diff --git a/CMakeLists.txt b/CMakeLists.txt
index 317f7f9eb4..b1d0abdf2c 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -24,6 +24,9 @@ message(STATUS "CXX compiler: ${CMAKE_CXX_COMPILER}, version: "
         "${CMAKE_CXX_COMPILER_ID} ${CMAKE_CXX_COMPILER_VERSION}")
 message(STATUS "C compiler: ${CMAKE_C_COMPILER}, version: "
         "${CMAKE_C_COMPILER_ID} ${CMAKE_C_COMPILER_VERSION}")
+if(WIN32)
+    set(CMAKE_STATIC_LIBRARY_PREFIX lib)
+endif(WIN32)
 
 if(NOT CMAKE_CROSSCOMPILING)
     find_package(CUDA QUIET)
@@ -165,7 +168,6 @@ include(external/python)    # download, build, install python
 include(external/openblas)  # download, build, install openblas
 include(external/mkldnn)    # download, build, install mkldnn
 include(external/swig)      # download, build, install swig
-include(external/warpctc)   # download, build, install warpctc
 include(external/boost)     # download boost
 include(external/any)       # download libn::any
 include(external/eigen)     # download eigen3
@@ -173,6 +175,14 @@ include(external/pybind11)  # download pybind11
 include(external/cares)
 include(external/cub)
 
+if (NOT WIN32)
+# there is no official support of snappystream, warpctc, nccl, cupti in windows
+include(external/snappy)    # download snappy
+include(external/snappystream) # download snappystream
+include(external/warpctc)   # download, build, install warpctc
+include(cupti)
+endif (NOT WIN32)
+
 if(WITH_DISTRIBUTE)
     if(WITH_GRPC)
         include(external/grpc)
@@ -194,13 +204,10 @@ if(WITH_BRPC_RDMA)
     endif()
 endif()
 
-include(external/snappy)    # download snappy
-include(external/snappystream)
-include(external/threadpool)
 
+include(external/threadpool)
 include(flags)              # set paddle compile flags
 include(cudnn)              # set cudnn libraries, must before configure
-include(cupti)
 include(configure)          # add paddle env configuration
 
 if(WITH_GPU)
diff --git a/cmake/configure.cmake b/cmake/configure.cmake
index e03e15bfc0..ce1857582b 100644
--- a/cmake/configure.cmake
+++ b/cmake/configure.cmake
@@ -61,6 +61,11 @@ if(NOT CMAKE_CROSSCOMPILING)
     endif()
 endif()
 
+if(WIN32)
+  # windows stupid compile option for all targets.
+  add_definitions(-D_XKEYCHECK_H)
+endif(WIN32)
+
 if(NOT WITH_GOLANG)
     add_definitions(-DPADDLE_WITHOUT_GOLANG)
 endif(NOT WITH_GOLANG)
diff --git a/cmake/external/boost.cmake b/cmake/external/boost.cmake
index 73713d93d5..ada61de8eb 100644
--- a/cmake/external/boost.cmake
+++ b/cmake/external/boost.cmake
@@ -28,7 +28,12 @@ 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()
-MESSAGE(STATUS "BOOST_TAR: ${BOOST_TAR}, BOOST_URL: ${BOOST_URL}")
+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)
+
 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)
@@ -36,12 +41,13 @@ set_directory_properties(PROPERTIES CLEAN_NO_CUSTOM 1)
 
 include_directories(${BOOST_INCLUDE_DIR})
 
+if (NOT WIN32)
 ExternalProject_Add(
     ${BOOST_PROJECT}
     ${EXTERNAL_PROJECT_LOG_ARGS}
     DOWNLOAD_DIR          ${BOOST_DOWNLOAD_DIR}
     DOWNLOAD_COMMAND      wget --no-check-certificate ${BOOST_URL} -c -q -O ${BOOST_TAR}.tar.gz
-                          && tar zxf ${BOOST_TAR}.tar.gz
+    && tar zxf ${BOOST_TAR}.tar.gz
     DOWNLOAD_NO_PROGRESS  1
     PREFIX                ${BOOST_SOURCES_DIR}
     CONFIGURE_COMMAND     ""
@@ -49,8 +55,9 @@ ExternalProject_Add(
     INSTALL_COMMAND       ""
     UPDATE_COMMAND        ""
 )
+endif(NOT WIN32)
 
-if (${CMAKE_VERSION} VERSION_LESS "3.3.0")
+if (${CMAKE_VERSION} VERSION_LESS "3.3.0" OR NOT WIN32)
     set(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/boost_dummy.c)
     file(WRITE ${dummyfile} "const char *dummy = \"${dummyfile}\";")
     add_library(boost STATIC ${dummyfile})
diff --git a/cmake/external/gflags.cmake b/cmake/external/gflags.cmake
index a1d2d0f446..cf58cc3976 100644
--- a/cmake/external/gflags.cmake
+++ b/cmake/external/gflags.cmake
@@ -18,7 +18,7 @@ SET(GFLAGS_SOURCES_DIR ${THIRD_PARTY_PATH}/gflags)
 SET(GFLAGS_INSTALL_DIR ${THIRD_PARTY_PATH}/install/gflags)
 SET(GFLAGS_INCLUDE_DIR "${GFLAGS_INSTALL_DIR}/include" CACHE PATH "gflags include directory." FORCE)
 IF(WIN32)
-  set(GFLAGS_LIBRARIES "${GFLAGS_INSTALL_DIR}/lib/gflags.lib" CACHE FILEPATH "GFLAGS_LIBRARIES" FORCE)
+  set(GFLAGS_LIBRARIES "${GFLAGS_INSTALL_DIR}/lib/libgflags.lib" CACHE FILEPATH "GFLAGS_LIBRARIES" FORCE)
 ELSE(WIN32)
   set(GFLAGS_LIBRARIES "${GFLAGS_INSTALL_DIR}/lib/libgflags.a" CACHE FILEPATH "GFLAGS_LIBRARIES" FORCE)
 ENDIF(WIN32)
@@ -45,7 +45,13 @@ ExternalProject_Add(
                      -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON
                      -DCMAKE_BUILD_TYPE:STRING=${THIRD_PARTY_BUILD_TYPE}
 )
-
+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
+  )
+  ENDIF()
+ENDIF(WIN32)
 ADD_LIBRARY(gflags STATIC IMPORTED GLOBAL)
 SET_PROPERTY(TARGET gflags PROPERTY IMPORTED_LOCATION ${GFLAGS_LIBRARIES})
 ADD_DEPENDENCIES(gflags extern_gflags)
@@ -60,3 +66,4 @@ IF(WITH_C_API)
     INSTALL(FILES ${GFLAGS_LIBRARIES} DESTINATION third_party/gflags/lib)
   ENDIF()
 ENDIF()
+
diff --git a/cmake/external/glog.cmake b/cmake/external/glog.cmake
index ac0181e69c..25ef2970ac 100644
--- a/cmake/external/glog.cmake
+++ b/cmake/external/glog.cmake
@@ -60,6 +60,13 @@ ExternalProject_Add(
                      -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON
                      -DCMAKE_BUILD_TYPE:STRING=${THIRD_PARTY_BUILD_TYPE}
 )
+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
+  )
+  ENDIF()
+ENDIF(WIN32)
 
