add mkl,ctc support for windows

5a6d7fe2 · peizhilin · 0f085f0a · 5a6d7fe2 · 5a6d7fe2 · 5a6d7fe2
40 changed file
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -125,16 +125,12 @@ if(ANDROID OR IOS)
    add_definitions(-DPADDLE_MOBILE_INFERENCE)
 endif()
-if (APPLE OR WIN32)
+if (APPLE)
    set(WITH_MKL OFF CACHE STRING
-        "Disable MKL for building on mac and windows" FORCE)
+        "Disable MKL for building on mac" FORCE)
 endif()
 if (WIN32)
-    set(WITH_DSO OFF CACHE STRING
-            "Disable DSO when compiling for Windows" FORCE)
-    set(WITH_MKL OFF CACHE STRING
-            "Disable MKL when compiling for Windows" FORCE)
    set(WITH_DISTRIBUTE OFF CACHE STRING
            "Disable DISTRIBUTE when compiling for Windows" FORCE)
    set(WITH_C_API OFF CACHE STRING
@@ -207,10 +203,10 @@ include(external/xxhash)    # download xxhash
 include(external/dlpack)
 include(external/snappy)    # download snappy
 include(external/snappystream) # download snappystream
+include(external/warpctc)   # download, build, install warpctc
 if (NOT WIN32)
-# there is no official support of warpctc, nccl, cupti in windows
+# there is no official support of nccl, cupti in windows
-include(external/warpctc)   # download, build, install warpctc
 include(cupti)
 include(external/gzstream)
 endif (NOT WIN32)

--- a/cmake/cuda.cmake
+++ b/cmake/cuda.cmake
@@ -139,10 +139,12 @@ endfunction()
 message(STATUS "CUDA detected: " ${CUDA_VERSION})
 if (${CUDA_VERSION} LESS 7.0)
  set(paddle_known_gpu_archs ${paddle_known_gpu_archs})
+  add_definitions("-DPADDLE_CUDA_BINVER=\"60\"")
 elseif (${CUDA_VERSION} LESS 8.0) # CUDA 7.x
  set(paddle_known_gpu_archs ${paddle_known_gpu_archs7})
  list(APPEND CUDA_NVCC_FLAGS "-D_MWAITXINTRIN_H_INCLUDED")
  list(APPEND CUDA_NVCC_FLAGS "-D__STRICT_ANSI__")
+  add_definitions("-DPADDLE_CUDA_BINVER=\"70\"")
 elseif (${CUDA_VERSION} LESS 9.0) # CUDA 8.x
  set(paddle_known_gpu_archs ${paddle_known_gpu_archs8})
  list(APPEND CUDA_NVCC_FLAGS "-D_MWAITXINTRIN_H_INCLUDED")
@@ -150,6 +152,7 @@ elseif (${CUDA_VERSION} LESS 9.0) # CUDA 8.x
  # CUDA 8 may complain that sm_20 is no longer supported. Suppress the
  # warning for now.
  list(APPEND CUDA_NVCC_FLAGS "-Wno-deprecated-gpu-targets")
+  add_definitions("-DPADDLE_CUDA_BINVER=\"80\"")
 endif()
 include_directories(${CUDA_INCLUDE_DIRS})

--- a/cmake/cudnn.cmake
+++ b/cmake/cudnn.cmake
@@ -89,6 +89,7 @@ if(CUDNN_FOUND)
        if(NOT CUDNN_MAJOR_VERSION)
            set(CUDNN_VERSION "???")
        else()
+            add_definitions("-DPADDLE_CUDNN_BINVER=\"${CUDNN_MAJOR_VERSION}\"")
            math(EXPR CUDNN_VERSION
                "${CUDNN_MAJOR_VERSION} * 1000 +
                 ${CUDNN_MINOR_VERSION} * 100 + ${CUDNN_PATCHLEVEL_VERSION}")

