merge develop

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)

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)

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)
@@ -36,6 +41,7 @@ set_directory_properties(PROPERTIES CLEAN_NO_CUSTOM 1)
 
 include_directories(${BOOST_INCLUDE_DIR})
 
+if (NOT WIN32)
 ExternalProject_Add(
     ${BOOST_PROJECT}
     ${EXTERNAL_PROJECT_LOG_ARGS}
@@ -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})

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()
+

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})

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)

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)
 

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})

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

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)

paddle/fluid/framework/CMakeLists.txt

-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

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);

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();

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

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 (!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, input.second[i]);
+          std::string dtype = GetDtype(*scope, var_name);
           ss << ":" << dtype;
-        ss << "[" << GetDims(*scope, input.second[i], true) << "]";
-        ss << "(" << GetLoD(*scope, input.second[i]) << ")";
+          ss << "[" << GetDims(*scope, var_name, true) << "]";
+          ss << "(" << GetLoD(*scope, var_name) << ")";
+        }
       }
       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) {
+        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, output.second[i], true) << "]";
-        ss << "(" << GetLoD(*scope, output.second[i]) << ")";
+          ss << "[" << GetDims(*scope, var_name, true) << "]";
+          ss << "(" << GetLoD(*scope, var_name) << ")";
+        }
       }
       if (i != output.second.size() - 1) {
         ss << ", ";

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

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")

paddle/fluid/memory/detail/system_allocator.cc

@@ -11,11 +11,17 @@ 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
+#endif
+#include <stdlib.h>   // for malloc and free
 #include <algorithm>  // for std::max
 
 #include "gflags/gflags.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);
 }

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()

paddle/fluid/operators/attention_lstm_op.cc

@@ -232,40 +232,28 @@ use lstm_x_t as input and compute as standard LSTM.
 template <typename T>
 inline void bias_relu(const int n, const T* x, const T* bias, T* y) {
   if (bias) {
-    for (int i = 0; i < n; ++i) {
-      y[i] = x[i] + bias[0];
-    }
-    math::vec_relu<T>(n, y, y);
+    math::vec_add_bias<T, platform::jit::avx>(n, *bias, x, y);
+    math::vec_relu<T, platform::jit::avx>(n, y, y);
   } else {
-    math::vec_relu<T>(n, x, y);
+    math::vec_relu<T, platform::jit::avx>(n, x, y);
   }
 }
 
-template <typename DeviceContext, typename T>
-inline void vec_softmax(const math::BlasT<DeviceContext, T>& blas, const int n,
-                        const T* x, T* y) {
+template <typename T>
+inline void vec_softmax(const int n, const T* x, T* y) {
   T scalar = x[0];
   // max
   for (int i = 1; i < n; ++i) {
     scalar = scalar < x[i] ? x[i] : scalar;
   }
-
-  // sub
-  for (int i = 0; i < n; ++i) {
-    y[i] = x[i] - scalar;
-  }
-
-  // exp
-  blas.VEXP(n, y, y);
-
+  math::vec_add_bias<T, platform::jit::avx>(n, -scalar, x, y);  // sub
+  math::vec_exp<T>(n, y, y);                                    // exp
   // sum
   scalar = T(0);
   for (int i = 0; i < n; ++i) {
     scalar += y[i];
   }
-
-  // scale
-  blas.SCAL(n, static_cast<T>(1) / scalar, y);
+  math::vec_scal<T>(n, static_cast<T>(1) / scalar, y);  // scale
 }
 
 template <typename T>
@@ -311,11 +299,21 @@ class AttentionLSTMKernel : public framework::OpKernel<T> {
     PADDLE_ENFORCE_EQ(c0->dims()[0], N, "C0 dims should be %d x %d.", N, D);
     fc_out->Resize({max_seq_len, 1});
 
-    math::VecActivations<T> act_functor;
     std::function<void(const int, const T *, T *)> act_gate, act_cell, act_cand;
-    act_gate = act_functor(ctx.Attr<std::string>("gate_activation"));
-    act_cell = act_functor(ctx.Attr<std::string>("cell_activation"));
-    act_cand = act_functor(ctx.Attr<std::string>("candidate_activation"));
+    auto& act_gate_str = ctx.Attr<std::string>("gate_activation");
+    auto& act_cell_str = ctx.Attr<std::string>("cell_activation");
+    auto& act_cand_str = ctx.Attr<std::string>("candidate_activation");
+    if (platform::jit::MayIUse(platform::jit::avx)) {
+      math::VecActivations<T, platform::jit::avx> act_functor;
+      act_gate = act_functor(act_gate_str);
+      act_cell = act_functor(act_cell_str);
+      act_cand = act_functor(act_cand_str);
+    } else {
+      math::VecActivations<T, platform::jit::isa_any> act_functor;
+      act_gate = act_functor(act_gate_str);
+      act_cell = act_functor(act_cell_str);
+      act_cand = act_functor(act_cand_str);
+    }
 
     const T* x_data = x->data<T>();
     const T* h0_data = h0 ? h0->data<T>() : NULL;
@@ -363,7 +361,7 @@ class AttentionLSTMKernel : public framework::OpKernel<T> {
                        fc_out_data);
         }
         // 1d. softmax
-        vec_softmax<DeviceContext, T>(blas, seq_len, fc_out_data, fc_out_data);
+        vec_softmax<T>(seq_len, fc_out_data, fc_out_data);
         // mul x(seq_len*M) and sum pool
         math::FCCompute<DeviceContext, T>(blas, 1, M, seq_len, fc_out_data,
                                           cur_x_data, lstm_x_data);

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());

paddle/fluid/operators/fusion_lstm_op.cc

@@ -15,10 +15,14 @@ limitations under the License. */
 #include "paddle/fluid/operators/fusion_lstm_op.h"
 #include <string>
 #include "paddle/fluid/operators/math/blas.h"
+#include "paddle/fluid/operators/math/cpu_vec.h"
 #include "paddle/fluid/operators/math/detail/activation_functions.h"
 #include "paddle/fluid/operators/math/fc_compute.h"
 #include "paddle/fluid/operators/math/lstm_compute.h"
 #include "paddle/fluid/operators/math/sequence2batch.h"
+#include "paddle/fluid/platform/cpu_info.h"
+
+DEFINE_bool(seq_mode, true, "Use sequence mode");
 
 namespace paddle {
 namespace operators {
@@ -98,7 +102,12 @@ void FusionLSTMOp::InferShape(framework::InferShapeContext* ctx) const {
   ctx->ShareLoD("X", "Hidden");
   ctx->ShareLoD("X", "Cell");
 
