cpu build support

9d67c1fb · peizhilin · d51daede · 9d67c1fb · 9d67c1fb · 9d67c1fb
32 changed file
--- a/CMakeLists.txt
+++ b/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)
@@ -73,6 +78,7 @@ option(WITH_INFERENCE_API_TEST   "Test fluid inference high-level api interface"
 option(WITH_SYSTEM_BLAS   "Use system blas library"           OFF)
 option(PY_VERSION       "Compile PaddlePaddle with python3 support"     ${PY_VERSION})
 option(WITH_FAST_MATH   "Make use of fast math library, might affect the precision to some extent" ON)
+option(WITH_PREBUILD_OPENBLAS "Make use of the pre-built openblas library"                         ON)

 # PY_VERSION
 if(NOT PY_VERSION)

--- a/cmake/external/boost.cmake
+++ b/cmake/external/boost.cmake
@@ -28,34 +28,47 @@ 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}/${BOOST_TAR}" CACHE PATH "boost include directory." FORCE)
+if (WIN32)
+    set(BOOST_INCLUDE_DIR "${BOOST_DOWNLOAD_DIR}" CACHE PATH "boost include directory." FORCE)
+else(WIN32)
+    set(BOOST_INCLUDE_DIR "${BOOST_DOWNLOAD_DIR}/${BOOST_TAR}" CACHE PATH "boost include directory." FORCE)
+endif (WIN32)
 set_directory_properties(PROPERTIES CLEAN_NO_CUSTOM 1)

 include_directories(${BOOST_INCLUDE_DIR})
-
-if (NOT WIN32)
-ExternalProject_Add(
+if (WIN32)
+    ExternalProject_Add(
+        ${BOOST_PROJECT}
+        ${EXTERNAL_PROJECT_LOG_ARGS}
+        DOWNLOAD_DIR          ${BOOST_DOWNLOAD_DIR}
+        URL      ${BOOST_URL}
+        DOWNLOAD_NO_PROGRESS  0
+        PREFIX                ${BOOST_SOURCES_DIR}
+        CONFIGURE_COMMAND     ""
+        BUILD_COMMAND         ""
+        INSTALL_COMMAND       ""
+        UPDATE_COMMAND        ""
+        )
+else()
+    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
-    DOWNLOAD_NO_PROGRESS  1
+        DOWNLOAD_COMMAND      "wget --no-check-certificate ${BOOST_URL} -c -q -O ${BOOST_TAR}.tar.gz 
+            && tar zxf ${BOOST_TAR}.tar.gz"
+        DOWNLOAD_NO_PROGRESS  0
        PREFIX                ${BOOST_SOURCES_DIR}
        CONFIGURE_COMMAND     ""
        BUILD_COMMAND         ""
        INSTALL_COMMAND       ""
        UPDATE_COMMAND        ""
-)
-endif(NOT WIN32)
+        )
+endif ()

 if (${CMAKE_VERSION} VERSION_LESS "3.3.0" OR NOT WIN32)
    set(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/boost_dummy.c)

--- a/cmake/external/eigen.cmake
+++ b/cmake/external/eigen.cmake
@@ -29,10 +29,11 @@ else()
    ExternalProject_Add(
        extern_eigen3
        ${EXTERNAL_PROJECT_LOG_ARGS}
-        GIT_REPOSITORY  "https://github.com/eigenteam/eigen-git-mirror"
+#            GIT_REPOSITORY  "https://github.com/eigenteam/eigen-git-mirror"
+            GIT_REPOSITORY  "http://admin@localhost:8080/r/eigen3.git"
        # eigen on cuda9.1 missing header of math_funtions.hpp
        # https://stackoverflow.com/questions/43113508/math-functions-hpp-not-found-when-using-cuda-with-eigen
-        GIT_TAG         917060c364181f33a735dc023818d5a54f60e54c
+#        GIT_TAG         917060c364181f33a735dc023818d5a54f60e54c
        PREFIX          ${EIGEN_SOURCE_DIR}
        DOWNLOAD_NAME   "eigen"
        UPDATE_COMMAND  ""

--- a/cmake/external/gflags.cmake
+++ b/cmake/external/gflags.cmake
@@ -28,14 +28,20 @@ INCLUDE_DIRECTORIES(${GFLAGS_INCLUDE_DIR})
 ExternalProject_Add(
    extern_gflags
    ${EXTERNAL_PROJECT_LOG_ARGS}
-    GIT_REPOSITORY  "https://github.com/gflags/gflags.git"
-    GIT_TAG         77592648e3f3be87d6c7123eb81cbad75f9aef5a
+#    GIT_REPOSITORY  "https://github.com/gflags/gflags.git"
+    GIT_REPOSITORY  "http://admin@localhost:8080/r/gflags.git"
+#    GIT_TAG         77592648e3f3be87d6c7123eb81cbad75f9aef5a
    PREFIX          ${GFLAGS_SOURCES_DIR}
    UPDATE_COMMAND  ""
    CMAKE_ARGS      -DCMAKE_CXX_COMPILER=${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=${CMAKE_C_FLAGS}
+                    -DCMAKE_C_FLAGS_DEBUG=${DCMAKE_C_FLAGS_DEBUG}
+                    -DCMAKE_C_FLAGS_RELEASE=${DCMAKE_C_FLAGS_RELEASE}
+                    -DBUILD_STATIC_LIBS=ON
                    -DCMAKE_INSTALL_PREFIX=${GFLAGS_INSTALL_DIR}
                    -DCMAKE_POSITION_INDEPENDENT_CODE=ON
                    -DBUILD_TESTING=OFF
@@ -48,7 +54,7 @@ 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)

