Merge branch 'develop' into ir5

CMakeLists.txt

@@ -24,6 +24,9 @@ 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}")
+if(WIN32)
+    set(CMAKE_STATIC_LIBRARY_PREFIX lib)
+endif(WIN32)
 
 if(NOT CMAKE_CROSSCOMPILING)
     find_package(CUDA QUIET)
@@ -138,12 +141,6 @@ else()
     set(THIRD_PARTY_BUILD_TYPE Release)
 endif()
 
-if(WITH_MKL)
-  option(MKL_SPLIT_GEMM "PaddlePaddle MKL gemm would split to small ones" OFF)
-  if (MKL_SPLIT_GEMM)
-    add_definitions(-DPADDLE_MKL_SPLIT_GEMM)
-  endif()
-endif()
 set(WITH_MKLML ${WITH_MKL})
 if (NOT DEFINED WITH_MKLDNN)
     if (WITH_MKL AND AVX2_FOUND)
@@ -171,7 +168,6 @@ include(external/python)    # download, build, install python
 include(external/openblas)  # download, build, install openblas
 include(external/mkldnn)    # download, build, install mkldnn
 include(external/swig)      # download, build, install swig
-include(external/warpctc)   # download, build, install warpctc
 include(external/boost)     # download boost
 include(external/any)       # download libn::any
 include(external/eigen)     # download eigen3
@@ -179,6 +175,14 @@ include(external/pybind11)  # download pybind11
 include(external/cares)
 include(external/cub)
 
+if (NOT WIN32)
+# there is no official support of snappystream, warpctc, nccl, cupti in windows
+include(external/snappy)    # download snappy
+include(external/snappystream) # download snappystream
+include(external/warpctc)   # download, build, install warpctc
+include(cupti)
+endif (NOT WIN32)
+
 if(WITH_DISTRIBUTE)
     if(WITH_GRPC)
         include(external/grpc)
@@ -200,13 +204,10 @@ if(WITH_BRPC_RDMA)
     endif()
 endif()
 
-include(external/snappy)    # download snappy
-include(external/snappystream)
-include(external/threadpool)
 
+include(external/threadpool)
 include(flags)              # set paddle compile flags
 include(cudnn)              # set cudnn libraries, must before configure
-include(cupti)
 include(configure)          # add paddle env configuration
 
 if(WITH_GPU)

cmake/configure.cmake

@@ -61,6 +61,11 @@ if(NOT CMAKE_CROSSCOMPILING)
     endif()
 endif()
 
+if(WIN32)
+  # windows stupid compile option for all targets.
+  add_definitions(-D_XKEYCHECK_H)
+endif(WIN32)
+
 if(NOT WITH_GOLANG)
     add_definitions(-DPADDLE_WITHOUT_GOLANG)
 endif(NOT WITH_GOLANG)

cmake/external/boost.cmake

@@ -28,7 +28,12 @@ if((NOT DEFINED BOOST_TAR) OR (NOT DEFINED BOOST_URL))
     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)
 endif()
-MESSAGE(STATUS "BOOST_TAR: ${BOOST_TAR}, BOOST_URL: ${BOOST_URL}")
+IF (WIN32)
+    MESSAGE(WARNING, "In windows, boost can not be downloaded automaticlly, please build it manually and put it at " ${THIRD_PARTY_PATH}install/boost)
+else()
+    MESSAGE(STATUS "BOOST_TAR: ${BOOST_TAR}, BOOST_URL: ${BOOST_URL}")
+ENDIF(WIN32)
+
 set(BOOST_SOURCES_DIR ${THIRD_PARTY_PATH}/boost)
 set(BOOST_DOWNLOAD_DIR  "${BOOST_SOURCES_DIR}/src/${BOOST_PROJECT}")
 set(BOOST_INCLUDE_DIR "${BOOST_DOWNLOAD_DIR}/${BOOST_TAR}" CACHE PATH "boost include directory." FORCE)
@@ -36,12 +41,13 @@ set_directory_properties(PROPERTIES CLEAN_NO_CUSTOM 1)
 
 include_directories(${BOOST_INCLUDE_DIR})
 
+if (NOT WIN32)
 ExternalProject_Add(
     ${BOOST_PROJECT}
     ${EXTERNAL_PROJECT_LOG_ARGS}
     DOWNLOAD_DIR          ${BOOST_DOWNLOAD_DIR}
     DOWNLOAD_COMMAND      wget --no-check-certificate ${BOOST_URL} -c -q -O ${BOOST_TAR}.tar.gz
-                          && tar zxf ${BOOST_TAR}.tar.gz
+    && tar zxf ${BOOST_TAR}.tar.gz
     DOWNLOAD_NO_PROGRESS  1
     PREFIX                ${BOOST_SOURCES_DIR}
     CONFIGURE_COMMAND     ""
@@ -49,8 +55,9 @@ ExternalProject_Add(
     INSTALL_COMMAND       ""
     UPDATE_COMMAND        ""
 )
+endif(NOT WIN32)
 
-if (${CMAKE_VERSION} VERSION_LESS "3.3.0")
+if (${CMAKE_VERSION} VERSION_LESS "3.3.0" OR NOT WIN32)
     set(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/boost_dummy.c)
     file(WRITE ${dummyfile} "const char *dummy = \"${dummyfile}\";")
     add_library(boost STATIC ${dummyfile})

cmake/external/gflags.cmake

@@ -18,7 +18,7 @@ SET(GFLAGS_SOURCES_DIR ${THIRD_PARTY_PATH}/gflags)
 SET(GFLAGS_INSTALL_DIR ${THIRD_PARTY_PATH}/install/gflags)
 SET(GFLAGS_INCLUDE_DIR "${GFLAGS_INSTALL_DIR}/include" CACHE PATH "gflags include directory." FORCE)
 IF(WIN32)
-  set(GFLAGS_LIBRARIES "${GFLAGS_INSTALL_DIR}/lib/gflags.lib" CACHE FILEPATH "GFLAGS_LIBRARIES" FORCE)
+  set(GFLAGS_LIBRARIES "${GFLAGS_INSTALL_DIR}/lib/libgflags.lib" CACHE FILEPATH "GFLAGS_LIBRARIES" FORCE)
 ELSE(WIN32)
   set(GFLAGS_LIBRARIES "${GFLAGS_INSTALL_DIR}/lib/libgflags.a" CACHE FILEPATH "GFLAGS_LIBRARIES" FORCE)
 ENDIF(WIN32)
@@ -45,7 +45,13 @@ ExternalProject_Add(
                      -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON
                      -DCMAKE_BUILD_TYPE:STRING=${THIRD_PARTY_BUILD_TYPE}
 )
-
+IF(WIN32)
+  IF(NOT EXISTS "${GFLAGS_INSTALL_DIR}/lib/libgflags.lib")
+    add_custom_command(TARGET extern_gflags POST_BUILD
+    COMMAND cmake -E rename ${GFLAGS_INSTALL_DIR}/lib/gflags_static.lib ${GFLAGS_INSTALL_DIR}/lib/libgflags.lib
+  )
+  ENDIF()
+ENDIF(WIN32)
 ADD_LIBRARY(gflags STATIC IMPORTED GLOBAL)
 SET_PROPERTY(TARGET gflags PROPERTY IMPORTED_LOCATION ${GFLAGS_LIBRARIES})
 ADD_DEPENDENCIES(gflags extern_gflags)
@@ -60,3 +66,4 @@ IF(WITH_C_API)
     INSTALL(FILES ${GFLAGS_LIBRARIES} DESTINATION third_party/gflags/lib)
   ENDIF()
 ENDIF()
+

cmake/external/glog.cmake

@@ -60,6 +60,13 @@ ExternalProject_Add(
                      -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON
                      -DCMAKE_BUILD_TYPE:STRING=${THIRD_PARTY_BUILD_TYPE}
 )
+IF(WIN32)
+  IF(NOT EXISTS "${GLOG_INSTALL_DIR}/lib/libglog.lib")
+    add_custom_command(TARGET extern_glog POST_BUILD
+    COMMAND cmake -E rename ${GLOG_INSTALL_DIR}/lib/glog.lib ${GLOG_INSTALL_DIR}/lib/libglog.lib
+  )
+  ENDIF()
+ENDIF(WIN32)
 
 ADD_LIBRARY(glog STATIC IMPORTED GLOBAL)
 SET_PROPERTY(TARGET glog PROPERTY IMPORTED_LOCATION ${GLOG_LIBRARIES})

cmake/external/mkldnn.cmake

@@ -54,7 +54,7 @@ ExternalProject_Add(
     ${EXTERNAL_PROJECT_LOG_ARGS}
     DEPENDS             ${MKLDNN_DEPENDS}
     GIT_REPOSITORY      "https://github.com/01org/mkl-dnn.git"
-    GIT_TAG             "a29d8487a63afca3d5b8c5bbdbb473cf8ccc6e51"
+    GIT_TAG             "64e03a1939e0d526aa8e9f2e3f7dc0ad8d372944"
     PREFIX              ${MKLDNN_SOURCES_DIR}
     UPDATE_COMMAND      ""
     CMAKE_ARGS          -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}

cmake/external/openblas.cmake

@@ -17,20 +17,29 @@ IF(USE_EIGEN_FOR_BLAS)
 ENDIF(USE_EIGEN_FOR_BLAS)
 
 INCLUDE(cblas)
+# IF(WIN32 AND NOT ${CBLAS_FOUND})
+
+
 
 IF(NOT ${CBLAS_FOUND})
+
     INCLUDE(ExternalProject)
 
     SET(CBLAS_SOURCES_DIR ${THIRD_PARTY_PATH}/openblas)
     SET(CBLAS_INSTALL_DIR ${THIRD_PARTY_PATH}/install/openblas)
-    SET(CBLAS_INC_DIR "${CBLAS_INSTALL_DIR}/include" CACHE PATH "openblas include directory." FORCE)
+    SET(CBLAS_INCLUDE_DIR "${CBLAS_INSTALL_DIR}/include" CACHE PATH "openblas include directory." FORCE)
 
     SET(CBLAS_LIBRARIES
         "${CBLAS_INSTALL_DIR}/lib/${CMAKE_STATIC_LIBRARY_PREFIX}openblas${CMAKE_STATIC_LIBRARY_SUFFIX}"
         CACHE FILEPATH "openblas library." FORCE)
 
     ADD_DEFINITIONS(-DPADDLE_USE_OPENBLAS)
+    IF (WIN32)
+        SET(CBLAS_FOUND true)
+        MESSAGE(WARNING, "In windows, openblas only support msvc build, please build it manually and put it at " ${CBLAS_INSTALL_DIR})
+    ENDIF(WIN32)
 
+    IF (NOT WIN32)
     SET(OPENBLAS_CC "${CMAKE_C_COMPILER} -Wno-unused-but-set-variable -Wno-unused-variable")
     SET(OPENBLAS_COMMIT "v0.2.20")
 
@@ -69,7 +78,6 @@ IF(NOT ${CBLAS_FOUND})
     ENDIF()
 
     SET(COMMON_ARGS CC=${OPENBLAS_CC} NO_SHARED=1 NO_LAPACK=1 libs)
-
     ExternalProject_Add(
         extern_openblas
         ${EXTERNAL_PROJECT_LOG_ARGS}
@@ -84,9 +92,11 @@ IF(NOT ${CBLAS_FOUND})
         UPDATE_COMMAND      ""
         CONFIGURE_COMMAND   ""
     )
+    ELSE()
+    ENDIF(NOT WIN32)
     SET(CBLAS_PROVIDER openblas)
     IF(WITH_C_API)
-        INSTALL(DIRECTORY ${CBLAS_INC_DIR} DESTINATION third_party/openblas)
+        INSTALL(DIRECTORY ${CBLAS_INCLUDE_DIR} DESTINATION third_party/openblas)
         # Because libopenblas.a is a symbolic link of another library, thus need to
         # install the whole directory.
         IF(ANDROID)
@@ -107,7 +117,8 @@ IF(NOT ${CBLAS_FOUND})
 ENDIF(NOT ${CBLAS_FOUND})
 
 MESSAGE(STATUS "BLAS library: ${CBLAS_LIBRARIES}")
-INCLUDE_DIRECTORIES(${CBLAS_INC_DIR})
+MESSAGE(STATUS "BLAS Include: ${CBLAS_INCLUDE_DIR}")
+INCLUDE_DIRECTORIES(${CBLAS_INCLUDE_DIR})
 
 # FIXME(gangliao): generate cblas target to track all high performance
 # linear algebra libraries for cc_library(xxx SRCS xxx.c DEPS cblas)

cmake/external/protobuf.cmake

@@ -14,11 +14,14 @@
 
 INCLUDE(ExternalProject)
 # Always invoke `FIND_PACKAGE(Protobuf)` for importing function protobuf_generate_cpp
+IF(NOT WIN32)
 FIND_PACKAGE(Protobuf QUIET)
+ENDIF(NOT WIN32)
 macro(UNSET_VAR VAR_NAME)
     UNSET(${VAR_NAME} CACHE)
     UNSET(${VAR_NAME})
 endmacro()
+
 UNSET_VAR(PROTOBUF_INCLUDE_DIR)
 UNSET_VAR(PROTOBUF_FOUND)
 UNSET_VAR(PROTOBUF_PROTOC_EXECUTABLE)
@@ -94,12 +97,14 @@ macro(PROMPT_PROTOBUF_LIB)
     SET(protobuf_DEPS ${ARGN})
 
     MESSAGE(STATUS "Protobuf protoc executable: ${PROTOBUF_PROTOC_EXECUTABLE}")
+    MESSAGE(STATUS "Protobuf-lite library: ${PROTOBUF_LITE_LIBRARY}")
     MESSAGE(STATUS "Protobuf library: ${PROTOBUF_LIBRARY}")
+    MESSAGE(STATUS "Protoc library: ${PROTOBUF_PROTOC_LIBRARY}")
     MESSAGE(STATUS "Protobuf version: ${PROTOBUF_VERSION}")
     INCLUDE_DIRECTORIES(${PROTOBUF_INCLUDE_DIR})
 
     # Assuming that all the protobuf libraries are of the same type.
-    IF(${PROTOBUF_LIBRARY} MATCHES "${CMAKE_STATIC_LIBRARY_SUFFIX}$")
+    IF(${PROTOBUF_LIBRARY} MATCHES ${CMAKE_STATIC_LIBRARY_SUFFIX})
         SET(protobuf_LIBTYPE STATIC)
     ELSEIF(${PROTOBUF_LIBRARY} MATCHES "${CMAKE_SHARED_LIBRARY_SUFFIX}$")
         SET(protobuf_LIBTYPE SHARED)
@@ -137,18 +142,25 @@ macro(SET_PROTOBUF_VERSION)
 endmacro()
 
 set(PROTOBUF_ROOT "" CACHE PATH "Folder contains protobuf")
+IF (WIN32)
+    SET(PROTOBUF_ROOT ${THIRD_PARTY_PATH}/install/protobuf)
+    MESSAGE(WARNING, "In windows, protobuf only support msvc build, please build it manually and put it at " ${PROTOBUF_ROOT})
+ENDIF(WIN32)
+
 if (NOT "${PROTOBUF_ROOT}" STREQUAL "")
+
     find_path(PROTOBUF_INCLUDE_DIR google/protobuf/message.h PATHS ${PROTOBUF_ROOT}/include NO_DEFAULT_PATH)
-    find_library(PROTOBUF_LIBRARY protobuf PATHS ${PROTOBUF_ROOT}/lib NO_DEFAULT_PATH)
-    find_library(PROTOBUF_LITE_LIBRARY protobuf-lite PATHS ${PROTOBUF_ROOT}/lib NO_DEFAULT_PATH)
-    find_library(PROTOBUF_PROTOC_LIBRARY protoc PATHS ${PROTOBUF_ROOT}/lib NO_DEFAULT_PATH)
+    find_library(PROTOBUF_LIBRARY protobuf libprotobuf.lib PATHS ${PROTOBUF_ROOT}/lib NO_DEFAULT_PATH)
+    find_library(PROTOBUF_LITE_LIBRARY protobuf-lite libprotobuf-lite.lib PATHS ${PROTOBUF_ROOT}/lib NO_DEFAULT_PATH)
+    find_library(PROTOBUF_PROTOC_LIBRARY protoc libprotoc.lib PATHS ${PROTOBUF_ROOT}/lib NO_DEFAULT_PATH)
     find_program(PROTOBUF_PROTOC_EXECUTABLE protoc PATHS ${PROTOBUF_ROOT}/bin NO_DEFAULT_PATH)
     if (PROTOBUF_INCLUDE_DIR AND PROTOBUF_LIBRARY AND PROTOBUF_LITE_LIBRARY AND PROTOBUF_PROTOC_LIBRARY AND PROTOBUF_PROTOC_EXECUTABLE)
         message(STATUS "Using custom protobuf library in ${PROTOBUF_ROOT}.")
+        SET(PROTOBUF_FOUND true)
         SET_PROTOBUF_VERSION()
         PROMPT_PROTOBUF_LIB()
     else()
-        message(WARNING "Cannot find protobuf library in ${PROTOBUF_ROOT}.")
+        message(WARNING "Cannot find protobuf library in ${PROTOBUF_ROOT}")
     endif()
 endif()
 
@@ -239,6 +251,7 @@ IF(CMAKE_CROSSCOMPILING)
         CACHE FILEPATH "protobuf executable." FORCE)
 ENDIF()
 
+
 IF(NOT PROTOBUF_FOUND)
     build_protobuf(extern_protobuf FALSE)
 

cmake/generic.cmake

@@ -148,7 +148,8 @@ function(merge_static_libs TARGET_NAME)
       COMMAND rm "${CMAKE_CURRENT_BINARY_DIR}/lib${TARGET_NAME}.a"
       COMMAND /usr/bin/libtool -static -o "${CMAKE_CURRENT_BINARY_DIR}/lib${TARGET_NAME}.a" ${libfiles}
       )
-  else() # general UNIX: use "ar" to extract objects and re-add to a common lib
+  endif(APPLE)
+  if(LINUX) # general UNIX: use "ar" to extract objects and re-add to a common lib
     set(target_DIR ${CMAKE_CURRENT_BINARY_DIR}/${TARGET_NAME}.dir)
 
     foreach(lib ${libs})
@@ -187,7 +188,36 @@ function(merge_static_libs TARGET_NAME)
         COMMAND ${CMAKE_AR} crs ${target_LIBNAME} `find ${target_DIR} -name '*.o'`
         COMMAND ${CMAKE_RANLIB} ${target_LIBNAME}
         WORKING_DIRECTORY ${target_DIR})
-  endif()
+  endif(LINUX)
+  if(WIN32) # windows do not support gcc/nvcc combined compiling. Use msvc lib.exe to merge libs.
+    # Make the generated dummy source file depended on all static input
+    # libs. If input lib changes,the source file is touched
+    # which causes the desired effect (relink).
+    add_custom_command(OUTPUT ${target_SRCS}
+      COMMAND ${CMAKE_COMMAND} -E touch ${target_SRCS}
+      DEPENDS ${libs})
+
+    # Generate dummy staic lib
+    file(WRITE ${target_SRCS} "const char *dummy_${TARGET_NAME} = \"${target_SRCS}\";")
+    add_library(${TARGET_NAME} STATIC ${target_SRCS})
+    target_link_libraries(${TARGET_NAME} ${libs_deps})
+
+    foreach(lib ${libs})
+      # Get the file names of the libraries to be merged
+      #if(NOT $<TARGET_FILE:${lib}> MATCHES "lib.*\\.lib")
+      #  message("library" ${lib})
+      #  set(libfiles ${libfiles} lib$<TARGET_FILE:${lib}>)
+      #else()
+      set(libfiles ${libfiles} $<TARGET_FILE:${lib}>)
+      #endif()
+    endforeach()
+   
+    # windows cmd return error in clean env.
+    # COMMAND del "${CMAKE_CURRENT_BINARY_DIR}/${CMAKE_BUILD_TYPE}/${TARGET_NAME}.lib"
+    add_custom_command(TARGET ${TARGET_NAME} POST_BUILD
+      COMMAND lib /OUT:${CMAKE_CURRENT_BINARY_DIR}/lib${TARGET_NAME}.lib ${libfiles}
+      )
+  endif(WIN32)
 endfunction(merge_static_libs)
 
 function(cc_library TARGET_NAME)
@@ -195,6 +225,10 @@ function(cc_library TARGET_NAME)
   set(oneValueArgs "")
   set(multiValueArgs SRCS DEPS)
   cmake_parse_arguments(cc_library "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
+  if(WIN32)
+      # add libxxx.lib prefix in windows
+      set(${TARGET_NAME}_LIB_NAME "${CMAKE_STATIC_LIBRARY_PREFIX}${TARGET_NAME}${CMAKE_STATIC_LIBRARY_SUFFIX}" CACHE STRING "output library name for target ${TARGET_NAME}")
+  endif(WIN32)
   if(cc_library_SRCS)
     if(cc_library_SHARED OR cc_library_shared) # build *.so
       add_library(${TARGET_NAME} SHARED ${cc_library_SRCS})

cmake/inference_lib.cmake

@@ -101,6 +101,7 @@ if(WITH_MKLDNN)
   )
 endif()
 
+if (NOT WIN32)
 if(NOT MOBILE_INFERENCE AND NOT RPI)
   set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/snappy")
   copy(snappy_lib
@@ -120,15 +121,23 @@ if(NOT MOBILE_INFERENCE AND NOT RPI)
     DSTS ${dst_dir} ${dst_dir}/lib
     DEPS zlib)
 endif()
+endif(NOT WIN32)
 
 # paddle fluid module
 set(src_dir "${PADDLE_SOURCE_DIR}/paddle/fluid")
 set(dst_dir "${FLUID_INSTALL_DIR}/paddle/fluid")
 set(module "framework")
+if (NOT WIN32)
 copy(framework_lib DEPS framework_py_proto 
   SRCS ${src_dir}/${module}/*.h ${src_dir}/${module}/details/*.h ${PADDLE_BINARY_DIR}/paddle/fluid/framework/framework.pb.h
   DSTS ${dst_dir}/${module} ${dst_dir}/${module}/details ${dst_dir}/${module}
 )
+else()
+copy(framework_lib
+  SRCS ${src_dir}/${module}/*.h ${src_dir}/${module}/details/*.h ${PADDLE_BINARY_DIR}/paddle/fluid/framework/framework.pb.h
+  DSTS ${dst_dir}/${module} ${dst_dir}/${module}/details ${dst_dir}/${module}
+)
+endif(NOT WIN32)
 
 set(module "memory")
 copy(memory_lib

doc/fluid/design/dist_train/dist_train_nccl2.md

 # Distributed Training with NCCL2
 
 We design a pattern that can enable training with `ParallelExecutor` and
-using [NCCL2](https://developer.nvidia.com/nccl) as it's collective
+use [NCCL2](https://developer.nvidia.com/nccl) as it's collective
 communication library.
 
 In `ParallelExecutor` we can use `AllReduce` or `Reduce` and `Broadcast`
@@ -9,14 +9,14 @@ to do multi GPU training. And if we initialize NCCL2 communicators as
 ranks in a distributed environment, we can simply run the `ParallelExecutor`
 as a distributed program! The only thing that may be different than in
 the single node version is that we need to broadcast the NCCL unique ID
-to all the nodes, and initialize communicators using that ID, so NCCL2
-will know each other as ranks.
+to all the nodes and initialize communicators using that ID, so NCCL2
+can know each other as ranks.
 
 To achieve this feature, we introduce a new operator: `gen_nccl_id` op,
 so we are ***not*** "bind to" running NCCL2 with MPI, we can run it in
-what ever platform you like.
+whatever platform you like.
 