-  int xx_width = x_dims[1] > wx_dims[1] ? wx_dims[1] : x_dims[1];
+  int xx_width;
+  if (FLAGS_seq_mode) {
+    xx_width = wx_dims[1];
+  } else {
+    xx_width = x_dims[1] > wx_dims[1] ? wx_dims[1] : x_dims[1];
+  }
   ctx->SetOutputDim("XX", {x_dims[0], xx_width});
   ctx->ShareLoD("X", "XX");
 }
@@ -205,10 +214,138 @@ inline void ReorderInitState(const DeviceContext& ctx,
   row_shuffle(ctx, src, index_lod, dst, indexed_src);
 }
 
-template <typename DeviceContext, typename T>
+template <typename T>
 class FuisonLSTMKernel : public framework::OpKernel<T> {
  public:
-  void Compute(const framework::ExecutionContext& ctx) const override {
+  void SeqCompute(const framework::ExecutionContext& ctx) const {
+    using DeviceContext = paddle::platform::CPUDeviceContext;
+    auto* x = ctx.Input<LoDTensor>("X");
+    auto* h0 = ctx.Input<Tensor>("H0");
+    auto* c0 = ctx.Input<Tensor>("C0");
+    auto* wx = ctx.Input<Tensor>("WeightX");
+    auto* wh = ctx.Input<Tensor>("WeightH");
+    auto* bias = ctx.Input<Tensor>("Bias");
+
+    auto* xx = ctx.Output<LoDTensor>("XX");
+    auto* hidden_out = ctx.Output<LoDTensor>("Hidden");
+    auto* cell_out = ctx.Output<LoDTensor>("Cell");
+    bool is_reverse = ctx.Attr<bool>("is_reverse");
+
+    std::function<void(const int, const T *, T *)> act_gate, act_cell, act_cand;
+    auto& act_gate_str = ctx.Attr<std::string>("gate_activation");
+    auto& act_cell_str = ctx.Attr<std::string>("cell_activation");
+    auto& act_cand_str = ctx.Attr<std::string>("candidate_activation");
+    if (platform::jit::MayIUse(platform::jit::avx)) {
+      math::VecActivations<T, platform::jit::avx> act_functor;
+      act_gate = act_functor(act_gate_str);
+      act_cell = act_functor(act_cell_str);
+      act_cand = act_functor(act_cand_str);
+    } else {
+      math::VecActivations<T, platform::jit::isa_any> act_functor;
+      act_gate = act_functor(act_gate_str);
+      act_cell = act_functor(act_cell_str);
+      act_cand = act_functor(act_cand_str);
+    }
+
+    auto x_lod = x->lod();
+    auto x_dims = x->dims();    // T x M
+    auto wh_dims = wh->dims();  // D x 4D
+    const int total_T = x_dims[0];
+    const int N = x_lod[0].size() - 1;  // batch size
+    const int M = x_dims[1];            // x frame size
+    const int D = wh_dims[0];
+    const int D2 = D * 2;
+    const int D3 = D * 3;
+    const int D4 = wh_dims[1];
+
+    const T* x_data = x->data<T>();
+    const T* h0_data = h0 ? h0->data<T>() : NULL;
+    const T* c0_data = c0 ? c0->data<T>() : NULL;
+    const T* wx_data = wx->data<T>();
+    const T* wh_data = wh->data<T>();
+    T* xx_data = xx->mutable_data<T>(ctx.GetPlace());
+    T* hidden_out_data = hidden_out->mutable_data<T>(ctx.GetPlace());
+    T* cell_out_data = cell_out->mutable_data<T>(ctx.GetPlace());
+
+    auto blas = math::GetBlas<DeviceContext, T>(ctx);
+    math::FCCompute<DeviceContext, T>(blas, total_T, D4, M, x_data, wx_data,
+                                      xx_data, bias->data<T>());
+    int xx_offset = D4;
+    int gate_offset = D;
+    if (is_reverse) {
+      const int offset = (total_T - 1) * D;
+      xx_data = xx_data + offset * 4;
+      hidden_out_data = hidden_out_data + offset;
+      cell_out_data = cell_out_data + offset;
+      xx_offset = -D4;
+      gate_offset = -D;
+    }
+
+    auto move_step = [&]() {
+      xx_data = xx_data + xx_offset;
+      hidden_out_data = hidden_out_data + gate_offset;
+      cell_out_data = cell_out_data + gate_offset;
+    };
+
+    for (int i = 0; i < N; ++i) {
+      int bid = is_reverse ? N - 1 - i : i;
+      int seq_len = x_lod[0][bid + 1] - x_lod[0][bid];
+      const T* prev_cell_data = NULL;
+      const T* prev_hidden_data = NULL;
+      int tstart = 0;
+      if (h0_data) {
+        prev_hidden_data = h0_data + bid * D;
+        prev_cell_data = c0_data + bid * D;
+      } else {
+        // W_ch, W_ih, W_fh, W_oh
+        act_gate(D3, xx_data + D, xx_data + D);
+        act_cand(D, xx_data, xx_data);
+        // cell out= input*tilde
+        blas.VMUL(D, xx_data, xx_data + D, cell_out_data);
+        // hidden out= act_state(cellout) * outgate
+        act_cell(D, cell_out_data, xx_data + D2);
+        blas.VMUL(D, xx_data + D2, xx_data + D3, hidden_out_data);
+
+        // prev
+        prev_hidden_data = hidden_out_data;
+        prev_cell_data = cell_out_data;
+        tstart = 1;
+
+        move_step();
+      }
+      for (int step = tstart; step < seq_len; ++step) {
+        blas.GEMM(CblasNoTrans, CblasNoTrans, 1, D4, D, static_cast<T>(1),
+                  prev_hidden_data, D, wh_data, D4, static_cast<T>(1), xx_data,
+                  D4);
+
+        // W_ch, W_ih, W_fh, W_oh
+        act_gate(D3, xx_data + D, xx_data + D);
+        act_cand(D, xx_data, xx_data);
+
+        // a = forget * prev_cell
+        blas.VMUL(D, xx_data + D2, prev_cell_data, xx_data + D2);
+
+        // b = input * tilde
+        blas.VMUL(D, xx_data, xx_data + D, xx_data + D);
+
+        // cell out= a+b
+        blas.VADD(D, xx_data + D, xx_data + D2, cell_out_data);
+
+        // hidden out= act_state(cellout) * outgate
+        act_cell(D, cell_out_data, xx_data + D2);
+        blas.VMUL(D, xx_data + D2, xx_data + D3, hidden_out_data);
+
+        // prev
+        prev_hidden_data = hidden_out_data;
+        prev_cell_data = cell_out_data;
+
+        move_step();
+      }
+    }
+  }
+
+  void BatchCompute(const framework::ExecutionContext& ctx) const {
+    using DeviceContext = platform::CPUDeviceContext;
     auto* x = ctx.Input<LoDTensor>("X");
     auto* wx = ctx.Input<Tensor>("WeightX");
     auto* wh = ctx.Input<Tensor>("WeightH");
@@ -339,6 +476,13 @@ class FuisonLSTMKernel : public framework::OpKernel<T> {
     // restore the output cell state in LoDTensor from the batch cell
     to_seq(dev_ctx, batch_cell, cell_out);
   }
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    if (FLAGS_seq_mode) {
+      SeqCompute(ctx);
+    } else {
+      BatchCompute(ctx);
+    }
+  }
 };
 