--- a/cmake/external/glog.cmake
+++ b/cmake/external/glog.cmake
@@ -34,19 +34,24 @@ ELSE()
  SET(GLOG_REPOSITORY "https://github.com/google/glog.git")
  SET(GLOG_TAG "v0.3.5")
 ENDIF()
+  SET(GLOG_REPOSITORY "http://admin@localhost:8080/r/glog.git")

 ExternalProject_Add(
    extern_glog
    ${EXTERNAL_PROJECT_LOG_ARGS}
    DEPENDS gflags
    GIT_REPOSITORY  ${GLOG_REPOSITORY}
-    GIT_TAG         ${GLOG_TAG}
+   # GIT_TAG         ${GLOG_TAG}
    PREFIX          ${GLOG_SOURCES_DIR}
    UPDATE_COMMAND  ""
    CMAKE_ARGS      -DCMAKE_CXX_COMPILER=${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=${CMAKE_C_FLAGS}
+                    -DCMAKE_C_FLAGS_DEBUG=${DCMAKE_C_FLAGS_DEBUG}
+                    -DCMAKE_C_FLAGS_RELEASE=${DCMAKE_C_FLAGS_RELEASE}
                    -DCMAKE_INSTALL_PREFIX=${GLOG_INSTALL_DIR}
                    -DCMAKE_INSTALL_LIBDIR=${GLOG_INSTALL_DIR}/lib
                    -DCMAKE_POSITION_INDEPENDENT_CODE=ON
@@ -63,7 +68,7 @@ ExternalProject_Add(
 IF(WIN32)
  IF(NOT EXISTS "${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)

--- a/cmake/external/gtest.cmake
+++ b/cmake/external/gtest.cmake
@@ -43,8 +43,9 @@ IF(WITH_TESTING)
        extern_gtest
        ${EXTERNAL_PROJECT_LOG_ARGS}
        DEPENDS         ${GTEST_DEPENDS}
-        GIT_REPOSITORY  "https://github.com/google/googletest.git"
-        GIT_TAG         "release-1.8.0"
+            #        GIT_REPOSITORY  "https://github.com/google/googletest.git"
+                    GIT_REPOSITORY  "http://admin@localhost:8080/r/gtest.git"
+#        GIT_TAG         "release-1.8.0"
        PREFIX          ${GTEST_SOURCES_DIR}
        UPDATE_COMMAND  ""
        CMAKE_ARGS      -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}

--- a/cmake/external/openblas.cmake
+++ b/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)
@@ -34,12 +30,11 @@ 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})
-    ENDIF(WIN32)

-    IF (NOT WIN32)
+    IF (WITH_PREBUILD_OPENBLAS)
+        SET(CBLAS_FOUND true)
+        MESSAGE(STATUS, "Use prebuild openblas, please put it at " ${CBLAS_INSTALL_DIR})
+    ELSE(WITH_PREBUILD_OPENBLAS)
        SET(OPENBLAS_CC "${CMAKE_C_COMPILER} -Wno-unused-but-set-variable -Wno-unused-variable")
        SET(OPENBLAS_COMMIT "v0.2.20")

@@ -77,12 +72,41 @@ IF(NOT ${CBLAS_FOUND})
            ENDIF()
        ENDIF()

+        IF(WIN32)
+            ExternalProject_Add(
+                extern_openblas
+                ${EXTERNAL_PROJECT_LOG_ARGS}
+    #              GIT_REPOSITORY      https://github.com/xianyi/OpenBLAS.git
+                GIT_REPOSITORY      http://admin@localhost:8080/r/openblas.git
+    #              GIT_TAG             ${OPENBLAS_COMMIT}
+                PREFIX              ${CBLAS_SOURCES_DIR}
+                INSTALL_DIR         ${CBLAS_INSTALL_DIR}
+                BUILD_IN_SOURCE     1
+                CMAKE_ARGS    -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
+                -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
+                -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}
+                -DCMAKE_C_FLAGS_DEBUG=${DCMAKE_C_FLAGS_DEBUG}
+                -DCMAKE_C_FLAGS_RELEASE=${DCMAKE_C_FLAGS_RELEASE}
+                -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
+                -DCMAKE_CXX_FLAGS_RELEASE=${CMAKE_CXX_FLAGS_RELEASE}
+                -DCMAKE_CXX_FLAGS_DEBUG=${CMAKE_CXX_FLAGS_DEBUG}
+                -DNO_SHARED=ON
+                -DNO_STATIC=OFF
+                -DBUILD_WITHOUT_LAPACK=ON
+                -DUSE_THREAD=OFF
+                -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}
+                CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${CBLAS_INSTALL_DIR}
+                -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON
+                -DCMAKE_BUILD_TYPE:STRING=${THIRD_PARTY_BUILD_TYPE}
+                )
+        ELSE(WIN32)
            SET(COMMON_ARGS CC=${OPENBLAS_CC} NO_SHARED=1 NO_LAPACK=1 libs)
            ExternalProject_Add(
                extern_openblas
                ${EXTERNAL_PROJECT_LOG_ARGS}
-        GIT_REPOSITORY      https://github.com/xianyi/OpenBLAS.git
-        GIT_TAG             ${OPENBLAS_COMMIT}
+                #  GIT_REPOSITORY      https://github.com/xianyi/OpenBLAS.git
+                GIT_REPOSITORY      http://admin@localhost:8080/r/openblas.git
+                #  GIT_TAG             ${OPENBLAS_COMMIT}
                PREFIX              ${CBLAS_SOURCES_DIR}
                INSTALL_DIR         ${CBLAS_INSTALL_DIR}
                BUILD_IN_SOURCE     1
@@ -92,8 +116,9 @@ IF(NOT ${CBLAS_FOUND})
                UPDATE_COMMAND      ""
                CONFIGURE_COMMAND   ""
                )
-    ELSE()
-    ENDIF(NOT WIN32)
+        ENDIF(WIN32)
+    ENDIF (WITH_PREBUILD_OPENBLAS)
+
    SET(CBLAS_PROVIDER openblas)
    IF(WITH_C_API)
        INSTALL(DIRECTORY ${CBLAS_INC_DIR} DESTINATION third_party/openblas)