--- a/cmake/external/cub.cmake
+++ b/cmake/external/cub.cmake
@@ -32,4 +32,4 @@ endif()
 add_dependencies(cub extern_cub)
-LIST(APPEND externl_project_dependencies cub)
+LIST(APPEND external_project_dependencies cub)
--- a/cmake/external/dlpack.cmake
+++ b/cmake/external/dlpack.cmake
@@ -28,4 +28,4 @@ endif()
 add_dependencies(dlpack extern_dlpack)
-LIST(APPEND externl_project_dependencies dlpack)
+LIST(APPEND external_project_dependencies dlpack)
--- a/cmake/external/mkldnn.cmake
+++ b/cmake/external/mkldnn.cmake
@@ -23,15 +23,14 @@ SET(MKLDNN_SOURCES_DIR    ${THIRD_PARTY_PATH}/mkldnn)
 SET(MKLDNN_INSTALL_DIR    ${THIRD_PARTY_PATH}/install/mkldnn)
 SET(MKLDNN_INC_DIR        "${MKLDNN_INSTALL_DIR}/include" CACHE PATH "mkldnn include directory." FORCE)
-IF(WIN32 OR APPLE)
+IF(APPLE)
    MESSAGE(WARNING
-        "Windows or Mac is not supported with MKLDNN in Paddle yet."
+        "Mac is not supported with MKLDNN in Paddle yet."
        "Force WITH_MKLDNN=OFF")
-    SET(WITH_MKLDNN OFF CACHE STRING "Disable MKLDNN in Windows and MacOS" FORCE)
+    SET(WITH_MKLDNN OFF CACHE STRING "Disable MKLDNN in MacOS" FORCE)
    return()
 ENDIF()
-SET(MKLDNN_LIB "${MKLDNN_INSTALL_DIR}/lib/libmkldnn.so" CACHE FILEPATH "mkldnn library." FORCE)
 MESSAGE(STATUS "Set ${MKLDNN_INSTALL_DIR}/lib to runtime path")
 SET(CMAKE_INSTALL_RPATH_USE_LINK_PATH TRUE)
 SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${MKLDNN_INSTALL_DIR}/lib")
@@ -44,10 +43,14 @@ IF(${CBLAS_PROVIDER} STREQUAL "MKLML")
 ELSE()
    MESSAGE(FATAL_ERROR "Should enable MKLML when build MKLDNN")
 ENDIF()
-SET(MKLDNN_FLAG "-Wno-error=strict-overflow -Wno-error=unused-result -Wno-error=array-bounds")
-SET(MKLDNN_FLAG "${MKLDNN_FLAG} -Wno-unused-result -Wno-unused-value")
+IF(NOT WIN32)
-SET(MKLDNN_CFLAG "${CMAKE_C_FLAGS} ${MKLDNN_FLAG}")
+    SET(MKLDNN_FLAG "-Wno-error=strict-overflow -Wno-error=unused-result -Wno-error=array-bounds")
-SET(MKLDNN_CXXFLAG "${CMAKE_CXX_FLAGS} ${MKLDNN_FLAG}")
+    SET(MKLDNN_FLAG "${MKLDNN_FLAG} -Wno-unused-result -Wno-unused-value")
+    SET(MKLDNN_CFLAG "${CMAKE_C_FLAGS} ${MKLDNN_FLAG}")
+    SET(MKLDNN_CXXFLAG "${CMAKE_CXX_FLAGS} ${MKLDNN_FLAG}")
+ENDIF(NOT WIN32)
 ExternalProject_Add(
    ${MKLDNN_PROJECT}
    ${EXTERNAL_PROJECT_LOG_ARGS}
@@ -58,8 +61,15 @@ ExternalProject_Add(
    UPDATE_COMMAND      ""
    CMAKE_ARGS          -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
    CMAKE_ARGS          -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
+    CMAKE_ARGS          -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
+    CMAKE_ARGS          -DCMAKE_CXX_FLAGS_RELEASE=${CMAKE_CXX_FLAGS_RELEASE}
+    CMAKE_ARGS          -DCMAKE_CXX_FLAGS_DEBUG=${CMAKE_CXX_FLAGS_DEBUG}
+    CMAKE_ARGS          -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}
+    CMAKE_ARGS          -DCMAKE_C_FLAGS_DEBUG=${CMAKE_C_FLAGS_DEBUG}
+    CMAKE_ARGS          -DCMAKE_C_FLAGS_RELEASE=${CMAKE_C_FLAGS_RELEASE}
    CMAKE_ARGS          -DCMAKE_INSTALL_PREFIX=${MKLDNN_INSTALL_DIR}
    CMAKE_ARGS          -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}
+    CMAKE_ARGS          -DCMAKE_POSITION_INDEPENDENT_CODE=ON
    CMAKE_ARGS          -DMKLROOT=${MKLML_ROOT}
    CMAKE_ARGS          -DCMAKE_C_FLAGS=${MKLDNN_CFLAG}
    CMAKE_ARGS          -DCMAKE_CXX_FLAGS=${MKLDNN_CXXFLAG}
@@ -67,6 +77,11 @@ ExternalProject_Add(
    CMAKE_CACHE_ARGS    -DCMAKE_INSTALL_PREFIX:PATH=${MKLDNN_INSTALL_DIR}
                        -DMKLROOT:PATH=${MKLML_ROOT}
 )
+if(WIN32)
+    SET(MKLDNN_LIB "${MKLDNN_INSTALL_DIR}/lib/mkldnn.lib" CACHE FILEPATH "mkldnn library." FORCE)
+else(WIN32)
+    SET(MKLDNN_LIB "${MKLDNN_INSTALL_DIR}/lib/libmkldnn.so" CACHE FILEPATH "mkldnn library." FORCE)
+endif(WIN32)
 ADD_LIBRARY(shared_mkldnn SHARED IMPORTED GLOBAL)
 SET_PROPERTY(TARGET shared_mkldnn PROPERTY IMPORTED_LOCATION ${MKLDNN_LIB})
@@ -85,10 +100,14 @@ ADD_DEPENDENCIES(mkldnn ${MKLDNN_PROJECT})
 # copy the real so.0 lib to install dir
 # it can be directly contained in wheel or capi
-SET(MKLDNN_SHARED_LIB ${MKLDNN_INSTALL_DIR}/libmkldnn.so.0)
+if(WIN32)
-ADD_CUSTOM_COMMAND(OUTPUT ${MKLDNN_SHARED_LIB}
+    SET(MKLDNN_SHARED_LIB ${MKLDNN_INSTALL_DIR}/lib/mkldnn.dll)
-    COMMAND cp ${MKLDNN_LIB} ${MKLDNN_SHARED_LIB}
+else(WIN32)
+    SET(MKLDNN_SHARED_LIB ${MKLDNN_INSTALL_DIR}/libmkldnn.so.0)
+    ADD_CUSTOM_COMMAND(OUTPUT ${MKLDNN_SHARED_LIB}
+            COMMAND ${CMAKE_COMMAND} -E copy ${MKLDNN_LIB} ${MKLDNN_SHARED_LIB}
            DEPENDS mkldnn)
+endif(WIN32)
 ADD_CUSTOM_TARGET(mkldnn_shared_lib ALL DEPENDS ${MKLDNN_SHARED_LIB})
 IF(WITH_C_API)

--- a/cmake/external/mklml.cmake
+++ b/cmake/external/mklml.cmake
@@ -16,44 +16,55 @@ IF(NOT ${WITH_MKLML})
  return()
 ENDIF(NOT ${WITH_MKLML})
-IF(WIN32 OR APPLE)
+IF(APPLE)
    MESSAGE(WARNING
-        "Windows or Mac is not supported with MKLML in Paddle yet."
+        "Mac is not supported with MKLML in Paddle yet."
        "Force WITH_MKLML=OFF")
    SET(WITH_MKLML OFF CACHE STRING "Disable MKLML package in Windows and MacOS" FORCE)
    return()
 ENDIF()
 INCLUDE(ExternalProject)
-SET(MKLML_PROJECT       "extern_mklml")
-IF((NOT DEFINED MKLML_VER) OR (NOT DEFINED MKLML_URL))
-  MESSAGE(STATUS "use pre defined download url")
-  SET(MKLML_VER "mklml_lnx_2019.0.20180710" CACHE STRING "" FORCE)
-  SET(MKLML_URL "http://paddlepaddledeps.cdn.bcebos.com/${MKLML_VER}.tgz" CACHE STRING "" FORCE)
-ENDIF()
-MESSAGE(STATUS "MKLML_VER: ${MKLML_VER}, MKLML_URL: ${MKLML_URL}")
-SET(MKLML_SOURCE_DIR    "${THIRD_PARTY_PATH}/mklml")
-SET(MKLML_DOWNLOAD_DIR  "${MKLML_SOURCE_DIR}/src/${MKLML_PROJECT}")
 SET(MKLML_DST_DIR       "mklml")
 SET(MKLML_INSTALL_ROOT  "${THIRD_PARTY_PATH}/install")
 SET(MKLML_INSTALL_DIR   ${MKLML_INSTALL_ROOT}/${MKLML_DST_DIR})
 SET(MKLML_ROOT          ${MKLML_INSTALL_DIR})
 SET(MKLML_INC_DIR       ${MKLML_ROOT}/include)
 SET(MKLML_LIB_DIR       ${MKLML_ROOT}/lib)
-SET(MKLML_LIB           ${MKLML_LIB_DIR}/libmklml_intel.so)
+if(WIN32)
-SET(MKLML_IOMP_LIB      ${MKLML_LIB_DIR}/libiomp5.so)
+    SET(MKLML_LIB                 ${MKLML_LIB_DIR}/mklml.lib)
+    SET(MKLML_IOMP_LIB            ${MKLML_LIB_DIR}/libiomp5md.lib)
+    SET(MKLML_SHARED_LIB          ${MKLML_LIB_DIR}/mklml.dll)
+    SET(MKLML_SHARED_IOMP_LIB     ${MKLML_LIB_DIR}/libiomp5md.dll)
+else()
+    SET(MKLML_LIB                 ${MKLML_LIB_DIR}/libmklml_intel.so)
+    SET(MKLML_IOMP_LIB            ${MKLML_LIB_DIR}/libiomp5.so)
+    SET(MKLML_SHARED_LIB          ${MKLML_LIB_DIR}/libmklml_intel.so)
+    SET(MKLML_SHARED_IOMP_LIB     ${MKLML_LIB_DIR}/libiomp5.so)
+endif()
 SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${MKLML_ROOT}/lib")
-INCLUDE_DIRECTORIES(${MKLML_INC_DIR})
+if(WIN32)
+    MESSAGE(WARNING
+    "Please download the MKLML and and put it at " ${THIRD_PARTY_PATH}/install/mklml)
+else()
+    SET(MKLML_PROJECT       "extern_mklml")
+    IF((NOT DEFINED MKLML_VER) OR (NOT DEFINED MKLML_URL))
+      MESSAGE(STATUS "use pre defined download url")
+      SET(MKLML_VER "mklml_lnx_2019.0.20180710" CACHE STRING "" FORCE)
+      SET(MKLML_URL "http://paddlepaddledeps.cdn.bcebos.com/${MKLML_VER}.tgz" CACHE STRING "" FORCE)
+    ENDIF()
+    MESSAGE(STATUS "MKLML_VER: ${MKLML_VER}, MKLML_URL: ${MKLML_URL}")
+    SET(MKLML_SOURCE_DIR    "${THIRD_PARTY_PATH}/mklml")
+    SET(MKLML_DOWNLOAD_DIR  "${MKLML_SOURCE_DIR}/src/${MKLML_PROJECT}")
-FILE(WRITE ${MKLML_DOWNLOAD_DIR}/CMakeLists.txt
+    FILE(WRITE ${MKLML_DOWNLOAD_DIR}/CMakeLists.txt
      "PROJECT(MKLML)\n"
      "cmake_minimum_required(VERSION 3.0)\n"
      "install(DIRECTORY ${MKLML_VER}/include ${MKLML_VER}/lib \n"
      "        DESTINATION ${MKLML_DST_DIR})\n")
-ExternalProject_Add(
+    ExternalProject_Add(
        ${MKLML_PROJECT}
        ${EXTERNAL_PROJECT_LOG_ARGS}
        PREFIX                ${MKLML_SOURCE_DIR}
@@ -64,11 +75,17 @@ ExternalProject_Add(
        UPDATE_COMMAND        ""
        CMAKE_ARGS            -DCMAKE_INSTALL_PREFIX=${MKLML_INSTALL_ROOT}
        CMAKE_CACHE_ARGS      -DCMAKE_INSTALL_PREFIX:PATH=${MKLML_INSTALL_ROOT}
-)
+    )
+endif()
+INCLUDE_DIRECTORIES(${MKLML_INC_DIR})
 ADD_LIBRARY(mklml SHARED IMPORTED GLOBAL)
 SET_PROPERTY(TARGET mklml PROPERTY IMPORTED_LOCATION ${MKLML_LIB})
-ADD_DEPENDENCIES(mklml ${MKLML_PROJECT})
+if(NOT WIN32)
+    ADD_DEPENDENCIES(mklml ${MKLML_PROJECT})
+endif()
 LIST(APPEND external_project_dependencies mklml)
 IF(WITH_C_API)

