Merge branch 'develop' into prv-calibration

.gitignore

@@ -4,6 +4,7 @@ paddle/operators/tensor.save
 python/paddle/v2/fluid/tests/book/image_classification_resnet.inference.model/
 python/paddle/v2/fluid/tests/book/image_classification_vgg.inference.model/
 python/paddle/v2/fluid/tests/book/label_semantic_roles.inference.model/
+paddle/fluid/operators/distributed/send_recv.proto
 *.DS_Store
 *.vs
 build/
@@ -28,4 +29,5 @@ third_party/
 build_*
 # clion workspace.
 cmake-build-*
+paddle/fluid/operators/distributed/send_recv.proto
 model_test

AUTHORS.md

@@ -43,6 +43,7 @@
 | qingqing01 | Qing-Qing Dang |
 | reyoung | Yang Yu |
 | Superjom | Chun-Wei Yan |
+| tensor-tang | Jian Tang |
 | tianbingsz | Tian-Bing Xu |
 | tpatejko | Tomasz Patejko |
 | typhoonzero | Yi Wu |

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)
@@ -41,6 +46,7 @@ option(WITH_GPU         "Compile PaddlePaddle with NVIDIA GPU"          ${CUDA_F
 option(WITH_AMD_GPU     "Compile PaddlePaddle with AMD GPU"             OFF)
 option(WITH_AVX         "Compile PaddlePaddle with AVX intrinsics"      ${AVX_FOUND})
 option(WITH_MKL         "Compile PaddlePaddle with MKL support."        ${AVX_FOUND})
+option(WITH_NGRAPH      "Compile PaddlePaddle with nGraph support."     OFF)
 option(WITH_DSO         "Compile PaddlePaddle with dynamic linked CUDA" ON)
 option(WITH_TESTING     "Compile PaddlePaddle with unit testing"        OFF)
 option(WITH_SWIG_PY     "Compile PaddlePaddle with inference api"       ON)
@@ -65,6 +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)
@@ -103,6 +111,8 @@ if(ANDROID OR IOS)
         "Disable RDMA when cross-compiling for Android and iOS" FORCE)
     set(WITH_MKL OFF CACHE STRING
         "Disable MKL when cross-compiling for Android and iOS" FORCE)
+    set(WITH_NGRAPH OFF CACHE STRING
+        "Disable nGraph when cross-compiling for Android and iOS" FORCE)
     set(WITH_GOLANG OFF CACHE STRING
         "Disable golang when cross-compiling for Android and iOS" FORCE)
 
@@ -171,6 +181,7 @@ include(external/protobuf)  # download, build, install protobuf
 include(external/python)    # download, build, install python
 include(external/openblas)  # download, build, install openblas
 include(external/mkldnn)    # download, build, install mkldnn
+include(external/ngraph)    # download, build, install nGraph
 include(external/swig)      # download, build, install swig
 include(external/boost)     # download boost
 include(external/any)       # download libn::any
@@ -304,7 +315,6 @@ endif()
 
 if (ON_INFER)
     message(STATUS "On inference mode, will take place some specific optimization.")
-    add_definitions(-DPADDLE_ON_INFERENCE)
 else()
     #TODO(luotao), combine this warning with `make inference_lib_dist` command.
     message(WARNING "On inference mode, will take place some specific optimization. Turn on the ON_INFER flag when building inference_lib only.")

cmake/configure.cmake

@@ -218,3 +218,7 @@ endif(WITH_GRPC)
 if(WITH_BRPC_RDMA)
     add_definitions(-DPADDLE_WITH_BRPC_RDMA)
 endif(WITH_BRPC_RDMA)
+
+if(ON_INFER)
+    add_definitions(-DPADDLE_ON_INFERENCE)
+endif(ON_INFER)

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/mkldnn.cmake

@@ -37,7 +37,6 @@ SET(CMAKE_INSTALL_RPATH_USE_LINK_PATH TRUE)
 SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${MKLDNN_INSTALL_DIR}/lib")
 
 INCLUDE_DIRECTORIES(${MKLDNN_INC_DIR}) # For MKLDNN code to include internal headers.
-INCLUDE_DIRECTORIES(${THIRD_PARTY_PATH}/install) # For Paddle code to include mkldnn.h
 
 IF(${CBLAS_PROVIDER} STREQUAL "MKLML")
     SET(MKLDNN_DEPENDS   ${MKLML_PROJECT})

cmake/external/ngraph.cmake

+# 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.
+
+add_library(ngraph INTERFACE)
+
+IF(WIN32 OR APPLE)
+    MESSAGE(WARNING
+        "Windows or Mac is not supported with nGraph in Paddle yet."
+        "Force WITH_NGRAPH=OFF")
+    SET(WITH_NGRAPH OFF CACHE STRING "Disable nGraph in Windows and MacOS" FORCE)
+ENDIF()
+
+IF(${WITH_NGRAPH} AND NOT ${WITH_MKLDNN})
+    MESSAGE(WARNING
+        "nGraph needs mkl-dnn to be enabled."
+        "Force WITH_NGRAPH=OFF")
+    SET(WITH_NGRAPH OFF CACHE STRING "Disable nGraph if mkl-dnn is disabled" FORCE)
+ENDIF()
+
+IF(NOT ${WITH_NGRAPH})
+    return()
+ENDIF()
+
+INCLUDE(ExternalProject)
+
+SET(NGRAPH_PROJECT         "extern_ngraph")
+SET(NGRAPH_VERSION         "0.9")
+SET(NGRAPH_GIT_TAG         "f9fd9d4cc318dc59dd4b68448e7fbb5f67a28bd0")
+SET(NGRAPH_SOURCES_DIR     ${THIRD_PARTY_PATH}/ngraph)
+SET(NGRAPH_INSTALL_DIR     ${THIRD_PARTY_PATH}/install/ngraph)
+SET(NGRAPH_INC_DIR         ${NGRAPH_INSTALL_DIR}/include)
+SET(NGRAPH_SHARED_LIB_NAME libngraph.so.${NGRAPH_VERSION})
+SET(NGRAPH_CPU_LIB_NAME    libcpu_backend.so)
+SET(NGRAPH_TBB_LIB_NAME    libtbb.so.2)
+SET(NGRAPH_GIT_REPO        "https://github.com/NervanaSystems/ngraph.git")
+
+ExternalProject_Add(
+    ${NGRAPH_PROJECT}
+    ${EXTERNAL_PROJECT_LOG_ARGS}
+    DEPENDS             ${MKLDNN_PROJECT} ${MKLML_PROJECT}
+    GIT_REPOSITORY      ${NGRAPH_GIT_REPO}
+    GIT_TAG             ${NGRAPH_GIT_TAG}
+    PREFIX              ${NGRAPH_SOURCES_DIR}
+    UPDATE_COMMAND      ""
+    CMAKE_ARGS          -DCMAKE_INSTALL_PREFIX=${NGRAPH_INSTALL_DIR}
+    CMAKE_ARGS          -DNGRAPH_UNIT_TEST_ENABLE=FALSE
+    CMAKE_ARGS          -DNGRAPH_TOOLS_ENABLE=FALSE
+    CMAKE_ARGS          -DNGRAPH_INTERPRETER_ENABLE=FALSE
+    CMAKE_ARGS          -DNGRAPH_DEX_ONLY=TRUE
+    CMAKE_ARGS          -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}
+    CMAKE_ARGS          -DMKLDNN_INCLUDE_DIR=${MKLDNN_INC_DIR}
+    CMAKE_ARGS          -DMKLDNN_LIB_DIR=${MKLDNN_INSTALL_DIR}/lib
+)
+
+if(UNIX AND NOT APPLE)
+    include(GNUInstallDirs)
+    SET(NGRAPH_LIB_DIR ${NGRAPH_INSTALL_DIR}/${CMAKE_INSTALL_LIBDIR})
+else()
+    SET(NGRAPH_LIB_DIR ${NGRAPH_INSTALL_DIR}/lib)
+endif()
+MESSAGE(STATUS "nGraph lib will be installed at: ${NGRAPH_LIB_DIR}")
+
+SET(NGRAPH_SHARED_LIB      ${NGRAPH_LIB_DIR}/${NGRAPH_SHARED_LIB_NAME})
+SET(NGRAPH_CPU_LIB         ${NGRAPH_LIB_DIR}/${NGRAPH_CPU_LIB_NAME})
+SET(NGRAPH_TBB_LIB         ${NGRAPH_LIB_DIR}/${NGRAPH_TBB_LIB_NAME})
+
+# Workaround for nGraph expecting mklml to be in mkldnn install directory.
+ExternalProject_Add_Step(
+    ${NGRAPH_PROJECT}
+    PrepareMKL
+    COMMAND ${CMAKE_COMMAND} -E create_symlink ${MKLML_LIB} ${MKLDNN_INSTALL_DIR}/lib/libmklml_intel.so
+    COMMAND ${CMAKE_COMMAND} -E create_symlink ${MKLML_IOMP_LIB} ${MKLDNN_INSTALL_DIR}/lib/libiomp5.so
+    DEPENDEES download
+    DEPENDERS configure
+)
+
+add_dependencies(ngraph ${NGRAPH_PROJECT})
+target_compile_definitions(ngraph INTERFACE -DPADDLE_WITH_NGRAPH)
+target_include_directories(ngraph INTERFACE ${NGRAPH_INC_DIR})
+target_link_libraries(ngraph INTERFACE ${NGRAPH_SHARED_LIB})
+LIST(APPEND external_project_dependencies ngraph)

cmake/external/openblas.cmake

@@ -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

@@ -30,66 +30,61 @@ UNSET_VAR(PROTOBUF_LITE_LIBRARY)
 UNSET_VAR(PROTOBUF_LIBRARY)
 UNSET_VAR(PROTOBUF_INCLUDE_DIR)
 UNSET_VAR(Protobuf_PROTOC_EXECUTABLE)
+function(protobuf_generate_python SRCS)
+    # shameless copy from https://github.com/Kitware/CMake/blob/master/Modules/FindProtobuf.cmake
+    if(NOT ARGN)
+        message(SEND_ERROR "Error: PROTOBUF_GENERATE_PYTHON() called without any proto files")
+        return()
+    endif()
 
-if(NOT COMMAND protobuf_generate_python)  # before cmake 3.4, protobuf_genrerate_python is not defined.
-    function(protobuf_generate_python SRCS)
-        # shameless copy from https://github.com/Kitware/CMake/blob/master/Modules/FindProtobuf.cmake
-        if(NOT ARGN)
-            message(SEND_ERROR "Error: PROTOBUF_GENERATE_PYTHON() called without any proto files")
-            return()
-        endif()
-
-        if(PROTOBUF_GENERATE_CPP_APPEND_PATH)
-            # Create an include path for each file specified
-            foreach(FIL ${ARGN})
-                get_filename_component(ABS_FIL ${FIL} ABSOLUTE)
-                get_filename_component(ABS_PATH ${ABS_FIL} PATH)
-                list(FIND _protobuf_include_path ${ABS_PATH} _contains_already)
-                if(${_contains_already} EQUAL -1)
-                    list(APPEND _protobuf_include_path -I ${ABS_PATH})
-                endif()
-            endforeach()
-        else()
-            set(_protobuf_include_path -I ${CMAKE_CURRENT_SOURCE_DIR})
-        endif()
-
-        if(DEFINED PROTOBUF_IMPORT_DIRS AND NOT DEFINED Protobuf_IMPORT_DIRS)
-            set(Protobuf_IMPORT_DIRS "${PROTOBUF_IMPORT_DIRS}")
-        endif()
-
-        if(DEFINED Protobuf_IMPORT_DIRS)
-            foreach(DIR ${Protobuf_IMPORT_DIRS})
-                get_filename_component(ABS_PATH ${DIR} ABSOLUTE)
-                list(FIND _protobuf_include_path ${ABS_PATH} _contains_already)
-                if(${_contains_already} EQUAL -1)
-                    list(APPEND _protobuf_include_path -I ${ABS_PATH})
-                endif()
-            endforeach()
-        endif()
-
-        set(${SRCS})
+    if(PROTOBUF_GENERATE_CPP_APPEND_PATH)
+        # Create an include path for each file specified
         foreach(FIL ${ARGN})
             get_filename_component(ABS_FIL ${FIL} ABSOLUTE)
-            get_filename_component(FIL_WE ${FIL} NAME_WE)
-            if(NOT PROTOBUF_GENERATE_CPP_APPEND_PATH)
-                get_filename_component(FIL_DIR ${FIL} DIRECTORY)
-                if(FIL_DIR)
-                    set(FIL_WE "${FIL_DIR}/${FIL_WE}")
-                endif()
+            get_filename_component(ABS_PATH ${ABS_FIL} PATH)
+            list(FIND _protobuf_include_path ${ABS_PATH} _contains_already)
+            if(${_contains_already} EQUAL -1)
+                list(APPEND _protobuf_include_path -I ${ABS_PATH})
             endif()
+        endforeach()
+    else()
+        set(_protobuf_include_path -I ${CMAKE_CURRENT_SOURCE_DIR})
+    endif()
+    if(DEFINED PROTOBUF_IMPORT_DIRS AND NOT DEFINED Protobuf_IMPORT_DIRS)
+        set(Protobuf_IMPORT_DIRS "${PROTOBUF_IMPORT_DIRS}")
+    endif()
 
-            list(APPEND ${SRCS} "${CMAKE_CURRENT_BINARY_DIR}/${FIL_WE}_pb2.py")
-            add_custom_command(
-                    OUTPUT "${CMAKE_CURRENT_BINARY_DIR}/${FIL_WE}_pb2.py"
-                    COMMAND  ${Protobuf_PROTOC_EXECUTABLE} --python_out ${CMAKE_CURRENT_BINARY_DIR} ${_protobuf_include_path} ${ABS_FIL}
-                    DEPENDS ${ABS_FIL} ${Protobuf_PROTOC_EXECUTABLE}
-                    COMMENT "Running Python protocol buffer compiler on ${FIL}"
-                    VERBATIM )
+    if(DEFINED Protobuf_IMPORT_DIRS)
+        foreach(DIR ${Protobuf_IMPORT_DIRS})
+            get_filename_component(ABS_PATH ${DIR} ABSOLUTE)
+            list(FIND _protobuf_include_path ${ABS_PATH} _contains_already)
+            if(${_contains_already} EQUAL -1)
+                list(APPEND _protobuf_include_path -I ${ABS_PATH})
+            endif()
         endforeach()
+    endif()
 
-        set(${SRCS} ${${SRCS}} PARENT_SCOPE)
-    endfunction()
-endif()
+    set(${SRCS})
+    foreach(FIL ${ARGN})
+        get_filename_component(ABS_FIL ${FIL} ABSOLUTE)
+        get_filename_component(FIL_WE ${FIL} NAME_WE)
+        if(NOT PROTOBUF_GENERATE_CPP_APPEND_PATH)
+            get_filename_component(FIL_DIR ${FIL} DIRECTORY)
+            if(FIL_DIR)
+                set(FIL_WE "${FIL_DIR}/${FIL_WE}")
+            endif()
+        endif()
+        list(APPEND ${SRCS} "${CMAKE_CURRENT_BINARY_DIR}/${FIL_WE}_pb2.py")
+        add_custom_command(
+                OUTPUT "${CMAKE_CURRENT_BINARY_DIR}/${FIL_WE}_pb2.py"
+                COMMAND  ${PROTOBUF_PROTOC_EXECUTABLE} --python_out ${CMAKE_CURRENT_BINARY_DIR} ${_protobuf_include_path} ${ABS_FIL}
+                DEPENDS ${ABS_FIL} ${PROTOBUF_PROTOC_EXECUTABLE}
+                COMMENT "Running Python protocol buffer compiler on ${FIL}"
+                VERBATIM )
+    endforeach()
+
+    set(${SRCS} ${${SRCS}} PARENT_SCOPE)
+endfunction()
 
 # Print and set the protobuf library information,
 # finish this cmake process and exit from this file.
@@ -126,6 +121,7 @@ macro(PROMPT_PROTOBUF_LIB)
     # FIND_Protobuf.cmake uses `Protobuf_PROTOC_EXECUTABLE`.
     # make `protobuf_generate_cpp` happy.
     SET(Protobuf_PROTOC_EXECUTABLE ${PROTOBUF_PROTOC_EXECUTABLE})
+
     FOREACH(dep ${protobuf_DEPS})
         ADD_DEPENDENCIES(protobuf ${dep})
         ADD_DEPENDENCIES(protobuf_lite ${dep})
@@ -144,7 +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 "${PROTOBUF_ROOT}" STREQUAL "")
@@ -192,13 +187,20 @@ FUNCTION(build_protobuf TARGET_NAME BUILD_FOR_HOST)
         SET(OPTIONAL_ARGS
             "-DCMAKE_CXX_COMPILER=${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

@@ -22,175 +22,196 @@ 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}
   SRCS ${src_dir}/${module}/*.h ${PADDLE_BINARY_DIR}/paddle/fluid/inference/libpaddle_fluid.*
-       ${src_dir}/${module}/api/paddle_inference_api.h
+       ${src_dir}/${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})
@@ -198,14 +219,14 @@ 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_inference_api.h for inference-only library
+# only need libpaddle_fluid.so/a and paddle_*.h for inference-only library
 copy(inference_api_lib DEPS fluid_lib_dist
   SRCS ${FLUID_INSTALL_DIR}/paddle/fluid/inference/libpaddle_fluid.*
-       ${FLUID_INSTALL_DIR}/paddle/fluid/inference/paddle_inference_api.h
+       ${FLUID_INSTALL_DIR}/paddle/fluid/inference/paddle_*.h
   DSTS ${FLUID_INFERENCE_INSTALL_DIR}/paddle/lib ${FLUID_INFERENCE_INSTALL_DIR}/paddle/include
 )
 
@@ -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)

