Merge branch 'develop' into dam_fc

CMakeLists.txt

@@ -26,6 +26,11 @@ message(STATUS "C compiler: ${CMAKE_C_COMPILER}, version: "
         "${CMAKE_C_COMPILER_ID} ${CMAKE_C_COMPILER_VERSION}")
 if(WIN32)
     set(CMAKE_STATIC_LIBRARY_PREFIX lib)
+    add_definitions("/DGOOGLE_GLOG_DLL_DECL=")
+    set(CMAKE_C_FLAGS_DEBUG   "${CMAKE_C_FLAGS_DEBUG} /bigobj /MTd")
+    set(CMAKE_C_FLAGS_RELEASE  "${CMAKE_C_FLAGS_RELEASE} /bigobj /MT")
+    set(CMAKE_CXX_FLAGS_DEBUG  "${CMAKE_CXX_FLAGS_DEBUG} /bigobj /MTd")
+    set(CMAKE_CXX_FLAGS_RELEASE   "${CMAKE_CXX_FLAGS_RELEASE} /bigobj /MT")
 endif(WIN32)
 
 if(NOT CMAKE_CROSSCOMPILING)
@@ -66,6 +71,8 @@ option(WITH_ARM_FP16    "Use half precision support on armv8.2-a cpu"   OFF)
 option(WITH_CONTRIB     "Compile the third-party contributation"        OFF)
 option(REPLACE_ENFORCE_GLOG "Replace PADDLE_ENFORCE with glog/CHECK for better debug." OFF)
 option(WITH_ANAKIN      "Compile with Anakin library"                   OFF)
+option(ANAKIN_BUILD_FAT_BIN "Build anakin cuda fat-bin lib for all device plantform, ignored when WITH_ANAKIN=OFF" OFF)
+option(ANAKIN_BUILD_CROSS_PLANTFORM "Build anakin lib for any nvidia device plantform. ignored when WITH_ANAKIN=OFF" ON)
 option(WITH_GRPC     "Use grpc as the default rpc framework"            ${WITH_DISTRIBUTE})
 option(WITH_BRPC_RDMA     "Use brpc rdma as the rpc protocal"           OFF)
 option(ON_INFER         "Turn on inference optimization."               OFF)

cmake/cuda.cmake

@@ -157,6 +157,9 @@ list(APPEND EXTERNAL_LIBS ${CUDA_LIBRARIES} ${CUDA_rt_LIBRARY})
 if(NOT WITH_DSO)
     # TODO(panyx0718): CUPTI only allows DSO?
     list(APPEND EXTERNAL_LIBS ${CUDNN_LIBRARY} ${CUPTI_LIBRARY} ${CUDA_CUBLAS_LIBRARIES} ${CUDA_curand_LIBRARY} ${NCCL_LIBRARY})
+    if(WIN32)
+      set_property(GLOBAL PROPERTY CUDA_MODULES ${CUDNN_LIBRARY} ${CUDA_CUBLAS_LIBRARIES} ${CUDA_curand_LIBRARY})
+    endif(WIN32)
 endif(NOT WITH_DSO)
 
 # setting nvcc arch flags
@@ -196,10 +199,12 @@ elseif(CMAKE_BUILD_TYPE  STREQUAL "MinSizeRel")
     list(APPEND CUDA_NVCC_FLAGS  ${CMAKE_CXX_FLAGS_RELEASE})
 endif()
 else(NOT WIN32)
-if(CMAKE_BUILD_TYPE STREQUAL "Release")
+if(CMAKE_BUILD_TYPE  STREQUAL "Debug")
+    list(APPEND CUDA_NVCC_FLAGS  "-g -G")
+elseif(CMAKE_BUILD_TYPE STREQUAL "Release")
   list(APPEND CUDA_NVCC_FLAGS "-O3 -DNDEBUG")
 else()
-  message(FATAL "Windows only support Release build now. Please set visual studio build type to Release, x64 build.")
+  message(FATAL "Windows only support Release or Debug build now. Please set visual studio build type to Release/Debug, x64 build.")
 endif()
 endif(NOT WIN32)
 

cmake/cudnn.cmake

@@ -2,7 +2,12 @@ if(NOT WITH_GPU)
     return()
 endif()
 
-set(CUDNN_ROOT "/usr" CACHE PATH "CUDNN ROOT")
+if(WIN32)
+    set(CUDNN_ROOT ${CUDA_TOOLKIT_ROOT_DIR})
+else(WIN32)
+    set(CUDNN_ROOT "/usr" CACHE PATH "CUDNN ROOT")
+endif(WIN32)
+
 find_path(CUDNN_INCLUDE_DIR cudnn.h
     PATHS ${CUDNN_ROOT} ${CUDNN_ROOT}/include
     $ENV{CUDNN_ROOT} $ENV{CUDNN_ROOT}/include ${CUDA_TOOLKIT_INCLUDE}

cmake/external/anakin.cmake

@@ -58,19 +58,21 @@ ExternalProject_Add(
                         -DPROTOBUF_ROOT=${THIRD_PARTY_PATH}/install/protobuf
                         -DMKLML_ROOT=${THIRD_PARTY_PATH}/install/mklml
                         -DENABLE_OP_TIMER=${ANAKIN_ENABLE_OP_TIMER}
+                        -DBUILD_FAT_BIN=${ANAKIN_BUILD_FAT_BIN}
+                        -DBUILD_CROSS_PLANTFORM=${ANAKIN_BUILD_CROSS_PLANTFORM}
                         ${EXTERNAL_OPTIONAL_ARGS}
     CMAKE_CACHE_ARGS    -DCMAKE_INSTALL_PREFIX:PATH=${ANAKIN_INSTALL_DIR}
 )
 
 message(STATUS "Anakin for inference is enabled")
 message(STATUS "Anakin is set INCLUDE:${ANAKIN_INCLUDE} LIBRARY:${ANAKIN_LIBRARY}")
-
+add_dependencies(extern_anakin protobuf mklml)
 add_library(anakin_shared SHARED IMPORTED GLOBAL)
 set_property(TARGET anakin_shared PROPERTY IMPORTED_LOCATION ${ANAKIN_SHARED_LIB})
-add_dependencies(anakin_shared extern_anakin protobuf mklml)
+add_dependencies(anakin_shared extern_anakin)
 
 add_library(anakin_saber SHARED IMPORTED GLOBAL)
 set_property(TARGET anakin_saber PROPERTY IMPORTED_LOCATION ${ANAKIN_SABER_LIB})
-add_dependencies(anakin_saber extern_anakin protobuf mklml)
+add_dependencies(anakin_saber extern_anakin)
 
 list(APPEND external_project_dependencies anakin_shared anakin_saber)

cmake/external/boost.cmake

@@ -28,34 +28,28 @@ 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()
-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)
+
+MESSAGE(STATUS "BOOST_TAR: ${BOOST_TAR}, BOOST_URL: ${BOOST_URL}")
 
 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)
-set_directory_properties(PROPERTIES CLEAN_NO_CUSTOM 1)
 
+set(BOOST_INCLUDE_DIR "${BOOST_DOWNLOAD_DIR}" CACHE PATH "boost include directory." FORCE)
+set_directory_properties(PROPERTIES CLEAN_NO_CUSTOM 1)
 include_directories(${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
+    URL      ${BOOST_URL}
     DOWNLOAD_NO_PROGRESS  1
     PREFIX                ${BOOST_SOURCES_DIR}
     CONFIGURE_COMMAND     ""
     BUILD_COMMAND         ""
     INSTALL_COMMAND       ""
     UPDATE_COMMAND        ""
-)
-endif(NOT WIN32)
+    )
 
 if (${CMAKE_VERSION} VERSION_LESS "3.3.0" OR NOT WIN32)
     set(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/boost_dummy.c)

cmake/external/gflags.cmake

@@ -35,7 +35,12 @@ ExternalProject_Add(
     CMAKE_ARGS      -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
                     -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
                     -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
+                    -DCMAKE_CXX_FLAGS_RELEASE=${CMAKE_CXX_FLAGS_RELEASE}
+                    -DCMAKE_CXX_FLAGS_DEBUG=${CMAKE_CXX_FLAGS_DEBUG}
                     -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}
+                    -DCMAKE_C_FLAGS_DEBUG=${CMAKE_C_FLAGS_DEBUG}
+                    -DCMAKE_C_FLAGS_RELEASE=${CMAKE_C_FLAGS_RELEASE}
+                    -DBUILD_STATIC_LIBS=ON
                     -DCMAKE_INSTALL_PREFIX=${GFLAGS_INSTALL_DIR}
                     -DCMAKE_POSITION_INDEPENDENT_CODE=ON
                     -DBUILD_TESTING=OFF
@@ -48,8 +53,8 @@ ExternalProject_Add(
 IF(WIN32)
   IF(NOT EXISTS "${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
-  )
+            COMMAND cmake -E copy ${GFLAGS_INSTALL_DIR}/lib/gflags_static.lib ${GFLAGS_INSTALL_DIR}/lib/libgflags.lib
+            )
   ENDIF()
 ENDIF(WIN32)
 ADD_LIBRARY(gflags STATIC IMPORTED GLOBAL)

cmake/external/glog.cmake

@@ -46,7 +46,11 @@ ExternalProject_Add(
     CMAKE_ARGS      -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
                     -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
                     -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
+                    -DCMAKE_CXX_FLAGS_RELEASE=${CMAKE_CXX_FLAGS_RELEASE}
+                    -DCMAKE_CXX_FLAGS_DEBUG=${CMAKE_CXX_FLAGS_DEBUG}
                     -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}
+                    -DCMAKE_C_FLAGS_DEBUG=${CMAKE_C_FLAGS_DEBUG}
+                    -DCMAKE_C_FLAGS_RELEASE=${CMAKE_C_FLAGS_RELEASE}
                     -DCMAKE_INSTALL_PREFIX=${GLOG_INSTALL_DIR}
                     -DCMAKE_INSTALL_LIBDIR=${GLOG_INSTALL_DIR}/lib
                     -DCMAKE_POSITION_INDEPENDENT_CODE=ON
@@ -63,7 +67,7 @@ ExternalProject_Add(
 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
+    COMMAND cmake -E copy ${GLOG_INSTALL_DIR}/lib/glog.lib ${GLOG_INSTALL_DIR}/lib/libglog.lib
   )
   ENDIF()
 ENDIF(WIN32)

cmake/external/openblas.cmake

@@ -17,12 +17,8 @@ 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)
@@ -34,6 +30,7 @@ IF(NOT ${CBLAS_FOUND})
         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})

cmake/external/protobuf.cmake

@@ -140,7 +140,6 @@ 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 "${PROTOBUF_ROOT}" STREQUAL "")
@@ -188,13 +187,20 @@ FUNCTION(build_protobuf TARGET_NAME BUILD_FOR_HOST)
         SET(OPTIONAL_ARGS
             "-DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}"
             "-DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}"
-            "-DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}"
             "-DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}"
+            "-DCMAKE_C_FLAGS_DEBUG=${CMAKE_C_FLAGS_DEBUG}"
+            "-DCMAKE_C_FLAGS_RELEASE=${CMAKE_C_FLAGS_RELEASE}"
+            "-DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}"
+            "-DCMAKE_CXX_FLAGS_RELEASE=${CMAKE_CXX_FLAGS_RELEASE}"
+            "-DCMAKE_CXX_FLAGS_DEBUG=${CMAKE_CXX_FLAGS_DEBUG}"
             "-Dprotobuf_WITH_ZLIB=ON"
             "-DZLIB_ROOT:FILEPATH=${ZLIB_ROOT}"
             ${EXTERNAL_OPTIONAL_ARGS})
         SET(OPTIONAL_CACHE_ARGS "-DZLIB_ROOT:STRING=${ZLIB_ROOT}")
     ENDIF()
+    IF(WIN32)
+        SET(OPTIONAL_ARGS ${OPTIONAL_ARGS} "-DCMAKE_GENERATOR_PLATFORM=x64")
+    ENDIF()
 
     SET(PROTOBUF_REPO "https://github.com/google/protobuf.git")
     SET(PROTOBUF_TAG "9f75c5aa851cd877fb0d93ccc31b8567a6706546")

cmake/external/python.cmake

@@ -21,6 +21,48 @@ INCLUDE(python_module)
 FIND_PACKAGE(PythonInterp ${PY_VERSION})
 FIND_PACKAGE(PythonLibs ${PY_VERSION})
 
+if(WIN32)
+    execute_process(COMMAND "${PYTHON_EXECUTABLE}" "-c"
+"from distutils import sysconfig as s;import sys;import struct;
+print(sys.prefix);
+print(s.get_config_var('LDVERSION') or s.get_config_var('VERSION'));
+"
+            RESULT_VARIABLE _PYTHON_SUCCESS
+            OUTPUT_VARIABLE _PYTHON_VALUES
+            ERROR_VARIABLE _PYTHON_ERROR_VALUE)
+
+    if(NOT _PYTHON_SUCCESS MATCHES 0)
+        set(PYTHONLIBS_FOUND FALSE)
+        return()
+    endif()
+
+    # Convert the process output into a list
+    string(REGEX REPLACE ";" "\\\\;" _PYTHON_VALUES ${_PYTHON_VALUES})
+    string(REGEX REPLACE "\n" ";" _PYTHON_VALUES ${_PYTHON_VALUES})
+    list(GET _PYTHON_VALUES 0 PYTHON_PREFIX)
+    list(GET _PYTHON_VALUES 1 PYTHON_LIBRARY_SUFFIX)
+
+    # Make sure all directory separators are '/'
+    string(REGEX REPLACE "\\\\" "/" PYTHON_PREFIX ${PYTHON_PREFIX})
+
+    set(PYTHON_LIBRARY
+            "${PYTHON_PREFIX}/libs/Python${PYTHON_LIBRARY_SUFFIX}.lib")
+
+    # when run in a venv, PYTHON_PREFIX points to it. But the libraries remain in the
+    # original python installation. They may be found relative to PYTHON_INCLUDE_DIR.
+    if(NOT EXISTS "${PYTHON_LIBRARY}")
+        get_filename_component(_PYTHON_ROOT ${PYTHON_INCLUDE_DIR} DIRECTORY)
+        set(PYTHON_LIBRARY
+                "${_PYTHON_ROOT}/libs/Python${PYTHON_LIBRARY_SUFFIX}.lib")
+    endif()
+
+    # raise an error if the python libs are still not found.
+    if(NOT EXISTS "${PYTHON_LIBRARY}")
+        message(FATAL_ERROR "Python libraries not found")
+    endif()
+    SET(PYTHON_LIBRARIES "${PYTHON_LIBRARY}")
+endif(WIN32)
+
 # Fixme: Maybe find a static library. Get SHARED/STATIC by FIND_PACKAGE.
 ADD_LIBRARY(python SHARED IMPORTED GLOBAL)
 SET_PROPERTY(TARGET python PROPERTY IMPORTED_LOCATION ${PYTHON_LIBRARIES})

cmake/external/xxhash.cmake

@@ -14,23 +14,52 @@ ELSE()
   ENDIF(APPLE)
 ENDIF()
 
-ExternalProject_Add(
-    extern_xxhash
-    ${EXTERNAL_PROJECT_LOG_ARGS}
-    GIT_REPOSITORY  "https://github.com/Cyan4973/xxHash"
-    GIT_TAG         "v0.6.5"
-    PREFIX          ${XXHASH_SOURCE_DIR}
-    DOWNLOAD_NAME   "xxhash"
-    UPDATE_COMMAND  ""
-    CONFIGURE_COMMAND ""
-    BUILD_IN_SOURCE 1
-    PATCH_COMMAND
-    BUILD_COMMAND     ${BUILD_CMD}
-    INSTALL_COMMAND   export PREFIX=${XXHASH_INSTALL_DIR}/ && make install
-    TEST_COMMAND      ""
-)
+if(WIN32)
+  ExternalProject_Add(
+          extern_xxhash
+          ${EXTERNAL_PROJECT_LOG_ARGS}
+          GIT_REPOSITORY  "https://github.com/Cyan4973/xxHash"
+          GIT_TAG         "v0.6.5"
+          PREFIX          ${XXHASH_SOURCE_DIR}
+          DOWNLOAD_NAME   "xxhash"
+          UPDATE_COMMAND  ""
+          BUILD_IN_SOURCE 1
+          PATCH_COMMAND
+          CONFIGURE_COMMAND
+          ${CMAKE_COMMAND} ${XXHASH_SOURCE_DIR}/src/extern_xxhash/cmake_unofficial
+          -DCMAKE_INSTALL_PREFIX:PATH=${XXHASH_INSTALL_DIR}
+          -DCMAKE_BUILD_TYPE:STRING=${CMAKE_BUILD_TYPE}
+          -DCMAKE_VERBOSE_MAKEFILE:BOOL=OFF
+          -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON
+          -DBUILD_XXHSUM=OFF
+          -DCMAKE_GENERATOR_PLATFORM=x64
+          -DBUILD_SHARED_LIBS=OFF
+          ${OPTIONAL_CACHE_ARGS}
+          TEST_COMMAND      ""
+  )
+else()
+  ExternalProject_Add(
+      extern_xxhash
+      ${EXTERNAL_PROJECT_LOG_ARGS}
+      GIT_REPOSITORY  "https://github.com/Cyan4973/xxHash"
+      GIT_TAG         "v0.6.5"
+      PREFIX          ${XXHASH_SOURCE_DIR}
+      DOWNLOAD_NAME   "xxhash"
+      UPDATE_COMMAND  ""
+      CONFIGURE_COMMAND ""
+      BUILD_IN_SOURCE 1
+      PATCH_COMMAND
+      BUILD_COMMAND     ${BUILD_CMD}
+      INSTALL_COMMAND   export PREFIX=${XXHASH_INSTALL_DIR}/ && make install
+      TEST_COMMAND      ""
+  )
+endif()
 
-set(XXHASH_LIBRARIES "${XXHASH_INSTALL_DIR}/lib/libxxhash.a")
+if (WIN32)
+  set(XXHASH_LIBRARIES "${XXHASH_INSTALL_DIR}/lib/xxhash.lib")
+else()
+  set(XXHASH_LIBRARIES "${XXHASH_INSTALL_DIR}/lib/libxxhash.a")
+endif ()
 INCLUDE_DIRECTORIES(${XXHASH_INCLUDE_DIR})
 
 add_library(xxhash STATIC IMPORTED GLOBAL)

cmake/generic.cmake

@@ -266,7 +266,11 @@ function(cc_library TARGET_NAME)
       if("${cc_library_DEPS};" MATCHES "python;")
         list(REMOVE_ITEM cc_library_DEPS python)
         add_dependencies(${TARGET_NAME} python)
-        target_link_libraries(${TARGET_NAME} "-Wl,-undefined,dynamic_lookup")
+        if(WIN32)
+          target_link_libraries(${TARGET_NAME} ${PYTHON_LIBRARIES})
+        else()
+          target_link_libraries(${TARGET_NAME} "-Wl,-undefined,dynamic_lookup")
+        endif(WIN32)
       endif()
       target_link_libraries(${TARGET_NAME} ${cc_library_DEPS})
       add_dependencies(${TARGET_NAME} ${cc_library_DEPS})
@@ -288,6 +292,45 @@ function(cc_library TARGET_NAME)
   endif(cc_library_SRCS)
 endfunction(cc_library)
 
+# The link operation under windows may exceeds the maximum characters limit, simply break the link command
+# into multiple link opeartion can fix that, say
+# original:
+#     lib /out:target.lib a.lib b.lib c.lib d.lib
+# after:
+#    1. lib /out:dummy_lib_1.lib a.lib b.lib
+#    2. lib /out:dummy_lib_2.lib c.lib d.lib
+#    1. lib /out:target.lib dummy_lib_1.lib dummy_lib_2.lib
+function(sep_library TARGET_NAME)
+  set(options STATIC static SHARED shared)
+  set(oneValueArgs "")
+  set(multiValueArgs SRCS DEPS)
+  cmake_parse_arguments(sep_library "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
+  set(dummy_index 1)
+  set(dummy_offset 1)
+  # the dummy target would be consisted of limit size libraries
+  set(dummy_limit 50)
+  list(LENGTH sep_library_DEPS sep_all_len)
+  foreach(v ${sep_library_DEPS})
+    list(APPEND dummy_list ${v})
+    list(LENGTH dummy_list listlen )
+    if ((${listlen} GREATER ${dummy_limit}) OR (${dummy_offset} EQUAL ${sep_all_len}))
+      message("create dummy library ${TARGET_NAME}_dummy_lib_${dummy_index} for ${TARGET_NAME}")
+      cc_library(${TARGET_NAME}_dummy_lib_${dummy_index} STATIC DEPS ${dummy_list})
+      foreach(i ${dummy_list})
+        list(REMOVE_AT dummy_list 0)
+      endforeach()
+      list(APPEND ${TARGET_NAME}_dummy_list ${TARGET_NAME}_dummy_lib_${dummy_index})
+      MATH(EXPR dummy_index "${dummy_index}+1")
+    endif()
+    MATH(EXPR dummy_offset "${dummy_offset}+1")
+  endforeach()
+  if(${sep_library_SHARED})
+    cc_library(${TARGET_NAME} SHARED SRCS ${sep_library_SRCS} DEPS ${${TARGET_NAME}_dummy_list})
+  else(${sep_library_SHARED})
+    cc_library(${TARGET_NAME} STATIC SRCS ${sep_library_SRCS} DEPS ${${TARGET_NAME}_dummy_list})
+  endif(${sep_library_SHARED})
+endfunction(sep_library)
+
 function(cc_binary TARGET_NAME)
   set(options "")
   set(oneValueArgs "")

cmake/inference_lib.cmake

@@ -22,144 +22,165 @@ function(copy TARGET)
 
     list(LENGTH copy_lib_SRCS copy_lib_SRCS_len)
     list(LENGTH copy_lib_DSTS copy_lib_DSTS_len)
-    if(NOT ${copy_lib_SRCS_len} EQUAL ${copy_lib_DSTS_len})
+    if (NOT ${copy_lib_SRCS_len} EQUAL ${copy_lib_DSTS_len})
         message(FATAL_ERROR "${TARGET} source numbers are not equal to destination numbers")
-    endif()
+    endif ()
     math(EXPR len "${copy_lib_SRCS_len} - 1")
 
     add_custom_target(${TARGET} DEPENDS ${copy_lib_DEPS})
-    foreach(index RANGE ${len})
+    foreach (index RANGE ${len})
         list(GET copy_lib_SRCS ${index} src)
         list(GET copy_lib_DSTS ${index} dst)
-        add_custom_command(TARGET ${TARGET} PRE_BUILD
-          COMMAND mkdir -p "${dst}"
-          COMMAND cp -r "${src}" "${dst}"
-          COMMENT "copying ${src} -> ${dst}")
-    endforeach()
+        if (WIN32)
+            # windows cmd shell will not expand wildcard automatically.
+            # below expand the files,libs and copy them by rules.
+            file(GLOB header_files ${src} "*.h")
+            file(GLOB static_lib_files ${src} "*.lib")
+            file(GLOB dll_lib_files ${src} "*.dll")
+            set(src_files ${header_files} ${static_lib_files} ${dll_lib_files})
+
+            if (NOT "${src_files}" STREQUAL "")
+                list(REMOVE_DUPLICATES src_files)
+            endif ()
+            add_custom_command(TARGET ${TARGET} PRE_BUILD
+                    COMMAND ${CMAKE_COMMAND} -E make_directory "${dst}"
+                    )
+            foreach (src_file ${src_files})
+                add_custom_command(TARGET ${TARGET} PRE_BUILD
+                        COMMAND ${CMAKE_COMMAND} -E copy "${src_file}" "${dst}"
+                        COMMENT "copying ${src_file} -> ${dst}")
+            endforeach ()
+        else (WIN32) # not windows
+            add_custom_command(TARGET ${TARGET} PRE_BUILD
+                    COMMAND mkdir -p "${dst}"
+                    COMMAND cp -r "${src}" "${dst}"
+                    COMMENT "copying ${src} -> ${dst}")
+        endif (WIN32) # not windows
+    endforeach ()
 endfunction()
 
