#include +#include + +namespace paddle { +namespace framework { + +namespace { // NOLINT +template +constexpr uint8_t GetDLDataTypeCode() { + return std::is_same::value || + std::is_floating_point::value + ? static_cast(kDLFloat) + : (std::is_unsigned::value + ? static_cast(kDLUInt) + : (std::is_integral::value ? static_cast(kDLInt) + : static_cast(-1))); +} +} // NOLINT + +template +void TestMain(const platform::Place &place, uint16_t lanes) { + DDim dims{4, 5, 6, 7}; + Tensor tensor; + tensor.Resize(dims); + void *p = tensor.mutable_data(place); + + DLPackTensor dlpack_tensor(tensor, lanes); + ::DLTensor &dl_tensor = dlpack_tensor; + + CHECK_EQ(p, dl_tensor.data); + if (platform::is_cpu_place(place)) { + CHECK_EQ(kDLCPU, dl_tensor.ctx.device_type); + CHECK_EQ(0, dl_tensor.ctx.device_id); + } else if (platform::is_gpu_place(place)) { + CHECK_EQ(kDLGPU, dl_tensor.ctx.device_type); + CHECK_EQ(boost::get(place).device, + dl_tensor.ctx.device_id); + } else if (platform::is_cuda_pinned_place(place)) { + CHECK_EQ(kDLCPUPinned, dl_tensor.ctx.device_type); + CHECK_EQ(0, dl_tensor.ctx.device_id); + } else { + CHECK_EQ(false, true); + } + + CHECK_EQ(dims.size(), dl_tensor.ndim); + for (auto i = 0; i < dims.size(); ++i) { + CHECK_EQ(dims[i], dl_tensor.shape[i]); + } + + CHECK_EQ(dl_tensor.strides == nullptr, true); + CHECK_EQ(static_cast(0), dl_tensor.byte_offset); + + CHECK_EQ(lanes, dl_tensor.dtype.lanes); + CHECK_EQ(sizeof(T) * 8, dl_tensor.dtype.bits); + + CHECK_EQ(GetDLDataTypeCode(), dl_tensor.dtype.code); +} + +template +void TestMainLoop() { +#ifdef PADDLE_WITH_CUDA + std::vector places{platform::CPUPlace(), + platform::CUDAPlace(0), + platform::CUDAPinnedPlace()}; + if (platform::GetCUDADeviceCount() > 1) { + places.emplace_back(platform::CUDAPlace(1)); + } +#else + std::vector places{platform::CPUPlace()}; +#endif + std::vector lanes{1, 2}; + for (auto &p : places) { + for (auto &l : lanes) { + TestMain(p, l); + } + } +} +TEST(dlpack, test_all) { +#define TestCallback(cpp_type, proto_type) TestMainLoop() + + _ForEachDataType_(TestCallback); +} + +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/executor.cc b/paddle/fluid/framework/executor.cc index 7ce08b728d943..da9556c6c1f34 100644 --- a/paddle/fluid/framework/executor.cc +++ b/paddle/fluid/framework/executor.cc @@ -13,17 +13,23 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/framework/executor.h" +#include #include "paddle/fluid/framework/feed_fetch_method.h" #include "paddle/fluid/framework/lod_rank_table.h" #include "paddle/fluid/framework/lod_tensor_array.h" -#include "paddle/fluid/framework/ngraph_operator.h" #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/reader.h" +#include "paddle/fluid/framework/transfer_scope_cache.h" +#include "paddle/fluid/framework/variable_helper.h" #include "paddle/fluid/operators/detail/macros.h" #include "paddle/fluid/platform/place.h" #include "paddle/fluid/platform/profiler.h" +#ifdef PADDLE_WITH_NGRAPH +#include "paddle/fluid/framework/ngraph_operator.h" +#endif + DECLARE_bool(benchmark); DEFINE_bool(use_mkldnn, false, "Use MKLDNN to run"); DEFINE_bool(use_ngraph, false, "Use NGRAPH to run"); @@ -36,40 +42,84 @@ namespace { int kProgramId = -1; } // namespace +static std::unordered_map GetNonPersistableReferenceCounts( + const BlockDesc& block, const std::vector& skip_var_list) { + std::unordered_map ref_cnts; + std::unordered_set skip_vars(skip_var_list.begin(), + skip_var_list.end()); + + auto update_ref_cnts = [&](OpDesc* op_desc, const VariableNameMap& name_map) { + for (auto& name_pair : name_map) { + for (auto& name : name_pair.second) { + if (skip_vars.count(name)) continue; + auto* var_desc = block.FindVar(name); + if (var_desc == nullptr || var_desc->Persistable()) continue; + auto type = var_desc->Proto()->type().type(); + if (type != proto::VarType::LOD_TENSOR && + type != proto::VarType::SELECTED_ROWS && + type != proto::VarType::LOD_TENSOR_ARRAY) { + continue; + } + ++ref_cnts[name]; + } + } + }; + + for (auto op_desc : block.AllOps()) { + update_ref_cnts(op_desc, op_desc->Inputs()); + update_ref_cnts(op_desc, op_desc->Outputs()); + } + return ref_cnts; +} + ExecutorPrepareContext::ExecutorPrepareContext( - const framework::ProgramDesc& prog, size_t block_id) + const framework::ProgramDesc& prog, size_t block_id, + const std::vector& skip_ref_cnt_vars) : prog_(prog), block_id_(block_id) { if (GetEagerDeletionThreshold() >= 0) { - ref_cnts_ = GetNonPersistableReferenceCount(prog_, block_id_); + global_ref_cnts_ = GetNonPersistableReferenceCounts(prog.Block(block_id), + skip_ref_cnt_vars); } } ExecutorPrepareContext::~ExecutorPrepareContext() { - VLOG(50) << "destroy ExecutorPrepareContext"; + VLOG(5) << "destroy ExecutorPrepareContext"; } -template -static void DeleteUnusedTensors(const Scope& scope, const OperatorBase* op, - GarbageCollector* gc, - RefCntMap* ref_cnts) { - std::unordered_set erase_tensors; +static void DeleteUnusedTensors( + const Scope& scope, const OperatorBase* op, GarbageCollector* gc, + std::unordered_map* ref_cnts) { + std::deque> garbages; auto handler = [&](const VariableNameMap& name_map) { for (auto& name_pair : name_map) { for (auto& name : name_pair.second) { auto it = ref_cnts->find(name); if (it == ref_cnts->end()) continue; - if ((it->second)-- == 1) { - auto* var = scope.FindVar(name); - if (var != nullptr) { - VLOG(100) << "Erase tensor \'" << name << "\'"; - if (var->IsType()) { - erase_tensors.insert(var->GetMutable()); - } else if (var->IsType()) { - erase_tensors.insert( - var->GetMutable()->mutable_value()); - } + if (--(it->second) != 0) { + continue; + } + auto* var = scope.FindVar(name); + if (var == nullptr) { + continue; + } + + VLOG(2) << "Erase variable " << name; + if (var->IsType()) { + garbages.emplace_back( + var->GetMutable()->MoveMemoryHolder()); + } else if (var->IsType()) { + garbages.emplace_back(var->GetMutable() + ->mutable_value() + ->MoveMemoryHolder()); + } else if (var->IsType()) { + auto* lod_tensor_arr = var->GetMutable(); + for (auto& t : *lod_tensor_arr) { + garbages.emplace_back(t.MoveMemoryHolder()); } + } else { + PADDLE_THROW("Type %s of %s is not supported eager deletion", + var->Type().name(), name); } } } @@ -78,19 +128,19 @@ static void DeleteUnusedTensors(const Scope& scope, const OperatorBase* op, handler(op->Inputs()); handler(op->Outputs()); - if (!erase_tensors.empty()) { - gc->Add(erase_tensors); + if (!garbages.empty()) { + gc->Add(std::move(garbages)); } } static void EnableFusedOp(ExecutorPrepareContext* ctx) { #ifdef PADDLE_WITH_NGRAPH VLOG(3) << "use_ngraph=True"; - auto intervals = FusedOperator::FusedOpIntervals(&ctx->ops_); + auto intervals = NgraphOperator::NgraphOpIntervals(&ctx->ops_); for (auto& interval : intervals) { - auto* fused_op = new FusedOperator(ctx->prog_, ctx->block_id_, - interval.at(0), interval.at(1)); - *interval[0] = std::unique_ptr(fused_op); + auto* ng_op = new NgraphOperator(ctx->prog_, ctx->block_id_, interval.at(0), + interval.at(1)); + *interval[0] = std::unique_ptr(ng_op); } for (auto it = intervals.rbegin(); it != intervals.rend(); ++it) { ctx->ops_.erase(it->at(0) + 1, it->at(1)); @@ -107,42 +157,12 @@ void Executor::Close() { #ifdef PADDLE_WITH_DISTRIBUTE // TODO(typhoonzero): complete message will need to use real trainer_id, // except 0. - ::paddle::operators::distributed::RPCClient::GetInstance< - ::paddle::operators::distributed::GRPCClient>(0) - ->SendComplete(); + auto client = + paddle::operators::distributed::RPCClient::GetInstance(0); + client->SendComplete(); #endif } -void InitializeVariable(Variable* var, proto::VarType::Type var_type) { - if (var_type == proto::VarType::LOD_TENSOR) { - var->GetMutable(); - } else if (var_type == proto::VarType::SELECTED_ROWS) { - var->GetMutable(); - } else if (var_type == proto::VarType::FEED_MINIBATCH) { - var->GetMutable(); - } else if (var_type == proto::VarType::FETCH_LIST) { - var->GetMutable(); - } else if (var_type == proto::VarType::STEP_SCOPES) { - var->GetMutable>(); - } else if (var_type == proto::VarType::LOD_RANK_TABLE) { - var->GetMutable(); - } else if (var_type == proto::VarType::LOD_TENSOR_ARRAY) { - var->GetMutable(); - } else if (var_type == proto::VarType::PLACE_LIST) { - var->GetMutable(); - } else if (var_type == proto::VarType::READER) { - var->GetMutable(); - } else if (var_type == proto::VarType::RAW) { - // GetMutable will be called in operator - } else { - PADDLE_THROW( - "Variable type %d is not in " - "[LOD_TENSOR, SELECTED_ROWS, FEED_MINIBATCH, FETCH_LIST, " - "LOD_RANK_TABLE, PLACE_LIST, READER, RAW]", - var_type); - } -} - void Executor::CreateVariables(const ProgramDesc& pdesc, Scope* scope, int block_id) { auto& global_block = pdesc.Block(block_id); @@ -161,21 +181,21 @@ void Executor::CreateVariables(const ProgramDesc& pdesc, Scope* scope, if (var->Persistable()) { auto* ptr = const_cast(ancestor_scope)->Var(var->Name()); InitializeVariable(ptr, var->GetType()); - VLOG(30) << "Create Variable " << var->Name() - << " global, which pointer is " << ptr; + VLOG(3) << "Create Variable " << var->Name() + << " global, which pointer is " << ptr; } else { auto* ptr = scope->Var(var->Name()); InitializeVariable(ptr, var->GetType()); - VLOG(30) << "Create Variable " << var->Name() - << " locally, which pointer is " << ptr; + VLOG(3) << "Create Variable " << var->Name() + << " locally, which pointer is " << ptr; } } } else { for (auto& var : global_block.AllVars()) { auto* ptr = scope->Var(var->Name()); InitializeVariable(ptr, var->GetType()); - VLOG(30) << "Create variable " << var->Name() << ", which pointer is " - << ptr; + VLOG(3) << "Create variable " << var->Name() << ", which pointer is " + << ptr; } } } @@ -306,7 +326,7 @@ void Executor::Run(const ProgramDesc& program, Scope* scope, int i = 0; for (auto& feed_target : (*feed_targets)) { std::string var_name = feed_target.first; - VLOG(30) << "feed target's name: " << var_name; + VLOG(3) << "feed target's name: " << var_name; // prepend feed op auto* op = global_block->PrependOp(); @@ -329,7 +349,7 @@ void Executor::Run(const ProgramDesc& program, Scope* scope, int i = 0; for (auto& fetch_target : (*fetch_targets)) { std::string var_name = fetch_target.first; - VLOG(30) << "fetch target's name: " << var_name; + VLOG(3) << "fetch target's name: " << var_name; // append fetch op auto* op = global_block->AppendOp(); @@ -350,9 +370,10 @@ void Executor::Run(const ProgramDesc& program, Scope* scope, } std::unique_ptr Executor::Prepare( - const ProgramDesc& program, int block_id) { + const ProgramDesc& program, int block_id, + const std::vector& skip_ref_cnt_vars) { std::unique_ptr ctx( - new ExecutorPrepareContext(program, block_id)); + new ExecutorPrepareContext(program, block_id, skip_ref_cnt_vars)); PADDLE_ENFORCE_LT(static_cast(block_id), program.Size()); auto& block = program.Block(block_id); for (auto& op_desc : block.AllOps()) { @@ -363,16 +384,28 @@ std::unique_ptr Executor::Prepare( } std::vector> Executor::Prepare( - const ProgramDesc& program, const std::vector& block_ids) { + const ProgramDesc& program, const std::vector& block_ids, + const std::vector>& skip_ref_cnt_vars) { + PADDLE_ENFORCE( + skip_ref_cnt_vars.empty() || skip_ref_cnt_vars.size() == block_ids.size(), + "skip_ref_cnt_vars should be either empty or equals to block number %d", + block_ids.size()); std::vector> result; + size_t idx = 0; for (auto& bid : block_ids) { - auto* ctx = new ExecutorPrepareContext(program, bid); + ExecutorPrepareContext* ctx; + if (skip_ref_cnt_vars.empty()) { + ctx = new ExecutorPrepareContext(program, bid); + } else { + ctx = new ExecutorPrepareContext(program, bid, skip_ref_cnt_vars[idx]); + } PADDLE_ENFORCE_LT(static_cast(bid), program.Size()); auto& block = program.Block(bid); for (auto& op_desc : block.AllOps()) { ctx->ops_.push_back(OpRegistry::CreateOp(*op_desc)); } result.push_back(std::shared_ptr(ctx)); + ++idx; } return result; } @@ -390,22 +423,23 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope, } int64_t max_memory_size = GetEagerDeletionThreshold(); - std::unique_ptr> gc; - // WhileOp would set keep_kids to false - // WhileGradOp would need the scopes created in WhileOp - // Perhaps, we should not perform eager deletion in WhileOp - // The scopes and variables created by WhileOp would be deleted - // in WhileGradOp. + std::unique_ptr gc; + // skip while_op and while_grad_op temporarily if (max_memory_size >= 0 && !keep_kids) { ctx->ResetReferenceCount(); #ifdef PADDLE_WITH_CUDA if (platform::is_gpu_place(place_)) { - gc.reset(new DefaultStreamGarbageCollector( - boost::get(place_), max_memory_size)); - } else { + if (IsFastEagerDeletionModeEnabled()) { + gc.reset(new UnsafeFastGPUGarbageCollector( + boost::get(place_), max_memory_size)); + } else { + gc.reset(new DefaultStreamGarbageCollector( + boost::get(place_), max_memory_size)); + } + } else if (platform::is_cpu_place(place_)) { #endif - gc.reset(new CPUGarbageCollector( - boost::get(place_), max_memory_size)); + gc.reset(new CPUGarbageCollector(boost::get(place_), + max_memory_size)); #ifdef PADDLE_WITH_CUDA } #endif @@ -414,17 +448,13 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope, for (auto& op : ctx->ops_) { op->Run(*local_scope, place_); - if (gc != nullptr) { + if (gc) { DeleteUnusedTensors(*local_scope, op.get(), gc.get(), - &(ctx->cur_ref_cnts_)); + &(ctx->runtime_ref_cnts_)); } } - if (gc != nullptr) { - gc->Wait(); - } else { - platform::DeviceContextPool::Instance().Get(place_)->Wait(); - } + platform::DeviceContextPool::Instance().Get(place_)->Wait(); if (local_scope != scope) { scope->DeleteScope(local_scope); @@ -481,7 +511,7 @@ void Executor::RunPreparedContext( void Executor::EnableMKLDNN(const ProgramDesc& program) { #ifdef PADDLE_WITH_MKLDNN - VLOG(30) << "use_mkldnn=True"; + VLOG(3) << "use_mkldnn=True"; for (size_t bid = 0; bid < program.Size(); ++bid) { auto* block = const_cast(program).MutableBlock(bid); for (auto* op : block->AllOps()) { diff --git a/paddle/fluid/framework/executor.h b/paddle/fluid/framework/executor.h index 36b36d49c2728..5a040ac641588 100644 --- a/paddle/fluid/framework/executor.h +++ b/paddle/fluid/framework/executor.h @@ -26,54 +26,22 @@ limitations under the License. */ namespace paddle { namespace framework { -extern void InitializeVariable(Variable* var, proto::VarType::Type var_type); - -template -std::unordered_map GetNonPersistableReferenceCount( - const ProgramDesc& prog, size_t block_id) { - auto& block = prog.Block(block_id); - std::unordered_map ref_cnts; - - auto update_ref_cnts = [&](OpDesc* op_desc, const VariableNameMap& name_map) { - for (auto& name_pair : name_map) { - for (auto& name : name_pair.second) { - auto* var_desc = block.FindVar(name); - if (var_desc == nullptr || var_desc->Persistable()) continue; - auto type = var_desc->Proto()->type().type(); - if (type != proto::VarType::LOD_TENSOR && - type != proto::VarType::SELECTED_ROWS) { - continue; - } - - auto it = ref_cnts.find(name); - if (it != ref_cnts.end()) { - ++it->second; - } else { - ref_cnts[name] = 1; - } - } - } - }; - - for (auto op_desc : block.AllOps()) { - update_ref_cnts(op_desc, op_desc->Inputs()); - update_ref_cnts(op_desc, op_desc->Outputs()); - } - return ref_cnts; -} struct ExecutorPrepareContext { - ExecutorPrepareContext(const framework::ProgramDesc& prog, size_t block_id); + ExecutorPrepareContext(const framework::ProgramDesc& prog, size_t block_id, + const std::vector& skip_ref_cnt_vars = + std::vector()); + ~ExecutorPrepareContext(); - void ResetReferenceCount() { cur_ref_cnts_ = ref_cnts_; } + void ResetReferenceCount() { runtime_ref_cnts_ = global_ref_cnts_; } const framework::ProgramDesc& prog_; size_t block_id_; std::vector> ops_; - std::unordered_map ref_cnts_; - std::unordered_map cur_ref_cnts_; + std::unordered_map global_ref_cnts_; + std::unordered_map runtime_ref_cnts_; }; class Executor { @@ -109,10 +77,14 @@ class Executor { const std::string& fetch_holder_name = "fetch"); static std::unique_ptr Prepare( - const ProgramDesc& program, int block_id); + const ProgramDesc& program, int block_id, + const std::vector& skip_ref_cnt_vars = + std::vector()); static std::vector> Prepare( - const ProgramDesc& program, const std::vector& block_ids); + const ProgramDesc& program, const std::vector& block_ids, + const std::vector>& skip_ref_cnt_vars = + std::vector>()); void CreateVariables(const ProgramDesc& pdesc, Scope* scope, int block_id); diff --git a/paddle/fluid/framework/executor_thread_worker.cc b/paddle/fluid/framework/executor_thread_worker.cc new file mode 100644 index 0000000000000..2eb9e564f8780 --- /dev/null +++ b/paddle/fluid/framework/executor_thread_worker.cc @@ -0,0 +1,643 @@ +/* Copyright (c) 2016 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/executor_thread_worker.h" +#include +#include "google/protobuf/io/zero_copy_stream_impl.h" +#include "google/protobuf/message.h" +#include "google/protobuf/text_format.h" + +#include "gflags/gflags.h" +#include "paddle/fluid/framework/feed_fetch_method.h" +#include "paddle/fluid/framework/feed_fetch_type.h" +#include "paddle/fluid/framework/lod_rank_table.h" +#include "paddle/fluid/framework/lod_tensor_array.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/framework/reader.h" +#include "paddle/fluid/framework/variable_helper.h" +#include "paddle/fluid/inference/io.h" +#include "paddle/fluid/platform/cpu_helper.h" +#include "paddle/fluid/platform/place.h" +#include "paddle/fluid/pybind/pybind.h" +namespace paddle { +namespace framework { + +#ifdef PADDLE_WITH_PSLIB +int DensePullThread::start() { + _running = true; + _t = std::thread(&DensePullThread::run, this); + return 0; +} + +void DensePullThread::run() { + while (_running) { + _pull_dense_status.resize(0); + for (auto& t : _dense_variable_name) { + if (check_update_param(t.first)) { + auto status = pull_dense(t.first); + _pull_dense_status.emplace_back(std::move(status)); + reset_thread_version(t.first); + } + } + if (_pull_dense_status.size() != 0) { + wait_all(); + } + + usleep(_sleep_time_ms * 1000); + } +} +bool DensePullThread::check_update_param(uint64_t table_id) { + { + std::lock_guard lock(_mutex_for_version); + auto& version = _training_versions[table_id]; + _current_version[table_id] = + *(std::min_element(version.begin(), version.end())); + } + if (_current_version[table_id] - _last_versions[table_id] < _threshold) { + return false; + } + return true; +} + +void DensePullThread::reset_thread_version(uint64_t table_id) { + std::lock_guard lock(_mutex_for_version); + _last_versions[table_id] = _current_version[table_id]; +} +std::future DensePullThread::pull_dense(uint64_t table_id) { + auto& regions = _regions[table_id]; + regions.clear(); + auto& variables = _dense_variable_name[table_id]; + regions.resize(variables.size()); + + for (auto i = 0u; i < variables.size(); ++i) { + auto& t = variables[i]; + Variable* var = _root_scope->FindVar(t); + LoDTensor* tensor = var->GetMutable(); + + float* w = tensor->data(); + paddle::ps::Region reg(w, tensor->numel()); + regions[i] = std::move(reg); + } + return _ps_client->pull_dense(regions.data(), regions.size(), table_id); +} + +void DensePullThread::wait_all() { + for (auto& t : _pull_dense_status) { + t.wait(); + auto status = t.get(); + if (status != 0) { + LOG(WARNING) << "pull dense failed times:" << ++_pull_dense_fail_times; + } + } + + if (_pull_dense_fail_times > 20) { + LOG(FATAL) << "pull dense failed times more than 20 times"; + exit(-1); + } + + _pull_dense_status.resize(0); +} + +void DensePullThread::increase_thread_version(int thread_id, + uint64_t table_id) { + std::lock_guard lock(_mutex_for_version); + _training_versions[table_id][thread_id]++; +} +#endif + +void ExecutorThreadWorker::CreateThreadOperators(const ProgramDesc& program) { + auto& block = program.Block(0); + op_names_.clear(); + for (auto& op_desc : block.AllOps()) { + std::unique_ptr local_op = OpRegistry::CreateOp(*op_desc); + op_names_.push_back(op_desc->Type()); + OperatorBase* local_op_ptr = local_op.release(); + ops_.push_back(local_op_ptr); + continue; + } +} + +void ExecutorThreadWorker::CreateThreadResource( + const framework::ProgramDesc& program, + const paddle::platform::Place& place) { + CreateThreadScope(program); + CreateThreadOperators(program); + SetMainProgram(program); + SetPlace(place); +} + +void ExecutorThreadWorker::CreateThreadScope(const ProgramDesc& program) { + auto& block = program.Block(0); + + PADDLE_ENFORCE_NOT_NULL( + root_scope_, "root_scope should be set before creating thread scope"); + + thread_scope_ = &root_scope_->NewScope(); + for (auto& var : block.AllVars()) { + if (var->Persistable()) { + auto* ptr = root_scope_->Var(var->Name()); + InitializeVariable(ptr, var->GetType()); + } else { + auto* ptr = thread_scope_->Var(var->Name()); + InitializeVariable(ptr, var->GetType()); + } + } +} + +void ExecutorThreadWorker::SetDataFeed( + const std::shared_ptr& datafeed) { + thread_reader_ = datafeed; +} + +void ExecutorThreadWorker::BindingDataFeedMemory() { + const std::vector& input_feed = + thread_reader_->GetUseSlotAlias(); + for (auto name : input_feed) { + thread_reader_->AddFeedVar(thread_scope_->Var(name), name); + } +} + +void ExecutorThreadWorker::SetFetchVarNames( + const std::vector& fetch_var_names) { + fetch_var_names_.clear(); + fetch_var_names_.insert(fetch_var_names_.end(), fetch_var_names.begin(), + fetch_var_names.end()); +} + +void ExecutorThreadWorker::SetDevice() { +#if defined _WIN32 || defined __APPLE__ + return; +#else + static unsigned concurrency_cap = std::thread::hardware_concurrency(); + int thread_id = this->thread_id_; + + if (static_cast(thread_id) < concurrency_cap) { + unsigned proc = thread_id; + + cpu_set_t mask; + CPU_ZERO(&mask); + CPU_SET(proc, &mask); + + if (-1 == sched_setaffinity(0, sizeof(mask), &mask)) { + VLOG(1) << "WARNING: Failed to set thread affinity for thread " + << thread_id; + } else { + CPU_ZERO(&mask); + if ((0 != sched_getaffinity(0, sizeof(mask), &mask)) || + (CPU_ISSET(proc, &mask) == 0)) { + VLOG(3) << "WARNING: Failed to set thread affinity for thread " + << thread_id; + } + } + } else { + VLOG(1) << "WARNING: Failed to set thread affinity for thread " + << thread_id; + } +#endif +} + +template +void print_lod_tensor(std::string var_name, const LoDTensor& lod_tensor) { + auto inspect = lod_tensor.data(); + auto element_num = lod_tensor.numel(); + + std::ostringstream sstream; + sstream << var_name << " (element num " << element_num << "): ["; + sstream << inspect[0]; + for (int j = 1; j < element_num; ++j) { + sstream << " " << inspect[j]; + } + sstream << "]"; + + std::cout << sstream.str() << std::endl; +} + +static void print_fetch_var(Scope* scope, const std::string& var_name) { + auto& tensor = scope->FindVar(var_name)->Get(); + +#define PrintLoDTensorCallback(cpp_type, proto_type) \ + do { \ + if (tensor.type() == proto_type) { \ + print_lod_tensor(var_name, tensor); \ + return; \ + } \ + } while (0) + + _ForEachDataType_(PrintLoDTensorCallback); + VLOG(1) << "print_fetch_var: unrecognized data type:" << tensor.type(); +} + +void ExecutorThreadWorker::TrainFiles() { + platform::SetNumThreads(1); + + // todo: configurable + SetDevice(); + + int fetch_var_num = fetch_var_names_.size(); + fetch_values_.clear(); + fetch_values_.resize(fetch_var_num); + + thread_reader_->Start(); + + int cur_batch; + int batch_cnt = 0; + while ((cur_batch = thread_reader_->Next()) > 0) { + // executor run here + for (auto& op : ops_) { + op->Run(*thread_scope_, place_); + } + + ++batch_cnt; + thread_scope_->DropKids(); + + if (debug_ == false || thread_id_ != 0) { + continue; + } + + for (int i = 0; i < fetch_var_num; ++i) { + print_fetch_var(thread_scope_, fetch_var_names_[i]); + } // end for (int i = 0...) + } // end while () +} + +void ExecutorThreadWorker::SetThreadId(int tid) { thread_id_ = tid; } + +void ExecutorThreadWorker::SetPlace(const platform::Place& place) { + place_ = place; +} + +void ExecutorThreadWorker::SetMainProgram( + const ProgramDesc& main_program_desc) { + main_program_.reset(new ProgramDesc(main_program_desc)); +} + +void ExecutorThreadWorker::SetRootScope(Scope* g_scope) { + root_scope_ = g_scope; +} + +#ifdef PADDLE_WITH_PSLIB +// AsyncExecutor +void AsyncExecutorThreadWorker::TrainFiles() { + SetDevice(); + + int fetch_var_num = fetch_var_names_.size(); + fetch_values_.clear(); + fetch_values_.resize(fetch_var_num); + + thread_reader_->Start(); + + int cur_batch; + int batch_cnt = 0; + while ((cur_batch = thread_reader_->Next()) > 0) { + // executor run here + TrainOneNetwork(); + + ++batch_cnt; + thread_scope_->DropKids(); + + if (debug_ == false || thread_id_ != 0) { + continue; + } + + for (int i = 0; i < fetch_var_num; ++i) { + print_fetch_var(thread_scope_, fetch_var_names_[i]); + } // end for (int i = 0...) + } // end while () +} + +void AsyncExecutorThreadWorker::SetPSlibPtr( + std::shared_ptr pslib_ptr) { + _pslib_ptr = pslib_ptr; +} +void AsyncExecutorThreadWorker::SetPullDenseThread( + std::shared_ptr dpt) { + _pull_dense_thread = dpt; +} +void AsyncExecutorThreadWorker::TrainOneNetwork() { + PrepareParams(); + + for (auto& op : ops_) { + if (op->Type().find("sgd") != std::string::npos) { + continue; + } + bool need_skip = false; + for (auto t = 0u; t < _param_config->skip_op.size(); ++t) { + if (op->Type().find(_param_config->skip_op[t]) != std::string::npos) { + need_skip = true; + break; + } + } + if (!need_skip) { + op->Run(*thread_scope_, place_); + } + } + UpdateParams(); +} + +void AsyncExecutorThreadWorker::SetParamConfig( + AsyncWorkerParamConfig* param_config) { + _param_config = param_config; +} + +void AsyncExecutorThreadWorker::PrepareParams() { + for (auto table_id : _param_config->sparse_table_id) { + PullSparse(table_id); + for (auto& t : _pull_sparse_status) { + t.wait(); + auto status = t.get(); + if (status != 0) { + LOG(ERROR) << "pull sparse failed, status[" << status << "]"; + exit(-1); + } + } + } + _pull_sparse_status.resize(0); + + for (auto table_id : _param_config->sparse_table_id) { + FillSparse(table_id); + } +} + +void AsyncExecutorThreadWorker::UpdateParams() { + for (auto i : _param_config->sparse_table_id) { + PushSparse(i); + } + for (auto i : _param_config->dense_table_id) { + PushDense(i); + } + int32_t tmp_push_dense_wait_times = -1; + int32_t tmp_push_sparse_wait_times = -1; + static uint32_t push_dense_wait_times = + static_cast(tmp_push_dense_wait_times); + static uint32_t push_sparse_wait_times = + static_cast(tmp_push_sparse_wait_times); + + if (_push_dense_status.size() >= push_dense_wait_times) { + for (auto& t : _push_dense_status) { + t.wait(); + } + _push_dense_status.resize(0); + } + if (tmp_push_dense_wait_times == -1) { + _push_dense_status.resize(0); + } + if (_push_sparse_status.size() >= push_sparse_wait_times) { + for (auto& t : _push_sparse_status) { + t.wait(); + } + _push_sparse_status.resize(0); + } + if (tmp_push_sparse_wait_times == -1) { + _push_sparse_status.resize(0); + } + for (auto dense_table_id : _param_config->dense_table_id) { + _pull_dense_thread->increase_thread_version(thread_id_, dense_table_id); + } +} + +void AsyncExecutorThreadWorker::PushDense(int table_id) { + std::vector regions; + for (auto& t : _param_config->dense_gradient_variable_name[table_id]) { + Variable* var = thread_scope_->FindVar(t); + CHECK(var != nullptr) << "var[" << t << "] not found"; + LoDTensor* tensor = var->GetMutable(); + int count = tensor->numel(); + float* g = tensor->data(); + paddle::ps::Region reg(g, count); + regions.emplace_back(std::move(reg)); + } + + auto status = _pslib_ptr->_worker_ptr->push_dense(regions.data(), + regions.size(), table_id); + _push_dense_status.push_back(std::move(status)); +} + +void AsyncExecutorThreadWorker::PullSparse(int table_id) { + auto& features = _features[table_id]; + auto& feature_value = _feature_value[table_id]; + auto fea_dim = _param_config->fea_dim; + // slot id starts from 1 + features.clear(); + features.resize(0); + features.reserve(MAX_FEASIGN_NUM); + const std::vector& feed_vec = thread_reader_->GetUseSlotAlias(); + // slot_idx = 0 is label TODO + for (auto slot_idx = 1u; slot_idx < feed_vec.size(); ++slot_idx) { + Variable* var = thread_scope_->FindVar(feed_vec[slot_idx]); + LoDTensor* tensor = var->GetMutable(); + int64_t* ids = tensor->data(); + int len = tensor->numel(); + for (auto i = 0u; i < len; ++i) { + // todo(colourful-tree): current trick - filter feasign=use_slot_mod( + // bug: datafeed fill use_slot_mod for empty slot) + if (ids[i] == 0u) { + continue; + } + features.push_back(static_cast(ids[i])); + } + } + check_pull_push_memory(features, &feature_value, fea_dim); + + std::vector pull_feature_value; + for (auto i = 0u; i < features.size(); ++i) { + pull_feature_value.push_back(feature_value[i].data()); + } + + auto status = _pslib_ptr->_worker_ptr->pull_sparse( + pull_feature_value.data(), table_id, features.data(), features.size()); + _pull_sparse_status.push_back(std::move(status)); + + auto& push_g = _feature_push_value[table_id]; + check_pull_push_memory(features, &push_g, fea_dim); + + collect_feasign_info(table_id); +} + +void AsyncExecutorThreadWorker::FillSparse(int table_id) { + auto slot_dim = _param_config->slot_dim; + auto fea_dim = _param_config->fea_dim; + auto& features = _features[table_id]; + auto& fea_value = _feature_value[table_id]; + + CHECK(features.size() > 0) << "feature size check failed"; + + auto fea_idx = 0u; + + std::vector init_value(fea_dim); + + const std::vector& feed_vec = thread_reader_->GetUseSlotAlias(); + // slot_idx = 0 is label TODO + for (auto slot_idx = 1u; slot_idx < feed_vec.size(); ++slot_idx) { + Variable* var = thread_scope_->FindVar(feed_vec[slot_idx]); + LoDTensor* tensor = var->GetMutable(); + int64_t* ids = tensor->data(); + int len = tensor->numel(); + Variable* var_emb = thread_scope_->FindVar( + _param_config->slot_input_vec[table_id][slot_idx - 1]); + LoDTensor* tensor_emb = var_emb->GetMutable(); + float* ptr = + tensor_emb->mutable_data({len, slot_dim}, platform::CPUPlace()); + memset(ptr, 0, sizeof(float) * len * slot_dim); + auto& tensor_lod = tensor->lod()[0]; + + LoD data_lod{tensor_lod}; + tensor_emb->set_lod(data_lod); + + for (auto index = 0u; index < len; ++index) { + if (ids[index] == 0u) { + memcpy(ptr + slot_dim * index, init_value.data() + 2, + sizeof(float) * slot_dim); + continue; + } + memcpy(ptr + slot_dim * index, fea_value[fea_idx].data() + 2, + sizeof(float) * slot_dim); + fea_idx++; + } + } +} + +void AsyncExecutorThreadWorker::PushSparse(int table_id) { + auto slot_dim = _param_config->slot_dim; + auto fea_dim = _param_config->fea_dim; + auto& features = _features[table_id]; + auto& push_g = _feature_push_value[table_id]; + check_pull_push_memory(features, &push_g, fea_dim); + CHECK(push_g.size() == features.size() + 1) + << "push_g size:" << push_g.size() + << " features size:" << features.size(); + uint64_t fea_idx = 0u; + auto& fea_info = _fea_info[table_id]; + int offset = 2; + const std::vector& feed_vec = thread_reader_->GetUseSlotAlias(); + // slot_idx = 0 is label + for (auto slot_idx = 1u; slot_idx < feed_vec.size(); ++slot_idx) { + if (_param_config->slot_alias_to_table.find(feed_vec[slot_idx]) == + _param_config->slot_alias_to_table.end()) { + LOG(ERROR) << "ERROR slot_idx:" << slot_idx + << " name:" << feed_vec[slot_idx]; + } else if (_param_config->slot_alias_to_table[feed_vec[slot_idx]] != + table_id) { + continue; + } + Variable* g_var = thread_scope_->FindVar( + _param_config->gradient_var[table_id][slot_idx - 1]); + CHECK(g_var != nullptr) + << "var[" << _param_config->gradient_var[table_id][slot_idx - 1] + << "] not found"; + LoDTensor* g_tensor = g_var->GetMutable(); + if (g_tensor == NULL) { + LOG(ERROR) << "var[" + << _param_config->gradient_var[table_id][slot_idx - 1] + << "] not found"; + exit(-1); + } + float* g = g_tensor->data(); + + Variable* var = thread_scope_->FindVar(feed_vec[slot_idx]); + CHECK(var != nullptr) << "var[" << feed_vec[slot_idx] << "] not found"; + LoDTensor* tensor = var->GetMutable(); + if (tensor == NULL) { + LOG(ERROR) << "var[" << feed_vec[slot_idx] << "] not found"; + exit(-1); + } + int len = tensor->numel(); + CHECK(slot_dim * len == g_tensor->numel()) + << "len:" << len << " g_numel:" << g_tensor->numel(); + CHECK(len == tensor->numel()) << "len:" << len + << "t_numel:" << tensor->numel(); + int64_t* ids = tensor->data(); + for (auto id_idx = 0u; id_idx < len; ++id_idx) { + if (ids[id_idx] == 0) { + g += slot_dim; + continue; + } + memcpy(push_g[fea_idx].data() + offset, g, sizeof(float) * slot_dim); + push_g[fea_idx][0] = 1.0f; + CHECK(fea_idx < fea_info.size()) << "fea_idx:" << fea_idx + << " size:" << fea_info.size(); + push_g[fea_idx][1] = static_cast(fea_info[fea_idx].label); + g += slot_dim; + fea_idx++; + } + } + CHECK(fea_idx == features.size()) << "fea_idx:" << fea_idx + << " features size:" << features.size(); + CHECK_GT(features.size(), 0); + + std::vector push_g_vec; + for (auto i = 0u; i < features.size(); ++i) { + push_g_vec.push_back(push_g[i].data()); + } + auto status = _pslib_ptr->_worker_ptr->push_sparse( + table_id, features.data(), (const float**)push_g_vec.data(), + features.size()); + _push_sparse_status.push_back(std::move(status)); +} + +void AsyncExecutorThreadWorker::collect_feasign_info(int table_id) { + auto& fea_info = _fea_info[table_id]; + auto& feature = _features[table_id]; + fea_info.resize(feature.size()); + const std::vector& feed_vec = thread_reader_->GetUseSlotAlias(); + Variable* var = thread_scope_->FindVar(feed_vec[0]); + LoDTensor* tensor = var->GetMutable(); + int64_t* label = tensor->data(); + + int global_index = 0; + for (auto slot_idx = 1u; slot_idx < feed_vec.size(); ++slot_idx) { + Variable* var = thread_scope_->FindVar(feed_vec[slot_idx]); + LoDTensor* tensor = var->GetMutable(); + int64_t* ids = tensor->data(); + + int fea_idx = 0; + for (auto ins_idx = 1u; ins_idx < tensor->lod()[0].size(); ++ins_idx) { + for (; fea_idx < tensor->lod()[0][ins_idx]; ++fea_idx) { + if (ids[fea_idx] == 0u) { + continue; + } + FeasignInfo info{slot_idx, ins_idx, label[ins_idx - 1]}; + + fea_info[global_index++] = std::move(info); + } + } + } + CHECK(global_index == feature.size()) + << "expect fea info size:" << feature.size() << " real:" << global_index; +} + +void AsyncExecutorThreadWorker::check_pull_push_memory( + const std::vector& features, + std::vector>* push_g, int dim) { + push_g->resize(features.size() + 1); + for (auto& t : *push_g) { + t.resize(dim); + } +} + +void AsyncExecutorThreadWorker::check_pull_push_memory( + const std::vector& features, std::vector* push_g, + int dim) { + if (features.size() > push_g->size()) { + push_g->reserve(features.size() + 1); + auto size = features.size() - push_g->size() + 1; + for (auto i = 0u; i < size; ++i) { + float* ptr = new float[dim]; + push_g->push_back(ptr); + } + } +} +#endif + +} // einit_modelnd namespace framework +} // end namespace paddle diff --git a/paddle/fluid/framework/executor_thread_worker.h b/paddle/fluid/framework/executor_thread_worker.h new file mode 100644 index 0000000000000..30b81ad88035e --- /dev/null +++ b/paddle/fluid/framework/executor_thread_worker.h @@ -0,0 +1,243 @@ +/* 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 +#include +#include // NOLINT +#include +#include +#include // NOLINT +#include +#include "paddle/fluid/framework/data_feed.h" +#include "paddle/fluid/framework/executor.h" +#include "paddle/fluid/framework/program_desc.h" +#include "paddle/fluid/framework/scope.h" +#ifdef PADDLE_WITH_PSLIB +#include +#endif + +namespace paddle { +namespace framework { + +void CreateTensor(Variable* var, proto::VarType::Type var_type); +#ifdef PADDLE_WITH_PSLIB +static const uint32_t MAX_FEASIGN_NUM = 1000 * 100 * 100; + +struct AsyncWorkerParamConfig { + int slot_dim; + int fea_dim; + int32_t tmp_push_dense_wait_times; + int32_t tmp_push_sparse_wait_times; + + std::vector skip_op; + + std::map> dense_variable_name; + std::map> dense_gradient_variable_name; + std::vector dense_table_id; + // fea_dim for each dense table + std::vector dense_table_size; + std::vector sparse_table_id; + std::map> slot_input_vec; + std::map> gradient_var; + std::map slot_alias_to_table; +}; + +struct DensePullThreadParam { + std::shared_ptr ps_client; + int threshold; + int training_thread_num; + Scope* root_scope; + std::map>* dense_params; + int sleep_time_ms = 2; +}; + +class DensePullThread { + public: + explicit DensePullThread(const DensePullThreadParam& param) + : _running(false) { + _ps_client = param.ps_client; + _threshold = param.threshold; + _thread_num = param.training_thread_num; + _root_scope = param.root_scope; + _sleep_time_ms = param.sleep_time_ms; + + for (auto& t : *param.dense_params) { + _dense_variable_name[t.first].insert(_dense_variable_name[t.first].end(), + t.second.begin(), t.second.end()); + _training_versions[t.first].resize(_thread_num, 0); + _last_versions[t.first] = 0; + _current_version[t.first] = 0; + } + } + + int start(); + + void stop() { + if (_running) { + _running = false; + _t.join(); + } + } + + void increase_thread_version(int thread_id, uint64_t table_id); + void reset_thread_version(uint64_t table_id); + std::future pull_dense(uint64_t table_id); + void pull_dense2(uint64_t table_id); + void wait_all(); + + private: + void run(); + bool check_update_param(uint64_t table_id); + + private: + std::shared_ptr _ps_client; + int _thread_num; + int _threshold; + int _sleep_time_ms; + Scope* _root_scope; + bool _running; + + std::map _last_versions; + std::map _current_version; + std::mutex _mutex_for_version; + std::map> _training_versions; + std::map> _dense_variable_name; + + std::thread _t; + + std::vector<::std::future> _pull_dense_status; + + std::map> _regions; + uint32_t _pull_dense_fail_times = 0; + + std::vector _base_norm_param; + std::vector _mean; + std::vector _scale; + float _squared_sum_epsilon = 1e-4; + std::mutex _mutex_for_mean_scale; + + float _total_batch_num = 0; +}; +#endif + +class ExecutorThreadWorker { + public: + ExecutorThreadWorker() + : thread_id_(-1), root_scope_(NULL), thread_scope_(NULL), debug_(false) {} + virtual ~ExecutorThreadWorker() {} + + void CreateThreadResource(const framework::ProgramDesc& program, + const paddle::platform::Place& place); + void SetThreadId(int tid); + void SetDebug(const bool debug) { debug_ = debug; } + void SetRootScope(Scope* g_scope); + // set cpu device in this function + // cpu binding is used by default + void SetDevice(); + // since we read data into memory that can not be accessed by program + // we need to bind memory of data with corresponding variables in program + // this function should be called after data feed is set + void BindingDataFeedMemory(); + // set data feed declared in executor + void SetDataFeed(const std::shared_ptr& datafeed); + // A multi-thread training function + virtual void TrainFiles(); + // set fetch variable names from python interface assigned by users + void SetFetchVarNames(const std::vector& fetch_var_names); +#ifdef PADDLE_WITH_PSLIB + virtual void SetPSlibPtr( + std::shared_ptr pslib_ptr) {} + virtual void SetPullDenseThread(std::shared_ptr dpt) {} + virtual void SetParamConfig(AsyncWorkerParamConfig* param_config) {} +#endif + + private: + void CreateThreadScope(const framework::ProgramDesc& program); + void CreateThreadOperators(const framework::ProgramDesc& program); + void SetMainProgram(const ProgramDesc& main_program_desc); + void SetPlace(const paddle::platform::Place& place); + + protected: + // thread index + std::shared_ptr thread_reader_; // shared queue, thread buffer + int thread_id_; + // operator name + std::vector op_names_; + // thread level, local operators for forward and backward + std::vector ops_; + // main program for training + std::unique_ptr main_program_; + // execution place + platform::Place place_; + // root scope for model parameters + Scope* root_scope_; + // a thread scope, father scope is global score which is shared + Scope* thread_scope_; + std::vector fetch_var_names_; + std::vector> fetch_values_; + bool debug_; +}; + +#ifdef PADDLE_WITH_PSLIB +class AsyncExecutorThreadWorker : public ExecutorThreadWorker { + public: + AsyncExecutorThreadWorker() {} + virtual ~AsyncExecutorThreadWorker() {} + void SetPSlibPtr(std::shared_ptr pslib_ptr); + void SetPullDenseThread(std::shared_ptr dpt); + void SetParamConfig(AsyncWorkerParamConfig* param_config); + void TrainFiles(); + void TrainOneNetwork(); + void PrepareParams(); + void UpdateParams(); + void PullSparse(int table_id); + void FillSparse(int table_id); + void PushSparse(int table_id); + void PushDense(int table_id); + + void check_pull_push_memory(const std::vector& features, + std::vector* push_g, int dim); + void check_pull_push_memory(const std::vector& features, + std::vector>* push_g, int dim); + void collect_feasign_info(int table_id); + + private: + struct FeasignInfo { + uint32_t slot; + uint32_t ins; + int64_t label; + }; + + std::map> _features; + std::map> _fea_info; + std::map>> _feature_value; + std::map>> _feature_push_value; + + std::shared_ptr _pslib_ptr; + + std::shared_ptr _pull_dense_thread; + + std::vector<::std::future> _pull_sparse_status; + std::vector<::std::future> _pull_dense_status; + std::vector<::std::future> _push_sparse_status; + std::vector<::std::future> _push_dense_status; + + AsyncWorkerParamConfig* _param_config; +}; +#endif + +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/feed_fetch_method.cc b/paddle/fluid/framework/feed_fetch_method.cc index 1f3c19c0d5901..6338be75a4b1d 100644 --- a/paddle/fluid/framework/feed_fetch_method.cc +++ b/paddle/fluid/framework/feed_fetch_method.cc @@ -16,7 +16,9 @@ limitations under the License. */ #include #include #include "glog/logging.h" +#include "paddle/fluid/framework/var_type.h" #include "paddle/fluid/framework/variable.h" +#include "paddle/fluid/platform/place.h" namespace paddle { namespace framework { @@ -25,7 +27,7 @@ void SetFeedVariable(Scope* scope, const LoDTensor& input, const std::string& var_name, size_t index) { // If var_name Variable is not found in GlobalScope, a new variable will // be created. - VLOG(30) << "SetFeedVariable name=" << var_name << " index=" << index; + VLOG(3) << "SetFeedVariable name=" << var_name << " index=" << index; Variable* g_feed_value = scope->Var(var_name); auto& feed_inputs = *(g_feed_value->GetMutable()); if (index >= feed_inputs.size()) { @@ -47,11 +49,18 @@ LoDTensor& GetFetchVariable(const Scope& scope, const std::string& var_name, typeid(FeedFetchList).name()); auto& fetch_outputs = *g_fetch_value->GetMutable(); auto& tensor = fetch_outputs[index]; - VLOG(30) << "Fetch " << var_name << " with index " << index - << " shape= " << tensor.dims(); + VLOG(3) << "Fetch " << var_name << " with index " << index + << " shape= " << tensor.dims(); PADDLE_ENFORCE_LT(index, fetch_outputs.size()); return tensor; } +LoDTensor& GetVariableTensor(const Scope& scope, const std::string& var_name) { + Variable* var = scope.FindVar(var_name); + PADDLE_ENFORCE(var, "%s no in scope", var_name); + PADDLE_ENFORCE(var->IsType(), "Only support lod tensor now."); + return *var->GetMutable(); +} + } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/feed_fetch_method.h b/paddle/fluid/framework/feed_fetch_method.h index 7f504bfd23286..031f8e01aa612 100644 --- a/paddle/fluid/framework/feed_fetch_method.h +++ b/paddle/fluid/framework/feed_fetch_method.h @@ -27,5 +27,7 @@ void SetFeedVariable(Scope* scope, const LoDTensor& input, LoDTensor& GetFetchVariable(const Scope& scope, const std::string& var_name, size_t index); +LoDTensor& GetVariableTensor(const Scope& scope, const std::string& var_name); + } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/garbage_collector.cc b/paddle/fluid/framework/garbage_collector.cc new file mode 100644 index 0000000000000..54d9d0dc018b0 --- /dev/null +++ b/paddle/fluid/framework/garbage_collector.cc @@ -0,0 +1,89 @@ +// 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 +#ifdef PADDLE_WITH_CUDA +#include "paddle/fluid/platform/cuda_device_guard.h" +#endif +#include "paddle/fluid/framework/garbage_collector.h" + +namespace paddle { +namespace framework { + +GarbageCollector::GarbageCollector(const platform::Place &place, + size_t max_memory_size) + : max_memory_size_((std::max)(max_memory_size, static_cast(1))) { + garbages_.reset(new GarbageQueue()); + dev_ctx_ = platform::DeviceContextPool::Instance().Get(place); +} + +CPUGarbageCollector::CPUGarbageCollector(const platform::CPUPlace &place, + size_t max_memory_size) + : GarbageCollector(place, max_memory_size) {} + +void CPUGarbageCollector::ClearCallback(const std::function &callback) { + callback(); +} + +#ifdef PADDLE_WITH_CUDA +UnsafeFastGPUGarbageCollector::UnsafeFastGPUGarbageCollector( + const platform::CUDAPlace &place, size_t max_memory_size) + : GarbageCollector(place, max_memory_size) {} + +void UnsafeFastGPUGarbageCollector::ClearCallback( + const std::function &callback) { + callback(); +} + +DefaultStreamGarbageCollector::DefaultStreamGarbageCollector( + const platform::CUDAPlace &place, size_t max_memory_size) + : GarbageCollector(place, max_memory_size) {} + +void DefaultStreamGarbageCollector::Wait() const { + static_cast(this->dev_ctx_) + ->WaitStreamCallback(); +} + +void DefaultStreamGarbageCollector::ClearCallback( + const std::function &callback) { + static_cast(this->dev_ctx_) + ->AddStreamCallback(callback); +} + +StreamGarbageCollector::StreamGarbageCollector(const platform::CUDAPlace &place, + size_t max_memory_size) + : GarbageCollector(place, max_memory_size) { + platform::CUDADeviceGuard guard(place.device); + PADDLE_ENFORCE(cudaStreamCreate(&stream_)); + callback_manager_.reset(new platform::StreamCallbackManager(stream_)); +} + +StreamGarbageCollector::~StreamGarbageCollector() { + auto place = boost::get(this->dev_ctx_->GetPlace()); + platform::CUDADeviceGuard guard(place.device); + PADDLE_ENFORCE(cudaStreamSynchronize(stream_)); + PADDLE_ENFORCE(cudaStreamDestroy(stream_)); +} + +cudaStream_t StreamGarbageCollector::stream() const { return stream_; } + +void StreamGarbageCollector::Wait() const { callback_manager_->Wait(); } + +void StreamGarbageCollector::ClearCallback( + const std::function &callback) { + callback_manager_->AddCallback(callback); +} +#endif +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/garbage_collector.h b/paddle/fluid/framework/garbage_collector.h index 818b3334ea417..2768671029c06 100644 --- a/paddle/fluid/framework/garbage_collector.h +++ b/paddle/fluid/framework/garbage_collector.h @@ -14,7 +14,6 @@ #pragma once -#include #include #include #include @@ -24,134 +23,74 @@ namespace paddle { namespace framework { -// T should have memory_size() and clear() method -template class GarbageCollector { public: - GarbageCollector(const platform::Place &place, size_t max_memory_size) - : max_memory_size_((std::max)(max_memory_size, static_cast(1))) { - garbages_.reset(new std::deque()); - dev_ctx_ = platform::DeviceContextPool::Instance().Get(place); - } + using GarbageQueue = std::deque>; - virtual ~GarbageCollector() {} + GarbageCollector(const platform::Place &place, size_t max_memory_size); - void Reset() { - std::lock_guard guard(mutex_); - garbages_.reset(new std::deque()); - cur_memory_size_ = 0; - } + virtual ~GarbageCollector() = default; + + virtual void Wait() const {} template - void Add(const Container &objs) { - Add(objs, []() {}); - } + void Add(Container &&objs); template - void Add(const Container &objs, Callback &&callback) { - std::shared_ptr> clear_deque; - { - std::lock_guard guard(mutex_); - for (auto *obj : objs) { - garbages_->push_back(obj); - cur_memory_size_ += obj->memory_size(); - } - if (cur_memory_size_ >= max_memory_size_) { - cur_memory_size_ = 0; - clear_deque = garbages_; - garbages_.reset(new std::deque()); - } - } - - if (clear_deque != nullptr) { - callback(); - ClearCallback([=]() { - for (auto *obj : *clear_deque) obj->clear(); - }); - } - } - - virtual void Wait() const {} + void Add(Container &&objs, Callback &&callback); protected: virtual void ClearCallback(const std::function &callback) = 0; platform::DeviceContext *dev_ctx_; - std::shared_ptr> garbages_; + std::unique_ptr garbages_; mutable std::mutex mutex_; const size_t max_memory_size_; - size_t cur_memory_size_ = 0; + size_t cur_memory_size_{0}; }; -template -class CPUGarbageCollector : public GarbageCollector { +class CPUGarbageCollector : public GarbageCollector { public: - CPUGarbageCollector(const platform::CPUPlace &place, size_t max_memory_size) - : GarbageCollector(place, max_memory_size) {} + CPUGarbageCollector(const platform::CPUPlace &place, size_t max_memory_size); protected: - void ClearCallback(const std::function &callback) override { - callback(); - } + void ClearCallback(const std::function &callback) override; }; #ifdef PADDLE_WITH_CUDA -template -class DefaultStreamGarbageCollector : public GarbageCollector { +class UnsafeFastGPUGarbageCollector : public GarbageCollector { public: - DefaultStreamGarbageCollector(const platform::CUDAPlace &place, - size_t max_memory_size) - : GarbageCollector(place, max_memory_size) {} + UnsafeFastGPUGarbageCollector(const platform::CUDAPlace &place, + size_t max_memory_size); - cudaStream_t stream() const { - return static_cast(this->dev_ctx_) - ->stream(); - } + protected: + void ClearCallback(const std::function &callback) override; +}; - void Wait() const override { - this->dev_ctx_->Wait(); - static_cast(this->dev_ctx_) - ->WaitStreamCallback(); - } +class DefaultStreamGarbageCollector : public GarbageCollector { + public: + DefaultStreamGarbageCollector(const platform::CUDAPlace &place, + size_t max_memory_size); + + void Wait() const override; protected: - void ClearCallback(const std::function &callback) override { - static_cast(this->dev_ctx_) - ->AddStreamCallback(callback); - } + void ClearCallback(const std::function &callback) override; }; -template -class StreamGarbageCollector : public GarbageCollector { +class StreamGarbageCollector : public GarbageCollector { public: StreamGarbageCollector(const platform::CUDAPlace &place, - size_t max_memory_size) - : GarbageCollector(place, max_memory_size) { - PADDLE_ENFORCE(cudaSetDevice(place.device)); - PADDLE_ENFORCE(cudaStreamCreate(&stream_)); - callback_manager_.reset(new platform::StreamCallbackManager(stream_)); - } + size_t max_memory_size); - ~StreamGarbageCollector() { - auto place = boost::get(this->dev_ctx_->GetPlace()); - PADDLE_ENFORCE(cudaSetDevice(place.device)); - PADDLE_ENFORCE(cudaStreamSynchronize(stream_)); - PADDLE_ENFORCE(cudaStreamDestroy(stream_)); - } + ~StreamGarbageCollector(); - void Wait() const override { - PADDLE_ENFORCE(cudaStreamSynchronize(stream_)); - std::lock_guard guard(this->mutex_); - callback_manager_->Wait(); - } + void Wait() const override; - cudaStream_t stream() const { return stream_; } + cudaStream_t stream() const; protected: - void ClearCallback(const std::function &callback) override { - std::lock_guard guard(this->mutex_); - callback_manager_->AddCallback(callback); - } + void ClearCallback(const std::function &callback) override; private: cudaStream_t stream_; @@ -159,5 +98,33 @@ class StreamGarbageCollector : public GarbageCollector { }; #endif +template +void GarbageCollector::Add(Container &&objs) { + Add(std::forward(objs), []() {}); +} + +template +void GarbageCollector::Add(Container &&objs, Callback &&callback) { + GarbageQueue *garbage_queue = nullptr; + { + std::lock_guard guard(mutex_); + for (auto &obj : objs) { + if (!obj) continue; + cur_memory_size_ += obj->size(); + garbages_->push_back(std::move(obj)); + } + if (cur_memory_size_ >= max_memory_size_) { + cur_memory_size_ = 0; + garbage_queue = garbages_.release(); + garbages_.reset(new GarbageQueue()); + } + } + + if (garbage_queue) { + callback(); + ClearCallback([garbage_queue]() { delete garbage_queue; }); + } +} + } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/ir/CMakeLists.txt b/paddle/fluid/framework/ir/CMakeLists.txt index 883575e41db2d..b7f7e2ee8ef59 100644 --- a/paddle/fluid/framework/ir/CMakeLists.txt +++ b/paddle/fluid/framework/ir/CMakeLists.txt @@ -42,6 +42,9 @@ pass_library(multi_batch_merge_pass base) pass_library(conv_bn_fuse_pass inference) pass_library(seqconv_eltadd_relu_fuse_pass inference) pass_library(is_test_pass base) +pass_library(conv_elementwise_add_act_fuse_pass inference) +pass_library(conv_elementwise_add2_act_fuse_pass inference) +pass_library(conv_elementwise_add_fuse_pass inference) if(WITH_MKLDNN) pass_library(mkldnn_placement_pass base) pass_library(depthwise_conv_mkldnn_pass base) diff --git a/paddle/fluid/framework/ir/attention_lstm_fuse_pass.cc b/paddle/fluid/framework/ir/attention_lstm_fuse_pass.cc index c436dd414d01a..a9897e0bb884c 100644 --- a/paddle/fluid/framework/ir/attention_lstm_fuse_pass.cc +++ b/paddle/fluid/framework/ir/attention_lstm_fuse_pass.cc @@ -147,19 +147,19 @@ void PrepareParameters(Graph* graph, const Param& param) { scope->Var(param.LSTMX)->GetMutable(); scope->Var(param.LSTMOUT)->GetMutable(); -#define GATE_W(name__) \ - auto* W_##name__##_w0 = scope->FindVar(#name__ ".w_0"); \ - auto* W_##name__##_w1 = scope->FindVar(#name__ ".w_1"); \ - auto* W_##name__##_b0 = scope->FindVar(#name__ ".b_0"); \ - CHECK_P3(W_##name__##_w0, W_##name__##_w1, W_##name__##_b0); \ - VLOG(40) << #name__ "_w0" \ - << " shape: " << W_##name__##_w0->Get().dims(); \ - VLOG(40) << #name__ "_w1" \ - << " shape: " << W_##name__##_w1->Get().dims(); \ - VLOG(40) << #name__ "_b0" \ - << " shape: " << W_##name__##_b0->Get().dims(); \ - auto& W_##name__##_w0_t = W_##name__##_w0->Get(); \ - auto& W_##name__##_w1_t = W_##name__##_w1->Get(); \ +#define GATE_W(name__) \ + auto* W_##name__##_w0 = scope->FindVar(#name__ ".w_0"); \ + auto* W_##name__##_w1 = scope->FindVar(#name__ ".w_1"); \ + auto* W_##name__##_b0 = scope->FindVar(#name__ ".b_0"); \ + CHECK_P3(W_##name__##_w0, W_##name__##_w1, W_##name__##_b0); \ + VLOG(4) << #name__ "_w0" \ + << " shape: " << W_##name__##_w0->Get().dims(); \ + VLOG(4) << #name__ "_w1" \ + << " shape: " << W_##name__##_w1->Get().dims(); \ + VLOG(4) << #name__ "_b0" \ + << " shape: " << W_##name__##_b0->Get().dims(); \ + auto& W_##name__##_w0_t = W_##name__##_w0->Get(); \ + auto& W_##name__##_w1_t = W_##name__##_w1->Get(); \ auto& W_##name__##_b0_t = W_##name__##_b0->Get(); GATE_W(forget); @@ -208,7 +208,7 @@ void PrepareLSTMWeight(const LoDTensor& W_forget_w0, int D = W_forget_w0.dims()[0]; int M = W_forget_w1.dims()[0]; out->Resize(make_ddim({D + M, 4 * D})); - VLOG(30) << "LSTMWeight resized to " << out->dims(); + VLOG(3) << "LSTMWeight resized to " << out->dims(); float* out_data = out->mutable_data(platform::CPUPlace()); std::array tensors{ diff --git a/paddle/fluid/framework/ir/conv_bias_mkldnn_fuse_pass.cc b/paddle/fluid/framework/ir/conv_bias_mkldnn_fuse_pass.cc index c9c4d5afe5a0c..d4a701e0b173a 100644 --- a/paddle/fluid/framework/ir/conv_bias_mkldnn_fuse_pass.cc +++ b/paddle/fluid/framework/ir/conv_bias_mkldnn_fuse_pass.cc @@ -46,18 +46,20 @@ std::unique_ptr ConvBiasFusePass::ApplyImpl( auto* scope = param_scope(); PADDLE_ENFORCE(scope); + std::string type = is_conv3d() ? "conv3d" : "conv2d"; + GraphPatternDetector gpd; auto* conv_input = gpd.mutable_pattern() ->NewNode(patterns::PDNodeName(name_scope_, "conv_input")) ->AsInput() - ->assert_is_op_input("conv2d", "Input"); + ->assert_is_op_input(type, "Input"); patterns::ConvBias conv_bias_pattern(gpd.mutable_pattern(), name_scope_); - conv_bias_pattern(conv_input); + conv_bias_pattern(conv_input, is_conv3d()); int found_conv_bias_count = 0; auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph, Graph* g) { - VLOG(40) << "handle ConvBias fuse"; + VLOG(4) << "handle ConvBias fuse"; GET_IR_NODE_FROM_SUBGRAPH(conv_weight, conv_weight, conv_bias_pattern); // Filter GET_IR_NODE_FROM_SUBGRAPH(conv_out, conv_out, conv_bias_pattern); // tmp @@ -74,7 +76,7 @@ std::unique_ptr ConvBiasFusePass::ApplyImpl( // check if fuse can be done and if MKL-DNN should be used FuseOptions fuse_option = FindFuseOption(*conv, *eltwise); if (fuse_option == DO_NOT_FUSE || fuse_option == FUSE_NATIVE) { - VLOG(30) << "do not perform conv+bias fuse"; + VLOG(3) << "do not perform conv+bias fuse"; return; } @@ -109,7 +111,7 @@ std::unique_ptr ConvBiasFusePass::ApplyImpl( desc.SetInput("Filter", std::vector({conv_weight->Name()})); desc.SetInput("Bias", std::vector({eltwise_bias->Name()})); desc.SetOutput("Output", std::vector({eltwise_out->Name()})); - desc.SetType("conv2d"); + desc.SetType(type); for (auto& attr : conv->Op()->GetAttrMap()) { desc.SetAttr(attr.first, attr.second); @@ -135,3 +137,5 @@ std::unique_ptr ConvBiasFusePass::ApplyImpl( } // namespace paddle REGISTER_PASS(conv_bias_mkldnn_fuse_pass, paddle::framework::ir::ConvBiasFusePass); +REGISTER_PASS(conv3d_bias_mkldnn_fuse_pass, + paddle::framework::ir::Conv3DBiasFusePass); diff --git a/paddle/fluid/framework/ir/conv_bias_mkldnn_fuse_pass.h b/paddle/fluid/framework/ir/conv_bias_mkldnn_fuse_pass.h index 5775b83b88730..f3ad9f1c2bf14 100644 --- a/paddle/fluid/framework/ir/conv_bias_mkldnn_fuse_pass.h +++ b/paddle/fluid/framework/ir/conv_bias_mkldnn_fuse_pass.h @@ -26,11 +26,19 @@ namespace ir { class ConvBiasFusePass : public FusePassBase { public: virtual ~ConvBiasFusePass() {} + virtual bool is_conv3d() const { return false; } protected: std::unique_ptr ApplyImpl(std::unique_ptr graph) const; const std::string name_scope_{"conv_bias_mkldnn_fuse"}; }; +/* +* Fuse the Conv3D and Elementwise_add to a Conv3DBiasOp. +*/ +class Conv3DBiasFusePass : public ConvBiasFusePass { + public: + bool is_conv3d() const override { return true; } +}; } // namespace ir } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/ir/conv_bn_fuse_pass.cc b/paddle/fluid/framework/ir/conv_bn_fuse_pass.cc index 34b4c26ae3a8c..846a14e365e6b 100644 --- a/paddle/fluid/framework/ir/conv_bn_fuse_pass.cc +++ b/paddle/fluid/framework/ir/conv_bn_fuse_pass.cc @@ -121,7 +121,7 @@ std::unique_ptr ConvBNFusePass::ApplyImpl( int found_conv_bn_count = 0; auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph, Graph* g) { - VLOG(40) << "handle ConvBN fuse"; + VLOG(4) << "handle ConvBN fuse"; // conv, batch_norm, // conv_weight, conv_out, @@ -133,7 +133,7 @@ std::unique_ptr ConvBNFusePass::ApplyImpl( // check if fuse can be done and if MKL-DNN should be used FuseOptions fuse_option = FindFuseOption(*conv, *batch_norm); if (fuse_option == DO_NOT_FUSE) { - VLOG(30) << "do not perform conv+bn fuse"; + VLOG(3) << "do not perform conv+bn fuse"; return; } @@ -241,7 +241,7 @@ std::unique_ptr ConvEltwiseAddBNFusePass::ApplyImpl( int found_conv_bn_count = 0; auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph, Graph* g) { - VLOG(40) << "handle ConvBN fuse"; + VLOG(4) << "handle ConvBN fuse"; // conv, batch_norm, // conv_weight, conv_out, diff --git a/paddle/fluid/framework/ir/conv_elementwise_add2_act_fuse.cc b/paddle/fluid/framework/ir/conv_elementwise_add2_act_fuse.cc new file mode 100644 index 0000000000000..6e9905b7ecdba --- /dev/null +++ b/paddle/fluid/framework/ir/conv_elementwise_add2_act_fuse.cc @@ -0,0 +1,106 @@ +// 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 +#include "paddle/fluid/framework/ir/conv_elementwise_add2_act_fuse_pass.h" + +namespace paddle { +namespace framework { +namespace ir { + +#define GET_IR_NODE(node__) GET_IR_NODE_FROM_SUBGRAPH(node__, node__, pattern); +#define GET_NODES \ + GET_IR_NODE(conv_op); \ + GET_IR_NODE(conv_out); \ + GET_IR_NODE(conv_filter); \ + GET_IR_NODE(elementwise_add_op); \ + GET_IR_NODE(elementwise_add_in_y); \ + GET_IR_NODE(elementwise_add_out); \ + GET_IR_NODE(elementwise_add_op_1); \ + GET_IR_NODE(elementwise_add_in_y_1); \ + GET_IR_NODE(elementwise_add_out_1); \ + GET_IR_NODE(act_op); \ + GET_IR_NODE(act_out); + +// Inherient the basic infomation from `base_desc`, and modify some fields. +framework::proto::OpDesc PrepareOpDesc( + const framework::proto::OpDesc& base_desc, const std::string& bias, + const std::string& bias1, const std::string& activation, + const std::string& output) { + auto proto = base_desc; + framework::OpDesc desc(proto, nullptr); + desc.SetInput("Bias", {bias}); + desc.SetInput("ResidualData", {bias1}); + desc.SetAttr("activation", activation); + desc.SetOutput("Output", {output}); + desc.SetAttr("is_test", true); + desc.SetAttr("use_cudnn", false); + + return *desc.Proto(); +} + +std::unique_ptr ConvElementwiseAddActFusePass::ApplyImpl( + std::unique_ptr graph) const { + const std::string pattern_name = "conv_elementwise_add_act_fuse"; + FusePassBase::Init(pattern_name, graph.get()); + + GraphPatternDetector gpd; + auto* x = gpd.mutable_pattern()->NewNode("x")->AsInput()->assert_is_op_input( + "conv2d", "Input"); + + patterns::ConvElementwiseaddAct pattern(gpd.mutable_pattern(), pattern_name); + pattern(x); + + auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph, + Graph* g) { + GET_NODES; + + auto base_op_desc = *conv_op->Op()->Proto(); + std::string bias_name = elementwise_add_in_y->Name(); + std::string bias1_name = elementwise_add_in_y_1->Name(); + std::string act_op_type = act_op->Op()->Type(); + std::string act_op_out = act_out->Name(); + + auto new_op_proto = PrepareOpDesc(base_op_desc, bias_name, bias1_name, + act_op_type, act_op_out); + framework::OpDesc new_op_desc(new_op_proto, nullptr); + + // Create a new node for the fused op. + auto new_conv_op = graph->CreateOpNode(&new_op_desc); + + // Link inputs and outputs. + PADDLE_ENFORCE(subgraph.count(x)); + auto* conv_in_node = subgraph.at(x); + + IR_NODE_LINK_TO(conv_in_node, new_conv_op); // Input + IR_NODE_LINK_TO(conv_filter, new_conv_op); // Filter + IR_NODE_LINK_TO(elementwise_add_in_y, new_conv_op); // Bias + IR_NODE_LINK_TO(elementwise_add_in_y_1, new_conv_op); // ResidualData + IR_NODE_LINK_TO(new_conv_op, act_out); // Output + + // Delete the unneeded nodes. + GraphSafeRemoveNodes(graph.get(), + {conv_op, elementwise_add_op, elementwise_add_op_1, + elementwise_add_out}); + }; + gpd(graph.get(), handler); + return graph; +} + +} // namespace ir +} // namespace framework +} // namespace paddle + +REGISTER_PASS(conv_elementwise_add2_act_fuse_pass, + paddle::framework::ir::ConvElementwiseAdd2ActFusePass); diff --git a/paddle/fluid/framework/ir/conv_elementwise_add2_act_fuse_pass.cc b/paddle/fluid/framework/ir/conv_elementwise_add2_act_fuse_pass.cc new file mode 100644 index 0000000000000..23f343f631628 --- /dev/null +++ b/paddle/fluid/framework/ir/conv_elementwise_add2_act_fuse_pass.cc @@ -0,0 +1,105 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/framework/ir/conv_elementwise_add2_act_fuse_pass.h" +#include + +namespace paddle { +namespace framework { +namespace ir { + +#define GET_IR_NODE(node__) GET_IR_NODE_FROM_SUBGRAPH(node__, node__, pattern); +#define GET_NODES \ + GET_IR_NODE(conv_op); \ + GET_IR_NODE(conv_out); \ + GET_IR_NODE(conv_filter); \ + GET_IR_NODE(elementwise_add_op); \ + GET_IR_NODE(elementwise_add_in_y); \ + GET_IR_NODE(elementwise_add_out); \ + GET_IR_NODE(elementwise_add_op_1); \ + GET_IR_NODE(elementwise_add_in_y_1); \ + GET_IR_NODE(elementwise_add_out_1); \ + GET_IR_NODE(act_op); \ + GET_IR_NODE(act_out); + +// Inherient the basic infomation from `base_desc`, and modify some fields. +framework::proto::OpDesc PrepareOpDesc( + const framework::proto::OpDesc& base_desc, const std::string& bias, + const std::string& bias1, const std::string& activation, + const std::string& output) { + auto proto = base_desc; + framework::OpDesc desc(proto, nullptr); + desc.SetInput("Bias", {bias}); + desc.SetInput("ResidualData", {bias1}); + desc.SetAttr("activation", activation); + desc.SetOutput("Output", {output}); + desc.SetAttr("is_test", true); + + return *desc.Proto(); +} + +std::unique_ptr ConvElementwiseAdd2ActFusePass::ApplyImpl( + std::unique_ptr graph) const { + const std::string pattern_name = "conv_elementwise_add_act_fuse"; + FusePassBase::Init(pattern_name, graph.get()); + + GraphPatternDetector gpd; + auto* x = gpd.mutable_pattern()->NewNode("x")->AsInput()->assert_is_op_input( + "conv2d", "Input"); + + patterns::ConvElementwiseadd2Act pattern(gpd.mutable_pattern(), pattern_name); + pattern(x); + + auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph, + Graph* g) { + GET_NODES; + + auto base_op_desc = *conv_op->Op()->Proto(); + std::string bias_name = elementwise_add_in_y->Name(); + std::string bias1_name = elementwise_add_in_y_1->Name(); + std::string act_op_type = act_op->Op()->Type(); + std::string act_op_out = act_out->Name(); + + auto new_op_proto = PrepareOpDesc(base_op_desc, bias_name, bias1_name, + act_op_type, act_op_out); + framework::OpDesc new_op_desc(new_op_proto, nullptr); + + // Create a new node for the fused op. + graph->CreateOpNode(&new_op_desc); + + // Link inputs and outputs. + PADDLE_ENFORCE(subgraph.count(x)); + auto* conv_in_node = subgraph.at(x); + + IR_NODE_LINK_TO(conv_in_node, conv_op); // Input + IR_NODE_LINK_TO(conv_filter, conv_op); // Filter + IR_NODE_LINK_TO(conv_op, conv_out); // Output + IR_NODE_LINK_TO(elementwise_add_in_y, conv_op); // Bias + IR_NODE_LINK_TO(elementwise_add_in_y_1, conv_op); // Bias + + // Delete the unneeded nodes. + GraphSafeRemoveNodes(graph.get(), + {conv_op, elementwise_add_op, elementwise_add_op_1, + elementwise_add_out}); + }; + gpd(graph.get(), handler); + return graph; +} + +} // namespace ir +} // namespace framework +} // namespace paddle + +REGISTER_PASS(conv_elementwise_add2_act_fuse_pass, + paddle::framework::ir::ConvElementwiseAdd2ActFusePass); diff --git a/paddle/fluid/framework/ir/conv_elementwise_add2_act_fuse_pass.h b/paddle/fluid/framework/ir/conv_elementwise_add2_act_fuse_pass.h new file mode 100644 index 0000000000000..3b40a5a92665c --- /dev/null +++ b/paddle/fluid/framework/ir/conv_elementwise_add2_act_fuse_pass.h @@ -0,0 +1,33 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once +#include "paddle/fluid/framework/ir/fuse_pass_base.h" +#include "paddle/fluid/framework/ir/graph_pattern_detector.h" + +namespace paddle { +namespace framework { +namespace ir { + +class ConvElementwiseAdd2ActFusePass : public FusePassBase { + public: + virtual ~ConvElementwiseAdd2ActFusePass() {} + + protected: + std::unique_ptr ApplyImpl(std::unique_ptr graph) const; +}; + +} // namespace ir +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/ir/conv_elementwise_add_act_fuse_pass.cc b/paddle/fluid/framework/ir/conv_elementwise_add_act_fuse_pass.cc new file mode 100644 index 0000000000000..fe3b4fca79f37 --- /dev/null +++ b/paddle/fluid/framework/ir/conv_elementwise_add_act_fuse_pass.cc @@ -0,0 +1,104 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/framework/ir/conv_elementwise_add_act_fuse_pass.h" +#include +#include "paddle/fluid/framework/ir/graph_viz_pass.h" + +namespace paddle { +namespace framework { +namespace ir { + +#define GET_IR_NODE(node__) GET_IR_NODE_FROM_SUBGRAPH(node__, node__, pattern); +#define GET_NODES \ + GET_IR_NODE(conv_op); \ + GET_IR_NODE(conv_out); \ + GET_IR_NODE(conv_filter); \ + GET_IR_NODE(elementwise_add_op); \ + GET_IR_NODE(elementwise_add_in_y); \ + GET_IR_NODE(elementwise_add_out); \ + GET_IR_NODE(act_op); \ + GET_IR_NODE(act_out); + +// Inherient the basic infomation from `base_desc`, and modify some fields. +framework::proto::OpDesc PrepareOpDesc( + const framework::proto::OpDesc& base_desc, const std::string& bias, + const std::string& activation, const std::string& output) { + auto proto = base_desc; + framework::OpDesc desc(proto, nullptr); + desc.SetType("conv2d_fusion"); + desc.SetInput("Bias", {bias}); + desc.SetInput("ResidualData", {}); + desc.SetAttr("activation", activation); + desc.SetOutput("Output", {output}); + desc.SetAttr("is_test", true); + desc.SetAttr("use_cudnn", false); + desc.Flush(); + return *desc.Proto(); +} + +std::unique_ptr ConvElementwiseAddActFusePass::ApplyImpl( + std::unique_ptr graph) const { + const std::string pattern_name = "conv_elementwise_add_act_fuse"; + FusePassBase::Init(pattern_name, graph.get()); + + GraphPatternDetector gpd; + auto* x = gpd.mutable_pattern() + ->NewNode("x") + ->assert_is_op_input("conv2d", "Input") + ->AsInput(); + + patterns::ConvElementwiseaddAct pattern(gpd.mutable_pattern(), pattern_name); + pattern(x); + + auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph, + Graph* g) { + GET_NODES; + + auto base_op_desc = *conv_op->Op()->Proto(); + std::string bias_name = elementwise_add_in_y->Name(); + std::string act_op_type = act_op->Op()->Type(); + std::string act_op_out = act_out->Name(); + + auto new_op_proto = + PrepareOpDesc(base_op_desc, bias_name, act_op_type, act_op_out); + framework::OpDesc new_op_desc(new_op_proto, nullptr); + + // Create a new node for the fused op. + auto* new_conv_op = graph->CreateOpNode(&new_op_desc); + + // Link inputs and outputs. + PADDLE_ENFORCE(subgraph.count(x)); + auto* conv_in_node = subgraph.at(x); + + IR_NODE_LINK_TO(conv_in_node, new_conv_op); // Input + IR_NODE_LINK_TO(conv_filter, new_conv_op); // Filter + IR_NODE_LINK_TO(elementwise_add_in_y, new_conv_op); // Bias + IR_NODE_LINK_TO(new_conv_op, act_out); // Output + + // Delete the unneeded nodes. + GraphSafeRemoveNodes(graph.get(), {conv_op, conv_out, elementwise_add_op, + elementwise_add_out, act_op}); + }; + + gpd(graph.get(), handler); + return graph; +} + +} // namespace ir +} // namespace framework +} // namespace paddle + +REGISTER_PASS(conv_elementwise_add_act_fuse_pass, + paddle::framework::ir::ConvElementwiseAddActFusePass); diff --git a/paddle/fluid/framework/ir/conv_elementwise_add_act_fuse_pass.h b/paddle/fluid/framework/ir/conv_elementwise_add_act_fuse_pass.h new file mode 100644 index 0000000000000..ac69aa6458fc8 --- /dev/null +++ b/paddle/fluid/framework/ir/conv_elementwise_add_act_fuse_pass.h @@ -0,0 +1,33 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once +#include "paddle/fluid/framework/ir/fuse_pass_base.h" +#include "paddle/fluid/framework/ir/graph_pattern_detector.h" + +namespace paddle { +namespace framework { +namespace ir { + +class ConvElementwiseAddActFusePass : public FusePassBase { + public: + virtual ~ConvElementwiseAddActFusePass() {} + + protected: + std::unique_ptr ApplyImpl(std::unique_ptr graph) const; +}; + +} // namespace ir +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/ir/conv_elementwise_add_fuse_pass.cc b/paddle/fluid/framework/ir/conv_elementwise_add_fuse_pass.cc new file mode 100644 index 0000000000000..476c9dbc353f8 --- /dev/null +++ b/paddle/fluid/framework/ir/conv_elementwise_add_fuse_pass.cc @@ -0,0 +1,91 @@ +// 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 + +#include "paddle/fluid/framework/ir/conv_elementwise_add_fuse_pass.h" +#include "paddle/fluid/framework/ir/graph_viz_pass.h" + +namespace paddle { +namespace framework { +namespace ir { + +#define GET_IR_NODE(node__) GET_IR_NODE_FROM_SUBGRAPH(node__, node__, pattern); +#define GET_NODES \ + GET_IR_NODE(conv_op); \ + GET_IR_NODE(conv_out); \ + GET_IR_NODE(conv_filter); \ + GET_IR_NODE(elementwise_add_op); \ + GET_IR_NODE(elementwise_add_in_y); \ + GET_IR_NODE(elementwise_add_out); + +std::unique_ptr ConvElementwiseAddFusePass::ApplyImpl( + std::unique_ptr graph) const { + const std::string pattern_name = "conv_elementwise_add_fuse"; + FusePassBase::Init(pattern_name, graph.get()); + + GraphPatternDetector gpd; + auto* x = gpd.mutable_pattern() + ->NewNode("x") + ->assert_is_op_input("conv2d", "Input") + ->AsInput(); + + patterns::ConvElementwiseadd pattern(gpd.mutable_pattern(), pattern_name); + pattern(x); + + auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph, + Graph* g) { + GET_NODES; + + auto base_op_desc = *conv_op->Op()->Proto(); + std::string bias_name = elementwise_add_in_y->Name(); + std::string output_name = elementwise_add_out->Name(); + + std::string act_type = "identity"; + framework::OpDesc new_op_desc(base_op_desc, nullptr); + new_op_desc.SetType("conv2d_fusion"); + new_op_desc.SetInput("Bias", {bias_name}); + new_op_desc.SetInput("ResidualData", {}); + new_op_desc.SetAttr("activation", act_type); + new_op_desc.SetOutput("Output", {output_name}); + new_op_desc.SetAttr("is_test", true); + new_op_desc.SetAttr("use_cudnn", false); + new_op_desc.Flush(); + + // Create a new node for the fused op. + auto* new_conv_op = graph->CreateOpNode(&new_op_desc); + + // Link inputs and outputs. + PADDLE_ENFORCE(subgraph.count(x)); + auto* conv_in_node = subgraph.at(x); + + IR_NODE_LINK_TO(conv_in_node, new_conv_op); // Input + IR_NODE_LINK_TO(conv_filter, new_conv_op); // Filter + IR_NODE_LINK_TO(elementwise_add_in_y, new_conv_op); // Bias + IR_NODE_LINK_TO(new_conv_op, elementwise_add_out); // Output + + // Delete the unneeded nodes. + GraphSafeRemoveNodes(graph.get(), {conv_op, conv_out, elementwise_add_op}); + }; + + gpd(graph.get(), handler); + return graph; +} + +} // namespace ir +} // namespace framework +} // namespace paddle + +REGISTER_PASS(conv_elementwise_add_fuse_pass, + paddle::framework::ir::ConvElementwiseAddFusePass); diff --git a/paddle/fluid/framework/ir/conv_elementwise_add_fuse_pass.h b/paddle/fluid/framework/ir/conv_elementwise_add_fuse_pass.h new file mode 100644 index 0000000000000..f234603f5856a --- /dev/null +++ b/paddle/fluid/framework/ir/conv_elementwise_add_fuse_pass.h @@ -0,0 +1,33 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once +#include "paddle/fluid/framework/ir/fuse_pass_base.h" +#include "paddle/fluid/framework/ir/graph_pattern_detector.h" + +namespace paddle { +namespace framework { +namespace ir { + +class ConvElementwiseAddFusePass : public FusePassBase { + public: + virtual ~ConvElementwiseAddFusePass() {} + + protected: + std::unique_ptr ApplyImpl(std::unique_ptr graph) const; +}; + +} // namespace ir +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.cc b/paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.cc index 5376fc163e259..a8029e67e659a 100644 --- a/paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.cc +++ b/paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.cc @@ -24,35 +24,6 @@ namespace paddle { namespace framework { namespace ir { -// The function keeps the graph consistent by replacing -// a node 'from' in the set of inputs nodes -// of the visited node by a node 'to'. -void CorrectGraphEdges(Graph* graph, Node* from, Node* to) { - for (auto& node : GraphTraits::DFS(*graph)) { - auto from_in_inputs = - std::find(std::begin(node.inputs), std::end(node.inputs), from); - - if (from_in_inputs != std::end(node.inputs)) { - IR_NODE_LINK_TO(to, (&node)); - - auto inputs = node.Op()->Inputs(); - - using input_type = VariableNameMap::value_type; - - std::for_each(std::begin(inputs), std::end(inputs), - [from, to, &node](const input_type& i) -> void { - auto param_names = i.second; - auto pi = std::find(std::begin(param_names), - std::end(param_names), from->Name()); - - if (pi != std::end(param_names)) { - node.Op()->SetInput(i.first, {to->Name()}); - } - }); - } - } -} - bool IsReachable(ir::Graph* graph, Node* from, Node* to) { auto find_node = [](ir::Graph* graph, const Node* node) -> Node* { for (auto n : graph->Nodes()) { @@ -99,25 +70,12 @@ bool IsReachable(ir::Graph* graph, Node* from, Node* to) { return false; } -boost::optional HasBias(const Node& op, const std::string& bias_name) { - auto bias_input_names = op.Op()->Inputs(); - auto bias_it = bias_input_names.find(bias_name); - - if (bias_it != std::end(bias_input_names)) { - bool has_bias = !bias_it->second.empty(); - - if (has_bias) { - auto bias_names = bias_it->second; - auto bias_names_it = - std::find_if(std::begin(op.inputs), std::end(op.inputs), - [&bias_names](Node* n) -> bool { - return n->Name() == bias_names[0]; - }); - return *bias_names_it; - } - } - - return boost::none; +template +boost::optional HasAttribute(const Node& op, const std::string& attr) { + if (op.Op()->HasAttr(attr)) + return boost::get(op.Op()->GetAttr(attr)); + else + return boost::none; } ResidualConnectionMKLDNNFusePass::IdentityFuseHandle::IdentityFuseHandle( @@ -151,40 +109,18 @@ void ResidualConnectionMKLDNNFusePass::IdentityFuseHandle::operator()( if (!IsReachable(graph, elementwise_add_identity, conv_output)) return; - OpDesc op_desc; - op_desc.SetType("conv2d"); - - op_desc.SetInput("Input", {conv_input->Name()}); - op_desc.SetInput("Filter", {conv_filter->Name()}); - op_desc.SetInput("ResidualData", {elementwise_add_identity->Name()}); - op_desc.SetOutput("Output", {conv_output->Name()}); + auto fuse_relu = HasAttribute(*conv_op, "fuse_relu"); + if (fuse_relu && *fuse_relu) return; - auto conv_bias = HasBias(*conv_op, "Bias"); + conv_op->Op()->SetInput("ResidualData", {elementwise_add_identity->Name()}); + conv_op->Op()->SetOutput("Output", {elementwise_add_out->Name()}); + conv_op->Op()->SetAttr("fuse_residual_connection", true); - if (conv_bias) { - op_desc.SetInput("Bias", {(*conv_bias)->Name()}); - } - - for (const auto& attr : conv_op->Op()->GetAttrMap()) { - op_desc.SetAttr(attr.first, attr.second); - } - - op_desc.SetAttr("fuse_residual_connection", true); + GraphSafeRemoveNodes(graph, {conv_output, elementwise_add_op}); - auto fused_conv_op = graph->CreateOpNode(&op_desc); - - IR_NODE_LINK_TO(conv_input, fused_conv_op); - IR_NODE_LINK_TO(conv_filter, fused_conv_op); - IR_NODE_LINK_TO(elementwise_add_identity, fused_conv_op); - IR_NODE_LINK_TO(fused_conv_op, conv_output); - - if (conv_bias) { - IR_NODE_LINK_TO((*conv_bias), fused_conv_op); - } + IR_NODE_LINK_TO(elementwise_add_identity, conv_op); + IR_NODE_LINK_TO(conv_op, elementwise_add_out); - CorrectGraphEdges(graph, elementwise_add_out, conv_output); - GraphSafeRemoveNodes(graph, - {elementwise_add_out, conv_op, elementwise_add_op}); (*fusion_stats)++; } @@ -229,60 +165,33 @@ void ResidualConnectionMKLDNNFusePass::ProjectionFuseHandle::operator()( Node* projection_node; Node* residual_conv_op; - Node* residual_conv_input; - Node* residual_conv_filter; Node* residual_conv_output; if (IsReachable(graph, conv_x_input, conv_y_output)) { projection_node = conv_x_output; residual_conv_op = conv_y_op; - residual_conv_input = conv_y_input; - residual_conv_filter = conv_y_filter; residual_conv_output = conv_y_output; } else if (IsReachable(graph, conv_y_input, conv_x_output)) { projection_node = conv_y_output; residual_conv_op = conv_x_op; - residual_conv_input = conv_x_input; - residual_conv_filter = conv_x_filter; residual_conv_output = conv_x_output; } else { return; } - OpDesc op_desc; - op_desc.SetType("conv2d"); + auto fuse_relu = HasAttribute(*residual_conv_op, "fuse_relu"); + if (fuse_relu && *fuse_relu) return; - op_desc.SetInput("Input", {residual_conv_input->Name()}); - op_desc.SetInput("Filter", {residual_conv_filter->Name()}); - op_desc.SetInput("ResidualData", {projection_node->Name()}); - op_desc.SetOutput("Output", {residual_conv_output->Name()}); + residual_conv_op->Op()->SetInput("ResidualData", {projection_node->Name()}); + residual_conv_op->Op()->SetOutput("Output", {elementwise_add_out->Name()}); - auto residual_conv_bias = HasBias(*residual_conv_op, "Bias"); + residual_conv_op->Op()->SetAttr("fuse_residual_connection", true); - if (residual_conv_bias) { - op_desc.SetInput("Bias", {(*residual_conv_bias)->Name()}); - } - - for (const auto& attr : residual_conv_op->Op()->GetAttrMap()) { - op_desc.SetAttr(attr.first, attr.second); - } - - op_desc.SetAttr("fuse_residual_connection", true); + GraphSafeRemoveNodes(graph, {residual_conv_output, elementwise_add_op}); - auto fused_conv_op = graph->CreateOpNode(&op_desc); - - IR_NODE_LINK_TO(residual_conv_input, fused_conv_op); - IR_NODE_LINK_TO(residual_conv_filter, fused_conv_op); - IR_NODE_LINK_TO(projection_node, fused_conv_op); - IR_NODE_LINK_TO(fused_conv_op, residual_conv_output); - - if (residual_conv_bias) { - IR_NODE_LINK_TO((*residual_conv_bias), fused_conv_op); - } + IR_NODE_LINK_TO(projection_node, residual_conv_op); + IR_NODE_LINK_TO(residual_conv_op, elementwise_add_out); - CorrectGraphEdges(graph, elementwise_add_out, residual_conv_output); - GraphSafeRemoveNodes( - graph, {elementwise_add_out, residual_conv_op, elementwise_add_op}); (*fusion_stats)++; } diff --git a/paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.cc b/paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.cc index 048868e1f913e..e359a3832ee8d 100644 --- a/paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.cc +++ b/paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.cc @@ -38,7 +38,7 @@ std::unique_ptr ConvReLUFusePass::ApplyImpl( int found_conv_relu_count = 0; auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph, Graph* g) { - VLOG(40) << "handle ConvReLU fuse"; + VLOG(4) << "handle ConvReLU fuse"; GET_IR_NODE_FROM_SUBGRAPH(conv_weight, conv_weight, conv_relu_pattern); // Filter GET_IR_NODE_FROM_SUBGRAPH(conv_out, conv_out, conv_relu_pattern); // tmp @@ -48,7 +48,7 @@ std::unique_ptr ConvReLUFusePass::ApplyImpl( FuseOptions fuse_option = FindFuseOption(*conv, *relu); if (fuse_option == DO_NOT_FUSE) { - VLOG(30) << "do not perform conv+relu fuse"; + VLOG(3) << "do not perform conv+relu fuse"; return; } diff --git a/paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass.cc b/paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass.cc index 5f3334578d10f..19056e18aa892 100644 --- a/paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass.cc +++ b/paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass.cc @@ -39,7 +39,7 @@ std::unique_ptr DepthwiseConvMKLDNNPass::ApplyImpl( int found_depthwise_conv_mkldnn_count = 0; auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph, Graph* g) { - VLOG(30) << "handle DepthwiseConvMKLDNN fuse"; + VLOG(3) << "handle DepthwiseConvMKLDNN fuse"; GET_NODE(depthwise_conv, (*pattern)); depthwise_conv->Op()->SetType("conv2d"); found_depthwise_conv_mkldnn_count++; diff --git a/paddle/fluid/framework/ir/fc_fuse_pass.cc b/paddle/fluid/framework/ir/fc_fuse_pass.cc index 7b6ce0da07309..26eac939054c1 100644 --- a/paddle/fluid/framework/ir/fc_fuse_pass.cc +++ b/paddle/fluid/framework/ir/fc_fuse_pass.cc @@ -39,7 +39,7 @@ std::unique_ptr FCFusePass::ApplyImpl( int found_fc_count = 0; auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph, Graph* g) { - VLOG(40) << "handle FC fuse"; + VLOG(4) << "handle FC fuse"; GET_IR_NODE_FROM_SUBGRAPH(w, w, fc_pattern); GET_IR_NODE_FROM_SUBGRAPH(fc_bias, bias, fc_pattern); GET_IR_NODE_FROM_SUBGRAPH(fc_out, Out, fc_pattern); diff --git a/paddle/fluid/framework/ir/fuse_elewise_add_act_pass.cc b/paddle/fluid/framework/ir/fuse_elewise_add_act_pass.cc index 8ed68905beed2..648acc4a75941 100644 --- a/paddle/fluid/framework/ir/fuse_elewise_add_act_pass.cc +++ b/paddle/fluid/framework/ir/fuse_elewise_add_act_pass.cc @@ -61,7 +61,7 @@ std::unique_ptr FuseElewiseAddActPass::FuseElewiseAddAct( auto handler = [&](const GraphPatternDetector::subgraph_t &subgraph, Graph *g) { - VLOG(40) << "handle FuseElewiseAddAct fuse"; + VLOG(4) << "handle FuseElewiseAddAct fuse"; GET_IR_NODE_FROM_SUBGRAPH(ele_y, ele_y, elewise_add_act_pattern); GET_IR_NODE_FROM_SUBGRAPH(ele_out, elewise_add_out, elewise_add_act_pattern); @@ -77,10 +77,10 @@ std::unique_ptr FuseElewiseAddActPass::FuseElewiseAddAct( Node *elewise_add_act_node = CreateFuseElewiseAddActNode( g, act, ele_add, ele_x_n, ele_y_n, ele_out_n, act_out_n); - VLOG(40) << "\n\t " << ele_x_n << " and " << ele_y_n << " -> " - << ele_add->Name() << " -> " << ele_out_n << "\n" - << "\t " << ele_out_n << " -> " << act->Name() << " -> " - << act_out_n; + VLOG(4) << "\n\t " << ele_x_n << " and " << ele_y_n << " -> " + << ele_add->Name() << " -> " << ele_out_n << "\n" + << "\t " << ele_out_n << " -> " << act->Name() << " -> " + << act_out_n; ReLinkNodes(g, ele_out, ele_add, act, elewise_add_act_node); found_elewise_add_act_count++; @@ -113,7 +113,7 @@ std::unique_ptr FuseElewiseAddActPass::FuseActElewiseAdd( auto handler = [&](const GraphPatternDetector::subgraph_t &subgraph, Graph *g) { - VLOG(40) << "handle FuseElewiseAddAct fuse"; + VLOG(4) << "handle FuseElewiseAddAct fuse"; GET_IR_NODE_FROM_SUBGRAPH(act_out, act_out, act_elewise_add_pattern); GET_IR_NODE_FROM_SUBGRAPH(ele_x, ele_x, act_elewise_add_pattern); GET_IR_NODE_FROM_SUBGRAPH(ele_out, elewise_add_out, @@ -129,9 +129,9 @@ std::unique_ptr FuseElewiseAddActPass::FuseActElewiseAdd( Node *elewise_add_act_node = CreateFuseElewiseAddActNode( g, ele_add, act, elewise_add_x_n, act_i_n, act_o_n, elewise_add_out_n); - VLOG(40) << "\n\t " << act_i_n << " -> " << act->Name() << " -> " << act_o_n - << "\n\t " << act_o_n << " and " << elewise_add_x_n << " -> " - << ele_add->Name() << " -> " << elewise_add_out_n; + VLOG(4) << "\n\t " << act_i_n << " -> " << act->Name() << " -> " << act_o_n + << "\n\t " << act_o_n << " and " << elewise_add_x_n << " -> " + << ele_add->Name() << " -> " << elewise_add_out_n; ReLinkNodes(g, act_out, act, ele_add, elewise_add_act_node); found_elewise_add_act_count++; @@ -165,7 +165,7 @@ std::unique_ptr FuseElewiseAddActPass::FuseElewiseAddActInplaceGrad( auto handler = [&](const GraphPatternDetector::subgraph_t &subgraph, Graph *g) { - VLOG(40) << "handle FuseElewiseAddActGrad1 fuse"; + VLOG(4) << "handle FuseElewiseAddActGrad1 fuse"; GET_IR_NODE_FROM_SUBGRAPH(act_out, act_out, elewise_add_act_grad_pattern); GET_IR_NODE_FROM_SUBGRAPH(act_grad, act_grad, elewise_add_act_grad_pattern); GET_IR_NODE_FROM_SUBGRAPH(d_itermediate_out, d_itermediate_out, @@ -208,10 +208,10 @@ std::unique_ptr FuseElewiseAddActPass::FuseElewiseAddActInplaceGrad( auto fused_node = g->CreateOpNode(&desc); - VLOG(40) << "\n\t " << d_act_out_n << " and " << act_out_n << " -> " - << act_grad->Name() << " -> " << d_itermediate_out_n << "\n\t " - << d_itermediate_out_n << " and " << act_out_n << " -> " - << ele_add_grad->Name() << " -> " << d_itermediate_out_n; + VLOG(4) << "\n\t " << d_act_out_n << " and " << act_out_n << " -> " + << act_grad->Name() << " -> " << d_itermediate_out_n << "\n\t " + << d_itermediate_out_n << " and " << act_out_n << " -> " + << ele_add_grad->Name() << " -> " << d_itermediate_out_n; ReLinkNodes(g, d_itermediate_out, act_grad, ele_add_grad, fused_node); found_elewise_add_act_count++; diff --git a/paddle/fluid/framework/ir/graph.cc b/paddle/fluid/framework/ir/graph.cc index ae0e42ff5e894..8670dcfed7e40 100644 --- a/paddle/fluid/framework/ir/graph.cc +++ b/paddle/fluid/framework/ir/graph.cc @@ -38,9 +38,8 @@ void CheckProgram(const ProgramDesc &program) { switch (role_id) { case _INT(OpRole::kForward): if (visit.find(_INT(OpRole::kBackward)) != visit.end()) { - LOG(ERROR) - << "Cannot add backward operator before forward operator %s." - << op->Type(); + LOG(ERROR) << "Cannot add backward operator before forward operator " + << op->Type(); } break; case _INT(OpRole::kBackward): @@ -90,7 +89,7 @@ Graph::Graph(const ProgramDesc &program) : program_(program) { std::map> Graph::InitFromProgram( const ProgramDesc &program) { - VLOG(30) << "block in program:" << program_.Size(); + VLOG(3) << "block in program:" << program_.Size(); std::unordered_map all_vars; // var nodes for each var name, will have multiple versions in SSA std::map> var_nodes; @@ -158,12 +157,15 @@ void Graph::ResolveHazard( auto it_old = versions.rbegin(); ++it_old; for (; it_old != versions.rend(); it_new = it_old, ++it_old) { - VLOG(30) << "deal with var: " << (*it_new)->Name(); + VLOG(3) << "deal with var: " << (*it_new)->Name(); ir::Node *write_op = (*it_new)->inputs.empty() ? nullptr : (*it_new)->inputs[0]; const auto &read_ops = (*it_old)->outputs; - PADDLE_ENFORCE(write_op, "The write_op should not be empty."); + PADDLE_ENFORCE( + write_op, + string::Sprintf("The write_op of var %s should not be empty.", + (*it_new)->Name())); // Add write after write dependence ir::Node *upstream_op = diff --git a/paddle/fluid/framework/ir/graph.h b/paddle/fluid/framework/ir/graph.h index 0c856f8e61007..47fcf96a3f92b 100644 --- a/paddle/fluid/framework/ir/graph.h +++ b/paddle/fluid/framework/ir/graph.h @@ -73,14 +73,21 @@ class Graph { } bool Has(const std::string &attr_name) const { - return attrs_.find(attr_name) != attrs_.end(); + return attrs_.count(attr_name) > 0; } template AttrType &Get(const std::string &attr_name) const { PADDLE_ENFORCE(Has(attr_name), "%s attr not registered for graph.", attr_name); - return *boost::any_cast(attrs_.at(attr_name)); + try { + return *boost::any_cast(attrs_.at(attr_name)); + } catch (boost::bad_any_cast &) { + PADDLE_THROW( + "Invalid attribute type of %s error, expected: %s, actual: %s", + attr_name, typeid(AttrType *).name(), + attrs_.at(attr_name).type().name()); + } } template @@ -89,7 +96,7 @@ class Graph { attr_name); attrs_[attr_name] = attr; attr_dels_[attr_name] = [attr, attr_name]() { - VLOG(30) << "deleting " << attr_name; + VLOG(3) << "deleting " << attr_name; delete attr; }; } @@ -177,14 +184,13 @@ class Graph { return nullptr; } - const ProgramDesc &program() const { return program_; } - std::map> InitFromProgram( - const ProgramDesc &program); - void ResolveHazard( const std::map> &var_nodes); private: + std::map> InitFromProgram( + const ProgramDesc &program); + // This method takes ownership of `node`. ir::Node *AddNode(ir::Node *node) { PADDLE_ENFORCE(node_set_.find(node) == node_set_.end()); diff --git a/paddle/fluid/framework/ir/graph_helper.cc b/paddle/fluid/framework/ir/graph_helper.cc index 963179192fa6c..d99f856d8f46e 100644 --- a/paddle/fluid/framework/ir/graph_helper.cc +++ b/paddle/fluid/framework/ir/graph_helper.cc @@ -18,6 +18,7 @@ limitations under the License. */ #include #include #include +#include #include DEFINE_string(print_sub_graph_dir, "", @@ -40,9 +41,8 @@ void SortHelper( } } - VLOG(30) << "topology sort insert: " << node->Name() - << reinterpret_cast(node) << " input " - << node->inputs.size(); + VLOG(3) << "topology sort insert: " << node->Name() + << reinterpret_cast(node) << " input " << node->inputs.size(); ret->push_back(node); } @@ -111,9 +111,9 @@ std::map> BuildOperationAdjList( for (auto &var : n->inputs) { for (auto &adj_n : var->inputs) { PADDLE_ENFORCE(adj_n->NodeType() == ir::Node::Type::kOperation); - VLOG(40) << "adj " << adj_n->Name() << reinterpret_cast(adj_n) - << " -> " << n->Name() << reinterpret_cast(n) - << " via " << var->Name() << reinterpret_cast(var); + VLOG(4) << "adj " << adj_n->Name() << reinterpret_cast(adj_n) + << " -> " << n->Name() << reinterpret_cast(n) + << " via " << var->Name() << reinterpret_cast(var); adj_list[n].insert(adj_n); } } @@ -122,7 +122,7 @@ std::map> BuildOperationAdjList( } size_t GraphNum(const Graph &graph) { - std::unordered_set nodes = graph.Nodes(); + std::unordered_set nodes(graph.Nodes()); std::unordered_set visited_nodes; visited_nodes.reserve(nodes.size()); std::deque q_nodes; diff --git a/paddle/fluid/framework/ir/graph_helper.h b/paddle/fluid/framework/ir/graph_helper.h index 8d92c406689ab..be525151f9f97 100644 --- a/paddle/fluid/framework/ir/graph_helper.h +++ b/paddle/fluid/framework/ir/graph_helper.h @@ -24,6 +24,7 @@ limitations under the License. */ namespace paddle { namespace framework { namespace ir { + // Test if the graph contains circle. bool HasCircle(const Graph &graph); diff --git a/paddle/fluid/framework/ir/graph_pattern_detector.cc b/paddle/fluid/framework/ir/graph_pattern_detector.cc index f1f971656ae6a..13d752e5167c0 100644 --- a/paddle/fluid/framework/ir/graph_pattern_detector.cc +++ b/paddle/fluid/framework/ir/graph_pattern_detector.cc @@ -25,6 +25,7 @@ #include "paddle/fluid/platform/enforce.h" #include "paddle/fluid/string/pretty_log.h" #include "paddle/fluid/string/printf.h" + namespace paddle { namespace framework { namespace ir { @@ -92,19 +93,19 @@ void GraphPatternDetector::operator()(Graph *graph, PrettyLogEndl(Style::detail(), "--- detect %d subgraphs", subgraphs.size()); int id = 0; for (auto &g : subgraphs) { - VLOG(30) << "optimizing #" << id++ << " subgraph"; + VLOG(3) << "optimizing #" << id++ << " subgraph"; handler(g, graph); } } bool GraphPatternDetector::MarkPDNodesInGraph(const ir::Graph &graph) { - VLOG(30) << "mark pdnodes in graph"; + VLOG(3) << "mark pdnodes in graph"; if (graph.Nodes().empty()) return false; for (auto &node : GraphTraits::DFS(graph)) { for (const auto &pdnode : pattern_.nodes()) { if (pdnode->Tell(&node)) { - VLOG(40) << "pdnode " << pdnode->name() << " marked"; + VLOG(4) << "Node " << node.Name() << " marked as " << pdnode->name(); pdnodes2nodes_[pdnode.get()].insert(&node); } } @@ -112,7 +113,7 @@ bool GraphPatternDetector::MarkPDNodesInGraph(const ir::Graph &graph) { // Check to early stop if some PDNode can't find matched Node. for (auto &pdnode : pattern_.nodes()) { if (!pdnodes2nodes_.count(pdnode.get())) { - VLOG(40) << pdnode->name() << " can't find matched Node, early stop"; + VLOG(4) << pdnode->name() << " can't find matched Node, early stop"; // return false; } } @@ -121,7 +122,7 @@ bool GraphPatternDetector::MarkPDNodesInGraph(const ir::Graph &graph) { GetMarkedNodes(const_cast(&graph)).insert(n); } } - VLOG(30) << pdnodes2nodes_.size() << " nodes marked"; + VLOG(3) << pdnodes2nodes_.size() << " nodes marked"; return !pdnodes2nodes_.empty(); } @@ -215,7 +216,7 @@ GraphPatternDetector::DetectPatterns() { // Extend a PDNode to subgraphs by deducing the connection relations defined // in edges of PDNodes. for (const auto &edge : pattern_.edges()) { - VLOG(40) << "check " << edge.first->name() << " -> " << edge.second->name(); + VLOG(4) << "check " << edge.first->name() << " -> " << edge.second->name(); // TODO(Superjomn) Fix bug here, the groups might be duplicate here. // Each role has two PDNodes, which indicates two roles. // Detect two Nodes that can match these two roles and they are connected. @@ -226,7 +227,7 @@ GraphPatternDetector::DetectPatterns() { // source -> target for (Node *source : pdnodes2nodes_[edge.first]) { for (Node *target : pdnodes2nodes_[edge.second]) { - VLOG(80) << "check " << source->id() << " -- " << target->id(); + VLOG(8) << "check " << source->id() << " -- " << target->id(); // TODO(Superjomn) add some prune strategies. for (const auto &group : pre_groups) { if (IsNodesLink(source, target)) { @@ -243,13 +244,12 @@ GraphPatternDetector::DetectPatterns() { } } } - VLOG(30) << "step " << step << " get records: " << cur_groups.size(); + VLOG(3) << "step " << step << " get records: " << cur_groups.size(); for (auto &group : cur_groups) { for (auto &item : group.roles) { - VLOG(40) << "node " << item.second->id() << " as " - << item.first->name(); + VLOG(4) << "node " << item.second->id() << " as " << item.first->name(); } - VLOG(40) << "========================================================="; + VLOG(4) << "========================================================="; } } @@ -1031,10 +1031,11 @@ PDNode *patterns::ElewiseAddActInplaceGrad::operator()( } PDNode *patterns::ConvBias::operator()( - paddle::framework::ir::PDNode *conv_input) { + paddle::framework::ir::PDNode *conv_input, bool is_conv3d) { + std::string type = is_conv3d ? "conv3d" : "conv2d"; // Create Operators - conv_input->assert_is_op_input("conv2d", "Input"); - auto *conv_op = pattern->NewNode(conv_repr())->assert_is_op("conv2d"); + conv_input->assert_is_op_input(type, "Input"); + auto *conv_op = pattern->NewNode(conv_repr())->assert_is_op(type); auto *eltiwse_op = pattern->NewNode(eltwise_repr())->assert_is_op("elementwise_add"); // Create variables @@ -1042,11 +1043,11 @@ PDNode *patterns::ConvBias::operator()( auto *conv_weight_var = pattern->NewNode(conv_weight_repr()) ->AsInput() ->assert_is_persistable_var() - ->assert_is_op_input("conv2d", "Filter"); + ->assert_is_op_input(type, "Filter"); // intermediate variable, will be removed in the IR after fuse. auto *conv_out_var = pattern->NewNode(conv_out_repr()) ->AsIntermediate() - ->assert_is_only_output_of_op("conv2d") + ->assert_is_only_output_of_op(type) ->assert_is_op_input("elementwise_add"); // Bias stored in elementwise_add auto *eltwise_bias_var = pattern->NewNode(eltwise_bias_repr()) @@ -1099,6 +1100,142 @@ PDNode *patterns::ElementwiseAdd::operator()(PDNode *x_var, PDNode *y_var) { return out_var; } + +std::unordered_set conv_act_set({"identity", "sigmoid", "relu", + "relu6", "relux", "tanh", + "band_pass"}); + +PDNode *patterns::ConvElementwiseaddAct::operator()(PDNode *conv_in) { + conv_in->AsInput(); + auto conv_op = pattern->NewNode(conv_op_repr())->assert_is_op("conv2d"); + auto conv_out = pattern->NewNode(conv_out_repr()) + ->assert_is_op_output("conv2d") + ->assert_is_op_input("elementwise_add", "X") + ->AsIntermediate(); + auto conv_filter = pattern->NewNode(conv_filter_repr()) + ->assert_is_op_input("conv2d", "Filter") + ->AsInput(); + auto elementwise_add_op = pattern->NewNode(elementwise_add_op_repr()) + ->assert_is_op("elementwise_add"); + auto elementwise_add_in_y = pattern->NewNode(elementwise_add_in_y_repr()) + ->assert_is_op_input("elementwise_add", "Y") + ->AsInput(); + auto elementwise_add_out = pattern->NewNode(elementwise_add_out_repr()) + ->assert_is_op_output("elementwise_add") + ->AsIntermediate(); + + auto act_op = pattern->NewNode(act_op_repr()) + ->assert_is_op() + ->assert_more([&](Node *node) { + auto op_type = node->Name(); + return conv_act_set.count(op_type); + }); + + auto act_out = pattern->NewNode(act_out_repr()) + ->assert_is_var() + // is activation op's output. + ->assert_more([&](Node *node) { + for (auto *in_op : node->inputs) { + if (conv_act_set.count(in_op->Name())) { + return true; + } + } + return false; + }) + ->AsOutput(); + + conv_op->LinksFrom({conv_in, conv_filter}); + conv_out->LinksFrom({conv_op}); + elementwise_add_op->LinksFrom({conv_out, elementwise_add_in_y}) + .LinksTo({elementwise_add_out}); + act_op->LinksFrom({elementwise_add_out}).LinksTo({act_out}); + + return act_out; +} + +PDNode *patterns::ConvElementwiseadd2Act::operator()(PDNode *conv_in) { + auto conv_op = pattern->NewNode(conv_op_repr())->assert_is_op("conv2d"); + auto conv_filter = pattern->NewNode(conv_filter_repr()) + ->assert_is_op_input("conv2d", "Filter") + ->AsInput(); + auto conv_out = pattern->NewNode(conv_out_repr()) + ->assert_is_op_output("conv2d") + ->assert_is_op_input("elementwise_add", "X") + ->AsIntermediate(); + auto elementwise_add_op = pattern->NewNode(elementwise_add_op_repr()) + ->assert_is_op("elementwise_add"); + auto elementwise_add_in_y = pattern->NewNode(elementwise_add_in_y_repr()) + ->assert_is_op_input("elementwise_add", "Y") + ->AsInput(); + auto elementwise_add_out = pattern->NewNode(elementwise_add_out_repr()) + ->assert_is_op_output("elementwise_add") + ->assert_is_op_input("elementwise_add", "X") + ->AsIntermediate(); + + auto elementwise_add_op_1 = pattern->NewNode(elementwise_add_op_1_repr()) + ->assert_is_op("elementwise_add"); + auto elementwise_add_in_y_1 = pattern->NewNode(elementwise_add_in_y_1_repr()) + ->assert_is_op_input("elementwise_add", "Y") + ->AsInput(); + auto elementwise_add_out_1 = pattern->NewNode(elementwise_add_out_1_repr()) + ->assert_is_op_output("elementwise_add") + ->AsIntermediate(); + + auto act_op = pattern->NewNode(act_op_repr()) + ->assert_is_op() + ->assert_more([&](Node *node) { + auto op_type = node->Name(); + return conv_act_set.count(op_type); + }); + auto act_out = pattern->NewNode(act_out_repr()) + ->assert_is_var() + // is activation op's output. + ->assert_more([&](Node *node) { + for (auto *in_op : node->inputs) { + if (conv_act_set.count(in_op->Name())) { + return true; + } + } + return false; + }) + ->AsOutput(); + + conv_op->LinksFrom({conv_in, conv_filter}).LinksTo({conv_out}); + elementwise_add_op->LinksFrom({conv_out, elementwise_add_in_y}) + .LinksTo({elementwise_add_out}); + elementwise_add_op_1->LinksFrom( + {elementwise_add_out, elementwise_add_in_y_1}); + act_op->LinksFrom({elementwise_add_out_1}).LinksTo({act_out}); + return act_out; +} + +PDNode *patterns::ConvElementwiseadd::operator()(PDNode *conv_in) { + conv_in->AsInput(); + auto conv_op = pattern->NewNode(conv_op_repr())->assert_is_op("conv2d"); + auto conv_out = pattern->NewNode(conv_out_repr()) + ->assert_is_op_output("conv2d") + ->assert_is_op_input("elementwise_add", "X") + ->AsIntermediate(); + auto conv_filter = pattern->NewNode(conv_filter_repr()) + ->assert_is_op_input("conv2d", "Filter") + ->AsInput(); + auto elementwise_add_op = pattern->NewNode(elementwise_add_op_repr()) + ->assert_is_op("elementwise_add"); + auto elementwise_add_in_y = pattern->NewNode(elementwise_add_in_y_repr()) + ->assert_is_op_input("elementwise_add", "Y") + ->AsInput(); + auto elementwise_add_out = pattern->NewNode(elementwise_add_out_repr()) + ->assert_is_op_output("elementwise_add") + ->AsOutput(); + + conv_op->LinksFrom({conv_in, conv_filter}); + conv_out->LinksFrom({conv_op}); + elementwise_add_op->LinksFrom({conv_out, elementwise_add_in_y}) + .LinksTo({elementwise_add_out}); + + return elementwise_add_out; +} + } // namespace ir } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/ir/graph_pattern_detector.h b/paddle/fluid/framework/ir/graph_pattern_detector.h index c12b9503fd817..eaedd9d08e0fa 100644 --- a/paddle/fluid/framework/ir/graph_pattern_detector.h +++ b/paddle/fluid/framework/ir/graph_pattern_detector.h @@ -623,7 +623,7 @@ struct ElewiseAddActInplaceGrad : public PatternBase { struct ConvBias : public PatternBase { ConvBias(PDPattern* pattern, const std::string& name_scope) : PatternBase(pattern, name_scope, "conv_bias") {} - PDNode* operator()(PDNode* conv_input); + PDNode* operator()(PDNode* conv_input, bool is_conv3d = false); // declare operator node's name PATTERN_DECL_NODE(conv); PATTERN_DECL_NODE(eltwise); @@ -671,6 +671,69 @@ struct ElementwiseAdd : public PatternBase { PATTERN_DECL_NODE(elementwise_add_y); PATTERN_DECL_NODE(elementwise_add_out); }; + +// Conv + ElementwiseAdd + an activation +// This pattern can futher fuse the conv related ops after the conv+bn fusion. +struct ConvElementwiseaddAct : public PatternBase { + ConvElementwiseaddAct(PDPattern* pattern, const std::string& name_scope) + : PatternBase(pattern, name_scope, "conv_elementwiseadd_act") {} + + PDNode* operator()(PDNode* conv_in); + + PATTERN_DECL_NODE(conv_op); + PATTERN_DECL_NODE(conv_out); + PATTERN_DECL_NODE(conv_filter); + + PATTERN_DECL_NODE(elementwise_add_op); + PATTERN_DECL_NODE(elementwise_add_in_y); // input + PATTERN_DECL_NODE(elementwise_add_out); + + PATTERN_DECL_NODE(act_op); + PATTERN_DECL_NODE(act_out); +}; + +// Conv + ElementwiseAdd + ElementwiseAdd + Activation +struct ConvElementwiseadd2Act : public PatternBase { + ConvElementwiseadd2Act(PDPattern* pattern, const std::string& name_scope) + : PatternBase(pattern, name_scope, + "conv_elementwiseadd2_elementwiseadd_act") {} + + PDNode* operator()(PDNode* conv_in); + + PATTERN_DECL_NODE(conv_op); + PATTERN_DECL_NODE(conv_filter); + PATTERN_DECL_NODE(conv_out); + + PATTERN_DECL_NODE(elementwise_add_op); + PATTERN_DECL_NODE(elementwise_add_in_y); // input + PATTERN_DECL_NODE(elementwise_add_out); + + PATTERN_DECL_NODE(elementwise_add_op_1); + PATTERN_DECL_NODE(elementwise_add_in_y_1); // input + PATTERN_DECL_NODE(elementwise_add_out_1); + + PATTERN_DECL_NODE(act_op); + PATTERN_DECL_NODE(act_out); +}; + +// Conv + ElementwiseAdd +// This pattern should be used after ConvElementwiseadd2Act or +// ConvElementwiseadd pass +struct ConvElementwiseadd : public PatternBase { + ConvElementwiseadd(PDPattern* pattern, const std::string& name_scope) + : PatternBase(pattern, name_scope, "conv_elementwiseadd") {} + + PDNode* operator()(PDNode* conv_in); + + PATTERN_DECL_NODE(conv_op); + PATTERN_DECL_NODE(conv_out); + PATTERN_DECL_NODE(conv_filter); + + PATTERN_DECL_NODE(elementwise_add_op); + PATTERN_DECL_NODE(elementwise_add_in_y); + PATTERN_DECL_NODE(elementwise_add_out); +}; + } // namespace patterns // Link two ir::Nodes from each other. diff --git a/paddle/fluid/framework/ir/graph_viz_pass.cc b/paddle/fluid/framework/ir/graph_viz_pass.cc index 13dd354dc59b2..31ed98db72c8f 100644 --- a/paddle/fluid/framework/ir/graph_viz_pass.cc +++ b/paddle/fluid/framework/ir/graph_viz_pass.cc @@ -41,7 +41,7 @@ std::string FormatName(const Node* node) { std::unique_ptr GraphVizPass::ApplyImpl( std::unique_ptr graph) const { const std::string graph_viz_path = Get(kGraphVizPath); - VLOG(30) << "draw IR graph viz to " << graph_viz_path; + VLOG(3) << "draw IR graph viz to " << graph_viz_path; std::unique_ptr fout(new std::ofstream(graph_viz_path)); PADDLE_ENFORCE(fout->good()); std::ostream& sout = *fout; diff --git a/paddle/fluid/framework/ir/is_test_pass.cc b/paddle/fluid/framework/ir/is_test_pass.cc index 292f232ffce48..57cc98e2ca017 100644 --- a/paddle/fluid/framework/ir/is_test_pass.cc +++ b/paddle/fluid/framework/ir/is_test_pass.cc @@ -38,7 +38,7 @@ std::unique_ptr IsTestPass::ApplyImpl( for (const Node* n : graph->Nodes()) { if (n->IsOp()) { auto* op = n->Op(); - if (op->HasAttr("is_test")) { + if (op->HasAttr("is_test") || op->HasProtoAttr("is_test")) { op->SetAttr("is_test", true); } else if (std::find(begin(op_list), end(op_list), op->Type()) != end(op_list)) { diff --git a/paddle/fluid/framework/ir/is_test_pass_tester.cc b/paddle/fluid/framework/ir/is_test_pass_tester.cc index cd2cb0c9f8a8e..9696441a21661 100644 --- a/paddle/fluid/framework/ir/is_test_pass_tester.cc +++ b/paddle/fluid/framework/ir/is_test_pass_tester.cc @@ -15,7 +15,10 @@ #include "paddle/fluid/framework/ir/is_test_pass.h" #include - +#ifdef _WIN32 +#undef FALSE +#undef TRUE +#endif namespace paddle { namespace framework { namespace ir { diff --git a/paddle/fluid/framework/ir/mkldnn_placement_pass.cc b/paddle/fluid/framework/ir/mkldnn_placement_pass.cc index 145a3a455c8ae..951fcb066ce75 100644 --- a/paddle/fluid/framework/ir/mkldnn_placement_pass.cc +++ b/paddle/fluid/framework/ir/mkldnn_placement_pass.cc @@ -13,6 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/framework/ir/mkldnn_placement_pass.h" +#include namespace paddle { namespace framework { @@ -20,10 +21,20 @@ namespace ir { std::unique_ptr MKLDNNPlacementPass::ApplyImpl( std::unique_ptr graph) const { - VLOG(30) << "Aplies MKL-DNN placement strategy."; + VLOG(3) << "Aplies MKL-DNN placement strategy."; + const auto& op_types_list = + Get>("mkldnn_enabled_op_types"); for (const Node* n : graph->Nodes()) { - if (n->IsOp() && n->Op()->HasAttr("use_mkldnn")) { - n->Op()->SetAttr("use_mkldnn", true); + if (n->IsOp()) { + auto* op = n->Op(); + if (op->HasAttr("use_mkldnn") || op->HasProtoAttr("use_mkldnn")) { + if (op_types_list.empty()) { + op->SetAttr("use_mkldnn", true); + } else if (std::find(op_types_list.begin(), op_types_list.end(), + n->Name()) != op_types_list.end()) { + op->SetAttr("use_mkldnn", true); + } + } } } return graph; @@ -33,5 +44,5 @@ std::unique_ptr MKLDNNPlacementPass::ApplyImpl( } // namespace framework } // namespace paddle -REGISTER_PASS(mkldnn_placement_pass, - paddle::framework::ir::MKLDNNPlacementPass); +REGISTER_PASS(mkldnn_placement_pass, paddle::framework::ir::MKLDNNPlacementPass) + .RequirePassAttr("mkldnn_enabled_op_types"); diff --git a/paddle/fluid/framework/ir/multi_batch_merge_pass.cc b/paddle/fluid/framework/ir/multi_batch_merge_pass.cc index 532961e4d59ad..bd5b76426eb55 100644 --- a/paddle/fluid/framework/ir/multi_batch_merge_pass.cc +++ b/paddle/fluid/framework/ir/multi_batch_merge_pass.cc @@ -62,7 +62,7 @@ VarDesc UpdateGradVarDesc( string::Sprintf("%s.repeat.%d", var_desc->Name(), repeat); VarDesc repeated_var = CopyVarDesc(var_desc); repeated_var.SetName(new_gname); - VLOG(30) << "update " << var_desc->Name() << " to repeat " << repeat; + VLOG(3) << "update " << var_desc->Name() << " to repeat " << repeat; return repeated_var; } return *var_desc; @@ -78,7 +78,7 @@ std::unique_ptr BatchMergePass::ApplyImpl( std::vector nodes = TopologySortOperations(*graph); auto origin_nodes = graph->ReleaseNodes(); - VLOG(30) << "origin nodes count: " << origin_nodes.size(); + VLOG(3) << "origin nodes count: " << origin_nodes.size(); ir::Graph& result = *graph; // 1. record op nodes of different roles @@ -137,8 +137,8 @@ std::unique_ptr BatchMergePass::ApplyImpl( "%s.repeat.%d", repeated_op.Input("Variance")[0], i); bn_vars_need_rename.insert(repeated_op.Input("Mean")[0]); bn_vars_need_rename.insert(repeated_op.Input("Variance")[0]); - VLOG(30) << "renaming " << repeated_op.Input("Mean")[0] << " to " - << new_mean_name; + VLOG(3) << "renaming " << repeated_op.Input("Mean")[0] << " to " + << new_mean_name; repeated_op.RenameInput(repeated_op.Input("Mean")[0], new_mean_name); repeated_op.RenameInput(repeated_op.Input("Variance")[0], new_var_name); repeated_op.RenameOutput(repeated_op.Output("MeanOut")[0], diff --git a/paddle/fluid/framework/ir/node.cc b/paddle/fluid/framework/ir/node.cc index 50d9113088903..45d81b9373922 100644 --- a/paddle/fluid/framework/ir/node.cc +++ b/paddle/fluid/framework/ir/node.cc @@ -13,6 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/framework/ir/node.h" +#include "paddle/fluid/framework/op_info.h" namespace paddle { namespace framework { @@ -24,10 +25,19 @@ constexpr char Node::kControlDepVarName[]; const char Node::kControlDepVarName[] = "__control_var"; #endif -std::unique_ptr CreateNodeForTest(const std::string& name, +std::unique_ptr CreateNodeForTest(const std::string &name, Node::Type type) { return std::unique_ptr(new Node(name, type)); } + +std::unique_ptr CreateNodeForTest(VarDesc *var_desc) { + return std::unique_ptr(new Node(var_desc)); +} + +std::unique_ptr CreateNodeForTest(OpDesc *op_desc) { + return std::unique_ptr(new Node(op_desc)); +} + } // namespace ir } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/ir/node.h b/paddle/fluid/framework/ir/node.h index d2a393b3f19e9..89dcc677b57eb 100644 --- a/paddle/fluid/framework/ir/node.h +++ b/paddle/fluid/framework/ir/node.h @@ -18,7 +18,6 @@ limitations under the License. */ #include #include #include - #include "paddle/fluid/framework/op_desc.h" #include "paddle/fluid/framework/var_desc.h" #include "paddle/fluid/platform/macros.h" @@ -125,6 +124,8 @@ class Node { friend class Graph; friend std::unique_ptr CreateNodeForTest(const std::string& name, Node::Type type); + friend std::unique_ptr CreateNodeForTest(VarDesc* var_desc); + friend std::unique_ptr CreateNodeForTest(OpDesc* op_desc); explicit Node(const std::string& name, Type type) : name_(name), var_desc_(nullptr), op_desc_(nullptr), type_(type) {} @@ -152,7 +153,9 @@ class Node { std::unique_ptr CreateNodeForTest(const std::string& name, Node::Type type); +std::unique_ptr CreateNodeForTest(VarDesc* var_desc); +std::unique_ptr CreateNodeForTest(OpDesc* op_desc); } // namespace ir } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/ir/pass.h b/paddle/fluid/framework/ir/pass.h index 615b539695de8..27746ff1453b1 100644 --- a/paddle/fluid/framework/ir/pass.h +++ b/paddle/fluid/framework/ir/pass.h @@ -51,11 +51,18 @@ class Pass { AttrType &Get(const std::string &attr_name) const { PADDLE_ENFORCE(attrs_.find(attr_name) != attrs_.end(), "%s attr not registered for pass.", attr_name); - return *boost::any_cast(attrs_.at(attr_name)); + try { + return *boost::any_cast(attrs_.at(attr_name)); + } catch (boost::bad_any_cast &) { + PADDLE_THROW( + "Invalid attribute type of %s error, expected: %s, actual: %s", + attr_name, typeid(AttrType *).name(), + attrs_.at(attr_name).type().name()); + } } bool Has(const std::string &attr_name) const { - return attrs_.find(attr_name) != attrs_.end(); + return attrs_.count(attr_name) > 0; } void Erase(const std::string &attr_name) { @@ -76,7 +83,7 @@ class Pass { attr_name); attrs_[attr_name] = attr; attr_dels_[attr_name] = [attr, attr_name]() { - VLOG(30) << "deleting " << attr_name; + VLOG(3) << "deleting " << attr_name; delete attr; }; } diff --git a/paddle/fluid/framework/ir/seq_concat_fc_fuse_pass.cc b/paddle/fluid/framework/ir/seq_concat_fc_fuse_pass.cc index b7687d61de3ea..012e68036c35c 100644 --- a/paddle/fluid/framework/ir/seq_concat_fc_fuse_pass.cc +++ b/paddle/fluid/framework/ir/seq_concat_fc_fuse_pass.cc @@ -196,7 +196,7 @@ std::unique_ptr SeqConcatFcFusePass::ApplyImpl( detector(graph.get(), [&](const GraphPatternDetector::subgraph_t& subgraph, Graph* graph) { - VLOG(40) << "get one concat pattern"; + VLOG(4) << "get one concat pattern"; // fc GET_NODE(fc_w, detector.pattern()); GET_NODE(fc_bias, detector.pattern()); diff --git a/paddle/fluid/framework/ir/seqconv_eltadd_relu_fuse_pass.cc b/paddle/fluid/framework/ir/seqconv_eltadd_relu_fuse_pass.cc index 015b5e3c6363c..0a1f65d274708 100644 --- a/paddle/fluid/framework/ir/seqconv_eltadd_relu_fuse_pass.cc +++ b/paddle/fluid/framework/ir/seqconv_eltadd_relu_fuse_pass.cc @@ -60,7 +60,7 @@ int BuildFusion(Graph* graph, const std::string& name_scope, Scope* scope) { auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph, Graph* g) { - VLOG(40) << "handle SeqConv EltAdd Relu fuse"; + VLOG(4) << "handle SeqConv EltAdd Relu fuse"; GET_IR_NODE_FROM_SUBGRAPH(seqconv, seqconv, fuse_pattern); GET_IR_NODE_FROM_SUBGRAPH(seqconv_weight, seqconv_weight, fuse_pattern); GET_IR_NODE_FROM_SUBGRAPH(seqconv_out, seqconv_out, fuse_pattern); diff --git a/paddle/fluid/framework/lod_rank_table.cc b/paddle/fluid/framework/lod_rank_table.cc index 660ce2ec85131..6bc795b642bf7 100644 --- a/paddle/fluid/framework/lod_rank_table.cc +++ b/paddle/fluid/framework/lod_rank_table.cc @@ -31,7 +31,7 @@ void LoDRankTable::Reset(const LoD& lod, size_t level) { TableItem item; item.index = i; item.length = vec[i + 1] - vec[i]; - VLOG(100) << "Add item to rank table " << item.index << " " << item.length; + VLOG(10) << "Add item to rank table " << item.index << " " << item.length; items_.emplace_back(item); } // NOTE(yuyang18): diff --git a/paddle/fluid/framework/lod_tensor.cc b/paddle/fluid/framework/lod_tensor.cc index 669d08c70c9b7..8fbbc6584e121 100644 --- a/paddle/fluid/framework/lod_tensor.cc +++ b/paddle/fluid/framework/lod_tensor.cc @@ -26,10 +26,8 @@ limitations under the License. */ #include "paddle/fluid/memory/memcpy.h" #include "paddle/fluid/memory/memory.h" -#if !defined(_WIN32) #include "paddle/fluid/recordio/scanner.h" #include "paddle/fluid/recordio/writer.h" -#endif // _WIN32 namespace paddle { namespace framework { @@ -72,9 +70,9 @@ std::ostream &operator<<(std::ostream &os, const LoDTensor &t) { // only print first ten elements int64_t size = t.numel() < 10 ? t.numel() : 10; for (int64_t i = 0; i < size; ++i) { - if (IsType(t.type())) { + if (t.type() == proto::VarType::FP32) { os << t.data()[i] << " "; - } else if (IsType(t.type())) { + } else if (t.type() == proto::VarType::INT64) { os << t.data()[i] << " "; } else { PADDLE_THROW("LoDTensor data type not in [float, int64_t]"); @@ -159,13 +157,8 @@ bool CheckLoD(const LoD &in, int tensor_height) { if (level.size() < 2) return false; // check: the first offset(the begin offset) of each level should be 0. if (level.front() != 0) return false; - // check: all the offsets in a level should be ascending(no same items - // allows). - if (!std::is_sorted(level.begin(), level.begin(), [](size_t a, size_t b) { - if (a < b) return true; - return false; - })) { - LOG(INFO) << "ascending error"; + // check: all the offsets in a level should be ascending(allow same items) + if (!std::is_sorted(level.begin(), level.end())) { return false; } } @@ -305,7 +298,6 @@ void DeserializeFromStream(std::istream &is, LoDTensor *tensor, TensorFromStream(is, static_cast(tensor), dev_ctx); } -#if !defined(_WIN32) void WriteToRecordIO(recordio::Writer *writer, const std::vector &tensor, const platform::DeviceContext &dev_ctx) { @@ -335,19 +327,7 @@ bool ReadFromRecordIO(recordio::Scanner *scanner, return true; } -#else -class Writer {}; -class Scanner {}; -void WriteToRecordIO(recordio::Writer *writer, - const std::vector &tensor, - const platform::DeviceContext &dev_ctx) {} -bool ReadFromRecordIO(recordio::Scanner *scanner, - const platform::DeviceContext &dev_ctx, - std::vector *result_ptr) { - PADDLE_ENFORCE("windows didn't supported recordio!."); - return true; -} -#endif // _WIN32 + std::vector LoDTensor::SplitLoDTensor( const std::vector places) const { check_memory_size(); @@ -402,7 +382,7 @@ void LoDTensor::MergeLoDTensor( PADDLE_ENFORCE(!lod_tensors.empty()); framework::DDim new_dim = lod_tensors[0]->dims(); - std::type_index new_type = lod_tensors[0]->type(); + auto new_type = lod_tensors[0]->type(); framework::DataLayout new_layout = lod_tensors[0]->layout(); LoD new_lod = lod_tensors[0]->lod(); for (size_t i = 1; i < lod_tensors.size(); ++i) { diff --git a/paddle/fluid/framework/lod_tensor_test.cc b/paddle/fluid/framework/lod_tensor_test.cc index cbf5fd04d7300..15928c18d38b8 100644 --- a/paddle/fluid/framework/lod_tensor_test.cc +++ b/paddle/fluid/framework/lod_tensor_test.cc @@ -217,6 +217,11 @@ TEST(LoD, CheckLoD) { // check with underlying tensor storage. ASSERT_TRUE(CheckLoD(relative_lod, 5)); ASSERT_FALSE(CheckLoD(relative_lod, 9)); + + // check whether lod is ascending-sorted (allow same items) + ASSERT_TRUE(CheckLoD({{0, 1, 2, 3, 4, 5}}, 5)); + ASSERT_TRUE(CheckLoD({{0, 1, 3, 3, 4, 5}}, 5)); + ASSERT_FALSE(CheckLoD({{0, 1, 3, 2, 5}}, 5)); } TEST(LoD, CheckAbsLoD) { @@ -274,7 +279,6 @@ TEST(LoD, ConvertToOffsetBasedLoD) { EXPECT_EQ(offset_lod, expected); } -#if !defined(_WIN32) template static void TestRecordIO() { LoDTensor tensor; @@ -321,7 +325,6 @@ TEST(LoDTensor, RecordIO) { TestRecordIO(); TestRecordIO(); } -#endif // !defined(_WIN32) } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/mixed_vector_test.cc b/paddle/fluid/framework/mixed_vector_test.cc index 0330cae377c32..0599c8d384641 100644 --- a/paddle/fluid/framework/mixed_vector_test.cc +++ b/paddle/fluid/framework/mixed_vector_test.cc @@ -51,7 +51,7 @@ TEST(mixed_vector, InitWithCount) { TEST(mixed_vector, ForEach) { vec tmp; for (auto& v : tmp) { - VLOG(30) << v; + VLOG(3) << v; } } diff --git a/paddle/fluid/framework/naive_executor.cc b/paddle/fluid/framework/naive_executor.cc index e8e53f988f926..f1642bc0d2b10 100644 --- a/paddle/fluid/framework/naive_executor.cc +++ b/paddle/fluid/framework/naive_executor.cc @@ -21,42 +21,11 @@ #include "paddle/fluid/framework/naive_executor.h" #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/reader.h" +#include "paddle/fluid/framework/variable_helper.h" #include "paddle/fluid/string/pretty_log.h" namespace paddle { namespace framework { - -// These code can be shared with Executor. -static void InitializeVariable(Variable *var, proto::VarType::Type var_type) { - if (var_type == proto::VarType::LOD_TENSOR) { - var->GetMutable(); - } else if (var_type == proto::VarType::SELECTED_ROWS) { - var->GetMutable(); - } else if (var_type == proto::VarType::FEED_MINIBATCH) { - var->GetMutable(); - } else if (var_type == proto::VarType::FETCH_LIST) { - var->GetMutable(); - } else if (var_type == proto::VarType::STEP_SCOPES) { - var->GetMutable>(); - } else if (var_type == proto::VarType::LOD_RANK_TABLE) { - var->GetMutable(); - } else if (var_type == proto::VarType::LOD_TENSOR_ARRAY) { - var->GetMutable(); - } else if (var_type == proto::VarType::PLACE_LIST) { - var->GetMutable(); - } else if (var_type == proto::VarType::READER) { - var->GetMutable(); - } else if (var_type == proto::VarType::RAW) { - // GetMutable will be called in operator - } else { - PADDLE_THROW( - "Variable type %d is not in " - "[LOD_TENSOR, SELECTED_ROWS, FEED_MINIBATCH, FETCH_LIST, " - "LOD_RANK_TABLE, PLACE_LIST, READER, CHANNEL, RAW]", - var_type); - } -} - void NaiveExecutor::Prepare(Scope *scope, const ProgramDesc &program_desc, int block_id, bool with_feed_fetch_ops) { if (!scope) { @@ -83,6 +52,7 @@ void NaiveExecutor::Run() { for (auto &op : ops_) { VLOG(3) << std::this_thread::get_id() << " run " << op->Type() << " on scope " << scope_; + op->SetIsCalledByExecutor(false); op->Run(*scope_, place_); } } diff --git a/paddle/fluid/framework/ngraph_bridge.cc b/paddle/fluid/framework/ngraph_bridge.cc index 8177436d0bd90..42190b52289bf 100644 --- a/paddle/fluid/framework/ngraph_bridge.cc +++ b/paddle/fluid/framework/ngraph_bridge.cc @@ -12,13 +12,16 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#ifdef PADDLE_WITH_NGRAPH #include #include - -#include "paddle/fluid/framework/ngraph_bridge.h" +#include #include "ngraph/ngraph.hpp" +#include "paddle/fluid/framework/ngraph_bridge.h" +#include "paddle/fluid/framework/operator.h" +#include "paddle/fluid/operators/ngraph/ngraph_ops.h" +#include "paddle/fluid/platform/enforce.h" +#include "paddle/fluid/platform/ngraph_helper.h" namespace paddle { namespace framework { @@ -27,13 +30,18 @@ std::map&, std::shared_ptr>>)>> - NgraphBridge::NG_NODE_MAP = {}; - -void NgraphBridge::build_graph(const std::shared_ptr& op) { + NgraphBridge::NG_NODE_MAP = { + {"fill_constant", paddle::operators::ngraphs::BuildFillConstantNode}, + {"mul", paddle::operators::ngraphs::BuildMulNode}, + {"mul_grad", paddle::operators::ngraphs::BuildMulGradNode}, + {"relu", paddle::operators::ngraphs::BuildUnaryNode}, + {"tanh", paddle::operators::ngraphs::BuildUnaryNode}, + {"top_k", paddle::operators::ngraphs::BuildTopKNode}}; + +void NgraphBridge::BuildNgNode(const std::shared_ptr& op) { auto& op_type = op->Type(); - NG_NODE_MAP[op_type](op, ngb_node_map); + NG_NODE_MAP[op_type](op, ngb_node_map_); } } // namespace framework } // namespace paddle -#endif diff --git a/paddle/fluid/framework/ngraph_bridge.h b/paddle/fluid/framework/ngraph_bridge.h index 55bf0d21f3471..5ad7b8daeb6a7 100644 --- a/paddle/fluid/framework/ngraph_bridge.h +++ b/paddle/fluid/framework/ngraph_bridge.h @@ -14,22 +14,18 @@ limitations under the License. */ #pragma once -#ifdef PADDLE_WITH_NGRAPH - #include #include #include #include -#include - -#include "paddle/fluid/framework/operator.h" -#include "paddle/fluid/platform/enforce.h" -#include "ngraph/ngraph.hpp" +#include "ngraph/node.hpp" namespace paddle { namespace framework { +class OperatorBase; + class NgraphBridge { public: static std::map< @@ -43,16 +39,15 @@ class NgraphBridge { std::shared_ptr< std::unordered_map>> var_node_map) - : ngb_node_map(var_node_map) {} + : ngb_node_map_(var_node_map) {} - void build_graph(const std::shared_ptr& op); + void BuildNgNode(const std::shared_ptr& op); private: std::shared_ptr< std::unordered_map>> - ngb_node_map; + ngb_node_map_; }; } // namespace framework } // namespace paddle -#endif diff --git a/paddle/fluid/framework/ngraph_operator.cc b/paddle/fluid/framework/ngraph_operator.cc index d967b2780c217..23f681ce886fd 100644 --- a/paddle/fluid/framework/ngraph_operator.cc +++ b/paddle/fluid/framework/ngraph_operator.cc @@ -12,21 +12,35 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#ifdef PADDLE_WITH_NGRAPH #include #include #include #include "paddle/fluid/framework/feed_fetch_type.h" +#include "paddle/fluid/framework/framework.pb.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/ngraph_bridge.h" #include "paddle/fluid/framework/ngraph_operator.h" -#include "paddle/fluid/framework/shape_inference.h" +#include "paddle/fluid/framework/tensor.h" #include "paddle/fluid/framework/var_desc.h" #include "paddle/fluid/framework/var_type.h" +#include "ngraph/ngraph.hpp" + namespace paddle { namespace framework { +static ngraph::Shape Ddim2Shape(const DDim& dims) { + ngraph::Shape sp; + for (int i = 0; i < dims.size(); ++i) { + int k = dims[i]; + k = k == 0 ? 1 : k; + sp.push_back(k); + } + return sp; +} + static std::map pd2ng_type_map = { {proto::VarType::FP32, ngraph::element::f32}, {proto::VarType::FP64, ngraph::element::f64}, @@ -42,16 +56,17 @@ typedef enum { /* nGraph support state on ops */ PARTIAL_TEST /* Support partial list of ops for test */ } op_state; -class NgraphOperator { +// perform graph build through bridge and execute computation +class NgraphEngine { public: - explicit NgraphOperator(const Scope& scope, const platform::Place& place, - const std::vector>& ops, - const std::unordered_map< - std::string, ngraph::element::Type>& var_type_map, - const std::unordered_set& persist, - const std::unordered_set& fetches, - const std::unordered_set& post_op_inputs, - op_state ng_op_state) + explicit NgraphEngine(const Scope& scope, const platform::Place& place, + const std::vector>& ops, + const std::unordered_map< + std::string, ngraph::element::Type>& var_type_map, + const std::unordered_set& persist, + const std::unordered_set& fetches, + const std::unordered_set& post_op_inputs, + op_state ng_op_state) : scope_(scope), place_(place), fused_ops_(ops), @@ -59,13 +74,23 @@ class NgraphOperator { persistables_(persist), fetches_(fetches), post_op_inputs_(post_op_inputs), - ng_op_state_(ng_op_state) {} + ng_op_state_(ng_op_state) { + var_in_node_map_ = std::make_shared< + std::unordered_map>>(); + + var_node_map_ = std::make_shared< + std::unordered_map>>(); + + BuildNgIO(); + + GetNgFunction(); + } void Run(const Scope& scope, const platform::Place& place) const; private: static std::unordered_map> - func_cache; + func_cache_; const Scope& scope_; const platform::Place& place_; std::vector> fused_ops_; @@ -74,10 +99,39 @@ class NgraphOperator { std::unordered_set fetches_; std::unordered_set post_op_inputs_; op_state ng_op_state_; + + // ngraph backend eg. CPU + static std::shared_ptr backend_; + // ngraph function to call and execute + std::shared_ptr ngraph_function_; + // var_name of inputs + std::vector var_in_; + // var_name of outputs from fetch in order + std::vector var_out_; + // map input vars to nodes + std::shared_ptr< + std::unordered_map>> + var_in_node_map_; + // map each var name with a ngraph node + std::shared_ptr< + std::unordered_map>> + var_node_map_; + // cache key to check if function is cached + std::shared_ptr GetCacheKey(); + // get ngraph input and define ngraph input parameters + void GetNgInputShape(std::shared_ptr op); + // Call ngraph bridge to map ops + void BuildNgNodes(); + // get the ngraph input and output var list + void BuildNgIO(); + // build ngraph function call + void BuildNgFunction(); + // Check cache for ngraph function or otherwise build the function + void GetNgFunction(); }; std::vector>::iterator>> -FusedOperator::FusedOpIntervals( +NgraphOperator::NgraphOpIntervals( std::vector>* ops) { std::vector>::iterator>> intervals; @@ -86,7 +140,7 @@ FusedOperator::FusedOpIntervals( } size_t size = ops->size(); size_t left = 0; - while (left < size && ops.at(left)->Type() != kFeedOpType) { + while (left < size && ops->at(left)->Type() != kFeedOpType) { ++left; } if (left == size) { @@ -116,7 +170,7 @@ FusedOperator::FusedOpIntervals( size_t start = pivot, end = start; while (pivot < right && (paddle::framework::NgraphBridge::NG_NODE_MAP.find( - ops.at(pivot)->Type()) != + ops->at(pivot)->Type()) != paddle::framework::NgraphBridge::NG_NODE_MAP.end())) { ++pivot; ++end; @@ -130,13 +184,15 @@ FusedOperator::FusedOpIntervals( return intervals; } -FusedOperator::FusedOperator( +NgraphOperator::NgraphOperator( const ProgramDesc& prog, size_t block_id, std::vector>::iterator start, std::vector>::iterator end, const std::string& type, const VariableNameMap& inputs, const VariableNameMap& outputs, const AttributeMap& attrs) - : OperatorBase(type, inputs, outputs, attrs), pdesc(prog), block(block_id) { + : OperatorBase(type, inputs, outputs, attrs), + pdesc_(prog), + block_(block_id) { for (std::vector>::iterator it = start; it != end; ++it) { fused_ops_.push_back(std::move(*it)); @@ -152,13 +208,13 @@ FusedOperator::FusedOperator( } if ((*(start - 1))->Type() == kFeedOpType && (*end)->Type() == kFetchOpType) { - is_complete = true; + is_full_ = true; } Process(); } -void FusedOperator::Process() { +void NgraphOperator::Process() { auto& bdesc = pdesc_.Block(block_); for (auto& var : bdesc.AllVars()) { if (!(var->GetType() == proto::VarType::SELECTED_ROWS || @@ -194,8 +250,8 @@ void FusedOperator::Process() { } } -void FusedOperator::RunImpl(const Scope& scope, - const platform::Place& place) const { +void NgraphOperator::RunImpl(const Scope& scope, + const platform::Place& place) const { op_state ng_op_state = PARTIAL_TEST; auto& bdesc = pdesc_.Block(block_); for (auto* op : bdesc.AllOps()) { @@ -205,16 +261,285 @@ void FusedOperator::RunImpl(const Scope& scope, } } - if (is_full) { + if (is_full_) { ng_op_state = ng_op_state == PARTIAL_TEST ? FULL_TEST : FULL_TRAIN; } - NgraphOperator ngraph_op(scope, place, fused_ops_, var_type_map_, - persistables_, fetches_, post_op_inputs_, - ng_op_state); - ngraph_op.Run(scope, place); + NgraphEngine ngraph_engine(scope, place, fused_ops_, var_type_map_, + persistables_, fetches_, post_op_inputs_, + ng_op_state); + ngraph_engine.Run(scope, place); +} + +std::unordered_map> + NgraphEngine::func_cache_ = {}; + +std::shared_ptr NgraphEngine::backend_ = + ngraph::runtime::Backend::create("CPU"); + +void NgraphEngine::GetNgInputShape(std::shared_ptr op) { + RuntimeContext ctx(op->Inputs(), op->Outputs(), scope_); + op->RuntimeInferShape(scope_, place_, ctx); + for (auto& var_name_item : op->Inputs()) { + for (auto& var_name : var_name_item.second) { + auto* var = scope_.FindVar(var_name); + if (var && var->IsType()) { + auto* tensor_pd = GetLoDTensorOrSelectedRowsValueFromVar(*var); + auto sp = Ddim2Shape(tensor_pd->dims()); + if (std::find(var_in_.begin(), var_in_.end(), var_name) != + var_in_.end()) { + if (var_node_map_->find(var_name) == var_node_map_->end()) { + auto ng_type = var_type_map_.at(var_name); + auto prm = + std::make_shared(ng_type, sp, true); + (*var_node_map_)[var_name] = prm; + (*var_in_node_map_)[var_name] = prm; + } + } + } + } + } +} + +void NgraphEngine::BuildNgNodes() { + for (auto& var_name : var_out_) { + if (var_node_map_->find(var_name) == var_node_map_->end()) { + auto* var = scope_.FindVar(var_name); + if (var && var->IsType()) { + auto* tensor_pd = GetLoDTensorOrSelectedRowsValueFromVar(*var); + auto& ddim = tensor_pd->dims(); + auto ng_shape = Ddim2Shape(ddim); + auto ng_type = var_type_map_.at(var_name); + auto prm = + std::make_shared(ng_type, ng_shape, true); + (*var_node_map_)[var_name] = prm; + } + } + } + + paddle::framework::NgraphBridge ngb(var_node_map_); + for (auto& op : fused_ops_) { + ngb.BuildNgNode(op); + } +} + +void NgraphEngine::BuildNgIO() { + std::unordered_set inputs; + std::unordered_set outputs; + + for (auto& op : fused_ops_) { + for (auto& var_name_item : op->Inputs()) { + for (auto& var_name : var_name_item.second) { + inputs.insert(var_name); + const bool is_output = outputs.find(var_name) != outputs.end(); + if (!is_output && + std::find(var_in_.begin(), var_in_.end(), var_name) == + var_in_.end()) { + // fill var_in here to keep lhs and rhs order + var_in_.push_back(var_name); + } + } + } + + if (op->Type() != "fill_constant") { + GetNgInputShape(op); + } + + for (auto& var_name_item : op->Outputs()) { + PADDLE_ENFORCE_LE(var_name_item.second.size(), 1, + "op %s has more than 1 output - Not handling yet", + op->Type()); + for (auto& var_name : var_name_item.second) { + outputs.insert(var_name); + } + } + } + + // var_out.clear(); + for (auto& op : fused_ops_) { + for (auto& var_name_item : op->Outputs()) { + PADDLE_ENFORCE_LE(var_name_item.second.size(), 1, + "op %s has more than 1 output - Not handling yet", + op->Type()); + for (auto& var_name : var_name_item.second) { + switch (ng_op_state_) { + case PARTIAL_TEST: + if (post_op_inputs_.find(var_name) != post_op_inputs_.end() || + fetches_.find(var_name) != fetches_.end()) { + var_out_.push_back(var_name); + } + break; + case FULL_TEST: + if (fetches_.find(var_name) != fetches_.end()) { + var_out_.push_back(var_name); + } + break; + case PARTIAL_TRAIN: + if (fetches_.find(var_name) != fetches_.end() || + post_op_inputs_.find(var_name) != post_op_inputs_.end() || + persistables_.find(var_name) != persistables_.end()) { + var_out_.push_back(var_name); + } + break; + case FULL_TRAIN: + if (fetches_.find(var_name) != fetches_.end() || + persistables_.find(var_name) != persistables_.end()) { + var_out_.push_back(var_name); + } + break; + default: + var_out_.push_back(var_name); + } + } + } + } } +void NgraphEngine::BuildNgFunction() { + BuildNgNodes(); + ngraph_function_ = nullptr; + ngraph::NodeVector func_outputs; + ngraph::op::ParameterVector func_inputs; + + for (auto& vo : var_out_) { + func_outputs.push_back(var_node_map_->at(vo)); + } + + for (auto& vi : var_in_) { + std::shared_ptr prm = + std::dynamic_pointer_cast( + var_in_node_map_->at(vi)); + func_inputs.push_back(prm); + } + + ngraph_function_ = + std::make_shared(func_outputs, func_inputs); +} + +std::shared_ptr NgraphEngine::GetCacheKey() { + auto cache_key = std::make_shared(""); + *cache_key += std::to_string(fused_ops_.size()); + for (auto& op : fused_ops_) { + *cache_key += op->Type(); + } + for (auto& var_name : var_in_) { + auto shape = var_node_map_->at(var_name)->get_shape(); + *cache_key += var_name; + *cache_key += var_type_map_.at(var_name).c_type_string(); + for (size_t i = 0; i < shape.size(); ++i) { + *cache_key += std::to_string(shape.at(i)); + } + } + + for (auto& var_name : var_out_) { + auto* var = scope_.FindVar(var_name); + if (var && var->IsType()) { + auto* tensor_pd = GetLoDTensorOrSelectedRowsValueFromVar(*var); + auto& ddim = tensor_pd->dims(); + for (int i = 0; i < ddim.size(); ++i) { + *cache_key += std::to_string(ddim[i]); + } + } + } + return cache_key; +} + +void NgraphEngine::GetNgFunction() { + bool cache_on = true; + if (cache_on) { + std::string cache_key_val = *GetCacheKey(); + if (func_cache_.find(cache_key_val) != func_cache_.end()) { + ngraph_function_ = func_cache_.at(cache_key_val); + } else { + BuildNgFunction(); + func_cache_[cache_key_val] = ngraph_function_; + } + } else { + BuildNgFunction(); + } +} + +void NgraphEngine::Run(const Scope& scope, const platform::Place& place) const { + std::vector> t_in; + std::vector> t_out; + + for (size_t i = 0; i < var_in_.size(); ++i) { + auto vi = var_in_.at(i); + auto sp = var_node_map_->at(vi)->get_shape(); + std::shared_ptr ti; + auto* var = scope.FindVar(vi); + if (var && var->IsType()) { + auto* tensor_pd = GetLoDTensorOrSelectedRowsValueFromVar(*var); + PADDLE_ENFORCE(sp == Ddim2Shape(tensor_pd->dims()), + "Ensure ngraph tensor layout align with paddle tensor"); + if (tensor_pd->type() == proto::VarType::FP32) { + const float* arr = tensor_pd->data(); + ti = backend_->create_tensor(ngraph::element::f32, sp, + const_cast(arr)); + } else if (tensor_pd->type() == proto::VarType::INT32) { + const int* arr = tensor_pd->data(); + ti = backend_->create_tensor(ngraph::element::i32, sp, + const_cast(arr)); + } else if (tensor_pd->type() == proto::VarType::INT64) { + const int64_t* arr = tensor_pd->data(); + ti = backend_->create_tensor(ngraph::element::i64, sp, + const_cast(arr)); + } else if (tensor_pd->type() == proto::VarType::FP64) { + const double* arr = tensor_pd->data(); + ti = backend_->create_tensor(ngraph::element::f64, sp, + const_cast(arr)); + } else if (tensor_pd->type() == proto::VarType::BOOL) { + const bool* arr = tensor_pd->data(); + ti = backend_->create_tensor(ngraph::element::boolean, sp, + const_cast(arr)); + } else { + PADDLE_THROW("Data type not handling for var %s", vi); + } + } else { + PADDLE_THROW("Cannot find var or tensor with var name %s", vi); + } + bool is_test = (ng_op_state_ == PARTIAL_TEST || ng_op_state_ == FULL_TEST) + ? true + : false; + bool is_persistable = + (persistables_.find(vi) != persistables_.end()) ? true : false; + if (is_test && is_persistable) { + ti->set_stale(false); + } + t_in.push_back(ti); + } + + for (size_t i = 0; i < var_out_.size(); ++i) { + auto var_name = var_out_[i]; + auto* var = scope.FindVar(var_name); + std::shared_ptr to; + if (var && var->IsType()) { + auto* tensor_pd = GetMutableLoDTensorOrSelectedRowsValueFromVar(var); + auto dd = tensor_pd->dims(); + ngraph::Shape sp = Ddim2Shape(dd); + auto ng_type = var_type_map_.at(var_name); + if (ng_type == ngraph::element::f32) { + auto pd_arr = tensor_pd->mutable_data(place); + to = backend_->create_tensor(ngraph::element::f32, sp, pd_arr); + } else if (ng_type == ngraph::element::i64) { + auto pd_arr = tensor_pd->mutable_data(place); + to = backend_->create_tensor(ngraph::element::i64, sp, pd_arr); + } else if (ng_type == ngraph::element::f64) { + auto pd_arr = tensor_pd->mutable_data(place); + to = backend_->create_tensor(ngraph::element::f64, sp, pd_arr); + } else if (ng_type == ngraph::element::boolean) { + auto pd_arr = tensor_pd->mutable_data(place); + to = backend_->create_tensor(ngraph::element::boolean, sp, pd_arr); + } else { + PADDLE_THROW("Data type not handled in for var %s", var_name); + } + t_out.push_back(to); + } else { + PADDLE_THROW("Cannot find var or tensor with var name %s", var_name); + } + } + + backend_->call(ngraph_function_, t_out, t_in); +} // NgraphEngine::RunImpl } // namespace framework } // namespace paddle -#endif diff --git a/paddle/fluid/framework/ngraph_operator.h b/paddle/fluid/framework/ngraph_operator.h index 0f655cef1dde6..ede80f44bea20 100644 --- a/paddle/fluid/framework/ngraph_operator.h +++ b/paddle/fluid/framework/ngraph_operator.h @@ -14,39 +14,32 @@ limitations under the License. */ #pragma once -#ifdef PADDLE_WITH_NGRAPH - #include -#include #include #include #include #include "paddle/fluid/framework/attribute.h" -#include "paddle/fluid/framework/framework.pb.h" -#include "paddle/fluid/framework/lod_tensor.h" -#include "paddle/fluid/framework/ngraph_bridge.h" #include "paddle/fluid/framework/op_info.h" #include "paddle/fluid/framework/op_kernel_type.h" #include "paddle/fluid/framework/operator.h" #include "paddle/fluid/framework/program_desc.h" #include "paddle/fluid/framework/scope.h" -#include "paddle/fluid/framework/tensor.h" #include "paddle/fluid/platform/variant.h" -#include "ngraph/ngraph.hpp" +#include "ngraph/type/element_type.hpp" namespace paddle { namespace framework { -class FusedOperator : public OperatorBase { +class NgraphOperator : public OperatorBase { public: static std::vector< std::vector>::iterator>> - FusedOpIntervals( + NgraphOpIntervals( std::vector>* ops); - explicit FusedOperator( + explicit NgraphOperator( const ProgramDesc& prog, size_t block_id, std::vector>::iterator start, std::vector>::iterator end, @@ -69,4 +62,3 @@ class FusedOperator : public OperatorBase { }; } // namespace framework } // namespace paddle -#endif diff --git a/paddle/fluid/framework/op_desc.cc b/paddle/fluid/framework/op_desc.cc index fbaa169df6324..2fe1c94ec02e8 100644 --- a/paddle/fluid/framework/op_desc.cc +++ b/paddle/fluid/framework/op_desc.cc @@ -81,29 +81,154 @@ class CompileTimeInferShapeContext : public InferShapeContext { "The %s[%d] is @EMPTY@", out, j); auto *in_var = block_.FindVarRecursive(Inputs(in)[i]); auto *out_var = block_.FindVarRecursive(Outputs(out)[j]); - if (in_var->GetType() != proto::VarType::LOD_TENSOR) { - VLOG(30) << "input " << in << " is not LodTensor"; + if (in_var->GetType() != proto::VarType::LOD_TENSOR && + in_var->GetType() != proto::VarType::LOD_TENSOR_ARRAY) { + VLOG(3) << "input " << in << " is not LodTensor or LodTensorArray."; return; } out_var->SetLoDLevel(in_var->GetLoDLevel()); } + void DecreaseLoDLevel(const std::string &in, const std::string &out, + size_t i = 0, size_t j = 0) const override { + PADDLE_ENFORCE_LT(i, Inputs(in).size()); + PADDLE_ENFORCE_LT(j, Outputs(out).size()); + PADDLE_ENFORCE(Inputs(in)[i] != framework::kEmptyVarName, + "The %s[%d] is @EMPTY@", in, i); + PADDLE_ENFORCE(Outputs(out)[j] != framework::kEmptyVarName, + "The %s[%d] is @EMPTY@", out, j); + auto *in_var = block_.FindVarRecursive(Inputs(in)[i]); + auto *out_var = block_.FindVarRecursive(Outputs(out)[j]); + PADDLE_ENFORCE(out_var->GetType() == proto::VarType::LOD_TENSOR_ARRAY || + out_var->GetType() == proto::VarType::LOD_TENSOR, + "The input %s should be LodTensorArray or LodTensor.", + out_var->Name()); + PADDLE_ENFORCE(in_var->GetType() == proto::VarType::LOD_TENSOR, + "The input %s should be LodTensor.", in_var->Name()); + if (in_var->GetLoDLevel() > 0) { + out_var->SetLoDLevel(in_var->GetLoDLevel() - 1); + } + } + + std::vector GetInputVarPtrs( + const std::string &name) override { + const std::vector arg_names = Inputs(name); + std::vector res; + res.reserve(arg_names.size()); + std::transform(arg_names.begin(), arg_names.end(), std::back_inserter(res), + [this](const std::string &name) { + return block_.FindVarRecursive(name); + }); + return res; + } + + std::vector GetOutputVarPtrs( + const std::string &name) override { + const std::vector arg_names = Outputs(name); + std::vector res; + res.reserve(arg_names.size()); + std::transform(arg_names.begin(), arg_names.end(), std::back_inserter(res), + [this](const std::string &name) { + return block_.FindVarRecursive(name); + }); + return res; + } + + DDim GetInputDim(const std::string &name) const override { + const std::vector &arg_names = Inputs(name); + PADDLE_ENFORCE_EQ(arg_names.size(), 1UL, + "Input(%s) should hold one element, but now it holds %d", + name, arg_names.size()); + return this->GetDim(arg_names[0]); + } + + std::vector GetInputsDim(const std::string &name) const override { + const std::vector &arg_names = Inputs(name); + return GetDims(arg_names); + } + bool IsRuntime() const override; + std::vector GetInputsVarType( + const std::string &name) const override { + return GetVarTypes(Inputs(name)); + } + + std::vector GetOutputsVarType( + const std::string &name) const override { + return GetVarTypes(Outputs(name)); + } + + void SetOutputDim(const std::string &name, const DDim &dim) override { + auto &arg_names = Outputs(name); + PADDLE_ENFORCE_EQ(arg_names.size(), 1UL, + "Output(%s) should hold one element, but now it holds %d", + name, arg_names.size()); + SetDim(arg_names[0], dim); + } + + void SetOutputsDim(const std::string &name, + const std::vector &dims) override { + auto &names = Outputs(name); + SetDims(names, dims); + } + protected: - proto::VarType::Type GetVarType(const std::string &name) const override; + std::vector GetVarTypes( + const std::vector &names) const { + std::vector retv; + retv.resize(names.size()); + std::transform( + names.begin(), names.end(), retv.begin(), + std::bind(std::mem_fn(&CompileTimeInferShapeContext::GetVarType), this, + std::placeholders::_1)); + return retv; + } + + proto::VarType::Type GetVarType(const std::string &name) const; + + DDim GetDim(const std::string &name) const { + auto var = block_.FindVarRecursive(name); + PADDLE_ENFORCE(var != nullptr, "Cannot find variable %s", name); + DDim res; + try { + auto shape = var->GetShape(); + res = shape.empty() ? make_ddim({0UL}) : make_ddim(shape); + } catch (...) { + VLOG(5) << "GetDim of variable " << name << " error"; + std::rethrow_exception(std::current_exception()); + } + return res; + } + + std::vector GetDims(const std::vector &names) const { + std::vector ret; + ret.reserve(names.size()); + std::transform( + names.begin(), names.end(), std::back_inserter(ret), + [this](const std::string &name) { return this->GetDim(name); }); + return ret; + } - DDim GetDim(const std::string &name) const override; + void SetDim(const std::string &name, const DDim &dim); - void SetDim(const std::string &name, const DDim &dim) override; + void SetDims(const std::vector &names, + const std::vector &dims) { + size_t length = names.size(); + PADDLE_ENFORCE_EQ(length, dims.size()); + for (size_t i = 0; i < length; ++i) { + if (names[i] == framework::kEmptyVarName) { + continue; + } + SetDim(names[i], dims[i]); + } + } std::vector GetRepeatedDims(const std::string &name) const override; void SetRepeatedDims(const std::string &name, const std::vector &dims) override; - InferShapeVarPtr GetVarPtr(const std::string &name) override; - const OpDesc &op_; const BlockDesc &block_; }; @@ -215,6 +340,23 @@ void OpDesc::SetOutput(const std::string ¶m_name, this->outputs_[param_name] = args; } +bool OpDesc::HasProtoAttr(const std::string &name) const { + auto &op_info = OpInfoMap::Instance(); + if (op_info.Has(desc_.type())) { + auto op_info_ptr = op_info.Get(desc_.type()); + if (op_info_ptr.HasOpProtoAndChecker()) { + const proto::OpProto &proto = op_info_ptr.Proto(); + for (int i = 0; i != proto.attrs_size(); ++i) { + const proto::OpProto::Attr &attr = proto.attrs(i); + if (attr.name() == name) { + return true; + } + } + } + } + return false; +} + proto::AttrType OpDesc::GetAttrType(const std::string &name) const { auto it = attrs_.find(name); PADDLE_ENFORCE(it != attrs_.end(), "Attribute %s is not found", name); @@ -241,32 +383,38 @@ void OpDesc::SetAttr(const std::string &name, const Attribute &v) { const proto::OpProto::Attr &attr = GetProtoAttr(name); switch (attr.type()) { case proto::AttrType::BOOLEANS: { - VLOG(110) << "SetAttr: " << Type() << ", " << name - << " from INTS to BOOLEANS"; + VLOG(11) << "SetAttr: " << Type() << ", " << name + << " from INTS to BOOLEANS"; this->attrs_[name] = std::vector(); break; } case proto::AttrType::INTS: { - VLOG(110) << "SetAttr: " << Type() << ", " << name - << " from INTS to INTS"; + VLOG(11) << "SetAttr: " << Type() << ", " << name + << " from INTS to INTS"; this->attrs_[name] = std::vector(); break; } + case proto::AttrType::LONGS: { + VLOG(11) << "SetAttr: " << Type() << ", " << name + << " from LONGS to LONGS"; + this->attrs_[name] = std::vector(); + break; + } case proto::AttrType::FLOATS: { - VLOG(110) << "SetAttr: " << Type() << ", " << name - << " from INTS to FLOATS"; + VLOG(11) << "SetAttr: " << Type() << ", " << name + << " from INTS to FLOATS"; this->attrs_[name] = std::vector(); break; } case proto::AttrType::STRINGS: { - VLOG(110) << "SetAttr: " << Type() << ", " << name - << " from INTS to STRINGS"; + VLOG(11) << "SetAttr: " << Type() << ", " << name + << " from INTS to STRINGS"; this->attrs_[name] = std::vector(); break; } case proto::AttrType::BLOCKS: { - VLOG(110) << "SetAttr: " << Type() << ", " << name - << " from INTS to BLOCKS"; + VLOG(11) << "SetAttr: " << Type() << ", " << name + << " from INTS to BLOCKS"; this->SetBlocksAttr(name, std::vector()); return; } @@ -499,13 +647,13 @@ void OpDesc::CheckAttrs() { } void OpDesc::InferShape(const BlockDesc &block) const { - VLOG(30) << "CompileTime infer shape on " << Type(); + VLOG(3) << "CompileTime infer shape on " << Type(); InitInferShapeFuncs(); auto &infer_shape = OpInfoMap::Instance().Get(this->Type()).infer_shape_; PADDLE_ENFORCE(static_cast(infer_shape), "%s's infer_shape has not been registered", this->Type()); CompileTimeInferShapeContext ctx(*this, block); - if (VLOG_IS_ON(100)) { + if (VLOG_IS_ON(10)) { std::ostringstream sout; auto inames = this->InputArgumentNames(); sout << " From ["; @@ -516,7 +664,7 @@ void OpDesc::InferShape(const BlockDesc &block) const { std::copy(onames.begin(), onames.end(), std::ostream_iterator(sout, ", ")); sout << "]"; - VLOG(100) << sout.str(); + VLOG(10) << sout.str(); } infer_shape(&ctx); } @@ -599,20 +747,6 @@ const std::vector &CompileTimeInferShapeContext::Outputs( return op_.Output(name); } -DDim CompileTimeInferShapeContext::GetDim(const std::string &name) const { - auto var = block_.FindVarRecursive(name); - PADDLE_ENFORCE(var != nullptr, "Cannot find variable %s", name); - DDim res; - try { - auto shape = var->GetShape(); - res = shape.empty() ? make_ddim({0UL}) : make_ddim(shape); - } catch (...) { - VLOG(50) << "GetDim of variable " << name << " error"; - std::rethrow_exception(std::current_exception()); - } - return res; -} - std::vector CompileTimeInferShapeContext::GetRepeatedDims( const std::string &name) const { auto var = block_.FindVarRecursive(name); @@ -624,7 +758,7 @@ std::vector CompileTimeInferShapeContext::GetRepeatedDims( res.push_back(s.empty() ? make_ddim({0UL}) : make_ddim(s)); } } catch (...) { - VLOG(50) << "GetRepeatedDim of variable " << name << " error."; + VLOG(5) << "GetRepeatedDim of variable " << name << " error."; std::rethrow_exception(std::current_exception()); } return res; @@ -651,10 +785,5 @@ proto::VarType::Type CompileTimeInferShapeContext::GetVarType( return block_.FindVarRecursive(name)->GetType(); } -InferShapeVarPtr CompileTimeInferShapeContext::GetVarPtr( - const std::string &name) { - return block_.FindVarRecursive(name); -} - } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/op_desc.h b/paddle/fluid/framework/op_desc.h index 30c8a26c3d2f0..d7352c5ee5a63 100644 --- a/paddle/fluid/framework/op_desc.h +++ b/paddle/fluid/framework/op_desc.h @@ -65,6 +65,8 @@ class OpDesc { return attrs_.find(name) != attrs_.end(); } + bool HasProtoAttr(const std::string &name) const; + proto::AttrType GetAttrType(const std::string &name) const; std::vector AttrNames() const; @@ -121,6 +123,8 @@ class OpDesc { BlockDesc *Block() { return this->block_; } + const BlockDesc *Block() const { return this->block_; } + private: template static std::vector MapKeys(const MapType &map) { diff --git a/paddle/fluid/framework/op_kernel_type.cc b/paddle/fluid/framework/op_kernel_type.cc new file mode 100644 index 0000000000000..6d4801e4a0eed --- /dev/null +++ b/paddle/fluid/framework/op_kernel_type.cc @@ -0,0 +1,54 @@ +/* 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/op_kernel_type.h" + +namespace paddle { +namespace framework { + +size_t OpKernelType::Hash::operator()(const OpKernelType& key) const { + int cur_loc = 0; + + int place = key.place_.which(); + cur_loc += OpKernelType::kPlaceBits; + + int data_type = static_cast(key.data_type_) << cur_loc; + cur_loc += OpKernelType::kPrimaryDTypeBits; + + int data_layout = static_cast(key.data_layout_) << cur_loc; + cur_loc += OpKernelType::kLayoutBits; + + int library_type = static_cast(key.library_type_) << cur_loc; + cur_loc += OpKernelType::kLibBits; + + int customized_value = key.customized_type_value_; + PADDLE_ENFORCE(customized_value < (1 << OpKernelType::kCustomizeBits)); + customized_value = customized_value << cur_loc; + cur_loc += OpKernelType::kCustomizeBits; + PADDLE_ENFORCE(cur_loc < 64); + + std::hash hasher; + return hasher(place + data_type + data_layout + library_type + + customized_value); +} + +bool OpKernelType::operator==(const OpKernelType& o) const { + return platform::places_are_same_class(place_, o.place_) && + data_type_ == o.data_type_ && data_layout_ == o.data_layout_ && + library_type_ == o.library_type_ && + customized_type_value_ == o.customized_type_value_; +} + +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/op_kernel_type.h b/paddle/fluid/framework/op_kernel_type.h index ac03302189731..9edc1a3e15002 100644 --- a/paddle/fluid/framework/op_kernel_type.h +++ b/paddle/fluid/framework/op_kernel_type.h @@ -24,54 +24,55 @@ limitations under the License. */ namespace paddle { namespace framework { -struct OpKernelType { - struct Hash { - size_t operator()(const OpKernelType& key) const { - int place = key.place_.which(); - int data_type = static_cast(key.data_type_) << LEFT_SHIFT; - int data_layout = static_cast(key.data_layout_) << (LEFT_SHIFT * 2); - int library_type = static_cast(key.library_type_) - << (LEFT_SHIFT * 3); - - std::hash hasher; - return hasher(place + data_type + data_layout + library_type); - } - }; +class OpKernelType { + public: + constexpr static int kDefaultCustomizedTypeValue = 0; - // place, data_type, library_type kinds less than 2^8 - constexpr static int LEFT_SHIFT = 8; - - proto::VarType::Type data_type_; - DataLayout data_layout_; - platform::Place place_; - LibraryType library_type_; + // In total should be smaller than 64. + constexpr static int kPlaceBits = 4; + constexpr static int kPrimaryDTypeBits = 8; + constexpr static int kLayoutBits = 4; + constexpr static int kLibBits = 4; + constexpr static int kCustomizeBits = 4; OpKernelType(proto::VarType::Type data_type, platform::Place place, DataLayout data_layout = DataLayout::kAnyLayout, - LibraryType library_type = LibraryType::kPlain) + LibraryType library_type = LibraryType::kPlain, + int customized_type_value = kDefaultCustomizedTypeValue) : data_type_(data_type), data_layout_(data_layout), place_(place), - library_type_(library_type) {} + library_type_(library_type), + customized_type_value_(customized_type_value) {} OpKernelType(proto::VarType::Type data_type, const platform::DeviceContext& dev_ctx, DataLayout data_layout = DataLayout::kAnyLayout, - LibraryType library_type = LibraryType::kPlain) + LibraryType library_type = LibraryType::kPlain, + int customized_type_value = kDefaultCustomizedTypeValue) : data_type_(data_type), data_layout_(data_layout), place_(dev_ctx.GetPlace()), - library_type_(library_type) {} + library_type_(library_type), + customized_type_value_(customized_type_value) {} + + virtual ~OpKernelType() {} + + struct Hash { + size_t operator()(const OpKernelType& key) const; + }; size_t hash_key() const { return Hash()(*this); } - bool operator==(const OpKernelType& o) const { - return platform::places_are_same_class(place_, o.place_) && - data_type_ == o.data_type_ && data_layout_ == o.data_layout_ && - library_type_ == o.library_type_; - } + bool operator==(const OpKernelType& o) const; bool operator!=(const OpKernelType& o) const { return !(*this == o); } + + proto::VarType::Type data_type_; + DataLayout data_layout_; + platform::Place place_; + LibraryType library_type_; + int customized_type_value_; }; inline std::ostream& operator<<(std::ostream& os, diff --git a/paddle/fluid/framework/op_kernel_type_test.cc b/paddle/fluid/framework/op_kernel_type_test.cc index 3e17a512ce154..40db85400d2c8 100644 --- a/paddle/fluid/framework/op_kernel_type_test.cc +++ b/paddle/fluid/framework/op_kernel_type_test.cc @@ -34,7 +34,8 @@ TEST(OpKernelType, ToString) { OpKernelType op_kernel_type2(DataType::FP16, CUDAPlace(0), DataLayout::kNCHW, LibraryType::kCUDNN); ASSERT_EQ(paddle::framework::KernelTypeToString(op_kernel_type2), - "data_type[float16]:data_layout[NCHW]:place[CUDAPlace(0)]:library_" + "data_type[::paddle::platform::float16]:data_layout[NCHW]:place[" + "CUDAPlace(0)]:library_" "type[CUDNN]"); } diff --git a/paddle/fluid/framework/op_registry.cc b/paddle/fluid/framework/op_registry.cc index 4a841bae8323f..bfc411ca2c4a4 100644 --- a/paddle/fluid/framework/op_registry.cc +++ b/paddle/fluid/framework/op_registry.cc @@ -46,9 +46,9 @@ static VariableNameMap ConvertOpDescVarsToVarNameMap( std::unique_ptr OpRegistry::CreateOp( const proto::OpDesc& op_desc) { - VLOG(10) << "CreateOp directly from OpDesc is deprecated. It should only be" - "used in unit tests. Use CreateOp(const OpDesc& op_desc) " - "instead."; + VLOG(1) << "CreateOp directly from OpDesc is deprecated. It should only be" + "used in unit tests. Use CreateOp(const OpDesc& op_desc) " + "instead."; VariableNameMap inputs = ConvertOpDescVarsToVarNameMap(op_desc.inputs()); VariableNameMap outputs = ConvertOpDescVarsToVarNameMap(op_desc.outputs()); AttributeMap attrs; diff --git a/paddle/fluid/framework/op_registry.h b/paddle/fluid/framework/op_registry.h index 0e6e74293c30d..6d39bb3c524b4 100644 --- a/paddle/fluid/framework/op_registry.h +++ b/paddle/fluid/framework/op_registry.h @@ -35,6 +35,7 @@ limitations under the License. */ namespace paddle { namespace framework { + class Registrar { public: // In our design, various kinds of classes, e.g., operators and kernels, @@ -78,7 +79,7 @@ struct OpKernelRegistrarFunctor; template inline void RegisterKernelClass(const char* op_type, const char* library_type, - Func func) { + int customized_type_value, Func func) { std::string library(library_type); std::string data_layout = "ANYLAYOUT"; if (library == "MKLDNN") { @@ -86,7 +87,7 @@ inline void RegisterKernelClass(const char* op_type, const char* library_type, } OpKernelType key(ToDataType(std::type_index(typeid(T))), PlaceType(), StringToDataLayout(data_layout), - StringToLibraryType(library_type)); + StringToLibraryType(library_type), customized_type_value); OperatorWithKernel::AllOpKernels()[op_type][key] = func; } @@ -95,22 +96,26 @@ struct OpKernelRegistrarFunctor { using KERNEL_TYPE = typename std::tuple_element>::type; - void operator()(const char* op_type, const char* library_type) const { + void operator()(const char* op_type, const char* library_type, + int customized_type_value) const { using T = typename KERNEL_TYPE::ELEMENT_TYPE; RegisterKernelClass( - op_type, library_type, [](const framework::ExecutionContext& ctx) { + op_type, library_type, customized_type_value, + + [](const framework::ExecutionContext& ctx) { KERNEL_TYPE().Compute(ctx); }); constexpr auto size = std::tuple_size>::value; OpKernelRegistrarFunctor func; - func(op_type, library_type); + func(op_type, library_type, customized_type_value); } }; template struct OpKernelRegistrarFunctor { - void operator()(const char* op_type, const char* library_type) const {} + void operator()(const char* op_type, const char* library_type, + int customized_type_value) const {} }; // User can register many kernel in one place. The data type could be @@ -118,9 +123,10 @@ struct OpKernelRegistrarFunctor { template class OpKernelRegistrar : public Registrar { public: - explicit OpKernelRegistrar(const char* op_type, const char* library_type) { + explicit OpKernelRegistrar(const char* op_type, const char* library_type, + int customized_type_value) { OpKernelRegistrarFunctor func; - func(op_type, library_type); + func(op_type, library_type, customized_type_value); } }; @@ -130,17 +136,19 @@ struct OpKernelRegistrarFunctorEx; template class OpKernelRegistrarEx : public Registrar { public: - explicit OpKernelRegistrarEx(const char* op_type, const char* library_type) { + explicit OpKernelRegistrarEx(const char* op_type, const char* library_type, + int customized_type_value) { OpKernelRegistrarFunctorEx func; - func(op_type, library_type); + func(op_type, library_type, customized_type_value); } }; template struct OpKernelRegistrarFunctorEx { - void operator()(const char* op_type, const char* library_type) const {} + void operator()(const char* op_type, const char* library_type, + int customized_type_value) const {} }; template @@ -153,18 +161,21 @@ struct OpKernelRegistrarFunctorEx>::type; - void operator()(const char* op_type, const char* library_type) const { - RegisterKernelClass(op_type, library_type, Functor()); + void operator()(const char* op_type, const char* library_type, + int customized_type_value) const { + RegisterKernelClass(op_type, library_type, + customized_type_value, Functor()); constexpr auto size = std::tuple_size>::value; OpKernelRegistrarFunctorEx= size, I + 2, DataTypeAndKernelType...> func; - func(op_type, library_type); + func(op_type, library_type, customized_type_value); } }; +// clang-format off /** * check if MACRO is used in GLOBAL NAMESPACE. */ @@ -199,42 +210,64 @@ struct OpKernelRegistrarFunctorEx \ - __op_kernel_registrar_##op_type##_##library_type##__(#op_type, \ - #library_type); \ - int TouchOpKernelRegistrar_##op_type##_##library_type() { \ - __op_kernel_registrar_##op_type##_##library_type##__.Touch(); \ - return 0; \ +#define REGISTER_OP_KERNEL_WITH_CUSTOM_TYPE(op_type, library_type, \ + place_class, customized_name, \ + customized_type_value, ...) \ + STATIC_ASSERT_GLOBAL_NAMESPACE( \ + __reg_op_kernel_##op_type##_##library_type##_##customized_name##__, \ + "REGISTER_OP_KERNEL must be called in " \ + "global namespace"); \ + static ::paddle::framework::OpKernelRegistrar \ + __op_kernel_registrar_##op_type##_##library_type##_##customized_name##__(\ + #op_type, #library_type, customized_type_value); \ + int TouchOpKernelRegistrar_##op_type##_##library_type##_##customized_name() {\ + __op_kernel_registrar_##op_type##_##library_type##_##customized_name##__ \ + .Touch(); \ + return 0; \ } +#define REGISTER_OP_KERNEL(op_type, library_type, place_class, ...) \ + REGISTER_OP_KERNEL_WITH_CUSTOM_TYPE( \ + op_type, library_type, place_class, DEFAULT_TYPE, \ + ::paddle::framework::OpKernelType::kDefaultCustomizedTypeValue, \ + __VA_ARGS__) + #define REGISTER_OP_CUDA_KERNEL(op_type, ...) \ REGISTER_OP_KERNEL(op_type, CUDA, ::paddle::platform::CUDAPlace, __VA_ARGS__) #define REGISTER_OP_CPU_KERNEL(op_type, ...) \ REGISTER_OP_KERNEL(op_type, CPU, ::paddle::platform::CPUPlace, __VA_ARGS__) -#define REGISTER_OP_KERNEL_EX(op_type, library_type, place_class, ...) \ - STATIC_ASSERT_GLOBAL_NAMESPACE( \ - __reg_op_kernel_##op_type##_##library_type##__, \ - "REGISTER_OP_KERNEL_EX must be called in global namespace"); \ - static ::paddle::framework::OpKernelRegistrarEx \ - __op_kernel_registrar_##op_type##_##library_type##__(#op_type, \ - #library_type); \ - int TouchOpKernelRegistrar_##op_type##_##library_type() { \ - __op_kernel_registrar_##op_type##_##library_type##__.Touch(); \ - return 0; \ +#define REGISTER_OP_KERNEL_EX(op_type, library_type, place_class, \ + customized_name, \ + customized_type_value, \ + ...) \ + STATIC_ASSERT_GLOBAL_NAMESPACE( \ + __reg_op_kernel_##op_type##_##library_type##_##customized_name##__, \ + "REGISTER_OP_KERNEL_EX must be called in " \ + "global namespace"); \ + static ::paddle::framework::OpKernelRegistrarEx \ + __op_kernel_registrar_##op_type##_##library_type##_##customized_name##__(\ + #op_type, #library_type, customized_type_value); \ + int TouchOpKernelRegistrar_##op_type##_##library_type##_##customized_name() {\ + __op_kernel_registrar_##op_type##_##library_type##_##customized_name##__ \ + .Touch(); \ + return 0; \ } #define REGISTER_OP_CUDA_KERNEL_FUNCTOR(op_type, ...) \ - REGISTER_OP_KERNEL_EX(op_type, CUDA, ::paddle::platform::CUDAPlace, \ - __VA_ARGS__) + REGISTER_OP_KERNEL_EX( \ + op_type, CUDA, ::paddle::platform::CUDAPlace, DEFAULT_TYPE, \ + ::paddle::framework::OpKernelType::kDefaultCustomizedTypeValue, \ + __VA_ARGS__) -#define REGISTER_OP_CPU_KERNEL_FUNCTOR(op_type, ...) \ - REGISTER_OP_KERNEL_EX(op_type, CPU, ::paddle::platform::CPUPlace, __VA_ARGS__) +#define REGISTER_OP_CPU_KERNEL_FUNCTOR(op_type, ...) \ + REGISTER_OP_KERNEL_EX( \ + op_type, CPU, ::paddle::platform::CPUPlace, DEFAULT_TYPE, \ + ::paddle::framework::OpKernelType::kDefaultCustomizedTypeValue, \ + __VA_ARGS__) /** * Macro to mark what Operator and Kernel @@ -248,13 +281,19 @@ struct OpKernelRegistrarFunctorEx> 2); -} - std::vector> kKernelPriority = { std::make_tuple(platform::CUDAPlace(0), LibraryType::kCUDNN), std::make_tuple(platform::CUDAPlace(0), LibraryType::kPlain), @@ -47,10 +43,9 @@ std::vector> kKernelPriority = { proto::VarType::Type GetDataTypeOfVar(const Variable* var) { if (var->IsType()) { - return framework::ToDataType(var->Get().type()); + return var->Get().type(); } else if (var->IsType()) { - return framework::ToDataType( - var->Get().value().type()); + return var->Get().value().type(); } else { PADDLE_THROW("Var should be LoDTensor or SelectedRows"); } @@ -97,13 +92,13 @@ static std::string GetDtype(const Scope& scope, const std::string& name) { if (UNLIKELY(!tensor.IsInitialized())) { return ""; } - return DataTypeToString(ToDataType(tensor.type())); + return DataTypeToString(tensor.type()); } else if (var->IsType()) { auto tensor = var->Get().value(); if (UNLIKELY(!tensor.IsInitialized())) { return "uninited"; } else { - return DataTypeToString(ToDataType(tensor.type())); + return DataTypeToString(tensor.type()); } } else { return ""; @@ -142,8 +137,27 @@ static LoD GetLoD(const Scope& scope, const std::string& name) { } } +RuntimeContext::RuntimeContext(const VariableNameMap& innames, + const VariableNameMap& outnames, + const Scope& scope) { + for (auto& var_name_item : innames) { + std::vector& input_vars = inputs[var_name_item.first]; + input_vars.reserve(var_name_item.second.size()); + for (auto& var_name : var_name_item.second) { + input_vars.push_back(scope.FindVar(var_name)); + } + } + for (auto& var_name_item : outnames) { + std::vector& output_vars = outputs[var_name_item.first]; + output_vars.reserve(var_name_item.second.size()); + for (auto& var_name : var_name_item.second) { + output_vars.push_back(scope.FindVar(var_name)); + } + } +} + void OperatorBase::Run(const Scope& scope, const platform::Place& place) { - VLOG(40) << place << " " << DebugStringEx(&scope); + VLOG(4) << place << " " << DebugStringEx(&scope); if (platform::is_gpu_place(place)) { #ifndef PADDLE_WITH_CUDA PADDLE_THROW("Cannot run operator on place %s", place); @@ -153,20 +167,17 @@ void OperatorBase::Run(const Scope& scope, const platform::Place& place) { #endif } -// The profile has a process-wide mutex, results in serious performance issue -// in concurrency scenerio. Here use an `if` to fix this issue. -// Please not remove the `if`, ask @Superjomn if there are any concern. -#ifndef _WIN32 + // The profile has a process-wide mutex, results in serious performance issue + // in concurrency scenerio. Here use an `if` to fix this issue. + // Please not remove the `if`, ask @Superjomn if there are any concern. if (platform::IsProfileEnabled()) { platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); platform::RecordEvent record_event(Type(), pool.Get(place)); RunImpl(scope, place); - } else -#endif - { + } else { RunImpl(scope, place); } - VLOG(30) << place << " " << DebugStringEx(&scope); + VLOG(3) << place << " " << DebugStringEx(&scope); } bool OperatorBase::HasInputs(const std::string& name) const { @@ -420,14 +431,73 @@ bool ExecutionContext::HasOutput(const std::string& name) const { return var != nullptr; } +const Variable* ExecutionContext::InputVar(const std::string& name) const { + auto it = ctx_.inputs.find(name); + if (it == ctx_.inputs.end()) return nullptr; + + PADDLE_ENFORCE_LE(it->second.size(), 1UL, + "Operator %s's input %s should contain only one variable.", + op_.Type(), name); + return it->second.empty() ? nullptr : it->second[0]; +} + +const Variable* ExecutionContext::LegacyInputVar( + const std::string& name) const { + auto ipt = op_.Input(name); + return ipt == kEmptyVarName ? nullptr : scope_.FindVar(ipt); +} + +Variable* ExecutionContext::OutputVar(const std::string& name) const { + auto it = ctx_.outputs.find(name); + if (it == ctx_.outputs.end()) return nullptr; + + PADDLE_ENFORCE_LE(it->second.size(), 1UL, + "Operator %s's output %s should contain only one variable.", + op_.Type(), name); + return it->second.empty() ? nullptr : it->second[0]; +} + +Variable* ExecutionContext::LegacyOutputVar(const std::string& name) const { + auto opt = op_.Output(name); + return opt == kEmptyVarName ? nullptr : scope_.FindVar(opt); +} + template <> const Tensor* ExecutionContext::Input(const std::string& name) const { return Input(name); } +template <> +const Tensor* ExecutionContext::LegacyInput( + const std::string& name) const { + return LegacyInput(name); +} + template <> const std::vector ExecutionContext::MultiInput( const std::string& name) const { + auto it = ctx_.inputs.find(name); + if (it == ctx_.inputs.end()) { + return {}; + } + const std::vector& vars = it->second; + std::vector res; + res.reserve(vars.size()); + std::transform(vars.begin(), vars.end(), std::back_inserter(res), + [&](Variable* var) -> const Tensor* { + if (var == nullptr) return nullptr; + PADDLE_ENFORCE( + var->IsType(), + "should be LoDTensor, but the received type is %s", + var->Type().name()); + return &(var->Get()); + }); + return res; +} + +template <> +const std::vector ExecutionContext::LegacyMultiInput( + const std::string& name) const { auto names = op().Inputs(name); std::vector res; res.reserve(names.size()); @@ -449,6 +519,11 @@ Tensor* ExecutionContext::Output(const std::string& name) const { return Output(name); } +template <> +Tensor* ExecutionContext::LegacyOutput(const std::string& name) const { + return LegacyOutput(name); +} + template <> std::vector ExecutionContext::MultiOutput( const std::string& name) const { @@ -485,51 +560,48 @@ bool OpSupportGPU(const std::string& op_type) { class RuntimeInferShapeContext : public InferShapeContext { public: - RuntimeInferShapeContext(const OperatorBase& op, const Scope& scope) - : op_(op), scope_(scope) {} + RuntimeInferShapeContext(const OperatorBase& op, const Scope& scope, + const RuntimeContext& ctx) + : op_(op), scope_(scope), ctx_(ctx) {} bool HasInput(const std::string& name) const override { // has only one input - const auto& ins = op_.Inputs(); + const auto& ins = ctx_.inputs; auto it = ins.find(name); if (it == ins.end()) { return false; } const auto& in = it->second; - if (in.size() == 0 || in[0] == kEmptyVarName) { - return false; - } + if (in.size() == 0) return false; PADDLE_ENFORCE_EQ(in.size(), 1UL, "Input %s should not have more than one inputs", name); - return scope_.FindVar(in[0]) != nullptr; + return in[0] != nullptr; } bool HasOutput(const std::string& name) const override { // has only one output - const auto& outs = op_.Outputs(); + const auto& outs = ctx_.outputs; auto it = outs.find(name); if (it == outs.end()) { return false; } const auto& out = it->second; - if (out.size() == 0 || out[0] == kEmptyVarName) { + if (out.size() == 0) { return false; } PADDLE_ENFORCE_EQ(out.size(), 1UL, "Output %s should not have more than one outputs", name); - return scope_.FindVar(out[0]) != nullptr; + return out[0] != nullptr; } bool HasInputs(const std::string& name) const override { - if (!op_.HasInputs(name)) { - return false; - } - auto inputs = op_.Inputs(name); - if (inputs.empty()) { + const auto& ins = ctx_.inputs; + auto it = ins.find(name); + if (it == ins.end() || it->second.empty()) { return false; } - for (auto& input : inputs) { - if (scope_.FindVar(input) == nullptr) { + for (auto& input : it->second) { + if (input == nullptr) { return false; } } @@ -537,15 +609,13 @@ class RuntimeInferShapeContext : public InferShapeContext { } bool HasOutputs(const std::string& name) const override { - if (!op_.HasOutputs(name)) { - return false; - } - auto outputs = op_.Outputs(name); - if (outputs.empty()) { + const auto& outs = ctx_.outputs; + auto it = outs.find(name); + if (it == outs.end() || it->second.empty()) { return false; } - for (auto& output : outputs) { - if (scope_.FindVar(output) == nullptr) { + for (auto& output : it->second) { + if (output == nullptr) { return false; } } @@ -566,16 +636,18 @@ class RuntimeInferShapeContext : public InferShapeContext { void ShareDim(const std::string& in, const std::string& out, size_t i = 0, size_t j = 0) override { - PADDLE_ENFORCE_LT(i, Inputs(in).size()); - PADDLE_ENFORCE_LT(j, Outputs(out).size()); - const std::string& input_n = Inputs(in)[i]; - const std::string& output_n = Outputs(out)[j]; + auto in_it = ctx_.inputs.find(in); + auto out_it = ctx_.outputs.find(out); + PADDLE_ENFORCE(in_it != ctx_.inputs.end() && in_it->second.size() > i, + "Inputs %s should have %llu argument", in, i); + PADDLE_ENFORCE(out_it != ctx_.outputs.end() && out_it->second.size() > j, + "Outputs %s should have %llu argument", out, j); + + Variable* in_var = in_it->second[i]; + Variable* out_var = out_it->second[j]; - Variable* in_var = scope_.FindVar(input_n); - Variable* out_var = scope_.FindVar(output_n); PADDLE_ENFORCE(in_var->Type() == out_var->Type(), - "The type of %s and %s is not the same.", output_n, - GetDim(input_n)); + "The type of %s and %s is not the same.", in, out); if (in_var->IsType()) { auto& in_sele_rows = in_var->Get(); @@ -596,13 +668,16 @@ class RuntimeInferShapeContext : public InferShapeContext { void ShareLoD(const std::string& in, const std::string& out, size_t i = 0, size_t j = 0) const override { - const std::vector& inputs = Inputs(in); - const std::vector& outputs = Outputs(out); - PADDLE_ENFORCE_LT(i, inputs.size()); - PADDLE_ENFORCE_LT(j, outputs.size()); - Variable* in_var = scope_.FindVar(inputs.at(i)); + auto in_it = ctx_.inputs.find(in); + auto out_it = ctx_.outputs.find(out); + PADDLE_ENFORCE(in_it != ctx_.inputs.end() && in_it->second.size() > i, + "Inputs %s should have %llu argument", in, i); + PADDLE_ENFORCE(out_it != ctx_.outputs.end() && out_it->second.size() > j, + "Outputs %s should have %llu argument", out, j); + + Variable* in_var = in_it->second.at(i); if (!in_var->IsType()) return; - Variable* out_var = scope_.FindVar(outputs.at(j)); + Variable* out_var = out_it->second.at(j); PADDLE_ENFORCE(out_var->IsType(), "The %d-th output of Output(%s) must be LoDTensor.", j, out); auto in_tensor = in_var->Get(); @@ -630,11 +705,71 @@ class RuntimeInferShapeContext : public InferShapeContext { out_tensor->set_layout(in_tensor.layout()); } + void DecreaseLoDLevel(const std::string& in, const std::string& out, + size_t i = 0, size_t j = 0) const override { + PADDLE_THROW("DecreaseLoDLevel is only used in compile time."); + } + bool IsRuntime() const override { return true; } + // TODO(paddle-dev): Can this be template? + std::vector GetInputVarPtrs( + const std::string& name) override { + const std::vector& vars = InputVars(name); + std::vector res; + res.reserve(vars.size()); + res.insert(res.begin(), vars.begin(), vars.end()); + return res; + } + + std::vector GetOutputVarPtrs( + const std::string& name) override { + const std::vector& vars = OutputVars(name); + std::vector res; + res.reserve(vars.size()); + res.insert(res.begin(), vars.begin(), vars.end()); + return res; + } + + DDim GetInputDim(const std::string& name) const override { + const std::vector& vars = InputVars(name); + PADDLE_ENFORCE_EQ(vars.size(), 1UL, + "Input(%s) should hold one element, but now it holds %d", + name, vars.size()); + return this->GetDim(vars[0]); + } + + std::vector GetInputsDim(const std::string& name) const override { + const std::vector& vars = InputVars(name); + return GetDims(vars); + } + + std::vector GetInputsVarType( + const std::string& name) const override { + return GetVarTypes(InputVars(name)); + } + + std::vector GetOutputsVarType( + const std::string& name) const override { + return GetVarTypes(OutputVars(name)); + } + + void SetOutputDim(const std::string& name, const DDim& dim) override { + auto& vars = OutputVars(name); + PADDLE_ENFORCE_EQ(vars.size(), 1UL, + "Output(%s) should hold one element, but now it holds %d", + name, vars.size()); + SetDim(vars[0], dim); + } + + void SetOutputsDim(const std::string& name, + const std::vector& dims) override { + auto& vars = OutputVars(name); + SetDims(vars, dims); + } + protected: - DDim GetDim(const std::string& name) const override { - Variable* var = scope_.FindVar(name); + DDim GetDim(Variable* var) const { PADDLE_ENFORCE_NOT_NULL(var); if (var->IsType()) { return var->Get().dims(); @@ -642,25 +777,44 @@ class RuntimeInferShapeContext : public InferShapeContext { return var->Get().GetCompleteDims(); } else { PADDLE_THROW( - "Only LoDTensor/SelectedRows support 'GetDim', but Variable %s's " + "Only LoDTensor/SelectedRows support 'GetDim', but Variables " "type_id is %s.", - name, var->Type().name()); + var->Type().name()); } } + std::vector GetDims(const std::vector& vars) const { + std::vector ret; + ret.reserve(vars.size()); + std::transform(vars.begin(), vars.end(), std::back_inserter(ret), + [this](Variable* var) { return this->GetDim(var); }); + return ret; + } + std::vector GetRepeatedDims(const std::string& name) const override { PADDLE_THROW("Only compile time support this method"); } - void SetDim(const std::string& name, const DDim& dim) override { - Variable* var = scope_.FindVar(name); + void SetDim(Variable* var, const DDim& dim) { if (var->IsType()) { var->GetMutable()->Resize(dim); } else if (var->IsType()) { var->GetMutable()->set_height(dim[0]); } else { - PADDLE_THROW("Variable %s type_id %s, expect LoDTensor/SelectedRows.", - name, var->Type().name()); + PADDLE_THROW("Variable type_id %s, expect LoDTensor/SelectedRows.", + var->Type().name()); + } + } + + void SetDims(const std::vector& vars, + const std::vector& dims) { + size_t length = vars.size(); + PADDLE_ENFORCE_EQ(length, dims.size()); + for (size_t i = 0; i < length; ++i) { + if (vars[i] == nullptr) { + continue; + } + SetDim(vars[i], dims[i]); } } @@ -669,18 +823,39 @@ class RuntimeInferShapeContext : public InferShapeContext { PADDLE_THROW("Only compile time support this method"); } - proto::VarType::Type GetVarType(const std::string& name) const override { - auto* var = scope_.FindVar(name); - return ToVarType(var->Type()); + std::vector GetVarTypes( + const std::vector& vars) const { + std::vector retv; + retv.resize(vars.size()); + std::transform(vars.begin(), vars.end(), retv.begin(), + std::bind(std::mem_fn(&RuntimeInferShapeContext::GetVarType), + this, std::placeholders::_1)); + return retv; } - InferShapeVarPtr GetVarPtr(const std::string& name) override { - return scope_.FindVar(name); + proto::VarType::Type GetVarType(Variable* var) const { + return ToVarType(var->Type()); } private: + const std::vector& InputVars(const std::string& name) const { + auto it = ctx_.inputs.find(name); + PADDLE_ENFORCE(it != ctx_.inputs.end(), + "Operator %s does not have the input %s.", op_.Type(), name); + return it->second; + } + + const std::vector& OutputVars(const std::string& name) const { + auto it = ctx_.outputs.find(name); + PADDLE_ENFORCE(it != ctx_.outputs.end(), + "Operator %s does not have the outputs %s.", op_.Type(), + name); + return it->second; + } + const OperatorBase& op_; const Scope& scope_; + const RuntimeContext& ctx_; }; static void CheckTensorNANOrInf(const std::string& name, @@ -688,7 +863,8 @@ static void CheckTensorNANOrInf(const std::string& name, if (tensor.memory_size() == 0) { return; } - if (!IsType(tensor.type()) && !IsType(tensor.type())) { + if (tensor.type() != proto::VarType::FP32 && + tensor.type() != proto::VarType::FP64) { return; } PADDLE_ENFORCE(!framework::TensorContainsInf(tensor), @@ -697,10 +873,16 @@ static void CheckTensorNANOrInf(const std::string& name, "Tensor %s contains