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

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into add-async_sparse_param_update_recorder

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)
@@ -62,6 +64,7 @@ option(WITH_DISTRIBUTE  "Compile with distributed support"              OFF)
 option(WITH_PSLIB       "Compile with pslib support"                    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)
@@ -188,7 +191,14 @@ 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()

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
 

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/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}")
 

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/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/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)
@@ -101,7 +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

@@ -34,7 +34,7 @@ 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          ${MKLML_LIB_DIR}/mklml.dll)
@@ -43,7 +43,7 @@ ELSE()
     #TODO(intel-huying):
     #  Now enable Erf function in mklml library temporarily, it will be updated as offical version later.
     SET(MKLML_VER "Glibc225_vsErf_mklml_lnx_${TIME_VERSION}" CACHE STRING "" FORCE)
-    SET(MKLML_URL "http://paddlepaddledeps.cdn.bcebos.com/${MKLML_VER}.tgz" CACHE STRING "" FORCE)
+    SET(MKLML_URL "http://paddlepaddledeps.bj.bcebos.com/${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)

cmake/external/ngraph.cmake

@@ -57,20 +57,25 @@ 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=${MKLDNN_INSTALL_DIR}/${CMAKE_INSTALL_LIBDIR}
-    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})

cmake/external/protobuf.cmake

@@ -201,7 +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(

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

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()

cmake/tensorrt.cmake

@@ -33,5 +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(${TENSORRT_INCLUDE_DIR})
+    link_directories(${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/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 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/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 +84,38 @@ 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) {}
+  // This function will do nothing at default
+  virtual void SetFleetSendBatchSize(int64_t size) {}
+  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 +129,9 @@ 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 +154,9 @@ 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 +179,7 @@ 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 +194,59 @@ 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 SetFleetSendBatchSize(int64_t size);
+  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 +268,7 @@ 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 +291,33 @@ 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 +346,37 @@ 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_layout_transform.cc

@@ -134,6 +134,11 @@ void TransDataLayoutFromMKLDNN(const OpKernelType& kernel_type_for_var,
   out_layout =
       out_layout == DataLayout::kAnyLayout ? DataLayout::kNCHW : out_layout;
 
+  auto& pool = platform::DeviceContextPool::Instance();
+  auto* dev_ctx = dynamic_cast<platform::MKLDNNDeviceContext*>(
+      pool.Get(expected_kernel_type.place_));
+  auto& cpu_engine = dev_ctx->GetEngine();
+
   std::vector<int> in_tz = paddle::framework::vectorize2int(in.dims());
   std::vector<int> out_tz = in_tz;
 
@@ -142,25 +147,29 @@ void TransDataLayoutFromMKLDNN(const OpKernelType& kernel_type_for_var,
                  "Input tensor type is not supported: %s", in.type());
   memory::data_type out_type = in_type;
 
+  auto in_format = platform::MKLDNNFormatForSize(in_tz.size(), in.format());
+  auto out_format =
+      platform::MKLDNNFormatForSize(in_tz.size(), ToMKLDNNFormat(out_layout));
+
   // output tensor has the same dims as input. Reorder don't change dims
   out->Resize(in.dims());
 
-  // tempory mem pd fr out , to make reorder
-  auto out_mem_pd = paddle::platform::create_prim_desc_from_dims(
-      paddle::framework::vectorize2int(out->dims()),
-      mkldnn::memory::format::blocked, out_type);
-  if (in.get_mkldnn_prim_desc() != out_mem_pd) {
+  if (in_format != out_format) {
     void* in_data = GetDataFromTensor(in, in_type);
     auto out_data = out->mutable_data(expected_kernel_type.place_, in.type());
 
-    auto in_memory = memory(in.get_mkldnn_prim_desc(), in_data);
-    auto out_memory = memory(out_mem_pd, out_data);
+    auto in_memory =
+        memory({{{in_tz}, in_type, in_format}, cpu_engine}, in_data);
+    auto out_memory =
+        memory({{{out_tz}, out_type, out_format}, cpu_engine}, out_data);
 
     platform::Reorder(in_memory, out_memory);
   } else {
     out->ShareDataWith(in);
   }
   out->set_layout(out_layout);
+  // reset format since the out tensor will be feed to non-MKLDNN OPkernel
+  out->set_format(memory::format::format_undef);
 #endif
 }
 

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);
+  }
+}
+
+// if you run distributed, and want to do global shuffle,
+// set this before global shuffle.
+// be sure you call CreateReaders before SetFleetSendBatchSize
+template <typename T>
+void DatasetImpl<T>::SetFleetSendBatchSize(int64_t size) {
+  fleet_send_batch_size_ = size;
+  for (auto reader : readers_) {
+    reader->SetFleetSendBatchSize(size);
+  }
+}
+
+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 fleet send batch size
+  virtual void SetFleetSendBatchSize(int64_t size) = 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 fleet send batch size
+  virtual int64_t GetFleetSendBatchSize() = 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 SetFleetSendBatchSize(int64_t size);
+  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 int64_t GetFleetSendBatchSize() { return fleet_send_batch_size_; }
+  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;
+  int64_t fleet_send_batch_size_;
+};
+
+// 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/data_transform.cc

@@ -51,31 +51,13 @@ void TransformData(const OpKernelType &expected_kernel_type,
 #ifdef PADDLE_WITH_MKLDNN
         // Case1 - transform from Non-MKLDNN OPKernel to MKLDNN OPKernel
         // Just set layout/format. No real transform occur
+
+        auto out_format = platform::MKLDNNFormatForSize(in.dims().size(),
+                                                        ToMKLDNNFormat(lin));
+
         out.ShareDataWith(input_tensor);
-        // TODO(jczaja): Remove that once all mkldnn ops
-        // are modified to work with mkldnn_blocked
-        auto mkldnn_fmt = [&](int rank) {
-          switch (rank) {
-            case 5:
-              return mkldnn::memory::format::ncdhw;
-            case 4:
-              return mkldnn::memory::format::nchw;
-            case 3:
-              return mkldnn::memory::format::ncw;
-            case 2:
-              return mkldnn::memory::format::nc;
-            case 1:
-              return mkldnn::memory::format::x;
-            default:
-              return mkldnn::memory::format::blocked;
-          }
-        };
-
-        auto out_mem_pd = paddle::platform::create_prim_desc_from_dims(
-            paddle::framework::vectorize2int(out.dims()),
-            mkldnn_fmt(out.dims().size()));
-
-        out.set_mkldnn_prim_desc(out_mem_pd);
+        out.set_layout(DataLayout::kMKLDNN);
+        out.set_format(out_format);
 #endif
       } else {
         // Case2 - transfrom from MKLDNN OPKernel to Non-MKLDNN OPKernel

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/framework/dataset_factory.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 <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)
@@ -46,9 +59,7 @@ 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)
 
 if(WITH_GPU)
 cc_library(memory_optimize_helper SRCS memory_optimize_helper.cc DEPS graph graph_helper gpu_info)
@@ -61,14 +72,17 @@ cc_library(inplace_op_pass SRCS inplace_op_pass.cc DEPS memory_optimize_pass op_
 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)
@@ -82,7 +96,11 @@ 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)
 
-cc_library(async_ssa_graph_executor SRCS async_ssa_graph_executor.cc DEPS threaded_ssa_graph_executor communicator)
+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)
@@ -100,4 +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

@@ -13,107 +13,186 @@
 // limitations under the License.
 
 #include <algorithm>
+#include <map>
 #include <string>
 #include <unordered_map>
 #include <unordered_set>
+#include <utility>
 #include <vector>
 
-#include "paddle/fluid/framework/details/all_reduce_deps_pass.h"
 #include "paddle/fluid/framework/details/all_reduce_op_handle.h"
+#include "paddle/fluid/framework/details/container_cast.h"
 #include "paddle/fluid/framework/details/multi_devices_helper.h"
 #include "paddle/fluid/framework/details/op_graph_view.h"
-#include "paddle/fluid/framework/details/var_handle.h"
+#include "paddle/fluid/framework/ir/graph.h"
 #include "paddle/fluid/framework/ir/graph_helper.h"
+#include "paddle/fluid/framework/ir/pass.h"
 #include "paddle/fluid/framework/op_proto_maker.h"
 
 namespace paddle {
 namespace framework {
 namespace details {
 
-VarHandle* GetValidInput(const OpHandleBase* a) {
-  for (auto p : a->Inputs()) {
-    VarHandle* b = dynamic_cast<VarHandle*>(p);
-    if (b) {
-      return b;
+class AllReduceDepsPass : public ir::Pass {
+ protected:
+  void ApplyImpl(ir::Graph* graph) const override {
+    std::vector<AllReduceOpHandle*> all_reduce_op_handles =
+        GetSortedAllReduceOps(*graph);
+
+    for (size_t i = 1; i < all_reduce_op_handles.size(); ++i) {
+      auto* dep_var = new DummyVarHandle(graph->CreateControlDepVar());
+      graph->Get<GraphDepVars>(kGraphDepVars).emplace(dep_var);
+      all_reduce_op_handles[i - 1]->AddOutput(dep_var);
+      all_reduce_op_handles[i]->AddInput(dep_var);
     }
-  }
 
-  return nullptr;
-}
-
-std::unique_ptr<ir::Graph> AllReduceDepsPass::ApplyImpl(
-    std::unique_ptr<ir::Graph> graph) const {
-  auto graph_ops = ir::FilterByNodeWrapper<OpHandleBase>(*graph);
-
-  // get vars order
-  int order = 0;
-  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 = 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;
-      }
+    if (VLOG_IS_ON(10)) {
+      DebugString(*graph, all_reduce_op_handles);
     }
-    order++;
   }
 