-It have two running modes:
+It has two running modes:
 
 1. Generate and broadcast mode, which should be used on trainer 0;
 1. Listen and fetch mode, which should be used on trainers other than 0.
@@ -29,7 +29,7 @@ initialize NCCL communicator objects.
 <img src="src/ncc2_design.png">
 
 The above figure indicates the general process when training with NCCL2
-distributed. Each trainer have the number of communicators equal to the
+distributed. Each trainer has the number of communicators equal to the
 number of GPUs, but the ranks should match the global ranks number: here
 we have total 8 GPUs, so `nranks==8`, for each trainer, the ranks should
 be from 0 ~ 3 on trainer 0 and 4 ~ 7 on trainer 1.

doc/fluid/dev/new_op_cn.md

@@ -36,19 +36,19 @@
 <tbody>
 <tr>
 <td>OpProtoMake定义 </td>
-<td>`.cc`文件，Backward Op不需要定义OpProtoMake </td>
+<td>.cc 文件，Backward Op不需要定义OpProtoMake </td>
 </tr>
 <tr>
 <td>Op定义 </td>
-<td> `.cc`文件</td>
+<td> .cc 文件</td>
 </tr>
 <tr>
 <td>Kernel实现 </td>
-<td> CPU、CUDA共享Kernel实现在`.h`文件中，否则，CPU 实现在`.cc`文件中，CUDA 实现在`.cu`文件中。</td>
+<td> CPU、CUDA共享Kernel实现在.h 文件中，否则，CPU 实现在.cc 文件中，CUDA 实现在.cu 文件中。</td>
 </tr>
 <tr>
 <td>注册Op </td>
-<td> Op注册实现在`.cc`文件；Kernel注册CPU实现在`.cc`文件中，CUDA实现在`.cu`文件中</td>
+<td> Op注册实现在.cc 文件；Kernel注册CPU实现在.cc 文件中，CUDA实现在.cu 文件中</td>
 </tr>
 </tbody>
 </table>
@@ -391,7 +391,7 @@ PADDLE_ENFORCE(ctx->HasInput("X"), "");
 ```
 问题示例2 ：提示信息过于简单
 ```
-PADDLE_ENFORCE(i != nullptr, "I must be set"); // I是什么？
+PADDLE_ENFORCE(i != nullptr, "i must be set"); // i是什么？
 ```
 
 2. 在报错信息中使用开发人员定义的变量缩写，不易理解！

doc/fluid/dev/releasing_process_en.md

@@ -50,6 +50,33 @@ pop-up box, choose the current release branch and click "Run Build" button. You
 * pypi does not allow overwrite the already uploaded version of wheel package, even if you delete the
   old version. you must change the version number before upload a new one.
 
+### Publish wheel Packages for MacOS
+
+You need to build the binary wheel package for MacOS before publishing, to
+make sure that the package can be used by many versions of MacOS
+(10.11, 10.12, 10.13) and different python installs (python.org, homebrew, etc.),
+you must build the package ***exactly*** following below steps:
+
+Build steps:
+
+1. install python from python.org downloads, and make sure it's currently in use
+   in your system.
+1. `export MACOSX_DEPLOYMENT_TARGET=10.11`, use `10.11` is enough for recent versions.
+1. `git clone https://github.com/PaddlePaddle/Paddle.git && cd Paddle && mkdir build && cd build`
+1. `cmake -DWITH_GPU=OFF -DWITH_MKL=OFF -DWITH_SYSTEM_BLAS=OFF  ..`, make sure the output of `cmake` command is using the correct python interpreter installed from python.org
+1. `make -j`
+1. `pip install delocate`
+1. `mkdir fixed_wheel && delocate-wheel -w fixed_wheel python/dist/*.whl`
+
+Then the whl under `fixed_wheel` is ready to upload.
+
+Install steps:
+
+1. run `pip install paddlepaddle...whl`
+1. find the `libpython.dylib` that are currently in use:
+    - for python.org package installs, do nothing.
+    - for other python installs, find the path of `libpython*.dylib` and `export LD_LIBRARY_PATH=you path && DYLD_LIBRARY_PATH=your path`
+
 ## Publish Docker Images
 
 Our CI tool will push latest images to DockerHub, so we only need to push a version tag like:

doc/fluid/howto/cluster/nccl2_rdma_training.md

 # Distributed Training with NCCL2 and RDMA
 
-When doing distributed multi-GPU training, network bandwith often becomes the
-bottle neck. We introduce a way to use NCCL2 to do such training job to
-achieve best performace.
+When doing distributed multi-GPU training, network bandwidth often becomes the
+bottleneck. We introduce a way to use NCCL2 to do such training job to
+achieve best performance.
 
-## Prepare Hardwares with RDMA and Multiple GPUs
+## Prepare Hardware with RDMA and Multiple GPUs
 
-I'm using two Linux servers each of them is installed with 8 GPUs and
+I'm using two Linux servers each of them installed with 8 GPUs and
 one 100Gb RDMA card.
 Base environment is:
 
@@ -25,7 +25,7 @@ In general, the steps including:
 1. Use docker to run tests and make sure GPUs and RDMA can work inside
    the container.
 
-I'll ommit section "Install GPU drivers" because we can find it easily
+I'll omit the section "Install GPU drivers" because we can find it easily
 somewhere else.
 
 ### Install RDMA drivers
@@ -33,7 +33,7 @@ somewhere else.
 For my case, I've got two machines with device
 "Mellanox Technologies MT27700 Family [ConnectX-4]" installed. The OS was
 "CentOS 7.4" and I updated the kernel to version 4.4 so that docker can
-work with latest overlay2 filesystem.
+work with the latest overlay2 filesystem.
 
 ***NOTE: before you start, make sure you have a way to get a console
 of the server other than ssh because we may need to re-configure the
@@ -45,14 +45,14 @@ network device.***
 1. Run `./mlnxofedinstall --add-kernel-support` in the software package.
 1. Run `/etc/init.d/openibd restart` to make everything work, note that
    this operation may cause the network goes down if you are using this
-   RDMA device as default network device and use ssh to login the server.
+   RDMA device as default network device and use ssh to log in the server.
 1. Re-configure the network interface, for example:
    `ifconfig eth2 192.168.16.30/20 up`, then add routes if needed:
    `ip route add default via 192.168.16.1 dev eth2`.
 1. Do the same thing on the other node.
 1. Use `ping` to test if the two nodes have typical ICMP connection.
 1. Use either `udaddy` or `ib_write_bw` to test the network connection is
-   ready and have the desired bandwith.
+   ready and have the desired bandwidth.
 
 ### Prepare Docker Image to Run RDMA Programs
 
@@ -60,7 +60,7 @@ network device.***
    package in it.
 1. Start a docker container and mount GPU driver libs into it (you can
    skip this step if you are using nvidia-docker).
-1. Mount RDMA dirvers and libs into the docker image (see below section),
+1. Mount RDMA drivers and libs into the docker image (see below section),
    also `udaddy` and `ib_write_bw` if needed.
 1. Mount GPU devices and RDMA devices into the container using `--device`
    or just use privileged mode `--privileged`.

doc/fluid/new_docs/advanced_usage/deploy/native_infer.rst

@@ -9,8 +9,6 @@ Paddle 预测 API
 
 -  头文件 ``paddle_inference_api.h`` 定义了所有的接口
 -  库文件\ ``libpaddle_fluid.so`` 或 ``libpaddle_fluid.a``
--  库文件 ``libpaddle_inference_api.so`` 或
-   ``libpaddle_inference_api.a``
 
 编译和依赖可以参考 :ref:`install_or_build_cpp_inference_lib` 。
 
@@ -97,8 +95,7 @@ engine
     CHECK(predictor->Run(slots, &outputs));
     // 获取 outputs ...
 
-编译时，联编 ``libpaddle_fluid.a/.so`` 和
-``libpaddle_inference_api.a/.so`` 便可。
+编译时，联编 ``libpaddle_fluid.a/.so`` 即可。
 
 详细代码参考
 ------------

paddle/fluid/API.spec

@@ -113,6 +113,7 @@ paddle.fluid.layers.beam_search_decode ArgSpec(args=['ids', 'scores', 'beam_size
 paddle.fluid.layers.conv2d_transpose ArgSpec(args=['input', 'num_filters', 'output_size', 'filter_size', 'padding', 'stride', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, None, 0, 1, 1, None, None, None, True, None, None))
 paddle.fluid.layers.conv3d_transpose ArgSpec(args=['input', 'num_filters', 'output_size', 'filter_size', 'padding', 'stride', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, None, 0, 1, 1, None, None, None, True, None, None))
 paddle.fluid.layers.sequence_expand ArgSpec(args=['x', 'y', 'ref_level', 'name'], varargs=None, keywords=None, defaults=(-1, None))
+paddle.fluid.layers.sequence_pad ArgSpec(args=['x', 'pad_value', 'maxlen'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.lstm_unit ArgSpec(args=['x_t', 'hidden_t_prev', 'cell_t_prev', 'forget_bias', 'param_attr', 'bias_attr', 'name'], varargs=None, keywords=None, defaults=(0.0, None, None, None))
 paddle.fluid.layers.reduce_sum ArgSpec(args=['input', 'dim', 'keep_dim', 'name'], varargs=None, keywords=None, defaults=(None, False, None))
 paddle.fluid.layers.reduce_mean ArgSpec(args=['input', 'dim', 'keep_dim', 'name'], varargs=None, keywords=None, defaults=(None, False, None))
@@ -146,6 +147,7 @@ paddle.fluid.layers.reshape ArgSpec(args=['x', 'shape', 'actual_shape', 'act', '
 paddle.fluid.layers.lod_reset ArgSpec(args=['x', 'y', 'target_lod'], varargs=None, keywords=None, defaults=(None, None))
 paddle.fluid.layers.lrn ArgSpec(args=['input', 'n', 'k', 'alpha', 'beta', 'name'], varargs=None, keywords=None, defaults=(5, 1.0, 0.0001, 0.75, None))
 paddle.fluid.layers.pad ArgSpec(args=['x', 'paddings', 'pad_value', 'name'], varargs=None, keywords=None, defaults=(0.0, None))
+paddle.fluid.layers.pad_constant_like ArgSpec(args=['x', 'y', 'pad_value', 'name'], varargs=None, keywords=None, defaults=(0.0, None))
 paddle.fluid.layers.label_smooth ArgSpec(args=['label', 'prior_dist', 'epsilon', 'dtype', 'name'], varargs=None, keywords=None, defaults=(None, 0.1, 'float32', None))
 paddle.fluid.layers.roi_pool ArgSpec(args=['input', 'rois', 'pooled_height', 'pooled_width', 'spatial_scale'], varargs=None, keywords=None, defaults=(1, 1, 1.0))
 paddle.fluid.layers.dice_loss ArgSpec(args=['input', 'label', 'epsilon'], varargs=None, keywords=None, defaults=(1e-05,))
@@ -162,6 +164,9 @@ paddle.fluid.layers.crop ArgSpec(args=['x', 'shape', 'offsets', 'name'], varargs
 paddle.fluid.layers.rank_loss ArgSpec(args=['label', 'left', 'right', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.prelu ArgSpec(args=['x', 'mode', 'param_attr', 'name'], varargs=None, keywords=None, defaults=(None, None))
 paddle.fluid.layers.flatten ArgSpec(args=['x', 'axis', 'name'], varargs=None, keywords=None, defaults=(1, None))
+paddle.fluid.layers.sequence_mask ArgSpec(args=['x', 'maxlen', 'dtype', 'name'], varargs=None, keywords=None, defaults=(None, 'int64', None))
+paddle.fluid.layers.stack ArgSpec(args=['x', 'axis'], varargs=None, keywords=None, defaults=(0,))
+paddle.fluid.layers.unstack ArgSpec(args=['x', 'axis', 'num'], varargs=None, keywords=None, defaults=(0, None))
 paddle.fluid.layers.data ArgSpec(args=['name', 'shape', 'append_batch_size', 'dtype', 'lod_level', 'type', 'stop_gradient'], varargs=None, keywords=None, defaults=(True, 'float32', 0, VarType.LOD_TENSOR, True))
 paddle.fluid.layers.open_recordio_file ArgSpec(args=['filename', 'shapes', 'lod_levels', 'dtypes', 'pass_num', 'for_parallel'], varargs=None, keywords=None, defaults=(1, True))
 paddle.fluid.layers.open_files ArgSpec(args=['filenames', 'shapes', 'lod_levels', 'dtypes', 'thread_num', 'buffer_size', 'pass_num', 'is_test'], varargs=None, keywords=None, defaults=(None, None, 1, None))
@@ -191,7 +196,7 @@ paddle.fluid.layers.argsort ArgSpec(args=['input', 'axis', 'name'], varargs=None
 paddle.fluid.layers.ones ArgSpec(args=['shape', 'dtype', 'force_cpu'], varargs=None, keywords=None, defaults=(False,))
 paddle.fluid.layers.zeros ArgSpec(args=['shape', 'dtype', 'force_cpu'], varargs=None, keywords=None, defaults=(False,))
 paddle.fluid.layers.reverse ArgSpec(args=['x', 'axis'], varargs=None, keywords=None, defaults=None)
-paddle.fluid.layers.While.__init__ ArgSpec(args=['self', 'cond', 'name'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.layers.While.__init__ ArgSpec(args=['self', 'cond', 'is_test', 'name'], varargs=None, keywords=None, defaults=(False, None))
 paddle.fluid.layers.While.block ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.Switch.__init__ ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.Switch.case ArgSpec(args=['self', 'condition'], varargs=None, keywords=None, defaults=None)
@@ -376,7 +381,7 @@ paddle.fluid.LoDTensor.__init__ 1. __init__(self: paddle.fluid.core.LoDTensor, a
 paddle.fluid.LoDTensor.has_valid_recursive_sequence_lengths has_valid_recursive_sequence_lengths(self: paddle.fluid.core.LoDTensor) -> bool
 paddle.fluid.LoDTensor.lod lod(self: paddle.fluid.core.LoDTensor) -> List[List[int]]
 paddle.fluid.LoDTensor.recursive_sequence_lengths recursive_sequence_lengths(self: paddle.fluid.core.LoDTensor) -> List[List[int]]
-paddle.fluid.LoDTensor.set 1. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float32], arg1: paddle::platform::CPUPlace) -> None  2. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int32], arg1: paddle::platform::CPUPlace) -> None  3. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float64], arg1: paddle::platform::CPUPlace) -> None  4. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int64], arg1: paddle::platform::CPUPlace) -> None  5. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[bool], arg1: paddle::platform::CPUPlace) -> None  6. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint16], arg1: paddle::platform::CPUPlace) -> None  7. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint8], arg1: paddle::platform::CPUPlace) -> None  8. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float32], arg1: paddle::platform::CUDAPlace) -> None  9. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int32], arg1: paddle::platform::CUDAPlace) -> None  10. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float64], arg1: paddle::platform::CUDAPlace) -> None  11. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int64], arg1: paddle::platform::CUDAPlace) -> None  12. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[bool], arg1: paddle::platform::CUDAPlace) -> None  13. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint16], arg1: paddle::platform::CUDAPlace) -> None  14. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint8], arg1: paddle::platform::CUDAPlace) -> None  15. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float32], arg1: paddle::platform::CUDAPinnedPlace) -> None  16. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int32], arg1: paddle::platform::CUDAPinnedPlace) -> None  17. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float64], arg1: paddle::platform::CUDAPinnedPlace) -> None  18. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int64], arg1: paddle::platform::CUDAPinnedPlace) -> None  19. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[bool], arg1: paddle::platform::CUDAPinnedPlace) -> None  20. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint16], arg1: paddle::platform::CUDAPinnedPlace) -> None  21. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint8], arg1: paddle::platform::CUDAPinnedPlace) -> None
+paddle.fluid.LoDTensor.set 1. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float32], arg1: paddle::platform::CPUPlace) -> None  2. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int32], arg1: paddle::platform::CPUPlace) -> None  3. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float64], arg1: paddle::platform::CPUPlace) -> None  4. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int64], arg1: paddle::platform::CPUPlace) -> None  5. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[bool], arg1: paddle::platform::CPUPlace) -> None  6. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint16], arg1: paddle::platform::CPUPlace) -> None  7. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint8], arg1: paddle::platform::CPUPlace) -> None  8. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int8], arg1: paddle::platform::CPUPlace) -> None  9. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float32], arg1: paddle::platform::CUDAPlace) -> None  10. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int32], arg1: paddle::platform::CUDAPlace) -> None  11. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float64], arg1: paddle::platform::CUDAPlace) -> None  12. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int64], arg1: paddle::platform::CUDAPlace) -> None  13. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[bool], arg1: paddle::platform::CUDAPlace) -> None  14. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint16], arg1: paddle::platform::CUDAPlace) -> None  15. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint8], arg1: paddle::platform::CUDAPlace) -> None  16. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int8], arg1: paddle::platform::CUDAPlace) -> None  17. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float32], arg1: paddle::platform::CUDAPinnedPlace) -> None  18. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int32], arg1: paddle::platform::CUDAPinnedPlace) -> None  19. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float64], arg1: paddle::platform::CUDAPinnedPlace) -> None  20. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int64], arg1: paddle::platform::CUDAPinnedPlace) -> None  21. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[bool], arg1: paddle::platform::CUDAPinnedPlace) -> None  22. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint16], arg1: paddle::platform::CUDAPinnedPlace) -> None  23. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint8], arg1: paddle::platform::CUDAPinnedPlace) -> None  24. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int8], arg1: paddle::platform::CUDAPinnedPlace) -> None
 paddle.fluid.LoDTensor.set_lod set_lod(self: paddle.fluid.core.LoDTensor, arg0: List[List[int]]) -> None
 paddle.fluid.LoDTensor.set_recursive_sequence_lengths set_recursive_sequence_lengths(self: paddle.fluid.core.LoDTensor, arg0: List[List[int]]) -> None
 paddle.fluid.LoDTensor.shape shape(self: paddle.fluid.core.Tensor) -> List[int]

paddle/fluid/CMakeLists.txt

@@ -2,9 +2,13 @@ add_subdirectory(memory)
 add_subdirectory(platform)
 add_subdirectory(framework)
 add_subdirectory(operators)
-add_subdirectory(pybind)
 add_subdirectory(string)
+
+if (NOT WIN32)
+add_subdirectory(pybind)
 add_subdirectory(recordio)
+endif(NOT WIN32)
+
 if(WITH_INFERENCE)
   # NOTE: please add subdirectory inference at last.
   add_subdirectory(inference)

paddle/fluid/framework/CMakeLists.txt

-add_subdirectory(details)
 add_subdirectory(ir)
+if (NOT WIN32)
+add_subdirectory(details)
+endif (NOT WIN32)
 # ddim lib
 proto_library(framework_proto SRCS framework.proto)
 
@@ -28,8 +30,12 @@ if(WITH_GPU)
 else()
   cc_test(mixed_vector_test SRCS mixed_vector_test.cc DEPS place memory device_context tensor)
 endif()
-
+if (NOT WIN32)
 cc_library(lod_tensor SRCS lod_tensor.cc DEPS ddim place tensor framework_proto recordio)
+else()
+cc_library(lod_tensor SRCS lod_tensor.cc DEPS ddim place tensor framework_proto)
+endif (NOT WIN32)
+
 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)
 
@@ -69,14 +75,22 @@ cc_library(op_proto_maker SRCS op_proto_maker.cc DEPS framework_proto attribute
 cc_test(op_proto_maker_test SRCS op_proto_maker_test.cc DEPS op_proto_maker)
 cc_library(op_info SRCS op_info.cc DEPS attribute framework_proto)
 cc_library(shape_inference SRCS shape_inference.cc DEPS ddim attribute device_context)
+
+if (NOT WIN32)
 cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope glog
     shape_inference data_transform lod_tensor profiler)
+else()
+cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope glog
+    shape_inference data_transform lod_tensor)
+endif(NOT WIN32)
+
 cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry device_context)
 cc_library(proto_desc SRCS var_desc.cc op_desc.cc block_desc.cc program_desc.cc DEPS shape_inference op_info operator glog)
 
 cc_library(op_registry SRCS op_registry.cc DEPS op_proto_maker op_info operator glog proto_desc)
 nv_test(op_registry_test SRCS op_registry_test.cc DEPS op_registry)
 
+if (NOT WIN32)
 py_proto_compile(framework_py_proto SRCS framework.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)
@@ -86,17 +100,18 @@ add_custom_command(TARGET framework_py_proto POST_BUILD
     COMMAND cp *.py ${PADDLE_BINARY_DIR}/python/paddle/fluid/proto/
     COMMENT "Copy generated python proto into directory paddle/fluid/proto."
     WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
+endif(NOT WIN32)
 
 cc_library(lod_rank_table SRCS lod_rank_table.cc DEPS lod_tensor)
 
 cc_library(feed_fetch_method SRCS feed_fetch_method.cc DEPS lod_tensor scope glog)
 
 if(WITH_DISTRIBUTE)
-  cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method sendrecvop_grpc cares grpc++_unsecure grpc_unsecure gpr)
+  cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method sendrecvop_grpc cares grpc++_unsecure grpc_unsecure gpr graph_to_program_pass)
   set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
   set_source_files_properties(executor.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
 else()
-  cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method)
+  cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass)
 endif()
 
 if (NOT WIN32)
@@ -120,7 +135,9 @@ cc_test(cow_ptr_tests SRCS details/cow_ptr_test.cc)
 # cc_test(channel_test SRCS channel_test.cc)
 cc_test(tuple_test SRCS tuple_test.cc )
 
+if (NOT WIN32)
 cc_test(rw_lock_test SRCS rw_lock_test.cc)
+endif (NOT WIN32)
 
 # disable test temporarily.
 # TODO https://github.com/PaddlePaddle/Paddle/issues/11971

paddle/fluid/framework/array.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 <cstdint>
+#include "paddle/fluid/platform/hostdevice.h"
+
+namespace paddle {
+namespace framework {
+template <typename T, size_t N>
+class Array {
+  static_assert(N > 0, "The size of array must be larger than 0");
+
+ public:
+  HOSTDEVICE Array() {}
+
+  HOSTDEVICE explicit Array(const T &val) {
+    for (size_t i = 0; i < N; ++i) data_[i] = val;
+  }
+
+  HOSTDEVICE const T *Get() const { return data_; }
+
+  HOSTDEVICE T *GetMutable() { return data_; }
+
+  HOSTDEVICE T &operator[](size_t index) { return data_[index]; }
+
+  HOSTDEVICE const T &operator[](size_t index) const { return data_[index]; }
+
+  HOSTDEVICE constexpr size_t size() const { return N; }
+
+ private:
+  T data_[N];
+};
+
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/data_type.cc

@@ -64,6 +64,7 @@ static DataTypeMap* InitDataTypeMap() {
   RegType(size_t, proto::VarType::SIZE_T);
   RegType(int16_t, proto::VarType::INT16);
   RegType(uint8_t, proto::VarType::UINT8);
+  RegType(int8_t, proto::VarType::INT8);
 
 #undef RegType
   return retv;

paddle/fluid/framework/data_type.h

@@ -26,6 +26,7 @@ namespace framework {
 extern proto::VarType::Type ToDataType(std::type_index type);
 extern std::type_index ToTypeIndex(proto::VarType::Type type);
 
+#if !defined(_WIN32)
 template <typename Visitor>
 inline void VisitDataType(proto::VarType::Type type, Visitor visitor) {
   switch (type) {
@@ -53,10 +54,47 @@ inline void VisitDataType(proto::VarType::Type type, Visitor visitor) {
     case proto::VarType::INT16:
       visitor.template operator()<int16_t>();
       break;
+    case proto::VarType::INT8:
+      visitor.template operator()<int8_t>();
+      break;
+    default:
+      PADDLE_THROW("Not supported %d", type);
+  }
+}
+#else
+// the msvc compiler do not implement two-stage name lookup correctly.
+template <typename Visitor>
+inline void VisitDataType(proto::VarType::Type type, Visitor visitor) {
+  switch (type) {
+    case proto::VarType::FP16:
+      visitor.operator()<platform::float16>();
+      break;
+    case proto::VarType::FP32:
+      visitor.operator()<float>();
+      break;
+    case proto::VarType::FP64:
+      visitor.operator()<double>();
+      break;
+    case proto::VarType::INT32:
+      visitor.operator()<int>();
+      break;
+    case proto::VarType::INT64:
+      visitor.operator()<int64_t>();
+      break;
+    case proto::VarType::BOOL:
+      visitor.operator()<bool>();
+      break;
+    case proto::VarType::UINT8:
+      visitor.operator()<uint8_t>();
+      break;
+    case proto::VarType::INT16:
+      visitor.operator()<int16_t>();
+      break;
     default:
       PADDLE_THROW("Not supported %d", type);
   }
 }
+#endif  // _WIN32
 
 extern std::string DataTypeToString(const proto::VarType::Type type);
 extern size_t SizeOfType(std::type_index type);

paddle/fluid/framework/details/multi_devices_graph_pass.cc

@@ -754,9 +754,20 @@ void MultiDevSSAGraphBuilder::CreateDistTrainOp(ir::Graph *result,
                  node->Op()->Type());
 
   CreateComputationalOp(result, node, op_dev_id);
-  if (node->Op()->Type() == "concat") {
-    ConnectOp(result, result->Get<GraphOps>(kGraphOps).back().get(),
-              "fetch_barrier");
+}
+
+void SetOpInputsAllPlaces(ir::Graph *result, ir::Node *node, int num_places) {
+  auto *op_handle = result->Get<GraphOps>(kGraphOps).back().get();
+  for (ir::Node *input : node->inputs) {
+    VarHandle *var = nullptr;
+    for (int place_offset = 0; place_offset < num_places; ++place_offset) {
+      auto &var_holders = result->Get<GraphVars>(kGraphVars)[place_offset];
+      auto &var_holder = var_holders[input->Name()];
+      if (!var_holder.empty()) {
+        var = var_holder.rbegin()->get();
+        op_handle->AddInput(var);
+      }
+    }
   }
 }
 
@@ -771,59 +782,83 @@ void MultiDevSSAGraphBuilder::CreateRPCOp(ir::Graph *result,
                    "This hack no longer holds, please fix.");
     // the variable name which contains .block means it was splited by
     // split_byref op
-    // so that we can balance the variable blocks to all the pserver
-    // instances.
     if (strategy_.reduce_ == BuildStrategy::ReduceStrategy::kAllReduce &&
         node->inputs[0]->Name().find(".block") == std::string::npos) {
       std::vector<std::string> input_var_names;
       for (ir::Node *n : node->inputs) {
         input_var_names.push_back(n->Name());
       }
-      op_dev_id = GetAppropriateDeviceID(input_var_names);
+      auto send_param_grad = boost::get<std::vector<std::string>>(
+          node->Op()->GetAttr(OpProtoAndCheckerMaker::OpRoleVarAttrName()));
+      PADDLE_ENFORCE_EQ(send_param_grad.size(), 2U);
+      op_dev_id = GetAppropriateDeviceID({send_param_grad[1]});
+      VLOG(10) << "send grad " << input_var_names[0] << " origin "
+               << send_param_grad[1] << " place: " << op_dev_id;
       for (auto &varname : input_var_names) {
         result->Get<ShardedVarDevice>(kShardedVarDevice)
             .emplace(varname, op_dev_id);
       }
+      result->Get<ShardedVarDevice>(kShardedVarDevice)
+          .emplace(send_param_grad[1], op_dev_id);
     }
   } else if (node->Op()->Type() == "recv") {
     std::vector<std::string> output_var_names;
     for (ir::Node *n : node->outputs) {
       output_var_names.push_back(n->Name());
     }
-    op_dev_id = GetAppropriateDeviceID(output_var_names);
+    auto recv_param_grad = boost::get<std::vector<std::string>>(
+        node->Op()->GetAttr(OpProtoAndCheckerMaker::OpRoleVarAttrName()));
+    if (recv_param_grad.size() == 2U) {
+      op_dev_id = GetVarDeviceID(*result, recv_param_grad[1]);
+      VLOG(10) << "recv param " << recv_param_grad[0]
+               << " get grad place: " << recv_param_grad[1]
+               << " place: " << op_dev_id;
+    } else {
+      op_dev_id = GetAppropriateDeviceID(output_var_names);
+    }
     for (auto &varname : output_var_names) {
       result->Get<ShardedVarDevice>(kShardedVarDevice)
           .emplace(varname, op_dev_id);
     }
   } else {
-    // send_barrier and fetch_barrier op can be scheduled on device 0
+    // send_barrier, fetch_barrier will run on place 0;
     op_dev_id = 0;
   }
 
   PADDLE_ENFORCE(op_dev_id != -1, "can not find the right place for rpc op: %s",
                  node->Op()->Type());
-
   result->Get<GraphOps>(kGraphOps).emplace_back(new RPCOpHandle(
       result->CreateOpNode(node->Op()), *node->Op(), local_scopes_[op_dev_id],
       node->Op()->Type(), places_[op_dev_id]));
 
-  // TODO(panyx0718): This might not be needed anymore.
-  if (node->Op()->Type() == "send_barrier") {
-    ConnectOp(result, result->Get<GraphOps>(kGraphOps).back().get(), "send");
-  } else if (node->Op()->Type() == "recv") {
-    ConnectOp(result, result->Get<GraphOps>(kGraphOps).back().get(),
-              "send_barrier");
-  } else if (node->Op()->Type() == "fetch_barrier") {
-    ConnectOp(result, result->Get<GraphOps>(kGraphOps).back().get(), "recv");
-  } else if (node->Op()->Type() == "send") {
-    // do nothing
+  if (node->Op()->Type() == "send") {
+    CreateOpHandleIOs(result, node, op_dev_id);
   } else {
-    PADDLE_THROW(
-        "rpc op should be in ["
-        "send, send_barrier. recv, fetch_barrier]");
-  }
+    // send_barrier, recv, fetch_barrier's inputs are deps var, get them from
+    // all places
+    auto p = places_[op_dev_id];
+    auto *op_handle = result->Get<GraphOps>(kGraphOps).back().get();
+    op_handle->SetDeviceContext(p,
+                                platform::DeviceContextPool::Instance().Get(p));
 
-  CreateOpHandleIOs(result, node, op_dev_id);
+    SetOpInputsAllPlaces(result, node, places_.size());
+    for (ir::Node *output : node->outputs) {
+      int outvar_dev_id = op_dev_id;
+      if (node->Op()->Type() == "fetch_barrier") {
+        outvar_dev_id = GetVarDeviceID(*result, output->Name());
+        PADDLE_ENFORCE_NE(outvar_dev_id, -1);
+      }
+      p = places_[outvar_dev_id];
+      ir::Node *new_node = nullptr;
+      if (output->Var()) {
+        new_node = result->CreateVarNode(output->Var());
+      } else {
+        new_node =
+            result->CreateEmptyNode(output->Name(), ir::Node::Type::kVariable);
+      }
+      CreateOpOutput(result, op_handle, new_node, p, outvar_dev_id);
+    }
+  }
 }
 
 bool MultiDevSSAGraphBuilder::IsScaleLossOp(ir::Node *node) const {

paddle/fluid/framework/details/multi_devices_graph_print_pass.cc

@@ -54,7 +54,8 @@ void GraphvizSSAGraphPrinter::Print(const ir::Graph &graph,
       sout << "var_" << cur_var_id << " [label=\"" << var_handle_ptr->name_
            << "\\n"
            << var_handle_ptr->place_ << "\\n"
-           << var_handle_ptr->version_ << "\"]" << std::endl;
+           << "scope: " << var_handle_ptr->scope_idx_ << "\\n"
+           << "v" << var_handle_ptr->version_ << "\"]" << std::endl;
     } else if (dummy_ptr) {
       sout << "var_" << cur_var_id << " [label=\"dummy\"]" << std::endl;
     }

paddle/fluid/framework/framework.proto

@@ -107,6 +107,7 @@ message VarType {
     // Tensor<size_t> is used in C++.
     SIZE_T = 19;
     UINT8 = 20;
+    INT8 = 21;
 
     // Other types that may need additional descriptions
     LOD_TENSOR = 7;

paddle/fluid/framework/ir/CMakeLists.txt

@@ -3,6 +3,7 @@ cc_library(graph SRCS graph.cc DEPS node)
 cc_library(graph_helper SRCS graph_helper.cc DEPS graph)
 cc_library(pass SRCS pass.cc DEPS graph node graph_helper)
 cc_library(graph_viz_pass SRCS graph_viz_pass.cc DEPS graph pass graph_helper)
+cc_library(graph_to_program_pass SRCS graph_to_program_pass.cc DEPS graph pass graph_helper)
 cc_library(graph_traits SRCS graph_traits.cc DEPS graph)
 cc_library(graph_pattern_detecter SRCS graph_pattern_detecter.cc DEPS graph graph_helper graph_traits)
 cc_library(fc_fuse_pass SRCS fc_fuse_pass.cc DEPS graph graph_pattern_detecter)
@@ -12,5 +13,6 @@ cc_library(infer_clean_graph_pass SRCS infer_clean_graph_pass.cc DEPS graph pass
 cc_test(pass_test SRCS pass_test.cc DEPS graph pass graph_helper)
 cc_test(graph_test SRCS graph_test.cc DEPS graph graph_helper op_registry)
 cc_test(graph_helper_test SRCS graph_helper_test.cc DEPS graph graph_helper op_registry)
+cc_test(graph_to_program_pass_test SRCS graph_to_program_pass_test.cc DEPS graph_to_program_pass)
 cc_test(test_graph_pattern_detecter SRCS graph_pattern_detecter_tester.cc DEPS graph_pattern_detecter)
 cc_test(test_fc_fuse_pass SRCS fc_fuse_pass_tester.cc DEPS fc_fuse_pass graph_pattern_detecter graph pass graph_traits framework_proto)

paddle/fluid/framework/ir/graph.cc

@@ -132,63 +132,6 @@ Graph::Graph(const ProgramDesc &program) : program_(program) {
     }
   }
 
-  std::vector<ir::Node *> send_ops;
-  ir::Node *send_bar = nullptr;
-  std::vector<ir::Node *> recv_ops;
-  ir::Node *fetch_bar = nullptr;
-  for (ir::Node *node : Nodes()) {
-    if (node->Name() == "send") {
-      send_ops.push_back(node);
-    } else if (node->Name() == "send_barrier") {
-      PADDLE_ENFORCE(!send_bar, "only has one send barrier");
-      send_bar = node;
-    } else if (node->Name() == "recv") {
-      recv_ops.push_back(node);
-    } else if (node->Name() == "fetch_barrier") {
-      PADDLE_ENFORCE(!fetch_bar, "only has one fetch barrier");
-      fetch_bar = node;
-    }
-  }
-  if (send_bar) {
-    for (ir::Node *send : send_ops) {
-      ir::Node *dep_var = CreateControlDepVar();
-      send->outputs.push_back(dep_var);
-      dep_var->inputs.push_back(send);
-      send_bar->inputs.push_back(dep_var);
-      dep_var->outputs.push_back(send_bar);
-    }
-    for (ir::Node *recv : recv_ops) {
-      ir::Node *dep_var = CreateControlDepVar();
-      recv->inputs.push_back(dep_var);
-      dep_var->outputs.push_back(recv);
-      send_bar->outputs.push_back(dep_var);
-      dep_var->inputs.push_back(send_bar);
-    }
-  }
-  if (fetch_bar) {
-    for (ir::Node *recv : recv_ops) {
-      ir::Node *dep_var = CreateControlDepVar();
-      recv->outputs.push_back(dep_var);
-      dep_var->inputs.push_back(recv);
-      fetch_bar->inputs.push_back(dep_var);
-      dep_var->outputs.push_back(fetch_bar);
-    }
-  }
-
-  std::vector<std::string> send_vars = FindDistTrainSendVars(send_ops);
-  std::vector<std::string> recv_vars = FindDistTrainRecvVars(recv_ops);
-  for (ir::Node *node : Nodes()) {
-    if (IsDistTrainOp(node, send_vars, recv_vars)) {
-      if (fetch_bar && node->Name() == "concat") {
-        ir::Node *dep_var = CreateControlDepVar();
-        fetch_bar->outputs.push_back(dep_var);
-        dep_var->inputs.push_back(fetch_bar);
-        node->inputs.push_back(dep_var);
-        dep_var->outputs.push_back(node);
-      }
-    }
-  }
-
   /**
    * We should handle write after read(WAR) and write after write(WAW) here.
    * Because some of the operators of the program can be executed parallelly.

paddle/fluid/framework/ir/graph.h

@@ -142,8 +142,6 @@ class Graph {
     nodes_.erase(node);
   }
 
-  const ProgramDesc &program() const { return program_; }
-
  private:
   // This method takes ownership of `node`.
   ir::Node *AddNode(ir::Node *node) {
@@ -154,7 +152,7 @@ class Graph {
   }
 
   // NOTE: program_ shouldn't be exposed to user.
-  const ProgramDesc &program_;
+  const ProgramDesc program_;
   std::map<std::string, boost::any> attrs_;
   std::map<std::string, std::function<void(void)>> attr_dels_;
   std::map<ir::Node *, std::unique_ptr<ir::Node>> nodes_;

paddle/fluid/framework/ir/graph_pattern_detecter_tester.cc

@@ -163,8 +163,8 @@ TEST(GraphPatternDetecter, MultiSubgraph) {
   // 3. Detect op2 -> var2 -> op4
   // 4. Detect op2 -> var3 -> op5
   // But 2 and 3 and 4 overlapped, so keep 2, so the final choices are 1 and 2
-  ASSERT_GE(count, 1UL);
-  ASSERT_LE(count, 2UL);
+  ASSERT_GE(count, 1);
+  ASSERT_LE(count, 2);
 }
 
 }  // namespace ir

paddle/fluid/framework/ir/graph_to_program_pass.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/ir/graph_to_program_pass.h"
+
+#include <map>
+#include <string>
+#include <vector>
+
+#include "paddle/fluid/framework/ir/graph.h"
+#include "paddle/fluid/framework/ir/graph_helper.h"
+
+#include "paddle/fluid/framework/program_desc.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+std::unique_ptr<Graph> GraphToProgramPass::ApplyImpl(
+    std::unique_ptr<Graph> graph) const {
+  ProgramDesc& program = Get<ProgramDesc>("program");
+
+  std::unique_ptr<proto::ProgramDesc> program_pb(
+      new proto::ProgramDesc(*program.Proto()));
+
+  auto block = program_pb->mutable_blocks(kRootBlockIndex);
+  block->clear_vars();
+  std::unordered_set<std::string> visited_vars;
+  for (ir::Node* n : graph->Nodes()) {
+    if (n->NodeType() == ir::Node::Type::kVariable) {
+      if (n->Var() && visited_vars.count(n->Var()->Name()) == 0) {
+        visited_vars.insert(n->Var()->Name());
+        block->add_vars()->MergeFrom(*n->Var()->Proto());
+      }
+    }
+  }
+
+  block->clear_ops();
+  std::vector<ir::Node*> nodes = TopologySortOperations(*graph);
+  for (ir::Node* n : nodes) {
+    if (!n->Op()) {
+      continue;
+    }
+    block->add_ops()->MergeFrom(*n->Op()->Proto());
+  }
+
+  program.CopyFrom(*program_pb);
+  return graph;
+}
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
+REGISTER_PASS(graph_to_program_pass, paddle::framework::ir::GraphToProgramPass);

paddle/fluid/framework/ir/graph_to_program_pass.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 "paddle/fluid/framework/ir/pass.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+class GraphToProgramPass : public Pass {
+ protected:
+  std::unique_ptr<Graph> ApplyImpl(std::unique_ptr<Graph> graph) const override;
+};
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/ir/graph_to_program_pass_test.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/ir/graph_to_program_pass.h"
+
+#include <string>
+#include <vector>
+#include "gtest/gtest.h"
+#include "paddle/fluid/framework/program_desc.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+void BuildNoCircleGraph(Graph* g) {
+  OpDesc op1;
+  op1.SetType("op1");
+  OpDesc op2;
+  op2.SetType("op2");
+  OpDesc op3;
+  op3.SetType("op3");
+  OpDesc op4;
+  op4.SetType("op4");
+  OpDesc op5;
+  op5.SetType("op5");
+  VarDesc var1("var1");
+  VarDesc var2("var2");
+  VarDesc var3("var3");
+  VarDesc var4("var4");
+
+  ir::Node* o1 = g->CreateOpNode(&op1);
+  ir::Node* o2 = g->CreateOpNode(&op2);
+  ir::Node* o3 = g->CreateOpNode(&op3);
+  ir::Node* o4 = g->CreateOpNode(&op4);
+  ir::Node* o5 = g->CreateOpNode(&op5);
+  ir::Node* v1 = g->CreateVarNode(&var1);
+  ir::Node* v2 = g->CreateVarNode(&var2);
+  ir::Node* v3 = g->CreateVarNode(&var3);
+  ir::Node* v4 = g->CreateVarNode(&var4);
+
+  // o1->v1->o2
+  o1->outputs.push_back(v1);
+  o2->inputs.push_back(v1);
+  v1->inputs.push_back(o1);
+  v1->outputs.push_back(o2);
+  // o2->v2->o3
+  // o2->v2->o4
+  o2->outputs.push_back(v2);
+  o3->inputs.push_back(v2);
+  o4->inputs.push_back(v2);
+  v2->outputs.push_back(o3);
+  v2->outputs.push_back(o4);
+  v2->inputs.push_back(o2);
+  // o2->v3->o5
+  o2->outputs.push_back(v3);
+  o5->inputs.push_back(v3);
+  v3->inputs.push_back(o2);
+  v3->outputs.push_back(o5);
+  // o3-v4->o5
+  o3->outputs.push_back(v4);
+  o5->inputs.push_back(v4);
+  v4->inputs.push_back(o3);
+  v4->outputs.push_back(o5);
+}
+
+TEST(GraphToProgramPass, Basic) {
+  ProgramDesc prog;
+  std::unique_ptr<Graph> g(new Graph(prog));
+  BuildNoCircleGraph(g.get());
+
+  auto pass = paddle::framework::ir::PassRegistry::Instance().Get(
+      "graph_to_program_pass");
+
+  ProgramDesc compiled_prog;
+  pass->SetNotOwned<paddle::framework::ProgramDesc>("program", &compiled_prog);
+  pass->Apply(std::move(g));
+  std::vector<OpDesc*> ops = compiled_prog.Block(0).AllOps();
+  EXPECT_EQ(ops[0]->Type(), "op1");
+  EXPECT_EQ(ops[1]->Type(), "op2");
+  if (ops[2]->Type() == "op3") {
+    EXPECT_EQ(ops[3]->Type(), "op4");
+  } else if (ops[2]->Type() == "op4") {
+    EXPECT_EQ(ops[3]->Type(), "op3");
+  }
+  EXPECT_EQ(ops[4]->Type(), "op5");
+
+  std::unordered_set<std::string> vars;
+  for (VarDesc* v : compiled_prog.Block(0).AllVars()) {
+    vars.insert(v->Name());
+  }
+  EXPECT_TRUE(vars.find("var1") != vars.end());
+  EXPECT_TRUE(vars.find("var2") != vars.end());
+  EXPECT_TRUE(vars.find("var3") != vars.end());
+}
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
+USE_PASS(graph_to_program_pass);

paddle/fluid/framework/ir/node.cc

@@ -17,7 +17,7 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 namespace ir {
-const char Node::kControlDepVarName[] = "__control_var";
+constexpr char Node::kControlDepVarName[];
 }  // namespace ir
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/ir/node.h

@@ -27,7 +27,7 @@ namespace ir {
 class Node {
  public:
   enum class Type { kOperation, kVariable };
-  static const char kControlDepVarName[];
+  static constexpr char kControlDepVarName[] = "__control_var";
 
   explicit Node(const std::string& name, Type type)
       : name_(name), var_desc_(nullptr), op_desc_(nullptr), type_(type) {}
@@ -41,8 +41,7 @@ class Node {
   explicit Node(OpDesc* op_desc)
       : name_(op_desc->Type()),
         var_desc_(nullptr),
-        op_desc_(new OpDesc(*op_desc)),  // TODO(panyx0718) the pointer in the
-                                         // original OpDesc might go out.
+        op_desc_(new OpDesc(*op_desc, op_desc->Block())),
         type_(Type::kOperation) {}
 
   Type NodeType() const { return type_; }

paddle/fluid/framework/lod_tensor.cc

@@ -25,8 +25,10 @@ limitations under the License. */
 #include "paddle/fluid/memory/memcpy.h"
 #include "paddle/fluid/memory/memory.h"
 
+#if !defined(_WIN32)
 #include "paddle/fluid/recordio/scanner.h"
 #include "paddle/fluid/recordio/writer.h"
+#endif  // _WIN32
 
 namespace paddle {
 namespace framework {
@@ -300,6 +302,7 @@ void DeserializeFromStream(std::istream &is, LoDTensor *tensor,
   TensorFromStream(is, static_cast<Tensor *>(tensor), dev_ctx);
 }
 
+#if !defined(_WIN32)
 void WriteToRecordIO(recordio::Writer *writer,
                      const std::vector<LoDTensor> &tensor,
                      const platform::DeviceContext &dev_ctx) {
@@ -329,7 +332,19 @@ bool ReadFromRecordIO(recordio::Scanner *scanner,
 
   return true;
 }
-
+#else
+class Writer {};
+class Scanner {};
+void WriteToRecordIO(recordio::Writer *writer,
+                     const std::vector<LoDTensor> &tensor,
+                     const platform::DeviceContext &dev_ctx) {}
+bool ReadFromRecordIO(recordio::Scanner *scanner,
+                      const platform::DeviceContext &dev_ctx,
+                      std::vector<LoDTensor> *result_ptr) {
+  PADDLE_ENFORCE("windows didn't supported recordio!.");
+  return true;
+}
+#endif  // _WIN32
 std::vector<LoDTensor> LoDTensor::SplitLoDTensor(
     const std::vector<platform::Place> places) const {
   check_memory_size();

paddle/fluid/framework/lod_tensor_test.cc

@@ -274,6 +274,7 @@ TEST(LoD, ConvertToOffsetBasedLoD) {
   EXPECT_EQ(offset_lod, expected);
 }
 
+#if !defined(_WIN32)
 template <typename T>
 static void TestRecordIO() {
   LoDTensor tensor;
@@ -320,6 +321,7 @@ TEST(LoDTensor, RecordIO) {
   TestRecordIO<float>();
   TestRecordIO<double>();
 }
+#endif  // !defined(_WIN32)
 
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/operator.cc

@@ -74,6 +74,12 @@ static DDim GetDims(const Scope& scope, const std::string& name,
   }
 }
 
+static bool VarInited(const Scope& scope, const std::string& name) {
+  Variable* var = scope.FindVar(name);
+  if (var == nullptr) return false;
+  return var->IsInitialized();
+}
+
 static std::string GetDtype(const Scope& scope, const std::string& name) {
   Variable* var = scope.FindVar(name);
   if (var == nullptr) {
@@ -87,8 +93,12 @@ static std::string GetDtype(const Scope& scope, const std::string& name) {
     }
     return DataTypeToString(ToDataType(tensor.type()));
   } else if (var->IsType<SelectedRows>()) {
-    return DataTypeToString(
-        ToDataType(var->Get<SelectedRows>().value().type()));
+    auto tensor = var->Get<SelectedRows>().value();
+    if (UNLIKELY(!tensor.IsInitialized())) {
+      return "uninited";
+    } else {
+      return DataTypeToString(ToDataType(tensor.type()));
+    }
   } else {
     return "";
   }
@@ -197,16 +207,21 @@ std::string OperatorBase::DebugStringEx(const Scope* scope) const {
     auto& input = *it;
     ss << input.first << "[";
     for (size_t i = 0; i < input.second.size(); ++i) {
-      ss << input.second[i];
+      auto var_name = input.second[i];
+      ss << var_name;
       if (scope) {
-        int row_size = GetRowSize(*scope, input.second[i]);
-        if (row_size >= 0) {
-          ss << "[row_size=" << row_size << "]";
+        if (!VarInited(*scope, var_name)) {
+          ss << "[uninited]";
+        } else {
+          int row_size = GetRowSize(*scope, var_name);
+          if (row_size >= 0) {
+            ss << "[row_size=" << row_size << "]";
+          }
+          std::string dtype = GetDtype(*scope, var_name);
+          ss << ":" << dtype;
+          ss << "[" << GetDims(*scope, var_name, true) << "]";
+          ss << "(" << GetLoD(*scope, var_name) << ")";
         }
-        std::string dtype = GetDtype(*scope, input.second[i]);
-        ss << ":" << dtype;
-        ss << "[" << GetDims(*scope, input.second[i], true) << "]";
-        ss << "(" << GetLoD(*scope, input.second[i]) << ")";
       }
       if (i != input.second.size() - 1) {
         ss << ", ";
@@ -223,14 +238,19 @@ std::string OperatorBase::DebugStringEx(const Scope* scope) const {
     auto& output = *it;
     ss << output.first << "[";
     for (size_t i = 0; i < output.second.size(); ++i) {
-      ss << output.second[i];
+      auto var_name = output.second[i];
+      ss << var_name;
       if (scope) {
-        int row_size = GetRowSize(*scope, output.second[i]);
-        if (row_size >= 0) {
-          ss << "[row_size=" << row_size << "]";
+        if (!VarInited(*scope, var_name)) {
+          ss << "[uninited]";
+        } else {
+          int row_size = GetRowSize(*scope, output.second[i]);
+          if (row_size >= 0) {
+            ss << "[row_size=" << row_size << "]";
+          }
+          ss << "[" << GetDims(*scope, var_name, true) << "]";
+          ss << "(" << GetLoD(*scope, var_name) << ")";
         }
-        ss << "[" << GetDims(*scope, output.second[i], true) << "]";
-        ss << "(" << GetLoD(*scope, output.second[i]) << ")";
       }
       if (i != output.second.size() - 1) {
         ss << ", ";

paddle/fluid/framework/program_desc.cc

@@ -80,6 +80,12 @@ ProgramDesc::ProgramDesc(const proto::ProgramDesc &desc) {
   InitFromProto();
 }
 
+void ProgramDesc::CopyFrom(const proto::ProgramDesc &desc) {
+  blocks_.clear();
+  desc_ = desc;
+  InitFromProto();
+}
+
 ProgramDesc::ProgramDesc(const std::string &binary_str) {
   PADDLE_ENFORCE(desc_.ParseFromString(binary_str),
                  "Fail to parse program_desc from binary string.");
@@ -111,10 +117,16 @@ void ProgramDesc::InitFromProto() {
 
 const std::vector<std::string> ProgramDesc::GetFeedTargetNames() {
   auto &global_block = Block(0);
+  // The order of feed_target_names must follow the index specified in `col`.
+  // since feed operator's order doesn't necessary follow 'col'.
   std::vector<std::string> feed_target_names;
   for (auto *op : global_block.AllOps()) {
     if (op->Type() == kFeedOpType) {
-      feed_target_names.insert(feed_target_names.begin(), op->Output("Out")[0]);
+      int col = boost::get<int>(op->GetAttr("col"));
+      if (col >= feed_target_names.size()) {
+        feed_target_names.resize(col + 1);
+      }
+      feed_target_names[col] = op->Output("Out")[0];
     }
   }
   return feed_target_names;
@@ -122,10 +134,16 @@ const std::vector<std::string> ProgramDesc::GetFeedTargetNames() {
 
 const std::vector<std::string> ProgramDesc::GetFetchTargetNames() {
   auto &global_block = Block(0);
+  // The order of fetch_target_names must follow the index specified in `col`.
+  // since fetch operator's order doesn't necessary follow 'col'.
   std::vector<std::string> fetch_target_names;
   for (auto *op : global_block.AllOps()) {
     if (op->Type() == kFetchOpType) {