 }  // namespace operators
@@ -348,7 +492,5 @@ namespace ops = paddle::operators;
 REGISTER_OPERATOR(fusion_lstm, ops::FusionLSTMOp, ops::FusionLSTMOpMaker,
                   paddle::framework::DefaultGradOpDescMaker<true>);
 
-REGISTER_OP_CPU_KERNEL(
-    fusion_lstm,
-    ops::FuisonLSTMKernel<paddle::platform::CPUDeviceContext, float>,
-    ops::FuisonLSTMKernel<paddle::platform::CPUDeviceContext, double>);
+REGISTER_OP_CPU_KERNEL(fusion_lstm, ops::FuisonLSTMKernel<float>,
+                       ops::FuisonLSTMKernel<double>);

paddle/fluid/operators/math/CMakeLists.txt

@@ -65,3 +65,4 @@ if(WITH_GPU)
     nv_test(selected_rows_functor_gpu_test SRCS selected_rows_functor_test.cu DEPS selected_rows_functor math_function)
 endif()
 cc_test(concat_test SRCS concat_test.cc DEPS concat)
+cc_test(cpu_vec_test SRCS cpu_vec_test.cc DEPS blas cpu_info)

paddle/fluid/operators/math/cpu_vec.h

@@ -13,8 +13,16 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
+#include <cmath>
 #include <string>
 #include "paddle/fluid/platform/cpu_info.h"
+#ifdef __AVX__
+#include <immintrin.h>
+#endif
+
+#ifdef PADDLE_WITH_MKLML
+#include "paddle/fluid/platform/dynload/mklml.h"
+#endif
 
 namespace paddle {
 namespace operators {
@@ -22,16 +30,161 @@ namespace math {
 
 #define SIGMOID_THRESHOLD_MIN -40.0
 #define SIGMOID_THRESHOLD_MAX 13.0
-#define EXP_MAX_INPUT 40.0
+
+#define AVX_FLOAT_BLOCK 8
+#define AVX_DOUBLE_BLOCK 4
+#define AVX2_FLOAT_BLOCK 8
+#define AVX2_DOUBLE_BLOCK 4
+#define AVX512_FLOAT_BLOCK 16
+#define AVX512_DOUBLE_BLOCK 8
 
 template <typename T>
-inline T sigmoid(T x) {
-  return 1. / (1. + exp(-x));
+inline void vec_exp(const int n, const T* x, T* y) {
+  for (int i = 0; i < n; ++i) {
+    y[i] = std::exp(x[i]);
+  }
 }
 