 ADD_LIBRARY(glog STATIC IMPORTED GLOBAL)
 SET_PROPERTY(TARGET glog PROPERTY IMPORTED_LOCATION ${GLOG_LIBRARIES})
diff --git a/cmake/external/openblas.cmake b/cmake/external/openblas.cmake
index 56024edf5b..c3fbe4dbdb 100644
--- a/cmake/external/openblas.cmake
+++ b/cmake/external/openblas.cmake
@@ -17,20 +17,29 @@ 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)
     SET(CBLAS_INSTALL_DIR ${THIRD_PARTY_PATH}/install/openblas)
-    SET(CBLAS_INC_DIR "${CBLAS_INSTALL_DIR}/include" CACHE PATH "openblas include directory." FORCE)
+    SET(CBLAS_INCLUDE_DIR "${CBLAS_INSTALL_DIR}/include" CACHE PATH "openblas include directory." FORCE)
 
     SET(CBLAS_LIBRARIES
         "${CBLAS_INSTALL_DIR}/lib/${CMAKE_STATIC_LIBRARY_PREFIX}openblas${CMAKE_STATIC_LIBRARY_SUFFIX}"
         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)
     SET(OPENBLAS_CC "${CMAKE_C_COMPILER} -Wno-unused-but-set-variable -Wno-unused-variable")
     SET(OPENBLAS_COMMIT "v0.2.20")
 
@@ -69,7 +78,6 @@ IF(NOT ${CBLAS_FOUND})
     ENDIF()
 
     SET(COMMON_ARGS CC=${OPENBLAS_CC} NO_SHARED=1 NO_LAPACK=1 libs)
-
     ExternalProject_Add(
         extern_openblas
         ${EXTERNAL_PROJECT_LOG_ARGS}
@@ -84,9 +92,11 @@ IF(NOT ${CBLAS_FOUND})
         UPDATE_COMMAND      ""
         CONFIGURE_COMMAND   ""
     )
+    ELSE()
+    ENDIF(NOT WIN32)
     SET(CBLAS_PROVIDER openblas)
     IF(WITH_C_API)
-        INSTALL(DIRECTORY ${CBLAS_INC_DIR} DESTINATION third_party/openblas)
+        INSTALL(DIRECTORY ${CBLAS_INCLUDE_DIR} DESTINATION third_party/openblas)
         # Because libopenblas.a is a symbolic link of another library, thus need to
         # install the whole directory.
         IF(ANDROID)
@@ -107,7 +117,8 @@ IF(NOT ${CBLAS_FOUND})
 ENDIF(NOT ${CBLAS_FOUND})
 
 MESSAGE(STATUS "BLAS library: ${CBLAS_LIBRARIES}")
-INCLUDE_DIRECTORIES(${CBLAS_INC_DIR})
+MESSAGE(STATUS "BLAS Include: ${CBLAS_INCLUDE_DIR}")
+INCLUDE_DIRECTORIES(${CBLAS_INCLUDE_DIR})
 
 # FIXME(gangliao): generate cblas target to track all high performance
 # linear algebra libraries for cc_library(xxx SRCS xxx.c DEPS cblas)
diff --git a/cmake/external/protobuf.cmake b/cmake/external/protobuf.cmake
index 2665996432..550b0dada8 100644
--- a/cmake/external/protobuf.cmake
+++ b/cmake/external/protobuf.cmake
@@ -14,11 +14,14 @@
 
 INCLUDE(ExternalProject)
 # Always invoke `FIND_PACKAGE(Protobuf)` for importing function protobuf_generate_cpp
+IF(NOT WIN32)
 FIND_PACKAGE(Protobuf QUIET)
+ENDIF(NOT WIN32)
 macro(UNSET_VAR VAR_NAME)
     UNSET(${VAR_NAME} CACHE)
     UNSET(${VAR_NAME})
 endmacro()
+
 UNSET_VAR(PROTOBUF_INCLUDE_DIR)
 UNSET_VAR(PROTOBUF_FOUND)
 UNSET_VAR(PROTOBUF_PROTOC_EXECUTABLE)
@@ -94,12 +97,14 @@ macro(PROMPT_PROTOBUF_LIB)
     SET(protobuf_DEPS ${ARGN})
 
     MESSAGE(STATUS "Protobuf protoc executable: ${PROTOBUF_PROTOC_EXECUTABLE}")
+    MESSAGE(STATUS "Protobuf-lite library: ${PROTOBUF_LITE_LIBRARY}")
     MESSAGE(STATUS "Protobuf library: ${PROTOBUF_LIBRARY}")
+    MESSAGE(STATUS "Protoc library: ${PROTOBUF_PROTOC_LIBRARY}")
     MESSAGE(STATUS "Protobuf version: ${PROTOBUF_VERSION}")
     INCLUDE_DIRECTORIES(${PROTOBUF_INCLUDE_DIR})
 
