Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into design/lite2

.pre-commit-config.yaml

@@ -42,12 +42,6 @@ repos:
         entry: bash ./tools/codestyle/pylint_pre_commit.hook
         language: system
         files: \.(py)$
--   repo: https://github.com/PaddlePaddle/pre-commit-golang
-    sha: 8337620115c25ff8333f1b1a493bd031049bd7c0
-    hooks:
-    -   id: go-fmt
-        types:
-        - go
 -   repo: local
     hooks:
     -   id: copyright_checker

AUTHORS.md

@@ -44,6 +44,7 @@
 | qingqing01 | Qing-Qing Dang |
 | reyoung | Yang Yu |
 | Sand3r- | Michal Gallus |
+| sfraczek | Sylwester Fraczek |
 | Superjom | Chun-Wei Yan |
 | tensor-tang | Jian Tang |
 | tianbingsz | Tian-Bing Xu |
@@ -54,6 +55,7 @@
 | wangyang59 | Yang Wang |
 | wangzhen-nlp | Zhen Wang |
 | wen-bo-yang | Wen-Bo Yang |
+| wojtuss | Wojciech Uss |
 | wwhu | Wei-Wei Hu |
 | xinghai-sun | Xing-Hai Sun |
 | Xreki | Yi-Qun Liu |

CMakeLists.txt

@@ -24,6 +24,8 @@ message(STATUS "CXX compiler: ${CMAKE_CXX_COMPILER}, version: "
         "${CMAKE_CXX_COMPILER_ID} ${CMAKE_CXX_COMPILER_VERSION}")
 message(STATUS "C compiler: ${CMAKE_C_COMPILER}, version: "
         "${CMAKE_C_COMPILER_ID} ${CMAKE_C_COMPILER_VERSION}")
+message(STATUS "AR tools: ${CMAKE_AR}")
+
 if(WIN32)
     set(CMAKE_SUPPRESS_REGENERATION ON)
     set(CMAKE_STATIC_LIBRARY_PREFIX lib)
@@ -54,25 +56,15 @@ option(WITH_NGRAPH      "Compile PaddlePaddle with nGraph support."     OFF)
 option(WITH_DSO         "Compile PaddlePaddle with dynamic linked CUDA" ON)
 option(WITH_TESTING     "Compile PaddlePaddle with unit testing"        OFF)
 option(WITH_PYTHON      "Compile PaddlePaddle with python interpreter"  ON)
-option(WITH_DOUBLE      "Compile PaddlePaddle with double precision"    OFF)
-option(WITH_RDMA        "Compile PaddlePaddle with RDMA support"        OFF)
-option(WITH_TIMER       "Compile PaddlePaddle with stats timer"         OFF)
 option(WITH_PROFILER    "Compile PaddlePaddle with GPU profiler and gperftools"        OFF)
 option(WITH_JEMALLOC    "Compile PaddlePaddle with jemalloc"            OFF)
-option(WITH_DOC         "Compile PaddlePaddle with documentation"       OFF)
 option(WITH_COVERAGE    "Compile PaddlePaddle with code coverage"       OFF)
 option(COVERALLS_UPLOAD "Package code coverage data to coveralls"       OFF)
-option(ON_TRAVIS        "Exclude special unit test on Travis CI"        OFF)
-option(WITH_FLUID_ONLY  "Compile PaddlePaddle fluid only"               OFF)
-option(WITH_GOLANG      "Compile PaddlePaddle with GOLANG"              OFF)
-option(GLIDE_INSTALL    "Download and install go dependencies "         ON)
 option(WITH_DISTRIBUTE  "Compile with distributed support"              OFF)
 option(WITH_PSLIB       "Compile with pslib support"                    OFF)
-option(USE_EIGEN_FOR_BLAS   "Use matrix multiplication in Eigen"        OFF)
-option(EIGEN_USE_THREADS "Compile with multi-threaded Eigen"            OFF)
-option(WITH_ARM_FP16    "Use half precision support on armv8.2-a cpu"   OFF)
 option(WITH_CONTRIB     "Compile the third-party contributation"        OFF)
 option(REPLACE_ENFORCE_GLOG "Replace PADDLE_ENFORCE with glog/CHECK for better debug." OFF)
+# TODO(Superjomn) Remove WITH_ANAKIN option if not needed latter.
 option(WITH_ANAKIN      "Compile with Anakin library"                   OFF)
 option(ANAKIN_BUILD_FAT_BIN "Build anakin cuda fat-bin lib for all device plantform, ignored when WITH_ANAKIN=OFF" OFF)
 option(ANAKIN_BUILD_CROSS_PLANTFORM "Build anakin lib for any nvidia device plantform. ignored when WITH_ANAKIN=OFF" ON)
@@ -83,6 +75,7 @@ option(WITH_INFERENCE_API_TEST   "Test fluid inference high-level api interface"
 option(WITH_SYSTEM_BLAS   "Use system blas library"           OFF)
 option(PY_VERSION       "Compile PaddlePaddle with python3 support"     ${PY_VERSION})
 option(WITH_FAST_MATH   "Make use of fast math library, might affect the precision to some extent" ON)
+option(WITH_WBAES       "Compile PaddlePaddle with WBAES support"       ON)
 
 # PY_VERSION
 if(NOT PY_VERSION)
@@ -105,8 +98,6 @@ endif()
 if (WIN32)
     set(WITH_DISTRIBUTE OFF CACHE STRING
             "Disable DISTRIBUTE when compiling for Windows" FORCE)
-    set(WITH_FLUID_ONLY ON CACHE STRING
-            "Enable FLUID_ONLY when compiling for Windows" FORCE)
 endif()
 
 set(THIRD_PARTY_PATH "${CMAKE_BINARY_DIR}/third_party" CACHE STRING
@@ -148,7 +139,6 @@ include(external/openblas)  # download, build, install openblas
 include(external/mkldnn)    # download, build, install mkldnn
 include(external/ngraph)    # download, build, install nGraph
 include(external/boost)     # download boost
-include(external/any)       # download libn::any
 include(external/eigen)     # download eigen3
 include(external/pybind11)  # download pybind11
 include(external/cares)
@@ -159,6 +149,7 @@ include(external/dlpack)
 include(external/snappy)    # download snappy
 include(external/snappystream) # download snappystream
 include(external/warpctc)   # download, build, install warpctc
+include(external/wbaes)     # download wbaes
 
 if (NOT WIN32)
 # there is no official support of nccl, cupti in windows
@@ -202,7 +193,14 @@ include(configure)          # add paddle env configuration
 if(WITH_GPU)
     include(cuda)
     include(tensorrt)
+    include(anakin_subgraph)
+endif()
+
+if(WITH_GPU AND NOT WIN32)
+    message(STATUS "add dgc lib.")
+    include(external/dgc)
 endif()
+
 if(WITH_MKL OR WITH_MKLML)
     include(external/anakin)
 elseif()
@@ -225,7 +223,6 @@ include(generic)            # simplify cmake module
 include(package)            # set paddle packages
 include(ccache)             # set ccache for compilation
 include(util)               # set unittest and link libs
-include(rdma)               # set rdma libraries
 include(version)            # set PADDLE_VERSION
 include(coveralls)          # set code coverage
 include(inference_lib)      # add paddle fluid inference libraries
@@ -233,38 +230,11 @@ include(inference_lib)      # add paddle fluid inference libraries
 
 include_directories("${PADDLE_SOURCE_DIR}")
 
-set(EXTERNAL_LIBS
-    gflags
-    glog
-    ${CBLAS_LIBRARIES}
-    protobuf
-    zlib
-    ${PYTHON_LIBRARIES}
-)
-
-if(WITH_PSLIB)
-    list(APPEND EXTERNAL_LIBS pslib)
-    list(APPEND EXTERNAL_LIBS pslib_brpc)
-    list(APPEND EXTERNAL_LIBS libmct)
-endif(WITH_PSLIB)
-
 if(WITH_AMD_GPU)
     find_package(HIP)
     include(hip)
 endif(WITH_AMD_GPU)
 
-if(WITH_MKLML)
-    list(APPEND EXTERNAL_LIBS ${MKLML_IOMP_LIB})
-endif()
-
-if(WITH_LIBXSMM)
-    list(APPEND EXTERNAL_LIBS ${LIBXSMM_LIBS})
-endif()
-
-if(WITH_MKLDNN)
-    list(APPEND EXTERNAL_LIBS ${MKLDNN_LIB})
-endif()
-
 set(PADDLE_PYTHON_BUILD_DIR "${CMAKE_CURRENT_BINARY_DIR}/python/build")
 
 set(CMAKE_CXX_FLAGS_RELWITHDEBINFO "-O3 -g -DNDEBUG")

CONTRIBUTING.md

@@ -156,7 +156,7 @@ python \
 
 This will enable VLOG messages generated by `buddy_allocator.{h,cc}` and in the verbose range of 0 to 3, so you will see above example VLOG message, which is in level 3.  This suggests that we output overall messages in lower verbose levels, so they display with higher probability.  When coding C++, please follow the verbose level convention as follows:
 
-- verbose level 1: [framework](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/framework)
-- verbose level 3: [operators](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators)
-- verbose level 5: [memory](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/memory), [platform](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/platform)
-- verbose level 7: [math](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/legacy/math)
+- verbose level 1: [framework](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid/framework)
+- verbose level 3: [operators](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid/operators)
+- verbose level 5: [memory](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid/memory), [platform](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid/platform)
+- verbose level 7: [math](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid/operators/math/)

Dockerfile

@@ -75,8 +75,9 @@ RUN curl -s -q https://glide.sh/get | sh
 #    and its size is only one-third of the official one.
 # 2. Manually add ~IPluginFactory() in IPluginFactory class of NvInfer.h, otherwise, it couldn't work in paddle.
 #    See https://github.com/PaddlePaddle/Paddle/issues/10129 for details.
-RUN wget -qO- http://paddlepaddledeps.cdn.bcebos.com/TensorRT-4.0.0.3.Ubuntu-16.04.4.x86_64-gnu.cuda-8.0.cudnn7.0.tar.gz | \
-    tar -xz -C /usr/local && \
+
+RUN wget -q https://paddlepaddledeps.cdn.bcebos.com/TensorRT-4.0.1.6-ubuntu14.04.x86_64-gnu.cuda.8.0.cudnn7.0.tar.gz --no-check-certificate && \
+    tar -zxf TensorRT-4.0.1.6-ubuntu14.04.x86_64-gnu.cuda.8.0.cudnn7.0.tar.gz -C /usr/local && \
     cp -rf /usr/local/TensorRT/include /usr && \
     cp -rf /usr/local/TensorRT/lib /usr
 

README.md

@@ -3,8 +3,8 @@
 English | [简体中文](./README_cn.md)
 
 [![Build Status](https://travis-ci.org/PaddlePaddle/Paddle.svg?branch=develop)](https://travis-ci.org/PaddlePaddle/Paddle)
-[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://paddlepaddle.org/documentation/docs/en/1.2/getstarted/index_en.html)
-[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/index.html)
+[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://paddlepaddle.org/documentation/docs/en/1.3/beginners_guide/index_en.html)
+[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://paddlepaddle.org/documentation/docs/zh/1.3/beginners_guide/index.html)
 [![Release](https://img.shields.io/github/release/PaddlePaddle/Paddle.svg)](https://github.com/PaddlePaddle/Paddle/releases)
 [![License](https://img.shields.io/badge/license-Apache%202-blue.svg)](LICENSE)
 
@@ -18,7 +18,7 @@ learning to many products at Baidu.
 Our vision is to enable deep learning for everyone via PaddlePaddle.
 Please refer to our [release announcement](https://github.com/PaddlePaddle/Paddle/releases) to track the latest feature of PaddlePaddle.
 
-### Latest PaddlePaddle Release: [Fluid 1.2.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.2)
+### Latest PaddlePaddle Release: [Fluid 1.3.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.3)
 ### Install Latest Stable Release:
 ```
 # Linux CPU
@@ -26,9 +26,9 @@ pip install paddlepaddle
 # Linux GPU cuda9cudnn7
 pip install paddlepaddle-gpu
 # Linux GPU cuda8cudnn7
-pip install paddlepaddle-gpu==1.2.0.post87
+pip install paddlepaddle-gpu==1.3.0.post87
 # Linux GPU cuda8cudnn5
-pip install paddlepaddle-gpu==1.2.0.post85
+pip install paddlepaddle-gpu==1.3.0.post85
 
 # For installation on other platform, refer to http://paddlepaddle.org/
 ```
@@ -75,26 +75,26 @@ pip install paddlepaddle-gpu==1.2.0.post85
 
 ## Installation
 
-It is recommended to read [this doc](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/install/index_cn.html) on our website.
+It is recommended to read [this doc](http://paddlepaddle.org/documentation/docs/en/1.3/beginners_guide/index_en.html) on our website.
 
 ## Documentation
 
-We provide [English](http://paddlepaddle.org/documentation/docs/en/1.2/getstarted/index_en.html) and
-[Chinese](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/index.html) documentation.
+We provide [English](http://paddlepaddle.org/documentation/docs/en/1.3/beginners_guide/index_en.html) and
+[Chinese](http://paddlepaddle.org/documentation/docs/zh/1.3/beginners_guide/index.html) documentation.
 
 - [Deep Learning 101](https://github.com/PaddlePaddle/book)
 
   You might want to start from this online interactive book that can run in a Jupyter Notebook.
 
-- [Distributed Training](http://paddlepaddle.org/documentation/docs/zh/1.2/user_guides/howto/training/cluster_howto.html)
+- [Distributed Training](http://paddlepaddle.org/documentation/docs/en/1.3/user_guides/howto/training/multi_node_en.html)
 
   You can run distributed training jobs on MPI clusters.
 
-- [Python API](http://paddlepaddle.org/documentation/docs/zh/1.2/api_cn/index_cn.html)
+- [Python API](http://paddlepaddle.org/documentation/docs/en/1.3/api/index_en.html)
 
    Our new API enables much shorter programs.
 
-- [How to Contribute](http://paddlepaddle.org/documentation/docs/zh/1.2/advanced_usage/development/contribute_to_paddle/index_cn.html)
+- [How to Contribute](http://paddlepaddle.org/documentation/docs/en/1.3/advanced_usage/development/contribute_to_paddle/index_en.html)
 
    We appreciate your contributions!
 

README_cn.md

@@ -3,8 +3,8 @@
 [English](./README.md) | 简体中文
 
 [![Build Status](https://travis-ci.org/PaddlePaddle/Paddle.svg?branch=develop)](https://travis-ci.org/PaddlePaddle/Paddle)
-[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://paddlepaddle.org/documentation/docs/en/1.2/getstarted/index_en.html)
-[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/index.html)
+[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://paddlepaddle.org/documentation/docs/en/1.3/beginners_guide/index_en.html)
+[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://paddlepaddle.org/documentation/docs/zh/1.3/beginners_guide/index.html)
 [![Release](https://img.shields.io/github/release/PaddlePaddle/Paddle.svg)](https://github.com/PaddlePaddle/Paddle/releases)
 [![License](https://img.shields.io/badge/license-Apache%202-blue.svg)](LICENSE)
 
@@ -16,7 +16,7 @@ PaddlePaddle (PArallel Distributed Deep LEarning) 是一个简单易用、高效
 
 跟进PaddlePaddle最新特性请参考我们的[版本说明](https://github.com/PaddlePaddle/Paddle/releases)
 
-### PaddlePaddle最新版本: [Fluid 1.2.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.2)
+### PaddlePaddle最新版本: [Fluid 1.3.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.3)
 ### 安装最新稳定版本:
 ```
 # Linux CPU
@@ -24,9 +24,9 @@ pip install paddlepaddle
 # Linux GPU cuda9cudnn7
 pip install paddlepaddle-gpu
 # Linux GPU cuda8cudnn7
-pip install paddlepaddle-gpu==1.2.0.post87
+pip install paddlepaddle-gpu==1.3.0.post87
 # Linux GPU cuda8cudnn5
-pip install paddlepaddle-gpu==1.2.0.post85
+pip install paddlepaddle-gpu==1.3.0.post85
 
 # 其他平台上的安装指引请参考 http://paddlepaddle.org/
 ```
@@ -57,26 +57,26 @@ pip install paddlepaddle-gpu==1.2.0.post85
 
 ## 安装
 
-推荐阅读官网上的[安装说明](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/install/index_cn.html)
+推荐阅读官网上的[安装说明](http://paddlepaddle.org/documentation/docs/zh/1.3/beginners_guide/install/index_cn.html)
 
 ## 文档
 
-我们提供[英文](http://paddlepaddle.org/documentation/docs/en/1.2/getstarted/index_en.html)和
-[中文](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/index.html) 文档
+我们提供[英文](http://paddlepaddle.org/documentation/docs/en/1.3/beginners_guide/index_en.html)和
+[中文](http://paddlepaddle.org/documentation/docs/zh/1.3/beginners_guide/index.html) 文档
 
 - [深度学习101](https://github.com/PaddlePaddle/book)
 
   或许您想从这个在线交互式书籍开始，可以在Jupyter Notebook中运行
 
-- [分布式训练](http://paddlepaddle.org/documentation/docs/zh/1.2/user_guides/howto/training/cluster_howto.html)
+- [分布式训练](http://paddlepaddle.org/documentation/docs/zh/1.3/user_guides/howto/training/multi_node.html)
 
   可以在MPI集群上运行分布式训练任务
 
-- [Python API](http://paddlepaddle.org/documentation/docs/zh/1.2/api_cn/index_cn.html)
+- [Python API](http://paddlepaddle.org/documentation/docs/zh/1.3/api_cn/index_cn.html)
 
    新的API支持代码更少更简洁的程序
 
-- [贡献方式](http://paddlepaddle.org/documentation/docs/zh/1.2/advanced_usage/development/contribute_to_paddle/index_cn.html)
+- [贡献方式](http://paddlepaddle.org/documentation/docs/zh/1.3/advanced_usage/development/contribute_to_paddle/index_cn.html)
 
    欢迎您的贡献!
 

benchmark/fluid/fluid_benchmark.py

@@ -179,7 +179,6 @@ def train_parallel(train_args, test_args, args, train_prog, test_prog,
     else:
         build_strategy.reduce_strategy = fluid.BuildStrategy(
         ).ReduceStrategy.AllReduce
-    build_strategy.fuse_broadcast_op = args.fuse_broadcast_op
 
     avg_loss = train_args[0]
 

cmake/anakin_subgraph.cmake

+if(NOT WITH_GPU)
+    return()
+endif()
+
+set(ANAKIN_ROOT "/usr" CACHE PATH "ANAKIN ROOT")
+find_path(ANAKIN_INCLUDE_DIR anakin_config.h
+    PATHS ${ANAKIN_ROOT} ${ANAKIN_ROOT}/include
+    $ENV{ANAKIN_ROOT} $ENV{ANAKIN_ROOT}/include
+    NO_DEFAULT_PATH
+)
+
+find_library(ANAKIN_LIBRARY NAMES libanakin_saber_common.so libanakin.so
+    PATHS ${ANAKIN_ROOT}
+    $ENV{ANAKIN_ROOT} $ENV{ANAKIN_ROOT}/lib
+    NO_DEFAULT_PATH
+    DOC "Path to ANAKIN library.")
+
+if(ANAKIN_INCLUDE_DIR AND ANAKIN_LIBRARY)
+  if(WITH_DSO)
+    set(ANAKIN_FOUND ON)
+  endif(WITH_DSO)
+else()
+    set(ANAKIN_FOUND OFF)
+endif()
+
+if(ANAKIN_FOUND)
+    message(STATUS "Current ANAKIN header is ${ANAKIN_INCLUDE_DIR}/anakin_config.h. ")
+    include_directories(${ANAKIN_ROOT}/include)
+    include_directories(${ANAKIN_ROOT}/include/saber)
+    link_directories(${ANAKIN_ROOT})
+    add_definitions(-DPADDLE_WITH_ANAKIN)
+endif()

cmake/configure.cmake

@@ -20,31 +20,10 @@ if(WITH_DSO)
     add_definitions(-DPADDLE_USE_DSO)
 endif(WITH_DSO)
 