 template <typename T>
-inline T tanh(T x) {
-  return 2. * sigmoid(2. * x) - 1.;
+inline void vec_scal(const int n, const T a, T* x) {
+  for (int i = 0; i < n; ++i) {
+    x[i] = a * x[i];
+  }
+}
+
+#ifdef PADDLE_WITH_MKLML
+template <>
+inline void vec_exp<float>(const int n, const float* x, float* y) {
+  platform::dynload::vsExp(n, x, y);
+}
+
+template <>
+inline void vec_exp<double>(const int n, const double* x, double* y) {
+  platform::dynload::vdExp(n, x, y);
+}
+
+template <>
+inline void vec_scal<float>(const int n, const float a, float* x) {
+  platform::dynload::cblas_sscal(n, a, x, 1);
+}
+
+template <>
+inline void vec_scal<double>(const int n, const double a, double* x) {
+  platform::dynload::cblas_dscal(n, a, x, 1);
+}
+#endif
+
+// MKL scal only support inplace, choose this if src and dst are not equal
+template <typename T, platform::jit::cpu_isa_t isa = platform::jit::isa_any>
+inline void vec_scal(const int n, const T a, const T* x, T* y) {
+  for (int i = 0; i < n; ++i) {
+    y[i] = a * x[i];
+  }
+}
+
+template <>
+inline void vec_scal<float, platform::jit::avx>(const int n, const float a,
+                                                const float* x, float* y) {
+#ifdef __AVX__
+  constexpr int block = AVX_FLOAT_BLOCK;
+  if (n < block) {
+    vec_scal<float, platform::jit::isa_any>(n, a, x, y);
+    return;
+  }
+  const int rest = n % block;
+  const int end = n - rest;
+  int i = 0;
+  __m256 scalar = _mm256_set1_ps(a);
+  __m256 tmp;
+#define MOVE_ONE_STEP               \
+  tmp = _mm256_loadu_ps(x + i);     \
+  tmp = _mm256_mul_ps(tmp, scalar); \
+  _mm256_storeu_ps(y + i, tmp)
+  for (i = 0; i < end; i += block) {
+    MOVE_ONE_STEP;
+  }
+#undef MOVE_ONE_STEP
+  if (rest == 0) {
+    return;
+  }
+  // can not continue move step if src and dst are inplace
+  for (i = n - rest; i < n; ++i) {
+    y[i] = a * x[i];
+  }
+#else
+  vec_scal<float, platform::jit::isa_any>(n, a, x, y);
+#endif
+}
+
+template <>
+inline void vec_scal<float, platform::jit::avx2>(const int n, const float a,
+                                                 const float* x, float* y) {
+  vec_scal<float, platform::jit::avx>(n, a, x, y);
+}
+
+template <>
+inline void vec_scal<float, platform::jit::avx512_common>(const int n,
+                                                          const float a,
+                                                          const float* x,
+                                                          float* y) {
+  // TODO(TJ): enable me
+  vec_scal<float, platform::jit::avx2>(n, a, x, y);
+}
+
+template <typename T, platform::jit::cpu_isa_t isa = platform::jit::isa_any>
+inline void vec_add_bias(const int n, const T a, const T* x, T* y) {
+  for (int i = 0; i < n; ++i) {
+    y[i] = x[i] + a;
+  }
+}
+
+template <>
+inline void vec_add_bias<float, platform::jit::avx>(const int n, const float a,
+                                                    const float* x, float* y) {
+#ifdef __AVX__
+  constexpr int block = AVX_FLOAT_BLOCK;
+  if (n < block) {
+    vec_add_bias<float, platform::jit::isa_any>(n, a, x, y);
+    return;
+  }
+  const int rest = n % block;
+  const int end = n - rest;
+  int i = 0;
+  __m256 bias = _mm256_set1_ps(a);
+  __m256 tmp;
+#define MOVE_ONE_STEP             \
+  tmp = _mm256_loadu_ps(x + i);   \
+  tmp = _mm256_add_ps(tmp, bias); \
+  _mm256_storeu_ps(y + i, tmp)
+  for (i = 0; i < end; i += block) {
+    MOVE_ONE_STEP;
+  }
+#undef MOVE_ONE_STEP
+  if (rest == 0) {
+    return;
+  }
+  // can not continue move step if src and dst are inplace
+  for (i = n - rest; i < n; ++i) {
+    y[i] = x[i] + a;
+  }
+#else
+  vec_add_bias<float, platform::jit::isa_any>(n, a, x, y);
+#endif
+}
+
+template <>
+inline void vec_add_bias<float, platform::jit::avx2>(const int n, const float a,
+                                                     const float* x, float* y) {
+  vec_add_bias<float, platform::jit::avx>(n, a, x, y);
+}
+
+template <>
+inline void vec_add_bias<float, platform::jit::avx512_common>(const int n,
+                                                              const float a,
+                                                              const float* x,
+                                                              float* y) {
+  // TODO(TJ): enable me
+  vec_add_bias<float, platform::jit::avx2>(n, a, x, y);
 }
 
 template <typename T, platform::jit::cpu_isa_t isa = platform::jit::isa_any>
@@ -45,18 +198,97 @@ inline void vec_sigmoid(const int n, const T* x, T* y) {
   const T min = SIGMOID_THRESHOLD_MIN;
   const T max = SIGMOID_THRESHOLD_MAX;
   for (int i = 0; i < n; ++i) {
-    T tmp = (x[i] < min) ? min : ((x[i] > max) ? max : x[i]);
-    y[i] = 1.0 / (1.0 + std::exp(-tmp));
+    y[i] = (x[i] < min) ? min : ((x[i] > max) ? max : x[i]);
+    y[i] = static_cast<T>(0) - y[i];
+  }
+  vec_exp<T>(n, y, y);
+  for (int i = 0; i < n; ++i) {
+    y[i] = static_cast<T>(1) / (static_cast<T>(1) + y[i]);
+  }
+}
+
+template <>
+inline void vec_sigmoid<float, platform::jit::avx>(const int n, const float* x,
+                                                   float* y) {
+#ifdef __AVX__
+  constexpr int block = AVX_FLOAT_BLOCK;
+  if (n < block) {
+    vec_sigmoid<float, platform::jit::isa_any>(n, x, y);
+    return;
   }
+  const int rest = n % block;
+  const int end = n - rest;
+  int i = 0;
+  __m256 max = _mm256_set1_ps(SIGMOID_THRESHOLD_MAX);
+  __m256 min = _mm256_set1_ps(SIGMOID_THRESHOLD_MIN);
+  __m256 zeros = _mm256_setzero_ps();
+  __m256 tmp;
+#define MOVE_ONE_STEP              \
+  tmp = _mm256_loadu_ps(x + i);    \
+  tmp = _mm256_max_ps(tmp, min);   \
+  tmp = _mm256_min_ps(tmp, max);   \
+  tmp = _mm256_sub_ps(zeros, tmp); \
+  _mm256_storeu_ps(y + i, tmp)
+  for (i = 0; i < end; i += block) {
+    MOVE_ONE_STEP;
+  }
+#undef MOVE_ONE_STEP
+  if (rest != 0) {
+    // can not continue move step since the src and dst address could be equal
+    const float xmin = SIGMOID_THRESHOLD_MIN;
+    const float xmax = SIGMOID_THRESHOLD_MAX;
+    for (i = n - rest; i < n; ++i) {
+      y[i] = 0.f - ((x[i] < xmin) ? xmin : ((x[i] > xmax) ? xmax : x[i]));
+    }
+  }
+
+  vec_exp<float>(n, y, y);
+
+  __m256 ones = _mm256_set1_ps(1.0f);
+#define MOVE_ONE_STEP             \
+  tmp = _mm256_loadu_ps(y + i);   \
+  tmp = _mm256_add_ps(ones, tmp); \
+  tmp = _mm256_div_ps(ones, tmp); \
+  _mm256_storeu_ps(y + i, tmp)
+  for (i = 0; i < end; i += block) {
+    MOVE_ONE_STEP;
+  }
+#undef MOVE_ONE_STEP
+  if (rest == 0) {
+    return;
+  }
+  // can not continue move step
+  for (i = n - rest; i < n; ++i) {
+    y[i] = 1.f / (1.f + y[i]);
+  }
+#else
+  vec_sigmoid<float, platform::jit::isa_any>(n, x, y);
+#endif
+}
+
+template <>
+inline void vec_sigmoid<float, platform::jit::avx2>(const int n, const float* x,
+                                                    float* y) {
+  vec_sigmoid<float, platform::jit::avx>(n, x, y);
+}
+
+template <>
+inline void vec_sigmoid<float, platform::jit::avx512_common>(const int n,
+                                                             const float* x,
+                                                             float* y) {
+  // TODO(TJ): enable me
+  vec_sigmoid<float, platform::jit::avx2>(n, x, y);
 }
 