-      fetch_target_names.push_back(op->Input("X")[0]);
+      int col = boost::get<int>(op->GetAttr("col"));
+      if (col >= fetch_target_names.size()) {
+        fetch_target_names.resize(col + 1);
+      }
+      fetch_target_names[col] = op->Input("X")[0];
     }
   }
   return fetch_target_names;

paddle/fluid/framework/program_desc.h

@@ -53,6 +53,8 @@ class ProgramDesc {
 
   void Flush();
 
+  void CopyFrom(const proto::ProgramDesc &desc);
+
   proto::ProgramDesc *Proto();
 
   // The output variable of feed_op is referenced as feed_target.

paddle/fluid/framework/rw_lock.h

@@ -14,13 +14,16 @@ limitations under the License. */
 
 #pragma once
 
+#if !defined(_WIN32)
 #include <pthread.h>
+#endif  // !_WIN32
 
 #include "paddle/fluid/platform/enforce.h"
 
 namespace paddle {
 namespace framework {
 
+#if !defined(_WIN32)
 struct RWLock {
   RWLock() { pthread_rwlock_init(&lock_, nullptr); }
 
@@ -43,6 +46,15 @@ struct RWLock {
  private:
   pthread_rwlock_t lock_;
 };
+#else
+// https://stackoverflow.com/questions/7125250/making-pthread-rwlock-wrlock-recursive
+// In windows, rw_lock seems like a hack. Use empty object and do nothing.
+struct RWLock {
+  void RDLock() {}
+  void WRLock() {}
+  void UNLock() {}
+};
+#endif
 
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/selected_rows.cc

@@ -139,7 +139,7 @@ int64_t SelectedRows::AutoGrownIndex(int64_t key, bool auto_grown) {
     }
     auto write_iter = id_to_index_.find(key);
     if (write_iter == id_to_index_.end()) {
-      size_t row_num = rows_.size();
+      int row_num = rows_.size();
       if (row_num == value_->dims()[0]) {
         rwlock_->UNLock();
         PADDLE_THROW("selected rows is full, then length exceed %d", row_num);
@@ -182,7 +182,7 @@ void SelectedRows::Get(const framework::Tensor& ids, framework::Tensor* value,
     PADDLE_ENFORCE_EQ(value_width, value->numel() / value->dims()[0],
                       "output tensor should have the same shape with table "
                       "except the dims[0].");
-    for (size_t i = 0; i < ids.numel(); ++i) {
+    for (int i = 0; i < ids.numel(); ++i) {
       int64_t index = AutoGrownIndex(ids.data<int64_t>()[i], auto_grown);
       framework::VisitDataType(
           framework::ToDataType(value_->type()),

paddle/fluid/framework/tensor.cc

@@ -31,7 +31,8 @@ size_t Tensor::memory_size() const {
   return holder_ == nullptr ? 0UL : holder_->size() - offset_;
 }
 
-void* Tensor::mutable_data(platform::Place place, std::type_index type) {
+void* Tensor::mutable_data(platform::Place place, std::type_index type,
+                           size_t requested_size) {
   if (holder_ != nullptr) {
     holder_->set_type(type);
   }
@@ -39,7 +40,11 @@ void* Tensor::mutable_data(platform::Place place, std::type_index type) {
                     "When calling this method, the Tensor's numel must be "
                     "equal or larger than zero. "
                     "Please check Tensor::Resize has been called first.");
-  int64_t size = numel() * SizeOfType(type);
+  size_t size = numel() * SizeOfType(type);
+  if (requested_size) {
+    PADDLE_ENFORCE_GE(requested_size, size);
+    size = requested_size;
+  }
   /* some versions of boost::variant don't have operator!= */
   if (holder_ == nullptr || !(holder_->place() == place) ||
       holder_->size() < size + offset_) {
@@ -68,10 +73,10 @@ void* Tensor::mutable_data(platform::Place place, std::type_index type) {
                                  offset_);
 }
 
-void* Tensor::mutable_data(platform::Place place) {
+void* Tensor::mutable_data(platform::Place place, size_t requested_size) {
   PADDLE_ENFORCE(this->holder_ != nullptr,
                  "Cannot invoke mutable data if current hold nothing.");
-  return mutable_data(place, holder_->type());
+  return mutable_data(place, holder_->type(), requested_size);
 }
 
 Tensor& Tensor::ShareDataWith(const Tensor& src) {

paddle/fluid/framework/tensor.h

@@ -89,22 +89,24 @@ class Tensor {
    * @note    If not exist, then allocation.
    */
   template <typename T>
-  T* mutable_data(platform::Place place);
+  T* mutable_data(platform::Place place, size_t requested_size = 0);
 
-  void* mutable_data(platform::Place place, std::type_index type);
+  void* mutable_data(platform::Place place, std::type_index type,
+                     size_t requested_size = 0);
 
-  void* mutable_data(platform::Place place);
+  void* mutable_data(platform::Place place, size_t requested_size = 0);
 
   /**
    * @brief     Return a pointer to mutable memory block.
    *
-   * @param[in] dims    The dimensions of the memory block.
-   * @param[in] place   The place of the memory block.
+   * @param[in] dims           The dimensions of the memory block.
+   * @param[in] place          The place of the memory block.
+   * @param[in] requested_size The size of the block in bytes.
    *
    * @note      If not exist, then allocation.
    */
   template <typename T>
-  T* mutable_data(DDim dims, platform::Place place);
+  T* mutable_data(DDim dims, platform::Place place, size_t requested_size = 0);
 
   /*! Return the dimensions of the memory block. */
   const DDim& dims() const;

paddle/fluid/framework/tensor_impl.h

@@ -46,16 +46,17 @@ inline T* Tensor::data() {
 }
 
 template <typename T>
-inline T* Tensor::mutable_data(DDim dims, platform::Place place) {
+inline T* Tensor::mutable_data(DDim dims, platform::Place place,
+                               size_t requested_size) {
   static_assert(std::is_pod<T>::value, "T must be POD");
   Resize(dims);
-  return mutable_data<T>(place);
+  return mutable_data<T>(place, requested_size);
 }
 
 template <typename T>
-inline T* Tensor::mutable_data(platform::Place place) {
+inline T* Tensor::mutable_data(platform::Place place, size_t requested_size) {
   static_assert(std::is_pod<T>::value, "T must be POD");
-  return reinterpret_cast<T*>(mutable_data(place, typeid(T)));
+  return reinterpret_cast<T*>(mutable_data(place, typeid(T), requested_size));
 }
 
 inline Tensor ReshapeToMatrix(const Tensor& src, int num_col_dims) {

paddle/fluid/framework/var_type.h

@@ -26,7 +26,7 @@ namespace paddle {
 namespace framework {
 
 template <typename T>
-bool IsType(const std::type_index& type_index) {
+inline bool IsType(const std::type_index& type_index) {
   return type_index == std::type_index(typeid(T));
 }
 

paddle/fluid/inference/CMakeLists.txt

@@ -10,7 +10,7 @@ set(FLUID_CORE_MODULES proto_desc memory lod_tensor executor)
 # TODO(panyx0718): Should this be called paddle_fluid_inference_api_internal?
 cc_library(paddle_fluid_api
     SRCS io.cc
-    DEPS ${FLUID_CORE_MODULES} ${GLOB_OP_LIB})
+    DEPS ${FLUID_CORE_MODULES} ${GLOB_OP_LIB} graph_to_program_pass)
 
 get_property(fluid_modules GLOBAL PROPERTY FLUID_MODULES)
 

paddle/fluid/inference/analysis/analyzer.cc

@@ -72,7 +72,7 @@ class DfgPassManagerImpl final : public DfgPassManager {
       auto trt_teller = [&](const Node* node) {
         std::unordered_set<std::string> teller_set(
             {"elementwise_add", "mul", "conv2d", "pool2d", "relu", "softmax",
-             "depthwise_conv2d", "batch_norm"});
+             "depthwise_conv2d", "batch_norm", "concat"});
         if (!node->IsFunction()) return false;
 
         const auto* func = static_cast<const Function*>(node);

paddle/fluid/inference/analysis/analyzer_tester.cc

@@ -23,6 +23,8 @@
 
 DEFINE_string(infer_ditu_rnn_model, "", "model path for ditu RNN");
 DEFINE_string(infer_ditu_rnn_data, "", "data path for ditu RNN");
+DEFINE_int32(batch_size, 10, "batch size.");
+DEFINE_int32(repeat, 1, "Running the inference program repeat times.");
 
 namespace paddle {
 namespace inference {
@@ -92,7 +94,7 @@ struct DataRecord {
   size_t batch_iter{0};
   size_t batch_size{1};
   DataRecord() = default;
-  DataRecord(const std::string &path, int batch_size = 1)
+  explicit DataRecord(const std::string &path, int batch_size = 1)
       : batch_size(batch_size) {
     Load(path);
   }
@@ -165,7 +167,6 @@ struct DataRecord {
 };
 void PrepareInputs(std::vector<PaddleTensor> *input_slots, DataRecord *data,
                    int batch_size) {
-  // DataRecord data(FLAGS_datapath, batch_size);
   PaddleTensor lod_attention_tensor, init_zero_tensor, lod_tensor_tensor,
       week_tensor, minute_tensor;
   lod_attention_tensor.name = "data_lod_attention";
@@ -174,28 +175,33 @@ void PrepareInputs(std::vector<PaddleTensor> *input_slots, DataRecord *data,
   week_tensor.name = "week";
   minute_tensor.name = "minute";
   auto one_batch = data->NextBatch();
-  // clang-format off
-  std::vector<int> rnn_link_data_shape
-      ({static_cast<int>(one_batch.rnn_link_data.size()), static_cast<int>(one_batch.rnn_link_data.front().size())});
+  std::vector<int> rnn_link_data_shape(
+      {static_cast<int>(one_batch.rnn_link_data.size()),
+       static_cast<int>(one_batch.rnn_link_data.front().size())});
   lod_attention_tensor.shape.assign({1, 2});
   lod_attention_tensor.lod.assign({one_batch.lod1, one_batch.lod2});
   init_zero_tensor.shape.assign({batch_size, 15});
   init_zero_tensor.lod.assign({one_batch.lod3});
   lod_tensor_tensor.shape = rnn_link_data_shape;
   lod_tensor_tensor.lod.assign({one_batch.lod1});
-  week_tensor.shape.assign({(int) one_batch.rnn_week_datas.size(), (int) one_batch.rnn_week_datas.front().size()});
+  // clang-format off
+  week_tensor.shape.assign(
+      {static_cast<int>(one_batch.rnn_week_datas.size()),
+       static_cast<int>(one_batch.rnn_week_datas.front().size())});
   week_tensor.lod.assign({one_batch.lod3});
-  minute_tensor.shape.assign({(int) one_batch.rnn_minute_datas.size(),
-                              (int) one_batch.rnn_minute_datas.front().size()});
+  minute_tensor.shape.assign(
+      {static_cast<int>(one_batch.rnn_minute_datas.size()),
+       static_cast<int>(one_batch.rnn_minute_datas.front().size())});
   minute_tensor.lod.assign({one_batch.lod3});
+  // clang-format on
   // assign data
-  TensorAssignData(&lod_attention_tensor, std::vector<std::vector<float>>({{0, 0}}));
+  TensorAssignData(&lod_attention_tensor,
+                   std::vector<std::vector<float>>({{0, 0}}));
   std::vector<float> tmp_zeros(batch_size * 15, 0.);
   TensorAssignData(&init_zero_tensor, {tmp_zeros});
   TensorAssignData(&lod_tensor_tensor, one_batch.rnn_link_data);
   TensorAssignData(&week_tensor, one_batch.rnn_week_datas);
   TensorAssignData(&minute_tensor, one_batch.rnn_minute_datas);
-  // clang-format on
   // Set inputs.
   auto init_zero_tensor1 = init_zero_tensor;
   init_zero_tensor1.name = "hidden_init";
@@ -231,12 +237,9 @@ std::string DescribeTensor(const PaddleTensor &tensor) {
   os << "\n";
   os << " - data: ";
 
-  // clang-format off
-  int dim = std::accumulate(tensor.shape.begin(),
-                            tensor.shape.end(),
-                            1,
-                            [](int a, int b) { return a * b; });  // clang-format on
-  for (size_t i = 0; i < dim; i++) {
+  int dim = std::accumulate(tensor.shape.begin(), tensor.shape.end(), 1,
+                            [](int a, int b) { return a * b; });
+  for (int i = 0; i < dim; i++) {
     os << static_cast<float *>(tensor.data.data())[i] << " ";
   }
   os << '\n';
@@ -300,13 +303,16 @@ void TestDituRNNPrediction(const std::string &model_path,
   for (int i = 0; i < num_times; i++) {
     predictor->Run(input_slots, &outputs);
   }
-  LOG(INFO) << "time/batch: " << timer.toc() / num_times;
+  LOG(INFO) << "===========profile result===========";
+  LOG(INFO) << "batch_size: " << batch_size << ", repeat: " << num_times
+            << ", latency: " << timer.toc() / num_times << "ms";
+  LOG(INFO) << "=====================================";
 
   for (auto &out : outputs) {
     size_t size = std::accumulate(out.shape.begin(), out.shape.end(), 1,
                                   [](int a, int b) { return a * b; });
     float *data = static_cast<float *>(out.data.data());
-    for (int i = 0;
+    for (size_t i = 0;
          i < std::min(sizeof(ditu_rnn_target_data) / sizeof(float), size);
          i++) {
       EXPECT_NEAR(data[i], ditu_rnn_target_data[i], 1e-3);
@@ -336,7 +342,7 @@ TEST(Analyzer, SupportIRPass) {
 // Directly infer with the original model.
 TEST(Analyzer, DituRNN_without_analysis) {
   TestDituRNNPrediction(FLAGS_infer_ditu_rnn_model, FLAGS_infer_ditu_rnn_data,
-                        10, false, false);
+                        FLAGS_batch_size, false, false, FLAGS_repeat);
 }
 
 // Inference with the original model with the analysis turned on, the analysis
@@ -344,14 +350,14 @@ TEST(Analyzer, DituRNN_without_analysis) {
 TEST(Analyzer, DituRNN_with_analysis) {
   LOG(INFO) << "ditu rnn with analysis";
   TestDituRNNPrediction(FLAGS_infer_ditu_rnn_model, FLAGS_infer_ditu_rnn_data,
-                        10, true, false, 1);
+                        FLAGS_batch_size, true, false, FLAGS_repeat);
 }
 
 // Inference with analysis and IR. The IR module will fuse some large kernels.
 TEST(Analyzer, DituRNN_with_analysis_with_IR) {
   LOG(INFO) << "ditu rnn with analysis and IR fuse";
   TestDituRNNPrediction(FLAGS_infer_ditu_rnn_model, FLAGS_infer_ditu_rnn_data,
-                        10, true, true, 1);
+                        FLAGS_batch_size, true, true, FLAGS_repeat);
 }
 
 }  // namespace analysis

paddle/fluid/inference/api/api_tensorrt_subgraph_engine.cc

@@ -32,6 +32,7 @@ class TensorRTSubgraphPredictor : public NativePaddlePredictor {
       : NativePaddlePredictor(config), config_(config) {}
 
   bool Init(const std::shared_ptr<framework::Scope>& parent_scope) {
+    FLAGS_IA_enable_tensorrt_subgraph_engine = true;
     VLOG(3) << "Predictor::init()";
     FLAGS_tensorrt_max_batch_size = config_.max_batch_size;
     FLAGS_tensorrt_workspace_size = config_.workspace_size;
@@ -161,3 +162,4 @@ USE_TRT_CONVERTER(fc);
 USE_TRT_CONVERTER(pool2d);
 USE_TRT_CONVERTER(softmax);
 USE_TRT_CONVERTER(batch_norm);
+USE_TRT_CONVERTER(concat);

paddle/fluid/inference/api/api_tensorrt_subgraph_engine_tester.cc

@@ -37,6 +37,7 @@ void CompareTensorRTWithFluid(bool enable_tensorrt) {
   config1.use_gpu = true;
   config1.fraction_of_gpu_memory = 0.3;
   config1.device = 0;
+  config1.max_batch_size = 10;
 
   auto predictor0 =
       CreatePaddlePredictor<NativeConfig, PaddleEngineKind::kNative>(config0);

paddle/fluid/inference/api/demo_ci/CMakeLists.txt

@@ -23,9 +23,11 @@ include_directories("${PADDLE_LIB}")
 include_directories("${PADDLE_LIB}/third_party/install/protobuf/include")
 include_directories("${PADDLE_LIB}/third_party/install/glog/include")
 include_directories("${PADDLE_LIB}/third_party/install/gflags/include")
+if (NOT WIN32)
 include_directories("${PADDLE_LIB}/third_party/install/snappy/include")
 include_directories("${PADDLE_LIB}/third_party/install/snappystream/include")
 include_directories("${PADDLE_LIB}/third_party/install/zlib/include")
+endif(NOT WIN32)
 
 include_directories("${PADDLE_LIB}/third_party/boost")
 include_directories("${PADDLE_LIB}/third_party/eigen3")

paddle/fluid/inference/tensorrt/convert/CMakeLists.txt

 # Add TRT tests
 nv_library(tensorrt_converter
   SRCS mul_op.cc conv2d_op.cc fc_op.cc pool2d_op.cc elementwise_op.cc
-batch_norm_op.cc activation_op.cc softmax_op.cc 
+batch_norm_op.cc activation_op.cc softmax_op.cc concat_op.cc
   DEPS tensorrt_engine operator scope framework_proto op_registry)
 
 nv_test(test_op_converter SRCS test_op_converter.cc DEPS
@@ -18,12 +18,12 @@ nv_test(test_trt_conv_op SRCS test_conv2d_op.cc conv2d_op.cc
         DEPS ${FLUID_CORE_MODULES} tensorrt_engine conv_op SERIAL)
 nv_test(test_trt_pool2d_op SRCS test_pool2d_op.cc pool2d_op.cc
         DEPS ${FLUID_CORE_MODULES} tensorrt_engine pool_op SERIAL)
-
 nv_test(test_trt_elementwise_op SRCS test_elementwise_op.cc elementwise_op.cc
         DEPS ${FLUID_CORE_MODULES} tensorrt_engine elementwise_add_op SERIAL)
-
 nv_test(test_trt_softmax_op SRCS test_softmax_op.cc softmax_op.cc
         DEPS ${FLUID_CORE_MODULES} tensorrt_engine softmax_op SERIAL)
-
 nv_test(test_trt_batch_norm_op SRCS test_batch_norm_op.cc batch_norm_op.cc
         DEPS ${FLUID_CORE_MODULES} tensorrt_engine batch_norm_op SERIAL)
+
+nv_test(test_trt_concat_op SRCS test_concat_op.cc concat_op.cc
+        DEPS ${FLUID_CORE_MODULES} tensorrt_engine concat_op SERIAL)

paddle/fluid/inference/tensorrt/convert/concat_op.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/inference/tensorrt/convert/op_converter.h"
+
+namespace paddle {
+namespace inference {
+namespace tensorrt {
+
+/*
+ * MulOp, IMatrixMultiplyLayer in TRT. This Layer doesn't has weights.
+ */
+class ConcatOpConverter : public OpConverter {
+ public:
+  void operator()(const framework::proto::OpDesc& op,
+                  const framework::Scope& scope, bool test_mode) override {
+    VLOG(4) << "convert a fluid mul op to tensorrt mul layer without bias";
+
+    framework::OpDesc op_desc(op, nullptr);
+    // Declare inputs
+    std::vector<nvinfer1::ITensor*> itensors;
+    for (auto& input_name : op_desc.Input("X")) {
+      itensors.push_back(engine_->GetITensor(input_name));
+    }
+    int axis = boost::get<int>(op_desc.GetAttr("axis"));
+    PADDLE_ENFORCE(axis > 0,
+                   "The axis attr of Concat op should be large than 0 for trt");
+
+    auto* layer = TRT_ENGINE_ADD_LAYER(engine_, Concatenation, itensors.data(),
+                                       itensors.size());
+    axis = axis - 1;  // Remove batch dim
+    layer->setAxis(axis);
+    auto output_name = op_desc.Output("Out")[0];
+    engine_->SetITensor(output_name, layer->getOutput(0));
+    if (test_mode) {  // the test framework can not determine which is the
+                      // output, so place the declaration inside.
+      engine_->DeclareOutput(output_name);
+    }
+  }
+};
+
+}  // namespace tensorrt
+}  // namespace inference
+}  // namespace paddle
+
+REGISTER_TRT_OP_CONVERTER(concat, ConcatOpConverter);

paddle/fluid/inference/tensorrt/convert/op_converter.h

@@ -79,6 +79,14 @@ class OpConverter {
         it =
             Registry<OpConverter>::Lookup("elementwise_" + op_type + "_tensor");
       }
+      PADDLE_ENFORCE_NOT_NULL(it, "no OpConverter for optype [%s]",
+                              op_desc.Type());
+    }
+
+    if (op_desc.Type() == "depthwise_conv2d") {
+      it = Registry<OpConverter>::Lookup("conv2d");
+      PADDLE_ENFORCE_NOT_NULL(it, "no OpConverter for optype [%s]",
+                              op_desc.Type());
     }
 
     if (!it) {

paddle/fluid/inference/tensorrt/convert/test_concat_op.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 <gtest/gtest.h>
+#include "paddle/fluid/inference/tensorrt/convert/op_converter.h"
+#include "paddle/fluid/inference/tensorrt/convert/ut_helper.h"
+
+namespace paddle {
+namespace inference {
+namespace tensorrt {
+
+TEST(concat_op, test) {
+  std::unordered_set<std::string> parameters({""});
+  framework::Scope scope;
+  TRTConvertValidation validator(10, parameters, scope, 1000);
+  validator.DeclInputVar("concat_x1", nvinfer1::DimsCHW(10, 3, 1));
+  validator.DeclInputVar("concat_x2", nvinfer1::DimsCHW(3, 3, 1));
+  validator.DeclInputVar("concat_x3", nvinfer1::DimsCHW(7, 3, 1));
+  validator.DeclOutputVar("concat_out", nvinfer1::DimsCHW(20, 3, 1));
+
+  // Prepare Op description
+  framework::OpDesc desc;
+  desc.SetType("concat");
+  desc.SetInput("X", {"concat_x1", "concat_x2", "concat_x3"});
+  desc.SetOutput("Out", {"concat_out"});
+
+  int axis = 1;
+  desc.SetAttr("axis", axis);
+
+  validator.SetOp(*desc.Proto());
+
+  validator.Execute(5);
+}
+
+}  // namespace tensorrt
+}  // namespace inference
+}  // namespace paddle
+USE_OP(concat);

paddle/fluid/inference/tests/test_helper.h

@@ -18,6 +18,7 @@ limitations under the License. */
 #include <string>
 #include <vector>
 
+#include "paddle/fluid/framework/ir/graph_to_program_pass.h"
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/inference/io.h"
 #include "paddle/fluid/platform/profiler.h"
@@ -135,6 +136,15 @@ std::vector<std::vector<int64_t>> GetFeedTargetShapes(
   return feed_target_shapes;
 }
 
+void Compile(paddle::framework::ProgramDesc* program) {
+  std::unique_ptr<paddle::framework::ir::Graph> g(
+      new paddle::framework::ir::Graph(*program));
+  auto pass = paddle::framework::ir::PassRegistry::Instance().Get(
+      "graph_to_program_pass");
+  pass->SetNotOwned<paddle::framework::ProgramDesc>("program", program);
+  pass->Apply(std::move(g));
+}
+
 template <typename Place, bool CreateVars = true, bool PrepareContext = false>
 void TestInference(const std::string& dirname,
                    const std::vector<paddle::framework::LoDTensor*>& cpu_feeds,
@@ -172,6 +182,8 @@ void TestInference(const std::string& dirname,
         paddle::platform::DeviceContextPool::Instance().Get(place));
     inference_program = InitProgram(&executor, scope, dirname, is_combined);
   }
+  Compile(inference_program.get());
+
   // Disable the profiler and print the timing information
   paddle::platform::DisableProfiler(paddle::platform::EventSortingKey::kDefault,
                                     "load_program_profiler");
@@ -249,3 +261,5 @@ void TestInference(const std::string& dirname,
 
   delete scope;
 }
+
+USE_PASS(graph_to_program_pass);

paddle/fluid/memory/detail/system_allocator.cc

@@ -11,12 +11,18 @@ distributed under the License is distributed on an "AS IS" BASIS,
 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
+#define GLOG_NO_ABBREVIATED_SEVERITIES
 
 #include "paddle/fluid/memory/detail/system_allocator.h"
 
-#include <stdlib.h>    // for malloc and free
+#ifdef _WIN32
+#include <malloc.h>
+#include <windows.h>  // VirtualLock/VirtualUnlock
+#else
 #include <sys/mman.h>  // for mlock and munlock
-#include <algorithm>   // for std::max
+#endif
+#include <stdlib.h>   // for malloc and free
+#include <algorithm>  // for std::max
 
 #include "gflags/gflags.h"
 #include "paddle/fluid/platform/assert.h"
@@ -35,31 +41,42 @@ namespace paddle {
 namespace memory {
 namespace detail {
 
-void* CPUAllocator::Alloc(size_t* index, size_t size) {
-  // According to http://www.cplusplus.com/reference/cstdlib/malloc/,
-  // malloc might not return nullptr if size is zero, but the returned
-  // pointer shall not be dereferenced -- so we make it nullptr.
-  if (size <= 0) return nullptr;
-
-  *index = 0;  // unlock memory
-
+void* AlignedMalloc(size_t size) {
   void* p = nullptr;
-
+  size_t alignment = 32ul;
 #ifdef PADDLE_WITH_MKLDNN
   // refer to https://github.com/01org/mkl-dnn/blob/master/include/mkldnn.hpp
   // memory alignment
-  PADDLE_ENFORCE_EQ(posix_memalign(&p, 4096ul, size), 0, "Alloc %ld error!",
-                    size);
+  alignment = 4096ul;
+#endif
+#ifdef _WIN32
+  p = _aligned_malloc(size, alignment);
 #else
-  PADDLE_ENFORCE_EQ(posix_memalign(&p, 32ul, size), 0, "Alloc %ld error!",
+  PADDLE_ENFORCE_EQ(posix_memalign(&p, alignment, size), 0, "Alloc %ld error!",
                     size);
 #endif
   PADDLE_ENFORCE(p, "Fail to allocate CPU memory: size = %d .", size);
+  return p;
+}
+
+void* CPUAllocator::Alloc(size_t* index, size_t size) {
+  // According to http://www.cplusplus.com/reference/cstdlib/malloc/,
+  // malloc might not return nullptr if size is zero, but the returned
+  // pointer shall not be dereferenced -- so we make it nullptr.
+  if (size <= 0) return nullptr;
+
+  *index = 0;  // unlock memory
+
+  void* p = AlignedMalloc(size);
 
   if (p != nullptr) {
     if (FLAGS_use_pinned_memory) {
       *index = 1;