-if(WITH_DOUBLE)
-    add_definitions(-DPADDLE_TYPE_DOUBLE)
-endif(WITH_DOUBLE)
-
-if(WITH_ARM_FP16)
-    add_definitions(-DPADDLE_ARM_FP16)
-    add_definitions("-march=armv8.2-a+fp16+simd")
-endif(WITH_ARM_FP16)
-
 if(WITH_TESTING)
     add_definitions(-DPADDLE_WITH_TESTING)
 endif(WITH_TESTING)
 
-if(NOT WITH_TIMER)
-    add_definitions(-DPADDLE_DISABLE_TIMER)
-endif(NOT WITH_TIMER)
-
-if(USE_EIGEN_FOR_BLAS)
-    add_definitions(-DPADDLE_USE_EIGEN_FOR_BLAS)
-endif(USE_EIGEN_FOR_BLAS)
-
-if(EIGEN_USE_THREADS)
-    add_definitions(-DEIGEN_USE_THREADS)
-endif(EIGEN_USE_THREADS)
-
 if(NOT WITH_PROFILER)
     add_definitions(-DPADDLE_DISABLE_PROFILER)
 endif(NOT WITH_PROFILER)
@@ -78,10 +57,6 @@ if(WIN32)
   endif(NOT MSVC)
 endif(WIN32)
 
-if(NOT WITH_GOLANG)
-    add_definitions(-DPADDLE_WITHOUT_GOLANG)
-endif(NOT WITH_GOLANG)
-
 if(WITH_PSLIB)
     add_definitions(-DPADDLE_WITH_PSLIB)
 endif()
@@ -171,55 +146,6 @@ if(WITH_DISTRIBUTE)
   add_definitions(-DPADDLE_WITH_DISTRIBUTE)
 endif()
 
-if(WITH_GOLANG)
-  # we need to symlink Paddle directory into GOPATH. If we
-  # don't do it and we have code that depends on Paddle, go
-  # get ./... will download a new Paddle repo from Github,
-  # without the changes in our current Paddle repo that we
-  # want to build.
-  set(GOPATH "${CMAKE_CURRENT_BINARY_DIR}/go")
-  file(MAKE_DIRECTORY ${GOPATH})
-  set(PADDLE_IN_GOPATH "${GOPATH}/src/github.com/PaddlePaddle/Paddle")
-  file(MAKE_DIRECTORY "${PADDLE_IN_GOPATH}")
-  set(PADDLE_GO_PATH "${CMAKE_SOURCE_DIR}/go")
-
-  add_custom_target(go_path)
-  add_custom_command(TARGET go_path
-    # Symlink Paddle directory into GOPATH
-    COMMAND mkdir -p ${PADDLE_IN_GOPATH}
-    COMMAND rm -rf ${PADDLE_IN_GOPATH}
-    COMMAND ln -sf ${CMAKE_SOURCE_DIR} ${PADDLE_IN_GOPATH}
-    # Automatically get all dependencies specified in the source code
-    # We can't run `go get -d ./...` for every target, because
-    # multiple `go get` can not run concurrently, but make need to be
-    # able to run with multiple jobs.
-    WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
-  )
-
-  if (GLIDE_INSTALL)
-    if(EXISTS $ENV{GOPATH}/bin/glide)
-      set(GLIDE "$ENV{GOPATH}/bin/glide")
-    else()
-      message(FATAL_ERROR "no glide executeble found: $ENV{GOPATH}/bin/glide")
-    endif()
-
-    # this command will only run when the file it depends is missing
-    # or has changed, or the output is missing.
-    add_custom_command(OUTPUT ${CMAKE_BINARY_DIR}/glide
-      COMMAND env GOPATH=${GOPATH} ${GLIDE} install
-      COMMAND touch ${CMAKE_BINARY_DIR}/glide
-      DEPENDS ${PADDLE_SOURCE_DIR}/go/glide.lock
-      WORKING_DIRECTORY "${PADDLE_IN_GOPATH}/go"
-      )
-
-    # depends on the custom command which outputs
-    # ${CMAKE_BINARY_DIR}/glide, the custom command does not need to
-    # run every time this target is built.
-    add_custom_target(go_vendor DEPENDS ${CMAKE_BINARY_DIR}/glide go_path)
-  endif()
-
-endif(WITH_GOLANG)
-
 if(WITH_GRPC)
     add_definitions(-DPADDLE_WITH_GRPC)
 endif(WITH_GRPC)
@@ -231,3 +157,7 @@ endif(WITH_BRPC_RDMA)
 if(ON_INFER)
     add_definitions(-DPADDLE_ON_INFERENCE)
 endif(ON_INFER)
+
+if(WITH_WBAES)
+    add_definitions(-DPADDLE_WITH_WBAES)
+endif(WITH_WBAES)

cmake/cuda.cmake

@@ -168,10 +168,7 @@ elseif (${CUDA_VERSION} LESS 11.0) # CUDA 10.x
 endif()
 
 include_directories(${CUDA_INCLUDE_DIRS})
-list(APPEND EXTERNAL_LIBS ${CUDA_LIBRARIES} ${CUDA_rt_LIBRARY})
 if(NOT WITH_DSO)
-    # TODO(panyx0718): CUPTI only allows DSO?
-    list(APPEND EXTERNAL_LIBS ${CUDNN_LIBRARY} ${CUPTI_LIBRARY} ${CUDA_CUBLAS_LIBRARIES} ${CUDA_curand_LIBRARY} ${NCCL_LIBRARY})
     if(WIN32)
       set_property(GLOBAL PROPERTY CUDA_MODULES ${CUDNN_LIBRARY} ${CUDA_CUBLAS_LIBRARIES} ${CUDA_curand_LIBRARY})
     endif(WIN32)

cmake/external/anakin.cmake

@@ -74,5 +74,3 @@ add_dependencies(anakin_shared extern_anakin)
 add_library(anakin_saber SHARED IMPORTED GLOBAL)
 set_property(TARGET anakin_saber PROPERTY IMPORTED_LOCATION ${ANAKIN_SABER_LIB})
 add_dependencies(anakin_saber extern_anakin)
-
-list(APPEND external_project_dependencies anakin_shared anakin_saber)

cmake/external/any.cmake

-INCLUDE(ExternalProject)
-
-SET(ANY_SOURCE_DIR ${THIRD_PARTY_PATH}/any)
-
-INCLUDE_DIRECTORIES(${ANY_SOURCE_DIR}/src/extern_lib_any)
-
-ExternalProject_Add(
-    extern_lib_any
-    ${EXTERNAL_PROJECT_LOG_ARGS}
-    GIT_REPOSITORY  "https://github.com/PaddlePaddle/any.git"
-    GIT_TAG         "15595d8324be9e8a9a80d9ae442fdd12bd66df5d"
-    PREFIX          ${ANY_SOURCE_DIR}
-    UPDATE_COMMAND  ""
-    CONFIGURE_COMMAND ""
-    BUILD_COMMAND     ""
-    INSTALL_COMMAND   ""
-    TEST_COMMAND      ""
-)
-
-if (${CMAKE_VERSION} VERSION_LESS "3.3.0")
-    set(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/lib_any_dummy.c)
-    file(WRITE ${dummyfile} "const char * dummy_any = \"${dummyfile}\";")
-    add_library(lib_any STATIC ${dummyfile})
-else()
-    add_library(lib_any INTERFACE)
-endif()
-
-add_dependencies(lib_any extern_lib_any)
-
-add_definitions(-DANY_IMPL_ANY_CAST_MOVEABLE)
-LIST(APPEND external_project_dependencies lib_any)

cmake/external/boost.cmake

@@ -24,7 +24,7 @@ set(BOOST_PROJECT       "extern_boost")
 # So we use 1.41.0 here.
 set(BOOST_VER           "1.41.0")
 set(BOOST_TAR "boost_1_41_0" CACHE STRING "" FORCE)
-set(BOOST_URL "http://paddlepaddledeps.cdn.bcebos.com/${BOOST_TAR}.tar.gz" CACHE STRING "" FORCE)
+set(BOOST_URL "http://paddlepaddledeps.bj.bcebos.com/${BOOST_TAR}.tar.gz" CACHE STRING "" FORCE)
 
 MESSAGE(STATUS "BOOST_TAR: ${BOOST_TAR}, BOOST_URL: ${BOOST_URL}")
 
@@ -57,5 +57,4 @@ else()
 endif()
 
 add_dependencies(boost ${BOOST_PROJECT})
-list(APPEND external_project_dependencies boost)
 set(Boost_INCLUDE_DIR ${BOOST_INCLUDE_DIR})

cmake/external/brpc.cmake

@@ -69,5 +69,3 @@ SET_PROPERTY(TARGET brpc PROPERTY IMPORTED_LOCATION ${BRPC_LIBRARIES})
 ADD_DEPENDENCIES(brpc extern_brpc)
 
 add_definitions(-DBRPC_WITH_GLOG)
-
-LIST(APPEND external_project_dependencies brpc)

cmake/external/cub.cmake

@@ -31,5 +31,3 @@ else()
 endif()
 
 add_dependencies(cub extern_cub)
-
-LIST(APPEND external_project_dependencies cub)

cmake/external/dgc.cmake

+# 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(ExternalProject)
+
+SET(DGC_SOURCES_DIR "${THIRD_PARTY_PATH}/dgc")
+SET(DGC_INSTALL_DIR "${THIRD_PARTY_PATH}/install/dgc")
+SET(DGC_INCLUDE_DIR "${DGC_INSTALL_DIR}/include" CACHE PATH "dgc include directory." FORCE)
+SET(DGC_LIBRARIES "${DGC_INSTALL_DIR}/lib/libdgc.a" CACHE FILEPATH "dgc library." FORCE)
+INCLUDE_DIRECTORIES(${DGC_INCLUDE_DIR})
+
+ExternalProject_Add(
+    extern_dgc
+    ${EXTERNAL_PROJECT_LOG_ARGS}
+    GIT_REPOSITORY "https://github.com/PaddlePaddle/Fleet"
+    GIT_TAG "2d04dc3800cdd0601f1b65d547dabcc60b0cf9dc"
+    SOURCE_DIR "${DGC_SOURCES_DIR}"
+    CONFIGURE_COMMAND ""
+    BUILD_COMMAND cd collective && make -j
+    INSTALL_COMMAND mkdir -p ${DGC_INSTALL_DIR}/lib/  ${DGC_INCLUDE_DIR}/dgc
+        && cp ${DGC_SOURCES_DIR}/collective/build/lib/libdgc.a ${DGC_LIBRARIES}
+        && cp ${DGC_SOURCES_DIR}/collective/build/include/dgc.h ${DGC_INCLUDE_DIR}/dgc/
+    BUILD_IN_SOURCE 1
+)
+
+ADD_LIBRARY(dgc STATIC IMPORTED GLOBAL)
+SET_PROPERTY(TARGET dgc PROPERTY IMPORTED_LOCATION ${DGC_LIBRARIES})
+ADD_DEPENDENCIES(dgc extern_dgc)
+
+LIST(APPEND external_project_dependencies dgc)
+

cmake/external/dlpack.cmake

@@ -27,5 +27,3 @@ else()
 endif()
 
 add_dependencies(dlpack extern_dlpack)
-
-LIST(APPEND external_project_dependencies dlpack)

cmake/external/eigen.cmake

@@ -52,5 +52,3 @@ else()
 endif()
 
 add_dependencies(eigen3 extern_eigen3)
-
-LIST(APPEND external_project_dependencies eigen3)

cmake/external/gflags.cmake

@@ -61,8 +61,6 @@ ADD_LIBRARY(gflags STATIC IMPORTED GLOBAL)
 SET_PROPERTY(TARGET gflags PROPERTY IMPORTED_LOCATION ${GFLAGS_LIBRARIES})
 ADD_DEPENDENCIES(gflags extern_gflags)
 
-LIST(APPEND external_project_dependencies gflags)
-
 # On Windows (including MinGW), the Shlwapi library is used by gflags if available.
 if (WIN32)
   include(CheckIncludeFileCXX)

cmake/external/glog.cmake

@@ -72,5 +72,3 @@ ADD_LIBRARY(glog STATIC IMPORTED GLOBAL)
 SET_PROPERTY(TARGET glog PROPERTY IMPORTED_LOCATION ${GLOG_LIBRARIES})
 ADD_DEPENDENCIES(glog extern_glog gflags)
 LINK_LIBRARIES(glog gflags)
-
-LIST(APPEND external_project_dependencies glog)

cmake/external/grpc.cmake

@@ -44,7 +44,7 @@ ExternalProject_Add(
     # 3. keep only zlib, cares, protobuf, boringssl under "third_party",
     #    checkout and clean other dirs under third_party
     # 4. remove .git, and package the directory.
-    URL "http://paddlepaddledeps.cdn.bcebos.com/grpc-v1.10.x.tar.gz"
+    URL "http://paddlepaddledeps.bj.bcebos.com/grpc-v1.10.x.tar.gz"
     URL_MD5  "1f268a2aff6759839dccd256adcc91cf"
     PREFIX          ${GRPC_SOURCES_DIR}
     UPDATE_COMMAND  ""

cmake/external/gtest.cmake

@@ -79,5 +79,4 @@ IF(WITH_TESTING OR (WITH_DISTRIBUTE AND NOT WITH_GRPC))
     SET_PROPERTY(TARGET gtest_main PROPERTY IMPORTED_LOCATION ${GTEST_MAIN_LIBRARIES})
     ADD_DEPENDENCIES(gtest_main extern_gtest)
 
-    LIST(APPEND external_project_dependencies gtest gtest_main)
 ENDIF(WITH_TESTING OR (WITH_DISTRIBUTE AND NOT WITH_GRPC))

cmake/external/leveldb.cmake

@@ -39,6 +39,3 @@ ADD_DEPENDENCIES(extern_leveldb snappy)
 ADD_LIBRARY(leveldb STATIC IMPORTED GLOBAL)
 SET_PROPERTY(TARGET leveldb PROPERTY IMPORTED_LOCATION ${LEVELDB_LIBRARIES})
 ADD_DEPENDENCIES(leveldb extern_leveldb)
-
-LIST(APPEND external_project_dependencies leveldb)
-

cmake/external/libmct.cmake

@@ -72,7 +72,4 @@ else()
     add_library(libmct INTERFACE)
 endif()
 
-#ADD_LIBRARY(libmct SHARED IMPORTED GLOBAL)
 ADD_DEPENDENCIES(libmct ${LIBMCT_PROJECT})
-LIST(APPEND external_project_dependencies libmct)
-

cmake/external/libxsmm.cmake

@@ -53,5 +53,3 @@ MESSAGE(STATUS "Libxsmm library: ${LIBXSMM_LIBS}")
 include_directories(${LIBXSMM_INCLUDE_DIR})
 ADD_DEFINITIONS(-DPADDLE_WITH_LIBXSMM)
 ADD_DEPENDENCIES(libxsmm extern_libxsmm)
-LIST(APPEND external_project_dependencies libxsmm)
-

cmake/external/mkldnn.cmake

@@ -31,9 +31,17 @@ IF(APPLE)
     return()
 ENDIF()
 
-MESSAGE(STATUS "Set ${MKLDNN_INSTALL_DIR}/lib to runtime path")
+# Introduce variables:
+# * CMAKE_INSTALL_LIBDIR
+INCLUDE(GNUInstallDirs)
+SET(LIBDIR "lib")
+if(CMAKE_INSTALL_LIBDIR MATCHES ".*lib64$")
+  SET(LIBDIR "lib64")
+endif()
+
+MESSAGE(STATUS "Set ${MKLDNN_INSTALL_DIR}/l${LIBDIR} to runtime path")
 SET(CMAKE_INSTALL_RPATH_USE_LINK_PATH TRUE)
-SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${MKLDNN_INSTALL_DIR}/lib")
+SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${MKLDNN_INSTALL_DIR}/${LIBDIR}")
 
 INCLUDE_DIRECTORIES(${MKLDNN_INC_DIR}) # For MKLDNN code to include internal headers.
 
@@ -58,7 +66,7 @@ ExternalProject_Add(
     ${EXTERNAL_PROJECT_LOG_ARGS}
     DEPENDS             ${MKLDNN_DEPENDS}
     GIT_REPOSITORY      "https://github.com/intel/mkl-dnn.git"
-    GIT_TAG             "830a10059a018cd2634d94195140cf2d8790a75a"
+    GIT_TAG             "863ff6e7042cec7d2e29897fe9f0872e0888b0fc"
     PREFIX              ${MKLDNN_SOURCES_DIR}
     UPDATE_COMMAND      ""
     CMAKE_ARGS          -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
@@ -79,9 +87,9 @@ ExternalProject_Add(
                         -DMKLROOT:PATH=${MKLML_ROOT}
 )
 if(WIN32)
-    SET(MKLDNN_LIB "${MKLDNN_INSTALL_DIR}/lib/mkldnn.lib" CACHE FILEPATH "mkldnn library." FORCE)
+    SET(MKLDNN_LIB "${MKLDNN_INSTALL_DIR}/${LIBDIR}/mkldnn.lib" CACHE FILEPATH "mkldnn library." FORCE)
 else(WIN32)
-    SET(MKLDNN_LIB "${MKLDNN_INSTALL_DIR}/lib/libmkldnn.so" CACHE FILEPATH "mkldnn library." FORCE)
+    SET(MKLDNN_LIB "${MKLDNN_INSTALL_DIR}/${LIBDIR}/libmkldnn.so" CACHE FILEPATH "mkldnn library." FORCE)
 endif(WIN32)
 
 ADD_LIBRARY(shared_mkldnn SHARED IMPORTED GLOBAL)
@@ -89,7 +97,6 @@ SET_PROPERTY(TARGET shared_mkldnn PROPERTY IMPORTED_LOCATION ${MKLDNN_LIB})
 ADD_DEPENDENCIES(shared_mkldnn ${MKLDNN_PROJECT})
 MESSAGE(STATUS "MKLDNN library: ${MKLDNN_LIB}")
 add_definitions(-DPADDLE_WITH_MKLDNN)
-LIST(APPEND external_project_dependencies shared_mkldnn)
 
 # generate a static dummy target to track mkldnn dependencies
 # for cc_library(xxx SRCS xxx.c DEPS mkldnn)
@@ -102,7 +109,7 @@ ADD_DEPENDENCIES(mkldnn ${MKLDNN_PROJECT})
 # copy the real so.0 lib to install dir
 # it can be directly contained in wheel or capi
 if(WIN32)
-    SET(MKLDNN_SHARED_LIB ${MKLDNN_INSTALL_DIR}/lib/mkldnn.dll)
+    SET(MKLDNN_SHARED_LIB ${MKLDNN_INSTALL_DIR}/bin/mkldnn.dll)
 else(WIN32)
     SET(MKLDNN_SHARED_LIB ${MKLDNN_INSTALL_DIR}/libmkldnn.so.0)
     ADD_CUSTOM_COMMAND(OUTPUT ${MKLDNN_SHARED_LIB}

cmake/external/mklml.cmake

@@ -34,14 +34,16 @@ SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${MKLML_ROOT}/lib")
 SET(TIME_VERSION "2019.0.1.20181227")
 IF(WIN32)
     SET(MKLML_VER "mklml_win_${TIME_VERSION}" CACHE STRING "" FORCE)
-    SET(MKLML_URL "https://paddlepaddledeps.cdn.bcebos.com/${MKLML_VER}.zip" CACHE STRING "" FORCE)
+    SET(MKLML_URL "https://paddlepaddledeps.bj.bcebos.com/${MKLML_VER}.zip" CACHE STRING "" FORCE)
     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     ${MKLML_LIB_DIR}/libiomp5md.dll)
-ELSE()  
-    SET(MKLML_VER "mklml_lnx_${TIME_VERSION}" CACHE STRING "" FORCE)
-    SET(MKLML_URL "http://paddlepaddledeps.cdn.bcebos.com/${MKLML_VER}.tgz" CACHE STRING "" FORCE)
+ELSE()
+    #TODO(intel-huying):
+    #  Now enable Erf function in mklml library temporarily, it will be updated as offical version later.
+    SET(MKLML_VER "Glibc225_vsErf_mklml_lnx_${TIME_VERSION}" CACHE STRING "" FORCE)
+    SET(MKLML_URL "http://paddlepaddledeps.bj.bcebos.com/${MKLML_VER}.tgz" CACHE STRING "" FORCE)
     SET(MKLML_LIB                 ${MKLML_LIB_DIR}/libmklml_intel.so)
     SET(MKLML_IOMP_LIB            ${MKLML_LIB_DIR}/libiomp5.so)
     SET(MKLML_SHARED_LIB          ${MKLML_LIB_DIR}/libmklml_intel.so)
@@ -73,4 +75,3 @@ INCLUDE_DIRECTORIES(${MKLML_INC_DIR})
 ADD_LIBRARY(mklml SHARED IMPORTED GLOBAL)
 SET_PROPERTY(TARGET mklml PROPERTY IMPORTED_LOCATION ${MKLML_LIB})
 ADD_DEPENDENCIES(mklml ${MKLML_PROJECT})
-LIST(APPEND external_project_dependencies mklml)

cmake/external/ngraph.cmake

@@ -37,7 +37,7 @@ INCLUDE(GNUInstallDirs)
 INCLUDE(ExternalProject)
 