-  std::vector<OpHandleBase*> dist_ops;
-  // get allreduce ops.
-  for (auto& op : graph_ops) {
-    // FIXME(gongwb):add broad cast.
-    if (op->Name() == "all_reduce" || op->Name() == "reduce") {
-      dist_ops.push_back(op);
+  std::vector<AllReduceOpHandle*> GetSortedAllReduceOps(
+      const ir::Graph& graph) const {
+    std::vector<AllReduceOpHandle*> all_reduce_op_handles;
+    std::unordered_map<OpHandleBase*, size_t> pending_ops;
+    std::unordered_set<OpHandleBase*> ready_ops;
+    std::unordered_set<OpHandleBase*> next_ready_ops;
+
+    auto op_handles = ir::FilterByNodeWrapper<OpHandleBase>(graph);
+    size_t num_of_ops = op_handles.size();
+    for (OpHandleBase* op : op_handles) {
+      size_t not_ready_vars = op->NotReadyInputSize();
+      if (not_ready_vars) {
+        pending_ops.insert({op, not_ready_vars});
+      } else {
+        ready_ops.insert(op);
+      }
     }
-  }
-
-  VLOG(10) << "dist_ops size:" << dist_ops.size() << std::endl;
-
-  std::sort(dist_ops.begin(), dist_ops.end(), [&](OpHandleBase* op1,
-                                                  OpHandleBase* op2) {
-    VarHandle* i0 = dynamic_cast<VarHandle*>(GetValidInput(op1));
-    VarHandle* i1 = dynamic_cast<VarHandle*>(GetValidInput(op2));
-
-    PADDLE_ENFORCE(i0 != nullptr && i1 != nullptr, "%s convert to %s error",
-                   op1->DebugString(), op2->DebugString());
 
-    auto l_it = vars.find(i0->name());
-    auto r_it = vars.find(i1->name());
-
-    if (l_it->second < r_it->second) return true;
+    GetSortedAllReduceOps(ready_ops, &all_reduce_op_handles);
+
+    size_t has_run_ops = ready_ops.size();
+    while (has_run_ops != num_of_ops) {
+      for (auto* op : ready_ops) {
+        for (auto& ready_var : op->Outputs()) {
+          for (auto* pend_op : ready_var->PendingOps()) {
+            auto& deps = --pending_ops[pend_op];
+            if (deps == 0) {
+              next_ready_ops.insert(pend_op);
+            }
+          }
+        }
+      }
 
-    if (l_it->second == r_it->second) {
-      return i0->name() < i1->name();
+      PADDLE_ENFORCE_NE(next_ready_ops.size(), 0, "There maybe have a cycle.");
+      ready_ops.clear();
+      std::swap(ready_ops, next_ready_ops);
+      GetSortedAllReduceOps(ready_ops, &all_reduce_op_handles);
+      has_run_ops += ready_ops.size();
     }
+    return all_reduce_op_handles;
+  }
 
-    return false;
-  });
-
-  // add dependency.
-  auto& sorted_ops = dist_ops;
-  for (size_t i = 1; i < sorted_ops.size(); ++i) {
-    auto* dep_var = new DummyVarHandle(graph->CreateControlDepVar());
-
-    auto* pre_op = sorted_ops[i - 1];
-    auto* op = sorted_ops[i];
-
-    pre_op->AddOutput(dep_var);
-    op->AddInput(dep_var);
-    graph->Get<GraphDepVars>(kGraphDepVars).emplace(dep_var);
-
-    VLOG(10) << "add all_reduce sequential dependencies between " << pre_op
-             << " and " << op;
+  void GetSortedAllReduceOps(
+      const std::unordered_set<OpHandleBase*>& ready_ops,
+      std::vector<AllReduceOpHandle*>* all_reduce_op_handles) const {
+    std::vector<AllReduceOpHandle*> current_all_reduce_op_handles;
+    for (auto& op_handle : ready_ops) {
+      auto all_reduce_op_handle = dynamic_cast<AllReduceOpHandle*>(op_handle);
+      if (all_reduce_op_handle) {
+        current_all_reduce_op_handles.emplace_back(all_reduce_op_handle);
+      }
+    }
 
-    VLOG(10) << "pre_op:" << pre_op->DebugString()
-             << ", op:" << op->DebugString();
+    // NOTE(zcd): For distributed training, it is important to keep the order of
+    // allReduce on each node consistent. Otherwise, hang may occur.
+    // Sort the current_all_reduce_op_handles according to the name of input.
+    sort(current_all_reduce_op_handles.begin(),
+         current_all_reduce_op_handles.end(),
+         [](const AllReduceOpHandle* left,
+            const AllReduceOpHandle* right) -> bool {
+           auto left_in_vars = DynamicCast<VarHandle>(left->Inputs());
+           auto right_in_vars = DynamicCast<VarHandle>(right->Inputs());
+           PADDLE_ENFORCE_GT(left_in_vars.size(), 0);
+           PADDLE_ENFORCE_EQ(left_in_vars.size(), right_in_vars.size());
+           return left_in_vars[0]->Name() > right_in_vars[0]->Name();
+         });
+
+    all_reduce_op_handles->insert(all_reduce_op_handles->end(),
+                                  current_all_reduce_op_handles.begin(),
+                                  current_all_reduce_op_handles.end());
   }
 
-  return graph;
-}
+  void DebugString(
+      const ir::Graph& graph,
+      const std::vector<AllReduceOpHandle*>& all_reduce_op_handles) const {
+    // get vars order
+    std::map<int, std::vector<std::string>> vars =
+        GetSoredGradientsFromStaleProgram(graph);
+    std::stringstream out;
+    size_t grads_of_stale_program = 0;
+    out << "Get Order From kStaleProgramOpDescs: ";
+    for (auto& var : vars) {
+      out << "Order " << var.first << " [";
+      for (auto& var_name : var.second) {
+        out << var_name << ", ";
+        ++grads_of_stale_program;
+      }
+      out << "], ";
+    }
+    VLOG(10) << out.str();
+
+    std::stringstream out2;
+    out2 << "Get Order From Topological order: ";
+    for (auto& op : all_reduce_op_handles) {
+      bool find_valid_input = false;
+      for (auto& in_var : op->Inputs()) {
+        if (dynamic_cast<VarHandle*>(in_var)) {
+          out2 << in_var->Name() << ", ";
+          find_valid_input = true;
+          break;
+        }
+      }
+      PADDLE_ENFORCE(find_valid_input, "Doesn't find valid input.");
+    }
+    VLOG(10) << out2.str();
+    if (grads_of_stale_program != all_reduce_op_handles.size()) {
+      VLOG(10)
+          << "The gradients number of stale program and graph is not equal.";
+    }
+  }
 
+  std::map<int, std::vector<std::string>> GetSoredGradientsFromStaleProgram(
+      const ir::Graph& graph) const {
+    std::map<int, std::vector<std::string>> vars;
+    auto ops = graph.Get<const std::vector<OpDesc*>>(kStaleProgramOpDescs);
+    int order = 0;
+    for (auto* op_desc : ops) {
+      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()));
+        if (backward_vars.empty()) continue;
+
+        PADDLE_ENFORCE_EQ(backward_vars.size() % 2, 0);
+        for (size_t i = 1; i < backward_vars.size(); i += 2) {
+          vars[order].emplace_back(backward_vars[i]);
+          VLOG(1) << "get parameter and gradient: " << backward_vars[i - 1]
+                  << ", " << backward_vars[i];
+        }
+        order++;
+      } catch (boost::bad_get e) {
+      }
+    }
+    return vars;
+  }
+};
 }  // namespace details
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/details/all_reduce_op_handle.cc

@@ -11,18 +11,24 @@
 // 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:
 // https://github.com/PaddlePaddle/Paddle/issues/15049
 DEFINE_bool(
-    sync_nccl_allreduce, false,
+    sync_nccl_allreduce, true,
     "If set true, will call `cudaStreamSynchronize(nccl_stream)`"
     "after allreduce, this mode can get better performance in some scenarios.");
 
@@ -34,16 +40,23 @@ 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));
     }
   }
+  // TODO(gongwb) :polish them!
+  if (is_encoded) {
+    VLOG(1) << "Use dgc allreduce mode";
+  }
 }
 #else
 AllReduceOpHandle::AllReduceOpHandle(ir::Node *node,
@@ -52,10 +65,189 @@ 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, "%s should not be null", encode_var_name);
+    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() {
+  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 +264,8 @@ 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 +293,17 @@ 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/alloc_continuous_space_for_grad_pass.cc