--- a/cmake/external/python.cmake
+++ b/cmake/external/python.cmake
@@ -23,9 +23,12 @@ FIND_PACKAGE(PythonLibs ${PY_VERSION})
 if(WIN32)
    execute_process(COMMAND "${PYTHON_EXECUTABLE}" "-c"
-"from distutils import sysconfig as s;import sys;import struct;
+"from distutils import sysconfig as s;import sys;import struct;import sysconfig;
 print(sys.prefix);
 print(s.get_config_var('LDVERSION') or s.get_config_var('VERSION'));
+print(sysconfig.get_platform());
+print(sysconfig.get_config_var('py_version_nodot'));
+print(sysconfig.get_config_var('SOABI'));
 "
            RESULT_VARIABLE _PYTHON_SUCCESS
            OUTPUT_VARIABLE _PYTHON_VALUES
@@ -41,6 +44,9 @@ print(s.get_config_var('LDVERSION') or s.get_config_var('VERSION'));
    string(REGEX REPLACE "\n" ";" _PYTHON_VALUES ${_PYTHON_VALUES})
    list(GET _PYTHON_VALUES 0 PYTHON_PREFIX)
    list(GET _PYTHON_VALUES 1 PYTHON_LIBRARY_SUFFIX)
+    list(GET _PYTHON_VALUES 2 SYS_PLATFORM)
+    list(GET _PYTHON_VALUES 3 PYTHON_SHORT_VERSION_NODOT)
+    list(GET _PYTHON_VALUES 4 PYTHON_SOABI)
    # Make sure all directory separators are '/'
    string(REGEX REPLACE "\\\\" "/" PYTHON_PREFIX ${PYTHON_PREFIX})

--- a/cmake/external/warpctc.cmake
+++ b/cmake/external/warpctc.cmake
@@ -26,25 +26,33 @@ SET(WARPCTC_INCLUDE_DIR "${WARPCTC_INSTALL_DIR}/include"
 # Used in unit test test_WarpCTCLayer
 SET(WARPCTC_LIB_DIR "${WARPCTC_INSTALL_DIR}/lib"
    CACHE PATH "Warp-ctc Library Directory" FORCE)
-SET(WARPCTC_LIBRARIES "${WARPCTC_INSTALL_DIR}/lib/libwarpctc${CMAKE_SHARED_LIBRARY_SUFFIX}"
-    CACHE FILEPATH "Warp-ctc Library" FORCE)
-IF(CMAKE_CXX_COMPILER_ID STREQUAL "Clang" OR CMAKE_CXX_COMPILER_ID STREQUAL "AppleClang" )
+IF(CMAKE_CXX_COMPILER_ID STREQUAL "Clang" OR CMAKE_CXX_COMPILER_ID STREQUAL "AppleClang" OR WIN32)
    SET(USE_OMP OFF)
 ELSE()
    SET(USE_OMP ON)
 ENDIF()
+IF(WIN32)
+    SET(WARPCTC_REPOSITORY "https://github.com/wopeizl/warp-ctc.git")
+ELSE()
+    SET(WARPCTC_REPOSITORY "https://github.com/dzhwinter/warp-ctc.git")
+ENDIF()
 ExternalProject_Add(
    extern_warpctc
    ${EXTERNAL_PROJECT_LOG_ARGS}
-    GIT_REPOSITORY  "https://github.com/dzhwinter/warp-ctc.git"
+    GIT_REPOSITORY ${WARPCTC_REPOSITORY}
    PREFIX          ${WARPCTC_SOURCES_DIR}
    UPDATE_COMMAND  ""
    CMAKE_ARGS      -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
                    -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
-                    -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
                    -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}
+                    -DCMAKE_C_FLAGS_DEBUG=${CMAKE_C_FLAGS_DEBUG}
+                    -DCMAKE_C_FLAGS_RELEASE=${CMAKE_C_FLAGS_RELEASE}
+                    -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
+                    -DCMAKE_CXX_FLAGS_RELEASE=${CMAKE_CXX_FLAGS_RELEASE}
+                    -DCMAKE_CXX_FLAGS_DEBUG=${CMAKE_CXX_FLAGS_DEBUG}
                    -DCMAKE_INSTALL_PREFIX=${WARPCTC_INSTALL_DIR}
                    -DWITH_GPU=${WITH_GPU}
                    -DWITH_OMP=${USE_OMP}
@@ -59,6 +67,18 @@ ExternalProject_Add(
                     -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON
                     -DCMAKE_INSTALL_PREFIX:PATH=${WARPCTC_INSTALL_DIR}
 )
+IF(WIN32)
+    IF(NOT EXISTS "${WARPCTC_INSTALL_DIR}/lib/warpctc${CMAKE_SHARED_LIBRARY_SUFFIX}")
+        add_custom_command(TARGET extern_warpctc POST_BUILD
+                COMMAND cmake -E copy ${WARPCTC_INSTALL_DIR}/bin/warpctc${CMAKE_SHARED_LIBRARY_SUFFIX} ${WARPCTC_INSTALL_DIR}/lib/warpctc${CMAKE_SHARED_LIBRARY_SUFFIX}
+                )
+    ENDIF()
+    SET(WARPCTC_LIBRARIES "${WARPCTC_INSTALL_DIR}/lib/warpctc${CMAKE_SHARED_LIBRARY_SUFFIX}"
+            CACHE FILEPATH "Warp-ctc Library" FORCE)
+else(WIN32)
+    SET(WARPCTC_LIBRARIES "${WARPCTC_INSTALL_DIR}/lib/libwarpctc${CMAKE_SHARED_LIBRARY_SUFFIX}"
+            CACHE FILEPATH "Warp-ctc Library" FORCE)
+ENDIF(WIN32)
 MESSAGE(STATUS "warp-ctc library: ${WARPCTC_LIBRARIES}")
 INCLUDE_DIRECTORIES(${WARPCTC_INCLUDE_DIR}) # For warpctc code to include its headers.