 SET(NGRAPH_PROJECT         "extern_ngraph")
-SET(NGRAPH_GIT_TAG         "20bd8bbc79ae3a81c57313846a2be7313e5d1dab")
+SET(NGRAPH_GIT_TAG         "a444f7a959b7d87f2c117c9b57a4c387759e481e")
 SET(NGRAPH_SOURCES_DIR     ${THIRD_PARTY_PATH}/ngraph)
 SET(NGRAPH_INSTALL_DIR     ${THIRD_PARTY_PATH}/install/ngraph)
 SET(NGRAPH_INC_DIR         ${NGRAPH_INSTALL_DIR}/include)
@@ -57,24 +57,28 @@ SET(NGRAPH_TBB_LIB         ${NGRAPH_LIB_DIR}/${NGRAPH_TBB_LIB_NAME})
 ExternalProject_Add(
     ${NGRAPH_PROJECT}
     ${EXTERNAL_PROJECT_LOG_ARGS}
-    DEPENDS             ${MKLDNN_PROJECT} ${MKLML_PROJECT}
-    GIT_REPOSITORY      ${NGRAPH_GIT_REPO}
-    GIT_TAG             ${NGRAPH_GIT_TAG}
-    PREFIX              ${NGRAPH_SOURCES_DIR}
-    UPDATE_COMMAND      ""
-    CMAKE_ARGS          -DCMAKE_INSTALL_PREFIX=${NGRAPH_INSTALL_DIR}
-    CMAKE_ARGS          -DNGRAPH_UNIT_TEST_ENABLE=FALSE
-    CMAKE_ARGS          -DNGRAPH_TOOLS_ENABLE=FALSE
-    CMAKE_ARGS          -DNGRAPH_INTERPRETER_ENABLE=FALSE
-    CMAKE_ARGS          -DNGRAPH_DEX_ONLY=TRUE
-    CMAKE_ARGS          -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}
-    CMAKE_ARGS          -DMKLDNN_INCLUDE_DIR=${MKLDNN_INC_DIR}
-    CMAKE_ARGS          -DMKLDNN_LIB_DIR=${MKLDNN_INSTALL_DIR}/lib
-    CMAKE_ARGS          -DMKLML_LIB_DIR=${MKLML_INSTALL_DIR}/lib
+    DEPENDS                  ${MKLDNN_PROJECT} ${MKLML_PROJECT}
+    GIT_REPOSITORY           ${NGRAPH_GIT_REPO}
+    GIT_TAG                  ${NGRAPH_GIT_TAG}
+    PREFIX                   ${NGRAPH_SOURCES_DIR}
+    UPDATE_COMMAND           ""
+    CMAKE_GENERATOR          ${CMAKE_GENERATOR}
+    CMAKE_GENERATOR_PLATFORM ${CMAKE_GENERATOR_PLATFORM}
+    CMAKE_GENERATOR_TOOLSET  ${CMAKE_GENERATOR_TOOLSET}
+    CMAKE_ARGS               -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
+    CMAKE_ARGS               -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
+    CMAKE_ARGS               -DCMAKE_INSTALL_PREFIX=${NGRAPH_INSTALL_DIR}
+    CMAKE_ARGS               -DNGRAPH_UNIT_TEST_ENABLE=FALSE
+    CMAKE_ARGS               -DNGRAPH_TOOLS_ENABLE=FALSE
+    CMAKE_ARGS               -DNGRAPH_INTERPRETER_ENABLE=FALSE
+    CMAKE_ARGS               -DNGRAPH_DEX_ONLY=TRUE
+    CMAKE_ARGS               -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}
+    CMAKE_ARGS               -DMKLDNN_INCLUDE_DIR=${MKLDNN_INC_DIR}
+    CMAKE_ARGS               -DMKLDNN_LIB_DIR=${MKLDNN_INSTALL_DIR}/${CMAKE_INSTALL_LIBDIR}
+    CMAKE_ARGS               -DMKLML_LIB_DIR=${MKLML_INSTALL_DIR}/lib
 )
 
 add_dependencies(ngraph ${NGRAPH_PROJECT})
 target_compile_definitions(ngraph INTERFACE -DPADDLE_WITH_NGRAPH)
 target_include_directories(ngraph INTERFACE ${NGRAPH_INC_DIR})
 target_link_libraries(ngraph INTERFACE ${NGRAPH_SHARED_LIB})
-LIST(APPEND external_project_dependencies ngraph)

cmake/external/openblas.cmake

@@ -11,11 +11,6 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-
-IF(USE_EIGEN_FOR_BLAS)
-    return()
-ENDIF(USE_EIGEN_FOR_BLAS)
-
 INCLUDE(cblas)
 
 IF(NOT ${CBLAS_FOUND})
@@ -91,7 +86,6 @@ ENDIF()
 
 IF(NOT ${CBLAS_FOUND})
     ADD_DEPENDENCIES(cblas extern_openblas)
-    LIST(APPEND external_project_dependencies cblas)
 ELSE()
     IF("${CBLAS_PROVIDER}" STREQUAL "MKLML")
         ADD_DEPENDENCIES(cblas mklml)

cmake/external/protobuf.cmake

@@ -129,7 +129,6 @@ macro(PROMPT_PROTOBUF_LIB)
         ADD_DEPENDENCIES(protoc ${dep})
     ENDFOREACH()
 
-    LIST(APPEND external_project_dependencies protobuf)
     RETURN()
 endmacro()
 macro(SET_PROTOBUF_VERSION)
@@ -202,7 +201,7 @@ FUNCTION(build_protobuf TARGET_NAME BUILD_FOR_HOST)
         SET(OPTIONAL_ARGS ${OPTIONAL_ARGS} "-DCMAKE_GENERATOR_PLATFORM=x64")
     ENDIF()
 
-    SET(PROTOBUF_REPO "https://github.com/google/protobuf.git")
+    SET(PROTOBUF_REPO "https://github.com/protocolbuffers/protobuf.git")
     SET(PROTOBUF_TAG "9f75c5aa851cd877fb0d93ccc31b8567a6706546")
 
     ExternalProject_Add(
@@ -231,7 +230,7 @@ FUNCTION(build_protobuf TARGET_NAME BUILD_FOR_HOST)
     )
 ENDFUNCTION()
 
-SET(PROTOBUF_VERSION 3.1)
+SET(PROTOBUF_VERSION 3.1.0)
 
 IF(NOT PROTOBUF_FOUND)
     build_protobuf(extern_protobuf FALSE)

cmake/external/pslib.cmake

@@ -70,4 +70,3 @@ ExternalProject_Add(
 ADD_LIBRARY(pslib SHARED IMPORTED GLOBAL)
 SET_PROPERTY(TARGET pslib PROPERTY IMPORTED_LOCATION ${PSLIB_LIB})
 ADD_DEPENDENCIES(pslib ${PSLIB_PROJECT})
-LIST(APPEND external_project_dependencies pslib)

cmake/external/pslib_brpc.cmake

@@ -70,4 +70,3 @@ ExternalProject_Add(
 ADD_LIBRARY(pslib_brpc SHARED IMPORTED GLOBAL)
 SET_PROPERTY(TARGET pslib_brpc PROPERTY IMPORTED_LOCATION ${PSLIB_BRPC_LIB})
 ADD_DEPENDENCIES(pslib_brpc ${PSLIB_BRPC_PROJECT})
-LIST(APPEND external_project_dependencies pslib_brpc)

cmake/external/threadpool.cmake

@@ -26,5 +26,3 @@ else()
 endif()
 
 add_dependencies(simple_threadpool extern_threadpool)
-
-LIST(APPEND external_project_dependencies simple_threadpool)

cmake/external/warpctc.cmake

@@ -83,5 +83,3 @@ INCLUDE_DIRECTORIES(${THIRD_PARTY_PATH}/install) # For Paddle code to include wa
 ADD_LIBRARY(warpctc SHARED IMPORTED GLOBAL)
 SET_PROPERTY(TARGET warpctc PROPERTY IMPORTED_LOCATION ${WARPCTC_LIBRARIES})
 ADD_DEPENDENCIES(warpctc extern_warpctc)
-
-LIST(APPEND external_project_dependencies warpctc)

cmake/external/wbaes.cmake

+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+IF(NOT ${WITH_WBAES})
+    return()
+ENDIF(NOT ${WITH_WBAES})
+
+INCLUDE(ExternalProject)
+SET(WBAES_DST_DIR       "wbaes")
+SET(WBAES_INSTALL_ROOT  "${THIRD_PARTY_PATH}/install")
+SET(WBAES_INSTALL_DIR   ${WBAES_INSTALL_ROOT}/${WBAES_DST_DIR})
+SET(WBAES_ROOT          ${WBAES_INSTALL_DIR})
+SET(WBAES_INC_DIR       ${WBAES_ROOT}/include)
+SET(WBAES_LIB_DIR       ${WBAES_ROOT}/lib)
+
+SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${WBAES_ROOT}/lib")
+SET(CMAKE_INSTALL_RPATH_USE_LINK_PATH TRUE)
+
+IF(APPLE)
+    SET(WBAES_TAG   "v1.0.0" CACHE STRING "" FORCE)
+    SET(WBAES_URL   "http://paddlepaddledeps.bj.bcebos.com/wbaes-sdk.mac.${WBAES_TAG}.tgz" CACHE STRING "" FORCE)
+    SET(WBAES_LIB   ${WBAES_LIB_DIR}/libwbaes.dylib)
+    SET(WBAES_SHARED_LIB   ${WBAES_LIB_DIR}/libwbaes.dylib)
+ELSEIF(WIN32)
+    SET(WBAES_TAG   "v1.0.0" CACHE STRING "" FORCE)
+    SET(WBAES_URL   "http://paddlepaddledeps.bj.bcebos.com/wbaes-sdk.windows-x64.${WBAES_TAG}.tgz" CACHE STRING "" FORCE)
+    SET(WBAES_LIB   ${WBAES_LIB_DIR}/libwbaes.lib)
+    SET(WBAES_SHARED_LIB   ${WBAES_LIB_DIR}/libwbaes.dll)
+ELSE()
+    SET(WBAES_TAG   "v1.0.2" CACHE STRING "" FORCE)
+    SET(WBAES_URL   "http://paddlepaddledeps.bj.bcebos.com/wbaes-sdk.linux-x86_64.${WBAES_TAG}.tgz" CACHE STRING "" FORCE)
+    SET(WBAES_LIB   ${WBAES_LIB_DIR}/libwbaes.so)
+    SET(WBAES_SHARED_LIB   ${WBAES_LIB_DIR}/libwbaes.so)
+ENDIF()
+
+SET(WBAES_PROJECT       "extern_wbaes")
+MESSAGE(STATUS "WBAES_URL: ${WBAES_URL}, WBAES_LIB: ${WBAES_LIB}")
+SET(WBAES_SOURCE_DIR    "${THIRD_PARTY_PATH}/wbaes") 
+SET(WBAES_DOWNLOAD_DIR  "${WBAES_SOURCE_DIR}/src/${WBAES_PROJECT}")
+
+ExternalProject_Add(
+    ${WBAES_PROJECT}
+    ${EXTERNAL_PROJECT_LOG_ARGS}
+    PREFIX                  ${WBAES_SOURCE_DIR}
+    URL                     ${WBAES_URL}
+    DOWNLOAD_DIR            ${WBAES_DOWNLOAD_DIR}
+    DOWNLOAD_NO_PROGRESS    1
+    CONFIGURE_COMMAND       ""
+    BUILD_COMMAND           ""
+    INSTALL_COMMAND         ""
+        ${CMAKE_COMMAND} -E copy_directory ${WBAES_DOWNLOAD_DIR}/include ${WBAES_INC_DIR} &&
+        ${CMAKE_COMMAND} -E copy_directory ${WBAES_DOWNLOAD_DIR}/lib ${WBAES_LIB_DIR}
+)
+
+INCLUDE_DIRECTORIES(${WBAES_INC_DIR})
+
+ADD_LIBRARY(wbaes SHARED IMPORTED GLOBAL)
+SET_PROPERTY(TARGET wbaes PROPERTY IMPORTED_LOCATION ${WBAES_LIB})
+SET_PROPERTY(TARGET wbaes PROPERTY IMPORTED_NO_SONAME 1)
+ADD_DEPENDENCIES(wbaes ${WBAES_PROJECT})

cmake/external/xbyak.cmake

@@ -55,4 +55,3 @@ else()
 endif()
 
 add_dependencies(xbyak ${XBYAK_PROJECT})
-list(APPEND external_project_dependencies xbyak)

cmake/external/xxhash.cmake

@@ -71,5 +71,3 @@ add_library(xxhash STATIC IMPORTED GLOBAL)
 set_property(TARGET xxhash PROPERTY IMPORTED_LOCATION ${XXHASH_LIBRARIES})
 include_directories(${XXHASH_INCLUDE_DIR})
 add_dependencies(xxhash extern_xxhash)
-
-LIST(APPEND external_project_dependencies xxhash)

cmake/external/zlib.cmake

@@ -57,5 +57,3 @@ ENDIF(WIN32)
 ADD_LIBRARY(zlib STATIC IMPORTED GLOBAL)
 SET_PROPERTY(TARGET zlib PROPERTY IMPORTED_LOCATION ${ZLIB_LIBRARIES})
 ADD_DEPENDENCIES(zlib extern_zlib)
-
-LIST(APPEND external_project_dependencies zlib)

cmake/generic.cmake

@@ -264,6 +264,14 @@ function(cc_library TARGET_NAME)
         list(REMOVE_ITEM cc_library_DEPS warpctc)
         add_dependencies(${TARGET_NAME} warpctc)
       endif()
+      # Only deps libwbaes.so, not link
+      if("${cc_library_DEPS};" MATCHES "wbaes;")
+        list(REMOVE_ITEM cc_library_DEPS wbaes)
+        if(NOT "${TARGET_NAME}" MATCHES "dynload_wbaes")
+          list(APPEND cc_library_DEPS dynload_wbaes)
+        endif()
+        add_dependencies(${TARGET_NAME} wbaes)
+      endif()
       # Only deps libmklml.so, not link
       if("${cc_library_DEPS};" MATCHES "mklml;")
         list(REMOVE_ITEM cc_library_DEPS mklml)

cmake/hip.cmake

@@ -11,8 +11,6 @@ include_directories("/opt/rocm/rocrand/include")
 include_directories("/opt/rocm/rccl/include")
 include_directories("/opt/rocm/thrust")
 
-list(APPEND EXTERNAL_LIBS "-L/opt/rocm/lib/ -lhip_hcc")
-
 set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -fPIC -DPADDLE_WITH_HIP -std=c++11" )
 
 if(WITH_DSO)
@@ -31,22 +29,12 @@ if(WITH_GRPC)
   set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DPADDLE_WITH_GRPC")
 endif(WITH_GRPC)
 
-if(NOT WITH_GOLANG)
-  set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DPADDLE_WITHOUT_GOLANG")
-endif(NOT WITH_GOLANG)
-
 if(WITH_MKLDNN)
   set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DPADDLE_WITH_MKLDNN")
 endif(WITH_MKLDNN)
 
 set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DANY_IMPL_ANY_CAST_MOVEABLE")
 
-if(NOT WITH_RDMA)
-  set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DPADDLE_DISABLE_RDMA")
-endif(NOT WITH_RDMA)
-
-
-
 if(CMAKE_BUILD_TYPE  STREQUAL "Debug")
     list(APPEND HIP_HCC_FLAGS  ${CMAKE_CXX_FLAGS_DEBUG})
 elseif(CMAKE_BUILD_TYPE  STREQUAL "RelWithDebInfo")

cmake/inference_lib.cmake

@@ -131,6 +131,15 @@ elseif (NOT CBLAS_FOUND OR WIN32)
             )
 endif ()
 
+if (WITH_GPU AND NOT WIN32)
+    set(dgc_dir "${FLUID_INSTALL_DIR}/third_party/install/dgc")
+    copy(dgc_lib
+            SRCS ${DGC_INSTALL_DIR}/lib ${DGC_INSTALL_DIR}/include
+            DSTS ${dgc_dir} ${dgc_dir}
+            DEPS dgc)
+endif()
+
+
 if (WITH_MKLDNN)
     set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/mkldnn")
     copy(mkldnn_lib
@@ -161,6 +170,14 @@ copy(snappystream_lib
         DSTS ${dst_dir} ${dst_dir}/lib
         DEPS snappystream)
 
+if (WITH_WBAES)
+    set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/wbaes")
+    copy(wbaes_lib
+            SRCS ${WBAES_INC_DIR} ${WBAES_LIB}
+            DSTS ${dst_dir} ${dst_dir}/lib
+            DEPS wbaes)
+endif ()
+
 set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/zlib")
 copy(zlib_lib
         SRCS ${ZLIB_INCLUDE_DIR} ${ZLIB_LIBRARIES}

cmake/operators.cmake

@@ -110,7 +110,7 @@ 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"
-"fusion_transpose_flatten_concat_op" "fusion_conv_inception_op")
+"fusion_transpose_flatten_concat_op" "fusion_conv_inception_op" "sync_batch_norm_op" "dgc_op")
         if ("${TARGET}" STREQUAL "${manual_pybind_op}")
             set(pybind_flag 1)
         endif()
@@ -153,7 +153,11 @@ function(op_library TARGET)
     # pybind USE_OP_DEVICE_KERNEL for CUDNN
     list(LENGTH cudnn_cu_cc_srcs cudnn_cu_cc_srcs_len)
     if (WITH_GPU AND ${cudnn_cu_cc_srcs_len} GREATER 0)
+      if(${TARGET} STREQUAL "activation")
+        file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(relu, CUDNN);\n")
+      else()
         file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, CUDNN);\n")
+      endif()
     endif()
 