+//   Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/fluid/framework/details/alloc_continuous_space_for_grad_pass.h"
+#include <algorithm>
+#include <string>
+#include <unordered_map>
+#include <utility>
+#include <vector>
+#include "paddle/fluid/framework/details/build_strategy.h"
+#include "paddle/fluid/framework/details/multi_devices_helper.h"
+#include "paddle/fluid/framework/ir/graph_helper.h"
+#include "paddle/fluid/framework/op_registry.h"
+
+DEFINE_uint64(fuse_parameter_memory_size, 0,  // 0 KB
+              "fuse_parameter_memory_size is up limited memory size "
+              "of one group parameters' gradient which is the input "
+              "of communication calling(e.g NCCLAllReduce). "
+              "The default value is 0, it means that "
+              "not set group according to memory_size.");
+DEFINE_int32(
+    fuse_parameter_groups_size, 3,
+    "fuse_parameter_groups_size is the size of one group parameters' gradient. "
+    "The default value is a experimental result. If the "
+    "fuse_parameter_groups_size is 1, it means that the groups size is "
+    "the number of parameters' gradient. If the fuse_parameter_groups_size is "
+    "-1, it means that there are only one group. The default value is 3, it is "
+    "an experimental value.");
+
+namespace paddle {
+namespace framework {
+namespace details {
+// SetFuseParameterGroupsSize and SetFuseParameterMemorySize are used in unit
+// test, because it is invalid that seting 'FLAGS_fuse_parameter_memory_size'
+// and 'FLAGS_fuse_parameter_groups_size' in unit test.
+void SetFuseParameterGroupsSize(int group_size) {
+  FLAGS_fuse_parameter_groups_size = group_size;
+}
+
+int GetFuseParameterGroupsSize() { return FLAGS_fuse_parameter_groups_size; }
+
+void SetFuseParameterMemorySize(uint64_t memory_size) {
+  FLAGS_fuse_parameter_memory_size = memory_size;
+}
+
+uint64_t GetFuseParameterMemorySize() {
+  return FLAGS_fuse_parameter_memory_size;
+}
+
+static const char kUnKnow[] = "@UNKNOW@";
+static framework::proto::VarType::Type kDefaultDtype =
+    framework::proto::VarType::Type::VarType_Type_BOOL;
+
+void AllocContinuousSpaceForGradPass::ApplyImpl(ir::Graph *graph) const {
+  ir::Graph &result = *graph;
+
+  auto &places = Get<const std::vector<platform::Place>>(kPlaces);
+  auto &local_scopes = Get<const std::vector<Scope *>>(kLocalScopes);
+
+  ResetAttribute<ParamsAndGrads>(kParamsAndGrads, &result);
+  ResetAttribute<GroupGradsAndParams>(kGroupGradsAndParams, &result);
+
+  // NOTE: The operator nodes should be in topology order.
+  std::vector<ir::Node *> topo_nodes = ir::TopologySortOperations(result);
+  auto &params_grads = result.Get<ParamsAndGrads>(kParamsAndGrads);
+  for (auto &node : topo_nodes) {
+    RecordParamsAndGrads(node, &params_grads);
+  }
+
+  if (params_grads.size() == 0) {
+    VLOG(10) << "Doesn't find gradients";
+    return;
+  }
+
+  std::unordered_map<std::string, ir::Node *> vars;
+  for (ir::Node *node : result.Nodes()) {
+    if (node->IsVar() && node->Var()) {
+      // Note: The graph may have the same name node. For example, parameter
+      // is the input of operator and it also is the output of optimizer;
+      vars.emplace(node->Var()->Name(), node);
+    }
+  }
+
+  auto &group_grads_params =
+      result.Get<GroupGradsAndParams>(kGroupGradsAndParams);
+
+  // Note: the order of params_grads may be changed by SetGroupGradsAndParams.
+  SetGroupGradsAndParams(vars, params_grads, &group_grads_params);
+
+  params_grads.clear();
+  for (auto &group_p_g : group_grads_params) {
+    params_grads.insert(params_grads.begin(), group_p_g.begin(),
+                        group_p_g.end());
+  }
+  for (auto &p_g : params_grads) {
+    std::swap(p_g.first, p_g.second);
+  }
+
+  // Set Gradients as Persistable to prevent this var becoming reusable.
+  auto dtype = kDefaultDtype;
+  for (auto &p_g : params_grads) {
+    // Get gradient var
+    auto iter = vars.find(p_g.second);
+    PADDLE_ENFORCE(iter != vars.end(), "%s is not found.", p_g.second);
+    iter->second->Var()->SetPersistable(true);
+
+    PADDLE_ENFORCE(IsSupportedVarType(iter->second->Var()->GetType()));
+
+    // Get Dtype
+    auto ele_dtype = iter->second->Var()->GetDataType();
+    if (dtype == kDefaultDtype) {
+      dtype = ele_dtype;
+      PADDLE_ENFORCE_NE(ele_dtype, kDefaultDtype,
+                        "The data type should not be bool.");
+    }
+    PADDLE_ENFORCE_EQ(ele_dtype, dtype,
+                      "The data type of input is not consistent.");
+  }
+
+  // Create a FusedVarsSet to avoid duplicating names for fused_var in other
+  // pass.
+  if (!result.Has(kFusedVars)) {
+    result.Set(kFusedVars, new FusedVars);
+  }
+  // the kFusedGrads is used be fuse_optimizer_op_pass.
+  result.Set(kFusedGrads, new FusedGrads);
+
+  // the fused_var_name should be unique, so it appends
+  // params_grads.begin()->second.
+  auto fused_var_name = std::string(kFusedVarNamePrefix) + "@GRAD@" +
+                        params_grads.begin()->second;
+  result.Get<FusedGrads>(kFusedGrads) = fused_var_name;
+  auto &fused_var_set = result.Get<FusedVars>(kFusedVars);
+  PADDLE_ENFORCE_EQ(fused_var_set.count(fused_var_name), 0,
+                    "%s is duplicate in FusedVars.", fused_var_name);
+  fused_var_set.insert(fused_var_name);
+
+  InitFusedVarsAndAllocSpaceForVars(places, local_scopes, vars, fused_var_name,
+                                    params_grads);
+}
+
+template <typename AttrType>
+void AllocContinuousSpaceForGradPass::ResetAttribute(
+    const std::string &attr_name, ir::Graph *graph) const {
+  if (graph->Has(attr_name)) {
+    VLOG(10) << attr_name << " is reset.";
+    graph->Erase(attr_name);
+  }
+  graph->Set(attr_name, new AttrType);
+}
+
+void AllocContinuousSpaceForGradPass::SetGroupGradsAndParams(
+    const std::unordered_map<std::string, ir::Node *> &var_nodes,
+    const ParamsAndGrads &params_grads,
+    GroupGradsAndParams *group_grads_params) const {
+  SetGroupAccordingToLayers(var_nodes, params_grads, group_grads_params);
+  SetGroupAccordingToMemorySize(var_nodes, group_grads_params);
+  SetGroupAccordingToGroupSize(var_nodes, group_grads_params);
+}
+
+void AllocContinuousSpaceForGradPass::SetGroupAccordingToLayers(
+    const std::unordered_map<std::string, ir::Node *> &var_nodes,
+    const ParamsAndGrads &params_grads,
+    GroupGradsAndParams *group_grads_params) const {
+  std::unordered_map<std::string, std::vector<int>> layer_params;
+
+  for (size_t i = 0; i < params_grads.size(); ++i) {
+    auto pos = params_grads[i].first.find_first_of(".");
+    if (pos == std::string::npos) {
+      layer_params[std::string(kUnKnow)].emplace_back(i);
+    } else {
+      layer_params[params_grads[i].first.substr(0, pos)].emplace_back(i);
+    }
+  }
+
+  group_grads_params->reserve(layer_params.size());
+  for (size_t i = 0; i < params_grads.size(); ++i) {
+    auto pos = params_grads[i].first.find_first_of(".");
+    std::string key = kUnKnow;
+    if (pos != std::string::npos) {
+      key = params_grads[i].first.substr(0, pos);
+    }
+    auto iter = layer_params.find(key);
+    if (iter == layer_params.end()) continue;
+
+    group_grads_params->emplace_back();
+    auto &local_group_grads_params = group_grads_params->back();
+    for (auto &idx : iter->second) {
+      local_group_grads_params.emplace_back(
+          std::make_pair(params_grads[idx].second, params_grads[idx].first));
+    }
+    layer_params.erase(iter);
+  }
+
+  VLOG(10) << "SetGroupAccordingToLayers: ";
+  for (size_t i = 0; i < group_grads_params->size(); ++i) {
+    VLOG(10) << "group " << i;
+    std::stringstream out;
+    for (auto &p_g : group_grads_params->at(i)) {
+      out << "(" << p_g.second << ", " << p_g.first << "), ";
+    }
+    VLOG(10) << out.str();
+  }
+}
+
+void AllocContinuousSpaceForGradPass::SetGroupAccordingToMemorySize(
+    const std::unordered_map<std::string, ir::Node *> &var_nodes,
+    GroupGradsAndParams *group_grads_params) const {
+  const uint64_t group_memory_size = GetFuseParameterMemorySize();
+  if (group_memory_size == 0) {
+    return;
+  }
+  GroupGradsAndParams local_group_grads_params;
+  size_t j = 0;
+  while (j < group_grads_params->size()) {
+    local_group_grads_params.emplace_back();
+    auto &group_p_g = local_group_grads_params.back();
+    size_t local_group_memory_size = 0;
+    while (j < group_grads_params->size()) {
+      std::for_each(
+          group_grads_params->at(j).begin(), group_grads_params->at(j).end(),
+          [&local_group_memory_size,
+           &var_nodes](const std::pair<std::string, std::string> &g_p) {
+            auto iter = var_nodes.find(g_p.second);
+            PADDLE_ENFORCE(iter != var_nodes.end(), "%s is not found.",
+                           g_p.second);
+            auto shape = iter->second->Var()->GetShape();
+            size_t size =
+                framework::SizeOfType(iter->second->Var()->GetDataType());
+            std::for_each(shape.begin(), shape.end(),
+                          [&size](const int64_t &n) { size *= n; });
+            local_group_memory_size += size;
+          });
+      group_p_g.insert(group_p_g.end(), group_grads_params->at(j).begin(),
+                       group_grads_params->at(j).end());
+      ++j;
+      if (local_group_memory_size >= group_memory_size) {
+        break;
+      }
+    }
+  }
+
+  std::swap(*group_grads_params, local_group_grads_params);
+
+  VLOG(10) << string::Sprintf("SetGroupAccordingToMemorySize(memory_size: %d):",
+                              group_memory_size);
+  for (size_t i = 0; i < group_grads_params->size(); ++i) {
+    VLOG(10) << "group " << i;
+    std::stringstream out;
+    for (auto &g_p : group_grads_params->at(i)) {
+      auto iter = var_nodes.find(g_p.second);
+      PADDLE_ENFORCE(iter != var_nodes.end(), "%s is not found.", g_p.second);
+      auto shape = iter->second->Var()->GetShape();
+      size_t size = framework::SizeOfType(iter->second->Var()->GetDataType());
+      std::for_each(shape.begin(), shape.end(),
+                    [&size](const int64_t &n) { size *= n; });
+      out << string::Sprintf("(%s(%d), %s)", g_p.second, size, g_p.first);
+    }
+    VLOG(10) << out.str();
+  }
+}
+
+void AllocContinuousSpaceForGradPass::SetGroupAccordingToGroupSize(
+    const std::unordered_map<std::string, ir::Node *> &var_nodes,
+    GroupGradsAndParams *group_grads_params) const {
+  if (GetFuseParameterGroupsSize() == 1) {
+    return;
+  }
+  const int group_size = GetFuseParameterGroupsSize() == -1
+                             ? static_cast<int>(group_grads_params->size())
+                             : GetFuseParameterGroupsSize();
+  PADDLE_ENFORCE_GT(group_size, 1);
+  size_t groups = (group_grads_params->size() + group_size - 1) / group_size;
+  GroupGradsAndParams local_group_grads_params;
+  local_group_grads_params.reserve(groups);
+
+  size_t j = 0;
+  for (size_t i = 0; i < groups; ++i) {
+    local_group_grads_params.emplace_back();
+    auto &group_p_g = local_group_grads_params.back();
+    group_p_g.reserve(group_size);
+    while (j < group_grads_params->size()) {
+      group_p_g.insert(group_p_g.end(), group_grads_params->at(j).begin(),
+                       group_grads_params->at(j).end());
+      ++j;
+      if (j % group_size == 0) break;
+    }
+  }
+  std::swap(*group_grads_params, local_group_grads_params);
+
+  VLOG(10) << string::Sprintf("SetGroupAccordingToGroupSize(group_size: %d):",
+                              group_size);
+  for (size_t i = 0; i < group_grads_params->size(); ++i) {
+    VLOG(10) << "group " << i;
+    std::stringstream out;
+    for (auto &p_g : group_grads_params->at(i)) {
+      out << "(" << p_g.second << ", " << p_g.first << "), ";
+    }
+    VLOG(10) << out.str();
+  }
+}
+
+bool AllocContinuousSpaceForGradPass::IsSupportedVarType(
+    const proto::VarType::Type &type) const {
+  // Current only support LOD_TENSOR.
+  return type == proto::VarType::LOD_TENSOR;
+}
+
+void AllocContinuousSpaceForGradPass::RecordParamsAndGrads(
+    ir::Node *node, ParamsAndGrads *params_grads) const {
+  try {
+    bool is_bk_op =
+        static_cast<bool>(boost::get<int>(node->Op()->GetAttr(
+                              OpProtoAndCheckerMaker::OpRoleAttrName())) &
+                          static_cast<int>(OpRole::kBackward));
+    if (!is_bk_op) return;
+
+    // Currently, we assume that once gradient is generated, it can be
+    // broadcast, and each gradient is only broadcast once.
+    auto backward_vars =
+        boost::get<std::vector<std::string>>(node->Op()->GetNullableAttr(
+            OpProtoAndCheckerMaker::OpRoleVarAttrName()));
+    PADDLE_ENFORCE_EQ(backward_vars.size() % 2, static_cast<size_t>(0));
+
+    for (size_t i = 0; i < backward_vars.size(); i += 2) {
+      VLOG(10) << "Trainable parameter: " << backward_vars[i]
+               << ", gradient: " << backward_vars[i + 1];
+
+      params_grads->emplace_back(std::make_pair(backward_vars[i] /*param*/,
+                                                backward_vars[i + 1] /*grad*/));
+    }
+  } catch (boost::bad_get e) {
+  }
+}
+
+void AllocContinuousSpaceForGradPass::InitFusedVarsAndAllocSpaceForVars(
+    const std::vector<platform::Place> &places,
+    const std::vector<Scope *> &local_scopes,
+    const std::unordered_map<std::string, ir::Node *> &vars,
+    const std::string &fused_var_name,
+    const ParamsAndGrads &params_grads) const {
+  //  Init Gradients and FusedVars
+  VLOG(10) << "Init FusedVars and Gradients.";
+  for (auto it = local_scopes.rbegin(); it != local_scopes.rend(); ++it) {
+    auto &scope = *it;
+
+    PADDLE_ENFORCE(scope->FindVar(fused_var_name) == nullptr,
+                   "%s has existed in scope.", fused_var_name);
+    scope->Var(fused_var_name)->GetMutable<LoDTensor>();
+
+    for (auto &p_g : params_grads) {
+      auto iter = vars.find(p_g.second);
+      PADDLE_ENFORCE(iter != vars.end());
+      PADDLE_ENFORCE_NOT_NULL(iter->second->Var());
+      PADDLE_ENFORCE_EQ(iter->second->Var()->GetType(),
+                        proto::VarType::LOD_TENSOR);
+      scope->Var(p_g.second)->GetMutable<LoDTensor>();
+    }
+  }
+
+  // Alloc continuous space for vars.
+  std::vector<std::string> grads_name;
+  std::vector<std::string> params_name;
+  grads_name.reserve(params_grads.size());
+  params_name.reserve(params_grads.size());
+  for (auto &p_g : params_grads) {
+    params_name.emplace_back(p_g.first);
+    grads_name.emplace_back(p_g.second);
+  }
+  framework::ProgramDesc program_desc;
+  AppendAllocSpaceForVarsOp(params_name, grads_name, fused_var_name,
+                            program_desc.MutableBlock(0));
+
+  for (size_t i = 0; i < local_scopes.size(); ++i) {
+    for (auto &op_desc : program_desc.Block(0).AllOps()) {
+      auto op = OpRegistry::CreateOp(*op_desc);
+      op->Run(*local_scopes[i], places[i]);
+    }
+  }
+}
+
+void AllocContinuousSpaceForGradPass::AppendAllocSpaceForVarsOp(
+    const std::vector<std::string> &params_name,
+    const std::vector<std::string> &grads_name,
+    const std::string &fused_var_name, BlockDesc *global_block) const {
+  auto op_desc = global_block->AppendOp();
+  op_desc->SetType("alloc_continuous_space");
+  op_desc->SetInput("Input", params_name);
+  op_desc->SetOutput("Output", grads_name);
+  op_desc->SetOutput("FusedOutput", {fused_var_name});
+}
+
+}  // namespace details
+}  // namespace framework
+}  // namespace paddle
+
+REGISTER_PASS(alloc_continuous_space_for_grad_pass,
+              paddle::framework::details::AllocContinuousSpaceForGradPass)
+    .RequirePassAttr(paddle::framework::details::kPlaces)
+    .RequirePassAttr(paddle::framework::details::kLocalScopes);