 # third party
 set(dst_dir "${FLUID_INSTALL_DIR}/third_party/eigen3")
 copy(eigen3_lib
-  SRCS ${EIGEN_INCLUDE_DIR}/Eigen/Core ${EIGEN_INCLUDE_DIR}/Eigen/src ${EIGEN_INCLUDE_DIR}/unsupported/Eigen
-  DSTS ${dst_dir}/Eigen ${dst_dir}/Eigen ${dst_dir}/unsupported
-  DEPS eigen3
-)
+        SRCS ${EIGEN_INCLUDE_DIR}/Eigen/Core ${EIGEN_INCLUDE_DIR}/Eigen/src ${EIGEN_INCLUDE_DIR}/unsupported/Eigen
+        DSTS ${dst_dir}/Eigen ${dst_dir}/Eigen ${dst_dir}/unsupported
+        DEPS eigen3
+        )
 
 set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/gflags")
 copy(gflags_lib
-  SRCS ${GFLAGS_INCLUDE_DIR} ${GFLAGS_LIBRARIES}
-  DSTS ${dst_dir} ${dst_dir}/lib
-  DEPS gflags
-)
+        SRCS ${GFLAGS_INCLUDE_DIR} ${GFLAGS_LIBRARIES}
+        DSTS ${dst_dir} ${dst_dir}/lib
+        DEPS gflags
+        )
 
 set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/glog")
 copy(glog_lib
-  SRCS ${GLOG_INCLUDE_DIR} ${GLOG_LIBRARIES}
-  DSTS ${dst_dir} ${dst_dir}/lib
-  DEPS glog
-)
+        SRCS ${GLOG_INCLUDE_DIR} ${GLOG_LIBRARIES}
+        DSTS ${dst_dir} ${dst_dir}/lib
+        DEPS glog
+        )
 
 set(dst_dir "${FLUID_INSTALL_DIR}/third_party/boost/")
 copy(boost_lib
-  SRCS ${BOOST_INCLUDE_DIR}/boost
-  DSTS ${dst_dir}
-  DEPS boost
-)
+        SRCS ${BOOST_INCLUDE_DIR}/boost
+        DSTS ${dst_dir}
+        DEPS boost
+        )
 
 set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/xxhash")
 copy(xxhash_lib
-  SRCS ${XXHASH_INCLUDE_DIR} ${XXHASH_LIBRARIES}
-  DSTS ${dst_dir} ${dst_dir}/lib
-  DEPS xxhash
-)
+        SRCS ${XXHASH_INCLUDE_DIR} ${XXHASH_LIBRARIES}
+        DSTS ${dst_dir} ${dst_dir}/lib
+        DEPS xxhash
+        )
 
-if(NOT PROTOBUF_FOUND)
+if (NOT PROTOBUF_FOUND)
     set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/protobuf")
     copy(protobuf_lib
-      SRCS ${PROTOBUF_INCLUDE_DIR} ${PROTOBUF_LIBRARY}
-      DSTS ${dst_dir} ${dst_dir}/lib
-      DEPS extern_protobuf
-    )
-endif()
+            SRCS ${PROTOBUF_INCLUDE_DIR} ${PROTOBUF_LIBRARY}
+            DSTS ${dst_dir} ${dst_dir}/lib
+            DEPS extern_protobuf
+            )
+endif ()
 
-if(NOT CBLAS_FOUND)
+if (NOT CBLAS_FOUND)
     set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/openblas")
     copy(openblas_lib
-      SRCS ${CBLAS_INSTALL_DIR}/lib ${CBLAS_INSTALL_DIR}/include
-      DSTS ${dst_dir} ${dst_dir}
-      DEPS extern_openblas
-    )
+            SRCS ${CBLAS_INSTALL_DIR}/lib ${CBLAS_INSTALL_DIR}/include
+            DSTS ${dst_dir} ${dst_dir}
+            DEPS extern_openblas
+            )
 elseif (WITH_MKLML)
     set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/mklml")
     copy(mklml_lib
-      SRCS ${MKLML_LIB} ${MKLML_IOMP_LIB} ${MKLML_INC_DIR}
-      DSTS ${dst_dir}/lib ${dst_dir}/lib ${dst_dir}
-      DEPS mklml
-    )
-endif()
-
-if(WITH_MKLDNN)
-  set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/mkldnn")
-  copy(mkldnn_lib
-    SRCS ${MKLDNN_INC_DIR} ${MKLDNN_SHARED_LIB}
-    DSTS ${dst_dir} ${dst_dir}/lib
-    DEPS mkldnn
-  )
-endif()
+            SRCS ${MKLML_LIB} ${MKLML_IOMP_LIB} ${MKLML_INC_DIR}
+            DSTS ${dst_dir}/lib ${dst_dir}/lib ${dst_dir}
+            DEPS mklml
+            )
+endif ()
+
+if (WITH_MKLDNN)
+    set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/mkldnn")
+    copy(mkldnn_lib
+            SRCS ${MKLDNN_INC_DIR} ${MKLDNN_SHARED_LIB}
+            DSTS ${dst_dir} ${dst_dir}/lib
+            DEPS mkldnn
+            )
+endif ()
 
 if (NOT WIN32)
-if(NOT MOBILE_INFERENCE AND NOT RPI)
-  set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/snappy")
-  copy(snappy_lib
-    SRCS ${SNAPPY_INCLUDE_DIR} ${SNAPPY_LIBRARIES}
-    DSTS ${dst_dir} ${dst_dir}/lib
-    DEPS snappy)
-
-  set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/snappystream")
-  copy(snappystream_lib
-    SRCS ${SNAPPYSTREAM_INCLUDE_DIR} ${SNAPPYSTREAM_LIBRARIES}
-    DSTS ${dst_dir} ${dst_dir}/lib
-    DEPS snappystream)
-
-  set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/zlib")
-  copy(zlib_lib
-    SRCS ${ZLIB_INCLUDE_DIR} ${ZLIB_LIBRARIES}
-    DSTS ${dst_dir} ${dst_dir}/lib
-    DEPS zlib)
-endif()
-endif(NOT WIN32)
+    if (NOT MOBILE_INFERENCE AND NOT RPI)
+        set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/snappy")
+        copy(snappy_lib
+                SRCS ${SNAPPY_INCLUDE_DIR} ${SNAPPY_LIBRARIES}
+                DSTS ${dst_dir} ${dst_dir}/lib
+                DEPS snappy)
+
+        set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/snappystream")
+        copy(snappystream_lib
+                SRCS ${SNAPPYSTREAM_INCLUDE_DIR} ${SNAPPYSTREAM_LIBRARIES}
+                DSTS ${dst_dir} ${dst_dir}/lib
+                DEPS snappystream)
+
+        set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/zlib")
+        copy(zlib_lib
+                SRCS ${ZLIB_INCLUDE_DIR} ${ZLIB_LIBRARIES}
+                DSTS ${dst_dir} ${dst_dir}/lib
+                DEPS zlib)
+    endif ()
+endif (NOT WIN32)
 
 # 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)
-set(framework_lib_deps framework_py_proto)
-endif(NOT WIN32)
+    set(framework_lib_deps framework_py_proto)
+endif (NOT WIN32)
 copy(framework_lib DEPS ${framework_lib_deps}
-  SRCS ${src_dir}/${module}/*.h ${src_dir}/${module}/details/*.h ${PADDLE_BINARY_DIR}/paddle/fluid/framework/framework.pb.h
-       ${src_dir}/${module}/ir/*.h
-  DSTS ${dst_dir}/${module} ${dst_dir}/${module}/details ${dst_dir}/${module} ${dst_dir}/${module}/ir
-)
+        SRCS ${src_dir}/${module}/*.h ${src_dir}/${module}/details/*.h ${PADDLE_BINARY_DIR}/paddle/fluid/framework/framework.pb.h
+        ${src_dir}/${module}/ir/*.h
+        DSTS ${dst_dir}/${module} ${dst_dir}/${module}/details ${dst_dir}/${module} ${dst_dir}/${module}/ir
+        )
 
 set(module "memory")
 copy(memory_lib
-  SRCS ${src_dir}/${module}/*.h ${src_dir}/${module}/detail/*.h
-  DSTS ${dst_dir}/${module} ${dst_dir}/${module}/detail
-)
+        SRCS ${src_dir}/${module}/*.h ${src_dir}/${module}/detail/*.h
+        DSTS ${dst_dir}/${module} ${dst_dir}/${module}/detail
+        )
 
 set(inference_deps paddle_fluid_shared paddle_fluid)
 
 set(module "inference/api")
 if (WITH_ANAKIN AND WITH_MKL)
     copy(anakin_inference_lib DEPS paddle_inference_api inference_anakin_api
-        SRCS
-        ${PADDLE_BINARY_DIR}/paddle/fluid/inference/api/libinference_anakin_api* # compiled anakin api
-        ${ANAKIN_INSTALL_DIR} # anakin release
-        DSTS ${FLUID_INSTALL_DIR}/third_party/install/anakin ${FLUID_INSTALL_DIR}/third_party/install/anakin)
-     list(APPEND inference_deps anakin_inference_lib)
-endif()
+            SRCS
+            ${PADDLE_BINARY_DIR}/paddle/fluid/inference/api/libinference_anakin_api* # compiled anakin api
+            ${ANAKIN_INSTALL_DIR} # anakin release
+            DSTS ${FLUID_INSTALL_DIR}/third_party/install/anakin ${FLUID_INSTALL_DIR}/third_party/install/anakin)
+    list(APPEND inference_deps anakin_inference_lib)
+endif ()
 
 set(module "inference")
 copy(inference_lib DEPS ${inference_deps}
@@ -167,30 +188,30 @@ copy(inference_lib DEPS ${inference_deps}
        ${src_dir}/${module}/api/paddle_*.h
        ${PADDLE_BINARY_DIR}/paddle/fluid/inference/api/paddle_inference_pass.h
   DSTS ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module}
-)
+        )
 
 set(module "platform")
 copy(platform_lib DEPS profiler_py_proto
-  SRCS ${src_dir}/${module}/*.h ${src_dir}/${module}/dynload/*.h ${src_dir}/${module}/details/*.h
-  DSTS ${dst_dir}/${module} ${dst_dir}/${module}/dynload ${dst_dir}/${module}/details
-)
+        SRCS ${src_dir}/${module}/*.h ${src_dir}/${module}/dynload/*.h ${src_dir}/${module}/details/*.h
+        DSTS ${dst_dir}/${module} ${dst_dir}/${module}/dynload ${dst_dir}/${module}/details
+        )
 
 set(module "string")
 copy(string_lib
-  SRCS ${src_dir}/${module}/*.h ${src_dir}/${module}/tinyformat/*.h
-  DSTS ${dst_dir}/${module} ${dst_dir}/${module}/tinyformat
-)
+        SRCS ${src_dir}/${module}/*.h ${src_dir}/${module}/tinyformat/*.h
+        DSTS ${dst_dir}/${module} ${dst_dir}/${module}/tinyformat
+        )
 
 set(module "pybind")
 copy(pybind_lib
-  SRCS ${CMAKE_CURRENT_BINARY_DIR}/paddle/fluid/${module}/pybind.h
-  DSTS ${dst_dir}/${module}
-)
+        SRCS ${CMAKE_CURRENT_BINARY_DIR}/paddle/fluid/${module}/pybind.h
+        DSTS ${dst_dir}/${module}
+        )
 
 # CMakeCache Info
 copy(cmake_cache
-  SRCS ${CMAKE_CURRENT_BINARY_DIR}/CMakeCache.txt
-  DSTS ${FLUID_INSTALL_DIR})
+        SRCS ${CMAKE_CURRENT_BINARY_DIR}/CMakeCache.txt
+        DSTS ${FLUID_INSTALL_DIR})
 
 # This command generates a complete fluid library for both train and inference
 add_custom_target(fluid_lib_dist DEPENDS ${fluid_lib_dist_dep})
@@ -198,9 +219,9 @@ add_custom_target(fluid_lib_dist DEPENDS ${fluid_lib_dist_dep})
 # Following commands generate a inference-only fluid library
 # third_party, version.txt and CMakeCache.txt are the same position with ${FLUID_INSTALL_DIR}
 copy(third_party DEPS fluid_lib_dist
-  SRCS ${FLUID_INSTALL_DIR}/third_party ${FLUID_INSTALL_DIR}/CMakeCache.txt
-  DSTS ${FLUID_INFERENCE_INSTALL_DIR} ${FLUID_INFERENCE_INSTALL_DIR}
-)
+        SRCS ${FLUID_INSTALL_DIR}/third_party ${FLUID_INSTALL_DIR}/CMakeCache.txt
+        DSTS ${FLUID_INFERENCE_INSTALL_DIR} ${FLUID_INFERENCE_INSTALL_DIR}
+        )
 
 # only need libpaddle_fluid.so/a and paddle_*.h for inference-only library
 copy(inference_api_lib DEPS fluid_lib_dist
@@ -213,20 +234,20 @@ add_custom_target(inference_lib_dist DEPENDS third_party inference_api_lib)
 
 # paddle fluid version
 function(version version_file)
-  execute_process(
-    COMMAND ${GIT_EXECUTABLE} log --pretty=format:%H -1
-    WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}
-    OUTPUT_VARIABLE PADDLE_GIT_COMMIT)
-  file(WRITE ${version_file}
-    "GIT COMMIT ID: ${PADDLE_GIT_COMMIT}\n"
-    "WITH_MKL: ${WITH_MKL}\n"
-    "WITH_MKLDNN: ${WITH_MKLDNN}\n"
-    "WITH_GPU: ${WITH_GPU}\n")
-  if(WITH_GPU)
-    file(APPEND ${version_file}
-      "CUDA version: ${CUDA_VERSION}\n"
-      "CUDNN version: v${CUDNN_MAJOR_VERSION}\n")
-  endif()
+    execute_process(
+            COMMAND ${GIT_EXECUTABLE} log --pretty=format:%H -1
+            WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}
+            OUTPUT_VARIABLE PADDLE_GIT_COMMIT)
+    file(WRITE ${version_file}
+            "GIT COMMIT ID: ${PADDLE_GIT_COMMIT}\n"
+            "WITH_MKL: ${WITH_MKL}\n"
+            "WITH_MKLDNN: ${WITH_MKLDNN}\n"
+            "WITH_GPU: ${WITH_GPU}\n")
+    if (WITH_GPU)
+        file(APPEND ${version_file}
+                "CUDA version: ${CUDA_VERSION}\n"
+                "CUDNN version: v${CUDNN_MAJOR_VERSION}\n")
+    endif ()
 endfunction()
 version(${FLUID_INSTALL_DIR}/version.txt)
 version(${FLUID_INFERENCE_INSTALL_DIR}/version.txt)

doc/v2/dev/contribute_to_paddle_en.md

-../../../CONTRIBUTING.md
+../../../CONTRIBUTING.md
\ No newline at end of file

paddle/fluid/CMakeLists.txt

@@ -4,11 +4,12 @@ add_subdirectory(framework)
 add_subdirectory(operators)
 add_subdirectory(string)
 
-if (NOT WIN32)
 add_subdirectory(pybind)
+if (NOT WIN32)
 add_subdirectory(recordio)
 endif(NOT WIN32)
 
 # NOTE: please add subdirectory inference at last.
 add_subdirectory(inference)
+
 add_subdirectory(train)

paddle/fluid/framework/CMakeLists.txt

@@ -136,20 +136,32 @@ cc_library(version SRCS version.cc)
 cc_test(version_test SRCS version_test.cc DEPS version)
 
 cc_library(proto_desc SRCS var_desc.cc op_desc.cc block_desc.cc program_desc.cc DEPS shape_inference op_info operator glog version)
+cc_library(ngraph_bridge SRCS ngraph_bridge.cc DEPS operator framework_proto)
+if(NOT WIN32)
+cc_library(ngraph_operator SRCS ngraph_operator.cc DEPS ngraph_bridge operator op_info device_context tensor scope glog
+  shape_inference data_transform lod_tensor profiler)
+endif(NOT WIN32)
 
 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)
 add_dependencies(framework_py_proto framework_py_proto_init)
-add_custom_command(TARGET framework_py_proto POST_BUILD
-    COMMAND ${CMAKE_COMMAND} -E make_directory ${PADDLE_BINARY_DIR}/python/paddle/fluid/proto
-    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})
+if (NOT WIN32)
+  add_custom_command(TARGET framework_py_proto POST_BUILD
+      COMMAND ${CMAKE_COMMAND} -E make_directory ${PADDLE_BINARY_DIR}/python/paddle/fluid/proto
+      COMMAND cp *.py ${PADDLE_BINARY_DIR}/python/paddle/fluid/proto/
+      COMMENT "Copy generated python proto into directory paddle/fluid/proto."
+      WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
+else(NOT WIN32)
+  string(REPLACE "/" "\\" proto_dstpath "${PADDLE_BINARY_DIR}/python/paddle/fluid/proto/")
+  add_custom_command(TARGET framework_py_proto POST_BUILD
+          COMMAND ${CMAKE_COMMAND} -E make_directory ${PADDLE_BINARY_DIR}/python/paddle/fluid/proto
+          COMMAND copy /Y *.py ${proto_dstpath}
+          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)
@@ -163,10 +175,14 @@ if(WITH_DISTRIBUTE)
   set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
   set_source_files_properties(executor.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
 else()
-  cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass)
+  if(NOT WIN32)
+    cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass ngraph_operator)
+  else(NOT WIN32)
+    cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass)
+  endif(NOT WIN32)
   cc_test(test_naive_executor SRCS naive_executor_test.cc DEPS naive_executor elementwise_add_op)
 endif()
- 
+
 if (NOT WIN32)
 cc_library(parallel_executor SRCS parallel_executor.cc DEPS
         threaded_ssa_graph_executor scope_buffered_ssa_graph_executor

paddle/fluid/framework/details/build_strategy.cc

@@ -79,9 +79,15 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
   BuildStrategy strategy_;
 };
 
-std::shared_ptr<ir::PassBuilder> BuildStrategy::CreatePassesFromStrategy()
-    const {
+std::shared_ptr<ir::PassBuilder> BuildStrategy::CreatePassesFromStrategy(
+    bool finalize_strategy) const {
+  if (is_finalized_) {
+    return pass_builder_;
+  }
   pass_builder_.reset(new ParallelExecutorPassBuilder(*this));
+  if (finalize_strategy) {
+    is_finalized_ = true;
+  }
   return pass_builder_;
 }
 
@@ -95,10 +101,8 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
 #else
     const bool use_cuda) const {
 #endif
-  // Create a default one if not initialized by user.
-  if (!pass_builder_) {
-    CreatePassesFromStrategy();
-  }
+  // Create a default one if not finalized by user.
+  CreatePassesFromStrategy(false);
 
   std::unique_ptr<ir::Graph> graph(new ir::Graph(main_program));
 

paddle/fluid/framework/details/build_strategy.h

@@ -75,12 +75,20 @@ struct BuildStrategy {
 
   bool remove_unnecessary_lock_{false};
 
+  // NOTE:
+  // Before you add new options, think if it's a general strategy that works
+  // with other strategy. If not, the strategy should be created through
+  // CreatePassesFromStrategy and the pass can be managed separately.
+
   // User normally doesn't need to call this API.
   // The PassBuilder allows for more customized insert, remove of passes
   // from python side.
   // A new PassBuilder is created based on configs defined above and
   // passes are owned by the PassBuilder.
-  std::shared_ptr<ir::PassBuilder> CreatePassesFromStrategy() const;
+  std::shared_ptr<ir::PassBuilder> CreatePassesFromStrategy(
+      bool finalize_strategy) const;
+
+  bool IsFinalized() const { return is_finalized_; }
 
   // Apply the passes built by the pass_builder_. The passes will be
   // applied to the Program and output an ir::Graph.
@@ -97,6 +105,7 @@ struct BuildStrategy {
 #endif
 
  private:
+  mutable bool is_finalized_ = false;
   mutable std::shared_ptr<ir::PassBuilder> pass_builder_;
 };
 

paddle/fluid/framework/executor.cc

@@ -17,6 +17,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/feed_fetch_method.h"
 #include "paddle/fluid/framework/lod_rank_table.h"
 #include "paddle/fluid/framework/lod_tensor_array.h"
+#include "paddle/fluid/framework/ngraph_operator.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/framework/reader.h"
 #include "paddle/fluid/operators/detail/macros.h"
@@ -25,6 +26,7 @@ limitations under the License. */
 
 DECLARE_bool(benchmark);
 DEFINE_bool(use_mkldnn, false, "Use MKLDNN to run");
+DEFINE_bool(use_ngraph, false, "Use NGRAPH to run");
 
 namespace paddle {
 namespace framework {
@@ -81,6 +83,24 @@ static void DeleteUnusedTensors(const Scope& scope, const OperatorBase* op,
   }
 }
 
+static void EnableFusedOp(ExecutorPrepareContext* ctx) {
+#ifdef PADDLE_WITH_NGRAPH
+  VLOG(3) << "use_ngraph=True";
+  auto intervals = FusedOperator::FusedOpIntervals(&ctx->ops_);
+  for (auto& interval : intervals) {
+    auto* fused_op = new FusedOperator(ctx->prog_, ctx->block_id_,
+                                       interval.at(0), interval.at(1));
+    *interval[0] = std::unique_ptr<OperatorBase>(fused_op);
+  }
+  for (auto it = intervals.rbegin(); it != intervals.rend(); ++it) {
+    ctx->ops_.erase(it->at(0) + 1, it->at(1));
+  }
+#else
+  LOG(WARNING)
+      << "'NGRAPH' is not supported, Please re-compile with WITH_NGRAPH option";
+#endif
+}
+
 Executor::Executor(const platform::Place& place) : place_(place) {}
 
 void Executor::Close() {
@@ -338,6 +358,7 @@ std::unique_ptr<ExecutorPrepareContext> Executor::Prepare(
   for (auto& op_desc : block.AllOps()) {
     ctx->ops_.push_back(OpRegistry::CreateOp(*op_desc));
   }
+  if (FLAGS_use_ngraph) EnableFusedOp(ctx.get());
   return ctx;
 }
 
@@ -486,6 +507,5 @@ void Executor::EnableMKLDNN(const ProgramDesc& program) {
       << "'MKLDNN' is not supported, Please re-compile with WITH_MKLDNN option";
 #endif
 }
-
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/garbage_collector.h

@@ -29,7 +29,7 @@ template <typename T>
 class GarbageCollector {
  public:
   GarbageCollector(const platform::Place &place, size_t max_memory_size)
-      : max_memory_size_(std::max(max_memory_size, static_cast<size_t>(1))) {
+      : max_memory_size_((std::max)(max_memory_size, static_cast<size_t>(1))) {
     garbages_.reset(new std::deque<T *>());
     dev_ctx_ = platform::DeviceContextPool::Instance().Get(place);
   }

paddle/fluid/framework/ir/attention_lstm_fuse_pass.cc

@@ -211,12 +211,12 @@ void PrepareLSTMWeight(const LoDTensor& W_forget_w0,
   VLOG(30) << "LSTMWeight resized to " << out->dims();
 
   float* out_data = out->mutable_data<float>(platform::CPUPlace());
-  std::array<const float*, 4> tensors(
-      {{W_forget_w0.data<float>(), W_input_w0.data<float>(),
-        W_output_w0.data<float>(), W_cell_w0.data<float>()}});
-  std::array<const float*, 4> tensors1(
-      {{W_forget_w1.data<float>(), W_input_w1.data<float>(),
-        W_output_w1.data<float>(), W_cell_w1.data<float>()}});
+  std::array<const float*, 4> tensors{
+      W_forget_w0.data<float>(), W_input_w0.data<float>(),
+      W_output_w0.data<float>(), W_cell_w0.data<float>()};
+  std::array<const float*, 4> tensors1{
+      W_forget_w1.data<float>(), W_input_w1.data<float>(),
+      W_output_w1.data<float>(), W_cell_w1.data<float>()};
 
   for (int row = 0; row < D; row++) {
     for (int col = 0; col < 4; col++) {
@@ -238,9 +238,9 @@ void PrepareLSTMWeight(const LoDTensor& W_forget_w0,
 void PrepareLSTMBias(const LoDTensor& B_forget, const LoDTensor& B_input,
                      const LoDTensor& B_output, const LoDTensor& B_cell,
                      LoDTensor* out) {
-  std::array<const float*, 4> tensors(
-      {{B_forget.data<float>(), B_input.data<float>(), B_output.data<float>(),
-        B_cell.data<float>()}});
+  std::array<const float*, 4> tensors{
+      B_forget.data<float>(), B_input.data<float>(), B_output.data<float>(),
+      B_cell.data<float>()};
 
   PADDLE_ENFORCE_EQ(B_forget.dims().size(), 1);
   int D = B_forget.dims()[0];

paddle/fluid/framework/ir/node.cc

@@ -17,7 +17,12 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 namespace ir {
+// msvc15 don't support constexpr in correct way.
+#if !defined(_WIN32)
 constexpr char Node::kControlDepVarName[];
+#else
+const char Node::kControlDepVarName[] = "__control_var";
+#endif
 
 std::unique_ptr<Node> CreateNodeForTest(const std::string& name,
                                         Node::Type type) {

paddle/fluid/framework/ir/node.h

@@ -55,7 +55,11 @@ class Node {
   }
 
   enum class Type { kOperation, kVariable };
+#if !defined(_WIN32)  // msvc not support constexpr correctly.
   static constexpr char kControlDepVarName[] = "__control_var";
+#else
+  static const char kControlDepVarName[];
+#endif
 
