Merge branch 'develop' into develop

6 years ago · a79a3ea2f0
parent cbc7208399 0f085f0a5a
commit a79a3ea2f0
471 changed files with 15446 additions and 3270 deletions
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@ -54,7 +54,7 @@ option(WITH_PYTHON      "Compile PaddlePaddle with python interpreter"  ON)
 option(WITH_DOUBLE      "Compile PaddlePaddle with double precision"    OFF)
 option(WITH_RDMA        "Compile PaddlePaddle with RDMA support"        OFF)
 option(WITH_TIMER       "Compile PaddlePaddle with stats timer"         OFF)
-option(WITH_PROFILER    "Compile PaddlePaddle with GPU profiler"        OFF)
+option(WITH_PROFILER    "Compile PaddlePaddle with GPU profiler and gperftools"        OFF)
 option(WITH_DOC         "Compile PaddlePaddle with documentation"       OFF)
 option(WITH_COVERAGE    "Compile PaddlePaddle with code coverage"       OFF)
 option(COVERALLS_UPLOAD "Package code coverage data to coveralls"       OFF)
@ -132,8 +132,6 @@ if (APPLE OR WIN32)
 endif()

 if (WIN32)
-    set(WITH_AVX OFF CACHE STRING
-            "Disable AVX when compiling for Windows" FORCE)
    set(WITH_DSO OFF CACHE STRING
            "Disable DSO when compiling for Windows" FORCE)
    set(WITH_MKL OFF CACHE STRING
@ -261,6 +259,12 @@ elseif()
    set(WITH_ANAKIN OFF CACHE STRING "Anakin is used in MKL only now." FORCE)
 endif()

+if (WITH_PROFILER)
+    find_package(Gperftools REQUIRED)
+    include_directories(${GPERFTOOLS_INCLUDE_DIR})
+    add_definitions(-DWITH_GPERFTOOLS)
+endif()
+
 include(generic)            # simplify cmake module
 include(package)            # set paddle packages
 include(ccache)             # set ccache for compilation
--- a/README.md
+++ b/README.md
@ -2,8 +2,8 @@


 [![Build Status](https://travis-ci.org/PaddlePaddle/Paddle.svg?branch=develop)](https://travis-ci.org/PaddlePaddle/Paddle)
-[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://paddlepaddle.org/documentation/docs/en/1.1/getstarted/index_en.html)
-[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://paddlepaddle.org/documentation/docs/zh/1.1/beginners_guide/index.html)
+[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://paddlepaddle.org/documentation/docs/en/1.2/getstarted/index_en.html)
+[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/index.html)
 [![Release](https://img.shields.io/github/release/PaddlePaddle/Paddle.svg)](https://github.com/PaddlePaddle/Paddle/releases)
 [![License](https://img.shields.io/badge/license-Apache%202-blue.svg)](LICENSE)

@ -19,7 +19,16 @@ Our vision is to enable deep learning for everyone via PaddlePaddle.
 Please refer to our [release announcement](https://github.com/PaddlePaddle/Paddle/releases) to track the latest feature of PaddlePaddle.


-### Latest PaddlePaddle Release: [Fluid 1.1.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.1)
+欢迎来到 PaddlePaddle GitHub
+
+PaddlePaddle (PArallel Distributed Deep LEarning) 是一个简单易用、高效灵活、可扩展的深度学习平台，最初由百度科学家和工程师共同开发，目的是将深度学习技术应用到百度的众多产品中。
+
+我们的愿景是让每个人都能通过PaddlePaddle接触深度学习
+
+跟进PaddlePaddle最新特性请参考我们的[版本说明](https://github.com/PaddlePaddle/Paddle/releases)
+
+
+### Latest PaddlePaddle Release: [Fluid 1.2.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.2)
 ### Install Latest Stable Release:
 ```
 # Linux CPU
@ -27,13 +36,30 @@ pip install paddlepaddle
 # Linux GPU cuda9cudnn7
 pip install paddlepaddle-gpu
 # Linux GPU cuda8cudnn7
-pip install paddlepaddle-gpu==1.1.0.post87
+pip install paddlepaddle-gpu==1.2.0.post87
 # Linux GPU cuda8cudnn5
-pip install paddlepaddle-gpu==1.1.0.post85
+pip install paddlepaddle-gpu==1.2.0.post85

 # For installation on other platform, refer to http://paddlepaddle.org/
 ```

+
+### PaddlePaddle最新版本: [Fluid 1.2.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.2)
+### 安装最新稳定版本:
+```
+# Linux CPU
+pip install paddlepaddle
+# Linux GPU cuda9cudnn7
+pip install paddlepaddle-gpu
+# Linux GPU cuda8cudnn7
+pip install paddlepaddle-gpu==1.2.0.post87
+# Linux GPU cuda8cudnn5
+pip install paddlepaddle-gpu==1.2.0.post85
+
+# 其他平台上的安装指引请参考 http://paddlepaddle.org/
+```
+
+
 ## Features

 - **Flexibility**
@ -74,35 +100,90 @@ pip install paddlepaddle-gpu==1.1.0.post85
    Baidu and it has achieved a significant impact. We hope you can also explore
    the capability of PaddlePaddle to make an impact on your product.

+## 特点
+
+- **灵活性**
+
+    PaddlePaddle支持丰富的神经网络架构和优化算法。易于配置复杂模型，例如带有注意力机制或复杂记忆连接的神经网络机器翻译模型。
+
+-  **高效性**
+
+    为了高效使用异步计算资源，PaddlePaddle对框架的不同层进行优化，包括计算、存储、架构和通信。下面是一些样例：
+    
+    - 通过SSE/AVX 内置函数、BLAS库(例如MKL、OpenBLAS、cuBLAS)或定制的CPU/GPU内核优化数学操作。
+    - 通过MKL-DNN库优化CNN网络
+    - 高度优化循环网络，无需执行 `padding` 操作即可处理 **变长** 序列
+    - 针对高维稀疏数据模型，优化了局部和分布式训练。
+     
+
+- **稳定性**
+
+    有了 PaddlePaddle，使得利用各种CPU/GPU和机器来加速训练变得简单。PaddlePaddle 通过优化通信可以实现巨大吞吐量和快速执行。
+
+- **连接产品**
+
+    另外，PaddlePaddle 的设计也易于部署。在百度，PaddlePaddle 已经部署到含有巨大用户量的产品和服务上，包括广告点击率（CTR）预测、大规模图像分类、光学字符识别（OCR）、搜索排序，计算机病毒检测、推荐系统等等。PaddlePaddle广泛应用于百度产品中，产生了非常重要的影响。我们希望您也能探索 PaddlePaddle 的能力，为您的产品创造新的影响力和效果。
+
 ## Installation

-It is recommended to read [this doc](http://paddlepaddle.org/documentation/docs/zh/1.1/beginners_guide/index.html) on our website.
+It is recommended to read [this doc](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/install/index_cn.html) on our website.
+
+## 安装
+
+推荐阅读官网上的[安装说明](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/install/index_cn.html) 

 ## Documentation

-We provide [English](http://paddlepaddle.org/documentation/docs/en/1.1/getstarted/index_en.html) and
-[Chinese](http://paddlepaddle.org/documentation/docs/zh/1.1/beginners_guide/index.html) documentation.
+We provide [English](http://paddlepaddle.org/documentation/docs/en/1.2/getstarted/index_en.html) and
+[Chinese](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/index.html) documentation.

 - [Deep Learning 101](https://github.com/PaddlePaddle/book)

  You might want to start from this online interactive book that can run in a Jupyter Notebook.

- [Distributed Training](http://paddlepaddle.org/documentation/docs/zh/1.1/user_guides/howto/training/cluster_howto.html)
+- [Distributed Training](http://paddlepaddle.org/documentation/docs/zh/1.2/user_guides/howto/training/cluster_howto.html)

  You can run distributed training jobs on MPI clusters.

- [Python API](http://paddlepaddle.org/documentation/api/zh/1.1/fluid.html)
+- [Python API](http://paddlepaddle.org/documentation/docs/zh/1.2/api_cn/index_cn.html)

   Our new API enables much shorter programs.

- [How to Contribute](http://paddlepaddle.org/documentation/docs/zh/1.1/advanced_usage/development/contribute_to_paddle.html)
+- [How to Contribute](http://paddlepaddle.org/documentation/docs/zh/1.2/advanced_usage/development/contribute_to_paddle/index_cn.html)

   We appreciate your contributions!

+## 文档
+
+我们提供[英文](http://paddlepaddle.org/documentation/docs/en/1.2/getstarted/index_en.html)和
+[中文](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/index.html) 文档
+
+- [深度学习101](https://github.com/PaddlePaddle/book)
+
+  或许您想从这个在线交互式书籍开始，可以在Jupyter Notebook中运行
+
+- [分布式训练](http://paddlepaddle.org/documentation/docs/zh/1.2/user_guides/howto/training/cluster_howto.html)
+
+  可以在MPI集群上运行分布式训练任务
+
+- [Python API](http://paddlepaddle.org/documentation/docs/zh/1.2/api_cn/index_cn.html)
+
+   新的API支持代码更少更简洁的程序
+
+- [贡献方式](http://paddlepaddle.org/documentation/docs/zh/1.2/advanced_usage/development/contribute_to_paddle/index_cn.html)
+
+   欢迎您的贡献!