cmake/operators.cmake

+set(PART_CUDA_KERNEL_FILES)
+function(op_library TARGET)
+    # op_library is a function to create op library. The interface is same as
+    # cc_library. But it handle split GPU/CPU code and link some common library
+    # for ops.
+    set(cc_srcs)
+    set(cu_srcs)
+    set(hip_cu_srcs)
+    set(miopen_hip_cc_srcs)
+    set(cu_cc_srcs)
+    set(cudnn_cu_cc_srcs)
+    set(CUDNN_FILE)
+    set(mkldnn_cc_srcs)
+    set(MKLDNN_FILE)
+    set(op_common_deps operator op_registry math_function)
+    set(options "")
+    set(oneValueArgs "")
+    set(multiValueArgs SRCS DEPS)
+    set(pybind_flag 0)
+    cmake_parse_arguments(op_library "${options}" "${oneValueArgs}"
+            "${multiValueArgs}" ${ARGN})
+
+    list(LENGTH op_library_SRCS op_library_SRCS_len)
+    if (${op_library_SRCS_len} EQUAL 0)
+        if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.cc)
+            list(APPEND cc_srcs ${TARGET}.cc)
+        endif()
+        if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.cu.cc)
+            list(APPEND cu_cc_srcs ${TARGET}.cu.cc)
+        endif()
+        if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.cu)
+            list(APPEND cu_srcs ${TARGET}.cu)
+        endif()
+        if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.part.cu)
+            set(PART_CUDA_KERNEL_FILES ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.part.cu
+                    ${PART_CUDA_KERNEL_FILES} PARENT_SCOPE)
+            list(APPEND cu_srcs ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.part.cu)
+        endif()
+
+        if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.hip.cu)
+            list(APPEND hip_cu_srcs ${TARGET}.hip.cu)
+        endif()
+        string(REPLACE "_op" "_cudnn_op" CUDNN_FILE "${TARGET}")
+        if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${CUDNN_FILE}.cu.cc)
+            list(APPEND cudnn_cu_cc_srcs ${CUDNN_FILE}.cu.cc)
+        endif()
+        if(WITH_AMD_GPU)
+            string(REPLACE "_op" "_miopen_op" MIOPEN_FILE "${TARGET}")
+            if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${MIOPEN_FILE}.hip.cc)
+                list(APPEND miopen_hip_cc_srcs ${MIOPEN_FILE}.hip.cc)
+            endif()
+        endif()
+        if(WITH_MKLDNN)
+            string(REPLACE "_op" "_mkldnn_op" MKLDNN_FILE "${TARGET}")
+            if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${MKLDNN_FILE}.cc)
+                list(APPEND mkldnn_cc_srcs ${MKLDNN_FILE}.cc)
+            endif()
+        endif()
+    else()
+        foreach(src ${op_library_SRCS})
+            if (${src} MATCHES ".*\\.hip.cu$")
+                list(APPEND hip_cu_srcs ${src})
+            elseif (${src} MATCHES ".*\\.cu$")
+                list(APPEND cu_srcs ${src})
+            elseif(${src} MATCHES ".*_cudnn_op.cu.cc$")
+                list(APPEND cudnn_cu_cc_srcs ${src})
+            elseif(WITH_AMD_GPU AND ${src} MATCHES ".*_miopen_op.hip.cc$")
+                list(APPEND miopen_hip_cc_srcs ${src})
+            elseif(WITH_MKLDNN AND ${src} MATCHES ".*_mkldnn_op.cc$")
+                list(APPEND mkldnn_cc_srcs ${src})
+            elseif(${src} MATCHES ".*\\.cu.cc$")
+                list(APPEND cu_cc_srcs ${src})
+            elseif(${src} MATCHES ".*\\.cc$")
+                list(APPEND cc_srcs ${src})
+            else()
+                message(FATAL_ERROR "${TARGET} Source file ${src} should only be .cc or .cu")
+            endif()
+        endforeach()
+    endif()
+
+    list(LENGTH cc_srcs cc_srcs_len)
+    if (${cc_srcs_len} EQUAL 0)
+        message(FATAL_ERROR "The op library ${TARGET} should contains at least one .cc file")
+    endif()
+    if (WIN32)
+    # remove windows unsupported op, because windows has no nccl, no warpctc such ops.
+    foreach(windows_unsupport_op "nccl_op" "gen_nccl_id_op" "warpctc_op" "hierarchical_sigmoid_op"
+     "crf_decoding_op" "select_op" "lstmp_op" "gru_op" "fusion_gru_op" "lstm_op" "fusion_lstm_op" "cumsum_op"
+      "fusion_seqconv_eltadd_relu_op" "channel_send_op" "channel_create_op" "channel_close_op" "channel_recv_op")
+        if ("${TARGET}" STREQUAL "${windows_unsupport_op}")
+          return()
+        endif()
+    endforeach()
+    endif(WIN32)
+    set(OP_LIBRARY ${TARGET} ${OP_LIBRARY} CACHE INTERNAL "op libs")
+
+    list(LENGTH op_library_DEPS op_library_DEPS_len)
+    if (${op_library_DEPS_len} GREATER 0)
+        set(DEPS_OPS ${TARGET} ${DEPS_OPS} PARENT_SCOPE)
+    endif()
+    if (WITH_GPU)
+        nv_library(${TARGET} SRCS ${cc_srcs} ${cu_cc_srcs} ${cudnn_cu_cc_srcs} ${mkldnn_cc_srcs} ${cu_srcs} DEPS ${op_library_DEPS}
+                ${op_common_deps})
+    elseif (WITH_AMD_GPU)
+        hip_library(${TARGET} SRCS ${cc_srcs} ${hip_cu_srcs} ${miopen_hip_cc_srcs} ${mkldnn_cc_srcs} DEPS ${op_library_DEPS}
+                ${op_common_deps})
+    else()
+        cc_library(${TARGET} SRCS ${cc_srcs} ${mkldnn_cc_srcs} DEPS ${op_library_DEPS}
+            ${op_common_deps})
+    endif()
+
+    # Define operators that don't need pybind here.
+    foreach(manual_pybind_op "compare_op" "logical_op" "nccl_op"
+"tensor_array_read_write_op" "tensorrt_engine_op" "conv_fusion_op")
+        if ("${TARGET}" STREQUAL "${manual_pybind_op}")
+            set(pybind_flag 1)
+        endif()
+    endforeach()
+
+    # The registration of USE_OP, please refer to paddle/fluid/framework/op_registry.h.
+    # Note that it's enough to just adding one operator to pybind in a *_op.cc file.
+    # And for detail pybind information, please see generated paddle/pybind/pybind.h.
+    file(READ ${TARGET}.cc TARGET_CONTENT)
+    string(REGEX MATCH "REGISTER_OPERATOR\\(.*REGISTER_OPERATOR\\(" multi_register "${TARGET_CONTENT}")
+    string(REGEX MATCH "REGISTER_OPERATOR\\([a-z0-9_]*," one_register "${multi_register}")
+    if (one_register STREQUAL "")
+        string(REPLACE "_op" "" TARGET "${TARGET}")
+    else ()
+        string(REPLACE "REGISTER_OPERATOR(" "" TARGET "${one_register}")
+        string(REPLACE "," "" TARGET "${TARGET}")
+    endif()
+
+    # pybind USE_NO_KERNEL_OP
+    # HACK: if REGISTER_OP_CPU_KERNEL presents the operator must have kernel
+    string(REGEX MATCH "REGISTER_OP_CPU_KERNEL" regex_result "${TARGET_CONTENT}")
+    string(REPLACE "_op" "" TARGET "${TARGET}")
+    if (${pybind_flag} EQUAL 0 AND regex_result STREQUAL "")
+        file(APPEND ${pybind_file} "USE_NO_KERNEL_OP(${TARGET});\n")
+        set(pybind_flag 1)
+    endif()
+
+    # pybind USE_CPU_ONLY_OP
+    list(LENGTH cu_srcs cu_srcs_len)
+    list(LENGTH cu_cc_srcs cu_cc_srcs_len)
+    list(LENGTH mkldnn_cc_srcs mkldnn_cc_srcs_len)
+    list(LENGTH hip_cu_srcs hip_cu_srcs_len)
+    list(LENGTH miopen_hip_cc_srcs miopen_hip_cc_srcs_len)
+    if (${pybind_flag} EQUAL 0 AND ${mkldnn_cc_srcs_len} EQUAL 0 AND ${cu_srcs_len} EQUAL 0 AND ${cu_cc_srcs_len} EQUAL 0 AND
+        ${hip_cu_srcs_len} EQUAL 0 AND ${miopen_hip_cc_srcs_len} EQUAL 0)
+        file(APPEND ${pybind_file} "USE_CPU_ONLY_OP(${TARGET});\n")
+        set(pybind_flag 1)
+    endif()
+
+    # pybind USE_OP_DEVICE_KERNEL for CUDNN
+    list(LENGTH cudnn_cu_cc_srcs cudnn_cu_cc_srcs_len)
+    if (WITH_GPU AND ${cudnn_cu_cc_srcs_len} GREATER 0)
+        file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, CUDNN);\n")
+    endif()
+
+    # pybind USE_OP_DEVICE_KERNEL for MIOPEN
+    if (WITH_AMD_GPU AND ${miopen_hip_cc_srcs_len} GREATER 0)
+        file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, MIOPEN);\n")
+    endif()
+
+    # pybind USE_OP_DEVICE_KERNEL for MKLDNN
+    if (WITH_MKLDNN AND ${mkldnn_cc_srcs_len} GREATER 0)
+      # Append first implemented MKLDNN activation operator
+      if (${MKLDNN_FILE} STREQUAL "activation_mkldnn_op")
+        file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(relu, MKLDNN);\n")
+      else()
+        file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, MKLDNN);\n")
+      endif()
+    endif()
+
+    # pybind USE_OP
+    if (${pybind_flag} EQUAL 0)
+      # NOTE(*): activation use macro to regist the kernels, set use_op manually.
+      if(${TARGET} STREQUAL "activation")
+        file(APPEND ${pybind_file} "USE_OP(relu);\n")
+      elseif(${TARGET} STREQUAL "fake_dequantize")
+        file(APPEND ${pybind_file} "USE_OP(fake_dequantize_max_abs);\n")
+      elseif(${TARGET} STREQUAL "fake_quantize")
+        file(APPEND ${pybind_file} "USE_OP(fake_quantize_abs_max);\n")
+      elseif(${TARGET} STREQUAL "tensorrt_engine_op")
+          message(STATUS "Pybind skips [tensorrt_engine_op], for this OP is only used in inference")
+      elseif(${TARGET} STREQUAL "fc")
+        # HACK: fc only have mkldnn and cpu, which would mismatch the cpu only condition
+        file(APPEND ${pybind_file} "USE_CPU_ONLY_OP(${TARGET});\n")
+      else()
+        file(APPEND ${pybind_file} "USE_OP(${TARGET});\n")
+      endif()
+    endif()
+endfunction()
+
+
+function(register_operators)
+    set(options "")
+    set(oneValueArgs "")
+    set(multiValueArgs EXCLUDES DEPS)
+    cmake_parse_arguments(register_operators "${options}" "${oneValueArgs}"
+            "${multiValueArgs}" ${ARGN})
+
+    file(GLOB OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "*_op.cc")
+    string(REPLACE "_mkldnn" "" OPS "${OPS}")
+    string(REPLACE ".cc" "" OPS "${OPS}")
+    list(REMOVE_DUPLICATES OPS)
+    list(LENGTH register_operators_DEPS register_operators_DEPS_len)
+
+    foreach(src ${OPS})
+        list(FIND register_operators_EXCLUDES ${src} _index)
+        if (${_index} EQUAL -1)
+            if (${register_operators_DEPS_len} GREATER 0)
+                op_library(${src} DEPS ${register_operators_DEPS})
+            else()
+                op_library(${src})
+            endif()
+        endif()
+    endforeach()
+endfunction()

cmake/tensorrt.cmake

@@ -34,4 +34,5 @@ if(TENSORRT_FOUND)
         "Current TensorRT version is v${TENSORRT_MAJOR_VERSION}. ")
     include_directories(${TENSORRT_INCLUDE_DIR})
     list(APPEND EXTERNAL_LIBS ${TENSORRT_LIBRARY})
+    add_definitions(-DPADDLE_WITH_TENSORRT)
 endif()

doc/v2/dev/contribute_to_paddle_en.md

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

paddle/fluid/API.spec

@@ -93,17 +93,17 @@ paddle.fluid.layers.edit_distance ArgSpec(args=['input', 'label', 'normalized',
 paddle.fluid.layers.l2_normalize ArgSpec(args=['x', 'axis', 'epsilon', 'name'], varargs=None, keywords=None, defaults=(1e-12, None))
 paddle.fluid.layers.matmul ArgSpec(args=['x', 'y', 'transpose_x', 'transpose_y', 'alpha', 'name'], varargs=None, keywords=None, defaults=(False, False, 1.0, None))
 paddle.fluid.layers.topk ArgSpec(args=['input', 'k', 'name'], varargs=None, keywords=None, defaults=(None,))
-paddle.fluid.layers.warpctc ArgSpec(args=['input', 'label', 'blank', 'norm_by_times'], varargs=None, keywords=None, defaults=(0, False))
+paddle.fluid.layers.warpctc ArgSpec(args=['input', 'label', 'blank', 'norm_by_times', 'use_cudnn'], varargs=None, keywords=None, defaults=(0, False, False))
 paddle.fluid.layers.sequence_reshape ArgSpec(args=['input', 'new_dim'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.transpose ArgSpec(args=['x', 'perm', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.im2sequence ArgSpec(args=['input', 'filter_size', 'stride', 'padding', 'input_image_size', 'out_stride', 'name'], varargs=None, keywords=None, defaults=(1, 1, 0, None, 1, None))
-paddle.fluid.layers.nce ArgSpec(args=['input', 'label', 'num_total_classes', 'sample_weight', 'param_attr', 'bias_attr', 'num_neg_samples', 'name'], varargs=None, keywords=None, defaults=(None, None, None, None, None))
+paddle.fluid.layers.nce ArgSpec(args=['input', 'label', 'num_total_classes', 'sample_weight', 'param_attr', 'bias_attr', 'num_neg_samples', 'name', 'sampler', 'custom_dist', 'seed'], varargs=None, keywords=None, defaults=(None, None, None, None, None, 'uniform', None, 0))
 paddle.fluid.layers.hsigmoid ArgSpec(args=['input', 'label', 'num_classes', 'param_attr', 'bias_attr', 'name'], varargs=None, keywords=None, defaults=(None, None, None))
 paddle.fluid.layers.beam_search ArgSpec(args=['pre_ids', 'pre_scores', 'ids', 'scores', 'beam_size', 'end_id', 'level', 'name'], varargs=None, keywords=None, defaults=(0, None))
 paddle.fluid.layers.row_conv ArgSpec(args=['input', 'future_context_size', 'param_attr', 'act'], varargs=None, keywords=None, defaults=(None, None))
 paddle.fluid.layers.multiplex ArgSpec(args=['inputs', 'index'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.layer_norm ArgSpec(args=['input', 'scale', 'shift', 'begin_norm_axis', 'epsilon', 'param_attr', 'bias_attr', 'act', 'name'], varargs=None, keywords=None, defaults=(True, True, 1, 1e-05, None, None, None, None))
-paddle.fluid.layers.softmax_with_cross_entropy ArgSpec(args=['logits', 'label', 'soft_label', 'ignore_index', 'numeric_stable_mode'], varargs=None, keywords=None, defaults=(False, -100, False))
+paddle.fluid.layers.softmax_with_cross_entropy ArgSpec(args=['logits', 'label', 'soft_label', 'ignore_index', 'numeric_stable_mode', 'return_softmax'], varargs=None, keywords=None, defaults=(False, -100, False, False))
 paddle.fluid.layers.smooth_l1 ArgSpec(args=['x', 'y', 'inside_weight', 'outside_weight', 'sigma'], varargs=None, keywords=None, defaults=(None, None, None))
 paddle.fluid.layers.one_hot ArgSpec(args=['input', 'depth'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.autoincreased_step_counter ArgSpec(args=['counter_name', 'begin', 'step'], varargs=None, keywords=None, defaults=(None, 1, 1))
@@ -128,6 +128,7 @@ paddle.fluid.layers.sequence_scatter ArgSpec(args=['input', 'index', 'updates',
 paddle.fluid.layers.random_crop ArgSpec(args=['x', 'shape', 'seed'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.mean_iou ArgSpec(args=['input', 'label', 'num_classes'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.relu ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.layers.selu ArgSpec(args=['x', 'scale', 'alpha', 'name'], varargs=None, keywords=None, defaults=(None, None, None))
 paddle.fluid.layers.log ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.crop ArgSpec(args=['x', 'shape', 'offsets', 'name'], varargs=None, keywords=None, defaults=(None, None, None))
 paddle.fluid.layers.rank_loss ArgSpec(args=['label', 'left', 'right', 'name'], varargs=None, keywords=None, defaults=(None,))
@@ -184,6 +185,7 @@ paddle.fluid.layers.hash ArgSpec(args=['input', 'hash_size', 'num_hash', 'name']
 paddle.fluid.layers.grid_sampler ArgSpec(args=['x', 'grid', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.log_loss ArgSpec(args=['input', 'label', 'epsilon', 'name'], varargs=None, keywords=None, defaults=(0.0001, None))
 paddle.fluid.layers.add_position_encoding ArgSpec(args=['input', 'alpha', 'beta', 'name'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.layers.bilinear_tensor_product ArgSpec(args=['x', 'y', 'size', 'act', 'name', 'param_attr', 'bias_attr'], varargs=None, keywords=None, defaults=(None, None, None, None))
 paddle.fluid.layers.data ArgSpec(args=['name', 'shape', 'append_batch_size', 'dtype', 'lod_level', 'type', 'stop_gradient'], varargs=None, keywords=None, defaults=(True, 'float32', 0, VarType.LOD_TENSOR, True))
 paddle.fluid.layers.open_files ArgSpec(args=['filenames', 'shapes', 'lod_levels', 'dtypes', 'thread_num', 'buffer_size', 'pass_num', 'is_test'], varargs=None, keywords=None, defaults=(None, None, 1, None))
 paddle.fluid.layers.read_file ArgSpec(args=['reader'], varargs=None, keywords=None, defaults=None)
@@ -273,6 +275,7 @@ paddle.fluid.layers.hard_shrink ArgSpec(args=['x', 'threshold'], varargs=None, k
 paddle.fluid.layers.cumsum ArgSpec(args=['x', 'axis', 'exclusive', 'reverse'], varargs=None, keywords=None, defaults=(None, None, None))
 paddle.fluid.layers.thresholded_relu ArgSpec(args=['x', 'threshold'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.prior_box ArgSpec(args=['input', 'image', 'min_sizes', 'max_sizes', 'aspect_ratios', 'variance', 'flip', 'clip', 'steps', 'offset', 'name', 'min_max_aspect_ratios_order'], varargs=None, keywords=None, defaults=(None, [1.0], [0.1, 0.1, 0.2, 0.2], False, False, [0.0, 0.0], 0.5, None, False))
+paddle.fluid.layers.density_prior_box ArgSpec(args=['input', 'image', 'densities', 'fixed_sizes', 'fixed_ratios', 'variance', 'clip', 'steps', 'offset', 'name'], varargs=None, keywords=None, defaults=(None, None, None, [0.1, 0.1, 0.2, 0.2], False, [0.0, 0.0], 0.5, None))
 paddle.fluid.layers.multi_box_head ArgSpec(args=['inputs', 'image', 'base_size', 'num_classes', 'aspect_ratios', 'min_ratio', 'max_ratio', 'min_sizes', 'max_sizes', 'steps', 'step_w', 'step_h', 'offset', 'variance', 'flip', 'clip', 'kernel_size', 'pad', 'stride', 'name', 'min_max_aspect_ratios_order'], varargs=None, keywords=None, defaults=(None, None, None, None, None, None, None, 0.5, [0.1, 0.1, 0.2, 0.2], True, False, 1, 0, 1, None, False))
 paddle.fluid.layers.bipartite_match ArgSpec(args=['dist_matrix', 'match_type', 'dist_threshold', 'name'], varargs=None, keywords=None, defaults=(None, None, None))
 paddle.fluid.layers.target_assign ArgSpec(args=['input', 'matched_indices', 'negative_indices', 'mismatch_value', 'name'], varargs=None, keywords=None, defaults=(None, None, None))

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/data_device_transform_test.cu

@@ -17,7 +17,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/op_info.h"
 #include "paddle/fluid/framework/op_registry.h"
-#include "paddle/fluid/operators/elementwise_op_function.h"
+#include "paddle/fluid/operators/elementwise/elementwise_op_function.h"
 #include "paddle/fluid/operators/math/math_function.h"
 #include "paddle/fluid/platform/device_context.h"
 #include "paddle/fluid/platform/init.h"

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/details/exception_holder.h

@@ -30,6 +30,8 @@ class ExceptionHolder {
       Catch(exp);
     } catch (platform::EnforceNotMet exp) {
       Catch(exp);
+    } catch (std::exception& ex) {
+      LOG(FATAL) << "std::exception caught, " << ex.what();
     } catch (...) {
       LOG(FATAL) << "Unknown exception caught";
     }

paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.cc

@@ -30,8 +30,8 @@ FastThreadedSSAGraphExecutor::FastThreadedSSAGraphExecutor(
       local_scopes_(local_scopes),
       places_(places),
       graph_(std::move(graph)),
-      pool_(strategy.num_threads_ +
-            1),  // add one more thread for generate op_deps
+      pool_(strategy.num_threads_),
+      prepare_pool_(1),  // add one more thread for generate op_deps
       fetch_ctxs_(places) {
   for (auto &op : ir::FilterByNodeWrapper<OpHandleBase>(*graph_)) {
     int dep = static_cast<int>(op->NotReadyInputSize());
@@ -160,7 +160,7 @@ void FastThreadedSSAGraphExecutor::RunOpAsync(
   });
 }
 void FastThreadedSSAGraphExecutor::PrepareAtomicOpDeps() {
-  atomic_op_deps_ = pool_.enqueue([&] {
+  atomic_op_deps_ = prepare_pool_.enqueue([&] {
     auto *op_deps = new std::unordered_map<OpHandleBase *, std::atomic<int>>;
     for (auto &pair : op_deps_) {
       (*op_deps)[pair.first] = pair.second;

paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h

@@ -46,6 +46,7 @@ class FastThreadedSSAGraphExecutor : public SSAGraphExecutor {
   std::vector<OpHandleBase *> bootstrap_ops_;
 
   ::ThreadPool pool_;
+  ::ThreadPool prepare_pool_;
   platform::DeviceContextPool fetch_ctxs_;
   std::atomic<int> remaining_;
 

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;
 }
 
@@ -359,6 +380,7 @@ std::vector<std::shared_ptr<ExecutorPrepareContext>> Executor::Prepare(
 void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
                                   bool create_local_scope, bool create_vars,
                                   bool keep_kids) {
+  PADDLE_ENFORCE_NOT_NULL(scope);
   Scope* local_scope = scope;
   if (create_vars) {
     if (create_local_scope) {
@@ -396,11 +418,6 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
       DeleteUnusedTensors(*local_scope, op.get(), gc.get(),
                           &(ctx->cur_ref_cnts_));
     }
-
-    if (FLAGS_benchmark) {
-      VLOG(20) << "Memory used after operator " + op->Type() + " running: "
-               << memory::memory_usage(place_);
-    }
   }
 
   if (gc != nullptr) {
@@ -422,13 +439,6 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
       scope->DropKids();
     }
   }
-
-  if (FLAGS_benchmark) {
-    VLOG(20) << "-------------------------------------------------------";
-    VLOG(20) << "Memory used after deleting local scope: "
-             << memory::memory_usage(place_);
-    VLOG(20) << "-------------------------------------------------------";
-  }
 }
 
 void Executor::RunPreparedContext(
@@ -485,6 +495,5 @@ 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/CMakeLists.txt

@@ -5,6 +5,7 @@ file(APPEND ${pass_file} "\#include \"paddle/fluid/framework/ir/pass.h\"\n")
 