 template <typename T, platform::jit::cpu_isa_t isa = platform::jit::isa_any>
 inline void vec_tanh(const int n, const T* x, T* y) {
-  for (int i = 0; i < n; ++i) {
-    y[i] = tanh<T>(x[i]);
-  }
+  vec_scal<T, isa>(n, static_cast<T>(2), x, y);
+  vec_sigmoid<T, isa>(n, y, y);
+  vec_scal<T>(n, static_cast<T>(2), y);
+  vec_add_bias<T, isa>(n, static_cast<T>(-1), y, y);
 }
 
+// TODO(TJ): make relu clip
 template <typename T, platform::jit::cpu_isa_t isa = platform::jit::isa_any>
 inline void vec_relu(const int n, const T* x, T* y) {
   for (int i = 0; i < n; ++i) {
@@ -65,23 +297,55 @@ inline void vec_relu(const int n, const T* x, T* y) {
 }
 
 template <>
-inline void vec_relu<float, platform::jit::avx2>(const int n, const float* x,
+inline void vec_relu<float, platform::jit::avx>(const int n, const float* x,
                                                 float* y) {
-  // TODO(TJ): complete me
-  for (int i = 0; i < n; ++i) {
-    y[i] = x[i] > 0 ? x[i] : 0;
+#ifdef __AVX__
+  constexpr int block = AVX_FLOAT_BLOCK;
+  if (n < block * 4) {
+    vec_relu<float, platform::jit::isa_any>(n, x, y);
+    return;
+  }
+
+  const int rest = n % block;
+  const int end = n - rest;
+  int i = 0;
+  __m256 zeros = _mm256_setzero_ps();
+  __m256 tmp;
+#define MOVE_ONE_STEP              \
+  tmp = _mm256_loadu_ps(x + i);    \
+  tmp = _mm256_max_ps(tmp, zeros); \
+  _mm256_storeu_ps(y + i, tmp)
+  for (i = 0; i < end; i += block) {
+    MOVE_ONE_STEP;
   }
+  if (rest == 0) {
+    return;
+  }
+  i = n - block;
+  MOVE_ONE_STEP;
+#undef MOVE_ONE_STEP
+
+#else
+  vec_relu<float, platform::jit::isa_any>(n, x, y);
+#endif
 }
 
 template <>
-inline void vec_relu<float, platform::jit::avx>(const int n, const float* x,
+inline void vec_relu<float, platform::jit::avx2>(const int n, const float* x,
                                                  float* y) {
-  // TODO(TJ): complete me
-  for (int i = 0; i < n; ++i) {
-    y[i] = x[i] > 0 ? x[i] : 0;
-  }
+  vec_relu<float, platform::jit::avx>(n, x, y);
+}
+
+template <>
+inline void vec_relu<float, platform::jit::avx512_common>(const int n,
+                                                          const float* x,
+                                                          float* y) {
+  // TODO(TJ): enable me
+  vec_relu<float, platform::jit::avx2>(n, x, y);
 }
 
+// TODO(TJ): optimize double of sigmoid, tanh and relu if necessary
+
 template <typename T, platform::jit::cpu_isa_t isa = platform::jit::isa_any>
 class VecActivations {
  public:
@@ -96,7 +360,7 @@ class VecActivations {
     } else if (type == "identity" || type == "") {
       return vec_identity<T, isa>;
     }
-    PADDLE_THROW("Not support type %s.", type);
+    LOG(FATAL) << "Not support type: " << type;
   }
 };
 