   Type NodeType() const { return type_; }
 

paddle/fluid/framework/ir/pass.h

@@ -197,26 +197,26 @@ struct PassRegistrar : public Registrar {
                 msg)
 
 // Register a new pass that can be applied on the IR.
-#define REGISTER_PASS(pass_type, pass_class)                          \
-  STATIC_ASSERT_PASS_GLOBAL_NAMESPACE(                                \
-      __reg_pass__##pass_type,                                        \
-      "REGISTER_PASS must be called in global namespace");            \
-  static ::paddle::framework::ir::PassRegistrar<pass_class>           \
-      __pass_registrar_##pass_type##__(#pass_type);                   \
-  int TouchPassRegistrar_##pass_type() {                              \
-    __pass_registrar_##pass_type##__.Touch();                         \
-    return 0;                                                         \
-  }                                                                   \
-  static ::paddle::framework::ir::PassRegistrar<pass_class>           \
-      &__pass_tmp_registrar_##pass_type##__ __attribute__((unused)) = \
+#define REGISTER_PASS(pass_type, pass_class)                \
+  STATIC_ASSERT_PASS_GLOBAL_NAMESPACE(                      \
+      __reg_pass__##pass_type,                              \
+      "REGISTER_PASS must be called in global namespace");  \
+  static ::paddle::framework::ir::PassRegistrar<pass_class> \
+      __pass_registrar_##pass_type##__(#pass_type);         \
+  int TouchPassRegistrar_##pass_type() {                    \
+    __pass_registrar_##pass_type##__.Touch();               \
+    return 0;                                               \
+  }                                                         \
+  static ::paddle::framework::ir::PassRegistrar<pass_class> \
+      &__pass_tmp_registrar_##pass_type##__ UNUSED =        \
           __pass_registrar_##pass_type##__
 
-#define USE_PASS(pass_type)                                           \
-  STATIC_ASSERT_PASS_GLOBAL_NAMESPACE(                                \
-      __use_pass_itself_##pass_type,                                  \
-      "USE_PASS must be called in global namespace");                 \
-  extern int TouchPassRegistrar_##pass_type();                        \
-  static int use_pass_itself_##pass_type##_ __attribute__((unused)) = \
+#define USE_PASS(pass_type)                           \
+  STATIC_ASSERT_PASS_GLOBAL_NAMESPACE(                \
+      __use_pass_itself_##pass_type,                  \
+      "USE_PASS must be called in global namespace"); \
+  extern int TouchPassRegistrar_##pass_type();        \
+  static int use_pass_itself_##pass_type##_ UNUSED =  \
       TouchPassRegistrar_##pass_type()
 
 }  // namespace ir

paddle/fluid/framework/ngraph_bridge.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. */
+
+#ifdef PADDLE_WITH_NGRAPH
+#include <algorithm>
+#include <functional>
+
+#include "paddle/fluid/framework/ngraph_bridge.h"
+
+#include "ngraph/ngraph.hpp"
+
+namespace paddle {
+namespace framework {
+
+std::map<std::string,
+         std::function<void(const std::shared_ptr<OperatorBase>&,
+                            std::shared_ptr<std::unordered_map<
+                                std::string, std::shared_ptr<ngraph::Node>>>)>>
+    NgraphBridge::NG_NODE_MAP = {};
+
+void NgraphBridge::build_graph(const std::shared_ptr<OperatorBase>& op) {
+  auto& op_type = op->Type();
+  NG_NODE_MAP[op_type](op, ngb_node_map);
+}
+
+}  // namespace framework
+}  // namespace paddle
+#endif

paddle/fluid/framework/ngraph_bridge.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
+
+#ifdef PADDLE_WITH_NGRAPH
+
+#include <algorithm>
+#include <map>
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+#include "paddle/fluid/framework/operator.h"
+#include "paddle/fluid/platform/enforce.h"
+
+#include "ngraph/ngraph.hpp"
+
+namespace paddle {
+namespace framework {
+
+class NgraphBridge {
+ public:
+  static std::map<
+      std::string,
+      std::function<void(const std::shared_ptr<OperatorBase>&,
+                         std::shared_ptr<std::unordered_map<
+                             std::string, std::shared_ptr<ngraph::Node>>>)>>
+      NG_NODE_MAP;
+
+  explicit NgraphBridge(
+      std::shared_ptr<
+          std::unordered_map<std::string, std::shared_ptr<ngraph::Node>>>
+          var_node_map)
+      : ngb_node_map(var_node_map) {}
+
+  void build_graph(const std::shared_ptr<OperatorBase>& op);
+
+ private:
+  std::shared_ptr<
+      std::unordered_map<std::string, std::shared_ptr<ngraph::Node>>>
+      ngb_node_map;
+};
+
+}  // namespace framework
+}  // namespace paddle
+#endif

paddle/fluid/framework/ngraph_operator.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. */
+
+#ifdef PADDLE_WITH_NGRAPH
+#include <glog/logging.h>
+
+#include <algorithm>
+#include <map>
+
+#include "paddle/fluid/framework/feed_fetch_type.h"
+#include "paddle/fluid/framework/ngraph_operator.h"
+#include "paddle/fluid/framework/shape_inference.h"
+#include "paddle/fluid/framework/var_desc.h"
+#include "paddle/fluid/framework/var_type.h"
+
+namespace paddle {
+namespace framework {
+
+static std::map<proto::VarType::Type, ngraph::element::Type> pd2ng_type_map = {
+    {proto::VarType::FP32, ngraph::element::f32},
+    {proto::VarType::FP64, ngraph::element::f64},
+    {proto::VarType::INT32, ngraph::element::i32},
+    {proto::VarType::INT64, ngraph::element::i64},
+    {proto::VarType::BOOL, ngraph::element::boolean},
+};
+
+typedef enum {                /* nGraph support state on ops          */
+               FULL_TRAIN,    /* Support full ops for train           */
+               PARTIAL_TRAIN, /* Support partial ops for train        */
+               FULL_TEST,     /* Support full list of ops for test    */
+               PARTIAL_TEST   /* Support partial list of ops for test */
+} op_state;
+
+class NgraphOperator {
+ public:
+  explicit NgraphOperator(const Scope& scope, const platform::Place& place,
+                          const std::vector<std::shared_ptr<OperatorBase>>& ops,
+                          const std::unordered_map<
+                              std::string, ngraph::element::Type>& var_type_map,
+                          const std::unordered_set<std::string>& persist,
+                          const std::unordered_set<std::string>& fetches,
+                          const std::unordered_set<std::string>& post_op_inputs,
+                          op_state ng_op_state)
+      : scope_(scope),
+        place_(place),
+        fused_ops_(ops),
+        var_type_map_(var_type_map),
+        persistables_(persist),
+        fetches_(fetches),
+        post_op_inputs_(post_op_inputs),
+        ng_op_state_(ng_op_state) {}
+
+  void Run(const Scope& scope, const platform::Place& place) const;
+
+ private:
+  static std::unordered_map<std::string, std::shared_ptr<ngraph::Function>>
+      func_cache;
+  const Scope& scope_;
+  const platform::Place& place_;
+  std::vector<std::shared_ptr<OperatorBase>> fused_ops_;
+  std::unordered_map<std::string, ngraph::element::Type> var_type_map_;
+  std::unordered_set<std::string> persistables_;
+  std::unordered_set<std::string> fetches_;
+  std::unordered_set<std::string> post_op_inputs_;
+  op_state ng_op_state_;
+};
+
+std::vector<std::vector<std::vector<std::unique_ptr<OperatorBase>>::iterator>>
+FusedOperator::FusedOpIntervals(
+    std::vector<std::unique_ptr<paddle::framework::OperatorBase>>* ops) {
+  std::vector<std::vector<std::vector<std::unique_ptr<OperatorBase>>::iterator>>
+      intervals;
+  if (ops->empty()) {
+    return intervals;
+  }
+  size_t size = ops->size();
+  size_t left = 0;
+  while (left < size && ops.at(left)->Type() != kFeedOpType) {
+    ++left;
+  }
+  if (left == size) {
+    return intervals;
+  }
+  while (left < size && ops->at(left)->Type() == kFeedOpType) {
+    ++left;
+  }
+
+  size_t right = left;
+  while (right < size && ops->at(right)->Type() != kFetchOpType) {
+    ++right;
+  }
+  if (right == size) {
+    return intervals;
+  }
+  if (left >= right) return intervals;
+
+  // (left, right - 1) represents indices between feed and fetch
+  size_t pivot = left;
+  while (pivot < right) {
+    auto op_type = ops->at(pivot)->Type();
+    if (paddle::framework::NgraphBridge::NG_NODE_MAP.find(op_type) ==
+        paddle::framework::NgraphBridge::NG_NODE_MAP.end()) {
+      ++pivot;
+    } else {
+      size_t start = pivot, end = start;
+      while (pivot < right &&
+             (paddle::framework::NgraphBridge::NG_NODE_MAP.find(
+                  ops.at(pivot)->Type()) !=
+              paddle::framework::NgraphBridge::NG_NODE_MAP.end())) {
+        ++pivot;
+        ++end;
+      }
+      std::vector<std::vector<std::unique_ptr<OperatorBase>>::iterator>
+          interval = {ops->begin() + start, ops->begin() + end};
+      intervals.push_back(interval);
+    }
+  }  // end while
+
+  return intervals;
+}
+
+FusedOperator::FusedOperator(
+    const ProgramDesc& prog, size_t block_id,
+    std::vector<std::unique_ptr<OperatorBase>>::iterator start,
+    std::vector<std::unique_ptr<OperatorBase>>::iterator end,
+    const std::string& type, const VariableNameMap& inputs,
+    const VariableNameMap& outputs, const AttributeMap& attrs)
+    : OperatorBase(type, inputs, outputs, attrs), pdesc(prog), block(block_id) {
+  for (std::vector<std::unique_ptr<OperatorBase>>::iterator it = start;
+       it != end; ++it) {
+    fused_ops_.push_back(std::move(*it));
+  }
+
+  for (std::vector<std::unique_ptr<OperatorBase>>::iterator it = end;
+       (*it)->Type() != kFetchOpType; ++it) {
+    for (auto& var_name_item : (*it)->Inputs()) {
+      for (auto& var_name : var_name_item.second) {
+        post_op_inputs_.insert(var_name);
+      }
+    }
+  }
+
+  if ((*(start - 1))->Type() == kFeedOpType && (*end)->Type() == kFetchOpType) {
+    is_complete = true;
+  }
+
+  Process();
+}
+
+void FusedOperator::Process() {
+  auto& bdesc = pdesc_.Block(block_);
+  for (auto& var : bdesc.AllVars()) {
+    if (!(var->GetType() == proto::VarType::SELECTED_ROWS ||
+          var->GetType() == proto::VarType::LOD_TENSOR ||
+          var->GetType() == proto::VarType::LOD_TENSOR_ARRAY)) {
+      continue;
+    }
+
+    auto var_name = var->Name();
+    if (var->Name() == framework::kEmptyVarName) {
+      continue;
+    }
+
+    if (var_name != "fetch" && var_name != "feed") {
+      auto pd_type = var->GetDataType();
+      if (pd2ng_type_map.find(pd_type) == pd2ng_type_map.end()) {
+        PADDLE_THROW("Data type of var %s not found in pd2ng_type_map",
+                     var_name);
+      }
+      var_type_map_[var_name] = pd2ng_type_map[pd_type];
+    }
+
+    if (var->Persistable()) {
+      persistables_.insert(var->Name());
+    }
+  }
+
+  for (auto* op : bdesc.AllOps()) {
+    if (op->Type() == kFetchOpType) {
+      std::string fetch_target_name = op->Input("X")[0];
+      fetches_.insert(fetch_target_name);
+    }
+  }
+}
+
+void FusedOperator::RunImpl(const Scope& scope,
+                            const platform::Place& place) const {
+  op_state ng_op_state = PARTIAL_TEST;
+  auto& bdesc = pdesc_.Block(block_);
+  for (auto* op : bdesc.AllOps()) {
+    if (op->Type().find("_grad") != std::string::npos) {
+      ng_op_state = PARTIAL_TRAIN;
+      break;
+    }
+  }
+
+  if (is_full) {
+    ng_op_state = ng_op_state == PARTIAL_TEST ? FULL_TEST : FULL_TRAIN;
+  }
+
+  NgraphOperator ngraph_op(scope, place, fused_ops_, var_type_map_,
+                           persistables_, fetches_, post_op_inputs_,
+                           ng_op_state);
+  ngraph_op.Run(scope, place);
+}
+
+}  // namespace framework
+}  // namespace paddle
+#endif

paddle/fluid/framework/ngraph_operator.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
+
+#ifdef PADDLE_WITH_NGRAPH
+
+#include <algorithm>
+#include <atomic>
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+#include "paddle/fluid/framework/attribute.h"
+#include "paddle/fluid/framework/framework.pb.h"
+#include "paddle/fluid/framework/lod_tensor.h"
+#include "paddle/fluid/framework/ngraph_bridge.h"
+#include "paddle/fluid/framework/op_info.h"
+#include "paddle/fluid/framework/op_kernel_type.h"
+#include "paddle/fluid/framework/operator.h"
+#include "paddle/fluid/framework/program_desc.h"
+#include "paddle/fluid/framework/scope.h"
+#include "paddle/fluid/framework/tensor.h"
+#include "paddle/fluid/platform/variant.h"
+
+#include "ngraph/ngraph.hpp"
+
+namespace paddle {
+namespace framework {
+
+class FusedOperator : public OperatorBase {
+ public:
+  static std::vector<
+      std::vector<std::vector<std::unique_ptr<OperatorBase>>::iterator>>
+  FusedOpIntervals(
+      std::vector<std::unique_ptr<paddle::framework::OperatorBase>>* ops);
+
+  explicit FusedOperator(
+      const ProgramDesc& prog, size_t block_id,
+      std::vector<std::unique_ptr<OperatorBase>>::iterator start,
+      std::vector<std::unique_ptr<OperatorBase>>::iterator end,
+      const std::string& type = "fused_op", const VariableNameMap& inputs = {},
+      const VariableNameMap& outputs = {}, const AttributeMap& attrs = {});
+
+  void RunImpl(const Scope& scope, const platform::Place& place) const final;
+
+ private:
+  const ProgramDesc pdesc_;
+  size_t block_;
+  std::vector<std::shared_ptr<OperatorBase>> fused_ops_;
+  std::unordered_map<std::string, ngraph::element::Type> var_type_map_;
+  std::unordered_set<std::string> persistables_;
+  std::unordered_set<std::string> fetches_;
+  std::unordered_set<std::string> post_op_inputs_;
+  bool is_full_ = false;
+
+  void Process();
+};
+}  // namespace framework
+}  // namespace paddle
+#endif

paddle/fluid/framework/operator.cc

@@ -150,14 +150,17 @@ void OperatorBase::Run(const Scope& scope, const platform::Place& place) {
 #endif
   }
 
-  // The profile has a process-wide mutex, results in serious performance issue
-  // in concurrency scenerio. Here use an `if` to fix this issue.
-  // Please not remove the `if`, ask @Superjomn if there are any concern.
+// The profile has a process-wide mutex, results in serious performance issue
+// in concurrency scenerio. Here use an `if` to fix this issue.
+// Please not remove the `if`, ask @Superjomn if there are any concern.
+#ifndef _WIN32
   if (platform::IsProfileEnabled()) {
     platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
     platform::RecordEvent record_event(Type(), pool.Get(place));
     RunImpl(scope, place);
-  } else {
+  } else
+#endif
+  {
     RunImpl(scope, place);
   }
   VLOG(30) << place << " " << DebugStringEx(&scope);

paddle/fluid/inference/CMakeLists.txt

@@ -16,9 +16,21 @@ cc_library(paddle_fluid_api
     DEPS ${FLUID_CORE_MODULES} ${GLOB_OP_LIB}) 
 
 get_property(fluid_modules GLOBAL PROPERTY FLUID_MODULES)
+get_property(cuda_modules GLOBAL PROPERTY CUDA_MODULES)
+get_property(fluid_third_partys GLOBAL PROPERTY FLUID_THRID_PARTYS)
+if (WIN32)
+list(APPEND fluid_third_partys gflags glog protobuf cblas)
+endif(WIN32)
 
 # paddle_fluid_origin exclude inference api interface
-cc_library(paddle_fluid_origin DEPS ${fluid_modules} paddle_fluid_api)
+if(WIN32)
+  sep_library(paddle_fluid_origin DEPS ${fluid_modules} paddle_fluid_api)
+  if(WITH_GPU AND NOT WITH_DSO)
+    target_link_libraries(paddle_fluid_origin ${cuda_modules})
+  endif(WITH_GPU AND NOT WITH_DSO)
+else(WIN32)
+  cc_library(paddle_fluid_origin DEPS ${fluid_modules} paddle_fluid_api)
+endif(WIN32)
 
 add_subdirectory(api)
 
@@ -28,8 +40,16 @@ set(SHARED_INFERENCE_SRCS
     ${CMAKE_CURRENT_SOURCE_DIR}/api/analysis_predictor.cc
     ${CMAKE_CURRENT_SOURCE_DIR}/api/details/zero_copy_tensor.cc)
 
-# Create static library
-cc_library(paddle_fluid DEPS ${fluid_modules} ${STATIC_INFERENCE_APIS} zero_copy_tensor reset_tensor_array analysis_config paddle_pass_builder)
+if(WIN32)
+  sep_library(paddle_fluid DEPS ${fluid_modules} ${STATIC_INFERENCE_APIS} zero_copy_tensor reset_tensor_array
+    analysis_config paddle_pass_builder)
+  if(WITH_GPU AND NOT WITH_DSO)
+    target_link_libraries(paddle_fluid ${cuda_modules})
+  endif(WITH_GPU AND NOT WITH_DSO)
+else(WIN32)
+  cc_library(paddle_fluid DEPS ${fluid_modules} ${STATIC_INFERENCE_APIS} zero_copy_tensor reset_tensor_array
+    analysis_config paddle_pass_builder)
+endif(WIN32)
 
 if(NOT APPLE)
   # TODO(liuyiqu: Temporarily disable the link flag because it is not support on Mac.
@@ -38,11 +58,20 @@ if(NOT APPLE)
 endif()
 
 # Create shared library
-cc_library(paddle_fluid_shared SHARED SRCS ${SHARED_INFERENCE_SRCS}
-    DEPS ${fluid_modules} paddle_fluid_api reset_tensor_array analysis_config paddle_pass_builder)
+if(WIN32)
+  sep_library(paddle_fluid_shared SHARED SRCS ${SHARED_INFERENCE_SRCS}
+          DEPS ${fluid_modules} paddle_fluid_api reset_tensor_array analysis_config paddle_pass_builder)
+  target_link_libraries(paddle_fluid_shared shlwapi)
+  if(WITH_GPU AND NOT WITH_DSO)
+    target_link_libraries(paddle_fluid_origin ${cuda_modules})
+  endif(WITH_GPU AND NOT WITH_DSO)
+else(WIN32)
+  cc_library(paddle_fluid_shared SHARED SRCS ${SHARED_INFERENCE_SRCS}
+      DEPS ${fluid_modules} paddle_fluid_api reset_tensor_array analysis_config paddle_pass_builder)
+endif()
 
 set_target_properties(paddle_fluid_shared PROPERTIES OUTPUT_NAME paddle_fluid)
-if(NOT APPLE)
+if(NOT APPLE AND NOT WIN32)
   # TODO(liuyiqun): Temporarily disable the link flag because it is not support on Mac.
   set(LINK_FLAGS "-Wl,--version-script ${CMAKE_CURRENT_SOURCE_DIR}/paddle_fluid.map")
   set_target_properties(paddle_fluid_shared PROPERTIES LINK_FLAGS "${LINK_FLAGS}")

paddle/fluid/inference/analysis/helper.h

@@ -26,6 +26,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/framework/variable.h"
 #include "paddle/fluid/platform/enforce.h"
+#include "paddle/fluid/platform/port.h"
 
 namespace paddle {
 namespace inference {
@@ -124,20 +125,6 @@ T &GetFromScope(const framework::Scope &scope, const std::string &name) {
   return *var->GetMutable<T>();
 }
 
-static void ExecShellCommand(const std::string &cmd, std::string *message) {
-  char buffer[128];
-  std::shared_ptr<FILE> pipe(popen(cmd.c_str(), "r"), pclose);
-  if (!pipe) {
-    LOG(ERROR) << "error running command: " << cmd;
-    return;
-  }
-  while (!feof(pipe.get())) {
-    if (fgets(buffer, 128, pipe.get()) != nullptr) {
-      *message += buffer;
-    }
-  }
-}
-
 static framework::proto::ProgramDesc LoadProgramDesc(
     const std::string &model_path) {
   std::ifstream fin(model_path, std::ios::in | std::ios::binary);

paddle/fluid/inference/analysis/passes/ir_analysis_compose_pass.cc

@@ -45,7 +45,7 @@ void IrAnalysisComposePass::InitTensorRTAttrs(Argument *argument) {
       std::unordered_set<std::string> teller_set(
           {"mul", "conv2d", "pool2d", "relu", "softmax", "sigmoid",
            "depthwise_conv2d", "batch_norm", "concat", "tanh", "pad",
-           "elementwise_add", "dropout"});
+           "elementwise_add", "dropout", "split"});
       if (!node->IsOp()) return false;
 
       if (teller_set.count(node->Op()->Type())) {

paddle/fluid/inference/api/analysis_predictor.cc

@@ -548,4 +548,5 @@ USE_TRT_CONVERTER(batch_norm);
 USE_TRT_CONVERTER(concat);
 USE_TRT_CONVERTER(dropout);
 USE_TRT_CONVERTER(pad);
+USE_TRT_CONVERTER(split);
 #endif

paddle/fluid/inference/api/analysis_predictor_tester.cc

@@ -15,7 +15,7 @@
 #include "paddle/fluid/inference/api/analysis_predictor.h"
 #include <glog/logging.h>
 #include <gtest/gtest.h>
-#include <thread>
+#include <thread>  // NOLINT
 #include "paddle/fluid/inference/api/helper.h"
 #include "paddle/fluid/inference/api/paddle_inference_api.h"
 

paddle/fluid/inference/api/demo_ci/simple_on_word2vec.cc

@@ -23,7 +23,7 @@ limitations under the License. */
 #include <memory>
 #include <thread>  //NOLINT
 
-#include "utils.h"
+#include "utils.h"  // NOLINT
 
 DEFINE_string(dirname, "", "Directory of the inference model.");
 DEFINE_bool(use_gpu, false, "Whether use gpu.");

paddle/fluid/inference/api/demo_ci/trt_mobilenet_demo.cc

@@ -4,7 +4,7 @@ Licensed under the Apache License, Version 2.0 (the "License");
 you may not use this file except in compliance with the License.
 You may obtain a copy of the License at
 
-http://www.apache.org/licenses/LICENSE-2.0
+    http://www.apache.org/licenses/LICENSE-2.0
 
 Unless required by applicable law or agreed to in writing, software
 distributed under the License is distributed on an "AS IS" BASIS,

paddle/fluid/inference/api/helper.h

@@ -15,9 +15,14 @@
 #pragma once
 
 #include <glog/logging.h>
+#if !defined(_WIN32)
 #include <sys/time.h>
+#else
+#endif
+
 #include <algorithm>
 #include <chrono>  // NOLINT
+#include <iterator>
 #include <numeric>
 #include <sstream>
 #include <string>

paddle/fluid/inference/api/paddle_analysis_config.h

@@ -49,6 +49,8 @@ struct AnalysisConfig : public NativeConfig {
 
   void EnableTensorRtEngine(int workspace_size = 1 << 20,
                             int max_batch_size = 1);
+  bool use_tensorrt() const { return use_tensorrt_; }
+
   // NOTE this is just for internal development, please not use it.
   // NOT stable yet.
   void EnableMKLDNN();

paddle/fluid/inference/api/paddle_pass_builder.h

@@ -91,7 +91,7 @@ class CpuPassStrategy : public PassStrategy {
 
   virtual ~CpuPassStrategy() = default;
 
-  virtual void EnableMKLDNN() override {
+  void EnableMKLDNN() override {
 // TODO(Superjomn) Consider the way to mix CPU with GPU.
 #ifdef PADDLE_WITH_MKLDNN
     passes_.insert(passes_.begin(), "mkldnn_placement_pass");
@@ -123,7 +123,7 @@ class GpuPassStrategy : public PassStrategy {
   GpuPassStrategy(const GpuPassStrategy &other)
       : PassStrategy(other.AllPasses()) {}
 