     # Assuming that all the protobuf libraries are of the same type.
-    IF(${PROTOBUF_LIBRARY} MATCHES "${CMAKE_STATIC_LIBRARY_SUFFIX}$")
+    IF(${PROTOBUF_LIBRARY} MATCHES ${CMAKE_STATIC_LIBRARY_SUFFIX})
         SET(protobuf_LIBTYPE STATIC)
     ELSEIF(${PROTOBUF_LIBRARY} MATCHES "${CMAKE_SHARED_LIBRARY_SUFFIX}$")
         SET(protobuf_LIBTYPE SHARED)
@@ -137,18 +142,25 @@ macro(SET_PROTOBUF_VERSION)
 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 "")
+
     find_path(PROTOBUF_INCLUDE_DIR google/protobuf/message.h PATHS ${PROTOBUF_ROOT}/include NO_DEFAULT_PATH)
-    find_library(PROTOBUF_LIBRARY protobuf PATHS ${PROTOBUF_ROOT}/lib NO_DEFAULT_PATH)
-    find_library(PROTOBUF_LITE_LIBRARY protobuf-lite PATHS ${PROTOBUF_ROOT}/lib NO_DEFAULT_PATH)
-    find_library(PROTOBUF_PROTOC_LIBRARY protoc PATHS ${PROTOBUF_ROOT}/lib NO_DEFAULT_PATH)
+    find_library(PROTOBUF_LIBRARY protobuf libprotobuf.lib PATHS ${PROTOBUF_ROOT}/lib NO_DEFAULT_PATH)
+    find_library(PROTOBUF_LITE_LIBRARY protobuf-lite libprotobuf-lite.lib PATHS ${PROTOBUF_ROOT}/lib NO_DEFAULT_PATH)
+    find_library(PROTOBUF_PROTOC_LIBRARY protoc libprotoc.lib PATHS ${PROTOBUF_ROOT}/lib NO_DEFAULT_PATH)
     find_program(PROTOBUF_PROTOC_EXECUTABLE protoc PATHS ${PROTOBUF_ROOT}/bin NO_DEFAULT_PATH)
     if (PROTOBUF_INCLUDE_DIR AND PROTOBUF_LIBRARY AND PROTOBUF_LITE_LIBRARY AND PROTOBUF_PROTOC_LIBRARY AND PROTOBUF_PROTOC_EXECUTABLE)
         message(STATUS "Using custom protobuf library in ${PROTOBUF_ROOT}.")
+        SET(PROTOBUF_FOUND true)
         SET_PROTOBUF_VERSION()
         PROMPT_PROTOBUF_LIB()
     else()
-        message(WARNING "Cannot find protobuf library in ${PROTOBUF_ROOT}.")
+        message(WARNING "Cannot find protobuf library in ${PROTOBUF_ROOT}")
     endif()
 endif()
 
@@ -239,6 +251,7 @@ IF(CMAKE_CROSSCOMPILING)
         CACHE FILEPATH "protobuf executable." FORCE)
 ENDIF()
 
+
 IF(NOT PROTOBUF_FOUND)
     build_protobuf(extern_protobuf FALSE)
 
diff --git a/cmake/generic.cmake b/cmake/generic.cmake
index 82c958073c..6d23094232 100644
--- a/cmake/generic.cmake
+++ b/cmake/generic.cmake
@@ -148,7 +148,8 @@ function(merge_static_libs TARGET_NAME)
       COMMAND rm "${CMAKE_CURRENT_BINARY_DIR}/lib${TARGET_NAME}.a"
       COMMAND /usr/bin/libtool -static -o "${CMAKE_CURRENT_BINARY_DIR}/lib${TARGET_NAME}.a" ${libfiles}
       )
-  else() # general UNIX: use "ar" to extract objects and re-add to a common lib
+  endif(APPLE)
+  if(LINUX) # general UNIX: use "ar" to extract objects and re-add to a common lib
     set(target_DIR ${CMAKE_CURRENT_BINARY_DIR}/${TARGET_NAME}.dir)
 
     foreach(lib ${libs})
@@ -187,7 +188,36 @@ function(merge_static_libs TARGET_NAME)
         COMMAND ${CMAKE_AR} crs ${target_LIBNAME} `find ${target_DIR} -name '*.o'`
         COMMAND ${CMAKE_RANLIB} ${target_LIBNAME}
         WORKING_DIRECTORY ${target_DIR})
-  endif()
+  endif(LINUX)
+  if(WIN32) # windows do not support gcc/nvcc combined compiling. Use msvc lib.exe to merge libs.
+    # Make the generated dummy source file depended on all static input
+    # libs. If input lib changes,the source file is touched
+    # which causes the desired effect (relink).
+    add_custom_command(OUTPUT ${target_SRCS}
+      COMMAND ${CMAKE_COMMAND} -E touch ${target_SRCS}
+      DEPENDS ${libs})
+
+    # Generate dummy staic lib
+    file(WRITE ${target_SRCS} "const char *dummy_${TARGET_NAME} = \"${target_SRCS}\";")
+    add_library(${TARGET_NAME} STATIC ${target_SRCS})
+    target_link_libraries(${TARGET_NAME} ${libs_deps})
+
+    foreach(lib ${libs})
+      # Get the file names of the libraries to be merged
+      #if(NOT $<TARGET_FILE:${lib}> MATCHES "lib.*\\.lib")
+      #  message("library" ${lib})
+      #  set(libfiles ${libfiles} lib$<TARGET_FILE:${lib}>)
+      #else()
+      set(libfiles ${libfiles} $<TARGET_FILE:${lib}>)
+      #endif()
+    endforeach()
+   
+    # windows cmd return error in clean env.
+    # COMMAND del "${CMAKE_CURRENT_BINARY_DIR}/${CMAKE_BUILD_TYPE}/${TARGET_NAME}.lib"
+    add_custom_command(TARGET ${TARGET_NAME} POST_BUILD
+      COMMAND lib /OUT:${CMAKE_CURRENT_BINARY_DIR}/lib${TARGET_NAME}.lib ${libfiles}
+      )
+  endif(WIN32)
 endfunction(merge_static_libs)
 
 function(cc_library TARGET_NAME)
@@ -195,6 +225,10 @@ function(cc_library TARGET_NAME)
   set(oneValueArgs "")
   set(multiValueArgs SRCS DEPS)
   cmake_parse_arguments(cc_library "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
+  if(WIN32)
+      # add libxxx.lib prefix in windows
+      set(${TARGET_NAME}_LIB_NAME "${CMAKE_STATIC_LIBRARY_PREFIX}${TARGET_NAME}${CMAKE_STATIC_LIBRARY_SUFFIX}" CACHE STRING "output library name for target ${TARGET_NAME}")
+  endif(WIN32)
   if(cc_library_SRCS)
     if(cc_library_SHARED OR cc_library_shared) # build *.so
       add_library(${TARGET_NAME} SHARED ${cc_library_SRCS})
diff --git a/cmake/inference_lib.cmake b/cmake/inference_lib.cmake
index 834ab5a9e5..bc36683a9f 100644
--- a/cmake/inference_lib.cmake
+++ b/cmake/inference_lib.cmake
@@ -101,6 +101,7 @@ 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
@@ -120,15 +121,23 @@ if(NOT MOBILE_INFERENCE AND NOT RPI)
     DSTS ${dst_dir} ${dst_dir}/lib
     DEPS zlib)
 endif()
+endif(NOT WIN32)
 