NAN", name); } +void OperatorWithKernel::RuntimeInferShape(const Scope& scope, + const platform::Place& place, + const RuntimeContext& ctx) const { + RuntimeInferShapeContext infer_shape_ctx(*this, scope, ctx); + this->InferShape(&infer_shape_ctx); +} + void OperatorWithKernel::RunImpl(const Scope& scope, const platform::Place& place) const { - RuntimeInferShapeContext infer_shape_ctx(*this, scope); - this->InferShape(&infer_shape_ctx); + RuntimeContext ctx(Inputs(), Outputs(), scope); platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); auto* dev_ctx = pool.Get(place); @@ -714,23 +896,16 @@ void OperatorWithKernel::RunImpl(const Scope& scope, OpKernelMap& kernels = kernels_iter->second; - // TODO(dzhwinter) : kernel fallback mechanism will be added when all the - // transform functions are ready. - - // for (auto& candidate : kKernelPriority) { - // Do selection - // } - - auto expected_kernel_key = - this->GetExpectedKernelType(ExecutionContext(*this, scope, *dev_ctx)); - VLOG(30) << "expected_kernel_key:" << expected_kernel_key; + auto expected_kernel_key = this->GetExpectedKernelType( + ExecutionContext(*this, scope, *dev_ctx, ctx)); + VLOG(3) << "expected_kernel_key:" << expected_kernel_key; auto kernel_iter = kernels.find(expected_kernel_key); #ifdef PADDLE_WITH_MKLDNN // workaround for missing MKLDNN kernel when FLAGS_use_mkldnn env var is set if (kernel_iter == kernels.end() && expected_kernel_key.library_type_ == LibraryType::kMKLDNN) { - VLOG(30) << "missing MKLDNN kernel: fallbacking to PLAIN one"; + VLOG(3) << "missing MKLDNN kernel: fallbacking to PLAIN one"; expected_kernel_key.library_type_ = LibraryType::kPlain; expected_kernel_key.data_layout_ = DataLayout::kAnyLayout; kernel_iter = kernels.find(expected_kernel_key); @@ -744,7 +919,7 @@ void OperatorWithKernel::RunImpl(const Scope& scope, // do data transformScope &transfer_scope; std::vector transfered_inplace_vars; auto* transfer_scope = - TryTransferData(scope, expected_kernel_key, &transfered_inplace_vars); + PrepareData(scope, expected_kernel_key, &transfered_inplace_vars, &ctx); // exec scope is the scope that kernel actually executed on. const Scope& exec_scope = @@ -754,7 +929,11 @@ void OperatorWithKernel::RunImpl(const Scope& scope, dev_ctx = pool.Get(expected_kernel_key.place_); } - kernel_iter->second(ExecutionContext(*this, exec_scope, *dev_ctx)); + RuntimeInferShapeContext infer_shape_ctx(*this, exec_scope, ctx); + this->InferShape(&infer_shape_ctx); + // TODO(panyx0718): ExecutionContext should only depend on RuntimeContext + // not Scope. Imperative mode only pass inputs and get outputs. + kernel_iter->second(ExecutionContext(*this, exec_scope, *dev_ctx, ctx)); if (!transfered_inplace_vars.empty()) { // there is inplace variable has been transfered. @@ -778,12 +957,12 @@ void OperatorWithKernel::RunImpl(const Scope& scope, } } } + void OperatorWithKernel::TransferInplaceVarsBack( const Scope& scope, const std::vector& inplace_vars, const Scope& transfer_scope) const { for (auto& var_name : inplace_vars) { - VLOG(30) << "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* original_tensor = GetMutableLoDTensorOrSelectedRowsValueFromVar(scope.FindVar(var_name)); auto* var = transfer_scope.FindVar(var_name); @@ -794,24 +973,18 @@ void OperatorWithKernel::TransferInplaceVarsBack( } } -Scope* OperatorWithKernel::TryTransferData( +Scope* OperatorWithKernel::PrepareData( const Scope& scope, const OpKernelType& expected_kernel_key, - std::vector* transfered_inplace_vars) const { -// In the inference scenerio, the scopes will be reused across the batches, so -// the `new_scope` here will result in GPU memroy explosion over the running of -// operators. -// We use a thread_local cache to fix that issue, the key in the cache is the -// combination of the `scope` argument, from_kernel_type, target_kernel_type. -// Have a discussion with @Superjomn or the inference developers if some changes -// on this logic for this macro might not tested on the other scenerios. -#ifdef PADDLE_ON_INFERENCE - thread_local std::unordered_map infer_transfer_scope_cache; -#endif - + std::vector* transfered_inplace_vars, + RuntimeContext* ctx) const { Scope* new_scope = nullptr; for (auto& var_name_item : Inputs()) { - for (auto& var_name : var_name_item.second) { - auto* var = scope.FindVar(var_name); + std::vector& input_vars = ctx->inputs[var_name_item.first]; + + for (size_t i = 0; i < var_name_item.second.size(); ++i) { + auto& var_name = var_name_item.second[i]; + auto* var = input_vars[i]; + // Only tensor can be tranfer to another device. if (var == nullptr || !VarIsTensor(*var)) { continue; @@ -835,30 +1008,31 @@ Scope* OperatorWithKernel::TryTransferData( transfered_inplace_vars->emplace_back(var_name); } - VLOG(30) << "Transform Variable " << var_name << " from " - << kernel_type_for_var << " to " << expected_kernel_key; - -#ifdef PADDLE_ON_INFERENCE - size_t infer_cache_key = - CombineHash(OpKernelType::Hash()(kernel_type_for_var), - OpKernelType::Hash()(expected_kernel_key)); - infer_cache_key = - CombineHash(infer_cache_key, std::hash()(&scope)); - - auto it = infer_transfer_scope_cache.find(infer_cache_key); - if (it != infer_transfer_scope_cache.end()) { - new_scope = infer_transfer_scope_cache[infer_cache_key]; - } else { - new_scope = &scope.NewScope(); - infer_transfer_scope_cache[infer_cache_key] = new_scope; + VLOG(3) << "Transform Variable " << var_name << " from " + << kernel_type_for_var << " to " << expected_kernel_key; + + // In the inference scenerio, the scopes will be reused across the + // batches, so the `new_scope` here will result in GPU memroy explosion + // over the running of operators. + // We use a thread_local cache to fix that issue, the key in the cache is + // the combination of the `scope` argument, from_kernel_type, + // target_kernel_type. + // Have a discussion with @Superjomn or the inference developers if some + // changes on this logic for this macro might not tested on the other + // scenerios. + // If this op is not called by an Executor or ParallelExecutor, it should + // called by a NaiveExecutor, the NaiveExecutor will cache the scopes and + // variables, that behavior a lot different. + if (!run_by_executor_) { + new_scope = TryCreateTransferScope(kernel_type_for_var, + expected_kernel_key, &scope); } -#endif - - if (new_scope == nullptr) { + if (!new_scope) { new_scope = &scope.NewScope(); } auto* trans_var = new_scope->Var(var_name); + input_vars[i] = trans_var; Tensor out; TransformData(expected_kernel_key, kernel_type_for_var, *tensor_in, &out); @@ -887,7 +1061,9 @@ proto::VarType::Type OperatorWithKernel::IndicateDataType( t = &(var->Get().value()); } if (t != nullptr) { - int tmp = static_cast(ToDataType(t->type())); + PADDLE_ENFORCE(t->IsInitialized(), "Input %s is not initialized: %s", + ipt_name, DebugString()); + int tmp = static_cast(t->type()); PADDLE_ENFORCE( tmp == data_type || data_type == -1, "DataType of Paddle Op %s must be the same. Get %s(%d) != %s(%d)", diff --git a/paddle/fluid/framework/operator.h b/paddle/fluid/framework/operator.h index ef838332177c0..1fe2daacf1369 100644 --- a/paddle/fluid/framework/operator.h +++ b/paddle/fluid/framework/operator.h @@ -70,8 +70,17 @@ Tensor* GetMutableLoDTensorOrSelectedRowsValueFromVar(Variable* var); class OperatorBase; class ExecutionContext; +class RuntimeContext { + public: + RuntimeContext(const VariableNameMap& innames, + const VariableNameMap& outnames, const Scope& scope); + + VariableValueMap inputs; + VariableValueMap outputs; +}; + /** - * OperatorBase has the basic element that Net will call to do computation. + * OperatorBase has the basic elements that Net will call to do computation. * Only CreateOperator from OpRegistry will new Operator directly. User * should always construct a proto message OpDesc and call * OpRegistry::CreateOp(op_desc) to get an Operator instance. @@ -127,6 +136,11 @@ class OperatorBase { //! Get all outputs variable names virtual std::vector OutputVars(bool has_intermediate) const; + void SetIsCalledByExecutor(bool x) { run_by_executor_ = x; } + virtual void RuntimeInferShape(const Scope& scope, + const platform::Place& place, + const RuntimeContext& ctx) const {} + protected: std::string type_; // NOTE: in case of OpGrad, inputs_ contains: @@ -139,6 +153,8 @@ class OperatorBase { // IG (Inputs Gradients) VariableNameMap outputs_; AttributeMap attrs_; + // Whether this operator executes in an Executor. + bool run_by_executor_{true}; private: void GenerateTemporaryNames(); @@ -150,8 +166,9 @@ class OperatorBase { class ExecutionContext { public: ExecutionContext(const OperatorBase& op, const Scope& scope, - const platform::DeviceContext& device_context) - : op_(op), scope_(scope), device_context_(device_context) {} + const platform::DeviceContext& device_context, + const RuntimeContext& ctx) + : op_(op), scope_(scope), device_context_(device_context), ctx_(ctx) {} const OperatorBase& op() const { return op_; } @@ -174,20 +191,37 @@ class ExecutionContext { return op_.Outputs(name).size(); } - const Variable* InputVar(const std::string& name) const { - auto ipt = op_.Input(name); - return ipt == kEmptyVarName ? nullptr : scope_.FindVar(ipt); + const Variable* InputVar(const std::string& name) const; + + Variable* OutputVar(const std::string& name) const; + + const std::vector MultiInputVar( + const std::string& name) const { + auto it = ctx_.inputs.find(name); + if (it == ctx_.inputs.end()) { + return {}; + } + std::vector res; + res.reserve(it->second.size()); + std::transform(it->second.begin(), it->second.end(), + std::back_inserter(res), + [this](Variable* var) { return var; }); + return res; } - Variable* OutputVar(const std::string& name) const { - auto opt = op_.Output(name); - return opt == kEmptyVarName ? nullptr : scope_.FindVar(opt); + std::vector MultiOutputVar(const std::string& name) const { + auto names = op_.Outputs(name); + auto it = ctx_.outputs.find(name); + if (it == ctx_.outputs.end()) { + return {}; + } + return it->second; } - const std::vector MultiInputVar( + const std::vector LegacyMultiInputVar( const std::string& name) const { auto names = op_.Inputs(name); - std::vector res; + std::vector res; res.reserve(names.size()); std::transform(names.begin(), names.end(), std::back_inserter(res), [this](const std::string& name) { @@ -197,7 +231,7 @@ class ExecutionContext { return res; } - std::vector MultiOutputVar(const std::string& name) const { + std::vector LegacyMultiOutputVar(const std::string& name) const { auto names = op_.Outputs(name); std::vector res; res.reserve(names.size()); @@ -221,8 +255,56 @@ class ExecutionContext { return var == nullptr ? nullptr : var->GetMutable(); } + template + const T* LegacyInput(const std::string& name) const { + auto* var = LegacyInputVar(name); + return var == nullptr ? nullptr : &var->Get(); + } + + template + T* LegacyOutput(const std::string& name) const { + auto var = LegacyOutputVar(name); + return var == nullptr ? nullptr : var->GetMutable(); + } + + const Variable* LegacyInputVar(const std::string& name) const; + + Variable* LegacyOutputVar(const std::string& name) const; + template const std::vector MultiInput(const std::string& name) const { + auto it = ctx_.inputs.find(name); + if (it == ctx_.inputs.end()) { + return {}; + } + const std::vector& vars = it->second; + std::vector res; + res.reserve(vars.size()); + std::transform(vars.begin(), vars.end(), std::back_inserter(res), + [&](Variable* var) -> const T* { + return var == nullptr ? nullptr : &var->Get(); + }); + return res; + } + + template + std::vector MultiOutput(const std::string& name) const { + auto it = ctx_.outputs.find(name); + if (it == ctx_.outputs.end()) { + return {}; + } + const std::vector& vars = it->second; + std::vector res; + res.reserve(vars.size()); + std::transform(vars.begin(), vars.end(), std::back_inserter(res), + [&](Variable* var) -> T* { + return var == nullptr ? nullptr : var->GetMutable(); + }); + return res; + } + + template + const std::vector LegacyMultiInput(const std::string& name) const { auto names = op_.Inputs(name); std::vector res; res.reserve(names.size()); @@ -235,7 +317,7 @@ class ExecutionContext { } template - std::vector MultiOutput(const std::string& name) const { + std::vector LegacyMultiOutput(const std::string& name) const { auto names = op_.Outputs(name); std::vector res; res.reserve(names.size()); @@ -280,18 +362,30 @@ class ExecutionContext { const OperatorBase& op_; const Scope& scope_; const platform::DeviceContext& device_context_; + const RuntimeContext& ctx_; }; template <> const Tensor* ExecutionContext::Input(const std::string& name) const; +template <> +const Tensor* ExecutionContext::LegacyInput( + const std::string& name) const; + template <> const std::vector ExecutionContext::MultiInput( const std::string& name) const; +template <> +const std::vector ExecutionContext::LegacyMultiInput( + const std::string& name) const; + template <> Tensor* ExecutionContext::Output(const std::string& name) const; +template <> +Tensor* ExecutionContext::LegacyOutput(const std::string& name) const; + template <> std::vector ExecutionContext::MultiOutput( const std::string& name) const; @@ -344,6 +438,9 @@ class OperatorWithKernel : public OperatorBase { OpInfoMap::Instance().Get(Type()).infer_shape_(ctx); } + void RuntimeInferShape(const Scope& scope, const platform::Place& place, + const RuntimeContext& ctx) const override; + protected: virtual OpKernelType GetExpectedKernelType(const ExecutionContext& ctx) const; virtual OpKernelType GetKernelTypeForVar( @@ -362,9 +459,10 @@ class OperatorWithKernel : public OperatorBase { * * * transfered_inplace_vars is a output vector. */ - Scope* TryTransferData( - const Scope& scope, const OpKernelType& expected_kernel_key, - std::vector* transfered_inplace_vars) const; + Scope* PrepareData(const Scope& scope, + const OpKernelType& expected_kernel_key, + std::vector* transfered_inplace_vars, + RuntimeContext* ctx) const; void TransferInplaceVarsBack(const Scope& scope, const std::vector& inplace_vars, diff --git a/paddle/fluid/framework/operator_test.cc b/paddle/fluid/framework/operator_test.cc index ac9dd8245ad4e..ab14732e4d6ea 100644 --- a/paddle/fluid/framework/operator_test.cc +++ b/paddle/fluid/framework/operator_test.cc @@ -50,6 +50,8 @@ class OpWithoutKernelCheckerMaker : public OpProtoAndCheckerMaker { AddInput("input", "input of test op"); AddOutput("output", "output of test op"); AddAttr("scale", "scale of cosine op"); + AddAttr("kernel_sub_type", "kernels with different implementations.") + .SetDefault(0); AddComment("This is test op"); } }; @@ -95,6 +97,8 @@ TEST(OperatorBase, all) { namespace paddle { namespace framework { +static int special_type_value = 1; + class OpKernelTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker { public: void Make() { @@ -103,11 +107,14 @@ class OpKernelTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker { AddAttr("scale", "scale of cosine op") .SetDefault(1.0) .GreaterThan(0.0); + AddAttr("kernel_sub_type", "kernels with different implementations.") + .SetDefault(0); AddComment("This is test op"); } }; static int cpu_kernel_run_num = 0; +static int cpu_kernel2_run_num = 0; class OpWithKernelTest : public OperatorWithKernel { public: @@ -117,7 +124,10 @@ class OpWithKernelTest : public OperatorWithKernel { void InferShape(framework::InferShapeContext* ctx) const override {} OpKernelType GetExpectedKernelType( const ExecutionContext& ctx) const override { - return OpKernelType(proto::VarType::FP32, ctx.GetPlace()); + int sub_type = ctx.Attr("kernel_sub_type"); + return OpKernelType(proto::VarType::FP32, ctx.GetPlace(), + framework::DataLayout::kAnyLayout, + framework::LibraryType::kPlain, sub_type); } }; @@ -132,6 +142,17 @@ class CPUKernelTest : public OpKernel { } }; +template +class CPUKernel2Test : public OpKernel { + public: + void Compute(const ExecutionContext& ctx) const { + std::cout << ctx.op().DebugString() << std::endl; + cpu_kernel2_run_num++; + ASSERT_EQ(ctx.op().Input("x"), "IN1"); + ASSERT_EQ(ctx.op().Output("y"), "OUT1"); + } +}; + class OpKernelTestMultiInputsProtoAndCheckerMaker : public OpProtoAndCheckerMaker { public: @@ -142,6 +163,8 @@ class OpKernelTestMultiInputsProtoAndCheckerMaker AddAttr("scale", "scale of cosine op") .SetDefault(1.0) .GreaterThan(0.0); + AddAttr("kernel_sub_type", "kernels with different implementations.") + .SetDefault(0); AddComment("This is test op"); } }; @@ -189,9 +212,15 @@ class CPUKernalMultiInputsTest : public OpKernel { REGISTER_OP_WITHOUT_GRADIENT( op_with_kernel, paddle::framework::OpWithKernelTest, paddle::framework::OpKernelTestProtoAndCheckerMaker); + REGISTER_OP_CPU_KERNEL(op_with_kernel, paddle::framework::CPUKernelTest); +REGISTER_OP_KERNEL_WITH_CUSTOM_TYPE( + op_with_kernel, CPU, paddle::platform::CPUPlace, MY_SPECIAL_NAME, + paddle::framework::special_type_value, + paddle::framework::CPUKernel2Test); + // test with single input TEST(OpKernel, all) { paddle::framework::InitDevices(true); @@ -211,7 +240,19 @@ TEST(OpKernel, all) { auto op = paddle::framework::OpRegistry::CreateOp(op_desc); ASSERT_EQ(paddle::framework::cpu_kernel_run_num, 0); op->Run(scope, cpu_place); + // kerne_sub_type = 0, hence cpu_kernel is called, cpu_kernel2 is not called. + ASSERT_EQ(paddle::framework::cpu_kernel_run_num, 1); + ASSERT_EQ(paddle::framework::cpu_kernel2_run_num, 0); + + attr = op_desc.mutable_attrs()->Add(); + attr->set_name("kernel_sub_type"); + attr->set_type(paddle::framework::proto::AttrType::INT); + attr->set_i(1); + auto op2 = paddle::framework::OpRegistry::CreateOp(op_desc); + op2->Run(scope, cpu_place); + // kerne_sub_type = 1, hence cpu_kernel2 is called, cpu_kernel is not called. ASSERT_EQ(paddle::framework::cpu_kernel_run_num, 1); + ASSERT_EQ(paddle::framework::cpu_kernel2_run_num, 1); } REGISTER_OP_WITHOUT_GRADIENT( diff --git a/paddle/fluid/framework/parallel_executor.cc b/paddle/fluid/framework/parallel_executor.cc index 2c6e337568306..a921f469f5e02 100644 --- a/paddle/fluid/framework/parallel_executor.cc +++ b/paddle/fluid/framework/parallel_executor.cc @@ -13,6 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/framework/parallel_executor.h" +#include #include #include #include @@ -20,23 +21,47 @@ limitations under the License. */ #include "paddle/fluid/framework/ir/graph.h" -#ifdef PADDLE_WITH_CUDA +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) #include "paddle/fluid/platform/nccl_helper.h" #endif #include "paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h" #include "paddle/fluid/framework/details/multi_devices_helper.h" +#include "paddle/fluid/framework/details/reference_count_pass_helper.h" #include "paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.h" #include "paddle/fluid/framework/details/threaded_ssa_graph_executor.h" #include "paddle/fluid/platform/profiler.h" +#ifdef WITH_GPERFTOOLS +#include "gperftools/profiler.h" +#endif +DEFINE_string(pe_profile_fname, "", + "Profiler filename for PE, which generated by gperftools." + "Only valid when compiled `WITH_PRIFILER=ON`. Empty if disable."); + namespace paddle { namespace framework { +static std::once_flag gProfileOnce; +#ifdef WITH_GPERFTOOLS +static bool gProfileStarted = false; +#endif class ParallelExecutorPrivate { public: explicit ParallelExecutorPrivate(const std::vector &places) - : places_(places) {} + : places_(places) { + if (!FLAGS_pe_profile_fname.empty()) { + std::call_once(gProfileOnce, [] { +#ifdef WITH_GPERFTOOLS + ProfilerStart(FLAGS_pe_profile_fname.c_str()); + gProfileStarted = true; +#else + LOG(WARNING) << "Paddle is not compiled with gperftools. " + "FLAGS_pe_profile_fname will be ignored"; +#endif + }); + } + } ~ParallelExecutorPrivate() { if (own_local_scope_) { @@ -49,26 +74,122 @@ class ParallelExecutorPrivate { } } } + + std::unique_ptr PrepareGCAndRefCnts( + std::unique_ptr graph, size_t max_memory_size); + + inline bool HasGarbageCollectors() const { return !gcs_.empty(); } + + void ResetRuntimeReferenceCount(const std::vector &fetch_tensors, + const std::string &fetched_var_name) { + for (size_t i = 0; i < runtime_ref_cnts_.size(); ++i) { + for (auto &pair : global_ref_cnts_[i]) { + runtime_ref_cnts_[i][pair.first] = pair.second; + } + + for (auto &fetch_name : fetch_tensors) { + runtime_ref_cnts_[i].erase(fetch_name); + } + runtime_ref_cnts_[i].erase(fetched_var_name); + } + } + + BuildStrategy build_strategy_; std::vector places_; std::vector local_scopes_; Scope *global_scope_; // not owned std::unique_ptr executor_; -#ifdef PADDLE_WITH_CUDA +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) std::unique_ptr nccl_ctxs_; #endif bool own_local_scope_; bool use_cuda_; bool use_all_reduce_; + + // global_ref_cnts_ is only initialized when ParallelExecutor constructs, and + // then keeps unchanged + // Before each iteration, runtime_ref_cnts_ is reset to global_ref_cnts_ + std::vector global_ref_cnts_; + std::vector runtime_ref_cnts_; + details::GarbageCollectorMap gcs_; }; +std::unique_ptr ParallelExecutorPrivate::PrepareGCAndRefCnts( + std::unique_ptr graph, size_t max_memory_size) { + for (size_t i = 0; i < places_.size(); ++i) { + auto &place = places_[i]; + if (gcs_.count(place) > 0) { + continue; + } + std::unique_ptr gc; +#ifdef PADDLE_WITH_CUDA + if (platform::is_gpu_place(place)) { + if (IsFastEagerDeletionModeEnabled()) { + gc.reset(new UnsafeFastGPUGarbageCollector( + boost::get(place), max_memory_size)); + } else { + gc.reset(new StreamGarbageCollector( + boost::get(place), max_memory_size)); + } + VLOG(10) << "Created " << i << "-th GarbageCollector at " << place; + } else { +#endif + if (platform::is_cpu_place(place)) { + gc.reset(new CPUGarbageCollector(boost::get(place), + max_memory_size)); + VLOG(10) << "Created GarbageCollector at " << place; + } else { + PADDLE_THROW("Unsupported place for garbage collection"); + } +#ifdef PADDLE_WITH_CUDA + } +#endif + + gcs_.emplace(place, std::move(gc)); + } + + if (!gcs_.empty()) { + std::vector last_live_ops_of_vars; + + auto ref_cnt_pass = + ir::PassRegistry::Instance().Get("reference_count_pass"); + ref_cnt_pass->SetNotOwned(details::kGlobalReferenceCount, + &global_ref_cnts_); + ref_cnt_pass->SetNotOwned(details::kLastLiveOpsOfVars, + &last_live_ops_of_vars); + graph = ref_cnt_pass->Apply(std::move(graph)); + VLOG(10) << "ReferenceCountPass Applied"; + + auto eager_deletion_pass = + ir::PassRegistry::Instance().Get("eager_deletion_pass"); + eager_deletion_pass->SetNotOwned(details::kRuntimeReferenceCount, + &runtime_ref_cnts_); + eager_deletion_pass->SetNotOwned(details::kGarbageCollector, &gcs_); + eager_deletion_pass->SetNotOwned(details::kLastLiveOpsOfVars, + &last_live_ops_of_vars); + eager_deletion_pass->SetNotOwned(details::kAllPlaces, &places_); + graph = eager_deletion_pass->Apply(std::move(graph)); + 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; +} + std::vector &ParallelExecutor::GetLocalScopes() { return member_->local_scopes_; } ParallelExecutor::ParallelExecutor( const std::vector &places, - const std::unordered_set ¶ms, const std::unordered_set &bcast_vars, const ProgramDesc &main_program, const std::string &loss_var_name, Scope *scope, const std::vector &local_scopes, @@ -77,6 +198,7 @@ ParallelExecutor::ParallelExecutor( : member_(new ParallelExecutorPrivate(places)) { member_->global_scope_ = scope; member_->use_cuda_ = exec_strategy.use_cuda_; + member_->build_strategy_ = build_strategy; member_->use_all_reduce_ = build_strategy.reduce_ == BuildStrategy::ReduceStrategy::kAllReduce; @@ -86,7 +208,7 @@ ParallelExecutor::ParallelExecutor( "the number of places must be greater than 1."); } - // Step 1. Bcast the params to devs. + // Step 1. Bcast the bcast_vars to devs. // Create local scopes if (local_scopes.empty()) { member_->own_local_scope_ = true; @@ -104,7 +226,7 @@ ParallelExecutor::ParallelExecutor( if (member_->use_cuda_) { // Bcast Parameters to all GPUs -#ifdef PADDLE_WITH_CUDA +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) auto *nccl_id_var = scope->FindVar(NCCL_ID_VARNAME); ncclUniqueId *nccl_id = nullptr; if (nccl_id_var != nullptr) { @@ -124,39 +246,20 @@ ParallelExecutor::ParallelExecutor( // Step 2. Convert main_program to SSA form and dependency graph. Also, insert // ncclOp -#ifdef PADDLE_WITH_CUDA +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) std::unique_ptr graph = build_strategy.Apply( - main_program, member_->places_, loss_var_name, params, - member_->local_scopes_, member_->use_cuda_, member_->nccl_ctxs_.get()); - - auto max_memory_size = GetEagerDeletionThreshold(); - if (max_memory_size >= 0) { - for (auto &place : member_->places_) { - if (!platform::is_gpu_place(place)) continue; - auto gpu_place = boost::get(place); - if (gcs_[gpu_place.device] == nullptr) { - ref_cnts_[gpu_place.device].reset(new details::ReferenceCountMap()); - cur_ref_cnts_[gpu_place.device].reset( - new details::AtomicReferenceCountMap()); - gcs_[gpu_place.device].reset( - new StreamGarbageCollector(gpu_place, max_memory_size)); - } - } - if (!gcs_.empty()) { - auto ref_cnt_pass = - ir::PassRegistry::Instance().Get("reference_count_pass"); - ref_cnt_pass->SetNotOwned(details::kGlobalReferenceCount, &ref_cnts_); - ref_cnt_pass->SetNotOwned(details::kCurReferenceCount, &cur_ref_cnts_); - ref_cnt_pass->SetNotOwned(details::kGarbageCollector, &gcs_); - graph = ref_cnt_pass->Apply(std::move(graph)); - graph->SetNotOwned("garbage_collector", &gcs_); - } - } + main_program, member_->places_, loss_var_name, member_->local_scopes_, + member_->use_cuda_, member_->nccl_ctxs_.get()); #else std::unique_ptr graph = build_strategy.Apply(main_program, member_->places_, loss_var_name, - params, member_->local_scopes_, member_->use_cuda_); + member_->local_scopes_, member_->use_cuda_); #endif + auto max_memory_size = GetEagerDeletionThreshold(); + if (max_memory_size >= 0) { + graph = member_->PrepareGCAndRefCnts(std::move(graph), + static_cast(max_memory_size)); + } // Step 3. Create vars in each scope. Passes may also create new vars. // skip control vars and empty vars @@ -186,10 +289,12 @@ ParallelExecutor::ParallelExecutor( if (exec_strategy.type_ == ExecutionStrategy::kDefault) { member_->executor_.reset(new details::ThreadedSSAGraphExecutor( - exec_strategy, member_->local_scopes_, places, std::move(graph))); + exec_strategy, member_->local_scopes_, member_->places_, + std::move(graph))); } else { member_->executor_.reset(new details::FastThreadedSSAGraphExecutor( - exec_strategy, member_->local_scopes_, places, std::move(graph))); + exec_strategy, member_->local_scopes_, member_->places_, + std::move(graph))); } member_->executor_.reset(new details::ScopeBufferedSSAGraphExecutor( @@ -208,12 +313,12 @@ void ParallelExecutor::BCastParamsToDevices( auto &main_tensor = main_var->Get(); if (!main_tensor.IsInitialized()) { - VLOG(30) << "one in var not inited, return!"; + VLOG(3) << "one in var not inited, return!"; continue; } auto &dims = main_tensor.dims(); if (paddle::platform::is_gpu_place(main_tensor.place())) { -#ifdef PADDLE_WITH_CUDA +#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) std::vector buffers; size_t numel = main_tensor.numel(); ncclDataType_t data_type = platform::ToNCCLDataType(main_tensor.type()); @@ -270,19 +375,16 @@ void ParallelExecutor::BCastParamsToDevices( void ParallelExecutor::Run(const std::vector &fetch_tensors, const std::string &fetched_var_name) { - platform::RecordBlock b(0); -#ifdef PADDLE_WITH_CUDA - if (!gcs_.empty()) { - ResetReferenceCount(); - for (auto &pair : cur_ref_cnts_) { - auto &name_map = *(pair.second); - for (auto &fetch_name : fetch_tensors) { - name_map.erase(fetch_name); - } - name_map.erase(fetched_var_name); - } +#ifdef WITH_GPERFTOOLS + if (gProfileStarted) { + ProfilerFlush(); } #endif + + platform::RecordBlock b(0); + if (member_->HasGarbageCollectors()) { + member_->ResetRuntimeReferenceCount(fetch_tensors, fetched_var_name); + } auto fetch_data = member_->executor_->Run(fetch_tensors); *member_->global_scope_->Var(fetched_var_name)->GetMutable() = fetch_data; @@ -326,13 +428,12 @@ ParallelExecutor::~ParallelExecutor() { for (auto &p : member_->places_) { platform::DeviceContextPool::Instance().Get(p)->Wait(); } - // member_ must be destructed before gcs_ since the destructor of - // ReferenceCountOpHandle use raw pointers of gcs_ inside. - member_.reset(); + delete member_; } } // namespace framework } // namespace paddle -#ifdef PADDLE_WITH_CUDA + +USE_PASS(memory_early_delete_pass); USE_PASS(reference_count_pass); -#endif +USE_PASS(eager_deletion_pass); diff --git a/paddle/fluid/framework/parallel_executor.h b/paddle/fluid/framework/parallel_executor.h index ef09b98b2aa91..5f6c2159aa2d9 100644 --- a/paddle/fluid/framework/parallel_executor.h +++ b/paddle/fluid/framework/parallel_executor.h @@ -14,7 +14,6 @@ limitations under the License. */ #pragma once -#include #include #include #include @@ -29,10 +28,6 @@ limitations under the License. */ #include "paddle/fluid/framework/tensor.h" #include "paddle/fluid/platform/device_context.h" -#ifdef PADDLE_WITH_CUDA -#include "paddle/fluid/framework/details/reference_count_pass.h" -#endif - namespace paddle { namespace framework { @@ -46,7 +41,6 @@ class ParallelExecutor { public: explicit ParallelExecutor(const std::vector &places, - const std::unordered_set ¶ms, const std::unordered_set &bcast_vars, const ProgramDesc &main_program, const std::string &loss_var_name, Scope *scope, @@ -75,24 +69,7 @@ class ParallelExecutor { private: void BCastParamsToDevices(const std::unordered_set &vars) const; - std::unique_ptr member_; - -#ifdef PADDLE_WITH_CUDA - // ref_cnts_ is only initialized when ParallelExecutor constructs, and then - // keeps unchanged - // Before each iteration, cur_ref_cnts_ is reset to ref_cnts_ - details::DeviceReferenceCountMap ref_cnts_; - details::AtomicDeviceReferenceCountMap cur_ref_cnts_; - details::DeviceGarbageCollectorMap gcs_; - - void ResetReferenceCount() { - for (auto &pair1 : ref_cnts_) { - for (auto &pair2 : *(pair1.second)) { - (*(cur_ref_cnts_[pair1.first]))[pair2.first] = pair2.second; - } - } - } -#endif + ParallelExecutorPrivate *member_; }; } // namespace framework diff --git a/paddle/fluid/framework/scope.cc b/paddle/fluid/framework/scope.cc index 26cb7d51a88af..6fa5e99f9f3a7 100644 --- a/paddle/fluid/framework/scope.cc +++ b/paddle/fluid/framework/scope.cc @@ -38,6 +38,10 @@ DEFINE_double( "Memory size threshold (GB) when the garbage collector clear tensors." "Disabled when this value is less than 0"); +DEFINE_bool(fast_eager_deletion_mode, false, + "Fast eager deletion mode. If enabled, memory would release " + "immediately without waiting GPU kernel ends."); + // When in inference scenario, the scopes will not be written by two threads in // a mean time, but a scope may be read by multiple threads concurrently, and // the mutex will cause serious performance issue. @@ -58,6 +62,8 @@ int64_t GetEagerDeletionThreshold() { (static_cast(1) << 30)); } +bool IsFastEagerDeletionModeEnabled() { return FLAGS_fast_eager_deletion_mode; } + Scope::~Scope() { DropKids(); } Scope& Scope::NewScope() const { @@ -162,7 +168,7 @@ Variable* Scope::VarInternal(const std::string& name) { v = new Variable(); vars_[name].reset(v); - VLOG(30) << "Create variable " << name; + VLOG(3) << "Create variable " << name; v->name_ = &(vars_.find(name)->first); return v; } diff --git a/paddle/fluid/framework/scope.h b/paddle/fluid/framework/scope.h index 1901ffbe57e0d..aded1f771cedb 100644 --- a/paddle/fluid/framework/scope.h +++ b/paddle/fluid/framework/scope.h @@ -27,6 +27,7 @@ namespace paddle { namespace framework { int64_t GetEagerDeletionThreshold(); +bool IsFastEagerDeletionModeEnabled(); class Scope; diff --git a/paddle/fluid/framework/selected_rows.cc b/paddle/fluid/framework/selected_rows.cc index f4f2b769d5e47..54a818250b45e 100644 --- a/paddle/fluid/framework/selected_rows.cc +++ b/paddle/fluid/framework/selected_rows.cc @@ -206,7 +206,7 @@ void SelectedRows::Get(const framework::Tensor& ids, framework::Tensor* value, PADDLE_ENFORCE(value->IsInitialized(), "The value tensor should be initialized."); if (ids.numel() == 0) { - VLOG(30) << "keys is empty, please check data!"; + VLOG(3) << "keys is empty, please check data!"; } else { int64_t value_width = value_->numel() / value_->dims()[0]; PADDLE_ENFORCE_EQ(value_width, value->numel() / value->dims()[0], @@ -218,11 +218,11 @@ void SelectedRows::Get(const framework::Tensor& ids, framework::Tensor* value, if (index < 0) { VLOG(5) << "id " << id << " not in the table, return 0"; framework::VisitDataType( - framework::ToDataType(value_->type()), + value_->type(), TensorFillVisitor(value, i * value_width, value_width, 0.0)); } else { framework::VisitDataType( - framework::ToDataType(value_->type()), + value_->type(), TensorCopyVisitor(value, i * value_width, *value_.get(), index * value_width, value_width)); } diff --git a/paddle/fluid/framework/selected_rows.h b/paddle/fluid/framework/selected_rows.h index 55ca02038e083..e1bdba9b46a4c 100644 --- a/paddle/fluid/framework/selected_rows.h +++ b/paddle/fluid/framework/selected_rows.h @@ -32,8 +32,7 @@ namespace framework { class SelectedRows { /* * @brief We can use the SelectedRows structure to reproduce a sparse table. - * A sparse table is a key-value structure that the key is an `int64_t` - * number, + * A sparse table is a key-value structure that the key is an `int64_t`, * and the value is a Tensor which the first dimension is 0. * You can use the following interface to operate the sparse table, and you * can find @@ -120,8 +119,22 @@ class SelectedRows { */ int64_t AutoGrownIndex(int64_t key, bool auto_grown, bool is_test = false); - void SyncIndex(); + /* + * @brief Get the index of the key from id_to_index_ map. + */ + inline int64_t GetIndexFromId(int64_t key) { + auto iter = id_to_index_.find(key); + if (iter == id_to_index_.end()) { + return -1; + } else { + return iter->second; + } + } + void SyncIndex(); + /* + * @brief Get complete Dims before + */ DDim GetCompleteDims() const { std::vector dims = vectorize(value_->dims()); dims[0] = height_; @@ -133,9 +146,10 @@ class SelectedRows { // SelectedRows are simply concated when adding together. Until a // SelectedRows add a Tensor, will the duplicate rows be handled. Vector rows_; - std::unordered_map id_to_index_; + std::unordered_map + id_to_index_; // should not be used when rows_ has duplicate member std::unique_ptr value_{nullptr}; - int64_t height_; + int64_t height_; // height indicates the underline tensor's height std::unique_ptr rwlock_{nullptr}; }; diff --git a/paddle/fluid/framework/shape_inference.cc b/paddle/fluid/framework/shape_inference.cc index ddff2c7c26174..4ac872ac3d3bf 100644 --- a/paddle/fluid/framework/shape_inference.cc +++ b/paddle/fluid/framework/shape_inference.cc @@ -22,20 +22,6 @@ limitations under the License. */ namespace paddle { namespace framework { -DDim InferShapeContext::GetInputDim(const std::string &name) const { - const std::vector &arg_names = Inputs(name); - PADDLE_ENFORCE_EQ(arg_names.size(), 1UL, - "Input(%s) should hold one element, but now it holds %d", - name, arg_names.size()); - return this->GetDim(arg_names[0]); -} - -std::vector InferShapeContext::GetInputsDim( - const std::string &name) const { - const std::vector &arg_names = Inputs(name); - return GetDims(arg_names); -} - std::vector InferShapeContext::GetReaderDims( const std::string &name) const { const std::vector &arg_names = Inputs(name); @@ -46,26 +32,6 @@ std::vector InferShapeContext::GetReaderDims( return this->GetRepeatedDims(arg_names[0]); } -DDim InferShapeContext::GetInputsElementDim(const std::string &name, - int idx) const { - const std::vector &names = Inputs(name); - return this->GetDim(names[idx]); -} - -void InferShapeContext::SetOutputDim(const std::string &name, const DDim &dim) { - auto &arg_names = Outputs(name); - PADDLE_ENFORCE_EQ(arg_names.size(), 1UL, - "Output(%s) should hold one element, but now it holds %d", - name, arg_names.size()); - SetDim(arg_names[0], dim); -} - -void InferShapeContext::SetOutputsDim(const std::string &name, - const std::vector &dims) { - auto &names = Outputs(name); - SetDims(names, dims); -} - void InferShapeContext::SetReaderDims(const std::string &name, const std::vector &dims) { const std::vector &arg_names = Outputs(name); @@ -76,69 +42,5 @@ void InferShapeContext::SetReaderDims(const std::string &name, return this->SetRepeatedDims(arg_names[0], dims); } -std::vector InferShapeContext::GetInputVarPtrs( - const std::string &name) { - const std::vector arg_names = Inputs(name); - std::vector res; - res.reserve(arg_names.size()); - std::transform( - arg_names.begin(), arg_names.end(), std::back_inserter(res), - [this](const std::string &name) { return this->GetVarPtr(name); }); - return res; -} - -std::vector InferShapeContext::GetOutputVarPtrs( - const std::string &name) { - const std::vector arg_names = Outputs(name); - std::vector res; - res.reserve(arg_names.size()); - std::transform( - arg_names.begin(), arg_names.end(), std::back_inserter(res), - [this](const std::string &name) { return this->GetVarPtr(name); }); - return res; -} - -std::vector InferShapeContext::GetDims( - const std::vector &names) const { - std::vector ret; - ret.reserve(names.size()); - std::transform( - names.begin(), names.end(), std::back_inserter(ret), - [this](const std::string &name) { return this->GetDim(name); }); - return ret; -} - -void InferShapeContext::SetDims(const std::vector &names, - const std::vector &dims) { - size_t length = names.size(); - PADDLE_ENFORCE_EQ(length, dims.size()); - for (size_t i = 0; i < length; ++i) { - if (names[i] == framework::kEmptyVarName) { - continue; - } - SetDim(names[i], dims[i]); - } -} - -std::vector InferShapeContext::GetInputsVarType( - const std::string &name) const { - return GetVarTypes(Inputs(name)); -} - -std::vector InferShapeContext::GetOutputsVarType( - const std::string &name) const { - return GetVarTypes(Outputs(name)); -} - -std::vector InferShapeContext::GetVarTypes( - const std::vector &names) const { - std::vector retv; - retv.resize(names.size()); - std::transform(names.begin(), names.end(), retv.begin(), - std::bind(std::mem_fn(&InferShapeContext::GetVarType), this, - std::placeholders::_1)); - return retv; -} - } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/shape_inference.h b/paddle/fluid/framework/shape_inference.h index 280bc19dce7b6..e0a848273b8d6 100644 --- a/paddle/fluid/framework/shape_inference.h +++ b/paddle/fluid/framework/shape_inference.h @@ -25,6 +25,8 @@ limitations under the License. */ namespace paddle { namespace framework { +class OperatorBase; + using InferShapeVarPtr = boost::variant; class InferShapeContext { @@ -33,22 +35,23 @@ class InferShapeContext { virtual bool HasInput(const std::string &name) const = 0; virtual bool HasOutput(const std::string &name) const = 0; - std::vector GetInputsVarType( - const std::string &name) const; - std::vector GetOutputsVarType( - const std::string &name) const; + virtual std::vector GetInputsVarType( + const std::string &name) const = 0; + virtual std::vector GetOutputsVarType( + const std::string &name) const = 0; virtual bool HasInputs(const std::string &name) const = 0; virtual bool HasOutputs(const std::string &name) const = 0; - DDim GetInputDim(const std::string &name) const; - std::vector GetInputsDim(const std::string &name) const; - std::vector GetReaderDims(const std::string &name) const; - DDim GetInputsElementDim(const std::string &name, int idx) const; + virtual DDim GetInputDim(const std::string &name) const = 0; + virtual std::vector GetInputsDim(const std::string &name) const = 0; + virtual std::vector GetReaderDims(const std::string &name) const; - void SetOutputDim(const std::string &name, const DDim &dim); - void SetOutputsDim(const std::string &name, const std::vector &dims); - void SetReaderDims(const std::string &name, const std::vector &dims); + virtual void SetOutputDim(const std::string &name, const DDim &dim) = 0; + virtual void SetOutputsDim(const std::string &name, + const std::vector &dims) = 0; + virtual void SetReaderDims(const std::string &name, + const std::vector &dims); virtual AttrReader Attrs() const = 0; virtual const std::vector &Inputs( @@ -62,29 +65,20 @@ class InferShapeContext { virtual void ShareLoD(const std::string &in, const std::string &out, size_t i = 0, size_t j = 0) const = 0; - virtual bool IsRuntime() const = 0; + virtual void DecreaseLoDLevel(const std::string &in, const std::string &out, + size_t i = 0, size_t j = 0) const = 0; - std::vector GetInputVarPtrs(const std::string &name); - std::vector GetOutputVarPtrs(const std::string &name); - virtual InferShapeVarPtr GetVarPtr(const std::string &name) = 0; + virtual bool IsRuntime() const = 0; - // Note: In while op, we need this to be public - void SetDims(const std::vector &names, - const std::vector &dims); + virtual std::vector GetInputVarPtrs( + const std::string &name) = 0; + virtual std::vector GetOutputVarPtrs( + const std::string &name) = 0; protected: - virtual DDim GetDim(const std::string &name) const = 0; - virtual void SetDim(const std::string &name, const DDim &dim) = 0; virtual std::vector GetRepeatedDims(const std::string &name) const = 0; virtual void SetRepeatedDims(const std::string &name, const std::vector &dims) = 0; - - std::vector GetDims(const std::vector &names) const; - - std::vector GetVarTypes( - const std::vector &names) const; - - virtual proto::VarType::Type GetVarType(const std::string &name) const = 0; }; } // namespace framework diff --git a/paddle/fluid/framework/tensor.cc b/paddle/fluid/framework/tensor.cc index 41566800e5781..5b09cad06c3f8 100644 --- a/paddle/fluid/framework/tensor.cc +++ b/paddle/fluid/framework/tensor.cc @@ -13,10 +13,11 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/framework/tensor.h" +#include "paddle/fluid/framework/var_type.h" namespace paddle { namespace framework { -extern size_t SizeOfType(std::type_index type); +extern size_t SizeOfType(proto::VarType::Type type); void Tensor::check_memory_size() const { PADDLE_ENFORCE_NOT_NULL( holder_, "Tensor holds no memory. Call Tensor::mutable_data first."); @@ -27,11 +28,14 @@ void Tensor::check_memory_size() const { "or maybe the required data-type mismatches the data already stored."); } +Tensor::Tensor(std::type_index type) + : type_(framework::ToDataType(type)), offset_(0) {} + size_t Tensor::memory_size() const { return holder_ == nullptr ? 0UL : holder_->size() - offset_; } -void* Tensor::mutable_data(platform::Place place, std::type_index type, +void* Tensor::mutable_data(platform::Place place, proto::VarType::Type type, memory::Allocator::Attr attr, size_t requested_size) { type_ = type; @@ -101,5 +105,12 @@ const DDim& Tensor::dims() const { return dims_; } int64_t Tensor::numel() const { return product(dims_); } +void Tensor::ResetHolder(std::shared_ptr holder) { + if (holder_) { + PADDLE_ENFORCE_EQ(numel() * SizeOfType(type()), holder->size()); + } + holder_ = holder; +} + } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/tensor.h b/paddle/fluid/framework/tensor.h index 71e8badd4b6b0..2e110133a33ed 100644 --- a/paddle/fluid/framework/tensor.h +++ b/paddle/fluid/framework/tensor.h @@ -19,9 +19,9 @@ limitations under the License. */ #include #include #include - #include "paddle/fluid/framework/data_layout.h" #include "paddle/fluid/framework/ddim.h" +#include "paddle/fluid/framework/framework.pb.h" #include "paddle/fluid/memory/memory.h" #include "paddle/fluid/platform/device_context.h" #include "paddle/fluid/platform/enforce.h" @@ -67,7 +67,9 @@ class Tensor { friend struct EigenVector; public: - Tensor() : type_(typeid(float)), offset_(0) {} + Tensor() : type_(proto::VarType::FP32), offset_(0) {} + + explicit Tensor(std::type_index type); /*! Return a pointer to mutable memory block. */ template @@ -88,7 +90,7 @@ class Tensor { memory::Allocator::Attr attr = memory::Allocator::kDefault, size_t requested_size = 0); - void* mutable_data(platform::Place place, std::type_index type, + void* mutable_data(platform::Place place, proto::VarType::Type type, memory::Allocator::Attr attr = memory::Allocator::kDefault, size_t requested_size = 0); @@ -138,7 +140,7 @@ class Tensor { return holder_->place(); } - std::type_index type() const { + proto::VarType::Type type() const { PADDLE_ENFORCE_NOT_NULL( holder_, "Tensor not initialized yet when Tensor::type() is called."); return type_; @@ -158,10 +160,16 @@ class Tensor { const std::shared_ptr& Holder() const { return holder_; } size_t offset() const { return offset_; } + std::shared_ptr MoveMemoryHolder() { + return std::move(holder_); + } + + void ResetHolder(std::shared_ptr holder); + private: /*! holds the memory block if allocated. */ std::shared_ptr holder_; - std::type_index type_; + proto::VarType::Type type_; /** * @brief points to elements dimensions. * diff --git a/paddle/fluid/framework/tensor_impl.h b/paddle/fluid/framework/tensor_impl.h index 0c9c0d782fc73..ce3ad18b1fb1c 100644 --- a/paddle/fluid/framework/tensor_impl.h +++ b/paddle/fluid/framework/tensor_impl.h @@ -24,9 +24,8 @@ template inline const T* Tensor::data() const { check_memory_size(); bool valid = - std::is_same::value || type_ == std::type_index(typeid(T)); - PADDLE_ENFORCE(valid, "Tensor holds the wrong type, it holds %s", - type_.name()); + std::is_same::value || type_ == DataTypeTrait::DataType; + PADDLE_ENFORCE(valid, "Tensor holds the wrong type, it holds %d", type_); return reinterpret_cast( reinterpret_cast(holder_->ptr()) + offset_); @@ -38,9 +37,8 @@ template inline T* Tensor::data() { check_memory_size(); bool valid = - std::is_same::value || type_ == std::type_index(typeid(T)); - PADDLE_ENFORCE(valid, "Tensor holds the wrong type, it holds %s", - type_.name()); + std::is_same::value || type_ == DataTypeTrait::DataType; + PADDLE_ENFORCE(valid, "Tensor holds the wrong type, it holds %s", type_); return reinterpret_cast(reinterpret_cast(holder_->ptr()) + offset_); } @@ -60,7 +58,7 @@ inline T* Tensor::mutable_data(platform::Place place, size_t requested_size) { static_assert(std::is_pod::value, "T must be POD"); return reinterpret_cast( - mutable_data(place, typeid(T), attr, requested_size)); + mutable_data(place, DataTypeTrait::DataType, attr, requested_size)); } inline Tensor ReshapeToMatrix(const Tensor& src, int num_col_dims) { diff --git a/paddle/fluid/framework/tensor_test.cc b/paddle/fluid/framework/tensor_test.cc index a0a9a573603ce..83dea8639010f 100644 --- a/paddle/fluid/framework/tensor_test.cc +++ b/paddle/fluid/framework/tensor_test.cc @@ -74,6 +74,22 @@ TEST(Tensor, MutableData) { p2 = src_tensor.mutable_data(framework::make_ddim({2, 2}), platform::CPUPlace()); EXPECT_EQ(p1, p2); + + float* p3 = nullptr; + float* p4 = nullptr; + // set src_tensor a different type but smaller size. + // memory block is supposed to be unchanged. + auto* tmp = src_tensor.mutable_data(framework::make_ddim({2, 2}), + platform::CPUPlace()); + p3 = reinterpret_cast(tmp); + EXPECT_EQ(p1, p3); + + // set src_tensor a different type but bigger size. + // memory block is supposed to be changed. + auto* tmp2 = src_tensor.mutable_data( + framework::make_ddim({2, 2, 3}), platform::CPUPlace()); + p4 = reinterpret_cast(tmp2); + EXPECT_NE(p1, p4); } // Not sure if it's desired, but currently, Tensor type can be changed. { diff --git a/paddle/fluid/framework/tensor_util.cc b/paddle/fluid/framework/tensor_util.cc index 8d8f07a1f52b3..85d15c5d3faa5 100644 --- a/paddle/fluid/framework/tensor_util.cc +++ b/paddle/fluid/framework/tensor_util.cc @@ -22,8 +22,8 @@ namespace framework { void TensorCopy(const Tensor& src, const platform::Place& dst_place, const platform::DeviceContext& ctx, Tensor* dst) { - VLOG(30) << "TensorCopy " << src.dims() << " from " << src.place() << " to " - << dst_place; + VLOG(3) << "TensorCopy " << src.dims() << " from " << src.place() << " to " + << dst_place; src.check_memory_size(); dst->Resize(src.dims()); @@ -37,8 +37,8 @@ void TensorCopy(const Tensor& src, const platform::Place& dst_place, if (platform::is_cpu_place(src_place) && platform::is_cpu_place(dst_place)) { if (src_ptr == dst_ptr) { - VLOG(30) << "Skip copy the same data async from " << src_place << " to " - << dst_place; + VLOG(3) << "Skip copy the same data async from " << src_place << " to " + << dst_place; return; } memory::Copy(boost::get(dst_place), dst_ptr, @@ -77,8 +77,8 @@ void TensorCopy(const Tensor& src, const platform::Place& dst_place, reinterpret_cast(ctx).stream(); if (platform::is_same_place(src_place, dst_place)) { if (src_ptr == dst_ptr) { - VLOG(30) << "Skip copy the same data async from " << src_place << " to " - << dst_place; + VLOG(3) << "Skip copy the same data async from " << src_place << " to " + << dst_place; return; } memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size, @@ -114,8 +114,8 @@ void TensorCopy(const Tensor& src, const platform::Place& dst_place, void TensorCopySync(const Tensor& src, const platform::Place& dst_place, Tensor* dst) { - VLOG(30) << "TensorCopySync " << src.dims() << " from " << src.place() - << " to " << dst_place; + VLOG(3) << "TensorCopySync " << src.dims() << " from " << src.place() + << " to " << dst_place; src.check_memory_size(); dst->Resize(src.dims()); dst->set_layout(src.layout()); @@ -125,8 +125,8 @@ void TensorCopySync(const Tensor& src, const platform::Place& dst_place, auto size = src.numel() * SizeOfType(src.type()); if (platform::is_cpu_place(src_place) && platform::is_cpu_place(dst_place)) { if (src_ptr == dst_ptr) { - VLOG(30) << "Skip copy the same data from " << src_place << " to " - << dst_place; + VLOG(3) << "Skip copy the same data from " << src_place << " to " + << dst_place; return; } memory::Copy(boost::get(dst_place), dst_ptr, @@ -146,8 +146,8 @@ void TensorCopySync(const Tensor& src, const platform::Place& dst_place, } else if (platform::is_gpu_place(src_place) && platform::is_gpu_place(dst_place)) { if (src_ptr == dst_ptr && platform::is_same_place(src_place, dst_place)) { - VLOG(30) << "Skip copy the same data from " << src_place << " to " - << dst_place; + VLOG(3) << "Skip copy the same data from " << src_place << " to " + << dst_place; return; } auto src_gpu_place = boost::get(src_place); @@ -186,8 +186,8 @@ struct AnyDTypeVisitor { template inline void AnyImpl(Predicate predicate, const framework::Tensor& tensor, const DevCtx& ctx, framework::Tensor* out) { - VisitDataType(ToDataType(tensor.type()), AnyDTypeVisitor( - predicate, tensor, ctx, out)); + VisitDataType(tensor.type(), AnyDTypeVisitor( + predicate, tensor, ctx, out)); } template @@ -379,7 +379,7 @@ void TensorToStream(std::ostream& os, const Tensor& tensor, // int32_t size // void* protobuf message proto::VarType::TensorDesc desc; - desc.set_data_type(framework::ToDataType(tensor.type())); + desc.set_data_type(tensor.type()); auto dims = framework::vectorize(tensor.dims()); auto* pb_dims = desc.mutable_dims(); pb_dims->Resize(static_cast(dims.size()), 0); @@ -461,9 +461,7 @@ void TensorFromStream(std::istream& is, Tensor* tensor, tensor->Resize(framework::make_ddim(dims)); void* buf; auto ctx = platform::CPUDeviceContext(); - size_t size = - tensor->numel() * - framework::SizeOfType(framework::ToTypeIndex(desc.data_type())); + size_t size = tensor->numel() * framework::SizeOfType(desc.data_type()); if (platform::is_gpu_place(dev_ctx.GetPlace())) { #ifdef PADDLE_WITH_CUDA Tensor cpu_tensor; diff --git a/paddle/fluid/framework/threadpool.cc b/paddle/fluid/framework/threadpool.cc index 2dab4e793eeac..fcec955360f1c 100644 --- a/paddle/fluid/framework/threadpool.cc +++ b/paddle/fluid/framework/threadpool.cc @@ -39,7 +39,7 @@ void ThreadPool::Init() { int num_threads = std::thread::hardware_concurrency(); if (FLAGS_dist_threadpool_size > 0) { num_threads = FLAGS_dist_threadpool_size; - VLOG(10) << "set dist_threadpool_size to " << num_threads; + VLOG(1) << "set dist_threadpool_size to " << num_threads; } PADDLE_ENFORCE_GT(num_threads, 0); threadpool_.reset(new ThreadPool(num_threads)); diff --git a/paddle/fluid/framework/transfer_scope_cache.cc b/paddle/fluid/framework/transfer_scope_cache.cc new file mode 100644 index 0000000000000..e52a8317e2113 --- /dev/null +++ b/paddle/fluid/framework/transfer_scope_cache.cc @@ -0,0 +1,72 @@ +// 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/transfer_scope_cache.h" + +namespace paddle { +namespace framework { + +std::unordered_map& global_transfer_data_cache() { + thread_local auto* x = new std::unordered_map; + return *x; +} + +std::unordered_set& global_transfer_scope_cache() { + thread_local auto* x = new std::unordered_set; + return *x; +} + +Scope* TryCreateTransferScope(OpKernelType type0, OpKernelType type1, + const Scope* scope) { + Scope* new_scope{nullptr}; + size_t infer_cache_key = + CombineHash(OpKernelType::Hash()(type0), OpKernelType::Hash()(type1)); + infer_cache_key = + CombineHash(infer_cache_key, std::hash()(scope)); + + auto it = global_transfer_data_cache().find(infer_cache_key); + if (it != global_transfer_data_cache().end()) { + new_scope = global_transfer_data_cache()[infer_cache_key]; + } else { + new_scope = &scope->NewScope(); + global_transfer_data_cache()[infer_cache_key] = new_scope; + } + global_transfer_scope_cache().insert(new_scope); + return new_scope; +} + +void RemoveKidsFromTransferScopeCache(Scope* scope) { + auto it = global_transfer_scope_cache().find(scope); + if (it != global_transfer_scope_cache().end()) { + global_transfer_scope_cache().erase(it); + } + for (auto* s : scope->kids()) { + auto it = global_transfer_scope_cache().find(s); + if (it != global_transfer_scope_cache().end()) { + global_transfer_scope_cache().erase(it); + } + } + + // remove global transfer data cache + auto& cache = global_transfer_data_cache(); + for (auto it = cache.begin(); it != cache.end();) { + if (it->second == scope) + it = cache.erase(it); + else + it++; + } +} + +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/transfer_scope_cache.h b/paddle/fluid/framework/transfer_scope_cache.h new file mode 100644 index 0000000000000..86fc0bf52972a --- /dev/null +++ b/paddle/fluid/framework/transfer_scope_cache.h @@ -0,0 +1,41 @@ +// 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 // NOLINT +#include +#include +#include "paddle/fluid/framework/op_kernel_type.h" +#include "paddle/fluid/framework/scope.h" + +namespace paddle { +namespace framework { + +std::unordered_map& global_transfer_data_cache(); + +std::unordered_set& global_transfer_scope_cache(); + +// Combine two hash values to a single hash. +static size_t CombineHash(size_t seed, size_t a) { + return (seed ^ a) + 0x9e3779b9 + (seed << 6) + (seed >> 2); +} + +Scope* TryCreateTransferScope(OpKernelType type0, OpKernelType type1, + const Scope* scope); + +void RemoveKidsFromTransferScopeCache(Scope* scope); + +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/type_defs.h b/paddle/fluid/framework/type_defs.h index 2de6233a9e0d3..938e2024c3359 100644 --- a/paddle/fluid/framework/type_defs.h +++ b/paddle/fluid/framework/type_defs.h @@ -28,8 +28,11 @@ class OperatorBase; class OpDesc; class InferShapeContext; class BlockDesc; +class Variable; using VariableNameMap = std::map>; +// TODO(panyx0718): Replace vector with something like gtl::Vector. +using VariableValueMap = std::map>; // The order should be as same as framework.proto using Attribute = diff --git a/paddle/fluid/framework/var_desc.cc b/paddle/fluid/framework/var_desc.cc index 29ef459b45407..7e3f002b53351 100644 --- a/paddle/fluid/framework/var_desc.cc +++ b/paddle/fluid/framework/var_desc.cc @@ -61,10 +61,10 @@ size_t VarDesc::GetTensorDescNum() const { void VarDesc::SetShapes( const std::vector> &multiple_dims) { if (multiple_dims.size() != GetTensorDescNum()) { - VLOG(30) << "WARNING: The number of given shapes(" << multiple_dims.size() - << ") doesn't match the existing tensor number(" - << GetTensorDescNum() - << "). The Reader is going to be reinitialized."; + VLOG(3) << "WARNING: The number of given shapes(" << multiple_dims.size() + << ") doesn't match the existing tensor number(" + << GetTensorDescNum() + << "). The Reader is going to be reinitialized."; SetTensorDescNum(multiple_dims.size()); } std::vector tensors = mutable_tensor_descs(); @@ -94,11 +94,11 @@ void VarDesc::SetDataType(proto::VarType::Type data_type) { void VarDesc::SetDataTypes( const std::vector &multiple_data_type) { if (multiple_data_type.size() != GetTensorDescNum()) { - VLOG(30) << "WARNING: The number of given data types(" - << multiple_data_type.size() - << ") doesn't match the existing tensor number(" - << GetTensorDescNum() - << "). The Reader is going to be reinitialized."; + VLOG(3) << "WARNING: The number of given data types(" + << multiple_data_type.size() + << ") doesn't match the existing tensor number(" + << GetTensorDescNum() + << "). The Reader is going to be reinitialized."; SetTensorDescNum(multiple_data_type.size()); } std::vector tensor_descs = @@ -139,11 +139,11 @@ void VarDesc::SetLoDLevel(int32_t lod_level) { void VarDesc::SetLoDLevels(const std::vector &multiple_lod_level) { if (multiple_lod_level.size() != GetTensorDescNum()) { - VLOG(30) << "WARNING: The number of given lod_levels(" - << multiple_lod_level.size() - << ") doesn't match the existing tensor number(" - << GetTensorDescNum() - << "). The Reader is going to be reinitialized."; + VLOG(3) << "WARNING: The number of given lod_levels(" + << multiple_lod_level.size() + << ") doesn't match the existing tensor number(" + << GetTensorDescNum() + << "). The Reader is going to be reinitialized."; SetTensorDescNum(multiple_lod_level.size()); } switch (desc_.type().type()) { diff --git a/paddle/fluid/framework/variable_helper.cc b/paddle/fluid/framework/variable_helper.cc new file mode 100644 index 0000000000000..fc4525549caee --- /dev/null +++ b/paddle/fluid/framework/variable_helper.cc @@ -0,0 +1,60 @@ +/* Copyright (c) 2016 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/variable_helper.h" + +#include + +#include "paddle/fluid/framework/feed_fetch_type.h" +#include "paddle/fluid/framework/lod_rank_table.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/lod_tensor_array.h" +#include "paddle/fluid/framework/reader.h" +#include "paddle/fluid/framework/scope.h" +#include "paddle/fluid/framework/selected_rows.h" +#include "paddle/fluid/platform/place.h" + +namespace paddle { +namespace framework { +void InitializeVariable(Variable* var, proto::VarType::Type var_type) { + if (var_type == proto::VarType::LOD_TENSOR) { + var->GetMutable(); + } else if (var_type == proto::VarType::SELECTED_ROWS) { + var->GetMutable(); + } else if (var_type == proto::VarType::FEED_MINIBATCH) { + var->GetMutable(); + } else if (var_type == proto::VarType::FETCH_LIST) { + var->GetMutable(); + } else if (var_type == proto::VarType::STEP_SCOPES) { + var->GetMutable>(); + } else if (var_type == proto::VarType::LOD_RANK_TABLE) { + var->GetMutable(); + } else if (var_type == proto::VarType::LOD_TENSOR_ARRAY) { + var->GetMutable(); + } else if (var_type == proto::VarType::PLACE_LIST) { + var->GetMutable(); + } else if (var_type == proto::VarType::READER) { + var->GetMutable(); + } else if (var_type == proto::VarType::RAW) { + // GetMutable will be called in operator + } else { + PADDLE_THROW( + "Variable type %d is not in " + "[LOD_TENSOR, SELECTED_ROWS, FEED_MINIBATCH, FETCH_LIST, " + "LOD_RANK_TABLE, PLACE_LIST, READER, RAW]", + var_type); + } +} +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/operators/tensorrt/tensorrt_engine_op.cu.cc b/paddle/fluid/framework/variable_helper.h similarity index 56% rename from paddle/fluid/operators/tensorrt/tensorrt_engine_op.cu.cc rename to paddle/fluid/framework/variable_helper.h index cbe1b426f6538..0e0c72c3621dc 100644 --- a/paddle/fluid/operators/tensorrt/tensorrt_engine_op.cu.cc +++ b/paddle/fluid/framework/variable_helper.h @@ -11,14 +11,12 @@ distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ +#pragma once -#include "paddle/fluid/operators/tensorrt/tensorrt_engine_op.h" - -namespace ops = paddle::operators; - -REGISTER_OP_CUDA_KERNEL( - tensorrt_engine, - ops::TensorRTEngineKernel, - ops::TensorRTEngineKernel, - ops::TensorRTEngineKernel, - ops::TensorRTEngineKernel); +#include "paddle/fluid/framework/framework.pb.h" +#include "paddle/fluid/framework/variable.h" +namespace paddle { +namespace framework { +void InitializeVariable(Variable *var, proto::VarType::Type var_type); +} +} diff --git a/paddle/fluid/imperative/CMakeLists.txt b/paddle/fluid/imperative/CMakeLists.txt new file mode 100644 index 0000000000000..373d292b443b7 --- /dev/null +++ b/paddle/fluid/imperative/CMakeLists.txt @@ -0,0 +1,3 @@ +cc_library(layer SRCS layer.cc DEPS proto_desc operator) +cc_library(tracer SRCS tracer.cc DEPS proto_desc) +cc_library(engine SRCS engine.cc) diff --git a/paddle/fluid/imperative/engine.cc b/paddle/fluid/imperative/engine.cc new file mode 100644 index 0000000000000..de7ab0e591828 --- /dev/null +++ b/paddle/fluid/imperative/engine.cc @@ -0,0 +1,53 @@ +// 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/imperative/engine.h" + +#include // NOLINT +#include + +#include "glog/logging.h" + +namespace paddle { +namespace imperative { + +static std::once_flag init_engine; +static Engine* engine; + +class DummyEngine : public Engine { + public: + void Enqueue(Runnable* runnable) override { + queued_runnables_.push_back(runnable); + } + + size_t Size() const override { return queued_runnables_.size(); } + + void Sync() override { + for (Runnable* l : queued_runnables_) { + LOG(INFO) << "running " << reinterpret_cast(l); + } + queued_runnables_.clear(); + } + + private: + std::vector queued_runnables_; +}; + +Engine* GetEngine() { + std::call_once(init_engine, []() { engine = new DummyEngine(); }); + return engine; +} + +} // namespace imperative +} // namespace paddle diff --git a/paddle/fluid/imperative/engine.h b/paddle/fluid/imperative/engine.h new file mode 100644 index 0000000000000..a1dfa5bda38d0 --- /dev/null +++ b/paddle/fluid/imperative/engine.h @@ -0,0 +1,39 @@ +// 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 +#include + +namespace paddle { +namespace imperative { + +struct Runnable {}; + +class Engine { + public: + virtual ~Engine() {} + + virtual void Enqueue(Runnable* runnable) = 0; + + virtual size_t Size() const = 0; + + virtual void Sync() = 0; +}; + +Engine* GetEngine(); + +} // namespace imperative +} // namespace paddle diff --git a/paddle/fluid/imperative/layer.cc b/paddle/fluid/imperative/layer.cc new file mode 100644 index 0000000000000..342cb68ab2bf8 --- /dev/null +++ b/paddle/fluid/imperative/layer.cc @@ -0,0 +1,223 @@ +// 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/imperative/layer.h" +#include +#include +#include +#include +#include + +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/string/printf.h" + +namespace paddle { +namespace imperative { + +using framework::Variable; + +void AddTo(Variable* src, Variable* dst) { + framework::LoDTensor* dst_tensor = dst->GetMutable(); + framework::LoDTensor* src_tensor = src->GetMutable(); + PADDLE_ENFORCE(dst_tensor->numel() == src_tensor->numel(), "%lld vs %lld", + dst_tensor->numel(), src_tensor->numel()); + float* dst_data = dst_tensor->mutable_data(platform::CPUPlace()); + const float* src_data = src_tensor->data(); + for (size_t i = 0; i < src_tensor->numel(); ++i) { + dst_data[i] += src_data[i]; + } +} + +class Autograd { + public: + explicit Autograd(framework::Scope* scope) : scope_(scope) {} + + void RunBackward(VarBase* var) { + PADDLE_ENFORCE(var->pre_op_->op_desc_); + // TODO(panyx0718): Only create for vars that "require_grad" + (*var->pre_op_->output_vars_)[var->pre_op_out_idx_]->grads_ = var->grads_; + + std::deque ready; + ready.push_back(var->pre_op_); + + std::map dep_counts = ComputeDepCounts(var->pre_op_); + + while (!ready.empty()) { + OpBase* ready_op = ready.front(); + ready.pop_front(); + std::vector input_grads = ready_op->ApplyGrad(scope_); + + for (size_t i = 0; i < input_grads.size(); ++i) { + if (!input_grads[i]) continue; + OpBase* pre_op = ready_op->pre_ops_->at(i); + if (!pre_op) continue; + + dep_counts[pre_op] -= 1; + PADDLE_ENFORCE(dep_counts[pre_op] >= 0); + bool pre_op_ready = dep_counts[pre_op] == 0; + if (pre_op_ready) { + ready.push_back(pre_op); + } + } + } + } + + private: + std::map ComputeDepCounts(OpBase* op) { + std::map ret; + + std::deque queue; + queue.push_back(op); + std::unordered_set visited; + visited.insert(op); + while (!queue.empty()) { + OpBase* candidate = queue.front(); + queue.pop_front(); + for (OpBase* pre_op : *(candidate->pre_ops_)) { + if (!pre_op) continue; + if (visited.find(pre_op) == visited.end()) { + visited.insert(pre_op); + queue.push_back(pre_op); + } + ret[pre_op] += 1; + } + } + + return ret; + } + + framework::Scope* scope_; +}; + +framework::Variable* CreateVariable(const std::string& name, + const framework::DDim& dim, float val, + framework::Scope* scope, + bool random_name = true) { + std::string varname = name; + if (random_name) { + std::mt19937 rng; + rng.seed(std::random_device()()); + std::uniform_int_distribution dist6( + 1, std::numeric_limits::max()); + int id = dist6(rng); + varname = string::Sprintf("%s@%d", varname, id); + } + + VLOG(3) << "creating var " << varname; + framework::Variable* var = scope->Var(varname); + framework::LoDTensor* tensor = var->GetMutable(); + + float* data = tensor->mutable_data(dim, platform::CPUPlace()); + std::fill(data, data + tensor->numel(), val); + return var; +} + +framework::LoDTensor& VarBase::Grad() { + VLOG(3) << "get var grad " << var_desc_->Name(); + return *grads_->GetMutable(); +} + +void VarBase::ApplyGrad(framework::Scope* scope, Variable* grad) { + VLOG(3) << "apply var grad " << var_desc_->Name() << " " + << grad->Get().data()[0]; + if (!grads_) { + grads_ = + CreateVariable(string::Sprintf("%s@IGrad", var_desc_->Name()), + var_->Get().dims(), 0.0, scope); + } + AddTo(grad, grads_); + VLOG(3) << "grad_ after apply var grad " << var_desc_->Name() << " " + << grads_->Get().data()[0]; +} + +std::vector OpBase::ApplyGrad(framework::Scope* scope) { + VLOG(3) << "op grad " << grad_op_desc_->Type(); + + for (const std::string& grad_invar : grad_op_desc_->InputArgumentNames()) { + if (grad_to_var_->find(grad_invar) == grad_to_var_->end()) { + // grad op inputs can be forward inputs, so not in grad_to_var. + continue; + } + VLOG(3) << "op grad in var " << grad_invar; + block_->FindRecursiveOrCreateVar(grad_invar); + framework::Variable* var = scope->Var(grad_invar); + const std::string& invar = grad_to_var_->at(grad_invar); + for (VarBase* varbase : *output_vars_) { + // Use the accumulated grads_ by sharing the input with grads_. + if (varbase->var_desc_->Name() == invar) { + var->GetMutable()->ShareDataWith( + varbase->grads_->Get()); + break; + } + } + } + + for (const std::string& outvar : grad_op_desc_->OutputArgumentNames()) { + VLOG(3) << "grad outvar " << outvar; + block_->FindRecursiveOrCreateVar(outvar); + framework::Variable* var = scope->Var(outvar); + if (!var->IsInitialized()) { + framework::VarDesc* var_desc = block_->FindVar(outvar); + if (var_desc->GetType() == framework::proto::VarType::LOD_TENSOR) { + var->GetMutable(); + } else { + LOG(ERROR) << "tracer doesn't support yet"; + } + } + } + grad_op_desc_->InferShape(*block_); + grad_op_desc_->InferVarType(block_); + std::unique_ptr opbase = + framework::OpRegistry::CreateOp(*grad_op_desc_); + + opbase->Run(*scope, platform::CPUPlace()); + + // `ret` matches exactly with `input_vars_` of forward op. + std::vector ret; + for (size_t i = 0; i < input_vars_->size(); ++i) { + bool found = false; + VarBase* origin_var = (*input_vars_)[i]; + for (const std::string& outvar : grad_op_desc_->OutputArgumentNames()) { + Variable* var = scope->FindVar(outvar); + std::string orig_var = grad_to_var_->at(outvar); + if (origin_var->var_desc_->Name() != orig_var) { + continue; + } + VLOG(3) << "apply grad " << outvar << " with origin " << orig_var; + origin_var->ApplyGrad(scope, var); + found = true; + ret.push_back(var); + // TODO(panyx0718): There might be another outvar with the same name. + // In that case, it doesn't matter the first one or the second one is + // used. + break; + } + if (!found) { + ret.push_back(nullptr); + } + } + return ret; +} + +void VarBase::RunBackward(framework::Scope* scope) { + grads_ = CreateVariable(framework::GradVarName(var_desc_->Name()), + var_->Get().dims(), 1.0, scope, + false); + if (!pre_op_) return; + Autograd(scope).RunBackward(this); +} + +} // namespace imperative +} // namespace paddle diff --git a/paddle/fluid/imperative/layer.h b/paddle/fluid/imperative/layer.h new file mode 100644 index 0000000000000..85a71ca83d21e --- /dev/null +++ b/paddle/fluid/imperative/layer.h @@ -0,0 +1,102 @@ +// 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 +#include +#include "paddle/fluid/framework/op_desc.h" +#include "paddle/fluid/framework/operator.h" +#include "paddle/fluid/framework/scope.h" +#include "paddle/fluid/framework/var_desc.h" +#include "paddle/fluid/platform/enforce.h" + +namespace paddle { +namespace imperative { + +class OpBase; + +class VarBase { + public: + VarBase() + : pre_op_(nullptr), + pre_op_out_idx_(-1), + var_desc_(nullptr), + var_(nullptr), + grads_(nullptr) {} + + virtual ~VarBase() {} + + void ApplyGrad(framework::Scope* scope, framework::Variable* grad); + + void RunBackward(framework::Scope* scope); + + framework::LoDTensor& Grad(); + + OpBase* pre_op_; + int pre_op_out_idx_; + + framework::VarDesc* var_desc_; + framework::Variable* var_; + framework::Variable* grads_; +}; + +class OpBase { + public: + OpBase() + : input_vars_(new std::vector()), + output_vars_(new std::vector()), + pre_ops_(new std::vector()), + pre_ops_out_idx_(new std::vector()), + op_desc_(nullptr), + grad_op_desc_(nullptr) {} + + virtual ~OpBase() { + delete input_vars_; + delete output_vars_; + + delete pre_ops_; + delete pre_ops_out_idx_; + + if (grad_op_desc_) delete grad_op_desc_; + if (grad_to_var_) delete grad_to_var_; + } + + std::vector ApplyGrad(framework::Scope* scope); + + std::vector* input_vars_; + std::vector* output_vars_; + std::vector* pre_ops_; + std::vector* pre_ops_out_idx_; + framework::OpDesc* op_desc_; + + framework::OpDesc* grad_op_desc_; + std::unordered_map* grad_to_var_; + framework::BlockDesc* block_; +}; + +class Layer { + public: + virtual ~Layer() {} + + virtual std::vector Forward(const std::vector& inputs) { + std::vector vars; + return vars; + } + + virtual void Backward() { LOG(ERROR) << "To support customize"; } +}; + +} // namespace imperative +} // namespace paddle diff --git a/paddle/fluid/imperative/tracer.cc b/paddle/fluid/imperative/tracer.cc new file mode 100644 index 0000000000000..f64f9e72c4a23 --- /dev/null +++ b/paddle/fluid/imperative/tracer.cc @@ -0,0 +1,19 @@ +// 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/imperative/tracer.h" + +namespace paddle { +namespace imperative {} // namespace imperative +} // namespace paddle diff --git a/paddle/fluid/imperative/tracer.h b/paddle/fluid/imperative/tracer.h new file mode 100644 index 0000000000000..97772dc110135 --- /dev/null +++ b/paddle/fluid/imperative/tracer.h @@ -0,0 +1,141 @@ +// 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 +#include +#include + +#include "paddle/fluid/framework/op_desc.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/framework/scope.h" +#include "paddle/fluid/imperative/engine.h" +#include "paddle/fluid/imperative/layer.h" + +namespace paddle { +namespace imperative { + +void CreateGradOp(const framework::OpDesc& op_desc, + const std::unordered_set& no_grad_set, + const std::vector& grad_sub_block, + framework::OpDesc** grad_op_desc, + std::unordered_map* grad_to_var) { + std::vector> grad_op_descs = + framework::OpInfoMap::Instance() + .Get(op_desc.Type()) + .GradOpMaker()(op_desc, no_grad_set, grad_to_var, grad_sub_block); + PADDLE_ENFORCE(grad_op_descs.size() == 1, "Only support 1 grad op now."); + // TODO(panyx0718): Leak? + *grad_op_desc = grad_op_descs[0].release(); +} + +class Tracer { + public: + explicit Tracer(framework::BlockDesc* root_block, + framework::BlockDesc* startup_block) + : root_block_(root_block), startup_block_(startup_block) { + root_scope_ = new framework::Scope(); + scopes_[root_block_] = root_scope_; + scopes_[startup_block_] = root_scope_; + } + + virtual ~Tracer() { delete root_scope_; } + + void Trace(OpBase* op, const std::vector& inputs, + const std::vector& outputs, + framework::BlockDesc* block) { + framework::Scope* scope = GetScope(block); + framework::OpDesc* op_desc = op->op_desc_; + VLOG(3) << "tracer tracing " << op_desc->Type(); + op_desc->InferShape(*block); + op_desc->InferVarType(block); + std::unique_ptr op_base = + framework::OpRegistry::CreateOp(*op_desc); + + *op->input_vars_ = inputs; + for (VarBase* input : inputs) { + const std::string vname = input->var_desc_->Name(); + framework::Variable* var = scope->Var(vname); + input->var_ = var; + if (!var->IsInitialized()) { + framework::VarDesc* var_desc = block->FindVar(vname); + if (var_desc->GetType() == framework::proto::VarType::LOD_TENSOR) { + var->GetMutable(); + } else { + LOG(ERROR) << "tracer doesn't support yet"; + } + } + if (input->pre_op_) { + op->pre_ops_->push_back(input->pre_op_); + op->pre_ops_out_idx_->push_back(input->pre_op_out_idx_); + } else { + op->pre_ops_->push_back(nullptr); + } + VLOG(3) << "input vname " << vname << " " + << var->Get().dims().size(); + } + + *op->output_vars_ = outputs; + for (size_t i = 0; i < outputs.size(); ++i) { + const std::string vname = outputs[i]->var_desc_->Name(); + framework::Variable* var = scope->Var(vname); + if (!var->IsInitialized()) { + framework::VarDesc* var_desc = block->FindVar(vname); + if (var_desc->GetType() == framework::proto::VarType::LOD_TENSOR) { + var->GetMutable(); + } else { + LOG(ERROR) << "tracer doesn't support yet"; + } + } + outputs[i]->var_ = var; + outputs[i]->pre_op_ = op; + outputs[i]->pre_op_out_idx_ = i; + } + + VLOG(3) << "tracer running " << op_desc->Type(); + op_base->Run(*scope, platform::CPUPlace()); + if (block == startup_block_) { + op->grad_op_desc_ = nullptr; + op->grad_to_var_ = nullptr; + } else { + framework::OpDesc* grad_op_desc; + auto grad_to_var = new std::unordered_map(); + CreateGradOp(*op_desc, {}, {block}, &grad_op_desc, grad_to_var); + op->grad_op_desc_ = grad_op_desc; + op->grad_to_var_ = grad_to_var; + } + op->block_ = block; + } + + framework::Scope* GetScope(framework::BlockDesc* block) { + if (scopes_.find(block) != scopes_.end()) { + return scopes_.at(block); + } + framework::BlockDesc* parent_block = block->ParentBlock(); + PADDLE_ENFORCE(scopes_.find(parent_block) != scopes_.end()); + framework::Scope* scope = &scopes_[parent_block]->NewScope(); + scopes_[block] = scope; + return scope; + } + + private: + std::map scopes_; + framework::BlockDesc* root_block_; + framework::BlockDesc* startup_block_; + framework::Scope* root_scope_; +}; + +} // namespace imperative +} // namespace paddle diff --git a/paddle/fluid/inference/CMakeLists.txt b/paddle/fluid/inference/CMakeLists.txt index 2c5364b72402b..b80e7ef752c52 100644 --- a/paddle/fluid/inference/CMakeLists.txt +++ b/paddle/fluid/inference/CMakeLists.txt @@ -4,6 +4,7 @@ endif() # analysis and tensorrt must be added before creating static library, # otherwise, there would be undefined reference to them in static library. add_subdirectory(analysis) +add_subdirectory(utils) if (TENSORRT_FOUND) add_subdirectory(tensorrt) endif() @@ -25,9 +26,6 @@ endif(WIN32) # paddle_fluid_origin exclude inference api interface if(WIN32) sep_library(paddle_fluid_origin DEPS ${fluid_modules} paddle_fluid_api) - if(WITH_GPU AND NOT WITH_DSO) - target_link_libraries(paddle_fluid_origin ${cuda_modules}) - endif(WITH_GPU AND NOT WITH_DSO) else(WIN32) cc_library(paddle_fluid_origin DEPS ${fluid_modules} paddle_fluid_api) endif(WIN32) @@ -43,9 +41,6 @@ set(SHARED_INFERENCE_SRCS if(WIN32) sep_library(paddle_fluid DEPS ${fluid_modules} ${STATIC_INFERENCE_APIS} zero_copy_tensor reset_tensor_array analysis_config paddle_pass_builder) - if(WITH_GPU AND NOT WITH_DSO) - target_link_libraries(paddle_fluid ${cuda_modules}) - endif(WITH_GPU AND NOT WITH_DSO) else(WIN32) cc_library(paddle_fluid DEPS ${fluid_modules} ${STATIC_INFERENCE_APIS} zero_copy_tensor reset_tensor_array analysis_config paddle_pass_builder) @@ -62,9 +57,6 @@ if(WIN32) sep_library(paddle_fluid_shared SHARED SRCS ${SHARED_INFERENCE_SRCS} DEPS ${fluid_modules} paddle_fluid_api reset_tensor_array analysis_config paddle_pass_builder) target_link_libraries(paddle_fluid_shared shlwapi) - if(WITH_GPU AND NOT WITH_DSO) - target_link_libraries(paddle_fluid_origin ${cuda_modules}) - endif(WITH_GPU AND NOT WITH_DSO) else(WIN32) cc_library(paddle_fluid_shared SHARED SRCS ${SHARED_INFERENCE_SRCS} DEPS ${fluid_modules} paddle_fluid_api reset_tensor_array analysis_config paddle_pass_builder) diff --git a/paddle/fluid/inference/analysis/CMakeLists.txt b/paddle/fluid/inference/analysis/CMakeLists.txt index 4bd3f93ef75ad..27b6b80955e45 100644 --- a/paddle/fluid/inference/analysis/CMakeLists.txt +++ b/paddle/fluid/inference/analysis/CMakeLists.txt @@ -35,4 +35,5 @@ function(inference_analysis_test TARGET) endif() endfunction(inference_analysis_test) -inference_analysis_test(test_analyzer SRCS analyzer_tester.cc EXTRA_DEPS reset_tensor_array paddle_inference_api) +inference_analysis_test(test_analyzer SRCS analyzer_tester.cc + EXTRA_DEPS reset_tensor_array paddle_inference_api) diff --git a/paddle/fluid/inference/analysis/analysis_pass.h b/paddle/fluid/inference/analysis/analysis_pass.h index 299f235a74ae0..d5a972fab3bea 100644 --- a/paddle/fluid/inference/analysis/analysis_pass.h +++ b/paddle/fluid/inference/analysis/analysis_pass.h @@ -46,8 +46,6 @@ class AnalysisPass { protected: // User should implement these. virtual void RunImpl(Argument* argument) = 0; - - Argument* argument_{nullptr}; }; } // namespace analysis diff --git a/paddle/fluid/inference/analysis/analyzer_tester.cc b/paddle/fluid/inference/analysis/analyzer_tester.cc index 84a0c3374c66f..cb88333d15703 100644 --- a/paddle/fluid/inference/analysis/analyzer_tester.cc +++ b/paddle/fluid/inference/analysis/analyzer_tester.cc @@ -19,6 +19,7 @@ #include "paddle/fluid/inference/analysis/ut_helper.h" #include "paddle/fluid/inference/api/paddle_inference_api.h" #include "paddle/fluid/inference/api/paddle_inference_pass.h" +#include "paddle/fluid/platform/port.h" namespace paddle { namespace inference { @@ -75,7 +76,8 @@ void TestWord2vecPrediction(const std::string& model_path) { 0.000932706}; const size_t num_elements = outputs.front().data.length() / sizeof(float); // The outputs' buffers are in CPU memory. - for (size_t i = 0; i < std::min(5UL, num_elements); i++) { + for (size_t i = 0; i < std::min(static_cast(5UL), num_elements); + i++) { LOG(INFO) << "data: " << static_cast(outputs.front().data.data())[i]; PADDLE_ENFORCE(static_cast(outputs.front().data.data())[i], diff --git a/paddle/fluid/inference/analysis/argument.h b/paddle/fluid/inference/analysis/argument.h index 21203e2d9f4e4..83d411eecf6d7 100644 --- a/paddle/fluid/inference/analysis/argument.h +++ b/paddle/fluid/inference/analysis/argument.h @@ -103,6 +103,7 @@ struct Argument { // Model specified with program and parameters files. DECL_ARGUMENT_FIELD(model_program_path, ModelProgramPath, std::string); DECL_ARGUMENT_FIELD(model_params_path, ModelParamsPath, std::string); + DECL_ARGUMENT_FIELD(model_from_memory, ModelFromMemory, bool); // The overall graph to work on. DECL_ARGUMENT_UNIQUE_FIELD(main_graph, MainGraph, framework::ir::Graph); @@ -115,6 +116,10 @@ struct Argument { DECL_ARGUMENT_FIELD(ir_analysis_passes, IrAnalysisPasses, std::vector); + // Pass a set of op types to enable its mkldnn kernel + DECL_ARGUMENT_FIELD(mkldnn_enabled_op_types, MKLDNNEnabledOpTypes, + std::unordered_set); + DECL_ARGUMENT_FIELD(use_gpu, UseGPU, bool); DECL_ARGUMENT_FIELD(gpu_device_id, GPUDeviceId, int); DECL_ARGUMENT_FIELD(use_tensorrt, UseTensorRT, bool); diff --git a/paddle/fluid/inference/analysis/ir_pass_manager.cc b/paddle/fluid/inference/analysis/ir_pass_manager.cc index fce5e1cac9206..51bca8039d453 100644 --- a/paddle/fluid/inference/analysis/ir_pass_manager.cc +++ b/paddle/fluid/inference/analysis/ir_pass_manager.cc @@ -63,6 +63,11 @@ void IRPassManager::CreatePasses(Argument *argument, pass->Set("graph_viz_path", new std::string(std::move(dot_file_path))); pass_num++; } + if (pass_name == "mkldnn_placement_pass") { + pass->Set("mkldnn_enabled_op_types", + new std::unordered_set( + argument->mkldnn_enabled_op_types())); + } if (pass_name == "tensorrt_subgraph_pass") { PADDLE_ENFORCE(argument->tensorrt_node_teller_valid()); diff --git a/paddle/fluid/inference/analysis/ir_passes/tensorrt_subgraph_pass.cc b/paddle/fluid/inference/analysis/ir_passes/tensorrt_subgraph_pass.cc index c6b7c05f784b7..9c42b83e7add3 100644 --- a/paddle/fluid/inference/analysis/ir_passes/tensorrt_subgraph_pass.cc +++ b/paddle/fluid/inference/analysis/ir_passes/tensorrt_subgraph_pass.cc @@ -63,7 +63,6 @@ std::unique_ptr analysis::TensorRtSubgraphPass::ApplyImpl( void TensorRtSubgraphPass::CreateTensorRTOp(framework::ir::Node *node, Graph *graph) const { auto *op_desc = node->Op(); - static int counter{0}; auto &subgraph = *Agent(node).subgraph(); PADDLE_ENFORCE(!subgraph.empty()); @@ -178,11 +177,12 @@ void TensorRtSubgraphPass::CreateTensorRTOp(framework::ir::Node *node, output_mapping.push_back(output_name_map[name]); } - *block_desc.Proto()->mutable_vars() = - const_cast(&graph->program()) - ->Proto() - ->blocks(0) - .vars(); + auto *vars = block_desc.Proto()->mutable_vars(); + for (framework::ir::Node *node : graph->Nodes()) { + if (node->IsVar() && node->Var()) { + *vars->Add() = *node->Var()->Proto(); + } + } PADDLE_ENFORCE(!block_desc.Proto()->vars().empty(), "the block has no var-desc"); PADDLE_ENFORCE(!output_mapping.empty()); @@ -191,8 +191,6 @@ void TensorRtSubgraphPass::CreateTensorRTOp(framework::ir::Node *node, block_desc.Proto()->SerializeAsString()); SetAttr(op_desc->Proto(), "max_batch_size", Get("max_batch_size")); SetAttr(op_desc->Proto(), "workspace_size", Get("workspace_size")); - SetAttr(op_desc->Proto(), "engine_uniq_key", - "trt-" + std::to_string(counter++)); SetAttr(op_desc->Proto(), "parameters", ExtractParameters(graph->Nodes())); SetAttr(op_desc->Proto(), "output_name_mapping", output_mapping); } diff --git a/paddle/fluid/inference/analysis/passes/CMakeLists.txt b/paddle/fluid/inference/analysis/passes/CMakeLists.txt index a30c27b1183a7..d3ea511d8f4d8 100644 --- a/paddle/fluid/inference/analysis/passes/CMakeLists.txt +++ b/paddle/fluid/inference/analysis/passes/CMakeLists.txt @@ -1,6 +1,7 @@ cc_library(ir_graph_build_pass SRCS ir_graph_build_pass.cc DEPS analysis_pass argument ir_pass_manager) cc_library(ir_analysis_pass SRCS ir_analysis_pass.cc DEPS analysis_pass argument ir_pass_manager) -cc_library(analysis_passes SRCS passes.cc DEPS ir_graph_build_pass ir_analysis_pass) +cc_library(ir_params_sync_among_devices_pass SRCS ir_params_sync_among_devices_pass.cc DEPS analysis_pass argument ir_pass_manager) +cc_library(analysis_passes SRCS passes.cc DEPS ir_graph_build_pass ir_analysis_pass ir_params_sync_among_devices_pass) set(analysis_deps ${analysis_deps} ir_graph_build_pass diff --git a/paddle/fluid/inference/analysis/passes/ir_analysis_compose_pass.cc b/paddle/fluid/inference/analysis/passes/ir_analysis_compose_pass.cc index 108cb6f74b120..c3a2b3ca1d3b0 100644 --- a/paddle/fluid/inference/analysis/passes/ir_analysis_compose_pass.cc +++ b/paddle/fluid/inference/analysis/passes/ir_analysis_compose_pass.cc @@ -61,6 +61,7 @@ void IrAnalysisComposePass::InitTensorRTAttrs(Argument *argument) { void IrAnalysisComposePass::ApplyIrPasses(Argument *argument) { std::vector passes({ "ir_graph_build_pass", "ir_analysis_pass", + "ir_params_sync_among_devices_pass", }); for (const auto &pass : passes) { VLOG(2) << "Run pass " << pass; diff --git a/paddle/fluid/inference/analysis/passes/ir_graph_build_pass.cc b/paddle/fluid/inference/analysis/passes/ir_graph_build_pass.cc index d5e0d90de1da8..c6e923c00484f 100644 --- a/paddle/fluid/inference/analysis/passes/ir_graph_build_pass.cc +++ b/paddle/fluid/inference/analysis/passes/ir_graph_build_pass.cc @@ -36,12 +36,7 @@ void IrGraphBuildPass::RunImpl(Argument *argument) { // so that the parameters will on the same device, or they will keep copying // between difference devices. platform::Place place; - if (argument->use_gpu()) { - PADDLE_ENFORCE(argument->gpu_device_id_valid()); - place = platform::CUDAPlace(argument->gpu_device_id()); - } else { - place = platform::CPUPlace(); - } + place = platform::CPUPlace(); if (argument->model_dir_valid()) { auto program = @@ -49,9 +44,10 @@ void IrGraphBuildPass::RunImpl(Argument *argument) { argument->SetMainProgram(program.release()); } else if (argument->model_program_path_valid() && argument->model_params_path_valid()) { - auto program = - LoadModel(argument->model_program_path(), argument->model_params_path(), - argument->scope_ptr(), place); + auto program = LoadModel( + argument->model_program_path(), argument->model_params_path(), + argument->scope_ptr(), place, + argument->model_from_memory_valid() && argument->model_from_memory()); argument->SetMainProgram(program.release()); } else { PADDLE_THROW( @@ -73,9 +69,14 @@ std::unique_ptr IrGraphBuildPass::LoadModel( std::unique_ptr IrGraphBuildPass::LoadModel( const std::string &program_path, const std::string ¶ms_path, - framework::Scope *scope, const platform::Place &place) { + framework::Scope *scope, const platform::Place &place, + bool model_from_memory) { framework::Executor exe(place); - return Load(&exe, scope, program_path, params_path); + if (!model_from_memory) { + return Load(&exe, scope, program_path, params_path); + } else { + return LoadFromMemory(&exe, scope, program_path, params_path); + } } std::string IrGraphBuildPass::repr() const { return "ir-graph-build-pass"; } diff --git a/paddle/fluid/inference/analysis/passes/ir_graph_build_pass.h b/paddle/fluid/inference/analysis/passes/ir_graph_build_pass.h index 271e64fce579b..adbde0433fad2 100644 --- a/paddle/fluid/inference/analysis/passes/ir_graph_build_pass.h +++ b/paddle/fluid/inference/analysis/passes/ir_graph_build_pass.h @@ -24,7 +24,7 @@ namespace inference { namespace analysis { /* - * Load program and parameter to memory from the disk. + * Load program and parameter to memory from the disk or directly from memory. */ class IrGraphBuildPass : public AnalysisPass { public: @@ -38,7 +38,8 @@ class IrGraphBuildPass : public AnalysisPass { const platform::Place &place); std::unique_ptr LoadModel( const std::string &program_path, const std::string ¶ms_path, - framework::Scope *scope, const platform::Place &place); + framework::Scope *scope, const platform::Place &place, + bool model_from_memory); std::string model_binary_str_; }; diff --git a/paddle/fluid/inference/analysis/passes/ir_params_sync_among_devices_pass.cc b/paddle/fluid/inference/analysis/passes/ir_params_sync_among_devices_pass.cc new file mode 100644 index 0000000000000..8be2d3ac0b105 --- /dev/null +++ b/paddle/fluid/inference/analysis/passes/ir_params_sync_among_devices_pass.cc @@ -0,0 +1,74 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/inference/analysis/passes/ir_params_sync_among_devices_pass.h" +#include "paddle/fluid/framework/data_layout.