-  virtual void EnableMKLDNN() override;
+  void EnableMKLDNN() override;
 
   virtual ~GpuPassStrategy() = default;
 };

paddle/fluid/inference/tensorrt/CMakeLists.txt

 nv_library(tensorrt_engine SRCS engine.cc DEPS framework_proto device_context)
 nv_test(test_tensorrt SRCS test_tensorrt.cc DEPS dynload_cuda device_context dynamic_loader)
 nv_test(test_tensorrt_engine SRCS test_engine.cc DEPS dynload_cuda tensorrt_engine)
+add_subdirectory(plugin)
 add_subdirectory(convert)

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 concat_op.cc dropout_op.cc pad_op.cc
-  DEPS tensorrt_engine operator scope framework_proto op_registry)
+batch_norm_op.cc activation_op.cc softmax_op.cc concat_op.cc dropout_op.cc
+pad_op.cc split_op.cc
+  DEPS tensorrt_engine tensorrt_plugin operator scope framework_proto op_registry)
 
 nv_test(test_op_converter SRCS test_op_converter.cc DEPS
   ${FLUID_CORE_MODULES} tensorrt_engine tensorrt_converter)
@@ -28,6 +29,8 @@ nv_test(test_trt_concat_op SRCS test_concat_op.cc concat_op.cc
         DEPS ${FLUID_CORE_MODULES} tensorrt_engine concat_op SERIAL)
 nv_test(test_trt_dropout_op SRCS test_dropout_op.cc dropout_op.cc
         DEPS ${FLUID_CORE_MODULES} tensorrt_engine dropout_op SERIAL)
-
 nv_test(test_trt_pad_op SRCS test_pad_op.cc pad_op.cc
         DEPS ${FLUID_CORE_MODULES} tensorrt_engine pad_op SERIAL)
+nv_test(test_trt_split_op SRCS test_split_op.cc split_op.cc
+        DEPS ${FLUID_CORE_MODULES} tensorrt_engine tensorrt_plugin
+split_op concat_op SERIAL)

paddle/fluid/inference/tensorrt/convert/concat_op.cc

@@ -19,7 +19,7 @@ namespace inference {
 namespace tensorrt {
 
 /*
- * MulOp, IMatrixMultiplyLayer in TRT. This Layer doesn't has weights.
+ * ConcatOp
  */
 class ConcatOpConverter : public OpConverter {
  public:

paddle/fluid/inference/tensorrt/convert/split_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"
+#include "paddle/fluid/inference/tensorrt/plugin/split_op_plugin.h"
+
+namespace paddle {
+namespace inference {
+namespace tensorrt {
+
+/*
+ * SplitOp.
+ */
+class SplitOpConverter : public OpConverter {
+ public:
+  void operator()(const framework::proto::OpDesc& op,
+                  const framework::Scope& scope, bool test_mode) override {
+    VLOG(40) << "convert a fluid split op to tensorrt split layer";
+
+    framework::OpDesc op_desc(op, nullptr);
+    // Declare inputs
+    auto* input = engine_->GetITensor(op_desc.Input("X")[0]);
+    auto input_dims = input->getDimensions();
+    int input_num = op_desc.Input("X").size();
+    size_t output_num = op_desc.Output("Out").size();
+
+    // Get Attrs
+    PADDLE_ENFORCE(input_num == 1);
+    int axis = boost::get<int>(op_desc.GetAttr("axis"));
+    std::vector<int> output_lengths =
+        boost::get<std::vector<int>>(op_desc.GetAttr("sections"));
+    PADDLE_ENFORCE(axis != 0);
+    if (axis < 0) {
+      axis += input_dims.nbDims;
+    } else {
+      axis -= 1;
+    }
+
+    PADDLE_ENFORCE(output_lengths.size() == output_num);
+
+    //
+    SplitPlugin* plugin = new SplitPlugin(axis, output_lengths);
+    nvinfer1::IPluginLayer* layer =
+        engine_->AddPlugin(&input, input_num, plugin);
+
+    std::string layer_name = "split (Output: ";
+    for (size_t i = 0; i < output_num; i++) {
+      auto output_name = op_desc.Output("Out")[i];
+      layer->getOutput(i)->setName(output_name.c_str());
+      engine_->SetITensor(output_name, layer->getOutput(i));
+      layer_name += output_name;
+      if (test_mode) {
+        engine_->DeclareOutput(output_name);
+      }
+    }
+    layer->setName((layer_name + ")").c_str());
+  }
+};
+
+}  // namespace tensorrt
+}  // namespace inference
+}  // namespace paddle
+
+REGISTER_TRT_OP_CONVERTER(split, SplitOpConverter);

paddle/fluid/inference/tensorrt/convert/test_split_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(split_op, test) {
+  std::unordered_set<std::string> parameters({""});
+  framework::Scope scope;
+  TRTConvertValidation validator(10, parameters, scope, 1000);
+  validator.DeclInputVar("split_input", nvinfer1::DimsCHW(3, 2, 2));
+  validator.DeclOutputVar("split_out1", nvinfer1::DimsCHW(2, 2, 2));
+  validator.DeclOutputVar("split_out2", nvinfer1::DimsCHW(1, 2, 2));
+
+  // Prepare Op description
+  framework::OpDesc desc;
+  desc.SetType("split");
+  desc.SetInput("X", {"split_input"});
+  desc.SetOutput("Out", {"split_out1", "split_out2"});
+
+  int num = 0;
+  int axis = 1;
+  std::vector<int> output_lengths = {2, 1};
+  desc.SetAttr("axis", axis);
+  desc.SetAttr("num", num);
+  desc.SetAttr("sections", output_lengths);
+
+  validator.SetOp(*desc.Proto());
+
+  validator.Execute(1);
+}
+
+}  // namespace tensorrt
+}  // namespace inference
+}  // namespace paddle
+
+USE_OP(split);

paddle/fluid/inference/tensorrt/engine.cc

@@ -255,6 +255,12 @@ void TensorRTEngine::freshDeviceId() {
   cudaSetDevice(device_);
 }
 
+nvinfer1::IPluginLayer *TensorRTEngine::AddPlugin(
+    nvinfer1::ITensor *const *inputs, int nbInputs, PluginTensorRT *plugin) {
+  owned_plugin_.emplace_back(plugin);
+  return infer_network_.get()->addPluginExt(inputs, nbInputs, *plugin);
+}
+
 }  // namespace tensorrt
 }  // namespace inference
 }  // namespace paddle

paddle/fluid/inference/tensorrt/engine.h

@@ -22,6 +22,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/tensor.h"
 #include "paddle/fluid/inference/engine.h"
 #include "paddle/fluid/inference/tensorrt/helper.h"
+#include "paddle/fluid/inference/tensorrt/plugin/trt_plugin.h"
 #include "paddle/fluid/inference/utils/singleton.h"
 
 namespace paddle {
@@ -125,6 +126,8 @@ class TensorRTEngine : public EngineBase {
   void SetRuntimeBatch(size_t batch_size);
   int GetRuntimeBatch();
   int GetDevice() { return device_; }
+  nvinfer1::IPluginLayer* AddPlugin(nvinfer1::ITensor* const* inputs,
+                                    int nbInputs, PluginTensorRT*);
 
   // A pointer to CPU memory is needed of the TRT weight.
   // Before TRT runs, fluid loads weight into GPU storage.
@@ -164,8 +167,10 @@ class TensorRTEngine : public EngineBase {
   std::unordered_map<std::string /*name*/, size_t /*max size*/> buffer_sizes_;
   std::unordered_map<std::string /*name*/, nvinfer1::ITensor* /*ITensor*/>
       itensor_map_;
+
   // The specific GPU id that the TensorRTEngine bounded to.
   int device_;
+  std::vector<std::unique_ptr<PluginTensorRT>> owned_plugin_;
 
   // TensorRT related internal members
   template <typename T>

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

+nv_library(tensorrt_plugin SRCS trt_plugin.cc split_op_plugin.cu DEPS enforce)

paddle/fluid/inference/tensorrt/plugin/serialize.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 <cassert>
+#include <cstring>
+#include <type_traits>
+#include <vector>
+
+template <typename T>
+inline void SerializeValue(void** buffer, T const& value);
+
+template <typename T>
+inline void DeserializeValue(void const** buffer, size_t* buffer_size,
+                             T* value);
+
+namespace {
+
+template <typename T, class Enable = void>
+struct Serializer {};
+
+template <typename T>
+struct Serializer<T, typename std::enable_if<std::is_arithmetic<T>::value ||
+                                             std::is_enum<T>::value ||
+                                             std::is_pod<T>::value>::type> {
+  static size_t SerializedSize(T const& value) { return sizeof(T); }
+  static void Serialize(void** buffer, T const& value) {
+    std::memcpy(*buffer, &value, sizeof(T));
+    reinterpret_cast<char*&>(*buffer) += sizeof(T);
+  }
+  static void Deserialize(void const** buffer, size_t* buffer_size, T* value) {
+    assert(*buffer_size >= sizeof(T));
+    std::memcpy(value, *buffer, sizeof(T));
+    reinterpret_cast<char const*&>(*buffer) += sizeof(T);
+    *buffer_size -= sizeof(T);
+  }
+};
+
+template <>
+struct Serializer<const char*> {
+  static size_t SerializedSize(const char* value) { return strlen(value) + 1; }
+  static void Serialize(void** buffer, const char* value) {
+    std::strcpy(static_cast<char*>(*buffer), value);
+    reinterpret_cast<char*&>(*buffer) += strlen(value) + 1;
+  }
+  static void Deserialize(void const** buffer, size_t* buffer_size,
+                          const char** value) {
+    *value = static_cast<char const*>(*buffer);
+    size_t data_size = strnlen(*value, *buffer_size) + 1;
+    assert(*buffer_size >= data_size);
+    reinterpret_cast<char const*&>(*buffer) += data_size;
+    *buffer_size -= data_size;
+  }
+};
+
+template <typename T>
+struct Serializer<std::vector<T>,
+                  typename std::enable_if<std::is_arithmetic<T>::value ||
+                                          std::is_enum<T>::value ||
+                                          std::is_pod<T>::value>::type> {
+  static size_t SerializedSize(std::vector<T> const& value) {
+    return sizeof(value.size()) + value.size() * sizeof(T);
+  }
+  static void Serialize(void** buffer, std::vector<T> const& value) {
+    SerializeValue(buffer, value.size());
+    size_t nbyte = value.size() * sizeof(T);
+    std::memcpy(*buffer, value.data(), nbyte);
+    reinterpret_cast<char*&>(*buffer) += nbyte;
+  }
+  static void Deserialize(void const** buffer, size_t* buffer_size,
+                          std::vector<T>* value) {
+    size_t size;
+    DeserializeValue(buffer, buffer_size, &size);
+    value->resize(size);
+    size_t nbyte = value->size() * sizeof(T);
+    assert(*buffer_size >= nbyte);
+    std::memcpy(value->data(), *buffer, nbyte);
+    reinterpret_cast<char const*&>(*buffer) += nbyte;
+    *buffer_size -= nbyte;
+  }
+};
+
+}  // namespace
+
+template <typename T>
+inline size_t SerializedSize(T const& value) {
+  return Serializer<T>::SerializedSize(value);
+}
+
+template <typename T>
+inline void SerializeValue(void** buffer, T const& value) {
+  return Serializer<T>::Serialize(buffer, value);
+}
+
+template <typename T>
+inline void DeserializeValue(void const** buffer, size_t* buffer_size,
+                             T* value) {
+  return Serializer<T>::Deserialize(buffer, buffer_size, value);
+}

paddle/fluid/inference/tensorrt/plugin/split_op_plugin.cu

+// 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 <stdio.h>
+#include <cassert>
+#include "paddle/fluid/inference/tensorrt/plugin/split_op_plugin.h"
+
+namespace paddle {
+namespace inference {
+namespace tensorrt {
+
+nvinfer1::Dims SplitPlugin::getOutputDimensions(int index,
+                                                const nvinfer1::Dims* inputDims,
+                                                int nbInputs) {
+  assert(nbInputs == 1);
+  assert(index < this->getNbOutputs());
+  nvinfer1::Dims const& input_dims = inputDims[0];
+  nvinfer1::Dims output_dims = input_dims;
+  output_dims.d[axis_] = output_length_.at(index);
+  return output_dims;
+}
+
+int SplitPlugin::initialize() {
+  std::vector<int> segment_offsets(1, 0);
+  for (int i = 0; i < this->getNbOutputs(); ++i) {
+    segment_offsets.push_back(segment_offsets.back() + output_length_[i]);
+  }
+  segment_offsets_ = segment_offsets;
+  nvinfer1::Dims dims = this->getInputDims(0);
+  nx_ = 1;
+  for (int i = dims.nbDims - 1; i > axis_; --i) {
+    nx_ *= dims.d[i];
+  }
+  ny_ = dims.d[axis_];
+  nz_ = 1;
+  for (int i = axis_ - 1; i >= 0; --i) {
+    nz_ *= dims.d[i];
+  }
+  return 0;
+}
+
+int SplitPlugin::enqueue(int batchSize, const void* const* inputs,
+                         void** outputs, void* workspace, cudaStream_t stream) {
+  auto const& input_dims = this->getInputDims(0);
+  int input_size = 0;
+  float const* idata = reinterpret_cast<float const*>(inputs[0]);
+  float** odatas = reinterpret_cast<float**>(outputs);
+
+  // kernel impl here.
+  int inputBatchOffset = nx_ * ny_ * nz_;
+  for (size_t i = 0; i < this->getNbOutputs(); i++) {
+    for (size_t j = 0; j < batchSize; j++) {
+      cudaMemcpyAsync(
+          odatas[i] +
+              j * (segment_offsets_[i + 1] - segment_offsets_[i]) * nx_ *
+                  sizeof(float),
+          inputs[0] +
+              (inputBatchOffset * j + segment_offsets_[i] * nx_) *
+                  sizeof(float),
+          (segment_offsets_[i + 1] - segment_offsets_[i]) * nx_ * sizeof(float),
+          cudaMemcpyDeviceToDevice, stream);
+    }
+  }
+
+  return cudaGetLastError() != cudaSuccess;
+}
+
+}  // tensorrt
+}  // inference
+}  // paddle

paddle/fluid/inference/tensorrt/plugin/split_op_plugin.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/inference/tensorrt/plugin/trt_plugin.h"
+
+namespace paddle {
+namespace inference {
+namespace tensorrt {
+
+class SplitPlugin : public PluginTensorRT {
+  int axis_;
+  std::vector<int> output_length_;
+  int nx_, ny_, nz_;
+  std::vector<int> segment_offsets_;
+
+ protected:
+  virtual size_t getSerializationSize() override {
+    return SerializedSize(axis_) + SerializedSize(output_length_) +
+           getBaseSerializationSize();
+  }
+
+  // TRT will call this func when we need to serialize the configuration of
+  // tensorrt.
+  // It should not be called by users.
+  virtual void serialize(void *buffer) override {
+    serializeBase(buffer);
+    SerializeValue(&buffer, axis_);
+    SerializeValue(&buffer, output_length_);
+  }
+
+ public:
+  SplitPlugin(int axis, std::vector<int> const &output_lengths)
+      : axis_(axis), output_length_(output_lengths) {
+    assert(axis <= nvinfer1::Dims::MAX_DIMS);
+  }
+
+  // It was used for tensorrt deserialization.
+  // It should not be called by users.
+  SplitPlugin(void const *serialData, size_t serialLength) {
+    deserializeBase(serialData, serialLength);
+    DeserializeValue(&serialData, &serialLength, &axis_);
+    DeserializeValue(&serialData, &serialLength, &output_length_);
+  }
+
+  SplitPlugin *clone() const override {
+    return new SplitPlugin(axis_, output_length_);
+  }
+
+  virtual const char *getPluginType() const override { return "split"; }
+  virtual int getNbOutputs() const override { return output_length_.size(); }
+  virtual nvinfer1::Dims getOutputDimensions(int index,
+                                             const nvinfer1::Dims *inputs,
+                                             int nbInputDims) override;
+  virtual int initialize() override;
+  virtual int enqueue(int batchSize, const void *const *inputs, void **outputs,
+                      void *workspace, cudaStream_t stream) override;
+};
+
+}  // tensorrt
+}  // inference
+}  // paddle

paddle/fluid/inference/tensorrt/plugin/trt_plugin.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/plugin/trt_plugin.h"
+
+namespace paddle {
+namespace inference {
+namespace tensorrt {
+
+void PluginTensorRT::serializeBase(void*& buffer) {
+  SerializeValue(&buffer, input_dims_);
+  SerializeValue(&buffer, max_batch_size_);
+  SerializeValue(&buffer, data_type_);
+  SerializeValue(&buffer, data_format_);
+}
+
+void PluginTensorRT::deserializeBase(void const*& serialData,
+                                     size_t& serialLength) {
+  DeserializeValue(&serialData, &serialLength, &input_dims_);
+  DeserializeValue(&serialData, &serialLength, &max_batch_size_);
+  DeserializeValue(&serialData, &serialLength, &data_type_);
+  DeserializeValue(&serialData, &serialLength, &data_format_);
+}
+
+size_t PluginTensorRT::getBaseSerializationSize() {
+  return (SerializedSize(input_dims_) + SerializedSize(max_batch_size_) +
+          SerializedSize(data_type_) + SerializedSize(data_format_));
+}
+
+bool PluginTensorRT::supportsFormat(nvinfer1::DataType type,
+                                    nvinfer1::PluginFormat format) const {
+  return ((type == nvinfer1::DataType::kFLOAT) &&
+          (format == nvinfer1::PluginFormat::kNCHW));
+}
+
+void PluginTensorRT::configureWithFormat(const nvinfer1::Dims* inputDims,
+                                         int nbInputs,
+                                         const nvinfer1::Dims* outputDims,
+                                         int nbOutputs, nvinfer1::DataType type,
+                                         nvinfer1::PluginFormat format,
+                                         int maxBatchSize) {
+  data_type_ = type;
+  data_format_ = format;
+  input_dims_.assign(inputDims, inputDims + nbInputs);
+  max_batch_size_ = maxBatchSize;
+}
+
+}  // namespace tensorrt
+}  // namespace inference
+}  // namespace paddle

paddle/fluid/inference/tensorrt/plugin/trt_plugin.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 <cassert>
+#include <cstring>
+#include <iostream>
+#include <unordered_map>
+#include <vector>
+#include "NvInfer.h"
+
+#include "paddle/fluid/inference/tensorrt/plugin/serialize.h"
+
+namespace paddle {
+namespace inference {
+namespace tensorrt {
+
+class PluginTensorRT : public nvinfer1::IPluginExt {
+ public:
+  PluginTensorRT() {}
+  PluginTensorRT(const void* serialized_data, size_t length) {}
+  nvinfer1::Dims const& getInputDims(int index) const {
+    return input_dims_.at(index);
+  }
+  size_t getMaxBatchSize() const { return max_batch_size_; }
+  nvinfer1::DataType getDataType() const { return data_type_; }
+  nvinfer1::PluginFormat getDataFormat() const { return data_format_; }
+  virtual const char* getPluginVersion() const { return "1"; }
+  size_t getWorkspaceSize(int) const override { return 0; }
+  void terminate() override {}
+  virtual ~PluginTensorRT() {}
+  // Check format support. The default is FLOAT32 and NCHW.
+  bool supportsFormat(nvinfer1::DataType type,
+                      nvinfer1::PluginFormat format) const override;
+  void configureWithFormat(const nvinfer1::Dims* inputDims, int nbInputs,
+                           const nvinfer1::Dims* outputDims, int nbOutputs,
+                           nvinfer1::DataType type,
+                           nvinfer1::PluginFormat format,
+                           int maxBatchSize) override;
+
+  // *NOTE* The following functions need to be overrided in the subclass.
+  virtual nvinfer1::IPluginExt* clone() const = 0;
+  virtual const char* getPluginType() const = 0;
+  // Initialize the layer for execution. This is called when the engine is
+  // created.
+  int initialize() override { return 0; }
+  // Serialize the layer config to buffer.
+  virtual void serialize(void* buffer) = 0;
+  virtual size_t getSerializationSize() = 0;
+  virtual int enqueue(int batchSize, const void* const* inputs, void** outputs,
+                      void* workspace, cudaStream_t stream) = 0;
+
+ protected:
+  // Deserialize input_dims, max_batch_size, data_type, data_format
+  void deserializeBase(void const*& serialData, size_t& serialLength);
+  size_t getBaseSerializationSize();
+  // Serialize input_dims, max_batch_size, data_type, data_format
+  void serializeBase(void*& buffer);
+
+  std::vector<nvinfer1::Dims> input_dims_;
+  size_t max_batch_size_;
+  nvinfer1::DataType data_type_;
+  nvinfer1::PluginFormat data_format_;
+};
+
+}  // namespace tensorrt
+}  // namespace inference
+}  // namespace paddle

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

@@ -104,8 +104,7 @@ if(WITH_GPU AND TENSORRT_FOUND)
    if (NOT EXISTS ${TRT_MODEL_INSTALL_DIR})
        inference_download_and_uncompress(${TRT_MODEL_INSTALL_DIR} ${INFERENCE_URL}/tensorrt_test "trt_test_models.tar.gz")
    endif()
-
    inference_analysis_test(test_trt_models SRCS trt_models_tester.cc
       EXTRA_DEPS ${INFERENCE_EXTRA_DEPS} analysis ${analysis_deps} ir_pass_manager analysis_predictor
-        ARGS --dirname=${TRT_MODEL_INSTALL_DIR}/trt_test_models SERIAL)
+        ARGS --infer_model=${TRT_MODEL_INSTALL_DIR}/trt_test_models SERIAL)
 endif()

paddle/fluid/inference/tests/api/analyzer_dam_tester.cc

@@ -178,7 +178,8 @@ TEST(Analyzer_dam, profile) {
   std::vector<PaddleTensor> outputs;
   std::vector<std::vector<PaddleTensor>> input_slots_all;
   SetInput(&input_slots_all);
-  TestPrediction(cfg, input_slots_all, &outputs, FLAGS_num_threads);
+  TestPrediction(reinterpret_cast<const PaddlePredictor::Config *>(&cfg),
+                 input_slots_all, &outputs, FLAGS_num_threads);
 
   if (FLAGS_num_threads == 1 && !FLAGS_test_all_data) {
     PADDLE_ENFORCE_GT(outputs.size(), 0);
@@ -213,7 +214,8 @@ TEST(Analyzer_dam, compare) {
   std::vector<std::vector<PaddleTensor>> input_slots_all;
   SetInput(&input_slots_all);
 
-  CompareNativeAndAnalysis(cfg, input_slots_all);
+  CompareNativeAndAnalysis(
+      reinterpret_cast<const PaddlePredictor::Config *>(&cfg), input_slots_all);
 }
 
 }  // namespace inference

paddle/fluid/inference/tests/api/analyzer_lac_tester.cc

@@ -133,7 +133,8 @@ TEST(Analyzer_LAC, profile) {
 
   std::vector<std::vector<PaddleTensor>> input_slots_all;
   SetInput(&input_slots_all);
-  TestPrediction(cfg, input_slots_all, &outputs, FLAGS_num_threads);
+  TestPrediction(reinterpret_cast<const PaddlePredictor::Config *>(&cfg),
+                 input_slots_all, &outputs, FLAGS_num_threads);
 
   if (FLAGS_num_threads == 1 && !FLAGS_test_all_data) {
     // the first inference result
@@ -175,7 +176,8 @@ TEST(Analyzer_LAC, compare) {
 
   std::vector<std::vector<PaddleTensor>> input_slots_all;
   SetInput(&input_slots_all);
-  CompareNativeAndAnalysis(cfg, input_slots_all);
+  CompareNativeAndAnalysis(
+      reinterpret_cast<const PaddlePredictor::Config *>(&cfg), input_slots_all);
 }
 
 }  // namespace analysis

paddle/fluid/inference/tests/api/analyzer_ner_tester.cc

@@ -121,7 +121,8 @@ TEST(Analyzer_Chinese_ner, profile) {
 
   std::vector<std::vector<PaddleTensor>> input_slots_all;
   SetInput(&input_slots_all);
-  TestPrediction(cfg, input_slots_all, &outputs, FLAGS_num_threads);
+  TestPrediction(reinterpret_cast<const PaddlePredictor::Config *>(&cfg),
+                 input_slots_all, &outputs, FLAGS_num_threads);
 
   if (FLAGS_num_threads == 1 && !FLAGS_test_all_data) {
     // the first inference result
@@ -160,7 +161,8 @@ TEST(Analyzer_Chinese_ner, compare) {
 
   std::vector<std::vector<PaddleTensor>> input_slots_all;
   SetInput(&input_slots_all);
-  CompareNativeAndAnalysis(cfg, input_slots_all);
+  CompareNativeAndAnalysis(
+      reinterpret_cast<const PaddlePredictor::Config *>(&cfg), input_slots_all);
 }
 