 ## Ask Questions

 You are welcome to submit questions and bug reports as [Github Issues](https://github.com/PaddlePaddle/Paddle/issues).

+## 答疑
+
+欢迎您将问题和bug报告以[Github Issues](https://github.com/PaddlePaddle/Paddle/issues)的形式提交
+
 ## Copyright and License
 PaddlePaddle is provided under the [Apache-2.0 license](LICENSE).
+
+## 版权和许可证
+PaddlePaddle由[Apache-2.0 license](LICENSE)提供
--- a/benchmark/fluid/fluid_benchmark.py
+++ b/benchmark/fluid/fluid_benchmark.py
@ -81,9 +81,11 @@ def dist_transpile(trainer_id, args, train_prog, startup_prog):
    # the role, should be either PSERVER or TRAINER
    training_role = os.getenv("PADDLE_TRAINING_ROLE")

-    config = distribute_transpiler.DistributeTranspilerConfig()
+    config = fluid.DistributeTranspilerConfig()
    config.slice_var_up = not args.no_split_var
+    config.min_block_size = 1048576
    t = distribute_transpiler.DistributeTranspiler(config=config)
+
    t.transpile(
        trainer_id,
        # NOTE: *MUST* use train_prog, for we are using with guard to
--- a/cmake/FindGperftools.cmake
+++ b/cmake/FindGperftools.cmake
@ -0,0 +1,63 @@
+# Tries to find Gperftools.
+#
+# Usage of this module as follows:
+#
+#     find_package(Gperftools)
+#
+# Variables used by this module, they can change the default behaviour and need
+# to be set before calling find_package:
+#
+#  Gperftools_ROOT_DIR  Set this variable to the root installation of
+#                       Gperftools if the module has problems finding
+#                       the proper installation path.
+#
+# Variables defined by this module:
+#
+#  GPERFTOOLS_FOUND              System has Gperftools libs/headers
+#  GPERFTOOLS_LIBRARIES          The Gperftools libraries (tcmalloc & profiler)
+#  GPERFTOOLS_INCLUDE_DIR        The location of Gperftools headers
+
+find_library(GPERFTOOLS_TCMALLOC
+  NAMES tcmalloc
+  HINTS ${Gperftools_ROOT_DIR}/lib)
+
+find_library(GPERFTOOLS_PROFILER
+  NAMES profiler
+  HINTS ${Gperftools_ROOT_DIR}/lib)
+
+find_library(GPERFTOOLS_TCMALLOC_AND_PROFILER
+  NAMES tcmalloc_and_profiler
+  HINTS ${Gperftools_ROOT_DIR}/lib)
+
+find_path(GPERFTOOLS_INCLUDE_DIR
+  NAMES gperftools/heap-profiler.h
+  HINTS ${Gperftools_ROOT_DIR}/include)
+
+set(GPERFTOOLS_LIBRARIES ${GPERFTOOLS_TCMALLOC_AND_PROFILER})
+
+include(FindPackageHandleStandardArgs)
+find_package_handle_standard_args(
+  Gperftools
+  DEFAULT_MSG
+  GPERFTOOLS_LIBRARIES
+  GPERFTOOLS_INCLUDE_DIR)
+
+mark_as_advanced(
+  Gperftools_ROOT_DIR
+  GPERFTOOLS_TCMALLOC
+  GPERFTOOLS_PROFILER
+  GPERFTOOLS_TCMALLOC_AND_PROFILER
+  GPERFTOOLS_LIBRARIES
+  GPERFTOOLS_INCLUDE_DIR)
+
+# create IMPORTED targets
+if (Gperftools_FOUND AND NOT TARGET gperftools::tcmalloc)
+  add_library(gperftools::tcmalloc UNKNOWN IMPORTED)
+  set_target_properties(gperftools::tcmalloc PROPERTIES
+    IMPORTED_LOCATION ${GPERFTOOLS_TCMALLOC}
+    INTERFACE_INCLUDE_DIRECTORIES "${GPERFTOOLS_INCLUDE_DIR}")
+  add_library(gperftools::profiler UNKNOWN IMPORTED)
+  set_target_properties(gperftools::profiler PROPERTIES
+    IMPORTED_LOCATION ${GPERFTOOLS_PROFILER}
+    INTERFACE_INCLUDE_DIRECTORIES "${GPERFTOOLS_INCLUDE_DIR}")
+endif()
--- a/cmake/configure.cmake
+++ b/cmake/configure.cmake
@ -90,6 +90,7 @@ endif()

 if(WITH_GPU)
    add_definitions(-DPADDLE_WITH_CUDA)
+    add_definitions(-DEIGEN_USE_GPU)

    FIND_PACKAGE(CUDA REQUIRED)

--- a/cmake/external/brpc.cmake
+++ b/cmake/external/brpc.cmake
@ -14,14 +14,16 @@

 INCLUDE(ExternalProject)

-find_library(SSL_LIBRARY NAMES ssl)
+find_package(OpenSSL REQUIRED) 
+
+message(STATUS "ssl:" ${OPENSSL_SSL_LIBRARY})
+message(STATUS "crypto:" ${OPENSSL_CRYPTO_LIBRARY})
+
 ADD_LIBRARY(ssl SHARED IMPORTED GLOBAL)
-SET_PROPERTY(TARGET ssl PROPERTY IMPORTED_LOCATION ${SSL_LIBRARY})
+SET_PROPERTY(TARGET ssl PROPERTY IMPORTED_LOCATION ${OPENSSL_SSL_LIBRARY})

-find_library(CRYPTO_LIBRARY NAMES crypto)
 ADD_LIBRARY(crypto SHARED IMPORTED GLOBAL)
-SET_PROPERTY(TARGET crypto PROPERTY IMPORTED_LOCATION ${CRYPTO_LIBRARY})
-
+SET_PROPERTY(TARGET crypto PROPERTY IMPORTED_LOCATION ${OPENSSL_CRYPTO_LIBRARY})

 SET(BRPC_SOURCES_DIR ${THIRD_PARTY_PATH}/brpc)
 SET(BRPC_INSTALL_DIR ${THIRD_PARTY_PATH}/install/brpc)
@ -31,14 +33,15 @@ SET(BRPC_LIBRARIES "${BRPC_INSTALL_DIR}/lib/libbrpc.a" CACHE FILEPATH "brpc libr
 INCLUDE_DIRECTORIES(${BRPC_INCLUDE_DIR})

 # Reference https://stackoverflow.com/questions/45414507/pass-a-list-of-prefix-paths-to-externalproject-add-in-cmake-args
-set(prefix_path "${THIRD_PARTY_PATH}/install/gflags|${THIRD_PARTY_PATH}/install/leveldb|${THIRD_PARTY_PATH}/install/snappy|${THIRD_PARTY_PATH}/install/gtest|${THIRD_PARTY_PATH}/install/protobuf|${THIRD_PARTY_PATH}/install/zlib")
+set(prefix_path "${THIRD_PARTY_PATH}/install/gflags|${THIRD_PARTY_PATH}/install/leveldb|${THIRD_PARTY_PATH}/install/snappy|${THIRD_PARTY_PATH}/install/gtest|${THIRD_PARTY_PATH}/install/protobuf|${THIRD_PARTY_PATH}/install/zlib|${THIRD_PARTY_PATH}/install/glog")

 # If minimal .a is need, you can set  WITH_DEBUG_SYMBOLS=OFF
 ExternalProject_Add(
    extern_brpc
    ${EXTERNAL_PROJECT_LOG_ARGS}
+    # TODO(gongwb): change to de newst repo when they changed.
    GIT_REPOSITORY  "https://github.com/gongweibao/brpc"
-    GIT_TAG         "7dc04defad1fd4173aae170c3fcbde131b65155a"
+    GIT_TAG         "e9b67ec1b7458f2af5fae76451afe1e27e01b4b4"
    PREFIX          ${BRPC_SOURCES_DIR}
    UPDATE_COMMAND  ""
    CMAKE_ARGS      -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
@ -50,7 +53,7 @@ ExternalProject_Add(
                    -DCMAKE_POSITION_INDEPENDENT_CODE=ON
                    -DCMAKE_BUILD_TYPE=${THIRD_PARTY_BUILD_TYPE}
                    -DCMAKE_PREFIX_PATH=${prefix_path}
-                    -DBRPC_WITH_GLOG=ON
+                    -DWITH_GLOG=ON
                    -DIOBUF_WITH_HUGE_BLOCK=ON
                    -DBRPC_WITH_RDMA=${WITH_BRPC_RDMA}
                    ${EXTERNAL_OPTIONAL_ARGS}
@ -65,5 +68,6 @@ ADD_LIBRARY(brpc STATIC IMPORTED GLOBAL)
 SET_PROPERTY(TARGET brpc PROPERTY IMPORTED_LOCATION ${BRPC_LIBRARIES})
 ADD_DEPENDENCIES(brpc extern_brpc)

+add_definitions(-DBRPC_WITH_GLOG)

 LIST(APPEND external_project_dependencies brpc)
--- a/cmake/external/gtest.cmake
+++ b/cmake/external/gtest.cmake
@ -12,8 +12,12 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.