paddle/fluid/operators/math/cpu_vec_test.cc

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include <sys/time.h>
+#include <cmath>
+#include <cstring>
+#include <random>
+#include <vector>
+#include "gflags/gflags.h"
+#include "glog/logging.h"
+#include "gtest/gtest.h"
+
+#include "paddle/fluid/operators/math/cpu_vec.h"
+
+inline double GetCurrentUS() {
+  struct timeval time;
+  gettimeofday(&time, NULL);
+  return 1e+6 * time.tv_sec + time.tv_usec;
+}
+constexpr int repeat = 1000;
+
+template <typename T>
+inline T _sigmoid(T x) {
+  const T min = SIGMOID_THRESHOLD_MIN;
+  const T max = SIGMOID_THRESHOLD_MAX;
+  T tmp = (x < min) ? min : ((x > max) ? max : x);
+  return static_cast<T>(1) / (static_cast<T>(1) + std::exp(-tmp));
+}
+
+template <typename T>
+inline T _tanh(T x) {
+  return static_cast<T>(2) * _sigmoid<T>(static_cast<T>(2) * x) -
+         static_cast<T>(1);
+}
+
+template <typename T>
+void ref_sigmoid(const int n, const T* x, T* y) {
+  for (int i = 0; i < n; ++i) {
+    y[i] = _sigmoid(x[i]);
+  }
+}
+
+template <typename T>
+void ref_tanh(const int n, const T* x, T* y) {
+  for (int i = 0; i < n; ++i) {
+    y[i] = _tanh(x[i]);
+  }
+}
+template <typename T>
+void ref_relu(const int n, const T* x, T* y) {
+  for (int i = 0; i < n; ++i) {
+    y[i] = x[i] > 0 ? x[i] : 0;
+  }
+}
+
+template <typename T>
+void RandomVec(const int n, T* a) {
+  static unsigned int seed = 100;
+  std::mt19937 rng(seed++);
+  std::uniform_real_distribution<double> uniform_dist(0, 1);
+  const T lower = static_cast<T>(-20.f);
+  const T upper = static_cast<T>(20.f);
+  for (int i = 0; i < n; ++i) {
+    a[i] = static_cast<T>(uniform_dist(rng) * (upper - lower) + lower);
+  }
+}
+
+template <typename T>
+void TestAndBench(const int n, std::function<void(const int, const T*, T*)> tgt,
+                  std::function<void(const int, const T*, T*)> ref) {
+  std::vector<T> x(n);
+  std::vector<T> ytgt(n), yref(n);
+  RandomVec<T>(n, x.data());
+
+  const T* x_data = x.data();
+  T* ytgt_data = ytgt.data();
+  T* yref_data = yref.data();
+  auto st = GetCurrentUS();
+  for (int i = 0; i < repeat; ++i) {
+    tgt(n, x_data, ytgt_data);
+  }
+  auto mt = GetCurrentUS();
+  for (int i = 0; i < repeat; ++i) {
+    ref(n, x_data, yref_data);
+  }
+  auto et = GetCurrentUS();
+
+  VLOG(3) << "Vec size " << n << ": refer takes: " << (et - mt) / repeat
+          << " us, tgt takes: " << (mt - st) / repeat;
+  for (int i = 0; i < n; ++i) {
+    EXPECT_NEAR(ytgt_data[i], yref_data[i], 1e-3);
+  }
+}
+
+TEST(CpuVecTest, sigmoid) {
+  namespace jit = paddle::platform::jit;
+  using namespace paddle::operators::math;  // NOLINT
+  for (auto sz : {1, 2, 15, 16, 30, 32, 128, 200, 512}) {
+    TestAndBench<float>(sz, vec_sigmoid<float>, ref_sigmoid<float>);
+    TestAndBench<float>(sz, vec_sigmoid<float, jit::avx>, ref_sigmoid<float>);
+    TestAndBench<float>(sz, vec_sigmoid<float, jit::avx2>, ref_sigmoid<float>);
+    TestAndBench<float>(sz, vec_sigmoid<float, jit::avx512_common>,
+                        ref_sigmoid<float>);
+  }
+  TestAndBench<double>(30, vec_sigmoid<double>, ref_sigmoid<double>);
+}
+
+TEST(CpuVecTest, tanh) {
+  namespace jit = paddle::platform::jit;
+  using namespace paddle::operators::math;  // NOLINT
+  for (auto sz : {1, 2, 15, 16, 30, 32, 128, 200, 512}) {
+    TestAndBench<float>(sz, vec_tanh<float>, ref_tanh<float>);
+    TestAndBench<float>(sz, vec_tanh<float, jit::avx>, ref_tanh<float>);
+    TestAndBench<float>(sz, vec_tanh<float, jit::avx2>, ref_tanh<float>);
+    TestAndBench<float>(sz, vec_tanh<float, jit::avx512_common>,
+                        ref_tanh<float>);
+  }
+  TestAndBench<double>(30, vec_tanh<double>, ref_tanh<double>);
+}
+
+TEST(CpuVecTest, relu) {
+  namespace jit = paddle::platform::jit;
+  using namespace paddle::operators::math;  // NOLINT
+  for (auto sz : {1, 2, 15, 16, 30, 32, 128, 200, 512}) {
+    TestAndBench<float>(sz, vec_relu<float>, ref_relu<float>);
+    TestAndBench<float>(sz, vec_relu<float, jit::avx>, ref_relu<float>);
+    TestAndBench<float>(sz, vec_relu<float, jit::avx2>, ref_relu<float>);
+    TestAndBench<float>(sz, vec_relu<float, jit::avx512_common>,
+                        ref_relu<float>);
+  }
+  TestAndBench<double>(30, vec_relu<double>, ref_relu<double>);
+}
+
+template <typename T>
+void TestInplace(const int n, std::function<void(const int, const T*, T*)> tgt,
+                 std::function<void(const int, const T*, T*)> ref) {
+  std::vector<T> x(n);
+  std::vector<T> ytgt(n), yref(n);
+  RandomVec<T>(n, x.data());
+
+  const T* x_data = x.data();
+  T* yref_data = yref.data();
+  T* ytgt_data = ytgt.data();
+  std::memcpy(yref_data, x_data, sizeof(T) * n);
+  std::memcpy(ytgt_data, x_data, sizeof(T) * n);
+
+  ref(n, yref_data, yref_data);
+  tgt(n, ytgt_data, ytgt_data);
+
+  for (int i = 0; i < n; ++i) {
+    EXPECT_NEAR(ytgt_data[i], yref_data[i], 1e-3);
+  }
+}
+
+TEST(CpuVecTest, inplace_sigmoid) {
+  namespace jit = paddle::platform::jit;
+  using namespace paddle::operators::math;  // NOLINT
+  for (auto sz : {1, 2, 15, 16, 30, 32, 128, 200, 512}) {
+    TestInplace<float>(sz, vec_sigmoid<float>, ref_sigmoid<float>);
+    TestInplace<float>(sz, vec_sigmoid<float, jit::avx>, ref_sigmoid<float>);
+    TestInplace<float>(sz, vec_sigmoid<float, jit::avx2>, ref_sigmoid<float>);
+    TestInplace<float>(sz, vec_sigmoid<float, jit::avx512_common>,
+                       ref_sigmoid<float>);
+  }
+  TestInplace<double>(30, vec_sigmoid<double>, ref_sigmoid<double>);
+}
+
+TEST(CpuVecTest, inplace_tanh) {
+  namespace jit = paddle::platform::jit;
+  using namespace paddle::operators::math;  // NOLINT
+  for (auto sz : {1, 2, 15, 16, 30, 32, 128, 200, 512}) {
+    TestInplace<float>(sz, vec_tanh<float>, ref_tanh<float>);
+    TestInplace<float>(sz, vec_tanh<float, jit::avx>, ref_tanh<float>);
+    TestInplace<float>(sz, vec_tanh<float, jit::avx2>, ref_tanh<float>);
+    TestInplace<float>(sz, vec_tanh<float, jit::avx512_common>,
+                       ref_tanh<float>);
+  }
+  TestInplace<double>(30, vec_tanh<double>, ref_tanh<double>);
+}
+
+TEST(CpuVecTest, inplace_relu) {
+  namespace jit = paddle::platform::jit;
+  using namespace paddle::operators::math;  // NOLINT
+  for (auto sz : {1, 2, 15, 16, 30, 32, 128, 200, 512}) {
+    TestInplace<float>(sz, vec_relu<float>, ref_relu<float>);
+    TestInplace<float>(sz, vec_relu<float, jit::avx>, ref_relu<float>);
+    TestInplace<float>(sz, vec_relu<float, jit::avx2>, ref_relu<float>);
+    TestInplace<float>(sz, vec_relu<float, jit::avx512_common>,
+                       ref_relu<float>);
+  }
+  TestInplace<double>(30, vec_relu<double>, ref_relu<double>);
+}