 }  // namespace inference

paddle/fluid/inference/tests/api/analyzer_resnet50_tester.cc

@@ -45,7 +45,8 @@ void profile(bool use_mkldnn = false) {
 
   std::vector<std::vector<PaddleTensor>> input_slots_all;
   SetInput(&input_slots_all);
-  TestPrediction(cfg, input_slots_all, &outputs, FLAGS_num_threads);
+  TestPrediction(reinterpret_cast<const PaddlePredictor::Config *>(&cfg),
+                 input_slots_all, &outputs, FLAGS_num_threads);
 }
 
 TEST(Analyzer_resnet50, profile) { profile(); }
@@ -74,7 +75,8 @@ void compare(bool use_mkldnn = false) {
 
   std::vector<std::vector<PaddleTensor>> input_slots_all;
   SetInput(&input_slots_all);
-  CompareNativeAndAnalysis(cfg, input_slots_all);
+  CompareNativeAndAnalysis(
+      reinterpret_cast<const PaddlePredictor::Config *>(&cfg), input_slots_all);
 }
 
 TEST(Analyzer_resnet50, compare) { compare(); }

paddle/fluid/inference/tests/api/analyzer_rnn1_tester.cc

@@ -233,8 +233,8 @@ TEST(Analyzer_rnn1, profile) {
 
   std::vector<std::vector<PaddleTensor>> input_slots_all;
   SetInput(&input_slots_all);
-  LOG(INFO) << "to test prediction";
-  TestPrediction(cfg, input_slots_all, &outputs, FLAGS_num_threads);
+  TestPrediction(reinterpret_cast<const PaddlePredictor::Config *>(&cfg),
+                 input_slots_all, &outputs, FLAGS_num_threads);
 }
 
 // Check the fuse status
@@ -261,7 +261,8 @@ TEST(Analyzer_rnn1, compare) {
 
   std::vector<std::vector<PaddleTensor>> input_slots_all;
   SetInput(&input_slots_all);
-  CompareNativeAndAnalysis(cfg, input_slots_all);
+  CompareNativeAndAnalysis(
+      reinterpret_cast<const PaddlePredictor::Config *>(&cfg), input_slots_all);
 }
 
 // Test Multi-Thread.
@@ -272,7 +273,8 @@ TEST(Analyzer_rnn1, multi_thread) {
 
   std::vector<std::vector<PaddleTensor>> input_slots_all;
   SetInput(&input_slots_all);
-  TestPrediction(cfg, input_slots_all, &outputs, 4 /* multi_thread */);
+  TestPrediction(reinterpret_cast<const PaddlePredictor::Config *>(&cfg),
+                 input_slots_all, &outputs, 4 /* multi_thread */);
 }
 
 // Validate that the AnalysisPredictor + ZeroCopyTensor really works by testing

paddle/fluid/inference/tests/api/analyzer_rnn2_tester.cc

@@ -132,7 +132,8 @@ TEST(Analyzer_rnn2, profile) {
 
   std::vector<std::vector<PaddleTensor>> input_slots_all;
   SetInput(&input_slots_all);
-  TestPrediction(cfg, input_slots_all, &outputs, FLAGS_num_threads);
+  TestPrediction(reinterpret_cast<const PaddlePredictor::Config *>(&cfg),
+                 input_slots_all, &outputs, FLAGS_num_threads);
 
   if (FLAGS_num_threads == 1 && !FLAGS_test_all_data) {
     // the first inference result
@@ -153,7 +154,8 @@ TEST(Analyzer_rnn2, compare) {
 
   std::vector<std::vector<PaddleTensor>> input_slots_all;
   SetInput(&input_slots_all);
-  CompareNativeAndAnalysis(cfg, input_slots_all);
+  CompareNativeAndAnalysis(
+      reinterpret_cast<const PaddlePredictor::Config *>(&cfg), input_slots_all);
 }
 
 }  // namespace inference

paddle/fluid/inference/tests/api/analyzer_seq_conv1_tester.cc

@@ -161,7 +161,8 @@ TEST(Analyzer_seq_conv1, profile) {
 
   std::vector<std::vector<PaddleTensor>> input_slots_all;
   SetInput(&input_slots_all);
-  TestPrediction(cfg, input_slots_all, &outputs, FLAGS_num_threads);
+  TestPrediction(reinterpret_cast<const PaddlePredictor::Config *>(&cfg),
+                 input_slots_all, &outputs, FLAGS_num_threads);
 
   if (FLAGS_num_threads == 1 && !FLAGS_test_all_data) {
     // the first inference result
@@ -198,7 +199,8 @@ TEST(Analyzer_seq_conv1, compare) {
 
   std::vector<std::vector<PaddleTensor>> input_slots_all;
   SetInput(&input_slots_all);
-  CompareNativeAndAnalysis(cfg, input_slots_all);
+  CompareNativeAndAnalysis(
+      reinterpret_cast<const PaddlePredictor::Config *>(&cfg), input_slots_all);
 }
 
 }  // namespace inference

paddle/fluid/inference/tests/api/analyzer_text_classification_tester.cc

@@ -74,7 +74,8 @@ TEST(Analyzer_Text_Classification, profile) {
 
   std::vector<std::vector<PaddleTensor>> input_slots_all;
   SetInput(&input_slots_all);
-  TestPrediction(cfg, input_slots_all, &outputs, FLAGS_num_threads);
+  TestPrediction(reinterpret_cast<const PaddlePredictor::Config *>(&cfg),
+                 input_slots_all, &outputs, FLAGS_num_threads);
 
   if (FLAGS_num_threads == 1) {
     // Get output
@@ -101,7 +102,8 @@ TEST(Analyzer_Text_Classification, compare) {
 
   std::vector<std::vector<PaddleTensor>> input_slots_all;
   SetInput(&input_slots_all);
-  CompareNativeAndAnalysis(cfg, input_slots_all);
+  CompareNativeAndAnalysis(
+      reinterpret_cast<const PaddlePredictor::Config *>(&cfg), input_slots_all);
 }
 
 TEST(Analyzer_Text_Classification, compare_against_embedding_fc_lstm_fused) {
@@ -112,7 +114,8 @@ TEST(Analyzer_Text_Classification, compare_against_embedding_fc_lstm_fused) {
 
   std::vector<std::vector<PaddleTensor>> input_slots_all;
   SetInput(&input_slots_all);
-  CompareNativeAndAnalysis(cfg, input_slots_all);
+  CompareNativeAndAnalysis(
+      reinterpret_cast<const PaddlePredictor::Config *>(&cfg), input_slots_all);
 }
 
 }  // namespace inference

paddle/fluid/inference/tests/api/analyzer_vis_tester.cc

@@ -91,7 +91,8 @@ void profile(bool use_mkldnn = false) {
 
   std::vector<std::vector<PaddleTensor>> input_slots_all;
   SetInput(&input_slots_all);
-  TestPrediction(cfg, input_slots_all, &outputs, FLAGS_num_threads);
+  TestPrediction(reinterpret_cast<const PaddlePredictor::Config *>(&cfg),
+                 input_slots_all, &outputs, FLAGS_num_threads);
 
   if (FLAGS_num_threads == 1 && !FLAGS_test_all_data) {
     const float ocr_result_data[] = {
@@ -133,7 +134,8 @@ void compare(bool use_mkldnn = false) {
 
   std::vector<std::vector<PaddleTensor>> input_slots_all;
   SetInput(&input_slots_all);
-  CompareNativeAndAnalysis(cfg, input_slots_all);
+  CompareNativeAndAnalysis(
+      reinterpret_cast<const PaddlePredictor::Config *>(&cfg), input_slots_all);
 }
 
 TEST(Analyzer_vis, compare) { compare(); }

paddle/fluid/inference/tests/api/config_printer.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 <ostream>
+#include <string>
+#include "paddle/fluid/inference/api/paddle_inference_api.h"
+
+namespace paddle {
+namespace inference {
+
+thread_local int num_spaces = 0;
+
+static std::string GenSpaces(int num_spaces) {
+  std::ostringstream os;
+  for (int i = 0; i < num_spaces; ++i) {
+    os << "  ";
+  }
+  return os.str();
+}
+
+std::ostream &operator<<(std::ostream &os,
+                         const PaddlePredictor::Config &config) {
+  os << GenSpaces(num_spaces) << "PaddlePredictor::Config {\n";
+  num_spaces++;
+  os << GenSpaces(num_spaces) << "model_dir: " << config.model_dir << "\n";
+  num_spaces--;
+  os << GenSpaces(num_spaces) << "}\n";
+  return os;
+}
+
+std::ostream &operator<<(std::ostream &os, const NativeConfig &config) {
+  os << GenSpaces(num_spaces) << "NativeConfig {\n";
+  num_spaces++;
+  os << *reinterpret_cast<const PaddlePredictor::Config *>(&config);
+  os << GenSpaces(num_spaces) << "use_gpu: " << config.use_gpu << "\n";
+  os << GenSpaces(num_spaces) << "device: " << config.device << "\n";
+  os << GenSpaces(num_spaces)
+     << "fraction_of_gpu_memory: " << config.fraction_of_gpu_memory << "\n";
+  os << GenSpaces(num_spaces) << "prog_file: " << config.prog_file << "\n";
+  os << GenSpaces(num_spaces) << "param_file: " << config.param_file << "\n";
+  os << GenSpaces(num_spaces)
+     << "specify_input_name: " << config.specify_input_name << "\n";
+  num_spaces--;
+  os << GenSpaces(num_spaces) << "}\n";
+  return os;
+}
+
+std::ostream &operator<<(std::ostream &os,
+                         const contrib::AnalysisConfig &config) {
+  os << GenSpaces(num_spaces) << "contrib::AnalysisConfig {\n";
+  num_spaces++;
+  os << *reinterpret_cast<const NativeConfig *>(&config);
+  os << GenSpaces(num_spaces) << "enable_ir_optim: " << config.enable_ir_optim
+     << "\n";
+  os << GenSpaces(num_spaces)
+     << "use_feed_fetch_ops: " << config.use_feed_fetch_ops << "\n";
+  os << GenSpaces(num_spaces) << "use_tensorrt: " << config.use_tensorrt()
+     << "\n";
+  os << GenSpaces(num_spaces) << "use_mkldnn: " << config.use_mkldnn() << "\n";
+  num_spaces--;
+  os << GenSpaces(num_spaces) << "}\n";
+  return os;
+}
+
+}  // namespace inference
+}  // namespace paddle

paddle/fluid/inference/tests/api/tester_helper.h

@@ -19,13 +19,16 @@
 #include <string>
 #include <thread>  // NOLINT
 #include <vector>
+
 #include "paddle/fluid/framework/ir/fuse_pass_base.h"
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/inference/analysis/analyzer.h"
 #include "paddle/fluid/inference/analysis/ut_helper.h"
 #include "paddle/fluid/inference/api/analysis_predictor.h"
-#include "paddle/fluid/inference/api/helper.h"
 #include "paddle/fluid/inference/api/paddle_inference_pass.h"
+
+#include "paddle/fluid/inference/api/helper.h"
+#include "paddle/fluid/inference/tests/api/config_printer.h"
 #include "paddle/fluid/inference/tests/test_helper.h"
 #include "paddle/fluid/platform/profiler.h"
 
@@ -38,10 +41,18 @@ DEFINE_int32(num_threads, 1, "Running the inference program in multi-threads.");
 DEFINE_bool(use_analysis, true,
             "Running the inference program in analysis mode.");
 
+DECLARE_bool(profile);
+
 namespace paddle {
 namespace inference {
 
-using contrib::AnalysisConfig;
+void PrintConfig(const PaddlePredictor::Config *config, bool use_analysis) {
+  if (use_analysis) {
+    LOG(INFO) << *reinterpret_cast<const contrib::AnalysisConfig *>(config);
+    return;
+  }
+  LOG(INFO) << *config;
+}
 
 void CompareResult(const std::vector<PaddleTensor> &outputs,
                    const std::vector<PaddleTensor> &ref_outputs) {
@@ -77,12 +88,13 @@ void CompareResult(const std::vector<PaddleTensor> &outputs,
 }
 
 std::unique_ptr<PaddlePredictor> CreateTestPredictor(
-    const AnalysisConfig &config, bool use_analysis = true) {
+    const PaddlePredictor::Config *config, bool use_analysis = true) {
   if (use_analysis) {
-    return CreatePaddlePredictor<contrib::AnalysisConfig>(config);
-  } else {
-    return CreatePaddlePredictor<NativeConfig>(config);
+    return CreatePaddlePredictor<contrib::AnalysisConfig>(
+        *(reinterpret_cast<const contrib::AnalysisConfig *>(config)));
   }
+  return CreatePaddlePredictor<NativeConfig>(
+      *(reinterpret_cast<const NativeConfig *>(config)));
 }
 
 size_t GetSize(const PaddleTensor &out) { return VecReduceToInt(out.shape); }
@@ -111,11 +123,23 @@ std::unordered_map<std::string, int> GetFuseStatis(PaddlePredictor *predictor,
 }
 
 void SetFakeImageInput(std::vector<std::vector<PaddleTensor>> *inputs,
-                       const std::string &dirname) {
+                       const std::string &dirname, bool is_combined = true,
+                       std::string model_filename = "model",
+                       std::string params_filename = "params") {
   // Set fake_image_data
   PADDLE_ENFORCE_EQ(FLAGS_test_all_data, 0, "Only have single batch of data.");
-  std::vector<std::vector<int64_t>> feed_target_shapes =
-      GetFeedTargetShapes(dirname, true, "model", "params");
+  std::vector<std::vector<int64_t>> feed_target_shapes = GetFeedTargetShapes(
+      dirname, is_combined, model_filename, params_filename);
+  std::ostringstream os;
+  for (size_t i = 0; i < feed_target_shapes.size(); ++i) {
+    os << "feed target " << i << ": {" << feed_target_shapes[i][0];
+    for (size_t j = 1; j < feed_target_shapes[i].size(); ++j) {
+      os << ", " << feed_target_shapes[i][j];
+    }
+    os << "}\n";
+  }
+  LOG(INFO) << os.str();
+
   int dim1 = feed_target_shapes[0][1];
   int dim2 = feed_target_shapes[0][2];
   int dim3 = feed_target_shapes[0][3];
@@ -139,25 +163,43 @@ void SetFakeImageInput(std::vector<std::vector<PaddleTensor>> *inputs,
 }
 
 void TestOneThreadPrediction(
-    const AnalysisConfig &config,
+    const PaddlePredictor::Config *config,
     const std::vector<std::vector<PaddleTensor>> &inputs,
     std::vector<PaddleTensor> *outputs, bool use_analysis = true) {
   int batch_size = FLAGS_batch_size;
   int num_times = FLAGS_repeat;
   auto predictor = CreateTestPredictor(config, use_analysis);
-  Timer timer;
-  timer.tic();
-  for (int i = 0; i < num_times; i++) {
-    for (size_t j = 0; j < inputs.size(); j++) {
-      predictor->Run(inputs[j], outputs);
+
+  // warmup run
+  LOG(INFO) << "Warm up run...";
+  {
+    Timer warmup_timer;
+    warmup_timer.tic();
+    predictor->Run(inputs[0], outputs, batch_size);
+    PrintTime(batch_size, 1, 1, 0, warmup_timer.toc(), 1);
+#if !defined(_WIN32)
+    if (FLAGS_profile) {
+      paddle::platform::ResetProfiler();
+    }
+#endif
+  }
+
+  LOG(INFO) << "Run " << num_times << " times...";
+  {
+    Timer run_timer;
+    run_timer.tic();
+    for (int i = 0; i < num_times; i++) {
+      for (size_t j = 0; j < inputs.size(); j++) {
+        predictor->Run(inputs[j], outputs, batch_size);
+      }
     }
+    PrintTime(batch_size, num_times, 1, 0, run_timer.toc() / num_times,
+              inputs.size());
   }
-  PrintTime(batch_size, num_times, 1, 0, timer.toc() / num_times,
-            inputs.size());
 }
 
 void TestMultiThreadPrediction(
-    const AnalysisConfig &config,
+    const PaddlePredictor::Config *config,
     const std::vector<std::vector<PaddleTensor>> &inputs,
     std::vector<PaddleTensor> *outputs, int num_threads,
     bool use_analysis = true) {
@@ -200,12 +242,11 @@ void TestMultiThreadPrediction(
   }
 }
 
-void TestPrediction(const AnalysisConfig &config,
+void TestPrediction(const PaddlePredictor::Config *config,
                     const std::vector<std::vector<PaddleTensor>> &inputs,
                     std::vector<PaddleTensor> *outputs, int num_threads,
                     bool use_analysis = FLAGS_use_analysis) {
-  LOG(INFO) << "use_analysis: " << use_analysis
-            << ", use_mkldnn: " << config.use_mkldnn();
+  PrintConfig(config, use_analysis);
   if (num_threads == 1) {
     TestOneThreadPrediction(config, inputs, outputs, use_analysis);
   } else {
@@ -215,9 +256,9 @@ void TestPrediction(const AnalysisConfig &config,
 }
 
 void CompareNativeAndAnalysis(
-    const AnalysisConfig &config,
+    const PaddlePredictor::Config *config,
     const std::vector<std::vector<PaddleTensor>> &inputs) {
-  LOG(INFO) << "use_mkldnn: " << config.use_mkldnn();
+  PrintConfig(config, true);
   std::vector<PaddleTensor> native_outputs, analysis_outputs;
   TestOneThreadPrediction(config, inputs, &native_outputs, false);
   TestOneThreadPrediction(config, inputs, &analysis_outputs, true);

paddle/fluid/inference/tests/api/trt_models_tester.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.
+/* 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 <gflags/gflags.h>
 #include <glog/logging.h>
 #include <gtest/gtest.h>
-#include "paddle/fluid/inference/analysis/analyzer.h"
-#include "paddle/fluid/inference/api/helper.h"
-#include "paddle/fluid/inference/api/paddle_inference_api.h"
-#include "paddle/fluid/inference/api/paddle_inference_pass.h"
+
 #include "paddle/fluid/inference/tests/api/tester_helper.h"
 
 namespace paddle {
-using paddle::contrib::AnalysisConfig;
-
-DEFINE_string(dirname, "", "Directory of the inference model.");
-
-NativeConfig GetConfigNative() {
-  NativeConfig config;
-  config.model_dir = FLAGS_dirname;
-  // LOG(INFO) << "dirname  " << config.model_dir;
-  config.fraction_of_gpu_memory = 0.15;
-  config.use_gpu = true;
-  config.device = 0;
-  return config;
-}
-
-void PrepareTRTConfig(AnalysisConfig *config) {
-  config->model_dir = FLAGS_dirname + "/" + "mobilenet";
-  config->fraction_of_gpu_memory = 0.15;
-  config->EnableTensorRtEngine(1 << 10, 5);
-  config->pass_builder()->DeletePass("conv_bn_fuse_pass");
-  config->pass_builder()->DeletePass("fc_fuse_pass");
-  config->pass_builder()->TurnOnDebug();
+namespace inference {
+
+DEFINE_bool(use_tensorrt, true, "Test the performance of TensorRT engine.");
+DEFINE_string(prog_filename, "", "Name of model file.");
+DEFINE_string(param_filename, "", "Name of parameters file.");
+
+template <typename ConfigType>
+void SetConfig(ConfigType* config, std::string model_dir, bool use_gpu,
+               bool use_tensorrt = false, int batch_size = -1) {
+  if (!FLAGS_prog_filename.empty() && !FLAGS_param_filename.empty()) {
+    config->prog_file = model_dir + "/" + FLAGS_prog_filename;
+    config->param_file = model_dir + "/" + FLAGS_param_filename;
+  } else {
+    config->model_dir = model_dir;
+  }
+  if (use_gpu) {
+    config->use_gpu = true;
+    config->device = 0;
+    config->fraction_of_gpu_memory = 0.15;
+  }
 }
 
-void PrepareInputs(std::vector<PaddleTensor> *tensors, int batch_size) {
-  PADDLE_ENFORCE_EQ(tensors->size(), 1UL);
-  auto &tensor = tensors->front();
-  int height = 224;
-  int width = 224;
-  float *data = new float[batch_size * 3 * height * width];
-  memset(data, 0, sizeof(float) * (batch_size * 3 * height * width));
-  data[0] = 1.0f;
-
-  // Prepare inputs
-  tensor.name = "input_0";
-  tensor.shape = std::vector<int>({batch_size, 3, height, width});
-  tensor.data = PaddleBuf(static_cast<void *>(data),
-                          sizeof(float) * (batch_size * 3 * height * width));
-  tensor.dtype = PaddleDType::FLOAT32;
+template <>
+void SetConfig<contrib::AnalysisConfig>(contrib::AnalysisConfig* config,
+                                        std::string model_dir, bool use_gpu,
+                                        bool use_tensorrt, int batch_size) {
+  if (!FLAGS_prog_filename.empty() && !FLAGS_param_filename.empty()) {
+    config->prog_file = model_dir + "/" + FLAGS_prog_filename;
+    config->param_file = model_dir + "/" + FLAGS_param_filename;
+  } else {
+    config->model_dir = model_dir;
+  }
+  if (use_gpu) {
+    config->use_gpu = true;
+    config->device = 0;
+    config->fraction_of_gpu_memory = 0.15;
+    if (use_tensorrt) {
+      config->EnableTensorRtEngine(1 << 10, batch_size);
+      config->pass_builder()->DeletePass("conv_bn_fuse_pass");
+      config->pass_builder()->DeletePass("fc_fuse_pass");
+      config->pass_builder()->TurnOnDebug();
+    } else {
+      config->enable_ir_optim = true;
+    }
+  }
 }
 
-void CompareTensorRTWithFluid(int batch_size, std::string model_dirname) {
-  auto config0 = GetConfigNative();
-  config0.model_dir = model_dirname;
-
-  AnalysisConfig config1(true);
-  PrepareTRTConfig(&config1);
-  config1.model_dir = model_dirname;
-
-  auto predictor0 = CreatePaddlePredictor<NativeConfig>(config0);
-  auto predictor1 = CreatePaddlePredictor(config1);
-
-  // Prepare inputs
-  std::vector<PaddleTensor> paddle_tensor_feeds(1);
-  PrepareInputs(&paddle_tensor_feeds, batch_size);
-
-  // Prepare outputs
-  std::vector<PaddleTensor> outputs0;
-  std::vector<PaddleTensor> outputs1;
-  CHECK(predictor0->Run(paddle_tensor_feeds, &outputs0));
-  CHECK(predictor1->Run(paddle_tensor_feeds, &outputs1, batch_size));
-
-  const size_t num_elements = outputs0.front().data.length() / sizeof(float);
-  const size_t num_elements1 = outputs1.front().data.length() / sizeof(float);
-  EXPECT_EQ(num_elements, num_elements1);
-
-  auto *data0 = static_cast<float *>(outputs0.front().data.data());
-  auto *data1 = static_cast<float *>(outputs1.front().data.data());
-
-  ASSERT_GT(num_elements, 0UL);
-  for (size_t i = 0; i < std::min(num_elements, num_elements1); i++) {
-    EXPECT_NEAR(data0[i], data1[i], 1e-3);
+void profile(std::string model_dir, bool use_analysis, bool use_tensorrt) {
+  std::vector<std::vector<PaddleTensor>> inputs_all;
+  if (!FLAGS_prog_filename.empty() && !FLAGS_param_filename.empty()) {
+    SetFakeImageInput(&inputs_all, model_dir, true, FLAGS_prog_filename,
+                      FLAGS_param_filename);
+  } else {
+    SetFakeImageInput(&inputs_all, model_dir, false, "__model__", "");
   }
-}
 
-TEST(trt_models_test, mobilenet) {
-  CompareTensorRTWithFluid(1, FLAGS_dirname + "/" + "mobilenet");
-}
-TEST(trt_models_test, resnet50) {
-  CompareTensorRTWithFluid(1, FLAGS_dirname + "/" + "resnet50");
-}
-TEST(trt_models_test, resnext50) {
-  CompareTensorRTWithFluid(1, FLAGS_dirname + "/" + "resnext50");
+  std::vector<PaddleTensor> outputs;
+  if (use_analysis || use_tensorrt) {
+    contrib::AnalysisConfig config(true);
+    SetConfig<contrib::AnalysisConfig>(&config, model_dir, true, use_tensorrt,
+                                       FLAGS_batch_size);
+    TestPrediction(reinterpret_cast<PaddlePredictor::Config*>(&config),
+                   inputs_all, &outputs, FLAGS_num_threads, true);
+  } else {
+    NativeConfig config;
+    SetConfig<NativeConfig>(&config, model_dir, true, false);
+    TestPrediction(reinterpret_cast<PaddlePredictor::Config*>(&config),
+                   inputs_all, &outputs, FLAGS_num_threads, false);
+  }
 }
 