 
 # Usage: pass_library(target inference) will append to paddle_inference_pass.h
+unset(INFER_IR_PASSES CACHE) # clear the global variable
 function(pass_library TARGET DEST)
     set(options "")
     set(oneValueArgs "")
@@ -15,10 +16,11 @@ function(pass_library TARGET DEST)
     if (${DEST} STREQUAL "base" OR ${DEST} STREQUAL "inference")
         message(STATUS "add pass ${TARGET} ${DEST}")
         file(APPEND ${pass_file} "USE_PASS(${TARGET});\n")
-        set(PASS_LIBRARY ${TARGET} ${PASS_LIBRARY} PARENT_SCOPE)
+        set(INFER_IR_PASSES ${INFER_IR_PASSES} ${TARGET} CACHE INTERNAL "")
     endif()
 endfunction()
 
+
 cc_library(node SRCS node.cc DEPS proto_desc)
 cc_library(graph SRCS graph.cc DEPS node pretty_log)
 cc_library(graph_helper SRCS graph_helper.cc DEPS graph)
@@ -39,6 +41,7 @@ pass_library(seq_concat_fc_fuse_pass inference)
 pass_library(multi_batch_merge_pass base)
 pass_library(conv_bn_fuse_pass inference)
 pass_library(seqconv_eltadd_relu_fuse_pass inference)
+pass_library(is_test_pass base)
 if(WITH_MKLDNN)
     pass_library(mkldnn_placement_pass base)
     pass_library(depthwise_conv_mkldnn_pass base)
@@ -63,6 +66,7 @@ cc_test(graph_helper_test SRCS graph_helper_test.cc DEPS graph graph_helper op_r
 cc_test(graph_to_program_pass_test SRCS graph_to_program_pass_test.cc DEPS graph_to_program_pass)
 cc_test(test_graph_pattern_detector SRCS graph_pattern_detector_tester.cc DEPS graph_pattern_detector)
 cc_test(test_fc_fuse_pass SRCS fc_fuse_pass_tester.cc DEPS fc_fuse_pass framework_proto)
+cc_test(test_is_test_pass SRCS is_test_pass_tester.cc DEPS is_test_pass)
 if (WITH_MKLDNN)
     cc_test(test_depthwise_conv_mkldnn_pass SRCS depthwise_conv_mkldnn_pass_tester.cc DEPS depthwise_conv_mkldnn_pass)
     cc_test(test_conv_relu_mkldnn_fuse_pass SRCS conv_relu_mkldnn_fuse_pass_tester.cc DEPS conv_relu_mkldnn_fuse_pass)

paddle/fluid/framework/ir/attention_lstm_fuse_pass.cc

@@ -91,10 +91,10 @@ void FindWhileOp(Graph* graph) {
 #undef OP_SET_IN
 #undef OP_SET_OUT
 
-  auto* X = graph->RetriveNode(34);
-  auto* LSTMOUT = graph->RetriveNode(81);
-  auto* cell_init = graph->RetriveNode(6);
-  auto* hidden_init = graph->RetriveNode(8);
+  auto* X = graph->RetrieveNode(34);
+  auto* LSTMOUT = graph->RetrieveNode(81);
+  auto* cell_init = graph->RetrieveNode(6);
+  auto* hidden_init = graph->RetrieveNode(8);
 
   auto* lstm_op = graph->CreateOpNode(&op_desc);
   PrepareParameters(graph, param);
@@ -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/conv_elementwise_add_mkldnn_fuse_pass.cc

@@ -14,14 +14,15 @@
 
 #include "paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.h"
 #include <functional>
-#include <utility>
+#include <list>
+#include <map>
+#include <tuple>
 
 #include "paddle/fluid/framework/ir/graph_traits.h"
 
 namespace paddle {
 namespace framework {
 namespace ir {
-namespace {
 
 // The function keeps the graph consistent by replacing
 // a node 'from' in the set of inputs nodes
@@ -52,97 +53,382 @@ void CorrectGraphEdges(Graph* graph, Node* from, Node* to) {
     }
   }
 }
-}  // namespace
-using graph_ptr = std::unique_ptr<ir::Graph>;
 
-graph_ptr ConvElementwiseAddMKLDNNFusePass::ApplyImpl(graph_ptr graph) const {
-  FusePassBase::Init(name_scope_, graph.get());
+bool IsReachable(ir::Graph* graph, Node* from, Node* to) {
+  auto find_node = [](ir::Graph* graph, const Node* node) -> Node* {
+    for (auto n : graph->Nodes()) {
+      if (n == node) {
+        return n;
+      }
+    }
 
-  GraphPatternDetector gpd;
-  auto pattern = gpd.mutable_pattern();
+    return nullptr;
+  };
 
-  patterns::Conv conv_pattern{pattern, name_scope_};
-  auto conv_output = conv_pattern();
+  if (from == to) {
+    return true;
+  }
+
+  std::map<Node*, bool> visited;
 
-  patterns::ElementwiseAdd elementwise_add_pattern{pattern, name_scope_};
+  for (auto& node : GraphTraits::DFS(*graph)) {
+    visited[&node] = false;
+  }
 
-  elementwise_add_pattern(conv_output);
+  visited[from] = true;
 
-  conv_output->AsIntermediate();
+  std::list<Node*> queue;
+  queue.push_back(from);
+
+  while (!queue.empty()) {
+    auto cur = find_node(graph, queue.front());
+    queue.pop_front();
+
+    if (!cur) return false;
 
-  auto conv_op_has_bias = [](const Node& conv_op) -> std::pair<bool, Node*> {
-    auto bias_input_names = conv_op.Op()->Inputs();
-    auto bias_it = bias_input_names.find("Bias");
-
-    if (bias_it != std::end(bias_input_names)) {
-      bool has_bias = !bias_it->second.empty();
-
-      if (has_bias) {
-        auto conv_bias_names = bias_it->second;
-        auto conv_bias_names_it =
-            std::find_if(std::begin(conv_op.inputs), std::end(conv_op.inputs),
-                         [&conv_bias_names](Node* n) -> bool {
-                           return n->Name() == conv_bias_names[0];
-                         });
-        return std::make_pair(has_bias, *conv_bias_names_it);
+    for (auto n : cur->outputs) {
+      if (n == to) {
+        return true;
       }
+
+      if (!visited[n]) {
+        visited[n] = true;
+        queue.push_back(n);
+      }
+    }
+  }
+  return false;
+}
+
+boost::optional<Node*> HasBias(const Node& op, const std::string& bias_name) {
+  auto bias_input_names = op.Op()->Inputs();
+  auto bias_it = bias_input_names.find(bias_name);
+
+  if (bias_it != std::end(bias_input_names)) {
+    bool has_bias = !bias_it->second.empty();
+
+    if (has_bias) {
+      auto bias_names = bias_it->second;
+      auto bias_names_it =
+          std::find_if(std::begin(op.inputs), std::end(op.inputs),
+                       [&bias_names](Node* n) -> bool {
+                         return n->Name() == bias_names[0];
+                       });
+      return *bias_names_it;
     }
+  }
 
-    return std::make_pair(false, nullptr);
-  };
+  return boost::none;
+}
 
+ResidualConnectionMKLDNNFusePass::IdentityFuseHandle::IdentityFuseHandle(
+    const ResidualConnectionMKLDNNFusePass::CanFuseFunc& can_fuse_func,
+    const ResidualConnectionMKLDNNFusePass::IdentityConvFunc&
+        get_node_from_conv_op,
+    const ResidualConnectionMKLDNNFusePass::IdentityElementwiseAddFunc&
+        get_node_from_elementwise_add_op)
+    : fusion_stats{std::make_shared<int>(0)},
+      can_fuse_func{can_fuse_func},
+      get_node_from_conv_op{get_node_from_conv_op},
+      get_node_from_elementwise_add_op{get_node_from_elementwise_add_op} {}
+
+void ResidualConnectionMKLDNNFusePass::IdentityFuseHandle::operator()(
+    const GraphPatternDetector::subgraph_t& subgraph, Graph* graph) {
+  Node* conv_op;
+  Node* conv_input;
+  Node* conv_filter;
+  Node* conv_output;
+
+  Node* elementwise_add_op;
+  Node* elementwise_add_identity;
+  Node* elementwise_add_out;
+
+  std::tie(conv_op, conv_input, conv_filter, conv_output) =
+      get_node_from_conv_op(subgraph);
+  std::tie(elementwise_add_op, elementwise_add_identity, elementwise_add_out) =
+      get_node_from_elementwise_add_op(subgraph);
+
+  if (!can_fuse_func(conv_op, elementwise_add_op)) return;
+
+  if (!IsReachable(graph, elementwise_add_identity, conv_output)) return;
+
+  OpDesc op_desc;
+  op_desc.SetType("conv2d");
+
+  op_desc.SetInput("Input", {conv_input->Name()});
+  op_desc.SetInput("Filter", {conv_filter->Name()});
+  op_desc.SetInput("ResidualData", {elementwise_add_identity->Name()});
+  op_desc.SetOutput("Output", {conv_output->Name()});
+
+  auto conv_bias = HasBias(*conv_op, "Bias");
+
+  if (conv_bias) {
+    op_desc.SetInput("Bias", {(*conv_bias)->Name()});
+  }
 
-  auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
-                     Graph* g) {
-    GET_IR_NODE_FROM_SUBGRAPH(conv_op, conv_op, conv_pattern);
-    GET_IR_NODE_FROM_SUBGRAPH(conv_input, conv_input, conv_pattern);
-    GET_IR_NODE_FROM_SUBGRAPH(conv_filter, conv_filter, conv_pattern);
-    GET_IR_NODE_FROM_SUBGRAPH(conv_output, conv_output, conv_pattern);
-    GET_IR_NODE_FROM_SUBGRAPH(elementwise_add_op, elementwise_add_op,
-                              elementwise_add_pattern);
-    GET_IR_NODE_FROM_SUBGRAPH(elementwise_add_x, elementwise_add_x,
-                              elementwise_add_pattern);
-    GET_IR_NODE_FROM_SUBGRAPH(elementwise_add_out, elementwise_add_out,
-                              elementwise_add_pattern);
+  for (const auto& attr : conv_op->Op()->GetAttrMap()) {
+    op_desc.SetAttr(attr.first, attr.second);
+  }
 
-    if (FindFuseOption(*conv_op, *elementwise_add_op) != FUSE_MKLDNN) return;
+  op_desc.SetAttr("fuse_residual_connection", true);
 
-    OpDesc op_desc;
-    op_desc.SetType("conv2d");
+  auto fused_conv_op = graph->CreateOpNode(&op_desc);
 
-    op_desc.SetInput("Input", {conv_input->Name()});
-    op_desc.SetInput("Filter", {conv_filter->Name()});
-    op_desc.SetInput("ResidualData", {elementwise_add_x->Name()});
-    op_desc.SetOutput("Output", {conv_output->Name()});
+  IR_NODE_LINK_TO(conv_input, fused_conv_op);
+  IR_NODE_LINK_TO(conv_filter, fused_conv_op);
+  IR_NODE_LINK_TO(elementwise_add_identity, fused_conv_op);
+  IR_NODE_LINK_TO(fused_conv_op, conv_output);
 
-    bool has_bias;
-    Node* conv_bias;
+  if (conv_bias) {
+    IR_NODE_LINK_TO((*conv_bias), fused_conv_op);
+  }
 
-    std::tie(has_bias, conv_bias) = conv_op_has_bias(*conv_op);
+  CorrectGraphEdges(graph, elementwise_add_out, conv_output);
+  GraphSafeRemoveNodes(graph,
+                       {elementwise_add_out, conv_op, elementwise_add_op});
+  (*fusion_stats)++;
+}
 
-    if (has_bias) {
-      op_desc.SetInput("Bias", {conv_bias->Name()});
-    }
+ResidualConnectionMKLDNNFusePass::ProjectionFuseHandle::ProjectionFuseHandle(
+    const ResidualConnectionMKLDNNFusePass::CanFuseFunc& can_fuse_func,
+    const ResidualConnectionMKLDNNFusePass::ProjectionConvFunc&
+        get_node_from_conv_x_op,
+    const ResidualConnectionMKLDNNFusePass::ProjectionConvFunc&
+        get_node_from_conv_y_op,
+    const ResidualConnectionMKLDNNFusePass::ProjectionElementwiseAddFunc&
+        get_node_from_elementwise_add_op)
+    : fusion_stats{std::make_shared<int>(0)},
+      can_fuse_func{can_fuse_func},
+      get_node_from_conv_x_op{get_node_from_conv_x_op},
+      get_node_from_conv_y_op{get_node_from_conv_y_op},
+      get_node_from_elementwise_add_op{get_node_from_elementwise_add_op} {}
+
+void ResidualConnectionMKLDNNFusePass::ProjectionFuseHandle::operator()(
+    const GraphPatternDetector::subgraph_t& subgraph, Graph* graph) {
+  Node* conv_x_op;
+  Node* conv_x_input;
+  Node* conv_x_filter;
+  Node* conv_x_output;
+
+  Node* conv_y_op;
+  Node* conv_y_input;
+  Node* conv_y_filter;
+  Node* conv_y_output;
+
+  Node* elementwise_add_op;
+  Node* elementwise_add_out;
+
+  std::tie(conv_x_op, conv_x_input, conv_x_filter, conv_x_output) =
+      get_node_from_conv_x_op(subgraph);
+  std::tie(conv_y_op, conv_y_input, conv_y_filter, conv_y_output) =
+      get_node_from_conv_y_op(subgraph);
+  std::tie(elementwise_add_op, elementwise_add_out) =
+      get_node_from_elementwise_add_op(subgraph);
+
+  if (!can_fuse_func(conv_x_op, elementwise_add_op)) return;
+  if (!can_fuse_func(conv_y_op, elementwise_add_op)) return;
+
+  Node* projection_node;
+  Node* residual_conv_op;
+  Node* residual_conv_input;
+  Node* residual_conv_filter;
+  Node* residual_conv_output;
+
+  if (IsReachable(graph, conv_x_input, conv_y_output)) {
+    projection_node = conv_x_output;
+    residual_conv_op = conv_y_op;
+    residual_conv_input = conv_y_input;
+    residual_conv_filter = conv_y_filter;
+    residual_conv_output = conv_y_output;
+  } else if (IsReachable(graph, conv_y_input, conv_x_output)) {
+    projection_node = conv_y_output;
+    residual_conv_op = conv_x_op;
+    residual_conv_input = conv_x_input;
+    residual_conv_filter = conv_x_filter;
+    residual_conv_output = conv_x_output;
+  } else {
+    return;
+  }
 
-    for (const auto& attr : conv_op->Op()->GetAttrMap()) {
-      op_desc.SetAttr(attr.first, attr.second);
-    }
+  OpDesc op_desc;
+  op_desc.SetType("conv2d");
 
-    op_desc.SetAttr("fuse_residual_connection", true);
+  op_desc.SetInput("Input", {residual_conv_input->Name()});
+  op_desc.SetInput("Filter", {residual_conv_filter->Name()});
+  op_desc.SetInput("ResidualData", {projection_node->Name()});
+  op_desc.SetOutput("Output", {residual_conv_output->Name()});
 
-    auto fused_conv_op = g->CreateOpNode(&op_desc);
+  auto residual_conv_bias = HasBias(*residual_conv_op, "Bias");
 
-    IR_NODE_LINK_TO(conv_input, fused_conv_op);
-    IR_NODE_LINK_TO(conv_filter, fused_conv_op);
-    IR_NODE_LINK_TO(elementwise_add_x, fused_conv_op);
-    IR_NODE_LINK_TO(fused_conv_op, conv_output);
+  if (residual_conv_bias) {
+    op_desc.SetInput("Bias", {(*residual_conv_bias)->Name()});
+  }
 
-   CorrectGraphEdges(g, elementwise_add_out, conv_output);
-    GraphSafeRemoveNodes(g, {elementwise_add_out, conv_op, elementwise_add_op});
-  };
+  for (const auto& attr : residual_conv_op->Op()->GetAttrMap()) {
+    op_desc.SetAttr(attr.first, attr.second);
+  }
+
+  op_desc.SetAttr("fuse_residual_connection", true);
+
+  auto fused_conv_op = graph->CreateOpNode(&op_desc);
+
+  IR_NODE_LINK_TO(residual_conv_input, fused_conv_op);
+  IR_NODE_LINK_TO(residual_conv_filter, fused_conv_op);
+  IR_NODE_LINK_TO(projection_node, fused_conv_op);
+  IR_NODE_LINK_TO(fused_conv_op, residual_conv_output);
+
+  if (residual_conv_bias) {
+    IR_NODE_LINK_TO((*residual_conv_bias), fused_conv_op);
+  }
+
+  CorrectGraphEdges(graph, elementwise_add_out, residual_conv_output);
+  GraphSafeRemoveNodes(
+      graph, {elementwise_add_out, residual_conv_op, elementwise_add_op});
+  (*fusion_stats)++;
+}
+
+std::tuple<Node*, Node*, Node*, Node*>
+ResidualConnectionMKLDNNFusePass::GetNodesFromConv(
+    const patterns::Conv& conv_pattern,
+    const GraphPatternDetector::subgraph_t& subgraph) const {
+  GET_IR_NODE_FROM_SUBGRAPH(conv_op, conv_op, conv_pattern);
+  GET_IR_NODE_FROM_SUBGRAPH(conv_input, conv_input, conv_pattern);
+  GET_IR_NODE_FROM_SUBGRAPH(conv_filter, conv_filter, conv_pattern);
+  GET_IR_NODE_FROM_SUBGRAPH(conv_output, conv_output, conv_pattern);
+
+  return std::make_tuple(conv_op, conv_input, conv_filter, conv_output);
+}
+
+GraphWithStats ResidualConnectionMKLDNNFusePass::FuseConvAsX(
+    const std::string& name_scope,
+    const GraphWithStats& graph_with_stats) const {
+  ir::Graph* graph;
+  int stats;
+
+  std::tie(graph, stats) = graph_with_stats;
+
+  GraphPatternDetector gpd;
+  auto pattern = gpd.mutable_pattern();
+
+  patterns::Conv conv_pattern{pattern, name_scope};
+  auto conv_output = conv_pattern();
 
-  gpd(graph.get(), handler);
+  patterns::ElementwiseAdd elementwise_add_pattern{pattern, name_scope};
+  elementwise_add_pattern(
+      conv_output,
+      pattern->NewNode(elementwise_add_pattern.elementwise_add_y_repr()));
+  conv_output->AsIntermediate();
+
+  auto get_node_from_elementwise_add = [&elementwise_add_pattern](
+      const GraphPatternDetector::subgraph_t& subgraph)
+      -> std::tuple<Node*, Node*, Node*> {
+        GET_IR_NODE_FROM_SUBGRAPH(elementwise_add_op, elementwise_add_op,
+                                  elementwise_add_pattern);
+        GET_IR_NODE_FROM_SUBGRAPH(elementwise_add_y, elementwise_add_y,
+                                  elementwise_add_pattern);
+        GET_IR_NODE_FROM_SUBGRAPH(elementwise_add_out, elementwise_add_out,
+                                  elementwise_add_pattern);
+
+        return std::make_tuple(elementwise_add_op, elementwise_add_y,
+                               elementwise_add_out);
+      };
+
+  return ExecuteHandleOnGraph<IdentityFuseHandle>(
+      &gpd, graph_with_stats,
+      [this, &conv_pattern](const GraphPatternDetector::subgraph_t& subgraph) {
+        return GetNodesFromConv(conv_pattern, subgraph);
+      },
+      get_node_from_elementwise_add);
+}
+
+GraphWithStats ResidualConnectionMKLDNNFusePass::FuseConvAsY(
+    const std::string& name_scope,
+    const GraphWithStats& graph_with_stats) const {
+  GraphPatternDetector gpd;
+  auto pattern = gpd.mutable_pattern();
+
+  patterns::Conv conv_pattern{pattern, name_scope};
+  auto conv_output = conv_pattern();
+
+  patterns::ElementwiseAdd elementwise_add_pattern{pattern, name_scope};
+  elementwise_add_pattern(
+      pattern->NewNode(elementwise_add_pattern.elementwise_add_x_repr()),
+      conv_output);
+  conv_output->AsIntermediate();
+
+  auto get_node_from_elementwise_add = [&elementwise_add_pattern](
+      const GraphPatternDetector::subgraph_t& subgraph)
+      -> std::tuple<Node*, Node*, Node*> {
+        GET_IR_NODE_FROM_SUBGRAPH(elementwise_add_op, elementwise_add_op,
+                                  elementwise_add_pattern);
+        GET_IR_NODE_FROM_SUBGRAPH(elementwise_add_x, elementwise_add_x,
+                                  elementwise_add_pattern);
+        GET_IR_NODE_FROM_SUBGRAPH(elementwise_add_out, elementwise_add_out,
+                                  elementwise_add_pattern);
+
+        return std::make_tuple(elementwise_add_op, elementwise_add_x,
+                               elementwise_add_out);
+      };
+
+  return ExecuteHandleOnGraph<IdentityFuseHandle>(
+      &gpd, graph_with_stats,
+      [this, &conv_pattern](const GraphPatternDetector::subgraph_t& subgraph) {
+        return GetNodesFromConv(conv_pattern, subgraph);
+      },
+      get_node_from_elementwise_add);
+}
 