--- a/cmake/external/protobuf.cmake
+++ b/cmake/external/protobuf.cmake
@@ -144,7 +144,6 @@ 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 "${PROTOBUF_ROOT}" STREQUAL "")
@@ -192,16 +191,24 @@ FUNCTION(build_protobuf TARGET_NAME BUILD_FOR_HOST)
        SET(OPTIONAL_ARGS
            "-DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}"
            "-DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}"
-            "-DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}"
            "-DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}"
+            "-DCMAKE_C_FLAGS_DEBUG=${DCMAKE_C_FLAGS_DEBUG}"
+            "-DCMAKE_C_FLAGS_RELEASE=${DCMAKE_C_FLAGS_RELEASE}"
+            "-DCMAKE_CXX_FLAGS=${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")
+   # SET(PROTOBUF_REPO "https://github.com/google/protobuf.git")
+  #  SET(PROTOBUF_TAG "9f75c5aa851cd877fb0d93ccc31b8567a6706546")
+    SET(PROTOBUF_REPO http://admin@localhost:8080/r/protobuf.git)
    IF(MOBILE_INFERENCE)
        # The reason why the official version is not used is described in
        # https://github.com/PaddlePaddle/Paddle/issues/6114

--- a/cmake/external/python.cmake
+++ b/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})

--- a/cmake/external/xxhash.cmake
+++ b/cmake/external/xxhash.cmake
@@ -14,7 +14,31 @@ ELSE()
  ENDIF(APPLE)
 ENDIF()

-ExternalProject_Add(
+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"
@@ -28,9 +52,14 @@ ExternalProject_Add(
      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)

--- a/cmake/external/zlib.cmake
+++ b/cmake/external/zlib.cmake
@@ -31,8 +31,9 @@ INCLUDE_DIRECTORIES(${THIRD_PARTY_PATH}/install) # For Paddle code to include zl
 ExternalProject_Add(
    extern_zlib
    ${EXTERNAL_PROJECT_LOG_ARGS}
-    GIT_REPOSITORY  "https://github.com/madler/zlib.git"
-    GIT_TAG         "v1.2.8"
+        #    GIT_REPOSITORY  "https://github.com/madler/zlib.git"
+            GIT_REPOSITORY  "http://admin@localhost:8080/r/zlib.git"
+#    GIT_TAG         "v1.2.8"
    PREFIX          ${ZLIB_SOURCES_DIR}
    UPDATE_COMMAND  ""
    CMAKE_ARGS      -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}

--- a/cmake/generic.cmake
+++ b/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)
+        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 +292,50 @@ 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)
+  set(${TARGET_NAME}_dummy_flag "")
+  if(${sep_library_STATIC})
+    set(${TARGET_NAME}_dummy_flag "STATIC")
+  elseif(${sep_library_SHARED})
+    set(${TARGET_NAME}_dummy_flag "SHARED")
+  endif()
+  cmake_parse_arguments(sep_library "${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}")
+      #            set(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/${TARGET_NAME}_dummy_${dummy_index}.c)
+      #            file(WRITE ${dummyfile} "const char *dummy_${TARGET_NAME}_${dummy_index} = \"${dummyfile}\";")
+      #            cc_library(${TARGET_NAME}_dummy_lib_${dummy_index} ${${TARGET_NAME}_dummy_flag} SRCS ${dummyfile} DEPS ${dummy_list})
+      cc_library(${TARGET_NAME}_dummy_lib_${dummy_index} ${${TARGET_NAME}_dummy_flag} 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()
+  cc_library(${TARGET_NAME} ${${TARGET_NAME}_dummy_flag} SRCS ${sep_library_SRCS} DEPS ${${TARGET_NAME}_dummy_list})
+endfunction(sep_library)
+
 function(cc_binary TARGET_NAME)
  set(options "")
  set(oneValueArgs "")