-TEST(trt_models_test, raw_gpu) {
-  std::string model_dir = FLAGS_dirname + "/" + "mobilenet";
-  auto config0 = GetConfigNative();
-  config0.model_dir = model_dir;
-  int batch_size = 2;
-
-  AnalysisConfig config1(true);
-  config1.fraction_of_gpu_memory = 0.1;
-  config1.enable_ir_optim = true;
-  config1.model_dir = model_dir;
+void compare(std::string model_dir, bool use_tensorrt) {
+  std::vector<std::vector<PaddleTensor>> inputs_all;
+  if (!FLAGS_prog_filename.empty() && !FLAGS_param_filename.empty()) {
+    SetFakeImageInput(&inputs_all, model_dir, true, FLAGS_prog_filename,
+                      FLAGS_param_filename);
+  } else {
+    SetFakeImageInput(&inputs_all, model_dir, false, "__model__", "");
+  }
 
-  auto predictor0 = CreatePaddlePredictor<NativeConfig>(config0);
-  auto predictor1 = CreatePaddlePredictor(config1);
+  std::vector<PaddleTensor> native_outputs;
+  NativeConfig native_config;
+  SetConfig<NativeConfig>(&native_config, model_dir, true, false,
+                          FLAGS_batch_size);
+  TestOneThreadPrediction(
+      reinterpret_cast<PaddlePredictor::Config*>(&native_config), inputs_all,
+      &native_outputs, false);
+
+  std::vector<PaddleTensor> analysis_outputs;
+  contrib::AnalysisConfig analysis_config(true);
+  SetConfig<contrib::AnalysisConfig>(&analysis_config, model_dir, true,
+                                     use_tensorrt, FLAGS_batch_size);
+  TestOneThreadPrediction(
+      reinterpret_cast<PaddlePredictor::Config*>(&analysis_config), inputs_all,
+      &analysis_outputs, true);
+
+  CompareResult(native_outputs, analysis_outputs);
+}
 
-  // Prepare inputs
-  std::vector<PaddleTensor> paddle_tensor_feeds(1);
-  PrepareInputs(&paddle_tensor_feeds, batch_size);
+TEST(TensorRT_mobilenet, compare) {
+  std::string model_dir = FLAGS_infer_model + "/mobilenet";
+  compare(model_dir, /* use_tensorrt */ true);
+}
 
-  // Prepare outputs
-  std::vector<PaddleTensor> outputs0;
-  std::vector<PaddleTensor> outputs1;
-  CHECK(predictor0->Run(paddle_tensor_feeds, &outputs0));
-  CHECK(predictor1->Run(paddle_tensor_feeds, &outputs1, batch_size));
+TEST(TensorRT_resnet50, compare) {
+  std::string model_dir = FLAGS_infer_model + "/resnet50";
+  compare(model_dir, /* use_tensorrt */ true);
+}
 
-  const size_t num_elements = outputs0.front().data.length() / sizeof(float);
-  const size_t num_elements1 = outputs1.front().data.length() / sizeof(float);
-  EXPECT_EQ(num_elements, num_elements1);
+TEST(TensorRT_resnext50, compare) {
+  std::string model_dir = FLAGS_infer_model + "/resnext50";
+  compare(model_dir, /* use_tensorrt */ true);
+}
 
-  auto *data0 = static_cast<float *>(outputs0.front().data.data());
-  auto *data1 = static_cast<float *>(outputs1.front().data.data());
+TEST(TensorRT_resnext50, profile) {
+  std::string model_dir = FLAGS_infer_model + "/resnext50";
+  profile(model_dir, /* use_analysis */ true, FLAGS_use_tensorrt);
+}
 
-  ASSERT_GT(num_elements, 0UL);
-  for (size_t i = 0; i < std::min(num_elements, num_elements1); i++) {
-    EXPECT_NEAR(data0[i], data1[i], 1e-3);
-  }
+TEST(TensorRT_mobilenet, analysis) {
+  std::string model_dir = FLAGS_infer_model + "/" + "mobilenet";
+  compare(model_dir, /* use_tensorrt */ false);
 }
 
+}  // namespace inference
 }  // namespace paddle
 
 USE_PASS(tensorrt_subgraph_pass);

paddle/fluid/operators/CMakeLists.txt

@@ -94,7 +94,8 @@ function(op_library TARGET)
     # remove windows unsupported op, because windows has no nccl, no warpctc such ops.
     foreach(windows_unsupport_op "nccl_op" "gen_nccl_id_op" "warpctc_op" "hierarchical_sigmoid_op"
      "crf_decoding_op" "select_op" "lstmp_op" "gru_op" "fusion_gru_op" "lstm_op" "fusion_lstm_op" "cumsum_op"
-      "fusion_seqconv_eltadd_relu_op" "channel_send_op" "channel_create_op" "channel_close_op" "channel_recv_op")
+      "fusion_seqconv_eltadd_relu_op" "channel_send_op" "channel_create_op" "channel_close_op" "channel_recv_op"
+            "fusion_seqexpand_concat_fc_op" "attention_lstm_op" "fused_embedding_fc_lstm_op" "fc_op")
         if ("${TARGET}" STREQUAL "${windows_unsupport_op}")
           return()
         endif()
@@ -308,8 +309,10 @@ op_library(flatten_op DEPS reshape_op)
 op_library(sequence_pad_op DEPS sequence_padding)
 op_library(unstack_op DEPS stack_op)
 op_library(fake_quantize_op DEPS memory)
+if (NOT WIN32)
 op_library(crf_decoding_op DEPS jit_kernel)
 op_library(fusion_lstm_op DEPS jit_kernel)
+endif(NOT WIN32)
 if (WITH_GPU)
     op_library(conv_op DEPS vol2col depthwise_conv im2col)
     op_library(layer_norm_op DEPS cub)
@@ -325,8 +328,8 @@ op_library(save_op DEPS lod_tensor)
 op_library(load_op DEPS lod_tensor)
 op_library(save_combine_op DEPS lod_tensor)
 op_library(load_combine_op DEPS lod_tensor)
-op_library(tensor_array_to_tensor_op DEPS concat_op)
 op_library(concat_op DEPS concat_and_split)
+op_library(tensor_array_to_tensor_op DEPS concat_op)
 
 list(REMOVE_ITEM GENERAL_OPS ${DEPS_OPS})
 

paddle/fluid/operators/conv_cudnn_op.cu.cc

@@ -50,12 +50,18 @@ static constexpr char kCUDNNBwdFilterAlgoCache[] = "kCUDNNBwdFilterAlgoCache";
 static constexpr size_t kCONV_CUDNN_WORKSPACE_LIMIT_BYTES =
     static_cast<size_t>(1024) * 1024 * 1024;
 
-static constexpr size_t kNUM_CUDNN_FWD_ALGS =
-    CUDNN_CONVOLUTION_BWD_FILTER_ALGO_COUNT;
+#if CUDNN_VERSION_MIN(6, 0, 5)
+static constexpr size_t kNUM_CUDNN_FWD_ALGS = CUDNN_CONVOLUTION_FWD_ALGO_COUNT;
 static constexpr size_t kNUM_CUDNN_BWD_FILTER_ALGS =
     CUDNN_CONVOLUTION_BWD_FILTER_ALGO_COUNT;
 static constexpr size_t kNUM_CUDNN_BWD_DATA_ALGS =
     CUDNN_CONVOLUTION_BWD_DATA_ALGO_COUNT;
+#else
+// cuDNN v5 has no CUDNN_CONVOLUTION_FWD_ALGO_COUNT etc.
+static constexpr size_t kNUM_CUDNN_FWD_ALGS = 7;
+static constexpr size_t kNUM_CUDNN_BWD_FILTER_ALGS = 4;
+static constexpr size_t kNUM_CUDNN_BWD_DATA_ALGS = 5;
+#endif
 
 template <typename T>
 class CUDNNConvOpKernel : public framework::OpKernel<T> {

paddle/fluid/operators/detection/roi_perspective_transform_op.cu

@@ -15,6 +15,10 @@ limitations under the License. */
 #include <algorithm>
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/platform/cuda_primitives.h"
+#include "paddle/fluid/platform/float16.h"
+
+using paddle::platform::PADDLE_CUDA_NUM_THREADS;
+using paddle::platform::float16;
 
 namespace paddle {
 namespace operators {

paddle/fluid/operators/elementwise_op_function.h

@@ -111,6 +111,17 @@ class RowwiseTransformIterator<T, platform::CPUDeviceContext>
     return *this;
   }
 
+  RowwiseTransformIterator<T, platform::CPUDeviceContext> &operator+(int n) {
+    while (n-- > 0) {
+      ++i_;
+      if (UNLIKELY(i_ == n_)) {
+        i_ = 0;
+      }
+    }
+
+    return *this;
+  }
+
   bool operator==(const RowwiseTransformIterator<T, platform::CPUDeviceContext>
                       &rhs) const {
     return (ptr_ + i_) == &(*rhs);
@@ -149,6 +160,21 @@ class MidWiseTransformIterator<T, platform::CPUDeviceContext>
     return *this;
   }
 
+  MidWiseTransformIterator<T, platform::CPUDeviceContext> &operator+(int n) {
+    while (n-- > 0) {
+      ++j_;
+      if (UNLIKELY(j_ == post_)) {
+        ++i_;
+        j_ = 0;
+        if (UNLIKELY(i_ == n_)) {
+          i_ = 0;
+        }
+      }
+    }
+
+    return *this;
+  }
+
   bool operator==(const MidWiseTransformIterator<T, platform::CPUDeviceContext>
                       &rhs) const {
     return (ptr_ + i_) == &(*rhs);

paddle/fluid/operators/grid_sampler_op.h

@@ -63,7 +63,8 @@ static void CalcGridLocations(const platform::CPUDeviceContext& ctx,
   Tensor ones;
   ones.mutable_data<T>({n, h, w}, ctx.GetPlace());
   auto ones_t = EigenTensor<T, 3>::From(ones).setConstant(1.0);
-  Tensor half_xmax, half_ymax;
+  Tensor half_xmax;
+  Tensor half_ymax;
   half_xmax.mutable_data<T>({n, h, w}, ctx.GetPlace());
   auto half_xmax_t =
       EigenTensor<T, 3>::From(half_xmax).setConstant(0.5 * x_max);

paddle/fluid/operators/math/CMakeLists.txt

@@ -75,12 +75,13 @@ if(WITH_GPU)
 endif()
 cc_test(concat_test SRCS concat_test.cc DEPS concat_and_split)
 cc_test(cpu_vec_test SRCS cpu_vec_test.cc DEPS blas cpu_info)
-
-set(JIT_KERNEL_SRCS jit_kernel.cc jit_kernel_blas.cc jit_kernel_exp.cc jit_kernel_rnn.cc jit_kernel_crf_decode.cc)
-set(JIT_KERNEL_DEPS cpu_info cblas gflags enforce)
-if(WITH_XBYAK)
-    list(APPEND JIT_KERNEL_SRCS jit_gen.cc jit_code.cc)
-    list(APPEND JIT_KERNEL_DEPS xbyak)
-endif()
-cc_library(jit_kernel SRCS ${JIT_KERNEL_SRCS} DEPS ${JIT_KERNEL_DEPS})
-cc_test(jit_kernel_test SRCS jit_kernel_test.cc DEPS jit_kernel)
+if (NOT WIN32)
+    set(JIT_KERNEL_SRCS jit_kernel.cc jit_kernel_blas.cc jit_kernel_exp.cc jit_kernel_rnn.cc jit_kernel_crf_decode.cc)
+    set(JIT_KERNEL_DEPS cpu_info cblas gflags enforce)
+    if(WITH_XBYAK)
+        list(APPEND JIT_KERNEL_SRCS jit_gen.cc jit_code.cc)
+        list(APPEND JIT_KERNEL_DEPS xbyak)
+    endif()
+    cc_library(jit_kernel SRCS ${JIT_KERNEL_SRCS} DEPS ${JIT_KERNEL_DEPS})
+    cc_test(jit_kernel_test SRCS jit_kernel_test.cc DEPS jit_kernel)
+endif (NOT WIN32)

paddle/fluid/operators/math/softmax.cc

@@ -19,8 +19,10 @@ namespace paddle {
 namespace operators {
 namespace math {
 
-template class SoftmaxFunctor<platform::CPUDeviceContext, float>;
-template class SoftmaxFunctor<platform::CPUDeviceContext, double>;
+template class SoftmaxFunctor<platform::CPUDeviceContext, float, true>;
+template class SoftmaxFunctor<platform::CPUDeviceContext, float, false>;
+template class SoftmaxFunctor<platform::CPUDeviceContext, double, true>;
+template class SoftmaxFunctor<platform::CPUDeviceContext, double, false>;
 template class SoftmaxGradFunctor<platform::CPUDeviceContext, float>;
 template class SoftmaxGradFunctor<platform::CPUDeviceContext, double>;
 

paddle/fluid/operators/math/softmax.cu

@@ -98,9 +98,14 @@ template class SoftmaxGradCUDNNFunctor<float>;
 template class SoftmaxGradCUDNNFunctor<double>;
 template class SoftmaxGradCUDNNFunctor<platform::float16>;
 
-template class SoftmaxFunctor<platform::CUDADeviceContext, platform::float16>;
-template class SoftmaxFunctor<platform::CUDADeviceContext, float>;
-template class SoftmaxFunctor<platform::CUDADeviceContext, double>;
+template class SoftmaxFunctor<platform::CUDADeviceContext, platform::float16,
+                              false>;
+template class SoftmaxFunctor<platform::CUDADeviceContext, platform::float16,
+                              true>;
+template class SoftmaxFunctor<platform::CUDADeviceContext, float, false>;
+template class SoftmaxFunctor<platform::CUDADeviceContext, double, false>;
+template class SoftmaxFunctor<platform::CUDADeviceContext, float, true>;
+template class SoftmaxFunctor<platform::CUDADeviceContext, double, true>;
 template class SoftmaxGradFunctor<platform::CUDADeviceContext, float>;
 template class SoftmaxGradFunctor<platform::CUDADeviceContext, double>;
 template class SoftmaxGradFunctor<platform::CUDADeviceContext,

paddle/fluid/operators/math/softmax.h

@@ -19,7 +19,7 @@ namespace paddle {
 namespace operators {
 namespace math {
 
-template <typename DeviceContext, typename T>
+template <typename DeviceContext, typename T, bool is_test>
 class SoftmaxFunctor {
  public:
   void operator()(const DeviceContext& context, const framework::Tensor* X,

paddle/fluid/operators/math/softmax_impl.h

@@ -32,10 +32,10 @@ struct ValueClip {
   }
 };
 
-template <typename DeviceContext, typename T>
-void SoftmaxFunctor<DeviceContext, T>::operator()(const DeviceContext& context,
-                                                  const framework::Tensor* X,
-                                                  framework::Tensor* Y) {
+template <typename DeviceContext, typename T, bool is_test>
+void SoftmaxFunctor<DeviceContext, T, is_test>::operator()(
+    const DeviceContext& context, const framework::Tensor* X,
+    framework::Tensor* Y) {
   auto logits = EigenMatrix<T>::From(*X);
   auto softmax = EigenMatrix<T>::From(*Y);
 
@@ -65,6 +65,39 @@ void SoftmaxFunctor<DeviceContext, T>::operator()(const DeviceContext& context,
                                                  .broadcast(one_by_class));
 }
 
+template <typename DeviceContext, typename T>
+class SoftmaxFunctor<DeviceContext, T, true> {
+  void operator()(const DeviceContext& context, const framework::Tensor* X,
+                  framework::Tensor* Y) {
+    auto logits = EigenMatrix<T>::From(*X);
+    auto softmax = EigenMatrix<T>::From(*Y);
+
+    const int kBatchDim = 0;
+    const int kClassDim = 1;
+
+    const int batch_size = logits.dimension(kBatchDim);
+    const int num_classes = logits.dimension(kClassDim);
+
+    Eigen::DSizes<int, 1> along_class(kClassDim);
+    Eigen::DSizes<int, 2> batch_by_one(batch_size, 1);
+    Eigen::DSizes<int, 2> one_by_class(1, num_classes);
+
+    auto shifted_logits = (logits -
+                           logits.maximum(along_class)
+                               .eval()
+                               .reshape(batch_by_one)
+                               .broadcast(one_by_class));
+
+    softmax.device(*context.eigen_device()) = shifted_logits.exp();
+    softmax.device(*context.eigen_device()) = (softmax *
+                                               softmax.sum(along_class)
+                                                   .inverse()
+                                                   .eval()
+                                                   .reshape(batch_by_one)
+                                                   .broadcast(one_by_class));
+  }
+};
+
 template <typename DeviceContext, typename T>
 void SoftmaxGradFunctor<DeviceContext, T>::operator()(
     const DeviceContext& context, const framework::Tensor* y,

paddle/fluid/operators/pad_constant_like_op.cc

@@ -74,7 +74,7 @@ PadConstantLikeOp Operator.
 
 Pad input(Y) with a pad_value, the number of values padded to the edges of each
 axis is specified by the difference of the shape of X and Y.
-((0, shape_x_0 - shape_y_0), … (0, shape_x_n - shape_y_n)) unique pad widths for
+((0, shape_x_0 - shape_y_0), ... (0, shape_x_n - shape_y_n)) unique pad widths for
 each axis.
 The input should be a k-D tensor(k > 0 and k < 7). As an example:
 

paddle/fluid/operators/roi_pool_op.cc

@@ -122,7 +122,7 @@ class ROIPoolOpMaker : public framework::OpProtoAndCheckerMaker {
               "(Tensor), "
               "Argmaxes corresponding to indices in X used "
               "for gradient computation. Only output "
-              "if arg “is_test” is false.")
+              "if arg \"is_test\" is false.")
         .AsIntermediate();
     AddAttr<float>("spatial_scale",
                    "(float, default 1.0), "

paddle/fluid/operators/softmax_op.h

@@ -35,8 +35,13 @@ class SoftmaxKernel : public framework::OpKernel<T> {
     Tensor X_2d = framework::ReshapeToMatrix(*X, rank - 1);
     Tensor Out_2d = framework::ReshapeToMatrix(*Out, rank - 1);
 
-    math::SoftmaxFunctor<DeviceContext, T>()(
+#ifdef ON_INFER
+    math::SoftmaxFunctor<DeviceContext, T, true>()(
         context.template device_context<DeviceContext>(), &X_2d, &Out_2d);
+#else
+    math::SoftmaxFunctor<DeviceContext, T, false>()(
+        context.template device_context<DeviceContext>(), &X_2d, &Out_2d);
+#endif
   }
 };
 

paddle/fluid/operators/softmax_with_cross_entropy_op.h

@@ -42,8 +42,8 @@ class SoftmaxWithCrossEntropyKernel : public framework::OpKernel<T> {
 
     auto& dev_ctx =
         context.template device_context<platform::CPUDeviceContext>();
-    math::SoftmaxFunctor<platform::CPUDeviceContext, T>()(dev_ctx, logits,
-                                                          softmax);
+    math::SoftmaxFunctor<platform::CPUDeviceContext, T, false>()(
+        dev_ctx, logits, softmax);
     math::CrossEntropyFunctor<platform::CPUDeviceContext, T>()(
         dev_ctx, loss, softmax, labels, context.Attr<bool>("soft_label"),
         context.Attr<int>("ignore_index"));

paddle/fluid/operators/unpool_op.cc

@@ -57,8 +57,8 @@ class Unpool2dOpMaker : public framework::OpProtoAndCheckerMaker {
 Input shape is: $(N, C_{in}, H_{in}, W_{in})$, Output shape is:
 $(N, C_{out}, H_{out}, W_{out})$, where
 $$
-H_{out} = (H_{in}−1) * strides[0] − 2 * paddings[0] + ksize[0] \\
-W_{out} = (W_{in}−1) * strides[1] − 2 * paddings[1] + ksize[1]
+H_{out} = (H_{in}-1) * strides[0] - 2 * paddings[0] + ksize[0] \\
+W_{out} = (W_{in}-1) * strides[1] - 2 * paddings[1] + ksize[1]
 $$
 Paper: http://www.matthewzeiler.com/wp-content/uploads/2017/07/iccv2011.pdf
 )DOC");

paddle/fluid/platform/init.cc

@@ -112,6 +112,14 @@ void InitDevices(bool init_p2p, const std::vector<int> devices) {
   }
   places.emplace_back(platform::CPUPlace());
   platform::DeviceContextPool::Init(places);
+
+// windows has no support for openblas multi-thread
+#ifdef _WIN32
+  if (FLAGS_paddle_num_threads > 1) {
+    FLAGS_paddle_num_threads = 1;
+  }
+#endif
+
 #ifndef PADDLE_WITH_MKLDNN
   platform::SetNumThreads(FLAGS_paddle_num_threads);
 #endif
@@ -167,7 +175,9 @@ void InitGLOG(const std::string &prog_name) {
   // glog will not hold the ARGV[0] inside.
   // Use strdup to alloc a new string.
   google::InitGoogleLogging(strdup(prog_name.c_str()));
+#ifndef _WIN32
   google::InstallFailureSignalHandler();
+#endif
 }
 
 }  // namespace framework

paddle/fluid/platform/nccl_helper.h

@@ -12,6 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+#ifndef _WIN32
 #pragma once
 
 #include <stdio.h>
@@ -149,3 +150,4 @@ struct NCCLContextMap {
 
 }  // namespace platform
 }  // namespace paddle
+#endif

paddle/fluid/platform/port.h

@@ -24,21 +24,16 @@
 #include "glog/logging.h"
 
 #if !defined(_WIN32)
-#define UNUSED __attribute__((unused))
 #include <dlfcn.h>     //  dladdr
 #include <execinfo.h>  // backtrace
 #include <sys/stat.h>
 #include <algorithm>  // std::accumulate
 #else
 #include <io.h>  // _popen, _pclose
+#include <stdio.h>
 #include <windows.h>
-#if defined(_WIN32)
 #include <numeric>  // std::accumulate in msvc
-#endif
-// windows version of __attribute__((unused))
-#define UNUSED __pragma(warning(suppress : 4100))
-
-#ifndef S_ISDIR  // windows port for sys/stat.h
+#ifndef S_ISDIR     // windows port for sys/stat.h
 #define S_ISDIR(mode) (((mode)&S_IFMT) == S_IFDIR)
 #endif  // S_ISDIR
 

paddle/fluid/platform/variant.h

@@ -42,3 +42,11 @@ limitations under the License. */
 #include <boost/mpl/comparison.hpp>
 #include <boost/mpl/less_equal.hpp>
 #include <boost/variant.hpp>
+
+// some platform-independent defintion
+#if defined(_WIN32)
+#define UNUSED
+#define __builtin_expect(EXP, C) (EXP)
+#else
+#define UNUSED __attribute__((unused))
+#endif

paddle/fluid/pybind/CMakeLists.txt

@@ -2,9 +2,9 @@
 set(PYBIND_DEPS pybind python proto_desc memory executor prune  feed_fetch_method pass_builder)
 set(PYBIND_SRCS pybind.cc exception.cc protobuf.cc const_value.cc)
 if(NOT WIN32)
-list(APPEND PYBIND_DEPS parallel_executor profiler)
-list(APPEND PYBIND_SRCS recordio.cc)
-endif()
+  list(APPEND PYBIND_DEPS parallel_executor profiler)
+  list(APPEND PYBIND_SRCS recordio.cc)
+endif(NOT WIN32)
 if(WITH_PYTHON)
   if(WITH_AMD_GPU)
     hip_library(paddle_pybind SHARED
@@ -21,5 +21,13 @@ if(WITH_PYTHON)
     endif(NOT APPLE AND NOT ANDROID AND NOT WIN32)
   endif(WITH_AMD_GPU)
 
+  if(WIN32)
+    if(WITH_GPU AND NOT WITH_DSO)
+      get_property(cuda_modules GLOBAL PROPERTY CUDA_MODULES)
+      target_link_libraries(paddle_pybind ${cuda_modules})
+    endif(WITH_GPU AND NOT WITH_DSO)
+    target_link_libraries(paddle_pybind shlwapi)
+  endif(WIN32)
+
   cc_test(tensor_py_test SRCS tensor_py_test.cc DEPS python)
 endif(WITH_PYTHON)

paddle/fluid/pybind/pybind.cc

@@ -21,6 +21,13 @@ limitations under the License. */
 #include <utility>
 #include <vector>
 