+GraphWithStats ResidualConnectionMKLDNNFusePass::FuseProjectionConv(
+    const std::string& name_scope,
+    const GraphWithStats& graph_with_stats) const {
+  GraphPatternDetector gpd;
+  auto pattern = gpd.mutable_pattern();
+
+  patterns::Conv conv_x_pattern{pattern, name_scope};
+  auto conv_x_output = conv_x_pattern();
+
+  patterns::Conv conv_y_pattern{pattern, name_scope};
+  auto conv_y_output = conv_y_pattern();
+
+  patterns::ElementwiseAdd elementwise_add_pattern{pattern, name_scope};
+  elementwise_add_pattern(conv_x_output, conv_y_output);
+  conv_x_output->AsIntermediate();
+  conv_y_output->AsIntermediate();
+
+  auto get_node_from_elementwise_add = [&elementwise_add_pattern](
+      const GraphPatternDetector::subgraph_t& subgraph)
+      -> std::tuple<Node*, Node*> {
+        GET_IR_NODE_FROM_SUBGRAPH(elementwise_add_op, elementwise_add_op,
+                                  elementwise_add_pattern);
+        GET_IR_NODE_FROM_SUBGRAPH(elementwise_add_out, elementwise_add_out,
+                                  elementwise_add_pattern);
+
+        return std::make_tuple(elementwise_add_op, elementwise_add_out);
+      };
+
+  return ExecuteHandleOnGraph<ProjectionFuseHandle>(
+      &gpd, graph_with_stats,
+      [this,
+       &conv_x_pattern](const GraphPatternDetector::subgraph_t& subgraph) {
+        return GetNodesFromConv(conv_x_pattern, subgraph);
+      },
+      [this,
+       &conv_y_pattern](const GraphPatternDetector::subgraph_t& subgraph) {
+        return GetNodesFromConv(conv_y_pattern, subgraph);
+      },
+      get_node_from_elementwise_add);
+}
+
+graph_ptr ResidualConnectionMKLDNNFusePass::ApplyImpl(graph_ptr graph) const {
+  FusePassBase::Init(name_scope_, graph.get());
+  auto fused_graph_with_stats = FuseConvAsY(
+      name_scope_,
+      FuseConvAsX(
+          name_scope_,
+          FuseProjectionConv(name_scope_, std::make_pair(graph.get(), 0))));
+
+  std::cout << "Fused graph " << fused_graph_with_stats.second << std::endl;
+  AddStatis(fused_graph_with_stats.second);
   return graph;
 }
 }  // namespace ir
@@ -150,4 +436,4 @@ graph_ptr ConvElementwiseAddMKLDNNFusePass::ApplyImpl(graph_ptr graph) const {
 }  // namespace paddle
 
 REGISTER_PASS(conv_elementwise_add_mkldnn_fuse_pass,
-              paddle::framework::ir::ConvElementwiseAddMKLDNNFusePass);
+              paddle::framework::ir::ResidualConnectionMKLDNNFusePass);

paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.h

@@ -15,24 +15,119 @@
 #pragma once
 
 #include <string>
+#include <tuple>
+#include <utility>
 #include "paddle/fluid/framework/ir/fuse_pass_base.h"
 #include "paddle/fluid/framework/ir/graph.h"
 #include "paddle/fluid/framework/ir/graph_pattern_detector.h"
 
+#include <boost/optional.hpp>
+
 namespace paddle {
 namespace framework {
 namespace ir {
 
-class ConvElementwiseAddMKLDNNFusePass : public FusePassBase {
+using graph_ptr = std::unique_ptr<ir::Graph>;
+using GraphWithStats = std::pair<ir::Graph*, int>;
+
+void CorrectGraphEdges(Graph* graph, Node* from, Node* to);
+bool IsReachable(ir::Graph* graph, Node* from, Node* to);
+boost::optional<Node*> HasBias(const Node& op, const std::string& bias_name);
+
+class ResidualConnectionMKLDNNFusePass : public FusePassBase {
+ private:
+  GraphWithStats FuseConvAsX(const std::string& name_scope,
+                             const GraphWithStats& graph_with_stats) const;
+  GraphWithStats FuseConvAsY(const std::string& name_scope,
+                             const GraphWithStats& graph_with_stats) const;
+  GraphWithStats FuseProjectionConv(
+      const std::string& name_scope,
+      const GraphWithStats& graph_with_stats) const;
+
+  template <typename RetType>
+  using GetNodeFunc =
+      std::function<RetType(const GraphPatternDetector::subgraph_t& subgraph)>;
+  using IdentityConvFunc = GetNodeFunc<std::tuple<Node*, Node*, Node*, Node*>>;
+  using IdentityElementwiseAddFunc =
+      GetNodeFunc<std::tuple<Node*, Node*, Node*>>;
+
+  using ProjectionConvFunc = IdentityConvFunc;
+  using ProjectionElementwiseAddFunc = GetNodeFunc<std::tuple<Node*, Node*>>;
+
+  using CanFuseFunc = std::function<bool(Node*, Node*)>;
+
+  std::tuple<Node*, Node*, Node*, Node*> GetNodesFromConv(
+      const patterns::Conv& conv_pattern,
+      const GraphPatternDetector::subgraph_t& subgraph) const;
+
+  std::tuple<Node*, Node*, Node*, Node*> GetNodesFromProjectionConv(
+      const patterns::Conv& conv_pattern,
+      const GraphPatternDetector::subgraph_t& subgraph) const;
+
+  template <typename HandleType, typename... OpFuncs>
+  GraphWithStats ExecuteHandleOnGraph(GraphPatternDetector* gpd,
+                                      const GraphWithStats& graph_with_stats,
+                                      OpFuncs&&... op_funcs) const {
+    ir::Graph* graph;
+    int stats;
+
+    std::tie(graph, stats) = graph_with_stats;
+
+    auto can_fuse = [this](Node* op1, Node* op2) -> bool {
+      return this->FindFuseOption(*op1, *op2) == FUSE_MKLDNN;
+    };
+
+    auto fuse_handle = HandleType{can_fuse, std::forward<OpFuncs>(op_funcs)...};
+
+    (*gpd)(graph, fuse_handle);
+
+    return std::make_pair(graph, stats + fuse_handle.get_stats());
+  }
+
+  struct IdentityFuseHandle {
+    IdentityFuseHandle(
+        const CanFuseFunc& can_fuse_func,
+        const IdentityConvFunc& get_node_from_conv_op,
+        const IdentityElementwiseAddFunc& get_node_from_elementwise_add_op);
+
+    void operator()(const GraphPatternDetector::subgraph_t& subgraph,
+                    Graph* graph);
+    int get_stats() const { return *fusion_stats; }
+
+   private:
+    std::shared_ptr<int> fusion_stats;
+    CanFuseFunc can_fuse_func;
+    IdentityConvFunc get_node_from_conv_op;
+    IdentityElementwiseAddFunc get_node_from_elementwise_add_op;
+  };
+
+  struct ProjectionFuseHandle {
+    ProjectionFuseHandle(
+        const CanFuseFunc& can_fuse_func,
+        const ProjectionConvFunc& get_node_from_conv_x_op,
+        const ProjectionConvFunc& get_node_from_conv_y_op,
+        const ProjectionElementwiseAddFunc& get_node_from_elementwise_add_op);
+
+    void operator()(const GraphPatternDetector::subgraph_t& subgraph,
+                    Graph* graph);
+    int get_stats() const { return *fusion_stats; }
+
+   private:
+    std::shared_ptr<int> fusion_stats;
+    CanFuseFunc can_fuse_func;
+    ProjectionConvFunc get_node_from_conv_x_op;
+    ProjectionConvFunc get_node_from_conv_y_op;
+    ProjectionElementwiseAddFunc get_node_from_elementwise_add_op;
+  };
+
  public:
-  virtual ~ConvElementwiseAddMKLDNNFusePass() {}
+  virtual ~ResidualConnectionMKLDNNFusePass() {}
 
  protected:
-  std::unique_ptr<ir::Graph> ApplyImpl(std::unique_ptr<ir::Graph> graph) const;
+  std::unique_ptr<ir::Graph> ApplyImpl(graph_ptr graph) const;
 
-  const std::string name_scope_{"residual_connections_fuse_pass"};
+  const std::string name_scope_{"residual_connection_fuse_pass"};
 };
-
 }  // namespace ir
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass_tester.cc

@@ -41,7 +41,7 @@ void SetOp(ProgramDesc* prog, const std::string& type,
 
 }
 
-struct IsReachable {
+struct TestIsReachable {
   using func = std::function<bool(const std::string&, const std::string&)>;
 
   auto operator()(const std::unique_ptr<ir::Graph>& graph) -> func {
@@ -90,7 +90,9 @@ struct IsReachable {
   }
 };
 
-void AssertOpsCount(const std::unique_ptr<ir::Graph>& graph) {
+void AssertOpsCount(const std::unique_ptr<ir::Graph>& graph,
+                    int expected_conv_count,
+                    int expected_elementwise_add_count = 0) {
   int conv_count = 0;
   int elementwise_add_count = 0;
 
@@ -102,8 +104,8 @@ void AssertOpsCount(const std::unique_ptr<ir::Graph>& graph) {
       ++elementwise_add_count;
     }
   }
-  EXPECT_EQ(conv_count, 1);
-  EXPECT_EQ(elementwise_add_count, 0);
+  EXPECT_EQ(conv_count, expected_conv_count);
+  EXPECT_EQ(elementwise_add_count, expected_elementwise_add_count);
 }
 
 ProgramDesc BuildProgramDesc(const std::vector<std::string>& transient_vars,
@@ -128,23 +130,13 @@ ProgramDesc BuildProgramDesc(const std::vector<std::string>& transient_vars,
 
   return prog;
 }
-}  // namespace
-
-TEST(ConvElementwiseAddMKLDNNFusePass, ConvolutionWithElementwiseAddRelu) {
-  auto prog =
-      BuildProgramDesc({"a", "b", "c", "d", "e", "f"}, {"bias", "weights"});
-
-  SetOp(&prog, "conv2d",
-        {{"Input", "a"}, {"Bias", "bias"}, {"Filter", "weights"}},
-        {"Output", "b"});
-  SetOp(&prog, "elementwise_add", {{"X", "b"}, {"Y", "c"}}, {"Out", "d"});
-  SetOp(&prog, "relu", {{"X", "d"}}, {"Out", "e"});
-
-  std::unique_ptr<ir::Graph> graph(new ir::Graph(prog));
 
-  IsReachable is_reachable;
+void RunPassAndAssert(ProgramDesc* prog, const std::string& from,
+                      const std::string& to, int expected_conv_num) {
+  std::unique_ptr<ir::Graph> graph(new ir::Graph(*prog));
 
-  EXPECT_TRUE(is_reachable(graph)("a", "relu"));
+  TestIsReachable is_reachable;
+  EXPECT_TRUE(is_reachable(graph)(from, to));
 
   auto pass =
       PassRegistry::Instance().Get("conv_elementwise_add_mkldnn_fuse_pass");
@@ -152,82 +144,87 @@ TEST(ConvElementwiseAddMKLDNNFusePass, ConvolutionWithElementwiseAddRelu) {
   graph = pass->Apply(std::move(graph));
   int current_nodes_num = graph->Nodes().size();
 
-  EXPECT_TRUE(is_reachable(graph)("a", "relu"));
+  EXPECT_TRUE(is_reachable(graph)(from, to));
 
   EXPECT_EQ(original_nodes_num - nodes_removed + nodes_added,
             current_nodes_num);
 
-  AssertOpsCount(graph);
+  AssertOpsCount(graph, expected_conv_num);
 }
+}  // namespace
 
-TEST(ConvElementwiseAddMKLDNNFusePass,
-     ConvolutionWithElementwiseAddReluNoBias) {
-  auto prog = BuildProgramDesc({"a", "b", "c", "d", "e"}, {"weights"});
-  SetOp(&prog, "conv2d", {{"Input", "a"}, {"Filter", "weights"}},
-        {"Output", "b"});
-  SetOp(&prog, "elementwise_add", {{"X", "b"}, {"Y", "c"}}, {"Out", "d"});
-  SetOp(&prog, "relu", {{"X", "d"}}, {"Out", "e"});
-
-  std::unique_ptr<ir::Graph> graph(new ir::Graph(prog));
+TEST(ConvElementwiseAddMKLDNNFusePass, ConvolutionAsYWithElementwiseAddRelu) {
+  auto prog = BuildProgramDesc({"a", "b", "c", "d", "e"}, {"bias", "weights"});
 
-  IsReachable is_reachable;
+  SetOp(&prog, "sigmoid", {{"X", "a"}}, {"Out", "b"});
+  SetOp(&prog, "conv2d",
+        {{"Input", "b"}, {"Bias", "bias"}, {"Filter", "weights"}},
+        {"Output", "c"});
 
-  EXPECT_TRUE(is_reachable(graph)("a", "relu"));
+  SetOp(&prog, "elementwise_add", {{"X", "a"}, {"Y", "c"}}, {"Out", "d"});
+  SetOp(&prog, "relu", {{"X", "d"}}, {"Out", "e"});
 
-  auto pass =
-      PassRegistry::Instance().Get("conv_elementwise_add_mkldnn_fuse_pass");
-  int original_nodes_num = graph->Nodes().size();
-  graph = pass->Apply(std::move(graph));
-  int current_nodes_num = graph->Nodes().size();
+  RunPassAndAssert(&prog, "a", "relu", 1);
+}
 
-  EXPECT_TRUE(is_reachable(graph)("a", "relu"));
+TEST(ConvElementwiseAddMKLDNNFusePass,
+     ConvolutionAsYWithElementwiseAddReluNoBias) {
+  auto prog = BuildProgramDesc({"a", "b", "c", "d", "e"}, {"weights"});
 
-  EXPECT_EQ(original_nodes_num - nodes_removed + nodes_added,
-            current_nodes_num);
+  SetOp(&prog, "sigmoid", {{"X", "a"}}, {"Out", "b"});
+  SetOp(&prog, "conv2d", {{"Input", "b"}, {"Filter", "weights"}},
+        {"Output", "c"});
+  SetOp(&prog, "elementwise_add", {{"X", "a"}, {"Y", "c"}}, {"Out", "d"});
+  SetOp(&prog, "relu", {{"X", "d"}}, {"Out", "e"});
 
-  AssertOpsCount(graph);
+  RunPassAndAssert(&prog, "a", "relu", 1);
 }
 
-TEST(ConvElementwiseAddMKLDNNFusePass, ConvolutionElementwiseAdd) {
-  auto prog = BuildProgramDesc({"a", "b", "c", "d"}, {"bias", "weights"});
+TEST(ConvElementwiseAddMKLDNNFusePass, ConvolutionAsXWithElementwiseAddRelu) {
+  auto prog = BuildProgramDesc({"a", "b", "c", "d", "e"}, {"bias", "weights"});
+
+  SetOp(&prog, "sigmoid", {{"X", "a"}}, {"Out", "b"});
   SetOp(&prog, "conv2d",
-        {{"Input", "a"}, {"Bias", "bias"}, {"Filter", "weights"}},
-        {"Output", "b"});
-  SetOp(&prog, "elementwise_add", {{"X", "b"}, {"Y", "c"}}, {"Out", "d"});
+        {{"Input", "b"}, {"Bias", "bias"}, {"Filter", "weights"}},
+        {"Output", "c"});
 
-  std::unique_ptr<ir::Graph> graph(new ir::Graph(prog));
+  SetOp(&prog, "elementwise_add", {{"X", "c"}, {"Y", "a"}}, {"Out", "d"});
+  SetOp(&prog, "relu", {{"X", "d"}}, {"Out", "e"});
 
-  IsReachable is_reachable;
-  EXPECT_TRUE(is_reachable(graph)("a", "d"));
+  RunPassAndAssert(&prog, "a", "relu", 1);
+}
 
-  auto pass =
-      PassRegistry::Instance().Get("conv_elementwise_add_mkldnn_fuse_pass");
-  int original_nodes_num = graph->Nodes().size();
-  graph = pass->Apply(std::move(graph));
-  int current_nodes_num = graph->Nodes().size();
+TEST(ConvElementwiseAddMKLDNNFusePass,
+     ConvolutionAsXWithElementwiseAddReluNoBias) {
+  auto prog = BuildProgramDesc({"a", "b", "c", "d", "e"}, {"weights"});
 
-  EXPECT_FALSE(is_reachable(graph)("a", "d"));
+  SetOp(&prog, "sigmoid", {{"X", "a"}}, {"Out", "b"});
+  SetOp(&prog, "conv2d", {{"Input", "b"}, {"Filter", "weights"}},
+        {"Output", "c"});
+  SetOp(&prog, "elementwise_add", {{"X", "c"}, {"Y", "a"}}, {"Out", "d"});
+  SetOp(&prog, "relu", {{"X", "d"}}, {"Out", "e"});
 
-  EXPECT_EQ(original_nodes_num - nodes_removed + nodes_added,
-            current_nodes_num);
-  AssertOpsCount(graph);
+  RunPassAndAssert(&prog, "a", "relu", 1);
 }
 
-TEST(ConvElementwiseAddMKLDNNFusePass, SigmoidConvolutionAddElementwiseRelu) {
+TEST(ConvElementwiseAddMKLDNNFusePass, NoFusion) {
   auto prog =
-      BuildProgramDesc({"a", "b", "c", "d", "e", "f"}, {"bias", "weights"});
+      BuildProgramDesc({"a", "b", "c", "d", "e", "f", "g"}, {"weights"});
+
   SetOp(&prog, "sigmoid", {{"X", "a"}}, {"Out", "b"});
-  SetOp(&prog, "conv2d",
-        {{"Input", "b"}, {"Bias", "bias"}, {"Filter", "weights"}},
+  SetOp(&prog, "conv2d", {{"Input", "b"}, {"Filter", "weights"}},
         {"Output", "c"});
-  SetOp(&prog, "elementwise_add", {{"X", "c"}, {"Y", "d"}}, {"Out", "e"});
-  SetOp(&prog, "relu", {{"X", "e"}}, {"Out", "f"});
 
-  std::unique_ptr<ir::Graph> graph(new ir::Graph(prog));
+  SetOp(&prog, "conv2d", {{"Input", "d"}, {"Filter", "weights"}},
+        {"Output", "e"});
 
-  IsReachable is_reachable;
+  SetOp(&prog, "elementwise_add", {{"X", "c"}, {"Y", "e"}}, {"Out", "f"});
+  SetOp(&prog, "relu", {{"X", "f"}}, {"Out", "g"});
 
-  EXPECT_TRUE(is_reachable(graph)("a", "f"));
+  std::unique_ptr<ir::Graph> graph(new ir::Graph(prog));
+
+  TestIsReachable is_reachable;
+  EXPECT_TRUE(is_reachable(graph)("a", "g"));
 
   auto pass =
       PassRegistry::Instance().Get("conv_elementwise_add_mkldnn_fuse_pass");
@@ -235,11 +232,10 @@ TEST(ConvElementwiseAddMKLDNNFusePass, SigmoidConvolutionAddElementwiseRelu) {
   graph = pass->Apply(std::move(graph));
   int current_nodes_num = graph->Nodes().size();
 
-  EXPECT_TRUE(is_reachable(graph)("a", "f"));
+  EXPECT_TRUE(is_reachable(graph)("a", "g"));
+  EXPECT_EQ(original_nodes_num, current_nodes_num);
 
-  EXPECT_EQ(original_nodes_num - nodes_removed + nodes_added,
-            current_nodes_num);
-  AssertOpsCount(graph);
+  AssertOpsCount(graph, 2, 1);
 }
 
 }  // namespace ir

paddle/fluid/framework/ir/fc_fuse_pass.cc

@@ -57,6 +57,7 @@ std::unique_ptr<ir::Graph> FCFusePass::ApplyImpl(
     desc.SetInput("W", std::vector<std::string>({fc_Y_in}));
     desc.SetInput("Bias", std::vector<std::string>({fc_bias_in}));
     desc.SetOutput("Out", std::vector<std::string>({fc_out_out}));
+    desc.SetAttr("in_num_col_dims", mul->Op()->GetAttr("x_num_col_dims"));
     desc.SetType("fc");
     auto fc_node = g->CreateOpNode(&desc);  // OpDesc will be copied.
     GraphSafeRemoveNodes(graph.get(), {mul, elementwise_add, mul_out});

paddle/fluid/framework/ir/fc_fuse_pass_tester.cc

@@ -29,6 +29,7 @@ void SetOp(ProgramDesc* prog, const std::string& type,
   if (type == "mul") {
     op->SetInput("X", {inputs[0]});
     op->SetInput("Y", {inputs[1]});
+    op->SetAttr("x_num_col_dims", {1});
   } else if (type == "elementwise_add") {
     op->SetInput("X", inputs);
   }

paddle/fluid/framework/ir/graph.cc

@@ -84,8 +84,6 @@ void CheckProgram(const ProgramDesc &program) {
 
 Graph::Graph(const ProgramDesc &program) : program_(program) {
   CheckProgram(program_);
-  // Make the nodes id start from 0.
-  Node::ResetId();
   auto var_nodes = InitFromProgram(program_);
   ResolveHazard(var_nodes);
 }

paddle/fluid/framework/ir/graph.h

@@ -116,13 +116,17 @@ class Graph {
   // Create a normal variable with non-null VarDesc.
   ir::Node *CreateVarNode(VarDesc *var_desc) {
     PADDLE_ENFORCE(var_desc);
-    return AddNode(new ir::Node(var_desc));
+    auto *x = AddNode(new ir::Node(var_desc));
+    x->SetId(num_node_created_++);
+    return x;
   }
 