+#ifdef _WIN32
+      VirtualLock(p, size);
+#else
       mlock(p, size);  // lock memory
+#endif
     }
   }
 
@@ -68,7 +85,11 @@ void* CPUAllocator::Alloc(size_t* index, size_t size) {
 
 void CPUAllocator::Free(void* p, size_t size, size_t index) {
   if (p != nullptr && index == 1) {
+#ifdef _WIN32
+    VirtualUnlock(p, size);
+#else
     munlock(p, size);
+#endif
   }
   free(p);
 }

paddle/fluid/operators/CMakeLists.txt

@@ -85,7 +85,7 @@ function(op_library TARGET)
 
     #remove windows unsupported op
     if (WIN32)
-    foreach(windows_unsupport_op "nccl_op" "gen_nccl_id_op")
+    foreach(windows_unsupport_op "nccl_op" "gen_nccl_id_op" "warpctc_op")
         if ("${TARGET}" STREQUAL "${windows_unsupport_op}")
           return()
         endif()
@@ -291,6 +291,8 @@ op_library(unsqueeze_op DEPS reshape_op)
 op_library(squeeze_op DEPS reshape_op)
 op_library(extract_rows_op DEPS memory)
 op_library(flatten_op DEPS reshape_op)
+op_library(sequence_pad_op DEPS sequence_padding)
+op_library(unstack_op DEPS stack_op)
 
 if (WITH_GPU)
     op_library(conv_op DEPS vol2col depthwise_conv im2col)
@@ -319,8 +321,9 @@ foreach(src ${GENERAL_OPS})
 endforeach()
 
 file(APPEND ${pybind_file} "USE_OP(less_than);\nUSE_OP(logical_and);\nUSE_NO_KERNEL_OP(read_from_array);\n")
-
+if (NOT WIN32)
 add_subdirectory(reader)
+endif(NOT WIN32)
 foreach(src ${READER_LIBRARY})
     set(OP_LIBRARY ${src} ${OP_LIBRARY})
 endforeach()

paddle/fluid/operators/attention_lstm_op.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/fluid/operators/attention_lstm_op.h"
+#include <sys/time.h>
+#include <string>
+#include "paddle/fluid/operators/math/blas.h"
+#include "paddle/fluid/operators/math/cpu_vec.h"
+#include "paddle/fluid/operators/math/fc_compute.h"
+#include "paddle/fluid/platform/cpu_info.h"
+
+namespace paddle {
+namespace operators {
+
+void AttentionLSTMOp::InferShape(framework::InferShapeContext* ctx) const {
+  PADDLE_ENFORCE(ctx->HasInput("X"),
+                 "Input(X) of AttentionLSTM should not be null.");
+  PADDLE_ENFORCE(ctx->HasInput("C0"),
+                 "Input(C0) of AttentionLSTM should not be null.");
+  PADDLE_ENFORCE(ctx->HasInput("LSTMWeight"),
+                 "Input(LSTMWeight) of AttentionLSTM should not be null.");
+  PADDLE_ENFORCE(ctx->HasInput("LSTMBias"),
+                 "Input(LSTMBias) of AttentionLSTM should not be null.");
+  PADDLE_ENFORCE(ctx->HasInput("AttentionWeight"),
+                 "Input(AttentionWeight) of AttentionLSTM should not be null.");
+
+  PADDLE_ENFORCE(ctx->HasOutput("Hidden"),
+                 "Output(Hidden) of AttentionLSTM should not be null.");
+  PADDLE_ENFORCE(ctx->HasOutput("Cell"),
+                 "Output(Cell) of AttentionLSTM should not be null.");
+  PADDLE_ENFORCE(ctx->HasOutput("AttentionedX"),
+                 "Output(AttentionedX) of AttentionLSTM should not be null.");
+  PADDLE_ENFORCE(ctx->HasOutput("AttentionFCOut"),
+                 "Output(AttentionFCOut) of AttentionLSTM should not be null.");
+  PADDLE_ENFORCE(ctx->HasOutput("LSTMX"),
+                 "Output(LSTMX) of AttentionLSTM should not be null.");
+  PADDLE_ENFORCE(ctx->HasOutput("LSTMOUT"),
+                 "Output(LSTMOUT) of AttentionLSTM should not be null.");
+
+  auto x_dims = ctx->GetInputDim("X");
+  const int M = x_dims[1];
+  PADDLE_ENFORCE_EQ(x_dims.size(), 2, "Input(X)'s rank must be 2.");
+
+  auto w_dims = ctx->GetInputDim("LSTMWeight");
+  const int D = w_dims[1] / 4;
+  PADDLE_ENFORCE_EQ(w_dims.size(), 2, "Input(LSTMWeight)'s rank must be 2.");
+  PADDLE_ENFORCE_EQ(w_dims[0], D + M,
+                    "LSTMWeight dims should be (%d + %d) * %d.", D + M, 4 * D);
+
+  auto b_dims = ctx->GetInputDim("LSTMBias");
+  PADDLE_ENFORCE_EQ(b_dims.size(), 2, "Input(LSTMBias)'s rank must be 2.");
+  PADDLE_ENFORCE_EQ(b_dims[0], 1, "LSTMBias dims should be 1 x %d.", 4 * D);
+  PADDLE_ENFORCE_EQ(b_dims[1], 4 * D, "LSTMBias dims should be 1 x %d.", 4 * D);
+
+  auto c_dims = ctx->GetInputDim("C0");
+  PADDLE_ENFORCE_EQ(c_dims.size(), 2, "Input(C0)'s rank must be 2.");
+  PADDLE_ENFORCE_EQ(c_dims[1], D, "C0 dims should be N x %d.", D);
+  if (ctx->HasInput("H0")) {
+    auto h_dims = ctx->GetInputDim("H0");
+    PADDLE_ENFORCE(h_dims == c_dims,
+                   "The dimension of Input(H0) and Input(C0) "
+                   "should be the same.");
+  }
+
+  auto atten_w_dims = ctx->GetInputDim("AttentionWeight");
+  PADDLE_ENFORCE_EQ(atten_w_dims.size(), 2,
+                    "Input(AttentionWeight)'s rank must be 2.");
+  PADDLE_ENFORCE_EQ(atten_w_dims[0], M + D,
+                    "AttentionWeight shapes must be (%d + %d) * 1.", M, D);
+  PADDLE_ENFORCE_EQ(atten_w_dims[1], 1,
+                    "AttentionWeight shapes must be (%d + %d) * 1.", M, D);
+  if (ctx->HasInput("AttentionBias")) {
+    auto atten_b_dims = ctx->GetInputDim("AttentionBias");
+    PADDLE_ENFORCE_EQ(atten_b_dims.size(), 2,
+                      "Input(AttentionBias)'s rank must be 2.");
+    PADDLE_ENFORCE_EQ(atten_b_dims[0], 1,
+                      "AttentionBias shapes must be 1 * 1.");
+    PADDLE_ENFORCE_EQ(atten_b_dims[1], 1,
+                      "AttentionBias shapes must be 1 * 1.");
+  }
+
+  if (ctx->HasInput("AttentionScalar")) {
+    auto dims = ctx->GetInputDim("AttentionScalar");
+    PADDLE_ENFORCE_EQ(dims.size(), 2,
+                      "Input(AttentionScalar)'s rank must be 2.");
+    PADDLE_ENFORCE_EQ(dims[0], 1, "AttentionScalar shapes must be 1 * 1.");
+    PADDLE_ENFORCE_EQ(dims[1], 1, "AttentionScalar shapes must be 1 * 1.");
+  }
+
+  if (ctx->HasInput("AttentionScalarBias")) {
+    auto dims = ctx->GetInputDim("AttentionScalarBias");
+    PADDLE_ENFORCE(
+        ctx->HasInput("AttentionScalar"),
+        "AttentionScalar should not be null when have AttentionScalarBias.");
+    PADDLE_ENFORCE_EQ(dims.size(), 2,
+                      "Input(AttentionScalarBias)'s rank must be 2.");
+    PADDLE_ENFORCE_EQ(dims[0], 1, "AttentionScalarBias shapes must be 1 * 1.");
+    PADDLE_ENFORCE_EQ(dims[1], 1, "AttentionScalarBias shapes must be 1 * 1.");
+  }
+
+  framework::DDim out_dims({x_dims[0], D});
+  ctx->SetOutputDim("Hidden", out_dims);
+  ctx->SetOutputDim("Cell", out_dims);
+  ctx->SetOutputDim("AttentionedX", {x_dims[0], 1});
+  ctx->SetOutputDim("LSTMX", {1, M});
+  ctx->SetOutputDim("LSTMOUT", {1, 4 * D});
+  // AttentionFCOut should be reshape as (maxseqlen,1) in runtime
+  ctx->ShareLoD("X", "Hidden");
+  ctx->ShareLoD("X", "Cell");
+}
+
+framework::OpKernelType AttentionLSTMOp::GetExpectedKernelType(
+    const framework::ExecutionContext& ctx) const {
+  return framework::OpKernelType(
+      framework::ToDataType(ctx.Input<framework::LoDTensor>("X")->type()),
+      ctx.device_context());
+}
+
+void AttentionLSTMOpMaker::Make() {
+  AddInput("X",
+           "(LoDTensor) the input is a LodTensor, which support "
+           "variable-time length input sequence. The underlying tensor in "
+           "this LoDTensor is a matrix with shape (T X M), where T is the "
+           "total time steps in this mini-batch, M is the dim size of x.");
+  AddInput("C0",
+           "(Tensor) LSTM C0"
+           "This is a tensor with shape (N x D), where N is the batch size, D "
+           "is the gate size."
+           "C0 is necessary because of attention.");
+  AddInput("H0",
+           "(Tensor, optional) LSTM H0"
+           "This is a tensor with shape (N x D), where N is the "
+           "batch size and D is the gate size.")
+      .AsDispensable();
+  AddInput("AttentionWeight",
+           "(Tensor) the weights of attention fc. Always relu the fc result."
+           "The shape is ((M+D) x 1), where M is the dim size of x, D is the "
+           "gate size of LSTM.");
+  AddInput("AttentionBias",
+           "(Tensor, optional) the bias of attention fc."
+           "The shape is (1 x 1)")
+      .AsDispensable();
+  AddInput("AttentionScalar",
+           "(Tensor, optional) the scalar on the result of attentioned fc. "
+           "Always relu the Scalar."
+           "The shape is (1 x 1)")
+      .AsDispensable();
+  AddInput("AttentionScalarBias",
+           "(Tensor, optional) the scalar bias of attention fc."
+           "The shape is (1 x 1)")
+      .AsDispensable();
+  AddInput("LSTMWeight",
+           "(Tensor) the combined weight of LSTM"
+           " - The shape is ((D+M) x 4D), where D is the hidden gate size, M "
+           "is the dim size of x"
+           " - Weight = {W_forget, W_input, W_output, W_cell}");
+  AddInput("LSTMBias",
+           "(Tensor) the combined bias of LSTM, shape (1x4D)."
+           "Note: we should add the bias of hidden and context accorindg to "
+           "the same gate: "
+           "{B_forget, B_input, B_output, B_cell}");
+  AddOutput("Hidden",
+            "(LoDTensor) (same as LSTMOp) the hidden state of LSTM operator. "
+            "The shape is (T x D), and lod is the same with the `Input`.");
+  AddOutput("Cell",
+            "(LoDTensor) (same as LSTMOp) the cell state of LSTM operator. "
+            "The shape is (T x D), and lod is the same with the `Input`.");
+  AddOutput("AttentionedX",
+            "(Tensor) shape is (T x 1), the result after X * AttentionWeight,"
+            " where T is the total time steps in this mini-batch,"
+            " D is the hidden size.")
+      .AsIntermediate();
+  AddOutput("AttentionFCOut",
+            "(Tensor) (max_seq_len, 1), compute at each step.")
+      .AsIntermediate();
+  AddOutput("LSTMX",
+            "(Tensor) the input X of LSTM for each step."
+            "Shape is (1 x M), where M is the x frame size")
+      .AsIntermediate();
+  AddOutput(
+      "LSTMOUT",
+      "(Tensor) the output of LSTM X(1*(D+M))* weight((D+M)*4D) for each step."
+      "Shape is (1 x 4D), where M is the x frame size")
+      .AsIntermediate();
+  AddAttr<std::string>("gate_activation",
+                       "(string, default: sigmoid)"
+                       "The activation for input gate, forget gate and output "
+                       "gate, `sigmoid` by default.")
+      .SetDefault("sigmoid")
+      .InEnum({"sigmoid", "tanh", "relu", "identity"});
+  AddAttr<std::string>("cell_activation",
+                       "(string, default: tanh)"
+                       "The activation for cell output, `tanh` by defalut.")
+      .SetDefault("tanh")
+      .InEnum({"sigmoid", "tanh", "relu", "identity"});
+  AddAttr<std::string>("candidate_activation",
+                       "(string, default: tanh)"
+                       "The activation for candidate hidden state, "
+                       "`tanh` by default.")
+      .SetDefault("tanh")
+      .InEnum({"sigmoid", "tanh", "relu", "identity"});
+  AddComment(R"DOC(
+Attention Long-Short Term Memory (LSTM) Operator.
+
+Attention part:
+concat( x(seqlen * M), expand( cell_t-1(1,D) ) ) => tmp(seqlen*(M+D))
+
+tmp(seqlen*(M+D)) * fc((M+D)*1) => fcout(seqlen*1) with bias, relu
+
+fcout(seqlen*1) * scalar => fcout(seqlen*1) with bias, relu
+
+dotmul and sum pool ( fcout(seqlen*1), x(seqlen * M) ) => lstm_x_t(1, M) 
+
+LSTM part:
+use lstm_x_t as input and compute as standard LSTM.
+
+)DOC");
+}
+
+// y[i] = (x[i] + bias[0]) > 0 ? (x[i] + bias[0]) : 0;
+template <typename T>
+inline void bias_relu(const int n, const T* x, const T* bias, T* y) {
+  if (bias) {
+    math::vec_add_bias<T, platform::jit::avx>(n, *bias, x, y);
+    math::vec_relu<T, platform::jit::avx>(n, y, y);
+  } else {
+    math::vec_relu<T, platform::jit::avx>(n, x, y);
+  }
+}
+
+template <typename T>
+inline void vec_softmax(const int n, const T* x, T* y) {
+  T scalar = x[0];
+  // max
+  for (int i = 1; i < n; ++i) {
+    scalar = scalar < x[i] ? x[i] : scalar;
+  }
+  math::vec_add_bias<T, platform::jit::avx>(n, -scalar, x, y);  // sub
+  math::vec_exp<T>(n, y, y);                                    // exp
+  // sum
+  scalar = T(0);
+  for (int i = 0; i < n; ++i) {
+    scalar += y[i];
+  }
+  math::vec_scal<T>(n, static_cast<T>(1) / scalar, y);  // scale
+}
+
+template <typename T>
+class AttentionLSTMKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    using DeviceContext = paddle::platform::CPUDeviceContext;
+
+    auto* x = ctx.Input<LoDTensor>("X");
+    auto* h0 = ctx.Input<Tensor>("H0");
+    auto* c0 = ctx.Input<Tensor>("C0");
+    auto* atten_w = ctx.Input<Tensor>("AttentionWeight");
+    auto* atten_b = ctx.Input<Tensor>("AttentionBias");
+    auto* atten_scalar = ctx.Input<Tensor>("AttentionScalar");
+    auto* atten_scalar_bias = ctx.Input<Tensor>("AttentionScalarBias");
+    auto* lstm_w = ctx.Input<Tensor>("LSTMWeight");
+    auto* lstm_b = ctx.Input<Tensor>("LSTMBias");
+
+    auto* hidden_out = ctx.Output<LoDTensor>("Hidden");
+    auto* cell_out = ctx.Output<LoDTensor>("Cell");
+    auto* atted_x = ctx.Output<Tensor>("AttentionedX");
+    auto* fc_out = ctx.Output<Tensor>("AttentionFCOut");
+    auto* lstm_x = ctx.Output<Tensor>("LSTMX");
+    auto* lstm_out = ctx.Output<Tensor>("LSTMOUT");
+
+    // some shape should be reshape here since infershape can not get lod info
+    auto x_lod = x->lod();
+    const int N = x_lod[0].size() - 1;  // batch size
+    auto x_dims = x->dims();            // T x M
+    auto w_dims = lstm_w->dims();       // (D+M) x 4D
+    const int total_T = x_dims[0];
+    const int M = x_dims[1];      // x frame size
+    const int D = w_dims[1] / 4;  // gate frame size
+    const int D2 = D * 2;
+    const int D3 = D * 3;
+    const int D4 = w_dims[1];
+    int max_seq_len = x_lod[0][1];
+    for (int i = 1; i < N; ++i) {
+      int len = x_lod[0][i + 1] - x_lod[0][i];
+      max_seq_len = max_seq_len < len ? len : max_seq_len;
+    }
+    PADDLE_ENFORCE_EQ(x_lod.size(), 1, "Input(X)'s lod size must be 1.");
+    PADDLE_ENFORCE_EQ(c0->dims()[0], N, "C0 dims should be %d x %d.", N, D);
+    fc_out->Resize({max_seq_len, 1});
+
+    std::function<void(const int, const T *, T *)> act_gate, act_cell, act_cand;
+    auto& act_gate_str = ctx.Attr<std::string>("gate_activation");
+    auto& act_cell_str = ctx.Attr<std::string>("cell_activation");
+    auto& act_cand_str = ctx.Attr<std::string>("candidate_activation");
+    if (platform::jit::MayIUse(platform::jit::avx)) {
+      math::VecActivations<T, platform::jit::avx> act_functor;
+      act_gate = act_functor(act_gate_str);
+      act_cell = act_functor(act_cell_str);
+      act_cand = act_functor(act_cand_str);
+    } else {
+      math::VecActivations<T, platform::jit::isa_any> act_functor;
+      act_gate = act_functor(act_gate_str);
+      act_cell = act_functor(act_cell_str);
+      act_cand = act_functor(act_cand_str);
+    }
+
+    const T* x_data = x->data<T>();
+    const T* h0_data = h0 ? h0->data<T>() : NULL;
+    const T* c0_data = c0->data<T>();
+    const T* lstm_w_data = lstm_w->data<T>();
+    const T* lstm_b_data = lstm_b->data<T>();
+    const T* atten_w_data = atten_w->data<T>();
+    const T* atten_b_data = atten_b ? atten_b->data<T>() : NULL;
+    const T* atten_scalar_data = atten_scalar ? atten_scalar->data<T>() : NULL;
+    const T* atten_scalar_bias_data =
+        atten_scalar_bias ? atten_scalar_bias->data<T>() : NULL;
+
+    T* hidden_out_data = hidden_out->mutable_data<T>(ctx.GetPlace());
+    T* cell_out_data = cell_out->mutable_data<T>(ctx.GetPlace());
+    T* atted_x_data = atted_x->mutable_data<T>(ctx.GetPlace());
+    T* fc_out_data = fc_out->mutable_data<T>(ctx.GetPlace());
+    T* lstm_x_data = lstm_x->mutable_data<T>(ctx.GetPlace());
+    T* lstm_out_data = lstm_out->mutable_data<T>(ctx.GetPlace());
+
+    // x(TxM) * fc (Mx1) part of atten_wgt(M+D)x1
+    auto blas = math::GetBlas<DeviceContext, T>(ctx);
+    math::FCCompute<DeviceContext, T>(blas, total_T, 1, M, x_data, atten_w_data,
+                                      atted_x_data, atten_b_data);
+
+    const T* cur_atten_x_data = atted_x_data;
+    const T* cur_x_data = x_data;
+    const T* prev_cell_data = NULL;
+    const T* prev_hidden_data = NULL;
+    T* cur_cell_out_data = cell_out_data;
+    T* cur_hidden_out_data = hidden_out_data;
+    for (int i = 0; i < N; ++i) {
+      int seq_len = x_lod[0][i + 1] - x_lod[0][i];
+      prev_cell_data = c0_data + i * D;
+      prev_hidden_data = h0_data ? h0_data + i * D : NULL;
+      for (int step = 0; step < seq_len; ++step) {
+        /// 1. compute attention vector
+        // 1a. prev_cell(1xD) * fc(D) rest part of atten_wgt
+        T prev_cell_bias = blas.DOT(D, prev_cell_data, atten_w_data + M);
+        // 1b. add cell bias and relu
+        bias_relu<T>(seq_len, cur_atten_x_data, &prev_cell_bias, fc_out_data);
+        // 1c. fc scalar
+        if (atten_scalar_data) {
+          blas.SCAL(seq_len, *atten_scalar_data, fc_out_data);
+          bias_relu<T>(seq_len, fc_out_data, atten_scalar_bias_data,
+                       fc_out_data);
+        }
+        // 1d. softmax
+        vec_softmax<T>(seq_len, fc_out_data, fc_out_data);
+        // mul x(seq_len*M) and sum pool
+        math::FCCompute<DeviceContext, T>(blas, 1, M, seq_len, fc_out_data,
+                                          cur_x_data, lstm_x_data);
+
+        /// 2. compute LSTM step
+        // lstm weight : concat[forget , input , output , tilde]
+        // shape : (D + M) x (4 * D)
+        // fc inputX(1xM) * weightX(M*(4D))  => 1 x 4D
+        blas.MatMul(1, D4, M, lstm_x_data, lstm_w_data + D * D4, lstm_out_data);
+        if (prev_hidden_data) {
+          blas.GEMM(CblasNoTrans, CblasNoTrans, 1, D4, D, static_cast<T>(1),
+                    prev_hidden_data, D, lstm_w_data, D4, static_cast<T>(1),
+                    lstm_out_data, D4);
+        }
+        // since input is 1xM, so can use add bias
+        blas.VADD(D4, lstm_b_data, lstm_out_data, lstm_out_data);
+
+        // gate act: sigmoid
+        act_gate(D3, lstm_out_data, lstm_out_data);
+        // candicate act: tanh
+        act_cand(D, lstm_out_data + D3, lstm_out_data + D3);
+
+        // a = forget * prev_cell
+        blas.VMUL(D, lstm_out_data, prev_cell_data, lstm_out_data);
+
+        // b = input * tilde
+        blas.VMUL(D, lstm_out_data + D, lstm_out_data + D3, lstm_out_data + D);
+
+        // cell_out = a + b
+        blas.VADD(D, lstm_out_data, lstm_out_data + D, cur_cell_out_data);
+
+        // state act tanh(cell_out) * output_gate
+        act_cell(D, cur_cell_out_data, lstm_out_data);
+        blas.VMUL(D, lstm_out_data, lstm_out_data + D2, cur_hidden_out_data);
+
+        prev_hidden_data = cur_hidden_out_data;
+        prev_cell_data = cur_cell_out_data;
+        cur_cell_out_data = cur_cell_out_data + D;
+        cur_hidden_out_data = cur_hidden_out_data + D;
+      }
+      cur_x_data = cur_x_data + seq_len * M;
+      cur_atten_x_data = cur_atten_x_data + seq_len;
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OPERATOR(attention_lstm, ops::AttentionLSTMOp,
+                  ops::AttentionLSTMOpMaker,
+                  paddle::framework::DefaultGradOpDescMaker<true>);
+
+REGISTER_OP_CPU_KERNEL(attention_lstm, ops::AttentionLSTMKernel<float>,
+                       ops::AttentionLSTMKernel<double>);

paddle/fluid/operators/attention_lstm_op.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+#include "paddle/fluid/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+using LoDTensor = framework::LoDTensor;
+using Tensor = framework::Tensor;
+
+class AttentionLSTMOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override;
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override;
+};
+
+class AttentionLSTMOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() override;
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/concat_op.h

@@ -62,9 +62,21 @@ class ConcatGradKernel : public framework::OpKernel<T> {
   void Compute(const framework::ExecutionContext& ctx) const {
     auto* out_grad =
         ctx.Input<framework::Tensor>(framework::GradVarName("Out"));
-    auto ins = ctx.MultiInput<framework::Tensor>("X");
+    auto ins = ctx.MultiInput<framework::LoDTensor>("X");
     auto out_var_names = ctx.Outputs(framework::GradVarName("X"));
-    auto outs = ctx.MultiOutput<framework::Tensor>(framework::GradVarName("X"));
+    auto outs =
+        ctx.MultiOutput<framework::LoDTensor>(framework::GradVarName("X"));
+
+    {
+      auto dx = outs;
+      auto x = ins;
+      for (size_t i = 0; i < dx.size(); ++i) {
+        if (dx[i] != nullptr) {
+          dx[i]->set_lod(x[i]->lod());
+        }
+      }
+    }
+
     int64_t axis = static_cast<int64_t>(ctx.Attr<int>("axis"));
 
     // get output tensor that the name is not kEmptyVarName

paddle/fluid/operators/conv_mkldnn_op.cc

@@ -53,6 +53,18 @@ class ConvMKLDNNHandler : public platform::MKLDNNHandler {
     key_ += "-BWD";
   }
 
+  size_t GetDstMemorySize() const {
+    return conv_pd_->dst_primitive_desc().get_size();
+  }
+
+  size_t GetDiffWeightsMemorySize() const {
+    return conv_bwd_weights_pd_->diff_weights_primitive_desc().get_size();
+  }
+
+  size_t GetDiffSourceMemorySize() const {
+    return conv_bwd_data_pd_->diff_src_primitive_desc().get_size();
+  }
+
   std::shared_ptr<mkldnn::memory> AcquireSrcMemoryFromWeightsPrimitive(
       const std::shared_ptr<mkldnn::memory> user_memory_p,
       std::vector<mkldnn::primitive>& pipeline) {  // NOLINT
@@ -294,7 +306,6 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
 
     const T* input_data = input->data<T>();
     const T* filter_data = filter->data<T>();
-    T* output_data = output->mutable_data<T>(ctx.GetPlace());
 
     std::vector<int> src_tz = paddle::framework::vectorize2int(input->dims());
     std::vector<int> weights_tz =
@@ -354,6 +365,8 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
     auto user_weights_memory_p = handler.AcquireWeightsMemory(
         user_weights_md, to_void_cast<T>(filter_data));
 
+    T* output_data =
+        output->mutable_data<T>(ctx.GetPlace(), handler.GetDstMemorySize());
     // create reorder primitive if the input format is not the preferred one
     auto src_memory_p =
         handler.AcquireSrcMemoryFromPrimitive(user_src_memory_p, pipeline);
@@ -476,13 +489,6 @@ class ConvMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
     T* input_grad_data = nullptr;
     T* filter_grad_data = nullptr;
 
-    if (input_grad) {
-      input_grad_data = input_grad->mutable_data<T>(ctx.GetPlace());
-    }
-    if (filter_grad) {
-      filter_grad_data = filter_grad->mutable_data<T>(ctx.GetPlace());
-    }
-
     std::vector<int> src_tz = paddle::framework::vectorize2int(input->dims());
     std::vector<int> weights_tz =
         paddle::framework::vectorize2int(filter->dims());
@@ -568,6 +574,9 @@ class ConvMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
           handler.AcquireDiffDstMemoryFromWeightsPrimitive(
               user_diff_dst_memory_p, pipeline);
 
+      const size_t size = handler.GetDiffWeightsMemorySize();
+      filter_grad_data = filter_grad->mutable_data<T>(ctx.GetPlace(), size);
+
       auto diff_weights_memory_p =
           handler.AcquireDiffWeightsMemoryFromWeightsPrimitive(
               reinterpret_cast<void*>(filter_grad_data));
@@ -590,6 +599,9 @@ class ConvMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
           handler.AcquireDiffDstMemoryFromDataPrimitive(user_diff_dst_memory_p,
                                                         pipeline);
 
+      const size_t size = handler.GetDiffSourceMemorySize();
+      input_grad_data = input_grad->mutable_data<T>(ctx.GetPlace(), size);
+