--- a/cmake/external/xbyak.cmake
+++ b/cmake/external/xbyak.cmake
@@ -13,8 +13,8 @@
 # limitations under the License.
 set(WITH_XBYAK ON)
-if(WIN32 OR APPLE)
+if(APPLE)
-    SET(WITH_XBYAK OFF CACHE STRING "Disable XBYAK in Windows and MacOS" FORCE)
+    SET(WITH_XBYAK OFF CACHE STRING "Disable XBYAK in MacOS" FORCE)
    return()
 endif()

--- a/cmake/generic.cmake
+++ b/cmake/generic.cmake
@@ -267,7 +267,11 @@ function(cc_library TARGET_NAME)
          list(APPEND cc_library_DEPS dynload_mklml)
        endif()
        add_dependencies(${TARGET_NAME} mklml)
+        if(WIN32)
+          target_link_libraries(${TARGET_NAME} ${MKLML_IOMP_LIB})
+        else(WIN32)
          target_link_libraries(${TARGET_NAME} "-L${MKLML_LIB_DIR} -liomp5 -Wl,--as-needed")
+        endif(WIN32)
      endif()
      # remove link to python, see notes at:
      # https://github.com/pybind/pybind11/blob/master/docs/compiling.rst#building-manually

--- a/cmake/inference_lib.cmake
+++ b/cmake/inference_lib.cmake
@@ -115,20 +115,20 @@ if (NOT PROTOBUF_FOUND OR WIN32)
            )
 endif ()
-if (NOT CBLAS_FOUND)
+if (WITH_MKLML)
-    set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/openblas")
-    copy(openblas_lib
-            SRCS ${CBLAS_INSTALL_DIR}/lib ${CBLAS_INSTALL_DIR}/include
-            DSTS ${dst_dir} ${dst_dir}
-            DEPS extern_openblas
-            )
-elseif (WITH_MKLML)
    set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/mklml")
    copy(mklml_lib
            SRCS ${MKLML_LIB} ${MKLML_IOMP_LIB} ${MKLML_INC_DIR}
            DSTS ${dst_dir}/lib ${dst_dir}/lib ${dst_dir}
            DEPS mklml
            )
+elseif (NOT CBLAS_FOUND OR WIN32)
+    set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/openblas")
+    copy(openblas_lib
+            SRCS ${CBLAS_INSTALL_DIR}/lib ${CBLAS_INSTALL_DIR}/include
+            DSTS ${dst_dir} ${dst_dir}
+            DEPS extern_openblas
+            )
 endif ()
 if (WITH_MKLDNN)

--- a/cmake/operators.cmake
+++ b/cmake/operators.cmake
@@ -84,7 +84,7 @@ function(op_library TARGET)
    endif()
    if (WIN32)
    # remove windows unsupported op, because windows has no nccl, no warpctc such ops.
-    foreach(windows_unsupport_op "nccl_op" "gen_nccl_id_op" "warpctc_op")
+    foreach(windows_unsupport_op "nccl_op" "gen_nccl_id_op")
        if ("${TARGET}" STREQUAL "${windows_unsupport_op}")
          return()
        endif()

--- a/cmake/simd.cmake
+++ b/cmake/simd.cmake
@@ -57,46 +57,43 @@ int main()
    return 0;
 }" SSE3_FOUND)
-# disable AVX by default on windows
+# Check AVX
-if(NOT WIN32)
+set(CMAKE_REQUIRED_FLAGS ${AVX_FLAG})
-    # Check AVX
+set(AVX_FOUND_EXITCODE 1 CACHE STRING "Result from TRY_RUN" FORCE)
-    set(CMAKE_REQUIRED_FLAGS ${AVX_FLAG})
+CHECK_CXX_SOURCE_RUNS("
-    set(AVX_FOUND_EXITCODE 1 CACHE STRING "Result from TRY_RUN" FORCE)
+#include <immintrin.h>
-    CHECK_CXX_SOURCE_RUNS("
+int main()
-    #include <immintrin.h>
+{
-    int main()
-    {
    __m256 a = _mm256_set_ps (-1.0f, 2.0f, -3.0f, 4.0f, -1.0f, 2.0f, -3.0f, 4.0f);
    __m256 b = _mm256_set_ps (1.0f, 2.0f, 3.0f, 4.0f, 1.0f, 2.0f, 3.0f, 4.0f);
    __m256 result = _mm256_add_ps (a, b);
    return 0;
-    }" AVX_FOUND)
+}" AVX_FOUND)
-    # Check AVX 2
+# Check AVX 2
-    set(CMAKE_REQUIRED_FLAGS ${AVX2_FLAG})
+set(CMAKE_REQUIRED_FLAGS ${AVX2_FLAG})
-    set(AVX2_FOUND_EXITCODE 1 CACHE STRING "Result from TRY_RUN" FORCE)
+set(AVX2_FOUND_EXITCODE 1 CACHE STRING "Result from TRY_RUN" FORCE)
-    CHECK_CXX_SOURCE_RUNS("
+CHECK_CXX_SOURCE_RUNS("
-    #include <immintrin.h>
+#include <immintrin.h>
-    int main()
+int main()
-    {
+{
    __m256i a = _mm256_set_epi32 (-1, 2, -3, 4, -1, 2, -3, 4);
    __m256i result = _mm256_abs_epi32 (a);
    return 0;
-    }" AVX2_FOUND)
+}" AVX2_FOUND)
-    # Check AVX512F
+# Check AVX512F
-    set(CMAKE_REQUIRED_FLAGS ${AVX512F_FLAG})
+set(CMAKE_REQUIRED_FLAGS ${AVX512F_FLAG})
-    set(AVX512F_FOUND_EXITCODE 1 CACHE STRING "Result from TRY_RUN" FORCE)
+set(AVX512F_FOUND_EXITCODE 1 CACHE STRING "Result from TRY_RUN" FORCE)
-    CHECK_CXX_SOURCE_RUNS("
+CHECK_CXX_SOURCE_RUNS("
-    #include <immintrin.h>
+#include <immintrin.h>
-    int main()
+int main()
-    {
+{
    __m512i a = _mm512_set_epi32 (-1, 2, -3, 4, -1, 2, -3, 4,
                                  13, -5, 6, -7, 9, 2, -6, 3);
    __m512i result = _mm512_abs_epi32 (a);
    return 0;
-    }" AVX512F_FOUND)
+}" AVX512F_FOUND)
-endif(NOT WIN32)
 set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_RETAINED})
 mark_as_advanced(MMX_FOUND SSE2_FOUND SSE3_FOUND AVX_FOUND AVX2_FOUND AVX512F_FOUND)
--- a/paddle/fluid/framework/CMakeLists.txt
+++ b/paddle/fluid/framework/CMakeLists.txt
@@ -15,8 +15,7 @@ function(windows_symbolic TARGET)
    file(GENERATE OUTPUT ${final_path}/.${src}.cu INPUT ${final_path}/${src}.cc)
    add_custom_command(OUTPUT ${final_path}/.${src}.cu
-            COMMAND ${CMAKE_COMMAND} -E remove ${final_path}/.${src}.cu
+            COMMAND ${CMAKE_COMMAND} -E copy_if_different "${final_path}/${src}.cc" "${final_path}/.${src}.cu"
-            COMMAND ${CMAKE_COMMAND} -E copy "${final_path}/${src}.cc" "${final_path}/.${src}.cu"
            COMMENT "create hidden file of ${src}.cu")
    add_custom_target(${TARGET} ALL DEPENDS .${src}.cu)
  endforeach()