paddle/fluid/framework/details/alloc_continuous_space_for_grad_pass.h

+//   Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+#pragma once
+#include <algorithm>
+#include <string>
+#include <unordered_map>
+#include <utility>
+#include <vector>
+#include "paddle/fluid/framework/details/build_strategy.h"
+#include "paddle/fluid/framework/details/multi_devices_helper.h"
+#include "paddle/fluid/framework/ir/graph_helper.h"
+
+namespace paddle {
+namespace framework {
+namespace details {
+
+void SetFuseParameterGroupsSize(int group_size);
+int GetFuseParameterGroupsSize();
+
+void SetFuseParameterMemorySize(uint64_t memory_size);
+uint64_t GetFuseParameterMemorySize();
+
+class AllocContinuousSpaceForGradPass : public ir::Pass {
+ protected:
+  void ApplyImpl(ir::Graph *graph) const override;
+
+  template <typename AttrType>
+  void ResetAttribute(const std::string &attr_name, ir::Graph *graph) const;
+
+  void SetGroupGradsAndParams(
+      const std::unordered_map<std::string, ir::Node *> &var_nodes,
+      const ParamsAndGrads &params_grads,
+      GroupGradsAndParams *group_grads_params) const;
+
+  void SetGroupAccordingToLayers(
+      const std::unordered_map<std::string, ir::Node *> &var_nodes,
+      const ParamsAndGrads &params_grads,
+      GroupGradsAndParams *group_grads_params) const;
+
+  void SetGroupAccordingToMemorySize(
+      const std::unordered_map<std::string, ir::Node *> &var_nodes,
+      GroupGradsAndParams *group_grads_params) const;
+
+  void SetGroupAccordingToGroupSize(
+      const std::unordered_map<std::string, ir::Node *> &var_nodes,
+      GroupGradsAndParams *group_grads_params) const;
+
+ private:
+  bool IsSupportedVarType(const proto::VarType::Type &type) const;
+
+  void RecordParamsAndGrads(ir::Node *node, ParamsAndGrads *params_grads) const;
+
+  void InitFusedVarsAndAllocSpaceForVars(
+      const std::vector<platform::Place> &places,
+      const std::vector<Scope *> &local_scopes,
+      const std::unordered_map<std::string, ir::Node *> &vars,
+      const std::string &fused_var_name,
+      const ParamsAndGrads &params_grads) const;
+
+  void AppendAllocSpaceForVarsOp(const std::vector<std::string> &params_name,
+                                 const std::vector<std::string> &grads_name,
+                                 const std::string &fused_var_name,
+                                 BlockDesc *global_block) const;
+};
+
+}  // namespace details
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/details/async_ssa_graph_executor.cc

@@ -15,7 +15,10 @@
 #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 {
@@ -43,6 +46,7 @@ inline void NewTempScopeAndInitVars(const std::vector<VarInfo> &var_infos,
 
 // 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;
@@ -85,33 +89,6 @@ void ProcessGraph(std::vector<ir::Graph *> graphs, Scope *scope) {
         }
       }
     }
-    /*
-    VLOG(3) << "delete all recv ops";
-    for (auto *node : nodes_to_delete) {
-      // delete input edge
-      for (auto *in : node->inputs) {
-        auto &in_outs = in->outputs;
-        for (auto iter = in_outs.begin(); iter != in_outs.end();) {
-          if (*iter == node) {
-            VLOG(3) << "delete input edge from " << in->Name() << " for "
-                    << node->Name();
-            iter = in_outs.erase(iter);
-          } else {
-            ++iter;
-          }
-        }
-      }
-      // delete output edge
-      for (auto *out : node->outputs) {
-        PADDLE_ENFORCE_EQ(out->outputs.size(), 0, "%s should have no outputs",
-                          out->Name());
-        VLOG(3) << "delete output edge to " << out->Name();
-        graphs[i]->RemoveNode(out);
-      }
-      VLOG(3) << "delete node " << node->Name();
-      graphs[i]->RemoveNode(node);
-    }
-    */
   }
   // init communicator here
   if (send_varname_to_ctx.size() > 0) {
@@ -120,6 +97,7 @@ void ProcessGraph(std::vector<ir::Graph *> graphs, Scope *scope) {
                                                recv_varname_to_ctx, scope);
     operators::distributed::Communicator::GetInstance()->Start();
   }
+#endif
 }
 
 AsyncSSAGraphExecutor::AsyncSSAGraphExecutor(

paddle/fluid/framework/details/broadcast_op_handle.cc

@@ -27,20 +27,17 @@ void BroadcastOpHandle::RunImpl() {
   if (places_.size() == 1) return;
 