     # pybind USE_OP_DEVICE_KERNEL for MIOPEN
@@ -168,6 +172,9 @@ function(op_library TARGET)
         file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(relu, MKLDNN);\n")
       elseif(${MKLDNN_FILE} STREQUAL "conv_mkldnn_op")
         file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL_WITH_CUSTOM_TYPE(conv2d, MKLDNN, FP32);\n")
+        file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL_WITH_CUSTOM_TYPE(conv2d, MKLDNN, S8);\n")
+        file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL_WITH_CUSTOM_TYPE(conv2d, MKLDNN, U8);\n")
+        
       else()
         file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, MKLDNN);\n")
       endif()

cmake/rdma.cmake

-# user should download rdma first from subversion repository
-
-# execute following instruction to download svn mannally
-# svn co https://svn.baidu.com/sys/ip/trunk/rdma/sockrdmav1 rdma/
-# svn co https://svn.baidu.com/sys/ip/trunk/rdma/thirdparty rdma/
-# we use static output in svn repositories to avoid implict bugs from not standard runtime env.
-
-if(WITH_RDMA)
-  set(RDMA_ROOT $ENV{RDMA_ROOT} CACHE PATH "Folder contains RDMA sock library and thirdparty library")
-
-  function(generate_rdma_links)
-    #redirect to current DIR to isolate the pollution from system runtime environment
-    #it can benifits unified control for different gcc environment.
-    #e.g, by default gcc48 did not refer /usr/lib64 which could contain low version
-    #runtime libraries that will crash process while loading it. That redirect trick
-    #can fix it.
-    execute_process(
-      COMMAND mkdir -p librdma
-      COMMAND ln -s -f /usr/lib64/libibverbs.so.1.0.0 librdma/libibverbs.so.1
-      COMMAND ln -s -f /usr/lib64/libibverbs.so.1.0.0 librdma/libibverbs.so
-      COMMAND ln -s -f /usr/lib64/librdmacm.so.1.0.0 librdma/librdmacm.so.1
-      COMMAND ln -s -f /usr/lib64/librdmacm.so.1.0.0 librdma/librdmacm.so
-      COMMAND ln -s -f /lib64/libnl.so.1.1.4 librdma/libnl.so.1
-      COMMAND ln -s -f /lib64/libnl.so.1.1.4 librdma/libnl.so
-      WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}
-    )
-  endfunction(generate_rdma_links)
-
-  #check and set headers
-  find_path(RDMA_INC_SXISOCK sxi_sock.h PATHS ${RDMA_ROOT}/sockrdmav1/output/include)
-  find_path(RDMA_INC_XIO libxio.h PATHS ${RDMA_ROOT}/thirdparty/output/accelio)
-  find_path(RDMA_INC_EVENT event2 PATHS ${RDMA_ROOT}/thirdparty/output/libevent)
-  find_path(RDMA_INC_NUMA numa.h PATHS ${RDMA_ROOT}/thirdparty/output/libnuma)
-
-  #check and set libs
-  find_library(RDMA_LIB_SXISOCK NAMES sxisock PATHS ${RDMA_ROOT}/sockrdmav1/output)
-  find_library(RDMA_LIB_XIO NAMES xio PATHS ${RDMA_ROOT}/thirdparty/output/accelio)
-  find_library(RDMA_LIB_EVENT NAMES event PATHS ${RDMA_ROOT}/thirdparty/output/libevent)
-  find_library(RDMA_LIB_EVENT_CORE NAMES event_core PATHS ${RDMA_ROOT}/thirdparty/output/libevent)
-  find_library(RDMA_LIB_EVENT_EXTRA NAMES event_extra PATHS ${RDMA_ROOT}/thirdparty/output/libevent)
-  find_library(RDMA_LIB_EVENT_PTHREADS NAMES event_pthreads PATHS ${RDMA_ROOT}/thirdparty/output/libevent)
-  find_library(RDMA_LIB_NUMA NAMES numa PATHS ${RDMA_ROOT}/thirdparty/output/libnuma)
-
-  if(
-      RDMA_INC_SXISOCK AND
-      RDMA_INC_XIO AND
-      RDMA_INC_EVENT AND
-      RDMA_INC_NUMA AND
-      RDMA_LIB_SXISOCK AND
-      RDMA_LIB_XIO AND
-      RDMA_LIB_EVENT AND
-      RDMA_LIB_EVENT_CORE AND
-      RDMA_LIB_EVENT_EXTRA AND
-      RDMA_LIB_EVENT_PTHREADS AND
-      RDMA_LIB_NUMA
-      )
-
-    set(RDMA_INC_DIR
-      ${RDMA_INC_SXISOCK}
-      ${RDMA_INC_XIO}
-      ${RDMA_INC_EVENT}
-      ${RDMA_INC_NUMA})
-    set(RDMA_LIBS
-      ${RDMA_LIB_SXISOCK}
-      ${RDMA_LIB_XIO}
-      ${RDMA_LIB_EVENT}
-      ${RDMA_LIB_EVENT_CORE}
-      ${RDMA_LIB_EVENT_EXTRA}
-      ${RDMA_LIB_EVENT_PTHREADS}
-      ${RDMA_LIB_NUMA}
-      )
-    set(RDMA_LD_FLAGS "-L./librdma -libverbs -lrdmacm -Xlinker -rpath ./librdma")
-    include_directories("${RDMA_INC_DIR}")
-  else()
-    #if this module is not called, RDMA_INC_DIR RDMA_LIBS will be null, so top module always refer this variable
-    message(FATAL_ERROR, "RDMA libraries are not found, try to set RDMA_ROOT or check all related libraries.")
-  endif()
-else(WITH_RDMA)
-  set(RDMA_LIBS "")
-  set(RDMA_LD_FLAGS "")
-  add_definitions(-DPADDLE_DISABLE_RDMA)
-endif(WITH_RDMA)

cmake/tensorrt.cmake

@@ -33,6 +33,6 @@ if(TENSORRT_FOUND)
     message(STATUS "Current TensorRT header is ${TENSORRT_INCLUDE_DIR}/NvInfer.h. "
         "Current TensorRT version is v${TENSORRT_MAJOR_VERSION}. ")
     include_directories(${TENSORRT_INCLUDE_DIR})
-    list(APPEND EXTERNAL_LIBS ${TENSORRT_LIBRARY})
+    link_directories(${TENSORRT_LIBRARY})
     add_definitions(-DPADDLE_WITH_TENSORRT)
 endif()

paddle/contrib/float16/README.md

@@ -5,13 +5,13 @@ Kexin Zhao <zhaokexin01@baidu.com>
 ## Introduction
 Deep learning is usually a two-stage work: training and inference. The training stage estimates model parameters (weights) from data.  The inference stage loads the weights and uses them to interpret inputs. Typically, weights are 32-bit float values (float32).  Some new devices, including NVIDIA Volta GPUs, support higher speed computation using 16-bit float values (float16).
 
-This article explains our efforts with PaddlePaddle to train using float32 and to inference using float16. We describe a [*transpiler*](https://github.com/PaddlePaddle/Paddle/blob/a4d3de0071e1f3912230c3ab3f9ac74cf06b093a/doc/fluid/design/motivation/fluid_compiler.md), which converts a PaddlePaddle Fluid model, which, to be precise, should be called a [Fluid *program*](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/program.md), into the inference program, and converts the weights from float32 into float16.
+This article explains our efforts with PaddlePaddle to train using float32 and to inference using float16. We describe a [*transpiler*](https://github.com/PaddlePaddle/Paddle/blob/a4d3de0071e1f3912230c3ab3f9ac74cf06b093a/doc/fluid/design/motivation/fluid_compiler.md), which converts a PaddlePaddle Fluid model, which, to be precise, should be called a [Fluid *program*](https://github.com/PaddlePaddle/FluidDoc/blob/develop/doc/fluid/design/concepts/program.md), into the inference program, and converts the weights from float32 into float16.
 
 
 ## What is float16?
 float16 (or FP16) is a half-precision floating-point format that uses 16 bits in memory to represent a value. The advantage over 32-bit single-precision floating-point format (commonly known as float or float32 data type) is that it requires half the storage and bandwidth at the expense of precision and range. Fortunately, DNN inference has a high tolerance for the loss of precision and range when using float16 to represent the weights, and the inference accuracy will only be minimally affected in most cases, which gives us the opportunity to use float16 data type to speed up the inference.
 
-Interested readers can refer to our [design doc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/data_type/float16.md) and [code](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/platform/float16.h) for more details on how we implement the float16 data type.
+Interested readers can refer to our [design doc](https://github.com/PaddlePaddle/FluidDoc/blob/develop/doc/fluid/design/data_type/float16.md) and [code](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/platform/float16.h) for more details on how we implement the float16 data type.
 
 ## Why float16?
 The trend in today's deep learning community is to use bigger and deeper model, which translates to larger memory footprint, higher computation demands, and as a result higher energy consumption on computing devices. The advantages of float16 over float32 are correspondingly three-fold:
@@ -24,12 +24,12 @@ The trend in today's deep learning community is to use bigger and deeper model,
 
 ## Fluid implementation of float16 inference
 ### Overview
-Fluid use [Program](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#program) instead of computation graph to describe a neural network model and the optimization procedure. Fluid program is a python wrapper around a protobuf message called [ProgramDesc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/program.md). Similar to programming languages, the basic structure of a Fluid program is some nested [blocks](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#block), where each block consists of some [variable](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#variable) definitions and a sequence of [operators](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#operator). An [executor](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/executor.md) will run a given program by sequentially executing the operators in the entrance block. 
+Fluid use [Program](https://github.com/PaddlePaddle/FluidDoc/blob/develop/doc/fluid/design/modules/python_api.md#program) instead of computation graph to describe a neural network model and the optimization procedure. Fluid program is a python wrapper around a protobuf message called [ProgramDesc](https://github.com/PaddlePaddle/FluidDoc/blob/develop/doc/fluid/design/concepts/program.md). Similar to programming languages, the basic structure of a Fluid program is some nested [blocks](https://github.com/PaddlePaddle/FluidDoc/blob/develop/doc/fluid/design/modules/python_api.md#block), where each block consists of some [variable](https://github.com/PaddlePaddle/FluidDoc/blob/develop/doc/fluid/design/modules/python_api.md#variable) definitions and a sequence of [operators](https://github.com/PaddlePaddle/FluidDoc/blob/develop/doc/fluid/design/modules/python_api.md#operator). An [executor](https://github.com/PaddlePaddle/FluidDoc/blob/develop/doc/fluid/design/concepts/executor.md) will run a given program by sequentially executing the operators in the entrance block.
 
 ### Basic requirement
 When an executor runs an operator, it uses a kernel to perform computations on tensors contained in the input variables, and then writes the results to the tensors in the output variables. Each operator has multiple kernels for different combinations of data types, devices, and library types, respectively. The operator will select the appropriate kernel to run based on, among other things, the data type of the input tensors. By default, every Fluid operator has a kernel for float data type that takes float inputs and generates float outputs.
 
-If we provide float input to the first operator in a program, then each operator will use float kernel to compute float output and send it as input to the next operator to trigger its float kernel. This chain effect will make the program run in float mode and gives us a final output of float data type. 
+If we provide float input to the first operator in a program, then each operator will use float kernel to compute float output and send it as input to the next operator to trigger its float kernel. This chain effect will make the program run in float mode and gives us a final output of float data type.
 
 The same principle applies if we want a program to run in float16 mode. We provide input variable of the float16 data type to the first operator, and every subsequent operator will invoke the float16 kernel until we get the final output in float16. So the preliminary requirements for float16 inference are to add float16 kernels to operators that are needed in a specific kind of neural networks. Our current focus is on Convolutional Neural Networks (CNN) and hence we have added float16 kernels to the following operators: convolution, pooling, GEMM, elementwise addition, batch norm, dropout, various activations including relu and tanh, and softmax.
 
@@ -75,7 +75,7 @@ In this scenario, we already have a float32 inference program and some associate
 
 We can then run various inference experiments in float16 mode and save the float16 program and weights on disk for future deployment. To enhance the code usability, we maintain a consistent API so that user can use the same float32 input data to run inference program in either float32 and float16 mode and obtain output data both of float32 data type. Consequently, we need to add cast operators in the float16 inference program for conversions between the float16 tensor and float32 tensor.
 
-The float16 transpiler is implemented to fulfill the requirements mentioned above. The details of the float16 transpiler can be found [here](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/data_type/float16.md#float16-inference).
+The float16 transpiler is implemented to fulfill the requirements mentioned above. The details of the float16 transpiler can be found [here](https://github.com/PaddlePaddle/FluidDoc/blob/develop/doc/fluid/design/data_type/float16.md#float16-inference).
 
 ### Experiment results
 Simply running the following commands to reproduce the experiment results presented in this section:
@@ -113,7 +113,7 @@ We repeat the test ten times and get the following results:
 | #10    | 62.53%  | 62.48%   |
 | average| 62.63%  | 62.62%   |
 
-We can see that the accuracy of float16 inference is very close to that of float32 inference in every experiment (within 0.05% difference) and is overall 0.01% better than its float32 counterpart averaged over ten tests. 
+We can see that the accuracy of float16 inference is very close to that of float32 inference in every experiment (within 0.05% difference) and is overall 0.01% better than its float32 counterpart averaged over ten tests.
 
 #### Performance benchmark
 Currently, Fluid only supports float16 inference on NVIDIA GPUs. There is no motivation to support float16 inference on non-ARM CPUs where float16 is not natively supported, and float16 calculation will only be slower than its float32 counterpart. 
@@ -132,7 +132,7 @@ Average inference time for one mini-batch on Vgg16 model tested on ImageNet data
 |float16|  3.32 | 4.11  |  5.88 |  9.41 | 16.54  | 30.47 |  60.23 |
 |Speedup|  4.22 | 2.36  |  3.91 |  3.00 |  3.26  |  2.77 |   2.97 |
 
-We can see that float16 inference provides **2x ~ 4x** speedup on different batch sizes. 
+We can see that float16 inference provides **2x ~ 4x** speedup on different batch sizes.
 
 Convolution operation is ususally the computational bottleneck of CNN, so we also check the average time spent on the Fluid convolution operators for one mini-batch as follows:
 
@@ -162,7 +162,7 @@ We find that the speedup provided by float16 inference starts relatively small a
 
 We also did the same benchmark on a single NVIDIA GeForce GTX 1080 Ti GPU that does not support Tensor Core. The results show that for Vgg16, float16 inference provides consistent small speedup (around 1.15x) for all mini-batch sizes, while for Resnet50, float16 inference is slower than its float32 counterpart in small batch sizes (mb = 1 and 2) and then delivers around 1.15x speedup for all larger batch sizes. By comparing the benchmarks on 1080 Ti and V100, we find that Tensor Core, which is specialized for float16 computations, is a critical component of high performance float16 inference.
 
-Please refer to [here](https://github.com/PaddlePaddle/Paddle/blob/develop/contrib/float16/float16_benchmark.md) for complete benchmark results.
+Please refer to [here](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/contrib/float16/float16_benchmark.md) for complete benchmark results.
 
 ### Summary
 1. Fluid is now able to run inference in float16 mode via a float16 transpiler. We currently support CNN programs, including Vgg and Resnet, to run in float16 inference mode.

paddle/contrib/float16/run_float16_demo.sh

@@ -14,9 +14,7 @@ cmake .. -DWITH_AVX=OFF \
          -DWITH_MKL=OFF \
          -DWITH_GPU=ON \
          -DWITH_TESTING=ON \
-         -DWITH_TIMER=ON \
          -DWITH_PROFILER=ON \
-         -DWITH_FLUID_ONLY=ON
 make -j `nproc`
 pip install -U "$WHEEL_PATH/$(ls $WHEEL_PATH)"
 

paddle/fluid/framework/CMakeLists.txt

+
 #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)
@@ -22,9 +23,13 @@ endfunction()
 
 add_subdirectory(ir)
 add_subdirectory(details)
+add_subdirectory(fleet)
+add_subdirectory(io)
 #ddim lib
 proto_library(framework_proto SRCS framework.proto)
+proto_library(data_feed_proto SRCS data_feed.proto)
 proto_library(async_executor_proto SRCS data_feed.proto)
+proto_library(trainer_desc_proto SRCS trainer_desc.proto data_feed.proto)
 
 cc_library(ddim SRCS ddim.cc DEPS eigen3 boost enforce)
 cc_test(ddim_test SRCS ddim_test.cc DEPS ddim)
@@ -38,10 +43,10 @@ if(WITH_GPU)
     nv_library(tensor SRCS tensor.cc .tensor_util.cu DEPS place memory data_type device_context)
     add_dependencies(tensor tensor_util)
   else()
-    nv_library(tensor SRCS tensor.cc tensor_util.cu DEPS place memory data_type device_context )
+    nv_library(tensor SRCS tensor.cc tensor_util.cu DEPS place memory data_type device_context profiler)
   endif(WIN32)
 else()
-  cc_library(tensor SRCS tensor.cc tensor_util.cc DEPS place memory data_type device_context )
+  cc_library(tensor SRCS tensor.cc tensor_util.cc DEPS place memory data_type device_context profiler)
 endif()
 
 cc_test(tensor_test SRCS tensor_test.cc DEPS tensor)
@@ -63,7 +68,7 @@ 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(garbage_collector SRCS garbage_collector.cc DEPS device_context memory gflags glog)
 
 cc_library(reader SRCS reader.cc DEPS lod_tensor ddim)
 cc_test(reader_test SRCS reader_test.cc DEPS reader)
@@ -129,9 +134,11 @@ 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(op_registry SRCS op_registry.cc DEPS op_proto_maker op_info operator glog proto_desc memory_optimize_helper)
+
 nv_test(op_registry_test SRCS op_registry_test.cc DEPS op_registry)
 
 py_proto_compile(framework_py_proto SRCS framework.proto data_feed.proto)
+py_proto_compile(trainer_py_proto SRCS trainer_desc.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)
@@ -164,28 +171,44 @@ else()
   set(NGRAPH_EXE_DEPS)
 endif()
 
+cc_library(executor_gc_helper SRCS executor_gc_helper.cc DEPS scope proto_desc operator garbage_collector)
 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_rpc  ${GLOB_DISTRIBUTE_DEPS} graph_to_program_pass variable_helper ${NGRAPH_EXE_DEPS})
-  set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
-  set_source_files_properties(executor.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
+  cc_library(executor SRCS executor.cc multi_trainer.cc dataset_factory.cc
+  dist_multi_trainer.cc trainer_factory.cc trainer.cc data_feed_factory.cc
+  data_feed.cc device_worker.cc hogwild_worker.cc downpour_worker.cc
+  pull_dense_worker.cc device_worker_factory.cc data_set.cc DEPS op_registry
+  device_context scope framework_proto trainer_desc_proto glog fs shell fleet_wrapper lodtensor_printer
+  lod_rank_table feed_fetch_method sendrecvop_rpc  ${GLOB_DISTRIBUTE_DEPS}
+  graph_to_program_pass variable_helper data_feed_proto ${NGRAPH_EXE_DEPS} timer)
+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()
-  cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass variable_helper ${NGRAPH_EXE_DEPS})
+  cc_library(executor SRCS executor.cc multi_trainer.cc dataset_factory.cc
+  dist_multi_trainer.cc trainer_factory.cc trainer.cc data_feed_factory.cc
+  data_feed.cc device_worker.cc hogwild_worker.cc downpour_worker.cc
+  pull_dense_worker.cc device_worker_factory.cc data_set.cc DEPS op_registry
+  device_context scope framework_proto data_feed_proto trainer_desc_proto glog
+  lod_rank_table fs shell fleet_wrapper lodtensor_printer feed_fetch_method
+  graph_to_program_pass variable_helper ${NGRAPH_EXE_DEPS} timer data_feed_proto)
   cc_test(test_naive_executor SRCS naive_executor_test.cc DEPS naive_executor elementwise_add_op)
 endif()
 
-target_link_libraries(executor garbage_collector)
+target_link_libraries(executor while_op_helper executor_gc_helper)
 
 cc_library(parallel_executor SRCS parallel_executor.cc DEPS
-        threaded_ssa_graph_executor scope_buffered_ssa_graph_executor parallel_ssa_graph_executor
+        threaded_ssa_graph_executor scope_buffered_ssa_graph_executor parallel_ssa_graph_executor async_ssa_graph_executor
         graph build_strategy
         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 timer)
-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 timer)
-endif(WITH_PSLIB)
+cc_library(async_executor SRCS async_executor.cc data_feed.cc data_feed_factory.cc
+           executor_thread_worker.cc multi_trainer.cc dist_multi_trainer.cc
+           trainer_factory.cc trainer.cc device_worker.cc hogwild_worker.cc
+           downpour_worker.cc pull_dense_worker.cc device_worker_factory.cc
+           data_set.cc dataset_factory.cc
+           DEPS op_registry device_context scope framework_proto
+           trainer_desc_proto glog lod_rank_table fleet_wrapper lodtensor_printer
+           feed_fetch_method graph_to_program_pass data_feed_proto
+           variable_helper timer fs shell)
 
 
 cc_test(data_feed_test SRCS data_feed_test.cc DEPS async_executor)
@@ -193,7 +216,7 @@ cc_library(prune SRCS prune.cc DEPS framework_proto)
 cc_test(prune_test SRCS prune_test.cc DEPS op_info prune recurrent_op device_context)
 cc_test(var_type_inference_test SRCS var_type_inference_test.cc DEPS op_registry
         proto_desc)
-cc_test(inplace_op_inference_test SRCS inplace_op_inference_test.cc DEPS op_registry proto_desc op_info memory_optimize_helper)
+cc_test(inplace_op_inference_test SRCS inplace_op_inference_test.cc DEPS inplace_op_pass op_registry proto_desc op_info memory_optimize_helper pass_builder)
 cc_library(selected_rows SRCS selected_rows.cc DEPS tensor)
 cc_test(selected_rows_test SRCS selected_rows_test.cc DEPS selected_rows)
 