       auto diff_src_memory_p = handler.AcquireDiffSrcMemoryFromDataPrimitive(
           reinterpret_cast<void*>(input_grad_data));
 

paddle/fluid/operators/distributed/variable_response.cc

@@ -151,6 +151,7 @@ bool VariableResponse::CopySelectRowsData(
     ::google::protobuf::io::CodedInputStream* input,
     const platform::DeviceContext& ctx, int length) {
   auto* slr = GetVar()->GetMutable<framework::SelectedRows>();
+  slr->mutable_rows()->clear();
   slr->mutable_rows()->resize(length /
                               framework::SizeOfType(typeid(int64_t)));  // int64
   int64_t* rows_data = slr->mutable_rows()->data();

paddle/fluid/operators/elementwise_add_mkldnn_op.cc

@@ -137,9 +137,10 @@ class EltwiseAddMKLDNNKernel : public framework::OpKernel<T> {
 };
 
 template <typename T>
-class EltwiseAddMKLDNNGradKernel : public framework::OpKernel<T> {
+class EltwiseAddMKLDNNGradKernel : public ElemwiseGradKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
+    ElemwiseGradKernel<T>::Compute(ctx);
     using Tensor = framework::Tensor;
 
     auto* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));

paddle/fluid/operators/elementwise_add_op.h

@@ -15,6 +15,7 @@ limitations under the License. */
 #pragma once
 
 #include "paddle/fluid/framework/eigen.h"
+#include "paddle/fluid/operators/elementwise_op.h"
 #include "paddle/fluid/operators/elementwise_op_function.h"
 #include "paddle/fluid/operators/math/blas.h"
 
@@ -136,9 +137,11 @@ elementwise_add_grad(const framework::ExecutionContext& ctx,
 }
 
 template <typename DeviceContext, typename T>
-class ElementwiseAddGradKernel : public framework::OpKernel<T> {
+class ElementwiseAddGradKernel : public ElemwiseGradKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
+    ElemwiseGradKernel<T>::Compute(ctx);
+
     using Tensor = framework::Tensor;
 
     auto* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));

paddle/fluid/operators/elementwise_div_op.h

@@ -14,8 +14,8 @@ limitations under the License. */
 
 #pragma once
 
+#include "paddle/fluid/operators/elementwise_op.h"
 #include "paddle/fluid/operators/elementwise_op_function.h"
-
 namespace paddle {
 namespace operators {
 
@@ -53,9 +53,10 @@ struct DivGradDY {
 };
 
 template <typename DeviceContext, typename T>
-class ElementwiseDivGradKernel : public framework::OpKernel<T> {
+class ElementwiseDivGradKernel : public ElemwiseGradKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
+    ElemwiseGradKernel<T>::Compute(ctx);
     using Tensor = framework::Tensor;
 
     auto* x = ctx.Input<Tensor>("X");

paddle/fluid/operators/elementwise_max_op.h

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #pragma once
 
+#include "paddle/fluid/operators/elementwise_op.h"
 #include "paddle/fluid/operators/elementwise_op_function.h"
 
 namespace paddle {
@@ -55,9 +56,10 @@ struct MaxGradDy {
 };
 
 template <typename DeviceContext, typename T>
-class ElementwiseMaxGradKernel : public framework::OpKernel<T> {
+class ElementwiseMaxGradKernel : public ElemwiseGradKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
+    ElemwiseGradKernel<T>::Compute(ctx);
     using Tensor = framework::Tensor;
 
     auto* x = ctx.Input<Tensor>("X");

paddle/fluid/operators/elementwise_min_op.h

@@ -14,8 +14,8 @@ limitations under the License. */
 
 #pragma once
 
+#include "paddle/fluid/operators/elementwise_op.h"
 #include "paddle/fluid/operators/elementwise_op_function.h"
-
 namespace paddle {
 namespace operators {
 
@@ -55,9 +55,10 @@ struct MinGradDy {
 };
 
 template <typename DeviceContext, typename T>
-class ElementwiseMinGradKernel : public framework::OpKernel<T> {
+class ElementwiseMinGradKernel : public ElemwiseGradKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
+    ElemwiseGradKernel<T>::Compute(ctx);
     using Tensor = framework::Tensor;
 
     auto* x = ctx.Input<Tensor>("X");

paddle/fluid/operators/elementwise_mul_op.h

@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
+#include "paddle/fluid/operators/elementwise_op.h"
 #include "paddle/fluid/operators/elementwise_op_function.h"
 #include "paddle/fluid/operators/math/blas.h"
 
@@ -84,9 +85,10 @@ struct MulGradDY {
 };
 
 template <typename DeviceContext, typename T>
-class ElementwiseMulGradKernel : public framework::OpKernel<T> {
+class ElementwiseMulGradKernel : public ElemwiseGradKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
+    ElemwiseGradKernel<T>::Compute(ctx);
     using Tensor = framework::Tensor;
 
     auto* x = ctx.Input<Tensor>("X");

paddle/fluid/operators/elementwise_op.h

@@ -205,6 +205,20 @@ class ElementwiseOpExplicitGrad : public ElementwiseOpGrad {
   }
 };
 
+template <typename T>
+class ElemwiseGradKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* dx =
+        context.Output<framework::LoDTensor>(framework::GradVarName("X"));
+    if (dx != nullptr) {
+      auto& dout =
+          *context.Input<framework::LoDTensor>(framework::GradVarName("Out"));
+      dx->set_lod(dout.lod());
+    }
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle
 

paddle/fluid/operators/elementwise_sub_op.h

@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
+#include "paddle/fluid/operators/elementwise_op.h"
 #include "paddle/fluid/operators/elementwise_op_function.h"
 
 namespace paddle {
@@ -50,9 +51,10 @@ struct SubGradDY {
 };
 
 template <typename DeviceContext, typename T>
-class ElementwiseSubGradKernel : public framework::OpKernel<T> {
+class ElementwiseSubGradKernel : public ElemwiseGradKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
+    ElemwiseGradKernel<T>::Compute(ctx);
     using Tensor = framework::Tensor;
 
     auto* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));

paddle/fluid/operators/fake_dequantize_op.cc

@@ -18,15 +18,32 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
+template <typename T>
+struct DequantizeFunctor<platform::CPUDeviceContext, T> {
+  void operator()(const platform::CPUDeviceContext& dev_ctx,
+                  const framework::Tensor* in, const framework::Tensor* scale,
+                  T max_range, framework::Tensor* out) {
+    auto in_e = framework::EigenVector<T>::Flatten(*in);
+    const T* scale_factor = scale->data<T>();
+    auto out_e = framework::EigenVector<T>::Flatten(*out);
+
+    auto& dev = *dev_ctx.eigen_device();
+    out_e.device(dev) = (scale_factor[0] / max_range) * in_e;
+  }
+};
+
+template struct DequantizeFunctor<platform::CPUDeviceContext, float>;
+template struct DequantizeFunctor<platform::CPUDeviceContext, double>;
+
 class FakeDequantizeMaxAbsOp : public framework::OperatorWithKernel {
  public:
-  FakeDequantizeMaxAbsOp(const std::string &type,
-                         const framework::VariableNameMap &inputs,
-                         const framework::VariableNameMap &outputs,
-                         const framework::AttributeMap &attrs)
+  FakeDequantizeMaxAbsOp(const std::string& type,
+                         const framework::VariableNameMap& inputs,
+                         const framework::VariableNameMap& outputs,
+                         const framework::AttributeMap& attrs)
       : OperatorWithKernel(type, inputs, outputs, attrs) {}
 
-  void InferShape(framework::InferShapeContext *ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("X"),
                    "Input(X) of FakeDequantizeMaxAbsOp should not be null.");
     PADDLE_ENFORCE(ctx->HasOutput("Out"),
@@ -42,21 +59,17 @@ class FakeDequantizeMaxAbsOpMaker : public framework::OpProtoAndCheckerMaker {
     AddInput("X",
              "(Tensor) The input with float-32/64 type is the "
              "low precision tensor.");
+    AddInput("Scale", "(float) The scale in quantization stage.");
     AddOutput("Out",
               "(Tensor) The output is the dequantized high "
               "precision tensor.");
-    AddAttr<int>("num_bits",
-                 "(int) `num_bits` is the quantization level bits, "
-                 "such as 2, 5, 8.");
-    AddAttr<float>("scale",
-                   "(float) The maximum absolute value of low precision tensor."
-                   "It is usually calculated by the fake_quantize_max_abs_op.");
+    AddAttr<float>("max_range", "(float) The max range in quantization stage.");
     AddComment(R"DOC(
 FakeDequantizeMaxAbsOp operator.
 
 This calculation is an opposite operation of FakeQuantizeMaxAbsOp:
 
-$$Out = \frac{scale*X}{2^{num_bits} - 1}$$
+$$Out = \frac{scale*X}{ max_range }$$
 
 )DOC");
   }

paddle/fluid/operators/fake_dequantize_op.cu

@@ -14,6 +14,42 @@ limitations under the License. */
 
 #include "paddle/fluid/operators/fake_dequantize_op.h"
 
+namespace paddle {
+namespace operators {
+
+template <typename T>
+__global__ void KeDequantize(const T* in, const T* scale, T max_range, int num,
+                             T* out) {
+  const int idx = threadIdx.x + blockIdx.x * blockDim.x;
+  if (idx < num) {
+    out[idx] = in[idx] * scale[0] / max_range;
+  }
+}
+
+template <typename T>
+struct DequantizeFunctor<platform::CUDADeviceContext, T> {
+  void operator()(const platform::CUDADeviceContext& dev_ctx,
+                  const framework::Tensor* in, const framework::Tensor* scale,
+                  T max_range, framework::Tensor* out) {
+    const T* in_data = in->data<T>();
+    const T* scale_factor = scale->data<T>();
+    T* out_data = out->mutable_data<T>(dev_ctx.GetPlace());
+
+    int num = in->numel();
+    int block = 512;
+    int grid = (num + block - 1) / block;
+
+    KeDequantize<T><<<grid, block, 0, dev_ctx.stream()>>>(
+        in_data, scale_factor, max_range, num, out_data);
+  }
+};
+
+template struct DequantizeFunctor<platform::CUDADeviceContext, float>;
+template struct DequantizeFunctor<platform::CUDADeviceContext, double>;
+
+}  // namespace operators
+}  // namespace paddle
+
 namespace ops = paddle::operators;
 using CUDA = paddle::platform::CUDADeviceContext;
 REGISTER_OP_CUDA_KERNEL(fake_dequantize_max_abs,

paddle/fluid/operators/fake_dequantize_op.h

@@ -19,22 +19,29 @@ limitations under the License. */
 
 namespace paddle {
 namespace operators {
+
+template <typename DeviceContext, typename T>
+struct DequantizeFunctor {
+  void operator()(const DeviceContext& dev_ctx, const framework::Tensor* in,
+                  const framework::Tensor* scale, T max_range,
+                  framework::Tensor* out);
+};
+
 template <typename DeviceContext, typename T>
 class FakeDequantizeMaxAbsKernel : public framework::OpKernel<T> {
  public:
   virtual void Compute(const framework::ExecutionContext& ctx) const {
     auto* in = ctx.Input<framework::Tensor>("X");
+    auto* scale = ctx.Input<framework::Tensor>("Scale");
     auto* out = ctx.Output<framework::Tensor>("Out");
-    out->mutable_data<T>(in->place());
 
-    int num_bits = ctx.Attr<int>("num_bits");
-    T scale = static_cast<T>(ctx.Attr<float>("scale"));
-    int range = std::pow(2, num_bits) - 1;
+    float max_range = ctx.Attr<float>("max_range");
+
+    auto& dev_ctx = ctx.template device_context<DeviceContext>();
+    out->mutable_data<T>(dev_ctx.GetPlace());
 
-    auto eigen_out = framework::EigenVector<T>::Flatten(*out);
-    auto eigen_in = framework::EigenVector<T>::Flatten(*in);
-    auto& dev = *ctx.template device_context<DeviceContext>().eigen_device();
-    eigen_out.device(dev) = (scale / range) * eigen_in;
+    DequantizeFunctor<DeviceContext, T>()(dev_ctx, in, scale,
+                                          static_cast<T>(max_range), out);
   }
 };
 

paddle/fluid/operators/fetch_barrier_op.cc

@@ -52,6 +52,8 @@ class FetchBarrierOp : public framework::OperatorBase {
 class FetchBarrierOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   void Make() {
+    AddOutput("Out", "(Any) Dummy outputs, used for control dependency")
+        .AsDuplicable();
     AddComment(R"DOC(
 SendBarrier operator
 

paddle/fluid/operators/fill_constant_op.cc

@@ -15,7 +15,6 @@ limitations under the License. */
 #include "paddle/fluid/framework/data_type.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/operators/math/math_function.h"
-#include "paddle/fluid/platform/device_context.h"
 
 namespace paddle {
 namespace operators {
@@ -41,19 +40,33 @@ class FillConstantOp : public framework::OperatorBase {
         static_cast<framework::proto::VarType::Type>(Attr<int>("dtype"));
     auto value = Attr<float>("value");
     auto force_cpu = Attr<bool>("force_cpu");
-    auto &out =
-        *scope.FindVar(Output("Out"))->GetMutable<framework::LoDTensor>();
-    out.Resize(framework::make_ddim(Attr<std::vector<int>>("shape")));
+
+    framework::Tensor *tensor = nullptr;
+
+    auto &out_var = *scope.FindVar(Output("Out"));
+
+    if (out_var.IsType<framework::LoDTensor>()) {
+      tensor = out_var.GetMutable<framework::LoDTensor>();
+      tensor->Resize(framework::make_ddim(Attr<std::vector<int>>("shape")));
+    } else if (out_var.IsType<framework::SelectedRows>()) {
+      tensor = out_var.GetMutable<framework::SelectedRows>()->mutable_value();
+      tensor->Resize(framework::make_ddim(Attr<std::vector<int>>("shape")));
+    } else {
+      PADDLE_THROW(
+          "fill constant op's output only"
+          "supports SelectedRows and LoDTensor");
+    }
+
     if (force_cpu) {
       auto cpu = platform::CPUPlace();
-      out.mutable_data(cpu, framework::ToTypeIndex(data_type));
+      tensor->mutable_data(cpu, framework::ToTypeIndex(data_type));
     } else {
-      out.mutable_data(dev_place, framework::ToTypeIndex(data_type));
+      tensor->mutable_data(dev_place, framework::ToTypeIndex(data_type));
     }
 
     platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
     auto &dev_ctx = *pool.Get(dev_place);
-    math::set_constant(dev_ctx, &out, value);
+    math::set_constant(dev_ctx, tensor, value);
   }
 };
 

paddle/fluid/operators/fusion_lstm_op.cc

@@ -15,10 +15,14 @@ limitations under the License. */
 #include "paddle/fluid/operators/fusion_lstm_op.h"
 #include <string>
 #include "paddle/fluid/operators/math/blas.h"
+#include "paddle/fluid/operators/math/cpu_vec.h"
 #include "paddle/fluid/operators/math/detail/activation_functions.h"
 #include "paddle/fluid/operators/math/fc_compute.h"
 #include "paddle/fluid/operators/math/lstm_compute.h"
 #include "paddle/fluid/operators/math/sequence2batch.h"
+#include "paddle/fluid/platform/cpu_info.h"
+
+DEFINE_bool(seq_mode, true, "Use sequence mode");
 
 namespace paddle {
 namespace operators {
@@ -98,7 +102,12 @@ void FusionLSTMOp::InferShape(framework::InferShapeContext* ctx) const {
   ctx->ShareLoD("X", "Hidden");
   ctx->ShareLoD("X", "Cell");
 
-  int xx_width = x_dims[1] > wx_dims[1] ? wx_dims[1] : x_dims[1];
+  int xx_width;
+  if (FLAGS_seq_mode) {
+    xx_width = wx_dims[1];
+  } else {
+    xx_width = x_dims[1] > wx_dims[1] ? wx_dims[1] : x_dims[1];
+  }
   ctx->SetOutputDim("XX", {x_dims[0], xx_width});
   ctx->ShareLoD("X", "XX");
 }
@@ -205,10 +214,138 @@ inline void ReorderInitState(const DeviceContext& ctx,
   row_shuffle(ctx, src, index_lod, dst, indexed_src);
 }
 
-template <typename DeviceContext, typename T>
+template <typename T>
 class FuisonLSTMKernel : public framework::OpKernel<T> {
  public:
-  void Compute(const framework::ExecutionContext& ctx) const override {
+  void SeqCompute(const framework::ExecutionContext& ctx) const {
+    using DeviceContext = paddle::platform::CPUDeviceContext;
+    auto* x = ctx.Input<LoDTensor>("X");
+    auto* h0 = ctx.Input<Tensor>("H0");
+    auto* c0 = ctx.Input<Tensor>("C0");
+    auto* wx = ctx.Input<Tensor>("WeightX");
+    auto* wh = ctx.Input<Tensor>("WeightH");
+    auto* bias = ctx.Input<Tensor>("Bias");
+
+    auto* xx = ctx.Output<LoDTensor>("XX");
+    auto* hidden_out = ctx.Output<LoDTensor>("Hidden");
+    auto* cell_out = ctx.Output<LoDTensor>("Cell");
+    bool is_reverse = ctx.Attr<bool>("is_reverse");
+
+    std::function<void(const int, const T *, T *)> act_gate, act_cell, act_cand;
+    auto& act_gate_str = ctx.Attr<std::string>("gate_activation");
+    auto& act_cell_str = ctx.Attr<std::string>("cell_activation");
+    auto& act_cand_str = ctx.Attr<std::string>("candidate_activation");
+    if (platform::jit::MayIUse(platform::jit::avx)) {
+      math::VecActivations<T, platform::jit::avx> act_functor;
+      act_gate = act_functor(act_gate_str);
+      act_cell = act_functor(act_cell_str);
+      act_cand = act_functor(act_cand_str);
+    } else {
+      math::VecActivations<T, platform::jit::isa_any> act_functor;
+      act_gate = act_functor(act_gate_str);
+      act_cell = act_functor(act_cell_str);
+      act_cand = act_functor(act_cand_str);
+    }
+
+    auto x_lod = x->lod();
+    auto x_dims = x->dims();    // T x M
+    auto wh_dims = wh->dims();  // D x 4D
+    const int total_T = x_dims[0];
+    const int N = x_lod[0].size() - 1;  // batch size
+    const int M = x_dims[1];            // x frame size
+    const int D = wh_dims[0];
+    const int D2 = D * 2;
+    const int D3 = D * 3;
+    const int D4 = wh_dims[1];
+
+    const T* x_data = x->data<T>();
+    const T* h0_data = h0 ? h0->data<T>() : NULL;
+    const T* c0_data = c0 ? c0->data<T>() : NULL;
+    const T* wx_data = wx->data<T>();
+    const T* wh_data = wh->data<T>();
+    T* xx_data = xx->mutable_data<T>(ctx.GetPlace());
+    T* hidden_out_data = hidden_out->mutable_data<T>(ctx.GetPlace());
+    T* cell_out_data = cell_out->mutable_data<T>(ctx.GetPlace());
+
+    auto blas = math::GetBlas<DeviceContext, T>(ctx);
+    math::FCCompute<DeviceContext, T>(blas, total_T, D4, M, x_data, wx_data,
+                                      xx_data, bias->data<T>());
+    int xx_offset = D4;
+    int gate_offset = D;
+    if (is_reverse) {
+      const int offset = (total_T - 1) * D;
+      xx_data = xx_data + offset * 4;
+      hidden_out_data = hidden_out_data + offset;
+      cell_out_data = cell_out_data + offset;
+      xx_offset = -D4;
+      gate_offset = -D;
+    }
+
+    auto move_step = [&]() {
+      xx_data = xx_data + xx_offset;
+      hidden_out_data = hidden_out_data + gate_offset;
+      cell_out_data = cell_out_data + gate_offset;
+    };
+
+    for (int i = 0; i < N; ++i) {
+      int bid = is_reverse ? N - 1 - i : i;
+      int seq_len = x_lod[0][bid + 1] - x_lod[0][bid];
+      const T* prev_cell_data = NULL;
+      const T* prev_hidden_data = NULL;
+      int tstart = 0;
+      if (h0_data) {
+        prev_hidden_data = h0_data + bid * D;
+        prev_cell_data = c0_data + bid * D;
+      } else {
+        // W_ch, W_ih, W_fh, W_oh
+        act_gate(D3, xx_data + D, xx_data + D);
+        act_cand(D, xx_data, xx_data);
+        // cell out= input*tilde
+        blas.VMUL(D, xx_data, xx_data + D, cell_out_data);
+        // hidden out= act_state(cellout) * outgate
+        act_cell(D, cell_out_data, xx_data + D2);
+        blas.VMUL(D, xx_data + D2, xx_data + D3, hidden_out_data);
+
+        // prev
+        prev_hidden_data = hidden_out_data;
+        prev_cell_data = cell_out_data;
+        tstart = 1;
+
+        move_step();
+      }
+      for (int step = tstart; step < seq_len; ++step) {
+        blas.GEMM(CblasNoTrans, CblasNoTrans, 1, D4, D, static_cast<T>(1),
+                  prev_hidden_data, D, wh_data, D4, static_cast<T>(1), xx_data,
+                  D4);
+
+        // W_ch, W_ih, W_fh, W_oh
+        act_gate(D3, xx_data + D, xx_data + D);
+        act_cand(D, xx_data, xx_data);
+
+        // a = forget * prev_cell
+        blas.VMUL(D, xx_data + D2, prev_cell_data, xx_data + D2);
+
+        // b = input * tilde
+        blas.VMUL(D, xx_data, xx_data + D, xx_data + D);
+
+        // cell out= a+b
+        blas.VADD(D, xx_data + D, xx_data + D2, cell_out_data);
+
+        // hidden out= act_state(cellout) * outgate
+        act_cell(D, cell_out_data, xx_data + D2);
+        blas.VMUL(D, xx_data + D2, xx_data + D3, hidden_out_data);
+
+        // prev
+        prev_hidden_data = hidden_out_data;
+        prev_cell_data = cell_out_data;
+
+        move_step();
+      }
+    }
+  }
+
+  void BatchCompute(const framework::ExecutionContext& ctx) const {
+    using DeviceContext = platform::CPUDeviceContext;
     auto* x = ctx.Input<LoDTensor>("X");
     auto* wx = ctx.Input<Tensor>("WeightX");
     auto* wh = ctx.Input<Tensor>("WeightH");
@@ -339,6 +476,13 @@ class FuisonLSTMKernel : public framework::OpKernel<T> {
     // restore the output cell state in LoDTensor from the batch cell
     to_seq(dev_ctx, batch_cell, cell_out);
   }
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    if (FLAGS_seq_mode) {
+      SeqCompute(ctx);
+    } else {
+      BatchCompute(ctx);
+    }
+  }
 };
 
 }  // namespace operators
@@ -348,7 +492,5 @@ namespace ops = paddle::operators;
 REGISTER_OPERATOR(fusion_lstm, ops::FusionLSTMOp, ops::FusionLSTMOpMaker,
                   paddle::framework::DefaultGradOpDescMaker<true>);
 
-REGISTER_OP_CPU_KERNEL(
-    fusion_lstm,
-    ops::FuisonLSTMKernel<paddle::platform::CPUDeviceContext, float>,
-    ops::FuisonLSTMKernel<paddle::platform::CPUDeviceContext, double>);
+REGISTER_OP_CPU_KERNEL(fusion_lstm, ops::FuisonLSTMKernel<float>,
+                       ops::FuisonLSTMKernel<double>);

paddle/fluid/operators/fusion_lstm_op.h

@@ -13,7 +13,6 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
-// #include <string>
 #include "paddle/fluid/framework/op_registry.h"
 
 namespace paddle {

paddle/fluid/operators/fusion_seqexpand_concat_fc_op.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/fluid/operators/fusion_seqexpand_concat_fc_op.h"
+#include <string>
+#include "paddle/fluid/operators/math/blas.h"
+#include "paddle/fluid/operators/math/cpu_vec.h"
+#include "paddle/fluid/operators/math/fc_compute.h"
+#include "paddle/fluid/platform/cpu_info.h"
+
+namespace paddle {
+namespace operators {
+
+void FusionSeqExpandConcatFCOp::InferShape(
+    framework::InferShapeContext* ctx) const {
+  PADDLE_ENFORCE_GT(
+      ctx->Inputs("X").size(), 1UL,
+      "Inputs(X) of FusionSeqExpandConcatFCOp should larger than 1.");
+  PADDLE_ENFORCE(
+      ctx->HasInput("FCWeight"),
+      "Input(FCWeight) of FusionSeqExpandConcatFCOp should not be null.");
+  PADDLE_ENFORCE(
+      ctx->HasOutput("Out"),
+      "Output(Out) of FusionSeqExpandConcatFCOp should not be null.");
+  PADDLE_ENFORCE(
+      ctx->HasOutput("FCOut"),
+      "Output(FCOut) of FusionSeqExpandConcatFCOp should not be null.");
+
+  auto ins_dims = ctx->GetInputsDim("X");
+  auto w_dims = ctx->GetInputDim("FCWeight");  // (M0+M1+M2+..) x D
+  PADDLE_ENFORCE_EQ(w_dims.size(), 2UL, "Input(FCWeight)'s rank must be 2.");
+  const int D = w_dims[1];
+  int sum = ins_dims[0][1];
+  for (size_t i = 1; i < ins_dims.size(); ++i) {
+    sum += ins_dims[i][1];
+  }
+  PADDLE_ENFORCE_EQ(sum, w_dims[0],
+                    "FC height should be sum of all inputs width.");
+  if (ctx->HasInput("FCBias")) {
+    auto b_dims = ctx->GetInputDim("FCBias");
+    PADDLE_ENFORCE_EQ(b_dims.size(), 2, "Input(FCBias)'s rank must be 2.");
+    PADDLE_ENFORCE_EQ(b_dims[0], 1, "FCBias shapes must be 1 * %d.", D);
+    PADDLE_ENFORCE_EQ(b_dims[1], D, "FCBias shapes must be 1 * %d.", D);
+  }
+
+  ctx->SetOutputDim("Out", {ins_dims[0][0], D});
+  // fcout should be reshape when run since can not get lod in infershape
+  // explicit share the ref lod
+  ctx->ShareLoD("X", "Out", 0);
+}
+
+framework::OpKernelType FusionSeqExpandConcatFCOp::GetExpectedKernelType(
+    const framework::ExecutionContext& ctx) const {
+  return framework::OpKernelType(
+      framework::ToDataType(ctx.MultiInput<LoDTensor>("X")[0]->type()),
+      ctx.device_context());