   // The input and output may have dummy vars.
-  VarHandle *in_var_handle;
-  {
-    auto in_var_handles = DynamicCast<VarHandle>(inputs_);
-    PADDLE_ENFORCE_EQ(in_var_handles.size(), 1UL,
-                      "The number of input should be one.");
-    in_var_handle = in_var_handles[0];
-  }
-
+  auto in_var_handles = DynamicCast<VarHandle>(inputs_);
   auto out_var_handles = DynamicCast<VarHandle>(outputs_);
 
+  PADDLE_ENFORCE_EQ(in_var_handles.size(), 1UL,
+                    "The number of input should be one.");
   PADDLE_ENFORCE_EQ(
       out_var_handles.size(), places_.size(),
       "The number of output should equal to the number of places.");
 
+  VarHandle *in_var_handle = in_var_handles[0];
+
   WaitInputVarGenerated();
 
   std::vector<const Scope *> var_scopes;

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/build_strategy.cc

@@ -17,7 +17,6 @@ limitations under the License. */
 #include <glog/logging.h>
 #include <memory>
 #include <utility>
-
 #include "paddle/fluid/framework/details/memory_optimize_helper.h"
 #include "paddle/fluid/framework/details/multi_devices_graph_pass.h"
 #include "paddle/fluid/framework/details/multi_devices_graph_print_pass.h"
@@ -46,12 +45,27 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
  public:
   explicit ParallelExecutorPassBuilder(const BuildStrategy &strategy)
       : ir::PassBuilder(), strategy_(strategy) {
+    // Add a graph viz pass to record a graph.
+    if (!strategy_.debug_graphviz_path_.empty()) {
+      auto viz_pass = AppendPass("graph_viz_pass");
+      const std::string graph_path = string::Sprintf(
+          "%s%s", strategy_.debug_graphviz_path_.c_str(), "_original_graph");
+      viz_pass->Set<std::string>("graph_viz_path", new std::string(graph_path));
+    }
+
     if (strategy_.enable_sequential_execution_) {
+      VLOG(10) << "Add sequential_execution_pass";
       AppendPass("sequential_execution_pass");
     }
 
+    // Add op fusion.
+    if (strategy.sync_batch_norm_) {
+      AppendPass("sync_batch_norm_pass");
+    }
+
     // Add op fusion.
     if (strategy.fuse_relu_depthwise_conv_) {
+      VLOG(10) << "Add fuse_relu_depthwise_conv_pass";
       AppendPass("fuse_relu_depthwise_conv_pass");
     }
 
@@ -63,29 +77,50 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
 
     // Add automatically inplace.
     if (strategy_.enable_inplace_) {
+      VLOG(10) << "Add inplace_pass";
       AppendPass("inplace_pass");
     }
 
+    if (strategy_.fuse_elewise_add_act_ops_) {
+      VLOG(10) << "Add fuse_elewise_add_act_pass";
+      AppendPass("fuse_elewise_add_act_pass");
+    }
+
+    // for single card training, fuse_all_reduce_ops is unnecessary.
+    // alloc_continuous_space_for_grad_pass should be before of MultiDevPass.
+    if (strategy_.fuse_all_reduce_ops_) {
+      VLOG(10) << "Add alloc_continuous_space_for_grad_pass";
+      AppendPass("alloc_continuous_space_for_grad_pass");
+    }
+
+    if (strategy_.fuse_all_optimizer_ops_) {
+      if (strategy_.reduce_ == BuildStrategy::ReduceStrategy::kReduce ||
+          strategy_.is_distribution_) {
+        VLOG(3)
+            << "Currently, fuse_all_optimizer_ops only works under AllReduce "
+               "mode.";
+        strategy_.fuse_all_optimizer_ops_ = false;
+      } else {
+        VLOG(10) << "Add alloc_continuous_space_for_grad_pass";
+        AppendPass("alloc_continuous_space_for_grad_pass");
+        // NOTE: fuse_all_xx_ops will count the number of xx operator first,
+        // if the number is zero, fuse_all_reduce_ops will do nothing.
+        // Currently, only one type of optimization algorithm can be fused.
+        VLOG(10) << "Add fuse_adam_op_pass";
+        AppendPass("fuse_adam_op_pass");
+        VLOG(10) << "Add fuse_sgd_op_pass";
+        AppendPass("fuse_sgd_op_pass");
+      }
+    }
+
     // Add a graph viz pass to record a graph.
-    if (!strategy_.debug_graphviz_path_.empty()) {
+    if (!strategy.debug_graphviz_path_.empty()) {
       auto viz_pass = AppendPass("graph_viz_pass");
       const std::string graph_path = string::Sprintf(
-          "%s%s", strategy_.debug_graphviz_path_.c_str(), "_original_graph");
+          "%s%s", strategy_.debug_graphviz_path_.c_str(), "_fused_graph");
       viz_pass->Set<std::string>("graph_viz_path", new std::string(graph_path));
     }
 
-    if (strategy.fuse_elewise_add_act_ops_) {
-      auto fuse_elewise_add_act_pass = AppendPass("fuse_elewise_add_act_pass");
-      // Add a graph viz pass to record a graph.
-      if (!strategy.debug_graphviz_path_.empty()) {
-        auto viz_pass = AppendPass("graph_viz_pass");
-        const std::string graph_path = string::Sprintf(
-            "%s%s", strategy.debug_graphviz_path_.c_str(), "_fused_graph");
-        viz_pass->Set<std::string>("graph_viz_path",
-                                   new std::string(graph_path));
-      }
-    }
-
     CollectiveContext *context = CollectiveContext::GetInstance();
     context->endpoints_ = strategy_.trainers_endpoints_;
     context->trainer_id_ = strategy_.trainer_id_;
@@ -102,11 +137,27 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
     // the de-fact IR, any reuse on Graph is meaningless.
     // A side-effect of that, memory optimize cannot forsee the fetched vars
     // , so fetchlist should be set persistable before call the Run interface.
-    if (strategy.memory_optimize_) {
-      auto memory_optimize_pass = AppendPass("memory_optimize_pass");
+    if (strategy_.memory_optimize_) {
+      VLOG(10) << "Add memory_optimize_pass";
+      AppendPass("memory_optimize_pass");
+    }
+
+    // runtime_context_cache pass should be the last pass to enable the attr of
+    // all original and fused operators. But no operators can be enabled this
+    // attr if putting it after MultiDevPass.
+    if (strategy_.cache_runtime_context_) {
+      VLOG(10) << "Add runtime_context_cache_pass";
+      AppendPass("runtime_context_cache_pass");
     }
 
-    AppendMultiDevPass(strategy);
+    AppendMultiDevPass(strategy_);
+
+    if (strategy_.fuse_all_reduce_ops_) {
+      // NOTE: fuse_all_reduce_ops will count the number of all_reduce operator
+      // first, if the number is zero, fuse_all_reduce_ops will do nothing.
+      VLOG(10) << "Add fuse_all_reduce_op_pass";
+      AppendPass("fuse_all_reduce_op_pass");
+    }
 
     // Add a graph print pass to record a graph with device info.
     if (!strategy_.debug_graphviz_path_.empty()) {
@@ -120,34 +171,41 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
           "graph_printer", new details::GraphvizSSAGraphPrinter);
     }
 
-    // Verify that the graph is correct for multi-device executor.
-    AppendPass("multi_devices_check_pass");
-
-    if (SeqOnlyAllReduceOps(strategy)) {
+    // experimental shows that the program will be faster if append
+    // all_reduce_deps_pass here.
+    if (!strategy_.enable_parallel_graph_ &&
+        (SeqOnlyAllReduceOps(strategy_) ||
+         strategy.reduce_ == BuildStrategy::ReduceStrategy::kAllReduce)) {
+      VLOG(10) << "Add all_reduce_deps_pass";
       AppendPass("all_reduce_deps_pass");
     }
 
     if (strategy_.remove_unnecessary_lock_) {
+      VLOG(10) << "Add modify_op_lock_and_record_event_pass";
       AppendPass("modify_op_lock_and_record_event_pass");
     }
+
+    // Verify that the graph is correct for multi-device executor.
+    AppendPass("multi_devices_check_pass");
   }
 
   // Convert graph to run on multi-devices.
   void AppendMultiDevPass(const BuildStrategy &strategy) {
-    ir::Pass *multi_devices_pass;
+    ir::Pass *multi_devices_pass = nullptr;
 
     if (strategy_.async_mode_) {
       multi_devices_pass = AppendPass("async_multi_devices_pass").get();
     } else if (strategy_.is_distribution_) {
-      VLOG(3) << "multi device parameter server mode";
+      VLOG(10)
+          << "Add dist_multi_devices_pass, multi device parameter server mode";
       multi_devices_pass = AppendPass("dist_multi_devices_pass").get();
     } else {
       if (strategy.reduce_ == BuildStrategy::ReduceStrategy::kAllReduce) {
-        VLOG(3) << "multi devices collective mode with allreduce";
+        VLOG(10) << "Add all_reduce_mode_multi_devices_pass";
         multi_devices_pass =
-            AppendPass("allreduce_mode_multi_devices_pass").get();
+            AppendPass("all_reduce_mode_multi_devices_pass").get();
       } else if (strategy.reduce_ == BuildStrategy::ReduceStrategy::kReduce) {
-        VLOG(3) << "multi deivces collective mode with reduce";
+        VLOG(10) << "Add reduce_mode_multi_devices_pass";
         multi_devices_pass = AppendPass("reduce_mode_multi_devices_pass").get();
       } else {
         PADDLE_THROW("Unknown reduce strategy.");
@@ -177,22 +235,23 @@ bool BuildStrategy::IsMultiDevPass(const std::string &pass_name) const {
   return framework::details::MultiDevSSAGraphBuilder().count(pass_name) > 0;
 }
 