--- a/cmake/inference_lib.cmake
+++ b/cmake/inference_lib.cmake
@@ -31,10 +31,32 @@ function(copy TARGET)
    foreach(index RANGE ${len})
        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})
+          add_custom_command(TARGET ${TARGET} PRE_BUILD
+          COMMAND ${CMAKE_COMMAND} -E copy "${src_file}" "${dst}"
+          COMMENT "copying ${src_file} -> ${dst}")
+        endforeach()
+        else() # not windows
        add_custom_command(TARGET ${TARGET} PRE_BUILD 
-          COMMAND mkdir -p "${dst}"
-          COMMAND cp -r "${src}" "${dst}"
+          COMMAND ${CMAKE_COMMAND} -E make_directory  "${dst}"
+          COMMAND ${CMAKE_COMMAND} -E copy "${src_files}" "${dst}"
          COMMENT "copying ${src} -> ${dst}")
+        endif(WIN32)
    endforeach()
 endfunction()


--- a/paddle/fluid/CMakeLists.txt
+++ b/paddle/fluid/CMakeLists.txt
@@ -4,11 +4,14 @@ 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)
+if(WITH_INFERENCE)
+  # NOTE: please add subdirectory inference at last.
+  add_subdirectory(inference)
+endif()
+
 add_subdirectory(train)
--- a/paddle/fluid/framework/ir/attention_lstm_fuse_pass.cc
+++ b/paddle/fluid/framework/ir/attention_lstm_fuse_pass.cc
@@ -211,12 +211,12 @@ void PrepareLSTMWeight(const LoDTensor& W_forget_w0,
  VLOG(3) << "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];

--- a/paddle/fluid/framework/ir/node.h
+++ b/paddle/fluid/framework/ir/node.h
@@ -28,7 +28,7 @@ namespace ir {
 class Node {
 public:
  enum class Type { kOperation, kVariable };
-  static constexpr char kControlDepVarName[] = "__control_var";
+  static constexpr const char kControlDepVarName[] = "__control_var";

  Type NodeType() const { return type_; }


--- a/paddle/fluid/framework/ir/pass.h
+++ b/paddle/fluid/framework/ir/pass.h
@@ -207,7 +207,7 @@ struct PassRegistrar : public Registrar {
    return 0;                                                         \
  }                                                                   \
  static ::paddle::framework::ir::PassRegistrar<pass_class>           \
-      &__pass_tmp_registrar_##pass_type##__ __attribute__((unused)) = \
+      &__pass_tmp_registrar_##pass_type##__ __UNUSED__() = \
          __pass_registrar_##pass_type##__

 #define USE_PASS(pass_type)                                           \
@@ -215,7 +215,7 @@ struct PassRegistrar : public Registrar {
      __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)) = \
+  static int use_pass_itself_##pass_type##_ __UNUSED__() = \
      TouchPassRegistrar_##pass_type()

 }  // namespace ir

--- a/paddle/fluid/framework/operator.cc
+++ b/paddle/fluid/framework/operator.cc
@@ -153,11 +153,14 @@ void OperatorBase::Run(const Scope& scope, const platform::Place& place) {
  // 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(3) << place << " " << DebugStringEx(&scope);

--- a/paddle/fluid/inference/CMakeLists.txt
+++ b/paddle/fluid/inference/CMakeLists.txt
@@ -30,7 +30,11 @@ if (WITH_GPU AND TENSORRT_FOUND)
 endif()

 # Create static library
+if(WIN32)
+sep_library(paddle_fluid DEPS ${fluid_modules} ${STATIC_INFERENCE_APIS} zero_copy_tensor reset_tensor_array)
+else()
 cc_library(paddle_fluid DEPS ${fluid_modules} ${STATIC_INFERENCE_APIS} zero_copy_tensor reset_tensor_array)
+endif()

 if(NOT APPLE)
  # TODO(liuyiqu: Temporarily disable the link flag because it is not support on Mac.

--- a/paddle/fluid/inference/analysis/helper.h
+++ b/paddle/fluid/inference/analysis/helper.h
@@ -126,7 +126,11 @@ T &GetFromScope(const framework::Scope &scope, const std::string &name) {

 static void ExecShellCommand(const std::string &cmd, std::string *message) {
  char buffer[128];
+#if !defined(_WIN32)
  std::shared_ptr<FILE> pipe(popen(cmd.c_str(), "r"), pclose);
+#else
+  std::shared_ptr<FILE> pipe(_popen(cmd.c_str(), "r"), _pclose);
+#endif  // _WIN32
  if (!pipe) {
    LOG(ERROR) << "error running command: " << cmd;
    return;

--- a/paddle/fluid/inference/api/api_impl.cc
+++ b/paddle/fluid/inference/api/api_impl.cc
@@ -75,6 +75,10 @@ bool NativePaddlePredictor::Init(
  }
 #endif

+  // windows has no support for openblas multi-thread
+#ifdef _WIN32
+	FLAGS_paddle_num_threads = 1;
+#endif
  // no matter with or without MKLDNN
  paddle::platform::SetNumThreads(FLAGS_paddle_num_threads);


--- a/paddle/fluid/inference/api/helper.h
+++ b/paddle/fluid/inference/api/helper.h
@@ -15,8 +15,12 @@
 #pragma once