--- a/paddle/fluid/framework/details/all_reduce_op_handle.cc
+++ b/paddle/fluid/framework/details/all_reduce_op_handle.cc
@@ -50,7 +50,7 @@ void AllReduceOpHandle::RunImpl() {
 // FIXME(typhoonzero): If scope0(global scope) have NCCL_ID_VAR,
 // this is a distributed or inter-process call, find a better way.
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
  if (NoDummyInputSize() == 1 &&
      local_scopes_[0]->FindLocalVar(NCCL_ID_VARNAME) == nullptr) {
 #else

--- a/paddle/fluid/framework/mixed_vector.h
+++ b/paddle/fluid/framework/mixed_vector.h
@@ -215,7 +215,7 @@ class Vector {
      auto stream = dev_ctx->stream();
      void *src = gpu_->ptr();
      void *dst = cpu_.data();
-      memory::Copy(platform::CPUPlace(), dst, CUDAPlace().get(), src,
+      paddle::memory::Copy(platform::CPUPlace(), dst, CUDAPlace().get(), src,
                           gpu_->size(), stream);
      dev_ctx->Wait();
    }
@@ -261,7 +261,7 @@ class Vector {
      auto *dev_ctx = static_cast<platform::CUDADeviceContext *>(
          platform::DeviceContextPool::Instance().Get(place));
      auto stream = dev_ctx->stream();
-      memory::Copy(CUDAPlace().get(), dst, platform::CPUPlace(), src,
+      paddle::memory::Copy(CUDAPlace().get(), dst, platform::CPUPlace(), src,
                           gpu_->size(), stream);
    }
@@ -284,7 +284,7 @@ class Vector {
    bool IsInCPU() const { return flag_ & kDataInCPU; }
    mutable std::vector<T> cpu_;
-    mutable memory::AllocationPtr gpu_;
+    mutable paddle::memory::AllocationPtr gpu_;
    mutable int flag_;
    mutable std::mutex mtx_;

--- a/paddle/fluid/framework/op_registry.h
+++ b/paddle/fluid/framework/op_registry.h
@@ -23,6 +23,7 @@ limitations under the License. */
 #include <unordered_map>
 #include <unordered_set>
+#define GLOG_NO_ABBREVIATED_SEVERITIES  // msvc conflict logging with windows.h
 #include "glog/logging.h"               // For VLOG()
 #include "paddle/fluid/framework/attribute.h"
 #include "paddle/fluid/framework/details/op_registry.h"

--- a/paddle/fluid/inference/api/demo_ci/CMakeLists.txt
+++ b/paddle/fluid/inference/api/demo_ci/CMakeLists.txt
@@ -89,12 +89,21 @@ endif()
 if(WITH_MKL)
  include_directories("${PADDLE_LIB}/third_party/install/mklml/include")
+  if(NOT WIN32)
    set(MATH_LIB ${PADDLE_LIB}/third_party/install/mklml/lib/libmklml_intel${CMAKE_SHARED_LIBRARY_SUFFIX}
                 ${PADDLE_LIB}/third_party/install/mklml/lib/libiomp5${CMAKE_SHARED_LIBRARY_SUFFIX})
+  else(WIN32)
+    set(MATH_LIB ${PADDLE_LIB}/third_party/install/mklml/lib/libmklml${CMAKE_SHARED_LIBRARY_SUFFIX}
+            ${PADDLE_LIB}/third_party/install/mklml/lib/libiomp5md${CMAKE_SHARED_LIBRARY_SUFFIX})
+  endif(WIN32)
  set(MKLDNN_PATH "${PADDLE_LIB}/third_party/install/mkldnn")
  if(EXISTS ${MKLDNN_PATH})
    include_directories("${MKLDNN_PATH}/include")
+    if(WIN32)
+      set(MKLDNN_LIB ${MKLDNN_PATH}/lib/mkldnn.lib)
+    else(WIN32)
      set(MKLDNN_LIB ${MKLDNN_PATH}/lib/libmkldnn.so.0)
+    endif(WIN32)
  endif()
 else()
  set(MATH_LIB ${PADDLE_LIB}/third_party/install/openblas/lib/libopenblas${CMAKE_STATIC_LIBRARY_SUFFIX})

--- a/paddle/fluid/memory/detail/system_allocator.cc
+++ b/paddle/fluid/memory/detail/system_allocator.cc
@@ -17,7 +17,6 @@ limitations under the License. */
 #ifdef _WIN32
 #include <malloc.h>
-#include <windows.h>  // VirtualLock/VirtualUnlock
 #else
 #include <sys/mman.h>  // for mlock and munlock
 #endif