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>

paddle/fluid/operators/stack_op.h

@@ -150,12 +150,7 @@ 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();
 #else
@@ -168,14 +163,6 @@ class StackKernel : public framework::OpKernel<T> {
     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);
-    }
-#endif
-  }
 };
 
 class StackOpGrad : public framework::OperatorWithKernel {
@@ -244,34 +231,19 @@ 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();
 #else
     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);
-    }
-#endif
-  }
 };
 
 }  // namespace operators

paddle/fluid/platform/CMakeLists.txt

+if (NOT WIN32)
 proto_library(profiler_proto SRCS profiler.proto DEPS framework_proto)
 py_proto_compile(profiler_py_proto SRCS profiler.proto)
 
@@ -10,6 +11,7 @@ add_custom_command(TARGET profiler_py_proto POST_BUILD
         COMMAND cp *.py ${PADDLE_BINARY_DIR}/python/paddle/fluid/proto/profiler
         COMMENT "Copy generated python proto into directory paddle/fluid/proto/profiler."
         WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
+endif(NOT WIN32)
 
 if(WITH_GPU)
   nv_library(enforce SRCS enforce.cc)
@@ -50,7 +52,7 @@ ENDIF()
 # memcpy depends on device_context, here add deps individually for
 # avoiding cycle dependencies
 cc_library(device_context SRCS device_context.cc init.cc DEPS malloc
-    place eigen3 stringpiece cpu_helper framework_proto ${GPU_CTX_DEPS} ${MKLDNN_CTX_DEPS})
+    place eigen3 stringpiece cpu_helper cpu_info framework_proto ${GPU_CTX_DEPS} ${MKLDNN_CTX_DEPS})
 nv_test(device_context_test SRCS device_context_test.cu DEPS device_context gpu_info)
 
 cc_test(init_test SRCS init_test.cc DEPS device_context)
@@ -58,9 +60,12 @@ cc_test(init_test SRCS init_test.cc DEPS device_context)
 nv_test(cudnn_helper_test SRCS cudnn_helper_test.cc DEPS dynload_cuda)
 nv_test(transform_test SRCS transform_test.cu DEPS memory place device_context)
 
+
+if (NOT WIN32)
 cc_library(device_tracer SRCS device_tracer.cc DEPS boost profiler_proto framework_proto ${GPU_CTX_DEPS})
 cc_library(profiler SRCS profiler.cc DEPS device_context device_tracer)
 cc_test(profiler_test SRCS profiler_test.cc DEPS profiler)
+endif(NOT WIN32)
 
 nv_test(float16_gpu_test SRCS float16_test.cu DEPS lod_tensor)
 cc_test(float16_test SRCS float16_test.cc DEPS lod_tensor)

paddle/fluid/platform/cpu_info.cc

@@ -22,9 +22,13 @@ 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
+#include <windows.h>
 #else
 #include <unistd.h>
-#endif
+#endif  // _WIN32
 
 #include <algorithm>
 #include "gflags/gflags.h"
@@ -32,16 +36,20 @@ limitations under the License. */
 DEFINE_double(fraction_of_cpu_memory_to_use, 1,
               "Default use 100% of CPU memory for PaddlePaddle,"
               "reserve the rest for page tables, etc");
-
+#if !defined(_WIN32)
 DEFINE_uint64(initial_cpu_memory_in_mb,
 #ifdef PADDLE_WITH_MKLDNN
               /* Aligned with mozga-intel, MKLDNN need at least 5000 MB
                * to obtain the best performance*/
-              5000,
+              5000ul,
 #else
-              500,
+              500ul,
 #endif
               "Initial CPU memory for PaddlePaddle, in MD unit.");
+#else
+DEFINE_uint64(initial_cpu_memory_in_mb, 500ul,
+              "Initial CPU memory for PaddlePaddle, in MD unit.");
+#endif  // !defined(_WIN32)
 
 DEFINE_double(
     fraction_of_cuda_pinned_memory_to_use, 0.5,
@@ -60,6 +68,11 @@ inline size_t CpuTotalPhysicalMemory() {
   size_t len = sizeof(size);
   if (sysctl(mib, 2, &size, &len, NULL, 0) == 0) return (size_t)size;
   return 0L;
+#elif defined(_WIN32)
+  MEMORYSTATUSEX sMeminfo;
+  sMeminfo.dwLength = sizeof(sMeminfo);
+  GlobalMemoryStatusEx(&sMeminfo);
+  return sMeminfo.ullTotalPhys;
 #else
   int64_t pages = sysconf(_SC_PHYS_PAGES);
   int64_t page_size = sysconf(_SC_PAGE_SIZE);

paddle/fluid/platform/cpu_info.h

@@ -51,7 +51,7 @@ typedef enum {
 } cpu_isa_t;  // Instruction set architecture
 
 // May I use some instruction
-inline bool MayIUse(const cpu_isa_t cpu_isa);
+bool MayIUse(const cpu_isa_t cpu_isa);
 