+#if defined(_WIN32)
+#define NOMINMAX
+#define GLOG_NO_ABBREVIATED_SEVERITIES  // msvc conflict logging with windows.h
+#define GOOGLE_GLOG_DLL_DECL
+#include <Windows.h>
+#endif
+
 #include "paddle/fluid/framework/executor.h"
 #include "paddle/fluid/framework/feed_fetch_method.h"
 #include "paddle/fluid/framework/framework.pb.h"
@@ -29,7 +36,9 @@ limitations under the License. */
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/lod_tensor_array.h"
 #include "paddle/fluid/framework/op_registry.h"
+#ifndef _WIN32
 #include "paddle/fluid/framework/parallel_executor.h"
+#endif
 #include "paddle/fluid/framework/prune.h"
 #include "paddle/fluid/framework/reader.h"
 #include "paddle/fluid/framework/selected_rows.h"
@@ -50,7 +59,9 @@ limitations under the License. */
 #include "paddle/fluid/string/to_string.h"
 
 #ifdef PADDLE_WITH_CUDA
+#ifndef _WIN32
 #include "paddle/fluid/operators/nccl/nccl_gpu_common.h"
+#endif
 #include "paddle/fluid/platform/cuda_profiler.h"
 #include "paddle/fluid/platform/gpu_info.h"
 #endif
@@ -340,22 +351,25 @@ All parameter, weight, gradient are variables in Paddle.
       .def("get_lod_tensor_array",
            [](Variable &self) { return self.GetMutable<LoDTensorArray>(); },
            py::return_value_policy::reference)
-#ifdef PADDLE_WITH_CUDA
+#if (defined(PADDLE_WITH_CUDA) && !defined(_WIN32))
       .def("get_communicator",
            [](Variable &self) -> platform::Communicator * {
              return self.GetMutable<platform::Communicator>();
            },
            py::return_value_policy::reference)
-#endif
       .def("get_reader",
            [](Variable &self) -> framework::ReaderHolder * {
              PADDLE_ENFORCE(self.IsType<framework::ReaderHolder>());
              return self.GetMutable<framework::ReaderHolder>();
            },
-           py::return_value_policy::reference);
+           py::return_value_policy::reference)
+#endif
+      ;
 
+#if !defined(_WIN32)
   py::class_<framework::ReaderHolder>(m, "Reader", "")
       .def("reset", &framework::ReaderHolder::ResetAll);
+#endif
 
   using LoDTensorBlockingQueue =
       ::paddle::operators::reader::LoDTensorBlockingQueue;
@@ -480,7 +494,7 @@ All parameter, weight, gradient are variables in Paddle.
 #endif
                 });;
 // clang-format on
-#ifdef PADDLE_WITH_CUDA
+#if (defined(PADDLE_WITH_CUDA) && !defined(_WIN32))
   py::class_<platform::Communicator>(m, "Communicator").def(py::init<>());
 #endif
   py::class_<platform::CUDAPlace>(m, "CUDAPlace")
@@ -617,11 +631,14 @@ All parameter, weight, gradient are variables in Paddle.
 #ifdef PADDLE_WITH_CUDA
   m.def("get_cuda_device_count", platform::GetCUDADeviceCount);
 
+#ifndef _WIN32
   m.def("nvprof_init", platform::CudaProfilerInit);
   m.def("nvprof_start", platform::CudaProfilerStart);
   m.def("nvprof_stop", platform::CudaProfilerStop);
 #endif
+#endif
 
+#ifndef _WIN32
   py::enum_<platform::ProfilerState>(m, "ProfilerState", py::arithmetic())
       .value("kDisabled", platform::ProfilerState::kDisabled)
       .value("kCPU", platform::ProfilerState::kCPU)
@@ -642,6 +659,7 @@ All parameter, weight, gradient are variables in Paddle.
   m.def("disable_profiler", platform::DisableProfiler);
   m.def("is_profiler_enabled", platform::IsProfileEnabled);
   m.def("reset_profiler", platform::ResetProfiler);
+#endif
 
   py::class_<ir::Pass, std::shared_ptr<ir::Pass>> pass(m, "Pass");
   pass.def(py::init())
@@ -650,9 +668,9 @@ All parameter, weight, gradient are variables in Paddle.
           [](ir::Pass &self, const std::string &name, const std::string &attr) {
             self.Set<std::string>(name, new std::string(attr));
           })
-      .def("set_int", [](ir::Pass &self, const std::string &name, int val) {
-        self.Set<const int>(name, new int(val));
-      });
+      .def("set_int", [](ir::Pass &self, const std::string &name,
+                         int val) { self.Set<const int>(name, new int(val)); })
+      .def("type", &ir::Pass::Type);
 
   py::class_<ir::PassBuilder, std::shared_ptr<ir::PassBuilder>> pb(
       m, "PassBuilder");
@@ -670,6 +688,7 @@ All parameter, weight, gradient are variables in Paddle.
       .def("remove_pass",
            [](ir::PassBuilder &self, size_t idx) { self.RemovePass(idx); });
 
+#ifndef _WIN32
   // -- python binds for parallel executor.
   py::class_<ParallelExecutor> pe(m, "ParallelExecutor");
   py::class_<ExecutionStrategy> exec_strategy(pe, "ExecutionStrategy", R"DOC(
@@ -791,6 +810,7 @@ All parameter, weight, gradient are variables in Paddle.
           "reduce_strategy",
           [](const BuildStrategy &self) { return self.reduce_; },
           [](BuildStrategy &self, BuildStrategy::ReduceStrategy strategy) {
+            PADDLE_ENFORCE(!self.IsFinalized(), "BuildStrategy is finlaized.");
             self.reduce_ = strategy;
           },
           R"DOC(The type is STR, there are two reduce strategies in ParallelExecutor,
@@ -804,6 +824,7 @@ All parameter, weight, gradient are variables in Paddle.
           [](const BuildStrategy &self) { return self.gradient_scale_; },
           [](BuildStrategy &self,
              BuildStrategy::GradientScaleStrategy strategy) {
+            PADDLE_ENFORCE(!self.IsFinalized(), "BuildStrategy is finlaized.");
             self.gradient_scale_ = strategy;
           },
           R"DOC(The type is STR, there are three ways of defining :math:`loss@grad` in
@@ -815,6 +836,7 @@ All parameter, weight, gradient are variables in Paddle.
           "debug_graphviz_path",
           [](const BuildStrategy &self) { return self.debug_graphviz_path_; },
           [](BuildStrategy &self, const std::string &path) {
+            PADDLE_ENFORCE(!self.IsFinalized(), "BuildStrategy is finlaized.");
             self.debug_graphviz_path_ = path;
           },
           R"DOC(The type is STR, debug_graphviz_path indicate the path that
@@ -824,6 +846,7 @@ All parameter, weight, gradient are variables in Paddle.
           "enable_data_balance",
           [](const BuildStrategy &self) { return self.enable_data_balance_; },
           [](BuildStrategy &self, bool b) {
+            PADDLE_ENFORCE(!self.IsFinalized(), "BuildStrategy is finlaized.");
             self.enable_data_balance_ = b;
           })  // FIXME(chengudo): enable_data_balance seems not important
       .def_property(
@@ -832,6 +855,7 @@ All parameter, weight, gradient are variables in Paddle.
             return self.enable_sequential_execution_;
           },
           [](BuildStrategy &self, bool b) {
+            PADDLE_ENFORCE(!self.IsFinalized(), "BuildStrategy is finlaized.");
             self.enable_sequential_execution_ = b;
           },
           R"DOC(The type is BOOL. If set True, the execution order of ops would be the same as what is in the program. Default False.)DOC")
@@ -841,6 +865,7 @@ All parameter, weight, gradient are variables in Paddle.
             return self.remove_unnecessary_lock_;
           },
           [](BuildStrategy &self, bool b) {
+            PADDLE_ENFORCE(!self.IsFinalized(), "BuildStrategy is finlaized.");
             self.remove_unnecessary_lock_ = b;
           },
           R"DOC(The type is BOOL. If set True, some locks in GPU ops would be released and ParallelExecutor would run faster. Default False.)DOC")
@@ -850,15 +875,19 @@ All parameter, weight, gradient are variables in Paddle.
             return self.fuse_elewise_add_act_ops_;
           },
           [](BuildStrategy &self, bool b) {
+            PADDLE_ENFORCE(!self.IsFinalized(), "BuildStrategy is finlaized.");
             self.fuse_elewise_add_act_ops_ = b;
           },
           R"DOC(The type is BOOL, fuse_elewise_add_act_ops indicate whether
                      to fuse elementwise_add_op and activation_op,
                      it may make the execution faster. Default False)DOC")
-      .def("_create_passes_from_strategy",
+      .def("_finalize_strategy_and_create_passes",
            [](BuildStrategy &self) -> std::shared_ptr<ir::PassBuilder> {
-             return self.CreatePassesFromStrategy();
-           });
+             return self.CreatePassesFromStrategy(true);
+           },
+           R"DOC(Allow user to customized passes. Normally model-specific
+                optimization passes should be defined in this way. BuildStrategy
+                cannot be updated after being finalized.)DOC");
 
   pe.def(py::init<const std::vector<platform::Place> &,
                   const std::unordered_set<std::string> &,
@@ -887,6 +916,7 @@ All parameter, weight, gradient are variables in Paddle.
       });
 
   BindRecordIOWriter(&m);
+#endif
   return m.ptr();
 }
 }  // namespace pybind

paddle/scripts/paddle_build.sh

@@ -156,6 +156,8 @@ function cmake_gen() {
         -DWITH_INFERENCE_API_TEST=${WITH_INFERENCE_API_TEST:-ON}
         -DINFERENCE_DEMO_INSTALL_DIR=${INFERENCE_DEMO_INSTALL_DIR}
         -DWITH_ANAKIN=${WITH_ANAKIN:-OFF}
+        -DANAKIN_BUILD_FAT_BIN=${ANAKIN_BUILD_FAT_BIN:OFF}
+        -DANAKIN_BUILD_CROSS_PLANTFORM=${ANAKIN_BUILD_CROSS_PLANTFORM:ON}
         -DPY_VERSION=${PY_VERSION:-2.7}
         -DCMAKE_INSTALL_PREFIX=${INSTALL_PREFIX:-/paddle/build}
     ========================================
@@ -188,6 +190,8 @@ EOF
         -DWITH_INFERENCE_API_TEST=${WITH_INFERENCE_API_TEST:-ON} \
         -DINFERENCE_DEMO_INSTALL_DIR=${INFERENCE_DEMO_INSTALL_DIR} \
         -DWITH_ANAKIN=${WITH_ANAKIN:-OFF} \
+        -DANAKIN_BUILD_FAT_BIN=${ANAKIN_BUILD_FAT_BIN:OFF}\
+        -DANAKIN_BUILD_CROSS_PLANTFORM=${ANAKIN_BUILD_CROSS_PLANTFORM:ON}\
         -DPY_VERSION=${PY_VERSION:-2.7} \
         -DCMAKE_INSTALL_PREFIX=${INSTALL_PREFIX:-/paddle/build}
 
@@ -777,6 +781,17 @@ function main() {
         test_fluid_lib
         assert_api_spec_approvals
         ;;
+      assert_api)
+        assert_api_not_changed ${PYTHON_ABI:-""}
+        ;;
+      test_inference)
+        gen_capi_package
+        gen_fluid_lib
+        test_fluid_lib
+        ;;
+      assert_api_approvals)
+        assert_api_spec_approvals
+        ;;
       maccheck)
         cmake_gen ${PYTHON_ABI:-""}
         build_mac

python/CMakeLists.txt

@@ -45,23 +45,42 @@ endif()
 
 configure_file(${CMAKE_CURRENT_SOURCE_DIR}/setup.py.in
     ${CMAKE_CURRENT_BINARY_DIR}/setup.py)
-
-set(FLUID_CORE ${PADDLE_BINARY_DIR}/python/paddle/fluid/core.so)
-add_custom_command(OUTPUT ${FLUID_CORE}
-        COMMAND cmake -E copy $<TARGET_FILE:paddle_pybind> ${FLUID_CORE}
-        DEPENDS paddle_pybind)
+IF(WIN32)
+    # Python would use the .pyd by default under Windows series platform
+    set(FLUID_DST_DIR ${PADDLE_BINARY_DIR}/python/paddle/fluid/)
+    get_filename_component(openblas_refpath ${CBLAS_LIBRARIES} DIRECTORY)
+    set(FLUID_CORE ${FLUID_DST_DIR}/core.pyd)
+    add_custom_command(OUTPUT ${FLUID_CORE}
+            COMMAND cmake -E copy $<TARGET_FILE:paddle_pybind> ${FLUID_CORE}
+            COMMAND cmake -E copy ${openblas_refpath}/openblas.dll ${FLUID_DST_DIR}
+            DEPENDS paddle_pybind)
+ELSE()
+    set(FLUID_CORE ${PADDLE_BINARY_DIR}/python/paddle/fluid/core.so)
+    add_custom_command(OUTPUT ${FLUID_CORE}
+            COMMAND cmake -E copy $<TARGET_FILE:paddle_pybind> ${FLUID_CORE}
+            DEPENDS paddle_pybind)
+ENDIF()
 add_custom_target(copy_paddle_pybind ALL DEPENDS ${FLUID_CORE})
 
-
-add_custom_command(OUTPUT ${PADDLE_PYTHON_BUILD_DIR}/.timestamp
-    COMMAND touch stub.cc
-    COMMAND cp -r ${PADDLE_SOURCE_DIR}/python/paddle ${PADDLE_BINARY_DIR}/python
-    COMMAND cp -r ${PADDLE_SOURCE_DIR}/paddle/py_paddle ${PADDLE_BINARY_DIR}/python/
-    COMMAND env ${py_env} ${PYTHON_EXECUTABLE} setup.py bdist_wheel
-    COMMAND ${CMAKE_COMMAND} -E touch ${PADDLE_PYTHON_BUILD_DIR}/.timestamp
-    COMMAND ${CMAKE_COMMAND} -E remove_directory ${PADDLE_PYTHON_BUILD_DIR}/lib-python
-    COMMAND ${CMAKE_COMMAND} -E copy_directory ${PADDLE_PYTHON_BUILD_DIR}/lib.* ${PADDLE_PYTHON_BUILD_DIR}/lib-python
-    DEPENDS gen_proto_py copy_paddle_pybind ${FLUID_CORE} framework_py_proto profiler_py_proto ${PY_FILES} ${external_project_dependencies} ${COPY_PADDLE_MASTER})
+IF(WIN32)
+    add_custom_command(OUTPUT ${PADDLE_PYTHON_BUILD_DIR}/.timestamp
+            COMMAND ${CMAKE_COMMAND} -E copy_directory ${PADDLE_SOURCE_DIR}/python/paddle ${PADDLE_BINARY_DIR}/python/paddle/
+            COMMAND ${CMAKE_COMMAND} -E copy_directory ${PADDLE_SOURCE_DIR}/paddle/py_paddle ${PADDLE_BINARY_DIR}/python/
+            COMMAND ${CMAKE_COMMAND} -E env ${py_env} ${PYTHON_EXECUTABLE} setup.py bdist_wheel
+            COMMAND ${CMAKE_COMMAND} -E touch ${PADDLE_PYTHON_BUILD_DIR}/.timestamp
+            COMMAND ${CMAKE_COMMAND} -E remove_directory ${PADDLE_PYTHON_BUILD_DIR}/lib-python
+            DEPENDS gen_proto_py copy_paddle_pybind ${FLUID_CORE} framework_py_proto profiler_py_proto ${PY_FILES} ${external_project_dependencies} ${COPY_PADDLE_MASTER})
+ELSE(WIN32)
+	add_custom_command(OUTPUT ${PADDLE_PYTHON_BUILD_DIR}/.timestamp
+		COMMAND touch stub.cc
+		COMMAND cp -r ${PADDLE_SOURCE_DIR}/python/paddle ${PADDLE_BINARY_DIR}/python
+		COMMAND cp -r ${PADDLE_SOURCE_DIR}/paddle/py_paddle ${PADDLE_BINARY_DIR}/python/
+		COMMAND env ${py_env} ${PYTHON_EXECUTABLE} setup.py bdist_wheel
+		COMMAND ${CMAKE_COMMAND} -E touch ${PADDLE_PYTHON_BUILD_DIR}/.timestamp
+		COMMAND ${CMAKE_COMMAND} -E remove_directory ${PADDLE_PYTHON_BUILD_DIR}/lib-python
+		COMMAND ${CMAKE_COMMAND} -E copy_directory ${PADDLE_PYTHON_BUILD_DIR}/lib* ${PADDLE_PYTHON_BUILD_DIR}/lib-python
+		DEPENDS gen_proto_py copy_paddle_pybind ${FLUID_CORE} framework_py_proto profiler_py_proto ${PY_FILES} ${external_project_dependencies} ${COPY_PADDLE_MASTER})
+ENDIF()
 
 set(paddle_python_deps ${PADDLE_PYTHON_BUILD_DIR}/.timestamp ${MKL_DEPENDS})
 if(NOT WITH_FLUID_ONLY)

python/paddle/fluid/__init__.py

@@ -13,6 +13,7 @@
 # limitations under the License.
 
 from __future__ import print_function
+import os
 # import all class inside framework into fluid module
 from . import framework
 from .framework import *
@@ -111,12 +112,16 @@ def __bootstrap__():
     os.environ['OMP_NUM_THREADS'] = str(num_threads)
 
     read_env_flags = [
-        'use_pinned_memory', 'check_nan_inf', 'benchmark', 'warpctc_dir',
-        'eager_delete_scope', 'use_mkldnn', 'initial_cpu_memory_in_mb',
+        'use_pinned_memory', 'check_nan_inf', 'benchmark', 'eager_delete_scope',
+        'use_mkldnn', 'use_ngraph', 'initial_cpu_memory_in_mb',
         'init_allocated_mem', 'free_idle_memory', 'paddle_num_threads',
-        'dist_threadpool_size', 'cpu_deterministic', 'eager_delete_tensor_gb',
+        'dist_threadpool_size', 'eager_delete_tensor_gb',
         'reader_queue_speed_test_mode'
     ]
+    if os.name != 'nt':
+        read_env_flags.append('warpctc_dir')
+        read_env_flags.append('cpu_deterministic')
+
     if core.is_compiled_with_dist():
         read_env_flags.append('rpc_deadline')
         read_env_flags.append('rpc_server_profile_path')

python/paddle/fluid/contrib/inferencer.py

@@ -15,13 +15,15 @@
 from __future__ import print_function
 
 import contextlib
+import os
 
 from .. import core
 
 from .. import executor
 from .. import framework
 from .. import io
-from .. import parallel_executor
+if os.name != 'nt':
+    from .. import parallel_executor
 from .. import unique_name
 from .trainer import check_and_get_place
 

python/paddle/fluid/contrib/trainer.py

@@ -28,7 +28,8 @@ from .. import framework
 from .. import io
 # optimizer is same as the parameter of Trainer.__init__. Rename it to opt_module
 from .. import optimizer as opt_module
-from .. import parallel_executor
+if os.name != 'nt':
+    from .. import parallel_executor
 from ..transpiler import distribute_transpiler
 
 __all__ = [

python/paddle/fluid/layers/io.py

@@ -15,6 +15,7 @@
 from __future__ import print_function
 import contextlib
 import multiprocessing
+import os
 import six
 import threading
 
@@ -346,70 +347,72 @@ def _copy_reader_create_op_(block, op):
     return new_op
 
 
-@templatedoc(op_type='create_recordio_file_reader')
-def open_recordio_file(filename,
-                       shapes,
-                       lod_levels,
-                       dtypes,
-                       pass_num=1,
-                       for_parallel=True):
-    """
-    ${comment}
-
-    Args:
-       filename(${filename_type}): ${filename_comment}.
-       shapes(list): List of tuples which declaring data shapes.
-       lod_levels(${lod_levels_type}): ${lod_levels_comment}.
-       dtypes(list): List of strs which declaring data type.
-       pass_num(int): Number of passes to run.
-       for_parallel(Bool): Set it as True if you are going to run
-            subsequent operators in parallel.
-
-    Returns:
-       ${out_comment}.
-
-    Examples:
-
-        >>> import paddle.fluid as fluid
-        >>> reader = fluid.layers.io.open_recordio_file(
-        >>>                               filename='./data.recordio',
-        >>>                               shapes=[(3,224,224), (1)],
-        >>>                               lod_levels=[0, 0],
-        >>>                               dtypes=['float32', 'int64'])
-        >>> # Via the reader, we can use 'read_file' layer to get data:
-        >>> image, label = fluid.layers.io.read_file(reader)
-    """
-    dtypes = [convert_np_dtype_to_dtype_(dt) for dt in dtypes]
-    shape_concat = []
-    ranks = []
+if os.name != 'nt':
+
+    @templatedoc(op_type='create_recordio_file_reader')
+    def open_recordio_file(filename,
+                           shapes,
+                           lod_levels,
+                           dtypes,
+                           pass_num=1,
+                           for_parallel=True):
+        """
+        ${comment}
+
+        Args:
+           filename(${filename_type}): ${filename_comment}.
+           shapes(list): List of tuples which declaring data shapes.
+           lod_levels(${lod_levels_type}): ${lod_levels_comment}.
+           dtypes(list): List of strs which declaring data type.
+           pass_num(int): Number of passes to run.
+           for_parallel(Bool): Set it as True if you are going to run
+                subsequent operators in parallel.
+
+        Returns:
+           ${out_comment}.
+
+        Examples:
+
+            >>> import paddle.fluid as fluid
+            >>> reader = fluid.layers.io.open_recordio_file(
+            >>>                               filename='./data.recordio',
+            >>>                               shapes=[(3,224,224), (1)],
+            >>>                               lod_levels=[0, 0],
+            >>>                               dtypes=['float32', 'int64'])
+            >>> # Via the reader, we can use 'read_file' layer to get data:
+            >>> image, label = fluid.layers.io.read_file(reader)
+        """
+        dtypes = [convert_np_dtype_to_dtype_(dt) for dt in dtypes]
+        shape_concat = []
+        ranks = []
 
-    for shape in shapes:
-        shape_concat.extend(shape)
-        ranks.append(len(shape))
+        for shape in shapes:
+            shape_concat.extend(shape)
+            ranks.append(len(shape))
 
-    var_name = unique_name('open_recordio_file')
+        var_name = unique_name('open_recordio_file')
 
-    startup_blk = default_startup_program().current_block()
-    startup_var = startup_blk.create_var(name=var_name)
-    startup_blk.append_op(
-        type='create_recordio_file_reader',
-        outputs={'Out': [startup_var]},
-        attrs={
-            'shape_concat': shape_concat,
-            'lod_levels': lod_levels,
-            'filename': filename,
-            'ranks': ranks
-        })
+        startup_blk = default_startup_program().current_block()
+        startup_var = startup_blk.create_var(name=var_name)
+        startup_blk.append_op(
+            type='create_recordio_file_reader',
+            outputs={'Out': [startup_var]},
+            attrs={
+                'shape_concat': shape_concat,
+                'lod_levels': lod_levels,
+                'filename': filename,
+                'ranks': ranks
+            })
 
-    startup_var.desc.set_dtypes(dtypes)
-    startup_var.persistable = True
-    main_prog_var = _copy_reader_var_(default_main_program().current_block(),
-                                      startup_var)
+        startup_var.desc.set_dtypes(dtypes)
+        startup_var.persistable = True
+        main_prog_var = _copy_reader_var_(
+            default_main_program().current_block(), startup_var)
 
-    if pass_num > 1:
-        main_prog_var = multi_pass(reader=main_prog_var, pass_num=pass_num)
+        if pass_num > 1:
+            main_prog_var = multi_pass(reader=main_prog_var, pass_num=pass_num)
 
-    return monkey_patch_reader_methods(main_prog_var)
+        return monkey_patch_reader_methods(main_prog_var)
 
 
 def random_data_generator(low, high, shapes, lod_levels, for_parallel=True):

python/paddle/fluid/layers/nn.py

@@ -18,6 +18,7 @@ All layers just related to the neural network.
 from __future__ import print_function
 
 import numpy as np
+import os
 from ..layer_helper import LayerHelper
 from ..initializer import Normal, Constant
 from ..framework import Variable, OpProtoHolder
@@ -341,126 +342,128 @@ def embedding(input,
     return tmp
 
 
-@templatedoc(op_type="lstm")
-def dynamic_lstm(input,
-                 size,
-                 h_0=None,
-                 c_0=None,
-                 param_attr=None,
-                 bias_attr=None,
-                 use_peepholes=True,
-                 is_reverse=False,
-                 gate_activation='sigmoid',
-                 cell_activation='tanh',
-                 candidate_activation='tanh',
-                 dtype='float32',
-                 name=None):
-    """
-    ${comment}
-
-    Args:
-        input (Variable): ${input_comment}
-        size (int): 4 * hidden size.
-        h_0(Variable): The initial hidden state is an optional input, default is zero.
-                       This is a tensor with shape (N x D), where N is the
-                       batch size and D is the hidden size.
-        c_0(Variable): The initial cell state is an optional input, default is zero.
-                       This is a tensor with shape (N x D), where N is the
-                       batch size. `h_0` and `c_0` can be NULL but only at the same time.
-        param_attr(ParamAttr|None): The parameter attribute for the learnable
-                               hidden-hidden weights.
+if os.name != 'nt':
 