-IF(WITH_TESTING)
-    ENABLE_TESTING()
+#FIXME:(gongwb) Move brpc's gtest dependency.
+IF(WITH_TESTING OR (WITH_DISTRIBUTE AND NOT WITH_GRPC))
+    IF(WITH_TESTING)
+        ENABLE_TESTING()
+    ENDIF(WITH_TESTING)
+
    INCLUDE(ExternalProject)

    SET(GTEST_SOURCES_DIR ${THIRD_PARTY_PATH}/gtest)
@ -76,4 +80,4 @@ IF(WITH_TESTING)
    ADD_DEPENDENCIES(gtest_main extern_gtest)

    LIST(APPEND external_project_dependencies gtest gtest_main)
-ENDIF(WITH_TESTING)
+ENDIF(WITH_TESTING OR (WITH_DISTRIBUTE AND NOT WITH_GRPC))
--- a/cmake/external/leveldb.cmake
+++ b/cmake/external/leveldb.cmake
@ -24,8 +24,8 @@ ExternalProject_Add(
    extern_leveldb
    ${EXTERNAL_PROJECT_LOG_ARGS}
    PREFIX ${LEVELDB_SOURCES_DIR}
-    URL "https://github.com/google/leveldb/archive/v1.18.tar.gz"
-    URL_MD5 "73770de34a2a5ab34498d2e05b2b7fa0"
+    GIT_REPOSITORY "https://github.com/google/leveldb"
+    GIT_TAG v1.18
    CONFIGURE_COMMAND ""
    BUILD_COMMAND CXXFLAGS=-fPIC make -j ${NUM_OF_PROCESSOR} libleveldb.a
    INSTALL_COMMAND mkdir -p ${LEVELDB_INSTALL_DIR}/lib/ 
--- a/cmake/external/ngraph.cmake
+++ b/cmake/external/ngraph.cmake
@ -32,6 +32,8 @@ IF(NOT ${WITH_NGRAPH})
    return()
 ENDIF()

+INCLUDE(GNUInstallDirs)
+
 INCLUDE(ExternalProject)

 SET(NGRAPH_PROJECT         "extern_ngraph")
@ -40,10 +42,14 @@ 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_LIB_DIR         ${NGRAPH_INSTALL_DIR}/${CMAKE_INSTALL_LIBDIR})
 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")
+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})

 ExternalProject_Add(
    ${NGRAPH_PROJECT}
@ -63,18 +69,6 @@ ExternalProject_Add(
    CMAKE_ARGS          -DMKLDNN_LIB_DIR=${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}
--- a/cmake/external/snappy.cmake
+++ b/cmake/external/snappy.cmake
@ -24,12 +24,6 @@ set(SNAPPY_SOURCES_DIR ${THIRD_PARTY_PATH}/snappy)
 set(SNAPPY_INSTALL_DIR ${THIRD_PARTY_PATH}/install/snappy)
 set(SNAPPY_INCLUDE_DIR "${SNAPPY_INSTALL_DIR}/include" CACHE PATH "snappy include directory." FORCE)

-if (WIN32)
-    set(SNAPPY_LIBRARIES "${SNAPPY_INSTALL_DIR}/lib/snappy.lib")
-else(WIN32)
-    set(SNAPPY_LIBRARIES "${SNAPPY_INSTALL_DIR}/lib/libsnappy.a")
-endif (WIN32)
-
 ExternalProject_Add(
    extern_snappy
    GIT_REPOSITORY "https://github.com/google/snappy"
@ -56,6 +50,16 @@ ExternalProject_Add(
                     -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON
                     -DCMAKE_BUILD_TYPE:STRING=${THIRD_PARTY_BUILD_TYPE}
 )
+IF(WIN32)
+    IF(NOT EXISTS "${SNAPPY_INSTALL_DIR}/lib/libsnappy.lib")
+        add_custom_command(TARGET extern_snappy POST_BUILD
+                COMMAND cmake -E copy ${SNAPPY_INSTALL_DIR}/lib/snappy.lib ${SNAPPY_INSTALL_DIR}/lib/libsnappy.lib
+                )
+    ENDIF()
+    set(SNAPPY_LIBRARIES "${SNAPPY_INSTALL_DIR}/lib/libsnappy.lib")
+else(WIN32)
+    set(SNAPPY_LIBRARIES "${SNAPPY_INSTALL_DIR}/lib/libsnappy.a")
+endif (WIN32)

 add_library(snappy STATIC IMPORTED GLOBAL)
 set_property(TARGET snappy PROPERTY IMPORTED_LOCATION ${SNAPPY_LIBRARIES})
--- a/cmake/external/xxhash.cmake
+++ b/cmake/external/xxhash.cmake
@ -56,7 +56,12 @@ else()
 endif()

 if (WIN32)
-  set(XXHASH_LIBRARIES "${XXHASH_INSTALL_DIR}/lib/xxhash.lib")
+  IF(NOT EXISTS "${XXHASH_INSTALL_DIR}/lib/libxxhash.lib")
+    add_custom_command(TARGET extern_xxhash POST_BUILD
+            COMMAND cmake -E copy ${XXHASH_INSTALL_DIR}/lib/xxhash.lib ${XXHASH_INSTALL_DIR}/lib/libxxhash.lib
+            )
+  ENDIF()
+  set(XXHASH_LIBRARIES "${XXHASH_INSTALL_DIR}/lib/libxxhash.lib")
 else()
  set(XXHASH_LIBRARIES "${XXHASH_INSTALL_DIR}/lib/libxxhash.a")
 endif ()
--- a/cmake/external/zlib.cmake
+++ b/cmake/external/zlib.cmake
@ -19,12 +19,6 @@ SET(ZLIB_INSTALL_DIR ${THIRD_PARTY_PATH}/install/zlib)
 SET(ZLIB_ROOT ${ZLIB_INSTALL_DIR} CACHE FILEPATH "zlib root directory." FORCE)
 SET(ZLIB_INCLUDE_DIR "${ZLIB_INSTALL_DIR}/include" CACHE PATH "zlib include directory." FORCE)

-IF(WIN32)
-  SET(ZLIB_LIBRARIES "${ZLIB_INSTALL_DIR}/lib/zlibstatic.lib" CACHE FILEPATH "zlib library." FORCE)
-ELSE(WIN32)
-  SET(ZLIB_LIBRARIES "${ZLIB_INSTALL_DIR}/lib/libz.a" CACHE FILEPATH "zlib library." FORCE)
-ENDIF(WIN32)
-
 INCLUDE_DIRECTORIES(${ZLIB_INCLUDE_DIR}) # For zlib code to include its own headers.
 INCLUDE_DIRECTORIES(${THIRD_PARTY_PATH}/install) # For Paddle code to include zlib.h.

@ -49,6 +43,16 @@ ExternalProject_Add(
                     -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON
                     -DCMAKE_BUILD_TYPE:STRING=${THIRD_PARTY_BUILD_TYPE}
 )
+IF(WIN32)
+  IF(NOT EXISTS "${ZLIB_INSTALL_DIR}/lib/libz.lib")
+    add_custom_command(TARGET extern_zlib POST_BUILD
+            COMMAND cmake -E copy ${ZLIB_INSTALL_DIR}/lib/zlibstatic.lib ${ZLIB_INSTALL_DIR}/lib/libz.lib
+            )
+  ENDIF()
+  SET(ZLIB_LIBRARIES "${ZLIB_INSTALL_DIR}/lib/libz.lib" CACHE FILEPATH "zlib library." FORCE)
+ELSE(WIN32)
+  SET(ZLIB_LIBRARIES "${ZLIB_INSTALL_DIR}/lib/libz.a" CACHE FILEPATH "zlib library." FORCE)
+ENDIF(WIN32)

 ADD_LIBRARY(zlib STATIC IMPORTED GLOBAL)
 SET_PROPERTY(TARGET zlib PROPERTY IMPORTED_LOCATION ${ZLIB_LIBRARIES})
--- a/cmake/generic.cmake
+++ b/cmake/generic.cmake
@ -110,6 +110,14 @@ function(find_fluid_modules TARGET_NAME)
  endif()
 endfunction(find_fluid_modules)

+
+function(common_link TARGET_NAME)
+  if (WITH_PROFILER)
+    target_link_libraries(${TARGET_NAME} gperftools::profiler)
+  endif()
+endfunction()
+
+
 # find all third_party modules is used for paddle static library
 # for reduce the dependency when building the inference libs.
 set_property(GLOBAL PROPERTY FLUID_THIRD_PARTY)
@ -274,6 +282,7 @@ function(cc_library TARGET_NAME)
      endif()
      target_link_libraries(${TARGET_NAME} ${cc_library_DEPS})
      add_dependencies(${TARGET_NAME} ${cc_library_DEPS})
+      common_link(${TARGET_NAME})
    endif()