 # paddle fluid module
 set(src_dir "${PADDLE_SOURCE_DIR}/paddle/fluid")
 set(dst_dir "${FLUID_INSTALL_DIR}/paddle/fluid")
 set(module "framework")
+if (NOT WIN32)
 copy(framework_lib DEPS framework_py_proto 
   SRCS ${src_dir}/${module}/*.h ${src_dir}/${module}/details/*.h ${PADDLE_BINARY_DIR}/paddle/fluid/framework/framework.pb.h
   DSTS ${dst_dir}/${module} ${dst_dir}/${module}/details ${dst_dir}/${module}
 )
+else()
+copy(framework_lib
+  SRCS ${src_dir}/${module}/*.h ${src_dir}/${module}/details/*.h ${PADDLE_BINARY_DIR}/paddle/fluid/framework/framework.pb.h
+  DSTS ${dst_dir}/${module} ${dst_dir}/${module}/details ${dst_dir}/${module}
+)
+endif(NOT WIN32)
 
 set(module "memory")
 copy(memory_lib
diff --git a/paddle/fluid/CMakeLists.txt b/paddle/fluid/CMakeLists.txt
index 2577e59d9c..ee1f655e25 100644
--- a/paddle/fluid/CMakeLists.txt
+++ b/paddle/fluid/CMakeLists.txt
@@ -2,9 +2,13 @@ add_subdirectory(memory)
 add_subdirectory(platform)
 add_subdirectory(framework)
 add_subdirectory(operators)
-add_subdirectory(pybind)
 add_subdirectory(string)
+
+if (NOT WIN32)
+add_subdirectory(pybind)
 add_subdirectory(recordio)
+endif(NOT WIN32)
+
 if(WITH_INFERENCE)
   # NOTE: please add subdirectory inference at last.
   add_subdirectory(inference)
diff --git a/paddle/fluid/framework/CMakeLists.txt b/paddle/fluid/framework/CMakeLists.txt
index 2ec422cc17..2c62d4ed6b 100644
--- a/paddle/fluid/framework/CMakeLists.txt
+++ b/paddle/fluid/framework/CMakeLists.txt
@@ -1,5 +1,7 @@
-add_subdirectory(details)
 add_subdirectory(ir)
+if (NOT WIN32)
+add_subdirectory(details)
+endif (NOT WIN32)
 # ddim lib
 proto_library(framework_proto SRCS framework.proto)
 
@@ -28,8 +30,12 @@ if(WITH_GPU)
 else()
   cc_test(mixed_vector_test SRCS mixed_vector_test.cc DEPS place memory device_context tensor)
 endif()
-
+if (NOT WIN32)
 cc_library(lod_tensor SRCS lod_tensor.cc DEPS ddim place tensor framework_proto recordio)
+else()
+cc_library(lod_tensor SRCS lod_tensor.cc DEPS ddim place tensor framework_proto)
+endif (NOT WIN32)
+
 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)
 
@@ -69,14 +75,22 @@ cc_library(op_proto_maker SRCS op_proto_maker.cc DEPS framework_proto attribute
 cc_test(op_proto_maker_test SRCS op_proto_maker_test.cc DEPS op_proto_maker)
 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)
+
+if (NOT WIN32)
 cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope glog
     shape_inference data_transform lod_tensor profiler)
+else()
+cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope glog
+    shape_inference data_transform lod_tensor)
+endif(NOT WIN32)
+
 cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry device_context)
 cc_library(proto_desc SRCS var_desc.cc op_desc.cc block_desc.cc program_desc.cc DEPS shape_inference op_info operator glog)
 
 cc_library(op_registry SRCS op_registry.cc DEPS op_proto_maker op_info operator glog proto_desc)
 nv_test(op_registry_test SRCS op_registry_test.cc DEPS op_registry)
 
+if (NOT WIN32)
 py_proto_compile(framework_py_proto SRCS framework.proto)
 # Generate an empty __init__.py to make framework_py_proto as a valid python module.
 add_custom_target(framework_py_proto_init ALL COMMAND ${CMAKE_COMMAND} -E touch __init__.py)
@@ -86,6 +100,7 @@ add_custom_command(TARGET framework_py_proto POST_BUILD
     COMMAND cp *.py ${PADDLE_BINARY_DIR}/python/paddle/fluid/proto/
     COMMENT "Copy generated python proto into directory paddle/fluid/proto."
     WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
+endif(NOT WIN32)
 
 cc_library(lod_rank_table SRCS lod_rank_table.cc DEPS lod_tensor)
 
@@ -120,7 +135,9 @@ cc_test(cow_ptr_tests SRCS details/cow_ptr_test.cc)
 # cc_test(channel_test SRCS channel_test.cc)
 cc_test(tuple_test SRCS tuple_test.cc )
 
+if (NOT WIN32)
 cc_test(rw_lock_test SRCS rw_lock_test.cc)
+endif (NOT WIN32)
 
 # disable test temporarily.
 # TODO https://github.com/PaddlePaddle/Paddle/issues/11971
diff --git a/paddle/fluid/framework/data_type.h b/paddle/fluid/framework/data_type.h
index 491413db8c..f8c72ffc89 100644
--- a/paddle/fluid/framework/data_type.h
+++ b/paddle/fluid/framework/data_type.h
@@ -26,6 +26,7 @@ namespace framework {
 extern proto::VarType::Type ToDataType(std::type_index type);
 extern std::type_index ToTypeIndex(proto::VarType::Type type);
 