@@ -212,18 +235,18 @@ cc_test(dlpack_tensor_test SRCS dlpack_tensor_test.cc DEPS dlpack_tensor glog)
 # Get the current working branch
 execute_process(
   COMMAND git rev-parse --abbrev-ref HEAD
-  WORKING_DIRECTORY ${CMAKE_SOURCE_DIR}
-  OUTPUT_VARIABLE PADDLE_BRANCH
-  OUTPUT_STRIP_TRAILING_WHITESPACE
-)
+    WORKING_DIRECTORY ${CMAKE_SOURCE_DIR}
+      OUTPUT_VARIABLE PADDLE_BRANCH
+        OUTPUT_STRIP_TRAILING_WHITESPACE
+	)
 
 # Get the latest abbreviated commit hash of the working branch
 execute_process(
   COMMAND git log -1 --format=%h
-  WORKING_DIRECTORY ${CMAKE_SOURCE_DIR}
-  OUTPUT_VARIABLE PADDLE_COMMIT
-  OUTPUT_STRIP_TRAILING_WHITESPACE
-)
+    WORKING_DIRECTORY ${CMAKE_SOURCE_DIR}
+      OUTPUT_VARIABLE PADDLE_COMMIT
+        OUTPUT_STRIP_TRAILING_WHITESPACE
+	)
 
 message(STATUS "commit: ${PADDLE_COMMIT}")
 message(STATUS "branch: ${PADDLE_BRANCH}")

paddle/fluid/framework/async_executor.cc

@@ -26,212 +26,44 @@ limitations under the License. */
 #include "paddle/fluid/framework/lod_tensor_array.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/framework/reader.h"
+#include "paddle/fluid/framework/trainer.h"
+#include "paddle/fluid/framework/trainer_desc.pb.h"
+#include "paddle/fluid/framework/trainer_factory.h"
 #include "paddle/fluid/inference/io.h"
 #include "paddle/fluid/platform/place.h"
 #include "paddle/fluid/pybind/pybind.h"
-#ifdef PADDLE_WITH_PSLIB
-#include <pslib.h>
-#endif
 
 namespace paddle {
 namespace framework {
 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<DataFeed>& reader,
-    const std::vector<std::string>& 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<std::shared_ptr<DataFeed>>& readers,  // NOLINT
-                    const int thread_num, const DataFeedDesc& data_feed_desc,
-                    const std::vector<std::string>& 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<paddle::distributed::PSlib>(
-      new paddle::distributed::PSlib());
-  _pslib_ptr->init_server(dist_desc, index);
-  InitParamConfig();
+  fleet_ptr_ = FleetWrapper::GetInstance();
+  fleet_ptr_->InitServer(dist_desc, index);
 }
 
 void AsyncExecutor::InitWorker(const std::string& dist_desc,
                                const std::vector<uint64_t>& host_sign_list,
                                int node_num, int index) {
-  _pslib_ptr = std::shared_ptr<paddle::distributed::PSlib>(
-      new paddle::distributed::PSlib());
-  _pslib_ptr->init_worker(
-      dist_desc, const_cast<uint64_t*>(host_sign_list.data()), node_num, index);
-
-  InitParamConfig();
+  fleet_ptr_ = FleetWrapper::GetInstance();
+  fleet_ptr_->InitWorker(dist_desc, host_sign_list, node_num, index);
 }
 
-uint64_t AsyncExecutor::StartServer() { return _pslib_ptr->run_server(); }
+uint64_t AsyncExecutor::StartServer() { return fleet_ptr_->RunServer(); }
 
-void AsyncExecutor::StopServer() { _pslib_ptr->stop_server(); }
+void AsyncExecutor::StopServer() { fleet_ptr_->StopServer(); }
 
 void AsyncExecutor::GatherServers(const std::vector<uint64_t>& host_sign_list,
                                   int node_num) {
-  _pslib_ptr->gather_servers(const_cast<uint64_t*>(host_sign_list.data()),
-                             node_num);
-}
-
-void AsyncExecutor::InitParamConfig() {
-  for (int i = 0; i < _pslib_ptr->get_param()
-                          ->server_param()
-                          .downpour_server_param()
-                          .downpour_table_param_size();
-       ++i) {
-    if (_pslib_ptr->get_param()
-            ->server_param()
-            .downpour_server_param()
-            .downpour_table_param(i)
-            .table_class()
-            .find("SparseTable") != -1) {
-      _param_config.fea_dim = _pslib_ptr->get_param()
-                                  ->server_param()
-                                  .downpour_server_param()
-                                  .downpour_table_param(i)
-                                  .accessor()
-                                  .fea_dim();
-      break;
-    }
-  }
-  _param_config.slot_dim = _param_config.fea_dim - 2;
-  _param_config.tmp_push_dense_wait_times = static_cast<int32_t>(
-      _pslib_ptr->get_param()->trainer_param().push_dense_per_batch());
-  _param_config.tmp_push_sparse_wait_times = static_cast<int32_t>(
-      _pslib_ptr->get_param()->trainer_param().push_sparse_per_batch());
-
-  for (auto t = 0u; t < _pslib_ptr->get_param()->trainer_param().skip_op_size();
-       ++t) {
-    _param_config.skip_op.push_back(
-        _pslib_ptr->get_param()->trainer_param().skip_op(t));
-  }
-
-  for (auto t = 0u;
-       t < _pslib_ptr->get_param()->trainer_param().sparse_table_size(); ++t) {
-    auto& table = _pslib_ptr->get_param()->trainer_param().sparse_table(t);
-    std::vector<std::string> tmp_sparse_variable_name;
-    for (int i = 0u; i < table.slot_value_size(); ++i) {
-      tmp_sparse_variable_name.push_back(table.slot_value(i));
-      _param_config.slot_alias_to_table[table.slot_key(i)] = table.table_id();
-    }
-    std::vector<std::string> tmp_sparse_gradient_variable_name;
-    for (auto i = 0u; i < table.slot_gradient_size(); ++i) {
-      tmp_sparse_gradient_variable_name.push_back(table.slot_gradient(i));
-    }
-    _param_config.slot_input_vec[table.table_id()] =
-        std::move(tmp_sparse_variable_name);
-    _param_config.gradient_var[table.table_id()] =
-        std::move(tmp_sparse_gradient_variable_name);
-    _param_config.sparse_table_id.push_back(table.table_id());
-  }
-
-  for (auto t = 0u;
-       t < _pslib_ptr->get_param()->trainer_param().dense_table_size(); ++t) {
-    auto& table = _pslib_ptr->get_param()->trainer_param().dense_table(t);
-    std::vector<std::string> tmp_dense_variable_name;
-    for (int i = 0u; i < table.dense_variable_name_size(); ++i) {
-      tmp_dense_variable_name.push_back(table.dense_variable_name(i));
-    }
-    std::vector<std::string> tmp_dense_gradient_variable_name;
-    for (auto i = 0u; i < table.dense_gradient_variable_name_size(); ++i) {
-      tmp_dense_gradient_variable_name.push_back(
-          table.dense_gradient_variable_name(i));
-    }
-    _param_config.dense_variable_name[table.table_id()] =
-        std::move(tmp_dense_variable_name);
-    _param_config.dense_gradient_variable_name[table.table_id()] =
-        std::move(tmp_dense_gradient_variable_name);
-    _param_config.dense_table_id.push_back(table.table_id());
-    _param_config.dense_table_size.push_back(table.fea_dim());
-  }
+  fleet_ptr_->GatherServers(host_sign_list, node_num);
 }
 
-void AsyncExecutor::InitModel() {
-  for (auto table_id : _param_config.dense_table_id) {
-    std::vector<paddle::ps::Region> regions;
-    for (auto& t : _param_config.dense_variable_name[table_id]) {
-      Variable* var = root_scope_->FindVar(t);
-      CHECK(var != nullptr) << "var[" << t << "] not found";
-      LoDTensor* tensor = var->GetMutable<LoDTensor>();
-
-      float* g = tensor->data<float>();
-      CHECK(g != nullptr) << "var[" << t << "] value not initialized";
-
-      float init_range = 0.2;
-      int rown = tensor->dims()[0];
-      init_range /= sqrt(rown);
-
-      std::normal_distribution<float> ndistr(0.0, 1.0);
-      for (auto i = 0u; i < tensor->numel(); ++i) {
-        g[i] = ndistr(local_random_engine()) * init_range;
-      }
-
-      paddle::ps::Region reg(g, tensor->numel());
-      regions.emplace_back(std::move(reg));
-    }
+// todo InitModel
+void AsyncExecutor::InitModel() {}
 
-    auto push_status = _pslib_ptr->_worker_ptr->push_dense_param(
-        regions.data(), regions.size(), table_id);
-    push_status.wait();
-    auto status = push_status.get();
-    if (status != 0) {
-      LOG(FATAL) << "push dense param failed, status[" << status << "]";
-      exit(-1);
-    }
-  }
-}
-
-void AsyncExecutor::SaveModel(const std::string& path) {
-  auto ret = _pslib_ptr->_worker_ptr->flush();
-  ret.wait();
-  ret = _pslib_ptr->_worker_ptr->save(path, 0);
-  ret.wait();
-  int32_t feasign_cnt = ret.get();
-  if (feasign_cnt == -1) {  // (colourful-tree) TODO should be feasign_cnt < 0
-    LOG(FATAL) << "save model failed";
-    exit(-1);
-  }
-}
-
-void AsyncExecutor::PrepareDenseThread(const std::string& mode) {
-  if (mode == "mpi") {
-    DensePullThreadParam param;
-    param.ps_client = _pslib_ptr->_worker_ptr;
-    param.threshold = 1;
-    param.training_thread_num = actual_thread_num;
-    param.root_scope = root_scope_;
-    param.dense_params = &_param_config.dense_variable_name;
-
-    _pull_dense_thread =
-        std::shared_ptr<DensePullThread>(new DensePullThread(param));
-    _pull_dense_thread->start();
-  }
-}
-#endif
+// todo SaveModel
+void AsyncExecutor::SaveModel(const std::string& path) {}
 
 void AsyncExecutor::RunFromFile(const ProgramDesc& main_program,
                                 const std::string& data_feed_desc_str,
@@ -256,14 +88,14 @@ void AsyncExecutor::RunFromFile(const ProgramDesc& main_program,
   google::protobuf::TextFormat::ParseFromString(data_feed_desc_str,
                                                 &data_feed_desc);
 
-  actual_thread_num = thread_num;
+  actual_thread_num_ = thread_num;
   int file_cnt = filelist.size();
   PADDLE_ENFORCE(file_cnt > 0, "File list cannot be empty");
 
-  if (actual_thread_num > file_cnt) {
+  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;
+    actual_thread_num_ = file_cnt;
   }
 
   /*
@@ -279,12 +111,14 @@ void AsyncExecutor::RunFromFile(const ProgramDesc& main_program,
    */
   // todo: should be factory method for creating datafeed
   std::vector<std::shared_ptr<DataFeed>> readers;
-  PrepareReaders(readers, actual_thread_num, data_feed_desc, filelist);
+  /*
+  PrepareReaders(readers, actual_thread_num_, data_feed_desc, filelist);
 #ifdef PADDLE_WITH_PSLIB
   PrepareDenseThread(mode);
 #endif
+  */
   std::vector<std::shared_ptr<ExecutorThreadWorker>> workers;
-  workers.resize(actual_thread_num);
+  workers.resize(actual_thread_num_);
   for (auto& worker : workers) {
 #ifdef PADDLE_WITH_PSLIB
     if (mode == "mpi") {
@@ -298,13 +132,15 @@ void AsyncExecutor::RunFromFile(const ProgramDesc& main_program,
   }
 
   // prepare thread resource here
-  for (int thidx = 0; thidx < actual_thread_num; ++thidx) {
+  /*
+  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) {
+  for (int thidx = 0; thidx < actual_thread_num_; ++thidx) {
     if (debug) {
       threads.push_back(std::thread(&ExecutorThreadWorker::TrainFilesWithTimer,
                                     workers[thidx].get()));
@@ -317,15 +153,19 @@ void AsyncExecutor::RunFromFile(const ProgramDesc& main_program,
   for (auto& th : threads) {
     th.join();
   }
+  // TODO(guru4elephant): we don't need this
+  /*
 #ifdef PADDLE_WITH_PSLIB
   if (mode == "mpi") {
     _pull_dense_thread->stop();
   }
 #endif
+  */
+  VLOG(3) << "start to run from files in async_executor";
+  VLOG(3) << "Drop current scope kids";
   root_scope_->DropKids();
-
   return;
 }
 
-}  // einit_modelnd namespace framework
+}  // end namespace framework
 }  // end namespace paddle

paddle/fluid/framework/async_executor.h

@@ -25,8 +25,10 @@ limitations under the License. */
 #include <typeinfo>
 #include <vector>
 #include "paddle/fluid/framework/data_feed.pb.h"
+#include "paddle/fluid/framework/data_set.h"
 #include "paddle/fluid/framework/executor.h"
 #include "paddle/fluid/framework/executor_thread_worker.h"
+#include "paddle/fluid/framework/fleet/fleet_wrapper.h"
 #include "paddle/fluid/framework/program_desc.h"
 #include "paddle/fluid/framework/scope.h"
 
@@ -65,9 +67,10 @@ class AsyncExecutor {
                    const std::string& data_feed_desc_str,
                    const std::vector<std::string>& filelist,
                    const int thread_num,
-                   const std::vector<std::string>& fetch_names,
-                   const std::string& mode, const bool debug = false);
-#ifdef PADDLE_WITH_PSLIB
+                   const std::vector<std::string>& fetch_var_names,
+                   const std::string& mode, const bool debug);
+
+  // TODO(guru4elephant): make init server decoupled from executor
   void InitServer(const std::string& dist_desc, int index);
   void InitWorker(const std::string& dist_desc,
                   const std::vector<uint64_t>& host_sign_list, int node_num,
@@ -77,31 +80,14 @@ class AsyncExecutor {
   void GatherServers(const std::vector<uint64_t>& host_sign_list, int node_num);
   void InitModel();
   void SaveModel(const std::string& path);
-  void InitParamConfig();
-#endif
-
- private:
-  void CreateThreads(ExecutorThreadWorker* worker,
-                     const ProgramDesc& main_program,
-                     const std::shared_ptr<DataFeed>& reader,
-                     const std::vector<std::string>& fetch_var_names,
-                     Scope* root_scope, const int thread_index,
-                     const bool debug);
-#ifdef PADDLE_WITH_PSLIB
-  void PrepareDenseThread(const std::string& mode);
-#endif
 
  public:
-#ifdef PADDLE_WITH_PSLIB
-  std::shared_ptr<paddle::distributed::PSlib> _pslib_ptr;
-  std::shared_ptr<DensePullThread> _pull_dense_thread;
-  AsyncWorkerParamConfig _param_config;
-#endif
+  std::shared_ptr<paddle::framework::FleetWrapper> fleet_ptr_;
   Scope* root_scope_;
   platform::Place place_;
 
  private:
-  int actual_thread_num;
+  int actual_thread_num_;
 };
 
 }  // namespace framework

paddle/fluid/framework/block_desc.cc

@@ -13,7 +13,11 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/framework/block_desc.h"
+
 #include <queue>
+#include <unordered_set>
+#include <utility>
+
 #include "paddle/fluid/framework/operator.h"
 #include "paddle/fluid/framework/program_desc.h"
 
@@ -155,6 +159,16 @@ void BlockDesc::RemoveOp(size_t s, size_t e) {
   ops_.erase(ops_.begin() + s, ops_.begin() + e);
 }
 
+void BlockDesc::RemoveOpInternal(const OpDesc *op_desc) {
+  // TODO(minqiyang): make this faster
+  for (auto it = ops_.begin(); it != ops_.end(); ++it) {
+    if (it->get() == op_desc) {
+      ops_.erase(it);
+      break;
+    }
+  }
+}
+
 std::vector<OpDesc *> BlockDesc::AllOps() const {
   std::vector<OpDesc *> res;
   for (const auto &op : ops_) {

paddle/fluid/framework/block_desc.h

@@ -93,6 +93,8 @@ class BlockDesc {
    */
   void RemoveOp(size_t s, size_t e);
 
+  void RemoveOpInternal(const OpDesc *op_desc);
+
   void RemoveVar(const std::string &name) { vars_.erase(name); }
 
   std::vector<OpDesc *> AllOps() const;

paddle/fluid/framework/blocking_queue.h

@@ -33,6 +33,14 @@ class BlockingQueue {
     cv_.notify_one();
   }
 
+  void Push(T &&item) {
+    {
+      std::lock_guard<std::mutex> g(mutex_);
+      q_.emplace_back(std::move(item));
+    }
+    cv_.notify_one();
+  }
+
   template <typename U>
   void Extend(const U &items) {
     {
@@ -44,6 +52,17 @@ class BlockingQueue {
     cv_.notify_all();
   }
 
+  template <typename U>
+  void Extend(U &&items) {
+    {
+      std::lock_guard<std::mutex> g(mutex_);
+      for (auto &item : items) {
+        q_.emplace_back(std::move(item));
+      }
+    }
+    cv_.notify_all();
+  }
+
   std::deque<T> PopAll(size_t ms, bool *timeout) {
     auto time =
         std::chrono::system_clock::now() + std::chrono::milliseconds(ms);
@@ -64,6 +83,18 @@ class BlockingQueue {
     return rc;
   }
 
+  void Pop(T *t) {
+    std::unique_lock<std::mutex> lock(mutex_);
+    cv_.wait(lock, [=] { return !q_.empty(); });
+    *t = std::move(q_.front());
+    q_.pop_front();
+  }
+
+  size_t Size() {
+    std::lock_guard<std::mutex> lock(mutex_);
+    return q_.size();
+  }
+
  private:
   std::mutex mutex_;
   std::condition_variable cv_;

paddle/fluid/framework/data_feed.h

@@ -15,17 +15,23 @@ limitations under the License. */
 #pragma once
 
 #include <fstream>
+#include <future>  // NOLINT
 #include <memory>
 #include <mutex>  // NOLINT
+#include <sstream>
 #include <string>
 #include <thread>  // NOLINT
+#include <utility>
 #include <vector>
 
+#include "paddle/fluid/framework/blocking_queue.h"
 #include "paddle/fluid/framework/data_feed.pb.h"
+#include "paddle/fluid/framework/fleet/fleet_wrapper.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"
+#include "paddle/fluid/string/string_helper.h"
 
 namespace paddle {
 namespace framework {
@@ -48,7 +54,10 @@ namespace framework {
 //   }
 class DataFeed {
  public:
-  DataFeed() {}
+  DataFeed() {
+    mutex_for_pick_file_ = nullptr;
+    file_idx_ = nullptr;
+  }
   virtual ~DataFeed() {}
   virtual void Init(const paddle::framework::DataFeedDesc& data_feed_desc) = 0;
   virtual bool CheckFile(const char* filename) {
@@ -59,6 +68,7 @@ class DataFeed {
   // Otherwise, Init() function will init finish_set_filelist_ flag.
   virtual bool SetFileList(const std::vector<std::string>& 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.
@@ -74,6 +84,36 @@ class DataFeed {
   // This function is used for binding feed_vec memory
   virtual void AddFeedVar(Variable* var, const std::string& name);
 