   // Create a normal runnable operator with OpDesc.
   ir::Node *CreateOpNode(OpDesc *op_desc) {
     PADDLE_ENFORCE(op_desc);
-    return AddNode(new ir::Node(op_desc));
+    auto *x = AddNode(new ir::Node(op_desc));
+    x->SetId(num_node_created_++);
+    return x;
   }
 
   // Create a control dependency var that connects 2 operations. The
@@ -132,13 +136,17 @@ class Graph {
     // TODO(panyx0718): control var name should be really unique.
     const std::string name = string::Sprintf(
         "%s@%llu", ir::Node::kControlDepVarName, node_set_.size());
-    return AddNode(new ir::Node(name, ir::Node::Type::kVariable));
+    auto *x = AddNode(new ir::Node(name, ir::Node::Type::kVariable));
+    x->SetId(num_node_created_++);
+    return x;
   }
 
   // A more free style way of creating a graph node. Mostly use for test
   // or "copy" from another node. Avoid using it if possible.
   ir::Node *CreateEmptyNode(const std::string &name, ir::Node::Type type) {
-    return AddNode(new ir::Node(name, type));
+    auto *x = AddNode(new ir::Node(name, type));
+    x->SetId(num_node_created_++);
+    return x;
   }
 
   // Clear all node information of the graph and return the ownership of the
@@ -160,7 +168,7 @@ class Graph {
   }
 
   // NOTE low performance, but simple and secure.
-  Node *RetriveNode(int id) {
+  Node *RetrieveNode(int id) {
     for (auto &node : nodes_) {
       if (node.second->id() == id) {
         return node.second.get();
@@ -169,6 +177,7 @@ class Graph {
     return nullptr;
   }
 
+  const ProgramDesc &program() const { return program_; }
   std::map<std::string, std::vector<ir::Node *>> InitFromProgram(
       const ProgramDesc &program);
 
@@ -190,6 +199,7 @@ class Graph {
   std::map<std::string, std::function<void(void)>> attr_dels_;
   std::map<ir::Node *, std::unique_ptr<ir::Node>> nodes_;
   std::unordered_set<ir::Node *> node_set_;
+  size_t num_node_created_{0};  // help to generate a unique node id.
 };
 
 bool IsControlDepVar(const ir::Node &var);

paddle/fluid/framework/ir/graph_helper.cc

@@ -15,8 +15,15 @@ limitations under the License. */
 #include "paddle/fluid/framework/ir/graph_helper.h"
 #include <algorithm>
 #include <deque>
+#include <fstream>
+#include <iosfwd>
+#include <ostream>
 #include <unordered_set>
 
+DEFINE_string(print_sub_graph_dir, "",
+              "FLAGS_print_sub_graph_dir is used "
+              "to print the nodes of sub_graphs.");
+
 namespace paddle {
 namespace framework {
 namespace ir {
@@ -164,12 +171,15 @@ size_t GraphNum(const Graph &graph) {
     graph_nodes.emplace_back(g_nodes);
   }
 
-  if (VLOG_IS_ON(100)) {
-    VLOG(100) << "graph_num: " << graph_nodes.size();
-    for (auto &g_n : graph_nodes) {
-      VLOG(100) << "graph_nodes: " << g_n.size();
-      if (g_n.size() < 10) {
-        std::stringstream out;
+  if (FLAGS_print_sub_graph_dir.size()) {
+    if (graph_nodes.size() > 1) {
+      std::stringstream out;
+      for (auto &g_n : graph_nodes) {
+        out << "graph_nodes: " << g_n.size() << "\n";
+      }
+      out << "\n\n";
+      for (auto &g_n : graph_nodes) {
+        out << "graph_nodes: " << g_n.size();
         for (auto &node : g_n) {
           out << "\nNode: " << node->Name() << " in [";
           for (auto &n : node->inputs) {
@@ -181,8 +191,12 @@ size_t GraphNum(const Graph &graph) {
           }
           out << "]";
         }
-        VLOG(100) << out.str();
+        out << "\n\n\n";
       }
+      std::unique_ptr<std::ostream> fout(
+          new std::ofstream(FLAGS_print_sub_graph_dir));
+      PADDLE_ENFORCE(fout->good());
+      *fout << out.str();
     }
   }
 

paddle/fluid/framework/ir/graph_pattern_detector.cc

@@ -167,10 +167,12 @@ struct HitGroup {
 
   bool Match(Node *node, PDNode *pat) {
     if (nodes_.count(node)) {
-      if (!roles.count(pat)) return false;
-      return roles[pat] == node;
+      if (roles.count(pat) && roles[pat] == node) return true;
+      return false;
+    } else {
+      if (roles.count(pat) && roles[pat] != node) return false;
+      return true;
     }
-    return !roles.count(pat) || roles.at(pat) == node;
   }
 
   void Register(Node *node, PDNode *pat) {
@@ -198,7 +200,6 @@ GraphPatternDetector::DetectPatterns() {
   std::vector<GraphPatternDetector::subgraph_t> result;
   std::vector<HitGroup> init_groups;
   std::array<std::vector<HitGroup>, 2> bi_records;
-  // PADDLE_ENFORCE(!pattern_.edges().empty(), "At least one edge is needed");
   auto *first_pnode = pattern_.edges().empty() ? pattern().nodes().front().get()
                                                : pattern_.edges().front().first;
   if (!pdnodes2nodes_.count(first_pnode)) return result;
@@ -228,11 +229,12 @@ GraphPatternDetector::DetectPatterns() {
         VLOG(80) << "check " << source->id() << " -- " << target->id();
         // TODO(Superjomn) add some prune strategies.
         for (const auto &group : pre_groups) {
-          HitGroup new_group = group;
-          if (IsNodesLink(source, target) &&
-              new_group.Match(source, edge.first)) {
-            new_group.Register(source, edge.first);
-            if (new_group.Match(target, edge.second)) {
+          if (IsNodesLink(source, target)) {
+            HitGroup new_group = group;
+            bool flag = new_group.Match(source, edge.first) &&
+                        new_group.Match(target, edge.second);
+            if (flag) {
+              new_group.Register(source, edge.first);
               new_group.Register(target, edge.second);
               cur_groups.push_back(new_group);
               // TODO(Superjomn) need to unique
@@ -261,14 +263,16 @@ GraphPatternDetector::DetectPatterns() {
   return result;
 }
 
-bool GraphItemCMP(const std::pair<PDNode *, Node *> &a,
+struct GraphItemLessThan {
+  bool operator()(const std::pair<PDNode *, Node *> &a,
                   const std::pair<PDNode *, Node *> &b) {
-  if (a.first != b.first) {
-    return a.first < b.first;
-  } else {
-    return a.second < b.second;
+    if (a.first != b.first) {
+      return a.first < b.first;
+    } else {
+      return a.second < b.second;
+    }
   }
-}
+};
 
 // TODO(Superjomn) enhance the function as it marks unique unique as duplicates
 // see https://github.com/PaddlePaddle/Paddle/issues/13550
@@ -282,7 +286,7 @@ void GraphPatternDetector::UniquePatterns(
   for (auto &g : *subgraphs) {
     // Sort the items in the sub-graph, and transform to a string key.
     std::vector<std::pair<PDNode *, Node *>> sorted_keys(g.begin(), g.end());
-    std::sort(sorted_keys.begin(), sorted_keys.end(), GraphItemCMP);
+    std::sort(sorted_keys.begin(), sorted_keys.end(), GraphItemLessThan());
     std::stringstream ss;
     for (auto &item : sorted_keys) {
       ss << item.first << ":" << item.second;
@@ -1080,16 +1084,12 @@ PDNode *patterns::Conv::operator()() {
   return output_var;
 }
 
-PDNode *patterns::ElementwiseAdd::operator()(PDNode *x_var) {
+PDNode *patterns::ElementwiseAdd::operator()(PDNode *x_var, PDNode *y_var) {
   auto elementwise_add_op = pattern->NewNode(elementwise_add_op_repr())
                                 ->assert_is_op("elementwise_add");
 
-  x_var->assert_is_op_input("elementwise_add", "X");
-
-  auto y_var = pattern->NewNode(elementwise_add_x_repr())
-                   ->AsInput()
-                   ->assert_is_op_input("elementwise_add", "Y");
-
+  x_var->AsInput()->assert_is_op_input("elementwise_add", "X");
+  y_var->AsInput()->assert_is_op_input("elementwise_add", "Y");
   auto out_var = pattern->NewNode(elementwise_add_out_repr())
                      ->AsOutput()
                      ->assert_is_op_output("elementwise_add", "Out");

paddle/fluid/framework/ir/graph_pattern_detector.h

@@ -310,8 +310,8 @@ void GraphSafeRemoveNodes(Graph* graph,
                           const std::unordered_set<const Node*>& nodes);
 
 // Some pre-defined patterns those can be reused in multiple passes.
-// The related Fluid Layer or Op should be one pattern here for better reusage
-// accross different fusion.
+// The related Fluid Layer or Op should be one pattern here for better re-usage
+// across different fusion.
 namespace patterns {
 
 struct KeyCounter {
@@ -664,7 +664,7 @@ struct ElementwiseAdd : public PatternBase {
   ElementwiseAdd(PDPattern* pattern, const std::string& name_scope)
       : PatternBase(pattern, name_scope, "elementwise_add") {}
 
-  PDNode* operator()(PDNode* x_var);
+  PDNode* operator()(PDNode* x_var, PDNode* y_var);
 
   PATTERN_DECL_NODE(elementwise_add_op);
   PATTERN_DECL_NODE(elementwise_add_x);

paddle/fluid/framework/ir/graph_to_program_pass.cc

@@ -35,10 +35,11 @@ std::unique_ptr<Graph> GraphToProgramPass::ApplyImpl(
       new proto::ProgramDesc(*program.Proto()));
 
   auto block = program_pb->mutable_blocks(kRootBlockIndex);
+  block->set_idx(kRootBlockIndex);
   block->clear_vars();
   std::unordered_set<std::string> visited_vars;
   for (ir::Node* n : graph->Nodes()) {
-    if (n->NodeType() == ir::Node::Type::kVariable) {
+    if (n->IsVar()) {
       if (n->Var() && visited_vars.count(n->Var()->Name()) == 0) {
         visited_vars.insert(n->Var()->Name());
         block->add_vars()->MergeFrom(*n->Var()->Proto());

paddle/fluid/framework/ir/graph_traits.cc

@@ -66,6 +66,76 @@ NodesDFSIterator &NodesDFSIterator::operator=(const NodesDFSIterator &other) {
 }
 Node *NodesDFSIterator::operator->() { return stack_.top(); }
 
+inline bool CheckNodeIndegreeEquals(const Node &node, size_t n) {
+  return node.inputs.size() == n;
+}
+
+NodesTSIterator::NodesTSIterator(const std::vector<Node *> &source) {
+  PADDLE_ENFORCE(!source.empty(),
+                 "Start points of topological sorting should not be empty!");
+  // CHECK all the inputs' in-degree is 0
+  for (auto *node : source) {
+    PADDLE_ENFORCE(CheckNodeIndegreeEquals(*node, 0));
+  }
+
+  std::unordered_set<Node *> visited;
+  std::unordered_set<Node *> to_visit{source.begin(), source.end()};
+
+  std::vector<Node *> inlink_visited;
+  while (!to_visit.empty()) {
+    std::vector<Node *> queue(to_visit.begin(), to_visit.end());
+    for (auto *p : queue) {
+      inlink_visited.clear();
+
+      std::copy_if(p->inputs.begin(), p->inputs.end(),
+                   std::back_inserter(inlink_visited),
+                   [&](Node *x) -> bool { return visited.count(x) != 0; });
+
+      if (inlink_visited.size() == p->inputs.size()) {
+        sorted_.push_back(p);
+        for (auto *_ : p->outputs) {
+          if (!visited.count(_)) {
+            to_visit.insert(_);
+          }
+        }
+
+        to_visit.erase(p);
+        visited.insert(p);
+      }
+    }
+  }
+}
+
+NodesTSIterator::NodesTSIterator(const NodesTSIterator &other)
+    : sorted_(other.sorted_), cursor_(other.cursor_) {}
+
+Node &NodesTSIterator::operator*() {
+  PADDLE_ENFORCE_LT(cursor_, sorted_.size());
+  return *sorted_[cursor_];
+}
+
+NodesTSIterator &NodesTSIterator::operator++() {
+  if (++cursor_ >= sorted_.size()) {
+    sorted_.clear();
+    cursor_ = 0;
+  }
+  return *this;
+}
+NodesTSIterator &NodesTSIterator::operator=(const NodesTSIterator &other) {
+  cursor_ = other.cursor_;
+  sorted_ = other.sorted_;
+  return *this;
+}
+
+bool NodesTSIterator::operator==(const NodesTSIterator &other) {
+  return sorted_ == other.sorted_ && cursor_ == other.cursor_;
+}
+
+Node *NodesTSIterator::operator->() {
+  PADDLE_ENFORCE_LT(cursor_, sorted_.size());
+  return sorted_[cursor_];
+}
+
 }  // namespace ir
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/ir/graph_traits.h

@@ -62,6 +62,32 @@ struct NodesDFSIterator
   std::unordered_set<Node *> visited_;
 };
 
+// Topological sorting iterator on nodes.
+struct NodesTSIterator
+    : public std::iterator<std::forward_iterator_tag, Node *> {
+  NodesTSIterator() = default;
+  NodesTSIterator(const std::vector<Node *> &source);
+  NodesTSIterator(NodesTSIterator &&other)
+      : sorted_(std::move(other.sorted_)), cursor_(other.cursor_) {
+    other.cursor_ = 0;
+  }
+  NodesTSIterator(const NodesTSIterator &other);
+
+  Node &operator*();
+  NodesTSIterator &operator++();
+  // TODO(Superjomn) current implementation just compare the first
+  // element, need to compare the graph and all the elements in the queue and
+  // set.
+  NodesTSIterator &operator=(const NodesTSIterator &other);
+  bool operator==(const NodesTSIterator &other);
+  bool operator!=(const NodesTSIterator &other) { return !(*this == other); }
+  Node *operator->();
+
+ private:
+  std::vector<Node *> sorted_;
+  size_t cursor_{0};
+};
+
 /*
  * GraphTraits contains some graph traversal algorithms.
  *
@@ -76,6 +102,14 @@ struct GraphTraits {
                                             NodesDFSIterator());
   }
 
+  static iterator_range<NodesTSIterator> TS(const Graph &g) {
+    auto start_points = ExtractStartPoints(g);
+    PADDLE_ENFORCE(!start_points.empty());
+    NodesTSIterator x(start_points);
+    return iterator_range<NodesTSIterator>(NodesTSIterator(start_points),
+                                           NodesTSIterator());
+  }
+
  private:
   // The nodes those have no input will be treated as start points.
   static std::vector<Node *> ExtractStartPoints(const Graph &g) {

paddle/fluid/framework/ir/is_test_pass.cc

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/fluid/framework/ir/is_test_pass.h"
+#include <string>
+#include <utility>
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+std::unique_ptr<ir::Graph> IsTestPass::ApplyImpl(
+    std::unique_ptr<ir::Graph> graph) const {
+  VLOG(3) << "Sets is_test attrbiute to true and if it is missing, inserts it "
+             "for activations and pooling.";
+  auto op_list = {"pool2d",      "sigmoid",      "logsigmoid",
+                  "softshrink",  "exp",          "brelu",
+                  "pow",         "leaky_relu",   "stanh",
+                  "relu",        "tanh",         "tanh_shrink",
+                  "sqrt",        "abs",          "ceil",
+                  "elu",         "floor",        "cos",
+                  "sin",         "round",        "reciprocal",
+                  "hard_shrink", "hard_sigmoid", "relu6",
+                  "soft_relu",   "swish",        "thresholded_relu",
+                  "log",         "square",       "softplus",
+                  "softsign"};
+  for (const Node* n : graph->Nodes()) {
+    if (n->IsOp()) {
+      auto* op = n->Op();
+      if (op->HasAttr("is_test")) {
+        op->SetAttr("is_test", true);
+      } else if (std::find(begin(op_list), end(op_list), op->Type()) !=
+                 end(op_list)) {
+        op->MutableAttrMap()->insert(
+            std::pair<std::string, Attribute>("is_test", true));
+      }
+    }
+  }
+  return graph;
+}
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
+REGISTER_PASS(is_test_pass, paddle::framework::ir::IsTestPass);

paddle/fluid/inference/analysis/graph_traits.cc → paddle/fluid/framework/ir/is_test_pass.h

-/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
-http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License. */
-
-#include "paddle/fluid/inference/analysis/graph_traits.h"
+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#include "paddle/fluid/framework/ir/pass.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+class IsTestPass : public Pass {
+ protected:
+  std::unique_ptr<ir::Graph> ApplyImpl(
+      std::unique_ptr<ir::Graph> graph) const override;
+};
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/ir/is_test_pass_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.
+
+#include "paddle/fluid/framework/ir/is_test_pass.h"
+
+#include <gtest/gtest.h>
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+enum class ISTEST_STATE { FALSE, TRUE, UNSET };
+
+void SetOp(ProgramDesc* prog, const std::string& type, const std::string& name,
+           const std::vector<std::string>& inputs,
+           const std::vector<std::string>& outputs, bool use_mkldnn = false,
+           ISTEST_STATE is_test = ISTEST_STATE::UNSET) {
+  auto* op = prog->MutableBlock(0)->AppendOp();
+  op->SetType(type);
+  op->SetAttr("name", name);
+  op->SetInput("X", inputs);
+  op->SetOutput("Out", outputs);
+  op->SetAttr("use_mkldnn", use_mkldnn);
+  if (is_test == ISTEST_STATE::UNSET)
+    op->MutableAttrMap()->erase("is_test");
+  else if (is_test == ISTEST_STATE::FALSE)
+    op->SetAttr("is_test", false);
+  else
+    op->SetAttr("is_test", true);
+}
+
+// a->pool2d->b
+// b->relu->c
+// c,weights1)->conv2d->d
+//
+// d->pool2d->e
+// e->hard_sigmoid->f
+// (f,weights2)->conv2d->g
+//
+// g->pool2d->h
+// h->tanh->i
+// (i,weights3)->conv2d->j
+ProgramDesc BuildProgramDesc() {
+  ProgramDesc prog;
+  for (auto& v :
+       std::vector<std::string>({"a", "b", "c", "d", "e", "f", "g", "h", "i",
+                                 "j", "weights1", "weights2", "weights3"})) {
+    auto* var = prog.MutableBlock(0)->Var(v);
+    var->SetType(proto::VarType::SELECTED_ROWS);
+    if (v == "weights1" || v == "weights2" || v == "weights3") {
+      var->SetPersistable(true);
+    }
+  }
+
+  SetOp(&prog, "pool2d", "pooling1", std::vector<std::string>({"a"}),
+        std::vector<std::string>({"b"}), true, ISTEST_STATE::TRUE);
+  SetOp(&prog, "relu", "activation1", std::vector<std::string>({"b"}),
+        std::vector<std::string>({"c"}), true, ISTEST_STATE::TRUE);
+  SetOp(&prog, "conv2d", "conv1", std::vector<std::string>({"c", "weights1"}),
+        std::vector<std::string>({"d"}), true, ISTEST_STATE::TRUE);
+
+  SetOp(&prog, "pool2d", "pooling2", std::vector<std::string>({"d"}),
+        std::vector<std::string>({"e"}), false, ISTEST_STATE::FALSE);
+  SetOp(&prog, "hard_sigmoid", "activation2", std::vector<std::string>({"e"}),
+        std::vector<std::string>({"f"}), false, ISTEST_STATE::FALSE);
+  SetOp(&prog, "conv2d", "conv2", std::vector<std::string>({"f", "weights2"}),
+        std::vector<std::string>({"g"}), false, ISTEST_STATE::FALSE);
+
+  SetOp(&prog, "pool2d", "pooling3", std::vector<std::string>({"g"}),
+        std::vector<std::string>({"h"}), false, ISTEST_STATE::UNSET);
+  SetOp(&prog, "tanh", "activation3", std::vector<std::string>({"h"}),
+        std::vector<std::string>({"i"}), true, ISTEST_STATE::UNSET);
+  SetOp(&prog, "conv2d", "conv3", std::vector<std::string>({"i", "weights3"}),
+        std::vector<std::string>({"j"}), false, ISTEST_STATE::UNSET);
+
+  return prog;
+}
+
+TEST(IsTestPass, basic) {
+  auto prog = BuildProgramDesc();
+
+  std::unique_ptr<ir::Graph> graph(new ir::Graph(prog));
+
+  auto pass = PassRegistry::Instance().Get("is_test_pass");
+
+  graph = pass->Apply(std::move(graph));
+
+  for (auto* node : graph->Nodes()) {
+    if (node->IsOp()) {
+      auto* op = node->Op();
+      auto op_name = boost::get<std::string>(op->GetAttr("name"));
+      if (op_name == "conv3") {
+        ASSERT_FALSE(op->HasAttr("is_test"));
+      } else {
+        ASSERT_TRUE(op->HasAttr("is_test"));
+        EXPECT_TRUE(boost::get<bool>(op->GetAttr("is_test")));
+      }
+    }
+  }
+}
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
+USE_PASS(is_test_pass);

paddle/fluid/framework/ir/node.cc

@@ -17,8 +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[];
-int Node::count_ = 0;
+#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_; }
 