    # cpplint code style
@ -340,6 +349,7 @@ function(cc_binary TARGET_NAME)
  if(cc_binary_DEPS)
    target_link_libraries(${TARGET_NAME} ${cc_binary_DEPS})
    add_dependencies(${TARGET_NAME} ${cc_binary_DEPS})
+    common_link(${TARGET_NAME})
  endif()
 endfunction(cc_binary)

@ -362,6 +372,7 @@ function(cc_test TARGET_NAME)
      target_link_libraries(${TARGET_NAME} ${win32_deps})
    endif(WIN32)
    add_dependencies(${TARGET_NAME} ${cc_test_DEPS} paddle_gtest_main lod_tensor memory gtest gflags glog)
+    common_link(${TARGET_NAME})
    add_test(NAME ${TARGET_NAME}
             COMMAND ${TARGET_NAME} ${cc_test_ARGS}
             WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
@ -420,6 +431,7 @@ function(nv_binary TARGET_NAME)
    if(nv_binary_DEPS)
      target_link_libraries(${TARGET_NAME} ${nv_binary_DEPS})
      add_dependencies(${TARGET_NAME} ${nv_binary_DEPS})
+      common_link(${TARGET_NAME})
    endif()
  endif()
 endfunction(nv_binary)
@ -433,6 +445,7 @@ function(nv_test TARGET_NAME)
    cuda_add_executable(${TARGET_NAME} ${nv_test_SRCS})
    target_link_libraries(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main lod_tensor memory gtest gflags glog)
    add_dependencies(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main lod_tensor memory gtest gflags glog)
+    common_link(${TARGET_NAME})
    add_test(${TARGET_NAME} ${TARGET_NAME})
    if (nv_test_SERIAL)
        set_property(TEST ${TARGET_NAME} PROPERTY RUN_SERIAL 1)
@ -499,6 +512,7 @@ function(hip_binary TARGET_NAME)
    if(hip_binary_DEPS)
      target_link_libraries(${TARGET_NAME} ${hip_binary_DEPS})
      add_dependencies(${TARGET_NAME} ${hip_binary_DEPS})
+      common_link(${TARGET_NAME})
    endif()
  endif()
 endfunction(hip_binary)
@ -518,6 +532,7 @@ function(hip_test TARGET_NAME)
    set_target_properties(${TARGET_NAME} PROPERTIES LINKER_LANGUAGE HIP)
    target_link_libraries(${TARGET_NAME} ${hip_test_DEPS} paddle_gtest_main memory gtest gflags)
    add_dependencies(${TARGET_NAME} ${hip_test_DEPS} paddle_gtest_main memory gtest gflags)
+    common_link(${TARGET_NAME})
    add_test(${TARGET_NAME} ${TARGET_NAME})
  endif()
 endfunction(hip_test)
@ -560,6 +575,7 @@ function(go_library TARGET_NAME)
  endif()
  if(go_library_DEPS)
    add_dependencies(${TARGET_NAME} ${go_library_DEPS})
+    common_link(${TARGET_NAME})
  endif(go_library_DEPS)

  # The "source file" of the library is `${dummyfile}` which never
--- a/cmake/inference_lib.cmake
+++ b/cmake/inference_lib.cmake
@ -32,24 +32,35 @@ function(copy TARGET)
        list(GET copy_lib_SRCS ${index} src)
        list(GET copy_lib_DSTS ${index} dst)
        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})
+            if(IS_DIRECTORY ${src})
+                get_filename_component(last_path ${src} NAME)
+                string(APPEND dst "/" ${last_path})
+                add_custom_command(TARGET ${TARGET} PRE_BUILD
+                        COMMAND ${CMAKE_COMMAND} -E make_directory "${dst}"
+                        )
+                if(EXISTS ${src})
+                    add_custom_command(TARGET ${TARGET} PRE_BUILD
+                            COMMAND cmake -E copy_directory "${src}" "${dst}"
+                            COMMENT "copying ${src} -> ${dst}")
+                else()
+                    message(WARNING "${src} not exist!")
+                endif()
+            else()
+                # windows cmd shell will not expand wildcard automatically.
+                # below expand the files, and copy them by rules.
+                file(GLOB src_files ${src})
+                if (NOT "${src_files}" STREQUAL "")
+                    list(REMOVE_DUPLICATES src_files)
+                endif ()
                add_custom_command(TARGET ${TARGET} PRE_BUILD
-                        COMMAND ${CMAKE_COMMAND} -E copy "${src_file}" "${dst}"
-                        COMMENT "copying ${src_file} -> ${dst}")
-            endforeach ()
+                        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 ()
+            endif()
        else (WIN32) # not windows
            add_custom_command(TARGET ${TARGET} PRE_BUILD
                    COMMAND mkdir -p "${dst}"
@ -95,7 +106,7 @@ copy(xxhash_lib
        DEPS xxhash
        )

-if (NOT PROTOBUF_FOUND)
+if (NOT PROTOBUF_FOUND OR WIN32)
    set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/protobuf")
    copy(protobuf_lib
            SRCS ${PROTOBUF_INCLUDE_DIR} ${PROTOBUF_LIBRARY}
@ -129,27 +140,34 @@ if (WITH_MKLDNN)
            )
 endif ()

-if (NOT WIN32)
-    if (NOT MOBILE_INFERENCE AND NOT RPI)
-        set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/snappy")
-        copy(snappy_lib
-                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} ${dst_dir}/lib
-                DEPS zlib)
-    endif ()
-endif (NOT WIN32)
+if (WITH_NGRAPH)
+    set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/ngraph")
+    copy(ngraph_lib
+            SRCS ${NGRAPH_INC_DIR} ${NGRAPH_LIB_DIR}
+            DSTS ${dst_dir} ${dst_dir}
+            DEPS ngraph
+            )
+endif ()
+
+if (NOT MOBILE_INFERENCE AND NOT RPI)
+    set(dst_dir "${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 ()

 # paddle fluid module
 set(src_dir "${PADDLE_SOURCE_DIR}/paddle/fluid")
@ -183,8 +201,13 @@ if (WITH_ANAKIN AND WITH_MKL)
 endif ()

 set(module "inference")
+if(WIN32)
+    set(paddle_fluid_lib ${PADDLE_BINARY_DIR}/paddle/fluid/inference/${CMAKE_BUILD_TYPE}/libpaddle_fluid.*)
+else(WIN32)
+    set(paddle_fluid_lib ${PADDLE_BINARY_DIR}/paddle/fluid/inference/libpaddle_fluid.*)
+endif(WIN32)
 copy(inference_lib DEPS ${inference_deps}
-  SRCS ${src_dir}/${module}/*.h ${PADDLE_BINARY_DIR}/paddle/fluid/inference/libpaddle_fluid.*
+  SRCS ${src_dir}/${module}/*.h ${paddle_fluid_lib}
       ${src_dir}/${module}/api/paddle_*.h
  DSTS ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module}
        )
@ -224,7 +247,7 @@ copy(third_party DEPS fluid_lib_dist