+  // This function will do nothing at default
+  virtual void SetMemoryData(void* memory_data) {}
+  // This function will do nothing at default
+  virtual void SetMemoryDataMutex(std::mutex* mutex) {}
+  // This function will do nothing at default
+  virtual void SetThreadId(int thread_id) {}
+  // This function will do nothing at default
+  virtual void SetThreadNum(int thread_num) {}
+  // This function will do nothing at default
+  virtual void SetTrainerNum(int trainer_num) {}
+  virtual void SetFileListMutex(std::mutex* mutex) {
+    mutex_for_pick_file_ = mutex;
+  }
+  virtual void SetFileListIndex(size_t* file_index) { file_idx_ = file_index; }
+  virtual void LoadIntoMemory() {
+    PADDLE_THROW("This function(LoadIntoMemory) is not implemented.");
+  }
+  virtual void LocalShuffle() {
+    PADDLE_THROW("This function(LocalShuffle) is not implemented.");
+  }
+  virtual void GlobalShuffle() {
+    PADDLE_THROW("This function(GlobalShuffle) is not implemented.");
+  }
+  // This function will do nothing at default
+  virtual void FillMemoryDataToChannel() {}
+  // This function will do nothing at default
+  virtual void FillChannelToMemoryData() {}
+  // This function will do nothing at default
+  virtual void PutInsToChannel(const std::string& ins_str) {}
+
  protected:
   // The following three functions are used to check if it is executed in this
   // order:
@@ -87,9 +127,9 @@ class DataFeed {
   // safe).
   virtual bool PickOneFile(std::string* filename);
 
-  static std::vector<std::string> filelist_;
-  static size_t file_idx_;
-  static std::mutex mutex_for_pick_file_;
+  std::vector<std::string> filelist_;
+  size_t* file_idx_;
+  std::mutex* mutex_for_pick_file_;
 
   // the alias of used slots, and its order is determined by
   // data_feed_desc(proto object)
@@ -112,8 +152,9 @@ class DataFeed {
   int batch_size_;
 
   bool finish_init_;
-  static bool finish_set_filelist_;
+  bool finish_set_filelist_;
   bool finish_start_;
+  std::string pipe_command_;
 };
 
 // PrivateQueueDataFeed is the base virtual class for ohther DataFeeds.
@@ -136,6 +177,7 @@ class PrivateQueueDataFeed : public DataFeed {
   virtual void SetQueueSize(int queue_size);
   // The reading and parsing method called in the ReadThread.
   virtual bool ParseOneInstance(T* instance) = 0;
+  virtual bool ParseOneInstanceFromPipe(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;
@@ -150,11 +192,58 @@ class PrivateQueueDataFeed : public DataFeed {
   //     ifstream one line and one line parse: 6034 ms
   //     fread one buffer and one buffer parse: 7097 ms
   std::ifstream file_;
+  std::shared_ptr<FILE> fp_;
   size_t queue_size_;
+  string::LineFileReader reader_;
   // The queue for store parsed data
   std::unique_ptr<paddle::operators::reader::BlockingQueue<T>> queue_;
 };
 
+template <typename T>
+class InMemoryDataFeed : public PrivateQueueDataFeed<T> {
+ public:
+  InMemoryDataFeed();
+  virtual ~InMemoryDataFeed() {}
+  virtual void Init(const paddle::framework::DataFeedDesc& data_feed_desc) = 0;
+  virtual bool Start();
+  virtual int Next();
+  virtual void SetMemoryData(void* memory_data);
+  virtual void SetMemoryDataMutex(std::mutex* mutex);
+  virtual void SetThreadId(int thread_id);
+  virtual void SetThreadNum(int thread_num);
+  virtual void SetTrainerNum(int trainer_num);
+  virtual void PutInsToChannel(const std::string& ins_str);
+  virtual void FillMemoryDataToChannel();
+  virtual void FillChannelToMemoryData();
+  virtual void LoadIntoMemory();
+  virtual void LocalShuffle();
+  virtual void GlobalShuffle();
+
+ protected:
+  virtual void AddInstanceToInsVec(T* vec_ins, const T& instance,
+                                   int index) = 0;
+  virtual bool ParseOneInstance(T* instance) = 0;
+  virtual bool ParseOneInstanceFromPipe(T* instance) = 0;
+  virtual void PutToFeedVec(const T& ins_vec) = 0;
+  virtual void SerializeIns(const std::vector<T*>& ins, std::string* str) = 0;
+  virtual void DeserializeIns(std::vector<T>* ins, const std::string& str) = 0;
+  virtual std::pair<int64_t, int64_t> GetMemoryDataInterval();
+
+  int thread_id_;
+  int thread_num_;
+  int trainer_num_;
+  uint32_t rand_seed;
+  std::vector<T>* memory_data_;
+  std::mutex* mutex_for_update_memory_data_;
+  // when read ins, we put ins from one channel to the other,
+  // and when finish reading, we set cur_channel = 1 - cur_channel,
+  // so if cur_channel=0, all data are in shuffled_ins_, else shuffled_ins_out_
+  int cur_channel_;
+  std::shared_ptr<paddle::framework::BlockingQueue<T>> shuffled_ins_;
+  std::shared_ptr<paddle::framework::BlockingQueue<T>> shuffled_ins_out_;
+  int64_t fleet_send_batch_size_;
+};
+
 // This class define the data type of instance(ins_vec) in MultiSlotDataFeed
 class MultiSlotType {
  public:
@@ -176,6 +265,7 @@ class MultiSlotType {
     offset_[0] = 0;
   }
   const std::vector<size_t>& GetOffset() const { return offset_; }
+  std::vector<size_t>& MutableOffset() { return offset_; }
   void AddValue(const float v) {
     CheckFloat();
     float_feasign_.push_back(v);
@@ -198,8 +288,33 @@ class MultiSlotType {
     }
   }
   const std::vector<float>& GetFloatData() const { return float_feasign_; }
+  std::vector<float>& MutableFloatData() { return float_feasign_; }
   const std::vector<uint64_t>& GetUint64Data() const { return uint64_feasign_; }
+  std::vector<uint64_t>& MutableUint64Data() { return uint64_feasign_; }
   const std::string& GetType() const { return type_; }
+  std::string& MutableType() { return type_; }
+
+  std::string DebugString() {
+    std::stringstream ss;
+    ss << "\ntype: " << type_ << "\n";
+    ss << "offset: ";
+    ss << "[";
+    for (const size_t& i : offset_) {
+      ss << offset_[i] << ",";
+    }
+    ss << "]\ndata: [";
+    if (type_[0] == 'f') {
+      for (const float& i : float_feasign_) {
+        ss << i << ",";
+      }
+    } else {
+      for (const uint64_t& i : uint64_feasign_) {
+        ss << i << ",";
+      }
+    }
+    ss << "]\n";
+    return ss.str();
+  }
 
  private:
   void CheckType(const std::string& type) const {
@@ -228,13 +343,37 @@ class MultiSlotDataFeed
   virtual ~MultiSlotDataFeed() {}
   virtual void Init(const paddle::framework::DataFeedDesc& data_feed_desc);
   virtual bool CheckFile(const char* filename);
+  // virtual void ReadThread();
 
  protected:
+  virtual void ReadThread();
   virtual void AddInstanceToInsVec(std::vector<MultiSlotType>* vec_ins,
                                    const std::vector<MultiSlotType>& instance,
                                    int index);
   virtual bool ParseOneInstance(std::vector<MultiSlotType>* instance);
+  virtual bool ParseOneInstanceFromPipe(std::vector<MultiSlotType>* instance);
   virtual void PutToFeedVec(const std::vector<MultiSlotType>& ins_vec);
 };
+
+class MultiSlotInMemoryDataFeed
+    : public InMemoryDataFeed<std::vector<MultiSlotType>> {
+ public:
+  MultiSlotInMemoryDataFeed() {}
+  virtual ~MultiSlotInMemoryDataFeed() {}
+  virtual void Init(const paddle::framework::DataFeedDesc& data_feed_desc);
+
+ protected:
+  virtual void AddInstanceToInsVec(std::vector<MultiSlotType>* vec_ins,
+                                   const std::vector<MultiSlotType>& instance,
+                                   int index);
+  virtual bool ParseOneInstance(std::vector<MultiSlotType>* instance);
+  virtual bool ParseOneInstanceFromPipe(std::vector<MultiSlotType>* instance);
+  virtual void PutToFeedVec(const std::vector<MultiSlotType>& ins_vec);
+  virtual void SerializeIns(const std::vector<std::vector<MultiSlotType>*>& ins,
+                            std::string* str);
+  virtual void DeserializeIns(std::vector<std::vector<MultiSlotType>>* ins,
+                              const std::string& str);
+};
+
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/data_feed.proto

@@ -27,4 +27,6 @@ message DataFeedDesc {
   optional string name = 1;
   optional int32 batch_size = 2 [ default = 32 ];
   optional MultiSlotDesc multi_slot_desc = 3;
+  optional string pipe_command = 4;
+  optional int32 thread_num = 5;
 }

paddle/fluid/framework/data_feed_factory.cc

@@ -54,11 +54,15 @@ std::string DataFeedFactory::DataFeedTypeList() {
 std::shared_ptr<DataFeed> DataFeedFactory::CreateDataFeed(
     std::string data_feed_class) {
   if (g_data_feed_map.count(data_feed_class) < 1) {
+    LOG(WARNING) << "Your DataFeed " << data_feed_class
+                 << "is not supported currently";
+    LOG(WARNING) << "Supported DataFeed: " << DataFeedTypeList();
     exit(-1);
   }
   return g_data_feed_map[data_feed_class]();
 }
 
 REGISTER_DATAFEED_CLASS(MultiSlotDataFeed);
+REGISTER_DATAFEED_CLASS(MultiSlotInMemoryDataFeed);
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/data_feed_test.cc

@@ -324,7 +324,7 @@ TEST(DataFeed, MultiSlotUnitTest) {
       load_datafeed_param_from_file(protofile);
   std::vector<MultiTypeSet> reader_elem_set;
   std::vector<MultiTypeSet> 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);
+  // GetElemSetFromReader(&reader_elem_set, data_feed_desc, filelist, 4);
+  // GetElemSetFromFile(&file_elem_set, data_feed_desc, filelist);
+  // CheckIsUnorderedSame(reader_elem_set, file_elem_set);
 }

paddle/fluid/framework/data_set.cc

+/* 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_set.h"
+#include <random>
+#include "google/protobuf/io/zero_copy_stream_impl.h"
+#include "google/protobuf/message.h"
+#include "google/protobuf/text_format.h"
+#include "paddle/fluid/framework/data_feed_factory.h"
+#include "paddle/fluid/framework/io/fs.h"
+#include "paddle/fluid/platform/timer.h"
+
+#if defined _WIN32 || defined __APPLE__
+#else
+#define _LINUX
+#endif
+
+namespace paddle {
+namespace framework {
+
+// constructor
+template <typename T>
+DatasetImpl<T>::DatasetImpl() {
+  thread_num_ = 1;
+  trainer_num_ = 1;
+  file_idx_ = 0;
+}
+
+// set filelist, file_idx_ will reset to zero.
+template <typename T>
+void DatasetImpl<T>::SetFileList(const std::vector<std::string>& filelist) {
+  VLOG(3) << "filelist size: " << filelist.size();
+  filelist_ = filelist;
+  file_idx_ = 0;
+}
+
+// set expect thread num. actually it may change
+template <typename T>
+void DatasetImpl<T>::SetThreadNum(int thread_num) {
+  VLOG(3) << "SetThreadNum thread_num=" << thread_num;
+  thread_num_ = thread_num;
+}
+
+// if you run distributed, and want to do global shuffle,
+// set this before global shuffle.
+// be sure you call CreateReaders before SetTrainerNum
+template <typename T>
+void DatasetImpl<T>::SetTrainerNum(int trainer_num) {
+  trainer_num_ = trainer_num;
+  // should inform reader of trainer_num directly
+  for (auto reader : readers_) {
+    reader->SetTrainerNum(trainer_num);
+  }
+}
+
+template <typename T>
+void DatasetImpl<T>::SetHdfsConfig(const std::string& fs_name,
+                                   const std::string& fs_ugi) {
+  fs_name_ = fs_name;
+  fs_ugi_ = fs_ugi;
+  std::string cmd = std::string("hadoop fs");
+  cmd += " -D fs.default.name=" + fs_name;
+  cmd += " -D hadoop.job.ugi=" + fs_ugi;
+  paddle::framework::hdfs_set_command(cmd);
+}
+
+template <typename T>
+void DatasetImpl<T>::SetDataFeedDesc(const std::string& data_feed_desc_str) {
+  google::protobuf::TextFormat::ParseFromString(data_feed_desc_str,
+                                                &data_feed_desc_);
+}
+
+// readers_.size() may not be equal to thread_num_,
+// it changes when filelist_.size() < thread_num_
+template <typename T>
+std::vector<std::shared_ptr<paddle::framework::DataFeed>>&
+DatasetImpl<T>::GetReaders() {
+  return readers_;
+}
+
+// if sent message between workers, should first call this function
+template <typename T>
+void DatasetImpl<T>::RegisterClientToClientMsgHandler() {
+  auto fleet_ptr = FleetWrapper::GetInstance();
+  VLOG(3) << "RegisterClientToClientMsgHandler";
+  fleet_ptr->RegisterClientToClientMsgHandler(
+      0, [this](int msg_type, int client_id, const std::string& msg) -> int {
+        return this->ReceiveFromClient(msg_type, client_id, msg);
+      });
+  VLOG(3) << "RegisterClientToClientMsgHandler done";
+}
+
+// load data into memory, Dataset hold this memory,
+// which will later be fed into readers' channel
+template <typename T>
+void DatasetImpl<T>::LoadIntoMemory() {
+  VLOG(3) << "DatasetImpl<T>::LoadIntoMemory() begin";
+  platform::Timer timeline;
+  timeline.Start();
+  if (readers_.size() == 0) {
+    CreateReaders();
+  }
+  std::vector<std::thread> load_threads;
+  for (int64_t i = 0; i < thread_num_; ++i) {
+    load_threads.push_back(std::thread(
+        &paddle::framework::DataFeed::LoadIntoMemory, readers_[i].get()));
+  }
+  for (std::thread& t : load_threads) {
+    t.join();
+  }
+  timeline.Pause();
+  VLOG(3) << "DatasetImpl<T>::LoadIntoMemory() end"
+          << ", memory data size=" << memory_data_.size()
+          << ", cost time=" << timeline.ElapsedSec() << " seconds";
+}
+
+// release memory data
+template <typename T>
+void DatasetImpl<T>::ReleaseMemory() {
+  VLOG(3) << "DatasetImpl<T>::ReleaseMemory() begin";
+  std::vector<T>().swap(memory_data_);
+  VLOG(3) << "DatasetImpl<T>::ReleaseMemory() end";
+}
+
+// do local shuffle
+template <typename T>
+void DatasetImpl<T>::LocalShuffle() {
+  VLOG(3) << "DatasetImpl<T>::LocalShuffle() begin";
+  platform::Timer timeline;
+  timeline.Start();
+  if (readers_.size() == 0) {
+    CreateReaders();
+  }
+  // if it is not InMemory, memory_data_ is empty
+  std::random_shuffle(memory_data_.begin(), memory_data_.end());
+
+  std::vector<std::thread> local_shuffle_threads;
+  for (int64_t i = 0; i < thread_num_; ++i) {
+    local_shuffle_threads.push_back(std::thread(
+        &paddle::framework::DataFeed::LocalShuffle, readers_[i].get()));
+  }
+  for (std::thread& t : local_shuffle_threads) {
+    t.join();
+  }
+  std::vector<T>().swap(memory_data_);
+  timeline.Pause();
+  VLOG(3) << "DatasetImpl<T>::LocalShuffle() end, cost time="
+          << timeline.ElapsedSec() << " seconds";
+}
+
+template <typename T>
+void DatasetImpl<T>::GlobalShuffle() {
+  VLOG(3) << "DatasetImpl<T>::GlobalShuffle() begin";
+  platform::Timer timeline;
+  timeline.Start();
+  if (readers_.size() == 0) {
+    CreateReaders();
+  }
+  // if it is not InMemory, memory_data_ is empty
+  std::random_shuffle(memory_data_.begin(), memory_data_.end());
+  VLOG(3) << "start global shuffle threads";
+  std::vector<std::thread> global_shuffle_threads;
+  for (int i = 0; i < thread_num_; ++i) {
+    global_shuffle_threads.push_back(std::thread(
+        &paddle::framework::DataFeed::GlobalShuffle, readers_[i].get()));
+  }
+  for (std::thread& t : global_shuffle_threads) {
+    t.join();
+  }
+  std::vector<T>().swap(memory_data_);
+  timeline.Pause();
+  VLOG(3) << "DatasetImpl<T>::GlobalShuffle() end, cost time="
+          << timeline.ElapsedSec() << " seconds";
+}
+
+template <typename T>
+void DatasetImpl<T>::CreateReaders() {
+  VLOG(3) << "Calling CreateReaders()";
+  CHECK(thread_num_ > 0) << "thread_num should > 0";
+  int file_cnt = filelist_.size();
+  int memory_data_size = memory_data_.size();
+  if (memory_data_size != 0 && thread_num_ > memory_data_size) {
+    VLOG(3) << "Dataset thread num = " << thread_num_
+            << ", memory data size = " << memory_data_size
+            << ". Changing Dataset thread num = " << memory_data_size;
+    thread_num_ = memory_data_size;
+  } else if (file_cnt != 0 && thread_num_ > file_cnt) {
+    VLOG(3) << "Dataset thread num = " << thread_num_
+            << ", file num = " << file_cnt
+            << ". Changing Dataset thread num = " << file_cnt;
+    thread_num_ = file_cnt;
+  }
+  VLOG(3) << "thread_num in Readers: " << thread_num_;
+  VLOG(3) << "readers size: " << readers_.size();
+  VLOG(3) << "Filelist size in readers: " << filelist_.size();
+  if (readers_.size() != 0) {
+    return;
+  }
+  VLOG(3) << "data feed class name: " << data_feed_desc_.name();
+  for (int i = 0; i < thread_num_; ++i) {
+    readers_.push_back(DataFeedFactory::CreateDataFeed(data_feed_desc_.name()));
+    readers_.back()->Init(data_feed_desc_);
+    readers_.back()->SetMemoryData(&memory_data_);
+    readers_.back()->SetMemoryDataMutex(&mutex_for_update_memory_data_);
+    readers_.back()->SetThreadId(i);
+    readers_.back()->SetThreadNum(thread_num_);
+    readers_.back()->SetTrainerNum(trainer_num_);
+    readers_.back()->SetFileListMutex(&mutex_for_pick_file_);
+    readers_.back()->SetFileListIndex(&file_idx_);
+    readers_.back()->SetFileList(filelist_);
+  }
+}
+
+template <typename T>
+void DatasetImpl<T>::DestroyReaders() {
+  VLOG(3) << "Calling DestroyReaders()";
+  // clear memory_data_ before fill it
+  // because if LoadIntoMemory but no Shuffle,
+  // memory_data_ has empty data which has been std::move to channel
+  if (memory_data_.size() != 0) {
+    std::vector<T>().swap(memory_data_);
+  }
+  std::vector<std::thread> fill_threads;
+  for (int i = 0; i < thread_num_; ++i) {
+    fill_threads.push_back(
+        std::thread(&paddle::framework::DataFeed::FillChannelToMemoryData,
+                    readers_[i].get()));
+  }
+  for (std::thread& t : fill_threads) {
+    t.join();
+  }
+  std::vector<std::shared_ptr<paddle::framework::DataFeed>>().swap(readers_);
+  VLOG(3) << "readers size: " << readers_.size();
+  // if memory_data_ is empty, which means it's not InMemory mode,
+  // so the next epoch should read all data again
+  if (memory_data_.size() == 0) {
+    file_idx_ = 0;
+  }
+}
+
+template <typename T>
+int DatasetImpl<T>::ReceiveFromClient(int msg_type, int client_id,
+                                      const std::string& msg) {
+#ifdef _LINUX
+  VLOG(3) << "ReceiveFromClient msg_type=" << msg_type
+          << ", client_id=" << client_id << ", msg length=" << msg.length();
+  auto fleet_ptr = FleetWrapper::GetInstance();
+  int64_t index = rand_r(&rand_seed) % thread_num_;
+  VLOG(3) << "ramdom index=" << index;
+  readers_[index]->PutInsToChannel(msg);
+#endif
+  return 0;
+}
+
+// explicit instantiation
+template class DatasetImpl<std::vector<MultiSlotType>>;
+
+}  // end namespace framework
+}  // end namespace paddle