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/tensor_util.h" +#include "paddle/fluid/platform/enforce.h" + +namespace paddle { +namespace inference { +namespace analysis { + +void IrParamsSyncAmongDevicesPass::RunImpl(Argument *argument) { + PADDLE_ENFORCE(argument->scope_valid()); + PADDLE_ENFORCE(argument->use_gpu_valid()); + + platform::Place place; + + // The parameters are on the cpu, therefore, synchronization is not necessary. + if (!argument->use_gpu()) return; + + LOG(INFO) << "Sync params from CPU to GPU"; + + PADDLE_ENFORCE(argument->gpu_device_id_valid()); + place = platform::CUDAPlace(argument->gpu_device_id()); + + auto *scope = argument->scope_ptr(); + std::vector all_vars = scope->LocalVarNames(); + + // We get all the vars from local_scope instead of the ProgramDesc. + // Because there exists the case that new parameter variables are not added to + // the program in the analysis pass. + for (auto &var_name : all_vars) { + auto *var = scope->FindLocalVar(var_name); + PADDLE_ENFORCE(var != nullptr); + if (var->IsType() || + var->IsType()) { + auto *t = var->GetMutable(); + + platform::CPUPlace cpu_place; + framework::LoDTensor temp_tensor; + temp_tensor.Resize(t->dims()); + temp_tensor.mutable_data(cpu_place); + + // Copy the parameter data to a tmp tensor. + TensorCopySync(*t, cpu_place, &temp_tensor); + // Reallocation the space on GPU + t->mutable_data(place); + + // Copy parameter data to newly allocated GPU space. + TensorCopySync(temp_tensor, place, t); + } + } +} + +std::string IrParamsSyncAmongDevicesPass::repr() const { + return "ir-params-sync-among-devices-pass"; +} + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/passes/ir_params_sync_among_devices_pass.h b/paddle/fluid/inference/analysis/passes/ir_params_sync_among_devices_pass.h new file mode 100644 index 0000000000000..a95f460df6f96 --- /dev/null +++ b/paddle/fluid/inference/analysis/passes/ir_params_sync_among_devices_pass.h @@ -0,0 +1,39 @@ +// 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 +#include + +#include "paddle/fluid/framework/scope.h" +#include "paddle/fluid/inference/analysis/analysis_pass.h" +#include "paddle/fluid/platform/place.h" + +namespace paddle { +namespace inference { +namespace analysis { + +/* + * Sync parameter from CPU to GPU. + */ +class IrParamsSyncAmongDevicesPass : public AnalysisPass { + public: + void RunImpl(Argument *argument) override; + std::string repr() const override; +}; + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/passes/passes.cc b/paddle/fluid/inference/analysis/passes/passes.cc index 2ef515f45f248..9245e32cee284 100644 --- a/paddle/fluid/inference/analysis/passes/passes.cc +++ b/paddle/fluid/inference/analysis/passes/passes.cc @@ -16,6 +16,7 @@ #include "paddle/fluid/inference/analysis/passes/ir_analysis_compose_pass.cc" #include "paddle/fluid/inference/analysis/passes/ir_analysis_pass.h" #include "paddle/fluid/inference/analysis/passes/ir_graph_build_pass.h" +#include "paddle/fluid/inference/analysis/passes/ir_params_sync_among_devices_pass.h" namespace paddle { namespace inference { @@ -27,6 +28,9 @@ PassRegistry::PassRegistry() { std::unique_ptr(new IrGraphBuildPass)); passes_.emplace("ir_analysis_compose_pass", std::unique_ptr(new IrAnalysisComposePass)); + passes_.emplace( + "ir_params_sync_among_devices_pass", + std::unique_ptr(new IrParamsSyncAmongDevicesPass)); } } // namespace analysis diff --git a/paddle/fluid/inference/api/CMakeLists.txt b/paddle/fluid/inference/api/CMakeLists.txt index e9969b84f3348..eda251c5346a6 100644 --- a/paddle/fluid/inference/api/CMakeLists.txt +++ b/paddle/fluid/inference/api/CMakeLists.txt @@ -30,7 +30,9 @@ cc_library(paddle_pass_builder SRCS paddle_pass_builder.cc) cc_library(analysis_predictor SRCS analysis_predictor.cc DEPS paddle_inference_api analysis naive_executor zero_copy_tensor reset_tensor_array analysis_config paddle_pass_builder ir_pass_manager) cc_library(zero_copy_tensor SRCS details/zero_copy_tensor.cc DEPS scope lod_tensor enforce) cc_library(zero_copy_tensor_dummy SRCS details/zero_copy_tensor_dummy.cc) -cc_library(paddle_inference_api SRCS api.cc api_impl.cc helper.cc DEPS lod_tensor scope paddle_pass_builder reset_tensor_array analysis_config analysis_config paddle_pass_builder DEPS zero_copy_tensor) +cc_library(paddle_inference_api SRCS api.cc api_impl.cc helper.cc DEPS + lod_tensor scope paddle_pass_builder reset_tensor_array analysis_config + analysis_config paddle_pass_builder zero_copy_tensor reset_tensor_array) cc_test(test_paddle_inference_api SRCS api_tester.cc diff --git a/paddle/fluid/inference/api/analysis_config.cc b/paddle/fluid/inference/api/analysis_config.cc index 5ccd2dc5ab353..dcefdd92f5157 100644 --- a/paddle/fluid/inference/api/analysis_config.cc +++ b/paddle/fluid/inference/api/analysis_config.cc @@ -46,12 +46,18 @@ contrib::AnalysisConfig::AnalysisConfig(const contrib::AnalysisConfig &other) { prog_file = other.prog_file; param_file = other.param_file; specify_input_name = other.specify_input_name; + cpu_math_library_num_threads_ = other.cpu_math_library_num_threads_; // fields from this. enable_ir_optim = other.enable_ir_optim; + // For mkldnn + use_mkldnn_ = other.use_mkldnn_; + mkldnn_enabled_op_types_ = other.mkldnn_enabled_op_types_; + use_feed_fetch_ops = other.use_feed_fetch_ops; use_tensorrt_ = other.use_tensorrt_; tensorrt_max_batchsize_ = other.tensorrt_max_batchsize_; tensorrt_workspace_size_ = other.tensorrt_workspace_size_; + model_from_memory_ = other.model_from_memory_; if (use_gpu) { pass_builder_.reset(new GpuPassStrategy( @@ -72,12 +78,19 @@ contrib::AnalysisConfig::AnalysisConfig(contrib::AnalysisConfig &&other) { prog_file = other.prog_file; param_file = other.param_file; specify_input_name = other.specify_input_name; + cpu_math_library_num_threads_ = other.cpu_math_library_num_threads_; // fields from this. enable_ir_optim = other.enable_ir_optim; + // For mkldnn + use_mkldnn_ = other.use_mkldnn_; + mkldnn_enabled_op_types_ = other.mkldnn_enabled_op_types_; + use_feed_fetch_ops = other.use_feed_fetch_ops; use_tensorrt_ = other.use_tensorrt_; tensorrt_max_batchsize_ = other.tensorrt_max_batchsize_; tensorrt_workspace_size_ = other.tensorrt_workspace_size_; + model_from_memory_ = other.model_from_memory_; + pass_builder_ = std::move(other.pass_builder_); } @@ -100,4 +113,13 @@ void contrib::AnalysisConfig::EnableTensorRtEngine(int workspace_size, pass_builder()->InsertPass(1, "tensorrt_subgraph_pass"); } +void contrib::AnalysisConfig::SetModelBuffer(const char *prog_buffer, + size_t prog_buffer_size, + const char *param_buffer, + size_t param_buffer_size) { + prog_file = std::string(prog_buffer, prog_buffer + prog_buffer_size); + param_file = std::string(param_buffer, param_buffer + param_buffer_size); + model_from_memory_ = true; +} + } // namespace paddle diff --git a/paddle/fluid/inference/api/analysis_predictor.cc b/paddle/fluid/inference/api/analysis_predictor.cc index cb14d2a260280..3937884ce4a5a 100644 --- a/paddle/fluid/inference/api/analysis_predictor.cc +++ b/paddle/fluid/inference/api/analysis_predictor.cc @@ -31,11 +31,11 @@ #include "paddle/fluid/inference/tensorrt/convert/op_converter.h" #endif #include "paddle/fluid/inference/utils/singleton.h" +#include "paddle/fluid/memory/memcpy.h" #include "paddle/fluid/platform/cpu_helper.h" #include "paddle/fluid/platform/profiler.h" DECLARE_bool(profile); -DECLARE_int32(paddle_num_threads); namespace paddle { @@ -55,8 +55,7 @@ bool IsPersistable(const framework::VarDesc *var) { bool AnalysisPredictor::Init( const std::shared_ptr &parent_scope, const std::shared_ptr &program) { - VLOG(30) << "Predictor::init()"; -#if !defined(_WIN32) + VLOG(3) << "Predictor::init()"; if (FLAGS_profile) { LOG(WARNING) << "Profiler is actived, might affect the performance"; LOG(INFO) << "You can turn off by set gflags '-profile false'"; @@ -64,10 +63,9 @@ bool AnalysisPredictor::Init( : platform::ProfilerState::kCPU; platform::EnableProfiler(tracking_device); } -#endif // no matter with or without MKLDNN - paddle::platform::SetNumThreads(FLAGS_paddle_num_threads); + paddle::platform::SetNumThreads(config_.cpu_math_library_num_threads()); if (!PrepareScope(parent_scope)) { return false; @@ -160,14 +158,21 @@ bool AnalysisPredictor::PrepareExecutor() { return true; } +void AnalysisPredictor::SetMkldnnThreadID(int tid) { +#ifdef PADDLE_WITH_MKLDNN + platform::set_cur_thread_id(tid); +#else + LOG(ERROR) << "Please compile with MKLDNN first to use MKLDNN"; +#endif +} + bool AnalysisPredictor::Run(const std::vector &inputs, std::vector *output_data, int batch_size) { - VLOG(30) << "Predictor::predict"; + VLOG(3) << "Predictor::predict"; inference::Timer timer; timer.tic(); // set feed variable - std::vector feeds; framework::Scope *scope = sub_scope_ ? sub_scope_ : scope_.get(); if (!SetFeed(inputs, scope)) { LOG(ERROR) << "fail to set feed"; @@ -183,17 +188,21 @@ bool AnalysisPredictor::Run(const std::vector &inputs, LOG(ERROR) << "fail to get fetches"; return false; } - VLOG(30) << "predict cost: " << timer.toc() << "ms"; - - // Fix TensorArray reuse not cleaned bug. - tensor_array_batch_cleaner_.CollectTensorArrays(scope_.get()); - tensor_array_batch_cleaner_.ResetTensorArray(); + VLOG(3) << "predict cost: " << timer.toc() << "ms"; + + // All the containers in the scope will be hold in inference, but the + // operators assume that the container will be reset after each batch. + // Here is a bugfix, collect all the container variables, and reset then to a + // bool; the next time, the operator will call MutableData and construct a new + // container again, so that the container will be empty for each batch. + tensor_array_batch_cleaner_.CollectNoTensorVars(sub_scope_); + tensor_array_batch_cleaner_.ResetNoTensorVars(); return true; } bool AnalysisPredictor::SetFeed(const std::vector &inputs, framework::Scope *scope) { - VLOG(30) << "Predictor::set_feed"; + VLOG(3) << "Predictor::set_feed"; if (inputs.size() != feeds_.size()) { LOG(ERROR) << "wrong feed input size, need " << feeds_.size() << " but get " << inputs.size(); @@ -208,17 +217,32 @@ bool AnalysisPredictor::SetFeed(const std::vector &inputs, framework::DDim ddim = framework::make_ddim(inputs[i].shape); void *input_ptr; if (inputs[i].dtype == PaddleDType::INT64) { - input_ptr = input.mutable_data(ddim, platform::CPUPlace()); + input_ptr = input.mutable_data(ddim, place_); } else if (inputs[i].dtype == PaddleDType::FLOAT32) { - input_ptr = input.mutable_data(ddim, platform::CPUPlace()); + input_ptr = input.mutable_data(ddim, place_); } else { LOG(ERROR) << "unsupported feed type " << inputs[i].dtype; return false; } - // TODO(panyx0718): Init LoDTensor from existing memcpy to save a copy. - std::memcpy(static_cast(input_ptr), inputs[i].data.data(), - inputs[i].data.length()); + if (platform::is_cpu_place(place_)) { + // TODO(panyx0718): Init LoDTensor from existing memcpy to save a copy. + std::memcpy(static_cast(input_ptr), inputs[i].data.data(), + inputs[i].data.length()); + } else { +#ifdef PADDLE_WITH_CUDA + platform::DeviceContextPool &pool = + platform::DeviceContextPool::Instance(); + auto *dev_ctx = + static_cast(pool.Get(place_)); + auto dst_gpu_place = boost::get(place_); + memory::Copy(dst_gpu_place, static_cast(input_ptr), + platform::CPUPlace(), inputs[i].data.data(), + inputs[i].data.length(), dev_ctx->stream()); +#else + PADDLE_THROW("Not compile with CUDA, should not reach here."); +#endif + } // TODO(Superjomn) Low performance, need optimization for heavy LoD copy. framework::LoD lod; for (auto &level : inputs[i].lod) { @@ -258,7 +282,7 @@ void AnalysisPredictor::GetFetchOne(const framework::LoDTensor &fetch, bool AnalysisPredictor::GetFetch(std::vector *outputs, framework::Scope *scope) { - VLOG(30) << "Predictor::get_fetch"; + VLOG(3) << "Predictor::get_fetch"; outputs->resize(fetchs_.size()); for (size_t i = 0; i < fetchs_.size(); ++i) { int idx = boost::get(fetchs_[i]->GetAttr("col")); @@ -267,10 +291,11 @@ bool AnalysisPredictor::GetFetch(std::vector *outputs, framework::GetFetchVariable(*scope, "fetch", idx); auto type = fetch.type(); auto output = &(outputs->at(i)); - if (type == typeid(float)) { + output->name = fetchs_[idx]->Input("X")[0]; + if (type == framework::proto::VarType::FP32) { GetFetchOne(fetch, output); output->dtype = PaddleDType::FLOAT32; - } else if (type == typeid(int64_t)) { + } else if (type == framework::proto::VarType::INT64) { GetFetchOne(fetch, output); output->dtype = PaddleDType::INT64; } else { @@ -286,6 +311,7 @@ void AnalysisPredictor::OptimizeInferenceProgram() { argument_.SetUseGPU(config_.use_gpu); argument_.SetGPUDeviceId(config_.device); + argument_.SetModelFromMemory(config_.model_from_memory_); // Analyze inference_program if (!config_.model_dir.empty()) { argument_.SetModelDir(config_.model_dir); @@ -304,6 +330,10 @@ void AnalysisPredictor::OptimizeInferenceProgram() { argument_.SetTensorRtMaxBatchSize(config_.tensorrt_max_batchsize_); } + if (config_.use_mkldnn_) { + argument_.SetMKLDNNEnabledOpTypes(config_.mkldnn_enabled_op_types_); + } + auto passes = config_.pass_builder()->AllPasses(); if (!config_.enable_ir_optim) passes.clear(); argument_.SetIrAnalysisPasses(passes); @@ -321,7 +351,7 @@ void AnalysisPredictor::OptimizeInferenceProgram() { template <> std::unique_ptr CreatePaddlePredictor< AnalysisConfig, PaddleEngineKind::kAnalysis>(const AnalysisConfig &config) { - VLOG(30) << "create AnalysisConfig"; + VLOG(3) << "create AnalysisConfig"; if (config.use_gpu) { // 1. GPU memeroy PADDLE_ENFORCE_GT( @@ -335,7 +365,7 @@ std::unique_ptr CreatePaddlePredictor< std::string flag = "--fraction_of_gpu_memory_to_use=" + std::to_string(config.fraction_of_gpu_memory); flags.push_back(flag); - VLOG(30) << "set flag: " << flag; + VLOG(3) << "set flag: " << flag; framework::InitGflags(flags); } } @@ -399,7 +429,7 @@ std::unique_ptr AnalysisPredictor::GetOutputTensor( bool AnalysisPredictor::ZeroCopyRun() { executor_->Run(); // Fix TensorArray reuse not cleaned bug. - tensor_array_batch_cleaner_.CollectTensorArrays(scope_.get()); + tensor_array_batch_cleaner_.CollectTensorArrays(sub_scope_); tensor_array_batch_cleaner_.ResetTensorArray(); return true; } @@ -426,20 +456,24 @@ bool AnalysisPredictor::LoadProgramDesc() { return false; } - std::string pb_content; - // Read binary - std::ifstream fin(filename, std::ios::in | std::ios::binary); - PADDLE_ENFORCE(static_cast(fin), "Cannot open file %s", filename); - fin.seekg(0, std::ios::end); - - pb_content.resize(fin.tellg()); - fin.seekg(0, std::ios::beg); - fin.read(&(pb_content.at(0)), pb_content.size()); - fin.close(); - // Create ProgramDesc framework::proto::ProgramDesc proto; - proto.ParseFromString(pb_content); + if (!config_.model_from_memory()) { + std::string pb_content; + // Read binary + std::ifstream fin(filename, std::ios::in | std::ios::binary); + PADDLE_ENFORCE(static_cast(fin.is_open()), "Cannot open file %s", + filename); + fin.seekg(0, std::ios::end); + pb_content.resize(fin.tellg()); + fin.seekg(0, std::ios::beg); + fin.read(&(pb_content.at(0)), pb_content.size()); + fin.close(); + + proto.ParseFromString(pb_content); + } else { + proto.ParseFromString(config_.prog_file); + } inference_program_.reset(new framework::ProgramDesc(proto)); return true; } @@ -447,6 +481,7 @@ bool AnalysisPredictor::LoadProgramDesc() { bool AnalysisPredictor::LoadParameters() { PADDLE_ENFORCE_NOT_NULL(inference_program_.get(), "The inference program should be loaded first."); + const auto &global_block = inference_program_->MutableBlock(0); // create a temporary program to load parameters. @@ -501,12 +536,10 @@ bool AnalysisPredictor::LoadParameters() { } AnalysisPredictor::~AnalysisPredictor() { -#if !defined(_WIN32) if (FLAGS_profile) { platform::DisableProfiler(platform::EventSortingKey::kTotal, "./profile.log"); } -#endif if (sub_scope_) { scope_->DeleteScope(sub_scope_); } diff --git a/paddle/fluid/inference/api/analysis_predictor.h b/paddle/fluid/inference/api/analysis_predictor.h index cf81b7db738d8..12ecb7c15e92c 100644 --- a/paddle/fluid/inference/api/analysis_predictor.h +++ b/paddle/fluid/inference/api/analysis_predictor.h @@ -69,6 +69,8 @@ class AnalysisPredictor : public PaddlePredictor { framework::Scope *scope() { return scope_.get(); } framework::ProgramDesc &program() { return *inference_program_; } + void SetMkldnnThreadID(int tid); + protected: bool PrepareProgram(const std::shared_ptr &program); bool PrepareScope(const std::shared_ptr &parent_scope); @@ -107,7 +109,7 @@ class AnalysisPredictor : public PaddlePredictor { std::map feed_names_; std::vector fetchs_; // Memory buffer for feed inputs. The temporary LoDTensor will cause serious - // concurrency problems, so cache them. + // concurrency problems, wrong results and memory leak, so cache them. std::vector feed_tensors_; details::TensorArrayBatchCleaner tensor_array_batch_cleaner_; diff --git a/paddle/fluid/inference/api/analysis_predictor_tester.cc b/paddle/fluid/inference/api/analysis_predictor_tester.cc index d67305670c91b..a361b34437ade 100644 --- a/paddle/fluid/inference/api/analysis_predictor_tester.cc +++ b/paddle/fluid/inference/api/analysis_predictor_tester.cc @@ -55,7 +55,12 @@ TEST(AnalysisPredictor, analysis_off) { } TEST(AnalysisPredictor, analysis_on) { - AnalysisConfig config(false); +#ifdef PADDLE_WITH_CUDA + AnalysisConfig config(true); + config.fraction_of_gpu_memory = 0.15; +#else + AnalysisConfig config; +#endif config.model_dir = FLAGS_dirname; config.enable_ir_optim = true; diff --git a/paddle/fluid/inference/api/api_impl.cc b/paddle/fluid/inference/api/api_impl.cc index fcbc3803d04de..102147a493ed1 100644 --- a/paddle/fluid/inference/api/api_impl.cc +++ b/paddle/fluid/inference/api/api_impl.cc @@ -24,11 +24,11 @@ limitations under the License. */ #include "paddle/fluid/inference/api/api_impl.h" #include "paddle/fluid/inference/api/details/reset_tensor_array.h" #include "paddle/fluid/inference/api/helper.h" +#include "paddle/fluid/memory/memcpy.h" #include "paddle/fluid/platform/cpu_helper.h" #include "paddle/fluid/platform/profiler.h" DEFINE_bool(profile, false, "Turn on profiler for fluid"); -DECLARE_int32(paddle_num_threads); namespace paddle { namespace { @@ -64,7 +64,6 @@ void NativePaddlePredictor::PrepareFeedFetch() { bool NativePaddlePredictor::Init( std::shared_ptr parent_scope) { VLOG(3) << "Predictor::init()"; -#if !defined(_WIN32) if (FLAGS_profile) { LOG(WARNING) << "Profiler is actived, might affect the performance"; LOG(INFO) << "You can turn off by set gflags '-profile false'"; @@ -73,10 +72,9 @@ bool NativePaddlePredictor::Init( : platform::ProfilerState::kCPU; platform::EnableProfiler(tracking_device); } -#endif // no matter with or without MKLDNN - paddle::platform::SetNumThreads(FLAGS_paddle_num_threads); + paddle::platform::SetNumThreads(config_.cpu_math_library_num_threads()); if (config_.use_gpu) { place_ = paddle::platform::CUDAPlace(config_.device); @@ -121,12 +119,10 @@ bool NativePaddlePredictor::Init( } NativePaddlePredictor::~NativePaddlePredictor() { -#if !defined(_WIN32) if (FLAGS_profile) { platform::DisableProfiler(platform::EventSortingKey::kTotal, "./profile.log"); } -#endif if (sub_scope_) { scope_->DeleteScope(sub_scope_); } @@ -139,7 +135,6 @@ bool NativePaddlePredictor::Run(const std::vector &inputs, Timer timer; timer.tic(); // set feed variable - std::vector feeds; framework::Scope *scope = sub_scope_ != nullptr ? sub_scope_ : scope_.get(); if (!SetFeed(inputs, scope)) { LOG(ERROR) << "fail to set feed"; @@ -157,11 +152,11 @@ bool NativePaddlePredictor::Run(const std::vector &inputs, LOG(ERROR) << "fail to get fetches"; return false; } - VLOG(30) << "predict cost: " << timer.toc() << "ms"; + VLOG(3) << "predict cost: " << timer.toc() << "ms"; - // Fix TensorArray reuse not cleaned bug. - tensor_array_batch_cleaner_.CollectTensorArrays(scope_.get()); - tensor_array_batch_cleaner_.ResetTensorArray(); + // For some other vector like containers not cleaned after each batch. + tensor_array_batch_cleaner_.CollectNoTensorVars(scope_.get()); + tensor_array_batch_cleaner_.ResetNoTensorVars(); return true; } @@ -190,22 +185,42 @@ bool NativePaddlePredictor::SetFeed(const std::vector &inputs, << inputs.size(); return false; } + + // Cache the inputs memory for better concurrency performance. + feed_tensors_.resize(inputs.size()); + for (size_t i = 0; i < inputs.size(); ++i) { - framework::LoDTensor input; + auto &input = feed_tensors_[i]; framework::DDim ddim = framework::make_ddim(inputs[i].shape); void *input_ptr; if (inputs[i].dtype == PaddleDType::INT64) { - input_ptr = input.mutable_data(ddim, platform::CPUPlace()); + input_ptr = input.mutable_data(ddim, place_); } else if (inputs[i].dtype == PaddleDType::FLOAT32) { - input_ptr = input.mutable_data(ddim, platform::CPUPlace()); + input_ptr = input.mutable_data(ddim, place_); } else { LOG(ERROR) << "unsupported feed type " << inputs[i].dtype; return false; } - // TODO(panyx0718): Init LoDTensor from existing memcpy to save a copy. - std::memcpy(static_cast(input_ptr), inputs[i].data.data(), - inputs[i].data.length()); + if (platform::is_cpu_place(place_)) { + // TODO(panyx0718): Init LoDTensor from existing memcpy to save a copy. + std::memcpy(static_cast(input_ptr), inputs[i].data.data(), + inputs[i].data.length()); + } else { +#ifdef PADDLE_WITH_CUDA + platform::DeviceContextPool &pool = + platform::DeviceContextPool::Instance(); + auto *dev_ctx = + static_cast(pool.Get(place_)); + auto dst_gpu_place = boost::get(place_); + memory::Copy(dst_gpu_place, static_cast(input_ptr), + platform::CPUPlace(), inputs[i].data.data(), + inputs[i].data.length(), dev_ctx->stream()); +#else + PADDLE_THROW("Not compile with CUDA, should not reach here."); +#endif + } + // TODO(Superjomn) Low performance, need optimization for heavy LoD copy. framework::LoD lod; for (auto &level : inputs[i].lod) { @@ -253,10 +268,11 @@ bool NativePaddlePredictor::GetFetch(std::vector *outputs, framework::GetFetchVariable(*scope, "fetch", idx); auto type = fetch.type(); auto output = &(outputs->at(i)); - if (type == typeid(float)) { + output->name = fetchs_[idx]->Input("X")[0]; + if (type == framework::DataTypeTrait::DataType) { GetFetchOne(fetch, output); output->dtype = PaddleDType::FLOAT32; - } else if (type == typeid(int64_t)) { + } else if (type == framework::DataTypeTrait::DataType) { GetFetchOne(fetch, output); output->dtype = PaddleDType::INT64; } else { diff --git a/paddle/fluid/inference/api/api_impl.h b/paddle/fluid/inference/api/api_impl.h index 9dfa48d501f17..c1fcd198ccda0 100644 --- a/paddle/fluid/inference/api/api_impl.h +++ b/paddle/fluid/inference/api/api_impl.h @@ -69,6 +69,9 @@ class NativePaddlePredictor : public PaddlePredictor { std::vector feeds_; std::map feed_names_; std::vector fetchs_; + // Memory buffer for feed inputs. The temporary LoDTensor will cause serious + // concurrency problems, wrong results and memory leak, so cache them. + std::vector feed_tensors_; // Do not use unique_ptr, use parent scope to delete framework::Scope *sub_scope_{nullptr}; details::TensorArrayBatchCleaner tensor_array_batch_cleaner_; diff --git a/paddle/fluid/inference/api/api_impl_tester.cc b/paddle/fluid/inference/api/api_impl_tester.cc index 014bdc6a37974..78396397397c3 100644 --- a/paddle/fluid/inference/api/api_impl_tester.cc +++ b/paddle/fluid/inference/api/api_impl_tester.cc @@ -36,10 +36,10 @@ namespace paddle { PaddleTensor LodTensorToPaddleTensor(framework::LoDTensor* t) { PaddleTensor pt; - if (t->type() == typeid(int64_t)) { + if (t->type() == framework::proto::VarType::INT64) { pt.data.Reset(t->data(), t->numel() * sizeof(int64_t)); pt.dtype = PaddleDType::INT64; - } else if (t->type() == typeid(float)) { + } else if (t->type() == framework::proto::VarType::FP32) { pt.data.Reset(t->data(), t->numel() * sizeof(float)); pt.dtype = PaddleDType::FLOAT32; } else { diff --git a/paddle/fluid/inference/api/demo_ci/CMakeLists.txt b/paddle/fluid/inference/api/demo_ci/CMakeLists.txt index 49683eab07a2f..f42ee9a697bfb 100644 --- a/paddle/fluid/inference/api/demo_ci/CMakeLists.txt +++ b/paddle/fluid/inference/api/demo_ci/CMakeLists.txt @@ -15,12 +15,43 @@ macro(safe_set_static_flag) endforeach(flag_var) endmacro() +if(NOT DEFINED PADDLE_LIB) + message(FATAL_ERROR "please set PADDLE_LIB with -DPADDLE_LIB=/path/paddle/lib") +endif() +if(NOT DEFINED DEMO_NAME) + message(FATAL_ERROR "please set DEMO_NAME with -DDEMO_NAME=demo_name") +endif() + +include_directories("${PADDLE_LIB}/") +include_directories("${PADDLE_LIB}/fluid_inference_install_dir/") +include_directories("${PADDLE_LIB}/third_party/install/protobuf/include") +include_directories("${PADDLE_LIB}/third_party/install/glog/include") +include_directories("${PADDLE_LIB}/third_party/install/gflags/include") +include_directories("${PADDLE_LIB}/third_party/install/xxhash/include") +include_directories("${PADDLE_LIB}/third_party/install/snappy/include") +include_directories("${PADDLE_LIB}/third_party/install/snappystream/include") +include_directories("${PADDLE_LIB}/third_party/install/zlib/include") +include_directories("${PADDLE_LIB}/third_party/boost") +include_directories("${PADDLE_LIB}/third_party/eigen3") + +link_directories("${PADDLE_LIB}/third_party/install/snappy/lib") +link_directories("${PADDLE_LIB}/third_party/install/snappystream/lib") +link_directories("${PADDLE_LIB}/third_party/install/zlib/lib") +link_directories("${PADDLE_LIB}/third_party/install/protobuf/lib") +link_directories("${PADDLE_LIB}/third_party/install/glog/lib") +link_directories("${PADDLE_LIB}/third_party/install/gflags/lib") +link_directories("${PADDLE_LIB}/third_party/install/xxhash/lib") +link_directories("${PADDLE_LIB}/paddle/lib") + if (WIN32) + add_definitions("/DGOOGLE_GLOG_DLL_DECL=") + set(CMAKE_C_FLAGS_DEBUG "${CMAKE_C_FLAGS_DEBUG} /bigobj /MTd") + set(CMAKE_C_FLAGS_RELEASE "${CMAKE_C_FLAGS_RELEASE} /bigobj /MT") + set(CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} /bigobj /MTd") + set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /bigobj /MT") if (WITH_STATIC_LIB) safe_set_static_flag() add_definitions(-DSTATIC_LIB) - set(CMAKE_CXX_FLAGS ${CMAKE_CXX_FLAGS} "/w") - set(CMAKE_CXX_FLAGS_RELEASE ${CMAKE_CXX_FLAGS_RELEASE} "/w") endif() set(CMAKE_STATIC_LIBRARY_PREFIX "lib") else() @@ -29,39 +60,16 @@ else() endif() message("flags" ${CMAKE_CXX_FLAGS}) -if(NOT DEFINED PADDLE_LIB) - message(FATAL_ERROR "please set PADDLE_LIB with -DPADDLE_LIB=/path/paddle/lib") -endif() -if(NOT DEFINED DEMO_NAME) - message(FATAL_ERROR "please set DEMO_NAME with -DDEMO_NAME=demo_name") -endif() - - if(WITH_GPU) if(NOT WIN32) set(CUDA_LIB "/usr/local/cuda/lib64/" CACHE STRING "CUDA Library") else() if(CUDA_LIB STREQUAL "") - set(CUDA_LIB "C:\\Program\ Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v8.0\\lib\\x64") + set(CUDA_LIB "C:\\Program\ Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v8.0\\lib\\x64") endif() endif(NOT WIN32) endif() -include_directories("D:/Paddle/") -include_directories("${PADDLE_LIB}") -include_directories("${PADDLE_LIB}/third_party/install/protobuf/include") -include_directories("${PADDLE_LIB}/third_party/install/glog/include") -include_directories("${PADDLE_LIB}/third_party/install/gflags/include") -include_directories("${PADDLE_LIB}/third_party/install/xxhash/include") -if (NOT WIN32) -include_directories("${PADDLE_LIB}/third_party/install/snappy/include") -include_directories("${PADDLE_LIB}/third_party/install/snappystream/include") -include_directories("${PADDLE_LIB}/third_party/install/zlib/include") -endif(NOT WIN32) - -include_directories("${PADDLE_LIB}/third_party/boost") -include_directories("${PADDLE_LIB}/third_party/eigen3") - if (NOT WIN32) if (USE_TENSORRT AND WITH_GPU) include_directories("${TENSORRT_INCLUDE_DIR}") @@ -70,27 +78,32 @@ if (NOT WIN32) endif(NOT WIN32) if (NOT WIN32) -link_directories("${PADDLE_LIB}/third_party/install/snappy/lib") -link_directories("${PADDLE_LIB}/third_party/install/snappystream/lib") -link_directories("${PADDLE_LIB}/third_party/install/zlib/lib") -endif(NOT WIN32) - -link_directories("${PADDLE_LIB}/third_party/install/protobuf/lib") -link_directories("${PADDLE_LIB}/third_party/install/glog/lib") -link_directories("${PADDLE_LIB}/third_party/install/gflags/lib") -link_directories("${PADDLE_LIB}/third_party/install/xxhash/lib") -link_directories("${PADDLE_LIB}/paddle/lib") - -add_executable(${DEMO_NAME} ${DEMO_NAME}.cc) + set(NGRAPH_PATH "${PADDLE_LIB}/third_party/install/ngraph") + if(EXISTS ${NGRAPH_PATH}) + include(GNUInstallDirs) + include_directories("${NGRAPH_PATH}/include") + link_directories("${NGRAPH_PATH}/${CMAKE_INSTALL_LIBDIR}") + set(NGRAPH_LIB ${NGRAPH_PATH}/${CMAKE_INSTALL_LIBDIR}/libngraph${CMAKE_SHARED_LIBRARY_SUFFIX}) + endif() +endif() if(WITH_MKL) include_directories("${PADDLE_LIB}/third_party/install/mklml/include") - set(MATH_LIB ${PADDLE_LIB}/third_party/install/mklml/lib/libmklml_intel${CMAKE_SHARED_LIBRARY_SUFFIX} - ${PADDLE_LIB}/third_party/install/mklml/lib/libiomp5${CMAKE_SHARED_LIBRARY_SUFFIX}) + if(NOT WIN32) + set(MATH_LIB ${PADDLE_LIB}/third_party/install/mklml/lib/libmklml_intel${CMAKE_SHARED_LIBRARY_SUFFIX} + ${PADDLE_LIB}/third_party/install/mklml/lib/libiomp5${CMAKE_SHARED_LIBRARY_SUFFIX}) + else(WIN32) + set(MATH_LIB ${PADDLE_LIB}/third_party/install/mklml/lib/libmklml${CMAKE_SHARED_LIBRARY_SUFFIX} + ${PADDLE_LIB}/third_party/install/mklml/lib/libiomp5md${CMAKE_SHARED_LIBRARY_SUFFIX}) + endif(WIN32) set(MKLDNN_PATH "${PADDLE_LIB}/third_party/install/mkldnn") if(EXISTS ${MKLDNN_PATH}) include_directories("${MKLDNN_PATH}/include") - set(MKLDNN_LIB ${MKLDNN_PATH}/lib/libmkldnn.so.0) + if(WIN32) + set(MKLDNN_LIB ${MKLDNN_PATH}/lib/mkldnn.lib) + else(WIN32) + set(MKLDNN_LIB ${MKLDNN_PATH}/lib/libmkldnn.so.0) + endif(WIN32) endif() else() set(MATH_LIB ${PADDLE_LIB}/third_party/install/openblas/lib/libopenblas${CMAKE_STATIC_LIBRARY_SUFFIX}) @@ -98,26 +111,25 @@ endif() # Note: libpaddle_inference_api.so/a must put before libpaddle_fluid.so/a if(WITH_STATIC_LIB) - set(DEPS - ${PADDLE_LIB}/paddle/lib/libpaddle_fluid${CMAKE_STATIC_LIBRARY_SUFFIX}) + set(DEPS ${PADDLE_LIB}/paddle/lib/libpaddle_fluid${CMAKE_STATIC_LIBRARY_SUFFIX}) else() - set(DEPS - ${PADDLE_LIB}/paddle/lib/libpaddle_fluid${CMAKE_SHARED_LIBRARY_SUFFIX}) + set(DEPS ${PADDLE_LIB}/paddle/lib/libpaddle_fluid${CMAKE_SHARED_LIBRARY_SUFFIX}) endif() if (NOT WIN32) -set(EXTERNAL_LIB "-lrt -ldl -lpthread") -set(DEPS ${DEPS} - ${MATH_LIB} ${MKLDNN_LIB} - glog gflags protobuf snappystream snappy z xxhash - ${EXTERNAL_LIB}) + set(EXTERNAL_LIB "-lrt -ldl -lpthread") + set(DEPS ${DEPS} + ${MATH_LIB} ${MKLDNN_LIB} ${NGRAPH_LIB} + glog gflags protobuf snappystream snappy z xxhash + ${EXTERNAL_LIB}) else() -set(DEPS ${DEPS} - ${MATH_LIB} ${MKLDNN_LIB} - ${CMAKE_STATIC_LIBRARY_PREFIX}glog ${CMAKE_STATIC_LIBRARY_PREFIX}gflags ${CMAKE_STATIC_LIBRARY_PREFIX}protobuf - ${EXTERNAL_LIB}) -# NOTE(dzhwinter) shlwapi is deprecated. -set(DEPS ${DEPS} libcmt shlwapi) + set(DEPS ${DEPS} + ${MATH_LIB} ${MKLDNN_LIB} + ${CMAKE_STATIC_LIBRARY_PREFIX}glog ${CMAKE_STATIC_LIBRARY_PREFIX}gflags ${CMAKE_STATIC_LIBRARY_PREFIX}protobuf + ${CMAKE_STATIC_LIBRARY_PREFIX}snappy ${CMAKE_STATIC_LIBRARY_PREFIX}z ${CMAKE_STATIC_LIBRARY_PREFIX}xxhash + snappystream ${EXTERNAL_LIB}) + # NOTE(dzhwinter) shlwapi is deprecated. + set(DEPS ${DEPS} libcmt shlwapi) endif(NOT WIN32) if(WITH_GPU) @@ -129,9 +141,10 @@ if(WITH_GPU) set(DEPS ${DEPS} ${CUDA_LIB}/libcudart${CMAKE_SHARED_LIBRARY_SUFFIX}) else() set(DEPS ${DEPS} ${CUDA_LIB}/cudart${CMAKE_STATIC_LIBRARY_SUFFIX} ) - set(DEPS ${DEPS} ${CUDA_LIB}/cublas${CMAKE_STATIC_LIBRARY_SUFFIX} ) - set(DEPS ${DEPS} ${CUDA_LIB}/cudnn${CMAKE_STATIC_LIBRARY_SUFFIX} ) + set(DEPS ${DEPS} ${CUDA_LIB}/cublas${CMAKE_STATIC_LIBRARY_SUFFIX} ) + set(DEPS ${DEPS} ${CUDA_LIB}/cudnn${CMAKE_STATIC_LIBRARY_SUFFIX} ) endif() endif() +add_executable(${DEMO_NAME} ${DEMO_NAME}.cc) target_link_libraries(${DEMO_NAME} ${DEPS}) diff --git a/paddle/fluid/inference/api/demo_ci/run.sh b/paddle/fluid/inference/api/demo_ci/run.sh index ff718077c1ba6..a94ccfa92439a 100755 --- a/paddle/fluid/inference/api/demo_ci/run.sh +++ b/paddle/fluid/inference/api/demo_ci/run.sh @@ -54,6 +54,9 @@ mkdir -p build cd build for WITH_STATIC_LIB in ON OFF; do +# TODO(Superjomn) reopen this +# something wrong with the TensorArray reset. +:< output; predictor->Run({input}, &output, 1); - VLOG(30) << "output.size " << output.size(); + VLOG(3) << "output.size " << output.size(); auto& tensor = output.front(); - VLOG(30) << "output: " << SummaryTensor(tensor); + VLOG(3) << "output: " << SummaryTensor(tensor); // compare with reference result CheckOutput(FLAGS_refer, tensor); diff --git a/paddle/fluid/inference/api/demo_ci/utils.h b/paddle/fluid/inference/api/demo_ci/utils.h index 664b9d01c7810..d70c6aea79121 100644 --- a/paddle/fluid/inference/api/demo_ci/utils.h +++ b/paddle/fluid/inference/api/demo_ci/utils.h @@ -47,7 +47,7 @@ static void split(const std::string& str, char sep, } Record ProcessALine(const std::string& line) { - VLOG(30) << "process a line"; + VLOG(3) << "process a line"; std::vector columns; split(line, '\t', &columns); CHECK_EQ(columns.size(), 2UL) @@ -65,8 +65,8 @@ Record ProcessALine(const std::string& line) { for (auto& s : shape_strs) { record.shape.push_back(std::stoi(s)); } - VLOG(30) << "data size " << record.data.size(); - VLOG(30) << "data shape size " << record.shape.size(); + VLOG(3) << "data size " << record.data.size(); + VLOG(3) << "data shape size " << record.shape.size(); return record; } @@ -78,8 +78,8 @@ void CheckOutput(const std::string& referfile, const PaddleTensor& output) { file.close(); size_t numel = output.data.length() / PaddleDtypeSize(output.dtype); - VLOG(30) << "predictor output numel " << numel; - VLOG(30) << "reference output numel " << refer.data.size(); + VLOG(3) << "predictor output numel " << numel; + VLOG(3) << "reference output numel " << refer.data.size(); CHECK_EQ(numel, refer.data.size()); switch (output.dtype) { case PaddleDType::INT64: { diff --git a/paddle/fluid/inference/api/details/reset_tensor_array.cc b/paddle/fluid/inference/api/details/reset_tensor_array.cc index 244b0b567b5df..569a487328e2f 100644 --- a/paddle/fluid/inference/api/details/reset_tensor_array.cc +++ b/paddle/fluid/inference/api/details/reset_tensor_array.cc @@ -26,7 +26,7 @@ void TensorArrayBatchCleaner::CollectTensorArrays(framework::Scope *scope) { // parameter. if (var_name == "feed" || var_name == "fetch") continue; if (var->Type() == typeid(framework::LoDTensorArray)) { - VLOG(40) << "collect " << var_name; + VLOG(4) << "collect " << var_name; arrays_.push_back(var->GetMutable()); } } @@ -34,7 +34,7 @@ void TensorArrayBatchCleaner::CollectTensorArrays(framework::Scope *scope) { CollectTensorArrays(kid); } - VLOG(30) << "Collect " << arrays_.size() << " arrays"; + VLOG(3) << "Collect " << arrays_.size() << " arrays"; flag_ = false; } } @@ -46,5 +46,28 @@ void TensorArrayBatchCleaner::ResetTensorArray() { } } +void TensorArrayBatchCleaner::CollectNoTensorVars(framework::Scope *scope) { + if (no_tensor_flag_) { + for (auto &var_name : scope->LocalVarNames()) { + auto *var = scope->FindVar(var_name); + if (!var->IsInitialized()) continue; + if (!valid_types_.count(var->Type())) { + no_tensor_vars_.insert(var); + } + } + + for (auto *kid : scope->kids()) { + CollectTensorArrays(kid); + } + no_tensor_flag_ = false; // Only collect one time. + } +} + +void TensorArrayBatchCleaner::ResetNoTensorVars() { + for (auto *var : no_tensor_vars_) { + var->Clear(); + } +} + } // namespace details } // namespace paddle diff --git a/paddle/fluid/inference/api/details/reset_tensor_array.h b/paddle/fluid/inference/api/details/reset_tensor_array.h index a39449ff0e677..6a5ea64de66fc 100644 --- a/paddle/fluid/inference/api/details/reset_tensor_array.h +++ b/paddle/fluid/inference/api/details/reset_tensor_array.h @@ -14,9 +14,11 @@ #pragma once +#include #include #include "paddle/fluid/framework/lod_tensor_array.h" #include "paddle/fluid/framework/scope.h" +#include "paddle/fluid/framework/variable.h" namespace paddle { namespace details { @@ -24,13 +26,28 @@ namespace details { // Clean the TensorArray each batch to make the behavior the same with the // training phase. struct TensorArrayBatchCleaner { + TensorArrayBatchCleaner() { + valid_types_.insert(typeid(framework::Tensor)); + valid_types_.insert(typeid(framework::LoDTensor)); + } + // Collect the variables that are not Tensor or LoDTensor, and reset them to a + // bool(trick), because some of them are containers, and some operators just + // keep inserting new items without clearing the containers first; So the + // memory grow larger and larger in inference service deployed online. + void CollectNoTensorVars(framework::Scope *scope); + void ResetNoTensorVars(); + // Fix the tensor array not clear in the inference scenarios. void CollectTensorArrays(framework::Scope *scope); void ResetTensorArray(); private: bool flag_{true}; + bool no_tensor_flag_{true}; std::vector arrays_; + + std::unordered_set valid_types_; + std::unordered_set no_tensor_vars_; }; } // namespace details diff --git a/paddle/fluid/inference/api/helper.h b/paddle/fluid/inference/api/helper.h index 6f9d663121004..9a393a61c4b45 100644 --- a/paddle/fluid/inference/api/helper.h +++ b/paddle/fluid/inference/api/helper.h @@ -15,10 +15,6 @@ #pragma once #include -#if !defined(_WIN32) -#include -#else -#endif #include #include // NOLINT @@ -28,6 +24,7 @@ #include #include #include "paddle/fluid/inference/api/paddle_inference_api.h" +#include "paddle/fluid/platform/port.h" #include "paddle/fluid/string/printf.h" namespace paddle { diff --git a/paddle/fluid/inference/api/paddle_analysis_config.h b/paddle/fluid/inference/api/paddle_analysis_config.h index 2ac736df7ccd5..f05b9832da55f 100644 --- a/paddle/fluid/inference/api/paddle_analysis_config.h +++ b/paddle/fluid/inference/api/paddle_analysis_config.h @@ -16,6 +16,7 @@ #include #include #include +#include #include // Here we include some header files with relative paths, for that in deploy, @@ -51,19 +52,27 @@ struct AnalysisConfig : public NativeConfig { int max_batch_size = 1); bool use_tensorrt() const { return use_tensorrt_; } - // NOTE this is just for internal development, please not use it. - // NOT stable yet. void EnableMKLDNN(); bool use_mkldnn() const { return use_mkldnn_; } + void SetMKLDNNOp(std::unordered_set op_list) { + mkldnn_enabled_op_types_ = op_list; + } + + // Specify the memory buffer of program and parameter + void SetModelBuffer(const char* prog_buffer, size_t prog_buffer_size, + const char* program_buffer, size_t program_buffer_size); + bool model_from_memory() const { return model_from_memory_; } friend class ::paddle::AnalysisPredictor; protected: bool use_tensorrt_{false}; bool use_mkldnn_{false}; + std::unordered_set mkldnn_enabled_op_types_; int tensorrt_workspace_size_; int tensorrt_max_batchsize_; std::unique_ptr pass_builder_; + bool model_from_memory_{false}; }; // Configurations for Anakin engine. diff --git a/paddle/fluid/inference/api/paddle_api.h b/paddle/fluid/inference/api/paddle_api.h index 0a2a2a1a23401..1513a4b3b4f66 100644 --- a/paddle/fluid/inference/api/paddle_api.h +++ b/paddle/fluid/inference/api/paddle_api.h @@ -186,6 +186,19 @@ struct NativeConfig : public PaddlePredictor::Config { // Specify the variable's name of each input if input tensors don't follow the // `feeds` and `fetches` of the phase `save_inference_model`. bool specify_input_name{false}; + + // Set and get the number of cpu math library threads. + void SetCpuMathLibraryNumThreads(int cpu_math_library_num_threads) { + cpu_math_library_num_threads_ = cpu_math_library_num_threads; + } + int cpu_math_library_num_threads() const { + return cpu_math_library_num_threads_; + } + + protected: + // number of cpu math library (such as MKL, OpenBlas) threads for each + // instance. + int cpu_math_library_num_threads_{1}; }; // A factory to help create different predictors. diff --git a/paddle/fluid/inference/api/paddle_pass_builder.h b/paddle/fluid/inference/api/paddle_pass_builder.h index 12e3a6f42e140..40ca0d287ccde 100644 --- a/paddle/fluid/inference/api/paddle_pass_builder.h +++ b/paddle/fluid/inference/api/paddle_pass_builder.h @@ -98,9 +98,10 @@ class CpuPassStrategy : public PassStrategy { passes_.insert(passes_.begin(), "mkldnn_placement_pass"); for (auto &pass : - std::vector({"depthwise_conv_mkldnn_pass", // - "conv_bias_mkldnn_fuse_pass", // - "conv_relu_mkldnn_fuse_pass", // + std::vector({"depthwise_conv_mkldnn_pass", // + "conv_bias_mkldnn_fuse_pass", // + "conv3d_bias_mkldnn_fuse_pass", // + "conv_relu_mkldnn_fuse_pass", // "conv_elementwise_add_mkldnn_fuse_pass"})) { passes_.push_back(pass); } @@ -116,12 +117,12 @@ class CpuPassStrategy : public PassStrategy { class GpuPassStrategy : public PassStrategy { public: GpuPassStrategy() : PassStrategy({}) { - // TODO(NHZlX) Problem with Data synchronization between GPU and CPU - // When running in GPU mode, the parameters are all on GPU. But the - // opearations of "conv_bn_fuse_pass" are on CPU. passes_.assign({ - "infer_clean_graph_pass", - // "infer_clean_graph_pass", "conv_bn_fuse_pass", + "infer_clean_graph_pass", // + "conv_bn_fuse_pass", // + "conv_elementwise_add_act_fuse_pass", // + "conv_elementwise_add2_act_fuse_pass", // + "conv_elementwise_add_fuse_pass", // }); } diff --git a/paddle/fluid/inference/io.cc b/paddle/fluid/inference/io.cc index bb749e8f8b0ba..ae72a74acce82 100644 --- a/paddle/fluid/inference/io.cc +++ b/paddle/fluid/inference/io.cc @@ -69,7 +69,8 @@ bool IsPersistable(const framework::VarDesc* var) { void LoadPersistables(framework::Executor* executor, framework::Scope* scope, const framework::ProgramDesc& main_program, const std::string& dirname, - const std::string& param_filename) { + const std::string& param_filename, + bool model_from_memory = false) { const framework::BlockDesc& global_block = main_program.Block(0); framework::ProgramDesc* load_program = new framework::ProgramDesc(); @@ -78,7 +79,7 @@ void LoadPersistables(framework::Executor* executor, framework::Scope* scope, for (auto* var : global_block.AllVars()) { if (IsPersistable(var)) { - VLOG(30) << "persistable variable's name: " << var->Name(); + VLOG(4) << "persistable variable's name: " << var->Name(); framework::VarDesc* new_var = load_block->Var(var->Name()); new_var->SetShape(var->GetShape()); @@ -108,6 +109,7 @@ void LoadPersistables(framework::Executor* executor, framework::Scope* scope, op->SetType("load_combine"); op->SetOutput("Out", paramlist); op->SetAttr("file_path", {param_filename}); + op->SetAttr("model_from_memory", {model_from_memory}); op->CheckAttrs(); } @@ -121,7 +123,7 @@ std::unique_ptr Load(framework::Executor* executor, const std::string& dirname) { std::string model_filename = dirname + "/__model__"; std::string program_desc_str; - VLOG(30) << "loading model from " << model_filename; + VLOG(3) << "loading model from " << model_filename; ReadBinaryFile(model_filename, &program_desc_str); std::unique_ptr main_program( @@ -130,16 +132,17 @@ std::unique_ptr Load(framework::Executor* executor, "model version %ld is not supported.", main_program->Version()); - LoadPersistables(executor, scope, *main_program, dirname, ""); + // model_from_memory is false in seperate parameters. + LoadPersistables(executor, scope, *main_program, dirname, "", + false /* model_from_memory */); return main_program; } std::unique_ptr Load( framework::Executor* executor, framework::Scope* scope, const std::string& prog_filename, const std::string& param_filename) { - std::string model_filename = prog_filename; std::string program_desc_str; - ReadBinaryFile(model_filename, &program_desc_str); + ReadBinaryFile(prog_filename, &program_desc_str); std::unique_ptr main_program( new framework::ProgramDesc(program_desc_str)); @@ -147,7 +150,22 @@ std::unique_ptr Load( "model version %ld is not supported.", main_program->Version()); - LoadPersistables(executor, scope, *main_program, "", param_filename); + LoadPersistables(executor, scope, *main_program, "", param_filename, + false /* model_from_memory */); + return main_program; +} + +std::unique_ptr LoadFromMemory( + framework::Executor* executor, framework::Scope* scope, + const std::string& prog_buffer, const std::string& param_buffer) { + std::unique_ptr main_program( + new framework::ProgramDesc(prog_buffer)); + PADDLE_ENFORCE(framework::IsProgramVersionSupported(main_program->Version()), + "model version %ld is not supported.", + main_program->Version()); + + LoadPersistables(executor, scope, *main_program, "", param_buffer, + true /* model_filename */); return main_program; } diff --git a/paddle/fluid/inference/io.h b/paddle/fluid/inference/io.h index ab492577c1476..317ef9d93acf3 100644 --- a/paddle/fluid/inference/io.h +++ b/paddle/fluid/inference/io.h @@ -30,7 +30,8 @@ void Init(const std::vector argv); void LoadPersistables(framework::Executor* executor, framework::Scope* scope, const framework::ProgramDesc& main_program, const std::string& dirname, - const std::string& param_filename); + const std::string& param_filename, + bool model_from_memory); std::unique_ptr Load(framework::Executor* executor, framework::Scope* scope, @@ -41,6 +42,10 @@ std::unique_ptr Load(framework::Executor* executor, const std::string& prog_filename, const std::string& param_filename); +std::unique_ptr LoadFromMemory( + framework::Executor* executor, framework::Scope* scope, + const std::string& prog_buffer, const std::string& param_buffer); + // Save the variables from a scope to disk. void SaveVars(const framework::Scope& scope, const std::vector& vars, const std::string& dirname, diff --git a/paddle/fluid/inference/tensorrt/convert/op_converter.h b/paddle/fluid/inference/tensorrt/convert/op_converter.h index d61d635ed707b..91670ba8ac533 100644 --- a/paddle/fluid/inference/tensorrt/convert/op_converter.h +++ b/paddle/fluid/inference/tensorrt/convert/op_converter.h @@ -103,6 +103,7 @@ class OpConverter { void ConvertBlock(const framework::proto::BlockDesc& block, const std::unordered_set& parameters, const framework::Scope& scope, TensorRTEngine* engine) { + std::unique_lock lk(mut_); for (int i = 0; i < block.ops_size(); i++) { const auto& op = block.ops(i); ConvertOp(op, parameters, scope, engine); @@ -125,6 +126,7 @@ class OpConverter { std::unordered_map converters_; // fluid inference scope framework::Scope* scope_{nullptr}; + std::mutex mut_; }; } // namespace tensorrt diff --git a/paddle/fluid/inference/tensorrt/convert/pool2d_op.cc b/paddle/fluid/inference/tensorrt/convert/pool2d_op.cc index d700e08590ec5..1d0d83d1f368f 100644 --- a/paddle/fluid/inference/tensorrt/convert/pool2d_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/pool2d_op.cc @@ -53,7 +53,7 @@ class Pool2dOpConverter : public OpConverter { public: void operator()(const framework::proto::OpDesc &op, const framework::Scope &scope, bool test_mode) override { - VLOG(40) + VLOG(4) << "convert a fluid pool2d op to tensorrt pool2d layer without bias"; framework::OpDesc op_desc(op, nullptr); // Declare inputs @@ -109,8 +109,12 @@ class Pool2dOpConverter : public OpConverter { } if (pool_type == "max") { - nvinfer1::DimsHW pre_pad(paddings[0], paddings[1]); - nvinfer1::DimsHW post_pad(paddings[0], paddings[1]); + // Under ceil mode, the pre_pad and post_pad are used to + // record the the padding size. In some ceil mode cases, + // we do not need padding, so we initialize the two vars to 0. + + nvinfer1::DimsHW pre_pad(0, 0); + nvinfer1::DimsHW post_pad(0, 0); if (ceil_mode) { // If ceil mode is true, we will pad the appropriate size to the input. DealCeilMode(input_shape, ksize, strides, paddings, &pre_pad, &post_pad, diff --git a/paddle/fluid/inference/tensorrt/convert/split_op.cc b/paddle/fluid/inference/tensorrt/convert/split_op.cc index 6620c76318f99..ae5b1b98060a4 100644 --- a/paddle/fluid/inference/tensorrt/convert/split_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/split_op.cc @@ -19,9 +19,6 @@ namespace paddle { namespace inference { namespace tensorrt { -/* - * SplitOp. - */ class SplitOpConverter : public OpConverter { public: void operator()(const framework::proto::OpDesc& op, @@ -40,16 +37,11 @@ class SplitOpConverter : public OpConverter { int axis = boost::get(op_desc.GetAttr("axis")); std::vector output_lengths = boost::get>(op_desc.GetAttr("sections")); + // split on batch is not supported in TensorRT PADDLE_ENFORCE(axis != 0); - if (axis < 0) { - axis += input_dims.nbDims; - } else { - axis -= 1; - } + axis += (axis < 0) ? input_dims.nbDims : -1; PADDLE_ENFORCE(output_lengths.size() == output_num); - - // plugin::SplitPlugin* plugin = new plugin::SplitPlugin(axis, output_lengths); nvinfer1::IPluginLayer* layer = engine_->AddPlugin(&input, input_num, plugin); diff --git a/paddle/fluid/inference/tensorrt/convert/test_prelu_op.cc b/paddle/fluid/inference/tensorrt/convert/test_prelu_op.cc index 453f222f1f1e3..b086c910d38a2 100644 --- a/paddle/fluid/inference/tensorrt/convert/test_prelu_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/test_prelu_op.cc @@ -90,5 +90,4 @@ TEST(prelu_op, test_scalar) { } // namespace inference } // namespace paddle -// USE_OP(prelu); -USE_CPU_ONLY_OP(prelu); +USE_OP(prelu); diff --git a/paddle/fluid/inference/tensorrt/convert/test_split_op.cc b/paddle/fluid/inference/tensorrt/convert/test_split_op.cc index f81d011552c15..5aacc5c600dd1 100644 --- a/paddle/fluid/inference/tensorrt/convert/test_split_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/test_split_op.cc @@ -20,30 +20,92 @@ namespace paddle { namespace inference { namespace tensorrt { -TEST(split_op, test) { +template +void TensorRTSplitTest(const std::vector &in_shape, + const std::vector §ions) { std::unordered_set parameters({""}); framework::Scope scope; - TRTConvertValidation validator(10, parameters, scope, 1000); - validator.DeclInputVar("split_input", nvinfer1::DimsCHW(3, 2, 2)); - validator.DeclOutputVar("split_out1", nvinfer1::DimsCHW(2, 2, 2)); - validator.DeclOutputVar("split_out2", nvinfer1::DimsCHW(1, 2, 2)); + TRTConvertValidation validator(BatchSize + 1, parameters, scope, 10000); + + auto make_dim = [](const std::vector &shape) { + nvinfer1::DimsCHW dim; + dim.c() = shape[0]; + dim.h() = shape[1]; + dim.w() = shape[2]; + return dim; + }; + validator.DeclInputVar("split_input", make_dim(in_shape)); + std::vector output_vars; + for (size_t i = 0; i < sections.size(); ++i) { + auto out_shape = in_shape; + out_shape[Axis - 1] = sections[i]; + std::string output_name = "split_out" + std::to_string(i); + validator.DeclOutputVar(output_name, make_dim(out_shape)); + output_vars.push_back(output_name); + } // Prepare Op description framework::OpDesc desc; desc.SetType("split"); desc.SetInput("X", {"split_input"}); - desc.SetOutput("Out", {"split_out1", "split_out2"}); + desc.SetOutput("Out", output_vars); - int num = 0; - int axis = 1; - std::vector output_lengths = {2, 1}; - desc.SetAttr("axis", axis); - desc.SetAttr("num", num); - desc.SetAttr("sections", output_lengths); + desc.SetAttr("axis", Axis); + desc.SetAttr("num", 0); + desc.SetAttr("sections", sections); validator.SetOp(*desc.Proto()); - validator.Execute(1); + validator.Execute(BatchSize); +} + +// batch = 0, axis = 1, same shape +TEST(split_op, test_same_shape_axis1_batch1) { + TensorRTSplitTest<1, 1>({4, 2, 2}, {2, 2}); +} +// batch = 0, axis = 1, different shape +TEST(split_op, test_different_shape_axis1_batch1) { + TensorRTSplitTest<1, 1>({3, 2, 2}, {2, 1}); +} +// batch = 10, axis = 1, same shape +TEST(split_op, test_same_shape_axis1_batch10) { + TensorRTSplitTest<10, 1>({4, 2, 2}, {2, 2}); +} +// batch = 10, axis = 1, different shape +TEST(split_op, test_different_shape_axis1_batch10) { + TensorRTSplitTest<10, 1>({3, 2, 2}, {2, 1}); +} +// batch = 0, axis = 2, same shape +TEST(split_op, test_same_shape_axis2_batch1) { + TensorRTSplitTest<1, 2>({3, 4, 2}, {2, 2}); +} +// batch = 0, axis = 2, different shape +TEST(split_op, test_different_shape_axis2_batch1) { + TensorRTSplitTest<1, 2>({3, 3, 2}, {2, 1}); +} +// batch = 10, axis = 2, same shape +TEST(split_op, test_same_shape_axis2_batch10) { + TensorRTSplitTest<10, 2>({3, 4, 2}, {2, 2}); +} +// batch = 10, axis = 2, different shape +TEST(split_op, test_different_shape_axis2_batch10) { + TensorRTSplitTest<10, 2>({3, 3, 2}, {2, 1}); +} +// batch = 0, axis = 3, same shape +TEST(split_op, test_same_shape_axis3_batch1) { + TensorRTSplitTest<1, 3>({3, 2, 4}, {2, 2}); +} +// batch = 0, axis = 3, different shape +TEST(split_op, test_different_shape_axis3_batch1) { + TensorRTSplitTest<1, 3>({3, 2, 3}, {2, 1}); +} +// batch = 10, axis = 3, same shape +TEST(split_op, test_same_shape_axis3_batch10) { + TensorRTSplitTest<10, 3>({3, 2, 4}, {2, 2}); +} +// batch = 10, axis = 3, different shape +TEST(split_op, test_different_shape_axis3_batch10) { + TensorRTSplitTest<10, 3>({3, 2, 3}, {2, 1}); } } // namespace tensorrt diff --git a/paddle/fluid/inference/tensorrt/plugin/CMakeLists.txt b/paddle/fluid/inference/tensorrt/plugin/CMakeLists.txt index e822785ad6f4f..95443e813327c 100644 --- a/paddle/fluid/inference/tensorrt/plugin/CMakeLists.txt +++ b/paddle/fluid/inference/tensorrt/plugin/CMakeLists.txt @@ -1,4 +1,4 @@ nv_library(tensorrt_plugin SRCS trt_plugin.cc split_op_plugin.cu elementwise_op_plugin.cu prelu_op_plugin.cu avg_pool_op_plugin.cu - DEPS enforce tensorrt_engine) + DEPS enforce tensorrt_engine prelu) diff --git a/paddle/fluid/inference/tensorrt/plugin/prelu_op_plugin.cu b/paddle/fluid/inference/tensorrt/plugin/prelu_op_plugin.cu index e8f4254402a5d..3075e87ea6d71 100644 --- a/paddle/fluid/inference/tensorrt/plugin/prelu_op_plugin.cu +++ b/paddle/fluid/inference/tensorrt/plugin/prelu_op_plugin.cu @@ -14,92 +14,16 @@ #include #include +#include #include "glog/logging.h" #include "paddle/fluid/inference/tensorrt/plugin/prelu_op_plugin.h" +#include "paddle/fluid/operators/math/prelu.h" namespace paddle { namespace inference { namespace tensorrt { namespace plugin { -static const int CUDA_NUM_THREADS = 1024; -static const int CUDA_MAX_NUM_BLOCKS = 65535; -inline static int GET_NUM_BLOCKS(const int N) { - return (N + CUDA_NUM_THREADS - 1) / CUDA_NUM_THREADS; -} - -__global__ void PReluChannelWiseKernel(const float *input, const float *alpha, - float *output, int channel, - size_t spatial_size) { - size_t offset = blockIdx.x * spatial_size; - const float *in = input + offset; - float *out = output + offset; - float scale = alpha[blockIdx.x % channel]; - - for (size_t i = threadIdx.x; i < spatial_size; i += blockDim.x) { - float x = in[i]; - out[i] = (x > 0) ? x : scale * x; - } -} - -__global__ void PReluElementWiseKernel(const float *input, const float *alpha, - float *output, size_t spatial_size) { - size_t offset = blockIdx.x * spatial_size; - const float *in = input + offset; - const float *scale = alpha + offset; - float *out = output + offset; - - for (size_t i = threadIdx.x; i < spatial_size; i += blockDim.x) { - float x = in[i]; - out[i] = (x > 0) ? x : scale[i] * x; - } -} - -__global__ void PReluScalarKernel(const float *input, const float *alpha, - float *output, size_t spatial_size) { - size_t offset = blockIdx.x * spatial_size; - const float *in = input + offset; - float scale = *alpha; - float *out = output + offset; - - for (size_t i = threadIdx.x; i < spatial_size; i += blockDim.x) { - float x = in[i]; - out[i] = (x > 0) ? x : scale * x; - } -} - -static inline void PReluChannelWise(cudaStream_t stream, const float *input, - const float *alpha, float *output, - int batch_size, - const nvinfer1::Dims &dims) { - size_t unroll = batch_size * dims.d[0]; - size_t spatial_size = dims.d[1] * dims.d[2]; - CHECK_LT(unroll, CUDA_MAX_NUM_BLOCKS); - PReluChannelWiseKernel<<>>( - input, alpha, output, dims.d[0], spatial_size); -} - -static inline void PReluElementWise(cudaStream_t stream, const float *input, - const float *alpha, float *output, - int batch_size, - const nvinfer1::Dims &dims) { - size_t unroll = batch_size * dims.d[0]; - size_t spatial_size = dims.d[1] * dims.d[2]; - CHECK_LT(unroll, CUDA_MAX_NUM_BLOCKS); - PReluElementWiseKernel<<>>( - input, alpha, output, spatial_size); -} - -static inline void PReluScalar(cudaStream_t stream, const float *input, - const float *alpha, float *output, - int batch_size, const nvinfer1::Dims &dims) { - size_t unroll = batch_size * dims.d[0]; - size_t spatial_size = dims.d[1] * dims.d[2]; - CHECK_LT(unroll, CUDA_MAX_NUM_BLOCKS); - PReluScalarKernel<<>>( - input, alpha, output, spatial_size); -} - nvinfer1::Dims PReluPlugin::getOutputDimensions(int index, const nvinfer1::Dims *inputDims, int nbInputs) { @@ -110,19 +34,31 @@ nvinfer1::Dims PReluPlugin::getOutputDimensions(int index, return output_dims; } -int PReluPlugin::enqueue(int batchSize, const void *const *inputs, +int PReluPlugin::enqueue(int batch_size, const void *const *inputs, void **outputs, void *workspace, cudaStream_t stream) { // input dims is CHW. const auto &input_dims = this->getInputDims(0); const float *input = reinterpret_cast(inputs[0]); const float *alpha = reinterpret_cast(alpha_.get().values); float *output = reinterpret_cast(outputs)[0]; + + std::vector input_shape; + input_shape.push_back(batch_size); + for (int i = 0; i < input_dims.nbDims; i++) { + input_shape.push_back(input_dims.d[i]); + } + if (mode_ == "channel") { - PReluChannelWise(stream, input, alpha, output, batchSize, input_dims); + operators::math::PreluChannelWiseDirectCUDAFunctor + prelu_channel_wise; + prelu_channel_wise(stream, input, alpha, output, input_shape); } else if (mode_ == "element") { - PReluElementWise(stream, input, alpha, output, batchSize, input_dims); + operators::math::PreluElementWiseDirectCUDAFunctor + prelu_element_wise; + prelu_element_wise(stream, input, alpha, output, input_shape); } else { - PReluScalar(stream, input, alpha, output, batchSize, input_dims); + operators::math::PreluScalarDirectCUDAFunctor prelu_scalar; + prelu_scalar(stream, input, alpha, output, input_shape); } return cudaGetLastError() != cudaSuccess; } diff --git a/paddle/fluid/inference/tensorrt/plugin/split_op_plugin.cu b/paddle/fluid/inference/tensorrt/plugin/split_op_plugin.cu index 4adea2db1ee80..de61ace59e299 100644 --- a/paddle/fluid/inference/tensorrt/plugin/split_op_plugin.cu +++ b/paddle/fluid/inference/tensorrt/plugin/split_op_plugin.cu @@ -12,6 +12,8 @@ // See the License for the specific language governing permissions and // limitations under the License. +#include +#include #include "paddle/fluid/inference/tensorrt/plugin/split_op_plugin.h" namespace paddle { @@ -19,6 +21,52 @@ namespace inference { namespace tensorrt { namespace plugin { +// copied from operators::math::SplitFunctor +template +__global__ void SplitKernel(const T* input_data, const int in_row, + const int in_col, const int* out_cols, + int out_cols_size, T** outputs_data) { + int tid_x = blockIdx.x * blockDim.x + threadIdx.x; + int curr_segment = 0; + int curr_offset = out_cols[0]; + for (; tid_x < in_col; tid_x += blockDim.x * gridDim.x) { + int curr_col_offset = out_cols[curr_segment + 1]; + while (curr_col_offset <= tid_x) { + curr_offset = curr_col_offset; + ++curr_segment; + curr_col_offset = out_cols[curr_segment + 1]; + } + + int local_col = tid_x - curr_offset; + int segment_width = curr_col_offset - curr_offset; + T* output_ptr = outputs_data[curr_segment]; + if (output_ptr != nullptr) { + int tid_y = blockIdx.y * blockDim.y + threadIdx.y; + for (; tid_y < in_row; tid_y += blockDim.y * gridDim.y) + output_ptr[tid_y * segment_width + local_col] = + input_data[tid_y * in_col + tid_x]; + } + } +} + +template +__global__ void SplitKernel(const T* input_data, const int in_row, + const int in_col, const int fixed_out_col, + T** outputs_data) { + int tid_x = blockIdx.x * blockDim.x + threadIdx.x; + for (; tid_x < in_col; tid_x += blockDim.x * gridDim.x) { + int split = tid_x / fixed_out_col; + int in_offset = tid_x - split * fixed_out_col; + T* output_ptr = outputs_data[split]; + if (output_ptr != nullptr) { + int tid_y = blockIdx.y * blockDim.y + threadIdx.y; + for (; tid_y < in_row; tid_y += blockDim.y * gridDim.y) + output_ptr[tid_y * fixed_out_col + in_offset] = + input_data[tid_y * in_col + tid_x]; + } + } +} + nvinfer1::Dims SplitPlugin::getOutputDimensions( int index, const nvinfer1::Dims* input_dims, int num_inputs) { PADDLE_ENFORCE_EQ(num_inputs, 1); @@ -31,48 +79,96 @@ nvinfer1::Dims SplitPlugin::getOutputDimensions( int SplitPlugin::initialize() { PADDLE_ENFORCE_LE(axis_, nvinfer1::Dims::MAX_DIMS); - + // notice input dims is [C, H, W] + nvinfer1::Dims dims = this->getInputDims(0); + outer_rows_ = 1; + inner_cols_ = 1; + for (int i = 0; i < axis_; ++i) { + outer_rows_ *= dims.d[i]; + } + for (int i = axis_ + 1; i < dims.nbDims; ++i) { + inner_cols_ *= dims.d[i]; + } + same_shape_ = true; std::vector segment_offsets(1, 0); for (int i = 0; i < this->getNbOutputs(); ++i) { - segment_offsets.push_back(segment_offsets.back() + output_length_[i]); + if (output_length_[i] != output_length_[0]) { + same_shape_ = false; + } + segment_offsets.push_back(segment_offsets.back() + + output_length_[i] * inner_cols_); } - segment_offsets_ = segment_offsets; - nvinfer1::Dims dims = this->getInputDims(0); - nx_ = 1; - for (int i = dims.nbDims - 1; i > axis_; --i) { - nx_ *= dims.d[i]; + inner_cols_ *= dims.d[axis_]; + d_segment_offsets_ = segment_offsets; + segment_offsets_ = std::move(segment_offsets); + d_output_ptrs_.resize(this->getNbOutputs(), nullptr); + return 0; +} + +template +inline void Split(cudaStream_t stream, const bool same_shape, + const int outer_rows, const int inner_cols, + const std::vector& segment_offsets, + const int* d_segment_offsets, const T* input, T** outputs) { + const int kThreadsPerBlock = 1024; + const int kMaxBlocks = 65535; + int block_cols = kThreadsPerBlock; + if (inner_cols < kThreadsPerBlock) { // block_cols is aligned by 32. + block_cols = ((inner_cols + 31) >> 5) << 5; } - ny_ = dims.d[axis_]; - nz_ = 1; - for (int i = axis_ - 1; i >= 0; --i) { - nz_ *= dims.d[i]; + int block_rows = kThreadsPerBlock / block_cols; + dim3 block_size = dim3(block_cols, block_rows, 1); + + int grid_cols = + std::min((inner_cols + block_cols - 1) / block_cols, kMaxBlocks); + int grid_rows = + std::min(kMaxBlocks / grid_cols, std::max(outer_rows / block_rows, 1)); + dim3 grid_size = dim3(grid_cols, grid_rows, 1); + + if (same_shape) { + SplitKernel<<>>( + input, outer_rows, inner_cols, segment_offsets[1], outputs); + } else { + SplitKernel<<>>( + input, outer_rows, inner_cols, d_segment_offsets, + static_cast(segment_offsets.size()), outputs); } - return 0; } int SplitPlugin::enqueue(int batchSize, const void* const* inputs, void** outputs, void* workspace, cudaStream_t stream) { - auto const& input_dims = this->getInputDims(0); - int input_size = 0; - float const* idata = reinterpret_cast(inputs[0]); - float** odatas = reinterpret_cast(outputs); - - // kernel impl here. - int inputBatchOffset = nx_ * ny_ * nz_; - for (size_t i = 0; i < this->getNbOutputs(); i++) { - for (size_t j = 0; j < batchSize; j++) { - cudaMemcpyAsync( - odatas[i] + - j * (segment_offsets_[i + 1] - segment_offsets_[i]) * nx_ * - sizeof(float), - inputs[0] + - (inputBatchOffset * j + segment_offsets_[i] * nx_) * - sizeof(float), - (segment_offsets_[i + 1] - segment_offsets_[i]) * nx_ * sizeof(float), - cudaMemcpyDeviceToDevice, stream); + float const* input_ptr = reinterpret_cast(inputs[0]); + if (((batchSize == 1 && axis_ == 0) || axis_ == -1) && + this->getNbOutputs() < 10) { + float** output_ptrs = reinterpret_cast(outputs); + int data_type_size = (this->getDataType() == nvinfer1::DataType::kFLOAT) + ? sizeof(float) + : sizeof(__half); + for (int i = 0; i < this->getNbOutputs(); ++i) { + PADDLE_ENFORCE( + cudaMemcpyAsync( + output_ptrs[i], input_ptr + segment_offsets_[i], + (segment_offsets_[i + 1] - segment_offsets_[i]) * data_type_size, + cudaMemcpyDeviceToDevice, stream) == cudaSuccess); + } + } else { + outer_rows_ *= batchSize; + const int* d_segment_offsets_ptr = + thrust::raw_pointer_cast(&d_segment_offsets_[0]); + float** output_ptrs = thrust::raw_pointer_cast(&d_output_ptrs_[0]); + PADDLE_ENFORCE(cudaMemcpyAsync(output_ptrs, outputs, + this->getNbOutputs() * sizeof(float*), + cudaMemcpyHostToDevice, + stream) == cudaSuccess); + if (this->getDataType() == nvinfer1::DataType::kFLOAT) { + Split(stream, same_shape_, outer_rows_, inner_cols_, segment_offsets_, + d_segment_offsets_ptr, input_ptr, output_ptrs); + } else { + Split(stream, same_shape_, outer_rows_, inner_cols_, segment_offsets_, + d_segment_offsets_ptr, (__half*)input_ptr, // NOLINT + (__half**)output_ptrs); // NOLINT } } - return cudaGetLastError() != cudaSuccess; } diff --git a/paddle/fluid/inference/tensorrt/plugin/split_op_plugin.h b/paddle/fluid/inference/tensorrt/plugin/split_op_plugin.h index b5b6e69992b05..6f028d3d72ae3 100644 --- a/paddle/fluid/inference/tensorrt/plugin/split_op_plugin.h +++ b/paddle/fluid/inference/tensorrt/plugin/split_op_plugin.h @@ -14,6 +14,7 @@ #pragma once +#include #include #include "paddle/fluid/inference/tensorrt/plugin/trt_plugin.h" @@ -25,7 +26,7 @@ namespace plugin { class SplitPlugin : public PluginTensorRT { public: SplitPlugin(int axis, std::vector const &output_lengths) - : axis_(axis), output_length_(output_lengths) {} + : axis_(axis), same_shape_(true), output_length_(output_lengths) {} SplitPlugin(void const *serial_data, size_t serial_length) { deserializeBase(serial_data, serial_length); @@ -60,9 +61,13 @@ class SplitPlugin : public PluginTensorRT { } int axis_; + int outer_rows_; + int inner_cols_; + bool same_shape_; std::vector output_length_; - int nx_, ny_, nz_; std::vector segment_offsets_; + thrust::device_vector d_segment_offsets_; + thrust::device_vector d_output_ptrs_; }; } // namespace plugin diff --git a/paddle/fluid/inference/tests/api/CMakeLists.txt b/paddle/fluid/inference/tests/api/CMakeLists.txt index e8bd13037ed6c..46ce61b73611d 100644 --- a/paddle/fluid/inference/tests/api/CMakeLists.txt +++ b/paddle/fluid/inference/tests/api/CMakeLists.txt @@ -1,4 +1,4 @@ -set(INFERENCE_EXTRA_DEPS paddle_inference_api paddle_fluid_api ir_pass_manager analysis_predictor) +set(INFERENCE_EXTRA_DEPS paddle_inference_api paddle_fluid_api ir_pass_manager analysis_predictor benchmark) if(WITH_GPU AND TENSORRT_FOUND) set(INFERENCE_EXTRA_DEPS ${INFERENCE_EXTRA_DEPS} analysis ${analysis_deps} ir_pass_manager analysis_predictor) @@ -30,6 +30,13 @@ function(inference_analysis_api_test_with_fake_data target install_dir filename ARGS --infer_model=${install_dir}/model) endfunction() +function(inference_analysis_api_test_with_refer_result target install_dir filename) + inference_analysis_test(${target} SRCS ${filename} + EXTRA_DEPS ${INFERENCE_EXTRA_DEPS} + ARGS --infer_model=${install_dir}/model --infer_data=${install_dir}/data.txt + --refer_result=${install_dir}/result.txt) +endfunction() + # RNN1 if(NOT APPLE AND WITH_MKLML) set(RNN1_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/rnn1") @@ -46,11 +53,18 @@ set(RNN2_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/rnn2") download_model_and_data(${RNN2_INSTALL_DIR} "rnn2_model.tar.gz" "rnn2_data.txt.tar.gz") inference_analysis_api_test(test_analyzer_rnn2 ${RNN2_INSTALL_DIR} analyzer_rnn2_tester.cc) -# DAM +# normal DAM set(DAM_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/dam") download_model_and_data(${DAM_INSTALL_DIR} "DAM_model.tar.gz" "DAM_data.txt.tar.gz") inference_analysis_api_test(test_analyzer_dam ${DAM_INSTALL_DIR} analyzer_dam_tester.cc) +# small DAM +set(DAM_SMALL_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/small_dam") +download_model_and_data(${DAM_SMALL_INSTALL_DIR} "dam_small_model.tar.gz" "dam_small_data.txt.tar.gz") +inference_analysis_test(test_analyzer_small_dam SRCS analyzer_dam_tester.cc + EXTRA_DEPS ${INFERENCE_EXTRA_DEPS} + ARGS --infer_model=${DAM_SMALL_INSTALL_DIR}/model --infer_data=${DAM_SMALL_INSTALL_DIR}/data.txt --max_turn_num=1) + # chinese_ner set(CHINESE_NER_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/chinese_ner") download_model_and_data(${CHINESE_NER_INSTALL_DIR} "chinese_ner_model.tar.gz" "chinese_ner-data.txt.tar.gz") @@ -74,45 +88,52 @@ inference_analysis_api_test(test_analyzer_seq_conv1 ${SEQ_CONV1_INSTALL_DIR} ana # ocr set(OCR_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/ocr") if (NOT EXISTS ${OCR_INSTALL_DIR}) - inference_download_and_uncompress(${OCR_INSTALL_DIR} "http://paddlemodels.cdn.bcebos.com/" "inference-vis-demos%2Focr.tar.gz") + inference_download_and_uncompress(${OCR_INSTALL_DIR} "http://paddlemodels.cdn.bcebos.com/" "inference-vis-demos%2Focr.tar.gz") +endif() +inference_analysis_api_test_with_refer_result(test_analyzer_ocr ${OCR_INSTALL_DIR} analyzer_vis_tester.cc) + +# mobilenet with transpose op +set(MOBILENET_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/mobilenet") +if (NOT EXISTS ${MOBILENET_INSTALL_DIR}) + inference_download_and_uncompress(${MOBILENET_INSTALL_DIR} "http://paddlemodels.cdn.bcebos.com/" "inference-vis-demos%2Fmobilenet.tar.gz") endif() -inference_analysis_api_test(test_analyzer_ocr ${OCR_INSTALL_DIR} analyzer_vis_tester.cc) +inference_analysis_api_test_with_refer_result(test_analyzer_mobilenet_transpose ${MOBILENET_INSTALL_DIR} analyzer_vis_tester.cc) # resnet50 inference_analysis_api_test_with_fake_data(test_analyzer_resnet50 "${INFERENCE_DEMO_INSTALL_DIR}/resnet50" analyzer_resnet50_tester.cc "resnet50_model.tar.gz") # mobilenet with depthwise_conv op -inference_analysis_api_test_with_fake_data(test_analyzer_mobilenet +inference_analysis_api_test_with_fake_data(test_analyzer_mobilenet_depthwise_conv "${INFERENCE_DEMO_INSTALL_DIR}/mobilenet_depthwise_conv" analyzer_resnet50_tester.cc "mobilenet_model.tar.gz") # anakin if (WITH_ANAKIN AND WITH_MKL) # only needed in CI - # anakin rnn1 - set(ANAKIN_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/anakin") - set(ANAKIN_RNN1_INSTALL_DIR "${ANAKIN_INSTALL_DIR}/rnn1") - inference_download(${ANAKIN_RNN1_INSTALL_DIR} ${INFERENCE_URL} "anakin_test%2Fditu_rnn.anakin2.model.bin") - inference_download(${ANAKIN_RNN1_INSTALL_DIR} ${INFERENCE_URL} "anakin_test%2Fditu_rnn_data.txt") - cc_test(test_anakin_rnn1 SRCS anakin_rnn1_tester.cc - ARGS --model=${ANAKIN_RNN1_INSTALL_DIR}/anakin_test%2Fditu_rnn.anakin2.model.bin - --datapath=${ANAKIN_RNN1_INSTALL_DIR}/anakin_test%2Fditu_rnn_data.txt - DEPS inference_anakin_api_shared SERIAL) - # anakin mobilenet - if(WITH_GPU) - set(ANAKIN_MOBILENET_INSTALL_DIR "${ANAKIN_INSTALL_DIR}/mobilenet") - inference_download(${ANAKIN_MOBILENET_INSTALL_DIR} ${INFERENCE_URL} "mobilenet_v2.anakin.bin") - cc_test(test_anakin_mobilenet SRCS anakin_mobilenet_tester.cc - ARGS --model=${ANAKIN_MOBILENET_INSTALL_DIR}/mobilenet_v2.anakin.bin - DEPS inference_anakin_api_shared dynload_cuda SERIAL) - endif() + # anakin rnn1 + set(ANAKIN_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/anakin") + set(ANAKIN_RNN1_INSTALL_DIR "${ANAKIN_INSTALL_DIR}/rnn1") + inference_download(${ANAKIN_RNN1_INSTALL_DIR} ${INFERENCE_URL} "anakin_test%2Fditu_rnn.anakin2.model.bin") + inference_download(${ANAKIN_RNN1_INSTALL_DIR} ${INFERENCE_URL} "anakin_test%2Fditu_rnn_data.txt") + cc_test(test_anakin_rnn1 SRCS anakin_rnn1_tester.cc + ARGS --model=${ANAKIN_RNN1_INSTALL_DIR}/anakin_test%2Fditu_rnn.anakin2.model.bin + --datapath=${ANAKIN_RNN1_INSTALL_DIR}/anakin_test%2Fditu_rnn_data.txt + DEPS inference_anakin_api_shared SERIAL) + # anakin mobilenet + if(WITH_GPU) + set(ANAKIN_MOBILENET_INSTALL_DIR "${ANAKIN_INSTALL_DIR}/mobilenet") + inference_download(${ANAKIN_MOBILENET_INSTALL_DIR} ${INFERENCE_URL} "mobilenet_v2.anakin.bin") + cc_test(test_anakin_mobilenet SRCS anakin_mobilenet_tester.cc + ARGS --model=${ANAKIN_MOBILENET_INSTALL_DIR}/mobilenet_v2.anakin.bin + DEPS inference_anakin_api_shared dynload_cuda SERIAL) + endif() endif() if(WITH_GPU AND TENSORRT_FOUND) - set(TRT_MODEL_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/trt") - if (NOT EXISTS ${TRT_MODEL_INSTALL_DIR}) - inference_download_and_uncompress(${TRT_MODEL_INSTALL_DIR} ${INFERENCE_URL}/tensorrt_test "trt_test_models.tar.gz") - endif() - inference_analysis_test(test_trt_models SRCS trt_models_tester.cc - EXTRA_DEPS ${INFERENCE_EXTRA_DEPS} - ARGS --infer_model=${TRT_MODEL_INSTALL_DIR}/trt_test_models SERIAL) + set(TRT_MODEL_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/trt") + if (NOT EXISTS ${TRT_MODEL_INSTALL_DIR}) + inference_download_and_uncompress(${TRT_MODEL_INSTALL_DIR} ${INFERENCE_URL}/tensorrt_test "trt_test_models.tar.gz") + endif() + inference_analysis_test(test_trt_models SRCS trt_models_tester.cc + EXTRA_DEPS ${INFERENCE_EXTRA_DEPS} + ARGS --infer_model=${TRT_MODEL_INSTALL_DIR}/trt_test_models SERIAL) endif() diff --git a/paddle/fluid/inference/tests/api/anakin_rnn1_tester.cc b/paddle/fluid/inference/tests/api/anakin_rnn1_tester.cc index c4022225fd452..da42688f29f04 100644 --- a/paddle/fluid/inference/tests/api/anakin_rnn1_tester.cc +++ b/paddle/fluid/inference/tests/api/anakin_rnn1_tester.cc @@ -13,7 +13,6 @@ See the License for the specific language governing permissions and limitations under the License. */ #include -#include #include #include #include diff --git a/paddle/fluid/inference/tests/api/analyzer_dam_tester.cc b/paddle/fluid/inference/tests/api/analyzer_dam_tester.cc index b369cba5c8b3f..12d61d06ce188 100644 --- a/paddle/fluid/inference/tests/api/analyzer_dam_tester.cc +++ b/paddle/fluid/inference/tests/api/analyzer_dam_tester.cc @@ -14,38 +14,54 @@ #include "paddle/fluid/inference/tests/api/tester_helper.h" +DEFINE_int32(max_turn_num, 9, + "The max turn number: 1 for the small and 9 for the normal."); + namespace paddle { namespace inference { using contrib::AnalysisConfig; -#define MAX_TURN_NUM 9 -#define MAX_TURN_LEN 50 + +constexpr int32_t kMaxTurnLen = 50; + static std::vector result_data; struct DataRecord { - std::vector> - turns[MAX_TURN_NUM]; // turns data : MAX_TURN_NUM - std::vector> - turns_mask[MAX_TURN_NUM]; // turns mask data : MAX_TURN_NUM - std::vector> response; // response data : 1 + std::vector> *turns; + std::vector> *turns_mask; + std::vector> response; // response data : 1 std::vector> response_mask; // response mask data : 1 size_t batch_iter{0}; size_t batch_size{1}; size_t num_samples; // total number of samples - DataRecord() = default; + + DataRecord() { + turns = new std::vector>[FLAGS_max_turn_num]; // turns data : FLAGS_max_turn_num + turns_mask = new std::vector>[FLAGS_max_turn_num]; // turns mask data : FLAGS_max_turn_num + } + explicit DataRecord(const std::string &path, int batch_size = 1) - : batch_size(batch_size) { + : DataRecord() { + this->batch_size = batch_size; Load(path); } + + ~DataRecord() { + delete[] turns; + delete[] turns_mask; + } + DataRecord NextBatch() { DataRecord data; size_t batch_end = batch_iter + batch_size; // NOTE skip the final batch, if no enough data is provided. if (batch_end <= response.size()) { - for (int i = 0; i < MAX_TURN_NUM; ++i) { + for (int i = 0; i < FLAGS_max_turn_num; ++i) { data.turns[i].assign(turns[i].begin() + batch_iter, turns[i].begin() + batch_end); } - for (int i = 0; i < MAX_TURN_NUM; ++i) { + for (int i = 0; i < FLAGS_max_turn_num; ++i) { data.turns_mask[i].assign(turns_mask[i].begin() + batch_iter, turns_mask[i].begin() + batch_end); } @@ -60,6 +76,7 @@ struct DataRecord { batch_iter += batch_size; return data; } + void Load(const std::string &path) { std::ifstream file(path); std::string line; @@ -69,30 +86,30 @@ struct DataRecord { num_lines++; std::vector data; split(line, ',', &data); - CHECK_EQ(data.size(), (size_t)(2 * MAX_TURN_NUM + 3)); + CHECK_EQ(data.size(), (size_t)(2 * FLAGS_max_turn_num + 3)); // load turn data - std::vector turns_tmp[MAX_TURN_NUM]; - for (int i = 0; i < MAX_TURN_NUM; ++i) { + std::vector turns_tmp[FLAGS_max_turn_num]; + for (int i = 0; i < FLAGS_max_turn_num; ++i) { split_to_int64(data[i], ' ', &turns_tmp[i]); turns[i].push_back(std::move(turns_tmp[i])); } // load turn_mask data - std::vector turns_mask_tmp[MAX_TURN_NUM]; - for (int i = 0; i < MAX_TURN_NUM; ++i) { - split_to_float(data[MAX_TURN_NUM + i], ' ', &turns_mask_tmp[i]); + std::vector turns_mask_tmp[FLAGS_max_turn_num]; + for (int i = 0; i < FLAGS_max_turn_num; ++i) { + split_to_float(data[FLAGS_max_turn_num + i], ' ', &turns_mask_tmp[i]); turns_mask[i].push_back(std::move(turns_mask_tmp[i])); } // load response data std::vector response_tmp; - split_to_int64(data[2 * MAX_TURN_NUM], ' ', &response_tmp); + split_to_int64(data[2 * FLAGS_max_turn_num], ' ', &response_tmp); response.push_back(std::move(response_tmp)); // load response_mask data std::vector response_mask_tmp; - split_to_float(data[2 * MAX_TURN_NUM + 1], ' ', &response_mask_tmp); + split_to_float(data[2 * FLAGS_max_turn_num + 1], ' ', &response_mask_tmp); response_mask.push_back(std::move(response_mask_tmp)); // load result data float result_tmp; - result_tmp = std::stof(data[2 * MAX_TURN_NUM + 2]); + result_tmp = std::stof(data[2 * FLAGS_max_turn_num + 2]); result_data.push_back(result_tmp); } num_samples = num_lines; @@ -101,8 +118,8 @@ struct DataRecord { void PrepareInputs(std::vector *input_slots, DataRecord *data, int batch_size) { - PaddleTensor turns_tensor[MAX_TURN_NUM]; - PaddleTensor turns_mask_tensor[MAX_TURN_NUM]; + PaddleTensor turns_tensor[FLAGS_max_turn_num]; + PaddleTensor turns_mask_tensor[FLAGS_max_turn_num]; PaddleTensor response_tensor; PaddleTensor response_mask_tensor; std::string turn_pre = "turn_"; @@ -110,16 +127,16 @@ void PrepareInputs(std::vector *input_slots, DataRecord *data, auto one_batch = data->NextBatch(); int size = one_batch.response[0].size(); - CHECK_EQ(size, MAX_TURN_LEN); + CHECK_EQ(size, kMaxTurnLen); // turn tensor assignment - for (int i = 0; i < MAX_TURN_NUM; ++i) { + for (int i = 0; i < FLAGS_max_turn_num; ++i) { turns_tensor[i].name = turn_pre + std::to_string(i); turns_tensor[i].shape.assign({batch_size, size, 1}); turns_tensor[i].dtype = PaddleDType::INT64; TensorAssignData(&turns_tensor[i], one_batch.turns[i]); } // turn mask tensor assignment - for (int i = 0; i < MAX_TURN_NUM; ++i) { + for (int i = 0; i < FLAGS_max_turn_num; ++i) { turns_mask_tensor[i].name = turn_mask_pre + std::to_string(i); turns_mask_tensor[i].shape.assign({batch_size, size, 1}); turns_mask_tensor[i].dtype = PaddleDType::FLOAT32; @@ -137,10 +154,10 @@ void PrepareInputs(std::vector *input_slots, DataRecord *data, TensorAssignData(&response_mask_tensor, one_batch.response_mask); // Set inputs. - for (int i = 0; i < MAX_TURN_NUM; ++i) { + for (int i = 0; i < FLAGS_max_turn_num; ++i) { input_slots->push_back(std::move(turns_tensor[i])); } - for (int i = 0; i < MAX_TURN_NUM; ++i) { + for (int i = 0; i < FLAGS_max_turn_num; ++i) { input_slots->push_back(std::move(turns_mask_tensor[i])); } input_slots->push_back(std::move(response_tensor)); @@ -171,10 +188,16 @@ void SetInput(std::vector> *inputs) { } // Easy for profiling independently. -TEST(Analyzer_dam, profile) { +void profile(bool use_mkldnn = false) { contrib::AnalysisConfig cfg; SetConfig(&cfg); + if (use_mkldnn) { + cfg.EnableMKLDNN(); + std::unordered_set op_list = {"conv3d"}; + cfg.SetMKLDNNOp(op_list); + } + std::vector outputs; std::vector> input_slots_all; SetInput(&input_slots_all); @@ -192,6 +215,11 @@ TEST(Analyzer_dam, profile) { } } +TEST(Analyzer_dam, profile) { profile(); } +#ifdef PADDLE_WITH_MKLDNN +TEST(Analyzer_dam, profile_mkldnn) { profile(true /* use_mkldnn */); } +#endif + // Check the fuse status TEST(Analyzer_dam, fuse_statis) { contrib::AnalysisConfig cfg; @@ -202,14 +230,17 @@ TEST(Analyzer_dam, fuse_statis) { auto fuse_statis = GetFuseStatis( static_cast(predictor.get()), &num_ops); ASSERT_TRUE(fuse_statis.count("fc_fuse")); - EXPECT_EQ(fuse_statis.at("fc_fuse"), 317); - EXPECT_EQ(num_ops, 2020); } // Compare result of NativeConfig and AnalysisConfig -TEST(Analyzer_dam, compare) { - contrib::AnalysisConfig cfg; +void compare(bool use_mkldnn = false) { + AnalysisConfig cfg; SetConfig(&cfg); + if (use_mkldnn) { + cfg.EnableMKLDNN(); + std::unordered_set op_list = {"conv3d"}; + cfg.SetMKLDNNOp(op_list); + } std::vector> input_slots_all; SetInput(&input_slots_all); @@ -218,5 +249,21 @@ TEST(Analyzer_dam, compare) { reinterpret_cast(&cfg), input_slots_all); } +TEST(Analyzer_dam, compare) { compare(); } +#ifdef PADDLE_WITH_MKLDNN +TEST(Analyzer_dam, compare_mkldnn) { compare(true /* use_mkldnn */); } +#endif + +// Compare Deterministic result +TEST(Analyzer_dam, compare_determine) { + AnalysisConfig cfg; + SetConfig(&cfg); + + std::vector> input_slots_all; + SetInput(&input_slots_all); + CompareDeterministic(reinterpret_cast(&cfg), + input_slots_all); +} + } // namespace inference } // namespace paddle diff --git a/paddle/fluid/inference/tests/api/analyzer_lac_tester.cc b/paddle/fluid/inference/tests/api/analyzer_lac_tester.cc index 310852e2f7cb2..142801382b4fd 100644 --- a/paddle/fluid/inference/tests/api/analyzer_lac_tester.cc +++ b/paddle/fluid/inference/tests/api/analyzer_lac_tester.cc @@ -180,6 +180,17 @@ TEST(Analyzer_LAC, compare) { reinterpret_cast(&cfg), input_slots_all); } +// Compare Deterministic result +TEST(Analyzer_LAC, compare_determine) { + AnalysisConfig cfg; + SetConfig(&cfg); + + std::vector> input_slots_all; + SetInput(&input_slots_all); + CompareDeterministic(reinterpret_cast(&cfg), + input_slots_all); +} + } // namespace analysis } // namespace inference } // namespace paddle diff --git a/paddle/fluid/inference/tests/api/analyzer_ner_tester.cc b/paddle/fluid/inference/tests/api/analyzer_ner_tester.cc index 3a5f844de3cae..f19a2ed59ef2f 100644 --- a/paddle/fluid/inference/tests/api/analyzer_ner_tester.cc +++ b/paddle/fluid/inference/tests/api/analyzer_ner_tester.cc @@ -93,9 +93,17 @@ void PrepareInputs(std::vector *input_slots, DataRecord *data, } } -void SetConfig(contrib::AnalysisConfig *cfg) { - cfg->prog_file = FLAGS_infer_model + "/__model__"; - cfg->param_file = FLAGS_infer_model + "/param"; +void SetConfig(contrib::AnalysisConfig *cfg, bool memory_load = false) { + if (memory_load) { + std::string buffer_prog, buffer_param; + ReadBinaryFile(FLAGS_infer_model + "/__model__", &buffer_prog); + ReadBinaryFile(FLAGS_infer_model + "/param", &buffer_param); + cfg->SetModelBuffer(&buffer_prog[0], buffer_prog.size(), &buffer_param[0], + buffer_param.size()); + } else { + cfg->prog_file = FLAGS_infer_model + "/__model__"; + cfg->param_file = FLAGS_infer_model + "/param"; + } cfg->use_gpu = false; cfg->device = 0; cfg->specify_input_name = true; @@ -114,9 +122,9 @@ void SetInput(std::vector> *inputs) { } // Easy for profiling independently. -TEST(Analyzer_Chinese_ner, profile) { +void profile(bool memory_load = false) { contrib::AnalysisConfig cfg; - SetConfig(&cfg); + SetConfig(&cfg, memory_load); std::vector outputs; std::vector> input_slots_all; @@ -138,6 +146,12 @@ TEST(Analyzer_Chinese_ner, profile) { } } +TEST(Analyzer_Chinese_ner, profile) { profile(); } + +TEST(Analyzer_Chinese_ner, profile_memory_load) { + profile(true /* memory_load */); +} + // Check the fuse status TEST(Analyzer_Chinese_ner, fuse_statis) { contrib::AnalysisConfig cfg; @@ -165,5 +179,16 @@ TEST(Analyzer_Chinese_ner, compare) { reinterpret_cast(&cfg), input_slots_all); } +// Compare Deterministic result +TEST(Analyzer_Chinese_ner, compare_determine) { + AnalysisConfig cfg; + SetConfig(&cfg); + + std::vector> input_slots_all; + SetInput(&input_slots_all); + CompareDeterministic(reinterpret_cast(&cfg), + input_slots_all); +} + } // namespace inference } // namespace paddle diff --git a/paddle/fluid/inference/tests/api/analyzer_resnet50_tester.cc b/paddle/fluid/inference/tests/api/analyzer_resnet50_tester.cc index 2b936175ed3f8..764ae5ed8506a 100644 --- a/paddle/fluid/inference/tests/api/analyzer_resnet50_tester.cc +++ b/paddle/fluid/inference/tests/api/analyzer_resnet50_tester.cc @@ -27,6 +27,7 @@ void SetConfig(AnalysisConfig *cfg) { cfg->device = 0; cfg->enable_ir_optim = true; cfg->specify_input_name = true; + cfg->SetCpuMathLibraryNumThreads(FLAGS_paddle_num_threads); } void SetInput(std::vector> *inputs) { @@ -84,6 +85,17 @@ TEST(Analyzer_resnet50, compare) { compare(); } TEST(Analyzer_resnet50, compare_mkldnn) { compare(true /* use_mkldnn */); } #endif +// Compare Deterministic result +TEST(Analyzer_resnet50, compare_determine) { + AnalysisConfig cfg; + SetConfig(&cfg); + + std::vector> input_slots_all; + SetInput(&input_slots_all); + CompareDeterministic(reinterpret_cast(&cfg), + input_slots_all); +} + } // namespace analysis } // namespace inference } // namespace paddle diff --git a/paddle/fluid/inference/tests/api/analyzer_rnn1_tester.cc b/paddle/fluid/inference/tests/api/analyzer_rnn1_tester.cc index 1ae2b4b03a1b2..17f4587a5093a 100644 --- a/paddle/fluid/inference/tests/api/analyzer_rnn1_tester.cc +++ b/paddle/fluid/inference/tests/api/analyzer_rnn1_tester.cc @@ -265,6 +265,17 @@ TEST(Analyzer_rnn1, compare) { reinterpret_cast(&cfg), input_slots_all); } +// Compare Deterministic result +TEST(Analyzer_rnn1, compare_determine) { + AnalysisConfig cfg; + SetConfig(&cfg); + + std::vector> input_slots_all; + SetInput(&input_slots_all); + CompareDeterministic(reinterpret_cast(&cfg), + input_slots_all); +} + // Test Multi-Thread. TEST(Analyzer_rnn1, multi_thread) { contrib::AnalysisConfig cfg; diff --git a/paddle/fluid/inference/tests/api/analyzer_rnn2_tester.cc b/paddle/fluid/inference/tests/api/analyzer_rnn2_tester.cc index e2985006f0ed8..f8354e76871e7 100644 --- a/paddle/fluid/inference/tests/api/analyzer_rnn2_tester.cc +++ b/paddle/fluid/inference/tests/api/analyzer_rnn2_tester.cc @@ -158,5 +158,16 @@ TEST(Analyzer_rnn2, compare) { reinterpret_cast(&cfg), input_slots_all); } +// Compare Deterministic result +TEST(Analyzer_rnn2, compare_determine) { + AnalysisConfig cfg; + SetConfig(&cfg); + + std::vector> input_slots_all; + SetInput(&input_slots_all); + CompareDeterministic(reinterpret_cast(&cfg), + input_slots_all); +} + } // namespace inference } // namespace paddle diff --git a/paddle/fluid/inference/tests/api/analyzer_seq_conv1_tester.cc b/paddle/fluid/inference/tests/api/analyzer_seq_conv1_tester.cc index 858191184a377..f5082cd60f1ae 100644 --- a/paddle/fluid/inference/tests/api/analyzer_seq_conv1_tester.cc +++ b/paddle/fluid/inference/tests/api/analyzer_seq_conv1_tester.cc @@ -204,5 +204,16 @@ TEST(Analyzer_seq_conv1, compare) { reinterpret_cast(&cfg), input_slots_all); } +// Compare Deterministic result +TEST(Analyzer_seq_conv1, compare_determine) { + AnalysisConfig cfg; + SetConfig(&cfg); + + std::vector> input_slots_all; + SetInput(&input_slots_all); + CompareDeterministic(reinterpret_cast(&cfg), + input_slots_all); +} + } // namespace inference } // namespace paddle diff --git a/paddle/fluid/inference/tests/api/analyzer_text_classification_tester.cc b/paddle/fluid/inference/tests/api/analyzer_text_classification_tester.cc index 34a241f070fdc..79f3c81ade450 100644 --- a/paddle/fluid/inference/tests/api/analyzer_text_classification_tester.cc +++ b/paddle/fluid/inference/tests/api/analyzer_text_classification_tester.cc @@ -106,6 +106,17 @@ TEST(Analyzer_Text_Classification, compare) { reinterpret_cast(&cfg), input_slots_all); } +// Compare Deterministic result +TEST(Analyzer_Text_Classification, compare_determine) { + AnalysisConfig cfg; + SetConfig(&cfg); + + std::vector> input_slots_all; + SetInput(&input_slots_all); + CompareDeterministic(reinterpret_cast(&cfg), + input_slots_all); +} + TEST(Analyzer_Text_Classification, compare_against_embedding_fc_lstm_fused) { AnalysisConfig cfg; SetConfig(&cfg); diff --git a/paddle/fluid/inference/tests/api/analyzer_vis_tester.cc b/paddle/fluid/inference/tests/api/analyzer_vis_tester.cc index 956a235edcefb..d73bccefd5fc8 100644 --- a/paddle/fluid/inference/tests/api/analyzer_vis_tester.cc +++ b/paddle/fluid/inference/tests/api/analyzer_vis_tester.cc @@ -27,7 +27,7 @@ struct Record { }; Record ProcessALine(const std::string &line) { - VLOG(30) << "process a line"; + VLOG(3) << "process a line"; std::vector columns; split(line, '\t', &columns); CHECK_EQ(columns.size(), 2UL) @@ -45,8 +45,8 @@ Record ProcessALine(const std::string &line) { for (auto &s : shape_strs) { record.shape.push_back(std::stoi(s)); } - VLOG(30) << "data size " << record.data.size(); - VLOG(30) << "data shape size " << record.shape.size(); + VLOG(3) << "data size " << record.data.size(); + VLOG(3) << "data shape size " << record.shape.size(); return record; } @@ -93,18 +93,20 @@ void profile(bool use_mkldnn = false) { SetInput(&input_slots_all); TestPrediction(reinterpret_cast(&cfg), input_slots_all, &outputs, FLAGS_num_threads); - if (FLAGS_num_threads == 1 && !FLAGS_test_all_data) { - const float ocr_result_data[] = { - 5.273636460856323538e-08, 3.296741795111302054e-07, - 1.873261190610264748e-08, 3.403730275408634043e-08, - 3.383312474625199684e-08}; - PADDLE_ENFORCE_EQ(outputs.size(), 1UL); - size_t size = GetSize(outputs[0]); - PADDLE_ENFORCE_GT(size, 0); - float *result = static_cast(outputs[0].data.data()); - for (size_t i = 0; i < std::min(5UL, size); i++) { - EXPECT_NEAR(result[i], ocr_result_data[i], 1e-3); + std::string line; + std::ifstream file(FLAGS_refer_result); + std::getline(file, line); + auto refer = ProcessALine(line); + file.close(); + + auto &output = outputs.front(); + size_t numel = output.data.length() / PaddleDtypeSize(output.dtype); + CHECK_EQ(numel, refer.data.size()); + for (size_t i = 0; i < numel; ++i) { + CHECK_LT( + fabs(static_cast(output.data.data())[i] - refer.data[i]), + 1e-5); } } } @@ -143,6 +145,17 @@ TEST(Analyzer_vis, compare) { compare(); } TEST(Analyzer_vis, compare_mkldnn) { compare(true /* use_mkldnn */); } #endif +// Compare Deterministic result +TEST(Analyzer_vis, compare_determine) { + AnalysisConfig cfg; + SetConfig(&cfg); + + std::vector> input_slots_all; + SetInput(&input_slots_all); + CompareDeterministic(reinterpret_cast(&cfg), + input_slots_all); +} + } // namespace analysis } // namespace inference } // namespace paddle diff --git a/paddle/fluid/inference/tests/api/config_printer.h b/paddle/fluid/inference/tests/api/config_printer.h index aa0c4b1d049bc..7046bce303e2b 100644 --- a/paddle/fluid/inference/tests/api/config_printer.h +++ b/paddle/fluid/inference/tests/api/config_printer.h @@ -49,10 +49,10 @@ std::ostream &operator<<(std::ostream &os, const NativeConfig &config) { os << GenSpaces(num_spaces) << "device: " << config.device << "\n"; os << GenSpaces(num_spaces) << "fraction_of_gpu_memory: " << config.fraction_of_gpu_memory << "\n"; - os << GenSpaces(num_spaces) << "prog_file: " << config.prog_file << "\n"; - os << GenSpaces(num_spaces) << "param_file: " << config.param_file << "\n"; os << GenSpaces(num_spaces) << "specify_input_name: " << config.specify_input_name << "\n"; + os << GenSpaces(num_spaces) + << "cpu_num_threads: " << config.cpu_math_library_num_threads() << "\n"; num_spaces--; os << GenSpaces(num_spaces) << "}\n"; return os; @@ -63,6 +63,13 @@ std::ostream &operator<<(std::ostream &os, os << GenSpaces(num_spaces) << "contrib::AnalysisConfig {\n"; num_spaces++; os << *reinterpret_cast(&config); + if (!config.model_from_memory()) { + os << GenSpaces(num_spaces) << "prog_file: " << config.prog_file << "\n"; + os << GenSpaces(num_spaces) << "param_file: " << config.param_file << "\n"; + } else { + os << GenSpaces(num_spaces) + << "prog_file and param_file: load from memory \n"; + } os << GenSpaces(num_spaces) << "enable_ir_optim: " << config.enable_ir_optim << "\n"; os << GenSpaces(num_spaces) diff --git a/paddle/fluid/inference/tests/api/tester_helper.h b/paddle/fluid/inference/tests/api/tester_helper.h index 7b686045a59c9..b0c8f395ce05f 100644 --- a/paddle/fluid/inference/tests/api/tester_helper.h +++ b/paddle/fluid/inference/tests/api/tester_helper.h @@ -30,18 +30,25 @@ #include "paddle/fluid/inference/api/helper.h" #include "paddle/fluid/inference/tests/api/config_printer.h" #include "paddle/fluid/inference/tests/test_helper.h" +#include "paddle/fluid/inference/utils/benchmark.h" #include "paddle/fluid/platform/profiler.h" +DEFINE_string(model_name, "", "model name"); DEFINE_string(infer_model, "", "model path"); DEFINE_string(infer_data, "", "data file"); +DEFINE_string(refer_result, "", "reference result for comparison"); DEFINE_int32(batch_size, 1, "batch size."); DEFINE_int32(repeat, 1, "Running the inference program repeat times."); DEFINE_bool(test_all_data, false, "Test the all dataset in data file."); DEFINE_int32(num_threads, 1, "Running the inference program in multi-threads."); DEFINE_bool(use_analysis, true, "Running the inference program in analysis mode."); +DEFINE_bool(record_benchmark, false, + "Record benchmark after profiling the model"); +DEFINE_double(accuracy, 1e-3, "Result Accuracy."); DECLARE_bool(profile); +DECLARE_int32(paddle_num_threads); namespace paddle { namespace inference { @@ -79,7 +86,7 @@ void CompareResult(const std::vector &outputs, float *pdata = static_cast(out.data.data()); float *pdata_ref = static_cast(ref_out.data.data()); for (size_t j = 0; j < size; ++j) { - EXPECT_NEAR(pdata_ref[j], pdata[j], 1e-3); + EXPECT_NEAR(pdata_ref[j], pdata[j], FLAGS_accuracy); } break; } @@ -177,11 +184,9 @@ void TestOneThreadPrediction( warmup_timer.tic(); predictor->Run(inputs[0], outputs, batch_size); PrintTime(batch_size, 1, 1, 0, warmup_timer.toc(), 1); -#if !defined(_WIN32) if (FLAGS_profile) { paddle::platform::ResetProfiler(); } -#endif } LOG(INFO) << "Run " << num_times << " times..."; @@ -193,8 +198,16 @@ void TestOneThreadPrediction( predictor->Run(inputs[j], outputs, batch_size); } } - PrintTime(batch_size, num_times, 1, 0, run_timer.toc() / num_times, - inputs.size()); + + double latency = run_timer.toc() / num_times; + PrintTime(batch_size, num_times, 1, 0, latency, inputs.size()); + if (FLAGS_record_benchmark) { + Benchmark benchmark; + benchmark.SetName(FLAGS_model_name); + benchmark.SetBatchSize(batch_size); + benchmark.SetLatency(latency); + benchmark.PersistToFile("benchmark_record.txt"); + } } } @@ -206,22 +219,23 @@ void TestMultiThreadPrediction( int batch_size = FLAGS_batch_size; int num_times = FLAGS_repeat; std::vector threads; - std::vector> predictors; - predictors.emplace_back(CreateTestPredictor(config, use_analysis)); - for (int tid = 1; tid < num_threads; ++tid) { - predictors.emplace_back(predictors.front()->Clone()); - } + auto main_predictor = CreateTestPredictor(config, use_analysis); size_t total_time{0}; for (int tid = 0; tid < num_threads; ++tid) { threads.emplace_back([&, tid]() { -#ifdef PADDLE_WITH_MKLDNN - platform::set_cur_thread_id(static_cast(tid) + 1); -#endif // Each thread should have local inputs and outputs. // The inputs of each thread are all the same. std::vector outputs_tid; - auto &predictor = predictors[tid]; + // To ensure the thread binding correctly, + // please clone inside the threadpool. + auto predictor = main_predictor->Clone(); +#ifdef PADDLE_WITH_MKLDNN + if (use_analysis) { + static_cast(predictor.get()) + ->SetMkldnnThreadID(static_cast(tid) + 1); + } +#endif // warmup run LOG(INFO) << "Running thread " << tid << ", warm up run..."; @@ -230,11 +244,9 @@ void TestMultiThreadPrediction( warmup_timer.tic(); predictor->Run(inputs[0], outputs, batch_size); PrintTime(batch_size, 1, num_threads, tid, warmup_timer.toc(), 1); -#if !defined(_WIN32) if (FLAGS_profile) { paddle::platform::ResetProfiler(); } -#endif } LOG(INFO) << "Thread " << tid << " run " << num_times << " times..."; @@ -272,6 +284,26 @@ void TestPrediction(const PaddlePredictor::Config *config, } } +void CompareDeterministic( + const PaddlePredictor::Config *config, + const std::vector> &inputs) { + int batch_size = FLAGS_batch_size; + int num_times = FLAGS_repeat; + auto predictor = CreateTestPredictor(config, FLAGS_use_analysis); + + // warmup run + std::vector warmup_outputs, outputs; + predictor->Run(inputs[0], &warmup_outputs, batch_size); + + // run num_times to Compare Deterministic Result. + for (int i = 0; i < num_times; i++) { + for (size_t j = 0; j < inputs.size(); j++) { + predictor->Run(inputs[j], &outputs, batch_size); + CompareResult(outputs, warmup_outputs); + } + } +} + void CompareNativeAndAnalysis( const PaddlePredictor::Config *config, const std::vector> &inputs) { @@ -363,7 +395,7 @@ static bool CompareTensorData(const framework::LoDTensor &a, } for (size_t i = 0; i < a_size; i++) { - if (a.type() == typeid(float)) { + if (a.type() == framework::proto::VarType::FP32) { const auto *a_data = a.data(); const auto *b_data = b.data(); if (std::abs(a_data[i] - b_data[i]) > 1e-3) { @@ -372,7 +404,7 @@ static bool CompareTensorData(const framework::LoDTensor &a, b_data[i]); return false; } - } else if (a.type() == typeid(int64_t)) { + } else if (a.type() == framework::proto::VarType::INT64) { const auto *a_data = a.data(); const auto *b_data = b.data(); if (std::abs(a_data[i] - b_data[i]) > 1e-3) { diff --git a/paddle/fluid/inference/tests/api/trt_models_tester.cc b/paddle/fluid/inference/tests/api/trt_models_tester.cc index ef612ce614832..d3bd035c1c49c 100644 --- a/paddle/fluid/inference/tests/api/trt_models_tester.cc +++ b/paddle/fluid/inference/tests/api/trt_models_tester.cc @@ -78,6 +78,7 @@ void profile(std::string model_dir, bool use_analysis, bool use_tensorrt) { std::vector outputs; if (use_analysis || use_tensorrt) { contrib::AnalysisConfig config(true); + config.pass_builder()->TurnOnDebug(); SetConfig(&config, model_dir, true, use_tensorrt, FLAGS_batch_size); TestPrediction(reinterpret_cast(&config), @@ -135,12 +136,37 @@ TEST(TensorRT_resnext50, compare) { TEST(TensorRT_resnext50, profile) { std::string model_dir = FLAGS_infer_model + "/resnext50"; + // Set FLAGS_record_benchmark to true to record benchmark to file. + // FLAGS_record_benchmark=true; + FLAGS_model_name = "resnext50"; profile(model_dir, /* use_analysis */ true, FLAGS_use_tensorrt); } +TEST(resnext50, compare_analysis_native) { + std::string model_dir = FLAGS_infer_model + "/resnext50"; + compare(model_dir, false /*use tensorrt*/); +} + TEST(TensorRT_mobilenet, analysis) { std::string model_dir = FLAGS_infer_model + "/" + "mobilenet"; - compare(model_dir, /* use_tensorrt */ false); + compare(model_dir, false /* use_tensorrt */); +} + +TEST(AnalysisPredictor, use_gpu) { + std::string model_dir = FLAGS_infer_model + "/" + "mobilenet"; + AnalysisConfig config(true); + config.model_dir = model_dir; + config.fraction_of_gpu_memory = 0.15; + config.pass_builder()->TurnOnDebug(); + + std::vector> inputs_all; + auto predictor = CreatePaddlePredictor(config); + SetFakeImageInput(&inputs_all, model_dir, false, "__model__", ""); + + std::vector outputs; + for (auto& input : inputs_all) { + ASSERT_TRUE(predictor->Run(input, &outputs)); + } } } // namespace inference diff --git a/paddle/fluid/inference/tests/book/test_inference_nlp.cc b/paddle/fluid/inference/tests/book/test_inference_nlp.cc index cbcfc964c91c3..5c1204b9e6b78 100644 --- a/paddle/fluid/inference/tests/book/test_inference_nlp.cc +++ b/paddle/fluid/inference/tests/book/test_inference_nlp.cc @@ -12,7 +12,6 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include #include #include #include // NOLINT diff --git a/paddle/fluid/inference/tests/test_helper.h b/paddle/fluid/inference/tests/test_helper.h index 2118fcfd4bb15..75fa611c0d701 100644 --- a/paddle/fluid/inference/tests/test_helper.h +++ b/paddle/fluid/inference/tests/test_helper.h @@ -20,6 +20,7 @@ limitations under the License. */ #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/inference/io.h" +#include "paddle/fluid/platform/port.h" #include "paddle/fluid/platform/profiler.h" DECLARE_bool(use_mkldnn); diff --git a/paddle/fluid/inference/utils/CMakeLists.txt b/paddle/fluid/inference/utils/CMakeLists.txt new file mode 100644 index 0000000000000..cfb80fe6ec11a --- /dev/null +++ b/paddle/fluid/inference/utils/CMakeLists.txt @@ -0,0 +1,7 @@ +cc_library(benchmark SRCS benchmark.cc DEPS enforce) +cc_test(test_benchmark SRCS benchmark_tester.cc DEPS benchmark) +cc_binary(visualizer SRCS visualizer.cc DEPS analysis + paddle_pass_builder ir_pass_manager pass graph_viz_pass analysis_passes) +if(WIN32) + target_link_libraries(visualizer shlwapi) +endif(WIN32) diff --git a/paddle/fluid/inference/utils/benchmark.cc b/paddle/fluid/inference/utils/benchmark.cc new file mode 100644 index 0000000000000..0bd526bcac2d9 --- /dev/null +++ b/paddle/fluid/inference/utils/benchmark.cc @@ -0,0 +1,49 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/inference/utils/benchmark.h" +#include +#include "paddle/fluid/platform/enforce.h" + +namespace paddle { +namespace inference { + +std::string Benchmark::SerializeToString() const { + std::stringstream ss; + ss << "-----------------------------------------------------\n"; + ss << "name\t"; + ss << "batch_size\t"; + ss << "num_threads\t"; + ss << "latency\t"; + ss << "qps"; + ss << '\n'; + + ss << name_ << "\t"; + ss << batch_size_ << "\t\t"; + ss << num_threads_ << "\t"; + ss << latency_ << "\t"; + ss << 1000.0 / latency_; + ss << '\n'; + return ss.str(); +} +void Benchmark::PersistToFile(const std::string &path) const { + std::ofstream file(path, std::ios::app); + PADDLE_ENFORCE(file.is_open(), "Can not open %s to add benchmark", path); + file << SerializeToString(); + file.flush(); + file.close(); +} + +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/utils/benchmark.h b/paddle/fluid/inference/utils/benchmark.h new file mode 100644 index 0000000000000..76a3dd2c2992e --- /dev/null +++ b/paddle/fluid/inference/utils/benchmark.h @@ -0,0 +1,54 @@ +// 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 +#include +#include + +namespace paddle { +namespace inference { + +/* + * Helper class to calculate the performance. + */ +struct Benchmark { + int batch_size() const { return batch_size_; } + void SetBatchSize(int x) { batch_size_ = x; } + + int num_threads() const { return num_threads_; } + void SetNumThreads(int x) { num_threads_ = x; } + + bool use_gpu() const { return use_gpu_; } + void SetUseGpu() { use_gpu_ = true; } + + float latency() const { return latency_; } + void SetLatency(float x) { latency_ = x; } + + const std::string& name() const { return name_; } + void SetName(const std::string& name) { name_ = name; } + + std::string SerializeToString() const; + void PersistToFile(const std::string& path) const; + + private: + bool use_gpu_{false}; + int batch_size_{0}; + float latency_; + int num_threads_{1}; + std::string name_; +}; + +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/utils/benchmark_tester.cc b/paddle/fluid/inference/utils/benchmark_tester.cc new file mode 100644 index 0000000000000..eb255474082b2 --- /dev/null +++ b/paddle/fluid/inference/utils/benchmark_tester.cc @@ -0,0 +1,39 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/inference/utils/benchmark.h" +#include +#include + +using namespace paddle::inference; +TEST(Benchmark, basic) { + Benchmark benchmark; + benchmark.SetName("key0"); + benchmark.SetBatchSize(10); + benchmark.SetUseGpu(); + benchmark.SetLatency(220); + LOG(INFO) << "benchmark:\n" << benchmark.SerializeToString(); +} + +TEST(Benchmark, PersistToFile) { + Benchmark benchmark; + benchmark.SetName("key0"); + benchmark.SetBatchSize(10); + benchmark.SetUseGpu(); + benchmark.SetLatency(220); + + benchmark.PersistToFile("1.log"); + benchmark.PersistToFile("1.log"); + benchmark.PersistToFile("1.log"); +} \ No newline at end of file diff --git a/paddle/fluid/inference/utils/visualizer.cc b/paddle/fluid/inference/utils/visualizer.cc new file mode 100644 index 0000000000000..7c0dd64dea88e --- /dev/null +++ b/paddle/fluid/inference/utils/visualizer.cc @@ -0,0 +1,92 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/inference/utils/visualizer.h" +#include +#include +#include +#include +#include "paddle/fluid/framework/ir/graph_viz_pass.h" +#include "paddle/fluid/inference/analysis/analyzer.h" +#include "paddle/fluid/inference/analysis/passes/ir_analysis_pass.h" +#include "paddle/fluid/platform/init.h" + +DEFINE_string(model_dir, "", "model directory"); +DEFINE_string(model_program_path, "", "model program path"); +DEFINE_string(model_params_path, "", "model params path"); + +using paddle::inference::analysis::Argument; + +namespace paddle { +namespace inference { +namespace utils { + +void Visualizer::SetArgument(Argument *argument) { argument_ = argument; } + +bool Visualizer::Run() { + paddle::framework::InitDevices(false); + paddle::inference::analysis::Analyzer().Run(argument_); + return true; +} + +} // namespace utils +} // namespace inference +} // namespace paddle + +// Generate a dot file describing the structure of graph. +// To use this tool, run command: ./visualizer [options...] +// Options: +// --model_dir: the directory of model +// --model_program_path: the path of program +// --model_params_path: the path of params +int main(int argc, char *argv[]) { + gflags::ParseCommandLineFlags(&argc, &argv, true); + google::InitGoogleLogging(argv[0]); + + paddle::inference::analysis::Argument argument; + argument.SetUseGPU(false); + argument.SetUseTensorRT(false); + + if (FLAGS_model_dir.empty()) { + if (FLAGS_model_program_path.empty() || FLAGS_model_params_path.empty()) { + LOG(ERROR) << "Please set model_dir" + " or model_program_path and model_params_path"; + return -1; + } else { + argument.SetModelProgramPath(FLAGS_model_program_path); + argument.SetModelParamsPath(FLAGS_model_params_path); + } + } else { + argument.SetModelDir(FLAGS_model_dir); + } + + // Only 1 pass, default filename is 0_ir_origin.dot + // For more details, looking for paddle::inference::analysis::IRPassManager + argument.SetIrAnalysisPasses({"infer_clean_graph_pass", "graph_viz_pass"}); + + std::unique_ptr scope{ + new paddle::framework::Scope()}; + argument.SetScopeNotOwned( + const_cast(scope.get())); + + paddle::inference::utils::Visualizer visualizer; + visualizer.SetArgument(&argument); + visualizer.Run(); + + return 0; +} + +USE_PASS(infer_clean_graph_pass); +USE_PASS(graph_viz_pass); +USE_PASS(graph_to_program_pass); diff --git a/paddle/fluid/inference/utils/visualizer.h b/paddle/fluid/inference/utils/visualizer.h new file mode 100644 index 0000000000000..be532f92cf60e --- /dev/null +++ b/paddle/fluid/inference/utils/visualizer.h @@ -0,0 +1,42 @@ +// 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 +#include "paddle/fluid/inference/analysis/argument.h" + +namespace paddle { +namespace inference { +namespace utils { + +using paddle::inference::analysis::Argument; + +class Visualizer final { + public: + Visualizer() = default; + ~Visualizer() = default; + Visualizer(const Visualizer &) = delete; + Visualizer &operator=(const Visualizer &) = delete; + + void SetArgument(Argument *); + bool Run(); + + private: + Argument *argument_; +}; + +} // namespace utils +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/memory/allocation/allocator_facade.cc b/paddle/fluid/memory/allocation/allocator_facade.cc index e207a853c8f78..794d729bdc1ad 100644 --- a/paddle/fluid/memory/allocation/allocator_facade.cc +++ b/paddle/fluid/memory/allocation/allocator_facade.cc @@ -76,12 +76,12 @@ class ChunkedAllocator : public Allocator { default_allocator_ = raw_allocator_; } else { if (capacity == 1) { - VLOG(10) << "Create BestFitAllocator with chunk_size " - << max_chunk_size_; + VLOG(1) << "Create BestFitAllocator with chunk_size " + << max_chunk_size_; default_allocator_ = CreateAllocatorWithChunk(); } else { - VLOG(10) << "Create AutoIncrementAllocator with chunk_size " - << max_chunk_size_ << " and capacity " << capacity; + VLOG(1) << "Create AutoIncrementAllocator with chunk_size " + << max_chunk_size_ << " and capacity " << capacity; default_allocator_ = std::make_shared( [this] { return std::move(CreateAllocatorWithChunk()); }, capacity); } diff --git a/paddle/fluid/memory/allocation/best_fit_allocator_test.cc b/paddle/fluid/memory/allocation/best_fit_allocator_test.cc index 20748a23a1951..b274b05562b15 100644 --- a/paddle/fluid/memory/allocation/best_fit_allocator_test.cc +++ b/paddle/fluid/memory/allocation/best_fit_allocator_test.cc @@ -99,9 +99,8 @@ TEST(BestFitAllocator, test_concurrent_cpu_allocation) { LockedAllocator locked_allocator(std::move(best_fit_allocator)); - auto th_main = [&] { - std::random_device dev; - std::default_random_engine engine(dev()); + auto th_main = [&](std::random_device::result_type seed) { + std::default_random_engine engine(seed); std::uniform_int_distribution dist(1U, 1024U); for (size_t i = 0; i < 128; ++i) { @@ -125,7 +124,8 @@ TEST(BestFitAllocator, test_concurrent_cpu_allocation) { { std::vector threads; for (size_t i = 0; i < 1024; ++i) { - threads.emplace_back(th_main); + std::random_device dev; + threads.emplace_back(th_main, dev()); } for (auto& th : threads) { th.join(); diff --git a/paddle/fluid/memory/allocation/best_fit_allocator_test.cu b/paddle/fluid/memory/allocation/best_fit_allocator_test.cu index f7f17e1d36e0a..fdd5b43ad4aa8 100644 --- a/paddle/fluid/memory/allocation/best_fit_allocator_test.cu +++ b/paddle/fluid/memory/allocation/best_fit_allocator_test.cu @@ -41,9 +41,8 @@ TEST(BestFitAllocator, concurrent_cuda) { LockedAllocator concurrent_allocator( std::unique_ptr(new BestFitAllocator(cuda_allocation.get()))); - auto th_main = [&] { - std::random_device dev; - std::default_random_engine engine(dev()); + auto th_main = [&](std::random_device::result_type seed) { + std::default_random_engine engine(seed); std::uniform_int_distribution dist(1U, 1024U); platform::CUDAPlace gpu(0); platform::CUDADeviceContext dev_ctx(gpu); @@ -75,7 +74,8 @@ TEST(BestFitAllocator, concurrent_cuda) { { std::vector threads; for (size_t i = 0; i < 1024; ++i) { - threads.emplace_back(th_main); + std::random_device dev; + threads.emplace_back(th_main, dev()); } for (auto& th : threads) { th.join(); diff --git a/paddle/fluid/memory/allocation/legacy_allocator.cc b/paddle/fluid/memory/allocation/legacy_allocator.cc index e66537272340e..64aa63ffe9705 100644 --- a/paddle/fluid/memory/allocation/legacy_allocator.cc +++ b/paddle/fluid/memory/allocation/legacy_allocator.cc @@ -14,11 +14,13 @@ #include "paddle/fluid/memory/allocation/legacy_allocator.h" #include +#include #include "glog/logging.h" #include "paddle/fluid/memory/detail/buddy_allocator.h" #include "paddle/fluid/memory/detail/system_allocator.h" #include "paddle/fluid/platform/gpu_info.h" #include "paddle/fluid/string/printf.h" +#include "paddle/fluid/string/split.h" DEFINE_bool(init_allocated_mem, false, "It is a mistake that the values of the memory allocated by " @@ -91,7 +93,7 @@ void *Alloc(const platform::CPUPlace &place, size_t size) { if (FLAGS_init_allocated_mem) { memset(p, 0xEF, size); } - VLOG(100) << " pointer=" << p; + VLOG(10) << " pointer=" << p; return p; } @@ -110,31 +112,35 @@ size_t Used(const platform::CPUPlace &place) { BuddyAllocator *GetGPUBuddyAllocator(int gpu_id) { static std::once_flag init_flag; static detail::BuddyAllocator **a_arr = nullptr; + static std::vector devices; std::call_once(init_flag, [gpu_id]() { - int gpu_num = platform::GetCUDADeviceCount(); - PADDLE_ENFORCE(gpu_id < gpu_num, "gpu_id:%d should < gpu_num:%d", gpu_id, - gpu_num); + devices = platform::GetSelectedDevices(); + int gpu_num = devices.size(); a_arr = new BuddyAllocator *[gpu_num]; - for (int i = 0; i < gpu_num; i++) { + for (size_t i = 0; i < devices.size(); ++i) { + int dev_id = devices[i]; a_arr[i] = nullptr; - platform::SetDeviceId(i); - a_arr[i] = new BuddyAllocator( - std::unique_ptr(new detail::GPUAllocator(i)), - platform::GpuMinChunkSize(), platform::GpuMaxChunkSize()); - - VLOG(100) << "\n\nNOTE: each GPU device use " - << FLAGS_fraction_of_gpu_memory_to_use * 100 - << "% of GPU memory.\n" - << "You can set GFlags environment variable '" - << "FLAGS_fraction_of_gpu_memory_to_use" - << "' to change the fraction of GPU usage.\n\n"; + platform::SetDeviceId(dev_id); + a_arr[i] = new BuddyAllocator(std::unique_ptr( + new detail::GPUAllocator(dev_id)), + platform::GpuMinChunkSize(), + platform::GpuMaxChunkSize()); + + VLOG(10) << "\n\nNOTE: each GPU device use " + << FLAGS_fraction_of_gpu_memory_to_use * 100 + << "% of GPU memory.\n" + << "You can set GFlags environment variable '" + << "FLAGS_fraction_of_gpu_memory_to_use" + << "' to change the fraction of GPU usage.\n\n"; } }); platform::SetDeviceId(gpu_id); - return a_arr[gpu_id]; + auto pos = std::distance(devices.begin(), + std::find(devices.begin(), devices.end(), gpu_id)); + return a_arr[pos]; } #endif diff --git a/paddle/fluid/memory/allocation/retry_allocator_test.cc b/paddle/fluid/memory/allocation/retry_allocator_test.cc index a0ce2875cb833..f0b215dac2524 100644 --- a/paddle/fluid/memory/allocation/retry_allocator_test.cc +++ b/paddle/fluid/memory/allocation/retry_allocator_test.cc @@ -41,7 +41,7 @@ TEST(RetryAllocator, RetryAllocator) { size_t thread_num = 32; size_t sleep_time = 40; - size_t extra_time = 2; + size_t extra_time = 10; // Reserve to perform more tests in the future std::vector> allocators; diff --git a/paddle/fluid/memory/detail/buddy_allocator.cc b/paddle/fluid/memory/detail/buddy_allocator.cc index dd7ffaa26426e..26ef27c3caafa 100644 --- a/paddle/fluid/memory/detail/buddy_allocator.cc +++ b/paddle/fluid/memory/detail/buddy_allocator.cc @@ -32,11 +32,11 @@ BuddyAllocator::BuddyAllocator( system_allocator_(std::move(system_allocator)) {} BuddyAllocator::~BuddyAllocator() { - VLOG(100) << "BuddyAllocator Disconstructor makes sure that all of these " - "have actually been freed"; + VLOG(10) << "BuddyAllocator Disconstructor makes sure that all of these " + "have actually been freed"; while (!pool_.empty()) { auto block = static_cast(std::get<2>(*pool_.begin())); - VLOG(100) << "Free from block (" << block << ", " << max_chunk_size_ << ")"; + VLOG(10) << "Free from block (" << block << ", " << max_chunk_size_ << ")"; system_allocator_->Free(block, max_chunk_size_, block->index(cache_)); cache_.invalidate(block); @@ -57,12 +57,12 @@ void* BuddyAllocator::Alloc(size_t unaligned_size) { // acquire the allocator lock std::lock_guard lock(mutex_); - VLOG(100) << "Allocate " << unaligned_size << " bytes from chunk size " - << size; + VLOG(10) << "Allocate " << unaligned_size << " bytes from chunk size " + << size; // if the allocation is huge, send directly to the system allocator if (size > max_chunk_size_) { - VLOG(100) << "Allocate from system allocator."; + VLOG(10) << "Allocate from system allocator."; return SystemAlloc(size); } @@ -77,9 +77,9 @@ void* BuddyAllocator::Alloc(size_t unaligned_size) { return nullptr; } } else { - VLOG(100) << "Allocation from existing memory block " << std::get<2>(*it) - << " at address " - << reinterpret_cast(std::get<2>(*it))->data(); + VLOG(10) << "Allocation from existing memory block " << std::get<2>(*it) + << " at address " + << reinterpret_cast(std::get<2>(*it))->data(); } total_used_ += size; @@ -96,10 +96,10 @@ void BuddyAllocator::Free(void* p) { // Acquire the allocator lock std::lock_guard lock(mutex_); - VLOG(100) << "Free from address " << block; + VLOG(10) << "Free from address " << block; if (block->type(cache_) == MemoryBlock::HUGE_CHUNK) { - VLOG(100) << "Free directly from system allocator"; + VLOG(10) << "Free directly from system allocator"; system_allocator_->Free(block, block->total_size(cache_), block->index(cache_)); @@ -116,8 +116,8 @@ void BuddyAllocator::Free(void* p) { // Trying to merge the right buddy if (block->has_right_buddy(cache_)) { - VLOG(100) << "Merging this block " << block << " with its right buddy " - << block->right_buddy(cache_); + VLOG(10) << "Merging this block " << block << " with its right buddy " + << block->right_buddy(cache_); auto right_buddy = block->right_buddy(cache_); @@ -134,8 +134,8 @@ void BuddyAllocator::Free(void* p) { // Trying to merge the left buddy if (block->has_left_buddy(cache_)) { - VLOG(100) << "Merging this block " << block << " with its left buddy " - << block->left_buddy(cache_); + VLOG(10) << "Merging this block " << block << " with its left buddy " + << block->left_buddy(cache_); auto left_buddy = block->left_buddy(cache_); @@ -151,8 +151,8 @@ void BuddyAllocator::Free(void* p) { } // Dumping this block into pool - VLOG(100) << "Inserting free block (" << block << ", " - << block->total_size(cache_) << ")"; + VLOG(10) << "Inserting free block (" << block << ", " + << block->total_size(cache_) << ")"; pool_.insert( IndexSizeAddress(block->index(cache_), block->total_size(cache_), block)); @@ -174,7 +174,7 @@ void* BuddyAllocator::SystemAlloc(size_t size) { size_t index = 0; void* p = system_allocator_->Alloc(&index, size); - VLOG(100) << "Allocated " << p << " from system allocator."; + VLOG(10) << "Allocated " << p << " from system allocator."; if (p == nullptr) return nullptr; @@ -200,8 +200,8 @@ BuddyAllocator::PoolSet::iterator BuddyAllocator::RefillPool() { if (p == nullptr) return pool_.end(); - VLOG(100) << "Creating and inserting new block " << p - << " from system allocator"; + VLOG(10) << "Creating and inserting new block " << p + << " from system allocator"; static_cast(p)->init(&cache_, MemoryBlock::FREE_CHUNK, index, max_chunk_size_, nullptr, nullptr); @@ -245,19 +245,19 @@ void* BuddyAllocator::SplitToAlloc(BuddyAllocator::PoolSet::iterator it, auto block = static_cast(std::get<2>(*it)); pool_.erase(it); - VLOG(100) << "Split block (" << block << ", " << block->total_size(cache_) - << ") into"; + VLOG(10) << "Split block (" << block << ", " << block->total_size(cache_) + << ") into"; block->split(&cache_, size); - VLOG(100) << "Left block (" << block << ", " << block->total_size(cache_) - << ")"; + VLOG(10) << "Left block (" << block << ", " << block->total_size(cache_) + << ")"; block->set_type(&cache_, MemoryBlock::ARENA_CHUNK); // the rest of memory if exist if (block->has_right_buddy(cache_)) { if (block->right_buddy(cache_)->type(cache_) == MemoryBlock::FREE_CHUNK) { - VLOG(100) << "Insert right block (" << block->right_buddy(cache_) << ", " - << block->right_buddy(cache_)->total_size(cache_) << ")"; + VLOG(10) << "Insert right block (" << block->right_buddy(cache_) << ", " + << block->right_buddy(cache_)->total_size(cache_) << ")"; pool_.insert( IndexSizeAddress(block->right_buddy(cache_)->index(cache_), @@ -284,7 +284,7 @@ void BuddyAllocator::CleanIdleFallBackAlloc() { return; } - VLOG(100) << "Return block " << block << " to fallback allocator."; + VLOG(10) << "Return block " << block << " to fallback allocator."; system_allocator_->Free(block, max_chunk_size_, block->index(cache_)); cache_.invalidate(block); @@ -320,7 +320,7 @@ void BuddyAllocator::CleanIdleNormalAlloc() { MemoryBlock* block = static_cast(std::get<2>(*pool)); - VLOG(100) << "Return block " << block << " to base allocator."; + VLOG(10) << "Return block " << block << " to base allocator."; system_allocator_->Free(block, max_chunk_size_, block->index(cache_)); cache_.invalidate(block); diff --git a/paddle/fluid/memory/detail/meta_cache.cc b/paddle/fluid/memory/detail/meta_cache.cc index 152e4e7f9fa2e..b86e4f38c42a2 100644 --- a/paddle/fluid/memory/detail/meta_cache.cc +++ b/paddle/fluid/memory/detail/meta_cache.cc @@ -29,7 +29,7 @@ MemoryBlock::Desc MetadataCache::load(const MemoryBlock* block) const { return existing_desc->second; } else { auto* desc = reinterpret_cast(block); - VLOG(100) << "Load MemoryBlock::Desc type=" << desc->type; + VLOG(10) << "Load MemoryBlock::Desc type=" << desc->type; PADDLE_ASSERT(desc->check_guards()); return *reinterpret_cast(block); } diff --git a/paddle/fluid/memory/detail/system_allocator.cc b/paddle/fluid/memory/detail/system_allocator.cc index 2019d1a14f6dd..3e8fb83e9d5ba 100644 --- a/paddle/fluid/memory/detail/system_allocator.cc +++ b/paddle/fluid/memory/detail/system_allocator.cc @@ -86,7 +86,11 @@ void CPUAllocator::Free(void* p, size_t size, size_t index) { munlock(p, size); #endif } +#ifdef _WIN32 + _aligned_free(p); +#else free(p); +#endif } bool CPUAllocator::UseGpu() const { return false; } diff --git a/paddle/fluid/operators/CMakeLists.txt b/paddle/fluid/operators/CMakeLists.txt index de4f23515d859..4a14eb941cd98 100644 --- a/paddle/fluid/operators/CMakeLists.txt +++ b/paddle/fluid/operators/CMakeLists.txt @@ -16,6 +16,7 @@ add_subdirectory(metrics) add_subdirectory(optimizers) add_subdirectory(reduce_ops) add_subdirectory(sequence_ops) +add_subdirectory(jit) if(WITH_DISTRIBUTE) add_subdirectory(distributed) @@ -37,7 +38,12 @@ if (WITH_GPU) SET(OP_HEADER_DEPS ${OP_HEADER_DEPS} cub) endif() -register_operators(EXCLUDES warpctc_op conv_fusion_op DEPS ${OP_HEADER_DEPS}) +SET(OP_PREFETCH_DEPS "") +if (WITH_DISTRIBUTE) + SET(OP_PREFETCH_DEPS ${OP_PREFETCH_DEPS} parameter_prefetch) +endif() + +register_operators(EXCLUDES py_func_op warpctc_op conv_fusion_op DEPS ${OP_HEADER_DEPS} ${OP_PREFETCH_DEPS}) # warpctc_op needs cudnn 7 above if (WITH_GPU AND NOT WIN32) @@ -58,13 +64,11 @@ endif() set(COMMON_OP_DEPS ${OP_HEADER_DEPS}) set(COMMON_OP_DEPS ${COMMON_OP_DEPS} selected_rows_functor selected_rows lod_tensor maxouting unpooling pooling lod_rank_table context_project sequence_pooling executor) -if (NOT WIN32) - set(COMMON_OP_DEPS ${COMMON_OP_DEPS} dynload_warpctc) -endif() -set(COMMON_OP_DEPS ${COMMON_OP_DEPS} sequence_padding sequence_scale cos_sim_functor memory jit_kernel concat_and_split cross_entropy softmax vol2col im2col sampler) +set(COMMON_OP_DEPS ${COMMON_OP_DEPS} dynload_warpctc) +set(COMMON_OP_DEPS ${COMMON_OP_DEPS} sequence_padding sequence_scale cos_sim_functor memory jit_kernel_helper concat_and_split cross_entropy softmax vol2col im2col sampler) set(COMMON_OP_DEPS ${COMMON_OP_DEPS} sequence2batch lstm_compute matrix_bit_code gru_compute activation_functions) if (WITH_GPU) - set(COMMON_OP_DEPS ${COMMON_OP_DEPS} depthwise_conv) + set(COMMON_OP_DEPS ${COMMON_OP_DEPS} depthwise_conv prelu) endif() # FIXME(typhoonzero): operator deps may not needed. @@ -88,4 +92,8 @@ cc_test(save_load_op_test SRCS save_load_op_test.cc DEPS save_op load_op) cc_test(save_load_combine_op_test SRCS save_load_combine_op_test.cc DEPS save_combine_op load_combine_op) nv_test(dropout_op_test SRCS dropout_op_test.cc DEPS dropout_op tensor) +if (WITH_PYTHON) + cc_library(py_func_op SRCS py_func_op.cc DEPS op_registry python pybind) +endif() + set(GLOB_OP_LIB ${OP_LIBRARY} CACHE INTERNAL "Global OP library") diff --git a/paddle/fluid/operators/activation_mkldnn_op.cc b/paddle/fluid/operators/activation_mkldnn_op.cc index 64649b1a5e471..e16b6f78d16ce 100644 --- a/paddle/fluid/operators/activation_mkldnn_op.cc +++ b/paddle/fluid/operators/activation_mkldnn_op.cc @@ -100,8 +100,9 @@ void eltwise_forward(const framework::ExecutionContext &ctx, const T *x_data = x->data(); T *y_data = y->mutable_data(ctx.GetPlace()); - PADDLE_ENFORCE(x->dims().size() == 2 || x->dims().size() == 4, - "Input dim must be with 2 or 4"); + PADDLE_ENFORCE( + x->dims().size() == 2 || x->dims().size() == 3 || x->dims().size() == 4, + "Input dim must be with 2, 3 or 4"); std::vector src_tz = framework::vectorize2int(x->dims()); diff --git a/paddle/fluid/operators/activation_op.cc b/paddle/fluid/operators/activation_op.cc index bb9ea3f3ba087..9c5b8604f40ae 100644 --- a/paddle/fluid/operators/activation_op.cc +++ b/paddle/fluid/operators/activation_op.cc @@ -76,8 +76,8 @@ framework::OpKernelType GetKernelType(const framework::ExecutionContext& ctx, } #endif return framework::OpKernelType( - framework::ToDataType(ctx.Input(name)->type()), - ctx.GetPlace(), layout, library); + framework::GetDataTypeOfVar(ctx.InputVar(name)), ctx.GetPlace(), layout, + library); } class ActivationOp : public framework::OperatorWithKernel { @@ -149,6 +149,13 @@ Relu Activation Operator. )DOC"; +UNUSED constexpr char GeluDoc[] = R"DOC( +Gelu Activation Operator. + +$out = \\frac{1 + erf(\\frac{x}{\\sqrt{2}})}{2} x$ + +)DOC"; + UNUSED constexpr char TanhDoc[] = R"DOC( Tanh Activation Operator. @@ -472,6 +479,7 @@ REGISTER_ACTIVATION_OP_MAKER(Sigmoid, SigmoidDoc); REGISTER_ACTIVATION_OP_MAKER(LogSigmoid, LogSigmoidDoc); REGISTER_ACTIVATION_OP_MAKER(Exp, ExpDoc); REGISTER_ACTIVATION_OP_MAKER(Relu, ReluDoc); +REGISTER_ACTIVATION_OP_MAKER(Gelu, GeluDoc); REGISTER_ACTIVATION_OP_MAKER(Tanh, TanhDoc); REGISTER_ACTIVATION_OP_MAKER(TanhShrink, TanhShrinkDoc); REGISTER_ACTIVATION_OP_MAKER(Sqrt, SqrtDoc); @@ -489,6 +497,7 @@ REGISTER_ACTIVATION_OP_MAKER(Softsign, SoftsignDoc); REGISTER_ACTIVATION_OP_GRAD_MAKER(Sigmoid, sigmoid); REGISTER_ACTIVATION_OP_GRAD_MAKER(Relu, relu); +REGISTER_ACTIVATION_OP_GRAD_MAKER(Gelu, gelu); REGISTER_ACTIVATION_OP_GRAD_MAKER(Exp, exp); REGISTER_ACTIVATION_OP_GRAD_MAKER(Tanh, tanh); REGISTER_ACTIVATION_OP_GRAD_MAKER(Ceil, ceil); @@ -525,6 +534,7 @@ namespace ops = paddle::operators; __macro(Round, round); \ __macro(Log, log); \ __macro(Square, square); \ + __macro(Gelu, gelu); \ __macro(BRelu, brelu); \ __macro(Pow, pow); \ __macro(STanh, stanh); \ diff --git a/paddle/fluid/operators/activation_op.h b/paddle/fluid/operators/activation_op.h index 4ffc7f364bcb9..c7df3ea58a915 100644 --- a/paddle/fluid/operators/activation_op.h +++ b/paddle/fluid/operators/activation_op.h @@ -16,6 +16,11 @@ limitations under the License. */ #include #include +#include +#ifndef _USE_MATH_DEFINES +#define _USE_MATH_DEFINES +#endif + #include "paddle/fluid/framework/eigen.h" #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/operators/detail/safe_ref.h" @@ -36,6 +41,12 @@ static std::unordered_set InplaceOpSet = { "floor", "reciprocal", "relu6", "soft_relu", "hard_sigmoid", }; +/* The following operator can be used to process SelectedRows, because the + * output of those operator for zero is zero too. + */ +static std::unordered_set CanBeUsedBySelectedRows = { + "abs", "abs_grad", "square", "square_grad", "sqrt", "sqrt_grad"}; + static bool IsInplace(std::string op) { return InplaceOpSet.count(op); } template @@ -45,16 +56,38 @@ class ActivationKernel using T = typename Functor::ELEMENT_TYPE; void Compute(const framework::ExecutionContext& context) const override { - auto& X = detail::Ref(context.Input("X"), - "Cannot get input tensor X, variable name = %s", - context.op().Input("X")); - - auto& Out = detail::Ref(context.Output("Out"), - "Cannot get output tensor Out, variable name = %s", - context.op().Output("Out")); - Out.mutable_data(context.GetPlace()); + auto x_var = context.InputVar("X"); + auto out_var = context.OutputVar("Out"); + PADDLE_ENFORCE(x_var != nullptr, + "Cannot get input Variable X, variable name = %s", + context.op().Input("X")); + PADDLE_ENFORCE(out_var != nullptr, + "Cannot get output Variable Out, variable name = %s", + context.op().Output("Out")); + + framework::Tensor X, *Out; + + if (CanBeUsedBySelectedRows.count(context.op().Type())) { + X = detail::Ref( + paddle::framework::GetLoDTensorOrSelectedRowsValueFromVar(*x_var), + "Cannot get input Tensor X, variable name = %s", + context.op().Input("X")); + Out = paddle::framework::GetMutableLoDTensorOrSelectedRowsValueFromVar( + out_var); + } else { + X = detail::Ref(context.Input("X"), + "Cannot get input Tensor X, variable name = %s", + context.op().Input("X")); + Out = context.Output("Out"); + } + + PADDLE_ENFORCE(Out != nullptr, + "Cannot get output tensor Out, variable name = %s", + context.op().Output("Out")); + + Out->mutable_data(context.GetPlace()); auto x = framework::EigenVector::Flatten(X); - auto out = framework::EigenVector::Flatten(Out); + auto out = framework::EigenVector::Flatten(*Out); auto* place = context.template device_context().eigen_device(); Functor functor; @@ -73,14 +106,54 @@ class ActivationGradKernel public: using T = typename Functor::ELEMENT_TYPE; void Compute(const framework::ExecutionContext& context) const override { - auto* Out = context.Input("Out"); - auto* dOut = - context.Input(framework::GradVarName("Out")); - auto* dX = context.Output(framework::GradVarName("X")); + auto out_var = context.InputVar("Out"); + auto out_grad_var = context.InputVar(framework::GradVarName("Out")); + auto x_grad_var = context.OutputVar(framework::GradVarName("X")); + PADDLE_ENFORCE(out_var != nullptr, + "Cannot get input Variable Out, variable name = %s", + context.op().Input("Out")); + PADDLE_ENFORCE(out_grad_var != nullptr, + "Cannot get input Variable %s, variable name = %s", + framework::GradVarName("Out"), + context.op().Input(framework::GradVarName("Out"))); + PADDLE_ENFORCE(x_grad_var != nullptr, + "Cannot get output Variable %s, variable name = %s", + framework::GradVarName("X"), + context.op().Output(framework::GradVarName("X"))); + + framework::Tensor Out, dOut, *dX; + if (CanBeUsedBySelectedRows.count(context.op().Type())) { + Out = detail::Ref( + paddle::framework::GetLoDTensorOrSelectedRowsValueFromVar(*out_var), + "Cannot get input Tensor Out, variable name = %s", + context.op().Input("Out")); + dOut = + detail::Ref(paddle::framework::GetLoDTensorOrSelectedRowsValueFromVar( + *out_grad_var), + "Cannot get input Tensor %s, variable name = %s", + framework::GradVarName("Out"), + context.op().Input(framework::GradVarName("Out"))); + dX = paddle::framework::GetMutableLoDTensorOrSelectedRowsValueFromVar( + x_grad_var); + } else { + Out = detail::Ref(context.Input("Out"), + "Cannot get input Tensor Out, variable name = %s", + context.op().Input("Out")); + dOut = detail::Ref( + context.Input(framework::GradVarName("Out")), + "Cannot get input Tensor %s, variable name = %s", + framework::GradVarName("Out"), + context.op().Input(framework::GradVarName("Out"))); + dX = context.Output(framework::GradVarName("X")); + } + PADDLE_ENFORCE(dX != nullptr, + "Cannot get output tensor %s, variable name = %s", + framework::GradVarName("X"), + context.op().Output(framework::GradVarName("X"))); dX->mutable_data(context.GetPlace()); - auto dout = framework::EigenVector::Flatten(*dOut); - auto out = framework::EigenVector::Flatten(*Out); + auto dout = framework::EigenVector::Flatten(dOut); + auto out = framework::EigenVector::Flatten(Out); auto dx = framework::EigenVector::Flatten(*dX); auto* place = context.template device_context().eigen_device(); @@ -91,11 +164,22 @@ class ActivationGradKernel } bool inplace = functor.Inplace(); if (!inplace) { - auto* X = context.Input("X"); - auto x = framework::EigenVector::Flatten(*X); + auto x_var = context.InputVar("X"); + PADDLE_ENFORCE(x_var != nullptr, + "Cannot get input tensor X, variable name = %s", + context.op().Input("X")); + framework::Tensor X; + if (CanBeUsedBySelectedRows.count(context.op().Type())) { + X = detail::Ref( + paddle::framework::GetLoDTensorOrSelectedRowsValueFromVar(*x_var)); + } else { + X = detail::Ref(context.Input("X")); + } + + auto x = framework::EigenVector::Flatten(X); functor(*place, x, out, dout, dx); } else { - VLOG(100) << " Inplace activation "; + VLOG(10) << " Inplace activation "; auto x = framework::EigenVector::Flatten(*dX); functor(*place, x, out, dout, dx); } @@ -212,6 +296,30 @@ struct ReluGradFunctor : public BaseActivationFunctor { } }; +// gelu(x) = 0.5 * x * (1 + erf(x / sqrt(2))) +template +struct GeluFunctor : public BaseActivationFunctor { + template + void operator()(Device d, X x, Out out) const { + auto temp = (x * static_cast(M_SQRT1_2)).erf(); + out.device(d) = x * static_cast(0.5) * (static_cast(1) + temp); + } +}; + +template +struct GeluGradFunctor : BaseActivationFunctor { + template + void operator()(Device d, X x, Out out, dOut dout, dX dx) const { + auto first = static_cast(0.5) * + (static_cast(1) + ((x * static_cast(M_SQRT1_2)).erf())); + + auto second = static_cast(0.5 * M_2_SQRTPI * M_SQRT1_2) * x * + (-static_cast(0.5) * x.square()).exp(); + dx.device(d) = dout * (first + second); + } +}; + // tanh(x) = (exp(x) - exp(-x)) / (exp(x) + exp(-x)) template struct TanhFunctor : public BaseActivationFunctor { @@ -877,6 +985,7 @@ struct SwishGradFunctor : public BaseActivationFunctor { __macro(logsigmoid, LogSigmoidFunctor, LogSigmoidGradFunctor); \ __macro(exp, ExpFunctor, ExpGradFunctor); \ __macro(relu, ReluFunctor, ReluGradFunctor); \ + __macro(gelu, GeluFunctor, GeluGradFunctor); \ __macro(tanh, TanhFunctor, TanhGradFunctor); \ __macro(softshrink, SoftShrinkFunctor, SoftShrinkGradFunctor); \ __macro(sqrt, SqrtFunctor, SqrtGradFunctor); \ diff --git a/paddle/fluid/operators/affine_grid_op.cc b/paddle/fluid/operators/affine_grid_op.cc index 6f7da445fc84f..1de59a5165c83 100644 --- a/paddle/fluid/operators/affine_grid_op.cc +++ b/paddle/fluid/operators/affine_grid_op.cc @@ -78,7 +78,7 @@ class AffineGridOp : public framework::OperatorWithKernel { library = framework::LibraryType::kCUDNN; } #endif - auto data_type = framework::ToDataType(ctx.Input("Theta")->type()); + auto data_type = ctx.Input("Theta")->type(); return framework::OpKernelType(data_type, ctx.GetPlace(), framework::DataLayout::kAnyLayout, library); } @@ -188,9 +188,9 @@ class AffineGridOpGrad : public framework::OperatorWithKernel { library_ = framework::LibraryType::kCUDNN; } #endif - return framework::OpKernelType( - framework::ToDataType(ctx.Input("Theta")->type()), - ctx.GetPlace(), framework::DataLayout::kAnyLayout, library_); + return framework::OpKernelType(ctx.Input("Theta")->type(), + ctx.GetPlace(), + framework::DataLayout::kAnyLayout, library_); } }; diff --git a/paddle/fluid/operators/arg_max_op.cc b/paddle/fluid/operators/arg_max_op.cc index 8174d3735859b..7fe9a0df74679 100644 --- a/paddle/fluid/operators/arg_max_op.cc +++ b/paddle/fluid/operators/arg_max_op.cc @@ -28,6 +28,5 @@ REGISTER_OP_CPU_KERNEL( int32_t>, paddle::operators::ArgMaxKernel, - paddle::operators::ArgMaxKernel, paddle::operators::ArgMaxKernel); diff --git a/paddle/fluid/operators/arg_max_op.cu b/paddle/fluid/operators/arg_max_op.cu index a147d77a9e9c5..85e4f98173511 100644 --- a/paddle/fluid/operators/arg_max_op.cu +++ b/paddle/fluid/operators/arg_max_op.cu @@ -25,7 +25,5 @@ REGISTER_OP_CUDA_KERNEL( int32_t>, paddle::operators::ArgMaxKernel, - paddle::operators::ArgMaxKernel, paddle::operators::ArgMaxKernel); diff --git a/paddle/fluid/operators/arg_min_op.cc b/paddle/fluid/operators/arg_min_op.cc index 41f188029f17d..23b24735cd0ba 100644 --- a/paddle/fluid/operators/arg_min_op.cc +++ b/paddle/fluid/operators/arg_min_op.cc @@ -28,6 +28,5 @@ REGISTER_OP_CPU_KERNEL( int32_t>, paddle::operators::ArgMinKernel, - paddle::operators::ArgMinKernel, paddle::operators::ArgMinKernel); diff --git a/paddle/fluid/operators/arg_min_op.cu b/paddle/fluid/operators/arg_min_op.cu index 4d020508505a6..47d7c8b12243c 100644 --- a/paddle/fluid/operators/arg_min_op.cu +++ b/paddle/fluid/operators/arg_min_op.cu @@ -25,7 +25,5 @@ REGISTER_OP_CUDA_KERNEL( int32_t>, paddle::operators::ArgMinKernel, - paddle::operators::ArgMinKernel, paddle::operators::ArgMinKernel); diff --git a/paddle/fluid/operators/array_operator.h b/paddle/fluid/operators/array_operator.h index eddf34494bdab..4309f0a549745 100644 --- a/paddle/fluid/operators/array_operator.h +++ b/paddle/fluid/operators/array_operator.h @@ -49,7 +49,7 @@ class ArrayOp : public framework::OperatorBase { } else { offset = static_cast(*i_tensor.data()); } - VLOG(100) << " Offset = " << offset; + VLOG(10) << " Offset = " << offset; return offset; } }; diff --git a/paddle/fluid/operators/array_to_lod_tensor_op.cc b/paddle/fluid/operators/array_to_lod_tensor_op.cc index 3c40135eca00f..d942391b86449 100644 --- a/paddle/fluid/operators/array_to_lod_tensor_op.cc +++ b/paddle/fluid/operators/array_to_lod_tensor_op.cc @@ -58,7 +58,7 @@ struct ArrayToLoDFunctor : public boost::static_visitor { ArrayToLoDFunctorImpl functor; functor.dev_ctx_ = dev_ctx; functor.prev_functor_ = this; - framework::VisitDataType(framework::ToDataType(out->type()), functor); + framework::VisitDataType(out->type(), functor); } }; @@ -91,7 +91,7 @@ class ArrayToLoDTensorOp : public framework::OperatorBase { PADDLE_ENFORCE(!x.empty(), "There's no element in the input array."); int rank = x[0].dims().size(); platform::Place place = x[0].place(); - std::type_index data_type = x[0].type(); + auto data_type = x[0].type(); int64_t batch_size = x[0].dims()[0]; framework::DDim ins_dims = rank > 1 ? framework::slice_ddim(x[0].dims(), 1, rank) @@ -148,8 +148,8 @@ class ArrayToLoDTensorOp : public framework::OperatorBase { size_t start_offset = lod_and_offset.second.first; size_t end_offset = lod_and_offset.second.second; - VLOG(100) << "idx=" << idx << " x_idx=" << x_idx << " [" - << ", " << end_offset << "]"; + VLOG(10) << "idx=" << idx << " x_idx=" << x_idx << " [" + << ", " << end_offset << "]"; // Copy data PADDLE_ENFORCE_GE(end_offset, start_offset); size_t len = end_offset - start_offset; diff --git a/paddle/fluid/operators/attention_lstm_op.cc b/paddle/fluid/operators/attention_lstm_op.cc index 9b943440a869e..b6996be4b0984 100644 --- a/paddle/fluid/operators/attention_lstm_op.cc +++ b/paddle/fluid/operators/attention_lstm_op.cc @@ -121,9 +121,8 @@ void AttentionLSTMOp::InferShape(framework::InferShapeContext* ctx) const { framework::OpKernelType AttentionLSTMOp::GetExpectedKernelType( const framework::ExecutionContext& ctx) const { - return framework::OpKernelType( - framework::ToDataType(ctx.Input("X")->type()), - ctx.device_context()); + return framework::OpKernelType(ctx.Input("X")->type(), + ctx.device_context()); } void AttentionLSTMOpMaker::Make() { @@ -231,10 +230,10 @@ use lstm_x_t as input and compute as standard LSTM. template inline void bias_relu(const int n, const T* x, const T* bias, T* y) { if (bias) { - math::vec_add_bias(n, *bias, x, y); - math::vec_relu(n, y, y); + math::vec_add_bias(n, *bias, x, y); + math::vec_relu(n, y, y); } else { - math::vec_relu(n, x, y); + math::vec_relu(n, x, y); } } @@ -245,8 +244,8 @@ inline void vec_softmax(const int n, const T* x, T* y) { for (int i = 1; i < n; ++i) { scalar = scalar < x[i] ? x[i] : scalar; } - math::vec_add_bias(n, -scalar, x, y); // sub - math::vec_exp(n, y, y); // exp + math::vec_add_bias(n, -scalar, x, y); // sub + math::vec_exp(n, y, y); // exp // sum scalar = T(0); for (int i = 0; i < n; ++i) { @@ -302,13 +301,13 @@ class AttentionLSTMKernel : public framework::OpKernel { auto& act_gate_str = ctx.Attr("gate_activation"); auto& act_cell_str = ctx.Attr("cell_activation"); auto& act_cand_str = ctx.Attr("candidate_activation"); - if (platform::jit::MayIUse(platform::jit::avx)) { - math::VecActivations act_functor; + if (platform::MayIUse(platform::avx)) { + math::VecActivations act_functor; act_gate = act_functor(act_gate_str); act_cell = act_functor(act_cell_str); act_cand = act_functor(act_cand_str); } else { - math::VecActivations act_functor; + math::VecActivations act_functor; act_gate = act_functor(act_gate_str); act_cell = act_functor(act_cell_str); act_cand = act_functor(act_cand_str); diff --git a/paddle/fluid/operators/average_accumulates_op.cc b/paddle/fluid/operators/average_accumulates_op.cc index f389eab605e08..0922b03b5f5fb 100644 --- a/paddle/fluid/operators/average_accumulates_op.cc +++ b/paddle/fluid/operators/average_accumulates_op.cc @@ -103,9 +103,8 @@ class AverageAccumulatesOp : public framework::OperatorWithKernel { protected: framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext& ctx) const override { - return framework::OpKernelType( - framework::ToDataType(ctx.Input("param")->type()), - ctx.GetPlace()); + return framework::OpKernelType(ctx.Input("param")->type(), + ctx.GetPlace()); } }; diff --git a/paddle/fluid/operators/batch_norm_mkldnn_op.cc b/paddle/fluid/operators/batch_norm_mkldnn_op.cc index de641cb08e4cc..bddca232e6c8a 100644 --- a/paddle/fluid/operators/batch_norm_mkldnn_op.cc +++ b/paddle/fluid/operators/batch_norm_mkldnn_op.cc @@ -14,7 +14,7 @@ limitations under the License. */ #include "mkldnn.hpp" #include "paddle/fluid/operators/batch_norm_op.h" -#include "paddle/fluid/platform/mkldnn_helper.h" +#include "paddle/fluid/platform/mkldnn_reuse.h" namespace paddle { namespace operators { @@ -146,7 +146,9 @@ class BatchNormMKLDNNOpKernel : public paddle::framework::OpKernel { const float epsilon = ctx.Attr("epsilon"); const float momentum = ctx.Attr("momentum"); const bool is_test = ctx.Attr("is_test"); + const bool use_global_stats = ctx.Attr("use_global_stats"); const bool fuse_with_relu = ctx.Attr("fuse_with_relu"); + bool global_stats = is_test || use_global_stats; const auto *x = ctx.Input("X"); const auto *mean = ctx.Input("Mean"); @@ -177,13 +179,14 @@ class BatchNormMKLDNNOpKernel : public paddle::framework::OpKernel { T *batch_mean_data = nullptr; T *batch_variance_data = nullptr; - if (!is_test) { + if (!global_stats) { batch_mean_data = batch_mean->mutable_data(ctx.GetPlace()); batch_variance_data = batch_variance->mutable_data(ctx.GetPlace()); } - auto propagation = is_test == true ? mkldnn::prop_kind::forward_scoring - : mkldnn::prop_kind::forward_training; + auto propagation = global_stats == true + ? mkldnn::prop_kind::forward_scoring + : mkldnn::prop_kind::forward_training; auto src_tz = paddle::framework::vectorize2int(x->dims()); auto scale_tz = paddle::framework::vectorize2int(scale->dims()); @@ -199,7 +202,7 @@ class BatchNormMKLDNNOpKernel : public paddle::framework::OpKernel { shift->data() + ic, &scaleshift_data); unsigned flags = mkldnn::use_scale_shift; - if (is_test) flags |= mkldnn::use_global_stats; + if (global_stats) flags |= mkldnn::use_global_stats; if (fuse_with_relu) flags |= mkldnn::fuse_bn_relu; // create mkldnn memory from input x tensor @@ -208,7 +211,7 @@ class BatchNormMKLDNNOpKernel : public paddle::framework::OpKernel { // keys for backward pass const std::string key = BatchNormMKLDNNHandler::GetHash( - src_tz, epsilon, flags, is_test, input_format, + src_tz, epsilon, flags, global_stats, input_format, ctx.op().Output("SavedMean")); const std::string key_batch_norm_fwd_pd = key + "@bn_fwd_pd"; @@ -239,7 +242,7 @@ class BatchNormMKLDNNOpKernel : public paddle::framework::OpKernel { batch_norm_fwd_pd->dst_primitive_desc().desc(), y_data); std::shared_ptr batch_norm_p; - if (is_test) { + if (global_stats) { // create mkldnn memory for stats (as input) std::shared_ptr mean_memory = handler.AcquireMeanMemoryFromPrimitive(to_void_cast(mean_data)); @@ -269,7 +272,7 @@ class BatchNormMKLDNNOpKernel : public paddle::framework::OpKernel { pipeline.push_back(*batch_norm_p); mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); - if (!is_test) { + if (!global_stats) { // mkldnn only compute stats for current batch // so we need compute momentum stats via Eigen lib EigenVectorArrayMap batch_mean_e(batch_mean_data, ic); diff --git a/paddle/fluid/operators/batch_norm_op.cc b/paddle/fluid/operators/batch_norm_op.cc index 2463c939bc5d1..8b672e09b2c5c 100644 --- a/paddle/fluid/operators/batch_norm_op.cc +++ b/paddle/fluid/operators/batch_norm_op.cc @@ -72,8 +72,7 @@ class BatchNormOp : public framework::OperatorWithKernel { protected: framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext &ctx) const override { - auto input_data_type = - framework::ToDataType(ctx.Input("X")->type()); + auto input_data_type = ctx.Input("X")->type(); // By default, the type of the scale, bias, mean, // and var tensors should both be float. (For float or float16 input tensor) // or double (For double input tensor). @@ -81,17 +80,13 @@ class BatchNormOp : public framework::OperatorWithKernel { if (input_data_type == framework::proto::VarType::FP64) { bn_param_type = framework::proto::VarType::FP64; } - PADDLE_ENFORCE_EQ(bn_param_type, - framework::ToDataType(ctx.Input("Scale")->type()), + PADDLE_ENFORCE_EQ(bn_param_type, ctx.Input("Scale")->type(), "Scale input should be of float type"); - PADDLE_ENFORCE_EQ(bn_param_type, - framework::ToDataType(ctx.Input("Bias")->type()), + PADDLE_ENFORCE_EQ(bn_param_type, ctx.Input("Bias")->type(), "Bias input should be of float type"); - PADDLE_ENFORCE_EQ(bn_param_type, - framework::ToDataType(ctx.Input("Mean")->type()), + PADDLE_ENFORCE_EQ(bn_param_type, ctx.Input("Mean")->type(), "Mean input should be of float type"); - PADDLE_ENFORCE_EQ(bn_param_type, framework::ToDataType( - ctx.Input("Variance")->type()), + PADDLE_ENFORCE_EQ(bn_param_type, ctx.Input("Variance")->type(), "Variance input should be of float type"); // TODO(pzelazko-intel): enable MKLDNN layout when it's ready @@ -159,6 +154,14 @@ class BatchNormOpMaker : public framework::OpProtoAndCheckerMaker { AddAttr("fuse_with_relu", "(bool, default false) Only used in mkldnn kernel") .SetDefault(false); + AddAttr("use_global_stats", + "(bool, default false) Whether to use global mean and " + "variance. In inference or test mode, set use_global_stats " + "to true or is_test true. the behavior is equivalent. " + "In train mode, when setting use_global_stats True, the " + "global mean and variance are also used during train time, " + "the BN acts as scaling and shiffting.") + .SetDefault(false); AddComment(R"DOC( Batch Normalization. @@ -190,6 +193,10 @@ class BatchNormKernel const float epsilon = ctx.Attr("epsilon"); const float momentum = ctx.Attr("momentum"); const bool is_test = ctx.Attr("is_test"); + const bool use_global_stats = ctx.Attr("use_global_stats"); + + bool global_stats = is_test || use_global_stats; + const std::string data_layout_str = ctx.Attr("data_layout"); const DataLayout data_layout = framework::StringToDataLayout(data_layout_str); @@ -217,7 +224,7 @@ class BatchNormKernel saved_mean->mutable_data(ctx.GetPlace()); saved_variance->mutable_data(ctx.GetPlace()); - if (!is_test) { + if (!global_stats) { // saved_xx is use just in this batch of data EigenVectorArrayMap saved_mean_e( saved_mean->mutable_data(ctx.GetPlace()), C); @@ -234,7 +241,7 @@ class BatchNormKernel if ((N * sample_size) == 1) { LOG(WARNING) << "Only 1 element in normalization dimension, " << "we skip the batch norm calculation, let y = x."; - framework::TensorCopySync(*x, ctx.GetPlace(), y); + framework::TensorCopy(*x, ctx.GetPlace(), y); return; } @@ -277,7 +284,7 @@ class BatchNormKernel // use SavedMean and SavedVariance to do normalize Eigen::Array inv_std(C); - if (is_test) { + if (global_stats) { ConstEigenVectorArrayMap var_arr( ctx.Input("Variance")->data(), C); inv_std = (var_arr + epsilon).sqrt().inverse(); @@ -289,8 +296,8 @@ class BatchNormKernel inv_std = saved_inv_std; } ConstEigenVectorArrayMap mean_arr( - is_test ? ctx.Input("Mean")->data() - : ctx.Output("SavedMean")->data(), + global_stats ? ctx.Input("Mean")->data() + : ctx.Output("SavedMean")->data(), C); // ((x - est_mean) * (inv_var) * scale + bias @@ -336,15 +343,27 @@ class BatchNormGradOp : public framework::OperatorWithKernel { void InferShape(framework::InferShapeContext *ctx) const override { // check input PADDLE_ENFORCE(ctx->HasInput("X")); - PADDLE_ENFORCE(ctx->HasInput("Scale"), ""); - PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Y")), ""); - PADDLE_ENFORCE(ctx->HasInput("SavedMean"), ""); - PADDLE_ENFORCE(ctx->HasInput("SavedVariance"), ""); + PADDLE_ENFORCE(ctx->HasInput("Scale"), "Input(scale) should not be null."); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Y")), + "Input(Y@GRAD) should not be null."); + PADDLE_ENFORCE(ctx->HasInput("SavedMean"), + "Input(SavedMean) should not be null."); + PADDLE_ENFORCE(ctx->HasInput("SavedVariance"), + "Input(SavedVariance) should not be null"); // check output PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("X")), ""); - PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("Scale")), ""); - PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("Bias")), ""); + if (ctx->HasOutput(framework::GradVarName("Scale"))) { + PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("Bias")), + "Output(Scale@GRAD) and Output(Bias@GRAD) should not be " + "null at same time"); + } + const bool use_global_stats = ctx->Attrs().Get("use_global_stats"); + if (use_global_stats) { + PADDLE_ENFORCE(!ctx->Attrs().Get("use_mkldnn"), + "Using global stats during training is not supported " + "in gradient op kernel of batch_norm_mkldnn_op now."); + } const auto x_dims = ctx->GetInputDim("X"); const DataLayout data_layout = framework::StringToDataLayout( @@ -354,8 +373,10 @@ class BatchNormGradOp : public framework::OperatorWithKernel { : x_dims[x_dims.size() - 1]); ctx->SetOutputDim(framework::GradVarName("X"), x_dims); - ctx->SetOutputDim(framework::GradVarName("Scale"), {C}); - ctx->SetOutputDim(framework::GradVarName("Bias"), {C}); + if (ctx->HasOutput(framework::GradVarName("Scale"))) { + ctx->SetOutputDim(framework::GradVarName("Scale"), {C}); + ctx->SetOutputDim(framework::GradVarName("Bias"), {C}); + } } protected: @@ -387,9 +408,8 @@ class BatchNormGradOp : public framework::OperatorWithKernel { } #endif - return framework::OpKernelType( - framework::ToDataType(ctx.Input("X")->type()), ctx.GetPlace(), - layout, library); + return framework::OpKernelType(ctx.Input("X")->type(), + ctx.GetPlace(), layout, library); } }; @@ -405,6 +425,8 @@ class BatchNormGradKernel // SavedVariance have been reverted in forward operator const auto *saved_inv_variance = ctx.Input("SavedVariance"); const std::string data_layout_str = ctx.Attr("data_layout"); + const bool use_global_stats = ctx.Attr("use_global_stats"); + const float epsilon = ctx.Attr("epsilon"); const DataLayout data_layout = framework::StringToDataLayout(data_layout_str); @@ -419,38 +441,60 @@ class BatchNormGradKernel : x_dims[x_dims.size() - 1]); const int sample_size = x->numel() / N / C; - ConstEigenVectorArrayMap scale_arr(scale->data(), C); - ConstEigenVectorArrayMap mean_arr(saved_mean->data(), C); - ConstEigenVectorArrayMap inv_var_arr(saved_inv_variance->data(), C); - // init output auto *d_x = ctx.Output(framework::GradVarName("X")); auto *d_scale = ctx.Output(framework::GradVarName("Scale")); auto *d_bias = ctx.Output(framework::GradVarName("Bias")); d_x->mutable_data(ctx.GetPlace()); - d_scale->mutable_data(ctx.GetPlace()); - d_bias->mutable_data(ctx.GetPlace()); + + const T *mean_data = saved_mean->data(); + const T *inv_var_data = saved_inv_variance->data(); + Tensor inv_var_tensor; + if (use_global_stats) { + const auto *running_mean = ctx.Input("Mean"); + const auto *running_variance = ctx.Input("Variance"); + mean_data = running_mean->data(); + T *running_inv_var_data = inv_var_tensor.mutable_data(ctx.GetPlace()); + EigenVectorArrayMap inv_var_tmp(running_inv_var_data, C); + ConstEigenVectorArrayMap var_arr(running_variance->data(), C); + + inv_var_tmp = (var_arr + epsilon).sqrt().inverse().eval(); + inv_var_data = running_inv_var_data; + } + + ConstEigenVectorArrayMap scale_arr(scale->data(), C); + ConstEigenVectorArrayMap mean_arr(mean_data, C); + ConstEigenVectorArrayMap inv_var_arr(inv_var_data, C); + + T *d_bias_data = nullptr; + T *d_scale_data = nullptr; + if (d_scale && d_bias) { + d_scale->mutable_data(ctx.GetPlace()); + d_bias->mutable_data(ctx.GetPlace()); + d_bias_data = d_bias->mutable_data(ctx.GetPlace()); + d_scale_data = d_scale->mutable_data(ctx.GetPlace()); + } // d_bias = np.sum(d_y, axis=0) // d_scale = np.sum((X - mean) / inv_std * dy, axis=0) // d_x = (1. / N) * scale * inv_var * (N * d_y - np.sum(d_y, axis=0) // - (X - mean) * inv_var * inv_var * np.sum(d_y * (X - mean), axis=0)) + EigenVectorArrayMap d_bias_arr(d_bias_data, C); + EigenVectorArrayMap d_scale_arr(d_scale_data, C); - EigenVectorArrayMap d_bias_arr(d_bias->mutable_data(ctx.GetPlace()), - C); - EigenVectorArrayMap d_scale_arr(d_scale->mutable_data(ctx.GetPlace()), - C); - - d_bias_arr.setZero(); - d_scale_arr.setZero(); + if (d_scale && d_bias) { + d_bias_arr.setZero(); + d_scale_arr.setZero(); + } - if ((N * sample_size) == 1) { - framework::TensorCopySync(*d_y, ctx.GetPlace(), d_x); + if ((N * sample_size) == 1 && !use_global_stats) { + framework::TensorCopy(*d_y, ctx.GetPlace(), d_x); return; } - const auto scale_inv_var_nhw = scale_arr * inv_var_arr / (N * sample_size); + int scale_coefff = use_global_stats ? 1 : N * sample_size; + const auto scale_inv_var_nhw = scale_arr * inv_var_arr / scale_coefff; switch (data_layout) { case DataLayout::kNCHW: { @@ -460,19 +504,29 @@ class BatchNormGradKernel sample_size, N * C); d_x_arr.setZero(); - for (int nc = 0; nc < N * C; ++nc) { - int c = nc % C; - d_bias_arr(c) += d_y_arr.col(nc).sum(); - d_scale_arr(c) += - ((x_arr.col(nc) - mean_arr(c)) * inv_var_arr(c) * d_y_arr.col(nc)) - .sum(); + if (d_scale && d_bias) { + for (int nc = 0; nc < N * C; ++nc) { + int c = nc % C; + d_bias_arr(c) += d_y_arr.col(nc).sum(); + d_scale_arr(c) += ((x_arr.col(nc) - mean_arr(c)) * inv_var_arr(c) * + d_y_arr.col(nc)) + .sum(); + } } - for (int nc = 0; nc < N * C; ++nc) { - int c = nc % C; - d_x_arr.col(nc) += - scale_inv_var_nhw(c) * - (d_y_arr.col(nc) * N * sample_size - d_bias_arr(c) - - (x_arr.col(nc) - mean_arr[c]) * d_scale_arr(c) * inv_var_arr(c)); + if (!use_global_stats) { + for (int nc = 0; nc < N * C; ++nc) { + int c = nc % C; + d_x_arr.col(nc) += + scale_inv_var_nhw(c) * + (d_y_arr.col(nc) * N * sample_size - d_bias_arr(c) - + (x_arr.col(nc) - mean_arr[c]) * d_scale_arr(c) * + inv_var_arr(c)); + } + } else { + for (int nc = 0; nc < N * C; ++nc) { + int c = nc % C; + d_x_arr.col(nc) += scale_inv_var_nhw(c) * d_y_arr.col(nc); + } } break; } @@ -488,15 +542,27 @@ class BatchNormGradKernel const auto d_y_mul_x_minus_mean_row_sum = (d_y_arr * x_minus_mean).rowwise().sum(); const auto inv_var_sqr = inv_var_arr * inv_var_arr; - for (int nhw = 0; nhw < N * sample_size; ++nhw) { - d_bias_arr += d_y_arr.col(nhw); - d_scale_arr += - (x_arr.col(nhw) - mean_arr) * inv_var_arr * d_y_arr.col(nhw); - d_x_arr.col(nhw) += - scale_inv_var_nhw * - (d_y_arr.col(nhw) * N * sample_size - d_y_row_sum - - x_minus_mean.col(nhw) * inv_var_sqr * - d_y_mul_x_minus_mean_row_sum); + + if (d_scale && d_bias) { + for (int nhw = 0; nhw < N * sample_size; ++nhw) { + d_bias_arr += d_y_arr.col(nhw); + d_scale_arr += + (x_arr.col(nhw) - mean_arr) * inv_var_arr * d_y_arr.col(nhw); + } + } + + if (!use_global_stats) { + for (int nhw = 0; nhw < N * sample_size; ++nhw) { + d_x_arr.col(nhw) += + scale_inv_var_nhw * + (d_y_arr.col(nhw) * N * sample_size - d_y_row_sum - + x_minus_mean.col(nhw) * inv_var_sqr * + d_y_mul_x_minus_mean_row_sum); + } + } else { + for (int nhw = 0; nhw < N * sample_size; ++nhw) { + d_x_arr.col(nhw) += scale_inv_var_nhw * d_y_arr.col(nhw); + } } break; } @@ -522,6 +588,10 @@ class BatchNormGradMaker : public framework::SingleGradOpDescMaker { op->SetInput("SavedMean", Output("SavedMean")); op->SetInput("SavedVariance", Output("SavedVariance")); + // used when setting use_global_stats True during training + op->SetInput("Mean", Output("MeanOut")); + op->SetInput("Variance", Output("VarianceOut")); + op->SetAttrMap(Attrs()); op->SetOutput(framework::GradVarName("X"), InputGrad("X")); diff --git a/paddle/fluid/operators/batch_norm_op.cu.cc b/paddle/fluid/operators/batch_norm_op.cu similarity index 57% rename from paddle/fluid/operators/batch_norm_op.cu.cc rename to paddle/fluid/operators/batch_norm_op.cu index 0609027c69405..1c45746a92ad0 100644 --- a/paddle/fluid/operators/batch_norm_op.cu.cc +++ b/paddle/fluid/operators/batch_norm_op.cu @@ -12,9 +12,13 @@ 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/operators/batch_norm_op.h" +#include #include +#include +#include +#include "cub/cub.cuh" #include "paddle/fluid/framework/data_layout.h" +#include "paddle/fluid/operators/batch_norm_op.h" #include "paddle/fluid/operators/math/math_function.h" #include "paddle/fluid/platform/cudnn_helper.h" #include "paddle/fluid/platform/float16.h" @@ -59,6 +63,7 @@ class BatchNormKernel double epsilon = static_cast(ctx.Attr("epsilon")); const float momentum = ctx.Attr("momentum"); const bool is_test = ctx.Attr("is_test"); + const bool use_global_stats = ctx.Attr("use_global_stats"); const std::string data_layout_str = ctx.Attr("data_layout"); const DataLayout data_layout = framework::StringToDataLayout(data_layout_str); @@ -96,7 +101,7 @@ class BatchNormKernel mode_ = CUDNN_BATCHNORM_SPATIAL; #endif - VLOG(30) << "Setting descriptors."; + VLOG(3) << "Setting descriptors."; std::vector dims; std::vector strides; if (data_layout == DataLayout::kNCHW) { @@ -121,7 +126,7 @@ class BatchNormKernel auto handle = dev_ctx.cudnn_handle(); // Now, depending on whether we are running test or not, we have two paths. - if (is_test) { + if (is_test || use_global_stats) { // only when test we use input to do computation. const auto *est_mean = ctx.Input("Mean"); const auto *est_var = ctx.Input("Variance"); @@ -163,7 +168,7 @@ class BatchNormKernel if ((N * H * W * D) == 1) { LOG(WARNING) << "Only 1 element in normalization dimension, " << "we skip the batch norm calculation, let y = x."; - framework::TensorCopySync(*x, ctx.GetPlace(), y); + framework::TensorCopy(*x, ctx.GetPlace(), y); } else { double this_factor = 1. - momentum; @@ -191,6 +196,58 @@ class BatchNormKernel } }; +template +static __global__ void KeBNBackwardData(const T *dy, + const BatchNormParamType *scale, + const BatchNormParamType *variance, + const double epsilon, const int C, + const int HxW, const int num, T *dx) { + int gid = blockIdx.x * blockDim.x + threadIdx.x; + int stride = blockDim.x * gridDim.x; + for (int i = gid; i < num; i += stride) { + const int c = layout == framework::DataLayout::kNCHW ? i / HxW % C : i % C; + BatchNormParamType inv_var = 1.0 / sqrt(variance[c] + epsilon); + dx[i] = static_cast(static_cast>(dy[i]) * + scale[c] * inv_var); + } +} + +template +static __global__ void KeBNBackwardScaleBias( + const T *dy, const T *x, const BatchNormParamType *mean, + const BatchNormParamType *variance, const double epsilon, const int N, + const int C, const int HxW, BatchNormParamType *dscale, + BatchNormParamType *dbias) { + const int outer_size = C; + const int inner_size = N * HxW; + typedef cub::BlockReduce, BlockDim> BlockReduce; + __shared__ typename BlockReduce::TempStorage ds_storage; + __shared__ typename BlockReduce::TempStorage db_storage; + + for (int i = blockIdx.x; i < outer_size; i += gridDim.x) { + BatchNormParamType ds_sum = static_cast>(0); + BatchNormParamType db_sum = static_cast>(0); + + BatchNormParamType inv_var_i = 1.0 / sqrt(variance[i] + epsilon); + BatchNormParamType mean_i = mean[i]; + for (int j = threadIdx.x; j < inner_size; j += blockDim.x) { + const int index = layout == framework::DataLayout::kNCHW + ? (j / HxW * C + i) * HxW + j % HxW + : j * outer_size + i; + ds_sum += static_cast>(dy[index]) * + (static_cast>(x[index]) - mean_i); + db_sum += static_cast>(dy[index]); + } + ds_sum = BlockReduce(ds_storage).Reduce(ds_sum, cub::Sum()); + db_sum = BlockReduce(db_storage).Reduce(db_sum, cub::Sum()); + if (threadIdx.x == 0) { + dscale[i] = ds_sum * inv_var_i; + dbias[i] = db_sum; + } + __syncthreads(); + } +} + template class BatchNormGradKernel : public framework::OpKernel { @@ -200,6 +257,8 @@ class BatchNormGradKernel "It must use CUDAPlace."); double epsilon = static_cast(ctx.Attr("epsilon")); const std::string data_layout_str = ctx.Attr("data_layout"); + const bool use_global_stats = ctx.Attr("use_global_stats"); + const DataLayout data_layout = framework::StringToDataLayout(data_layout_str); const auto *x = ctx.Input("X"); @@ -219,42 +278,13 @@ class BatchNormGradKernel auto *d_bias = ctx.Output(framework::GradVarName("Bias")); d_x->mutable_data(ctx.GetPlace()); - d_scale->mutable_data>(ctx.GetPlace()); - d_bias->mutable_data>(ctx.GetPlace()); - - auto &dev_ctx = ctx.template device_context(); - if ((N * H * W * D) == 1) { - framework::TensorCopySync(*d_y, ctx.GetPlace(), d_x); - math::SetConstant> - functor; - functor(dev_ctx, d_scale, static_cast>(0)); - functor(dev_ctx, d_bias, static_cast>(0)); - return; + if (d_scale && d_bias) { + d_scale->mutable_data>(ctx.GetPlace()); + d_bias->mutable_data>(ctx.GetPlace()); } - PADDLE_ENFORCE_EQ(scale->dims().size(), 1UL); PADDLE_ENFORCE_EQ(scale->dims()[0], C); - // ------------------- cudnn descriptors --------------------- - cudnnTensorDescriptor_t data_desc_; - cudnnTensorDescriptor_t bn_param_desc_; - cudnnBatchNormMode_t mode_; - - CUDNN_ENFORCE(platform::dynload::cudnnCreateTensorDescriptor(&data_desc_)); - CUDNN_ENFORCE( - platform::dynload::cudnnCreateTensorDescriptor(&bn_param_desc_)); - if (epsilon <= CUDNN_BN_MIN_EPSILON - FLT_EPSILON) { - LOG(ERROR) << "Provided epsilon is smaller than " - << "CUDNN_BN_MIN_EPSILON. Setting it to " - << "CUDNN_BN_MIN_EPSILON instead."; - } - epsilon = std::max(epsilon, CUDNN_BN_MIN_EPSILON); -#if CUDNN_VERSION_MIN(7, 0, 0) - mode_ = CUDNN_BATCHNORM_SPATIAL_PERSISTENT; -#else - mode_ = CUDNN_BATCHNORM_SPATIAL; -#endif - std::vector dims; std::vector strides; if (data_layout == DataLayout::kNCHW) { @@ -264,34 +294,114 @@ class BatchNormGradKernel dims = {N, C, H, W, D}; strides = {H * W * C * D, 1, W * D * C, D * C, C}; } - CUDNN_ENFORCE(platform::dynload::cudnnSetTensorNdDescriptor( - data_desc_, CudnnDataType::type, - x_dims.size() > 3 ? x_dims.size() : 4, dims.data(), strides.data())); - CUDNN_ENFORCE(platform::dynload::cudnnDeriveBNTensorDescriptor( - bn_param_desc_, data_desc_, mode_)); - - const auto *saved_mean = ctx.Input("SavedMean"); - const auto *saved_var = ctx.Input("SavedVariance"); - const void *saved_mean_data = - saved_mean->template data>(); - const void *saved_var_data = - saved_var->template data>(); - - CUDNN_ENFORCE(platform::dynload::cudnnBatchNormalizationBackward( - dev_ctx.cudnn_handle(), mode_, CudnnDataType::kOne(), - CudnnDataType::kZero(), CudnnDataType::kOne(), - CudnnDataType::kZero(), data_desc_, x->template data(), - data_desc_, d_y->template data(), data_desc_, - d_x->template mutable_data(ctx.GetPlace()), bn_param_desc_, - scale->template data>(), - d_scale->template mutable_data>(ctx.GetPlace()), - d_bias->template mutable_data>(ctx.GetPlace()), - epsilon, saved_mean_data, saved_var_data)); - // clean when exit. - CUDNN_ENFORCE(platform::dynload::cudnnDestroyTensorDescriptor(data_desc_)); - CUDNN_ENFORCE( - platform::dynload::cudnnDestroyTensorDescriptor(bn_param_desc_)); + auto &dev_ctx = ctx.template device_context(); + if (!use_global_stats) { + if ((N * H * W * D) == 1) { + framework::TensorCopy(*d_y, ctx.GetPlace(), d_x); + math::SetConstant> + functor; + functor(dev_ctx, d_scale, static_cast>(0)); + functor(dev_ctx, d_bias, static_cast>(0)); + return; + } + + // ------------------- cudnn descriptors --------------------- + cudnnTensorDescriptor_t data_desc_; + cudnnTensorDescriptor_t bn_param_desc_; + cudnnBatchNormMode_t mode_; + + CUDNN_ENFORCE( + platform::dynload::cudnnCreateTensorDescriptor(&data_desc_)); + CUDNN_ENFORCE( + platform::dynload::cudnnCreateTensorDescriptor(&bn_param_desc_)); + if (epsilon <= CUDNN_BN_MIN_EPSILON - FLT_EPSILON) { + LOG(ERROR) << "Provided epsilon is smaller than " + << "CUDNN_BN_MIN_EPSILON. Setting it to " + << "CUDNN_BN_MIN_EPSILON instead."; + } + epsilon = std::max(epsilon, CUDNN_BN_MIN_EPSILON); +#if CUDNN_VERSION_MIN(7, 0, 0) + mode_ = CUDNN_BATCHNORM_SPATIAL_PERSISTENT; +#else + mode_ = CUDNN_BATCHNORM_SPATIAL; +#endif + + CUDNN_ENFORCE(platform::dynload::cudnnSetTensorNdDescriptor( + data_desc_, CudnnDataType::type, + x_dims.size() > 3 ? x_dims.size() : 4, dims.data(), strides.data())); + CUDNN_ENFORCE(platform::dynload::cudnnDeriveBNTensorDescriptor( + bn_param_desc_, data_desc_, mode_)); + + const auto *saved_mean = ctx.Input("SavedMean"); + const auto *saved_var = ctx.Input("SavedVariance"); + const void *saved_mean_data = + saved_mean->template data>(); + const void *saved_var_data = + saved_var->template data>(); + + CUDNN_ENFORCE(platform::dynload::cudnnBatchNormalizationBackward( + dev_ctx.cudnn_handle(), mode_, CudnnDataType::kOne(), + CudnnDataType::kZero(), CudnnDataType::kOne(), + CudnnDataType::kZero(), data_desc_, x->template data(), + data_desc_, d_y->template data(), data_desc_, + d_x->template mutable_data(ctx.GetPlace()), bn_param_desc_, + scale->template data>(), + d_scale->template mutable_data>(ctx.GetPlace()), + d_bias->template mutable_data>(ctx.GetPlace()), + epsilon, saved_mean_data, saved_var_data)); + + // clean when exit. + CUDNN_ENFORCE( + platform::dynload::cudnnDestroyTensorDescriptor(data_desc_)); + CUDNN_ENFORCE( + platform::dynload::cudnnDestroyTensorDescriptor(bn_param_desc_)); + } else { + const auto *running_mean = ctx.Input("Mean"); + const auto *running_var = ctx.Input("Variance"); + + const auto *running_mean_data = + running_mean->template data>(); + const auto *running_var_data = + running_var->template data>(); + + const int num = x->numel(); + const int block = 512; + int max_threads = dev_ctx.GetMaxPhysicalThreadCount(); + const int max_blocks = std::max(max_threads / block, 1); + int grid1 = (num + block - 1) / block; + int grid2 = std::min(C, max_blocks); + + if (data_layout == framework::DataLayout::kNCHW) { + if (d_x) { + KeBNBackwardData<<< + grid1, block, 0, dev_ctx.stream()>>>( + d_y->data(), scale->data>(), + running_var_data, epsilon, C, H * W, num, d_x->data()); + } + if (d_scale && d_bias) { + KeBNBackwardScaleBias<<< + grid2, block, 0, dev_ctx.stream()>>>( + d_y->data(), x->data(), running_mean_data, running_var_data, + epsilon, C, H * W, num, d_scale->data>(), + d_bias->data>()); + } + } else { + if (d_x) { + KeBNBackwardData<<< + grid1, block, 0, dev_ctx.stream()>>>( + d_y->data(), scale->data>(), + running_var_data, epsilon, C, H * W, num, d_x->data()); + } + if (d_scale && d_bias) { + KeBNBackwardScaleBias<<< + grid2, block, 0, dev_ctx.stream()>>>( + d_y->data(), x->data(), running_mean_data, running_var_data, + epsilon, C, H * W, num, d_scale->data>(), + d_bias->data>()); + } + } + } } }; diff --git a/paddle/fluid/operators/beam_search_decode_op.cc b/paddle/fluid/operators/beam_search_decode_op.cc index 0d32cae0e1e5f..7f2bde55c9827 100644 --- a/paddle/fluid/operators/beam_search_decode_op.cc +++ b/paddle/fluid/operators/beam_search_decode_op.cc @@ -122,7 +122,8 @@ class BeamSearchDecodeOp : public framework::OperatorBase { platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); auto& dev_ctx = *pool.Get(dev_place); - framework::ExecutionContext ctx(*this, scope, dev_ctx); + framework::RuntimeContext run_ctx(Inputs(), Outputs(), scope); + framework::ExecutionContext ctx(*this, scope, dev_ctx, run_ctx); const LoDTensorArray* ids = ctx.Input("Ids"); const LoDTensorArray* scores = ctx.Input("Scores"); @@ -145,7 +146,7 @@ class BeamSearchDecodeOp : public framework::OperatorBase { LoDTensor* sentenceScores = ctx.Output("SentenceScores"); framework::VisitDataType( - framework::ToDataType(scores->at(0).type()), + scores->at(0).type(), BeamSearchDecodeFunctor(*ids, *scores, sentenceIds, sentenceScores, beam_size, end_id)); } diff --git a/paddle/fluid/operators/beam_search_op.cc b/paddle/fluid/operators/beam_search_op.cc index 791f8a4d3be67..30f700f1d91c5 100644 --- a/paddle/fluid/operators/beam_search_op.cc +++ b/paddle/fluid/operators/beam_search_op.cc @@ -33,11 +33,11 @@ void BeamSearch::operator()(const framework::LoDTensor &pre_ids, auto items = SelectTopBeamSizeItems(pre_ids, pre_scores); auto selected_items = ToMap(items, high_level.back()); - VLOG(30) << "selected_items:"; + VLOG(3) << "selected_items:"; for (size_t i = 0; i < selected_items.size(); ++i) { - VLOG(30) << "offset:" << i; + VLOG(3) << "offset:" << i; for (auto &item : selected_items[i]) { - VLOG(30) << ItemToString(item); + VLOG(3) << ItemToString(item); } } @@ -138,11 +138,11 @@ std::vector> BeamSearch::SelectTopBeamSizeItems( } result.emplace_back(items); } - VLOG(30) << "SelectTopBeamSizeItems result size " << result.size(); + VLOG(3) << "SelectTopBeamSizeItems result size " << result.size(); for (auto &items : result) { - VLOG(30) << "item set:"; + VLOG(3) << "item set:"; for (auto &item : items) { - VLOG(30) << ItemToString(item); + VLOG(3) << ItemToString(item); } } @@ -282,8 +282,7 @@ class BeamSearchOp : public framework::OperatorWithKernel { framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext &ctx) const override { framework::OpKernelType kt = framework::OpKernelType( - framework::ToDataType( - ctx.Input("pre_ids")->type()), + ctx.Input("pre_ids")->type(), platform::CPUPlace()); return kt; } diff --git a/paddle/fluid/operators/beam_search_op_test.cc b/paddle/fluid/operators/beam_search_op_test.cc index 501807e7f3e04..40b46781daa98 100644 --- a/paddle/fluid/operators/beam_search_op_test.cc +++ b/paddle/fluid/operators/beam_search_op_test.cc @@ -46,7 +46,7 @@ void CreateInput(LoDTensor* ids, LoDTensor* scores) { auto* scores_data = scores->mutable_data(place); vector _ids({4, 2, 5, 2, 1, 3, 3, 5, 2, 8, 2, 1}); vector _scores( - {0.5, 0.3, 0.2, 0.6, 0.3, 0.1, 0.9, 0.5, 0.1, 0.7, 0.5, 0.1}); + {0.5f, 0.3f, 0.2f, 0.6f, 0.3f, 0.1f, 0.9f, 0.5f, 0.1f, 0.7f, 0.5f, 0.1f}); for (int i = 0; i < 12; i++) { ids_data[i] = _ids[i]; @@ -80,7 +80,7 @@ TEST(DISABLED_beam_search_op, run) { ASSERT_EQ(sids.lod(), sscores.lod()); vector tids({4, 2, 3, 8}); - vector tscores({0.5, 0.6, 0.9, 0.7}); + vector tscores({0.5f, 0.6f, 0.9f, 0.7f}); for (int i = 0; i < 4; i++) { ASSERT_EQ(tids[i], sids.data()[i]); diff --git a/paddle/fluid/operators/bilinear_tensor_product_op.cu b/paddle/fluid/operators/bilinear_tensor_product_op.cu index 9426ffbe174c7..c2b4f69e68545 100644 --- a/paddle/fluid/operators/bilinear_tensor_product_op.cu +++ b/paddle/fluid/operators/bilinear_tensor_product_op.cu @@ -12,7 +12,6 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#define EIGEN_USE_GPU #include "paddle/fluid/operators/bilinear_tensor_product_op.h" namespace ops = paddle::operators; diff --git a/paddle/fluid/operators/bilinear_tensor_product_op.h b/paddle/fluid/operators/bilinear_tensor_product_op.h index f23336f7b98d6..5017c3a457abc 100644 --- a/paddle/fluid/operators/bilinear_tensor_product_op.h +++ b/paddle/fluid/operators/bilinear_tensor_product_op.h @@ -70,7 +70,7 @@ class BilinearTensorProductKernel : public framework::OpKernel { if (bias) { auto bias_vec = EigenMatrix::From(*bias); Eigen::DSizes bcast(batch_size, 1); - output_mat.device(place) = bias_vec.broadcast(bcast) + output_mat; + output_mat.device(place) = bias_vec.broadcast(bcast).eval() + output_mat; } } }; @@ -99,13 +99,13 @@ class BilinearTensorProductGradKernel : public framework::OpKernel { auto d_out_mat = EigenMatrix::From(*d_out); auto& place = *ctx.template device_context().eigen_device(); auto& dev_ctx = ctx.template device_context(); - // Create the intermediate variable to caculate the Output(Y@Grad). + // Create the intermediate variable to calculate the Output(Y@Grad). Tensor x_scale; x_scale.mutable_data(framework::make_ddim({batch_size, x_dim}), ctx.GetPlace()); auto x_scale_mat = EigenMatrix::From(x_scale); - // Create the intermediate variable to caculate the Output(X@Grad). + // Create the intermediate variable to calculate the Output(X@Grad). Tensor y_scale; y_scale.mutable_data(framework::make_ddim({batch_size, y_dim}), ctx.GetPlace()); @@ -113,65 +113,64 @@ class BilinearTensorProductGradKernel : public framework::OpKernel { math::SetConstant set_zero; - // Set Output(X@Grad) be zero. if (d_x) { d_x->mutable_data(ctx.GetPlace()); set_zero(dev_ctx, d_x, static_cast(0)); } - // Set Output(Y@Grad) be zero. if (d_y) { d_y->mutable_data(ctx.GetPlace()); set_zero(dev_ctx, d_y, static_cast(0)); } + if (d_weight) { + d_weight->mutable_data(ctx.GetPlace()); + } + auto blas = math::GetBlas(ctx); // Caculate the Output(X@Grad) and Output(Y@Grad). - if (d_x || d_y) { + if (d_x || d_y || d_weight) { Eigen::DSizes bcast_for_x(1, y_dim); Eigen::DSizes bcast_for_y(1, x_dim); + Eigen::DSizes bcast_for_weight(1, x_dim); + for (int i = 0; i < out_dim; ++i) { Tensor weight_i = weight->Slice(i, i + 1).Resize( framework::make_ddim({x_dim, y_dim})); auto output_vec = d_out_mat.chip(i, 1); + if (d_x) { y_scale_mat.device(place) = output_vec.reshape(Eigen::DSizes(batch_size, 1)) - .broadcast(bcast_for_x) * + .broadcast(bcast_for_x) + .eval() * y_mat; blas.GEMM(CblasNoTrans, CblasTrans, batch_size, x_dim, y_dim, 1, y_scale.data(), weight_i.data(), 1, d_x->data()); } - if (d_y) { - x_scale_mat.device(place) = + + if (d_y || d_weight) { + auto output_vec_y = output_vec.reshape(Eigen::DSizes(batch_size, 1)) - .broadcast(bcast_for_y) * - x_mat; - blas.GEMM(CblasNoTrans, CblasNoTrans, batch_size, y_dim, x_dim, 1, - x_scale.data(), weight_i.data(), 1, d_y->data()); + .broadcast(bcast_for_y) + .eval(); + x_scale_mat.device(place) = output_vec_y * x_mat; + if (d_y) { + blas.GEMM(CblasNoTrans, CblasNoTrans, batch_size, y_dim, x_dim, 1, + x_scale.data(), weight_i.data(), 1, d_y->data()); + } + if (d_weight) { + Tensor d_weight_i = d_weight->Slice(i, i + 1).Resize( + framework::make_ddim({x_dim, y_dim})); + blas.GEMM(CblasTrans, CblasNoTrans, x_dim, y_dim, batch_size, 1, + x_scale.data(), y->data(), 0, d_weight_i.data()); + } } } } - // Caculate the gradient of Input(Weight). - if (d_weight) { - d_weight->mutable_data(ctx.GetPlace()); - Eigen::DSizes bcast_for_weight(1, x_dim); - for (int i = 0; i < out_dim; ++i) { - Tensor d_weight_i = d_weight->Slice(i, i + 1).Resize( - framework::make_ddim({x_dim, y_dim})); - auto output_vec = d_out_mat.chip(i, 1); - x_scale_mat.device(place) = - output_vec.reshape(Eigen::DSizes(batch_size, 1)) - .broadcast(bcast_for_weight) * - x_mat; - blas.GEMM(CblasTrans, CblasNoTrans, x_dim, y_dim, batch_size, 1, - x_scale.data(), y->data(), 0, d_weight_i.data()); - } - } - - // Caculate the gradient of Input(Bias). + // calculate the gradient of Input(Bias). if (d_bias) { d_bias->mutable_data(ctx.GetPlace()); auto d_bias_mat = framework::EigenVector::Flatten(*d_bias); diff --git a/paddle/fluid/operators/bpr_loss_op.cc b/paddle/fluid/operators/bpr_loss_op.cc new file mode 100644 index 0000000000000..f349c51d8a99a --- /dev/null +++ b/paddle/fluid/operators/bpr_loss_op.cc @@ -0,0 +1,143 @@ +/* Copyright (c) 2016 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/operators/bpr_loss_op.h" + +namespace paddle { +namespace operators { + +class BprLossOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should be not null."); + PADDLE_ENFORCE(ctx->HasInput("Label"), "Input(Label) should be not null."); + PADDLE_ENFORCE(ctx->HasOutput("Y"), "Output(Y) should be not null."); + + auto x_dims = ctx->GetInputDim("X"); + auto label_dims = ctx->GetInputDim("Label"); + int rank = x_dims.size(); + PADDLE_ENFORCE_EQ(rank, label_dims.size(), + "Input(X) and Input(Label) shall have the same rank."); + PADDLE_ENFORCE_EQ(framework::slice_ddim(x_dims, 0, rank - 1), + framework::slice_ddim(label_dims, 0, rank - 1), + "Input(X) and Input(Label) shall have the same shape " + "except the last dimension."); + + auto y_dims = x_dims; + y_dims[rank - 1] = 1; + ctx->SetOutputDim("Y", y_dims); + ctx->ShareLoD("X", /*->*/ "Y"); + } + + protected: + // Explicitly set that the data type of computation kernel of Seq-bpr + // is determined by its input "X". + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + return framework::OpKernelType(ctx.Input("X")->type(), + platform::CPUPlace()); + } +}; + +class BprLossGradientOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should be not null."); + PADDLE_ENFORCE(ctx->HasInput("Label"), "Input(Label) should be not null."); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Y")), + "Input(Y@GRAD) shoudl be not null."); + PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("X")), + "Output(X@GRAD) should be not null."); + + auto x_dims = ctx->GetInputDim("X"); + auto label_dims = ctx->GetInputDim("Label"); + auto dy_dims = ctx->GetInputDim(framework::GradVarName("Y")); + int rank = x_dims.size(); + PADDLE_ENFORCE_EQ(dy_dims.size(), rank, + "Input(Y@Grad) and Input(X) should have the same rank."); + PADDLE_ENFORCE_EQ(label_dims.size(), rank, + "Input(Label) and Input(X) should have the same rank."); + PADDLE_ENFORCE_EQ(framework::slice_ddim(x_dims, 0, rank - 1), + framework::slice_ddim(label_dims, 0, rank - 1), + "The Input(X) and Input(Label) should have the same " + "shape except the last dimension."); + PADDLE_ENFORCE_EQ(framework::slice_ddim(x_dims, 0, rank - 1), + framework::slice_ddim(dy_dims, 0, rank - 1), + "The Input(X) and Input(Y@Grad) should have the same " + "shape except the last dimension."); + PADDLE_ENFORCE_EQ(dy_dims[rank - 1], 1, + "The last dimension of Input(Y@Grad) should be 1."); + PADDLE_ENFORCE_EQ(label_dims[rank - 1], 1, + " the last dimension of Input(Label) should be 1."); + ctx->SetOutputDim(framework::GradVarName("X"), x_dims); + ctx->ShareLoD("X", framework::GradVarName("X")); + } + + protected: + // Explicitly set that the data type of computation kernel of cross_entropy + // is determined by its input "X". + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + return framework::OpKernelType(ctx.Input("X")->type(), + platform::CPUPlace()); + } +}; + +class BprLossOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("X", + "(Tensor, default Tensor), a tensor whose last dimension " + "size is equal to the number of classes. This input is a " + "real number."); + AddInput( + "Label", + "(Tensor), the tensor which represents the ground truth. It has the " + "same shape with 'X' except the last dimension. the last dimension " + "size is 1."); + AddOutput("Y", + "(Tensor, default Tensor), a tensor whose shape is same " + "with 'X' except that the last dimension size is 1. It " + "represents the sequence bpr loss."); + AddComment(R"DOC( +Bayesian Personalized Ranking Loss Operator. + +This operator belongs to pairwise ranking loss. Label is the desired item. +The loss at a given point in one session is defined as: +$Y[i] = -\frac{1}{N_{i}} * \sum_{j=0}^{N_{i}}\log(\sigma(X[i, Label[i]]-X[i, j]))$ + +Learn more details by reading paper (https://arxiv.org/abs/1511.06939) + +)DOC"); + } +}; +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +using CPUCtx = paddle::platform::CPUDeviceContext; + +REGISTER_OPERATOR(bpr_loss, ops::BprLossOp, ops::BprLossOpMaker, + paddle::framework::DefaultGradOpDescMaker); +REGISTER_OPERATOR(bpr_loss_grad, ops::BprLossGradientOp); +REGISTER_OP_CPU_KERNEL(bpr_loss, ops::BprLossOpKernel, + ops::BprLossOpKernel); +REGISTER_OP_CPU_KERNEL(bpr_loss_grad, + ops::BprLossGradientOpKernel, + ops::BprLossGradientOpKernel); diff --git a/paddle/fluid/operators/bpr_loss_op.h b/paddle/fluid/operators/bpr_loss_op.h new file mode 100644 index 0000000000000..e223be7af8214 --- /dev/null +++ b/paddle/fluid/operators/bpr_loss_op.h @@ -0,0 +1,118 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include "paddle/fluid/framework/eigen.