@@ -115,37 +119,30 @@ class Node {
   int id_;
 
  private:
+  // ID can only set by a Graph.
+  void SetId(int id) { id_ = id; }
+
   friend class Graph;
   friend std::unique_ptr<Node> CreateNodeForTest(const std::string& name,
                                                  Node::Type type);
 
   explicit Node(const std::string& name, Type type)
-      : name_(name),
-        var_desc_(nullptr),
-        op_desc_(nullptr),
-        type_(type),
-        id_(count_++) {}
+      : name_(name), var_desc_(nullptr), op_desc_(nullptr), type_(type) {}
 
   explicit Node(VarDesc* var_desc)
       : name_(var_desc->Name()),
         var_desc_(new VarDesc(*var_desc)),
         op_desc_(nullptr),
-        type_(Type::kVariable),
-        id_(count_++) {}
+        type_(Type::kVariable) {}
 
   explicit Node(OpDesc* op_desc)
       : name_(op_desc->Type()),
         var_desc_(nullptr),
         op_desc_(new OpDesc(*op_desc, op_desc->Block())),
-        type_(Type::kOperation),
-        id_(count_++) {}
+        type_(Type::kOperation) {}
 
   Node() = delete;
 
-  static int count_;
-  // Please don't use this API or make this public.
-  static void ResetId() { count_ = 0; }
-
   boost::any wrapper_;
   std::function<void(void)> wrapper_deleter_;
   std::type_index wrapper_type_ = std::type_index(typeid(void));

paddle/fluid/framework/ir/pass.h

@@ -93,6 +93,7 @@ class Pass {
  protected:
   virtual std::unique_ptr<Graph> ApplyImpl(std::unique_ptr<Graph> graph) const {
     LOG(FATAL) << "Calling virtual Pass not implemented.";
+    return graph;
   }
 
  private:
@@ -196,26 +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/lod_tensor.h

@@ -111,9 +111,6 @@ class LoDTensor : public Tensor {
  public:
   LoDTensor() : Tensor() {}
 
-  /* Constructor with place should only be used in pybind */
-  explicit LoDTensor(const platform::Place& place) : Tensor(place) {}
-
   explicit LoDTensor(const LoD& lod) : lod_(lod) {}
 
   void set_lod(const LoD& lod) { lod_ = lod; }

paddle/fluid/framework/mixed_vector.h

@@ -23,6 +23,7 @@
 #include "paddle/fluid/framework/details/cow_ptr.h"
 #include "paddle/fluid/framework/tensor.h"
 #include "paddle/fluid/framework/tensor_util.h"
+#include "paddle/fluid/memory/malloc.h"
 #include "paddle/fluid/memory/memcpy.h"
 
 #include "glog/logging.h"
@@ -31,46 +32,6 @@ namespace paddle {
 namespace framework {
 
 #if defined(PADDLE_WITH_CUDA)
-namespace details {
-struct CUDABuffer {
-  void *data_{nullptr};
-  size_t size_{0};
-  platform::CUDAPlace place_;
-
-  CUDABuffer() {}
-  CUDABuffer(platform::Place place, size_t size)
-      : size_(size), place_(boost::get<platform::CUDAPlace>(place)) {
-    data_ = memory::Alloc(place_, size);
-  }
-
-  ~CUDABuffer() { ClearMemory(); }
-
-  CUDABuffer(const CUDABuffer &o) = delete;
-  CUDABuffer &operator=(const CUDABuffer &o) = delete;
-
-  void Resize(platform::Place place, size_t size) {
-    ClearMemory();
-    place_ = boost::get<platform::CUDAPlace>(place);
-    data_ = memory::Alloc(place_, size);
-    PADDLE_ENFORCE_NOT_NULL(data_);
-    size_ = size;
-  }
-
-  void Swap(CUDABuffer &o) {
-    std::swap(data_, o.data_);
-    std::swap(place_, o.place_);
-    std::swap(size_, o.size_);
-  }
-
- private:
-  void ClearMemory() const {
-    if (data_ != nullptr) {
-      memory::Free(place_, data_);
-    }
-  }
-};
-}  // namespace details
-
 // Vector<T> implements the std::vector interface, and can get Data or
 // MutableData from any place. The data will be synced implicitly inside.
 template <typename T>
@@ -103,8 +64,6 @@ class Vector {
       o.ImmutableCPU();
       cpu_ = o.cpu_;
       flag_ = kDataInCPU;
-      details::CUDABuffer null;
-      gpu_.Swap(null);
       return *this;
     }
 
@@ -199,7 +158,7 @@ class Vector {
       PADDLE_ENFORCE(platform::is_gpu_place(place),
                      "CUDA Data must on CUDA place");
       ImmutableCUDA(place);
-      return reinterpret_cast<T *>(gpu_.data_);
+      return reinterpret_cast<T *>(gpu_->ptr());
     }
 
     // get cuda ptr. mutable
@@ -234,13 +193,11 @@ class Vector {
 
     std::mutex &Mutex() const { return mtx_; }
 
-    std::unique_ptr<platform::CUDAPlace> CUDAPlace() const {
-      if (gpu_.data_ == nullptr) {
-        return nullptr;
-      } else {
-        return std::unique_ptr<platform::CUDAPlace>(
-            new platform::CUDAPlace(gpu_.place_));
-      }
+    boost::optional<platform::CUDAPlace> CUDAPlace() const {
+      return gpu_ == nullptr
+                 ? boost::none
+                 : boost::optional<platform::CUDAPlace>(
+                       boost::get<platform::CUDAPlace>(gpu_->place()));
     }
 
    private:
@@ -254,13 +211,12 @@ class Vector {
     void CopyToCPU() const {
       // COPY GPU Data To CPU
       auto *dev_ctx = static_cast<platform::CUDADeviceContext *>(
-          platform::DeviceContextPool::Instance().Get(
-              platform::Place(gpu_.place_)));
+          platform::DeviceContextPool::Instance().Get(gpu_->place()));
       auto stream = dev_ctx->stream();
-      void *src = gpu_.data_;
+      void *src = gpu_->ptr();
       void *dst = cpu_.data();
-      memory::Copy(platform::CPUPlace(), dst, gpu_.place_, src, gpu_.size_,
-                   stream);
+      memory::Copy(platform::CPUPlace(), dst, CUDAPlace().get(), src,
+                   gpu_->size(), stream);
       dev_ctx->Wait();
     }
 
@@ -277,8 +233,7 @@ class Vector {
           CopyCPUDataToCUDA(place);
           UnsetFlag(kDirty);
           SetFlag(kDataInCUDA);
-        } else if (IsInCUDA() &&
-                   !(boost::get<platform::CUDAPlace>(place) == gpu_.place_)) {
+        } else if (IsInCUDA() && !(place == gpu_->place())) {
           PADDLE_THROW("This situation should not happen");
           // Still dirty
         } else {
@@ -290,7 +245,7 @@ class Vector {
           // Even data is not dirty. However, data is not in CUDA. Copy data.
           CopyCPUDataToCUDA(place);
           SetFlag(kDataInCUDA);
-        } else if (!(boost::get<platform::CUDAPlace>(place) == gpu_.place_)) {
+        } else if (!(place == gpu_->place())) {
           PADDLE_THROW("This situation should not happen.");
         } else {
           // Not Dirty && DataInCUDA && Device is same
@@ -301,13 +256,13 @@ class Vector {
 
     void CopyCPUDataToCUDA(const platform::Place &place) const {
       void *src = cpu_.data();
-      gpu_.Resize(place, cpu_.size() * sizeof(T));
-      void *dst = gpu_.data_;
+      gpu_ = memory::Alloc(place, cpu_.size() * sizeof(T));
+      void *dst = gpu_->ptr();
       auto *dev_ctx = static_cast<platform::CUDADeviceContext *>(
           platform::DeviceContextPool::Instance().Get(place));
       auto stream = dev_ctx->stream();
-      memory::Copy(gpu_.place_, dst, platform::CPUPlace(), src, gpu_.size_,
-                   stream);
+      memory::Copy(CUDAPlace().get(), dst, platform::CPUPlace(), src,
+                   gpu_->size(), stream);
     }
 
     void ImmutableCPU() const {
@@ -329,7 +284,7 @@ class Vector {
     bool IsInCPU() const { return flag_ & kDataInCPU; }
 
     mutable std::vector<T> cpu_;
-    mutable details::CUDABuffer gpu_;
+    mutable memory::AllocationPtr gpu_;
     mutable int flag_;
 
     mutable std::mutex mtx_;
@@ -428,8 +383,8 @@ class Vector {
       auto &mtx = m_.Data().Mutex();
       std::lock_guard<std::mutex> guard(mtx);
       auto cuda_place = m_.Data().CUDAPlace();
-      if (cuda_place == nullptr ||
-          *cuda_place == boost::get<platform::CUDAPlace>(place)) {
+      if (cuda_place == boost::none ||
+          cuda_place == boost::get<platform::CUDAPlace>(place)) {
         return m_.Data().CUDAData(place);
       }
     }
@@ -444,8 +399,8 @@ class Vector {
       auto &mtx = m_.Data().Mutex();
       std::lock_guard<std::mutex> guard(mtx);
       auto cuda_place = m_.Data().CUDAPlace();
-      if (cuda_place == nullptr ||
-          *cuda_place == boost::get<platform::CUDAPlace>(place)) {
+      if (cuda_place == boost::none ||
+          cuda_place == boost::get<platform::CUDAPlace>(place)) {
         return m_.MutableData()->CUDAMutableData(place);
       }
     }

paddle/fluid/framework/naive_executor.cc

@@ -57,60 +57,68 @@ static void InitializeVariable(Variable *var, proto::VarType::Type var_type) {
   }
 }
 
-void NaiveExecutor::Prepare(Scope *parent_scope,
-                            const ProgramDesc &program_desc, int block_id,
-                            bool with_feed_fetch_ops) {
-  if (!parent_scope) {
+void NaiveExecutor::Prepare(Scope *scope, const ProgramDesc &program_desc,
+                            int block_id, bool with_feed_fetch_ops) {
+  if (!scope) {
     scope_ = new framework::Scope;
   } else {
-    scope_ = &parent_scope->NewScope();
+    scope_ = scope;
   }
-  CreateVariables(program_desc, scope_, block_id);
+
+  VLOG(3) << "NaiveExecutor init with scope " << scope;
   CreateOps(program_desc, block_id, with_feed_fetch_ops);
 }
 
 void NaiveExecutor::Run() {
+#ifndef PADDLE_ON_INFERENCE
+  LOG_FIRST_N(WARNING, 15) << "The NaiveExecutor can not work properly if the "
+                              "cmake flag ON_INFER is not set.";
+  LOG_FIRST_N(WARNING, 15) << "Unlike the training phase, all the scopes and "
+                              "variables will be reused to save the allocation "
+                              "overhead.";
+  LOG_FIRST_N(WARNING, 15) << "Please re-compile the inference library by "
+                              "setting the cmake flag ON_INFER=ON if you are "
+                              "running Paddle Inference";
+#endif  // PADDLE_ON_INFERENCE
   for (auto &op : ops_) {
-    VLOG(40) << "run " << op->Type();
+    VLOG(3) << std::this_thread::get_id() << " run " << op->Type()
+            << " on scope " << scope_;
     op->Run(*scope_, place_);
   }
 }
 
-void NaiveExecutor::CreateVariables(const ProgramDesc &desc, Scope *scope,
-                                    int block_id) {
-  PADDLE_ENFORCE(scope);
+void NaiveExecutor::CreateVariables(const ProgramDesc &desc, int block_id,
+                                    bool persistable, Scope *scope) {
+  PADDLE_ENFORCE_NOT_NULL(scope);
+
   auto &global_block = desc.Block(block_id);
 
-  const Scope *ancestor_scope = scope;
-  while (ancestor_scope->parent()) {
-    ancestor_scope = ancestor_scope->parent();
+  const auto *anc = scope;
+  PADDLE_ENFORCE(anc->parent() != anc);
+  while (anc->parent()) {
+    anc = anc->parent();
   }
 
-  if (ancestor_scope != scope) {
-    for (auto &var : global_block.AllVars()) {
-      if (var->Name() == framework::kEmptyVarName) {
-        continue;
-      }
-      // Create persistable vars in ancestor scope.
-      if (var->Persistable()) {
-        auto *ptr = const_cast<Scope *>(ancestor_scope)->Var(var->Name());
-        InitializeVariable(ptr, var->GetType());
-        VLOG(30) << "Create Variable " << var->Name()
-                 << " global, which pointer is " << ptr;
-      } else {  // Create temporary variables in local scope.
-        auto *ptr = scope->Var(var->Name());
+  for (auto &var : global_block.AllVars()) {
+    if (var->Name() == framework::kEmptyVarName) {
+      continue;
+    }
+
+    if (persistable == var->Persistable()) {
+      if (persistable) {
+        if (!anc->FindVar(var->Name())) {
+          auto *ptr = const_cast<Scope *>(anc)->Var(var->Name());
+          VLOG(3) << scope << " Create persistable variable " << var->Name()
+                  << ", which pointer is " << ptr;
+          InitializeVariable(ptr, var->GetType());
+        }
+      } else {
+        auto *ptr = const_cast<Scope *>(scope)->Var(var->Name());
+        VLOG(3) << scope << " Create variable " << var->Name()
+                << ", which pointer is " << ptr;
         InitializeVariable(ptr, var->GetType());
-        VLOG(30) << "Create Variable " << var->Name()
-                 << " locally, which pointer is " << ptr;
       }
     }
-  } else {
-    for (auto &var : global_block.AllVars()) {
-      auto *ptr = scope->Var(var->Name());
-      InitializeVariable(ptr, var->GetType());
-      VLOG(30) << "Create variable " << var->Name() << ", which pointer is "
-               << ptr;
-    }
   }
 }
 

paddle/fluid/framework/naive_executor.h

@@ -35,8 +35,14 @@ class NaiveExecutor {
   // Create child scope.
   // Create variables.
   // @with_feed_fetch_ops: whether to work with the feed and fetch operators.
-  void Prepare(Scope* parent_scope, const ProgramDesc& program_desc,
-               int block_id, bool with_feed_fetch_ops);
+  void Prepare(Scope* scope, const ProgramDesc& program_desc, int block_id,
+               bool with_feed_fetch_ops);
+
+  // Create variables before head.
+  // Create parameters if persistable is ture, or create the temporary variables
+  // instead.
+  void CreateVariables(const ProgramDesc& desc, int block_id, bool persistable,
+                       Scope* scope);
 
   // Run all the operators.
   void Run();
@@ -49,8 +55,6 @@ class NaiveExecutor {
   void CleanFeedFetchOps();
 
  protected:
-  void CreateVariables(const ProgramDesc& desc, Scope* scope, int block_id);
-
   void CreateOps(const ProgramDesc& desc, int block_id,
                  bool with_feed_fetch_ops);
 

paddle/fluid/framework/naive_executor_test.cc

@@ -39,7 +39,7 @@ TEST(NaiveExecutor, Basic) {
 
   auto place = platform::CPUPlace();
   NaiveExecutor exe(place);
-  exe.Prepare(nullptr, program, 0, false /*with feed fetch ops*/);
+  exe.Prepare(nullptr, program, 0, false);
   auto* a_tensor = exe.FindTensor("a");
   auto* b_tensor = exe.FindTensor("b");
   auto* c_tensor = exe.FindTensor("c");

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/op_kernel_type.h

@@ -63,6 +63,8 @@ struct OpKernelType {
         place_(dev_ctx.GetPlace()),
         library_type_(library_type) {}
 
+  size_t hash_key() const { return Hash()(*this); }
+
   bool operator==(const OpKernelType& o) const {
     return platform::places_are_same_class(place_, o.place_) &&
            data_type_ == o.data_type_ && data_layout_ == o.data_layout_ &&

paddle/fluid/framework/operator.cc

@@ -35,6 +35,11 @@ DEFINE_bool(check_nan_inf, false,
 namespace paddle {
 namespace framework {
 
+// Combine two hash values to a single hash.
+inline size_t CombineHash(size_t seed, size_t a) {
+  return (seed ^ a) + 0x9e3779b9 + (seed << 6) + (seed >> 2);
+}
+
 std::vector<std::tuple<platform::Place, LibraryType>> kKernelPriority = {
     std::make_tuple(platform::CUDAPlace(0), LibraryType::kCUDNN),
     std::make_tuple(platform::CUDAPlace(0), LibraryType::kPlain),
@@ -150,14 +155,17 @@ 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);
@@ -791,6 +799,17 @@ void OperatorWithKernel::TransferInplaceVarsBack(
 Scope* OperatorWithKernel::TryTransferData(
     const Scope& scope, const OpKernelType& expected_kernel_key,
     std::vector<std::string>* transfered_inplace_vars) const {
+// In the inference scenerio, the scopes will be reused across the batches, so
+// the `new_scope` here will result in GPU memroy explosion over the running of
+// operators.
+// We use a thread_local cache to fix that issue, the key in the cache is the
+// combination of the `scope` argument, from_kernel_type, target_kernel_type.
+// Have a discussion with @Superjomn or the inference developers if some changes
+// on this logic for this macro might not tested on the other scenerios.
+#ifdef PADDLE_ON_INFERENCE
+  thread_local std::unordered_map<size_t, Scope*> infer_transfer_scope_cache;
+#endif
+
   Scope* new_scope = nullptr;
   for (auto& var_name_item : Inputs()) {
     for (auto& var_name : var_name_item.second) {
@@ -821,11 +840,28 @@ Scope* OperatorWithKernel::TryTransferData(
       VLOG(30) << "Transform Variable " << var_name << " from "
                << kernel_type_for_var << " to " << expected_kernel_key;
 
+#ifdef PADDLE_ON_INFERENCE
+      size_t infer_cache_key =
+          CombineHash(OpKernelType::Hash()(kernel_type_for_var),
+                      OpKernelType::Hash()(expected_kernel_key));
+      infer_cache_key =
+          CombineHash(infer_cache_key, std::hash<const Scope*>()(&scope));
+
+      auto it = infer_transfer_scope_cache.find(infer_cache_key);
+      if (it != infer_transfer_scope_cache.end()) {
+        new_scope = infer_transfer_scope_cache[infer_cache_key];
+      } else {
+        new_scope = &scope.NewScope();
+        infer_transfer_scope_cache[infer_cache_key] = new_scope;
+      }
+#endif
+
       if (new_scope == nullptr) {
         new_scope = &scope.NewScope();
       }
 
       auto* trans_var = new_scope->Var(var_name);
+
       Tensor out;
       TransformData(expected_kernel_key, kernel_type_for_var, *tensor_in, &out);
       SetTensorToVariable(*var, out, trans_var);

paddle/fluid/framework/parallel_executor.cc

@@ -171,8 +171,17 @@ ParallelExecutor::ParallelExecutor(
   }
   // If the loss_var_name is given, the number of graph should be only one.
   if (loss_var_name.size()) {
-    PADDLE_ENFORCE_EQ(ir::GraphNum(*graph), 1,
-                      "The number of graph should be only one");
+    size_t graph_num = ir::GraphNum(*graph);
+    if (graph_num > 1) {
+      LOG(WARNING)
+          << "The number of graph should be only one, "
+             "but the current graph has "
+          << ir::GraphNum(*graph)
+          << " sub_graphs. If you want to see the nodes of the "
+             "sub_graphs, you should use 'FLAGS_print_sub_graph_dir' "
+             "to specify the output dir. NOTES: if you not do training, "
+             "please don't pass loss_var_name.";
+    }
   }
 
   if (exec_strategy.type_ == ExecutionStrategy::kDefault) {

paddle/fluid/framework/scope.cc

@@ -15,7 +15,9 @@ limitations under the License. */
 #include "paddle/fluid/framework/scope.h"
 
 #include <memory>  // for unique_ptr
+#include <queue>
 #include <set>
+#include <unordered_set>
 #include "glog/logging.h"
 #include "paddle/fluid/framework/threadpool.h"
 #include "paddle/fluid/string/printf.h"
@@ -36,6 +38,16 @@ DEFINE_double(
     "Memory size threshold (GB) when the garbage collector clear tensors."
     "Disabled when this value is less than 0");
 
+// When in inference scenario, the scopes will not be written by two threads in
+// a mean time, but a scope may be read by multiple threads concurrently, and
+// the mutex will cause serious performance issue.
+// So the mutex is disabled when `ON_INFER`.
+#ifdef PADDLE_ON_INFERENCE
+#define SCOPE_LOCK_GUARD
+#else
+#define SCOPE_LOCK_GUARD std::lock_guard<std::mutex> lock(mutex_);
+#endif
+
 namespace paddle {
 namespace framework {
 