-                               - Weights = {:math:`W_{ch}, W_{ih}, \
-                                                W_{fh}, W_{oh}`}
-                               - The shape is (D x 4D), where D is the hidden
-                                 size.
-
-                               If it is set to None or one attribute of ParamAttr,
-                               dynamic_lstm will create ParamAttr as param_attr.
-                               If the Initializer of the param_attr is not set, the
-                               parameter is initialized with Xavier. Default: None.
-        bias_attr (ParamAttr|None): The bias attribute for the learnable bias
-                              weights, which contains two parts, input-hidden
-                              bias weights and peephole connections weights if
-                              setting `use_peepholes` to `True`.
-
-                              1. `use_peepholes = False`
-                                 - Biases = {:math:`b_c, b_i, b_f, b_o`}.
-                                 - The shape is (1 x 4D).
-                              2. `use_peepholes = True`
-                                 - Biases = { :math:`b_c, b_i, b_f, b_o, W_{ic}, \
-                                                 W_{fc}, W_{oc}`}.
-                                 - The shape is (1 x 7D).
-
-                              If it is set to None or one attribute of ParamAttr,
-                              dynamic_lstm will create ParamAttr as bias_attr.
-                              If the Initializer of the bias_attr is not set,
-                              the bias is initialized zero. Default: None.
-        use_peepholes (bool): ${use_peepholes_comment}
-        is_reverse (bool): ${is_reverse_comment}
-        gate_activation (str): ${gate_activation_comment}
-        cell_activation (str): ${cell_activation_comment}
-        candidate_activation (str): ${candidate_activation_comment}
-        dtype (str): Data type. Choices = ["float32", "float64"], default "float32".
-        name (str|None): A name for this layer(optional). If set None, the layer
-                         will be named automatically.
-
-    Returns:
-        tuple: The hidden state, and cell state of LSTM. The shape of both \
-        is (T x D), and lod is the same with the `input`.
-
-    Examples:
-        .. code-block:: python
-
-            hidden_dim = 512
-            forward_proj = fluid.layers.fc(input=input_seq, size=hidden_dim * 4,
-                                           bias_attr=False)
-            forward, _ = fluid.layers.dynamic_lstm(
-                input=forward_proj, size=hidden_dim * 4, use_peepholes=False)
-    """
-    assert bias_attr is not False, "bias_attr should not be False in dynamic_lstmp."
-    helper = LayerHelper('lstm', **locals())
-    size = size // 4
-    weight = helper.create_parameter(
-        attr=helper.param_attr, shape=[size, 4 * size], dtype=dtype)
-    bias_size = [1, 7 * size]
-    if not use_peepholes:
-        bias_size[1] = 4 * size
-    bias = helper.create_parameter(
-        attr=helper.bias_attr, shape=bias_size, dtype=dtype, is_bias=True)
+    @templatedoc(op_type="lstm")
+    def dynamic_lstm(input,
+                     size,
+                     h_0=None,
+                     c_0=None,
+                     param_attr=None,
+                     bias_attr=None,
+                     use_peepholes=True,
+                     is_reverse=False,
+                     gate_activation='sigmoid',
+                     cell_activation='tanh',
+                     candidate_activation='tanh',
+                     dtype='float32',
+                     name=None):
+        """
+        ${comment}
+
+        Args:
+            input (Variable): ${input_comment}
+            size (int): 4 * hidden size.
+            h_0(Variable): The initial hidden state is an optional input, default is zero.
+                           This is a tensor with shape (N x D), where N is the
+                           batch size and D is the hidden size.
+            c_0(Variable): The initial cell state is an optional input, default is zero.
+                           This is a tensor with shape (N x D), where N is the
+                           batch size. `h_0` and `c_0` can be NULL but only at the same time.
+            param_attr(ParamAttr|None): The parameter attribute for the learnable
+                                   hidden-hidden weights.
+
+                                   - Weights = {:math:`W_{ch}, W_{ih}, \
+                                                    W_{fh}, W_{oh}`}
+                                   - The shape is (D x 4D), where D is the hidden
+                                     size.
+
+                                   If it is set to None or one attribute of ParamAttr,
+                                   dynamic_lstm will create ParamAttr as param_attr.
+                                   If the Initializer of the param_attr is not set, the
+                                   parameter is initialized with Xavier. Default: None.
+            bias_attr (ParamAttr|None): The bias attribute for the learnable bias
+                                  weights, which contains two parts, input-hidden
+                                  bias weights and peephole connections weights if
+                                  setting `use_peepholes` to `True`.
+
+                                  1. `use_peepholes = False`
+                                     - Biases = {:math:`b_c, b_i, b_f, b_o`}.
+                                     - The shape is (1 x 4D).
+                                  2. `use_peepholes = True`
+                                     - Biases = { :math:`b_c, b_i, b_f, b_o, W_{ic}, \
+                                                     W_{fc}, W_{oc}`}.
+                                     - The shape is (1 x 7D).
+
+                                  If it is set to None or one attribute of ParamAttr,
+                                  dynamic_lstm will create ParamAttr as bias_attr.
+                                  If the Initializer of the bias_attr is not set,
+                                  the bias is initialized zero. Default: None.
+            use_peepholes (bool): ${use_peepholes_comment}
+            is_reverse (bool): ${is_reverse_comment}
+            gate_activation (str): ${gate_activation_comment}
+            cell_activation (str): ${cell_activation_comment}
+            candidate_activation (str): ${candidate_activation_comment}
+            dtype (str): Data type. Choices = ["float32", "float64"], default "float32".
+            name (str|None): A name for this layer(optional). If set None, the layer
+                             will be named automatically.
+
+        Returns:
+            tuple: The hidden state, and cell state of LSTM. The shape of both \
+            is (T x D), and lod is the same with the `input`.
+
+        Examples:
+            .. code-block:: python
+
+                hidden_dim = 512
+                forward_proj = fluid.layers.fc(input=input_seq, size=hidden_dim * 4,
+                                               bias_attr=False)
+                forward, _ = fluid.layers.dynamic_lstm(
+                    input=forward_proj, size=hidden_dim * 4, use_peepholes=False)
+        """
+        assert bias_attr is not False, "bias_attr should not be False in dynamic_lstmp."
+        helper = LayerHelper('lstm', **locals())
+        size = size // 4
+        weight = helper.create_parameter(
+            attr=helper.param_attr, shape=[size, 4 * size], dtype=dtype)
+        bias_size = [1, 7 * size]
+        if not use_peepholes:
+            bias_size[1] = 4 * size
+        bias = helper.create_parameter(
+            attr=helper.bias_attr, shape=bias_size, dtype=dtype, is_bias=True)
 
-    hidden = helper.create_variable_for_type_inference(dtype)
-    cell = helper.create_variable_for_type_inference(dtype)
-    batch_gate = helper.create_variable_for_type_inference(dtype)
-    batch_cell_pre_act = helper.create_variable_for_type_inference(dtype)
-    inputs = {'Input': input, 'Weight': weight, 'Bias': bias}
-    batch_size = input.shape[0]
-    if h_0:
-        assert h_0.shape == (batch_size, size), \
-            'The shape of h0 should be (batch_size, %d)' % size
-        inputs['H0'] = h_0
-    if c_0:
-        assert c_0.shape == (batch_size, size), \
-            'The shape of c0 should be (batch_size, %d)' % size
-        inputs['C0'] = c_0
+        hidden = helper.create_variable_for_type_inference(dtype)
+        cell = helper.create_variable_for_type_inference(dtype)
+        batch_gate = helper.create_variable_for_type_inference(dtype)
+        batch_cell_pre_act = helper.create_variable_for_type_inference(dtype)
+        inputs = {'Input': input, 'Weight': weight, 'Bias': bias}
+        batch_size = input.shape[0]
+        if h_0:
+            assert h_0.shape == (batch_size, size), \
+                'The shape of h0 should be (batch_size, %d)' % size
+            inputs['H0'] = h_0
+        if c_0:
+            assert c_0.shape == (batch_size, size), \
+                'The shape of c0 should be (batch_size, %d)' % size
+            inputs['C0'] = c_0
 
-    helper.append_op(
-        type='lstm',
-        inputs=inputs,
-        outputs={
-            'Hidden': hidden,
-            'Cell': cell,
-            'BatchGate': batch_gate,
-            'BatchCellPreAct': batch_cell_pre_act
-        },
-        attrs={
-            'use_peepholes': use_peepholes,
-            'is_reverse': is_reverse,
-            'gate_activation': gate_activation,
-            'cell_activation': cell_activation,
-            'candidate_activation': candidate_activation
-        })
-    return hidden, cell
+        helper.append_op(
+            type='lstm',
+            inputs=inputs,
+            outputs={
+                'Hidden': hidden,
+                'Cell': cell,
+                'BatchGate': batch_gate,
+                'BatchCellPreAct': batch_cell_pre_act
+            },
+            attrs={
+                'use_peepholes': use_peepholes,
+                'is_reverse': is_reverse,
+                'gate_activation': gate_activation,
+                'cell_activation': cell_activation,
+                'candidate_activation': candidate_activation
+            })
+        return hidden, cell
 
 
 def dynamic_lstmp(input,
@@ -959,39 +962,43 @@ def linear_chain_crf(input, label, param_attr=None):
     return log_likelihood
 
 
-@templatedoc()
-def crf_decoding(input, param_attr, label=None):
-    """
-    ${comment}
+if os.name != 'nt':
 
-    Args:
-        input(${emission_type}): ${emission_comment}
+    @templatedoc()
+    def crf_decoding(input, param_attr, label=None):
+        """
+        ${comment}
 
-        param_attr(ParamAttr): The parameter attribute for training.
+        Args:
+            input(${emission_type}): ${emission_comment}
 
-        label(${label_type}): ${label_comment}
+            param_attr(ParamAttr): The parameter attribute for training.
 
-    Returns:
-        Variable: ${viterbi_path_comment}
+            label(${label_type}): ${label_comment}
 
-    Examples:
-        .. code-block:: python
+        Returns:
+            Variable: ${viterbi_path_comment}
 
-           crf_decode = layers.crf_decoding(
-                input=hidden, param_attr=ParamAttr(name="crfw"))
-    """
-    helper = LayerHelper('crf_decoding', **locals())
-    transition = helper.get_parameter(param_attr.name)
-    viterbi_path = helper.create_variable_for_type_inference(
-        dtype=helper.input_dtype())
-    helper.append_op(
-        type='crf_decoding',
-        inputs={"Emission": [input],
+        Examples:
+            .. code-block:: python
+
+               crf_decode = layers.crf_decoding(
+                    input=hidden, param_attr=ParamAttr(name="crfw"))
+        """
+        helper = LayerHelper('crf_decoding', **locals())
+        transition = helper.get_parameter(param_attr.name)
+        viterbi_path = helper.create_variable_for_type_inference(
+            dtype=helper.input_dtype())
+        helper.append_op(
+            type='crf_decoding',
+            inputs={
+                "Emission": [input],
                 "Transition": transition,
-                "Label": label},
-        outputs={"ViterbiPath": [viterbi_path]})
+                "Label": label
+            },
+            outputs={"ViterbiPath": [viterbi_path]})
 
-    return viterbi_path
+        return viterbi_path
 
 
 @templatedoc()
@@ -5538,42 +5545,48 @@ def label_smooth(label,
     return smooth_label
 
 
-@templatedoc()
-def roi_pool(input, rois, pooled_height=1, pooled_width=1, spatial_scale=1.0):
-    """
-    ${comment}
-
-    Args:
-        input (Variable): ${x_comment}
-        rois (Variable): ROIs (Regions of Interest) to pool over.
-        pooled_height (integer): ${pooled_height_comment} Default: 1
-        pooled_width (integer): ${pooled_width_comment} Default: 1
-        spatial_scale (float): ${spatial_scale_comment} Default: 1.0
-
-    Returns:
-        Variable: ${out_comment}.
-
-    Examples:
-        .. code-block:: python
-
-            pool_out = fluid.layers.roi_pool(input=x, rois=rois, 7, 7, 1.0)
-    """
-    helper = LayerHelper('roi_pool', **locals())
-    dtype = helper.input_dtype()
-    pool_out = helper.create_variable_for_type_inference(dtype)
-    argmaxes = helper.create_variable_for_type_inference(dtype='int32')
-    helper.append_op(
-        type="roi_pool",
-        inputs={"X": input,
-                "ROIs": rois},
-        outputs={"Out": pool_out,
-                 "Argmax": argmaxes},
-        attrs={
-            "pooled_height": pooled_height,
-            "pooled_width": pooled_width,
-            "spatial_scale": spatial_scale
-        })
-    return pool_out
+if os.name != 'nt':
+
+    @templatedoc()
+    def roi_pool(input,
+                 rois,
+                 pooled_height=1,
+                 pooled_width=1,
+                 spatial_scale=1.0):
+        """
+        ${comment}
+
+        Args:
+            input (Variable): ${x_comment}
+            rois (Variable): ROIs (Regions of Interest) to pool over.
+            pooled_height (integer): ${pooled_height_comment} Default: 1
+            pooled_width (integer): ${pooled_width_comment} Default: 1
+            spatial_scale (float): ${spatial_scale_comment} Default: 1.0
+
+        Returns:
+            Variable: ${out_comment}.
+
+        Examples:
+            .. code-block:: python
+
+                pool_out = fluid.layers.roi_pool(input=x, rois=rois, 7, 7, 1.0)
+        """
+        helper = LayerHelper('roi_pool', **locals())
+        dtype = helper.input_dtype()
+        pool_out = helper.create_variable_for_type_inference(dtype)
+        argmaxes = helper.create_variable_for_type_inference(dtype='int32')
+        helper.append_op(
+            type="roi_pool",
+            inputs={"X": input,
+                    "ROIs": rois},
+            outputs={"Out": pool_out,
+                     "Argmax": argmaxes},
+            attrs={
+                "pooled_height": pooled_height,
+                "pooled_width": pooled_width,
+                "spatial_scale": spatial_scale
+            })
+        return pool_out
 
 
 @templatedoc()
@@ -6822,7 +6835,7 @@ def prelu(x, mode, param_attr=None, name=None):
         alpha_shape = x.shape
     dtype = helper.input_dtype(input_param_name='x')
     alpha = helper.create_parameter(
-        attr=param_attr,
+        attr=helper.param_attr,
         shape=alpha_shape,
         dtype='float32',
         is_bias=False,

python/paddle/fluid/layers/ops.py

@@ -13,6 +13,7 @@
 # limitations under the License.
 
 from __future__ import print_function
+import os
 from .layer_function_generator import generate_layer_fn, generate_layer_fn_noattr
 from .. import core
 from ..framework import convert_np_dtype_to_dtype_
@@ -99,27 +100,26 @@ Examples:
     >>> result = fluid.layers.hard_shrink(x=data, threshold=0.3)
 """
 
-__all__ += ['cumsum']
-
-_cum_sum_ = generate_layer_fn('cumsum')
-
-
-def cumsum(x, axis=None, exclusive=None, reverse=None):
-    locals_var = locals().keys()
-    kwargs = dict()
-    for name in locals_var:
-        val = locals()[name]
-        if val is not None:
-            kwargs[name] = val
-    return _cum_sum_(**kwargs)
-
-
-cumsum.__doc__ = _cum_sum_.__doc__ + """
-Examples:
-
-    >>> data = fluid.layers.data(name="input", shape=[32, 784])
-    >>> result = fluid.layers.cumsum(data, axis=0)
-"""
+if os.name != 'nt':
+    __all__ += ['cumsum']
+
+    _cum_sum_ = generate_layer_fn('cumsum')
+
+    def cumsum(x, axis=None, exclusive=None, reverse=None):
+        locals_var = locals().keys()
+        kwargs = dict()
+        for name in locals_var:
+            val = locals()[name]
+            if val is not None:
+                kwargs[name] = val
+        return _cum_sum_(**kwargs)
+
+    cumsum.__doc__ = _cum_sum_.__doc__ + """
+    Examples:
+    
+        >>> data = fluid.layers.data(name="input", shape=[32, 784])
+        >>> result = fluid.layers.cumsum(data, axis=0)
+    """
 
 __all__ += ['thresholded_relu']
 

python/paddle/fluid/tests/unittests/test_dist_base.py

@@ -105,7 +105,7 @@ class TestDistRunnerBase(object):
             build_stra.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.AllReduce
 
         if args.batch_merge_repeat > 1:
-            pass_builder = build_stra._create_passes_from_strategy()
+            pass_builder = build_stra._finalize_strategy_and_create_passes()
             mypass = pass_builder.insert_pass(
                 len(pass_builder.all_passes()) - 2, "multi_batch_merge_pass")
             mypass.set_int("num_repeats", args.batch_merge_repeat)

python/paddle/fluid/tests/unittests/test_pass_builder.py

@@ -94,7 +94,12 @@ class TestPassBuilder(unittest.TestCase):
 
     def test_parallel_testing_with_new_strategy(self):
         build_strategy = fluid.BuildStrategy()
-        pass_builder = build_strategy._create_passes_from_strategy()
+        self.assertFalse(build_strategy.fuse_elewise_add_act_ops)
+        build_strategy.fuse_elewise_add_act_ops = True
+        pass_builder = build_strategy._finalize_strategy_and_create_passes()
+        self.assertTrue("fuse_elewise_add_act_pass" in
+                        [p.type() for p in pass_builder.all_passes()])
+
         origin_len = len(pass_builder.all_passes())
 
         viz_pass = pass_builder.append_pass("graph_viz_pass")

python/requirements.txt

 requests==2.9.2
 numpy>=1.12,<=1.14 #TODO:change to ">=1.12" when numpy fix bug in 1.15 and higher version
 protobuf==3.1
-recordio>=0.1.0
+recordio>=0.1.0; sys_platform != 'win32'
 matplotlib==2.2.3 # TODO: let python3 paddlepaddle package use latest matplotlib
 rarfile
 scipy>=0.19.0

python/setup.py.in

@@ -9,7 +9,7 @@ class BinaryDistribution(Distribution):
 
 RC      = 0
 
-
+ext_name = '.dll' if os.name == 'nt' else '.so'
 
 def git_commit():
     try:
@@ -136,10 +136,13 @@ if '${WITH_FLUID_ONLY}'== 'OFF':
                    '${PADDLE_BINARY_DIR}/paddle/legacy/pserver/paddle_pserver_main',
                    '${PADDLE_BINARY_DIR}/paddle/scripts/paddle']
 
-package_data={'paddle.fluid': ['core.so']}
+package_data={'paddle.fluid': ['core' + (ext_name if os.name != 'nt' else '.pyd')]}
+if os.name == 'nt':
+    package_data['paddle.fluid'] += ['openblas' + ext_name]
+
 if '${WITH_FLUID_ONLY}'== 'OFF':
-    package_data['paddle.v2.master']=['libpaddle_master.so']
-    package_data['py_paddle']=['*.py','_swig_paddle.so']
+    package_data['paddle.v2.master']=['libpaddle_master' + ext_name]
+    package_data['py_paddle']=['*.py','_swig_paddle' +  + ext_name]
 
 package_dir={
     '': '${PADDLE_BINARY_DIR}/python',
@@ -153,13 +156,15 @@ if '${WITH_FLUID_ONLY}'== 'OFF':
     package_dir['py_paddle']='${PADDLE_BINARY_DIR}/python/py_paddle'
 
 # put all thirdparty libraries in paddle.libs
-package_data['paddle.libs']=['libwarpctc.so']
 libs_path='${PADDLE_BINARY_DIR}/python/paddle/libs'
-shutil.copy('${WARPCTC_LIBRARIES}', libs_path)
+if os.name != 'nt':
+    package_data['paddle.libs']= []
+    package_data['paddle.libs']=['libwarpctc' + ext_name]
+    shutil.copy('${WARPCTC_LIBRARIES}', libs_path)
 if '${WITH_MKL}' == 'ON':
     shutil.copy('${MKLML_LIB}', libs_path)
     shutil.copy('${MKLML_IOMP_LIB}', libs_path)
-    package_data['paddle.libs']+=['libmklml_intel.so','libiomp5.so']
+    package_data['paddle.libs']+=['libmklml_intel' + ext_name,'libiomp5' + ext_name]
 if '${CMAKE_BUILD_TYPE}' == 'Release':
     # only change rpath in Release mode.
     if '${WITH_MKLDNN}' == 'ON':
@@ -187,36 +192,47 @@ if '${WITH_NGRAPH}' == 'ON':
                                   '${NGRAPH_CPU_LIB_NAME}',
                                   '${NGRAPH_TBB_LIB_NAME}']
 # remove unused paddle/libs/__init__.py
-os.remove(libs_path+'/__init__.py')
+if os.path.isfile(libs_path+'/__init__.py'):
+    os.remove(libs_path+'/__init__.py')
 package_dir['paddle.libs']=libs_path
 
-# change rpath of core.so, add $ORIGIN/../libs/ to it.
-# The reason is that libwarpctc.so, libiomp5.so etc are in paddle.libs, and
-# core.so is in paddle.fluid, thus paddle/fluid/../libs will pointer to above libraries.
+# change rpath of core.ext, add $ORIGIN/../libs/ to it.
+# The reason is that libwarpctc.ext, libiomp5.ext etc are in paddle.libs, and
+# core.ext is in paddle.fluid, thus paddle/fluid/../libs will pointer to above libraries.
 # This operation will fix https://github.com/PaddlePaddle/Paddle/issues/3213
 if '${CMAKE_BUILD_TYPE}' == 'Release':
-    # only change rpath in Release mode, since in Debug mode, core.so is too large to be changed.
-    if "@APPLE@" == "1":
-        command = "install_name_tool -id \"@loader_path/../libs/\" ${PADDLE_BINARY_DIR}/python/paddle/fluid/core.so"
-    else:
-        command = "patchelf --set-rpath '$ORIGIN/../libs/' ${PADDLE_BINARY_DIR}/python/paddle/fluid/core.so"
-    if os.system(command) != 0:
-        raise Exception("patch core.so failed, command: %s" % command)
-    if '${WITH_FLUID_ONLY}'== 'OFF':
-        # change rpath of _swig_paddle.so.
+    if os.name != 'nt':
+        # only change rpath in Release mode, since in Debug mode, core.xx is too large to be changed.
         if "@APPLE@" == "1":
-            command = "install_name_tool -id \"@loader_path/../paddle/libs/\" ${PADDLE_BINARY_DIR}/python/py_paddle/_swig_paddle.so"
+            command = "install_name_tool -id \"@loader_path/../libs/\" ${PADDLE_BINARY_DIR}/python/paddle/fluid/core" + ext_name
         else:
-            command = "patchelf --set-rpath '$ORIGIN/../paddle/libs/' ${PADDLE_BINARY_DIR}/python/py_paddle/_swig_paddle.so"
+            command = "patchelf --set-rpath '$ORIGIN/../libs/' ${PADDLE_BINARY_DIR}/python/paddle/fluid/core" + ext_name
         if os.system(command) != 0:
-            raise Exception("patch _swig_paddle.so failed, command: %s" % command)
+            raise Exception("patch core.%s failed, command: %s" % (ext_name, command))
+        if '${WITH_FLUID_ONLY}'== 'OFF':
+            # change rpath of _swig_paddle.xx.
+            if "@APPLE@" == "1":
+                command = "install_name_tool -id \"@loader_path/../paddle/libs/\" ${PADDLE_BINARY_DIR}/python/py_paddle/_swig_paddle" + ext_name
+            else:
+                command = "patchelf --set-rpath '$ORIGIN/../paddle/libs/' ${PADDLE_BINARY_DIR}/python/py_paddle/_swig_paddle" + ext_name
+            if os.system(command) != 0:
+                raise Exception("patch _swig_paddle.%s failed, command: %s" % (ext_name, command))
+
+ext_modules = [Extension('_foo', ['stub.cc'])]
+if os.name == 'nt':
+    # fix the path separator under windows
+    fix_package_dir = {}
+    for k, v in package_dir.items():
+        fix_package_dir[k] = v.replace('/', '\\')
+    package_dir = fix_package_dir
+    ext_modules = []
 
 setup(name='${PACKAGE_NAME}',
       version='${PADDLE_VERSION}',
       description='Parallel Distributed Deep Learning',
       install_requires=setup_requires,
       packages=packages,
-      ext_modules=[Extension('_foo', ['stub.cc'])],
+      ext_modules=ext_modules,
       package_data=package_data,
       package_dir=package_dir,
       scripts=paddle_bins