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/math/math_function.h" +#include "paddle/fluid/platform/for_range.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; +/*Todo: + *Find a way to adapt TolerableValue, using blas or eigen. + */ +template +struct TolerableValue { + HOSTDEVICE T operator()(const T& x) const { + PADDLE_ASSERT(std::is_floating_point::value); + const T kApproInf = 1e20; + if (x == INFINITY) return kApproInf; + if (x == -INFINITY) return -kApproInf; + return x; + } +}; + +template +class BprLossOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* x = ctx.Input("X"); + auto* label = ctx.Input("Label"); + auto* y = ctx.Output("Y"); + y->mutable_data(ctx.GetPlace()); + int rank = x->dims().size(); + + Tensor x_2d = framework::ReshapeToMatrix(*x, rank - 1); + Tensor labels_2d = framework::ReshapeToMatrix(*label, rank - 1); + Tensor y_2d = framework::ReshapeToMatrix(*y, rank - 1); + + const framework::Tensor* logits = &x_2d; + const framework::Tensor* labels = &labels_2d; + framework::Tensor* out = &y_2d; + + const int step_size = logits->dims()[0]; + const int class_num = logits->dims()[1]; + const T* logits_data = logits->data(); + T* loss_data = out->data(); + + const int64_t* label_data = labels->data(); + for (int i = 0; i < step_size; ++i) { + int lbl_pos = label_data[i]; + PADDLE_ENFORCE_GE(lbl_pos, 0); + PADDLE_ENFORCE_LT(lbl_pos, class_num); + int index_pos = i * class_num + lbl_pos; + T sum = static_cast(0); + for (int j = 0; j < class_num; j++) { + if (j == lbl_pos) continue; + int index_neg = i * class_num + j; + sum += TolerableValue()(-std::log( + 1.0f + TolerableValue()(std::exp(logits_data[index_neg] - + logits_data[index_pos])))); + } + loss_data[i] = -sum / (class_num - 1); + } + } +}; + +template +class BprLossGradientOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* x = ctx.Input("X"); + auto* dy = ctx.Input(framework::GradVarName("Y")); + auto* label = ctx.Input("Label"); + auto* dx = ctx.Output(framework::GradVarName("X")); + + const int step_size = x->dims()[0]; + const int num_classes = x->dims()[1]; + T* dx_data = dx->mutable_data(ctx.GetPlace()); + const T* dy_data = dy->data(); + const T* x_data = x->data(); + const int64_t* label_data = label->data(); + + for (size_t sample_id = 0; sample_id < step_size; sample_id++) { + for (size_t x_offset = sample_id * num_classes; + x_offset < (sample_id + 1) * num_classes; x_offset++) { + dx_data[x_offset] = static_cast(0); + } + auto p_index = sample_id * num_classes + label_data[sample_id]; + for (size_t ni = 0; ni < num_classes; ni++) { + if (label_data[sample_id] == ni) continue; + auto n_index = sample_id * num_classes + ni; + auto grad_ = -dy_data[sample_id] / + ((num_classes - 1) * + (1.0f + TolerableValue()(std::exp(x_data[p_index] - + x_data[n_index])))); + dx_data[p_index] += grad_; + dx_data[n_index] -= grad_; + } + } + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/concat_mkldnn_op.cc b/paddle/fluid/operators/concat_mkldnn_op.cc new file mode 100644 index 0000000000000..7ad674056f0d7 --- /dev/null +++ b/paddle/fluid/operators/concat_mkldnn_op.cc @@ -0,0 +1,152 @@ +/* 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 +#include "paddle/fluid/operators/concat_op.h" +#include "paddle/fluid/platform/mkldnn_helper.h" + +namespace paddle { +namespace operators { + +using framework::DataLayout; +using framework::Tensor; +using mkldnn::memory; +using mkldnn::primitive; +using mkldnn::concat; +using mkldnn::stream; +using platform::to_void_cast; + +static void EnforceLayouts(const std::vector inputs) { + for (auto* input : inputs) { + const bool is_layout_correct = input->layout() == DataLayout::kMKLDNN; + const bool is_format_defined = + input->format() != memory::format::format_undef; + PADDLE_ENFORCE(is_layout_correct && is_format_defined, + "Wrong layout/format set for Input tensor"); + } +} + +static memory::primitive_desc CreateMemPrimDesc(const Tensor& input, + const mkldnn::engine& engine) { + constexpr auto data_type = mkldnn::memory::f32; + const auto dims = paddle::framework::vectorize2int(input.dims()); + const auto format = input.format(); + auto description = memory::desc(dims, data_type, format); + auto mem_prim_desc = memory::primitive_desc(description, engine); + return mem_prim_desc; +} + +static mkldnn::memory::format GetDstMemFormat( + const concat::primitive_desc& concat_pd) { + return (memory::format)concat_pd.dst_primitive_desc().desc().data.format; +} + +static platform::CPUPlace GetCpuPlace( + const paddle::framework::ExecutionContext& ctx) { + auto place = ctx.GetPlace(); + PADDLE_ENFORCE(paddle::platform::is_cpu_place(place), + "It must use CPUPlace."); + return boost::get(place); +} + +static const mkldnn::engine& GetMKLDNNEngine( + const paddle::framework::ExecutionContext& ctx) { + auto& dev_ctx = ctx.template device_context(); + return dev_ctx.GetEngine(); +} + +template +class ConcatPrimitiveFactory { + public: + concat::primitive_desc CreateConcatPrimDescriptor( + const std::vector multi_input, Tensor* output, + int concat_axis, const mkldnn::engine& mkldnn_engine) { + CreateSourcesDescriptors(multi_input, mkldnn_engine); + auto dst_desc = CreateDstMemDescriptor(output); + return concat::primitive_desc(dst_desc, concat_axis, srcs_pd); + } + + concat CreateConcatPrimitive(const concat::primitive_desc& concat_pd, + Tensor* output, platform::CPUPlace place) { + CreateSourcePrimitiveAts(); + dst_mem = CreateDstMemory(concat_pd, output, place); + return concat(concat_pd, inputs, dst_mem.get()); + } + + private: + memory::desc CreateDstMemDescriptor(Tensor* output) { + auto dst_dims = paddle::framework::vectorize2int(output->dims()); + return memory::desc(dst_dims, platform::MKLDNNGetDataType(), + memory::format::any); + } + + mkldnn::memory CreateDstMemory(const concat::primitive_desc& concat_pd, + Tensor* output, platform::CPUPlace place) { + return memory(concat_pd.dst_primitive_desc(), + output->mutable_data(place)); + } + + void CreateSourcesDescriptors(const std::vector multi_input, + const mkldnn::engine& mkldnn_engine) { + for (size_t i = 0; i < multi_input.size(); i++) { + auto mem_prim_desc = CreateMemPrimDesc(*multi_input[i], mkldnn_engine); + srcs_pd.push_back(mem_prim_desc); + srcs.push_back( + memory(mem_prim_desc, to_void_cast(multi_input[i]->data()))); + } + } + + void CreateSourcePrimitiveAts() { + inputs.reserve(srcs.size()); + for (size_t i = 0; i < srcs.size(); i++) { + inputs.push_back(srcs[i]); + } + } + + private: + std::vector srcs_pd; + std::vector srcs; + std::vector inputs; + boost::optional dst_mem; // TODO(mgallus): change to std::optional +}; // upon introduction of C++17 to paddle + +template +class ConcatMKLDNNOpKernel : public paddle::framework::OpKernel { + public: + void Compute(const paddle::framework::ExecutionContext& ctx) const override { + auto place = GetCpuPlace(ctx); + const auto& mkldnn_engine = GetMKLDNNEngine(ctx); + + auto multi_input = ctx.MultiInput("X"); + EnforceLayouts(multi_input); + Tensor* output = ctx.Output("Out"); + int64_t concat_axis = static_cast(ctx.Attr("axis")); + + ConcatPrimitiveFactory prim_creator; + auto concat_pd = prim_creator.CreateConcatPrimDescriptor( + multi_input, output, static_cast(concat_axis), mkldnn_engine); + auto concat = prim_creator.CreateConcatPrimitive(concat_pd, output, place); + stream(stream::kind::eager).submit({concat}).wait(); + + output->set_layout(DataLayout::kMKLDNN); + output->set_format(GetDstMemFormat(concat_pd)); + } +}; +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; + +REGISTER_OP_KERNEL(concat, MKLDNN, ::paddle::platform::CPUPlace, + ops::ConcatMKLDNNOpKernel) diff --git a/paddle/fluid/operators/concat_op.cc b/paddle/fluid/operators/concat_op.cc index 093b0a9a1f9ac..194f9cf5033a3 100644 --- a/paddle/fluid/operators/concat_op.cc +++ b/paddle/fluid/operators/concat_op.cc @@ -13,10 +13,13 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/operators/concat_op.h" - #include #include +#ifdef PADDLE_WITH_MKLDNN +#include +#endif + namespace paddle { namespace operators { using framework::Tensor; @@ -37,7 +40,7 @@ class ConcatOp : public framework::OperatorWithKernel { PADDLE_ENFORCE_GT(n, 0, "Input tensors count should > 0."); if (n == 1) { - VLOG(30) << "Warning: concat op have only one input, may waste memory"; + VLOG(3) << "Warning: concat op have only one input, may waste memory"; } auto out_dims = ins[0]; @@ -59,6 +62,22 @@ class ConcatOp : public framework::OperatorWithKernel { ctx->SetOutputDim("Out", out_dims); ctx->ShareLoD("X", /*->*/ "Out"); } + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext &ctx) const override { + auto input_data_type = + framework::GetDataTypeOfVar(ctx.MultiInputVar("X")[0]); + +#ifdef PADDLE_WITH_MKLDNN + if (platform::CanMKLDNNBeUsed(ctx)) { + return framework::OpKernelType(input_data_type, ctx.GetPlace(), + framework::DataLayout::kMKLDNN, + framework::LibraryType::kMKLDNN); + } +#endif + return framework::OpKernelType(input_data_type, ctx.GetPlace()); + } }; class ConcatOpMaker : public framework::OpProtoAndCheckerMaker { @@ -66,6 +85,10 @@ class ConcatOpMaker : public framework::OpProtoAndCheckerMaker { void Make() override { AddInput("X", "Input tensors of concat operator.").AsDuplicable(); AddOutput("Out", "Output tensor of concat operator."); + AddAttr( + "use_mkldnn", + "(bool, default false) Indicates if MKL-DNN kernel will be used") + .SetDefault(false); AddAttr("axis", "The axis along which the input tensors will be concatenated.") .SetDefault(0); diff --git a/paddle/fluid/operators/controlflow/CMakeLists.txt b/paddle/fluid/operators/controlflow/CMakeLists.txt index b1c2ee22951a3..b614e9b035026 100644 --- a/paddle/fluid/operators/controlflow/CMakeLists.txt +++ b/paddle/fluid/operators/controlflow/CMakeLists.txt @@ -1,4 +1,4 @@ include(operators) -register_operators() +register_operators(DEPS naive_executor) file(APPEND ${pybind_file} "USE_OP(less_than);\nUSE_OP(logical_and);\nUSE_NO_KERNEL_OP(read_from_array);\n") diff --git a/paddle/fluid/operators/controlflow/conditional_block_op.cc b/paddle/fluid/operators/controlflow/conditional_block_op.cc index 135254ce6b6bf..dd28f82b65403 100644 --- a/paddle/fluid/operators/controlflow/conditional_block_op.cc +++ b/paddle/fluid/operators/controlflow/conditional_block_op.cc @@ -48,13 +48,12 @@ class ConditionalOp : public framework::OperatorBase { if (!(ips.size() == 1UL && ips[0]->IsInitialized())) { PADDLE_THROW("should have one initialized input as condition"); } - if (!(framework::IsType(ips[0]->type()) && // NOLINT - ips[0]->numel() == 1)) { - PADDLE_THROW( - "condition input's data type should be bool, " - "numel should be 1, actual numel is %d", - ips[0]->numel()); - } + + PADDLE_ENFORCE(ips[0]->type() == framework::proto::VarType::BOOL && + ips[0]->numel() == 1, + "condition input's data type should be bool, " + "numel should be 1, actual numel is %d", + ips[0]->numel()); bool res = false; if (platform::is_gpu_place(ips[0]->place())) { #ifdef PADDLE_WITH_CUDA diff --git a/paddle/fluid/operators/controlflow/feed_op.cc b/paddle/fluid/operators/controlflow/feed_op.cc index 5da0a536d96e5..dc7ef66495823 100644 --- a/paddle/fluid/operators/controlflow/feed_op.cc +++ b/paddle/fluid/operators/controlflow/feed_op.cc @@ -47,8 +47,8 @@ class FeedOp : public framework::OperatorBase { auto col = Attr("col"); - VLOG(30) << "Feed Var " << feed_var_name << "'s " << col - << " column to var " << out_name; + VLOG(3) << "Feed Var " << feed_var_name << "'s " << col << " column to var " + << out_name; auto &feed_list = feed_var->Get(); auto &feed_item = feed_list.at(static_cast(col)); diff --git a/paddle/fluid/operators/controlflow/fetch_op.cc b/paddle/fluid/operators/controlflow/fetch_op.cc index c9e759ebff639..c197b45e8196a 100644 --- a/paddle/fluid/operators/controlflow/fetch_op.cc +++ b/paddle/fluid/operators/controlflow/fetch_op.cc @@ -57,7 +57,7 @@ class FetchOp : public framework::OperatorBase { TensorCopySync(src_item, platform::CPUPlace(), &dst_item); dst_item.set_lod(src_item.lod()); - VLOG(30) << "Fetch variable " << fetch_var_name << " to " << out_name; + VLOG(3) << "Fetch variable " << fetch_var_name << " to " << out_name; } }; diff --git a/paddle/fluid/operators/controlflow/parallel_do_op.cc b/paddle/fluid/operators/controlflow/parallel_do_op.cc deleted file mode 100644 index c795d4bdd10c0..0000000000000 --- a/paddle/fluid/operators/controlflow/parallel_do_op.cc +++ /dev/null @@ -1,426 +0,0 @@ -/* Copyright (c) 2016 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 - -#include "paddle/fluid/framework/executor.h" -#include "paddle/fluid/framework/op_registry.h" -#include "paddle/fluid/framework/threadpool.h" -#include "paddle/fluid/operators/detail/safe_ref.h" - -namespace paddle { -namespace operators { - -static constexpr char kInputs[] = "inputs"; -static constexpr char kParameters[] = "parameters"; -static constexpr char kPlaces[] = "places"; - -static constexpr char kOutputs[] = "outputs"; -static constexpr char kParallelScopes[] = "parallel_scopes"; - -static constexpr char kParallelBlock[] = "sub_block"; -static constexpr char kUseNCCL[] = "use_nccl"; - -using LoDTensor = framework::LoDTensor; -using SelectedRows = framework::SelectedRows; - -static void SplitTensorAndMoveTensorToScopes( - const framework::Scope &scope, std::vector *sub_scopes, - const std::vector &places, - const std::vector &names) { - size_t num_sub_scopes = 0; - for (auto &argu : names) { - const auto &tensor = - detail::Ref(scope.FindVar(argu), - "Cannot find variable %s in the parent scope", argu) - .Get(); - auto lod_tensors = tensor.SplitLoDTensor(places); - - for (auto &lod : lod_tensors) { - VLOG(30) << lod.dims(); - } - if (num_sub_scopes == 0) { - num_sub_scopes = lod_tensors.size(); - } else { - PADDLE_ENFORCE_EQ(num_sub_scopes, lod_tensors.size()); - } - PADDLE_ENFORCE_NE(num_sub_scopes, 0); - if (sub_scopes->size() == 0) { - sub_scopes->reserve(num_sub_scopes); - for (size_t i = 0; i < num_sub_scopes; ++i) { - sub_scopes->emplace_back(&scope.NewScope()); - } - } - - for (size_t i = 0; i < lod_tensors.size(); ++i) { - *detail::Ref(sub_scopes->at(i)->Var(argu), - "Cannot find variable in the sub-scope", argu) - .GetMutable() = lod_tensors[i]; - } - } -} - -inline void CopyOrShare(const framework::Variable &src, - const platform::Place &dst_place, - framework::Variable *dst) { - if (src.IsType()) { - if (src.Get().place() == dst_place) { - dst->GetMutable()->ShareDataWith(src.Get()); - dst->GetMutable()->set_lod(src.Get().lod()); - } else { - TensorCopy(src.Get(), dst_place, dst->GetMutable()); - } - } else if (src.IsType()) { - auto &src_sr = src.Get(); - auto *dst_sr = dst->GetMutable(); - dst_sr->set_height(src_sr.height()); - if (src_sr.value().place() == dst_place) { - dst_sr->mutable_value()->ShareDataWith(src_sr.value()); - dst_sr->set_rows(src_sr.rows()); - } else { - TensorCopy(src_sr.value(), dst_place, dst_sr->mutable_value()); - } - } else { - PADDLE_THROW("Expect LoDTensor/SelectedRows, get %s", src.Type().name()); - } -} - -void WaitOnPlace(const platform::Place place) { - platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance(); - auto &dev_ctx = *pool.Get(place); - dev_ctx.Wait(); -} - -void WaitOnPlaces(const std::vector places) { - platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance(); - - for (auto &place : places) { - auto &dev_ctx = *pool.Get(place); - dev_ctx.Wait(); - } -} - -class ParallelDoOp : public framework::OperatorBase { - public: - ParallelDoOp(const std::string &type, - const framework::VariableNameMap &inputs, - const framework::VariableNameMap &outputs, - const framework::AttributeMap &attrs) - : framework::OperatorBase(type, inputs, outputs, attrs) {} - - private: - void RunImpl(const framework::Scope &scope, - const platform::Place &place) const override { - // get device context from pool - platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance(); - auto &dev_ctx = *pool.Get(place); - - auto *block = Attr(kParallelBlock); - auto *program = block->Program(); - - auto &places = scope.FindVar(Input(kPlaces))->Get(); - - auto &sub_scopes = *scope.FindVar(Output(kParallelScopes)) - ->GetMutable>(); - - // split input - SplitTensorAndMoveTensorToScopes(scope, &sub_scopes, places, - Inputs(kInputs)); - - // copy parameter - for (auto ¶m : Inputs(kParameters)) { - PADDLE_ENFORCE(scope.FindVar(param)->IsType(), - "Only support parameter type as LoDTensor"); - auto &src = scope.FindVar(param)->Get(); - - auto *sub_scope0 = sub_scopes[0]; - auto *dst0 = sub_scope0->Var(param)->GetMutable(); - dst0->ShareDataWith(src); - - for (size_t i = 1; i < sub_scopes.size(); ++i) { - auto &place = places[i]; - auto *sub_scope = sub_scopes[i]; - auto *dst = sub_scope->Var(param)->GetMutable(); - framework::TensorCopy(src, place, dst); - } - } - WaitOnPlaces(places); - - std::vector> workers; - workers.reserve(places.size()); - for (size_t place_idx = 0; place_idx < sub_scopes.size(); ++place_idx) { - auto &place = places[place_idx]; - auto *cur_scope = sub_scopes[place_idx]; - - workers.emplace_back(framework::Async([program, cur_scope, place, block] { - framework::Executor executor(place); - executor.Run(*program, cur_scope, block->ID(), - false /*create_local_scope*/); - })); - } - for (auto &worker : workers) { - worker.wait(); - } - WaitOnPlaces(places); - - // merge output - for (auto &o_name : Outputs(kOutputs)) { - std::vector lod_tensors; - lod_tensors.reserve(sub_scopes.size()); - for (auto *sub_scope : sub_scopes) { - lod_tensors.emplace_back(&sub_scope->FindVar(o_name)->Get()); - } - - auto *lod_tensor_to_be_merged = - scope.FindVar(o_name)->GetMutable(); - lod_tensor_to_be_merged->MergeLoDTensor(lod_tensors, dev_ctx.GetPlace()); - } - WaitOnPlaces(places); - } -}; - -class ParallelDoOpProtoMaker : public framework::OpProtoAndCheckerMaker { - public: - void Make() override { - AddInput(kInputs, "").AsDuplicable(); - AddInput(kParameters, "").AsDuplicable(); - AddInput(kPlaces, ""); - AddOutput(kOutputs, "").AsDuplicable(); - AddOutput(kParallelScopes, ""); - AddAttr(kParallelBlock, ""); - AddAttr(kUseNCCL, "true if we use nccl on backward") - .SetDefault(false); - AddComment(R"DOC( -ParallelDo Operator. -)DOC"); - } -}; - -class ParallelDoGradOp : public framework::OperatorBase { - public: - ParallelDoGradOp(const std::string &type, - const framework::VariableNameMap &inputs, - const framework::VariableNameMap &outputs, - const framework::AttributeMap &attrs) - : framework::OperatorBase(type, inputs, outputs, attrs) {} - - private: - void RunImpl(const framework::Scope &scope, - const platform::Place &place) const override { - auto *block = Attr(kParallelBlock); - auto *program = block->Program(); - - auto &sub_scopes = scope.FindVar(Input(kParallelScopes)) - ->Get>(); - auto &places = scope.FindVar(Input(kPlaces))->Get(); - - // feed output@grad - SplitTensorAndMoveTensorToScopes( - scope, const_cast *>(&sub_scopes), - places, Inputs(framework::GradVarName(kOutputs))); - WaitOnPlaces(places); - - // exe run - std::vector> workers; - for (size_t i = 0; i < sub_scopes.size(); ++i) { - auto &place = places[i]; - auto *cur_scope = sub_scopes[i]; - - // execute - workers.emplace_back(framework::Async([program, cur_scope, place, block] { - framework::Executor executor(place); - executor.Run(*program, cur_scope, block->ID(), - false /*create_local_scope*/); - })); - } - for (auto &worker : workers) { - worker.wait(); - } - WaitOnPlaces(places); - - // NCCL allreduce op will be added by backward, - // so no need to explicitly accumulate grad - if (!(Attr(kUseNCCL))) { - AccumulateGrad(scope, place, sub_scopes, places); - } else { - for (auto &place : places) { - PADDLE_ENFORCE(platform::is_gpu_place(place), - "NCCL only supports cuda place"); - } - } - for (auto &s : Outputs(framework::GradVarName(kParameters))) { - if (s == framework::kEmptyVarName) { - continue; - } - VLOG(30) << "Moving " << s; - CopyOrShare(*sub_scopes[0]->FindVar(s), place, scope.FindVar(s)); - } - WaitOnPlaces(places); - } - - void AccumulateGrad(const framework::Scope &scope, - const platform::Place &place, - const std::vector &sub_scopes, - const platform::PlaceList &places) const { - for (auto &s : Outputs(framework::GradVarName(kParameters))) { - if (s == framework::kEmptyVarName) { - continue; - } - VLOG(30) << "Accumulating " << s; - if (s == framework::kEmptyVarName) continue; - std::string tmp_name; - auto *tmp = sub_scopes[0]->Var(&tmp_name); - - for (size_t i = 1; i < sub_scopes.size(); ++i) { - CopyOrShare(*sub_scopes[i]->FindVar(s), places[0], tmp); - WaitOnPlaces(places); - - auto sum_op = framework::OpRegistry::CreateOp( - "sum", {{"X", {s, tmp_name}}}, {{"Out", {s}}}, - framework::AttributeMap{{"use_mkldnn", {false}}}); - VLOG(100) << sum_op->DebugStringEx(sub_scopes[0]); - sum_op->Run(*sub_scopes[0], places[0]); - WaitOnPlace(places[0]); - } - - CopyOrShare(*sub_scopes[0]->FindVar(s), place, scope.FindVar(s)); - } - WaitOnPlaces(places); - } -}; - -std::ostream &operator<<(std::ostream &sout, - const std::vector &strs) { - std::copy(strs.begin(), strs.end(), - std::ostream_iterator(sout, ",")); - return sout; -} - -class ParallelDoGradOpDescMaker : public framework::SingleGradOpDescMaker { - public: - using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; - - protected: - virtual std::unique_ptr Apply() const { - auto *grad = new framework::OpDesc(); - grad->SetType("parallel_do_grad"); - for (auto &input_param : this->InputNames()) { - VLOG(30) << input_param; - grad->SetInput(input_param, this->Input(input_param)); - if (input_param != kPlaces) { - grad->SetOutput(framework::GradVarName(input_param), - this->InputGrad(input_param, false)); - } - } - auto *g_block = this->grad_block_[0]; - - // All variable name that needed by gradient operators - std::unordered_set all_inputs_in_grad_blocks; - - for (size_t i = 0; i < g_block->OpSize(); ++i) { - auto *op = g_block->Op(i); - for (auto &var_name : op->InputArgumentNames()) { - all_inputs_in_grad_blocks.insert(var_name); - } - } - - for (auto &output_param : this->OutputNames()) { - if (output_param == kParallelScopes) { - grad->SetInput(output_param, this->Output(output_param)); - grad->SetInput(framework::GradVarName(output_param), - this->Output(output_param)); - } else { - grad->SetInput(output_param, this->Output(output_param)); - std::vector og_names; - for (auto &og_name : this->OutputGrad(output_param)) { - if (all_inputs_in_grad_blocks.count(og_name) != 0) { - // there are some gradient operators who need the OG. So make this - // OG as an input of parallel.do - og_names.push_back(og_name); - } - // else, there is no operator who need the OG. Do not use this OG as - // an input - } - grad->SetInput(framework::GradVarName(output_param), og_names); - } - } - grad->SetInput("Communicator", {"nccl_com__do_not_change_"}); - grad->SetAttrMap(this->Attrs()); - grad->SetBlockAttr(kParallelBlock, grad_block_[0]); - - return std::unique_ptr(grad); - } -}; - -class ParallelDoGradOpShapeInference : public framework::InferShapeBase { - public: - void operator()(framework::InferShapeContext *ctx) const override { - PADDLE_ENFORCE(ctx->HasInputs(kParameters)); - PADDLE_ENFORCE(ctx->HasInputs(kInputs)); - PADDLE_ENFORCE(ctx->HasInputs(kOutputs)); - - ctx->SetOutputsDim(framework::GradVarName(kParameters), - ctx->GetInputsDim(kParameters)); - - auto i_dims = ctx->GetInputsDim(kInputs); - auto ig_names = ctx->Outputs(framework::GradVarName(kInputs)); - - for (size_t i = 0; i < ig_names.size(); ++i) { - auto &ig_name = ig_names[i]; - if (ig_name == framework::kEmptyVarName) { - continue; - } - - ctx->SetDims({ig_name}, {i_dims[i]}); - } - - auto p_dims = ctx->GetInputsDim(kParameters); - auto pg_names = ctx->Outputs(framework::GradVarName(kParameters)); - for (size_t i = 0; i < pg_names.size(); ++i) { - auto &pg_name = pg_names[i]; - if (pg_name == framework::kEmptyVarName) { - continue; - } - ctx->SetDims({pg_name}, {p_dims[i]}); - } - } -}; - -class ParallelDoGradOpVarTypeInference : public framework::VarTypeInference { - public: - void operator()(const framework::OpDesc &op_desc, - framework::BlockDesc *block) const override { - framework::BlockDesc *sub_block = - boost::get(op_desc.GetAttr(kParallelBlock)); - for (auto &out_vars : op_desc.Outputs()) { - for (auto &out_var : out_vars.second) { - auto &var = block->FindRecursiveOrCreateVar(out_var); - auto sub_var = sub_block->FindRecursiveOrCreateVar(out_var); - if (sub_var.GetType() != var.GetType()) { - var.SetType(sub_var.GetType()); - } - } - } - } -}; - -} // namespace operators -} // namespace paddle - -REGISTER_OPERATOR(parallel_do, paddle::operators::ParallelDoOp, - paddle::operators::ParallelDoOpProtoMaker, - paddle::operators::ParallelDoGradOpDescMaker); -REGISTER_OPERATOR(parallel_do_grad, paddle::operators::ParallelDoGradOp, - paddle::operators::ParallelDoGradOpShapeInference, - paddle::operators::ParallelDoGradOpVarTypeInference); diff --git a/paddle/fluid/operators/controlflow/tensor_array_read_write_op.cc b/paddle/fluid/operators/controlflow/tensor_array_read_write_op.cc index 484160aeb8de5..fa18ade3234ed 100644 --- a/paddle/fluid/operators/controlflow/tensor_array_read_write_op.cc +++ b/paddle/fluid/operators/controlflow/tensor_array_read_write_op.cc @@ -34,8 +34,8 @@ class WriteToArrayOp : public ArrayOp { auto *out = scope.FindVar(Output("Out"))->GetMutable(); if (offset >= out->size()) { - VLOG(100) << "Resize " << Output("Out") << " from " << out->size() - << " to " << offset + 1; + VLOG(10) << "Resize " << Output("Out") << " from " << out->size() + << " to " << offset + 1; out->resize(offset + 1); } auto *out_tensor = &out->at(offset); @@ -47,9 +47,9 @@ class WriteToArrayOp : public ArrayOp { TensorCopy(x_tensor, place, dev_ctx, out_tensor); } else { - VLOG(100) << "WARNING: The input tensor 'x_tensor' holds no memory, so " - "nothing has been written to output array[" - << offset << "]."; + VLOG(10) << "WARNING: The input tensor 'x_tensor' holds no memory, so " + "nothing has been written to output array[" + << offset << "]."; } } }; @@ -104,7 +104,7 @@ class WriteToArrayInferVarType : public framework::VarTypeInference { framework::BlockDesc *block) const override { auto x_name = op_desc.Input("X")[0]; auto out_name = op_desc.Output("Out")[0]; - VLOG(100) << "Set Variable " << out_name << " as LOD_TENSOR_ARRAY"; + VLOG(10) << "Set Variable " << out_name << " as LOD_TENSOR_ARRAY"; auto &out = block->FindRecursiveOrCreateVar(out_name); out.SetType(framework::proto::VarType::LOD_TENSOR_ARRAY); auto *x = block->FindVarRecursive(x_name); @@ -139,7 +139,7 @@ class ReadFromArrayOp : public ArrayOp { framework::TensorCopy(x_array[offset], place, dev_ctx, out_tensor); out_tensor->set_lod(x_array[offset].lod()); } else { - VLOG(100) << "offset " << offset << " >= " << x_array.size(); + VLOG(10) << "offset " << offset << " >= " << x_array.size(); } } }; @@ -167,6 +167,19 @@ equation is }; class ReadFromArrayInferShape : public WriteToArrayInferShape { + public: + void operator()(framework::InferShapeContext *context) const override { + WriteToArrayInferShape::operator()(context); + if (!context->HasInput("X")) { + return; + } + + // FIXME: just for compile time. + if (!context->IsRuntime()) { + context->ShareLoD("X", /*->*/ "Out"); + } + } + protected: const char *NotHasXError() const override { return "The input array X must be set"; diff --git a/paddle/fluid/operators/controlflow/while_op.cc b/paddle/fluid/operators/controlflow/while_op.cc index 2b56514fe086d..48800947fd387 100644 --- a/paddle/fluid/operators/controlflow/while_op.cc +++ b/paddle/fluid/operators/controlflow/while_op.cc @@ -32,6 +32,20 @@ static constexpr char kStepScopes[] = "StepScopes"; static constexpr char kX[] = "X"; static constexpr char kXGRAD[] = "X@GRAD"; static constexpr char kOutputs[] = "Out"; +static constexpr char kSkipEagerDeletionVars[] = "skip_eager_deletion_vars"; + +namespace { // NOLINT +static std::string GetSkipEagerDeletionVarsDebugString( + const std::vector &vars) { + std::string str = "Skip " + std::to_string(vars.size()) + + " var(s) in eager deletion mode: "; + for (auto &var : vars) { + str.append(var); + str.push_back(' '); + } + return str; +} +} // NOLINT class WhileOp : public framework::OperatorBase { public: @@ -59,7 +73,10 @@ class WhileOp : public framework::OperatorBase { "Condition of while op must in CPU memory."); bool is_test = Attr("is_test"); - auto ctx = executor.Prepare(*program, block->ID()); + auto &skip_vars = Attr>(kSkipEagerDeletionVars); + VLOG(2) << GetSkipEagerDeletionVarsDebugString(skip_vars); + + auto ctx = executor.Prepare(*program, block->ID(), skip_vars); while (cond.data()[0]) { auto ¤t_scope = scope.NewScope(); step_scopes->push_back(¤t_scope); @@ -96,6 +113,10 @@ class WhileOpMaker : public framework::OpProtoAndCheckerMaker { "(bool, default false) Set to true for inference only, false " "for training. Some layers may run faster when this is true.") .SetDefault(false); + AddAttr>(kSkipEagerDeletionVars, + "Vars that would skip eager deletion." + "Users should not set this manually.") + .SetDefault(std::vector()); AddComment(R"DOC( )DOC"); } @@ -119,7 +140,10 @@ class WhileGradOp : public framework::OperatorBase { framework::Executor executor(dev_place); auto *block = Attr(kStepBlock); auto *program = block->Program(); - auto ctx = executor.Prepare(*program, block->ID()); + + auto &skip_vars = Attr>(kSkipEagerDeletionVars); + VLOG(2) << GetSkipEagerDeletionVarsDebugString(skip_vars); + auto ctx = executor.Prepare(*program, block->ID(), skip_vars); auto *step_scopes = scope.FindVar(Input(kStepScopes))->GetMutable(); @@ -132,15 +156,15 @@ class WhileGradOp : public framework::OperatorBase { for (auto cur_scope_iter = step_scopes->rbegin(); cur_scope_iter != step_scopes->rend(); ++cur_scope_iter) { - VLOG(30) << "Start backward at time_step " - << cur_scope_iter - step_scopes->rbegin(); + VLOG(3) << "Start backward at time_step " + << cur_scope_iter - step_scopes->rbegin(); framework::Scope &cur_scope = **cur_scope_iter; // Link OG from outside to inside for (size_t i = 0; i < outside_og_names.size(); ++i) { auto outside_og_name = outside_og_names[i]; auto inside_og_name = inside_og_names[i]; - VLOG(80) << "Linking outside " << outside_og_name << " --> inside " - << inside_og_name; + VLOG(8) << "Linking outside " << outside_og_name << " --> inside " + << inside_og_name; if (scope.FindVar(outside_og_name) == nullptr) { continue; } @@ -162,11 +186,11 @@ class WhileGradOp : public framework::OperatorBase { auto &outside_array = og_outside.Get(); auto &inside_array = detail::Ref(og_inside.GetMutable()); - VLOG(80) << outside_og_name << " size = " << outside_array.size(); + VLOG(8) << outside_og_name << " size = " << outside_array.size(); inside_array.resize(outside_array.size()); for (size_t j = 0; j < inside_array.size(); ++j) { - VLOG(80) << j << " " << outside_array[j].numel(); + VLOG(8) << j << " " << outside_array[j].numel(); if (outside_array[j].numel() != 0) { inside_array[j].set_lod(outside_array[j].lod()); inside_array[j].ShareDataWith(outside_array[j]); @@ -237,7 +261,7 @@ class WhileGradOp : public framework::OperatorBase { if (var->IsType()) { auto &inside_tensor = var->Get(); framework::AttributeMap attrs; - attrs["dtype"] = framework::ToDataType(inside_tensor.type()); + attrs["dtype"] = inside_tensor.type(); attrs["shape"] = framework::vectorize2int(inside_tensor.dims()); attrs["value"] = 0.0f; @@ -292,7 +316,7 @@ class WhileGradOpDescMaker : public framework::SingleGradOpDescMaker { auto igs = InputGrad(kX, /*do not drop empty gradient*/ false); for (auto &each_ig : igs) { if (inner_op_outputs.find(each_ig) == inner_op_outputs.end()) { - VLOG(80) << "Ignore " << each_ig; + VLOG(8) << "Ignore " << each_ig; each_ig = framework::kEmptyVarName; } } @@ -341,6 +365,8 @@ class WhileGradOpDescMaker : public framework::SingleGradOpDescMaker { // while operator could be renamed. while_grad->SetAttr("original_output_grad", output_grads_list); + while_grad->SetAttr(kSkipEagerDeletionVars, std::vector()); + return std::unique_ptr(while_grad); } }; @@ -356,8 +382,8 @@ class WhileGradOpVarTypeInference : public framework::VarTypeInference { auto &p_var = detail::Ref(block->FindVarRecursive(p_names[i])); auto *g_var = block->FindVarRecursive(pg_ig_names[i]); if (g_var != nullptr) { // Gradient could be @EMPTY@ - VLOG(50) << "Setting " << pg_ig_names[i] << " following " << p_names[i] - << " type: " << p_var.GetType(); + VLOG(5) << "Setting " << pg_ig_names[i] << " following " << p_names[i] + << " type: " << p_var.GetType(); g_var->SetType(p_var.GetType()); g_var->SetDataType(p_var.GetDataType()); } @@ -373,26 +399,41 @@ class WhileGradOpShapeInference : public framework::InferShapeBase { ctx->HasInputs(kOutputs); ctx->HasInputs(framework::GradVarName(kOutputs)); - auto p_names = ctx->Inputs(kX); auto pg_ig_names = ctx->Outputs(kXGRAD); - auto var_types = ctx->GetInputsVarType(kX); - std::vector names_to_set; - std::vector dims_to_set; - for (size_t i = 0; i < p_names.size(); ++i) { + std::vector in_var_ptrs = + ctx->GetInputVarPtrs(kX); + std::vector out_var_ptrs = + ctx->GetOutputVarPtrs(kXGRAD); + PADDLE_ENFORCE(in_var_ptrs.size() == out_var_ptrs.size()); + + for (size_t i = 0; i < in_var_ptrs.size(); ++i) { if (pg_ig_names[i] == framework::kEmptyVarName) { continue; } - auto dims = ctx->GetInputsElementDim(kX, i); - if (var_types[i] == framework::proto::VarType::LOD_TENSOR) { - names_to_set.push_back(pg_ig_names[i]); - dims_to_set.push_back(dims); - } else if (var_types[i] == framework::proto::VarType::LOD_TENSOR_ARRAY) { - // not sure how to set the dim of LOD_TENSOR_ARRAY - names_to_set.push_back(pg_ig_names[i]); - dims_to_set.push_back(dims); + if (ctx->IsRuntime()) { + framework::Variable *in_var = + boost::get(in_var_ptrs[i]); + framework::Variable *out_var = + boost::get(out_var_ptrs[i]); + + auto type = framework::ToVarType(in_var->Type()); + if (type == framework::proto::VarType::LOD_TENSOR) { + out_var->GetMutable()->Resize( + in_var->Get().dims()); + } else if (type == framework::proto::VarType::SELECTED_ROWS) { + out_var->GetMutable()->set_height( + in_var->Get().GetCompleteDims()[0]); + } else if (type == framework::proto::VarType::LOD_TENSOR_ARRAY) { + PADDLE_THROW("WhileGradOp doesn't support type %d", + static_cast(type)); + } + } else { + framework::VarDesc *in_var = + boost::get(in_var_ptrs[i]); + boost::get(out_var_ptrs[i]) + ->SetShape(in_var->GetShape()); } } - ctx->SetDims(names_to_set, dims_to_set); } }; diff --git a/paddle/fluid/operators/conv_cudnn_op.cu.cc b/paddle/fluid/operators/conv_cudnn_op.cu.cc index 42c2b3a24c116..dbb6ffd5e29d7 100644 --- a/paddle/fluid/operators/conv_cudnn_op.cu.cc +++ b/paddle/fluid/operators/conv_cudnn_op.cu.cc @@ -151,11 +151,11 @@ class CUDNNConvOpKernel : public framework::OpKernel { // Currently tensor core is only enabled using this algo algo = CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM; half_float = true; - VLOG(50) << "use cudnn_tensor_op_math"; + VLOG(5) << "use cudnn_tensor_op_math"; } else { CUDNN_ENFORCE(platform::dynload::cudnnSetConvolutionMathType( cudnn_conv_desc, CUDNN_DEFAULT_MATH)); - VLOG(50) << "NOT use cudnn_tensor_op_math"; + VLOG(5) << "NOT use cudnn_tensor_op_math"; } #endif diff --git a/paddle/fluid/operators/conv_fusion_op.cu.cc b/paddle/fluid/operators/conv_fusion_op.cu.cc index 2c09ee7394ad6..3235ad52b999e 100644 --- a/paddle/fluid/operators/conv_fusion_op.cu.cc +++ b/paddle/fluid/operators/conv_fusion_op.cu.cc @@ -110,11 +110,7 @@ class CUDNNConvFusionOpKernel : public framework::OpKernel { auto x_dims = framework::vectorize(input->dims()); auto f_dims = framework::vectorize(filter->dims()); - if (activation == "identity") { - // Only the CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM algo is - // enabled with CUDNN_ACTIVATION_IDENTITY in cuDNN lib. - algo = CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM; - } else if (!exhaustive_search) { + if (!exhaustive_search) { CUDNN_ENFORCE(platform::dynload::cudnnGetConvolutionForwardAlgorithm( handle, cudnn_input_desc, cudnn_filter_desc, cudnn_conv_desc, cudnn_output_desc, CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT, @@ -165,18 +161,42 @@ class CUDNNConvFusionOpKernel : public framework::OpKernel { PADDLE_ENFORCE_LE(workspace_size_in_bytes, workspace_size_limit, "workspace_size to be allocated exceeds the limit"); - // ------------------- cudnn conv+bias+act forward -------------------- - ScalingParamType alpha1 = 1.0f; - ScalingParamType alpha2 = residual ? 1.0f : 0.0f; - auto cudnn_func = [&](void* cudnn_workspace) { - CUDNN_ENFORCE(platform::dynload::cudnnConvolutionBiasActivationForward( - handle, &alpha1, cudnn_input_desc, input_data, cudnn_filter_desc, - filter_data, cudnn_conv_desc, algo, cudnn_workspace, - workspace_size_in_bytes, &alpha2, cudnn_output_desc, residual_data, - cudnn_bias_desc, bias_data, cudnn_act_desc, cudnn_output_desc, + if ((activation == "identity") && + (algo != CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM) && + (!residual)) { + // Only the CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM algo is + // enabled with CUDNN_ACTIVATION_IDENTITY in cuDNN lib. + // But test in some case, the speed is slower, change to use + // cudnnConvolutionForward and cudnnAddTensor + // ------------- cudnn conv forward and bias add --------------------- + ScalingParamType alpha = 1.0f, beta = 0.0f; + auto cudnn_func = [&](void* cudnn_workspace) { + CUDNN_ENFORCE(platform::dynload::cudnnConvolutionForward( + handle, &alpha, cudnn_input_desc, input_data, cudnn_filter_desc, + filter_data, cudnn_conv_desc, algo, cudnn_workspace, + workspace_size_in_bytes, &beta, cudnn_output_desc, output_data)); + }; + workspace_handle.RunFunc(cudnn_func, workspace_size_in_bytes); + CUDNN_ENFORCE(platform::dynload::cudnnAddTensor( + handle, &alpha, cudnn_bias_desc, bias_data, &alpha, cudnn_output_desc, output_data)); - }; - workspace_handle.RunFunc(cudnn_func, workspace_size_in_bytes); + } else { + if (activation == "identity") { + algo = CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM; + } + // ------------------- cudnn conv+bias+act forward -------------------- + ScalingParamType alpha1 = 1.0f; + ScalingParamType alpha2 = residual ? 1.0f : 0.0f; + auto cudnn_func = [&](void* cudnn_workspace) { + CUDNN_ENFORCE(platform::dynload::cudnnConvolutionBiasActivationForward( + handle, &alpha1, cudnn_input_desc, input_data, cudnn_filter_desc, + filter_data, cudnn_conv_desc, algo, cudnn_workspace, + workspace_size_in_bytes, &alpha2, cudnn_output_desc, residual_data, + cudnn_bias_desc, bias_data, cudnn_act_desc, cudnn_output_desc, + output_data)); + }; + workspace_handle.RunFunc(cudnn_func, workspace_size_in_bytes); + } } }; #endif diff --git a/paddle/fluid/operators/conv_mkldnn_op.cc b/paddle/fluid/operators/conv_mkldnn_op.cc index 9e2e2cf818000..8c116c4abfe42 100644 --- a/paddle/fluid/operators/conv_mkldnn_op.cc +++ b/paddle/fluid/operators/conv_mkldnn_op.cc @@ -15,7 +15,7 @@ #include "paddle/fluid/framework/data_layout_transform.h" #include "paddle/fluid/memory/malloc.h" #include "paddle/fluid/operators/conv_op.h" -#include "paddle/fluid/platform/mkldnn_helper.h" +#include "paddle/fluid/platform/mkldnn_reuse.h" namespace paddle { namespace operators { @@ -28,258 +28,45 @@ using mkldnn::stream; using platform::to_void_cast; using platform::GetMKLDNNFormat; -class ConvMKLDNNHandler : public platform::MKLDNNHandler { - public: - ConvMKLDNNHandler( - std::shared_ptr conv_pd, - const platform::MKLDNNDeviceContext& dev_ctx, mkldnn::engine engine, - const std::string& base_key) - : platform::MKLDNNHandler(dev_ctx, engine, base_key) { - conv_pd_ = conv_pd; - } - - ConvMKLDNNHandler( - std::shared_ptr conv_pd, - std::shared_ptr - conv_bwd_data_pd, - std::shared_ptr - conv_bwd_weights_pd, - const platform::MKLDNNDeviceContext& dev_ctx, mkldnn::engine engine, - const std::string& base_key) - : platform::MKLDNNHandler(dev_ctx, engine, base_key), - conv_pd_(conv_pd), - conv_bwd_weights_pd_(conv_bwd_weights_pd), - conv_bwd_data_pd_(conv_bwd_data_pd) { - // If we are in Grad operatgor then update a key with BWD suffix to - // distinguish from FWD memory primitives - key_ += "-BWD"; - } - - size_t GetDstMemorySize() const { - return conv_pd_->dst_primitive_desc().get_size(); - } - - mkldnn::memory::format GetDstFormat() const { - return static_cast( - conv_pd_->dst_primitive_desc().desc().data.format); - } - - size_t GetDiffWeightsMemorySize() const { - return conv_bwd_weights_pd_->diff_weights_primitive_desc().get_size(); - } - - size_t GetDiffSourceMemorySize() const { - return conv_bwd_data_pd_->diff_src_primitive_desc().get_size(); - } - - std::shared_ptr AcquireSrcMemoryFromWeightsPrimitive( - const std::shared_ptr user_memory_p, - std::vector& pipeline) { // NOLINT - auto src_pd = conv_bwd_weights_pd_->src_primitive_desc(); - auto user_pd = user_memory_p->get_primitive_desc(); - return this->AcquireMemory(src_pd, user_pd, user_memory_p, - "@weights-src_mem_p", pipeline); - } - - std::shared_ptr AcquireDiffDstMemoryFromWeightsPrimitive( - const std::shared_ptr user_memory_p, - std::vector& pipeline) { // NOLINT - auto diff_dst_pd = conv_bwd_weights_pd_->diff_dst_primitive_desc(); - auto user_pd = user_memory_p->get_primitive_desc(); - return this->AcquireMemory(diff_dst_pd, user_pd, user_memory_p, - "@weights-diff_dst_mem_p", pipeline); - } - - std::shared_ptr AcquireDiffWeightsMemoryFromWeightsPrimitive( - void* ptr) { - return this->AcquireMemoryFromPrimitive( - conv_bwd_weights_pd_->diff_weights_primitive_desc(), ptr, - "@diff_weights_mem_p"); - } - - std::shared_ptr AcquireDiffDstMemoryFromDataPrimitive( - const std::shared_ptr user_memory_p, - std::vector& pipeline) { // NOLINT - auto diff_dst_pd = conv_bwd_data_pd_->diff_dst_primitive_desc(); - auto user_pd = user_memory_p->get_primitive_desc(); - return this->AcquireMemory(diff_dst_pd, user_pd, user_memory_p, - "@data-diff_dst_mem_p", pipeline); - } - - std::shared_ptr AcquireWeightsMemoryFromDataPrimitive( - const std::shared_ptr user_weights_memory_p, - std::vector& pipeline) { // NOLINT - auto weights_pd = conv_bwd_data_pd_->weights_primitive_desc(); - auto user_pd = user_weights_memory_p->get_primitive_desc(); - return this->AcquireMemory(weights_pd, user_pd, user_weights_memory_p, - "@data-weights_mem_p", pipeline); - } - - std::shared_ptr AcquireResidualDataMemory( - const mkldnn::memory::desc& md, void* ptr) { - return this->AcquireMemory(md, ptr, "@user_residual_data_mem_p"); - } - - std::shared_ptr AcquireDstMemoryFromResidualDataMemory( - const std::shared_ptr& user_residual_memory_p, - void* dst_ptr, - std::vector& pipeline) { // NOLINT - return this->AcquireMemory(user_residual_memory_p, - this->AcquireDstMemoryFromPrimitive(dst_ptr), - "@residual_data_mem_p", pipeline); - } - - std::shared_ptr AcquireDiffSrcMemoryFromDataPrimitive( - void* ptr) { - return this->AcquireMemoryFromPrimitive( - conv_bwd_data_pd_->diff_src_primitive_desc(), ptr, "@diff_src_mem_p"); - } - - std::shared_ptr AcquireDstMemoryFromPrimitive(void* ptr) { - return this->AcquireMemoryFromPrimitive(conv_pd_->dst_primitive_desc(), ptr, - "@dst_mem_p"); - } - - std::shared_ptr AcquireSrcMemoryFromPrimitive( - const std::shared_ptr user_memory_p, - std::vector& pipeline) { // NOLINT - auto src_pd = conv_pd_->src_primitive_desc(); - auto user_pd = user_memory_p->get_primitive_desc(); - return this->AcquireMemory(src_pd, user_pd, user_memory_p, "@src_mem_p", - pipeline); - } - - std::shared_ptr AcquireWeightsMemoryFromPrimitive( - const std::shared_ptr user_weights_memory_p, - std::vector& pipeline, // NOLINT - bool is_persistent = false) { - auto user_weights_pd = user_weights_memory_p->get_primitive_desc(); - auto weights_pd = conv_pd_->weights_primitive_desc(); - return this->AcquireMemory(weights_pd, user_weights_pd, - user_weights_memory_p, "@weights_mem_p", - pipeline, is_persistent); - } - - std::shared_ptr AcquireBiasMemoryFromPrimitive( - const std::shared_ptr user_bias_memory_p, - std::vector& pipeline) { // NOLINT - auto user_bias_pd = user_bias_memory_p->get_primitive_desc(); - auto bias_pd = conv_pd_->bias_primitive_desc(); - return this->AcquireMemory(bias_pd, user_bias_pd, user_bias_memory_p, - "@bias_mem_p", pipeline); - } - - std::shared_ptr AcquireConvolution( - std::shared_ptr src_memory_p, - std::shared_ptr weights_memory_p, - std::shared_ptr dst_memory_p) { - auto prim_key = key_ + "@conv_p"; - auto conv_p = std::static_pointer_cast( - dev_ctx_.GetBlob(prim_key)); - PADDLE_ENFORCE((conv_p != nullptr) || (is_reusing_ == false), - "Fail to find convolution primitive in device context"); - if (conv_p == nullptr) { - conv_p = std::make_shared( - *conv_pd_, *(src_memory_p), *(weights_memory_p.get()), - *(dst_memory_p.get())); - - dev_ctx_.SetBlob(prim_key, conv_p); +inline void GetWeightsTz(std::vector& weights_tz, int groups, // NOLINT + bool is_conv3d) { + if (groups > 1) { + if (is_conv3d) { + int output = weights_tz[0]; + int input = weights_tz[1]; + int dimension = weights_tz[2]; + int height = weights_tz[3]; + int width = weights_tz[4]; + weights_tz.resize(6); + weights_tz[0] = groups; + weights_tz[1] = output / groups; + weights_tz[2] = input; + weights_tz[3] = dimension; + weights_tz[4] = height; + weights_tz[5] = width; } else { - is_reusing_ = true; - } - return conv_p; - } - - std::shared_ptr AcquireConvolution( - std::shared_ptr src_memory_p, - std::shared_ptr weights_memory_p, - std::shared_ptr bias_memory_p, - std::shared_ptr dst_memory_p) { - auto prim_key = key_ + "@conv_p"; - auto conv_p = std::static_pointer_cast( - dev_ctx_.GetBlob(prim_key)); - PADDLE_ENFORCE((conv_p != nullptr) || (is_reusing_ == false), - "Fail to find convolution primitive in device context"); - if (conv_p == nullptr) { - conv_p = std::make_shared( - *conv_pd_, *(src_memory_p), *(weights_memory_p.get()), - *(bias_memory_p.get()), *(dst_memory_p.get())); - - dev_ctx_.SetBlob(prim_key, conv_p); - } else { - is_reusing_ = true; - } - return conv_p; - } - - std::shared_ptr - AcquireConvolutionBackwardWeights( - std::shared_ptr src_memory_p, - std::shared_ptr diff_dst_memory_p, - std::shared_ptr diff_weights_memory_p) { - auto prim_key = key_ + "@conv_bwd_weights_p"; - auto conv_bwd_weights_p = - std::static_pointer_cast( - dev_ctx_.GetBlob(prim_key)); - PADDLE_ENFORCE( - (conv_bwd_weights_p != nullptr) || (is_reusing_ == false), - "Fail to find convolution bwd weights primitive in device context"); - if (conv_bwd_weights_p == nullptr) { - // create backward conv primitive for weights - conv_bwd_weights_p = - std::make_shared( - *conv_bwd_weights_pd_, *src_memory_p, *diff_dst_memory_p, - *diff_weights_memory_p); - dev_ctx_.SetBlob(prim_key, conv_bwd_weights_p); - } else { - is_reusing_ = true; - } - return conv_bwd_weights_p; - } - - std::shared_ptr - AcquireConvolutionBackwardData( - std::shared_ptr diff_dst_memory_p, - std::shared_ptr weights_memory_p, - std::shared_ptr diff_src_memory_p) { - auto prim_key = key_ + "@conv_bwd_data_p"; - auto conv_bwd_data_p = - std::static_pointer_cast( - dev_ctx_.GetBlob(prim_key)); - PADDLE_ENFORCE( - (conv_bwd_data_p != nullptr) || (is_reusing_ == false), - "Fail to find convolution bwd data primitive in device context"); - if (conv_bwd_data_p == nullptr) { - conv_bwd_data_p = std::make_shared( - *conv_bwd_data_pd_, *diff_dst_memory_p, *weights_memory_p, - *diff_src_memory_p); - dev_ctx_.SetBlob(prim_key, conv_bwd_data_p); - } else { - is_reusing_ = true; + int output = weights_tz[0]; + int input = weights_tz[1]; + int height = weights_tz[2]; + int width = weights_tz[3]; + weights_tz.resize(5); + weights_tz[0] = groups; + weights_tz[1] = output / groups; + weights_tz[2] = input; + weights_tz[3] = height; + weights_tz[4] = width; } - return conv_bwd_data_p; } - - // Generate keys for storing/retriving primitives for this operator - // TODO(jczaja): Make hashing function more optimial - static std::string GetHash(memory::dims& input_dims, // NOLINT - memory::dims& weights_dims, // NOLINT - std::vector& strides, // NOLINT - std::vector& paddings, // NOLINT - std::vector& dilations, // NOLINT - int groups, const std::string& suffix) { - return dims2str(input_dims) + dims2str(weights_dims) + dims2str(strides) + - dims2str(paddings) + dims2str(dilations) + std::to_string(groups) + - suffix; +} + +inline mkldnn::memory::format GetWeightsFormat(mkldnn::memory::format format, + int groups, bool is_conv3d) { + if (is_conv3d) { + return (groups == 1) ? format : mkldnn::memory::format::goidhw; + } else { + return (groups == 1) ? format : mkldnn::memory::format::goihw; } - - private: - std::shared_ptr conv_pd_; - std::shared_ptr - conv_bwd_weights_pd_; - std::shared_ptr - conv_bwd_data_pd_; -}; +} template class ConvMKLDNNOpKernel : public paddle::framework::OpKernel { @@ -305,10 +92,10 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel { PADDLE_ENFORCE(filter->layout() == DataLayout::kMKLDNN && filter->format() != memory::format::format_undef, "Wrong layout/format set for Filter tensor"); - PADDLE_ENFORCE(input->dims().size() == 4, - "Input must be with 4 dimensions, i.e. NCHW"); - PADDLE_ENFORCE(filter->dims().size() == 4, - "Filter must be with 4 dimensions, i.e. OIHW"); + PADDLE_ENFORCE(input->dims().size() == 4 || input->dims().size() == 5, + "Input must be with 4 or 5 dimensions, i.e. NCHW or NCDHW"); + PADDLE_ENFORCE(filter->dims().size() == 4 || filter->dims().size() == 5, + "Filter must be with 4 or 5 dimensions, i.e. OIHW or OIDHW"); if (bias) { PADDLE_ENFORCE(bias->layout() == DataLayout::kMKLDNN && bias->format() != memory::format::format_undef, @@ -324,9 +111,13 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel { bool fuse_residual_conn = ctx.Attr("fuse_residual_connection"); int groups = ctx.Attr("groups"); + bool is_conv3d = strides.size() == 3U; // TODO(tpatejko): add support for dilation PADDLE_ENFORCE( - dilations.size() == 2 && dilations[0] == 1 && dilations[1] == 1, + is_conv3d + ? dilations.size() == 3 && dilations[0] == 1 && dilations[1] == 1 && + dilations[2] == 1 + : dilations.size() == 2 && dilations[0] == 1 && dilations[1] == 1, "dilation in convolution is not implemented yet"); const T* input_data = input->data(); @@ -336,33 +127,25 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel { std::vector weights_tz = paddle::framework::vectorize2int(filter->dims()); int g = std::max(groups, 1); - if (g > 1) { - int o = weights_tz[0]; - int i = weights_tz[1]; - int h = weights_tz[2]; - int w = weights_tz[3]; - weights_tz.resize(5); - weights_tz[0] = g; - weights_tz[1] = o / g; - weights_tz[2] = i; - weights_tz[3] = h; - weights_tz[4] = w; - } + GetWeightsTz(weights_tz, g, is_conv3d); std::vector dst_tz = paddle::framework::vectorize2int(output->dims()); // Get unique name for storing MKLDNN primitives - const std::string key = ConvMKLDNNHandler::GetHash( + const std::string key = platform::ConvMKLDNNHandler::GetHash( src_tz, weights_tz, strides, paddings, dilations, groups, ctx.op().Output("Output")); const std::string key_conv_pd = key + "@conv_pd"; std::vector pipeline; + auto src_format = input->format(); + mkldnn::memory::format weights_format = + GetWeightsFormat(filter->format(), g, is_conv3d); + auto user_src_md = platform::MKLDNNMemDesc( - {src_tz}, platform::MKLDNNGetDataType(), input->format()); + {src_tz}, platform::MKLDNNGetDataType(), src_format); auto user_weights_md = platform::MKLDNNMemDesc( - {weights_tz}, platform::MKLDNNGetDataType(), - (g == 1) ? filter->format() : mkldnn::memory::format::goihw); + {weights_tz}, platform::MKLDNNGetDataType(), weights_format); /* create memory descriptor for convolution without specified format * ('any') which lets a primitive (convolution in this case) choose @@ -372,10 +155,19 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel { auto chosen_memory_format = platform::data_format_to_memory_format(data_format); + weights_format = mkldnn::memory::format::any; + // Check the format for user's special output + if (chosen_memory_format != mkldnn::memory::format::any) { + if (is_conv3d) { + chosen_memory_format = + platform::MKLDNNFormatForSize(src_tz.size(), chosen_memory_format); + } + } + auto src_md = platform::MKLDNNMemDesc( src_tz, platform::MKLDNNGetDataType(), chosen_memory_format); auto weights_md = platform::MKLDNNMemDesc( - weights_tz, platform::MKLDNNGetDataType(), chosen_memory_format); + weights_tz, platform::MKLDNNGetDataType(), weights_format); std::vector bias_tz; // TODO(mgallus): avoid empty vector creation. // Currently used whenever bias is != nullptr. auto dst_md = platform::MKLDNNMemDesc( @@ -400,7 +192,7 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel { // Save conv_pd/src_memory/weights_memory for backward pass if (!is_test) dev_ctx.SetBlob(key_conv_pd, conv_pd); - ConvMKLDNNHandler handler(conv_pd, dev_ctx, mkldnn_engine, key); + platform::ConvMKLDNNHandler handler(conv_pd, dev_ctx, mkldnn_engine, key); // create mkldnn memory from input tensors (data/weights) auto user_src_memory_p = @@ -516,8 +308,8 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel { const mkldnn::engine& engine, const bool fuse_relu, const bool fuse_residual_conn, mkldnn::prop_kind fwd_prop_kind) const { - memory::dims stride_dims = {strides[0], strides[1]}; - memory::dims padding_dims = {paddings[0], paddings[1]}; + memory::dims stride_dims = strides; + memory::dims padding_dims = paddings; auto conv_desc = mkldnn::convolution_forward::desc( fwd_prop_kind, mkldnn::convolution_direct, src, weights, dst, @@ -541,8 +333,8 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel { const mkldnn::engine& engine, const bool fuse_relu, const bool fuse_residual_conn, mkldnn::prop_kind fwd_prop_kind) const { - memory::dims stride_dims = {strides[0], strides[1]}; - memory::dims padding_dims = {paddings[0], paddings[1]}; + memory::dims stride_dims = strides; + memory::dims padding_dims = paddings; auto conv_desc = mkldnn::convolution_forward::desc( fwd_prop_kind, mkldnn::convolution_direct, src, weights, bias, dst, @@ -602,6 +394,7 @@ class ConvMKLDNNGradOpKernel : public paddle::framework::OpKernel { std::vector dilations = ctx.Attr>("dilations"); int groups = ctx.Attr("groups"); + bool is_conv3d = strides.size() == 3U; const T* input_data = input->data(); const T* filter_data = filter->data(); const T* output_grad_data = output_grad->data(); @@ -611,23 +404,29 @@ class ConvMKLDNNGradOpKernel : public paddle::framework::OpKernel { std::vector src_tz = paddle::framework::vectorize2int(input->dims()); std::vector weights_tz = paddle::framework::vectorize2int(filter->dims()); + int g = std::max(groups, 1); + GetWeightsTz(weights_tz, g, is_conv3d); std::vector dst_tz = paddle::framework::vectorize2int(output->dims()); + auto src_format = input->format(); + mkldnn::memory::format weights_format = + GetWeightsFormat(filter->format(), g, is_conv3d); + // Get an unique name from "argument" name of "Output" variable // as well as attributes of primitive to be created // This name will be used as key when saving info into device context - const std::string key = - ConvMKLDNNHandler::GetHash(src_tz, weights_tz, strides, paddings, - dilations, groups, ctx.op().Input("Output")); + const std::string key = platform::ConvMKLDNNHandler::GetHash( + src_tz, weights_tz, strides, paddings, dilations, groups, + ctx.op().Input("Output")); const std::string key_conv_pd = key + "@conv_pd"; std::vector pipeline; // Create user memory descriptors auto user_src_md = platform::MKLDNNMemDesc( - {src_tz}, platform::MKLDNNGetDataType(), input->format()); + {src_tz}, platform::MKLDNNGetDataType(), src_format); auto user_weights_md = platform::MKLDNNMemDesc( - {weights_tz}, platform::MKLDNNGetDataType(), filter->format()); + {weights_tz}, platform::MKLDNNGetDataType(), weights_format); auto user_diff_dst_md = platform::MKLDNNMemDesc( {dst_tz}, platform::MKLDNNGetDataType(), output_grad->format()); @@ -639,14 +438,23 @@ class ConvMKLDNNGradOpKernel : public paddle::framework::OpKernel { auto chosen_memory_format = platform::data_format_to_memory_format(data_format); + weights_format = mkldnn::memory::format::any; + // Check the format for user's special output + if (chosen_memory_format != mkldnn::memory::format::any) { + if (is_conv3d) { + chosen_memory_format = + platform::MKLDNNFormatForSize(src_tz.size(), chosen_memory_format); + } + } + auto src_md = platform::MKLDNNMemDesc( src_tz, platform::MKLDNNGetDataType(), chosen_memory_format); auto diff_src_md = platform::MKLDNNMemDesc( src_tz, platform::MKLDNNGetDataType(), chosen_memory_format); auto weights_md = platform::MKLDNNMemDesc( - weights_tz, platform::MKLDNNGetDataType(), chosen_memory_format); + weights_tz, platform::MKLDNNGetDataType(), weights_format); auto diff_weights_md = platform::MKLDNNMemDesc( - weights_tz, platform::MKLDNNGetDataType(), chosen_memory_format); + weights_tz, platform::MKLDNNGetDataType(), weights_format); auto diff_dst_md = platform::MKLDNNMemDesc( dst_tz, platform::MKLDNNGetDataType(), chosen_memory_format); @@ -673,8 +481,9 @@ class ConvMKLDNNGradOpKernel : public paddle::framework::OpKernel { std::make_shared( conv_bwd_data_desc, mkldnn_engine, *conv_pd); - ConvMKLDNNHandler handler(conv_pd, conv_bwd_data_pd, conv_bwd_weights_pd, - dev_ctx, mkldnn_engine, key); + platform::ConvMKLDNNHandler handler(conv_pd, conv_bwd_data_pd, + conv_bwd_weights_pd, dev_ctx, + mkldnn_engine, key); // create mkldnn memory from input tensors (data/weights) auto user_src_memory_p = @@ -743,8 +552,22 @@ class ConvMKLDNNGradOpKernel : public paddle::framework::OpKernel { namespace ops = paddle::operators; -REGISTER_OP_KERNEL(conv2d, MKLDNN, ::paddle::platform::CPUPlace, - ops::ConvMKLDNNOpKernel); - -REGISTER_OP_KERNEL(conv2d_grad, MKLDNN, ::paddle::platform::CPUPlace, - ops::ConvMKLDNNGradOpKernel); +REGISTER_OP_KERNEL_WITH_CUSTOM_TYPE(conv2d, MKLDNN, + ::paddle::platform::CPUPlace, FP32, + ops::kConvMKLDNNFP32, + ops::ConvMKLDNNOpKernel); + +REGISTER_OP_KERNEL_WITH_CUSTOM_TYPE(conv2d_grad, MKLDNN, + ::paddle::platform::CPUPlace, FP32, + ops::kConvMKLDNNFP32, + ops::ConvMKLDNNGradOpKernel); + +REGISTER_OP_KERNEL_WITH_CUSTOM_TYPE(conv3d, MKLDNN, + ::paddle::platform::CPUPlace, FP32, + ops::kConvMKLDNNFP32, + ops::ConvMKLDNNOpKernel); + +REGISTER_OP_KERNEL_WITH_CUSTOM_TYPE(conv3d_grad, MKLDNN, + ::paddle::platform::CPUPlace, FP32, + ops::kConvMKLDNNFP32, + ops::ConvMKLDNNGradOpKernel); diff --git a/paddle/fluid/operators/conv_op.cc b/paddle/fluid/operators/conv_op.cc index 342525be49e28..8e0d2824953a3 100644 --- a/paddle/fluid/operators/conv_op.cc +++ b/paddle/fluid/operators/conv_op.cc @@ -44,7 +44,9 @@ void ConvOp::InferShape(framework::InferShapeContext* ctx) const { std::vector dilations = ctx->Attrs().Get>("dilations"); PADDLE_ENFORCE(in_dims.size() == 4 || in_dims.size() == 5, - "Conv intput should be 4-D or 5-D tensor."); + "Conv intput should be 4-D or 5-D tensor, get %u", + in_dims.size()); + PADDLE_ENFORCE_EQ( in_dims.size(), filter_dims.size(), "Conv input dimension and filter dimension should be the same."); @@ -74,6 +76,8 @@ void ConvOp::InferShape(framework::InferShapeContext* ctx) const { framework::OpKernelType ConvOp::GetExpectedKernelType( const framework::ExecutionContext& ctx) const { + int customized_type_value = + framework::OpKernelType::kDefaultCustomizedTypeValue; framework::LibraryType library{framework::LibraryType::kPlain}; // TODO(pzelazko-intel): enable MKLDNN layout when it's ready std::string data_format = ctx.Attr("data_format"); @@ -89,13 +93,12 @@ framework::OpKernelType ConvOp::GetExpectedKernelType( platform::CanMKLDNNBeUsed(ctx)) { library = framework::LibraryType::kMKLDNN; layout = framework::DataLayout::kMKLDNN; + customized_type_value = kConvMKLDNNFP32; } #endif - auto input_data_type = - framework::ToDataType(ctx.Input("Input")->type()); - auto filter_data_type = - framework::ToDataType(ctx.Input("Filter")->type()); + auto input_data_type = ctx.Input("Input")->type(); + auto filter_data_type = ctx.Input("Filter")->type(); PADDLE_ENFORCE_EQ(input_data_type, filter_data_type, "input and filter data type should be consistent"); @@ -105,7 +108,7 @@ framework::OpKernelType ConvOp::GetExpectedKernelType( } return framework::OpKernelType(input_data_type, ctx.GetPlace(), layout, - library); + library, customized_type_value); } void Conv2DOpMaker::Make() { @@ -131,14 +134,14 @@ void Conv2DOpMaker::Make() { "The format of output tensor is X (one-dimensional) of size equal" "to the number of output channels. Only used with MKL-DNN.") .AsDispensable(); - AddOutput("Output", - "(Tensor) The output tensor of convolution operator. " - "The format of output tensor is also NCHW."); AddInput("ResidualData", "(Tensor) Tensor with residual data " "to which convolution output will be added." "Used with fuse_residual_connection fusion.") .AsDispensable(); + AddOutput("Output", + "(Tensor) The output tensor of convolution operator. " + "The format of output tensor is also NCHW."); AddAttr>("strides", "(vector default:{1, 1}), the " "strides(h_stride, w_stride) of " @@ -229,6 +232,10 @@ The input(X) size and output(Out) size may be different. } void Conv3DOpMaker::Make() { + AddAttr("is_test", + "(bool, default false) Set to true for inference only, false " + "for training. Some layers may run faster when this is true.") + .SetDefault(false); AddInput( "Input", "(Tensor) The input tensor of convolution operator. " @@ -244,6 +251,11 @@ void Conv3DOpMaker::Make() { "is the width of the filter." "If the groups attribute is greater than 1, C equals the number of " "input image channels divided by the groups."); + AddInput("ResidualData", + "(Tensor) Tensor with residual data " + "to which convolution output will be added." + "Used with fuse_residual_connection fusion.") + .AsDispensable(); AddOutput("Output", "(Tensor) The output tensor of convolution operator." "The format of output tensor is also NCDHW."); @@ -277,6 +289,13 @@ void Conv3DOpMaker::Make() { AddAttr("use_mkldnn", "(bool, default false) Only used in mkldnn kernel") .SetDefault(false); + AddAttr("fuse_relu", "(bool, default false) Only used in mkldnn kernel") + .SetDefault(false); + AddAttr("fuse_residual_connection", + "(bool, default false) Only used in mkldnn kernel. Used " + "whenever convolution output is as an input to residual " + "connection.") + .SetDefault(false); AddAttr( "data_format", "(string, default NCHW) Only used in " @@ -342,6 +361,8 @@ void ConvOpGrad::InferShape(framework::InferShapeContext* ctx) const { framework::OpKernelType ConvOpGrad::GetExpectedKernelType( const framework::ExecutionContext& ctx) const { + int customized_type_value = + framework::OpKernelType::kDefaultCustomizedTypeValue; framework::LibraryType library_{framework::LibraryType::kPlain}; // TODO(pzelazko-intel): enable MKLDNN layout when it's ready std::string data_format = ctx.Attr("data_format"); @@ -357,12 +378,13 @@ framework::OpKernelType ConvOpGrad::GetExpectedKernelType( platform::CanMKLDNNBeUsed(ctx)) { library_ = framework::LibraryType::kMKLDNN; layout_ = framework::DataLayout::kMKLDNN; + customized_type_value = kConvMKLDNNFP32; } #endif - return framework::OpKernelType( - framework::ToDataType(ctx.Input("Input")->type()), ctx.GetPlace(), - layout_, library_); + return framework::OpKernelType(ctx.Input("Input")->type(), + ctx.GetPlace(), layout_, library_, + customized_type_value); } } // namespace operators diff --git a/paddle/fluid/operators/conv_op.h b/paddle/fluid/operators/conv_op.h index e69814001e4da..4a7b31c7d491f 100644 --- a/paddle/fluid/operators/conv_op.h +++ b/paddle/fluid/operators/conv_op.h @@ -22,11 +22,14 @@ limitations under the License. */ #include "paddle/fluid/operators/math/depthwise_conv.h" #include "paddle/fluid/operators/math/im2col.h" #include "paddle/fluid/operators/math/vol2col.h" +#include "paddle/fluid/platform/create_tensor_with_allocationptr.h" namespace paddle { namespace operators { using Tensor = framework::Tensor; +constexpr int kConvMKLDNNFP32 = 1; +constexpr int kConvMKLDNNINT8 = 2; // Base convolution operator definations for other conv // like operators to reuse the implementation. @@ -121,6 +124,8 @@ class GemmConvKernel : public framework::OpKernel { std::vector paddings = context.Attr>("paddings"); std::vector dilations = context.Attr>("dilations"); + auto& dev_ctx = context.template device_context(); + const int batch_size = static_cast(input->dims()[0]); // filter_shape_vec: {k_o, k_i, k_h, k_w} or {k_o, k_i, k_d, k_h, k_w} @@ -153,13 +158,19 @@ class GemmConvKernel : public framework::OpKernel { // to call the matrix multiplication interface. Tensor col_matrix; if (is_expand) { - col.mutable_data(col_shape, context.GetPlace()); + auto tmp_allocation_ptr = + platform::DeviceTemporaryAllocator::Instance().Get(dev_ctx).Allocate( + framework::product(col_shape) * sizeof(T)); + Tensor tep_tensor = + platform::GetTensor(std::move(tmp_allocation_ptr), col_shape); + + col.ShareDataWith(tep_tensor); col_matrix.ShareDataWith(col); col_matrix.Resize(col_matrix_shape); } - framework::DDim input_shape = framework::slice_ddim( - input->dims(), 1, static_cast(input->dims().size())); + framework::DDim input_shape = + framework::slice_ddim(input->dims(), 1, input->dims().size()); framework::DDim filter_matrix_shape = {filter.dims()[0], filter.numel() / filter.dims()[0]}; @@ -176,7 +187,6 @@ class GemmConvKernel : public framework::OpKernel { math::Vol2ColFunctor vol2col; math::Im2ColFunctor im2col; - auto& dev_ctx = context.template device_context(); auto blas = math::GetBlas(dev_ctx); for (int i = 0; i < batch_size; i++) { Tensor in_batch = input->Slice(i, i + 1).Resize(input_shape); @@ -235,6 +245,8 @@ class GemmConvGradKernel : public framework::OpKernel { const int batch_size = static_cast(input->dims()[0]); + auto& dev_ctx = context.template device_context(); + // filter_shape_vec: {k_o, k_i, k_h, k_w} or {k_o, k_i, k_d, k_h, k_w} std::vector filter_shape_vec(framework::vectorize(filter.dims())); // output_shape_vec: {o_n, o_c, o_h, o_w} or {o_n, o_c, o_d, o_h, o_w} @@ -260,8 +272,8 @@ class GemmConvGradKernel : public framework::OpKernel { framework::DDim col_matrix_shape = framework::flatten_to_2d(col_shape, data_dim + 1); - framework::DDim input_shape = framework::slice_ddim( - input->dims(), 1, static_cast(input->dims().size())); + framework::DDim input_shape = + framework::slice_ddim(input->dims(), 1, input->dims().size()); framework::DDim filter_matrix_shape = {filter.dims()[0], filter.numel() / filter.dims()[0]}; @@ -284,13 +296,18 @@ class GemmConvGradKernel : public framework::OpKernel { // to call the matrix multiplication interface. Tensor col_matrix; if (is_expand) { - col.mutable_data(col_shape, context.GetPlace()); + auto tmp_allocation_ptr = + platform::DeviceTemporaryAllocator::Instance().Get(dev_ctx).Allocate( + framework::product(col_shape) * sizeof(T)); + Tensor tep_tensor = + platform::GetTensor(std::move(tmp_allocation_ptr), col_shape); + + col.ShareDataWith(tep_tensor); col_matrix.ShareDataWith(col); col_matrix.Resize(col_matrix_shape); } math::SetConstant set_zero; - auto& dev_ctx = context.template device_context(); auto blas = math::GetBlas(dev_ctx); if (input_grad) { diff --git a/paddle/fluid/operators/conv_transpose_mkldnn_op.cc b/paddle/fluid/operators/conv_transpose_mkldnn_op.cc new file mode 100644 index 0000000000000..317d4cebe26b8 --- /dev/null +++ b/paddle/fluid/operators/conv_transpose_mkldnn_op.cc @@ -0,0 +1,299 @@ +/* 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/data_layout_transform.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/memory/malloc.h" +#include "paddle/fluid/platform/mkldnn_reuse.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; +using framework::DataLayout; + +template +class ConvTransposeMKLDNNOpKernel : public paddle::framework::OpKernel { + public: + void Compute(const paddle::framework::ExecutionContext& ctx) const override { + PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()), + "It must use CPUPlace."); + + const bool is_test = ctx.Attr("is_test"); + PADDLE_ENFORCE( + is_test == true, + "ConvTransposeMKLDNN works only for inference!. Set is_test = True"); + + auto& dev_ctx = + ctx.template device_context(); + const auto& mkldnn_engine = dev_ctx.GetEngine(); + + auto* input = ctx.Input("Input"); + auto* filter = ctx.Input("Filter"); + auto* bias = ctx.HasInput("Bias") ? ctx.Input("Bias") : nullptr; + auto* output = ctx.Output("Output"); + + PADDLE_ENFORCE(input->layout() == DataLayout::kMKLDNN && + input->format() != mkldnn::memory::format::format_undef, + "Wrong layout/format set for Input tensor"); + PADDLE_ENFORCE(filter->layout() == DataLayout::kMKLDNN && + filter->format() != mkldnn::memory::format::format_undef, + "Wrong layout/format set for Filter tensor"); + PADDLE_ENFORCE(input->dims().size() == 4, + "Input must be with 4 dimensions, i.e. NCHW"); + PADDLE_ENFORCE(filter->dims().size() == 4, + "Filter must be with 4 dimensions, i.e. OIHW"); + + if (bias) { + PADDLE_ENFORCE(bias->layout() == DataLayout::kMKLDNN && + bias->format() != mkldnn::memory::format::format_undef, + "Wrong layout/format set for Bias tensor"); + PADDLE_ENFORCE(bias->dims().size() == 1, + "Bias must only have 1 dimension, i.e. X"); + } + + std::vector strides = ctx.Attr>("strides"); + std::vector paddings = ctx.Attr>("paddings"); + std::vector dilations = ctx.Attr>("dilations"); + int groups = ctx.Attr("groups"); + + // TODO(tpatejko): add support for dilation + PADDLE_ENFORCE( + dilations.size() == 2 && dilations[0] == 1 && dilations[1] == 1, + "dilation in convolution is not implemented yet"); + + const T* input_data = input->data(); + const T* filter_data = filter->data(); + + std::vector src_tz = paddle::framework::vectorize2int(input->dims()); + std::vector iohw_weights_tz = + paddle::framework::vectorize2int(filter->dims()); + std::vector weights_tz = iohw_weights_tz; + // IOHW -> OIHW + weights_tz[0] = iohw_weights_tz[1]; + weights_tz[1] = iohw_weights_tz[0]; + + // Custom Reorder from IOHW to OIHW + auto iohw2oihw_reorder = + [&iohw_weights_tz](const T* filter_data) -> std::shared_ptr { + int o = iohw_weights_tz[1]; + int c = iohw_weights_tz[0]; + int h = iohw_weights_tz[2]; + int w = iohw_weights_tz[3]; + std::shared_ptr reordered_filter_data(new T[o * c * h * w](), + std::default_delete()); + for (int i = 0; i < c; ++i) { + for (int j = 0; j < o; ++j) { + int in_offset = j * h * w + i * o * h * w; + int out_offset = j * c * h * w + i * h * w; + std::memcpy(&(reordered_filter_data.get())[out_offset], + &filter_data[in_offset], h * w * sizeof(T)); + } + } + + return reordered_filter_data; + }; + + int g = std::max(groups, 1); + if (g > 1) { + int o = weights_tz[0]; + int i = weights_tz[1]; + int h = weights_tz[2]; + int w = weights_tz[3]; + weights_tz.resize(5); + weights_tz[0] = g; + weights_tz[1] = o / g; + weights_tz[2] = i; + weights_tz[3] = h; + weights_tz[4] = w; + } + std::vector dst_tz = paddle::framework::vectorize2int(output->dims()); + + // Get unique name for storing MKLDNN primitives + const std::string key = platform::ConvTransposeMKLDNNHandler::GetHash( + src_tz, weights_tz, strides, paddings, dilations, groups, + ctx.op().Output("Output")); + const std::string key_conv_transpose_pd = key + "@conv_transpose_pd"; + + std::vector pipeline; + + auto user_src_md = platform::MKLDNNMemDesc( + {src_tz}, platform::MKLDNNGetDataType(), input->format()); + auto user_weights_md = + platform::MKLDNNMemDesc({weights_tz}, platform::MKLDNNGetDataType(), + (g == 1) ? mkldnn::memory::format::oihw + : mkldnn::memory::format::goihw); + + /* create memory descriptor for convolution without specified format + * ('any') which lets a primitive (convolution in this case) choose + * the memory format preferred for best performance + */ + std::string data_format = ctx.Attr("data_format"); + auto chosen_memory_format = + platform::data_format_to_memory_format(data_format); + bool fuse_relu = ctx.Attr("fuse_relu"); + + auto src_md = platform::MKLDNNMemDesc( + src_tz, platform::MKLDNNGetDataType(), chosen_memory_format); + auto weights_md = platform::MKLDNNMemDesc( + weights_tz, platform::MKLDNNGetDataType(), chosen_memory_format); + std::vector bias_tz; // TODO(mgallus): avoid empty vector creation. + // Currently used whenever bias is != nullptr. + auto dst_md = platform::MKLDNNMemDesc( + dst_tz, platform::MKLDNNGetDataType(), chosen_memory_format); + + // create a deconv(conv transpose) primitive descriptor and save it for + // usage in backward + std::shared_ptr + conv_transpose_pd; + auto fwd_prop_kind = is_test ? mkldnn::prop_kind::forward_inference + : mkldnn::prop_kind::forward_training; + if (bias) { + bias_tz = paddle::framework::vectorize2int(bias->dims()); + auto bias_md = platform::MKLDNNMemDesc( + bias_tz, platform::MKLDNNGetDataType(), mkldnn::memory::format::x); + conv_transpose_pd = ConvTransposeFwdPrimitiveDesc( + src_md, weights_md, bias_md, dst_md, strides, paddings, mkldnn_engine, + fuse_relu, fwd_prop_kind); + } else { + conv_transpose_pd = ConvTransposeFwdPrimitiveDesc( + src_md, weights_md, dst_md, strides, paddings, mkldnn_engine, + fuse_relu, fwd_prop_kind); + } + // Save conv_pd/src_memory/weights_memory for backward pass + if (!is_test) dev_ctx.SetBlob(key_conv_transpose_pd, conv_transpose_pd); + + platform::ConvTransposeMKLDNNHandler handler(conv_transpose_pd, dev_ctx, + mkldnn_engine, key); + + // create mkldnn memory from input tensors (data/weights) + auto user_src_memory_p = handler.AcquireSrcMemory( + user_src_md, platform::to_void_cast(input_data)); + auto user_weights_memory_p = handler.AcquireWeightsMemory( + user_weights_md, platform::to_void_cast(filter_data), + is_test ? iohw2oihw_reorder : platform::user_function()); + + // create reorder primitive if the input format is not the preferred one + auto src_memory_p = + handler.AcquireSrcMemoryFromPrimitive(user_src_memory_p, pipeline); + auto weights_memory_p = handler.AcquireWeightsMemoryFromPrimitive( + user_weights_memory_p, pipeline, is_test); + + std::shared_ptr dst_memory_p; + + auto output_data = output->mutable_data( + ctx.GetPlace(), paddle::memory::Allocator::kDefault, + handler.GetDstMemorySize()); + dst_memory_p = handler.AcquireDstMemoryFromPrimitive( + platform::to_void_cast(output_data)); + + // create convolution op primitive + std::shared_ptr conv_p; + if (bias) { + const T* bias_data = bias->data(); + auto user_bias_md = + platform::MKLDNNMemDesc({bias_tz}, platform::MKLDNNGetDataType(), + mkldnn::memory::format::x); + auto user_bias_memory_p = handler.AcquireBiasMemory( + user_bias_md, platform::to_void_cast(bias_data)); + + auto bias_memory_p = + handler.AcquireBiasMemoryFromPrimitive(user_bias_memory_p, pipeline); + conv_p = handler.AcquireConvolution(src_memory_p, weights_memory_p, + bias_memory_p, dst_memory_p); + } else { + conv_p = handler.AcquireConvolution(src_memory_p, weights_memory_p, + dst_memory_p); + } + + // push primitive to stream and wait until it's executed + pipeline.push_back(*conv_p); + mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); + + output->set_layout(DataLayout::kMKLDNN); + output->set_format(platform::GetMKLDNNFormat(*dst_memory_p)); + } + + private: + mkldnn::primitive_attr CreatePostOps(bool fuse_relu) const { + mkldnn::primitive_attr conv_attr; + mkldnn::post_ops post_operations; + // Fusion with ReLU layer is executed through the PostOps feature. Create a + // PostOps object and configure it to execute an eltwise relu operation. + if (fuse_relu) { + constexpr float scale = 1.0f; + constexpr float negative_slope = 0.0f; + constexpr float placeholder = 0.0f; + post_operations.append_eltwise(scale, mkldnn::algorithm::eltwise_relu, + negative_slope, placeholder); + } + conv_attr.set_post_ops(post_operations); + return conv_attr; + } + + std::unique_ptr + ConvTransposeFwdPrimitiveDesc( + const mkldnn::memory::desc& src, const mkldnn::memory::desc& weights, + const mkldnn::memory::desc& dst, const std::vector& strides, + const std::vector& paddings, const mkldnn::engine& engine, + const bool fuse_relu, mkldnn::prop_kind fwd_prop_kind) const { + mkldnn::memory::dims stride_dims = {strides[0], strides[1]}; + mkldnn::memory::dims padding_dims = {paddings[0], paddings[1]}; + + auto deconv_desc = mkldnn::deconvolution_forward::desc( + fwd_prop_kind, mkldnn::deconvolution_direct, src, weights, dst, + stride_dims, padding_dims, padding_dims, mkldnn::padding_kind::zero); + + mkldnn::primitive_attr deconv_attr = CreatePostOps(fuse_relu); + + auto p_conv_transpose_pd = + new mkldnn::deconvolution_forward::primitive_desc(deconv_desc, + deconv_attr, engine); + + return std::unique_ptr( + p_conv_transpose_pd); + } + + std::unique_ptr + ConvTransposeFwdPrimitiveDesc( + const mkldnn::memory::desc& src, const mkldnn::memory::desc& weights, + const mkldnn::memory::desc& bias, const mkldnn::memory::desc& dst, + const std::vector& strides, const std::vector& paddings, + const mkldnn::engine& engine, const bool fuse_relu, + mkldnn::prop_kind fwd_prop_kind) const { + mkldnn::memory::dims stride_dims = {strides[0], strides[1]}; + mkldnn::memory::dims padding_dims = {paddings[0], paddings[1]}; + + auto deconv_desc = mkldnn::deconvolution_forward::desc( + fwd_prop_kind, mkldnn::deconvolution_direct, src, weights, bias, dst, + stride_dims, padding_dims, padding_dims, mkldnn::padding_kind::zero); + + mkldnn::primitive_attr deconv_attr = CreatePostOps(fuse_relu); + + auto p_conv_transpose_pd = + new mkldnn::deconvolution_forward::primitive_desc(deconv_desc, + deconv_attr, engine); + + return std::unique_ptr( + p_conv_transpose_pd); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; + +REGISTER_OP_KERNEL(conv2d_transpose, MKLDNN, ::paddle::platform::CPUPlace, + ops::ConvTransposeMKLDNNOpKernel); diff --git a/paddle/fluid/operators/conv_transpose_op.cc b/paddle/fluid/operators/conv_transpose_op.cc index a916dd3496ffa..86a140f152190 100644 --- a/paddle/fluid/operators/conv_transpose_op.cc +++ b/paddle/fluid/operators/conv_transpose_op.cc @@ -16,6 +16,10 @@ limitations under the License. */ #include #include +#ifdef PADDLE_WITH_MKLDNN +#include "paddle/fluid/platform/mkldnn_helper.h" +#endif + namespace paddle { namespace operators { @@ -78,29 +82,37 @@ void ConvTransposeOp::InferShape(framework::InferShapeContext* ctx) const { framework::OpKernelType ConvTransposeOp::GetExpectedKernelType( const framework::ExecutionContext& ctx) const { + framework::LibraryType library_{framework::LibraryType::kPlain}; + std::string data_format = ctx.Attr("data_format"); + framework::DataLayout layout_ = framework::StringToDataLayout(data_format); bool use_cudnn = ctx.Attr("use_cudnn"); use_cudnn &= platform::is_gpu_place(ctx.GetPlace()); #ifdef PADDLE_WITH_CUDA if (platform::is_gpu_place(ctx.GetPlace())) { auto& dev_ctx = ctx.template device_context(); use_cudnn &= dev_ctx.cudnn_handle() != nullptr; + if (use_cudnn) { + library_ = framework::LibraryType::kCUDNN; + } } #endif - framework::LibraryType library_; - if (use_cudnn) { - library_ = framework::LibraryType::kCUDNN; - } else { - library_ = framework::LibraryType::kPlain; +#ifdef PADDLE_WITH_MKLDNN + if (library_ == framework::LibraryType::kPlain && + platform::CanMKLDNNBeUsed(ctx)) { + library_ = framework::LibraryType::kMKLDNN; + layout_ = framework::DataLayout::kMKLDNN; } +#endif - std::string data_format = ctx.Attr("data_format"); - framework::DataLayout layout_ = framework::StringToDataLayout(data_format); - return framework::OpKernelType( - framework::ToDataType(ctx.Input("Input")->type()), ctx.GetPlace(), - layout_, library_); + return framework::OpKernelType(ctx.Input("Input")->type(), + ctx.GetPlace(), layout_, library_); } void Conv2DTransposeOpMaker::Make() { + AddAttr("is_test", + "(bool, default false) Set to true for inference only, false " + "for training. Some layers may run faster when this is true.") + .SetDefault(false); AddInput( "Input", "(Tensor) The input tensor of convolution transpose operator. " @@ -145,6 +157,11 @@ void Conv2DTransposeOpMaker::Make() { "use_cudnn", "(bool, default false) Only used in cudnn kernel, need install cudnn") .SetDefault(false); + AddAttr("use_mkldnn", + "(bool, default false) Only used in mkldnn kernel") + .SetDefault(false); + AddAttr("fuse_relu", "(bool, default false) Only used in mkldnn kernel") + .SetDefault(false); AddAttr( "data_format", "(string, default NCHW) Only used in " @@ -238,6 +255,9 @@ void Conv3DTransposeOpMaker::Make() { "use_cudnn", "(bool, default false) Only used in cudnn kernel, need install cudnn") .SetDefault(false); + AddAttr("use_mkldnn", + "(bool, default false) Only used in mkldnn kernel") + .SetDefault(false); AddAttr( "data_format", "(string, default NCHW) Only used in " @@ -314,9 +334,8 @@ framework::OpKernelType ConvTransposeOpGrad::GetExpectedKernelType( std::string data_format = ctx.Attr("data_format"); framework::DataLayout layout_ = framework::StringToDataLayout(data_format); - return framework::OpKernelType( - framework::ToDataType(ctx.Input("Input")->type()), ctx.GetPlace(), - layout_, library_); + return framework::OpKernelType(ctx.Input("Input")->type(), + ctx.GetPlace(), layout_, library_); } } // namespace operators diff --git a/paddle/fluid/operators/cos_sim_op.cu b/paddle/fluid/operators/cos_sim_op.cu index 82205e9c75402..3d144ca29d998 100644 --- a/paddle/fluid/operators/cos_sim_op.cu +++ b/paddle/fluid/operators/cos_sim_op.cu @@ -11,8 +11,6 @@ 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. */ - -#define EIGEN_USE_GPU #include "paddle/fluid/operators/cos_sim_op.h" namespace ops = paddle::operators; diff --git a/paddle/fluid/operators/crf_decoding_op.cc b/paddle/fluid/operators/crf_decoding_op.cc index c27befe1143ba..81c9e9e543191 100644 --- a/paddle/fluid/operators/crf_decoding_op.cc +++ b/paddle/fluid/operators/crf_decoding_op.cc @@ -118,9 +118,8 @@ class CRFDecodingOp : public framework::OperatorWithKernel { protected: framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext& ctx) const override { - return framework::OpKernelType( - framework::ToDataType(ctx.Input("Emission")->type()), - platform::CPUPlace()); + return framework::OpKernelType(ctx.Input("Emission")->type(), + platform::CPUPlace()); } }; } // namespace operators diff --git a/paddle/fluid/operators/crf_decoding_op.h b/paddle/fluid/operators/crf_decoding_op.h index e9d2e84a434d7..72774a878d98b 100644 --- a/paddle/fluid/operators/crf_decoding_op.h +++ b/paddle/fluid/operators/crf_decoding_op.h @@ -16,7 +16,7 @@ limitations under the License. */ #include #include "paddle/fluid/framework/eigen.h" #include "paddle/fluid/framework/op_registry.h" -#include "paddle/fluid/operators/math/jit_kernel.h" +#include "paddle/fluid/operators/jit/kernels.h" #include "paddle/fluid/operators/math/math_function.h" namespace paddle { @@ -82,10 +82,9 @@ class CRFDecodingOpKernel : public framework::OpKernel { Tensor track; int* track_value = track.mutable_data(emission_dims, platform::CPUPlace()); - const auto& ker = math::jitkernel::KernelPool::Instance() - .template Get>( - static_cast(tag_num)); - ker->Compute(static_cast(seq_len), x, w, alpha_value, track_value); + auto ker = jit::Get, + platform::CPUPlace>(tag_num); + ker(static_cast(seq_len), x, w, alpha_value, track_value, tag_num); T max_score = -std::numeric_limits::max(); int max_i = 0; for (size_t i = 0; i < tag_num; ++i) { diff --git a/paddle/fluid/operators/crop_op.cc b/paddle/fluid/operators/crop_op.cc index a2a871efa850d..97d20681b8136 100644 --- a/paddle/fluid/operators/crop_op.cc +++ b/paddle/fluid/operators/crop_op.cc @@ -51,9 +51,8 @@ class CropOp : public framework::OperatorWithKernel { framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext& ctx) const override { - return framework::OpKernelType( - framework::ToDataType(ctx.Input("X")->type()), - ctx.device_context()); + return framework::OpKernelType(ctx.Input("X")->type(), + ctx.device_context()); } }; @@ -174,9 +173,7 @@ class CropOpGrad : public framework::OperatorWithKernel { framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext& ctx) const override { return framework::OpKernelType( - framework::ToDataType( - ctx.Input(framework::GradVarName("Out")) - ->type()), + ctx.Input(framework::GradVarName("Out"))->type(), ctx.device_context()); } }; diff --git a/paddle/fluid/operators/crop_op.cu b/paddle/fluid/operators/crop_op.cu index b75678217e36a..66cb5c452de4b 100644 --- a/paddle/fluid/operators/crop_op.cu +++ b/paddle/fluid/operators/crop_op.cu @@ -11,8 +11,6 @@ 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. */ - -#define EIGEN_USE_GPU #include "paddle/fluid/operators/crop_op.h" namespace ops = paddle::operators; diff --git a/paddle/fluid/operators/cross_entropy_op.cc b/paddle/fluid/operators/cross_entropy_op.cc index a904dd91302c9..1968e54b00601 100644 --- a/paddle/fluid/operators/cross_entropy_op.cc +++ b/paddle/fluid/operators/cross_entropy_op.cc @@ -57,9 +57,8 @@ class CrossEntropyOp : public framework::OperatorWithKernel { // is determined by its input "X". framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext& ctx) const override { - return framework::OpKernelType( - framework::ToDataType(ctx.Input("X")->type()), - ctx.device_context()); + return framework::OpKernelType(ctx.Input("X")->type(), + ctx.device_context()); } }; @@ -111,9 +110,8 @@ class CrossEntropyGradientOp : public framework::OperatorWithKernel { // is determined by its input "X". framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext& ctx) const override { - return framework::OpKernelType( - framework::ToDataType(ctx.Input("X")->type()), - ctx.device_context()); + return framework::OpKernelType(ctx.Input("X")->type(), + ctx.device_context()); } }; diff --git a/paddle/fluid/operators/ctc_align_op.cc b/paddle/fluid/operators/ctc_align_op.cc index d2b440d9d2e50..e7c472f8c0ce2 100644 --- a/paddle/fluid/operators/ctc_align_op.cc +++ b/paddle/fluid/operators/ctc_align_op.cc @@ -36,9 +36,8 @@ class CTCAlignOp : public framework::OperatorWithKernel { protected: framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext& ctx) const override { - return framework::OpKernelType( - framework::ToDataType(ctx.Input("Input")->type()), - ctx.device_context()); + return framework::OpKernelType(ctx.Input("Input")->type(), + ctx.device_context()); } }; diff --git a/paddle/fluid/operators/cudnn_lstm_op.cc b/paddle/fluid/operators/cudnn_lstm_op.cc new file mode 100644 index 0000000000000..e63d57be57a66 --- /dev/null +++ b/paddle/fluid/operators/cudnn_lstm_op.cc @@ -0,0 +1,218 @@ +/* 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 +#include "paddle/fluid/framework/op_registry.h" + +namespace paddle { +namespace operators { + +class CudnnLSTMOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Input"), + "Input(Input) of LSTM should not be null."); + PADDLE_ENFORCE(ctx->HasInput("W"), + "Input(Weight) of LSTM should not be null."); + + PADDLE_ENFORCE(ctx->HasInput("InitH"), + "Input(init_h) of LSTM should not be null."); + PADDLE_ENFORCE(ctx->HasInput("InitC"), + "Input(init_c) of LSTM should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Cache"), + "Input(Cache) of LSTM should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of LSTM should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("last_h"), + "Output(last_h) of LSTM should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("last_c"), + "Output(last_c) of LSTM should not be null."); + + auto in_dims = ctx->GetInputDim("Input"); + PADDLE_ENFORCE_EQ(in_dims.size(), 3, "Input(X)'s rank must be 3."); + + ctx->SetOutputDim("Out", ctx->GetInputDim("Input")); + ctx->SetOutputDim("last_h", ctx->GetInputDim("InitH")); + ctx->SetOutputDim("last_c", ctx->GetInputDim("InitC")); + } +}; + +class CudnnLSTMOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput( + "Input", + "(Tensor) RNN input tensor, which support variable-time length input " + "sequence." + "The shape of the Tensor MUST be ( seq_len * batch_size * input_size)" + "seq_len is the total time step in this mini-batch (CAN be change in " + "different batch)" + "batch_size is the instance number of this batch" + "input_size is the hidden size of the input." + "input_hidden_size and the hidden_size in the next may not be same"); + AddInput("InitH", + "(Tensor) the initial hidden state of the LSTM" + "input. This is a tensor with shape (num_layers x batch_size x " + "hidden_size)" + "and When is_bidirec is True, the shape will be (num_layers*2 x " + "batch_size x hidden_size)"); + AddInput("InitC", + "(Tensor) the initial cell state of the LSTm " + "input. This is a tensor with shape (num_layers x batch_size x " + "hidden_size)" + "and When is_bidirec is True, the shape will be (num_layers*2 x " + "batch_size x hidden_size)"); + AddInput("W", + "(Tensor) the learnable hidden-hidden weights." + " The shape is (N), where N is total weight size of the LSTM. " + " cudnn concatenate all the weight to one Tensor"); + AddInput("Cache", + "The cache of dropout op, a RAW type variable including random " + "number generator states and some descriptors, which is used in " + "cudnn kernel.") + .AsDispensable(); + AddOutput("Out", + "(Tensor) the hidden state of LSTM operator. " + "The shape is ( seq_len x batch_size x hidden_size) if " + "is_bidirec is False" + "and When is_bidirec is True, the shape will be ( seq_len x " + "batch_size x hidden_size * 2) "); + AddOutput("last_h", + "(Tensor) the hidden state of the last step. " + "The shape is ( num_layers x batch_size x hidden_size) if " + "is_bidirec is False" + "and When is_bidirec is True, the shape will be (num_layers*2 x " + "batch_size x hidden_size)"); + AddOutput("last_c", + "(Tensor) the cell state of the last step" + "The shape is ( num_layers x batch_size x hidden_size) if " + "is_bidirec is False" + "and When is_bidirect is True, the shape will be (num_layers*2 x " + "batch_size x hidden_size*2)"); + AddAttr("max_len", + "max length of the LSTM op" + "the first dim of the Input can NOT be greater than max_len") + .SetDefault(20); + AddAttr( + "dropout_prob", + "dropout prob of the dropout op" + "the dropout ONLY work between lstm layers, not between time steps" + "There is no dropout work on the Out tensor") + .SetDefault(0.0); + AddAttr("is_bidirec", + "is_bidirec" + "if it is bidirection rnn" + "The will affect the shape of the Out, last_h, and last_c") + .SetDefault(false); + AddAttr("input_size", "input size ot the Input Tensor").SetDefault(10); + AddAttr("hidden_size", "hidden size of the LSTM").SetDefault(100); + AddAttr("num_layers", "the total layer number of the LSTM") + .SetDefault(1); + AddAttr("is_test", "True if in test phase.").SetDefault(false); + AddAttr("seed", "seed to used if fix_seed is True").SetDefault(-1); + AddComment(R"DOC( +CUDNN LSTM implementation + +A four-gate Long Short-Term Memory network with no peephole connections. +In the forward pass the output ht and cell output ct for a given iteration can be computed from the recurrent input ht-1, +the cell input ct-1 and the previous layer input xt given matrices W, R and biases bW, bR from the following equations: + +$$ i_t = sigmoid(W_{ix}x_{t} + W_{ih}h_{t-1} + bx_i + bh_i) $$ + +$$ f_t = sigmoid(W_{fx}x_{t} + W_{fh}h_{t-1} + bx_f + bh_f) $$ + +$$ o_t = sigmoid(W_{ox}x_{t} + W_{oh}h_{t-1} + bx_o + bh_o) $$ + +$$ \\tilde{c_t} = tanh(W_{cx}x_t + W_{ch}h_{t-1} + bx_c + bh_c) $$ + +$$ c_t = f_t \\odot c_{t-1} + i_t \\odot \\tilde{c_t} $$ + +$$ h_t = o_t \\odot tanh(c_t) $$ + +- W terms denote weight matrices (e.g. $W_{ix}$ is the matrix + of weights from the input gate to the input) +- The b terms denote bias vectors ($bx_i$ and $bh_i$ are the input gate bias vector). +- sigmoid is the logistic sigmoid function. +- $i, f, o$ and $c$ are the input gate, forget gate, output gate, + and cell activation vectors, respectively, all of which have the same size as + the cell output activation vector $h$. +- The $\odot$ is the element-wise product of the vectors. +- `tanh` is the activation functions. +- $\tilde{c_t}$ is also called candidate hidden state, + which is computed based on the current input and the previous hidden state. + +Where sigmoid is the sigmoid operator: sigmoid(x) = 1 / (1 + e^-x), * represents a point-wise multiplication, +X represensts a matrix multiplication + + +)DOC"); + } +}; + +class CudnnLSTMGradOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Input"), + "Input(Input) of LSTM should not be null."); + PADDLE_ENFORCE(ctx->HasInput("W"), "Input(W) of LSTM should not be null."); + PADDLE_ENFORCE(ctx->HasInput("last_h"), + "Input(last_h) of LSTM should not be null."); + PADDLE_ENFORCE(ctx->HasInput("last_c"), + "Input(last_c) of LSTM should not be null."); + + PADDLE_ENFORCE(ctx->HasInput("Cache"), + "Input(last_c) of LSTM should not be null."); + PADDLE_ENFORCE(ctx->HasInput("InitH"), + "Input(init_h) of LSTM should not be null."); + + PADDLE_ENFORCE(ctx->HasInput("InitC"), + "Input(init_c) of LSTM should not be null."); + + auto SetOutGradDim = [&ctx](const std::string& name) { + auto g_name = framework::GradVarName(name); + if (ctx->HasOutput(g_name)) { + ctx->SetOutputDim(g_name, ctx->GetInputDim(name)); + } + }; + + SetOutGradDim("Input"); + SetOutGradDim("W"); + SetOutGradDim("InitH"); + SetOutGradDim("InitC"); + } +}; + +template +class NotImpleKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + PADDLE_THROW( + "CPU is not support for this kernel now. Will be add in the future"); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OPERATOR(cudnn_lstm, ops::CudnnLSTMOp, ops::CudnnLSTMOpMaker, + paddle::framework::DefaultGradOpDescMaker); +REGISTER_OPERATOR(cudnn_lstm_grad, ops::CudnnLSTMGradOp); + +REGISTER_OP_CPU_KERNEL(cudnn_lstm, ops::NotImpleKernel); +REGISTER_OP_CPU_KERNEL(cudnn_lstm_grad, ops::NotImpleKernel); diff --git a/paddle/fluid/operators/cudnn_lstm_op.cu.cc b/paddle/fluid/operators/cudnn_lstm_op.cu.cc new file mode 100644 index 0000000000000..f2ba75485c587 --- /dev/null +++ b/paddle/fluid/operators/cudnn_lstm_op.cu.cc @@ -0,0 +1,495 @@ +/* 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/op_registry.h" +#include "paddle/fluid/operators/math/math_function.h" +#include "paddle/fluid/platform/cudnn_helper.h" + +namespace paddle { +namespace operators { + +using LoDTensor = framework::LoDTensor; +using Tensor = framework::Tensor; + +struct CudnnRNNCache { + CudnnRNNCache() { + x_desc_ = NULL; + y_desc_ = NULL; + dx_desc_ = NULL; + dy_desc_ = NULL; + } + ~CudnnRNNCache() { release(); } + + cudnnRNNDescriptor_t rnn_desc_; + cudnnTensorDescriptor_t *x_desc_; + cudnnTensorDescriptor_t *y_desc_; + cudnnTensorDescriptor_t *dx_desc_; + cudnnTensorDescriptor_t *dy_desc_; + + cudnnTensorDescriptor_t hx_desc_; + cudnnTensorDescriptor_t cx_desc_; + cudnnTensorDescriptor_t hy_desc_; + cudnnTensorDescriptor_t cy_desc_; + + cudnnTensorDescriptor_t dhx_desc_; + cudnnTensorDescriptor_t dcx_desc_; + cudnnTensorDescriptor_t dhy_desc_; + cudnnTensorDescriptor_t dcy_desc_; + + cudnnTensorDescriptor_t output_x_desc_; + cudnnTensorDescriptor_t output_y_desc_; + + cudnnDropoutDescriptor_t dropout_desc_; + + size_t weights_size_; + cudnnFilterDescriptor_t w_desc_; + cudnnFilterDescriptor_t dw_desc_; + + size_t workspace_size_; + size_t reserve_size_; + Tensor reserve_data_; + Tensor workspace_data_; + + Tensor dropout_state_; + + size_t max_length_; + + float dropout_prob_; + bool is_bidirec_; + + int batch_size_; + int input_size_; + int hidden_size_; + int num_layers_; + int seed_; + + void init(cudnnHandle_t handle, const framework::ExecutionContext &ctx, + size_t max_len, int batch_size, int input_size, int hidden_size, + int num_layers, float dropout_prob, bool is_bidirec, int seed, + int weight_numel) { + max_length_ = max_len; + batch_size_ = batch_size; + input_size_ = input_size; + hidden_size_ = hidden_size; + num_layers_ = num_layers; + dropout_prob_ = dropout_prob; + is_bidirec_ = is_bidirec; + seed_ = seed; + + x_desc_ = new cudnnTensorDescriptor_t[max_length_]; + y_desc_ = new cudnnTensorDescriptor_t[max_length_]; + dx_desc_ = new cudnnTensorDescriptor_t[max_length_]; + dy_desc_ = new cudnnTensorDescriptor_t[max_length_]; + int dim_a[3]; + int stride_a[3]; + + for (size_t i = 0; i < max_length_; ++i) { + CUDNN_ENFORCE( + platform::dynload::cudnnCreateTensorDescriptor(&x_desc_[i])); + CUDNN_ENFORCE( + platform::dynload::cudnnCreateTensorDescriptor(&y_desc_[i])); + CUDNN_ENFORCE( + platform::dynload::cudnnCreateTensorDescriptor(&dx_desc_[i])); + CUDNN_ENFORCE( + platform::dynload::cudnnCreateTensorDescriptor(&dy_desc_[i])); + dim_a[0] = batch_size_; + dim_a[1] = input_size_; + dim_a[2] = 1; + + stride_a[0] = dim_a[2] * dim_a[1]; + stride_a[1] = dim_a[2]; + stride_a[2] = 1; + CUDNN_ENFORCE(platform::dynload::cudnnSetTensorNdDescriptor( + x_desc_[i], CUDNN_DATA_FLOAT, 3, dim_a, stride_a)); + CUDNN_ENFORCE(platform::dynload::cudnnSetTensorNdDescriptor( + dx_desc_[i], CUDNN_DATA_FLOAT, 3, dim_a, stride_a)); + + dim_a[0] = batch_size_; + dim_a[1] = is_bidirec_ ? hidden_size_ * 2 : hidden_size_; + dim_a[2] = 1; + + stride_a[0] = dim_a[2] * dim_a[1]; + stride_a[1] = dim_a[2]; + stride_a[2] = 1; + + CUDNN_ENFORCE(platform::dynload::cudnnSetTensorNdDescriptor( + y_desc_[i], CUDNN_DATA_FLOAT, 3, dim_a, stride_a)); + CUDNN_ENFORCE(platform::dynload::cudnnSetTensorNdDescriptor( + dy_desc_[i], CUDNN_DATA_FLOAT, 3, dim_a, stride_a)); + } + + dim_a[0] = num_layers_ * (is_bidirec_ ? 