 # 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.*
+  SRCS ${paddle_fluid_lib}
       ${FLUID_INSTALL_DIR}/paddle/fluid/inference/paddle_*.h
  DSTS ${FLUID_INFERENCE_INSTALL_DIR}/paddle/lib ${FLUID_INFERENCE_INSTALL_DIR}/paddle/include
 )
--- a/cmake/operators.cmake
+++ b/cmake/operators.cmake
@ -166,6 +166,8 @@ function(op_library TARGET)
      # Append first implemented MKLDNN activation operator
      if (${MKLDNN_FILE} STREQUAL "activation_mkldnn_op")
        file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(relu, MKLDNN);\n")
+      elseif(${MKLDNN_FILE} STREQUAL "conv_mkldnn_op")
+        file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL_WITH_CUSTOM_TYPE(conv2d, MKLDNN, FP32);\n")
      else()
        file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, MKLDNN);\n")
      endif()
--- a/paddle/fluid/API.spec
+++ b/paddle/fluid/API.spec
@ -74,6 +74,7 @@ paddle.fluid.layers.linear_chain_crf ArgSpec(args=['input', 'label', 'param_attr
 paddle.fluid.layers.crf_decoding ArgSpec(args=['input', 'param_attr', 'label'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.cos_sim ArgSpec(args=['X', 'Y'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.cross_entropy ArgSpec(args=['input', 'label', 'soft_label', 'ignore_index'], varargs=None, keywords=None, defaults=(False, -100))
+paddle.fluid.layers.bpr_loss ArgSpec(args=['input', 'label', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.square_error_cost ArgSpec(args=['input', 'label'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.chunk_eval ArgSpec(args=['input', 'label', 'chunk_scheme', 'num_chunk_types', 'excluded_chunk_types'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.sequence_conv ArgSpec(args=['input', 'num_filters', 'filter_size', 'filter_stride', 'padding', 'bias_attr', 'param_attr', 'act', 'name'], varargs=None, keywords=None, defaults=(3, 1, None, None, None, None, None))
@ -84,6 +85,8 @@ paddle.fluid.layers.sequence_softmax ArgSpec(args=['input', 'use_cudnn', 'name']
 paddle.fluid.layers.softmax ArgSpec(args=['input', 'use_cudnn', 'name'], varargs=None, keywords=None, defaults=(True, None))
 paddle.fluid.layers.pool2d ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name', 'exclusive'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None, True))
 paddle.fluid.layers.pool3d ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name', 'exclusive'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None, True))
+paddle.fluid.layers.adaptive_pool2d ArgSpec(args=['input', 'pool_size', 'pool_type', 'require_index', 'name'], varargs=None, keywords=None, defaults=('max', False, None))
+paddle.fluid.layers.adaptive_pool3d ArgSpec(args=['input', 'pool_size', 'pool_type', 'require_index', 'name'], varargs=None, keywords=None, defaults=('max', False, None))
 paddle.fluid.layers.batch_norm ArgSpec(args=['input', 'act', 'is_test', 'momentum', 'epsilon', 'param_attr', 'bias_attr', 'data_layout', 'in_place', 'name', 'moving_mean_name', 'moving_variance_name', 'do_model_average_for_mean_and_var', 'fuse_with_relu', 'use_global_stats'], varargs=None, keywords=None, defaults=(None, False, 0.9, 1e-05, None, None, 'NCHW', False, None, None, None, False, False, False))
 paddle.fluid.layers.beam_search_decode ArgSpec(args=['ids', 'scores', 'beam_size', 'end_id', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.conv2d_transpose ArgSpec(args=['input', 'num_filters', 'output_size', 'filter_size', 'padding', 'stride', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, None, 0, 1, 1, None, None, None, True, None, None))
@ -190,7 +193,7 @@ paddle.fluid.layers.clip ArgSpec(args=['x', 'min', 'max', 'name'], varargs=None,
 paddle.fluid.layers.clip_by_norm ArgSpec(args=['x', 'max_norm', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.mean ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.mul ArgSpec(args=['x', 'y', 'x_num_col_dims', 'y_num_col_dims', 'name'], varargs=None, keywords=None, defaults=(1, 1, None))
-paddle.fluid.layers.sigmoid_cross_entropy_with_logits ArgSpec(args=['x', 'label', 'name'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.layers.sigmoid_cross_entropy_with_logits ArgSpec(args=['x', 'label', 'ignore_index', 'name'], varargs=None, keywords=None, defaults=(-100, None))
 paddle.fluid.layers.maxout ArgSpec(args=['x', 'groups', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.space_to_depth ArgSpec(args=['x', 'blocksize', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.affine_grid ArgSpec(args=['theta', 'out_shape', 'name'], varargs=None, keywords=None, defaults=(None,))
@ -202,6 +205,10 @@ paddle.fluid.layers.grid_sampler ArgSpec(args=['x', 'grid', 'name'], varargs=Non
 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.merge_selected_rows ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.layers.get_tensor_from_selected_rows ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.layers.lstm ArgSpec(args=['input', 'init_h', 'init_c', 'max_len', 'hidden_size', 'num_layers', 'dropout_prob', 'is_bidirec', 'is_test', 'name', 'default_initializer', 'seed'], varargs=None, keywords=None, defaults=(0.0, False, False, None, None, -1))
+paddle.fluid.layers.psroi_pool ArgSpec(args=['input', 'rois', 'output_channels', 'spatial_scale', 'pooled_height', 'pooled_width', 'name'], varargs=None, keywords=None, defaults=(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)
@ -306,6 +313,7 @@ paddle.fluid.layers.generate_proposals ArgSpec(args=['scores', 'bbox_deltas', 'i
 paddle.fluid.layers.iou_similarity ArgSpec(args=['x', 'y', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.box_coder ArgSpec(args=['prior_box', 'prior_box_var', 'target_box', 'code_type', 'box_normalized', 'name'], varargs=None, keywords=None, defaults=('encode_center_size', True, None))
 paddle.fluid.layers.polygon_box_transform ArgSpec(args=['input', 'name'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.layers.yolov3_loss ArgSpec(args=['x', 'gtbox', 'gtlabel', 'anchors', 'class_num', 'ignore_thresh', 'loss_weight_xy', 'loss_weight_wh', 'loss_weight_conf_target', 'loss_weight_conf_notarget', 'loss_weight_class', 'name'], varargs=None, keywords=None, defaults=(None, None, None, None, None, None))
 paddle.fluid.layers.accuracy ArgSpec(args=['input', 'label', 'k', 'correct', 'total'], varargs=None, keywords=None, defaults=(1, None, None))
 paddle.fluid.layers.auc ArgSpec(args=['input', 'label', 'curve', 'num_thresholds', 'topk', 'slide_steps'], varargs=None, keywords=None, defaults=('ROC', 4095, 1, 1))
 paddle.fluid.layers.exponential_decay ArgSpec(args=['learning_rate', 'decay_steps', 'decay_rate', 'staircase'], varargs=None, keywords=None, defaults=(False,))
@ -426,3 +434,17 @@ paddle.fluid.Scope.drop_kids drop_kids(self: paddle.fluid.core.Scope) -> None
 paddle.fluid.Scope.find_var find_var(self: paddle.fluid.core.Scope, arg0: unicode) -> paddle.fluid.core.Variable
 paddle.fluid.Scope.new_scope new_scope(self: paddle.fluid.core.Scope) -> paddle.fluid.core.Scope
 paddle.fluid.Scope.var var(self: paddle.fluid.core.Scope, arg0: unicode) -> paddle.fluid.core.Variable
+paddle.reader.map_readers ArgSpec(args=['func'], varargs='readers', keywords=None, defaults=None)
+paddle.reader.buffered ArgSpec(args=['reader', 'size'], varargs=None, keywords=None, defaults=None)
+paddle.reader.compose ArgSpec(args=[], varargs='readers', keywords='kwargs', defaults=None)
+paddle.reader.chain ArgSpec(args=[], varargs='readers', keywords=None, defaults=None)
+paddle.reader.shuffle ArgSpec(args=['reader', 'buf_size'], varargs=None, keywords=None, defaults=None)
+paddle.reader.firstn ArgSpec(args=['reader', 'n'], varargs=None, keywords=None, defaults=None)
+paddle.reader.xmap_readers ArgSpec(args=['mapper', 'reader', 'process_num', 'buffer_size', 'order'], varargs=None, keywords=None, defaults=(False,))
+paddle.reader.PipeReader.__init__ ArgSpec(args=['self', 'command', 'bufsize', 'file_type'], varargs=None, keywords=None, defaults=(8192, 'plain'))
+paddle.reader.PipeReader.get_line ArgSpec(args=['self', 'cut_lines', 'line_break'], varargs=None, keywords=None, defaults=(True, '\n'))
+paddle.reader.multiprocess_reader ArgSpec(args=['readers', 'use_pipe', 'queue_size'], varargs=None, keywords=None, defaults=(True, 1000))
+paddle.reader.Fake.__init__ ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
+paddle.reader.creator.np_array ArgSpec(args=['x'], varargs=None, keywords=None, defaults=None)
+paddle.reader.creator.text_file ArgSpec(args=['path'], varargs=None, keywords=None, defaults=None)
+paddle.reader.creator.recordio ArgSpec(args=['paths', 'buf_size'], varargs=None, keywords=None, defaults=(100,))
--- a/paddle/fluid/CMakeLists.txt
+++ b/paddle/fluid/CMakeLists.txt
@ -1,6 +1,7 @@
 add_subdirectory(memory)
 add_subdirectory(platform)
 add_subdirectory(framework)
+add_subdirectory(imperative)
 add_subdirectory(operators)
 add_subdirectory(string)
 add_subdirectory(recordio)
--- a/paddle/fluid/framework/CMakeLists.txt
+++ b/paddle/fluid/framework/CMakeLists.txt
@ -3,8 +3,9 @@
 #We create a hidden file and compile it instead of origin source file.
 function(windows_symbolic TARGET)
  set(oneValueArgs "")
-  set(multiValueArgs SRCS DEPS)
+  set(multiValueArgs SRCS PATH)
  cmake_parse_arguments(windows_symbolic "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
+  set(final_path ${CMAKE_CURRENT_SOURCE_DIR}/${windows_symbolic_PATH})
  foreach(src ${windows_symbolic_SRCS})
  get_filename_component(src ${src} NAME_WE)
  if (NOT EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${src}.cc OR NOT EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${src}.cu)
@ -72,6 +73,8 @@ cc_library(lod_tensor SRCS lod_tensor.cc DEPS ddim place tensor framework_proto
 cc_test(lod_tensor_test SRCS lod_tensor_test.cc DEPS lod_tensor memory)
 nv_test(lod_tensor_gpu_test SRCS lod_tensor_test.cu DEPS lod_tensor)