+#if !defined(_WIN32)
 template <typename Visitor>
 inline void VisitDataType(proto::VarType::Type type, Visitor visitor) {
   switch (type) {
@@ -57,6 +58,40 @@ inline void VisitDataType(proto::VarType::Type type, Visitor visitor) {
       PADDLE_THROW("Not supported %d", type);
   }
 }
+#else
+// the msvc compiler do not implement two-stage name lookup correctly.
+template <typename Visitor>
+inline void VisitDataType(proto::VarType::Type type, Visitor visitor) {
+  switch (type) {
+    case proto::VarType::FP16:
+      visitor.operator()<platform::float16>();
+      break;
+    case proto::VarType::FP32:
+      visitor.operator()<float>();
+      break;
+    case proto::VarType::FP64:
+      visitor.operator()<double>();
+      break;
+    case proto::VarType::INT32:
+      visitor.operator()<int>();
+      break;
+    case proto::VarType::INT64:
+      visitor.operator()<int64_t>();
+      break;
+    case proto::VarType::BOOL:
+      visitor.operator()<bool>();
+      break;
+    case proto::VarType::UINT8:
+      visitor.operator()<uint8_t>();
+      break;
+    case proto::VarType::INT16:
+      visitor.operator()<int16_t>();
+      break;
+    default:
+      PADDLE_THROW("Not supported %d", type);
+  }
+}
+#endif  // _WIN32
 
 extern std::string DataTypeToString(const proto::VarType::Type type);
 extern size_t SizeOfType(std::type_index type);
diff --git a/paddle/fluid/framework/lod_tensor.cc b/paddle/fluid/framework/lod_tensor.cc
index 919029c38f..adeb26e4e7 100644
--- a/paddle/fluid/framework/lod_tensor.cc
+++ b/paddle/fluid/framework/lod_tensor.cc
@@ -25,8 +25,10 @@ limitations under the License. */
 #include "paddle/fluid/memory/memcpy.h"
 #include "paddle/fluid/memory/memory.h"
 
+#if !defined(_WIN32)
 #include "paddle/fluid/recordio/scanner.h"
 #include "paddle/fluid/recordio/writer.h"
+#endif  // _WIN32
 
 namespace paddle {
 namespace framework {
@@ -300,6 +302,7 @@ void DeserializeFromStream(std::istream &is, LoDTensor *tensor,
   TensorFromStream(is, static_cast<Tensor *>(tensor), dev_ctx);
 }
 
+#if !defined(_WIN32)
 void WriteToRecordIO(recordio::Writer *writer,
                      const std::vector<LoDTensor> &tensor,
                      const platform::DeviceContext &dev_ctx) {
@@ -329,7 +332,19 @@ bool ReadFromRecordIO(recordio::Scanner *scanner,
 
   return true;
 }
-
+#else
+class Writer {};
+class Scanner {};
+void WriteToRecordIO(recordio::Writer *writer,
+                     const std::vector<LoDTensor> &tensor,
+                     const platform::DeviceContext &dev_ctx) {}
+bool ReadFromRecordIO(recordio::Scanner *scanner,
+                      const platform::DeviceContext &dev_ctx,
+                      std::vector<LoDTensor> *result_ptr) {
+  PADDLE_ENFORCE("windows didn't supported recordio!.");
+  return true;
+}
+#endif  // _WIN32
 std::vector<LoDTensor> LoDTensor::SplitLoDTensor(
     const std::vector<platform::Place> places) const {
   check_memory_size();
diff --git a/paddle/fluid/framework/lod_tensor_test.cc b/paddle/fluid/framework/lod_tensor_test.cc
index cd50aaa260..cbf5fd04d7 100644
--- a/paddle/fluid/framework/lod_tensor_test.cc
+++ b/paddle/fluid/framework/lod_tensor_test.cc
@@ -274,6 +274,7 @@ TEST(LoD, ConvertToOffsetBasedLoD) {
   EXPECT_EQ(offset_lod, expected);
 }
 
+#if !defined(_WIN32)
 template <typename T>
 static void TestRecordIO() {
   LoDTensor tensor;
@@ -320,6 +321,7 @@ TEST(LoDTensor, RecordIO) {
   TestRecordIO<float>();
   TestRecordIO<double>();
 }
+#endif  // !defined(_WIN32)
 
 }  // namespace framework
 }  // namespace paddle
diff --git a/paddle/fluid/framework/operator.cc b/paddle/fluid/framework/operator.cc
index d04f774496..d58d6e4f3e 100644
--- a/paddle/fluid/framework/operator.cc
+++ b/paddle/fluid/framework/operator.cc
@@ -74,6 +74,12 @@ static DDim GetDims(const Scope& scope, const std::string& name,
   }
 }
 
+static bool VarInited(const Scope& scope, const std::string& name) {
+  Variable* var = scope.FindVar(name);
+  if (var == nullptr) return false;
+  return var->IsInitialized();
+}
+
 static std::string GetDtype(const Scope& scope, const std::string& name) {
   Variable* var = scope.FindVar(name);
   if (var == nullptr) {
@@ -87,8 +93,12 @@ static std::string GetDtype(const Scope& scope, const std::string& name) {
     }
     return DataTypeToString(ToDataType(tensor.type()));
   } else if (var->IsType<SelectedRows>()) {
-    return DataTypeToString(
-        ToDataType(var->Get<SelectedRows>().value().type()));
+    auto tensor = var->Get<SelectedRows>().value();
+    if (UNLIKELY(!tensor.IsInitialized())) {
+      return "uninited";
+    } else {
+      return DataTypeToString(ToDataType(tensor.type()));
+    }
   } else {
     return "";
   }
@@ -197,16 +207,21 @@ std::string OperatorBase::DebugStringEx(const Scope* scope) const {
     auto& input = *it;
     ss << input.first << "[";
     for (size_t i = 0; i < input.second.size(); ++i) {
-      ss << input.second[i];
+      auto var_name = input.second[i];
+      ss << var_name;
       if (scope) {
-        int row_size = GetRowSize(*scope, input.second[i]);
-        if (row_size >= 0) {
-          ss << "[row_size=" << row_size << "]";
+        if (!VarInited(*scope, var_name)) {
+          ss << "[uninited]";
+        } else {
+          int row_size = GetRowSize(*scope, var_name);
+          if (row_size >= 0) {
+            ss << "[row_size=" << row_size << "]";
+          }
+          std::string dtype = GetDtype(*scope, var_name);
+          ss << ":" << dtype;
+          ss << "[" << GetDims(*scope, var_name, true) << "]";
+          ss << "(" << GetLoD(*scope, var_name) << ")";
         }
-        std::string dtype = GetDtype(*scope, input.second[i]);
-        ss << ":" << dtype;
-        ss << "[" << GetDims(*scope, input.second[i], true) << "]";
-        ss << "(" << GetLoD(*scope, input.second[i]) << ")";
       }
       if (i != input.second.size() - 1) {
         ss << ", ";
@@ -223,14 +238,19 @@ std::string OperatorBase::DebugStringEx(const Scope* scope) const {
     auto& output = *it;
     ss << output.first << "[";
     for (size_t i = 0; i < output.second.size(); ++i) {
-      ss << output.second[i];
+      auto var_name = output.second[i];
+      ss << var_name;
       if (scope) {
-        int row_size = GetRowSize(*scope, output.second[i]);
-        if (row_size >= 0) {
-          ss << "[row_size=" << row_size << "]";
+        if (!VarInited(*scope, var_name)) {
+          ss << "[uninited]";
+        } else {
+          int row_size = GetRowSize(*scope, output.second[i]);
+          if (row_size >= 0) {
+            ss << "[row_size=" << row_size << "]";
+          }
+          ss << "[" << GetDims(*scope, var_name, true) << "]";
+          ss << "(" << GetLoD(*scope, var_name) << ")";
         }
-        ss << "[" << GetDims(*scope, output.second[i], true) << "]";
-        ss << "(" << GetLoD(*scope, output.second[i]) << ")";
       }
       if (i != output.second.size() - 1) {
         ss << ", ";
diff --git a/paddle/fluid/framework/rw_lock.h b/paddle/fluid/framework/rw_lock.h
index 1418fb5134..a068d3543d 100644
--- a/paddle/fluid/framework/rw_lock.h
+++ b/paddle/fluid/framework/rw_lock.h
@@ -14,13 +14,16 @@ limitations under the License. */
 