@@ -49,18 +61,18 @@ int64_t GetEagerDeletionThreshold() {
 Scope::~Scope() { DropKids(); }
 
 Scope& Scope::NewScope() const {
-  std::lock_guard<std::mutex> lock(mutex_);
+  SCOPE_LOCK_GUARD
   kids_.push_back(new Scope(this));
   return *kids_.back();
 }
 
 Variable* Scope::Var(const std::string& name) {
-  std::lock_guard<std::mutex> lock(mutex_);
+  SCOPE_LOCK_GUARD
   return VarInternal(name);
 }
 
 Variable* Scope::Var(std::string* name) {
-  std::lock_guard<std::mutex> lock(mutex_);
+  SCOPE_LOCK_GUARD
   auto new_name = string::Sprintf("%p.%d", this, vars_.size());
   if (name != nullptr) {
     *name = new_name;
@@ -69,34 +81,34 @@ Variable* Scope::Var(std::string* name) {
 }
 
 Variable* Scope::FindVar(const std::string& name) const {
-  std::lock_guard<std::mutex> lock(mutex_);
+  SCOPE_LOCK_GUARD
   return FindVarInternal(name);
 }
 
 Variable* Scope::FindLocalVar(const std::string& name) const {
-  std::lock_guard<std::mutex> lock(mutex_);
+  SCOPE_LOCK_GUARD
   return FindVarLocally(name);
 }
 
 const Scope* Scope::FindScope(const Variable* var) const {
-  std::lock_guard<std::mutex> lock(mutex_);
+  SCOPE_LOCK_GUARD
   return FindScopeInternal(var);
 }
 
 void Scope::DropKids() {
-  std::lock_guard<std::mutex> lock(mutex_);
+  SCOPE_LOCK_GUARD
   for (Scope* s : kids_) delete s;
   kids_.clear();
 }
 
 bool Scope::HasKid(const Scope* scope) const {
-  std::lock_guard<std::mutex> lock(mutex_);
+  SCOPE_LOCK_GUARD
   auto it = std::find(this->kids_.begin(), this->kids_.end(), scope);
   return it != this->kids_.end();
 }
 
 std::vector<std::string> Scope::LocalVarNames() const {
-  std::lock_guard<std::mutex> lock(mutex_);
+  SCOPE_LOCK_GUARD
   std::vector<std::string> known_vars;
   known_vars.reserve(this->vars_.size());
   for (auto& p : vars_) {
@@ -106,9 +118,10 @@ std::vector<std::string> Scope::LocalVarNames() const {
 }
 
 void Scope::DeleteScope(Scope* scope) const {
-  std::lock_guard<std::mutex> lock(mutex_);
+  SCOPE_LOCK_GUARD
   auto it = std::find(this->kids_.begin(), this->kids_.end(), scope);
-  PADDLE_ENFORCE(it != this->kids_.end(), "Cannot find %p as kid scope", scope);
+  PADDLE_ENFORCE(it != this->kids_.end(), "%p Cannot find %p as kid scope",
+                 this, scope);
   this->kids_.erase(it);
   // When making memory benchmark on Fluid, we have to delete scope sync.
   if (FLAGS_benchmark || FLAGS_eager_delete_scope) {
@@ -119,7 +132,7 @@ void Scope::DeleteScope(Scope* scope) const {
 }
 
 void Scope::EraseVars(const std::vector<std::string>& var_names) {
-  std::lock_guard<std::mutex> lock(mutex_);
+  SCOPE_LOCK_GUARD
   std::set<std::string> var_set(var_names.begin(), var_names.end());
   for (auto it = vars_.begin(); it != vars_.end();) {
     if (var_set.find(it->first) != var_set.end()) {
@@ -132,12 +145,12 @@ void Scope::EraseVars(const std::vector<std::string>& var_names) {
 
 void Scope::Rename(const std::string& origin_name,
                    const std::string& new_name) const {
-  std::lock_guard<std::mutex> lock(mutex_);
+  SCOPE_LOCK_GUARD
   RenameInternal(origin_name, new_name);
 }
 
 std::string Scope::Rename(const std::string& origin_name) const {
-  std::lock_guard<std::mutex> lock(mutex_);
+  SCOPE_LOCK_GUARD
   auto new_name = string::Sprintf("%p.%d", this, vars_.size());
   RenameInternal(origin_name, new_name);
   return new_name;
@@ -189,5 +202,46 @@ Variable* Scope::FindVarLocally(const std::string& name) const {
   return nullptr;
 }
 
+std::string GenScopeTreeDebugInfo(Scope* root) {
+  std::stringstream os;
+
+  if (!root) return "";
+
+  // level traversal
+  std::queue<Scope*> queue;
+  queue.push(root);
+
+  std::vector<Scope*> scopes;
+
+  while (!queue.empty()) {
+    auto* end = queue.back();
+    Scope* q = nullptr;
+    while (q != end) {
+      q = queue.front();
+      queue.pop();
+      os << q << " ";
+      scopes.push_back(q);
+
+      for (auto* c : q->kids()) {
+        queue.push(c);
+      }
+    }
+    // end of a level
+    os << "\n------------------------------------------\n";
+  }
+
+  os << "\nDetails:\n\n";
+
+  for (Scope* q : scopes) {
+    os << "====\n";
+    os << q << ":\n";
+    for (auto& var : q->LocalVarNames()) {
+      os << "  - " << var << "\n";
+    }
+  }
+
+  return os.str();
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/scope.h

@@ -78,11 +78,11 @@ class Scope {
   /// Drop all kids scopes belonged to this scope.
   void DropKids();
 
-  std::list<Scope*>& kids() const { return kids_; }
-
   /// Find if a scope exists in the kid scopes
   bool HasKid(const Scope* scope) const;
 
+  const std::list<Scope*>& kids() const { return kids_; }
+
   // enumerate all the variables current contains.
   std::vector<std::string> LocalVarNames() const;
 
@@ -118,12 +118,17 @@ class Scope {
 
   // Scope in `kids_` are owned by this class.
   mutable std::list<Scope*> kids_;
-  Scope const* parent_{nullptr};
+  const Scope* parent_{nullptr};
 
   DISABLE_COPY_AND_ASSIGN(Scope);
 
  private:
   mutable std::mutex mutex_;
 };
+
+// Generate some debug string about the inherience structure of scope, quite
+// naive.
+std::string GenScopeTreeDebugInfo(Scope*);
+
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/selected_rows.cc

@@ -63,6 +63,26 @@ struct TensorCopyVisitor {
   int64_t size_;
 };
 
+struct TensorFillVisitor {
+  TensorFillVisitor(framework::Tensor* dst, int64_t dst_offset, int64_t size,
+                    float value)
+      : dst_(dst), dst_offset_(dst_offset), size_(size) {}
+
+  template <typename T>
+  void apply() const {
+    // TODO(qiao): support other place
+    platform::CPUPlace cpu;
+    auto* tensor_data = dst_->mutable_data<T>(cpu);
+    auto* start = tensor_data + dst_offset_;
+    auto* end = start + size_;
+    std::fill(start, end, static_cast<T>(0.0));
+  }
+
+  framework::Tensor* dst_;
+  int64_t dst_offset_;
+  int64_t size_;
+};
+
 void SerializeToStream(std::ostream& os, const SelectedRows& selected_rows,
                        const platform::DeviceContext& dev_ctx) {
   {  // the 1st field, uint32_t version
@@ -120,7 +140,17 @@ bool SelectedRows::HasKey(int64_t key) const {
                                                                    : true;
 }
 
-int64_t SelectedRows::AutoGrownIndex(int64_t key, bool auto_grown) {
+int64_t SelectedRows::AutoGrownIndex(int64_t key, bool auto_grown,
+                                     bool is_test) {
+  if (is_test) {
+    auto iter = id_to_index_.find(key);
+    if (iter == id_to_index_.end()) {
+      return -1;
+    } else {
+      return iter->second;
+    }
+  }
+
   rwlock_->RDLock();
   auto iter = id_to_index_.find(key);
   if (iter == id_to_index_.end()) {
@@ -172,7 +202,7 @@ void SelectedRows::SyncIndex() {
 }
 
 void SelectedRows::Get(const framework::Tensor& ids, framework::Tensor* value,
-                       bool auto_grown) {
+                       bool auto_grown, bool is_test) {
   PADDLE_ENFORCE(value->IsInitialized(),
                  "The value tensor should be initialized.");
   if (ids.numel() == 0) {
@@ -183,11 +213,19 @@ void SelectedRows::Get(const framework::Tensor& ids, framework::Tensor* value,
                       "output tensor should have the same shape with table "
                       "except the dims[0].");
     for (int i = 0; i < ids.numel(); ++i) {
-      int64_t index = AutoGrownIndex(ids.data<int64_t>()[i], auto_grown);
-      framework::VisitDataType(
-          framework::ToDataType(value_->type()),
-          TensorCopyVisitor(value, i * value_width, *value_.get(),
-                            index * value_width, value_width));
+      auto id = ids.data<int64_t>()[i];
+      int64_t index = AutoGrownIndex(id, auto_grown, is_test);
+      if (index < 0) {
+        VLOG(5) << "id " << id << " not in the table, return 0";
+        framework::VisitDataType(
+            framework::ToDataType(value_->type()),
+            TensorFillVisitor(value, i * value_width, value_width, 0.0));
+      } else {
+        framework::VisitDataType(
+            framework::ToDataType(value_->type()),
+            TensorCopyVisitor(value, i * value_width, *value_.get(),
+                              index * value_width, value_width));
+      }
     }
   }
 }

paddle/fluid/framework/selected_rows.h

@@ -105,7 +105,7 @@ class SelectedRows {
    * the value
    */
   void Get(const framework::Tensor& ids, framework::Tensor* value,
-           bool auto_grown = false);
+           bool auto_grown = false, bool is_test = false);
 
   /*
    * @brief Get the index of the key from id_to_index_ map. If the key not
@@ -118,7 +118,7 @@ class SelectedRows {
    *
    * @return index of the key.
    */
-  int64_t AutoGrownIndex(int64_t key, bool auto_grown);
+  int64_t AutoGrownIndex(int64_t key, bool auto_grown, bool is_test = false);
 
   void SyncIndex();
 

paddle/fluid/framework/selected_rows_test.cc

@@ -84,10 +84,14 @@ TEST(SelectedRows, SparseTable) {
       data[i * embedding_width + j] = static_cast<float>(i);
     }
   }
-  ASSERT_EQ(table.AutoGrownIndex(10, true), 0);
-  ASSERT_EQ(table.AutoGrownIndex(8, true), 1);
-  ASSERT_EQ(table.AutoGrownIndex(8, true), 1);
-  ASSERT_EQ(table.AutoGrownIndex(6, true), 2);
+  ASSERT_EQ(table.AutoGrownIndex(10, true, false), 0);
+  ASSERT_EQ(table.AutoGrownIndex(8, true, false), 1);
+  ASSERT_EQ(table.AutoGrownIndex(8, true, false), 1);
+  ASSERT_EQ(table.AutoGrownIndex(6, true, false), 2);
+  for (int64_t i = 11; i < 20; i++) {
+    ASSERT_EQ(table.AutoGrownIndex(i, true, true), -1);
+    ASSERT_TRUE(!table.HasKey(i));
+  }
   ASSERT_TRUE(table.HasKey(10));
   ASSERT_TRUE(table.HasKey(8));
   ASSERT_TRUE(table.HasKey(6));

paddle/fluid/framework/tensor.cc

@@ -32,10 +32,9 @@ size_t Tensor::memory_size() const {
 }
 
 void* Tensor::mutable_data(platform::Place place, std::type_index type,
+                           memory::Allocator::Attr attr,
                            size_t requested_size) {
-  if (holder_ != nullptr) {
-    holder_->set_type(type);
-  }
+  type_ = type;
   PADDLE_ENFORCE_GE(numel(), 0,
                     "When calling this method, the Tensor's numel must be "
                     "equal or larger than zero. "
@@ -48,35 +47,18 @@ void* Tensor::mutable_data(platform::Place place, std::type_index type,
   /* some versions of boost::variant don't have operator!= */
   if (holder_ == nullptr || !(holder_->place() == place) ||
       holder_->size() < size + offset_) {
-    if (platform::is_cpu_place(place)) {
-      holder_.reset(new PlaceholderImpl<platform::CPUPlace>(
-          boost::get<platform::CPUPlace>(place), size, type));
-    } else if (platform::is_gpu_place(place) ||
-               platform::is_cuda_pinned_place(place)) {
-#ifndef PADDLE_WITH_CUDA
-      PADDLE_THROW(
-          "CUDAPlace or CUDAPinnedPlace is not supported in CPU-only mode.");
-    }
-#else
-      if (platform::is_gpu_place(place)) {
-        holder_.reset(new PlaceholderImpl<platform::CUDAPlace>(
-            boost::get<platform::CUDAPlace>(place), size, type));
-      } else if (platform::is_cuda_pinned_place(place)) {
-        holder_.reset(new PlaceholderImpl<platform::CUDAPinnedPlace>(
-            boost::get<platform::CUDAPinnedPlace>(place), size, type));
-      }
-    }
-#endif
+    holder_ = memory::AllocShared(place, size, attr);
     offset_ = 0;
   }
   return reinterpret_cast<void*>(reinterpret_cast<uintptr_t>(holder_->ptr()) +
                                  offset_);
 }
 
-void* Tensor::mutable_data(platform::Place place, size_t requested_size) {
+void* Tensor::mutable_data(platform::Place place, memory::Allocator::Attr attr,
+                           size_t requested_size) {
   PADDLE_ENFORCE(this->holder_ != nullptr,
                  "Cannot invoke mutable data if current hold nothing.");
-  return mutable_data(place, holder_->type(), requested_size);
+  return mutable_data(place, type_, attr, requested_size);
 }
 
 Tensor& Tensor::ShareDataWith(const Tensor& src) {
@@ -101,6 +83,7 @@ Tensor Tensor::Slice(int begin_idx, int end_idx) const {
     Tensor dst;
     dst.holder_ = holder_;
     dst.set_layout(layout_);
+    dst.type_ = type_;
     DDim dst_dims = dims_;
     dst_dims[0] = end_idx - begin_idx;
     dst.Resize(dst_dims);

paddle/fluid/framework/tensor.h

@@ -67,12 +67,7 @@ class Tensor {
   friend struct EigenVector;
 
  public:
-  Tensor() : offset_(0) {}
-
-  /*! Constructor with place should only be used in pybind. */
-  explicit Tensor(const platform::Place& place) : offset_(0) {
-    holder_->set_place(place);
-  }
+  Tensor() : type_(typeid(float)), offset_(0) {}
 
   /*! Return a pointer to mutable memory block. */
   template <typename T>
@@ -89,12 +84,17 @@ class Tensor {
    * @note    If not exist, then allocation.
    */
   template <typename T>
-  T* mutable_data(platform::Place place, size_t requested_size = 0);
+  T* mutable_data(platform::Place place,
+                  memory::Allocator::Attr attr = memory::Allocator::kDefault,
+                  size_t requested_size = 0);
 
   void* mutable_data(platform::Place place, std::type_index type,
+                     memory::Allocator::Attr attr = memory::Allocator::kDefault,
                      size_t requested_size = 0);
 
-  void* mutable_data(platform::Place place, size_t requested_size = 0);
+  void* mutable_data(platform::Place place,
+                     memory::Allocator::Attr attr = memory::Allocator::kDefault,
+                     size_t requested_size = 0);
 
   /**
    * @brief     Return a pointer to mutable memory block.
@@ -106,7 +106,9 @@ class Tensor {
    * @note      If not exist, then allocation.
    */
   template <typename T>
-  T* mutable_data(DDim dims, platform::Place place, size_t requested_size = 0);
+  T* mutable_data(DDim dims, platform::Place place,
+                  memory::Allocator::Attr attr = memory::Allocator::kDefault,
+                  size_t requested_size = 0);
 
   /*! Return the dimensions of the memory block. */
   const DDim& dims() const;
@@ -139,7 +141,7 @@ class Tensor {
   std::type_index type() const {
     PADDLE_ENFORCE_NOT_NULL(
         holder_, "Tensor not initialized yet when Tensor::type() is called.");
-    return holder_->type();
+    return type_;
   }
 
   // memory size returns the holding memory size in byte.
@@ -153,56 +155,13 @@ class Tensor {
 
   void clear() { holder_ = nullptr; }
 
- private:
-  /**
-   * @note    Placeholder hides type T, so it doesn't appear as a template
-   *          parameter of Variable.
-   */
-  struct Placeholder {
-    virtual ~Placeholder() = default;
-    virtual void* ptr() const = 0;
-    virtual size_t size() const = 0;
-    virtual std::type_index type() const = 0;
-    virtual platform::Place place() const = 0;
-    virtual void set_type(std::type_index type) = 0;
-    virtual void set_place(platform::Place place) = 0;
-  };
-
-  template <typename Place>
-  struct PlaceholderImpl : public Placeholder {
-    PlaceholderImpl(Place place, size_t size, std::type_index type)
-        : ptr_(static_cast<uint8_t*>(memory::Alloc(place, size)),
-               memory::PODDeleter<uint8_t, Place>(place)),
-          place_(place),
-          size_(size),
-          type_(type) {
-      PADDLE_ENFORCE_NOT_NULL(ptr_, "Insufficient %s memory to allocation.",
-                              (is_cpu_place(place_) ? "CPU" : "GPU"));
-    }
-
-    virtual size_t size() const { return size_; }
-    virtual platform::Place place() const { return place_; }
-    virtual void* ptr() const { return static_cast<void*>(ptr_.get()); }
-    virtual std::type_index type() const { return type_; }
-    virtual void set_type(std::type_index type) { type_ = type; }
-    virtual void set_place(platform::Place place) { place_ = place; }
-
-    /*! the pointer of memory block. */
-    std::unique_ptr<uint8_t, memory::PODDeleter<uint8_t, Place>> ptr_;
-
-    /*! the place of memory block. */
-    platform::Place place_;
-
-    /*! the size of memory block. */
-    size_t size_;
-
-    /* the current type of memory */
-    std::type_index type_;
-  };
+  const std::shared_ptr<memory::Allocation>& Holder() const { return holder_; }
+  size_t offset() const { return offset_; }
 
+ private:
   /*! holds the memory block if allocated. */
-  std::shared_ptr<Placeholder> holder_;
-
+  std::shared_ptr<memory::Allocation> holder_;
+  std::type_index type_;
   /**
    * @brief points to elements dimensions.
    *

paddle/fluid/framework/tensor_impl.h

@@ -23,10 +23,10 @@ namespace framework {
 template <typename T>
 inline const T* Tensor::data() const {
   check_memory_size();
-  bool valid = std::is_same<T, void>::value ||
-               holder_->type() == std::type_index(typeid(T));
-  PADDLE_ENFORCE(valid, "Tensor holds the wrong type, it holds %d",
-                 this->holder_->type());
+  bool valid =
+      std::is_same<T, void>::value || type_ == std::type_index(typeid(T));
+  PADDLE_ENFORCE(valid, "Tensor holds the wrong type, it holds %s",
+                 type_.name());
 
   return reinterpret_cast<const T*>(
       reinterpret_cast<uintptr_t>(holder_->ptr()) + offset_);
@@ -37,26 +37,30 @@ inline bool Tensor::IsInitialized() const { return holder_ != nullptr; }
 template <typename T>
 inline T* Tensor::data() {
   check_memory_size();
-  bool valid = std::is_same<T, void>::value ||
-               holder_->type() == std::type_index(typeid(T));
+  bool valid =
+      std::is_same<T, void>::value || type_ == std::type_index(typeid(T));
   PADDLE_ENFORCE(valid, "Tensor holds the wrong type, it holds %s",
-                 this->holder_->type().name());
+                 type_.name());
   return reinterpret_cast<T*>(reinterpret_cast<uintptr_t>(holder_->ptr()) +
                               offset_);
 }
 
 template <typename T>
 inline T* Tensor::mutable_data(DDim dims, platform::Place place,
+                               memory::Allocator::Attr attr,
                                size_t requested_size) {
   static_assert(std::is_pod<T>::value, "T must be POD");
   Resize(dims);
-  return mutable_data<T>(place, requested_size);
+  return mutable_data<T>(place, attr, requested_size);
 }
 
 template <typename T>
-inline T* Tensor::mutable_data(platform::Place place, size_t requested_size) {
+inline T* Tensor::mutable_data(platform::Place place,
+                               memory::Allocator::Attr attr,
+                               size_t requested_size) {
   static_assert(std::is_pod<T>::value, "T must be POD");
-  return reinterpret_cast<T*>(mutable_data(place, typeid(T), requested_size));
+  return reinterpret_cast<T*>(
+      mutable_data(place, typeid(T), attr, requested_size));
 }
 
 inline Tensor ReshapeToMatrix(const Tensor& src, int num_col_dims) {

paddle/fluid/framework/tensor_util_test.cc

@@ -379,7 +379,9 @@ TEST(Tensor, FromAndToStream) {
     TensorToStream(oss, gpu_tensor, gpu_ctx);
 
     std::istringstream iss(oss.str());
-    TensorFromStream(iss, &dst_tensor, gpu_ctx);
+    TensorFromStream(
+        iss, &dst_tensor,
+        *platform::DeviceContextPool::Instance().Get(platform::CPUPlace()));
 
     int* dst_ptr = dst_tensor.mutable_data<int>(platform::CPUPlace());
     for (int i = 0; i < 6; ++i) {

paddle/fluid/inference/CMakeLists.txt

@@ -13,12 +13,24 @@ set(FLUID_CORE_MODULES proto_desc memory lod_tensor executor)
 # TODO(panyx0718): Should this be called paddle_fluid_inference_api_internal?
 cc_library(paddle_fluid_api
     SRCS io.cc
-    DEPS ${FLUID_CORE_MODULES} ${GLOB_OP_LIB}) 
+    DEPS ${FLUID_CORE_MODULES} ${GLOB_OP_LIB} ${GLOB_OPERATOR_DEPS})
 