+cc_library(garbage_collector SRCS garbage_collector.cc DEPS device_context memory)
+
 cc_library(reader SRCS reader.cc DEPS lod_tensor ddim)
 cc_test(reader_test SRCS reader_test.cc DEPS reader)

@ -118,8 +121,9 @@ cc_library(op_info SRCS op_info.cc DEPS attribute framework_proto)
 cc_library(shape_inference SRCS shape_inference.cc DEPS ddim attribute device_context)

 cc_library(transfer_scope_cache SRCS transfer_scope_cache.cc DEPS scope framework_proto device_context)
+cc_library(op_kernel_type SRCS op_kernel_type.cc DEPS device_context place)
 cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope glog
-    shape_inference data_transform lod_tensor profiler transfer_scope_cache)
+    shape_inference data_transform lod_tensor profiler transfer_scope_cache op_kernel_type)

 cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry device_context)

@ -127,11 +131,14 @@ 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)
+
+if(WITH_NGRAPH)
+  if(NOT WIN32)
+    cc_library(ngraph_bridge SRCS ngraph_bridge.cc DEPS operator framework_proto ngraph)
+    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 ngraph)
+  endif(NOT WIN32)
+endif(WITH_NGRAPH)

 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)
@ -164,18 +171,27 @@ cc_library(variable_helper SRCS variable_helper.cc DEPS lod_tensor)
 cc_library(naive_executor SRCS naive_executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass variable_helper)

 if(WITH_DISTRIBUTE)
-  cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method sendrecvop_grpc cares grpc++_unsecure grpc_unsecure gpr graph_to_program_pass variable_helper)
-  set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
-  set_source_files_properties(executor.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
+    cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog
+        lod_rank_table feed_fetch_method sendrecvop_rpc  ${GLOB_DISTRIBUTE_DEPS} graph_to_program_pass variable_helper)
+
+   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()
-  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 variable_helper)
-  else(NOT WIN32)
+  if(WITH_NGRAPH)
+    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 ngraph_operator variable_helper)
+    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 variable_helper)
+    endif(NOT WIN32)
+  else(WITH_NGRAPH)
    cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass variable_helper)
-  endif(NOT WIN32)
+  endif(WITH_NGRAPH)
  cc_test(test_naive_executor SRCS naive_executor_test.cc DEPS naive_executor elementwise_add_op)
 endif()

+target_link_libraries(executor garbage_collector)
+
 cc_library(parallel_executor SRCS parallel_executor.cc DEPS
        threaded_ssa_graph_executor scope_buffered_ssa_graph_executor
        graph build_strategy
@ -196,7 +212,7 @@ cc_test(var_type_inference_test SRCS var_type_inference_test.cc DEPS op_registry
 cc_library(selected_rows SRCS selected_rows.cc DEPS tensor)
 cc_test(selected_rows_test SRCS selected_rows_test.cc DEPS selected_rows)

-cc_test(op_kernel_type_test SRCS op_kernel_type_test.cc DEPS place device_context framework_proto)
+cc_test(op_kernel_type_test SRCS op_kernel_type_test.cc DEPS place device_context framework_proto op_kernel_type)
 cc_test(cow_ptr_tests SRCS details/cow_ptr_test.cc)

 cc_test(tuple_test SRCS tuple_test.cc )
--- a/paddle/fluid/framework/data_feed.cc
+++ b/paddle/fluid/framework/data_feed.cc
@ -33,11 +33,7 @@ void DataFeed::AddFeedVar(Variable* var, const std::string& name) {
  CheckInit();
  for (size_t i = 0; i < use_slots_.size(); ++i) {
    if (name == use_slots_[i]) {
-      if (use_slots_is_dense_[i]) {
-        feed_vec_[i] = MixTensor(var->GetMutable<Tensor>());
-      } else {
-        feed_vec_[i] = MixTensor(var->GetMutable<LoDTensor>());
-      }
+      feed_vec_[i] = var->GetMutable<LoDTensor>();
    }
  }
 }
@ -201,22 +197,22 @@ bool MultiSlotDataFeed::CheckFile(const char* filename) {
    for (size_t i = 0; i < all_slots_.size(); ++i) {
      int num = strtol(endptr, &endptr, 10);
      if (num < 0) {
-        VLOG(1) << "error: the number of ids is a negative number: " << num;
-        VLOG(1) << "please check line<" << instance_cout << "> in file<"
+        VLOG(0) << "error: the number of ids is a negative number: " << num;
+        VLOG(0) << "please check line<" << instance_cout << "> in file<"
                << filename << ">";
        return false;
      } else if (num == 0) {
-        VLOG(1)
+        VLOG(0)
            << "error: the number of ids can not be zero, you need "
               "padding it in data generator; or if there is something wrong"
               " with the data, please check if the data contains unresolvable "
               "characters.";
-        VLOG(1) << "please check line<" << instance_cout << "> in file<"
+        VLOG(0) << "please check line<" << instance_cout << "> in file<"
                << filename << ">";
        return false;
      } else if (errno == ERANGE || num > INT_MAX) {
-        VLOG(1) << "error: the number of ids greater than INT_MAX";
-        VLOG(1) << "please check line<" << instance_cout << "> in file<"
+        VLOG(0) << "error: the number of ids greater than INT_MAX";
+        VLOG(0) << "please check line<" << instance_cout << "> in file<"
                << filename << ">";
        return false;
      }
@ -224,15 +220,15 @@ bool MultiSlotDataFeed::CheckFile(const char* filename) {
        for (int i = 0; i < num; ++i) {
          strtof(endptr, &endptr);
          if (errno == ERANGE) {
-            VLOG(1) << "error: the value is out of the range of "
+            VLOG(0) << "error: the value is out of the range of "
                       "representable values for float";
-            VLOG(1) << "please check line<" << instance_cout << "> in file<"
+            VLOG(0) << "please check line<" << instance_cout << "> in file<"
                    << filename << ">";
            return false;
          }
          if (i + 1 != num && endptr - str == len) {
-            VLOG(1) << "error: there is a wrong with the number of ids.";
-            VLOG(1) << "please check line<" << instance_cout << "> in file<"
+            VLOG(0) << "error: there is a wrong with the number of ids.";
+            VLOG(0) << "please check line<" << instance_cout << "> in file<"
                    << filename << ">";
            return false;
          }
@ -241,30 +237,41 @@ bool MultiSlotDataFeed::CheckFile(const char* filename) {
        for (int i = 0; i < num; ++i) {
          strtoull(endptr, &endptr, 10);
          if (errno == ERANGE) {
-            VLOG(1) << "error: the value is out of the range of "
+            VLOG(0) << "error: the value is out of the range of "
                       "representable values for uint64_t";
-            VLOG(1) << "please check line<" << instance_cout << "> in file<"
+            VLOG(0) << "please check line<" << instance_cout << "> in file<"
                    << filename << ">";
            return false;
          }
          if (i + 1 != num && endptr - str == len) {
-            VLOG(1) << "error: there is a wrong with the number of ids.";
-            VLOG(1) << "please check line<" << instance_cout << "> in file<"
+            VLOG(0) << "error: there is a wrong with the number of ids.";
+            VLOG(0) << "please check line<" << instance_cout << "> in file<"
                    << filename << ">";
            return false;
          }
        }
      } else {
-        VLOG(1) << "error: this type<" << all_slots_type_[i]
+        VLOG(0) << "error: this type<" << all_slots_type_[i]
                << "> is not supported";
        return false;
      }
    }
-    if (endptr - str != len) {
-      VLOG(1) << "error: there is some data at the end of the line.";
-      VLOG(1) << "please check line<" << instance_cout << "> in file<"
-              << filename << ">";
-      return false;
+    // It may be added '\t' character to the end of the output of reduce
+    // task when processes data by Hadoop(when the output of the reduce
+    // task of Hadoop has only one field, it will add a '\t' at the end
+    // of the line by default, and you can use this option to avoid it:
+    // `-D mapred.textoutputformat.ignoreseparator=true`), which does
+    // not affect the correctness of the data. Therefore, it should be
+    // judged that the data is not normal when the end of each line of
+    // data contains characters which are not spaces.
+    while (endptr - str != len) {
+      if (!isspace(*(endptr++))) {
+        VLOG(0)
+            << "error: there is some extra characters at the end of the line.";
+        VLOG(0) << "please check line<" << instance_cout << "> in file<"
+                << filename << ">";
+        return false;
+      }
    }
  }
  VLOG(3) << "instances cout: " << instance_cout;
@ -291,6 +298,7 @@ bool MultiSlotDataFeed::ParseOneInstance(std::vector<MultiSlotType>* instance) {
          "the data, please check if the data contains unresolvable "
          "characters.\nplease check this error line: %s",
          str);
+
      if (idx != -1) {
        (*instance)[idx].Init(all_slots_type_[i]);
        if ((*instance)[idx].GetType()[0] == 'f') {  // float
@ -327,6 +335,7 @@ void MultiSlotDataFeed::AddInstanceToInsVec(
      (*ins_vec)[i].InitOffset();
    }
  }
+
  for (size_t i = 0; i < instance.size(); ++i) {
    (*ins_vec)[i].AddIns(instance[i]);
  }
@ -338,36 +347,25 @@ void MultiSlotDataFeed::PutToFeedVec(
    const auto& type = ins_vec[i].GetType();
    const auto& offset = ins_vec[i].GetOffset();
    int total_instance = static_cast<int>(offset.back());
+
    if (type[0] == 'f') {  // float
      const auto& feasign = ins_vec[i].GetFloatData();
-      if (feed_vec_[i].IsDense()) {
-        int size_in_each_batch = total_instance / batch_size_;
-        float* tensor_ptr = feed_vec_[i].GetTensor()->mutable_data<float>(
-            {batch_size_, size_in_each_batch}, platform::CPUPlace());
-        memcpy(tensor_ptr, &feasign[0], total_instance * sizeof(float));
-      } else {
-        float* tensor_ptr = feed_vec_[i].GetLoDTensor()->mutable_data<float>(
-            {total_instance, 1}, platform::CPUPlace());
-        memcpy(tensor_ptr, &feasign[0], total_instance * sizeof(float));
-        LoD data_lod{offset};
-        feed_vec_[i].GetLoDTensor()->set_lod(data_lod);
-      }
+      float* tensor_ptr = feed_vec_[i]->mutable_data<float>(
+          {total_instance, 1}, platform::CPUPlace());
+      memcpy(tensor_ptr, &feasign[0], total_instance * sizeof(float));
    } else if (type[0] == 'u') {  // uint64
      // no uint64_t type in paddlepaddle
      const auto& feasign = ins_vec[i].GetUint64Data();
-      if (feed_vec_[i].IsDense()) {
-        int size_in_each_batch = total_instance / batch_size_;
-        int64_t* tensor_ptr = feed_vec_[i].GetTensor()->mutable_data<int64_t>(
-            {batch_size_, size_in_each_batch}, platform::CPUPlace());
-        memcpy(tensor_ptr, &feasign[0], total_instance * sizeof(int64_t));
-      } else {
-        int64_t* tensor_ptr =
-            feed_vec_[i].GetLoDTensor()->mutable_data<int64_t>(
-                {total_instance, 1}, platform::CPUPlace());
-        memcpy(tensor_ptr, &feasign[0], total_instance * sizeof(int64_t));
-        LoD data_lod{offset};
-        feed_vec_[i].GetLoDTensor()->set_lod(data_lod);
-      }
+      int64_t* tensor_ptr = feed_vec_[i]->mutable_data<int64_t>(
+          {total_instance, 1}, platform::CPUPlace());
+      memcpy(tensor_ptr, &feasign[0], total_instance * sizeof(int64_t));
+    }
+
+    LoD data_lod{offset};
+    feed_vec_[i]->set_lod(data_lod);
+    if (use_slots_is_dense_[i]) {
+      int dim = total_instance / batch_size_;
+      feed_vec_[i]->Resize({batch_size_, dim});
    }
  }
 }
--- a/paddle/fluid/framework/data_feed.h
+++ b/paddle/fluid/framework/data_feed.h
@ -30,35 +30,6 @@ limitations under the License. */
 namespace paddle {
 namespace framework {