--- a/paddle/fluid/operators/CMakeLists.txt
+++ b/paddle/fluid/operators/CMakeLists.txt
@@ -44,9 +44,8 @@ endif()
 register_operators(EXCLUDES warpctc_op conv_fusion_op DEPS ${OP_HEADER_DEPS} ${OP_PREFETCH_DEPS})
 # warpctc_op needs cudnn 7 above
-if (WITH_GPU AND NOT WIN32)
+if (WITH_GPU)
    if (${CUDNN_MAJOR_VERSION} VERSION_LESS 7)
        op_library(warpctc_op DEPS dynload_warpctc sequence_padding sequence_scale SRCS warpctc_op.cc warpctc_op.cu.cc)
    else()
@@ -64,9 +63,7 @@ endif()
 set(COMMON_OP_DEPS ${OP_HEADER_DEPS})
 set(COMMON_OP_DEPS ${COMMON_OP_DEPS} selected_rows_functor selected_rows lod_tensor maxouting unpooling pooling lod_rank_table context_project sequence_pooling executor)
-if (NOT WIN32)
+set(COMMON_OP_DEPS ${COMMON_OP_DEPS} dynload_warpctc)
-    set(COMMON_OP_DEPS ${COMMON_OP_DEPS} dynload_warpctc)
-endif()
 set(COMMON_OP_DEPS ${COMMON_OP_DEPS} sequence_padding sequence_scale cos_sim_functor memory jit_kernel concat_and_split cross_entropy softmax vol2col im2col sampler)
 set(COMMON_OP_DEPS ${COMMON_OP_DEPS} sequence2batch lstm_compute matrix_bit_code gru_compute activation_functions)
 if (WITH_GPU)

--- a/paddle/fluid/operators/cum_op.h
+++ b/paddle/fluid/operators/cum_op.h
@@ -13,6 +13,8 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 #pragma once
+#include <array>
 #include "paddle/fluid/framework/eigen.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/framework/operator.h"

--- a/paddle/fluid/operators/elementwise/elementwise_mul_mkldnn_op.cc
+++ b/paddle/fluid/operators/elementwise/elementwise_mul_mkldnn_op.cc
@@ -19,6 +19,9 @@ limitations under the License. */
 #include "paddle/fluid/platform/mkldnn_helper.h"
 #include "paddle/fluid/operators/math/jit_kernel.h"
+#if defined(_WIN32) && defined(_WINSOCKAPI_)
+#define _WINSOCK2API_ /* Prevent inclusion of winsock2.h */
+#endif
 #include "xbyak/xbyak.h"
 #include "xbyak/xbyak_util.h"

--- a/paddle/fluid/operators/math/detail/lstm_cpu_kernel.h
+++ b/paddle/fluid/operators/math/detail/lstm_cpu_kernel.h
@@ -17,6 +17,12 @@ limitations under the License. */
 #include "paddle/fluid/operators/math/detail/activation_functions.h"
 #include "paddle/fluid/operators/math/lstm_compute.h"
+#if defined(_WIN32)
+#if defined(__AVX2__) || defined(__AVX__)
+inline __m256 operator+=(__m256 a, __m256 b) { return _mm256_add_ps(a, b); }
+#endif
+#endif
 namespace paddle {
 namespace operators {
 namespace math {

--- a/paddle/fluid/operators/math/jit_gen.h
+++ b/paddle/fluid/operators/math/jit_gen.h
@@ -18,6 +18,9 @@ limitations under the License. */
 #include <type_traits>
 #include "paddle/fluid/platform/macros.h"
+#if defined(_WIN32) && defined(_WINSOCKAPI_)
+#define _WINSOCK2API_ /* Prevent inclusion of winsock2.h */
+#endif
 #define XBYAK_USE_MMAP_ALLOCATOR
 #include "xbyak/xbyak.h"
 #include "xbyak/xbyak_util.h"

--- a/paddle/fluid/platform/cpu_info.cc
+++ b/paddle/fluid/platform/cpu_info.cc
@@ -14,6 +14,10 @@ limitations under the License. */
 #include "paddle/fluid/platform/cpu_info.h"
+#if defined(_WIN32)
+#define NOMINMAX  // msvc max/min macro conflict with std::min/max
+#endif
 #ifdef PADDLE_WITH_XBYAK
 #include "xbyak/xbyak.h"
 #include "xbyak/xbyak_util.h"
@@ -22,9 +26,8 @@ limitations under the License. */
 #ifdef __APPLE__
 #include <sys/sysctl.h>
 #include <sys/types.h>
 #elif defined(_WIN32)
-#define NOMINMAX  // msvc max/min macro conflict with std::min/max
+#define WIN32_LEAN_AND_MEAN
 #include <windows.h>
 #else
 #include <unistd.h>

--- a/paddle/fluid/platform/dynload/CMakeLists.txt
+++ b/paddle/fluid/platform/dynload/CMakeLists.txt
@@ -16,9 +16,7 @@ if (CUPTI_FOUND)
    list(APPEND CUDA_SRCS cupti.cc)
 endif(CUPTI_FOUND)
 nv_library(dynload_cuda SRCS ${CUDA_SRCS} DEPS dynamic_loader)
-if (NOT WIN32)
 cc_library(dynload_warpctc SRCS warpctc.cc DEPS dynamic_loader warpctc)
-endif(NOT WIN32)
 if (WITH_MKLML)
    cc_library(dynload_mklml SRCS mklml.cc DEPS dynamic_loader mklml)
 endif()

--- a/paddle/fluid/platform/dynload/cudnn.cc
+++ b/paddle/fluid/platform/dynload/cudnn.cc
@@ -38,6 +38,10 @@ CUDNN_DNN_ROUTINE_EACH_AFTER_R4(DEFINE_WRAP);
 CUDNN_DNN_ROUTINE_EACH_R5(DEFINE_WRAP);
 #endif
+#ifdef CUDNN_DNN_ROUTINE_EACH_R6
+CUDNN_DNN_ROUTINE_EACH_R6(DEFINE_WRAP);
+#endif
 #ifdef CUDNN_DNN_ROUTINE_EACH_R7
 CUDNN_DNN_ROUTINE_EACH_R7(DEFINE_WRAP);
 #endif

--- a/paddle/fluid/platform/dynload/cudnn.h
+++ b/paddle/fluid/platform/dynload/cudnn.h
@@ -34,7 +34,7 @@ extern void EnforceCUDNNLoaded(const char* fn_name);
 #define DECLARE_DYNAMIC_LOAD_CUDNN_WRAP(__name)                            \
  struct DynLoad__##__name {                                               \
    template <typename... Args>                                            \
-    auto operator()(Args... args) -> decltype(__name(args...)) {           \
+    auto operator()(Args... args) -> DECLARE_TYPE(__name, args...) {       \
      using cudnn_func = decltype(&::__name);                              \
      std::call_once(cudnn_dso_flag, []() {                                \
        cudnn_dso_handle = paddle::platform::dynload::GetCUDNNDsoHandle(); \