2 : 1); + dim_a[1] = batch_size_; + dim_a[2] = hidden_size_; + + stride_a[0] = dim_a[2] * dim_a[1]; + stride_a[1] = dim_a[2]; + stride_a[2] = 1; + + CUDNN_ENFORCE(platform::dynload::cudnnCreateTensorDescriptor(&hx_desc_)); + CUDNN_ENFORCE(platform::dynload::cudnnCreateTensorDescriptor(&cx_desc_)); + CUDNN_ENFORCE(platform::dynload::cudnnCreateTensorDescriptor(&hy_desc_)); + CUDNN_ENFORCE(platform::dynload::cudnnCreateTensorDescriptor(&cy_desc_)); + CUDNN_ENFORCE(platform::dynload::cudnnCreateTensorDescriptor(&dhx_desc_)); + CUDNN_ENFORCE(platform::dynload::cudnnCreateTensorDescriptor(&dcx_desc_)); + CUDNN_ENFORCE(platform::dynload::cudnnCreateTensorDescriptor(&dhy_desc_)); + CUDNN_ENFORCE(platform::dynload::cudnnCreateTensorDescriptor(&dcy_desc_)); + + CUDNN_ENFORCE(platform::dynload::cudnnSetTensorNdDescriptor( + hx_desc_, CUDNN_DATA_FLOAT, 3, dim_a, stride_a)); + CUDNN_ENFORCE(platform::dynload::cudnnSetTensorNdDescriptor( + cx_desc_, CUDNN_DATA_FLOAT, 3, dim_a, stride_a)); + CUDNN_ENFORCE(platform::dynload::cudnnSetTensorNdDescriptor( + hy_desc_, CUDNN_DATA_FLOAT, 3, dim_a, stride_a)); + CUDNN_ENFORCE(platform::dynload::cudnnSetTensorNdDescriptor( + cy_desc_, CUDNN_DATA_FLOAT, 3, dim_a, stride_a)); + CUDNN_ENFORCE(platform::dynload::cudnnSetTensorNdDescriptor( + dhx_desc_, CUDNN_DATA_FLOAT, 3, dim_a, stride_a)); + CUDNN_ENFORCE(platform::dynload::cudnnSetTensorNdDescriptor( + dcx_desc_, CUDNN_DATA_FLOAT, 3, dim_a, stride_a)); + CUDNN_ENFORCE(platform::dynload::cudnnSetTensorNdDescriptor( + dhy_desc_, CUDNN_DATA_FLOAT, 3, dim_a, stride_a)); + CUDNN_ENFORCE(platform::dynload::cudnnSetTensorNdDescriptor( + dcy_desc_, CUDNN_DATA_FLOAT, 3, dim_a, stride_a)); + + CUDNN_ENFORCE( + platform::dynload::cudnnCreateDropoutDescriptor(&dropout_desc_)); + + size_t state_size; + CUDNN_ENFORCE( + platform::dynload::cudnnDropoutGetStatesSize(handle, &state_size); + dropout_state_.Resize({static_cast(state_size)})); + auto *dropout_state_data = + dropout_state_.mutable_data(ctx.GetPlace()); + CUDNN_ENFORCE(platform::dynload::cudnnSetDropoutDescriptor( + dropout_desc_, handle, dropout_prob_, dropout_state_data, state_size, + seed_)); + + CUDNN_ENFORCE(platform::dynload::cudnnCreateRNNDescriptor(&rnn_desc_)); + +#if CUDNN_VERSION >= 6000 + CUDNN_ENFORCE(platform::dynload::cudnnSetRNNDescriptor_v6( + handle, rnn_desc_, hidden_size_, num_layers_, dropout_desc_, + CUDNN_LINEAR_INPUT, + is_bidirec_ ? CUDNN_BIDIRECTIONAL : CUDNN_UNIDIRECTIONAL, CUDNN_LSTM, + CUDNN_RNN_ALGO_STANDARD, CUDNN_DATA_FLOAT)); +#else + CUDNN_ENFORCE(platform::dynload::cudnnSetRNNDescriptor( + rnn_desc_, hidden_size_, num_layers_, dropout_desc_, CUDNN_LINEAR_INPUT, + is_bidirec_ ? CUDNN_BIDIRECTIONAL : CUDNN_UNIDIRECTIONAL, CUDNN_LSTM, + CUDNN_DATA_FLOAT)); +#endif + + CUDNN_ENFORCE(platform::dynload::cudnnCreateFilterDescriptor(&w_desc_)); + CUDNN_ENFORCE(platform::dynload::cudnnCreateFilterDescriptor(&dw_desc_)); + + CUDNN_ENFORCE(platform::dynload::cudnnGetRNNParamsSize( + handle, rnn_desc_, x_desc_[0], &weights_size_, CUDNN_DATA_FLOAT)); + + PADDLE_ENFORCE_EQ(weights_size_, sizeof(float) * weight_numel, + "cudnn lstm weight size should be SAME"); + int dim_w[3]; + dim_w[0] = weights_size_ / sizeof(float); + dim_w[1] = 1; + dim_w[2] = 1; + CUDNN_ENFORCE(platform::dynload::cudnnSetFilterNdDescriptor( + w_desc_, CUDNN_DATA_FLOAT, CUDNN_TENSOR_NCHW, 3, dim_w)); + CUDNN_ENFORCE(platform::dynload::cudnnSetFilterNdDescriptor( + dw_desc_, CUDNN_DATA_FLOAT, CUDNN_TENSOR_NCHW, 3, dim_w)); + + CUDNN_ENFORCE(platform::dynload::cudnnGetRNNWorkspaceSize( + handle, rnn_desc_, max_length_, x_desc_, &workspace_size_)); + CUDNN_ENFORCE(platform::dynload::cudnnGetRNNTrainingReserveSize( + handle, rnn_desc_, max_length_, x_desc_, &reserve_size_)); + + reserve_data_.Resize({static_cast(reserve_size_)}); + reserve_data_.mutable_data(ctx.GetPlace()); + + workspace_data_.Resize({static_cast(workspace_size_)}); + workspace_data_.mutable_data(ctx.GetPlace()); + } + + void release() { + for (size_t i = 0; i < max_length_; ++i) { + CUDNN_ENFORCE( + platform::dynload::cudnnDestroyTensorDescriptor(x_desc_[i])); + CUDNN_ENFORCE( + platform::dynload::cudnnDestroyTensorDescriptor(y_desc_[i])); + CUDNN_ENFORCE( + platform::dynload::cudnnDestroyTensorDescriptor(dx_desc_[i])); + CUDNN_ENFORCE( + platform::dynload::cudnnDestroyTensorDescriptor(dy_desc_[i])); + } + + delete[] x_desc_; + delete[] y_desc_; + delete[] dx_desc_; + delete[] dy_desc_; + + CUDNN_ENFORCE(platform::dynload::cudnnDestroyTensorDescriptor(hx_desc_)); + CUDNN_ENFORCE(platform::dynload::cudnnDestroyTensorDescriptor(cx_desc_)); + CUDNN_ENFORCE(platform::dynload::cudnnDestroyTensorDescriptor(hy_desc_)); + CUDNN_ENFORCE(platform::dynload::cudnnDestroyTensorDescriptor(cy_desc_)); + CUDNN_ENFORCE(platform::dynload::cudnnDestroyTensorDescriptor(dhx_desc_)); + CUDNN_ENFORCE(platform::dynload::cudnnDestroyTensorDescriptor(dcx_desc_)); + CUDNN_ENFORCE(platform::dynload::cudnnDestroyTensorDescriptor(dhy_desc_)); + CUDNN_ENFORCE(platform::dynload::cudnnDestroyTensorDescriptor(dcy_desc_)); + + CUDNN_ENFORCE( + platform::dynload::cudnnDestroyDropoutDescriptor(dropout_desc_)); + CUDNN_ENFORCE(platform::dynload::cudnnDestroyRNNDescriptor(rnn_desc_)); + + CUDNN_ENFORCE(platform::dynload::cudnnDestroyFilterDescriptor(w_desc_)); + CUDNN_ENFORCE(platform::dynload::cudnnDestroyFilterDescriptor(dw_desc_)); + } +}; + +template +class CudnnLSTMGPUKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &ctx) const override { + const Tensor *x = ctx.Input("Input"); + const Tensor *init_h = ctx.Input("InitH"); + const Tensor *init_c = ctx.Input("InitC"); + + auto w = ctx.Input("W"); + + Tensor *out = ctx.Output("Out"); + Tensor *last_h = ctx.Output("last_h"); + Tensor *last_c = ctx.Output("last_c"); + + const T *x_data = x->data(); + const T *init_h_data = init_h->data(); + const T *init_c_data = init_c->data(); + + const T *w_data = w->data(); + + T *out_data = out->mutable_data(ctx.GetPlace()); + T *last_h_data = last_h->mutable_data(ctx.GetPlace()); + T *last_c_data = last_c->mutable_data(ctx.GetPlace()); + + size_t max_len = ctx.Attr("max_len"); + float dropout_prob = ctx.Attr("dropout_prob"); + bool is_bidirec = ctx.Attr("is_bidirec"); + int input_size = ctx.Attr("input_size"); + int hidden_size = ctx.Attr("hidden_size"); + int num_layers = ctx.Attr("num_layers"); + bool is_test = ctx.Attr("is_test"); + + auto &dev_ctx = ctx.template device_context(); + auto handle = dev_ctx.cudnn_handle(); + auto *cache_var = ctx.InputVar("Cache"); + if (!cache_var) { + // The RAW type cache variable wouldn't be created and broadcasted on + // multi-devices before the first running. + // use parent scope to make cache persistable + auto *scope = const_cast(ctx.scope().parent()); + auto cache_var_name = ctx.Inputs("Cache")[0]; + cache_var = scope->Var(cache_var_name); + } + CudnnRNNCache *cudnn_rnn_cache = nullptr; + if (cache_var->IsInitialized()) { + // const_cast is usually bad. + cudnn_rnn_cache = const_cast(cache_var) + ->GetMutable(); + } else { + // const_cast is usually bad. + cudnn_rnn_cache = const_cast(cache_var) + ->GetMutable(); + std::random_device rnd; + int seed = ctx.Attr("seed"); + if (seed == -1) { + seed = rnd(); + } + + auto input_w_numel = w->numel(); + auto batch_size = x->dims()[1]; + cudnn_rnn_cache->init(handle, ctx, max_len, batch_size, input_size, + hidden_size, num_layers, dropout_prob, is_bidirec, + seed, input_w_numel); + } + + auto run_seq_len = x->dims()[0]; + + if (is_test) { + // for inference + CUDNN_ENFORCE(platform::dynload::cudnnRNNForwardInference( + handle, cudnn_rnn_cache->rnn_desc_, run_seq_len, + cudnn_rnn_cache->x_desc_, x_data, cudnn_rnn_cache->hx_desc_, + init_h_data, cudnn_rnn_cache->cx_desc_, init_c_data, + cudnn_rnn_cache->w_desc_, w_data, cudnn_rnn_cache->y_desc_, out_data, + cudnn_rnn_cache->hy_desc_, last_h_data, cudnn_rnn_cache->cy_desc_, + last_c_data, cudnn_rnn_cache->workspace_data_.data(), + cudnn_rnn_cache->workspace_size_)); + } else { + // for train + CUDNN_ENFORCE(platform::dynload::cudnnRNNForwardTraining( + handle, cudnn_rnn_cache->rnn_desc_, run_seq_len, + cudnn_rnn_cache->x_desc_, x_data, cudnn_rnn_cache->hx_desc_, + init_h_data, cudnn_rnn_cache->cx_desc_, init_c_data, + cudnn_rnn_cache->w_desc_, w_data, cudnn_rnn_cache->y_desc_, out_data, + cudnn_rnn_cache->hy_desc_, last_h_data, cudnn_rnn_cache->cy_desc_, + last_c_data, cudnn_rnn_cache->workspace_data_.data(), + cudnn_rnn_cache->workspace_size_, + cudnn_rnn_cache->reserve_data_.data(), + cudnn_rnn_cache->reserve_size_)); + } + } +}; + +template +class CudnnLSTMGPUGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &ctx) const override { + auto *input = ctx.Input("Input"); + auto *weight = ctx.Input("W"); + auto *init_h = ctx.Input("InitH"); + auto *init_c = ctx.Input("InitC"); + // auto * last_h = ctx.Input("last_h"); + // auto * last_c = ctx.Input("last_c"); + auto *out = ctx.Input("Out"); + auto *out_grad = ctx.Input(framework::GradVarName("Out")); + auto *last_h_grad = ctx.Input(framework::GradVarName("last_h")); + auto *last_c_grad = ctx.Input(framework::GradVarName("last_c")); + + // auto* init_h = ctx.Input("init_h"); + // auto* init_c = ctx.Input("init_c"); + + auto *in_grad = ctx.Output(framework::GradVarName("Input")); + auto *weight_grad = ctx.Output(framework::GradVarName("W")); + auto *init_h_grad = ctx.Output(framework::GradVarName("InitH")); + auto *init_c_grad = ctx.Output(framework::GradVarName("InitC")); + + auto &dev_ctx = ctx.template device_context(); + auto handle = dev_ctx.cudnn_handle(); + auto *cache_var = ctx.InputVar("Cache"); + PADDLE_ENFORCE(cache_var->IsInitialized()); + CudnnRNNCache *cudnn_rnn_cache = + const_cast(cache_var) + ->GetMutable(); + + auto input_dims = input->dims(); + auto weight_dims = weight->dims(); + auto init_h_dims = init_h->dims(); + auto init_c_dims = init_c->dims(); + in_grad->mutable_data(ctx.GetPlace()); + weight_grad->mutable_data(ctx.GetPlace()); + math::SetConstant zero; + zero(dev_ctx, in_grad, static_cast(0.0)); + zero(dev_ctx, weight_grad, static_cast(0.0)); + + T *init_h_grad_data = NULL; + if (init_h_grad == nullptr) { + Tensor init_h_grad_temp; + init_h_grad_temp.mutable_data(init_h_dims, ctx.GetPlace()); + zero(dev_ctx, &init_h_grad_temp, static_cast(0.0)); + + init_h_grad_data = init_h_grad_temp.data(); + } else { + init_h_grad->mutable_data(init_h_dims, ctx.GetPlace()); + zero(dev_ctx, init_h_grad, static_cast(0.0)); + init_h_grad_data = init_h_grad->data(); + } + + T *init_c_grad_data = NULL; + if (init_c_grad == nullptr) { + Tensor init_c_grad_temp; + init_c_grad_temp.mutable_data(init_c_dims, ctx.GetPlace()); + zero(dev_ctx, &init_c_grad_temp, static_cast(0.0)); + + init_c_grad_data = init_c_grad_temp.data(); + } else { + init_c_grad->mutable_data(init_c_dims, ctx.GetPlace()); + zero(dev_ctx, init_c_grad, static_cast(0.0)); + init_c_grad_data = init_c_grad->data(); + } + + const T *last_h_grad_data = NULL; + if (last_h_grad == nullptr) { + Tensor last_h_grad_temp; + last_h_grad_temp.mutable_data(init_h_dims, ctx.GetPlace()); + zero(dev_ctx, &last_h_grad_temp, static_cast(0.0)); + + last_h_grad_data = (const T *)last_h_grad_temp.data(); + } else { + last_h_grad_data = last_h_grad->data(); + } + + const T *last_c_grad_data = NULL; + if (last_c_grad == nullptr) { + Tensor last_c_grad_temp; + last_c_grad_temp.mutable_data(init_c_dims, ctx.GetPlace()); + zero(dev_ctx, &last_c_grad_temp, static_cast(0.0)); + + last_c_grad_data = (const T *)last_c_grad_temp.data(); + } else { + last_c_grad_data = last_c_grad->data(); + } + + const T *out_grad_data = NULL; + if (out_grad == nullptr) { + Tensor out_grad_temp; + out_grad_temp.mutable_data(out->dims(), ctx.GetPlace()); + zero(dev_ctx, &out_grad_temp, static_cast(0.0)); + + out_grad_data = (const T *)out_grad_temp.data(); + } else { + out_grad_data = out_grad->data(); + } + + // zero( dev_ctx, last_h_grad, static_cast(0.0)); + // zero( dev_ctx, last_c_grad, static_cast(0.0)); + + auto out_data = out->data(); + // auto out_grad_data = out_grad->data(); + auto weight_data = weight->data(); + auto init_h_data = init_h->data(); + auto init_c_data = init_c->data(); + auto in_grad_data = in_grad->data(); + + auto work_data = cudnn_rnn_cache->workspace_data_.data(); + auto reserve_data = cudnn_rnn_cache->reserve_data_.data(); + + auto run_seq_len = input_dims[0]; + PADDLE_ENFORCE_LE((size_t)run_seq_len, cudnn_rnn_cache->max_length_, + "cudnn running seq_len CAN not greater max_lengh"); + CUDNN_ENFORCE(platform::dynload::cudnnRNNBackwardData( + handle, cudnn_rnn_cache->rnn_desc_, run_seq_len, + cudnn_rnn_cache->y_desc_, out_data, cudnn_rnn_cache->dy_desc_, + out_grad_data, cudnn_rnn_cache->dhy_desc_, last_h_grad_data, + cudnn_rnn_cache->dcy_desc_, last_c_grad_data, cudnn_rnn_cache->w_desc_, + weight_data, cudnn_rnn_cache->hx_desc_, init_h_data, + cudnn_rnn_cache->cx_desc_, init_c_data, cudnn_rnn_cache->dx_desc_, + in_grad_data, cudnn_rnn_cache->dhx_desc_, init_h_grad_data, + cudnn_rnn_cache->dcx_desc_, init_c_grad_data, work_data, + cudnn_rnn_cache->workspace_size_, reserve_data, + cudnn_rnn_cache->reserve_size_)); + + CUDNN_ENFORCE(platform::dynload::cudnnRNNBackwardWeights( + handle, cudnn_rnn_cache->rnn_desc_, run_seq_len, + cudnn_rnn_cache->x_desc_, input->data(), cudnn_rnn_cache->hx_desc_, + init_h->data(), cudnn_rnn_cache->y_desc_, out->data(), + cudnn_rnn_cache->workspace_data_.data(), + cudnn_rnn_cache->workspace_size_, cudnn_rnn_cache->dw_desc_, + weight_grad->data(), cudnn_rnn_cache->reserve_data_.data(), + cudnn_rnn_cache->reserve_size_)); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP_CUDA_KERNEL(cudnn_lstm, ops::CudnnLSTMGPUKernel); +REGISTER_OP_CUDA_KERNEL(cudnn_lstm_grad, ops::CudnnLSTMGPUGradKernel); diff --git a/paddle/fluid/operators/dequantize_mkldnn_op.cc b/paddle/fluid/operators/dequantize_mkldnn_op.cc new file mode 100644 index 0000000000000..262b7408a7f5f --- /dev/null +++ b/paddle/fluid/operators/dequantize_mkldnn_op.cc @@ -0,0 +1,88 @@ +/* Copyright (c) 2016 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 "mkldnn.hpp" +#include "paddle/fluid/framework/data_layout_transform.h" +#include "paddle/fluid/framework/tensor.h" +#include "paddle/fluid/operators/dequantize_op.h" +#include "paddle/fluid/platform/mkldnn_helper.h" + +namespace paddle { +namespace operators { + +using mkldnn::memory; +using mkldnn::primitive; +using mkldnn::reorder; +using platform::to_void_cast; +using Tensor = framework::Tensor; +using framework::DataLayout; +using mkldnn::stream; +using platform::GetMKLDNNFormat; + +template +class DeQuantOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* input = ctx.Input("Input"); + auto scale_data = ctx.Attr("Scale"); + auto* output = ctx.Output("Output"); + auto& dev_ctx = + ctx.template device_context(); + const auto& engine = dev_ctx.GetEngine(); + + const T* input_data = input->data(); + float* output_data = output->mutable_data(ctx.GetPlace()); + std::vector reorder_scale = {1.0f / scale_data}; + + std::vector pipeline; + std::vector src_tz = paddle::framework::vectorize2int(input->dims()); + std::vector dst_tz = paddle::framework::vectorize2int(output->dims()); + mkldnn::memory::data_type src_dt = + paddle::framework::ToMKLDNNDataType(input->type()); + mkldnn::memory::format src_fmt = input->format(); + + mkldnn::primitive_attr attri; + int mask = 0; + attri.set_output_scales(mask, reorder_scale); + + auto src_md = platform::MKLDNNMemDesc({src_tz}, src_dt, src_fmt); + auto src_pd = mkldnn::memory::primitive_desc(src_md, engine); + auto src_memory = + std::make_shared(src_pd, to_void_cast(input_data)); + std::shared_ptr src_memory_p = + std::shared_ptr(new primitive::at(*src_memory)); + + auto dst_md = platform::MKLDNNMemDesc({dst_tz}, memory::data_type::f32, + memory::format::nchw); + auto dst_pd = mkldnn::memory::primitive_desc(dst_md, engine); + auto dst_memory = mkldnn::memory(dst_pd, to_void_cast(output_data)); + + auto reorder_pd = std::shared_ptr( + new reorder::primitive_desc(src_pd, dst_pd, attri)); + auto reorder_p = std::shared_ptr( + new reorder(*reorder_pd, *src_memory_p, dst_memory)); + pipeline.push_back(*reorder_p); + stream(stream::kind::eager).submit(pipeline).wait(); + + output->set_format(GetMKLDNNFormat(dst_memory)); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; + +REGISTER_OP_KERNEL(dequantize, MKLDNN, ::paddle::platform::CPUPlace, + ops::DeQuantOpKernel, ops::DeQuantOpKernel); diff --git a/paddle/fluid/operators/dequantize_op.cc b/paddle/fluid/operators/dequantize_op.cc new file mode 100644 index 0000000000000..38159f84a0d56 --- /dev/null +++ b/paddle/fluid/operators/dequantize_op.cc @@ -0,0 +1,45 @@ +/* Copyright (c) 2016 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/operators/dequantize_op.h" +#ifdef PADDLE_WITH_MKLDNN +#include "paddle/fluid/platform/mkldnn_helper.h" +#endif + +namespace paddle { +namespace operators { + +framework::OpKernelType DeQuantOp::GetExpectedKernelType( + const framework::ExecutionContext& ctx) const { + framework::LibraryType library_ = framework::LibraryType::kMKLDNN; + framework::DataLayout layout_ = framework::DataLayout::kMKLDNN; + + return framework::OpKernelType(ctx.Input("Input")->type(), + ctx.GetPlace(), layout_, library_); +} + +void DeQuantOpMaker::Make() { + AddInput("Input", "input data"); + AddOutput("Output", "output data"); + AddAttr("Scale", "scale data").SetDefault({1.0f}); + AddComment(R"DOC(This op will dequantize data from INT8 to FP32)DOC"); +} + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; + +REGISTER_OPERATOR(dequantize, ops::DeQuantOp, ops::DeQuantOpMaker, + paddle::framework::DefaultGradOpDescMaker); diff --git a/paddle/fluid/operators/dequantize_op.h b/paddle/fluid/operators/dequantize_op.h new file mode 100644 index 0000000000000..75c27a06c210f --- /dev/null +++ b/paddle/fluid/operators/dequantize_op.h @@ -0,0 +1,54 @@ +/* Copyright (c) 2016 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 +#include +#include "paddle/fluid/framework/op_registry.h" + +namespace paddle { +namespace operators { + +using framework::OpKernelType; +using framework::Tensor; + +class DeQuantOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override { + ctx->SetOutputDim("Output", ctx->GetInputDim("Input")); + ctx->ShareLoD("Input", /*->*/ "Output"); + } + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override; +}; + +class DeQuantOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override; +}; + +class DeQuantGradOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override {} +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/detection/CMakeLists.txt b/paddle/fluid/operators/detection/CMakeLists.txt index 58f6f48467310..6c85f1577e0c4 100644 --- a/paddle/fluid/operators/detection/CMakeLists.txt +++ b/paddle/fluid/operators/detection/CMakeLists.txt @@ -22,7 +22,7 @@ iou_similarity_op.cu) detection_library(mine_hard_examples_op SRCS mine_hard_examples_op.cc) detection_library(multiclass_nms_op SRCS multiclass_nms_op.cc poly_util.cc gpc.cc) detection_library(prior_box_op SRCS prior_box_op.cc prior_box_op.cu) -detection_library(density_prior_box_op SRCS density_prior_box_op.cc) +detection_library(density_prior_box_op SRCS density_prior_box_op.cc density_prior_box_op.cu) detection_library(anchor_generator_op SRCS anchor_generator_op.cc anchor_generator_op.cu) detection_library(target_assign_op SRCS target_assign_op.cc diff --git a/paddle/fluid/operators/detection/anchor_generator_op.cc b/paddle/fluid/operators/detection/anchor_generator_op.cc index 0c0155a0a9778..f2984d1af2f26 100644 --- a/paddle/fluid/operators/detection/anchor_generator_op.cc +++ b/paddle/fluid/operators/detection/anchor_generator_op.cc @@ -53,8 +53,7 @@ class AnchorGeneratorOp : public framework::OperatorWithKernel { framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext& ctx) const override { return framework::OpKernelType( - framework::ToDataType(ctx.Input("Input")->type()), - ctx.device_context()); + ctx.Input("Input")->type(), ctx.device_context()); } }; diff --git a/paddle/fluid/operators/detection/bipartite_match_op.cc b/paddle/fluid/operators/detection/bipartite_match_op.cc index c23b65fe4dead..b7da1261a8f97 100644 --- a/paddle/fluid/operators/detection/bipartite_match_op.cc +++ b/paddle/fluid/operators/detection/bipartite_match_op.cc @@ -45,9 +45,8 @@ class BipartiteMatchOp : public framework::OperatorWithKernel { protected: framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext& ctx) const override { - return framework::OpKernelType( - framework::ToDataType(ctx.Input("DistMat")->type()), - platform::CPUPlace()); + return framework::OpKernelType(ctx.Input("DistMat")->type(), + platform::CPUPlace()); } }; diff --git a/paddle/fluid/operators/detection/box_coder_op.h b/paddle/fluid/operators/detection/box_coder_op.h index 5ed8520acddfa..b2a2bcdce9320 100644 --- a/paddle/fluid/operators/detection/box_coder_op.h +++ b/paddle/fluid/operators/detection/box_coder_op.h @@ -43,6 +43,9 @@ class BoxCoderKernel : public framework::OpKernel { const T* prior_box_var_data = nullptr; if (prior_box_var) prior_box_var_data = prior_box_var->data(); +#ifdef PADDLE_WITH_MKLML +#pragma omp parallel for collapse(2) +#endif for (int64_t i = 0; i < row; ++i) { for (int64_t j = 0; j < col; ++j) { T prior_box_width = prior_box_data[j * len + 2] - @@ -96,6 +99,9 @@ class BoxCoderKernel : public framework::OpKernel { const T* prior_box_var_data = nullptr; if (prior_box_var) prior_box_var_data = prior_box_var->data(); +#ifdef PADDLE_WITH_MKLML +#pragma omp parallel for collapse(2) +#endif for (int64_t i = 0; i < row; ++i) { for (int64_t j = 0; j < col; ++j) { size_t offset = i * col * len + j * len; diff --git a/paddle/fluid/operators/detection/density_prior_box_op.cc b/paddle/fluid/operators/detection/density_prior_box_op.cc index 99df15c3226b4..cacd47ed4a804 100644 --- a/paddle/fluid/operators/detection/density_prior_box_op.cc +++ b/paddle/fluid/operators/detection/density_prior_box_op.cc @@ -39,32 +39,34 @@ class DensityPriorBoxOp : public framework::OperatorWithKernel { auto fixed_sizes = ctx->Attrs().Get>("fixed_sizes"); auto fixed_ratios = ctx->Attrs().Get>("fixed_ratios"); auto densities = ctx->Attrs().Get>("densities"); + bool flatten = ctx->Attrs().Get("flatten_to_2d"); PADDLE_ENFORCE_EQ(fixed_sizes.size(), densities.size(), "The number of fixed_sizes and densities must be equal."); size_t num_priors = 0; - if ((fixed_sizes.size() > 0) && (densities.size() > 0)) { - for (size_t i = 0; i < densities.size(); ++i) { - if (fixed_ratios.size() > 0) { - num_priors += (fixed_ratios.size()) * (pow(densities[i], 2)); - } - } + for (size_t i = 0; i < densities.size(); ++i) { + num_priors += (fixed_ratios.size()) * (pow(densities[i], 2)); + } + if (!flatten) { + std::vector dim_vec(4); + dim_vec[0] = input_dims[2]; + dim_vec[1] = input_dims[3]; + dim_vec[2] = num_priors; + dim_vec[3] = 4; + ctx->SetOutputDim("Boxes", framework::make_ddim(dim_vec)); + ctx->SetOutputDim("Variances", framework::make_ddim(dim_vec)); + } else { + int64_t dim0 = input_dims[2] * input_dims[3] * num_priors; + ctx->SetOutputDim("Boxes", {dim0, 4}); + ctx->SetOutputDim("Variances", {dim0, 4}); } - std::vector dim_vec(4); - dim_vec[0] = input_dims[2]; - dim_vec[1] = input_dims[3]; - dim_vec[2] = num_priors; - dim_vec[3] = 4; - ctx->SetOutputDim("Boxes", framework::make_ddim(dim_vec)); - ctx->SetOutputDim("Variances", framework::make_ddim(dim_vec)); } protected: framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext& ctx) const override { return framework::OpKernelType( - framework::ToDataType(ctx.Input("Input")->type()), - platform::CPUPlace()); + ctx.Input("Input")->type(), ctx.GetPlace()); } }; @@ -101,7 +103,10 @@ class DensityPriorBoxOpMaker : public framework::OpProtoAndCheckerMaker { }); AddAttr("clip", "(bool) Whether to clip out-of-boundary boxes.") .SetDefault(true); - + AddAttr("flatten_to_2d", + "(bool) Whether to flatten to 2D and " + "the second dim is 4.") + .SetDefault(false); AddAttr( "step_w", "Density prior boxes step across width, 0.0 for auto calculation.") diff --git a/paddle/fluid/operators/detection/density_prior_box_op.cu b/paddle/fluid/operators/detection/density_prior_box_op.cu new file mode 100644 index 0000000000000..acd5993154ed0 --- /dev/null +++ b/paddle/fluid/operators/detection/density_prior_box_op.cu @@ -0,0 +1,172 @@ +/* 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/operators/detection/density_prior_box_op.h" + +namespace paddle { +namespace operators { + +template +static __device__ inline T Clip(T in) { + return min(max(in, 0.), 1.); +} + +template +static __global__ void GenDensityPriorBox( + const int height, const int width, const int im_height, const int im_width, + const T offset, const T step_width, const T step_height, + const int num_priors, const T* ratios_shift, bool is_clip, const T var_xmin, + const T var_ymin, const T var_xmax, const T var_ymax, T* out, T* var) { + int gidx = blockIdx.x * blockDim.x + threadIdx.x; + int gidy = blockIdx.y * blockDim.y + threadIdx.y; + int step_x = blockDim.x * gridDim.x; + int step_y = blockDim.y * gridDim.y; + + const T* width_ratio = ratios_shift; + const T* height_ratio = ratios_shift + num_priors; + const T* width_shift = ratios_shift + 2 * num_priors; + const T* height_shift = ratios_shift + 3 * num_priors; + + for (int j = gidy; j < height; j += step_y) { + for (int i = gidx; i < width * num_priors; i += step_x) { + int h = j; + int w = i / num_priors; + int k = i % num_priors; + + T center_x = (w + offset) * step_width; + T center_y = (h + offset) * step_height; + + T center_x_temp = center_x + width_shift[k]; + T center_y_temp = center_y + height_shift[k]; + + T box_width_ratio = width_ratio[k] / 2.; + T box_height_ratio = height_ratio[k] / 2.; + + T xmin = max((center_x_temp - box_width_ratio) / im_width, 0.); + T ymin = max((center_y_temp - box_height_ratio) / im_height, 0.); + T xmax = min((center_x_temp + box_width_ratio) / im_width, 1.); + T ymax = min((center_y_temp + box_height_ratio) / im_height, 1.); + + int out_offset = (j * width * num_priors + i) * 4; + out[out_offset] = is_clip ? Clip(xmin) : xmin; + out[out_offset + 1] = is_clip ? Clip(ymin) : ymin; + out[out_offset + 2] = is_clip ? Clip(xmax) : xmax; + out[out_offset + 3] = is_clip ? Clip(ymax) : ymax; + + var[out_offset] = var_xmin; + var[out_offset + 1] = var_ymin; + var[out_offset + 2] = var_xmax; + var[out_offset + 3] = var_ymax; + } + } +} + +template +class DensityPriorBoxOpCUDAKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* input = ctx.Input("Input"); + auto* image = ctx.Input("Image"); + auto* boxes = ctx.Output("Boxes"); + auto* vars = ctx.Output("Variances"); + + auto variances = ctx.Attr>("variances"); + auto is_clip = ctx.Attr("clip"); + + auto fixed_sizes = ctx.Attr>("fixed_sizes"); + auto fixed_ratios = ctx.Attr>("fixed_ratios"); + auto densities = ctx.Attr>("densities"); + + T step_w = static_cast(ctx.Attr("step_w")); + T step_h = static_cast(ctx.Attr("step_h")); + T offset = static_cast(ctx.Attr("offset")); + + auto img_width = image->dims()[3]; + auto img_height = image->dims()[2]; + + auto feature_width = input->dims()[3]; + auto feature_height = input->dims()[2]; + + T step_width, step_height; + if (step_w == 0 || step_h == 0) { + step_width = static_cast(img_width) / feature_width; + step_height = static_cast(img_height) / feature_height; + } else { + step_width = step_w; + step_height = step_h; + } + + int num_priors = 0; + for (size_t i = 0; i < densities.size(); ++i) { + num_priors += (fixed_ratios.size()) * (pow(densities[i], 2)); + } + int step_average = static_cast((step_width + step_height) * 0.5); + + framework::Tensor h_temp; + T* tdata = h_temp.mutable_data({num_priors * 4}, platform::CPUPlace()); + int idx = 0; + for (size_t s = 0; s < fixed_sizes.size(); ++s) { + auto fixed_size = fixed_sizes[s]; + int density = densities[s]; + for (size_t r = 0; r < fixed_ratios.size(); ++r) { + float ar = fixed_ratios[r]; + int shift = step_average / density; + float box_width_ratio = fixed_size * sqrt(ar); + float box_height_ratio = fixed_size / sqrt(ar); + for (int di = 0; di < density; ++di) { + for (int dj = 0; dj < density; ++dj) { + float center_x_temp = shift / 2. + dj * shift - step_average / 2.; + float center_y_temp = shift / 2. + di * shift - step_average / 2.; + tdata[idx] = box_width_ratio; + tdata[num_priors + idx] = box_height_ratio; + tdata[2 * num_priors + idx] = center_x_temp; + tdata[3 * num_priors + idx] = center_y_temp; + idx++; + } + } + } + } + + boxes->mutable_data(ctx.GetPlace()); + vars->mutable_data(ctx.GetPlace()); + + framework::Tensor d_temp; + framework::TensorCopy(h_temp, ctx.GetPlace(), &d_temp); + + // At least use 32 threads, at most 512 threads. + // blockx is multiple of 32. + int blockx = std::min( + static_cast(((feature_width * num_priors + 31) >> 5) << 5), + 512L); + int gridx = (feature_width * num_priors + blockx - 1) / blockx; + dim3 threads(blockx, 1); + dim3 grids(gridx, feature_height); + + auto stream = + ctx.template device_context().stream(); + GenDensityPriorBox<<>>( + feature_height, feature_width, img_height, img_width, offset, + step_width, step_height, num_priors, d_temp.data(), is_clip, + variances[0], variances[1], variances[2], variances[3], + boxes->data(), vars->data()); + } +}; // namespace operators + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP_CUDA_KERNEL(density_prior_box, + ops::DensityPriorBoxOpCUDAKernel, + ops::DensityPriorBoxOpCUDAKernel); diff --git a/paddle/fluid/operators/detection/density_prior_box_op.h b/paddle/fluid/operators/detection/density_prior_box_op.h index 9a52077e9cf90..ed2f5df80cf4d 100644 --- a/paddle/fluid/operators/detection/density_prior_box_op.h +++ b/paddle/fluid/operators/detection/density_prior_box_op.h @@ -1,4 +1,4 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +/* 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 @@ -52,18 +52,16 @@ class DensityPriorBoxOpKernel : public framework::OpKernel { step_height = step_h; } int num_priors = 0; - if (fixed_sizes.size() > 0 && densities.size() > 0) { - for (size_t i = 0; i < densities.size(); ++i) { - if (fixed_ratios.size() > 0) { - num_priors += (fixed_ratios.size()) * (pow(densities[i], 2)); - } - } + for (size_t i = 0; i < densities.size(); ++i) { + num_priors += (fixed_ratios.size()) * (pow(densities[i], 2)); } boxes->mutable_data(ctx.GetPlace()); vars->mutable_data(ctx.GetPlace()); - auto e_boxes = framework::EigenTensor::From(*boxes).setConstant(0.0); + auto box_dim = vars->dims(); + boxes->Resize({feature_height, feature_width, num_priors, 4}); + auto e_boxes = framework::EigenTensor::From(*boxes).setConstant(0.0); int step_average = static_cast((step_width + step_height) * 0.5); for (int h = 0; h < feature_height; ++h) { @@ -76,36 +74,34 @@ class DensityPriorBoxOpKernel : public framework::OpKernel { auto fixed_size = fixed_sizes[s]; int density = densities[s]; // Generate density prior boxes with fixed ratios. - if (fixed_ratios.size() > 0) { - for (size_t r = 0; r < fixed_ratios.size(); ++r) { - float ar = fixed_ratios[r]; - int shift = step_average / density; - float box_width_ratio = fixed_size * sqrt(ar); - float box_height_ratio = fixed_size / sqrt(ar); - for (int di = 0; di < density; ++di) { - for (int dj = 0; dj < density; ++dj) { - float center_x_temp = - center_x - step_average / 2. + shift / 2. + dj * shift; - float center_y_temp = - center_y - step_average / 2. + shift / 2. + di * shift; - e_boxes(h, w, idx, 0) = - (center_x_temp - box_width_ratio / 2.) / img_width >= 0 - ? (center_x_temp - box_width_ratio / 2.) / img_width - : 0; - e_boxes(h, w, idx, 1) = - (center_y_temp - box_height_ratio / 2.) / img_height >= 0 - ? (center_y_temp - box_height_ratio / 2.) / img_height - : 0; - e_boxes(h, w, idx, 2) = - (center_x_temp + box_width_ratio / 2.) / img_width <= 1 - ? (center_x_temp + box_width_ratio / 2.) / img_width - : 1; - e_boxes(h, w, idx, 3) = - (center_y_temp + box_height_ratio / 2.) / img_height <= 1 - ? (center_y_temp + box_height_ratio / 2.) / img_height - : 1; - idx++; - } + for (size_t r = 0; r < fixed_ratios.size(); ++r) { + float ar = fixed_ratios[r]; + int shift = step_average / density; + float box_width_ratio = fixed_size * sqrt(ar); + float box_height_ratio = fixed_size / sqrt(ar); + for (int di = 0; di < density; ++di) { + for (int dj = 0; dj < density; ++dj) { + float center_x_temp = + center_x - step_average / 2. + shift / 2. + dj * shift; + float center_y_temp = + center_y - step_average / 2. + shift / 2. + di * shift; + e_boxes(h, w, idx, 0) = + (center_x_temp - box_width_ratio / 2.) / img_width >= 0 + ? (center_x_temp - box_width_ratio / 2.) / img_width + : 0; + e_boxes(h, w, idx, 1) = + (center_y_temp - box_height_ratio / 2.) / img_height >= 0 + ? (center_y_temp - box_height_ratio / 2.) / img_height + : 0; + e_boxes(h, w, idx, 2) = + (center_x_temp + box_width_ratio / 2.) / img_width <= 1 + ? (center_x_temp + box_width_ratio / 2.) / img_width + : 1; + e_boxes(h, w, idx, 3) = + (center_y_temp + box_height_ratio / 2.) / img_height <= 1 + ? (center_y_temp + box_height_ratio / 2.) / img_height + : 1; + idx++; } } } @@ -139,6 +135,7 @@ class DensityPriorBoxOpKernel : public framework::OpKernel { e_vars = var_et.broadcast(Eigen::DSizes(box_num, 1)); vars->Resize(var_dim); + boxes->Resize(box_dim); } }; // namespace operators diff --git a/paddle/fluid/operators/detection/generate_proposals_op.cc b/paddle/fluid/operators/detection/generate_proposals_op.cc index 709c2dfc4b7c6..2c46803fd00e4 100644 --- a/paddle/fluid/operators/detection/generate_proposals_op.cc +++ b/paddle/fluid/operators/detection/generate_proposals_op.cc @@ -66,9 +66,8 @@ class GenerateProposalsOp : public framework::OperatorWithKernel { protected: framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext &ctx) const override { - return framework::OpKernelType( - framework::ToDataType(ctx.Input("Anchors")->type()), - ctx.device_context()); + return framework::OpKernelType(ctx.Input("Anchors")->type(), + ctx.device_context()); } }; diff --git a/paddle/fluid/operators/detection/mine_hard_examples_op.cc b/paddle/fluid/operators/detection/mine_hard_examples_op.cc index 54a4b87ec8f13..f70e6adb5b4ae 100644 --- a/paddle/fluid/operators/detection/mine_hard_examples_op.cc +++ b/paddle/fluid/operators/detection/mine_hard_examples_op.cc @@ -249,8 +249,7 @@ class MineHardExamplesOp : public framework::OperatorWithKernel { framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext& ctx) const override { return framework::OpKernelType( - framework::ToDataType(ctx.Input("ClsLoss")->type()), - platform::CPUPlace()); + ctx.Input("ClsLoss")->type(), platform::CPUPlace()); } }; diff --git a/paddle/fluid/operators/detection/multiclass_nms_op.cc b/paddle/fluid/operators/detection/multiclass_nms_op.cc index f0f8851be0ec2..2395b18148542 100644 --- a/paddle/fluid/operators/detection/multiclass_nms_op.cc +++ b/paddle/fluid/operators/detection/multiclass_nms_op.cc @@ -65,8 +65,7 @@ class MultiClassNMSOp : public framework::OperatorWithKernel { framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext& ctx) const override { return framework::OpKernelType( - framework::ToDataType( - ctx.Input("Scores")->type()), + ctx.Input("Scores")->type(), platform::CPUPlace()); } }; diff --git a/paddle/fluid/operators/detection/prior_box_op.cc b/paddle/fluid/operators/detection/prior_box_op.cc index b5cb6a724c095..3e75c0394f971 100644 --- a/paddle/fluid/operators/detection/prior_box_op.cc +++ b/paddle/fluid/operators/detection/prior_box_op.cc @@ -72,8 +72,7 @@ class PriorBoxOp : public framework::OperatorWithKernel { framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext& ctx) const override { return framework::OpKernelType( - framework::ToDataType(ctx.Input("Input")->type()), - ctx.device_context()); + ctx.Input("Input")->type(), ctx.device_context()); } }; diff --git a/paddle/fluid/operators/detection/roi_perspective_transform_op.cc b/paddle/fluid/operators/detection/roi_perspective_transform_op.cc index 42c720e701fba..3796854fe6738 100644 --- a/paddle/fluid/operators/detection/roi_perspective_transform_op.cc +++ b/paddle/fluid/operators/detection/roi_perspective_transform_op.cc @@ -498,9 +498,8 @@ class ROIPerspectiveTransformOp : public framework::OperatorWithKernel { protected: framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext& ctx) const override { - return framework::OpKernelType( - framework::ToDataType(ctx.Input("X")->type()), - ctx.device_context()); + return framework::OpKernelType(ctx.Input("X")->type(), + ctx.device_context()); } }; @@ -519,9 +518,8 @@ class ROIPerspectiveTransformGradOp : public framework::OperatorWithKernel { protected: framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext& ctx) const override { - return framework::OpKernelType( - framework::ToDataType(ctx.Input("X")->type()), - ctx.device_context()); + return framework::OpKernelType(ctx.Input("X")->type(), + ctx.device_context()); } }; diff --git a/paddle/fluid/operators/detection/roi_perspective_transform_op.cu b/paddle/fluid/operators/detection/roi_perspective_transform_op.cu index 2d262f932aed9..862d664d42e03 100644 --- a/paddle/fluid/operators/detection/roi_perspective_transform_op.cu +++ b/paddle/fluid/operators/detection/roi_perspective_transform_op.cu @@ -35,12 +35,12 @@ namespace operators { template __device__ bool GT_E(T a, T b) { - return (a > b) || fabs(a - b) < 1e-4; + return (a > b) || Eigen::numext::abs(a - b) < 1e-4; } template __device__ bool LT_E(T a, T b) { - return (a < b) || fabs(a - b) < 1e-4; + return (a < b) || Eigen::numext::abs(a - b) < 1e-4; } template diff --git a/paddle/fluid/operators/detection/rpn_target_assign_op.cc b/paddle/fluid/operators/detection/rpn_target_assign_op.cc index 46fff9d338b77..dc6c3d5a668f9 100644 --- a/paddle/fluid/operators/detection/rpn_target_assign_op.cc +++ b/paddle/fluid/operators/detection/rpn_target_assign_op.cc @@ -78,8 +78,7 @@ class RpnTargetAssignOp : public framework::OperatorWithKernel { framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext& ctx) const override { return framework::OpKernelType( - framework::ToDataType( - ctx.Input("Anchor")->type()), + ctx.Input("Anchor")->type(), platform::CPUPlace()); } }; diff --git a/paddle/fluid/operators/detection/target_assign_op.cc b/paddle/fluid/operators/detection/target_assign_op.cc index 3670019392511..c057c82ce0f5e 100644 --- a/paddle/fluid/operators/detection/target_assign_op.cc +++ b/paddle/fluid/operators/detection/target_assign_op.cc @@ -57,9 +57,8 @@ class TargetAssignOp : public framework::OperatorWithKernel { protected: framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext& ctx) const override { - return framework::OpKernelType( - framework::ToDataType(ctx.Input("X")->type()), - ctx.device_context()); + return framework::OpKernelType(ctx.Input("X")->type(), + ctx.device_context()); } }; diff --git a/paddle/fluid/operators/detection_map_op.cc b/paddle/fluid/operators/detection_map_op.cc index d7f49a9590e4e..e1d113f8542da 100644 --- a/paddle/fluid/operators/detection_map_op.cc +++ b/paddle/fluid/operators/detection_map_op.cc @@ -71,8 +71,7 @@ class DetectionMAPOp : public framework::OperatorWithKernel { framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext& ctx) const override { return framework::OpKernelType( - framework::ToDataType( - ctx.Input("DetectRes")->type()), + ctx.Input("DetectRes")->type(), platform::CPUPlace()); } }; diff --git a/paddle/fluid/operators/distributed/CMakeLists.txt b/paddle/fluid/operators/distributed/CMakeLists.txt index 21db93958a4a5..eab4297c737bb 100644 --- a/paddle/fluid/operators/distributed/CMakeLists.txt +++ b/paddle/fluid/operators/distributed/CMakeLists.txt @@ -9,36 +9,54 @@ else() endif() configure_file(send_recv.proto.in ${CMAKE_CURRENT_SOURCE_DIR}/send_recv.proto @ONLY) +set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor") + if(WITH_GRPC) - grpc_library(sendrecvop_grpc SRCS grpc_bytebuffer_stream.cc sendrecvop_utils.cc grpc_client.cc - request_handler_impl.cc rpc_client.cc rpc_server.cc grpc_server.cc variable_response.cc grpc_variable_response.cc grpc_serde.cc + grpc_library(sendrecvop_rpc SRCS grpc_bytebuffer_stream.cc sendrecvop_utils.cc grpc_client.cc + request_handler_impl.cc rpc_client.cc rpc_server.cc grpc_server.cc variable_response.cc grpc_variable_response.cc grpc_serde.cc collective_client.cc collective_server.cc PROTO send_recv.proto - DEPS lod_tensor selected_rows memory) - set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor") + DEPS lod_tensor selected_rows_functor memory) + set_source_files_properties(grpc_serde_test.cc rpc_server_test.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) + cc_test(grpc_serde_test SRCS grpc_serde_test.cc - DEPS grpc++_unsecure grpc_unsecure gpr cares zlib protobuf sendrecvop_grpc scope profiler math_function SERIAL) + DEPS grpc++_unsecure grpc_unsecure gpr cares zlib protobuf sendrecvop_rpc scope profiler math_function SERIAL) + cc_test(rpc_server_test SRCS rpc_server_test.cc - DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib protobuf executor proto_desc lookup_sparse_table_op SERIAL) + DEPS sendrecvop_rpc grpc++_unsecure grpc_unsecure gpr cares zlib protobuf executor proto_desc lookup_sparse_table_op SERIAL) + cc_test(varhandle_test SRCS varhandle_test.cc DEPS profiler) - return() -endif() + if(WITH_GPU) + cc_test(collective_server_test SRCS collective_server_test.cc + DEPS sendrecvop_rpc grpc++_unsecure grpc_unsecure gpr cares zlib protobuf executor + selected_rows_functor scope math_function SERIAL) + endif() -set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor") + cc_library(parameter_prefetch SRCS parameter_prefetch.cc DEPS sendrecvop_rpc memory) +else() + set_source_files_properties(brpc_server.cc parameter_prefetch.cc brpc_client.cc rpc_server_test.cc brpc_serde_test.cc + brpc_variable_response.cc brpc_sendrecvop_utils.cc brpc_rdma_pool.cc collective_server.cc collective_server_test.cc + collective_client.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) -set_source_files_properties(brpc_server.cc brpc_client.cc rpc_server_test.cc brpc_serde_test.cc - brpc_variable_response.cc brpc_sendrecvop_utils.cc brpc_rdma_pool.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) + brpc_library(sendrecvop_rpc SRCS brpc_client.cc brpc_server.cc rpc_server.cc rpc_client.cc request_handler_impl.cc brpc_sendrecvop_utils.cc + brpc_variable_response.cc variable_response.cc sendrecvop_utils.cc brpc_rdma_pool.cc collective_client.cc collective_server.cc + PROTO send_recv.proto + DEPS lod_tensor selected_rows memory) -brpc_library(sendrecvop_brpc SRCS brpc_client.cc brpc_server.cc rpc_server.cc rpc_client.cc request_handler_impl.cc brpc_sendrecvop_utils.cc - brpc_variable_response.cc variable_response.cc sendrecvop_utils.cc brpc_rdma_pool.cc - PROTO send_recv.proto - DEPS lod_tensor selected_rows memory) + cc_library(parameter_prefetch SRCS parameter_prefetch.cc DEPS sendrecvop_rpc memory) -set(brpc_test_depends sendrecvop_brpc brpc ssl crypto protobuf leveldb gflags glog executor proto_desc lookup_table_op snappystream snappy) + set(brpc_test_depends sendrecvop_rpc brpc ssl crypto protobuf leveldb gflags glog executor + proto_desc lookup_sparse_table_op snappystream snappy zlib) -cc_test(brpc_server_test SRCS rpc_server_test.cc - DEPS ${brpc_test_depends} SERIAL) + cc_test(rpc_server_test SRCS rpc_server_test.cc + DEPS ${brpc_test_depends} SERIAL) + + cc_test(brpc_serde_test SRCS brpc_serde_test.cc + DEPS ${brpc_test_depends} SERIAL) -cc_test(brpc_serde_test SRCS brpc_serde_test.cc - DEPS ${brpc_test_depends} SERIAL) + if(WITH_GPU) + cc_test(collective_server_test SRCS collective_server_test.cc + DEPS ${brpc_test_depends} selected_rows_functor scope math_function SERIAL) + endif() +endif() diff --git a/paddle/fluid/operators/distributed/brpc_client.cc b/paddle/fluid/operators/distributed/brpc_client.cc index b394c678fb650..62e32977b8cd7 100644 --- a/paddle/fluid/operators/distributed/brpc_client.cc +++ b/paddle/fluid/operators/distributed/brpc_client.cc @@ -14,135 +14,316 @@ #include "paddle/fluid/operators/distributed/brpc_client.h" #include "paddle/fluid/framework/threadpool.h" +#include "paddle/fluid/operators/distributed/brpc_sendrecvop_utils.h" +#include "paddle/fluid/platform/profiler.h" namespace paddle { namespace operators { namespace distributed { -DEFINE_int32(brpc_channel_num, 24, - "Number of channels to send requests connected to one server"); DEFINE_int32(timeout_ms, 30000, "RPC timeout in milliseconds"); DEFINE_int32(max_retry, 3, "Max retries(not including the first RPC)"); BRPCClient::~BRPCClient() { Wait(); } -void HandleSendResponse(brpc::Controller* cntl, - sendrecv::VoidMessage* response) { +void HandleSendResponse(brpc::Controller* cntl, sendrecv::VoidMessage* response, + VarHandlePtr var_h, ChannelQueuePtr ch_ptr, + ChannelContextPtr ch_ctx, BRPCClient* cls) { // std::unique_ptr makes sure cntl/response will be deleted before returning. std::unique_ptr cntl_guard(cntl); std::unique_ptr response_guard(response); + // this channel can be used by other now. + ch_ptr->Push(ch_ctx); + if (cntl->Failed()) { - LOG(WARNING) << "Fail to send EchoRequest, " << cntl->ErrorText(); + LOG(FATAL) << "Fail to send SendVar: " << var_h->name() + << ", error text: " << cntl->ErrorText(); + var_h->Finish(false); + cls->DecreaseReqCount(); return; } - LOG(INFO) << "Received response from " << cntl->remote_side() - << " latency=" << cntl->latency_us() << "us"; + var_h->Finish(true); + cls->DecreaseReqCount(); + + VLOG(4) << "HandleSendResponse from: " << cntl->remote_side() + << ", varname: " << var_h->name() + << ", latency: " << cntl->latency_us() << "us"; + VLOG(4) << "Finish HandleSendResponse"; } -bool BRPCClient::AsyncSendVar(const std::string& ep, - const platform::DeviceContext& ctx, - const framework::Scope& scope, - const std::string& var_name, int64_t time_out) { +VarHandlePtr BRPCClient::AsyncSendVar(const std::string& ep, + const platform::DeviceContext& ctx, + const framework::Scope& scope, + const std::string& var_name, + int64_t time_out) { const platform::DeviceContext* p_ctx = &ctx; const std::string ep_val = ep; const std::string var_name_val = var_name; const framework::Scope* p_scope = &scope; const auto ch_ptr = GetChannel(ep_val); + const std::string method = "SendRPC"; + VarHandlePtr var_h(new VarHandle(ep, method, var_name_val, p_ctx, p_scope)); + + framework::AsyncIO([=] { + auto ch_ctx = ch_ptr->Pop(); + brpc::Controller* cntl = new brpc::Controller(); + sendrecv::VoidMessage* response = new sendrecv::VoidMessage(); + cntl->set_timeout_ms(time_out); - framework::AsyncIO( - [var_name_val, p_ctx, ep_val, p_scope, time_out, ch_ptr, this] { - auto ch_ctx = ch_ptr->Pop(); - brpc::Controller* cntl = new brpc::Controller(); - sendrecv::VoidMessage* response = new sendrecv::VoidMessage(); - cntl->set_timeout_ms(time_out); + auto* var = p_scope->FindVar(var_name_val); + sendrecv::VariableMessage request; + distributed::SerializeToIOBuf(var_name_val, var, *p_ctx, &request, + &cntl->request_attachment(), "", false, + trainer_id_); - google::protobuf::Closure* done = - brpc::NewCallback(&HandleSendResponse, cntl, response); + google::protobuf::Closure* done = brpc::NewCallback( + &HandleSendResponse, cntl, response, var_h, ch_ptr, ch_ctx, this); - sendrecv::VariableMessage request; - ch_ctx->stub->SendVariable(cntl, &request, response, done); - }); + platform::RecordRPCEvent record_event(method, p_ctx); + + ch_ctx->stub->SendVariable(cntl, &request, response, done); + + if (UNLIKELY(platform::IsProfileEnabled())) { + var_h->Wait(); + } + }); req_count_++; - return true; + return var_h; } +void HandleFetchBarrierResponse(brpc::Controller* cntl, + sendrecv::VariableMessage* response, + VarHandlePtr var_h, ChannelQueuePtr ch_ptr, + ChannelContextPtr ch_ctx, BRPCClient* cls) { + // std::unique_ptr makes sure cntl/response will be deleted before returning. + std::unique_ptr cntl_guard(cntl); + std::unique_ptr response_guard(response); + + // this channel can be used other now. + ch_ptr->Push(ch_ctx); + if (cntl->Failed()) { + LOG(FATAL) << "Fail to get HandleFetchBarrierResponse: " << var_h->name() + << ", error text: " << cntl->ErrorText(); + var_h->Finish(false); + cls->DecreaseReqCount(); + return; + } + + var_h->Finish(true); + cls->DecreaseReqCount(); + + VLOG(4) << "HandleFetchBarrierResponse from: " << cntl->remote_side() + << ", varname: " << var_h->name() + << ", latency: " << cntl->latency_us() << "us"; + VLOG(4) << "Finish HandleFetchBarrierResponse"; +} void HandleGetResponse(brpc::Controller* cntl, - sendrecv::VariableMessage* response) { + sendrecv::VariableMessage* response, VarHandlePtr var_h, + ChannelQueuePtr ch_ptr, ChannelContextPtr ch_ctx, + BRPCClient* cls) { // std::unique_ptr makes sure cntl/response will be deleted before returning. std::unique_ptr cntl_guard(cntl); std::unique_ptr response_guard(response); + // this channel can be used other now. + ch_ptr->Push(ch_ctx); + if (cntl->Failed()) { - LOG(WARNING) << "Fail to send EchoRequest, " << cntl->ErrorText(); + LOG(FATAL) << "Fail to GetVar: " << var_h->name() + << ", error text: " << cntl->ErrorText(); + cls->DecreaseReqCount(); + var_h->Finish(false); return; } - LOG(INFO) << "Received response from " << cntl->remote_side() - << " latency=" << cntl->latency_us() << "us"; - // framework::Variable* outvar = nullptr; - // DeserializeFromByteBuffer(ret_msg, *var_h.ctx, var_h.scope, &outvar); + VLOG(4) << "HandleGetResponse from: " << cntl->remote_side() + << ", varname: " << var_h->name() + << ", latency: " << cntl->latency_us() << "us"; + + framework::Variable* outvar = nullptr; + int trainer_id; + distributed::DeserializeFromIOBuf(*response, cntl->response_attachment(), + *var_h->ctx(), var_h->scope(), &outvar, + &trainer_id); + VLOG(4) << "Finish HandleGetResponse"; + cls->DecreaseReqCount(); + var_h->Finish(true); } -bool BRPCClient::AsyncGetVar(const std::string& ep, - const platform::DeviceContext& ctx, - const framework::Scope& scope, - const std::string& var_name, int64_t time_out) { +VarHandlePtr BRPCClient::_AsyncGetVar(const std::string& ep, + const platform::DeviceContext& ctx, + const framework::Scope& scope, + const std::string& var_name, + const std::string& method_name, + int64_t time_out) { const platform::DeviceContext* p_ctx = &ctx; const std::string ep_val = ep; const std::string var_name_val = var_name; const framework::Scope* p_scope = &scope; - const auto ch = GetChannel(ep_val); + const auto ch_ptr = GetChannel(ep_val); + const std::string method = "GetRPC"; + VarHandlePtr var_h(new VarHandle(ep, method, var_name_val, p_ctx, p_scope)); + + framework::AsyncIO([=] { + auto ch_ctx = ch_ptr->Pop(); + + brpc::Controller* cntl = new brpc::Controller(); + sendrecv::VariableMessage* response = new sendrecv::VariableMessage(); + cntl->set_timeout_ms(time_out); - framework::AsyncIO( - [var_name_val, ep_val, p_scope, p_ctx, time_out, ch, this] {}); + sendrecv::VariableMessage req; + req.set_varname(var_name_val); + req.set_trainer_id(trainer_id_); + + google::protobuf::Closure* done = brpc::NewCallback( + &HandleGetResponse, cntl, response, var_h, ch_ptr, ch_ctx, this); + + platform::RecordRPCEvent record_event(method, p_ctx); + + if (method_name == "GetMonomerVariable") { + ch_ctx->stub->GetMonomerVariable(cntl, &req, response, done); + } else { + ch_ctx->stub->GetVariable(cntl, &req, response, done); + } + + if (UNLIKELY(platform::IsProfileEnabled())) { + var_h->Wait(); + } + }); req_count_++; - return true; + return var_h; +} + +VarHandlePtr BRPCClient::AsyncGetMonomerVariable( + const std::string& ep, const platform::DeviceContext& ctx, + const framework::Scope& scope, const std::string& var_name, + int64_t time_out) { + return _AsyncGetVar(ep, ctx, scope, var_name, "GetMonomerVariable", time_out); +} + +VarHandlePtr BRPCClient::AsyncGetMonomerBarrier(const std::string& ep, + const std::string& var_name, + int64_t time_out) { + return AsyncSendMessage(ep, "GetMonomerBarrier", var_name, time_out); } -bool BRPCClient::AsyncPrefetchVar(const std::string& ep, - const platform::DeviceContext& ctx, - const framework::Scope& scope, - const std::string& in_var_name, - const std::string& out_var_name, - int64_t time_out) { +VarHandlePtr BRPCClient::AsyncGetVar(const std::string& ep, + const platform::DeviceContext& ctx, + const framework::Scope& scope, + const std::string& var_name, + int64_t time_out) { + return _AsyncGetVar(ep, ctx, scope, var_name, "GetVariable", time_out); +} + +VarHandlePtr BRPCClient::AsyncPrefetchVar(const std::string& ep, + const platform::DeviceContext& ctx, + const framework::Scope& scope, + const std::string& in_var_name, + const std::string& out_var_name, + const std::string& table_name, + int64_t time_out) { const platform::DeviceContext* p_ctx = &ctx; const std::string ep_val = ep; const std::string in_var_name_val = in_var_name; const std::string out_var_name_val = out_var_name; + const std::string table_name_val = table_name; const framework::Scope* p_scope = &scope; - const auto ch = GetChannel(ep_val); + const auto ch_ptr = GetChannel(ep_val); + + const std::string method = "PrefetchRPC"; + + VarHandlePtr var_h( + new VarHandle(ep, method, out_var_name_val, p_ctx, p_scope)); + + framework::AsyncIO([=] { + auto ch_ctx = ch_ptr->Pop(); + + brpc::Controller* cntl = new brpc::Controller(); + sendrecv::VariableMessage* response = new sendrecv::VariableMessage(); + cntl->set_timeout_ms(time_out); + + auto* var = p_scope->FindVar(in_var_name_val); + sendrecv::VariableMessage req; + distributed::SerializeToIOBuf(in_var_name_val, var, *p_ctx, &req, + &cntl->request_attachment(), out_var_name_val, + false, 0, table_name_val); + + platform::RecordRPCEvent record_event(method, p_ctx); + + google::protobuf::Closure* done = brpc::NewCallback( + &HandleGetResponse, cntl, response, var_h, ch_ptr, ch_ctx, this); - framework::AsyncIO([in_var_name_val, out_var_name_val, ep_val, p_scope, p_ctx, - time_out, ch, this] {}); + ch_ctx->stub->PrefetchVariable(cntl, &req, response, done); + + if (UNLIKELY(platform::IsProfileEnabled())) { + var_h->Wait(); + } + }); req_count_++; - return true; + return var_h; } -void BRPCClient::AsyncSendBatchBarrier(const std::string& ep, - int64_t time_out) { - req_count_++; +VarHandlePtr BRPCClient::AsyncSendBatchBarrier(const std::string& ep, + int64_t time_out) { + return AsyncSendMessage(ep, "BatchBarrierRPC", BATCH_BARRIER_MESSAGE, + time_out); } -void BRPCClient::AsyncSendFetchBarrier(const std::string& ep, - int64_t time_out) { +VarHandlePtr BRPCClient::AsyncSendFetchBarrier(const std::string& ep, + int64_t time_out) { + auto ch_ptr = GetChannel(ep); + auto ch_ctx = ch_ptr->Pop(); + + brpc::Controller* cntl = new brpc::Controller(); + sendrecv::VariableMessage* response = new sendrecv::VariableMessage(); + cntl->set_timeout_ms(time_out); + + sendrecv::VariableMessage req; + req.set_varname(FETCH_BARRIER_MESSAGE); + + const std::string method = "FetchBarrierRPC"; + // var handle + VarHandlePtr var_h( + new VarHandle(ep, method, FETCH_BARRIER_MESSAGE, nullptr, nullptr)); + + platform::RecordRPCEvent record_event(method, nullptr); + + google::protobuf::Closure* done = brpc::NewCallback( + &HandleFetchBarrierResponse, cntl, response, var_h, ch_ptr, ch_ctx, this); + + ch_ctx->stub->GetVariable(cntl, &req, response, done); + req_count_++; + + if (UNLIKELY(platform::IsProfileEnabled())) { + var_h->Wait(); + } + + return var_h; } -void BRPCClient::Wait() { - std::unique_lock lk(sync_mutex_); - sync_cond_.wait(lk, [this] { return req_count_ == 0; }); +bool BRPCClient::Wait() { + VLOG(9) << "begin to brpcclient wait"; + { + std::unique_lock lk(sync_mutex_); + sync_cond_.wait(lk, [this] { return req_count_ == 0; }); + } + VLOG(9) << "end to brpcclient wait"; + return true; } ChannelQueuePtr BRPCClient::GetChannel(const std::string& ep) { + VLOG(4) << "begin to GetChannel:" << ep; { std::lock_guard guard(chan_mutex_); auto it = channels_.find(ep); if (it != channels_.end()) { + VLOG(4) << "end to GetChannel:" << ep; return it->second; } } @@ -150,15 +331,23 @@ ChannelQueuePtr BRPCClient::GetChannel(const std::string& ep) { ChannelQueuePtr q(new framework::BlockingQueue()); brpc::ChannelOptions options; +#ifdef PADDLE_WITH_BRPC_RDMA + options.use_rdma = true; +#endif options.protocol = "baidu_std"; - options.connection_type = "pooled"; - options.connect_timeout_ms = 100; + // don't use pooled type. the server can't afford that. + options.connection_type = "single"; + options.connect_timeout_ms = 1000; options.timeout_ms = FLAGS_timeout_ms /*milliseconds*/; options.max_retry = FLAGS_max_retry; - for (int i = 0; i < FLAGS_brpc_channel_num; ++i) { + + VLOG(1) << "create " << brpc_channel_num_per_server_ + << " brpc channels to pserver:" << ep; + + for (int i = 0; i < brpc_channel_num_per_server_; ++i) { std::shared_ptr c(new ChannelContext()); if (c->channel.Init(ep.c_str(), &options) != 0) { - LOG(ERROR) << "Fail to initialize channel"; + LOG(FATAL) << "Fail to initialize channel"; return nullptr; } @@ -172,9 +361,75 @@ ChannelQueuePtr BRPCClient::GetChannel(const std::string& ep) { channels_[ep] = q; } + VLOG(4) << "end to GetChannel:" << ep; return q; } +VarHandlePtr BRPCClient::AsyncSendComplete(const std::string& ep, + int64_t time_out) { + return AsyncSendMessage(ep, "SendCompleteRPC", COMPLETE_MESSAGE, time_out); +} + +void BRPCClient::SendComplete() { + for (auto& kv : channels_) { + AsyncSendComplete(kv.first); + } +} + +VarHandlePtr BRPCClient::AsyncSendVarMessage( + const std::string& ep, const std::string& method_name, + const sendrecv::VariableMessage& req, int64_t time_out) { + auto ch_ptr = GetChannel(ep); + auto ch_ctx = ch_ptr->Pop(); + + brpc::Controller* cntl = new brpc::Controller(); + sendrecv::VoidMessage* response = new sendrecv::VoidMessage(); + cntl->set_timeout_ms(time_out); + + platform::RecordRPCEvent record_event(method_name, nullptr); + + VarHandlePtr var_h( + new VarHandle(ep, method_name, req.varname(), nullptr, nullptr)); + + google::protobuf::Closure* done = brpc::NewCallback( + &HandleSendResponse, cntl, response, var_h, ch_ptr, ch_ctx, this); + + if (method_name == "CheckPointNotifyRPC") { + ch_ctx->stub->CheckpointNotify(cntl, &req, response, done); + } else if (method_name == "GetMonomerBarrier") { + ch_ctx->stub->GetMonomerBarrier(cntl, &req, response, done); + } else { + ch_ctx->stub->SendVariable(cntl, &req, response, done); + } + req_count_++; + + if (UNLIKELY(platform::IsProfileEnabled())) { + var_h->Wait(); + } + + return var_h; +} + +VarHandlePtr BRPCClient::AsyncSendMessage(const std::string& ep, + const std::string& method_name, + const std::string& message, + int64_t time_out) { + sendrecv::VariableMessage req; + req.set_varname(message); + + return AsyncSendVarMessage(ep, method_name, req, time_out); +} + +VarHandlePtr BRPCClient::AsyncCheckpointNotify(const std::string& ep, + const std::string& dir, + int64_t time_out) { + sendrecv::VariableMessage req; + req.set_varname(CHECKPOINT_SAVE_MESSAGE); + req.set_out_varname(dir); + + return AsyncSendVarMessage(ep, "CheckPointNotifyRPC", req, time_out); +} + } // namespace distributed } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/distributed/brpc_client.h b/paddle/fluid/operators/distributed/brpc_client.h index 8ff1f0a6076b3..80cc81bff3791 100644 --- a/paddle/fluid/operators/distributed/brpc_client.h +++ b/paddle/fluid/operators/distributed/brpc_client.h @@ -31,6 +31,8 @@ limitations under the License. */ #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/scope.h" #include "paddle/fluid/framework/selected_rows.h" +#include "paddle/fluid/operators/distributed/brpc_sendrecvop_utils.h" +#include "paddle/fluid/operators/distributed/request_handler.h" #include "paddle/fluid/operators/distributed/rpc_client.h" #include "paddle/fluid/operators/distributed/send_recv.pb.h" #include "paddle/fluid/platform/macros.h" // for DISABLE_COPY_AND_ASSIGN @@ -53,33 +55,94 @@ class BRPCClient : public RPCClient { BRPCClient() {} virtual ~BRPCClient(); - bool AsyncSendVar(const std::string& ep, const platform::DeviceContext& ctx, - const framework::Scope& scope, const std::string& var_name, - int64_t time_out = FLAGS_rpc_deadline) override; + VarHandlePtr AsyncSendVar(const std::string& ep, + const platform::DeviceContext& ctx, + const framework::Scope& scope, + const std::string& var_name, + int64_t time_out = FLAGS_rpc_deadline) override; - bool AsyncGetVar(const std::string& ep, const platform::DeviceContext& ctx, - const framework::Scope& scope, const std::string& var_name, - int64_t time_out = FLAGS_rpc_deadline) override; + VarHandlePtr AsyncGetVar(const std::string& ep, + const platform::DeviceContext& ctx, + const framework::Scope& scope, + const std::string& var_name, + int64_t time_out = FLAGS_rpc_deadline) override; - bool AsyncPrefetchVar(const std::string& ep, - const platform::DeviceContext& ctx, - const framework::Scope& scope, - const std::string& in_var_name, - const std::string& out_var_name, - int64_t time_out = FLAGS_rpc_deadline) override; + VarHandlePtr AsyncGetMonomerBarrier( + const std::string& ep, const std::string& var_name, + int64_t time_out = FLAGS_rpc_deadline) override; - void AsyncSendBatchBarrier(const std::string& ep, - int64_t time_out = FLAGS_rpc_deadline) override; + VarHandlePtr AsyncGetMonomerVariable( + const std::string& ep, const platform::DeviceContext& ctx, + const framework::Scope& scope, const std::string& var_name, + int64_t time_out = FLAGS_rpc_deadline) override; - void AsyncSendFetchBarrier(const std::string& ep, - int64_t time_out = FLAGS_rpc_deadline) override; + VarHandlePtr AsyncPrefetchVar(const std::string& ep, + const platform::DeviceContext& ctx, + const framework::Scope& scope, + const std::string& in_var_name, + const std::string& out_var_name, + const std::string& table_name = "", + int64_t time_out = FLAGS_rpc_deadline) override; - void Wait() override; + VarHandlePtr AsyncSendBatchBarrier( + const std::string& ep, int64_t time_out = FLAGS_rpc_deadline) override; + + VarHandlePtr AsyncSendFetchBarrier( + const std::string& ep, int64_t time_out = FLAGS_rpc_deadline) override; + + VarHandlePtr AsyncCheckpointNotify( + const std::string& ep, const std::string& dir, + int64_t time_out = FLAGS_rpc_deadline) override; + + bool Wait() override; + + void SendComplete() override; private: + VarHandlePtr _AsyncGetVar(const std::string& ep, + const platform::DeviceContext& ctx, + const framework::Scope& scope, + const std::string& var_name, + const std::string& method_name, + int64_t time_out = FLAGS_rpc_deadline); + void Proceed(); ChannelQueuePtr GetChannel(const std::string& ep); + VarHandlePtr AsyncSendComplete(const std::string& ep, + int64_t time_out = FLAGS_rpc_deadline); + + VarHandlePtr AsyncSendMessage(const std::string& ep, + const std::string& method_name, + const std::string& message, int64_t time_out); + + VarHandlePtr AsyncSendVarMessage(const std::string& ep, + const std::string& method_name, + const sendrecv::VariableMessage& req, + int64_t time_out); + + friend void HandleSendResponse(brpc::Controller* cntl, + sendrecv::VoidMessage* response, + VarHandlePtr var_h, ChannelQueuePtr ch_ptr, + ChannelContextPtr ch_ctx, BRPCClient* cls); + + friend void HandleGetResponse(brpc::Controller* cntl, + sendrecv::VariableMessage* response, + VarHandlePtr var_h, ChannelQueuePtr ch_ptr, + ChannelContextPtr ch_ctx, BRPCClient* cls); + + friend void HandleFetchBarrierResponse(brpc::Controller* cntl, + sendrecv::VariableMessage* response, + VarHandlePtr var_h, + ChannelQueuePtr ch_ptr, + ChannelContextPtr ch_ctx, + BRPCClient* cls); + void DecreaseReqCount() { + if (--req_count_ <= 0) { + sync_cond_.notify_all(); + } + } + private: std::unordered_map channels_; @@ -88,6 +151,8 @@ class BRPCClient : public RPCClient { std::condition_variable sync_cond_; std::atomic req_count_{0}; + static constexpr int brpc_channel_num_per_server_ = 4; + // mutex for GetChannel thread safety std::mutex chan_mutex_; DISABLE_COPY_AND_ASSIGN(BRPCClient); diff --git a/paddle/fluid/operators/distributed/brpc_rdma_pool.cc b/paddle/fluid/operators/distributed/brpc_rdma_pool.cc new file mode 100644 index 0000000000000..e1be5673dfbc5 --- /dev/null +++ b/paddle/fluid/operators/distributed/brpc_rdma_pool.cc @@ -0,0 +1,84 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#ifdef PADDLE_WITH_BRPC_RDMA + +#include "paddle/fluid/operators/distributed/brpc_rdma_pool.h" +#include "brpc/channel.h" +#include "brpc/rdma/rdma_helper.h" +#include "paddle/fluid/platform/enforce.h" + +namespace paddle { +namespace operators { +namespace distributed { + +RdmaMemPool& RdmaMemPool::Instance() { + static RdmaMemPool* g_rdma_mem_pool = new RdmaMemPool(); + return *g_rdma_mem_pool; +} + +void* RdmaMemPool::Find(const std::string& varname, int64_t size) { + pthread_rwlock_rdlock(&access_); + auto it = pool_.find(varname); + if (it == pool_.end()) { + pthread_rwlock_unlock(&access_); + return nullptr; + } + + auto info = it->second; + if (info.data_size != size) { + pthread_rwlock_unlock(&access_); + PADDLE_ENFORCE(false, "var:%s size:%ld != %ld", varname, size, + info.data_size); + return nullptr; + } + + pthread_rwlock_unlock(&access_); + return info.data; +} + +void RdmaMemPool::Register(const std::string& varname, void* data, + int64_t data_size) { + void* old = Find(varname, data_size); + if (old != nullptr) { + if (data != old) { + PADDLE_ENFORCE(false, "var:%s data:%ld != %ld", varname, data, old); + } + VLOG(7) << "Find on rdma:" << varname << " data:" << data + << " data_size:" << data_size; + return; + } + + VarInfo info; + info.data = data; + info.data_size = data_size; + + pthread_rwlock_wrlock(&access_); + pool_[varname] = info; + pthread_rwlock_unlock(&access_); + + if (brpc::rdma::RegisterMemoryForRdma(data, data_size)) { + LOG(FATAL) << "register " << varname << " data:" << data + << " data_size:" << data_size << " error"; + } + + VLOG(4) << "register on rdma:" << varname << " data:" << data + << " data_size:" << data_size; +} + +} // namespace distributed +} // namespace operators +} // namespace paddle + +#endif diff --git a/paddle/fluid/operators/distributed/brpc_rdma_pool.h b/paddle/fluid/operators/distributed/brpc_rdma_pool.h new file mode 100644 index 0000000000000..156a93ec57847 --- /dev/null +++ b/paddle/fluid/operators/distributed/brpc_rdma_pool.h @@ -0,0 +1,56 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once +#ifdef PADDLE_WITH_BRPC_RDMA + +#include // NOLINT +#include +#include + +namespace paddle { +namespace operators { +namespace distributed { + +/* + * This class is used to avoid duplicated registion of brpc::rdma. + */ +class RdmaMemPool { + public: + static RdmaMemPool& Instance(); + RdmaMemPool() : access_(PTHREAD_RWLOCK_INITIALIZER) {} + + virtual ~RdmaMemPool() { pthread_rwlock_destroy(&access_); } + + void Register(const std::string& varname, void* data, int64_t size); + void* Find(const std::string& varname, int64_t size); + + private: + struct VarInfo { + void* data; + int64_t data_size; + + VarInfo() : data(nullptr), data_size(0) {} + }; + + private: + std::unordered_map pool_; + pthread_rwlock_t access_; +}; + +} // namespace distributed +} // namespace operators +} // namespace paddle + +#endif diff --git a/paddle/fluid/operators/distributed/brpc_sendrecvop_utils.cc b/paddle/fluid/operators/distributed/brpc_sendrecvop_utils.cc new file mode 100644 index 0000000000000..e4604db3a3816 --- /dev/null +++ b/paddle/fluid/operators/distributed/brpc_sendrecvop_utils.cc @@ -0,0 +1,207 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#ifdef PADDLE_WITH_CUDA +#include +#endif +#include +#include +#include // NOLINT + +#include "paddle/fluid/framework/data_type.h" +#include "paddle/fluid/operators/distributed/brpc_rdma_pool.h" +#include "paddle/fluid/operators/distributed/brpc_sendrecvop_utils.h" +#include "paddle/fluid/operators/distributed/brpc_variable_response.h" +#include "paddle/fluid/operators/distributed/send_recv.pb.h" +#include "paddle/fluid/platform/profiler.h" + +namespace paddle { +namespace operators { +namespace distributed { + +class IOBufWriter { + public: + static void Append(const std::string& varname, butil::IOBuf* iobuf, int k, + const char* v, int64_t vlen) { + if (vlen >= std::numeric_limits::max() || vlen < 0) { + LOG(FATAL) << "AppendZeroCopy varname:" << varname << ", vlen:" << vlen; + } + + iobuf->append(reinterpret_cast(&k), 4); + iobuf->append(reinterpret_cast(&vlen), 8); + iobuf->append(v, vlen); + } + + static void AppendTCPZeroCopy(butil::IOBuf* iobuf, int k, const char* v, + int64_t vlen, bool in_cuda_pinned, + void (*destroy)(void*), void* user_data) { + VLOG(7) << "AppendTCPZeroCopy " + << " k:" << k + << " data:" << static_cast(const_cast(v)) + << " data_size:" << vlen << " in_cuda_pinned:" << in_cuda_pinned; + + iobuf->append(reinterpret_cast(&k), 4); + iobuf->append(reinterpret_cast(&vlen), 8); + + // FIXME(gongwb): use append_zerocopy + /* + if (in_cuda_pinned) { + iobuf->append_zerocopy(v, vlen, IOBufWriter::FreeMemory); + } else { + iobuf->append_zerocopy(v, vlen, nullptr); + } + */ + iobuf->append(v, vlen); + destroy(user_data); + } + +#ifdef PADDLE_WITH_BRPC_RDMA + static void AppendRdmaZeroCopy(const std::string varname, butil::IOBuf* iobuf, + int k, const char* v, int64_t vlen, + bool in_cuda_pinned, void (*destroy)(void*), + void* user_data) { + VLOG(7) << "AppendRdmaZeroCopy varname:" << varname << " k:" << k + << " data:" << static_cast(const_cast(v)) + << " data_size:" << vlen << " in_cuda_pinned:" << in_cuda_pinned; + + iobuf->append(reinterpret_cast(&k), 4); + iobuf->append(reinterpret_cast(&vlen), 8); + + RdmaMemPool::Instance().Register( + varname, static_cast(const_cast(v)), vlen); + + // FIXME(gongwb): use append_zerocopy + // iobuf->append_zerocopy(v, vlen, nullptr); + iobuf->append(v, vlen); + destroy(user_data); + return; + } +#endif + + static void AppendZeroCopy(const std::string varname, butil::IOBuf* iobuf, + int k, const char* v, int64_t vlen, + bool in_cuda_pinned, void (*destroy)(void*), + void* user_data) { + if (vlen >= std::numeric_limits::max() || vlen < 0) { + LOG(FATAL) << "AppendZeroCopy varname:" << varname << ", vlen:" << vlen; + } + +#ifdef PADDLE_WITH_BRPC_RDMA + IOBufWriter::AppendRdmaZeroCopy(varname, iobuf, k, v, vlen, in_cuda_pinned, + destroy, user_data); +#else + IOBufWriter::AppendTCPZeroCopy(iobuf, k, v, vlen, in_cuda_pinned, destroy, + user_data); +#endif + } +}; + +void SerializeToIOBuf(const std::string& name, framework::Variable* var, + const platform::DeviceContext& ctx, VarMsg* request, + butil::IOBuf* iobuf, const std::string& out_varname, + bool var_is_not_stable, int trainer_id, + const std::string& table_name) { + std::unique_ptr payload; + + request->set_varname(name); + request->set_trainer_id(trainer_id); + // Note: normally the profiler is enabled in 1 trainer, hence only + // 1 trainer returns true for ShouldSendProfileState(). It tells PS + // servers the trainer's profiling state so that PS can follow the + // trainer. + if (platform::ShouldSendProfileState()) { + if (platform::IsProfileEnabled()) { + request->set_profile(platform::kEnableProfiler); + } else { + request->set_profile(platform::kDisableProfiler); + } + } + if (!out_varname.empty()) { + request->set_out_varname(out_varname); + } + if (!table_name.empty()) { + request->set_table_name(table_name); + } + if (var->IsType()) { + request->set_type(::sendrecv::LOD_TENSOR); + payload.reset(new TensorPayload(GetTensorPayload(var, ctx, request))); + } else if (var->IsType()) { + request->set_type(::sendrecv::SELECTED_ROWS); + payload.reset(new TensorPayload(GetSelectedRowsPayload(var, ctx, request))); +#ifdef PADDLE_WITH_CUDA + } else if (var->IsType()) { + request->set_type(::sendrecv::NCCL_ID); + const ncclUniqueId& uid = var->Get(); + // TODO(gongwb): use append_zero to avoid data copy. + IOBufWriter::Append(name, iobuf, + sendrecv::VariableMessage::kSerializedFieldNumber, + uid.internal, NCCL_UNIQUE_ID_BYTES); + return; +#endif + } else { + PADDLE_THROW("Serialize does not support type: %s", + typeid(var->Type()).name()); + } + + PADDLE_ENFORCE_NOT_NULL(payload); + + // FIXME(gongwb): it seems that can use zero copy. + if (var_is_not_stable) { + IOBufWriter::Append( + name, iobuf, ::sendrecv::VariableMessage::kSerializedFieldNumber, + static_cast(payload->ptr()), payload->memory_size()); + } else { + if (platform::is_gpu_place(ctx.GetPlace())) { +#ifdef PADDLE_WITH_CUDA + IOBufWriter::AppendZeroCopy( + name, iobuf, ::sendrecv::VariableMessage::kSerializedFieldNumber, + static_cast(payload->ptr()), payload->memory_size(), + true, SerializeDestroyCallback, static_cast(payload.get())); + payload.release(); +#endif + } else { + IOBufWriter::AppendZeroCopy( + name, iobuf, ::sendrecv::VariableMessage::kSerializedFieldNumber, + static_cast(payload->ptr()), payload->memory_size(), + false, SerializeDestroyCallback, static_cast(payload.get())); + payload.release(); + } + } + + if (var->IsType()) { + auto* slr = var->GetMutable(); + PADDLE_ENFORCE(VectorElemName(slr->rows()) == typeid(int64_t).name()); + size_t rows_memory_size = slr->rows().size() * sizeof(int64_t); + + IOBufWriter::Append(name, iobuf, + ::sendrecv::VariableMessage::kRowsFieldNumber, + reinterpret_cast(slr->rows().data()), + static_cast(rows_memory_size)); + } +} + +void DeserializeFromIOBuf(const ::sendrecv::VariableMessage& meta, + const butil::IOBuf& iobuf, + const platform::DeviceContext& ctx, + const framework::Scope* scope, + framework::Variable** var, int* trainer_id) { + operators::distributed::BRPCVariableResponse resp(scope, &ctx); + PADDLE_ENFORCE(resp.Parse(iobuf, meta) == 0, "parse iobuf to tensor error!"); + *var = resp.GetVar(); + *trainer_id = resp.GetTrainerId(); +} + +} // namespace distributed +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/distributed/brpc_sendrecvop_utils.h b/paddle/fluid/operators/distributed/brpc_sendrecvop_utils.h new file mode 100644 index 0000000000000..ffaf442224228 --- /dev/null +++ b/paddle/fluid/operators/distributed/brpc_sendrecvop_utils.h @@ -0,0 +1,49 @@ +/* Copyright (c) 2016 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 +#include +#include +#include + +#include "brpc/channel.h" +#include "paddle/fluid/framework/data_type.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/scope.h" +#include "paddle/fluid/framework/selected_rows.h" +#include "paddle/fluid/framework/tensor_util.h" +#include "paddle/fluid/framework/var_type.h" +#include "paddle/fluid/operators/distributed/send_recv.pb.h" +#include "paddle/fluid/operators/distributed/sendrecvop_utils.h" + +namespace paddle { +namespace operators { +namespace distributed { + +void SerializeToIOBuf(const std::string& name, framework::Variable* var, + const platform::DeviceContext& ctx, VarMsg* request, + butil::IOBuf* iobuf, const std::string& out_varname, + bool var_is_not_stable, const int trainer_id = 0, + const std::string& table_name = std::string()); + +void DeserializeFromIOBuf(const VarMsg& meta, const butil::IOBuf& iobuf, + const platform::DeviceContext& ctx, + const framework::Scope* scope, + framework::Variable** var, int* trainer_id); + +} // namespace distributed +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/distributed/brpc_serde_test.cc b/paddle/fluid/operators/distributed/brpc_serde_test.cc new file mode 100644 index 0000000000000..2a2dc72150a32 --- /dev/null +++ b/paddle/fluid/operators/distributed/brpc_serde_test.cc @@ -0,0 +1,175 @@ +/* Copyright (c) 2016 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 +#include +#include // NOLINT + +#include "brpc/channel.h" +#include "google/protobuf/text_format.h" +#include "gtest/gtest.