paddle/fluid/framework/data_set.h

+/* 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 <fstream>
+#include <memory>
+#include <mutex>  // NOLINT
+#include <string>
+#include <thread>  // NOLINT
+#include <utility>
+#include <vector>
+
+#include "paddle/fluid/framework/data_feed.h"
+
+namespace paddle {
+namespace framework {
+
+// Dataset is a abstract class, which defines user interfaces
+// Example Usage:
+//    Dataset* dataset = DatasetFactory::CreateDataset("InMemoryDataset")
+//    dataset->SetFileList(std::vector<std::string>{"a.txt", "b.txt"})
+//    dataset->SetThreadNum(1)
+//    dataset->CreateReaders();
+//    dataset->SetDataFeedDesc(your_data_feed_desc);
+//    dataset->LoadIntoMemory();
+//    dataset->SetTrainerNum(2);
+//    dataset->GlobalShuffle();
+class Dataset {
+ public:
+  Dataset() {}
+  virtual ~Dataset() {}
+  // set file list
+  virtual void SetFileList(const std::vector<std::string>& filelist) = 0;
+  // set readers' num
+  virtual void SetThreadNum(int thread_num) = 0;
+  // set workers' num
+  virtual void SetTrainerNum(int trainer_num) = 0;
+  // set fs name and ugi
+  virtual void SetHdfsConfig(const std::string& fs_name,
+                             const std::string& fs_ugi) = 0;
+  // set data fedd desc, which contains:
+  //   data feed name, batch size, slots
+  virtual void SetDataFeedDesc(const std::string& data_feed_desc_str) = 0;
+  // get file list
+  virtual const std::vector<std::string>& GetFileList() = 0;
+  // get thread num
+  virtual int GetThreadNum() = 0;
+  // get worker num
+  virtual int GetTrainerNum() = 0;
+  // get hdfs config
+  virtual std::pair<std::string, std::string> GetHdfsConfig() = 0;
+  // get data fedd desc
+  virtual const paddle::framework::DataFeedDesc& GetDataFeedDesc() = 0;
+  // get readers, the reader num depend both on thread num
+  // and filelist size
+  virtual std::vector<std::shared_ptr<paddle::framework::DataFeed>>&
+  GetReaders() = 0;
+  // register message handler between workers
+  virtual void RegisterClientToClientMsgHandler() = 0;
+  // load all data into memory
+  virtual void LoadIntoMemory() = 0;
+  // release all memory data
+  virtual void ReleaseMemory() = 0;
+  // local shuffle data
+  virtual void LocalShuffle() = 0;
+  // global shuffle data
+  virtual void GlobalShuffle() = 0;
+  // create readers
+  virtual void CreateReaders() = 0;
+  // destroy readers
+  virtual void DestroyReaders() = 0;
+
+ protected:
+  virtual int ReceiveFromClient(int msg_type, int client_id,
+                                const std::string& msg) = 0;
+};
+
+// DatasetImpl is the implementation of Dataset,
+// it holds memory data if user calls load_into_memory
+template <typename T>
+class DatasetImpl : public Dataset {
+ public:
+  DatasetImpl();
+  virtual ~DatasetImpl() {}
+
+  virtual void SetFileList(const std::vector<std::string>& filelist);
+  virtual void SetThreadNum(int thread_num);
+  virtual void SetTrainerNum(int trainer_num);
+  virtual void SetHdfsConfig(const std::string& fs_name,
+                             const std::string& fs_ugi);
+  virtual void SetDataFeedDesc(const std::string& data_feed_desc_str);
+
+  virtual const std::vector<std::string>& GetFileList() { return filelist_; }
+  virtual int GetThreadNum() { return thread_num_; }
+  virtual int GetTrainerNum() { return trainer_num_; }
+  virtual std::pair<std::string, std::string> GetHdfsConfig() {
+    return std::make_pair(fs_name_, fs_ugi_);
+  }
+  virtual const paddle::framework::DataFeedDesc& GetDataFeedDesc() {
+    return data_feed_desc_;
+  }
+  virtual std::vector<std::shared_ptr<paddle::framework::DataFeed>>&
+  GetReaders();
+
+  virtual void RegisterClientToClientMsgHandler();
+  virtual void LoadIntoMemory();
+  virtual void ReleaseMemory();
+  virtual void LocalShuffle();
+  virtual void GlobalShuffle();
+  virtual void CreateReaders();
+  virtual void DestroyReaders();
+
+ protected:
+  virtual int ReceiveFromClient(int msg_type, int client_id,
+                                const std::string& msg);
+  std::vector<std::shared_ptr<paddle::framework::DataFeed>> readers_;
+  std::vector<T> memory_data_;
+  std::mutex mutex_for_update_memory_data_;
+  int thread_num_;
+  paddle::framework::DataFeedDesc data_feed_desc_;
+  int trainer_num_;
+  std::vector<std::string> filelist_;
+  size_t file_idx_;
+  std::mutex mutex_for_pick_file_;
+  std::string fs_name_;
+  std::string fs_ugi_;
+  unsigned int rand_seed;
+};
+
+// use std::vector<MultiSlotType> as data type
+class MultiSlotDataset : public DatasetImpl<std::vector<MultiSlotType>> {
+ public:
+  MultiSlotDataset() {}
+  virtual ~MultiSlotDataset() {}
+};
+
+}  // end namespace framework
+}  // end namespace paddle

paddle/fluid/framework/dataset_factory.cc

+/* 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/dataset_factory.h"
+#include <memory>
+#include <string>
+#include <unordered_map>
+
+#include "paddle/fluid/framework/data_set.h"
+
+namespace paddle {
+namespace framework {
+typedef std::shared_ptr<Dataset> (*CreateDatasetFunction)();
+typedef std::unordered_map<std::string, CreateDatasetFunction> datasetMap;
+datasetMap g_dataset_map;
+
+#define REGISTER_DATASET_CLASS(dataset_class)                   \
+  namespace {                                                   \
+  std::shared_ptr<Dataset> Creator_##dataset_class() {          \
+    return std::shared_ptr<Dataset>(new dataset_class);         \
+  }                                                             \
+  class __Registerer_##dataset_class {                          \
+   public:                                                      \
+    __Registerer_##dataset_class() {                            \
+      g_dataset_map[#dataset_class] = &Creator_##dataset_class; \
+    }                                                           \
+  };                                                            \
+  __Registerer_##dataset_class g_registerer_##dataset_class;    \
+  }  // namespace
+
+std::string DatasetFactory::DatasetTypeList() {
+  std::string dataset_types;
+  for (auto iter = g_dataset_map.begin(); iter != g_dataset_map.end(); ++iter) {
+    if (iter != g_dataset_map.begin()) {
+      dataset_types += ", ";
+    }
+    dataset_types += iter->first;
+  }
+  return dataset_types;
+}
+
+std::shared_ptr<Dataset> DatasetFactory::CreateDataset(
+    std::string dataset_class) {
+  if (g_dataset_map.count(dataset_class) < 1) {
+    LOG(WARNING) << "Your Dataset " << dataset_class
+                 << "is not supported currently";
+    LOG(WARNING) << "Supported Dataset: " << DatasetTypeList();
+    exit(-1);
+  }
+  return g_dataset_map[dataset_class]();
+}
+
+REGISTER_DATASET_CLASS(MultiSlotDataset);
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/operators/ngraph/ngraph_ops.h → paddle/fluid/framework/dataset_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,
@@ -12,26 +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. */
 
-/*
- * This file contains the list of the ngraph operators for Paddle.
- *
- * ATTENTION: It requires some C++11 features, for lower version C++ or C, we
- * might release another API.
- */
-
 #pragma once
 
-#include "ops/accuracy_op.h"
-#include "ops/activation_op.h"
-#include "ops/batch_norm_op.h"
-#include "ops/binary_unary_op.h"
-#include "ops/conv2d_op.h"
-#include "ops/elementwise_add_op.h"
-#include "ops/fill_constant_op.h"
-#include "ops/mean_op.h"
-#include "ops/mul_op.h"
-#include "ops/pool2d_op.h"
-#include "ops/scale_op.h"
-#include "ops/softmax_op.h"
-#include "ops/sum_op.h"
-#include "ops/top_k_op.h"
+#include <memory>
+#include <string>
+#include "paddle/fluid/framework/data_set.h"
+
+namespace paddle {
+namespace framework {
+class DatasetFactory {
+ public:
+  static std::string DatasetTypeList();
+  static std::shared_ptr<Dataset> CreateDataset(std::string dataset_class);
+};
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/details/CMakeLists.txt

@@ -5,11 +5,16 @@ cc_library(scale_loss_grad_op_handle SRCS scale_loss_grad_op_handle.cc DEPS op_h
 cc_library(fetch_op_handle SRCS fetch_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory)
 cc_library(computation_op_handle SRCS computation_op_handle.cc DEPS framework_proto scope place operator op_registry)
 cc_library(rpc_op_handle SRCS rpc_op_handle.cc DEPS framework_proto scope place operator op_registry)
+cc_library(fetch_barrier_op_handle SRCS fetch_barrier_op_handle.cc DEPS framework_proto scope place operator op_registry)
 
 cc_library(multi_devices_helper SRCS multi_devices_helper.cc DEPS graph graph_helper)
 cc_library(multi_devices_graph_print_pass SRCS multi_devices_graph_print_pass.cc DEPS multi_devices_helper)
 cc_library(multi_devices_graph_check_pass SRCS multi_devices_graph_check_pass.cc DEPS multi_devices_helper)
 
+cc_library(alloc_continuous_space_for_grad_pass SRCS alloc_continuous_space_for_grad_pass.cc DEPS graph graph_helper)
+cc_library(fuse_adam_op_pass SRCS fuse_adam_op_pass.cc fuse_optimizer_op_pass.cc DEPS graph graph_helper)
+cc_library(fuse_sgd_op_pass SRCS fuse_sgd_op_pass.cc fuse_optimizer_op_pass.cc DEPS graph graph_helper)
+
 cc_library(variable_visitor SRCS variable_visitor.cc DEPS lod_tensor selected_rows)
 
 if(WITH_DISTRIBUTE)
@@ -20,7 +25,13 @@ if(WITH_DISTRIBUTE)
 endif()
 
 if(WITH_GPU)
+    set(dgc_deps "")
+    if(NOT WIN32)
+        set(dgc_deps dgc)
+    endif()
     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 ${dgc_deps})
+    nv_library(fused_all_reduce_op_handle SRCS fused_all_reduce_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory
             dynload_cuda variable_visitor)
     if(WITH_DISTRIBUTE)
         nv_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope
@@ -35,6 +46,8 @@ if(WITH_GPU)
 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(fused_all_reduce_op_handle SRCS fused_all_reduce_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory
+            variable_visitor)
     if(WITH_DISTRIBUTE)
         cc_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope
             ddim selected_rows_functor sendrecvop_rpc)
@@ -46,24 +59,30 @@ else()
     cc_library(fused_broadcast_op_handle SRCS fused_broadcast_op_handle.cc DEPS broadcast_op_handle)
 endif()
 
-cc_library(data_balance_op_handle SRCS data_balance_op_handle.cc DEPS op_handle_base scope lod_tensor)
 cc_library(gather_op_handle SRCS gather_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor)
-cc_library(fuse_vars_op_handle SRCS fuse_vars_op_handle.cc DEPS op_handle_base scope)
 
-cc_library(memory_optimize_helper SRCS memory_optimize_helper.cc DEPS graph graph_helper)
+if(WITH_GPU)
+cc_library(memory_optimize_helper SRCS memory_optimize_helper.cc DEPS graph graph_helper gpu_info)
+else()
+cc_library(memory_optimize_helper SRCS memory_optimize_helper.cc DEPS graph graph_helper cpu_info)
+endif()
+
 cc_library(memory_optimize_pass SRCS memory_optimize_pass.cc DEPS memory_optimize_helper pass)
 cc_library(inplace_op_pass SRCS inplace_op_pass.cc DEPS memory_optimize_pass op_info)
 cc_library(modify_op_lock_and_record_event_pass SRCS modify_op_lock_and_record_event_pass.cc DEPS computation_op_handle op_graph_view multi_devices_helper)
 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(while_op_eager_deletion_pass SRCS while_op_eager_deletion_pass.cc DEPS while_op_helper graph_helper pass computation_op_handle)
+cc_library(eager_deletion_pass SRCS eager_deletion_pass.cc DEPS computation_op_handle eager_deletion_op_handle graph graph_helper pass while_op_eager_deletion_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)
+        scale_loss_grad_op_handle rpc_op_handle fetch_barrier_op_handle all_reduce_op_handle reduce_op_handle broadcast_op_handle fused_broadcast_op_handle)
+
+cc_library(fuse_all_reduce_op_pass SRCS fuse_all_reduce_op_pass.cc DEPS graph graph_helper fused_all_reduce_op_handle)
 
 set(SSA_GRAPH_EXECUTOR_DEPS graph framework_proto sequential_execution_pass modify_op_lock_and_record_event_pass all_reduce_deps_pass reference_count_pass eager_deletion_pass memory_optimize_pass inplace_op_pass)
 if (WITH_GPU)
@@ -77,6 +96,12 @@ cc_library(threaded_ssa_graph_executor SRCS threaded_ssa_graph_executor.cc DEPS
 
 cc_library(parallel_ssa_graph_executor SRCS parallel_ssa_graph_executor.cc DEPS threaded_ssa_graph_executor)
 
+set(ASYNC_SSA_GRAPH_EXECUTOR_DEPS threaded_ssa_graph_executor)
+if(WITH_DISTRIBUTE)
+    list(APPEND ASYNC_SSA_GRAPH_EXECUTOR_DEPS communicator)
+endif()
+cc_library(async_ssa_graph_executor SRCS async_ssa_graph_executor.cc DEPS ${ASYNC_SSA_GRAPH_EXECUTOR_DEPS})
+
 cc_test(broadcast_op_test SRCS broadcast_op_handle_test.cc DEPS var_handle op_handle_base scope ddim memory
         device_context broadcast_op_handle)
 cc_test(gather_op_test SRCS gather_op_handle_test.cc DEPS var_handle op_handle_base scope ddim memory
@@ -93,4 +118,6 @@ cc_library(build_strategy SRCS build_strategy.cc DEPS
         multi_devices_graph_print_pass multi_devices_graph_check_pass
         fuse_elewise_add_act_pass multi_batch_merge_pass 
         fuse_relu_depthwise_conv_pass
-        memory_optimize_pass lock_free_optimize_pass)
+        memory_optimize_pass lock_free_optimize_pass
+        alloc_continuous_space_for_grad_pass fuse_all_reduce_op_pass
+        fuse_adam_op_pass fuse_sgd_op_pass)

paddle/fluid/framework/details/all_reduce_deps_pass.cc

@@ -13,6 +13,7 @@
 // limitations under the License.
 
 #include <algorithm>
+#include <memory>
 #include <string>
 #include <unordered_map>
 #include <unordered_set>
@@ -30,8 +31,6 @@ namespace paddle {
 namespace framework {
 namespace details {
 
-static constexpr char kAllOpDescs[] = "all_op_descs";
-
 VarHandle* GetValidInput(const OpHandleBase* a) {
   for (auto p : a->Inputs()) {
     VarHandle* b = dynamic_cast<VarHandle*>(p);
@@ -43,8 +42,7 @@ VarHandle* GetValidInput(const OpHandleBase* a) {
   return nullptr;
 }
 
-std::unique_ptr<ir::Graph> AllReduceDepsPass::ApplyImpl(
-    std::unique_ptr<ir::Graph> graph) const {
+void AllReduceDepsPass::ApplyImpl(ir::Graph* graph) const {
   auto graph_ops = ir::FilterByNodeWrapper<OpHandleBase>(*graph);
 
   // get vars order
@@ -52,15 +50,30 @@ std::unique_ptr<ir::Graph> AllReduceDepsPass::ApplyImpl(
   std::unordered_map<std::string, int> vars;
   // TODO(gongwb): use graph topology sort to find the order of operators.
   //               Note that must assert topology sort is stable
-  auto& ops = Get<const std::vector<OpDesc*>>(kAllOpDescs);
+  auto& ops = graph->Get<const std::vector<OpDesc*>>(kStaleProgramOpDescs);
   for (auto* op_desc : ops) {
-    auto outputs = op_desc->Outputs();
-    for (auto& o_it : outputs) {
-      for (auto& v : o_it.second) {  // values
-        vars[v] = order;
+    try {
+      bool is_bk_op =
+          static_cast<bool>(boost::get<int>(op_desc->GetAttr(
+                                OpProtoAndCheckerMaker::OpRoleAttrName())) &
+                            static_cast<int>(OpRole::kBackward));
+      if (!is_bk_op) continue;
+
+      auto backward_vars =
+          boost::get<std::vector<std::string>>(op_desc->GetNullableAttr(
+              OpProtoAndCheckerMaker::OpRoleVarAttrName()));
+      PADDLE_ENFORCE_EQ(backward_vars.size() % 2, 0);
+
+      auto outputs = op_desc->Outputs();
+      for (auto& o_it : outputs) {
+        for (auto& v : o_it.second) {  // values
+          vars[v] = order;
+          VLOG(10) << "in all_reduce_deps_pass:" << v;
+        }
       }
+      order++;
+    } catch (boost::bad_get e) {
     }
-    order++;
   }
 
   std::vector<OpHandleBase*> dist_ops;
@@ -72,7 +85,8 @@ std::unique_ptr<ir::Graph> AllReduceDepsPass::ApplyImpl(
     }
   }
 
-  VLOG(10) << "dist_ops size:" << dist_ops.size() << std::endl;
+  VLOG(10) << "dist_ops size:" << dist_ops.size()
+           << ", outputs size:" << vars.size() << ", ops size:" << ops.size();
 
   std::sort(dist_ops.begin(), dist_ops.end(), [&](OpHandleBase* op1,
                                                   OpHandleBase* op2) {
@@ -85,6 +99,10 @@ std::unique_ptr<ir::Graph> AllReduceDepsPass::ApplyImpl(
     auto l_it = vars.find(i0->name());
     auto r_it = vars.find(i1->name());
 
+    PADDLE_ENFORCE(l_it != vars.end() && r_it != vars.end(),
+                   "can't find var's name %s and %s in opdesc", i0->name(),
+                   i1->name());
+
     if (l_it->second < r_it->second) return true;
 
     if (l_it->second == r_it->second) {
@@ -112,8 +130,6 @@ std::unique_ptr<ir::Graph> AllReduceDepsPass::ApplyImpl(
     VLOG(10) << "pre_op:" << pre_op->DebugString()
              << ", op:" << op->DebugString();
   }
-
-  return graph;
 }
 
 }  // namespace details
@@ -122,4 +138,4 @@ std::unique_ptr<ir::Graph> AllReduceDepsPass::ApplyImpl(
 
 REGISTER_PASS(all_reduce_deps_pass,
               paddle::framework::details::AllReduceDepsPass)
-    .RequirePassAttr(paddle::framework::details::kAllOpDescs);
+    .RequireGraphAttr(paddle::framework::details::kStaleProgramOpDescs);

paddle/fluid/framework/details/all_reduce_deps_pass.h

@@ -24,8 +24,7 @@ namespace details {
 // TODO(gongwb): overlap allreduce with backward computation.
 class AllReduceDepsPass : public ir::Pass {
  protected:
-  std::unique_ptr<ir::Graph> ApplyImpl(
-      std::unique_ptr<ir::Graph> graph) const override;
+  void ApplyImpl(ir::Graph* graph) const override;
 };
 
 }  // namespace details

paddle/fluid/framework/details/all_reduce_op_handle.cc

@@ -11,12 +11,18 @@
 // 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 <algorithm>
-
 #include "paddle/fluid/framework/details/all_reduce_op_handle.h"
+#include <algorithm>
 #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"
+#include "paddle/fluid/framework/operator.h"
+
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+#include "dgc/dgc.h"
+#endif
+
+#include "paddle/fluid/platform/gpu_info.h"
 #include "paddle/fluid/platform/profiler.h"
 