--- a/paddle/fluid/platform/dynload/dynamic_loader.cc
+++ b/paddle/fluid/platform/dynload/dynamic_loader.cc
@@ -53,6 +53,12 @@ namespace platform {
 namespace dynload {
 static constexpr char cupti_lib_path[] = CUPTI_LIB_PATH;
+#if defined(_WIN32) && defined(PADDLE_WITH_CUDA)
+static constexpr char* win_cublas_lib = "cublas64_" PADDLE_CUDA_BINVER ".dll";
+static constexpr char* win_curand_lib = "curand64_" PADDLE_CUDA_BINVER ".dll";
+static constexpr char* win_cudnn_lib = "cudnn64_" PADDLE_CUDNN_BINVER ".dll";
+#endif
 static inline std::string join(const std::string& part1,
                               const std::string& part2) {
  // directory separator
@@ -165,6 +171,8 @@ static inline void* GetDsoHandleFromSearchPath(const std::string& search_root,
 void* GetCublasDsoHandle() {
 #if defined(__APPLE__) || defined(__OSX__)
  return GetDsoHandleFromSearchPath(FLAGS_cuda_dir, "libcublas.dylib");
+#elif defined(_WIN32) && defined(PADDLE_WITH_CUDA)
+  return GetDsoHandleFromSearchPath(FLAGS_cuda_dir, win_cublas_lib);
 #else
  return GetDsoHandleFromSearchPath(FLAGS_cuda_dir, "libcublas.so");
 #endif
@@ -173,6 +181,8 @@ void* GetCublasDsoHandle() {
 void* GetCUDNNDsoHandle() {
 #if defined(__APPLE__) || defined(__OSX__)
  return GetDsoHandleFromSearchPath(FLAGS_cudnn_dir, "libcudnn.dylib", false);
+#elif defined(_WIN32) && defined(PADDLE_WITH_CUDA)
+  return GetDsoHandleFromSearchPath(FLAGS_cudnn_dir, win_cudnn_lib);
 #else
  return GetDsoHandleFromSearchPath(FLAGS_cudnn_dir, "libcudnn.so", false);
 #endif
@@ -193,6 +203,8 @@ void* GetCUPTIDsoHandle() {
 void* GetCurandDsoHandle() {
 #if defined(__APPLE__) || defined(__OSX__)
  return GetDsoHandleFromSearchPath(FLAGS_cuda_dir, "libcurand.dylib");
+#elif defined(_WIN32) && defined(PADDLE_WITH_CUDA)
+  return GetDsoHandleFromSearchPath(FLAGS_cuda_dir, win_curand_lib);
 #else
  return GetDsoHandleFromSearchPath(FLAGS_cuda_dir, "libcurand.so");
 #endif
@@ -201,6 +213,8 @@ void* GetCurandDsoHandle() {
 void* GetWarpCTCDsoHandle() {
 #if defined(__APPLE__) || defined(__OSX__)
  return GetDsoHandleFromSearchPath(FLAGS_warpctc_dir, "libwarpctc.dylib");
+#elif defined(_WIN32)
+  return GetDsoHandleFromSearchPath(FLAGS_warpctc_dir, "warpctc.dll");
 #else
  return GetDsoHandleFromSearchPath(FLAGS_warpctc_dir, "libwarpctc.so");
 #endif
@@ -225,6 +239,8 @@ void* GetTensorRtDsoHandle() {
 void* GetMKLMLDsoHandle() {
 #if defined(__APPLE__) || defined(__OSX__)
  return GetDsoHandleFromSearchPath(FLAGS_mklml_dir, "libmklml_intel.dylib");
+#elif defined(_WIN32)
+  return GetDsoHandleFromSearchPath(FLAGS_mklml_dir, "mklml.dll");
 #else
  return GetDsoHandleFromSearchPath(FLAGS_mklml_dir, "libmklml_intel.so");
 #endif

--- a/paddle/fluid/platform/dynload/dynamic_loader.h
+++ b/paddle/fluid/platform/dynload/dynamic_loader.h
@@ -18,6 +18,12 @@ namespace paddle {
 namespace platform {
 namespace dynload {
+#ifndef _WIN32
+#define DECLARE_TYPE(__name, ...) decltype(__name(__VA_ARGS__))
+#else
+#define DECLARE_TYPE(__name, ...) decltype(auto)
+#endif
 void* GetCublasDsoHandle();
 void* GetCUDNNDsoHandle();
 void* GetCUPTIDsoHandle();

--- a/paddle/fluid/platform/dynload/mklml.h
+++ b/paddle/fluid/platform/dynload/mklml.h
@@ -34,7 +34,7 @@ extern void* mklml_dso_handle;
 #define DYNAMIC_LOAD_MKLML_WRAP(__name)                                    \
  struct DynLoad__##__name {                                               \
    template <typename... Args>                                            \
-    auto operator()(Args... args) -> decltype(__name(args...)) {           \
+    auto operator()(Args... args) -> DECLARE_TYPE(__name, args...) {       \
      using mklmlFunc = decltype(&::__name);                               \
      std::call_once(mklml_dso_flag, []() {                                \
        mklml_dso_handle = paddle::platform::dynload::GetMKLMLDsoHandle(); \

--- a/paddle/fluid/platform/dynload/tensorrt.h
+++ b/paddle/fluid/platform/dynload/tensorrt.h
@@ -33,7 +33,7 @@ extern void* tensorrt_dso_handle;
 #define DECLARE_DYNAMIC_LOAD_TENSORRT_WRAP(__name)                      \
  struct DynLoad__##__name {                                            \
    template <typename... Args>                                         \
-    auto operator()(Args... args) -> decltype(__name(args...)) {        \
+    auto operator()(Args... args) -> DECLARE_TYPE(__name, args...) {    \
      using tensorrt_func = decltype(__name(args...)) (*)(Args...);     \
      std::call_once(tensorrt_dso_flag, []() {                          \
        tensorrt_dso_handle =                                           \

--- a/paddle/fluid/platform/dynload/warpctc.h
+++ b/paddle/fluid/platform/dynload/warpctc.h
@@ -34,7 +34,7 @@ extern void* warpctc_dso_handle;
 #define DYNAMIC_LOAD_WARPCTC_WRAP(__name)                                      \
  struct DynLoad__##__name {                                                   \
    template <typename... Args>                                                \
-    auto operator()(Args... args) -> decltype(__name(args...)) {               \
+    auto operator()(Args... args) -> DECLARE_TYPE(__name, args...) {           \
      using warpctcFunc = decltype(&::__name);                                 \
      std::call_once(warpctc_dso_flag, []() {                                  \
        warpctc_dso_handle = paddle::platform::dynload::GetWarpCTCDsoHandle(); \

--- a/paddle/fluid/platform/port.h
+++ b/paddle/fluid/platform/port.h
@@ -37,6 +37,10 @@
 #define GOOGLE_GLOG_DLL_DECL
 #include <io.h>  // _popen, _pclose
 #include <stdio.h>
+#ifdef _WINSOCKAPI_
+/* Prevent inclusion of winsock.h in windows.h */
+#define WIN32_LEAN_AND_MEAN
+#endif
 #include <windows.h>
 #include <numeric>  // std::accumulate in msvc
 #ifndef S_ISDIR     // windows port for sys/stat.h
@@ -55,7 +59,6 @@ static void *dlsym(void *handle, const char *symbol_name) {
 static void *dlopen(const char *filename, int flag) {
  std::string file_name(filename);
-  file_name.replace(0, file_name.size() - 1, '/', '\\');
  HMODULE hModule = LoadLibrary(file_name.c_str());
  if (!hModule) {
    throw std::runtime_error(file_name + " not found.");