-std::unique_ptr<ir::Graph> BuildStrategy::Apply(
-    std::unique_ptr<ir::Graph> graph,
-    const std::vector<platform::Place> &places,
-    const std::string &loss_var_name, const std::vector<Scope *> &local_scopes,
-    const size_t &nranks,
+ir::Graph *BuildStrategy::Apply(ir::Graph *graph,
+                                const std::vector<platform::Place> &places,
+                                const std::string &loss_var_name,
+                                const std::vector<Scope *> &local_scopes,
+                                const size_t &nranks,
 #if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
-    const bool use_cuda, platform::NCCLContextMap *nccl_ctxs) const {
+                                const bool use_cuda,
+                                platform::NCCLContextMap *nccl_ctxs) const {
 #else
-    const bool use_cuda) const {
+                                const bool use_cuda) const {
 #endif
   VLOG(3) << "apply all passes";
   // Create a default one if not finalized by user.
   CreatePassesFromStrategy(false);
 
   for (std::shared_ptr<ir::Pass> &pass : pass_builder_->AllPasses()) {
-    VLOG(3) << "apply " << pass->Type();
+    VLOG(3) << "BuildStrategy::Apply pass:" << pass->Type();
     if (IsMultiDevPass(pass->Type())) {
       pass->Erase(kPlaces);
       pass->SetNotOwned<const std::vector<platform::Place>>(kPlaces, &places);
@@ -206,9 +265,31 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
 
 #if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
       platform::NCCLContextMap *nctx = use_cuda ? nccl_ctxs : nullptr;
-      pass->Erase("nccl_ctxs");
-      pass->SetNotOwned<platform::NCCLContextMap>("nccl_ctxs", nctx);
+      pass->Erase(kNCCLCtxs);
+      pass->SetNotOwned<platform::NCCLContextMap>(kNCCLCtxs, nctx);
+#endif
+    } else if (pass->Type() == "alloc_continuous_space_for_grad_pass" ||
+               pass->Type() == "fuse_adam_op_pass" ||
+               pass->Type() == "fuse_sgd_op_pass" ||
+               pass->Type() == "fuse_all_reduce_op_pass") {
+      pass->Erase(kPlaces);
+      pass->SetNotOwned<const std::vector<platform::Place>>(kPlaces, &places);
+      pass->Erase(kLocalScopes);
+      pass->SetNotOwned<const std::vector<Scope *>>(kLocalScopes,
+                                                    &local_scopes);
+      if (pass->Type() == "fuse_all_reduce_op_pass") {
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+        platform::NCCLContextMap *nctx = use_cuda ? nccl_ctxs : nullptr;
+        pass->Erase(kNCCLCtxs);
+        pass->SetNotOwned<platform::NCCLContextMap>(kNCCLCtxs, nctx);
 #endif
+      }
+    } else if (pass->Type() == "alloc_continuous_space_for_grad_pass") {
+      pass->Erase(kPlaces);
+      pass->SetNotOwned<const std::vector<platform::Place>>(kPlaces, &places);
+      pass->Erase(kLocalScopes);
+      pass->SetNotOwned<const std::vector<Scope *>>(kLocalScopes,
+                                                    &local_scopes);
     } else if (pass->Type() == "sequential_execution_pass") {
       LOG(INFO) << "set enable_sequential_execution:"
                 << enable_sequential_execution_;
@@ -223,7 +304,7 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
       }
     }
     VLOG(3) << "Start Apply Pass " << pass->Type();
-    graph = pass->Apply(std::move(graph));
+    graph = pass->Apply(graph);
     VLOG(3) << "Finish Apply Pass " << pass->Type();
   }
   VLOG(3) << "All Passes Applied";
@@ -234,12 +315,13 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
 }  // namespace framework
 }  // namespace paddle
 
+USE_PASS(sync_batch_norm_pass);
 USE_PASS(fuse_relu_depthwise_conv_pass);
 USE_PASS(fuse_elewise_add_act_pass);
 USE_PASS(graph_viz_pass);
 USE_PASS(multi_batch_merge_pass);
 USE_PASS(reduce_mode_multi_devices_pass);
-USE_PASS(allreduce_mode_multi_devices_pass);
+USE_PASS(all_reduce_mode_multi_devices_pass);
 USE_PASS(dist_multi_devices_pass);
 USE_PASS(multi_devices_check_pass);
 USE_PASS(multi_devices_print_pass);
@@ -249,4 +331,9 @@ USE_PASS(all_reduce_deps_pass);
 USE_PASS(modify_op_lock_and_record_event_pass);
 USE_PASS(inplace_pass);
 USE_PASS(lock_free_optimize_pass);
+USE_PASS(alloc_continuous_space_for_grad_pass);
 USE_PASS(graph_to_program_pass);
+USE_PASS(fuse_adam_op_pass);
+USE_PASS(fuse_sgd_op_pass);
+USE_PASS(fuse_all_reduce_op_pass);
+USE_PASS(runtime_context_cache_pass);

paddle/fluid/framework/details/build_strategy.h

@@ -16,8 +16,8 @@
 
 #include <memory>
 #include <string>
+#include <utility>
 #include <vector>
-
 #include "paddle/fluid/framework/ir/pass_builder.h"
 #include "paddle/fluid/framework/program_desc.h"
 #include "paddle/fluid/framework/scope.h"
@@ -75,17 +75,27 @@ struct BuildStrategy {
 
   bool fuse_elewise_add_act_ops_{false};
 
+  bool fuse_all_optimizer_ops_{false};
+
+  bool fuse_all_reduce_ops_{false};
+
   bool fuse_relu_depthwise_conv_{false};
 
-  bool memory_optimize_{false};
+  bool sync_batch_norm_{false};
+
+  bool memory_optimize_{true};
   // TODO(dzhwinter):
   // make enable_inplace, memory_optimize_
   // memory_early_delete_ true by default
-  bool enable_inplace_{false};
+  bool enable_inplace_{true};
 
   bool enable_sequential_execution_{false};
 
-  bool fuse_broadcast_op_{false};
+  // NOTE(zcd): In reduce mode, fusing broadcast ops may make the program
+  // faster. Because fusing broadcast OP equals delaying the execution of all
+  // broadcast Ops, in this case, all nccl streams are used only for reduce
+  // operations for a period of time.
+  bool fuse_broadcast_ops_{false};
 
   // FIXME(zcd): is_distribution_ is a temporary field, because in pserver mode,
   // num_trainers is 1, so the current fields of build_strategy doesn't tell if
@@ -97,6 +107,8 @@ struct BuildStrategy {
   std::vector<std::string> trainers_endpoints_;
   bool remove_unnecessary_lock_{true};
 
+  bool cache_runtime_context_{false};
+
   // NOTE:
   // Before you add new options, think if it's a general strategy that works
   // with other strategy. If not, the strategy should be created through
@@ -116,16 +128,15 @@ struct BuildStrategy {
 
   // Apply the passes built by the pass_builder_. The passes will be
   // applied to the Program and output an ir::Graph.
-  std::unique_ptr<ir::Graph> Apply(std::unique_ptr<ir::Graph> graph,
-                                   const std::vector<platform::Place> &places,
-                                   const std::string &loss_var_name,
-                                   const std::vector<Scope *> &local_scopes,
-                                   const size_t &nranks,
+  ir::Graph *Apply(ir::Graph *graph, const std::vector<platform::Place> &places,
+                   const std::string &loss_var_name,
+                   const std::vector<Scope *> &local_scopes,
+                   const size_t &nranks,
 #if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
-                                   const bool use_cuda,
-                                   platform::NCCLContextMap *nccl_ctxs) const;
+                   const bool use_cuda,
+                   platform::NCCLContextMap *nccl_ctxs) const;
 #else
-                                   const bool use_cuda) const;
+                   const bool use_cuda) const;
 #endif
 
   // If set true, ParallelExecutor would build the main_program into multiple

paddle/fluid/framework/details/computation_op_handle.h

@@ -14,6 +14,7 @@
 
 #pragma once
 
+#include <memory>
 #include <string>
 #include <vector>
 
@@ -31,6 +32,8 @@ class ComputationOpHandle : public OpHandleBase {
   ComputationOpHandle(ir::Node *node, Scope *scope, platform::Place place,
                       size_t scope_idx);
 
+  OperatorBase *GetOp() { return op_.get(); }
+
   std::string Name() const override;
 