 #pragma once
 
+#if !defined(_WIN32)
 #include <pthread.h>
+#endif  // !_WIN32
 
 #include "paddle/fluid/platform/enforce.h"
 
 namespace paddle {
 namespace framework {
 
+#if !defined(_WIN32)
 struct RWLock {
   RWLock() { pthread_rwlock_init(&lock_, nullptr); }
 
@@ -43,6 +46,15 @@ struct RWLock {
  private:
   pthread_rwlock_t lock_;
 };
+#else
+// https://stackoverflow.com/questions/7125250/making-pthread-rwlock-wrlock-recursive
+// In windows, rw_lock seems like a hack. Use empty object and do nothing.
+struct RWLock {
+  void RDLock() {}
+  void WRLock() {}
+  void UNLock() {}
+};
+#endif
 
 }  // namespace framework
 }  // namespace paddle
diff --git a/paddle/fluid/inference/api/demo_ci/CMakeLists.txt b/paddle/fluid/inference/api/demo_ci/CMakeLists.txt
index ba73a6eaa6..a697218377 100644
--- a/paddle/fluid/inference/api/demo_ci/CMakeLists.txt
+++ b/paddle/fluid/inference/api/demo_ci/CMakeLists.txt
@@ -23,9 +23,11 @@ include_directories("${PADDLE_LIB}")
 include_directories("${PADDLE_LIB}/third_party/install/protobuf/include")
 include_directories("${PADDLE_LIB}/third_party/install/glog/include")
 include_directories("${PADDLE_LIB}/third_party/install/gflags/include")
+if (NOT WIN32)
 include_directories("${PADDLE_LIB}/third_party/install/snappy/include")
 include_directories("${PADDLE_LIB}/third_party/install/snappystream/include")
 include_directories("${PADDLE_LIB}/third_party/install/zlib/include")
+endif(NOT WIN32)
 
 include_directories("${PADDLE_LIB}/third_party/boost")
 include_directories("${PADDLE_LIB}/third_party/eigen3")
diff --git a/paddle/fluid/memory/detail/system_allocator.cc b/paddle/fluid/memory/detail/system_allocator.cc
index 9b1ab1e228..1b96798d23 100644
--- a/paddle/fluid/memory/detail/system_allocator.cc
+++ b/paddle/fluid/memory/detail/system_allocator.cc
@@ -11,12 +11,18 @@ distributed under the License is distributed on an "AS IS" BASIS,
 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
+#define GLOG_NO_ABBREVIATED_SEVERITIES
 
 #include "paddle/fluid/memory/detail/system_allocator.h"
 
-#include <stdlib.h>    // for malloc and free
+#ifdef _WIN32
+#include <malloc.h>
+#include <windows.h>  // VirtualLock/VirtualUnlock
+#else
 #include <sys/mman.h>  // for mlock and munlock
-#include <algorithm>   // for std::max
+#endif
+#include <stdlib.h>   // for malloc and free
+#include <algorithm>  // for std::max
 
 #include "gflags/gflags.h"
 #include "paddle/fluid/platform/assert.h"
@@ -35,31 +41,42 @@ namespace paddle {
 namespace memory {
 namespace detail {
 
-void* CPUAllocator::Alloc(size_t* index, size_t size) {
-  // According to http://www.cplusplus.com/reference/cstdlib/malloc/,
-  // malloc might not return nullptr if size is zero, but the returned
-  // pointer shall not be dereferenced -- so we make it nullptr.
-  if (size <= 0) return nullptr;
-
-  *index = 0;  // unlock memory
-
+void* AlignedMalloc(size_t size) {
   void* p = nullptr;
-
+  size_t alignment = 32ul;
 #ifdef PADDLE_WITH_MKLDNN
   // refer to https://github.com/01org/mkl-dnn/blob/master/include/mkldnn.hpp
   // memory alignment
-  PADDLE_ENFORCE_EQ(posix_memalign(&p, 4096ul, size), 0, "Alloc %ld error!",
-                    size);
+  alignment = 4096ul;
+#endif
+#ifdef _WIN32
+  p = _aligned_malloc(size, alignment);
 #else
-  PADDLE_ENFORCE_EQ(posix_memalign(&p, 32ul, size), 0, "Alloc %ld error!",
+  PADDLE_ENFORCE_EQ(posix_memalign(&p, alignment, size), 0, "Alloc %ld error!",
                     size);
 #endif
   PADDLE_ENFORCE(p, "Fail to allocate CPU memory: size = %d .", size);
+  return p;
+}
+
+void* CPUAllocator::Alloc(size_t* index, size_t size) {
+  // According to http://www.cplusplus.com/reference/cstdlib/malloc/,
+  // malloc might not return nullptr if size is zero, but the returned
+  // pointer shall not be dereferenced -- so we make it nullptr.
+  if (size <= 0) return nullptr;
+
+  *index = 0;  // unlock memory
+
+  void* p = AlignedMalloc(size);
 
   if (p != nullptr) {
     if (FLAGS_use_pinned_memory) {
       *index = 1;
+#ifdef _WIN32
+      VirtualLock(p, size);
+#else
       mlock(p, size);  // lock memory
+#endif
     }
   }
 