-// Pack Tensor type and LoDTensor type into MixTensor type, in order
-// to record either Tensor or LoDTensor information at the same time.
-class MixTensor {
- public:
-  MixTensor() {}
-  explicit MixTensor(LoDTensor* lodtensor) {
-    is_dense_ = false;
-    lodtensor_ = lodtensor;
-  }
-  explicit MixTensor(Tensor* tensor) {
-    is_dense_ = true;
-    tensor_ = tensor;
-  }
-  bool IsDense() { return is_dense_; }
-  LoDTensor* GetLoDTensor() {
-    PADDLE_ENFORCE(!is_dense_, "Let a dense var return a LoDTensor ptr.");
-    return lodtensor_;
-  }
-  Tensor* GetTensor() {
-    PADDLE_ENFORCE(is_dense_, "Let a sparse var return a Tensor ptr.");
-    return tensor_;
-  }
-
- private:
-  bool is_dense_;
-  LoDTensor* lodtensor_;
-  Tensor* tensor_;
-};
-
 // DataFeed is the base virtual class for all ohther DataFeeds.
 // It is used to read files and parse the data for subsequent trainer.
 // Example:
@ -133,7 +104,7 @@ class DataFeed {
      use_slots_index_;  // -1: not used; >=0: the index of use_slots_

  // The data read by DataFeed will be stored here
-  std::vector<MixTensor> feed_vec_;
+  std::vector<LoDTensor*> feed_vec_;

  // the batch size defined by user
  int default_batch_size_;
--- a/paddle/fluid/framework/data_feed_test.cc
+++ b/paddle/fluid/framework/data_feed_test.cc
@ -152,19 +152,13 @@ void GetElemSetFromReader(std::vector<MultiTypeSet>* reader_elem_set,
      const auto& multi_slot_desc = data_feed_desc.multi_slot_desc();
      std::map<std::string, const paddle::framework::LoDTensor*>
          lodtensor_targets;
-      std::map<std::string, const paddle::framework::Tensor*> tensor_targets;
      for (int i = 0; i < multi_slot_desc.slots_size(); ++i) {
        const auto& slot = multi_slot_desc.slots(i);
        if (slot.is_used()) {
          const auto& name = slot.name();
          readers[idx]->AddFeedVar(scope->Var(name), name);
-          if (slot.is_dense()) {
-            tensor_targets[name] =
-                &scope->FindVar(name)->Get<paddle::framework::Tensor>();
-          } else {
-            lodtensor_targets[name] =
-                &scope->FindVar(name)->Get<paddle::framework::LoDTensor>();
-          }
+          lodtensor_targets[name] =
+              &scope->FindVar(name)->Get<paddle::framework::LoDTensor>();
        }
      }
      readers[idx]->Start();
@ -175,8 +169,9 @@ void GetElemSetFromReader(std::vector<MultiTypeSet>* reader_elem_set,
          if (!slot.is_used()) {
            continue;
          }
+          const paddle::framework::LoDTensor* tens =
+              lodtensor_targets[slot.name()];
          if (slot.is_dense()) {  // dense branch
-            const paddle::framework::Tensor* tens = tensor_targets[slot.name()];
            if (slot.type() == "uint64") {
              const int64_t* data = tens->data<int64_t>();
              int batch_size = tens->dims()[0];
@ -202,8 +197,6 @@ void GetElemSetFromReader(std::vector<MultiTypeSet>* reader_elem_set,
              PADDLE_THROW("Error type in proto file.");
            }
          } else {  // sparse branch
-            const paddle::framework::LoDTensor* tens =
-                lodtensor_targets[slot.name()];
            if (slot.type() == "uint64") {
              const int64_t* data = tens->data<int64_t>();
              for (size_t i = 0; i < tens->NumElements(); ++i) {
--- a/paddle/fluid/framework/data_layout_transform.cc
+++ b/paddle/fluid/framework/data_layout_transform.cc
@ -85,7 +85,7 @@ void TransDataLayout(const OpKernelType& kernel_type_for_var,
  out->mutable_data(expected_kernel_type.place_, in.type());

  framework::VisitDataType(
-      framework::ToDataType(in.type()),
+      in.type(),
      CastDataLayout(pool.Get(expected_kernel_type.place_), axis, in, out));

  out->set_layout(expected_kernel_type.data_layout_);
@ -101,7 +101,7 @@ void* GetDataFromTensor(const Tensor& tensor, mkldnn::memory::data_type type) {
    case mkldnn::memory::data_type::f32:
      return platform::to_void_cast(tensor.data<float>());
    case mkldnn::memory::data_type::s8:
-      return platform::to_void_cast(tensor.data<char>());
+      return platform::to_void_cast(tensor.data<int8_t>());
    case mkldnn::memory::data_type::u8:
      return platform::to_void_cast(tensor.data<unsigned char>());
    case mkldnn::memory::data_type::s16:
@ -144,26 +144,29 @@ void TransDataLayoutFromMKLDNN(const OpKernelType& kernel_type_for_var,

  memory::data_type in_type = ToMKLDNNDataType(in.type());
  PADDLE_ENFORCE(in_type != memory::data_type::data_undef,
-                 "Input tensor type is not supported: ", in.type().name());
+                 "Input tensor type is not supported: %s", in.type());
  memory::data_type out_type = in_type;

  auto in_format = platform::MKLDNNFormatForSize(in_tz.size(), in.format());
  auto out_format =
      platform::MKLDNNFormatForSize(in_tz.size(), ToMKLDNNFormat(out_layout));

-  void* in_data = GetDataFromTensor(in, in_type);
-
  // output tensor has the same dims as input. Reorder don't change dims
  out->Resize(in.dims());