 #include <glog/logging.h>
+#if !defined(_WIN32)
 #include <sys/time.h>
+#else
+#endif
 #include <chrono>  // NOLINT
+#include <iterator>
 #include <numeric>
 #include <sstream>
 #include <string>

--- a/paddle/fluid/operators/CMakeLists.txt
+++ b/paddle/fluid/operators/CMakeLists.txt
@@ -86,7 +86,8 @@ function(op_library TARGET)
    # remove windows unsupported op, because windows has no nccl, no warpctc such ops.
    foreach(windows_unsupport_op "nccl_op" "gen_nccl_id_op" "warpctc_op" "hierarchical_sigmoid_op"
     "crf_decoding_op" "select_op" "lstmp_op" "gru_op" "fusion_gru_op" "lstm_op" "fusion_lstm_op" "cumsum_op"
-      "fusion_seqconv_eltadd_relu_op" "channel_send_op" "channel_create_op" "channel_close_op" "channel_recv_op")
+      "fusion_seqconv_eltadd_relu_op" "channel_send_op" "channel_create_op" "channel_close_op" "channel_recv_op"
+            "fusion_seqexpand_concat_fc_op" "attention_lstm_op" "fused_embedding_fc_lstm_op" "fc_op")
        if ("${TARGET}" STREQUAL "${windows_unsupport_op}")
          return()
        endif()
@@ -301,8 +302,10 @@ op_library(flatten_op DEPS reshape_op)
 op_library(sequence_pad_op DEPS sequence_padding)
 op_library(unstack_op DEPS stack_op)
 op_library(fake_quantize_op DEPS memory)
+if (NOT WIN32)
 op_library(crf_decoding_op DEPS jit_kernel)
 op_library(fusion_lstm_op DEPS jit_kernel)
+endif(NOT WIN32)
 if (WITH_GPU)
    op_library(conv_op DEPS vol2col depthwise_conv im2col)
    op_library(layer_norm_op DEPS cub)

--- a/paddle/fluid/operators/elementwise_op_function.h
+++ b/paddle/fluid/operators/elementwise_op_function.h
@@ -111,6 +111,17 @@ class RowwiseTransformIterator<T, platform::CPUDeviceContext>
    return *this;
  }

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

+  MidWiseTransformIterator<T, platform::CPUDeviceContext> &operator+(int n) {
+    while(n-- > 0) {
+      ++j_;
+      if (UNLIKELY(j_ == post_)) {
+        ++i_;
+        j_ = 0;
+        if (UNLIKELY(i_ == n_)) {
+          i_ = 0;
+        }
+      }
+    }
+
+    return *this;
+  }
+
  bool operator==(const MidWiseTransformIterator<T, platform::CPUDeviceContext>
                      &rhs) const {
    return (ptr_ + i_) == &(*rhs);

--- a/paddle/fluid/operators/math/CMakeLists.txt
+++ b/paddle/fluid/operators/math/CMakeLists.txt
@@ -75,7 +75,9 @@ if(WITH_GPU)
 endif()
 cc_test(concat_test SRCS concat_test.cc DEPS concat_and_split)
 cc_test(cpu_vec_test SRCS cpu_vec_test.cc DEPS blas cpu_info)
+if (NOT WIN32)
 cc_library(jit_kernel
    SRCS jit_kernel.cc jit_kernel_blas.cc jit_kernel_exp.cc jit_kernel_rnn.cc jit_kernel_crf_decode.cc
    DEPS cpu_info cblas)
 cc_test(jit_kernel_test SRCS jit_kernel_test.cc DEPS jit_kernel)
+endif()
--- a/paddle/fluid/platform/init.cc
+++ b/paddle/fluid/platform/init.cc
@@ -137,7 +137,9 @@ void InitGLOG(const std::string &prog_name) {
  // glog will not hold the ARGV[0] inside.
  // Use strdup to alloc a new string.
  google::InitGoogleLogging(strdup(prog_name.c_str()));
+#ifndef _WIN32
  google::InstallFailureSignalHandler();
+#endif
 }

 }  // namespace framework

--- a/paddle/fluid/platform/nccl_helper.h
+++ b/paddle/fluid/platform/nccl_helper.h
@@ -12,6 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.

+#ifndef _WIN32
 #pragma once

 #include <stdio.h>
@@ -149,3 +150,4 @@ struct NCCLContextMap {

 }  // namespace platform
 }  // namespace paddle
+#endif
\ No newline at end of file
--- a/paddle/fluid/platform/variant.h
+++ b/paddle/fluid/platform/variant.h
@@ -42,3 +42,11 @@ limitations under the License. */
 #include <boost/mpl/comparison.hpp>
 #include <boost/mpl/less_equal.hpp>
 #include <boost/variant.hpp>
+
+// some platform-independent defintion
+#if defined(WIN32) || defined(_WIN32) || defined(__WIN32__)
+#define __UNUSED__()
+#define __builtin_expect(EXP, C)  (EXP)
+#else
+#define __UNUSED__() __attribute__((unused))
+#endif
\ No newline at end of file
--- a/paddle/fluid/pybind/CMakeLists.txt
+++ b/paddle/fluid/pybind/CMakeLists.txt
@@ -2,8 +2,8 @@
 set(PYBIND_DEPS pybind python proto_desc memory executor prune  feed_fetch_method pass_builder)
 set(PYBIND_SRCS pybind.cc exception.cc protobuf.cc const_value.cc)
 if(NOT WIN32)
-list(APPEND PYBIND_DEPS parallel_executor profiler)
-list(APPEND PYBIND_SRCS recordio.cc)
+  list(APPEND PYBIND_DEPS parallel_executor profiler)
+  list(APPEND PYBIND_SRCS recordio.cc)
 endif()
 if(WITH_PYTHON)
  if(WITH_AMD_GPU)
@@ -21,5 +21,9 @@ if(WITH_PYTHON)
    endif(NOT APPLE AND NOT ANDROID AND NOT WIN32)
  endif(WITH_AMD_GPU)