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/tensor_util.h" +#include "paddle/fluid/framework/variable.h" +#include "paddle/fluid/operators/distributed/brpc_sendrecvop_utils.h" +#include "paddle/fluid/operators/distributed/brpc_variable_response.h" +#include "paddle/fluid/operators/distributed/sendrecvop_utils.h" +#include "paddle/fluid/operators/distributed/variable_response.h" +#include "paddle/fluid/operators/math/math_function.h" +#include "paddle/fluid/platform/place.h" +#include "paddle/fluid/string/printf.h" + +namespace framework = paddle::framework; +namespace platform = paddle::platform; +namespace operators = paddle::operators; +namespace math = paddle::operators::math; +namespace memory = paddle::memory; + +void RunSerdeTestSelectedRows(platform::Place place) { + platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); + auto& ctx = *pool.Get(place); + + butil::IOBuf iobuf; + sendrecv::VariableMessage msg; + int tensor_numel = 564 * 128; + + // serialize var to IOBuf + { + framework::Variable var; + auto* slr = var.GetMutable(); + slr->set_height(1000); + auto* tensor = slr->mutable_value(); + auto* rows = slr->mutable_rows(); + tensor->Resize(framework::make_ddim({564, 128})); + tensor->mutable_data(place); + math::set_constant(ctx, tensor, 32.7); + for (int i = 0; i < 564; ++i) rows->push_back(i); + + operators::distributed::SerializeToIOBuf("myvar", &var, ctx, &msg, &iobuf, + "", false); + } + + // desrialize + { + framework::Scope scope; + scope.Var("myvar"); + operators::distributed::BRPCVariableResponse resp(&scope, &ctx); + EXPECT_EQ(resp.Parse(iobuf, msg), 0); + + framework::Variable* var2 = resp.GetVar(); + + auto* slr2 = var2->GetMutable(); + auto* tensor2 = slr2->mutable_value(); + auto* rows2 = slr2->mutable_rows(); + float* tensor_data2 = nullptr; + framework::Tensor tmp_tensor; + + if (platform::is_gpu_place(ctx.GetPlace())) { + platform::CPUPlace cpu; + framework::TensorCopy(*tensor2, cpu, &tmp_tensor); + tensor_data2 = tmp_tensor.data(); + } else { + tensor_data2 = const_cast(tensor2->data()); + } + const int64_t* rows_data2 = rows2->data(); + + for (int i = 0; i < tensor_numel; ++i) { + EXPECT_FLOAT_EQ(tensor_data2[i], 32.7); + } + for (size_t i = 0; i < rows2->size(); ++i) { + EXPECT_EQ(rows_data2[i], static_cast(i)); + } + EXPECT_EQ(slr2->height(), 1000); + } +} + +void RunTestLodTensor(platform::Place place) { + platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); + auto& ctx = *pool.Get(place); + + // serialize var to ByteBuffer + butil::IOBuf iobuf; + sendrecv::VariableMessage msg; + int tensor_numel = 512 * 8 * 4 * 2; + { + framework::Variable var; + auto* tensor = var.GetMutable(); + tensor->Resize(framework::make_ddim({512, 8, 4, 2})); + framework::LoD lod; + lod.push_back(framework::Vector({1, 3, 8})); + tensor->set_lod(lod); + tensor->mutable_data(place); + math::set_constant(ctx, tensor, 31.9); + + operators::distributed::SerializeToIOBuf("myvar", &var, ctx, &msg, &iobuf, + "", false); + } + + // check sendrecv::VariableMessage meta data + { + EXPECT_EQ(msg.varname(), "myvar"); + EXPECT_EQ(msg.type(), 0); + EXPECT_EQ(msg.dims()[0], 512); + EXPECT_EQ(msg.dims()[1], 8); + EXPECT_EQ(msg.dims()[2], 4); + EXPECT_EQ(msg.dims()[3], 2); + EXPECT_EQ(msg.lod_level(), 1); + EXPECT_EQ(msg.lod(0).lod_data(0), 1); + EXPECT_EQ(msg.lod(0).lod_data(1), 3); + EXPECT_EQ(msg.lod(0).lod_data(2), 8); + } + + // deserialize + { + framework::Scope scope; + scope.Var("myvar"); + operators::distributed::BRPCVariableResponse resp(&scope, &ctx); + EXPECT_EQ(resp.Parse(iobuf, msg), 0); + + framework::Variable* var2 = resp.GetVar(); + + auto tensor2 = var2->Get(); + float* tensor_data2 = nullptr; + framework::Tensor tmp_tensor; + + if (platform::is_gpu_place(ctx.GetPlace())) { + platform::CPUPlace cpu; + framework::TensorCopy(tensor2, cpu, &tmp_tensor); + tensor_data2 = tmp_tensor.data(); + } else { + tensor_data2 = const_cast(tensor2.data()); + } + + for (int i = 0; i < tensor_numel; ++i) + EXPECT_FLOAT_EQ(tensor_data2[i], 31.9); + } +} + +TEST(LodTensor, Run) { + platform::CPUPlace place; + RunTestLodTensor(place); +#ifdef PADDLE_WITH_CUDA + platform::CUDAPlace gpu(0); + RunTestLodTensor(gpu); +#endif +} + +TEST(SelectedRows, Run) { + platform::CPUPlace place; + RunSerdeTestSelectedRows(place); +#ifdef PADDLE_WITH_CUDA + platform::CUDAPlace gpu; + RunSerdeTestSelectedRows(gpu); +#endif +} diff --git a/paddle/fluid/operators/distributed/brpc_server.cc b/paddle/fluid/operators/distributed/brpc_server.cc index 47a06dd0f378f..78d41aeac50a3 100644 --- a/paddle/fluid/operators/distributed/brpc_server.cc +++ b/paddle/fluid/operators/distributed/brpc_server.cc @@ -13,84 +13,287 @@ // limitations under the License. #include "paddle/fluid/operators/distributed/brpc_server.h" +#include "paddle/fluid/framework/threadpool.h" +#include "paddle/fluid/operators/distributed/brpc_sendrecvop_utils.h" +#include "paddle/fluid/operators/distributed/brpc_variable_response.h" #include "paddle/fluid/operators/distributed/request_handler.h" namespace sendrecv { -typedef std::unordered_map +namespace distributed = paddle::operators::distributed; + +typedef std::unordered_map HandlerMap; class BRPCServiceImpl : public SendRecvService { public: - explicit BRPCServiceImpl(const HandlerMap& rpc_call_map) - : request_send_h_(nullptr), - request_get_h_(nullptr), - request_prefetch_h_(nullptr) { - auto it = rpc_call_map.find(paddle::operators::distributed::kRequestSend); + explicit BRPCServiceImpl(const HandlerMap& rpc_call_map, + distributed::RPCServer* rpc_server) + : rpc_server_(rpc_server) { + VLOG(3) << "BRPCServiceImpl size: " << rpc_call_map.size(); + auto it = rpc_call_map.find(distributed::kRequestSend); if (it != rpc_call_map.end()) { request_send_h_ = it->second; + send_threads_.reset(new paddle::framework::ThreadPool( + rpc_server_->GetThreadNum(distributed::kRequestSend))); } - it = rpc_call_map.find(paddle::operators::distributed::kRequestSend); + it = rpc_call_map.find(distributed::kRequestGet); if (it != rpc_call_map.end()) { request_get_h_ = it->second; + get_threads_.reset(new paddle::framework::ThreadPool( + rpc_server_->GetThreadNum(distributed::kRequestGet))); } - it = rpc_call_map.find(paddle::operators::distributed::kRequestPrefetch); + it = rpc_call_map.find(distributed::kRequestPrefetch); if (it != rpc_call_map.end()) { request_prefetch_h_ = it->second; + prefetch_threads_.reset(new paddle::framework::ThreadPool( + rpc_server_->GetThreadNum(distributed::kRequestPrefetch))); + } + + it = rpc_call_map.find(distributed::kRequestCheckpoint); + if (it != rpc_call_map.end()) { + request_checkpoint_h_ = it->second; + checkpoint_notify_threads_.reset(new paddle::framework::ThreadPool( + rpc_server_->GetThreadNum(distributed::kRequestPrefetch))); + } + + it = rpc_call_map.find(distributed::kRequestGetMonomerVariable); + if (it != rpc_call_map.end()) { + request_get_monomer_handler_h_ = it->second; + } + + it = rpc_call_map.find(distributed::kRequestGetMonomerBarrier); + if (it != rpc_call_map.end()) { + request_get_monomer_barrier_handler_h_ = it->second; } } virtual ~BRPCServiceImpl() {} - void SendVariable(google::protobuf::RpcController* cntl_butil, const VariableMessage* request, VoidMessage* response, google::protobuf::Closure* done) override { + send_threads_->Run( + [=] { _SendVariable(cntl_butil, request, response, done); }); + } + + void _SendVariable(google::protobuf::RpcController* cntl_butil, + const VariableMessage* request, VoidMessage* response, + google::protobuf::Closure* done) { PADDLE_ENFORCE(request_send_h_ != nullptr, "RequestSend handler should be registed first!"); brpc::ClosureGuard done_guard(done); - - paddle::framework::Scope* local_scope = request_send_h_->scope(); - paddle::framework::Variable* outvar = nullptr; - paddle::framework::Variable* invar = nullptr; + brpc::Controller* cntl = static_cast(cntl_butil); std::string varname = request->varname(); + VLOG(3) << "RequestSend var_name:" << varname + << ", trainer_id:" << request->trainer_id() + << ", from:" << cntl->remote_side(); - if (!request_send_h_->sync_mode()) { - local_scope = &request_send_h_->scope()->NewScope(); - invar = local_scope->Var(varname); - } else { - invar = local_scope->FindVar(varname); - } + distributed::BRPCVariableResponse resp(request_send_h_->scope(), + request_send_h_->dev_ctx(), + !request_send_h_->sync_mode()); + PADDLE_ENFORCE(resp.Parse(cntl->request_attachment(), *request) == 0, + "parse iobuf to tensor error!"); - request_send_h_->Handle(varname, local_scope, invar, &outvar); + auto scope = resp.GetMutableLocalScope(); + auto invar = resp.GetVar(); + int trainer_id = request->trainer_id(); + paddle::framework::Variable* outvar = nullptr; - if (!request_send_h_->sync_mode()) { - request_send_h_->scope()->DeleteScope(local_scope); - } + request_send_h_->Handle(varname, scope, invar, &outvar, trainer_id); } void GetVariable(google::protobuf::RpcController* cntl_butil, const VariableMessage* request, VariableMessage* response, google::protobuf::Closure* done) override { + get_threads_->Run( + [=] { _GetVariable(cntl_butil, request, response, done); }); + } + + void _GetVariable(google::protobuf::RpcController* cntl_butil, + const VariableMessage* request, VariableMessage* response, + google::protobuf::Closure* done) { PADDLE_ENFORCE(request_get_h_ != nullptr, "RequestGet handler should be registed first!"); - } + brpc::ClosureGuard done_guard(done); + brpc::Controller* cntl = static_cast(cntl_butil); + + std::string varname = request->varname(); + VLOG(3) << "RequestGet varname:" << varname + << ", trainer_id:" << request->trainer_id() + << ", from:" << cntl->remote_side(); + + auto scope = request_get_h_->scope(); + auto invar = scope->FindVar(varname); + int trainer_id = request->trainer_id(); + paddle::framework::Variable* outvar = nullptr; + + request_get_h_->Handle(varname, scope, invar, &outvar, trainer_id); + + if (outvar) { + distributed::SerializeToIOBuf(varname, outvar, *request_get_h_->dev_ctx(), + response, &cntl->response_attachment(), "", + false); + } + } void PrefetchVariable(google::protobuf::RpcController* cntl_butil, const VariableMessage* request, VariableMessage* response, google::protobuf::Closure* done) override { + prefetch_threads_->Run( + [=] { _PrefetchVariable(cntl_butil, request, response, done); }); + } + + void _PrefetchVariable(google::protobuf::RpcController* cntl_butil, + const VariableMessage* request, + VariableMessage* response, + google::protobuf::Closure* done) { PADDLE_ENFORCE(request_prefetch_h_ != nullptr, "kRequestPrefetch handler should be registed first!"); + + brpc::ClosureGuard done_guard(done); + brpc::Controller* cntl = static_cast(cntl_butil); + + // prefetch process... + std::string in_var_name = request->varname(); + std::string out_var_name = request->out_varname(); + VLOG(3) << "RequestPrefetch, in_var_name: " << in_var_name + << ", out_var_name: " << out_var_name + << ", trainer_id:" << request->trainer_id() + << ", from:" << cntl->remote_side(); + + distributed::BRPCVariableResponse resp( + request_prefetch_h_->scope(), request_prefetch_h_->dev_ctx(), true); + + PADDLE_ENFORCE(resp.Parse(cntl->request_attachment(), *request) == 0, + "parse iobuf to tensor error!"); + + auto scope = resp.GetMutableLocalScope(); + auto invar = scope->FindVar(in_var_name); + std::string table_name = request->table_name(); + int trainer_id = request->trainer_id(); + paddle::framework::Variable* outvar = scope->Var(out_var_name); + + request_prefetch_h_->Handle(in_var_name, scope, invar, &outvar, trainer_id, + out_var_name, table_name); + + distributed::SerializeToIOBuf(out_var_name, outvar, + *request_prefetch_h_->dev_ctx(), response, + &cntl->response_attachment(), "", true); + } + + void CheckpointNotify(google::protobuf::RpcController* cntl_butil, + const VariableMessage* request, VoidMessage* response, + google::protobuf::Closure* done) override { + checkpoint_notify_threads_->Run( + [=] { _CheckpointNotify(cntl_butil, request, response, done); }); + } + + void _CheckpointNotify(google::protobuf::RpcController* cntl_butil, + const VariableMessage* request, VoidMessage* response, + google::protobuf::Closure* done) { + PADDLE_ENFORCE( + request_checkpoint_h_ != nullptr, + "kRequestCheckpointNotify handler should be registed first!"); + + brpc::ClosureGuard done_guard(done); + brpc::Controller* cntl = static_cast(cntl_butil); + + distributed::BRPCVariableResponse resp(request_checkpoint_h_->scope(), + request_checkpoint_h_->dev_ctx()); + + auto scope = resp.GetMutableLocalScope(); + + std::string checkpoint_notify = request->varname(); + std::string checkpoint_dir = request->out_varname(); + int trainer_id = request->trainer_id(); + + VLOG(4) << "RequestCheckpointNotify notify: " << checkpoint_notify + << ", dir: " << checkpoint_dir + << ", trainer_id:" << request->trainer_id() + << ", from:" << cntl->remote_side(); + + request_checkpoint_h_->Handle(checkpoint_notify, scope, nullptr, nullptr, + trainer_id, checkpoint_dir); + } + + void GetMonomerVariable(google::protobuf::RpcController* cntl_butil, + const VariableMessage* request, + VariableMessage* response, + google::protobuf::Closure* done) override { + PADDLE_ENFORCE( + request_get_monomer_handler_h_ != nullptr, + "kRequestGetMonomerVariable handler should be registed first!"); + + brpc::ClosureGuard done_guard(done); + brpc::Controller* cntl = static_cast(cntl_butil); + + // proc request. + std::string varname = request->varname(); + VLOG(3) << "GetMonomerVariable " << varname + << ", trainer_id:" << request->trainer_id() + << ", from:" << cntl->remote_side(); + + rpc_server_->WaitVarCond(varname); + distributed::MonomerHandle h = rpc_server_->GetMonomer(varname); + + auto scope = h.scope_; + auto invar = scope->FindVar(varname); + paddle::framework::Variable* outvar = nullptr; + + request_get_monomer_handler_h_->Handle(varname, scope, invar, &outvar, + request->trainer_id()); + + if (outvar) { + distributed::SerializeToIOBuf(varname, outvar, *h.dev_ctx_, response, + &cntl->response_attachment(), "", false); + } + } + + void GetMonomerBarrier(google::protobuf::RpcController* cntl_butil, + const VariableMessage* request, VoidMessage* response, + google::protobuf::Closure* done) override { + PADDLE_ENFORCE( + request_get_monomer_barrier_handler_h_ != nullptr, + "RequestGetMonomerBarrier handler should be registed first!"); + + brpc::ClosureGuard done_guard(done); + brpc::Controller* cntl = static_cast(cntl_butil); + + std::string varname = request->varname(); + VLOG(3) << "RequestGetMonomerBarrier var_name:" << varname + << ", trainer_id:" << request->trainer_id() + << ", from:" << cntl->remote_side(); + + rpc_server_->WaitVarCond(varname); + distributed::MonomerHandle h = rpc_server_->GetMonomer(varname); + + paddle::framework::Scope* scope = nullptr; + paddle::framework::Variable* invar = nullptr; + paddle::framework::Variable* outvar = nullptr; + + request_get_monomer_barrier_handler_h_->Handle( + varname, scope, invar, &outvar, request->trainer_id()); } private: - paddle::operators::distributed::RequestHandler* request_send_h_; - paddle::operators::distributed::RequestHandler* request_get_h_; - paddle::operators::distributed::RequestHandler* request_prefetch_h_; + distributed::RequestHandler* request_send_h_{nullptr}; + distributed::RequestHandler* request_get_h_{nullptr}; + distributed::RequestHandler* request_prefetch_h_{nullptr}; + distributed::RequestHandler* request_checkpoint_h_{nullptr}; + distributed::RequestHandler* request_get_monomer_handler_h_{nullptr}; + distributed::RequestHandler* request_get_monomer_barrier_handler_h_{nullptr}; + + distributed::RPCServer* rpc_server_{nullptr}; + + // FIXME(gongwb): brpc should support process one rpce use one threadpool. + std::unique_ptr send_threads_; + std::unique_ptr get_threads_; + std::unique_ptr prefetch_threads_; + std::unique_ptr checkpoint_notify_threads_; }; } // namespace sendrecv @@ -100,7 +303,7 @@ namespace distributed { void AsyncBRPCServer::StartServer() { // Instance of your service. - sendrecv::BRPCServiceImpl service_impl(rpc_call_map_); + sendrecv::BRPCServiceImpl service_impl(rpc_call_map_, this); // Add the service into server. Notice the second parameter, because the // service is put on stack, we don't want server to delete it, otherwise @@ -111,6 +314,9 @@ void AsyncBRPCServer::StartServer() { } brpc::ServerOptions options; +#ifdef PADDLE_WITH_BRPC_RDMA + options.use_rdma = true; +#endif options.idle_timeout_sec = idle_timeout_s_; options.max_concurrency = max_concurrency_; if (server_.Start(bind_address_.c_str(), &options) != 0) { @@ -133,10 +339,10 @@ void AsyncBRPCServer::StartServer() { void AsyncBRPCServer::ShutDownImpl() { server_.Stop(1000); } void AsyncBRPCServer::WaitServerReady() { - VLOG(30) << "AsyncGRPCServer is wait server ready"; + VLOG(3) << "AsyncGRPCServer is wait server ready"; std::unique_lock lock(this->mutex_ready_); condition_ready_.wait(lock, [=] { return this->ready_ == 1; }); - VLOG(30) << "AsyncGRPCServer WaitSeverReady"; + VLOG(3) << "AsyncGRPCServer WaitSeverReady"; } }; // namespace distributed diff --git a/paddle/fluid/operators/distributed/brpc_variable_response.cc b/paddle/fluid/operators/distributed/brpc_variable_response.cc new file mode 100644 index 0000000000000..75306d72334ab --- /dev/null +++ b/paddle/fluid/operators/distributed/brpc_variable_response.cc @@ -0,0 +1,73 @@ +// 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/operators/distributed/brpc_variable_response.h" +#include "paddle/fluid/operators/distributed/send_recv.pb.h" + +namespace paddle { +namespace operators { +namespace distributed { + +namespace pb = ::google::protobuf; +using vr = ::sendrecv::VariableMessage; + +int BRPCVariableResponse::Parse(Source* source) { + pb::io::ZeroCopyInputStream* input_stream = source->contents(); + pb::io::CodedInputStream input(input_stream); + input.SetTotalBytesLimit(INT_MAX, INT_MAX); + + while (1) { + unsigned int tag = 0; + if (!input.ReadLittleEndian32(&tag)) { + break; + } + + uint64_t num_bytes = 0; + if (!input.ReadLittleEndian64(&num_bytes)) { + break; + } + + int field = static_cast(tag); + int ret = field == 0 ? -1 : field; + switch (field) { + case vr::kSerializedFieldNumber: { + if (!ProcSerializedField(field, &input, num_bytes)) { + return ret; + } + break; + } + case vr::kRowsFieldNumber: { + PADDLE_ENFORCE((meta_.type() == sendrecv::SELECTED_ROWS || + meta_.type() == sendrecv::LOD_TENSOR) && + meta_.varname() != "", + "meta info should be got first!"); + + if (!CopySelectRowsData(&input, *dev_ctx_, num_bytes)) { + return ret; + } + break; + } + default: { + PADDLE_ENFORCE(false, "not surpported %u fieldnumber", field); + return ret; + } + } + } + + return 0; +} +} // namespace distributed +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/distributed/brpc_variable_response.h b/paddle/fluid/operators/distributed/brpc_variable_response.h new file mode 100644 index 0000000000000..b0b91a42a01c7 --- /dev/null +++ b/paddle/fluid/operators/distributed/brpc_variable_response.h @@ -0,0 +1,67 @@ +// 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 + +#include "brpc/channel.h" +#include "paddle/fluid/framework/data_type.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/scope.h" +#include "paddle/fluid/framework/selected_rows.h" +#include "paddle/fluid/framework/var_type.h" + +#include "paddle/fluid/operators/distributed/send_recv.pb.h" + +#include "google/protobuf/io/coded_stream.h" +#include "google/protobuf/io/zero_copy_stream.h" +#include "paddle/fluid/framework/tensor.h" +#include "paddle/fluid/operators/distributed/variable_response.h" + +namespace paddle { +namespace operators { +namespace distributed { + +class BRPCSourceWrapper : public Source { + public: + explicit BRPCSourceWrapper(const butil::IOBuf& iobuf) : source_(iobuf) {} + ::google::protobuf::io::ZeroCopyInputStream* contents() override { + return &source_; + } + + private: + butil::IOBufAsZeroCopyInputStream source_; +}; + +class BRPCVariableResponse : public VariableResponse { + public: + BRPCVariableResponse(const framework::Scope* scope, + const platform::DeviceContext* dev_ctx, + bool create_scope = false) + : VariableResponse(scope, dev_ctx, create_scope) {} + + virtual ~BRPCVariableResponse() {} + + // parse attachment from iobuf + int Parse(Source* source) override; + int Parse(const butil::IOBuf& iobuf, const sendrecv::VariableMessage& meta) { + BRPCSourceWrapper wrapper(iobuf); + return VariableResponse::Parse(&wrapper, meta); + } +}; + +}; // namespace distributed +}; // namespace operators +}; // namespace paddle diff --git a/paddle/fluid/operators/distributed/collective_client.cc b/paddle/fluid/operators/distributed/collective_client.cc new file mode 100644 index 0000000000000..6d3f534311136 --- /dev/null +++ b/paddle/fluid/operators/distributed/collective_client.cc @@ -0,0 +1,59 @@ +// 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 // NOLINT +#include +#include "gflags/gflags.h" + +#include "paddle/fluid/operators/distributed/collective_client.h" + +DECLARE_int32(rpc_deadline); + +namespace paddle { +namespace operators { +namespace distributed { +std::once_flag CollectiveClient::init_flag_; +std::unique_ptr CollectiveClient::client_(nullptr); + +bool CollectiveClient::Gather(const std::vector& remote_vars, + std::vector* dst, + const platform::DeviceContext& ctx, + framework::Scope* scope, int64_t time_out) { + for (auto r : remote_vars) { + VLOG(50) << "begin gather from ep:" << r.String(); + scope->Var(r.var_name_)->GetMutable(); + VarHandlePtr ptr = rpc_client_->AsyncGetMonomerVariable( + r.ep_, ctx, *scope, r.var_name_, time_out); + } + + rpc_client_->Wait(); + + for (auto r : remote_vars) { + auto select_rows = + scope->FindVar(r.var_name_)->GetMutable(); + dst->push_back(select_rows); + + VLOG(4) << "gather from ep:" << r.String() + << ", select_rows:" << GetSelectedRowsInfo(*select_rows); + + rpc_client_->AsyncGetMonomerBarrier(r.ep_, r.var_name_); + } + + rpc_client_->Wait(); + return true; +} + +} // namespace distributed +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/distributed/collective_client.h b/paddle/fluid/operators/distributed/collective_client.h new file mode 100644 index 0000000000000..53b03c531a2b8 --- /dev/null +++ b/paddle/fluid/operators/distributed/collective_client.h @@ -0,0 +1,93 @@ +// 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 // NOLINT +#include +#include +#include "gflags/gflags.h" + +#include "paddle/fluid/framework/data_type.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/scope.h" +#include "paddle/fluid/operators/detail/macros.h" +#include "paddle/fluid/operators/distributed/request_handler.h" + +DECLARE_int32(rpc_deadline); + +namespace paddle { +namespace operators { +namespace distributed { + +inline std::string GetSelectedRowsInfo(const framework::SelectedRows& slr) { + std::stringstream ss; + ss << ", height:" << slr.height() << ", rows:["; + for (unsigned int i = 0; i < slr.rows().size(); i++) { + if (i != slr.rows().size() - 1) { + ss << slr.rows()[i] << ","; + } else { + ss << slr.rows()[i]; + } + } + ss << "], dims:" << slr.value().dims(); + return ss.str(); +} + +struct RemoteVar { + std::string ep_; + std::string var_name_; + int trainer_id_{0}; + + std::string String() { + std::stringstream ss; + ss << "ep:" << ep_ << ", var_name:" << var_name_ + << ", trainer_id:" << trainer_id_; + + return ss.str(); + } +}; + +class CollectiveClient { + public: + CollectiveClient() { + rpc_client_.reset(new RPCCLIENT_T()); + rpc_client_->InitImpl(); + } + virtual ~CollectiveClient() {} + + // note this function will retain the rank order. + bool Gather(const std::vector& remote_vars, + std::vector* dst, + const platform::DeviceContext& ctx, framework::Scope* scope, + int64_t time_out = FLAGS_rpc_deadline); + + static CollectiveClient* GetInstance() { + std::call_once(init_flag_, [&]() { + if (client_.get() == nullptr) { + client_.reset(new CollectiveClient()); + } + }); + return client_.get(); + } + + private: + std::unique_ptr rpc_client_; + + static std::once_flag init_flag_; + static std::unique_ptr client_; +}; +} // namespace distributed +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/distributed/collective_server.cc b/paddle/fluid/operators/distributed/collective_server.cc new file mode 100644 index 0000000000000..c95652400c27a --- /dev/null +++ b/paddle/fluid/operators/distributed/collective_server.cc @@ -0,0 +1,74 @@ +/* Copyright (c) 2016 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 // for removing the port file +#include +#include +#include +#include // NOLINT +#include + +#include "paddle/fluid/operators/distributed/collective_server.h" + +DEFINE_int32(collective_get_thread_num, 5, "number of threads for rpc get"); + +namespace paddle { +namespace operators { +namespace distributed { + +std::once_flag CollectiveServer::init_flag_; +std::shared_ptr CollectiveServer::collective_server_(nullptr); + +CollectiveServer::CollectiveServer(const std::string& end_point, int fan_in) { + VLOG(1) << "Create colllective server:" << end_point << ", fan_in:" << fan_in; + rpc_server_.reset(new RPCSERVER_T(end_point, fan_in)); +} + +void CollectiveServer::Stop() { + rpc_server_->ShutDown(); + server_thread_->join(); + loop_thread_->join(); +} + +void CollectiveServer::StartServer() { + get_monomer_handler_.reset(new GetMonomerHandler()); + get_monomer_handler_->SetRPCServer(rpc_server_.get()); + + get_barrier_handler_.reset(new GetMonomerBarrierHandler()); + get_barrier_handler_->SetRPCServer(rpc_server_.get()); + + rpc_server_->RegisterRPC(distributed::kRequestGetMonomerVariable, + get_monomer_handler_.get(), + FLAGS_collective_get_thread_num); + rpc_server_->RegisterRPC(distributed::kRequestGetMonomerBarrier, + get_barrier_handler_.get(), 1); + + server_thread_.reset(new std::thread([&]() { rpc_server_->StartServer(); })); + rpc_server_->WaitServerReady(); + + loop_thread_.reset(new std::thread([&]() { + while (true) { + if (rpc_server_->IsExit()) { + LOG(WARNING) << "get exit!rpc_processor break!"; + break; + } + sleep(1); + } + VLOG(1) << "CollectiveServer loop_thread end"; + })); +} + +}; // namespace distributed +}; // namespace operators +}; // namespace paddle diff --git a/paddle/fluid/operators/distributed/collective_server.h b/paddle/fluid/operators/distributed/collective_server.h new file mode 100644 index 0000000000000..a23dc18b4de86 --- /dev/null +++ b/paddle/fluid/operators/distributed/collective_server.h @@ -0,0 +1,110 @@ +/* Copyright (c) 2016 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 +#include +#include +#include // NOLINT +#include +#include + +#include "gflags/gflags.h" + +#include "paddle/fluid/operators/detail/macros.h" +#include "paddle/fluid/operators/distributed/request_handler.h" +#include "paddle/fluid/operators/distributed/request_handler_impl.h" +#include "paddle/fluid/operators/distributed/rpc_server.h" + +namespace paddle { +namespace operators { +namespace distributed { + +class CollectiveServer; + +class GetMonomerHandler final : public RequestHandler { + public: + GetMonomerHandler() : RequestHandler(true) {} + virtual ~GetMonomerHandler() {} + bool Handle(const std::string& var_name, framework::Scope* scope, + framework::Variable* var, framework::Variable** outvar, + const int trainer_id, const std::string& out_var_name = "", + const std::string& table_name = "") override { + VLOG(50) << "GetMonomerHandler recv " << var_name; + + *outvar = scope->FindVar(var_name); + PADDLE_ENFORCE(outvar != nullptr, "%s not found", var_name); + + return true; + } +}; + +class GetMonomerBarrierHandler final : public RequestHandler { + public: + GetMonomerBarrierHandler() : RequestHandler(true) {} + virtual ~GetMonomerBarrierHandler() {} + bool Handle(const std::string& var_name, framework::Scope* scope, + framework::Variable* var, framework::Variable** outvar, + const int trainer_id, const std::string& out_var_name = "", + const std::string& table_name = "") override { + VLOG(50) << "GetMonomerHandler recv " << var_name; + + rpc_server_->IncreaseVarBarrier(var_name); + + return true; + } +}; + +class CollectiveServer final { + public: + explicit CollectiveServer(const std::string& end_point, int fan_in); + + virtual ~CollectiveServer() {} + + void StartServer(); + + static CollectiveServer* GetInstance(const std::string& end_point, + int fan_in) { + std::call_once(init_flag_, [&]() { + if (collective_server_.get() == nullptr) { + collective_server_.reset(new CollectiveServer(end_point, fan_in)); + collective_server_->StartServer(); + } + }); + + return collective_server_.get(); + } + + std::shared_ptr GetRPCServer() { return rpc_server_; } + + void Stop(); + + private: + std::unique_ptr get_monomer_handler_; + std::unique_ptr get_barrier_handler_; + + std::shared_ptr rpc_server_; + std::shared_ptr server_thread_; + std::shared_ptr loop_thread_; + + bool ready_{false}; + + static std::once_flag init_flag_; + static std::shared_ptr collective_server_; +}; + +}; // namespace distributed +}; // namespace operators +}; // namespace paddle diff --git a/paddle/fluid/operators/distributed/collective_server_test.cc b/paddle/fluid/operators/distributed/collective_server_test.cc new file mode 100644 index 0000000000000..0a9c69e393257 --- /dev/null +++ b/paddle/fluid/operators/distributed/collective_server_test.cc @@ -0,0 +1,115 @@ +/* Copyright (c) 2016 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 +#include +#include // NOLINT + +#include "gtest/gtest.h" +#include "paddle/fluid/framework/block_desc.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/framework/operator.h" + +#include "paddle/fluid/operators/detail/macros.h" +#include "paddle/fluid/operators/distributed/collective_client.h" +#include "paddle/fluid/operators/distributed/collective_server.h" +#include "paddle/fluid/operators/distributed/request_handler_impl.h" +#include "paddle/fluid/operators/math/math_function.h" + +namespace framework = paddle::framework; +namespace platform = paddle::platform; +namespace distributed = paddle::operators::distributed; + +std::unique_ptr StartServer( + const std::string& ep, int fan_in, framework::Scope* scope, + platform::DeviceContext* dev_ctx) { + distributed::CollectiveServer* server = + distributed::CollectiveServer::GetInstance(ep, fan_in); + + auto rpc_server = server->GetRPCServer(); + rpc_server->RegisterVar("var1", distributed::kRequestGetMonomerVariable, + scope, dev_ctx); + + std::cout << "StartServer return" << std::endl; + return std::unique_ptr(server); +} + +std::unique_ptr GenerateVars(platform::Place place) { + platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); + auto& ctx = *pool.Get(place); + + framework::Scope* scope = new framework::Scope(); + framework::Variable* var = scope->Var("var1"); + auto* slr = var->GetMutable(); + slr->set_height(1000); + + auto* tensor = slr->mutable_value(); + auto* rows = slr->mutable_rows(); + + tensor->Resize(framework::make_ddim({3, 5})); + tensor->mutable_data(place); + + paddle::operators::math::set_constant(ctx, tensor, 32.7); + for (int i = 0; i < 3; ++i) rows->push_back(i); + + std::cout << "src:" << distributed::GetSelectedRowsInfo(*slr); + + return std::unique_ptr(scope); +} + +void Gather(const std::vector& vars, + platform::DeviceContext* dev_ctx) { + distributed::CollectiveClient* client = + distributed::CollectiveClient::GetInstance(); + + framework::Scope* scope = new framework::Scope(); + framework::Variable* var = scope->Var("var1"); + var->GetMutable(); + + std::vector dst; + client->Gather(vars, &dst, *dev_ctx, scope); + std::cout << "dst:" << distributed::GetSelectedRowsInfo(*dst[0]); +} + +TEST(PREFETCH, GPU) { + platform::CUDAPlace place; + platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); + auto& ctx = *pool.Get(place); + + std::string ep = "127.0.0.1:7164"; + auto scope = GenerateVars(place); + + auto* v1 = scope->FindVar("var1"); + std::cout << "var1:" << v1 << std::endl; + + auto server = StartServer(ep, 2, scope.get(), &ctx); + auto rpc_server = server->GetRPCServer(); + + distributed::RemoteVar var; + var.ep_ = ep; + var.var_name_ = "var1"; + var.trainer_id_ = 0; + + std::vector vars{var}; + Gather(vars, &ctx); + Gather(vars, &ctx); + + std::cout << "begin WaitVarBarrier" << std::endl; + rpc_server->WaitVarBarrier("var1"); + rpc_server->ClearRegisteredVars(); + server->Stop(); + + scope.release(); + server.release(); +} diff --git a/paddle/fluid/operators/distributed/grpc_client.cc b/paddle/fluid/operators/distributed/grpc_client.cc index c28f86146d304..8c54159a41e33 100644 --- a/paddle/fluid/operators/distributed/grpc_client.cc +++ b/paddle/fluid/operators/distributed/grpc_client.cc @@ -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 limitations under the License. */ -#include +#include #include #include "glog/logging.h" // For VLOG @@ -20,17 +20,20 @@ limitations under the License. */ #include "paddle/fluid/operators/distributed/grpc_client.h" #include "paddle/fluid/operators/distributed/grpc_serde.h" #include "paddle/fluid/operators/distributed/request_handler.h" +#include "paddle/fluid/platform/port.h" #include "paddle/fluid/platform/profiler.h" +DECLARE_bool(rpc_disable_reuse_port); + namespace paddle { namespace operators { namespace distributed { -void GRPCClient::InitImpl() { InitEventLoop(); } - -void GRPCClient::InitEventLoop() { +void GRPCClient::InitImpl() { // start the client process thread // TODO(wuyi): can make this in a threadpool + PADDLE_ENFORCE(client_thread_ == nullptr, + "please not re init proceed thread"); client_thread_.reset(new std::thread(std::bind(&GRPCClient::Proceed, this))); } @@ -38,7 +41,7 @@ void GRPCClient::SendComplete() { std::unique_lock lk(completed_mutex_); if (!completed_) { for (auto& it : channels_) { - VLOG(30) << "send complete message to " << it.first; + VLOG(3) << "send complete message to " << it.first; this->AsyncSendComplete(it.first); } PADDLE_ENFORCE(this->Wait(), "internal grpc error"); @@ -81,7 +84,7 @@ VarHandlePtr GRPCClient::AsyncSendVar(const std::string& ep, ::grpc::ByteBuffer req; SerializeToByteBuffer(var_name_val, var, *p_ctx, &req, "", trainer_id_); - VLOG(30) << s->GetVarHandlePtr()->String() << " begin"; + VLOG(3) << s->GetVarHandlePtr()->String() << " begin"; // stub context s->response_call_back_ = nullptr; @@ -104,6 +107,7 @@ VarHandlePtr GRPCClient::AsyncSendVar(const std::string& ep, void ProcGetResponse(const VarHandle& var_h, const ::grpc::ByteBuffer& ret_msg) { + VLOG(100) << "ProcGetResponse"; framework::Variable* outvar = nullptr; // get response's trainer_id is not used int trainer_id; @@ -124,6 +128,24 @@ VarHandlePtr GRPCClient::AsyncGetVar(const std::string& ep, const framework::Scope& scope, const std::string& var_name, int64_t time_out) { + return _AsyncGetVar(ep, ctx, scope, var_name, + "/sendrecv.SendRecvService/GetVariable", time_out); +} + +VarHandlePtr GRPCClient::AsyncGetMonomerVariable( + const std::string& ep, const platform::DeviceContext& ctx, + const framework::Scope& scope, const std::string& var_name, + int64_t time_out) { + return _AsyncGetVar(ep, ctx, scope, var_name, + "/sendrecv.SendRecvService/GetMonomerVariable", time_out); +} + +VarHandlePtr GRPCClient::_AsyncGetVar(const std::string& ep, + const platform::DeviceContext& ctx, + const framework::Scope& scope, + const std::string& var_name, + const std::string& rpc_path, + int64_t time_out) { const platform::DeviceContext* p_ctx = &ctx; const std::string ep_val = ep; const std::string var_name_val = var_name; @@ -134,7 +156,7 @@ VarHandlePtr GRPCClient::AsyncGetVar(const std::string& ep, VarHandlePtr h(new VarHandle(ep, method, var_name_val, p_ctx, p_scope)); s->Prepare(h, time_out); - framework::AsyncIO([var_name_val, s, method, p_ctx, h, this] { + framework::AsyncIO([var_name_val, s, method, p_ctx, h, rpc_path, this] { // prepare input sendrecv::VariableMessage req; req.set_varname(var_name_val); @@ -142,15 +164,15 @@ VarHandlePtr GRPCClient::AsyncGetVar(const std::string& ep, ::grpc::ByteBuffer buf; RequestToByteBuffer(req, &buf); - VLOG(30) << s->GetVarHandlePtr()->String() << " begin"; + VLOG(3) << s->GetVarHandlePtr()->String() << " begin"; // stub context s->response_call_back_ = ProcGetResponse; platform::RecordRPCEvent record_event(method, p_ctx); - auto call = s->stub_g_.PrepareUnaryCall( - s->context_.get(), "/sendrecv.SendRecvService/GetVariable", buf, &cq_); + auto call = + s->stub_g_.PrepareUnaryCall(s->context_.get(), rpc_path, buf, &cq_); call->StartCall(); call->Finish(&s->reply_, &s->status_, reinterpret_cast(s)); @@ -169,11 +191,13 @@ VarHandlePtr GRPCClient::AsyncPrefetchVar(const std::string& ep, const framework::Scope& scope, const std::string& in_var_name, const std::string& out_var_name, + const std::string& table_name, int64_t time_out) { const platform::DeviceContext* p_ctx = &ctx; const std::string ep_val = ep; const std::string in_var_name_val = in_var_name; const std::string out_var_name_val = out_var_name; + const std::string table_name_val = table_name; const framework::Scope* p_scope = &scope; const auto ch = GetChannel(ep_val); GetProcessor* s = new GetProcessor(ch); @@ -184,13 +208,14 @@ VarHandlePtr GRPCClient::AsyncPrefetchVar(const std::string& ep, s->Prepare(h, time_out); framework::AsyncIO([in_var_name_val, out_var_name_val, ep_val, p_scope, p_ctx, - s, method, h, this] { + s, method, h, table_name_val, this] { auto* var = p_scope->FindVar(in_var_name_val); ::grpc::ByteBuffer req; - SerializeToByteBuffer(in_var_name_val, var, *p_ctx, &req, out_var_name_val); + SerializeToByteBuffer(in_var_name_val, var, *p_ctx, &req, out_var_name_val, + 0, table_name_val); - VLOG(30) << s->GetVarHandlePtr()->String() << " begin"; + VLOG(3) << s->GetVarHandlePtr()->String() << " begin"; // stub context s->response_call_back_ = ProcGetResponse; @@ -263,6 +288,33 @@ VarHandlePtr GRPCClient::AsyncSendFetchBarrier(const std::string& ep, return h; } +VarHandlePtr GRPCClient::AsyncGetMonomerBarrier(const std::string& ep, + const std::string& var_name, + int64_t time_out) { + const auto ch = GetChannel(ep); + BatchBarrierProcessor* s = new BatchBarrierProcessor(ch); + const std::string method = "SendMonomerFetchBarrierRPC"; + VarHandlePtr h(new VarHandle(ep, method, var_name, nullptr, nullptr)); + s->Prepare(h, time_out); + + VLOG(30) << s->GetVarHandlePtr()->String() << " begin"; + + sendrecv::VariableMessage req; + req.set_varname(var_name); + + platform::RecordRPCEvent record_event(method, nullptr); + + auto rpc = s->stub_->AsyncGetMonomerBarrier(s->context_.get(), req, &cq_); + rpc->Finish(&s->reply_, &s->status_, reinterpret_cast(s)); + req_count_++; + + if (UNLIKELY(platform::IsProfileEnabled())) { + h->Wait(); + } + + return h; +} + VarHandlePtr GRPCClient::AsyncSendComplete(const std::string& ep, int64_t time_out) { const auto ch = GetChannel(ep); @@ -328,18 +380,17 @@ void GRPCClient::Proceed() { void* tag = nullptr; bool ok = false; - VLOG(30) << "GRPCClient Proceed begin"; + VLOG(3) << "GRPCClient Proceed begin"; while (!stopped_ && cq_.Next(&tag, &ok)) { BaseProcessor* c = static_cast(tag); GPR_ASSERT(ok); PADDLE_ENFORCE(c); if (c->status_.ok()) { - VLOG(30) << c->GetVarHandlePtr()->String() << " process"; + VLOG(3) << c->GetVarHandlePtr()->String() << " process"; c->Process(); } else if (c->status_.error_code() == grpc::StatusCode::DEADLINE_EXCEEDED) { - // FIXME(gongwb): parse error_details? - LOG(ERROR) << c->GetVarHandlePtr()->String() + LOG(FATAL) << c->GetVarHandlePtr()->String() << " meets grpc error, error_code:" << c->status_.error_code() << " error_message:" << c->status_.error_message() << " error_details:" << c->status_.error_details(); @@ -370,7 +421,15 @@ void GRPCClient::Proceed() { sync_cond_.notify_all(); } } - VLOG(30) << "GRPCClient Proceed end"; + + // Last log message + // Avoid using VLOG() and LOG(): in the destructor of google::LogMessage() a + // static Mutex log_mutex is used for synchronization, which might have been + // destructed at this moment. + if (FLAGS_v >= 3) { + std::string msg("GRPCClient Proceed end"); + fwrite(msg.c_str(), msg.length(), 1, stdout); + } } std::shared_ptr GRPCClient::GetChannel(const std::string& ep) { @@ -383,6 +442,9 @@ std::shared_ptr GRPCClient::GetChannel(const std::string& ep) { // Channel configurations: grpc::ChannelArguments args; args.SetInt(GRPC_ARG_MAX_RECONNECT_BACKOFF_MS, 2000); + if (FLAGS_rpc_disable_reuse_port) { + args.SetInt(GRPC_ARG_ALLOW_REUSEPORT, 0); + } args.SetCompressionAlgorithm(GRPC_COMPRESS_NONE); args.SetMaxSendMessageSize(std::numeric_limits::max()); args.SetMaxReceiveMessageSize(std::numeric_limits::max()); diff --git a/paddle/fluid/operators/distributed/grpc_client.h b/paddle/fluid/operators/distributed/grpc_client.h index d8e9cee85bd73..01bf46cc313b4 100644 --- a/paddle/fluid/operators/distributed/grpc_client.h +++ b/paddle/fluid/operators/distributed/grpc_client.h @@ -189,18 +189,28 @@ class GRPCClient : public RPCClient { const std::string& var_name, int64_t time_out = FLAGS_rpc_deadline) override; + VarHandlePtr AsyncGetMonomerVariable( + const std::string& ep, const platform::DeviceContext& ctx, + const framework::Scope& scope, const std::string& var_name, + int64_t time_out = FLAGS_rpc_deadline) override; + VarHandlePtr AsyncPrefetchVar(const std::string& ep, const platform::DeviceContext& ctx, const framework::Scope& scope, const std::string& in_var_name, const std::string& out_var_name, + const std::string& table_name = "", int64_t time_out = FLAGS_rpc_deadline) override; VarHandlePtr AsyncSendBatchBarrier( const std::string& ep, int64_t time_out = FLAGS_rpc_deadline) override; - VarHandlePtr AsyncSendFetchBarrier( - const std::string& ep, int64_t time_out = FLAGS_rpc_deadline) override; + VarHandlePtr AsyncSendFetchBarrier(const std::string& ep, + int64_t time_out) override; + + VarHandlePtr AsyncGetMonomerBarrier( + const std::string& ep, const std::string& var_name, + int64_t time_out = FLAGS_rpc_deadline) override; VarHandlePtr AsyncCheckpointNotify( const std::string& ep, const std::string& dir, @@ -213,21 +223,22 @@ class GRPCClient : public RPCClient { void SendComplete() override; - protected: void InitImpl() override; private: - // InitEventLoop should only be called by Init() - void InitEventLoop(); - void Proceed(); std::shared_ptr GetChannel(const std::string& ep); + VarHandlePtr _AsyncGetVar(const std::string& ep, + const platform::DeviceContext& ctx, + const framework::Scope& scope, + const std::string& var_name, const std::string& rpc, + int64_t time_out); private: grpc::CompletionQueue cq_; std::unordered_map> channels_; - std::unique_ptr client_thread_; + std::unique_ptr client_thread_{nullptr}; // mutex for Wait client sync std::mutex sync_mutex_; diff --git a/paddle/fluid/operators/distributed/grpc_serde.cc b/paddle/fluid/operators/distributed/grpc_serde.cc index f27b70a5a3dd2..a9dea9cfd2eea 100644 --- a/paddle/fluid/operators/distributed/grpc_serde.cc +++ b/paddle/fluid/operators/distributed/grpc_serde.cc @@ -15,7 +15,7 @@ limitations under the License. */ #ifdef PADDLE_WITH_CUDA #include #endif -#include +#include #include // NOLINT #include "google/protobuf/io/coded_stream.h" @@ -26,23 +26,18 @@ limitations under the License. */ #include "paddle/fluid/operators/distributed/grpc_variable_response.h" #include "paddle/fluid/operators/distributed/proto_encoder_helper.h" #include "paddle/fluid/operators/distributed/sendrecvop_utils.h" +#include "paddle/fluid/platform/port.h" #include "paddle/fluid/platform/profiler.h" namespace paddle { namespace operators { namespace distributed { -static void SerializeDestroyCallback(void* payload) { - if (payload != nullptr) { - auto* shared_payload = reinterpret_cast(payload); - delete shared_payload; - } -} - void SerializeToByteBuffer(const std::string& name, framework::Variable* var, const platform::DeviceContext& ctx, ::grpc::ByteBuffer* msg, const std::string& out_name, - const int trainer_id) { + const int trainer_id, + const std::string& table_name) { platform::RecordRPCEvent record_event("serial", &ctx); VarMsg request; TensorPayload* payload = nullptr; @@ -63,6 +58,9 @@ void SerializeToByteBuffer(const std::string& name, framework::Variable* var, if (!out_name.empty()) { request.set_out_varname(out_name); } + if (!table_name.empty()) { + request.set_table_name(table_name); + } if (var->IsType()) { request.set_type(::sendrecv::LOD_TENSOR); payload = new TensorPayload(GetTensorPayload(var, ctx, &request)); @@ -105,6 +103,10 @@ void SerializeToByteBuffer(const std::string& name, framework::Variable* var, e.WriteVarlengthBeginning(VarMsg::kSerializedFieldNumber, payload->memory_size()); + if (payload->memory_size() >= std::numeric_limits::max()) { + LOG(FATAL) << "AppendZeroCopy varname:" << name + << ", vlen:" << payload->memory_size(); + } // steal reference of tensor data ::grpc::Slice slices[4]; // metadata, tensor, rows meta, rows int num_slices = 2; // only SelectedRows have rows buffer @@ -118,8 +120,10 @@ void SerializeToByteBuffer(const std::string& name, framework::Variable* var, if (var->IsType()) { auto* slr = var->GetMutable(); ProtoEncodeHelper e2(static_cast(buf), 128); - size_t rows_memory_size = - slr->rows().size() * framework::SizeOfType(typeid(int64_t)); + + PADDLE_ENFORCE(VectorElemName(slr->rows()) == typeid(int64_t).name()); + size_t rows_memory_size = slr->rows().size() * sizeof(int64_t); + e2.WriteVarlengthBeginning(VarMsg::kRowsFieldNumber, rows_memory_size); slices[2] = ::grpc::Slice(e2.size()); memcpy(const_cast(slices[2].begin()), e2.data(), e2.size()); diff --git a/paddle/fluid/operators/distributed/grpc_serde.h b/paddle/fluid/operators/distributed/grpc_serde.h index 7ec489e961630..16f5293b0eb41 100644 --- a/paddle/fluid/operators/distributed/grpc_serde.h +++ b/paddle/fluid/operators/distributed/grpc_serde.h @@ -13,7 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once -#include + #include #include #include @@ -25,6 +25,7 @@ limitations under the License. */ #include "paddle/fluid/framework/tensor_util.h" #include "paddle/fluid/framework/var_type.h" #include "paddle/fluid/operators/distributed/sendrecvop_utils.h" +#include "paddle/fluid/platform/port.h" #include "paddle/fluid/operators/distributed/send_recv.grpc.pb.h" #include "paddle/fluid/operators/distributed/send_recv.pb.h" @@ -39,7 +40,8 @@ void SerializeToByteBuffer(const std::string& name, framework::Variable* var, const platform::DeviceContext& ctx, ::grpc::ByteBuffer* msg, const std::string& out_varname = std::string(), - const int trainer_id = 0); + const int trainer_id = 0, + const std::string& table_name = std::string()); void DeserializeFromByteBuffer(const ::grpc::ByteBuffer& msg, const platform::DeviceContext& ctx, diff --git a/paddle/fluid/operators/distributed/grpc_serde_test.cc b/paddle/fluid/operators/distributed/grpc_serde_test.cc index 96ea05e74ed76..1936c2c623a77 100644 --- a/paddle/fluid/operators/distributed/grpc_serde_test.cc +++ b/paddle/fluid/operators/distributed/grpc_serde_test.cc @@ -130,7 +130,8 @@ void RunTestLodTensor(platform::Place place, int from_type = 0) { math::set_constant(ctx, tensor, 31.9); ::grpc::ByteBuffer msg; - operators::distributed::SerializeToByteBuffer("myvar", &var, ctx, &msg); + operators::distributed::SerializeToByteBuffer("myvar", &var, ctx, &msg, + "outvar", 0, "table_name"); EXPECT_GT(msg.Length(), static_cast(0)); // deserialize diff --git a/paddle/fluid/operators/distributed/grpc_server.cc b/paddle/fluid/operators/distributed/grpc_server.cc index ffd2b1707bea6..cda102e78d2de 100644 --- a/paddle/fluid/operators/distributed/grpc_server.cc +++ b/paddle/fluid/operators/distributed/grpc_server.cc @@ -20,6 +20,8 @@ limitations under the License. */ using ::grpc::ServerAsyncResponseWriter; +DECLARE_bool(rpc_disable_reuse_port); + namespace paddle { namespace operators { namespace distributed { @@ -98,7 +100,7 @@ class RequestSend final : public RequestBase { void Process() override { std::string varname = GetReqName(); - VLOG(40) << "RequestSend var_name:" << varname; + VLOG(4) << "RequestSend var_name:" << varname; auto scope = request_->GetMutableLocalScope(); auto invar = request_->GetVar(); @@ -135,7 +137,7 @@ class RequestGet final : public RequestBase { // proc request. std::string varname = request_.varname(); int trainer_id = request_.trainer_id(); - VLOG(40) << "RequestGet " << varname; + VLOG(4) << "RequestGet " << varname; auto scope = request_handler_->scope(); auto invar = scope->FindVar(varname); @@ -156,6 +158,98 @@ class RequestGet final : public RequestBase { ServerAsyncResponseWriter<::grpc::ByteBuffer> responder_; }; +class RequestGetMonomerVariable final : public RequestBase { + public: + explicit RequestGetMonomerVariable(GrpcService::AsyncService* service, + ::grpc::ServerCompletionQueue* cq, + RequestHandler* request_handler, + int req_id, RPCServer* rpc_server) + : RequestBase(service, cq, request_handler, req_id), + responder_(&ctx_), + rpc_server_(rpc_server) { + auto method_id = + static_cast(distributed::GrpcMethod::kGetMonomerVariable); + service_->RequestAsyncUnary( + method_id, &ctx_, &request_, &responder_, cq_, cq_, + reinterpret_cast(static_cast(req_id))); + } + + virtual ~RequestGetMonomerVariable() {} + + std::string GetReqName() override { return request_.varname(); } + + void Process() override { + // proc request. + std::string varname = request_.varname(); + + rpc_server_->WaitVarCond(varname); + MonomerHandle h = rpc_server_->GetMonomer(varname); + + auto scope = h.scope_; + auto invar = scope->FindVar(varname); + framework::Variable* outvar = nullptr; + + request_handler_->Handle(varname, scope, invar, &outvar, + request_.trainer_id()); + + if (outvar) { + SerializeToByteBuffer(varname, outvar, *h.dev_ctx_, &reply_); + } + Finish(reply_, &responder_); + } + + protected: + sendrecv::VariableMessage request_; + ::grpc::ByteBuffer reply_; + ServerAsyncResponseWriter<::grpc::ByteBuffer> responder_; + RPCServer* rpc_server_{nullptr}; +}; + +class RequestGetMonomerBarrier final : public RequestBase { + public: + explicit RequestGetMonomerBarrier(GrpcService::AsyncService* service, + ::grpc::ServerCompletionQueue* cq, + RequestHandler* request_handler, int req_id, + RPCServer* rpc_server) + : RequestBase(service, cq, request_handler, req_id), + responder_(&ctx_), + rpc_server_(rpc_server) { + auto method_id = + static_cast(distributed::GrpcMethod::kGetMonomerBarrier); + service_->RequestAsyncUnary( + method_id, &ctx_, &request_, &responder_, cq_, cq_, + reinterpret_cast(static_cast(req_id))); + } + + virtual ~RequestGetMonomerBarrier() {} + + std::string GetReqName() override { return request_.varname(); } + + void Process() override { + // proc request. + std::string varname = request_.varname(); + VLOG(4) << "RequestGetMonomerBarrier " << varname; + + rpc_server_->WaitVarCond(varname); + MonomerHandle h = rpc_server_->GetMonomer(varname); + + framework::Scope* scope = nullptr; + framework::Variable* invar = nullptr; + framework::Variable* outvar = nullptr; + + request_handler_->Handle(varname, scope, invar, &outvar, + request_.trainer_id()); + + Finish(reply_, &responder_); + } + + protected: + sendrecv::VariableMessage request_; + sendrecv::VoidMessage reply_; + ServerAsyncResponseWriter responder_; + RPCServer* rpc_server_{nullptr}; +}; + class RequestPrefetch final : public RequestBase { public: explicit RequestPrefetch(GrpcService::AsyncService* service, @@ -181,9 +275,10 @@ class RequestPrefetch final : public RequestBase { // prefetch process... std::string in_var_name = request_->Varname(); std::string out_var_name = request_->OutVarname(); + std::string table_name = request_->TableName(); int trainer_id = request_->GetTrainerId(); - VLOG(40) << "RequestPrefetch, in_var_name: " << in_var_name - << " out_var_name: " << out_var_name; + VLOG(4) << "RequestPrefetch, in_var_name: " << in_var_name + << " out_var_name: " << out_var_name; auto scope = request_->GetMutableLocalScope(); auto invar = scope->FindVar(in_var_name); @@ -191,7 +286,7 @@ class RequestPrefetch final : public RequestBase { framework::Variable* outvar = scope->Var(out_var_name); request_handler_->Handle(in_var_name, scope, invar, &outvar, trainer_id, - out_var_name); + out_var_name, table_name); SerializeToByteBuffer(out_var_name, outvar, *request_handler_->dev_ctx(), &reply_); @@ -231,8 +326,8 @@ class RequestCheckpointNotify final : public RequestBase { std::string checkpoint_dir = request_->OutVarname(); int trainer_id = request_->GetTrainerId(); - VLOG(40) << "RequestCheckpointNotify notify: " << checkpoint_notify - << ", dir: " << checkpoint_dir; + VLOG(4) << "RequestCheckpointNotify notify: " << checkpoint_notify + << ", dir: " << checkpoint_dir; request_handler_->Handle(checkpoint_notify, scope, nullptr, nullptr, trainer_id, checkpoint_dir); @@ -246,12 +341,26 @@ class RequestCheckpointNotify final : public RequestBase { }; void AsyncGRPCServer::WaitServerReady() { - VLOG(40) << "AsyncGRPCServer is wait server ready"; + VLOG(4) << "AsyncGRPCServer is waiting server ready"; std::unique_lock lock(this->mutex_ready_); condition_ready_.wait(lock, [=] { return this->ready_ == 1; }); - VLOG(40) << "AsyncGRPCServer WaitSeverReady"; + VLOG(4) << "AsyncGRPCServer WaitSeverReady"; } +// Define an option subclass in order to disable SO_REUSEPORT for the +// server socket. +// Come from: +// https://github.com/tensorflow/tensorflow/blob/master/tensorflow/core/distributed_runtime/rpc/grpc_server_lib.cc +class NoReusePortOption : public ::grpc::ServerBuilderOption { + public: + void UpdateArguments(::grpc::ChannelArguments* args) override { + args->SetInt(GRPC_ARG_ALLOW_REUSEPORT, 0); + } + + void UpdatePlugins(std::vector>* + plugins) override {} +}; + void AsyncGRPCServer::StartServer() { ::grpc::ServerBuilder builder; builder.AddListeningPort(bind_address_, ::grpc::InsecureServerCredentials(), @@ -259,6 +368,10 @@ void AsyncGRPCServer::StartServer() { builder.SetMaxSendMessageSize(std::numeric_limits::max()); builder.SetMaxReceiveMessageSize(std::numeric_limits::max()); + if (FLAGS_rpc_disable_reuse_port) { + builder.SetOption( + std::unique_ptr<::grpc::ServerBuilderOption>(new NoReusePortOption)); + } builder.RegisterService(&service_); for (auto t : rpc_call_map_) { @@ -282,15 +395,14 @@ void AsyncGRPCServer::StartServer() { reqs.reserve(kRequestBufSize); for (int i = 0; i < kRequestBufSize; i++) { - VLOG(60) << "TryToRegisterNewOne on RPC NAME: " << rpc_name - << " I: " << i; + VLOG(6) << "TryToRegisterNewOne on RPC NAME: " << rpc_name << " I: " << i; TryToRegisterNewOne(rpc_name, i); } for (int i = 0; i < threadnum; i++) { rpc_threads_[rpc_name].emplace_back(new std::thread(std::bind( &AsyncGRPCServer::HandleRequest, this, cq.get(), rpc_name, f))); - VLOG(40) << t.first << " creates threads!"; + VLOG(4) << t.first << " creates threads!"; } } @@ -307,7 +419,7 @@ void AsyncGRPCServer::StartServer() { auto& threads = t.second; for (size_t i = 0; i < threads.size(); ++i) { threads[i]->join(); - VLOG(40) << t.first << " threads ends!"; + VLOG(4) << t.first << " threads ends!"; } } } @@ -315,7 +427,7 @@ void AsyncGRPCServer::StartServer() { void AsyncGRPCServer::ShutdownQueue() { for (auto& t : rpc_cq_) { t.second->Shutdown(); - VLOG(40) << t.first << " queue shutdown!"; + VLOG(4) << t.first << " queue shutdown!"; } } @@ -324,7 +436,7 @@ void AsyncGRPCServer::ShutDownImpl() { is_shut_down_ = true; ShutdownQueue(); - VLOG(40) << "server_ shutdown!"; + VLOG(4) << "server_ shutdown!"; server_->Shutdown(); } @@ -332,12 +444,12 @@ void AsyncGRPCServer::TryToRegisterNewOne(const std::string& rpc_name, int req_id) { std::unique_lock lock(cq_mutex_); if (is_shut_down_) { - VLOG(40) << "shutdown, do not TryToRegisterNewSendOne"; + VLOG(4) << "shutdown, do not TryToRegisterNewSendOne"; return; } - VLOG(40) << "TryToRegisterNewOne on RPC NAME: " << rpc_name - << " REQ ID: " << req_id; + VLOG(4) << "TryToRegisterNewOne on RPC NAME: " << rpc_name + << " REQ ID: " << req_id; auto& reqs = rpc_reqs_[rpc_name]; auto& handler = rpc_call_map_[rpc_name]; @@ -348,6 +460,12 @@ void AsyncGRPCServer::TryToRegisterNewOne(const std::string& rpc_name, b = new RequestSend(&service_, cq.get(), handler, req_id); } else if (rpc_name == kRequestGet) { b = new RequestGet(&service_, cq.get(), handler, req_id); + } else if (rpc_name == kRequestGetMonomerVariable) { + b = new RequestGetMonomerVariable(&service_, cq.get(), handler, req_id, + this); + } else if (rpc_name == kRequestGetMonomerBarrier) { + b = new RequestGetMonomerBarrier(&service_, cq.get(), handler, req_id, + this); } else if (rpc_name == kRequestPrefetch) { b = new RequestPrefetch(&service_, cq.get(), handler, req_id); } else if (rpc_name == kRequestCheckpoint) { @@ -358,7 +476,7 @@ void AsyncGRPCServer::TryToRegisterNewOne(const std::string& rpc_name, reqs[req_id] = b; - VLOG(40) << "Create RequestSend status:" << b->Status(); + VLOG(4) << "TryToRegisterNewOne status:" << b->Status(); } void AsyncGRPCServer::HandleRequest( @@ -368,15 +486,15 @@ void AsyncGRPCServer::HandleRequest( bool ok = false; while (true) { - VLOG(40) << "HandleRequest " << rpc_name << " wait next"; + VLOG(4) << "HandleRequest " << rpc_name << " wait next"; if (!cq->Next(&tag, &ok)) { - VLOG(30) << "CompletionQueue " << rpc_name << " shutdown!"; + LOG(WARNING) << "CompletionQueue " << rpc_name << " shutdown!"; break; } int req_id = static_cast(reinterpret_cast(tag)); - VLOG(40) << "HandleRequest " << rpc_name << ", req_id:" << req_id - << " get next"; + VLOG(4) << "HandleRequest " << rpc_name << ", req_id:" << req_id + << " get next"; auto& reqs = rpc_reqs_[rpc_name]; RequestBase* base = nullptr; @@ -386,16 +504,15 @@ void AsyncGRPCServer::HandleRequest( base = reqs[req_id]; } - VLOG(30) << base->Status2String(rpc_name); + VLOG(3) << base->Status2String(rpc_name); // reference: // https://github.com/tensorflow/tensorflow/issues/5596 // https://groups.google.com/forum/#!topic/grpc-io/xftlRy-IQwM // https://groups.google.com/forum/#!topic/grpc-io/ywATt88Ef_I if (!ok) { - LOG(WARNING) << "completion queue:" << rpc_name - << " recv no regular event" - << " context:" << base->Status2String(rpc_name); + VLOG(4) << "completion queue:" << rpc_name << " recv no regular event" + << " context:" << base->Status2String(rpc_name); TryToRegisterNewOne(rpc_name, req_id); delete base; continue; diff --git a/paddle/fluid/operators/distributed/grpc_service.h b/paddle/fluid/operators/distributed/grpc_service.h index 9ae9a31a003cb..537429b5fe989 100644 --- a/paddle/fluid/operators/distributed/grpc_service.h +++ b/paddle/fluid/operators/distributed/grpc_service.h @@ -81,10 +81,12 @@ enum class GrpcMethod { kGetVariable, kPrefetchVariable, kCheckpointNotify, + kGetMonomerVariable, + kGetMonomerBarrier, }; static const int kGrpcNumMethods = - static_cast(GrpcMethod::kCheckpointNotify) + 1; + static_cast(GrpcMethod::kGetMonomerBarrier) + 1; inline const char* GrpcMethodName(GrpcMethod id) { switch (id) { @@ -92,6 +94,10 @@ inline const char* GrpcMethodName(GrpcMethod id) { return "/sendrecv.SendRecvService/SendVariable"; case GrpcMethod::kGetVariable: return "/sendrecv.SendRecvService/GetVariable"; + case GrpcMethod::kGetMonomerVariable: + return "/sendrecv.SendRecvService/GetMonomerVariable"; + case GrpcMethod::kGetMonomerBarrier: + return "/sendrecv.SendRecvService/GetMonomerBarrier"; case GrpcMethod::kPrefetchVariable: return "/sendrecv.SendRecvService/PrefetchVariable"; case GrpcMethod::kCheckpointNotify: diff --git a/paddle/fluid/operators/distributed/grpc_variable_response.cc b/paddle/fluid/operators/distributed/grpc_variable_response.cc index d6d219d4369ba..76ad02b0300a5 100644 --- a/paddle/fluid/operators/distributed/grpc_variable_response.cc +++ b/paddle/fluid/operators/distributed/grpc_variable_response.cc @@ -301,6 +301,20 @@ int GRPCVariableResponse::Parse(Source* source) { meta_.set_trainer_id(trainer_id); break; } + case sendrecv::VariableMessage::kTableNameFieldNumber: { + uint32_t length; + if ((wt != WIRETYPE_LENGTH_DELIMITED) || !input.ReadVarint32(&length)) { + return tag; + } + + std::string temp; + if (!input.ReadString(&temp, length)) { + return tag; + } + + meta_.set_table_name(temp); + break; + } default: { // Unknown tag, return unknown error. return -1; diff --git a/paddle/fluid/operators/distributed/parameter_prefetch.cc b/paddle/fluid/operators/distributed/parameter_prefetch.cc new file mode 100644 index 0000000000000..cf14538b1c284 --- /dev/null +++ b/paddle/fluid/operators/distributed/parameter_prefetch.cc @@ -0,0 +1,255 @@ +// 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 +#include +#include + +#include "paddle/fluid/operators/distributed/parameter_prefetch.h" + +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/scope.h" +#include "paddle/fluid/framework/selected_rows.h" +#include "paddle/fluid/framework/tensor.h" + +#include "paddle/fluid/operators/detail/macros.h" +#include "paddle/fluid/operators/distributed/rpc_client.h" +#include "paddle/fluid/operators/distributed/variable_response.h" +#include "paddle/fluid/operators/distributed_ops/send_recv_util.h" + +namespace paddle { +namespace operators { +namespace distributed { + +using Tensor = framework::Tensor; +using LoDTensor = framework::LoDTensor; +using SelectedRows = framework::SelectedRows; +using DDim = framework::DDim; + +static size_t GetSectionIndex(int64_t id, + const std::vector& abs_sections) { + for (size_t i = 1; i < abs_sections.size(); ++i) { + if (id < abs_sections[i]) { + return i - 1; + } + } + return abs_sections.size() - 1; +} + +static std::vector ToAbsoluteSection( + const std::vector& height_sections) { + std::vector abs_sections; + abs_sections.resize(height_sections.size()); + abs_sections[0] = 0; + for (size_t i = 1; i < height_sections.size(); ++i) { + abs_sections[i] = height_sections[i - 1] + abs_sections[i - 1]; + } + return abs_sections; +} + +static std::vector> SplitIds( + const std::vector& ids_vector, + const std::vector& height_section, framework::Scope* scope) { + std::set all_ids; + for (auto id : ids_vector) { + all_ids.insert(id); + } + + auto abs_sections = ToAbsoluteSection(height_section); + std::vector> splited_ids; + splited_ids.resize(height_section.size() + 1); + for (auto& id : all_ids) { + auto section_index = GetSectionIndex(id, abs_sections); + splited_ids[section_index].push_back(id - abs_sections[section_index]); + } + return splited_ids; +} + +static void SplitIdsIntoMultipleVarsBySection( + const std::vector& in_var_names, + const std::vector& height_section, + const std::vector>& splited_ids, + framework::Scope* scope) { + PADDLE_ENFORCE_EQ(in_var_names.size(), height_section.size(), ""); + + auto place = platform::CPUPlace(); + + for (size_t i = 0; i < in_var_names.size(); ++i) { + auto* id_tensor = + scope->Var(in_var_names[i])->GetMutable(); + auto& ids = splited_ids[i]; + if (!ids.empty()) { + auto* id_tensor_data = id_tensor->mutable_data( + framework::make_ddim({static_cast(ids.size()), 1}), place); + memcpy(id_tensor_data, ids.data(), sizeof(int64_t) * ids.size()); + } + } +} + +static void MergeMultipleVarsIntoOneBySection( + const std::string& id_name, const std::vector& ids_vector, + const std::string& out_name, const std::vector& out_var_names, + const std::vector& height_section, + const std::vector>& splited_ids, + const framework::ExecutionContext& context, framework::Scope* scope, + platform::DeviceContext* actual_ctx) { + PADDLE_ENFORCE_EQ(out_var_names.size(), height_section.size(), ""); + + auto cpu_place = platform::CPUPlace(); + + auto abs_sections = ToAbsoluteSection(height_section); + std::unordered_map> id_to_offset; + for (size_t i = 0; i < ids_vector.size(); ++i) { + id_to_offset[ids_vector[i]].push_back(i); + } + + auto& id_tensor = scope->FindVar(id_name)->Get(); + auto* out_tensor = + scope->FindVar(out_name)->GetMutable(); + auto* out_tensor_data = out_tensor->mutable_data(id_tensor.place()); + + bool is_on_cpu_place = true; + if (!platform::is_cpu_place(id_tensor.place())) { + is_on_cpu_place = false; + } + + for (size_t section_idx = 0; section_idx < out_var_names.size(); + ++section_idx) { + auto& ids_in_this_section = splited_ids[section_idx]; + if (!ids_in_this_section.empty()) { + auto& prefetch_out_var = + scope->Var(out_var_names[section_idx])->Get(); + const auto* out_var_data = prefetch_out_var.data(); + auto& dims = prefetch_out_var.dims(); + + PADDLE_ENFORCE_EQ(dims.size(), 2, ""); + PADDLE_ENFORCE_EQ(ids_in_this_section.size(), dims[0]); + + auto row_numel = dims[1]; + + for (size_t i = 0; i < dims[0]; ++i) { + auto id = ids_in_this_section[i]; + auto origin_id = id + abs_sections[section_idx]; + auto& offsets = id_to_offset[origin_id]; + for (auto& offset : offsets) { + // should support GPU tensor + if (is_on_cpu_place) { + memory::Copy(cpu_place, out_tensor_data + offset * row_numel, + cpu_place, out_var_data + i * row_numel, + sizeof(float) * row_numel); + } else { +#ifndef PADDLE_WITH_CUDA + PADDLE_THROW("paddle is not compiled with CUDA!"); +#else + auto stream = + static_cast(actual_ctx)->stream(); + memory::Copy(boost::get(id_tensor.place()), + out_tensor_data + offset * row_numel, cpu_place, + out_var_data + i * row_numel, + sizeof(float) * row_numel, stream); +#endif + } + } + } + } else { + VLOG(3) << "ids in this section is empty"; + } + } +} + +void prefetch(const std::string& id_name, const std::string& out_name, + const std::vector& table_names, + const std::vector& epmap, + const std::vector& height_sections, + const framework::ExecutionContext& context) { + auto& local_scope = context.scope().NewScope(); + + platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); + auto& cpu_ctx = *pool.Get(platform::CPUPlace()); + auto& actual_ctx = *pool.Get(context.GetPlace()); + + distributed::RPCClient* rpc_client = + distributed::RPCClient::GetInstance( + context.Attr("trainer_id")); + + std::vector in_var_names; + std::vector out_var_names; + for (size_t i = 0; i < epmap.size(); ++i) { + in_var_names.push_back(id_name + "@" + epmap[i]); + out_var_names.push_back(out_name + "@" + epmap[i]); + } + + auto& id_tensor = local_scope.FindVar(id_name)->Get(); + std::vector ids_vector; + if (platform::is_cpu_place(id_tensor.place())) { + auto* id_data = id_tensor.data(); + for (size_t i = 0; i < id_tensor.numel(); ++i) { + ids_vector.push_back(id_data[i]); + } + } else { +#ifndef PADDLE_WITH_CUDA + PADDLE_THROW("paddle is not compiled with CUDA!"); +#else + auto cpu_place = platform::CPUPlace(); + framework::Tensor cpu_tensor; + auto* cpu_tensor_data = + cpu_tensor.mutable_data(id_tensor.dims(), cpu_place); + auto stream = + static_cast(&actual_ctx)->stream(); + memory::Copy(cpu_place, cpu_tensor_data, + boost::get(id_tensor.place()), + id_tensor.data(), sizeof(int64_t) * id_tensor.numel(), + stream); + for (size_t i = 0; i < cpu_tensor.numel(); ++i) { + ids_vector.push_back(cpu_tensor_data[i]); + } +#endif + } + + auto splited_ids = SplitIds(ids_vector, height_sections, &local_scope); + SplitIdsIntoMultipleVarsBySection(in_var_names, height_sections, splited_ids, + &local_scope); + + // create output var in local scope + for (auto& name : out_var_names) { + local_scope.Var(name)->GetMutable(); + } + + std::vector rets; + for (size_t i = 0; i < in_var_names.size(); i++) { + if (NeedSend(local_scope, in_var_names[i])) { + VLOG(3) << "sending " << in_var_names[i] << " to " << epmap[i] + << " to get " << out_var_names[i] << " back"; + rets.push_back(rpc_client->AsyncPrefetchVar( + epmap[i], cpu_ctx, local_scope, in_var_names[i], out_var_names[i], + table_names[i])); + } else { + VLOG(3) << "don't send no-initialied variable: " << out_var_names[i]; + } + } + + for (size_t i = 0; i < rets.size(); i++) { + PADDLE_ENFORCE(rets[i]->Wait(), "internal error in RPCClient"); + } + + MergeMultipleVarsIntoOneBySection(id_name, ids_vector, out_name, + out_var_names, height_sections, splited_ids, + context, &local_scope, &actual_ctx); + + context.scope().DeleteScope(&local_scope); +} + +}; // namespace distributed +}; // namespace operators +}; // namespace paddle diff --git a/paddle/fluid/operators/distributed/parameter_prefetch.h b/paddle/fluid/operators/distributed/parameter_prefetch.h new file mode 100644 index 0000000000000..53b0fbfb51f60 --- /dev/null +++ b/paddle/fluid/operators/distributed/parameter_prefetch.h @@ -0,0 +1,34 @@ +// 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 +#include + +#include "paddle/fluid/framework/operator.h" + +namespace paddle { +namespace operators { +namespace distributed { + +void prefetch(const std::string& id_name, const std::string& out_name, + const std::vector& table_names, + const std::vector& epmap, + const std::vector& height_sections, + const framework::ExecutionContext& context); + +}; // namespace distributed +}; // namespace operators +}; // namespace paddle diff --git a/paddle/fluid/operators/distributed/request_handler.h b/paddle/fluid/operators/distributed/request_handler.h index 3bcc59a47ba5f..62b24f150b41e 100644 --- a/paddle/fluid/operators/distributed/request_handler.h +++ b/paddle/fluid/operators/distributed/request_handler.h @@ -37,6 +37,8 @@ namespace distributed { constexpr char kRequestSend[] = "RequestSend"; constexpr char kRequestGet[] = "RequestGet"; +constexpr char kRequestGetMonomerVariable[] = "RequestGetMonomerVariable"; +constexpr char kRequestGetMonomerBarrier[] = "RequestGetMonomerBarrier"; constexpr char kRequestPrefetch[] = "RequestPrefetch"; constexpr char kRequestCheckpoint[] = "RequestCheckpoint"; constexpr char kRequestPassBarrier[] = "RequestPassBarrier"; @@ -75,7 +77,7 @@ class VarHandle { wait_cond_.wait(lk, [this] { return status_ != kDefaultState; }); ret = status_; } - VLOG(70) << "VarHandle wait:" << ret; + VLOG(7) << "VarHandle wait:" << ret; return ret != kErrorState; } @@ -84,7 +86,7 @@ class VarHandle { std::unique_lock lk(sync_mutex_); status_ = ok ? kFinishState : kErrorState; } - VLOG(70) << "VarHandle finish:" << ok; + VLOG(7) << "VarHandle finish:" << ok; wait_cond_.notify_all(); } @@ -191,7 +193,8 @@ class RequestHandler { virtual bool Handle(const std::string& varname, framework::Scope* scope, framework::Variable* var, framework::Variable** outvar, const int trainer_id, - const std::string& out_var_name = "") = 0; + const std::string& out_var_name = "", + const std::string& table_name = "") = 0; protected: const bool sync_mode_; diff --git a/paddle/fluid/operators/distributed/request_handler_impl.cc b/paddle/fluid/operators/distributed/request_handler_impl.cc index dae56cc8436c2..9722f8c96e91d 100644 --- a/paddle/fluid/operators/distributed/request_handler_impl.cc +++ b/paddle/fluid/operators/distributed/request_handler_impl.cc @@ -12,6 +12,7 @@ // See the License for the specific language governing permissions and // limitations under the License. +#include "paddle/fluid/operators/distributed/request_handler_impl.h" #include #include #include @@ -20,7 +21,7 @@ #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/scope.h" #include "paddle/fluid/framework/selected_rows.h" -#include "paddle/fluid/operators/distributed/request_handler_impl.h" +#include "paddle/fluid/framework/variable_helper.h" #include "paddle/fluid/operators/distributed/rpc_server.h" #include "paddle/fluid/string/printf.h" @@ -37,20 +38,21 @@ bool RequestSendHandler::Handle(const std::string& varname, framework::Variable* invar, framework::Variable** outvar, const int trainer_id, - const std::string& out_var_name) { - VLOG(40) << "RequestSendHandler:" << varname; + const std::string& out_var_name, + const std::string& table_name) { + VLOG(4) << "RequestSendHandler:" << varname; // Sync if (varname == BATCH_BARRIER_MESSAGE) { - VLOG(30) << "sync: recv BATCH_BARRIER_MESSAGE"; + VLOG(3) << "sync: recv BATCH_BARRIER_MESSAGE"; rpc_server_->IncreaseBatchBarrier(kRequestSend); } else if (varname == COMPLETE_MESSAGE) { - VLOG(30) << "sync: recv complete message"; + VLOG(3) << "sync: recv complete message"; rpc_server_->Complete(); } else { // Async if (!sync_mode_) { - VLOG(30) << "async process var: " << varname; + VLOG(3) << "async process var: " << varname; try { executor_->RunPreparedContext((*grad_to_prepared_ctx_)[varname].get(), scope); @@ -61,7 +63,7 @@ bool RequestSendHandler::Handle(const std::string& varname, return true; } else { // sync rpc_server_->WaitCond(kRequestSend); - VLOG(30) << "sync: processing received var: " << varname; + VLOG(3) << "sync: processing received var: " << varname; if (invar == nullptr) { LOG(FATAL) << "sync: Can not find server side var: " << varname; @@ -77,11 +79,13 @@ bool RequestGetHandler::Handle(const std::string& varname, framework::Variable* invar, framework::Variable** outvar, const int trainer_id, - const std::string& out_var_name) { - VLOG(40) << "RequestGetHandler:" << varname; + const std::string& out_var_name, + const std::string& table_name) { + VLOG(4) << "RequestGetHandler:" << varname; + if (sync_mode_) { if (varname == FETCH_BARRIER_MESSAGE) { - VLOG(30) << "sync: recv fetch barrier message"; + VLOG(3) << "sync: recv fetch barrier message"; rpc_server_->IncreaseBatchBarrier(kRequestGet); } else { rpc_server_->WaitCond(kRequestGet); @@ -93,14 +97,13 @@ bool RequestGetHandler::Handle(const std::string& varname, // NOTE: the format is determined by distributed_transpiler.py std::string param_bak_name = string::Sprintf("%s.trainer_%d_bak", varname, trainer_id); - VLOG(30) << "getting " << param_bak_name << " trainer_id " - << trainer_id; + VLOG(3) << "getting " << param_bak_name << " trainer_id " << trainer_id; auto var = scope_->FindVar(varname); auto t_orig = var->Get(); auto param_bak = scope_->Var(param_bak_name); auto t = param_bak->GetMutable(); t->mutable_data(dev_ctx_->GetPlace(), t_orig.type()); - VLOG(30) << "copying " << varname << " to " << param_bak_name; + VLOG(3) << "copying " << varname << " to " << param_bak_name; framework::TensorCopy(t_orig, dev_ctx_->GetPlace(), t); } *outvar = scope_->FindVar(varname); @@ -114,14 +117,22 @@ bool RequestPrefetchHandler::Handle(const std::string& varname, framework::Variable* invar, framework::Variable** outvar, const int trainer_id, - const std::string& out_var_name) { - VLOG(40) << "RequestPrefetchHandler " << varname; - - auto var_desc = program_->Block(0).FindVar(out_var_name); - InitializeVariable(*outvar, var_desc->GetType()); - executor_->RunPreparedContext( - (*prefetch_var_name_to_prepared_ctx_)[varname].get(), scope); + const std::string& out_var_name, + const std::string& table_name) { + VLOG(4) << "RequestPrefetchHandler " << varname; + if (table_name.empty()) { + auto var_desc = program_->Block(0).FindVar(out_var_name); + InitializeVariable(*outvar, var_desc->GetType()); + executor_->RunPreparedContext( + (*prefetch_var_name_to_prepared_ctx_)[varname].get(), scope); + } else { + (*outvar)->GetMutable(); + auto lookup_table_op = + BuildLookupTableOp(table_name, varname, out_var_name); + paddle::platform::CPUPlace cpu_place; + lookup_table_op->Run(*scope, cpu_place); + } return true; } @@ -130,7 +141,8 @@ bool RequestCheckpointHandler::Handle(const std::string& varname, framework::Variable* invar, framework::Variable** outvar, const int trainer_id, - const std::string& out_var_name) { + const std::string& out_var_name, + const std::string& table_name) { PADDLE_ENFORCE( checkpoint_notify_id != -1, "when checkpoint_notify_id = -1, there should be no RPC invoke."); @@ -139,8 +151,8 @@ bool RequestCheckpointHandler::Handle(const std::string& varname, auto* lt_var = scope_->FindVar(LOOKUP_TABLE_PATH)->GetMutable(); lt_var->clear(); lt_var->append(out_var_name); - VLOG(40) << "RequestCheckpointHandler update var kLookupTablePath to: " - << out_var_name; + VLOG(4) << "RequestCheckpointHandler update var kLookupTablePath to: " + << out_var_name; executor_->RunPreparedContext(checkpoint_prepared_ctx_.get(), scope_); return true; } diff --git a/paddle/fluid/operators/distributed/request_handler_impl.h b/paddle/fluid/operators/distributed/request_handler_impl.h index c1afda9dd2445..5e0b25c5c2ce1 100644 --- a/paddle/fluid/operators/distributed/request_handler_impl.h +++ b/paddle/fluid/operators/distributed/request_handler_impl.h @@ -24,6 +24,7 @@ #include "paddle/fluid/framework/data_type.h" #include "paddle/fluid/framework/executor.h" #include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/program_desc.h" #include "paddle/fluid/framework/scope.h" #include "paddle/fluid/framework/selected_rows.h" @@ -43,8 +44,8 @@ class RequestSendHandler final : public RequestHandler { virtual ~RequestSendHandler() {} bool Handle(const std::string& varname, framework::Scope* scope, framework::Variable* var, framework::Variable** outvar, - const int trainer_id, - const std::string& out_var_name = "") override; + const int trainer_id, const std::string& out_var_name = "", + const std::string& table_name = "") override; private: bool enable_dc_asgd_; @@ -59,21 +60,44 @@ class RequestGetHandler final : public RequestHandler { virtual ~RequestGetHandler() {} bool Handle(const std::string& varname, framework::Scope* scope, framework::Variable* var, framework::Variable** outvar, - const int trainer_id, - const std::string& out_var_name = "") override; + const int trainer_id, const std::string& out_var_name = "", + const std::string& table_name = "") override; private: bool enable_dc_asgd_; }; +static inline void BuildVar(const std::string& param_name, + std::initializer_list arguments, + paddle::framework::proto::OpDesc::Var* var) { + var->set_parameter(param_name); + for (auto& arg_name : arguments) { + *var->mutable_arguments()->Add() = arg_name; + } +} + class RequestPrefetchHandler final : public RequestHandler { public: explicit RequestPrefetchHandler(bool sync_mode) : RequestHandler(sync_mode) {} virtual ~RequestPrefetchHandler() {} bool Handle(const std::string& varname, framework::Scope* scope, framework::Variable* var, framework::Variable** outvar, - const int trainer_id, - const std::string& out_var_name = "") override; + const int trainer_id, const std::string& out_var_name = "", + const std::string& table_name = "") override; + + private: + std::unique_ptr BuildLookupTableOp( + const std::string& table_name, const std::string& id_name, + const std::string& out_name) { + paddle::framework::proto::OpDesc op_desc; + op_desc.set_type("lookup_table"); + BuildVar("W", {table_name.data()}, op_desc.add_inputs()); + BuildVar("Ids", {id_name.data()}, op_desc.add_inputs()); + BuildVar("Out", {out_name.data()}, op_desc.add_outputs()); + + auto op = paddle::framework::OpRegistry::CreateOp(op_desc); + return op; + } }; class RequestCheckpointHandler final : public RequestHandler { @@ -85,8 +109,8 @@ class RequestCheckpointHandler final : public RequestHandler { virtual ~RequestCheckpointHandler() {} bool Handle(const std::string& varname, framework::Scope* scope, framework::Variable* var, framework::Variable** outvar, - const int trainer_id, - const std::string& out_var_name = "") override; + const int trainer_id, const std::string& out_var_name = "", + const std::string& table_name = "") override; private: int checkpoint_notify_id; diff --git a/paddle/fluid/operators/distributed/rpc_client.h b/paddle/fluid/operators/distributed/rpc_client.h index 1983802e49506..b668d869787a4 100644 --- a/paddle/fluid/operators/distributed/rpc_client.h +++ b/paddle/fluid/operators/distributed/rpc_client.h @@ -45,10 +45,15 @@ class RPCClient { const std::string& var_name, int64_t time_out = FLAGS_rpc_deadline) = 0; + virtual VarHandlePtr AsyncGetMonomerVariable( + const std::string& ep, const platform::DeviceContext& ctx, + const framework::Scope& scope, const std::string& var_name, + int64_t time_out = FLAGS_rpc_deadline) = 0; + virtual VarHandlePtr AsyncPrefetchVar( const std::string& ep, const platform::DeviceContext& ctx, const framework::Scope& scope, const std::string& in_var_name, - const std::string& out_var_name, + const std::string& out_var_name, const std::string& table_name = "", int64_t time_out = FLAGS_rpc_deadline) = 0; virtual VarHandlePtr AsyncSendBatchBarrier( @@ -57,6 +62,10 @@ class RPCClient { virtual VarHandlePtr AsyncSendFetchBarrier( const std::string& ep, int64_t time_out = FLAGS_rpc_deadline) = 0; + virtual VarHandlePtr AsyncGetMonomerBarrier( + const std::string& ep, const std::string& var_name, + int64_t time_out = FLAGS_rpc_deadline) = 0; + virtual VarHandlePtr AsyncCheckpointNotify( const std::string& ep, const std::string& dir, int64_t time_out = FLAGS_rpc_deadline) = 0; @@ -87,8 +96,9 @@ class RPCClient { } } - protected: virtual void InitImpl() {} + + protected: // each trainer have exact one trainer id, it should be static static int trainer_id_; diff --git a/paddle/fluid/operators/distributed/rpc_server.cc b/paddle/fluid/operators/distributed/rpc_server.cc index 4055091104f2f..122619d41b25d 100644 --- a/paddle/fluid/operators/distributed/rpc_server.cc +++ b/paddle/fluid/operators/distributed/rpc_server.cc @@ -39,7 +39,7 @@ void RPCServer::SavePort() const { port_file.open(file_path); port_file << selected_port_; port_file.close(); - VLOG(40) << "selected port written to " << file_path; + VLOG(4) << "selected port written to " << file_path; } void RPCServer::WaitBarrier(const std::string& rpc_name) { @@ -49,12 +49,12 @@ void RPCServer::WaitBarrier(const std::string& rpc_name) { exit_flag_.load()); }); - VLOG(30) << "batch_barrier_: " << rpc_name << " " - << barrier_counter_[rpc_name]; + VLOG(3) << "batch_barrier_: " << rpc_name << " " + << barrier_counter_[rpc_name]; } void RPCServer::IncreaseBatchBarrier(const std::string rpc_name) { - VLOG(40) << "RPCServer begin IncreaseBatchBarrier " << rpc_name; + VLOG(4) << "RPCServer begin IncreaseBatchBarrier " << rpc_name; int b = 0; std::unique_lock lock(mutex_); b = ++barrier_counter_[rpc_name]; @@ -71,7 +71,7 @@ void RPCServer::Complete() { client_num_--; need_reset_all_vars_ = true; - VLOG(40) << "decrease client_num to: " << client_num_; + VLOG(4) << "decrease client_num to: " << client_num_; if (cur_cond_.load() == rpc_cond_map_[kRequestGet]) { barrier_counter_[kRequestGet]--; } @@ -90,7 +90,7 @@ int RPCServer::GetClientNum() { } void RPCServer::ResetBarrierCounter() { - VLOG(30) << "RPCServer ResetBarrierCounter "; + VLOG(3) << "RPCServer ResetBarrierCounter "; std::unique_lock lock(mutex_); for (auto& t : barrier_counter_) { t.second = 0; @@ -105,12 +105,12 @@ void RPCServer::RegisterRPC(const std::string& rpc_name, static int cond = -1; rpc_cond_map_[rpc_name] = ++cond; - VLOG(40) << "RegisterRPC rpc_name:" << rpc_name << ", handler:" << handler - << ", cond:" << rpc_cond_map_[rpc_name]; + VLOG(4) << "RegisterRPC rpc_name:" << rpc_name << ", handler:" << handler + << ", cond:" << rpc_cond_map_[rpc_name]; } void RPCServer::SetCond(const std::string& rpc_name) { - VLOG(30) << "RPCServer SetCond " << rpc_name; + VLOG(3) << "RPCServer SetCond " << rpc_name; { std::unique_lock lock(mutex_); cur_cond_ = rpc_cond_map_[rpc_name]; @@ -120,7 +120,7 @@ void RPCServer::SetCond(const std::string& rpc_name) { } void RPCServer::WaitCond(const std::string& rpc_name) { - VLOG(40) << "RPCServer WaitCond " << rpc_name; + VLOG(4) << "RPCServer WaitCond " << rpc_name; int cond = 0; { std::unique_lock lock(mutex_); @@ -132,6 +132,96 @@ void RPCServer::WaitCond(const std::string& rpc_name) { lock, [=] { return (cur_cond_.load() == cond || exit_flag_.load()); }); } +void RPCServer::RegisterVar(const std::string& var_name, + const std::string& rpc_name, + framework::Scope* scope, + platform::DeviceContext* dev_ctx) { + MonomerHandle h; + h.var_name_ = var_name; + h.rpc_name_ = rpc_name; + h.scope_ = scope; + h.dev_ctx_ = dev_ctx; + + { + std::unique_lock lock(mutex_); + if (var_map_.find(var_name) != var_map_.end()) { + PADDLE_ENFORCE(false, "%s alreay in var_map", var_name); + } + var_map_[var_name] = h; + } + + rpc_cond_.notify_all(); + VLOG(4) << "RegisterVar context:" << h.String(); +} + +void RPCServer::IncreaseVarBarrier(const std::string& var_name) { + int b = 0; + MonomerHandle h; + { + std::unique_lock lock(mutex_); + b = ++var_map_[var_name].barrier_; + h = var_map_[var_name]; + } + + if (b >= client_num_) { + barrier_cond_.notify_all(); + } + + VLOG(4) << "IncreaseVarBarrier context:" << h.String(); +} + +void RPCServer::WaitVarBarrier(const std::string& var_name) { + VLOG(4) << "WaitBarrier var_name:" << var_name; + + std::unique_lock lock(mutex_); + barrier_cond_.wait(lock, [&]() { + return ((var_map_[var_name].barrier_ >= client_num_ && client_num_ != 0) || + exit_flag_.load()); + }); + + VLOG(4) << "WaitBarrier context: " << var_map_[var_name].String(); +} + +void RPCServer::SetVarCond(const std::string& var_name) { + VLOG(4) << "SetVarCond var_name:" << var_name; + { + std::unique_lock lock(mutex_); + if (var_map_.find(var_name) != var_map_.end()) { + rpc_cond_.notify_all(); + } + } +} + +void RPCServer::WaitVarCond(const std::string& var_name) { + VLOG(4) << "WaitVarCond var_name:" << var_name; + + std::unique_lock lock(mutex_); + rpc_cond_.wait(lock, [=] { + return (var_map_.find(var_name) != var_map_.end() || exit_flag_.load()); + }); + + VLOG(4) << "WaitVarCond var_name:" << var_name << " end"; +} + +MonomerHandle RPCServer::GetMonomer(const std::string& var_name) { + MonomerHandle h; + { + std::unique_lock lock(mutex_); + h = var_map_[var_name]; + } + + return h; +} + +void RPCServer::ClearRegisteredVars() { + std::unique_lock lock(mutex_); + var_map_.clear(); +} + +void RPCServer::ClearVar(const std::string& var_name) { + std::unique_lock lock(mutex_); + var_map_.erase(var_name); +} } // namespace distributed } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/distributed/rpc_server.h b/paddle/fluid/operators/distributed/rpc_server.h index c78c5007a7f26..8c7b7f1d7eeec 100644 --- a/paddle/fluid/operators/distributed/rpc_server.h +++ b/paddle/fluid/operators/distributed/rpc_server.h @@ -21,12 +21,30 @@ #include #include +#include "paddle/fluid/framework/scope.h" #include "paddle/fluid/operators/distributed/request_handler.h" +#include "paddle/fluid/platform/device_context.h" namespace paddle { namespace operators { namespace distributed { +struct MonomerHandle { + std::string var_name_; + std::string rpc_name_; + framework::Scope* scope_{nullptr}; + platform::DeviceContext* dev_ctx_{nullptr}; + int64_t barrier_{0}; + + std::string String() { + std::stringstream ss; + ss << "var_name:" << var_name_ << ", rpc_name:" << rpc_name_ + << ", scope:" << scope_ << ", dev_ctx:" << dev_ctx_ + << ", barrier_:" << barrier_; + return ss.str(); + } +}; + class RPCServer { public: explicit RPCServer(const std::string& address, int client_num) @@ -57,6 +75,10 @@ class RPCServer { void RegisterRPC(const std::string& rpc_name, RequestHandler* handler, int thread_num = 5); + int GetThreadNum(const std::string& rpc_name) { + return rpc_thread_num_[rpc_name]; + } + // Wait util all the clients have reached the barrier for one // rpc method. This function should be called in the // RequestHandler if you want to run the server/client in a @@ -67,6 +89,16 @@ class RPCServer { void WaitCond(const std::string& rpc_name); void IncreaseBatchBarrier(const std::string rpc_name); + void RegisterVar(const std::string& var_name, const std::string& rpc_name, + framework::Scope* scope, platform::DeviceContext* dev_ctx); + void IncreaseVarBarrier(const std::string& var_name); + void WaitVarBarrier(const std::string& var_name); + void SetVarCond(const std::string& var_name); + void WaitVarCond(const std::string& var_name); + void ClearRegisteredVars(); + void ClearVar(const std::string& var_name); + MonomerHandle GetMonomer(const std::string& var_name); + void Complete(); void ResetBarrierCounter(); @@ -95,6 +127,9 @@ class RPCServer { std::unordered_map rpc_call_map_; std::unordered_map rpc_thread_num_; friend class RequestHandler; + + // TODO(gongwb): use more cond to notify or wait; + std::unordered_map var_map_; }; }; // namespace distributed diff --git a/paddle/fluid/operators/distributed/send_recv.proto.in b/paddle/fluid/operators/distributed/send_recv.proto.in index 55820c980e813..2637619f304d2 100644 --- a/paddle/fluid/operators/distributed/send_recv.proto.in +++ b/paddle/fluid/operators/distributed/send_recv.proto.in @@ -28,6 +28,9 @@ service SendRecvService { rpc PrefetchVariable(VariableMessage) returns (VariableMessage) {} rpc CheckpointNotify(VariableMessage) returns (VoidMessage) {} + + rpc GetMonomerVariable(VariableMessage) returns (VariableMessage) {} + rpc GetMonomerBarrier(VariableMessage) returns (VoidMessage) {} } // VariableMessage is serialized paddle variable message. @@ -80,6 +83,7 @@ message VariableMessage { // when profile switches from 1 to 2. int64 profile = 11; int64 trainer_id = 12; + string table_name = 13; } message VoidMessage {} diff --git a/paddle/fluid/operators/distributed/sendrecvop_utils.cc b/paddle/fluid/operators/distributed/sendrecvop_utils.cc index 374fa680e3681..25e2f77fb74f2 100644 --- a/paddle/fluid/operators/distributed/sendrecvop_utils.cc +++ b/paddle/fluid/operators/distributed/sendrecvop_utils.cc @@ -15,12 +15,15 @@ limitations under the License. */ #ifdef PADDLE_WITH_CUDA #include #endif -#include #include // NOLINT #include "paddle/fluid/framework/data_type.h" +#include "paddle/fluid/operators/distributed/brpc_rdma_pool.h" #include "paddle/fluid/operators/distributed/sendrecvop_utils.h" #include "paddle/fluid/operators/distributed/variable_response.h" +#include "paddle/fluid/platform/port.h" + +DEFINE_bool(rpc_disable_reuse_port, false, "Disable SO_REUSEPORT or not."); namespace paddle { namespace operators { @@ -43,7 +46,6 @@ static TensorPayload GetCommunicationAllocationFromTensor( memory::Copy(cuda_pinned, result->ptr(), boost::get(tensor.place()), tensor.data(), copy_size, gpu_dev_ctx.stream()); - ctx.Wait(); return TensorPayload(result); #else @@ -59,8 +61,7 @@ TensorPayload GetTensorPayload(framework::Variable* var, auto tensor = var->Get(); // FIXME(wuyi): data types in send_recv.proto is copied from // framework.proto - request->set_data_type( - static_cast(framework::ToDataType(tensor.type()))); + request->set_data_type(static_cast(tensor.type())); for (auto& dim : framework::vectorize(tensor.dims())) { request->add_dims(dim); } @@ -81,8 +82,7 @@ TensorPayload GetSelectedRowsPayload(framework::Variable* var, const platform::DeviceContext& ctx, VarMsg* request) { auto* slr = var->GetMutable(); - request->set_data_type( - static_cast(framework::ToDataType(slr->value().type()))); + request->set_data_type(static_cast(slr->value().type())); request->set_lod_level(0); request->set_slr_height(slr->height()); diff --git a/paddle/fluid/operators/distributed/sendrecvop_utils.h b/paddle/fluid/operators/distributed/sendrecvop_utils.h index 480fc59c4281e..6a87178be5daa 100644 --- a/paddle/fluid/operators/distributed/sendrecvop_utils.h +++ b/paddle/fluid/operators/distributed/sendrecvop_utils.h @@ -13,9 +13,9 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once -#include #include #include +#include #include #include "paddle/fluid/framework/data_type.h" @@ -24,8 +24,8 @@ limitations under the License. */ #include "paddle/fluid/framework/selected_rows.h" #include "paddle/fluid/framework/tensor_util.h" #include "paddle/fluid/framework/var_type.h" - #include "paddle/fluid/operators/distributed/send_recv.pb.h" +#include "paddle/fluid/platform/port.h" namespace paddle { namespace operators { @@ -50,6 +50,13 @@ class TensorPayload final { size_t memory_size_; }; +inline void SerializeDestroyCallback(void* payload) { + if (payload != nullptr) { + auto* shared_payload = reinterpret_cast(payload); + delete shared_payload; + } +} + TensorPayload GetTensorPayload(framework::Variable* var, const platform::DeviceContext& ctx, VarMsg* request); @@ -58,23 +65,29 @@ TensorPayload GetSelectedRowsPayload(framework::Variable* var, const platform::DeviceContext& ctx, VarMsg* request); -inline std::type_index ToTypeIndex(sendrecv::VariableMessage::Type type) { +inline framework::proto::VarType::Type ToVarType( + sendrecv::VariableMessage::Type type) { switch (type) { case sendrecv::VariableMessage::FP32: - return typeid(float); // NOLINT + return framework::proto::VarType::FP32; // NOLINT case sendrecv::VariableMessage::FP64: - return typeid(double); // NOLINT + return framework::proto::VarType::FP64; // NOLINT case sendrecv::VariableMessage::INT32: - return typeid(int); // NOLINT + return framework::proto::VarType::INT32; // NOLINT case sendrecv::VariableMessage::INT64: - return typeid(int64_t); // NOLINT + return framework::proto::VarType::INT64; // NOLINT case sendrecv::VariableMessage::BOOL: - return typeid(bool); // NOLINT + return framework::proto::VarType::BOOL; // NOLINT default: PADDLE_THROW("Not support type %d", type); } } +template