   const Scope *GetScope() const { return scope_; }

paddle/fluid/framework/details/data_balance_op_handle.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/data_balance_op_handle.h"
-#include <algorithm>
-#include "paddle/fluid/framework/details/container_cast.h"
-
-namespace paddle {
-namespace framework {
-namespace details {
-
-#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
-DataBalanceOpHandle::DataBalanceOpHandle(
-    ir::Node *node, const std::vector<Scope *> &local_scopes,
-    const std::vector<platform::Place> &places,
-    const platform::NCCLContextMap *ctxs)
-    : OpHandleBase(node), local_scopes_(local_scopes), places_(places) {
-  if (ctxs) {
-    for (auto &p : places_) {
-      this->SetDeviceContext(p, ctxs->DevCtx(p));
-    }
-  }
-}
-#else
-DataBalanceOpHandle::DataBalanceOpHandle(
-    ir::Node *node, const std::vector<Scope *> &local_scopes,
-    const std::vector<platform::Place> &places)
-    : OpHandleBase(node), local_scopes_(local_scopes), places_(places) {}
-#endif
-
-std::string DataBalanceOpHandle::Name() const { return "data balance"; }
-
-std::vector<std::array<int, 3>> DataBalanceOpHandle::GetBalancePlan(
-    const std::vector<int> &device_sizes) {
-  int device_num = device_sizes.size();
-  int total_size = 0;
-  int empty_num = 0;
-  std::vector<std::array<int, 2>> size_device_vec;
-  size_device_vec.reserve(device_num);
-  for (int i = 0; i < device_num; ++i) {
-    if (device_sizes[i] == 0) {
-      ++empty_num;
-    }
-    total_size += device_sizes[i];
-    size_device_vec.push_back({{device_sizes[i], i}});
-  }
-  std::vector<std::array<int, 3>> res;
-  if (empty_num == 0) {
-    // No need to do data balance.
-    return res;
-  }
-  if (total_size < device_num) {
-    // No enough data.
-    PADDLE_THROW_EOF();
-  }
-  std::sort(size_device_vec.begin(), size_device_vec.end(),
-            [](const std::array<int, 2> &a, const std::array<int, 2> &b) {
-              return a[0] > b[0];
-            });
-  int expected_device_size = total_size / device_num;
-  int src_idx = 0;
-  for (int dst_idx = device_num - empty_num; dst_idx < device_num; ++dst_idx) {
-    if (size_device_vec[src_idx][0] <= expected_device_size) {
-      ++src_idx;
-      PADDLE_ENFORCE_LT(
-          src_idx, device_num - empty_num,
-          "In current srategy an empty tensor should not be copy source.");
-    }
-    size_device_vec[src_idx][0] -= expected_device_size;
-    size_device_vec[dst_idx][0] += expected_device_size;
-    res.push_back({{size_device_vec[src_idx][1], size_device_vec[dst_idx][1],
-                    expected_device_size}});
-  }
-  return res;
-}
-
-void DataBalanceOpHandle::RunImpl() {
-  PADDLE_ENFORCE_GT(places_.size(), 1UL,
-                    "Data balance can only be enabled when the number of "
-                    "places to run larger than 1.");
-  auto in_var_handles = DynamicCast<VarHandle>(this->Inputs());
-  auto out_var_handles = DynamicCast<VarHandle>(this->Outputs());
-  PADDLE_ENFORCE(in_var_handles.size() % places_.size() == 0);
-  PADDLE_ENFORCE_EQ(
-      in_var_handles.size(), out_var_handles.size(),
-      "The NoDummyInputSize and NoDummyOutputSize should be equal.");
-  int data_num = in_var_handles.size() / places_.size();
-  WaitInputVarGenerated();
-  std::vector<std::vector<LoDTensor *>> lod_tensors(data_num);
-  std::vector<int> device_sizes;
-  for (int i = 0; i < static_cast<int>(in_var_handles.size()); ++i) {
-    PADDLE_ENFORCE_EQ(in_var_handles[i]->name(), out_var_handles[i]->name(),
-                      "The name of input and output should be equal.");
-    int place_idx = i / data_num;
-    int data_idx = i % data_num;
-    auto *local_scope =
-        local_scopes_[place_idx]->FindVar(kLocalExecScopeName)->Get<Scope *>();
-    auto *tensor_var = local_scope->FindVar(in_var_handles[i]->name());
-    PADDLE_ENFORCE(tensor_var->IsType<LoDTensor>());
-    auto *tensor = tensor_var->GetMutable<LoDTensor>();
-    lod_tensors[data_idx].push_back(tensor);
-    int ins_size =
-        tensor->lod().empty() ? tensor->dims()[0] : tensor->NumElements();
-    if (data_idx == 0) {
-      device_sizes.emplace_back(ins_size);
-    } else {
-      PADDLE_ENFORCE_EQ(
-          ins_size, device_sizes.at(place_idx),
-          "All data on the same device shall have the same batch size.");
-    }
-  }
-  const auto &balance_plan = GetBalancePlan(device_sizes);
-
-  for (const auto &trans : balance_plan) {
-    for (int data_idx = 0; data_idx < data_num; ++data_idx) {
-      LoDTensor *src_tensor = lod_tensors[data_idx][trans[0]];
-      LoDTensor *dst_tensor = lod_tensors[data_idx][trans[1]];
-      int trans_ins_size = trans[2];
-      LoD src_lod = src_tensor->lod();
-      int src_ins_size =
-          src_lod.empty() ? src_tensor->dims()[0] : src_tensor->NumElements();
-      int cut_point = src_ins_size - trans_ins_size;
-      if (!src_lod.empty()) {
-        for (auto &level : src_lod) {
-          cut_point = level[cut_point];
-        }
-      }
-      TensorCopySync(src_tensor->Slice(cut_point, src_tensor->dims()[0]),
-                     dst_tensor->place(), dst_tensor);
-      src_tensor->ShareDataWith(src_tensor->Slice(0, cut_point));
-      if (!src_lod.empty()) {
-        dst_tensor->set_lod(SliceInLevel(
-            src_lod, 0, src_ins_size - trans_ins_size, src_ins_size));
-        src_tensor->set_lod(
-            SliceInLevel(src_lod, 0, 0, src_ins_size - trans_ins_size));
-      }
-    }
-  }
-}
-
-}  // namespace details
-}  // namespace framework
-}  // namespace paddle

paddle/fluid/framework/details/eager_deletion_op_handle.cc

@@ -12,6 +12,10 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+#include <memory>
+#include <unordered_set>
+#include <utility>
+
 #include "paddle/fluid/framework/details/eager_deletion_op_handle.h"
 #include "paddle/fluid/framework/lod_tensor_array.h"
 #include "paddle/fluid/framework/scope.h"
@@ -45,6 +49,7 @@ EagerDeletionOpHandle::EagerDeletionOpHandle(
     }
   }
 #endif
+  PADDLE_ENFORCE(!var_names_.empty(), "Var names cannot be empty");
 }
 
 EagerDeletionOpHandle::~EagerDeletionOpHandle() {
@@ -60,15 +65,20 @@ EagerDeletionOpHandle::~EagerDeletionOpHandle() {
 std::string EagerDeletionOpHandle::Name() const { return "eager_deletion"; }
 
 void EagerDeletionOpHandle::RunImpl() {
-  auto *exec_scope = scope_->FindVar(kLocalExecScopeName)->Get<Scope *>();
+  Scope *exec_scope = nullptr;
   std::deque<std::shared_ptr<memory::Allocation>> garbages;
   for (auto &name : var_names_) {
     auto it = ref_cnts_->find(name);
-    // Var not found, not reference count has not decreased to 0
+    // Reference count has not decreased to 0
     if (it == ref_cnts_->end() || it->second.fetch_sub(1) != 1) {
       continue;
     }
 
+    if (!exec_scope) {
+      exec_scope = scope_->FindVar(kLocalExecScopeName)->Get<Scope *>();
+    }
+
+    // Var not found
     auto *var = exec_scope->FindVar(name);
     if (var == nullptr) {
       continue;

paddle/fluid/framework/details/eager_deletion_pass.cc

@@ -12,22 +12,168 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+#include <algorithm>
+#include <functional>
 #include <queue>
 #include <string>
+#include <tuple>
 #include <vector>
 
 #include "paddle/fluid/framework/details/computation_op_handle.h"
 #include "paddle/fluid/framework/details/eager_deletion_op_handle.h"
-#include "paddle/fluid/framework/details/eager_deletion_pass.h"
 #include "paddle/fluid/framework/details/multi_devices_helper.h"
+#include "paddle/fluid/framework/garbage_collector.h"
 #include "paddle/fluid/framework/ir/graph_helper.h"
 
 namespace paddle {
 namespace framework {
 namespace details {
 
-std::unique_ptr<ir::Graph> EagerDeletionPass::ApplyImpl(
-    std::unique_ptr<ir::Graph> graph) const {
+// op -> variables which can be deleted after op runs
+using OpToVarNameSetMap =
+    std::unordered_map<ComputationOpHandle *, std::unordered_set<std::string>>;
+
+// Check whether the variable is LoDTensor based on static VarDesc info
+static bool IsLoDTensor(VarDesc *var) {
+  return var->Proto()->type().type() == proto::VarType::LOD_TENSOR;
+}
+
+// Get memory size of LoDTensor
+static int64_t GetMemorySize(
+    const std::unordered_map<std::string, std::vector<VarHandle *>> &vars,
+    const std::string &var_name) {
+  auto *var_desc = TryGetLatestVarDesc(vars.at(var_name));
+  PADDLE_ENFORCE_NOT_NULL(var_desc);
+  PADDLE_ENFORCE(IsLoDTensor(var_desc));
+  auto dims = var_desc->GetShape();
+  return SizeOfType(var_desc->GetDataType()) *
+         std::accumulate(dims.begin(), dims.end(), static_cast<int64_t>(1),
+                         std::multiplies<int64_t>());
+}
+
+// Split all variables in the graph into LoDTensor and Non-LoDTensor (e.g.
+// SelectedRows, LoDTensorArray)
+// Since partial GC is based on static analysis of memory size of each variable
+// So we should skip SelectedRows and LoDTensorArray here
+static void SplitIntoLoDTensorAndNonLoDTensorVars(
+    const OpToVarNameSetMap &m, const GraphVars &vars,
+    OpToVarNameSetMap *lod_tensors, OpToVarNameSetMap *other_vars) {
+  lod_tensors->clear();
+  other_vars->clear();
+
+  for (auto &op_vars_pair : m) {
+    for (auto &var_name : op_vars_pair.second) {
+      auto *var_desc = TryGetLatestVarDesc(
+          vars[op_vars_pair.first->GetScopeIdx()].at(var_name));
+      if (IsLoDTensor(var_desc)) {
+        (*lod_tensors)[op_vars_pair.first].insert(var_name);
+      } else {
+        (*other_vars)[op_vars_pair.first].insert(var_name);
+      }
+    }
+  }
+}
+
+struct GCVarInfo {
+  GCVarInfo(const std::string &name, int64_t memory_size,
+            ComputationOpHandle *op, size_t scope_idx)
+      : name_(name),
+        memory_size_(memory_size),
+        op_(op),
+        scope_idx_(scope_idx) {}
+
+  std::string name_;         // variable name
+  int64_t memory_size_;      // memory size
+  ComputationOpHandle *op_;  // op after which the variable could be deleted
+  size_t scope_idx_;         // scope index where the variable locates
+
+  int64_t AbsMemorySize() const { return std::abs(memory_size_); }
+};
+
+// Delete delete_lod_tensor_only is not used currently
+static OpToVarNameSetMap ShrinkGCVars(
+    const OpToVarNameSetMap &m, const GraphVars &vars,
+    const std::vector<platform::Place> &places, double fraction_of_memory_size,
+    bool delete_lod_tensor_only = false) {
+  // Do not perform gc when fraction_of_memory_size = 0
+  if (fraction_of_memory_size <= 0.0) return {};
+
+  /**
+   * Step 1: Split all variables into LoDTensor and Non-LoDTensor.
+   * We can only calculate memory size of LoDTensors
+   */
+  OpToVarNameSetMap lod_tensors, other_vars;
+  SplitIntoLoDTensorAndNonLoDTensorVars(m, vars, &lod_tensors, &other_vars);
+
+  // Perform complete gc when fraction_of_memory_size >= 1
+  if (fraction_of_memory_size >= 1.0) {
+    return delete_lod_tensor_only ? lod_tensors : m;
+  }
+
+  /**
+   * Step 2: build GCVarInfos, and calculate total memory sizes of each device
+   */
+
+  // place -> variable info (name, memory size, place, scope_idx)
+  std::map<platform::Place, std::vector<GCVarInfo>> place_to_vars;
+
+  // place -> total memory sizes
+  std::map<platform::Place, int64_t> place_to_size;
+  for (auto &op_vars_pair : lod_tensors) {
+    auto *op = op_vars_pair.first;
+    auto &var_names = op_vars_pair.second;
+    auto scope_idx = op->GetScopeIdx();
+    auto &place = places[scope_idx];
+
+    for (auto &var_name : var_names) {
+      auto var_size = GetMemorySize(vars[scope_idx], var_name);
+      GCVarInfo var_info(var_name, var_size, op, scope_idx);
+      place_to_size[place] += var_info.AbsMemorySize();
+      place_to_vars[place].emplace_back(std::move(var_info));
+    }
+  }
+
+  /**
+   * Step 3: sort GCVarInfos, and only delete the largest variables.
+   */
+  OpToVarNameSetMap partial_vars;
+  for (auto &place_to_var_pair : place_to_vars) {
+    auto &place = place_to_var_pair.first;
+    auto &gc_vars = place_to_var_pair.second;
+    std::sort(gc_vars.begin(), gc_vars.end(),
+              [](const GCVarInfo &var1, const GCVarInfo &var2) {
+                return var1.AbsMemorySize() > var2.AbsMemorySize();
+              });
+
+    int64_t accumulated_size = 0;
+    int64_t size_threshold =
+        static_cast<int64_t>(fraction_of_memory_size * place_to_size[place]);
+    for (size_t i = 0; i < gc_vars.size() && accumulated_size < size_threshold;
+         ++i) {
+      partial_vars[gc_vars[i].op_].insert(gc_vars[i].name_);
+      accumulated_size += gc_vars[i].AbsMemorySize();
+    }
+  }
+
+  /**
+   * Step 4: Combine other vars (SelectedRows, LoDTensorArray)
+   */
+  if (!delete_lod_tensor_only) {
+    for (auto &op_vars_pair : other_vars) {
+      partial_vars[op_vars_pair.first].insert(op_vars_pair.second.begin(),
+                                              op_vars_pair.second.end());
+    }
+  }
+
+  return partial_vars;
+}
+
+class EagerDeletionPass : public ir::Pass {
+ protected:
+  void ApplyImpl(ir::Graph *graph) const override;
+};
+
+void EagerDeletionPass::ApplyImpl(ir::Graph *graph) const {
   auto &ref_cnts =
       Get<std::vector<AtomicReferenceCountMap>>(kRuntimeReferenceCount);
   PADDLE_ENFORCE(ref_cnts.empty(),
@@ -43,9 +189,7 @@ std::unique_ptr<ir::Graph> EagerDeletionPass::ApplyImpl(
 