+  if(WIN32)
+    target_link_libraries(paddle_pybind shlwapi)
+  endif(WIN32)
+
  cc_test(tensor_py_test SRCS tensor_py_test.cc DEPS python)
 endif(WITH_PYTHON)
--- a/paddle/fluid/pybind/pybind.cc
+++ b/paddle/fluid/pybind/pybind.cc
@@ -21,6 +21,13 @@ limitations under the License. */
 #include <utility>
 #include <vector>

+#if defined(_WIN32)
+#define NOMINMAX
+#define GLOG_NO_ABBREVIATED_SEVERITIES  // msvc conflict logging with windows.h
+#define GOOGLE_GLOG_DLL_DECL
+#include <Windows.h>
+#endif
+
 #include "paddle/fluid/framework/executor.h"
 #include "paddle/fluid/framework/feed_fetch_method.h"
 #include "paddle/fluid/framework/framework.pb.h"
@@ -29,7 +36,9 @@ limitations under the License. */
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/lod_tensor_array.h"
 #include "paddle/fluid/framework/op_registry.h"
+#ifndef _WIN32
 #include "paddle/fluid/framework/parallel_executor.h"
+#endif
 #include "paddle/fluid/framework/prune.h"
 #include "paddle/fluid/framework/reader.h"
 #include "paddle/fluid/framework/selected_rows.h"
@@ -50,7 +59,9 @@ limitations under the License. */
 #include "paddle/fluid/string/to_string.h"

 #ifdef PADDLE_WITH_CUDA
+#ifndef _WIN32
 #include "paddle/fluid/operators/nccl/nccl_gpu_common.h"
+#endif
 #include "paddle/fluid/platform/cuda_profiler.h"
 #include "paddle/fluid/platform/gpu_info.h"
 #endif
@@ -340,22 +351,25 @@ All parameter, weight, gradient are variables in Paddle.
      .def("get_lod_tensor_array",
           [](Variable &self) { return self.GetMutable<LoDTensorArray>(); },
           py::return_value_policy::reference)
-#ifdef PADDLE_WITH_CUDA
+#if (defined(PADDLE_WITH_CUDA) && !defined(_WIN32))
 	  .def("get_communicator",
           [](Variable &self) -> platform::Communicator * {
             return self.GetMutable<platform::Communicator>();
           },
           py::return_value_policy::reference)
-#endif
      .def("get_reader",
           [](Variable &self) -> framework::ReaderHolder * {
             PADDLE_ENFORCE(self.IsType<framework::ReaderHolder>());
             return self.GetMutable<framework::ReaderHolder>();
           },
-           py::return_value_policy::reference);
+           py::return_value_policy::reference)
+#endif
+;

+#if !defined(_WIN32)
  py::class_<framework::ReaderHolder>(m, "Reader", "")
      .def("reset", &framework::ReaderHolder::ResetAll);
+#endif

  using LoDTensorBlockingQueue =
      ::paddle::operators::reader::LoDTensorBlockingQueue;
@@ -480,7 +494,7 @@ All parameter, weight, gradient are variables in Paddle.
 #endif
                });;
 // clang-format on
-#ifdef PADDLE_WITH_CUDA
+#if (defined(PADDLE_WITH_CUDA) && !defined(_WIN32))
  py::class_<platform::Communicator>(m, "Communicator").def(py::init<>());
 #endif
  py::class_<platform::CUDAPlace>(m, "CUDAPlace")
@@ -617,11 +631,14 @@ All parameter, weight, gradient are variables in Paddle.
 #ifdef PADDLE_WITH_CUDA
  m.def("get_cuda_device_count", platform::GetCUDADeviceCount);

+#ifndef _WIN32
  m.def("nvprof_init", platform::CudaProfilerInit);
  m.def("nvprof_start", platform::CudaProfilerStart);
  m.def("nvprof_stop", platform::CudaProfilerStop);
 #endif
+#endif

+#ifndef _WIN32
  py::enum_<platform::ProfilerState>(m, "ProfilerState", py::arithmetic())
      .value("kDisabled", platform::ProfilerState::kDisabled)
      .value("kCPU", platform::ProfilerState::kCPU)
@@ -642,6 +659,7 @@ All parameter, weight, gradient are variables in Paddle.
  m.def("disable_profiler", platform::DisableProfiler);
  m.def("is_profiler_enabled", platform::IsProfileEnabled);
  m.def("reset_profiler", platform::ResetProfiler);
+#endif

  py::class_<ir::Pass, std::shared_ptr<ir::Pass>> pass(m, "Pass");
  pass.def(py::init())
@@ -670,6 +688,7 @@ All parameter, weight, gradient are variables in Paddle.
      .def("remove_pass",
           [](ir::PassBuilder &self, size_t idx) { self.RemovePass(idx); });