@@ -68,7 +85,11 @@ void* CPUAllocator::Alloc(size_t* index, size_t size) {
 
 void CPUAllocator::Free(void* p, size_t size, size_t index) {
   if (p != nullptr && index == 1) {
+#ifdef _WIN32
+    VirtualUnlock(p, size);
+#else
     munlock(p, size);
+#endif
   }
   free(p);
 }
diff --git a/paddle/fluid/operators/CMakeLists.txt b/paddle/fluid/operators/CMakeLists.txt
index 68fbde2c09..8da0aaaafe 100644
--- a/paddle/fluid/operators/CMakeLists.txt
+++ b/paddle/fluid/operators/CMakeLists.txt
@@ -85,7 +85,7 @@ function(op_library TARGET)
 
     #remove windows unsupported op
     if (WIN32)
-    foreach(windows_unsupport_op "nccl_op" "gen_nccl_id_op")
+    foreach(windows_unsupport_op "nccl_op" "gen_nccl_id_op" "warpctc_op")
         if ("${TARGET}" STREQUAL "${windows_unsupport_op}")
           return()
         endif()
@@ -319,8 +319,9 @@ foreach(src ${GENERAL_OPS})
 endforeach()
 
 file(APPEND ${pybind_file} "USE_OP(less_than);\nUSE_OP(logical_and);\nUSE_NO_KERNEL_OP(read_from_array);\n")
-
+if (NOT WIN32)
 add_subdirectory(reader)
+endif(NOT WIN32)
 foreach(src ${READER_LIBRARY})
     set(OP_LIBRARY ${src} ${OP_LIBRARY})
 endforeach()
diff --git a/paddle/fluid/operators/auc_op.h b/paddle/fluid/operators/auc_op.h
index 0a18585edb..0651203286 100644
--- a/paddle/fluid/operators/auc_op.h
+++ b/paddle/fluid/operators/auc_op.h
@@ -60,6 +60,20 @@ class AucKernel : public framework::OpKernel<T> {
     const T* inference_data = predict->data<T>();
     const auto* label_data = label->data<int64_t>();
 
+    // check if states are inited.
+    auto* tp_in = ctx.Input<Tensor>("TP");
+    auto* fp_in = ctx.Input<Tensor>("FP");
+    auto* tn_in = ctx.Input<Tensor>("TN");
+    auto* fn_in = ctx.Input<Tensor>("FN");
+    PADDLE_ENFORCE(tp_in->IsInitialized(), "true_positive is not inited!");
+    PADDLE_ENFORCE(fp_in->IsInitialized(), "false_negative is not inited!");
+    PADDLE_ENFORCE(tn_in->IsInitialized(), "true_negative is not inited!");
+    PADDLE_ENFORCE(fn_in->IsInitialized(), "false_positive is not inited!");
+    PADDLE_ENFORCE_EQ(tp_in->numel(), num_thresholds, "");
+    PADDLE_ENFORCE_EQ(fp_in->numel(), num_thresholds, "");
+    PADDLE_ENFORCE_EQ(tn_in->numel(), num_thresholds, "");
+    PADDLE_ENFORCE_EQ(fn_in->numel(), num_thresholds, "");
+
     auto* tp_data = true_positive->mutable_data<int64_t>(ctx.GetPlace());
     auto* fn_data = false_negative->mutable_data<int64_t>(ctx.GetPlace());
     auto* tn_data = true_negative->mutable_data<int64_t>(ctx.GetPlace());
diff --git a/paddle/fluid/operators/math/cpu_vec_test.cc b/paddle/fluid/operators/math/cpu_vec_test.cc
index bf6481c5cc..3ce66f49ed 100644
--- a/paddle/fluid/operators/math/cpu_vec_test.cc
+++ b/paddle/fluid/operators/math/cpu_vec_test.cc
@@ -15,6 +15,7 @@ limitations under the License. */
 #include <sys/time.h>
 #include <cmath>
 #include <cstring>
+#include <random>
 #include <vector>
 #include "gflags/gflags.h"
 #include "glog/logging.h"
diff --git a/paddle/fluid/operators/math/math_function.h b/paddle/fluid/operators/math/math_function.h
index 7ec78d9ef8..c63ad89e46 100644
--- a/paddle/fluid/operators/math/math_function.h
+++ b/paddle/fluid/operators/math/math_function.h
@@ -19,6 +19,10 @@ limitations under the License. */
 
 #ifdef PADDLE_USE_OPENBLAS
 #include <cblas.h>
+// remove typedef in openblas
+#undef FLOAT
+#undef INT
+#undef SIZE
 #endif
 
 #include <cmath>
diff --git a/paddle/fluid/operators/stack_op.h b/paddle/fluid/operators/stack_op.h
index c777d5feae..d236c5b943 100644
--- a/paddle/fluid/operators/stack_op.h
+++ b/paddle/fluid/operators/stack_op.h
@@ -150,30 +150,17 @@ class StackKernel : public framework::OpKernel<T> {
     int total_num = pre * n * post;
 