   // a reverse map of last_live_ops
   //   i.e., last op --> variable names which can be deleted.
-  std::unordered_map<ComputationOpHandle *, std::unordered_set<std::string>>
-      op_vars_map;
-
+  OpToVarNameSetMap op_vars_map;
   for (auto &var_ops_map : last_live_ops) {
     for (auto &var_ops_pair : var_ops_map) {
       const std::string &var_name = var_ops_pair.first;
@@ -55,6 +199,10 @@ std::unique_ptr<ir::Graph> EagerDeletionPass::ApplyImpl(
     }
   }
 
+  double memory_fraction = framework::GetEagerDeletionMemoryFraction();
+
+  op_vars_map = ShrinkGCVars(op_vars_map, vars, places, memory_fraction);
+
   for (auto &pair : op_vars_map) {
     auto *op = pair.first;
     auto &var_names = pair.second;
@@ -85,8 +233,12 @@ std::unique_ptr<ir::Graph> EagerDeletionPass::ApplyImpl(
     eager_deletion_op->AddOutput(dummy_leaf);
   }
 
+  VLOG(10) << "FLAGS_memory_fraction_of_eager_deletion = " << memory_fraction;
   VLOG(10) << "Create " << op_vars_map.size() << " EagerDeletionOpHandle(s)";
-  return graph;
+
+  auto while_op_eager_deletion_pass =
+      ir::PassRegistry::Instance().Get("while_op_eager_deletion_pass");
+  while_op_eager_deletion_pass->Apply(graph);
 }
 
 }  // namespace details
@@ -99,3 +251,5 @@ REGISTER_PASS(eager_deletion_pass,
     .RequirePassAttr(paddle::framework::details::kLastLiveOpsOfVars)
     .RequirePassAttr(paddle::framework::details::kAllPlaces)
     .RequirePassAttr(paddle::framework::details::kGarbageCollector);
+
+USE_PASS(while_op_eager_deletion_pass);

paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.cc

@@ -31,9 +31,10 @@ FastThreadedSSAGraphExecutor::FastThreadedSSAGraphExecutor(
       local_scopes_(local_scopes),
       places_(places),
       graph_(graph),
+      fetch_ctxs_(places),
       pool_(strategy.num_threads_),
-      prepare_pool_(1),  // add one more thread for generate op_deps
-      fetch_ctxs_(places) {
+      // add one more thread for generate op_deps
+      prepare_pool_(1) {
   for (auto &op : ir::FilterByNodeWrapper<OpHandleBase>(*graph_)) {
     int dep = static_cast<int>(op->NotReadyInputSize());
     op_deps_.emplace(op, dep);
@@ -55,6 +56,7 @@ FeedFetchList FastThreadedSSAGraphExecutor::Run(
   fetches.resize(fetch_tensors.size());
   std::unordered_map<std::string, std::vector<VarHandleBase *>> fetched_vars;
   std::vector<FetchOpHandle *> fetch_ops;
+  std::vector<OpHandleBase *> ready_fetch_ops;
 
   for (auto &fetch_var_name : fetch_tensors) {
     for (auto &var_map : graph_->Get<details::GraphVars>(details::kGraphVars)) {
@@ -69,8 +71,9 @@ FeedFetchList FastThreadedSSAGraphExecutor::Run(
     auto &var_name = fetch_tensors[i];
     auto fetched_var_it = fetched_vars.find(var_name);
     PADDLE_ENFORCE(fetched_var_it != fetched_vars.end(),
-                   "Cannot find fetched variable.(Perhaps the main_program "
-                   "is not set to ParallelExecutor)");
+                   "Cannot find fetched variable(%s).(Perhaps the main_program "
+                   "is not set to ParallelExecutor)",
+                   var_name);
 
     auto &vars = fetched_var_it->second;
 
@@ -87,7 +90,11 @@ FeedFetchList FastThreadedSSAGraphExecutor::Run(
       op->AddInput(var);
     }
 
-    (*op_deps)[op] = static_cast<int>(op->NotReadyInputSize());
+    int dep = static_cast<int>(op->NotReadyInputSize());
+    (*op_deps)[op] = dep;
+    if (dep == 0) {
+      ready_fetch_ops.emplace_back(op);
+    }
   }
 
   size_t num_complete = 0;
@@ -96,7 +103,9 @@ FeedFetchList FastThreadedSSAGraphExecutor::Run(
   for (auto op : bootstrap_ops_) {
     RunOpAsync(op_deps.get(), op, complete_q);
   }
-
+  for (auto op : ready_fetch_ops) {
+    RunOpAsync(op_deps.get(), op, complete_q);
+  }
   while (num_complete != op_deps->size()) {
     size_t num_comp = complete_q->Pop();
     if (num_comp == -1UL) {

paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h

@@ -14,7 +14,9 @@
 
 #pragma once
 #include <ThreadPool.h>
+#include <memory>
 #include <string>
+#include <unordered_map>
 #include <vector>
 #include "paddle/fluid/framework/blocking_queue.h"
 #include "paddle/fluid/framework/details/exception_holder.h"
@@ -37,6 +39,8 @@ class FastThreadedSSAGraphExecutor : public SSAGraphExecutor {
   const ir::Graph &Graph() const override;
 
  private:
+  // Note(zcd): the ThreadPool should be placed last so that ThreadPool should
+  // be destroyed first.
   ExecutionStrategy strategy_;
   std::vector<Scope *> local_scopes_;
   std::vector<platform::Place> places_;
@@ -45,21 +49,22 @@ class FastThreadedSSAGraphExecutor : public SSAGraphExecutor {
   std::unordered_map<OpHandleBase *, int> op_deps_;
   std::vector<OpHandleBase *> bootstrap_ops_;
 
-  ::ThreadPool pool_;
-  ::ThreadPool prepare_pool_;
   platform::DeviceContextPool fetch_ctxs_;
   std::atomic<int> remaining_;
 
+  std::future<
+      std::unique_ptr<std::unordered_map<OpHandleBase *, std::atomic<int>>>>
+      atomic_op_deps_;
+  ExceptionHolder exception_;
+
+  ::ThreadPool pool_;
+  ::ThreadPool prepare_pool_;
+
   void RunOpAsync(std::unordered_map<OpHandleBase *, std::atomic<int>> *op_deps,
                   OpHandleBase *op,
                   const std::shared_ptr<BlockingQueue<size_t>> &complete_q);
 
   void PrepareAtomicOpDeps();
-
-  std::future<
-      std::unique_ptr<std::unordered_map<OpHandleBase *, std::atomic<int>>>>
-      atomic_op_deps_;
-  ExceptionHolder exception_;
 };
 }  // namespace details
 }  // namespace framework

paddle/fluid/framework/details/fetch_op_handle.h

@@ -39,6 +39,8 @@ struct FetchOpHandle : public OpHandleBase {
 
   std::string Name() const override;
 
+  bool IsMultiDeviceTransfer() override;
+
  protected:
   void RunImpl() override;
 

paddle/fluid/framework/details/modify_op_lock_and_record_event_pass.h

@@ -23,8 +23,7 @@ namespace details {
 
 class ModifyOpLockAndRecordEventPass : 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