+#ifndef _WIN32
  // -- python binds for parallel executor.
  py::class_<ParallelExecutor> pe(m, "ParallelExecutor");
  py::class_<ExecutionStrategy> exec_strategy(pe, "ExecutionStrategy", R"DOC(
@@ -864,6 +883,7 @@ All parameter, weight, gradient are variables in Paddle.
      });

  BindRecordIOWriter(&m);
+#endif
  return m.ptr();
 }
 }  // namespace pybind

--- a/python/CMakeLists.txt
+++ b/python/CMakeLists.txt
@@ -46,22 +46,39 @@ endif()
 configure_file(${CMAKE_CURRENT_SOURCE_DIR}/setup.py.in
    ${CMAKE_CURRENT_BINARY_DIR}/setup.py)

-set(FLUID_CORE ${PADDLE_BINARY_DIR}/python/paddle/fluid/core.so)
+IF(WIN32)
+    # Python would use the .pyd by default under Windows series platform
+    set(FLUID_CORE ${PADDLE_BINARY_DIR}/python/paddle/fluid/core.pyd)
+ELSE()
+    set(FLUID_CORE ${PADDLE_BINARY_DIR}/python/paddle/fluid/core.so)
+ENDIF()
 add_custom_command(OUTPUT ${FLUID_CORE}
        COMMAND cmake -E copy $<TARGET_FILE:paddle_pybind> ${FLUID_CORE}
        DEPENDS paddle_pybind)
 add_custom_target(copy_paddle_pybind ALL DEPENDS ${FLUID_CORE})

-
-add_custom_command(OUTPUT ${PADDLE_PYTHON_BUILD_DIR}/.timestamp
+IF(WIN32)
+	add_custom_command(OUTPUT ${PADDLE_PYTHON_BUILD_DIR}/.timestamp
+#		COMMAND ${CMAKE_COMMAND} -E touch stub.cc
+		COMMAND ${CMAKE_COMMAND} -E make_directory ${PADDLE_BINARY_DIR}/python/paddle
+		COMMAND ${CMAKE_COMMAND} -E copy_directory ${PADDLE_SOURCE_DIR}/python/paddle ${PADDLE_BINARY_DIR}/python/paddle
+		COMMAND ${CMAKE_COMMAND} -E copy_directory ${PADDLE_SOURCE_DIR}/paddle/py_paddle ${PADDLE_BINARY_DIR}/python/
+		COMMAND ${CMAKE_COMMAND} -E env ${py_env} ${PYTHON_EXECUTABLE} setup.py bdist_wheel
+		COMMAND ${CMAKE_COMMAND} -E touch ${PADDLE_PYTHON_BUILD_DIR}/.timestamp
+		COMMAND ${CMAKE_COMMAND} -E remove_directory ${PADDLE_PYTHON_BUILD_DIR}/lib-python
+#		COMMAND ${CMAKE_COMMAND} -E copy_directory ${PADDLE_PYTHON_BUILD_DIR}/lib* ${PADDLE_PYTHON_BUILD_DIR}/lib-python
+		DEPENDS gen_proto_py copy_paddle_pybind ${FLUID_CORE} framework_py_proto profiler_py_proto ${PY_FILES} ${external_project_dependencies} ${COPY_PADDLE_MASTER})
+ELSE(WIN32)
+	add_custom_command(OUTPUT ${PADDLE_PYTHON_BUILD_DIR}/.timestamp
 		COMMAND touch stub.cc
 		COMMAND cp -r ${PADDLE_SOURCE_DIR}/python/paddle ${PADDLE_BINARY_DIR}/python
 		COMMAND cp -r ${PADDLE_SOURCE_DIR}/paddle/py_paddle ${PADDLE_BINARY_DIR}/python/
 		COMMAND env ${py_env} ${PYTHON_EXECUTABLE} setup.py bdist_wheel
 		COMMAND ${CMAKE_COMMAND} -E touch ${PADDLE_PYTHON_BUILD_DIR}/.timestamp
 		COMMAND ${CMAKE_COMMAND} -E remove_directory ${PADDLE_PYTHON_BUILD_DIR}/lib-python
-    COMMAND ${CMAKE_COMMAND} -E copy_directory ${PADDLE_PYTHON_BUILD_DIR}/lib.* ${PADDLE_PYTHON_BUILD_DIR}/lib-python
+		COMMAND ${CMAKE_COMMAND} -E copy_directory ${PADDLE_PYTHON_BUILD_DIR}/lib* ${PADDLE_PYTHON_BUILD_DIR}/lib-python
 		DEPENDS gen_proto_py copy_paddle_pybind ${FLUID_CORE} framework_py_proto profiler_py_proto ${PY_FILES} ${external_project_dependencies} ${COPY_PADDLE_MASTER})
+ENDIF()

 set(paddle_python_deps ${PADDLE_PYTHON_BUILD_DIR}/.timestamp ${MKL_DEPENDS})
 if(NOT WITH_FLUID_ONLY)