-  auto out_data = out->mutable_data(expected_kernel_type.place_, in.type());
-
-  auto in_memory = memory({{{in_tz}, in_type, in_format}, cpu_engine}, in_data);
-  auto out_memory =
-      memory({{{out_tz}, out_type, out_format}, cpu_engine}, out_data);
+  if (in_format != out_format) {
+    void* in_data = GetDataFromTensor(in, in_type);
+    auto out_data = out->mutable_data(expected_kernel_type.place_, in.type());

-  platform::Reorder(in_memory, out_memory);
+    auto in_memory =
+        memory({{{in_tz}, in_type, in_format}, cpu_engine}, in_data);
+    auto out_memory =
+        memory({{{out_tz}, out_type, out_format}, cpu_engine}, out_data);

+    platform::Reorder(in_memory, out_memory);
+  } else {
+    out->ShareDataWith(in);
+  }
  out->set_layout(out_layout);
  // reset format since the out tensor will be feed to non-MKLDNN OPkernel
  out->set_format(memory::format::format_undef);
--- a/paddle/fluid/framework/data_layout_transform.h
+++ b/paddle/fluid/framework/data_layout_transform.h
@ -50,14 +50,14 @@ inline DataLayout ToPaddleLayout(const MKLDNNFormat& format) {
  }
 }

-inline MKLDNNDataType ToMKLDNNDataType(const std::type_index type) {
-  static const std::map<std::type_index, MKLDNNDataType> dict{
-      {std::type_index(typeid(float)), MKLDNNDataType::f32},  // NOLINT
-      {std::type_index(typeid(char)), MKLDNNDataType::s8},    // NOLINT
-      {std::type_index(typeid(unsigned char)), MKLDNNDataType::u8},
-      {std::type_index(typeid(int16_t)), MKLDNNDataType::s16},
-      {std::type_index(typeid(int32_t)), MKLDNNDataType::s32}};
-  auto iter = dict.find(type);
+inline MKLDNNDataType ToMKLDNNDataType(proto::VarType::Type type) {
+  static std::unordered_map<int, MKLDNNDataType> dict{
+      {DataTypeTrait<float>::DataType, MKLDNNDataType::f32},
+      {DataTypeTrait<int8_t>::DataType, MKLDNNDataType::s8},
+      {DataTypeTrait<uint8_t>::DataType, MKLDNNDataType::u8},
+      {DataTypeTrait<int16_t>::DataType, MKLDNNDataType::s16},
+      {DataTypeTrait<int32_t>::DataType, MKLDNNDataType::s32}};
+  auto iter = dict.find(static_cast<int>(type));
  if (iter != dict.end()) return iter->second;
  return MKLDNNDataType::data_undef;
 }
--- a/paddle/fluid/framework/data_type.cc
+++ b/paddle/fluid/framework/data_type.cc
@ -26,7 +26,7 @@ struct DataTypeMap {
  std::unordered_map<std::type_index, proto::VarType::Type> cpp_to_proto_;
  std::unordered_map<int, std::type_index> proto_to_cpp_;
  std::unordered_map<int, std::string> proto_to_str_;
-  std::unordered_map<std::type_index, size_t> cpp_to_size_;
+  std::unordered_map<int, size_t> proto_to_size_;
 };

 static DataTypeMap* InitDataTypeMap();
@ -45,7 +45,7 @@ static inline void RegisterType(DataTypeMap* map,
  map->proto_to_cpp_.emplace(static_cast<int>(proto_type), typeid(T));
  map->cpp_to_proto_.emplace(typeid(T), proto_type);
  map->proto_to_str_.emplace(static_cast<int>(proto_type), name);
-  map->cpp_to_size_.emplace(typeid(T), sizeof(T));
+  map->proto_to_size_.emplace(static_cast<int>(proto_type), sizeof(T));
 }

 static DataTypeMap* InitDataTypeMap() {
@ -54,17 +54,7 @@ static DataTypeMap* InitDataTypeMap() {
 #define RegType(cc_type, proto_type) \
  RegisterType<cc_type>(retv, proto_type, #cc_type)

-  // NOTE: Add your customize type here.
-  RegType(float16, proto::VarType::FP16);
-  RegType(float, proto::VarType::FP32);
-  RegType(double, proto::VarType::FP64);
-  RegType(int, proto::VarType::INT32);
-  RegType(int64_t, proto::VarType::INT64);
-  RegType(bool, proto::VarType::BOOL);
-  RegType(size_t, proto::VarType::SIZE_T);
-  RegType(int16_t, proto::VarType::INT16);
-  RegType(uint8_t, proto::VarType::UINT8);
-  RegType(int8_t, proto::VarType::INT8);
+  _ForEachDataType_(RegType);

 #undef RegType
  return retv;
@ -96,12 +86,12 @@ std::string DataTypeToString(const proto::VarType::Type type) {
               static_cast<int>(type));
 }

-size_t SizeOfType(std::type_index type) {
-  auto it = gDataTypeMap().cpp_to_size_.find(type);
-  if (it != gDataTypeMap().cpp_to_size_.end()) {
+size_t SizeOfType(proto::VarType::Type type) {
+  auto it = gDataTypeMap().proto_to_size_.find(static_cast<int>(type));
+  if (it != gDataTypeMap().proto_to_size_.end()) {
    return it->second;
  }
-  PADDLE_THROW("Not support %s as tensor type", type.name());
+  PADDLE_THROW("Not support %s as tensor type", DataTypeToString(type));
 }

 }  // namespace framework
--- a/paddle/fluid/framework/data_type.h
+++ b/paddle/fluid/framework/data_type.h
@ -22,46 +22,59 @@ limitations under the License. */
 namespace paddle {
 namespace framework {

+template <typename T>
+struct DataTypeTrait {};
+
+// Stub handle for void
+template <>
+struct DataTypeTrait<void> {
+  constexpr static auto DataType = proto::VarType::RAW;
+};
+
+#define _ForEachDataTypeHelper_(callback, cpp_type, proto_type) \
+  callback(cpp_type, ::paddle::framework::proto::VarType::proto_type);
+
+#define _ForEachDataType_(callback)                                     \
+  _ForEachDataTypeHelper_(callback, float, FP32);                       \
+  _ForEachDataTypeHelper_(callback, ::paddle::platform::float16, FP16); \
+  _ForEachDataTypeHelper_(callback, double, FP64);                      \
+  _ForEachDataTypeHelper_(callback, int, INT32);                        \
+  _ForEachDataTypeHelper_(callback, int64_t, INT64);                    \
+  _ForEachDataTypeHelper_(callback, bool, BOOL);                        \
+  _ForEachDataTypeHelper_(callback, uint8_t, UINT8);                    \
+  _ForEachDataTypeHelper_(callback, int16_t, INT16);                    \
+  _ForEachDataTypeHelper_(callback, int8_t, INT8)
+
+#define DefineDataTypeTrait(cpp_type, proto_type) \
+  template <>                                     \
+  struct DataTypeTrait<cpp_type> {                \
+    constexpr static auto DataType = proto_type;  \
+  }
+
+_ForEachDataType_(DefineDataTypeTrait);
+
+#undef DefineDataTypeTrait
+
 extern proto::VarType::Type ToDataType(std::type_index type);
 extern std::type_index ToTypeIndex(proto::VarType::Type type);

 template <typename Visitor>
 inline void VisitDataType(proto::VarType::Type type, Visitor visitor) {
-  switch (type) {
-    case proto::VarType::FP16:
-      visitor.template apply<platform::float16>();
-      break;
-    case proto::VarType::FP32:
-      visitor.template apply<float>();
-      break;
-    case proto::VarType::FP64:
-      visitor.template apply<double>();
-      break;
-    case proto::VarType::INT32:
-      visitor.template apply<int>();
-      break;
-    case proto::VarType::INT64:
-      visitor.template apply<int64_t>();
-      break;
-    case proto::VarType::BOOL:
-      visitor.template apply<bool>();
-      break;
-    case proto::VarType::UINT8:
-      visitor.template apply<uint8_t>();
-      break;
-    case proto::VarType::INT16:
-      visitor.template apply<int16_t>();
-      break;
-    case proto::VarType::INT8:
-      visitor.template apply<int8_t>();
-      break;
-    default:
-      PADDLE_THROW("Not supported %d", type);
-  }
+#define VisitDataTypeCallback(cpp_type, proto_type) \
+  do {                                              \
+    if (type == proto_type) {                       \
+      visitor.template apply<cpp_type>();           \
+      return;                                       \
+    }                                               \
+  } while (0)
+
+  _ForEachDataType_(VisitDataTypeCallback);
+#undef VisitDataTypeCallback
+  PADDLE_THROW("Not supported %d", type);
 }

 extern std::string DataTypeToString(const proto::VarType::Type type);
-extern size_t SizeOfType(std::type_index type);
+extern size_t SizeOfType(proto::VarType::Type type);
 inline std::ostream& operator<<(std::ostream& out,
                                const proto::VarType::Type& type) {
  out << DataTypeToString(type);
--- a/Show More
+++ b/Show More