     auto &dev_ctx = ctx.template device_context<DeviceContext>();
-    constexpr auto kMaxThreshold = 16;
-    if (std::is_same<DeviceContext, platform::CPUDeviceContext>::value ||
-        n > kMaxThreshold) {
 #ifdef __NVCC__
-      VLOG(10) << "Stack more than " << kMaxThreshold
-               << " tensors on GPU may be slow.";
-      thrust::device_vector<const T *> device_x_vec(x_datas);
-      auto x_data_arr = device_x_vec.data().get();
+    thrust::device_vector<const T *> device_x_vec(x_datas);
+    auto x_data_arr = device_x_vec.data().get();
 #else
-      auto x_data_arr = x_datas.data();
+    auto x_data_arr = x_datas.data();
 #endif
-      StackFunctorForRange(dev_ctx, x_data_arr, y_data, total_num, n, post);
+    StackFunctorForRange(dev_ctx, x_data_arr, y_data, total_num, n, post);
 #ifdef __NVCC__
-      // Wait() must be called because device_x_vec may be destructed before
-      // kernel ends
-      dev_ctx.Wait();
-#endif
-    }
-#ifdef __NVCC__
-    else {  // NOLINT
-      framework::Array<const T *, kMaxThreshold> x_data_arr;
-      for (int i = 0; i < n; ++i) x_data_arr[i] = x_datas[i];
-      StackFunctorForRange(dev_ctx, x_data_arr, y_data, total_num, n, post);
-    }
+    // Wait() must be called because device_x_vec may be destructed before
+    // kernel ends
+    dev_ctx.Wait();
 #endif
   }
 };
@@ -244,32 +231,17 @@ class StackGradKernel : public framework::OpKernel<T> {
     int post = total_num / (n * pre);
 
     auto &dev_ctx = ctx.template device_context<DeviceContext>();
-    constexpr auto kMaxThreshold = 16;
-    if (std::is_same<DeviceContext, platform::CPUDeviceContext>::value ||
-        n > kMaxThreshold) {
 #ifdef __NVCC__
-      VLOG(10) << "Stack more than " << kMaxThreshold
-               << " tensors on GPU may be slow.";
-      thrust::device_vector<T *> device_dx_vec(dx_datas);
-      auto dx_data_arr = device_dx_vec.data().get();
+    thrust::device_vector<T *> device_dx_vec(dx_datas);
+    auto dx_data_arr = device_dx_vec.data().get();
 #else
-      auto dx_data_arr = dx_datas.data();
+    auto dx_data_arr = dx_datas.data();
 #endif
-      StackGradFunctorForRange(dev_ctx, dx_data_arr, dy_data, total_num, n,
-                               post);
+    StackGradFunctorForRange(dev_ctx, dx_data_arr, dy_data, total_num, n, post);
 #ifdef __NVCC__
-      // Wait() must be called because device_dx_vec may be destructed before
-      // kernel ends
-      dev_ctx.Wait();
-#endif
-    }
-#ifdef __NVCC__
-    else {  // NOLINT
-      framework::Array<T *, kMaxThreshold> dx_data_arr;
-      for (int i = 0; i < n; ++i) dx_data_arr[i] = dx_datas[i];
-      StackGradFunctorForRange(dev_ctx, dx_data_arr, dy_data, total_num, n,
-                               post);
-    }
+    // Wait() must be called because device_dx_vec may be destructed before
+    // kernel ends
+    dev_ctx.Wait();
 #endif
   }
 };
diff --git a/paddle/fluid/platform/float16.h b/paddle/fluid/platform/float16.h
index efb021c838..ee16fc66e4 100644
--- a/paddle/fluid/platform/float16.h
+++ b/paddle/fluid/platform/float16.h
@@ -56,7 +56,11 @@ limitations under the License. */
 #include <immintrin.h>
 #endif  // PADDLE_ARM
 
+#if !defined(_WIN32)
 #define PADDLE_ALIGN(x) __attribute__((aligned(x)))
+#else
+#define PADDLE_ALIGN(x) /*do nothing*/
+#endif
 
 namespace paddle {
 namespace platform {
diff --git a/paddle/fluid/pybind/CMakeLists.txt b/paddle/fluid/pybind/CMakeLists.txt
index d6a14b3305..b5bd07d401 100644
--- a/paddle/fluid/pybind/CMakeLists.txt
+++ b/paddle/fluid/pybind/CMakeLists.txt
@@ -1,17 +1,19 @@
-set(PYBIND_DEPS pybind python proto_desc memory executor prune profiler feed_fetch_method
-          )
+
+set(PYBIND_DEPS pybind python proto_desc memory executor prune  feed_fetch_method)
+set(PYBIND_SRCS pybind.cc exception.cc protobuf.cc const_value.cc)
 if(NOT WIN32)
-list(APPEND PYBIND_DEPS parallel_executor)
+list(APPEND PYBIND_DEPS parallel_executor profiler)
+list(APPEND PYBIND_SRCS recordio.cc)
 endif()
 if(WITH_PYTHON)
   if(WITH_AMD_GPU)
     hip_library(paddle_pybind SHARED
-      SRCS pybind.cc exception.cc protobuf.cc const_value.cc recordio.cc
+      SRCS ${PYBIND_SRCS}
       DEPS ${PYBIND_DEPS}
       ${GLOB_OP_LIB})
   else()
     cc_library(paddle_pybind SHARED
-      SRCS pybind.cc exception.cc protobuf.cc const_value.cc recordio.cc
+      SRCS ${PYBIND_SRCS}
       DEPS ${PYBIND_DEPS}
       ${GLOB_OP_LIB})
     if(NOT APPLE AND NOT ANDROID AND NOT WIN32)
diff --git a/python/paddle/fluid/layers/metric_op.py b/python/paddle/fluid/layers/metric_op.py
index 2c3bdd77e1..0182bbeb63 100644
--- a/python/paddle/fluid/layers/metric_op.py
+++ b/python/paddle/fluid/layers/metric_op.py
@@ -119,10 +119,14 @@ def auc(input, label, curve='ROC', num_thresholds=200, topk=1):
     helper = LayerHelper("auc", **locals())
     auc_out = helper.create_tmp_variable(dtype="float64")
     # make tp, tn, fp, fn persistable, so that can accumulate all batches.
-    tp = helper.create_global_variable(persistable=True, dtype='int64')
-    tn = helper.create_global_variable(persistable=True, dtype='int64')
-    fp = helper.create_global_variable(persistable=True, dtype='int64')
-    fn = helper.create_global_variable(persistable=True, dtype='int64')
+    tp = helper.create_global_variable(
+        persistable=True, dtype='int64', shape=[num_thresholds])
+    tn = helper.create_global_variable(
+        persistable=True, dtype='int64', shape=[num_thresholds])
+    fp = helper.create_global_variable(
+        persistable=True, dtype='int64', shape=[num_thresholds])
+    fn = helper.create_global_variable(
+        persistable=True, dtype='int64', shape=[num_thresholds])
     for var in [tp, tn, fp, fn]:
         helper.set_variable_initializer(
             var, Constant(