+}
+
+void FusionSeqExpandConcatFCOpMaker::Make() {
+  AddInput("X",
+           "(LoDTensor) input LodDTensors, the first one must be have ref lod "
+           "for sequence expand, and the rest input should have same lod.")
+      .AsDuplicable();
+  AddInput("FCWeight", "(Tensor) the weights of fc.");
+  AddInput("FCBias", "(Tensor, optional) the bias of fc.").AsDispensable();
+  AddOutput("Out", "(LoDTensor) Output LodTensor.");
+  AddOutput(
+      "FCOut",
+      "(Tensor) the intermediate tensor to keep the result of fc."
+      "Shape is (N x D), where N is the batch size, D is the output dim of fc")
+      .AsIntermediate();
+  AddAttr<std::string>("fc_activation",
+                       "(string, default: identity)"
+                       "The activation for the result of fc."
+                       "`identity` by default.")
+      .SetDefault("identity")
+      .InEnum({"sigmoid", "tanh", "relu", "identity"});
+  AddComment(R"DOC(
+Fusion Sequence expand + concat + fc Operator.
+
+All below conditions should be meet:
+
+The ref_level of seq_expand should be 0.
+
+The ref lod of seq_expand level is the first input of concat.
+
+The other inputs should have same lod and same batch size of ref lod.
+
+The seq len of other inputs should be 1.
+
+The concat axis should be 1.
+
+)DOC");
+}
+
+template <typename T>
+class FusionSeqExpandConcatFCOpKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    using DeviceContext = paddle::platform::CPUDeviceContext;
+    auto ins = ctx.MultiInput<LoDTensor>("X");
+    auto* w = ctx.Input<Tensor>("FCWeight");
+    auto* b = ctx.Input<Tensor>("FCBias");
+    auto* out = ctx.Output<LoDTensor>("Out");
+    auto* fc_out = ctx.Output<Tensor>("FCOut");
+
+    auto* ref_in = ins[0];
+    auto ref_lod = ref_in->lod();
+    auto in1_lod = ins[1]->lod();
+    auto ref_dims = ref_in->dims();  // T x M0
+    auto in1_dims = ins[1]->dims();  // N x M1
+    auto w_dims = w->dims();
+    const int N = ref_lod[0].size() - 1;
+    const int total_T = ref_dims[0];
+    const int M0 = ref_dims[1];
+    const int M1 = in1_dims[1];
+    const int D = w_dims[1];
+
+    // some check and fcout should be reshape here
+    // since infershape can not get lod info
+    PADDLE_ENFORCE_EQ(ref_lod.size(), 1UL, "Only support input lod size is 1.");
+    PADDLE_ENFORCE_EQ(in1_lod.size(), 1UL, "Only support input lod size is 1.");
+    PADDLE_ENFORCE_EQ(in1_lod[0].size() - 1, N,
+                      "Batch size of all inputs should be equal.");
+    PADDLE_ENFORCE_EQ(in1_lod[0][N], N,
+                      "Seq_length of other inputs should be 1.");
+    PADDLE_ENFORCE_EQ(in1_dims[0], N, "input height should be batch size.");
+    for (size_t i = 2; i < ins.size(); ++i) {
+      PADDLE_ENFORCE_EQ(ins[i]->dims()[0], N,
+                        "All other inputs height should be equal");
+      PADDLE_ENFORCE_EQ(ins[i]->lod(), in1_lod,
+                        "All other inputs should have same lod");
+    }
+    fc_out->Resize({N, D});
+
+    std::function<void(const int, const T*, T*)> fc_act;
+    auto& fc_act_str = ctx.Attr<std::string>("fc_activation");
+    if (platform::jit::MayIUse(platform::jit::avx)) {
+      math::VecActivations<T, platform::jit::avx> act_functor;
+      fc_act = act_functor(fc_act_str);
+    } else {
+      math::VecActivations<T, platform::jit::isa_any> act_functor;
+      fc_act = act_functor(fc_act_str);
+    }
+
+    const T* ref_in_data = ref_in->data<T>();
+    const T* in1_data = ins[1]->data<T>();
+    const T* w_data = w->data<T>();
+    T* out_data = out->mutable_data<T>(ctx.GetPlace());
+    T* fc_out_data = fc_out->mutable_data<T>(ctx.GetPlace());
+
+    auto blas = math::GetBlas<DeviceContext, T>(ctx);
+    math::FCCompute<DeviceContext, T>(blas, total_T, D, M0, ref_in_data, w_data,
+                                      out_data, b ? b->data<T>() : NULL);
+    w_data = w_data + M0 * D;
+    // first write on
+    blas.MatMul(N, D, M1, in1_data, w_data, fc_out_data);
+    w_data = w_data + M1 * D;
+    for (size_t i = 2; i < ins.size(); ++i) {
+      // add on
+      const T* in_data = ins[i]->data<T>();
+      const int K = ins[i]->dims()[1];
+      blas.GEMM(CblasNoTrans, CblasNoTrans, N, D, K, static_cast<T>(1), in_data,
+                K, w_data, D, static_cast<T>(1), fc_out_data, D);
+      w_data = w_data + K * D;
+    }
+    T* cur_out_data = out_data;
+    for (int i = 0; i < N; ++i) {
+      int seq_len = ref_lod[0][i + 1] - ref_lod[0][i];
+      T* src = fc_out_data + i * D;
+      for (int step = 0; step < seq_len; ++step) {
+        blas.VADD(D, cur_out_data, src, cur_out_data);
+        cur_out_data = cur_out_data + D;
+      }
+    }
+    fc_act(total_T * D, out_data, out_data);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OPERATOR(fusion_seqexpand_concat_fc, ops::FusionSeqExpandConcatFCOp,
+                  ops::FusionSeqExpandConcatFCOpMaker,
+                  paddle::framework::DefaultGradOpDescMaker<true>);
+
+REGISTER_OP_CPU_KERNEL(fusion_seqexpand_concat_fc,
+                       ops::FusionSeqExpandConcatFCOpKernel<float>,
+                       ops::FusionSeqExpandConcatFCOpKernel<double>);

paddle/fluid/operators/fusion_seqexpand_concat_fc_op.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 "paddle/fluid/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+using LoDTensor = framework::LoDTensor;
+using Tensor = framework::Tensor;
+
+class FusionSeqExpandConcatFCOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override;
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override;
+};
+
+class FusionSeqExpandConcatFCOpMaker
+    : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() override;
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/listen_and_serv_op.cc

@@ -165,12 +165,13 @@ void ListenAndServOp::RunSyncLoop(
                           recv_scope);
     VLOG(2) << "run all blocks spent " << GetTimestamp() - ts << "(ms)";
 
-    rpc_service_->SetCond(distributed::kRequestGet);
-    rpc_service_->WaitBarrier(distributed::kRequestGet);
-    rpc_service_->ResetBarrierCounter();
     // reset received sparse vars to avoid reuse it in the next mini-batch
     dynamic_cast<distributed::RequestSendHandler *>(request_send_handler_.get())
         ->ResetSparseVarRecorder();
+
+    rpc_service_->SetCond(distributed::kRequestGet);
+    rpc_service_->WaitBarrier(distributed::kRequestGet);
+    rpc_service_->ResetBarrierCounter();
   }  // while(true)
 }
 

paddle/fluid/operators/math/CMakeLists.txt

@@ -65,3 +65,4 @@ if(WITH_GPU)
     nv_test(selected_rows_functor_gpu_test SRCS selected_rows_functor_test.cu DEPS selected_rows_functor math_function)
 endif()
 cc_test(concat_test SRCS concat_test.cc DEPS concat)
+cc_test(cpu_vec_test SRCS cpu_vec_test.cc DEPS blas cpu_info)

paddle/fluid/operators/math/blas.h

@@ -90,6 +90,11 @@ class Blas {
   void GEMM(bool transA, bool transB, int M, int N, int K, T alpha, const T* A,
             int lda, const T* B, int ldb, T beta, T* C, int ldc) const;
 
+  template <typename T>
+  void GEMM(CBLAS_TRANSPOSE transA, CBLAS_TRANSPOSE transB, int M, int N, int K,
+            T alpha, const T* A, int lda, const T* B, int ldb, T beta, T* C,
+            int ldc) const;
+
 #ifdef PADDLE_WITH_MKLML
   template <typename T>
   T* GEMM_ALLOC(const CBLAS_IDENTIFIER id, const int M, const int N,
@@ -109,6 +114,10 @@ class Blas {
   void GEMM_FREE(T* data) const;
 #endif
 
+  template <typename T>
+  void MatMul(const int M, const int N, const int K, const T* A, const T* B,
+              T* C) const;
+
   template <typename T>
   void MatMul(const framework::Tensor& mat_a, bool trans_a,
               const framework::Tensor& mat_b, bool trans_b, T alpha,
@@ -140,10 +149,19 @@ class Blas {
   template <typename T>
   void VCOPY(int n, const T* x, T* y) const;
 
+  template <typename T>
+  void VEXP(int n, const T* x, T* y) const;
+
   template <typename T>
   void GEMV(bool trans_a, int M, int N, T alpha, const T* A, const T* B, T beta,
             T* C) const;
 
+  template <typename T>
+  T DOT(int n, const T* x, const T* y) const;
+
+  template <typename T>
+  void SCAL(int n, const T a, T* x) const;
+
   template <typename T>
   void BatchedGEMM(CBLAS_TRANSPOSE transA, CBLAS_TRANSPOSE transB, int M, int N,
                    int K, T alpha, const T* A, const T* B, T beta, T* C,
@@ -215,11 +233,26 @@ class BlasT : private Blas<DeviceContext> {
     Base()->template VCOPY<T>(args...);
   }
 
+  template <typename... ARGS>
+  void VEXP(ARGS... args) const {
+    Base()->template VEXP<T>(args...);
+  }
+
   template <typename... ARGS>
   void GEMV(ARGS... args) const {
     Base()->template GEMV<T>(args...);
   }
 
+  template <typename... ARGS>
+  T DOT(ARGS... args) const {
+    return Base()->template DOT<T>(args...);
+  }
+
+  template <typename... ARGS>
+  void SCAL(ARGS... args) const {
+    Base()->template SCAL<T>(args...);
+  }
+
   template <typename... ARGS>
   void BatchedGEMM(ARGS... args) const {
     Base()->template BatchedGEMM<T>(args...);

paddle/fluid/operators/math/blas_impl.h

@@ -73,6 +73,16 @@ struct CBlas<float> {
     platform::dynload::cblas_sgemv(args...);
   }
 
+  template <typename... ARGS>
+  static float DOT(ARGS... args) {
+    return platform::dynload::cblas_sdot(args...);
+  }
+
+  template <typename... ARGS>
+  static void SCAL(ARGS... args) {
+    platform::dynload::cblas_sscal(args...);
+  }
+
   template <typename... ARGS>
   static void GEMM_BATCH(ARGS... args) {
     platform::dynload::cblas_sgemm_batch(args...);
@@ -87,6 +97,11 @@ struct CBlas<float> {
   static void VMUL(ARGS... args) {
     platform::dynload::vsMul(args...);
   }
+
+  template <typename... ARGS>
+  static void VEXP(ARGS... args) {
+    platform::dynload::vsExp(args...);
+  }
 };
 
 template <>
@@ -138,6 +153,16 @@ struct CBlas<double> {
     platform::dynload::cblas_dgemv(args...);
   }
 
+  template <typename... ARGS>
+  static double DOT(ARGS... args) {
+    return platform::dynload::cblas_ddot(args...);
+  }
+
+  template <typename... ARGS>
+  static void SCAL(ARGS... args) {
+    platform::dynload::cblas_dscal(args...);
+  }
+
   template <typename... ARGS>
   static void GEMM_BATCH(ARGS... args) {
     platform::dynload::cblas_dgemm_batch(args...);
@@ -152,6 +177,11 @@ struct CBlas<double> {
   static void VMUL(ARGS... args) {
     platform::dynload::vdMul(args...);
   }
+
+  template <typename... ARGS>
+  static void VEXP(ARGS... args) {
+    platform::dynload::vdExp(args...);
+  }
 };
 
 #else
@@ -210,6 +240,9 @@ struct CBlas<platform::float16> {
     PADDLE_THROW("float16 SMM_GEMM not supported on CPU");
   }
   static void VMUL(...) { PADDLE_THROW("float16 VMUL not supported on CPU"); }
+  static void VEXP(...) { PADDLE_THROW("float16 VEXP not supported on CPU"); }
+  static void DOT(...) { PADDLE_THROW("float16 DOT not supported on CPU"); };
+  static void SCAL(...) { PADDLE_THROW("float16 SCAL not supported on CPU"); };
 #ifdef PADDLE_WITH_MKLML
   static void GEMM_BATCH(...) {
     PADDLE_THROW("float16 GEMM_BATCH not supported on CPU");
@@ -217,64 +250,6 @@ struct CBlas<platform::float16> {
 #endif
 };
 
-template <typename T>
-inline bool UseXSMM(const int &m, const int &n, const int &k, bool transa,
-                    bool transb, const T &alpha, const T &beta) {
-#ifdef PADDLE_WITH_LIBXSMM
-  // Refer to https://github.com/hfp/libxsmm/blob/master/README.md
-  // But the threshold is custom
-  constexpr int LIBXSMM_THRESHOLD = 20 * 20 * 20;
-  if (m * n * k > LIBXSMM_THRESHOLD || transa || transb ||
-      std::abs<T>(alpha - static_cast<T>(1) >
-                  std::numeric_limits<T>::epsilon()) ||
-      std::abs<T>(beta) > std::numeric_limits<T>::epsilon()) {
-    return false;
-  } else {
-    return true;
-  }
-#endif
-  return false;
-}
-
-template <>
-inline bool UseXSMM<platform::float16>(const int &m, const int &n, const int &k,
-                                       bool transa, bool transb,
-                                       const platform::float16 &alpha,
-                                       const platform::float16 &beta) {
-  return false;
-}
-
-template <typename T>
-inline void GEMM_WARP(CBLAS_ORDER order, CBLAS_TRANSPOSE transA,
-                      CBLAS_TRANSPOSE transB, int M, int N, int K, T alpha,
-                      const T *A, int lda, const T *B, int ldb, T beta, T *C,
-                      int ldc) {
-#ifdef PADDLE_WITH_LIBXSMM
-  if (UseXSMM<T>(M, N, K, transA != CblasNoTrans, transB != CblasNoTrans, alpha,
-                 beta)) {
-    // Note: SMM use ColMajor
-    const char transa = 'N';
-    const char transb = 'N';
-    CBlas<T>::SMM_GEMM(&transa, &transb, &N, &M, &K, &alpha, B, &ldb, A, &lda,
-                       &beta, C, &ldc);
-    return;
-  }
-#endif
-
-#ifdef PADDLE_MKL_SPLIT_GEMM
-  constexpr int bs = 2;
-  if (M % bs == 0 && transA == CblasNoTrans && transB == CblasNoTrans) {
-    for (int off = 0; off < M; off += bs) {
-      CBlas<T>::GEMM(CblasRowMajor, CblasNoTrans, CblasNoTrans, bs, N, K, alpha,
-                     A + off * lda, lda, B, ldb, beta, C + off * ldb, ldc);
-    }
-    return;
-  }
-#endif
-  CBlas<T>::GEMM(CblasRowMajor, transA, transB, M, N, K, alpha, A, lda, B, ldb,
-                 beta, C, ldc);
-}
-
 #ifdef PADDLE_WITH_MKLML
 template <>
 template <typename T>
@@ -319,8 +294,8 @@ void Blas<platform::CPUDeviceContext>::GEMM(CBLAS_TRANSPOSE transA,
   int lda = (transA == CblasNoTrans) ? K : M;
   int ldb = (transB == CblasNoTrans) ? N : K;
   int ldc = N;
-  GEMM_WARP<T>(CblasRowMajor, transA, transB, M, N, K, alpha, A, lda, B, ldb,
-               beta, C, ldc);
+  CBlas<T>::GEMM(CblasRowMajor, transA, transB, M, N, K, alpha, A, lda, B, ldb,
+                 beta, C, ldc);
 }
 
 template <>
@@ -329,9 +304,20 @@ void Blas<platform::CPUDeviceContext>::GEMM(bool transA, bool transB, int M,
                                             int N, int K, T alpha, const T *A,
                                             int lda, const T *B, int ldb,
                                             T beta, T *C, int ldc) const {
-  GEMM_WARP<T>(CblasRowMajor, transA == false ? CblasNoTrans : CblasTrans,
-               transB == false ? CblasNoTrans : CblasTrans, M, N, K, alpha, A,
-               lda, B, ldb, beta, C, ldc);
+  CBlas<T>::GEMM(CblasRowMajor, transA == false ? CblasNoTrans : CblasTrans,
+                 transB == false ? CblasNoTrans : CblasTrans, M, N, K, alpha, A,
+                 lda, B, ldb, beta, C, ldc);
+}
+
+template <>
+template <typename T>
+void Blas<platform::CPUDeviceContext>::GEMM(CBLAS_TRANSPOSE transA,
+                                            CBLAS_TRANSPOSE transB, int M,
+                                            int N, int K, T alpha, const T *A,
+                                            int lda, const T *B, int ldb,
+                                            T beta, T *C, int ldc) const {
+  CBlas<T>::GEMM(CblasRowMajor, transA, transB, M, N, K, alpha, A, lda, B, ldb,
+                 beta, C, ldc);
 }
 
 template <typename DeviceContext>
@@ -399,6 +385,47 @@ void Blas<platform::CPUDeviceContext>::VMUL(int n, const T *x, const T *y,
 #endif
 }
 
+template <>
+template <typename T>
+void Blas<platform::CPUDeviceContext>::VEXP(int n, const T *x, T *y) const {
+#ifdef PADDLE_WITH_MKLML
+  CBlas<T>::VEXP(n, x, y);
+#else
+  // try to find if openblas support vexp
+  for (int i = 0; i < n; ++i) {
+    y[i] = std::exp(x[i]);
+  }
+#endif
+}
+
+template <>
+template <typename T>
+T Blas<platform::CPUDeviceContext>::DOT(int n, const T *x, const T *y) const {
+#ifdef PADDLE_WITH_MKLML
+  return CBlas<T>::DOT(n, x, 1, y, 1);
+#else
+  // try to find if openblas support cblas_dot
+  T sum = 0;
+  for (int i = 0; i < n; ++i) {
+    sum += x[i] * y[i];
+  }
+  return sum;
+#endif
+}
+
+template <>
+template <typename T>
+void Blas<platform::CPUDeviceContext>::SCAL(int n, const T a, T *x) const {
+#ifdef PADDLE_WITH_MKLML
+  CBlas<T>::SCAL(n, a, x, 1);
+#else
+  // try to find if openblas support cblas_scal
+  for (int i = 0; i < n; ++i) {
+    x[i] = a * x[i];
+  }
+#endif
+}
+
 template <>
 template <typename T>
 void Blas<platform::CPUDeviceContext>::GEMV(bool trans_a, int M, int N, T alpha,
@@ -440,6 +467,42 @@ void Blas<platform::CPUDeviceContext>::BatchedGEMM(
 #endif
 }
 
+template <typename DeviceContext>
+template <typename T>
+void Blas<DeviceContext>::MatMul(const int M, const int N, const int K,
+                                 const T *A, const T *B, T *C) const {
+  this->template GEMM<T>(CblasRowMajor, CblasNoTrans, CblasNoTrans, M, N, K,
+                         static_cast<T>(1), A, K, B, N, static_cast<T>(0), C,
+                         N);
+}
+
+template <>
+template <typename T>
+void Blas<platform::CPUDeviceContext>::MatMul(const int M, const int N,
+                                              const int K, const T *A,
+                                              const T *B, T *C) const {
+#ifdef PADDLE_WITH_LIBXSMM
+  // Refer to https://github.com/hfp/libxsmm/blob/master/README.md
+  // But the threshold is custom constexpr int LIBXSMM_THRESHOLD = 20 * 20 * 20;
+
+  // Since the matrix is very small,
+  // so the unit of calculation is already very fast,
+  // and the if( M*N*K < LIBXSMM_THRESHOLD) would be overhead,
+  // use xsmm directly.
+  // Note: SMM use ColMajor
+  const char transa = 'N';
+  const char transb = 'N';
+  const T alpha = static_cast<T>(1);
+  const T beta = static_cast<T>(0);
+  CBlas<T>::SMM_GEMM(&transa, &transb, &N, &M, &K, &alpha, B, &N, A, &K, &beta,
+                     C, &N);
+  return;
+#endif
+
+  CBlas<T>::GEMM(CblasRowMajor, CblasNoTrans, CblasNoTrans, M, N, K,
+                 static_cast<T>(1), A, K, B, N, static_cast<T>(0), C, N);
+}
+
 template <typename DeviceContext>
 template <typename T>
 void Blas<DeviceContext>::MatMul(const framework::Tensor &mat_a,

paddle/fluid/operators/math/concat.cc

@@ -48,16 +48,16 @@ class ConcatFunctor<platform::CPUDeviceContext, T> {
     auto cpu_place = boost::get<platform::CPUPlace>(context.GetPlace());
 
     // computation
-    for (int k = 0; k < out_rows; ++k) {
-      T* dst_ptr = output->data<T>() + k * out_cols;
-      int col_idx = 0;
-      for (int j = 0; j < num; ++j) {
-        int col_len = input_cols[j];
-        const T* src_prt = input[j].data<T>() + k * col_len;
-        memory::Copy(cpu_place, dst_ptr + col_idx, cpu_place, src_prt,
-                     sizeof(T) * col_len);
-        col_idx += col_len;
+    auto output_data = output->data<T>();
+    int col_idx = 0;
+    for (int j = 0; j < num; ++j) {
+      int col_len = input_cols[j];
+      auto input_data = input[j].data<T>();
+      for (int k = 0; k < out_rows; ++k) {
+        memory::Copy(cpu_place, output_data + k * out_cols + col_idx, cpu_place,
+                     input_data + k * col_len, sizeof(T) * col_len);
       }
+      col_idx += col_len;
     }
   }
 };
@@ -71,7 +71,7 @@ class ConcatGradFunctor<platform::CPUDeviceContext, T> {
  public:
   void operator()(const platform::CPUDeviceContext& context,
                   const framework::Tensor& input,
-                  const std::vector<const framework::Tensor*>& ref_inputs,
+                  const std::vector<const framework::LoDTensor*>& ref_inputs,
                   const int axis, std::vector<framework::Tensor*>* outputs) {
     // TODO(zcd): Add input data validity checking
     size_t num = outputs->size();

paddle/fluid/operators/math/concat.cu

@@ -177,6 +177,9 @@ class ConcatFunctor<platform::CUDADeviceContext, T> {
           dev_ins_data, dev_ins_col_data, static_cast<int>(inputs_col.size()),
           out_row, out_col, output->data<T>());
     }
+    // Wait() must be called because `inputs_data` may be destructed before
+    // kernel ends
+    context.Wait();
   }
 };
 
@@ -189,7 +192,7 @@ class ConcatGradFunctor<platform::CUDADeviceContext, T> {
  public:
   void operator()(const platform::CUDADeviceContext& context,
                   const framework::Tensor& input,
-                  const std::vector<const framework::Tensor*>& ref_inputs,
+                  const std::vector<const framework::LoDTensor*>& ref_inputs,
                   const int axis, std::vector<framework::Tensor*>* outputs) {
     // TODO(zcd): Add input data validity checking
     int o_num = outputs->size();
@@ -252,6 +255,9 @@ class ConcatGradFunctor<platform::CUDADeviceContext, T> {
           input.data<T>(), in_row, in_col, dev_outs_col_data,
           static_cast<int>(outputs_cols.size()), dev_out_gpu_data);
     }
+    // Wait() must be called because `outputs_data` may be destructed before
+    // kernel ends
+    context.Wait();
   }
 };
 

paddle/fluid/operators/math/concat.h

@@ -15,7 +15,7 @@ limitations under the License. */
 #pragma once
 #include <vector>
 #include "paddle/fluid/framework/data_type.h"
-#include "paddle/fluid/framework/tensor.h"
+#include "paddle/fluid/framework/lod_tensor.h"
 
 namespace paddle {
 namespace operators {
@@ -57,7 +57,7 @@ template <typename DeviceContext, typename T>
 class ConcatGradFunctor {
  public:
   void operator()(const DeviceContext& context, const framework::Tensor& input,
-                  const std::vector<const framework::Tensor*>& ref_inputs,
+                  const std::vector<const framework::LoDTensor*>& ref_inputs,
                   const int axis, std::vector<framework::Tensor*>* outputs);
 };
 

paddle/fluid/operators/math/cpu_vec.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+#include <cmath>
+#include <string>
+#include "paddle/fluid/platform/cpu_info.h"
+#ifdef __AVX__
+#include <immintrin.h>
+#endif
+
+#ifdef PADDLE_WITH_MKLML
+#include "paddle/fluid/platform/dynload/mklml.h"
+#endif
+
+namespace paddle {
+namespace operators {
+namespace math {
+
+#define SIGMOID_THRESHOLD_MIN -40.0
+#define SIGMOID_THRESHOLD_MAX 13.0
+
+#define AVX_FLOAT_BLOCK 8
+#define AVX_DOUBLE_BLOCK 4
+#define AVX2_FLOAT_BLOCK 8
+#define AVX2_DOUBLE_BLOCK 4
+#define AVX512_FLOAT_BLOCK 16
+#define AVX512_DOUBLE_BLOCK 8
+
+template <typename T>
+inline void vec_exp(const int n, const T* x, T* y) {
+  for (int i = 0; i < n; ++i) {
+    y[i] = std::exp(x[i]);
+  }
+}
+
+template <typename T>
+inline void vec_scal(const int n, const T a, T* x) {
+  for (int i = 0; i < n; ++i) {
+    x[i] = a * x[i];
+  }
+}
+
+#ifdef PADDLE_WITH_MKLML
+template <>
+inline void vec_exp<float>(const int n, const float* x, float* y) {
+  platform::dynload::vsExp(n, x, y);
+}
+
+template <>
+inline void vec_exp<double>(const int n, const double* x, double* y) {
+  platform::dynload::vdExp(n, x, y);
+}
+
+template <>
+inline void vec_scal<float>(const int n, const float a, float* x) {
+  platform::dynload::cblas_sscal(n, a, x, 1);
+}
+
+template <>
+inline void vec_scal<double>(const int n, const double a, double* x) {
+  platform::dynload::cblas_dscal(n, a, x, 1);
+}
+#endif
+
+// MKL scal only support inplace, choose this if src and dst are not equal
+template <typename T, platform::jit::cpu_isa_t isa = platform::jit::isa_any>
+inline void vec_scal(const int n, const T a, const T* x, T* y) {
+  for (int i = 0; i < n; ++i) {
+    y[i] = a * x[i];
+  }
+}
+
+template <>
+inline void vec_scal<float, platform::jit::avx>(const int n, const float a,
+                                                const float* x, float* y) {
+#ifdef __AVX__
+  constexpr int block = AVX_FLOAT_BLOCK;
+  if (n < block) {
+    vec_scal<float, platform::jit::isa_any>(n, a, x, y);
+    return;
+  }