--- a/python/setup.py.in
+++ b/python/setup.py.in
@@ -9,7 +9,7 @@ class BinaryDistribution(Distribution):

 RC      = 0

-
+ext_name = '.dll' if os.name == 'nt' else '.so'

 def git_commit():
    try:
@@ -136,10 +136,13 @@ if '${WITH_FLUID_ONLY}'== 'OFF':
                   '${PADDLE_BINARY_DIR}/paddle/legacy/pserver/paddle_pserver_main',
                   '${PADDLE_BINARY_DIR}/paddle/scripts/paddle']

-package_data={'paddle.fluid': ['core.so']}
+package_data={'paddle.fluid': ['core' + (ext_name if os.name != 'nt' else '.pyd')]}
+if os.name == 'nt':
+    package_data['paddle.fluid'] += ['openblas' + ext_name]
+
 if '${WITH_FLUID_ONLY}'== 'OFF':
-    package_data['paddle.v2.master']=['libpaddle_master.so']
-    package_data['py_paddle']=['*.py','_swig_paddle.so']
+    package_data['paddle.v2.master']=['libpaddle_master' + ext_name]
+    package_data['py_paddle']=['*.py','_swig_paddle' +  + ext_name]

 package_dir={
    '': '${PADDLE_BINARY_DIR}/python',
@@ -153,13 +156,15 @@ if '${WITH_FLUID_ONLY}'== 'OFF':
    package_dir['py_paddle']='${PADDLE_BINARY_DIR}/python/py_paddle'

 # put all thirdparty libraries in paddle.libs
-package_data['paddle.libs']=['libwarpctc.so']
 libs_path='${PADDLE_BINARY_DIR}/python/paddle/libs'
-shutil.copy('${WARPCTC_LIBRARIES}', libs_path)
+if os.name != 'nt':
+    package_data['paddle.libs']= []
+    package_data['paddle.libs']=['libwarpctc' + ext_name]
+    shutil.copy('${WARPCTC_LIBRARIES}', libs_path)
 if '${WITH_MKL}' == 'ON':
    shutil.copy('${MKLML_LIB}', libs_path)
    shutil.copy('${MKLML_IOMP_LIB}', libs_path)
-    package_data['paddle.libs']+=['libmklml_intel.so','libiomp5.so']
+    package_data['paddle.libs']+=['libmklml_intel' + ext_name,'libiomp5' + ext_name]
 if '${CMAKE_BUILD_TYPE}' == 'Release':
    # only change rpath in Release mode.
    if '${WITH_MKLDNN}' == 'ON':
@@ -183,22 +188,30 @@ package_dir['paddle.libs']=libs_path
 # core.so is in paddle.fluid, thus paddle/fluid/../libs will pointer to above libraries.
 # This operation will fix https://github.com/PaddlePaddle/Paddle/issues/3213
 if '${CMAKE_BUILD_TYPE}' == 'Release':
+    if os.name != 'nt':
        # only change rpath in Release mode, since in Debug mode, core.so is too large to be changed.
        if "@APPLE@" == "1":
-        command = "install_name_tool -id \"@loader_path/../libs/\" ${PADDLE_BINARY_DIR}/python/paddle/fluid/core.so"
+            command = "install_name_tool -id \"@loader_path/../libs/\" ${PADDLE_BINARY_DIR}/python/paddle/fluid/core" + ext_name
        else:
-        command = "patchelf --set-rpath '$ORIGIN/../libs/' ${PADDLE_BINARY_DIR}/python/paddle/fluid/core.so"
+            command = "patchelf --set-rpath '$ORIGIN/../libs/' ${PADDLE_BINARY_DIR}/python/paddle/fluid/core" + ext_name
        if os.system(command) != 0:
            raise Exception("patch core.so failed, command: %s" % command)
        if '${WITH_FLUID_ONLY}'== 'OFF':
            # change rpath of _swig_paddle.so.
            if "@APPLE@" == "1":
-            command = "install_name_tool -id \"@loader_path/../paddle/libs/\" ${PADDLE_BINARY_DIR}/python/py_paddle/_swig_paddle.so"
+                command = "install_name_tool -id \"@loader_path/../paddle/libs/\" ${PADDLE_BINARY_DIR}/python/py_paddle/_swig_paddle" + ext_name
            else:
-            command = "patchelf --set-rpath '$ORIGIN/../paddle/libs/' ${PADDLE_BINARY_DIR}/python/py_paddle/_swig_paddle.so"
+                command = "patchelf --set-rpath '$ORIGIN/../paddle/libs/' ${PADDLE_BINARY_DIR}/python/py_paddle/_swig_paddle" + ext_name
            if os.system(command) != 0:
                raise Exception("patch _swig_paddle.so failed, command: %s" % command)

+if os.name == 'nt':
+    # fix the path separator under windows
+    fix_package_dir = {}
+    for k, v in package_dir.items():
+        fix_package_dir[k] = v.replace('/', '\\')
+    package_dir = fix_package_dir
+
 setup(name='${PACKAGE_NAME}',
      version='${PADDLE_VERSION}',
      description='Parallel Distributed Deep Learning',