 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)
 
@@ -27,13 +39,17 @@ set(SHARED_INFERENCE_SRCS
     io.cc ${CMAKE_CURRENT_SOURCE_DIR}/api/api.cc ${CMAKE_CURRENT_SOURCE_DIR}/api/api_impl.cc
     ${CMAKE_CURRENT_SOURCE_DIR}/api/analysis_predictor.cc
     ${CMAKE_CURRENT_SOURCE_DIR}/api/details/zero_copy_tensor.cc)
-if (WITH_GPU AND TENSORRT_FOUND)
-  set(STATIC_INFERENCE_APIS ${STATIC_INFERENCE_APIS} paddle_inference_tensorrt_subgraph_engine)
-  set(SHARED_INFERENCE_SRCS ${SHARED_INFERENCE_SRCS} ${CMAKE_CURRENT_SOURCE_DIR}/api/api_tensorrt_subgraph_engine.cc)
-endif()
 
-# Create static library
-cc_library(paddle_fluid DEPS ${fluid_modules} ${STATIC_INFERENCE_APIS} zero_copy_tensor reset_tensor_array)
+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.
@@ -42,11 +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)
+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/CMakeLists.txt

-cc_library(ir_pass_manager SRCS ir_pass_manager.cc DEPS graph pass)
-set(analysis_deps
-        framework_proto proto_desc ir_pass_manager graph pass paddle_fluid_api executor pretty_log)
+unset(analysis_deps CACHE)
+set(analysis_deps # analysis_deps can be extended accross the project
+        framework_proto proto_desc graph pass paddle_fluid_api executor pretty_log
+        ir_pass_manager
+        CACHE INTERNAL "")
 
-cc_library(analysis SRCS pass_manager.cc node.cc data_flow_graph.cc graph_traits.cc subgraph_splitter.cc
+add_subdirectory(ir_passes)
+add_subdirectory(passes)
+
+cc_library(ir_pass_manager SRCS ir_pass_manager.cc DEPS graph pass ${INFER_IR_PASSES})
+
+cc_library(argument SRCS argument.cc DEPS scope proto_desc)
+cc_library(analysis_pass SRCS analysis_pass.cc DEPS proto_desc)
+
+cc_library(analysis SRCS
   analyzer.cc
   helper.cc
-  # passes
-  analysis_pass.cc
-  fluid_to_data_flow_graph_pass.cc
-  data_flow_graph_to_fluid_pass.cc
-  dfg_graphviz_draw_pass.cc
-  tensorrt_subgraph_pass.cc
-  tensorrt_subgraph_node_mark_pass.cc
-  fluid_to_ir_pass.cc
-  model_store_pass.cc
-  DEPS ${analysis_deps})
+  analysis_pass
+  DEPS ${analysis_deps}
+  )
 
-cc_test(test_node SRCS node_tester.cc DEPS analysis)
 cc_test(test_dot SRCS dot_tester.cc DEPS analysis)
-cc_binary(inference_analyzer SRCS analyzer_main.cc DEPS analysis paddle_fluid)
 
 function(inference_analysis_test TARGET)
   if(WITH_TESTING)
@@ -34,13 +35,3 @@ function(inference_analysis_test TARGET)
 endfunction(inference_analysis_test)
 
 inference_analysis_test(test_analyzer SRCS analyzer_tester.cc EXTRA_DEPS paddle_inference_api)
-inference_analysis_test(test_data_flow_graph SRCS data_flow_graph_tester.cc)
-inference_analysis_test(test_data_flow_graph_to_fluid_pass SRCS data_flow_graph_to_fluid_pass_tester.cc)
-inference_analysis_test(test_fluid_to_ir_pass SRCS fluid_to_ir_pass_tester.cc)
-inference_analysis_test(test_fluid_to_data_flow_graph_pass SRCS fluid_to_data_flow_graph_pass_tester.cc)
-inference_analysis_test(test_subgraph_splitter SRCS subgraph_splitter_tester.cc)
-inference_analysis_test(test_dfg_graphviz_draw_pass SRCS dfg_graphviz_draw_pass_tester.cc)
-inference_analysis_test(test_tensorrt_subgraph_pass SRCS tensorrt_subgraph_pass_tester.cc)
-inference_analysis_test(test_pass_manager SRCS pass_manager_tester.cc)
-inference_analysis_test(test_tensorrt_subgraph_node_mark_pass SRCS tensorrt_subgraph_node_mark_pass_tester.cc)
-inference_analysis_test(test_model_store_pass SRCS model_store_pass_tester.cc)

paddle/fluid/inference/analysis/analysis_pass.h

@@ -19,42 +19,36 @@ limitations under the License. */
 #include <string>
 
 #include "paddle/fluid/framework/framework.pb.h"
+#include "paddle/fluid/framework/ir/graph.h"
 #include "paddle/fluid/inference/analysis/argument.h"
-#include "paddle/fluid/inference/analysis/data_flow_graph.h"
 #include "paddle/fluid/inference/analysis/helper.h"
-#include "paddle/fluid/inference/analysis/node.h"
 
 namespace paddle {
 namespace inference {
 namespace analysis {
 
+/*
+ * AnalysisPass is a pass used to control the IR passes.
+ */
 class AnalysisPass {
  public:
   AnalysisPass() = default;
   virtual ~AnalysisPass() = default;
-  // Mutable Pass.
-  virtual bool Initialize(Argument *argument) { return false; }
-  // Readonly Pass.
-  virtual bool Initialize(const Argument &argument) { return false; }
 
-  // Virtual method overriden by subclasses to do any necessary clean up after
-  // all passes have run.
-  virtual bool Finalize() { return false; }
-
-  // Create a debugger Pass that draw the DFG by graphviz toolkit.
-  virtual AnalysisPass *CreateGraphvizDebugerPass() const { return nullptr; }
-
-  // Run on a single DataFlowGraph.
-  virtual void Run(DataFlowGraph *x) = 0;
+  // Run on a single Graph.
+  void Run(Argument* argument) { RunImpl(argument); }
 
   // Human-readable short representation.
   virtual std::string repr() const = 0;
   // Human-readable long description.
   virtual std::string description() const { return "No DOC"; }
-};
 
-// GraphPass processes on any GraphType.
-class DataFlowGraphPass : public AnalysisPass {};
+ protected:
+  // User should implement these.
+  virtual void RunImpl(Argument* argument) = 0;
+
+  Argument* argument_{nullptr};
+};
 
 }  // namespace analysis
 }  // namespace inference

paddle/fluid/inference/analysis/analyzer.cc

@@ -15,138 +15,23 @@
 #include "paddle/fluid/inference/analysis/analyzer.h"
 #include <string>
 #include <vector>
-
-#include "paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.h"
-#include "paddle/fluid/inference/analysis/dfg_graphviz_draw_pass.h"
-#include "paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.h"
-#include "paddle/fluid/inference/analysis/fluid_to_ir_pass.h"
-#include "paddle/fluid/inference/analysis/model_store_pass.h"
-#include "paddle/fluid/inference/analysis/pass_manager.h"
-#include "paddle/fluid/inference/analysis/tensorrt_subgraph_node_mark_pass.h"
-#include "paddle/fluid/inference/analysis/tensorrt_subgraph_pass.h"
-
-DEFINE_bool(IA_enable_tensorrt_subgraph_engine, false,
-            "Enable subgraph to TensorRT engine for acceleration");
-
-DEFINE_bool(IA_enable_ir, false, "Turn on IR support");
-
-DEFINE_string(IA_graphviz_log_root, "./",
-              "Graphviz debuger for data flow graphs.");
-
-DEFINE_string(IA_output_storage_path, "", "optimized model output path");
+#include "paddle/fluid/inference/analysis/passes/ir_analysis_compose_pass.h"
+#include "paddle/fluid/inference/analysis/passes/passes.h"
 
 namespace paddle {
 namespace inference {
 namespace analysis {
 
-class DfgPassManagerImpl final : public DfgPassManager {
- public:
-  DfgPassManagerImpl() {
-    // TODO(Superjomn) set the key with pass reprs.
-    if (!FLAGS_IA_enable_ir) {
-      AddPass("fluid-to-data-flow-graph", new FluidToDataFlowGraphPass);
-    } else {
-      AddPass("fluid-to-ir-pass", new FluidToIrPass);
-    }
-    TryAddTensorRtPass();
-    AddPass("data-flow-graph-to-fluid", new DataFlowGraphToFluidPass);
-    if (!FLAGS_IA_output_storage_path.empty()) {
-      AddPass("model-store-pass", new ModelStorePass);
-    }
-  }
+Analyzer::Analyzer() {}
 
-  std::string repr() const override { return "dfg-pass-manager"; }
-  std::string description() const override { return "DFG pass manager."; }
+void Analyzer::Run(Argument *argument) { RunIrAnalysis(argument); }
 
- private:
-  void AddPass(const std::string& name, AnalysisPass* pass) {
-    VLOG(30) << "Adding pass " << name;
-    Register(name, pass);
-    AddGraphvizDebugerPass(pass);
-  }
+void Analyzer::RunIrAnalysis(Argument *argument) {
+  std::vector<std::string> passes({"ir_analysis_compose_pass"});
 
-  void TryAddTensorRtPass() {
-    if (FLAGS_IA_enable_tensorrt_subgraph_engine) {
-      auto trt_teller = [&](const Node* node) {
-        std::unordered_set<std::string> teller_set(
-            {"mul", "conv2d", "pool2d", "relu", "softmax", "sigmoid",
-             "depthwise_conv2d", "batch_norm", "concat", "tanh", "pad",
-             "elementwise_add", "dropout"});
-        if (!node->IsFunction()) return false;
-
-        const auto* func = static_cast<const Function*>(node);
-        if (teller_set.count(func->func_type())) {
-          return true;
-        } else {
-          return false;
-        }
-      };
-
-      AddPass("tensorrt-subgraph-marker",
-              new TensorRTSubgraphNodeMarkPass(trt_teller));
-      AddPass("tensorrt-subgraph", new TensorRTSubGraphPass(trt_teller));
-    }
-  }
-
-  // Add the graphviz debuger pass if the parent pass has one.
-  void AddGraphvizDebugerPass(AnalysisPass* pass) {
-    auto* debuger_pass = pass->CreateGraphvizDebugerPass();
-    if (debuger_pass) {
-      Register(debuger_pass->repr(), debuger_pass);
-    }
+  for (auto &pass : passes) {
+    PassRegistry::Global().Retreive(pass)->Run(argument);
   }
-};
-
-Analyzer::Analyzer() { Register("manager1", new DfgPassManagerImpl); }
-
-void Analyzer::Run(Argument* argument) {
-  std::vector<std::string> passes;
-  passes.push_back("graph_viz_pass");  // add graphviz for debug.
-#ifdef PADDLE_WITH_MKLDNN
-  if (use_mkldnn_) {
-    VLOG(30) << "Adding MKL-DNN placement pass";
-    passes.push_back("mkldnn_placement_pass");
-  }
-#endif
-  // infer_clean_graph_pass should be the first default pass
-  // after mkldnn_placement_pass.
-  passes.push_back("infer_clean_graph_pass");
-  passes.push_back("graph_viz_pass");  // add graphviz for debug.
-  for (auto& pass : ir_passes_) {
-    // skip mkldnn pass when use_mkldnn_ = false;
-    bool skip_pass = (!use_mkldnn_) && pass.find("mkldnn") != std::string::npos;
-    if (!disabled_ir_passes_.count(pass) && !skip_pass) {
-      passes.push_back(pass);
-      passes.push_back("graph_viz_pass");  // add graphviz for debug.
-    }
-  }
-  argument->Set(kFluidToIrPassesAttr, new std::vector<std::string>(passes));
-
-  for (auto& x : data_) {
-    PADDLE_ENFORCE(x->Initialize(argument));
-    x->RunAll();
-    PADDLE_ENFORCE(x->Finalize());
-  }
-}
-
-Analyzer& Analyzer::IncludeAllIrPasses() {
-  ir_passes_ = all_ir_passes_;
-  return *this;
-}
-
-Analyzer& Analyzer::DisableIrPasses(const std::vector<std::string>& passes) {
-  disabled_ir_passes_.insert(passes.begin(), passes.end());
-  return *this;
-}
-
-Analyzer& Analyzer::IncludeIrPasses(const std::vector<std::string>& passes) {
-  ir_passes_ = passes;
-  return *this;
-}
-
-Analyzer& Analyzer::SetUseMkldnn(bool use_mkldnn) {
-  use_mkldnn_ = use_mkldnn;
-  return *this;
 }
 
 }  // namespace analysis

paddle/fluid/inference/analysis/analyzer.h

@@ -40,56 +40,21 @@ limitations under the License. */
 #include <vector>
 #include "paddle/fluid/inference/analysis/analysis_pass.h"
 #include "paddle/fluid/inference/analysis/flags.h"
-#include "paddle/fluid/inference/analysis/pass_manager.h"
 
 namespace paddle {
 namespace inference {
 namespace analysis {
 
-class Analyzer : public OrderedRegistry<PassManager> {
+class Analyzer final {
  public:
-  // Register all the pass-managers.
   Analyzer();
 
   void Run(Argument* argument);
 
-  Analyzer& DisableIrPasses(const std::vector<std::string>& passes);
-  Analyzer& IncludeIrPasses(const std::vector<std::string>& passes);
-  Analyzer& IncludeAllIrPasses();
-  Analyzer& SetUseMkldnn(bool use_mkldnn);
-
   DISABLE_COPY_AND_ASSIGN(Analyzer);
 
- private:
-  // All avaiable IR passes.
-  // The bigger fuse comes first, so that the small operators prefer to be
-  // merged in a larger fuse op. The small fusion will not break the pattern of
-  // larger fusion.
-  const std::vector<std::string> all_ir_passes_{{
-      // Manual update the passes here.
-      "attention_lstm_fuse_pass",       //
-      "seqconv_eltadd_relu_fuse_pass",  //
-      "embedding_fc_lstm_fuse_pass",    //
-      "fc_lstm_fuse_pass",              //
-      "mul_lstm_fuse_pass",             //
-      "fc_gru_fuse_pass",               //
-      "mul_gru_fuse_pass",              //
-      "seq_concat_fc_fuse_pass",        //
-      "fc_fuse_pass",                   //
-      "conv_bn_fuse_pass",              //
-      "conv_eltwiseadd_bn_fuse_pass",   //
-#ifdef PADDLE_WITH_MKLDNN
-      "depthwise_conv_mkldnn_pass",             //
-      "conv_bias_mkldnn_fuse_pass",             //
-      "conv_relu_mkldnn_fuse_pass",             //
-      "conv_elementwise_add_mkldnn_fuse_pass",  //
-#endif
-  }};
-
-  std::unordered_set<std::string> disabled_ir_passes_;
-  // Ir passes to run
-  std::vector<std::string> ir_passes_;
-  bool use_mkldnn_;
+ protected:
+  void RunIrAnalysis(Argument* argument);
 };
 
 }  // namespace analysis

paddle/fluid/inference/analysis/analyzer_tester.cc

@@ -27,21 +27,21 @@ namespace analysis {
 using namespace framework;  // NOLINT
 
 TEST(Analyzer, analysis_without_tensorrt) {
-  FLAGS_IA_enable_tensorrt_subgraph_engine = false;
   Argument argument;
-  argument.fluid_model_dir.reset(new std::string(FLAGS_inference_model_dir));
+  argument.SetModelDir(FLAGS_inference_model_dir);
+  argument.SetIrAnalysisPasses({"infer_clean_graph_pass"});
+
   Analyzer analyser;
   analyser.Run(&argument);
 }
 
 TEST(Analyzer, analysis_with_tensorrt) {
-  FLAGS_IA_enable_tensorrt_subgraph_engine = true;
   Argument argument;
-  argument.Set<int>("minimum_subgraph_size", new int(0));
-  argument.Set<int>("max_batch_size", new int(3));
-  argument.Set<int>("workspace_size", new int(1 << 20));
-  argument.Set<std::string>("precision_mode", new std::string("FP32"));
-  argument.fluid_model_dir.reset(new std::string(FLAGS_inference_model_dir));
+  argument.SetTensorRtMaxBatchSize(3);
+  argument.SetTensorRtWorkspaceSize(1 << 20);
+  argument.SetModelDir(FLAGS_inference_model_dir);
+  argument.SetIrAnalysisPasses({"infer_clean_graph_pass"});
+
   Analyzer analyser;
   analyser.Run(&argument);
 }

paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.h

-/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
-
-   Licensed under the Apache License, Version 2.0 (the "License");
-   you may not use this file except in compliance with the License.
-   You may obtain a copy of the License at
-
-   http://www.apache.org/licenses/LICENSE-2.0
-
-   Unless required by applicable law or agreed to in writing, software
-   distributed under the License is distributed on an "AS IS" BASIS,
-   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-   See the License for the specific language governing permissions and
-   limitations under the License. */
-
-/*
- * This file implements the transformation from fluid ProgramDesc to data flow
- * graph.
- */
-
-#pragma once
-
-#include <string>
-#include "paddle/fluid/framework/program_desc.h"
-#include "paddle/fluid/inference/analysis/analysis_pass.h"
-#include "paddle/fluid/inference/analysis/data_flow_graph.h"
-
-namespace paddle {
-namespace inference {
-
-namespace analysis {
-class DataFlowGraphToFluidPass final : public DataFlowGraphPass {
- public:
-  DataFlowGraphToFluidPass() = default;
-
-  bool Initialize(Argument *argument) override;
-  bool Finalize() override;
-
-  void Run(DataFlowGraph *graph) override;
-
-  std::string repr() const override { return "DFG to fluid"; }
-  std::string description() const override {
-    return "Transform a DFG to a Fluid ProgramDesc";
-  }
-
-  AnalysisPass *CreateGraphvizDebugerPass() const override;
-
- protected:
-  // Add a Fluid Op into the ProgramDesc.
-  void AddFluidOp(Node *node);
-  // Add a EngineOp into the ProgramDesc.
-  void AddEngineOp(Node *node);
-
- private:
-  framework::proto::ProgramDesc *desc_;
-  Argument *argument_;
-};
-}  // namespace analysis
-}  // namespace inference
-}  // namespace paddle

paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass_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.
-
-#include "paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.h"
-
-#include <glog/logging.h>
-#include <google/protobuf/text_format.h>
-#include <gtest/gtest.h>
-#include "paddle/fluid/framework/executor.h"
-#include "paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.h"
-#include "paddle/fluid/inference/analysis/ut_helper.h"
-#include "paddle/fluid/inference/io.h"
-
-namespace paddle {
-namespace inference {
-namespace analysis {
-
-TEST(DataFlowGraph, Test) {
-  Argument argument(FLAGS_inference_model_dir);
-
-  FluidToDataFlowGraphPass pass0;
-  DataFlowGraphToFluidPass pass1;
-  ASSERT_TRUE(pass0.Initialize(&argument));
-  ASSERT_TRUE(pass1.Initialize(&argument));
-
-  pass0.Run(argument.main_dfg.get());
-  pass1.Run(argument.main_dfg.get());
-
-  pass0.Finalize();
-  pass1.Finalize();
-
-  LOG(INFO) << argument.main_dfg->nodes.size();
-}
-
-};  // namespace analysis
-};  // namespace inference
-};  // namespace paddle

paddle/fluid/inference/analysis/dot_tester.cc

@@ -16,7 +16,6 @@
 
 #include <gtest/gtest.h>
 #include <memory>
-#include "paddle/fluid/inference/analysis/data_flow_graph.h"
 
 namespace paddle {
 namespace inference {

paddle/fluid/inference/analysis/ir_passes/CMakeLists.txt

+cc_library(subgraph_detector SRCS subgraph_detector.cc DEPS proto_desc)
+cc_library(tensorrt_subgraph_pass SRCS tensorrt_subgraph_pass.cc DEPS subgraph_detector)
+set(analysis_deps ${analysis_deps}
+        subgraph_detector tensorrt_subgraph_pass
+        CACHE INTERNAL "")
+
+set(INFER_IR_PASSES ${INFER_IR_PASSES} tensorrt_subgraph_pass CACHE INTERNAL "")