--- a/paddle/fluid/train/demo/CMakeLists.txt
+++ b/paddle/fluid/train/demo/CMakeLists.txt
@@ -35,16 +35,26 @@ add_executable(demo_trainer demo_trainer.cc)
 if(WITH_MKLDNN)
  include_directories("${PADDLE_LIB}/third_party/install/mkldnn/include")
+  if(WIN32)
+    set(MKLDNN_LIB ${PADDLE_LIB}/third_party/install/mkldnn/lib/mkldnn.lib)
+  else(WIN32)
    set(MKLDNN_LIB ${PADDLE_LIB}/third_party/install/mkldnn/lib/libmkldnn.so.0)
-endif()
+  endif(WIN32)
+endif(WITH_MKLDNN)
 if(WITH_MKL)
  include_directories("${PADDLE_LIB}/third_party/install/mklml/include")
+  if(WIN32)
+    set(MATH_LIB ${PADDLE_LIB}/third_party/install/mklml/lib/mklml.lib)
+  else(WIN32)
    set(MATH_LIB ${PADDLE_LIB}/third_party/install/mklml/lib/libmklml_intel.so)
+  endif(WIN32)
 else()
  if(APPLE)
    set(MATH_LIB cblas)
-  else(APPLE)
+  elseif(WIN32)
+    set(MATH_LIB ${PADDLE_LIB}/third_party/install/openblas/lib/libopenblas.lib)
+  else()
    set(MATH_LIB ${PADDLE_LIB}/third_party/install/openblas/lib/libopenblas.a)
  endif(APPLE)
 endif()

--- a/python/CMakeLists.txt
+++ b/python/CMakeLists.txt
@@ -48,12 +48,18 @@ configure_file(${CMAKE_CURRENT_SOURCE_DIR}/setup.py.in
 IF(WIN32)
    # Python would use the .pyd by default under Windows series platform
    set(FLUID_DST_DIR ${PADDLE_BINARY_DIR}/python/paddle/fluid/)
-    get_filename_component(openblas_refpath ${CBLAS_LIBRARIES} DIRECTORY)
    set(FLUID_CORE ${FLUID_DST_DIR}/core.pyd)
+    if(NOT WITH_MKLDNN)
+        get_filename_component(openblas_refpath ${CBLAS_LIBRARIES} DIRECTORY)
        add_custom_command(OUTPUT ${FLUID_CORE}
                COMMAND cmake -E copy $<TARGET_FILE:paddle_pybind> ${FLUID_CORE}
                COMMAND cmake -E copy ${openblas_refpath}/openblas.dll ${FLUID_DST_DIR}
                DEPENDS paddle_pybind)
+    else(NOT WITH_MKLDNN)
+        add_custom_command(OUTPUT ${FLUID_CORE}
+                COMMAND cmake -E copy $<TARGET_FILE:paddle_pybind> ${FLUID_CORE}
+                DEPENDS paddle_pybind)
+    endif(NOT WITH_MKLDNN)
 ELSE()
    set(FLUID_CORE ${PADDLE_BINARY_DIR}/python/paddle/fluid/core.so)
    add_custom_command(OUTPUT ${FLUID_CORE}

--- a/python/paddle/fluid/__init__.py
+++ b/python/paddle/fluid/__init__.py
@@ -102,6 +102,12 @@ def __bootstrap__():
    import sys
    import os
    import platform
+    if os.name == 'nt':
+        third_lib_path = os.path.abspath(os.path.dirname(__file__)) + os.sep + '..' + os.sep + 'libs'
+        os.environ['path'] += ';' + third_lib_path
+        sys.path.append(third_lib_path)
    from . import core
    in_test = 'unittest' in sys.modules
@@ -128,13 +134,12 @@ def __bootstrap__():
        'free_idle_memory', 'paddle_num_threads', "dist_threadpool_size",
        'eager_delete_tensor_gb', 'fast_eager_deletion_mode',
        'allocator_strategy', 'reader_queue_speed_test_mode',
-        'print_sub_graph_dir', 'pe_profile_fname'
+        'print_sub_graph_dir', 'pe_profile_fname', 'warpctc_dir'
    ]
    if 'Darwin' not in sysstr:
        read_env_flags.append('use_pinned_memory')
    if os.name != 'nt':
-        read_env_flags.append('warpctc_dir')
        read_env_flags.append('cpu_deterministic')
    if core.is_compiled_with_dist():

--- a/python/paddle/fluid/framework.py
+++ b/python/paddle/fluid/framework.py
@@ -16,6 +16,7 @@ from __future__ import print_function
 import collections
 import contextlib
+import os
 import re
 import six
 import sys
@@ -27,6 +28,13 @@ from .proto import framework_pb2
 try:
    from . import core
 except ImportError as e:
+    if os.name == 'nt':
+        raise ImportError(
+            """NOTE: You may need to run \"set PATH=c:\python27\lib:%PATH%\"
+        if you encounters \"mkldnn.dll not found\" errors. If you have python
+        installed in other directory, replace \"c:\python27\lib" with your own
+        directory. The original error is: \n""" + cpt.get_exception_message(e))
+    else:
        raise ImportError(
            """NOTE: You may need to run \"export LD_LIBRARY_PATH=/usr/local/lib:$LD_LIBRARY_PATH\"
        if you encounters \"libmkldnn.so not found\" errors. If you have python

--- a/python/setup.py.in
+++ b/python/setup.py.in
@@ -158,16 +158,18 @@ if '${WITH_FLUID_ONLY}'== 'OFF':
 # put all thirdparty libraries in paddle.libs
 libs_path='${PADDLE_BINARY_DIR}/python/paddle/libs'
-if os.name != 'nt':
-    package_data['paddle.libs']= []
+package_data['paddle.libs']= []
-    package_data['paddle.libs']=['libwarpctc' + ext_name]
+package_data['paddle.libs']=[('libwarpctc' if os.name != 'nt' else 'warpctc') + ext_name]
-    shutil.copy('${WARPCTC_LIBRARIES}', libs_path)
+shutil.copy('${WARPCTC_LIBRARIES}', libs_path)
 if '${WITH_MKL}' == 'ON':
-    shutil.copy('${MKLML_LIB}', libs_path)
+    shutil.copy('${MKLML_SHARED_LIB}', libs_path)
-    shutil.copy('${MKLML_IOMP_LIB}', libs_path)
+    shutil.copy('${MKLML_SHARED_IOMP_LIB}', libs_path)
-    package_data['paddle.libs']+=['libmklml_intel' + ext_name,'libiomp5' + ext_name]
+    package_data['paddle.libs']+=[('libmklml_intel' if os.name != 'nt' else 'mklml') + ext_name, ('libiomp5' if os.name != 'nt' else 'libiomp5md') + ext_name]
 if '${WITH_MKLDNN}' == 'ON':
    if '${CMAKE_BUILD_TYPE}' == 'Release':
+        if os.name != 'nt':
            # only change rpath in Release mode.
            # TODO(typhoonzero): use install_name_tool to patch mkl libs once
            # we can support mkl on mac.
@@ -178,7 +180,7 @@ if '${WITH_MKLDNN}' == 'ON':
            command = "patchelf --set-rpath '$ORIGIN/' ${MKLDNN_SHARED_LIB}"
            if os.system(command) != 0:
                raise Exception("patch libmkldnn.so failed, command: %s" % command)
-    package_data['paddle.libs']+=['libmkldnn.so.0']
+    package_data['paddle.libs']+=['libmkldnn.so.0' if os.name != 'nt' else ('mkldnn' + ext_name)]
    shutil.copy('${MKLDNN_SHARED_LIB}', libs_path)
 if '${WITH_NGRAPH}' == 'ON':
    # only change rpath in Release mode,