 // asynchronous nccl allreduce or synchronous issue:
@@ -34,11 +40,14 @@ namespace details {
 AllReduceOpHandle::AllReduceOpHandle(ir::Node *node,
                                      const std::vector<Scope *> &local_scopes,
                                      const std::vector<platform::Place> &places,
-                                     const platform::NCCLContextMap *ctxs)
+                                     const platform::NCCLContextMap *ctxs,
+                                     bool is_encoded, int nranks)
     : OpHandleBase(node),
       local_scopes_(local_scopes),
       places_(places),
-      nccl_ctxs_(ctxs) {
+      nccl_ctxs_(ctxs),
+      is_encoded_(is_encoded),
+      nranks_(nranks) {
   if (nccl_ctxs_) {
     for (auto &p : places_) {
       this->SetDeviceContext(p, nccl_ctxs_->DevCtx(p));
@@ -52,10 +61,189 @@ AllReduceOpHandle::AllReduceOpHandle(ir::Node *node,
     : OpHandleBase(node), local_scopes_(local_scopes), places_(places) {}
 #endif
 
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+void AllReduceOpHandle::RunImplEncoded() {
+  platform::RecordEvent record_event(Name());
+
+  WaitInputVarGenerated();
+
+  auto in_var_handles = DynamicCast<VarHandle>(this->Inputs());
+  auto out_var_handles = DynamicCast<VarHandle>(this->Outputs());
+  PADDLE_ENFORCE_EQ(
+      in_var_handles.size(), places_.size(),
+      "The NoDummyInputSize should be equal to the number of places.");
+  PADDLE_ENFORCE_EQ(
+      in_var_handles.size(), out_var_handles.size(),
+      "The NoDummyInputSize and NoDummyOutputSize should be equal.");
+
+  std::vector<const LoDTensor *> ins;
+  std::vector<LoDTensor *> outs;
+  int k = -1;
+  for (size_t i = 0; i < local_scopes_.size(); ++i) {
+    auto &local_scope =
+        local_scopes_[i]->FindVar(kLocalExecScopeName)->Get<Scope *>();
+    auto original_name =
+        paddle::framework::GradOriginalVarName(in_var_handles[i]->name());
+    auto encode_var_name = original_name + g_dgc_encoded;
+    auto *in_var = local_scope->FindVar(encode_var_name);
+    PADDLE_ENFORCE_NOT_NULL(in_var);
+    auto &in = in_var->Get<LoDTensor>();
+    ins.emplace_back(&in);
+
+    auto *out = local_scope->FindVar(out_var_handles[i]->name())
+                    ->GetMutable<LoDTensor>();
+    outs.emplace_back(out);
+
+    if (k < 0) {
+      k = GetKValue(in_var_handles[i]->name());
+    }
+  }
+
+  PADDLE_ENFORCE(platform::is_gpu_place(ins[0]->place()));
+  PADDLE_ENFORCE(platform::is_gpu_place(outs[0]->place()));
+  PADDLE_ENFORCE(nccl_ctxs_, "nccl_ctxs should not be nullptr.");
+
+  int dtype = -1;
+  size_t in_numel = 0;
+  size_t out_numel = 0;
+  PADDLE_ENFORCE(nranks_ > 1);
+  std::vector<std::function<void()>> all_reduce_calls;
+
+  for (size_t i = 0; i < local_scopes_.size(); ++i) {
+    auto &place = places_[i];
+    auto &in = *ins[i];
+    void *in_tensor_buf = const_cast<void *>(in.data<void>());
+
+    auto &out = *outs[i];
+    float *out_tensor_buf = out.data<float>();
+
+    dtype = (dtype == -1) ? platform::ToNCCLDataType(in.type()) : dtype;
+    in_numel = (in_numel == 0) ? static_cast<size_t>(in.numel()) : in_numel;
+    PADDLE_ENFORCE(in_numel % 2 == 0);
+    PADDLE_ENFORCE(in_numel / 2 == static_cast<size_t>(k));
+    out_numel = (out_numel == 0) ? static_cast<size_t>(out.numel()) : out_numel;
+
+    int dev_id = boost::get<platform::CUDAPlace>(place).device;
+    auto &nccl_ctx = nccl_ctxs_->at(dev_id);
+    auto stream = nccl_ctx.stream();
+    auto comm = nccl_ctx.comm_;
+
+    auto &allocator =
+        platform::DeviceTemporaryAllocator::Instance().Get(place, stream);
+    int encode_size = 2 * k * sizeof(int);
+    // dgc use ncclAllGather to get all the encoded data
+    // so the buffer need nranks.
+    int buf_size = nranks_ * encode_size;
+    auto tmp_ious_data = allocator.Allocate(buf_size);
+    void *gather_buff = reinterpret_cast<void *>(tmp_ious_data->ptr());
+
+    VLOG(10) << "in_numel:" << in_numel << ", out_numel:" << out_numel
+             << ", nranks:" << nranks_ << ", gather_buf size:" << buf_size
+             << ", k:" << k << ", place:" << place << ", dtype:" << dtype;
+
+    all_reduce_calls.emplace_back([=] {
+      PADDLE_ENFORCE(paddle::communication::dgc::sparseAllGReduce(
+          in_tensor_buf, gather_buff, k, out_tensor_buf, out_numel, comm,
+          stream));
+    });
+  }
+
+  this->RunAndRecordEvent([&] {
+    if (all_reduce_calls.size() == 1UL) {
+      // Do not use NCCLGroup when manage NCCL by per thread per device
+      all_reduce_calls[0]();
+    } else {
+      platform::NCCLGroupGuard guard;
+      for (auto &call : all_reduce_calls) {
+        call();
+      }
+    }
+  });
+
+  if (FLAGS_sync_nccl_allreduce) {
+    for (auto &p : places_) {
+      int dev_id = boost::get<platform::CUDAPlace>(p).device;
+      auto &nccl_ctx = nccl_ctxs_->at(dev_id);
+      auto stream = nccl_ctx.stream();
+      cudaError_t e_sync = cudaStreamSynchronize(stream);
+      if (e_sync != 0) {
+        LOG(FATAL) << "cudaStreamSynchronize " << cudaGetErrorString(e_sync);
+      }
+
+      cudaError_t e_get = cudaGetLastError();
+      if (e_get != 0) {
+        LOG(FATAL) << "cudaGetLastError  " << cudaGetErrorString(e_get)
+                   << " errno:" << e_get;
+      }
+    }
+  }
+}
+
+int AllReduceOpHandle::GetKValue(const std::string &grad_name) {
+  auto original_name = paddle::framework::GradOriginalVarName(grad_name);
+  auto var_name = original_name + g_dgc_k;
+  PADDLE_ENFORCE(local_scopes_.size() > 0);
+
+  auto *scope = local_scopes_[0];
+  auto &local_scope = scope->FindVar(kLocalExecScopeName)->Get<Scope *>();
+  auto var = local_scope->FindVar(var_name);
+  PADDLE_ENFORCE_NOT_NULL(var);
+  auto tensor = var->Get<LoDTensor>().data<float>();
+  return *tensor;
+}
+#endif
+
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+bool AllReduceOpHandle::IsEncoded() {
+  if (!is_encoded_) {
+    return false;
+  }
+  auto counter_name = g_dgc_counter_name;
+  auto step_name = g_dgc_rampup_begin_step;
+  PADDLE_ENFORCE(local_scopes_.size() > 0);
+
+  auto *scope = local_scopes_[0];
+  auto &local_scope = scope->FindVar(kLocalExecScopeName)->Get<Scope *>();
+  auto count_var = local_scope->FindVar(counter_name);
+  auto step_var = local_scope->FindVar(step_name);
+  if (count_var == nullptr || step_var == nullptr) {
+    PADDLE_THROW("not find count_var:%s or step_var:%s", counter_name,
+                 step_var);
+  }
+
+  float count = *count_var->Get<LoDTensor>().data<float>();
+  float step = *step_var->Get<LoDTensor>().data<float>();
+  if (static_cast<int>(count) < static_cast<int>(step)) {
+    VLOG(10) << "in all_reduce currentstep:" << count
+             << " < rampup_begin_step:" << step
+             << " so not use sparse all reduce";
+    return false;
+  }
+
+  return true;
+}
+#else
+bool AllReduceOpHandle::IsEncoded() { return false; }
+#endif
+
 void AllReduceOpHandle::RunImpl() {
-  platform::RecordEvent record_event(Name(), dev_ctxes_.cbegin()->second);
+  if (!IsEncoded()) {
+    RunImplNormal();
+    return;
+  }
+
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+  RunImplEncoded();
+#else
+  PADDLE_THROW("Not compiled with CUDA");
+#endif
+}
+
+void AllReduceOpHandle::RunImplNormal() {
+  platform::RecordEvent record_event(Name());
 
   WaitInputVarGenerated();
+
   auto in_var_handles = DynamicCast<VarHandle>(this->Inputs());
   auto out_var_handles = DynamicCast<VarHandle>(this->Outputs());
   PADDLE_ENFORCE_EQ(
@@ -72,6 +260,8 @@ void AllReduceOpHandle::RunImpl() {
     auto &lod_tensor =
         local_scope.FindVar(in_var_handles[i]->name())->Get<LoDTensor>();
     lod_tensors.emplace_back(&lod_tensor);
+    VLOG(10) << "place:" << i << ", input_name:" << in_var_handles[i]->name()
+             << ", out_name:" << out_var_handles[i]->name();
     PADDLE_ENFORCE_EQ(in_var_handles[i]->name(), out_var_handles[i]->name(),
                       "The name of input and output should be equal.");
   }
@@ -99,13 +289,17 @@ void AllReduceOpHandle::RunImpl() {
       auto &nccl_ctx = nccl_ctxs_->at(dev_id);
       auto stream = nccl_ctx.stream();
       auto comm = nccl_ctx.comm_;
+
+      VLOG(10) << "before all reduce buffer:" << buffer << ", numel:" << numel
+               << ", dev_id:" << dev_id << ", dtype:" << dtype
+               << ", place:" << p;
+
       all_reduce_calls.emplace_back([=] {
         PADDLE_ENFORCE(platform::dynload::ncclAllReduce(
             buffer, buffer, numel, static_cast<ncclDataType_t>(dtype), ncclSum,
             comm, stream));
       });
     }
-
     this->RunAndRecordEvent([&] {
       if (all_reduce_calls.size() == 1UL) {
         // Do not use NCCLGroup when manage NCCL by per thread per device

paddle/fluid/framework/details/all_reduce_op_handle.h

@@ -28,11 +28,19 @@ namespace paddle {
 namespace framework {
 namespace details {
 
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+constexpr char g_dgc_counter_name[] = "__g_dgc_counter__";
+constexpr char g_dgc_rampup_begin_step[] = "__g_rampup_begin_step__";
+constexpr char g_dgc_encoded[] = "__dgc_encoded__";
+constexpr char g_dgc_k[] = "__dgc_k__";
+#endif
+
 struct AllReduceOpHandle : public OpHandleBase {
 #if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   AllReduceOpHandle(ir::Node *node, const std::vector<Scope *> &local_scopes,
                     const std::vector<platform::Place> &places,
-                    const platform::NCCLContextMap *ctxs);
+                    const platform::NCCLContextMap *ctxs,
+                    bool is_encoded = false, int nranks = -1);
 #else
   AllReduceOpHandle(ir::Node *node, const std::vector<Scope *> &local_scopes,
                     const std::vector<platform::Place> &places);
@@ -50,8 +58,14 @@ struct AllReduceOpHandle : public OpHandleBase {
   std::vector<Scope *> local_scopes_;
   std::vector<platform::Place> places_;
 #if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+  void RunImplEncoded();
   const platform::NCCLContextMap *nccl_ctxs_;
+  bool is_encoded_{false};
+  int nranks_{-1};
+  int GetKValue(const std::string &grad_name);
 #endif
+  void RunImplNormal();
+  bool IsEncoded();
 };
 
 }  // namespace details

paddle/fluid/framework/details/async_ssa_graph_executor.cc

+//   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/async_ssa_graph_executor.h"
+
+#include "paddle/fluid/framework/variable_helper.h"
+
+#ifdef PADDLE_WITH_DISTRIBUTE
+#include "paddle/fluid/operators/distributed/communicator.h"
+#endif
+
+namespace paddle {
+namespace framework {
+namespace details {
+
+inline void NewTempScopeAndInitVars(const std::vector<VarInfo> &var_infos,
+                                    Scope *scope) {
+  VLOG(3) << "NewTempScopeAndInitVars";
+  Scope &local_scope = scope->NewScope();
+  *scope->Var(details::kLocalExecScopeName)->GetMutable<Scope *>() =
+      &local_scope;
+
+  for (auto &info : var_infos) {
+    if (scope->FindVar(info.name_) != nullptr) {
+      continue;
+    }
+
+    if (info.persistable_) {  // Persistable
+      InitializeVariable(scope->Var(info.name_), info.type_);
+    } else {
+      InitializeVariable(local_scope.Var(info.name_), info.type_);
+    }
+  }
+}
+
+// get RpcContext and remote send and recv op
+void ProcessGraph(std::vector<ir::Graph *> graphs, Scope *scope) {
+#ifdef PADDLE_WITH_DISTRIBUTE
+  using RpcCtxMap = operators::distributed::RpcCtxMap;
+  VLOG(3) << "ProcessGraph";
+  RpcCtxMap send_varname_to_ctx;
+  RpcCtxMap recv_varname_to_ctx;
+  for (auto i = 0; i < graphs.size(); ++i) {
+    std::vector<ir::Node *> nodes_to_delete;
+    for (auto &node : graphs[i]->Nodes()) {
+      VLOG(3) << "node name " << node->Name();
+      if (node && node->IsOp()) {
+        if (node->Name() == "send") {
+          auto send_var_name = node->Op()->Input("X")[0];
+          auto send_varnames = boost::get<std::vector<std::string>>(
+              node->Op()->GetNullableAttr("send_varnames"));
+          auto epmap = boost::get<std::vector<std::string>>(
+              node->Op()->GetNullableAttr("epmap"));
+          auto height_section = boost::get<std::vector<int64_t>>(
+              node->Op()->GetNullableAttr("sections"));
+          send_varname_to_ctx[send_var_name] =
+              operators::distributed::RpcContext(send_var_name, send_varnames,
+                                                 epmap, height_section);
+          VLOG(3) << "find and init an send op: "
+                  << send_varname_to_ctx[send_var_name];
+        } else if (node->Name() == "recv") {
+          auto recv_var_name = node->Op()->Output("Out")[0];
+          auto recv_varnames = boost::get<std::vector<std::string>>(
+              node->Op()->GetNullableAttr("recv_varnames"));
+          auto epmap = boost::get<std::vector<std::string>>(
+              node->Op()->GetNullableAttr("epmap"));
+          recv_varname_to_ctx[recv_var_name] =
+              operators::distributed::RpcContext(recv_var_name, recv_varnames,
+                                                 epmap, {});
+          nodes_to_delete.push_back(node);
+          VLOG(3) << "find and remove an recv op: "
+                  << recv_varname_to_ctx[recv_var_name];
+        }
+      }
+    }
+  }
+  // init communicator here
+  if (send_varname_to_ctx.size() > 0) {
+    VLOG(3) << "this is distribute mode, will use communicator";
+    operators::distributed::Communicator::Init(send_varname_to_ctx,
+                                               recv_varname_to_ctx, scope);
+    operators::distributed::Communicator::GetInstance()->Start();
+  }
+#endif
+}
+
+AsyncSSAGraphExecutor::AsyncSSAGraphExecutor(
+    const ExecutionStrategy &strategy, const std::vector<Scope *> &local_scopes,
+    const std::vector<platform::Place> &places, std::vector<ir::Graph *> graphs)
+    : strategy_(std::move(strategy)),
+      local_scopes_(std::move(local_scopes)),
+      pool_(places.size() >= 2 ? new ::ThreadPool(places.size()) : nullptr),
+      places_(std::move(places)),
+      graphs_(std::move(graphs)) {
+  VLOG(3) << "build AsyncSSAGraphExecutor";
+  PADDLE_ENFORCE_EQ(places_.size(), local_scopes_.size());
+
+  // set the correct size of thread pool to each device.
+  strategy_.num_threads_ = strategy_.num_threads_ < places_.size()
+                               ? 1UL
+                               : strategy_.num_threads_ / places_.size();
+  VLOG(1) << "set num_threads: " << strategy_.num_threads_
+          << " to run the operators of the graph on each device.";
+  for (size_t i = 0; i < places.size(); ++i) {
+    executors_.emplace_back(new details::ThreadedSSAGraphExecutor(
+        strategy_, {local_scopes_[i]}, {places_[i]}, graphs_[i]));
+  }
+
+  for (auto &node : graphs_[0]->Nodes()) {
+    if (node->IsVar() && !node->IsCtrlVar() && node->Var()) {
+      var_infos_.emplace_back();
+      var_infos_.back().name_ = node->Var()->Name();
+      var_infos_.back().type_ = node->Var()->GetType();
+      var_infos_.back().persistable_ = node->Var()->Persistable();
+    }
+  }
+  for (auto *scope : local_scopes_) {
+    NewTempScopeAndInitVars(var_infos_, scope);
+  }
+  ProcessGraph(graphs_, local_scopes_[0]);
+}
+
+void AsyncSSAGraphExecutor::StartOffPythonTrainLoop() {
+  VLOG(3) << "StartOffPythonTrainLoop size = " << places_.size();
+  for (size_t i = 1; i < places_.size(); ++i) {
+    auto call = [this, i]() -> void {
+      VLOG(3) << "start off python thread " << i;
+      try {
+        while (true) {
+          executors_[i]->Run({});
+        }
+      } catch (...) {
+        exception_holder_.Catch(std::current_exception());
+        VLOG(3) << "get exception type = " << exception_holder_.Type();
+      }
+      VLOG(3) << "thread " << i << " exited!";
+    };
+    run_futures_.emplace_back(pool_->enqueue(std::move(call)));
+  }
+}
+
+void AsyncSSAGraphExecutor::HandleException() {
+  if (exception_holder_.IsCaught()) {
+    for (auto &f : run_futures_) {
+      VLOG(3) << "wait future";
+      f.wait();
+    }
+    VLOG(3) << "caught exception " << exception_holder_.Type()
+            << ", rethrow it";
+    run_futures_.clear();
+    exception_holder_.ReThrow();
+  }
+}
+
+FeedFetchList AsyncSSAGraphExecutor::Run(
+    const std::vector<std::string> &fetch_tensors) {
+  // init once
+  if (run_futures_.size() == 0 && places_.size() > 1) {
+    exception_holder_.Clear();
+    StartOffPythonTrainLoop();
+  }
+
+  if (places_.size() == 1) {
+    exception_holder_.Clear();
+  } else {
+    HandleException();
+  }
+
+  FeedFetchList fetch_data;
+  fetch_data.reserve(fetch_tensors.size());
+
+  try {
+    fetch_data = executors_[0]->Run(fetch_tensors);
+  } catch (...) {
+    exception_holder_.Catch(std::current_exception());
+  }
+
+  HandleException();
+
+  FeedFetchList ret;
+  for (size_t fetch_idx = 0; fetch_idx < fetch_tensors.size(); ++fetch_idx) {
+    std::vector<const LoDTensor *> lodtensor_ptrs;
+    lodtensor_ptrs.push_back(&fetch_data.at(fetch_idx));
+    ret.emplace_back();
+    ret.back().MergeLoDTensor(lodtensor_ptrs, platform::CPUPlace());
+  }
+  return ret;
+}
+
+}  // namespace details
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/details/broadcast_op_handle.h

@@ -57,7 +57,7 @@ struct BroadcastOpHandle : public OpHandleBase {
 
   std::string Name() const override;
 
-  bool IsMultiDeviceTransfer() override { return false; };
+  bool IsMultiDeviceTransfer() override { return true; };
 
  protected:
   void RunImpl() override;

paddle/fluid/framework/details/execution_strategy.h

@@ -31,6 +31,8 @@ struct ExecutionStrategy {
   size_t num_iteration_per_drop_scope_{1};
   ExecutorType type_{kDefault};
   bool dry_run_{false};
+  size_t num_iteration_per_run_{1};  // only use with async_ssa_graph_executor
+                                     // and pyreader with data queue
 };
 
 }  //  namespace details