+  const int rest = n % block;
+  const int end = n - rest;
+  int i = 0;
+  __m256 scalar = _mm256_set1_ps(a);
+  __m256 tmp;
+#define MOVE_ONE_STEP               \
+  tmp = _mm256_loadu_ps(x + i);     \
+  tmp = _mm256_mul_ps(tmp, scalar); \
+  _mm256_storeu_ps(y + i, tmp)
+  for (i = 0; i < end; i += block) {
+    MOVE_ONE_STEP;
+  }
+#undef MOVE_ONE_STEP
+  if (rest == 0) {
+    return;
+  }
+  // can not continue move step if src and dst are inplace
+  for (i = n - rest; i < n; ++i) {
+    y[i] = a * x[i];
+  }
+#else
+  vec_scal<float, platform::jit::isa_any>(n, a, x, y);
+#endif
+}
+
+template <>
+inline void vec_scal<float, platform::jit::avx2>(const int n, const float a,
+                                                 const float* x, float* y) {
+  vec_scal<float, platform::jit::avx>(n, a, x, y);
+}
+
+template <>
+inline void vec_scal<float, platform::jit::avx512_common>(const int n,
+                                                          const float a,
+                                                          const float* x,
+                                                          float* y) {
+  // TODO(TJ): enable me
+  vec_scal<float, platform::jit::avx2>(n, a, x, y);
+}
+
+template <typename T, platform::jit::cpu_isa_t isa = platform::jit::isa_any>
+inline void vec_add_bias(const int n, const T a, const T* x, T* y) {
+  for (int i = 0; i < n; ++i) {
+    y[i] = x[i] + a;
+  }
+}
+
+template <>
+inline void vec_add_bias<float, platform::jit::avx>(const int n, const float a,
+                                                    const float* x, float* y) {
+#ifdef __AVX__
+  constexpr int block = AVX_FLOAT_BLOCK;
+  if (n < block) {
+    vec_add_bias<float, platform::jit::isa_any>(n, a, x, y);
+    return;
+  }
+  const int rest = n % block;
+  const int end = n - rest;
+  int i = 0;
+  __m256 bias = _mm256_set1_ps(a);
+  __m256 tmp;
+#define MOVE_ONE_STEP             \
+  tmp = _mm256_loadu_ps(x + i);   \
+  tmp = _mm256_add_ps(tmp, bias); \
+  _mm256_storeu_ps(y + i, tmp)
+  for (i = 0; i < end; i += block) {
+    MOVE_ONE_STEP;
+  }
+#undef MOVE_ONE_STEP
+  if (rest == 0) {
+    return;
+  }
+  // can not continue move step if src and dst are inplace
+  for (i = n - rest; i < n; ++i) {
+    y[i] = x[i] + a;
+  }
+#else
+  vec_add_bias<float, platform::jit::isa_any>(n, a, x, y);
+#endif
+}
+
+template <>
+inline void vec_add_bias<float, platform::jit::avx2>(const int n, const float a,
+                                                     const float* x, float* y) {
+  vec_add_bias<float, platform::jit::avx>(n, a, x, y);
+}
+
+template <>
+inline void vec_add_bias<float, platform::jit::avx512_common>(const int n,
+                                                              const float a,
+                                                              const float* x,
+                                                              float* y) {
+  // TODO(TJ): enable me
+  vec_add_bias<float, platform::jit::avx2>(n, a, x, y);
+}
+
+template <typename T, platform::jit::cpu_isa_t isa = platform::jit::isa_any>
+inline void vec_identity(const int n, const T* x, T* y) {
+  // do nothing
+  return;
+}
+
+template <typename T, platform::jit::cpu_isa_t isa = platform::jit::isa_any>
+inline void vec_sigmoid(const int n, const T* x, T* y) {
+  const T min = SIGMOID_THRESHOLD_MIN;
+  const T max = SIGMOID_THRESHOLD_MAX;
+  for (int i = 0; i < n; ++i) {
+    y[i] = (x[i] < min) ? min : ((x[i] > max) ? max : x[i]);
+    y[i] = static_cast<T>(0) - y[i];
+  }
+  vec_exp<T>(n, y, y);
+  for (int i = 0; i < n; ++i) {
+    y[i] = static_cast<T>(1) / (static_cast<T>(1) + y[i]);
+  }
+}
+
+template <>
+inline void vec_sigmoid<float, platform::jit::avx>(const int n, const float* x,
+                                                   float* y) {
+#ifdef __AVX__
+  constexpr int block = AVX_FLOAT_BLOCK;
+  if (n < block) {
+    vec_sigmoid<float, platform::jit::isa_any>(n, x, y);
+    return;
+  }
+  const int rest = n % block;
+  const int end = n - rest;
+  int i = 0;
+  __m256 max = _mm256_set1_ps(SIGMOID_THRESHOLD_MAX);
+  __m256 min = _mm256_set1_ps(SIGMOID_THRESHOLD_MIN);
+  __m256 zeros = _mm256_setzero_ps();
+  __m256 tmp;
+#define MOVE_ONE_STEP              \
+  tmp = _mm256_loadu_ps(x + i);    \
+  tmp = _mm256_max_ps(tmp, min);   \
+  tmp = _mm256_min_ps(tmp, max);   \
+  tmp = _mm256_sub_ps(zeros, tmp); \
+  _mm256_storeu_ps(y + i, tmp)
+  for (i = 0; i < end; i += block) {
+    MOVE_ONE_STEP;
+  }
+#undef MOVE_ONE_STEP
+  if (rest != 0) {
+    // can not continue move step since the src and dst address could be equal
+    const float xmin = SIGMOID_THRESHOLD_MIN;
+    const float xmax = SIGMOID_THRESHOLD_MAX;
+    for (i = n - rest; i < n; ++i) {
+      y[i] = 0.f - ((x[i] < xmin) ? xmin : ((x[i] > xmax) ? xmax : x[i]));
+    }
+  }
+
+  vec_exp<float>(n, y, y);
+
+  __m256 ones = _mm256_set1_ps(1.0f);
+#define MOVE_ONE_STEP             \
+  tmp = _mm256_loadu_ps(y + i);   \
+  tmp = _mm256_add_ps(ones, tmp); \
+  tmp = _mm256_div_ps(ones, tmp); \
+  _mm256_storeu_ps(y + i, tmp)
+  for (i = 0; i < end; i += block) {
+    MOVE_ONE_STEP;
+  }
+#undef MOVE_ONE_STEP
+  if (rest == 0) {
+    return;
+  }
+  // can not continue move step
+  for (i = n - rest; i < n; ++i) {
+    y[i] = 1.f / (1.f + y[i]);
+  }
+#else
+  vec_sigmoid<float, platform::jit::isa_any>(n, x, y);
+#endif
+}
+
+template <>
+inline void vec_sigmoid<float, platform::jit::avx2>(const int n, const float* x,
+                                                    float* y) {
+  vec_sigmoid<float, platform::jit::avx>(n, x, y);
+}
+
+template <>
+inline void vec_sigmoid<float, platform::jit::avx512_common>(const int n,
+                                                             const float* x,
+                                                             float* y) {
+  // TODO(TJ): enable me
+  vec_sigmoid<float, platform::jit::avx2>(n, x, y);
+}
+
+template <typename T, platform::jit::cpu_isa_t isa = platform::jit::isa_any>
+inline void vec_tanh(const int n, const T* x, T* y) {
+  vec_scal<T, isa>(n, static_cast<T>(2), x, y);
+  vec_sigmoid<T, isa>(n, y, y);
+  vec_scal<T>(n, static_cast<T>(2), y);
+  vec_add_bias<T, isa>(n, static_cast<T>(-1), y, y);
+}
+
+// TODO(TJ): make relu clip
+template <typename T, platform::jit::cpu_isa_t isa = platform::jit::isa_any>
+inline void vec_relu(const int n, const T* x, T* y) {
+  for (int i = 0; i < n; ++i) {
+    y[i] = x[i] > 0 ? x[i] : 0;
+  }
+}
+
+template <>
+inline void vec_relu<float, platform::jit::avx>(const int n, const float* x,
+                                                float* y) {
+#ifdef __AVX__
+  constexpr int block = AVX_FLOAT_BLOCK;
+  if (n < block * 4) {
+    vec_relu<float, platform::jit::isa_any>(n, x, y);
+    return;
+  }
+
+  const int rest = n % block;
+  const int end = n - rest;
+  int i = 0;
+  __m256 zeros = _mm256_setzero_ps();
+  __m256 tmp;
+#define MOVE_ONE_STEP              \
+  tmp = _mm256_loadu_ps(x + i);    \
+  tmp = _mm256_max_ps(tmp, zeros); \
+  _mm256_storeu_ps(y + i, tmp)
+  for (i = 0; i < end; i += block) {
+    MOVE_ONE_STEP;
+  }
+  if (rest == 0) {
+    return;
+  }
+  i = n - block;
+  MOVE_ONE_STEP;
+#undef MOVE_ONE_STEP
+
+#else
+  vec_relu<float, platform::jit::isa_any>(n, x, y);
+#endif
+}
+
+template <>
+inline void vec_relu<float, platform::jit::avx2>(const int n, const float* x,
+                                                 float* y) {
+  vec_relu<float, platform::jit::avx>(n, x, y);
+}
+
+template <>
+inline void vec_relu<float, platform::jit::avx512_common>(const int n,
+                                                          const float* x,
+                                                          float* y) {
+  // TODO(TJ): enable me
+  vec_relu<float, platform::jit::avx2>(n, x, y);
+}
+
+// TODO(TJ): optimize double of sigmoid, tanh and relu if necessary
+
+template <typename T, platform::jit::cpu_isa_t isa = platform::jit::isa_any>
+class VecActivations {
+ public:
+  std::function<void(const int, const T*, T*)> operator()(
+      const std::string& type) {
+    if (type == "sigmoid") {
+      return vec_sigmoid<T, isa>;
+    } else if (type == "relu") {
+      return vec_relu<T, isa>;
+    } else if (type == "tanh") {
+      return vec_tanh<T, isa>;
+    } else if (type == "identity" || type == "") {
+      return vec_identity<T, isa>;
+    }
+    LOG(FATAL) << "Not support type: " << type;
+  }
+};
+
+}  // namespace math
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/math/cpu_vec_test.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 <sys/time.h>
+#include <cmath>
+#include <cstring>
+#include <random>
+#include <vector>
+#include "gflags/gflags.h"
+#include "glog/logging.h"
+#include "gtest/gtest.h"
+
+#include "paddle/fluid/operators/math/cpu_vec.h"
+
+inline double GetCurrentUS() {
+  struct timeval time;
+  gettimeofday(&time, NULL);
+  return 1e+6 * time.tv_sec + time.tv_usec;
+}
+constexpr int repeat = 1000;
+
+template <typename T>
+inline T _sigmoid(T x) {
+  const T min = SIGMOID_THRESHOLD_MIN;
+  const T max = SIGMOID_THRESHOLD_MAX;
+  T tmp = (x < min) ? min : ((x > max) ? max : x);
+  return static_cast<T>(1) / (static_cast<T>(1) + std::exp(-tmp));
+}
+
+template <typename T>
+inline T _tanh(T x) {
+  return static_cast<T>(2) * _sigmoid<T>(static_cast<T>(2) * x) -
+         static_cast<T>(1);
+}
+
+template <typename T>
+void ref_sigmoid(const int n, const T* x, T* y) {
+  for (int i = 0; i < n; ++i) {
+    y[i] = _sigmoid(x[i]);
+  }
+}
+
+template <typename T>
+void ref_tanh(const int n, const T* x, T* y) {
+  for (int i = 0; i < n; ++i) {
+    y[i] = _tanh(x[i]);
+  }
+}
+template <typename T>
+void ref_relu(const int n, const T* x, T* y) {
+  for (int i = 0; i < n; ++i) {
+    y[i] = x[i] > 0 ? x[i] : 0;
+  }
+}
+
+template <typename T>
+void RandomVec(const int n, T* a) {
+  static unsigned int seed = 100;
+  std::mt19937 rng(seed++);
+  std::uniform_real_distribution<double> uniform_dist(0, 1);
+  const T lower = static_cast<T>(-20.f);
+  const T upper = static_cast<T>(20.f);
+  for (int i = 0; i < n; ++i) {
+    a[i] = static_cast<T>(uniform_dist(rng) * (upper - lower) + lower);
+  }
+}
+
+template <typename T>
+void TestAndBench(const int n, std::function<void(const int, const T*, T*)> tgt,
+                  std::function<void(const int, const T*, T*)> ref) {
+  std::vector<T> x(n);
+  std::vector<T> ytgt(n), yref(n);
+  RandomVec<T>(n, x.data());
+
+  const T* x_data = x.data();
+  T* ytgt_data = ytgt.data();
+  T* yref_data = yref.data();
+  auto st = GetCurrentUS();
+  for (int i = 0; i < repeat; ++i) {
+    tgt(n, x_data, ytgt_data);
+  }
+  auto mt = GetCurrentUS();
+  for (int i = 0; i < repeat; ++i) {
+    ref(n, x_data, yref_data);
+  }
+  auto et = GetCurrentUS();
+
+  VLOG(3) << "Vec size " << n << ": refer takes: " << (et - mt) / repeat
+          << " us, tgt takes: " << (mt - st) / repeat;
+  for (int i = 0; i < n; ++i) {
+    EXPECT_NEAR(ytgt_data[i], yref_data[i], 1e-3);
+  }
+}
+
+TEST(CpuVecTest, sigmoid) {
+  namespace jit = paddle::platform::jit;
+  using namespace paddle::operators::math;  // NOLINT
+  for (auto sz : {1, 2, 15, 16, 30, 32, 128, 200, 512}) {
+    TestAndBench<float>(sz, vec_sigmoid<float>, ref_sigmoid<float>);
+    TestAndBench<float>(sz, vec_sigmoid<float, jit::avx>, ref_sigmoid<float>);
+    TestAndBench<float>(sz, vec_sigmoid<float, jit::avx2>, ref_sigmoid<float>);
+    TestAndBench<float>(sz, vec_sigmoid<float, jit::avx512_common>,
+                        ref_sigmoid<float>);
+  }
+  TestAndBench<double>(30, vec_sigmoid<double>, ref_sigmoid<double>);
+}
+
+TEST(CpuVecTest, tanh) {
+  namespace jit = paddle::platform::jit;
+  using namespace paddle::operators::math;  // NOLINT
+  for (auto sz : {1, 2, 15, 16, 30, 32, 128, 200, 512}) {
+    TestAndBench<float>(sz, vec_tanh<float>, ref_tanh<float>);
+    TestAndBench<float>(sz, vec_tanh<float, jit::avx>, ref_tanh<float>);
+    TestAndBench<float>(sz, vec_tanh<float, jit::avx2>, ref_tanh<float>);
+    TestAndBench<float>(sz, vec_tanh<float, jit::avx512_common>,
+                        ref_tanh<float>);
+  }
+  TestAndBench<double>(30, vec_tanh<double>, ref_tanh<double>);
+}
+
+TEST(CpuVecTest, relu) {
+  namespace jit = paddle::platform::jit;
+  using namespace paddle::operators::math;  // NOLINT
+  for (auto sz : {1, 2, 15, 16, 30, 32, 128, 200, 512}) {
+    TestAndBench<float>(sz, vec_relu<float>, ref_relu<float>);
+    TestAndBench<float>(sz, vec_relu<float, jit::avx>, ref_relu<float>);
+    TestAndBench<float>(sz, vec_relu<float, jit::avx2>, ref_relu<float>);
+    TestAndBench<float>(sz, vec_relu<float, jit::avx512_common>,
+                        ref_relu<float>);
+  }
+  TestAndBench<double>(30, vec_relu<double>, ref_relu<double>);
+}
+
+template <typename T>
+void TestInplace(const int n, std::function<void(const int, const T*, T*)> tgt,
+                 std::function<void(const int, const T*, T*)> ref) {
+  std::vector<T> x(n);
+  std::vector<T> ytgt(n), yref(n);
+  RandomVec<T>(n, x.data());
+
+  const T* x_data = x.data();
+  T* yref_data = yref.data();
+  T* ytgt_data = ytgt.data();
+  std::memcpy(yref_data, x_data, sizeof(T) * n);
+  std::memcpy(ytgt_data, x_data, sizeof(T) * n);
+
+  ref(n, yref_data, yref_data);
+  tgt(n, ytgt_data, ytgt_data);
+
+  for (int i = 0; i < n; ++i) {
+    EXPECT_NEAR(ytgt_data[i], yref_data[i], 1e-3);
+  }
+}
+
+TEST(CpuVecTest, inplace_sigmoid) {
+  namespace jit = paddle::platform::jit;
+  using namespace paddle::operators::math;  // NOLINT
+  for (auto sz : {1, 2, 15, 16, 30, 32, 128, 200, 512}) {
+    TestInplace<float>(sz, vec_sigmoid<float>, ref_sigmoid<float>);
+    TestInplace<float>(sz, vec_sigmoid<float, jit::avx>, ref_sigmoid<float>);
+    TestInplace<float>(sz, vec_sigmoid<float, jit::avx2>, ref_sigmoid<float>);
+    TestInplace<float>(sz, vec_sigmoid<float, jit::avx512_common>,
+                       ref_sigmoid<float>);
+  }
+  TestInplace<double>(30, vec_sigmoid<double>, ref_sigmoid<double>);
+}
+
+TEST(CpuVecTest, inplace_tanh) {
+  namespace jit = paddle::platform::jit;
+  using namespace paddle::operators::math;  // NOLINT
+  for (auto sz : {1, 2, 15, 16, 30, 32, 128, 200, 512}) {
+    TestInplace<float>(sz, vec_tanh<float>, ref_tanh<float>);
+    TestInplace<float>(sz, vec_tanh<float, jit::avx>, ref_tanh<float>);
+    TestInplace<float>(sz, vec_tanh<float, jit::avx2>, ref_tanh<float>);
+    TestInplace<float>(sz, vec_tanh<float, jit::avx512_common>,
+                       ref_tanh<float>);
+  }
+  TestInplace<double>(30, vec_tanh<double>, ref_tanh<double>);
+}
+
+TEST(CpuVecTest, inplace_relu) {
+  namespace jit = paddle::platform::jit;
+  using namespace paddle::operators::math;  // NOLINT
+  for (auto sz : {1, 2, 15, 16, 30, 32, 128, 200, 512}) {
+    TestInplace<float>(sz, vec_relu<float>, ref_relu<float>);
+    TestInplace<float>(sz, vec_relu<float, jit::avx>, ref_relu<float>);
+    TestInplace<float>(sz, vec_relu<float, jit::avx2>, ref_relu<float>);
+    TestInplace<float>(sz, vec_relu<float, jit::avx512_common>,
+                       ref_relu<float>);
+  }
+  TestInplace<double>(30, vec_relu<double>, ref_relu<double>);
+}

paddle/fluid/operators/math/fc_compute.h

@@ -25,17 +25,25 @@ namespace math {
 template <typename DeviceContext, typename T>
 inline void FCCompute(const BlasT<DeviceContext, T>& blas, const int M,
                       const int N, const int K, const T* X, const T* W, T* Y,
-                      const T* B = NULL) {
-  blas.GEMM(CblasNoTrans, CblasNoTrans, M, N, K, static_cast<T>(1), X, W,
-            static_cast<T>(0), Y);
-  if (B) {
+                      const T* B = NULL, bool relu = false) {
+  blas.MatMul(M, N, K, X, W, Y);
+  if (B == NULL) {
+    return;
+  }
+
 #ifdef PADDLE_WITH_MKLML
 #pragma omp parallel for if (FLAGS_paddle_num_threads > 1)
 #endif
-    for (int i = 0; i < M; i++) {
-      blas.AXPY(N, static_cast<T>(1), B, Y + i * N);
-    }
+  for (int i = 0; i < M; i++) {
+    blas.AXPY(N, static_cast<T>(1), B, Y + i * N);
   }
+
+  if (!relu) {
+    return;
+  }
+
+  // TODO(TJ): fuse relu
+  LOG(FATAL) << "Not implemented!";
 }
 
 }  // namespace math

paddle/fluid/operators/math/math_function.cc

@@ -41,7 +41,8 @@ template struct SetConstant<platform::CPUDeviceContext, uint8_t>;
   template struct Transpose<platform::CPUDeviceContext, int64_t, RANK>;    \
   template struct Transpose<platform::CPUDeviceContext, bool, RANK>;       \
   template struct Transpose<platform::CPUDeviceContext, int16_t, RANK>;    \
-  template struct Transpose<platform::CPUDeviceContext, uint8_t, RANK>;
+  template struct Transpose<platform::CPUDeviceContext, uint8_t, RANK>;    \
+  template struct Transpose<platform::CPUDeviceContext, int8_t, RANK>;
 
 DEFINE_CPU_TRANS(1);
 DEFINE_CPU_TRANS(2);

paddle/fluid/operators/math/math_function.cu

@@ -33,10 +33,11 @@ template struct SetConstant<platform::CUDADeviceContext, int>;
 template struct SetConstant<platform::CUDADeviceContext, int64_t>;
 template struct SetConstant<platform::CUDADeviceContext, bool>;
 
-#define DEFINE_GPU_TRANS(RANK)                                          \
-  template struct Transpose<platform::CUDADeviceContext, float, RANK>;  \
-  template struct Transpose<platform::CUDADeviceContext, double, RANK>; \
-  template struct Transpose<platform::CUDADeviceContext, float16, RANK>;
+#define DEFINE_GPU_TRANS(RANK)                                           \
+  template struct Transpose<platform::CUDADeviceContext, float, RANK>;   \
+  template struct Transpose<platform::CUDADeviceContext, double, RANK>;  \
+  template struct Transpose<platform::CUDADeviceContext, float16, RANK>; \
+  template struct Transpose<platform::CUDADeviceContext, int8_t, RANK>;
 
 DEFINE_GPU_TRANS(1);
 DEFINE_GPU_TRANS(2);

paddle/fluid/operators/math/math_function.h

@@ -19,6 +19,10 @@ limitations under the License. */
 
 #ifdef PADDLE_USE_OPENBLAS
 #include <cblas.h>
+// remove typedef in openblas
+#undef FLOAT
+#undef INT
+#undef SIZE
 #endif
 
 #include <cmath>

paddle/fluid/operators/math/sequence_padding.h

@@ -15,6 +15,7 @@ limitations under the License. */
 #pragma once
 
 #include <algorithm>
+#include <vector>
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/platform/device_context.h"
 
@@ -22,17 +23,33 @@ namespace paddle {
 namespace operators {
 namespace math {
 
-inline static size_t MaximumSequenceLength(const framework::LoD& lod,
-                                           const size_t level) {
-  const size_t num_sequences = lod[level].size() - 1;
-  size_t max_sequence_length = 0;
-  framework::LoD abs_offset_lod = framework::ToAbsOffset(lod);
-  for (size_t i = 0; i < num_sequences; ++i) {
-    max_sequence_length =
-        std::max(max_sequence_length,
-                 abs_offset_lod[level][i + 1] - abs_offset_lod[level][i]);
+enum PadLayout { kBatchLengthWidth = 0, kLengthBatchWidth };
+
+enum CopyType { kSeqToPad, kPadToSeq };
+
+inline static size_t MaximumSequenceLength(
+    const framework::Vector<size_t>& seq_offset) {
+  size_t seq_num = seq_offset.size() - 1;
+  size_t max_seq_len = 0;
+  for (size_t i = 0; i < seq_num; ++i) {
+    max_seq_len = std::max(max_seq_len, seq_offset[i + 1] - seq_offset[i]);
   }
-  return max_sequence_length;
+  return max_seq_len;
+}
+
+inline static void CheckDims(const framework::DDim& seq_tensor_dims,
+                             const framework::DDim& pad_tensor_dims,
+                             const framework::Vector<size_t>& seq_offset,
+                             int64_t padded_seq_len, int64_t step_width,
+                             const PadLayout& layout) {
+  PADDLE_ENFORCE_EQ(static_cast<size_t>(seq_tensor_dims[0]), seq_offset.back(),
+                    "Value of 1st dimension of the sequence tensor should be "
+                    "equal to sum of lengths of all sequences.");
+
+  PADDLE_ENFORCE(seq_tensor_dims.size() + 1 == pad_tensor_dims.size() ||
+                     seq_tensor_dims.size() == pad_tensor_dims.size(),
+                 "pad_tensor's rank should be 1 greater than seq_tensor's "
+                 "rank, or be equal with it.");
 }
 
 /*
@@ -64,15 +81,22 @@ inline static size_t MaximumSequenceLength(const framework::LoD& lod,
 template <typename DeviceContext, typename T>
 class PaddingLoDTensorFunctor {
  public:
-  void operator()(const DeviceContext& context, const framework::LoDTensor& seq,
-                  framework::Tensor* padding, bool norm_by_times);
+  void operator()(const DeviceContext& context,
+                  const framework::LoDTensor& seq_tensor,
+                  framework::LoDTensor* pad_tensor,
+                  const framework::LoDTensor& pad_value, int pad_seq_len = -1,
+                  int lod_level = 0, bool norm_by_times = false,
+                  const PadLayout layout = kBatchLengthWidth);
 };
 
 template <typename DeviceContext, typename T>
 class UnpaddingLoDTensorFunctor {
  public:
-  void operator()(const DeviceContext& context, framework::LoDTensor* seq,
-                  const framework::Tensor& padding, bool norm_by_times);
+  void operator()(const DeviceContext& context,
+                  const framework::LoDTensor& pad_tensor,
+                  framework::LoDTensor* seq_tensor, int pad_seq_len = -1,
+                  int lod_level = 0, bool norm_by_times = false,
+                  const PadLayout layout = kBatchLengthWidth);
 };
 
 }  // namespace math

paddle/fluid/operators/mul_op.h

@@ -62,23 +62,31 @@ class MulGradKernel : public framework::OpKernel<T> {
   void Compute(const framework::ExecutionContext& ctx) const override {
     int x_num_col_dims = ctx.template Attr<int>("x_num_col_dims");
     int y_num_col_dims = ctx.template Attr<int>("y_num_col_dims");
-    const Tensor* x = ctx.Input<Tensor>("X");
-    const Tensor* y = ctx.Input<Tensor>("Y");
-    const Tensor x_matrix = x->dims().size() > 2
-                                ? framework::ReshapeToMatrix(*x, x_num_col_dims)
-                                : *x;
-    const Tensor y_matrix = y->dims().size() > 2
-                                ? framework::ReshapeToMatrix(*y, y_num_col_dims)
-                                : *y;
-    const Tensor* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));
+    auto* x = ctx.Input<framework::LoDTensor>("X");
+    auto* y = ctx.Input<framework::LoDTensor>("Y");
+    auto x_matrix = x->dims().size() > 2
+                        ? framework::ReshapeToMatrix(*x, x_num_col_dims)
+                        : static_cast<const Tensor&>(*x);
+    auto y_matrix = y->dims().size() > 2
+                        ? framework::ReshapeToMatrix(*y, y_num_col_dims)
+                        : static_cast<const Tensor&>(*y);
+    auto* dout = ctx.Input<framework::LoDTensor>(framework::GradVarName("Out"));
 
     Tensor dout_mat;
     dout_mat.ShareDataWith(*dout);
     dout_mat.Resize({framework::flatten_to_2d(x->dims(), x_num_col_dims)[0],
                      framework::flatten_to_2d(y->dims(), y_num_col_dims)[1]});
 
-    Tensor* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
-    Tensor* dy = ctx.Output<Tensor>(framework::GradVarName("Y"));
+    auto* dx = ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
+    auto* dy = ctx.Output<framework::LoDTensor>(framework::GradVarName("Y"));
+
+    if (dx != nullptr) {
+      dx->set_lod(x->lod());
+    }
+    if (dy != nullptr) {
+      dy->set_lod(y->lod());
+    }
+
     auto& dev_ctx = ctx.template device_context<DeviceContext>();
     auto blas = math::GetBlas<DeviceContext, T>(dev_ctx);
     if (dx) {