 }  // namespace jit
 

paddle/fluid/platform/device_tracer.h

@@ -13,7 +13,12 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 #pragma once
 
+#if !defined(_WIN32)
 #include <sys/time.h>
+#else
+#include <windows.h>
+#endif  // !_WIN32
+
 #include <time.h>
 #include <chrono>  // NOLINT
 #include <string>
@@ -27,12 +32,15 @@ namespace platform {
 ///////////////////////
 // WARN: Under Development. Don't depend on it yet.
 //////////////////////
-
+#if !defined(_WIN32)
 inline uint64_t PosixInNsec() {
   struct timeval tv;
   gettimeofday(&tv, nullptr);
   return 1000 * (static_cast<uint64_t>(tv.tv_sec) * 1000000 + tv.tv_usec);
 }
+#else
+inline uint64_t PosixInNsec() { return static_cast<uint64_t>(0); }
+#endif  // !_WIN32
 
 // DeviceTracer performs the following tasks:
 // 1. Register cuda callbacks for various events: kernel, memcpy, etc.

paddle/fluid/platform/dynload/CMakeLists.txt

@@ -16,7 +16,9 @@ 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()

paddle/fluid/platform/dynload/dynamic_loader.cc

@@ -13,8 +13,6 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 #include "paddle/fluid/platform/dynload/dynamic_loader.h"
 
-#include <dlfcn.h>
-
 #include <memory>
 #include <mutex>  // NOLINT
 #include <string>
@@ -23,6 +21,7 @@ limitations under the License. */
 #include "glog/logging.h"
 #include "paddle/fluid/platform/dynload/cupti_lib_path.h"
 #include "paddle/fluid/platform/enforce.h"
+#include "paddle/fluid/platform/port.h"
 
 DEFINE_string(cudnn_dir, "",
               "Specify path for loading libcudnn.so. For instance, "

paddle/fluid/platform/enforce.h

@@ -18,6 +18,11 @@ limitations under the License. */
 #include <cxxabi.h>  // for __cxa_demangle
 #endif               // __GNUC__
 
+#if defined(_WIN32)
+#define NOMINMAX  // msvc max/min macro conflict with std::min/max
+#define GLOG_NO_ABBREVIATED_SEVERITIES  // msvc conflict logging with windows.h
+#endif
+
 #ifdef PADDLE_WITH_CUDA
 #include <cublas_v2.h>
 #include <cudnn.h>
@@ -117,7 +122,12 @@ struct EOFException : public std::exception {
 // always forces branch prediction of true.
 // This generates faster binary code. __builtin_expect is since C++11.
 // For more details, please check https://stackoverflow.com/a/43870188/724872.
+#if !defined(_WIN32)
 #define UNLIKELY(condition) __builtin_expect(static_cast<bool>(condition), 0)
+#else
+// there is no equivalent intrinsics in msvc.
+#define UNLIKELY(condition) (condition == 0)
+#endif
 
 template <typename... Args>
 inline typename std::enable_if<sizeof...(Args) != 0, void>::type throw_on_error(
@@ -230,6 +240,7 @@ inline void throw_on_error(T e) {
   throw_on_error(e, "");
 }
 
+#if !defined(_WIN32)
 #define PADDLE_THROW(...)                                              \
   do {                                                                 \
     throw ::paddle::platform::EnforceNotMet(                           \
@@ -248,15 +259,28 @@ inline void throw_on_error(T e) {
                                               __FILE__, __LINE__);      \
     }                                                                   \
   } while (false)
-#else
-#define PADDLE_ENFORCE(...) ::paddle::platform::throw_on_error(__VA_ARGS__);
-#endif
 
 #define PADDLE_THROW_EOF()                                                     \
   do {                                                                         \
     throw ::paddle::platform::EOFException("There is no next data.", __FILE__, \
                                            __LINE__);                          \
   } while (false)
+
+#else
+#define PADDLE_ENFORCE(...) ::paddle::platform::throw_on_error(__VA_ARGS__)
+#endif  // REPLACE_ENFORCE_GLOG
+
+#else  // !_WIN32
+// disable enforce, caused by the varardic macro exception error
+#define PADDLE_THROW(x)                                      \
+  do {                                                       \
+    throw std::make_exception_ptr(                           \
+        std::runtime_error("Windows disable the enforce.")); \
+  } while (false)
+
+#define PADDLE_ENFORCE(x, ...) x
+#endif  // !_WIN32
+
 /*
  * Some enforce helpers here, usage:
  *    int a = 1;

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 {

paddle/fluid/platform/init.cc

@@ -18,6 +18,7 @@ limitations under the License. */
 
 #include "paddle/fluid/framework/operator.h"
 #include "paddle/fluid/platform/cpu_helper.h"
+#include "paddle/fluid/platform/cpu_info.h"
 #include "paddle/fluid/platform/device_context.h"
 #include "paddle/fluid/platform/init.h"
 #include "paddle/fluid/platform/place.h"
@@ -120,6 +121,22 @@ void InitDevices(bool init_p2p, const std::vector<int> devices) {
 #ifndef PADDLE_WITH_MKLDNN
   platform::SetNumThreads(FLAGS_paddle_num_threads);
 #endif
+
+  if (platform::jit::MayIUse(platform::jit::avx512_common)) {
+#ifndef __AVX512F__
+    LOG(WARNING) << "AVX512F is available, Please re-compile on local machine";
+#endif
+  }
+  if (platform::jit::MayIUse(platform::jit::avx2)) {
+#ifndef __AVX2__
+    LOG(WARNING) << "AVX2 is available, Please re-compile on local machine";
+#endif
+  }
+  if (platform::jit::MayIUse(platform::jit::avx)) {
+#ifndef __AVX__
+    LOG(WARNING) << "AVX is available, Please re-compile on local machine";
+#endif
+  }
 }
 
 void InitGLOG(const std::string &prog_name) {

paddle/fluid/platform/profiler.h

@@ -69,6 +69,7 @@ void PushEvent(const std::string& name, const DeviceContext* dev_ctx);
 
 void PopEvent(const std::string& name, const DeviceContext* dev_ctx);
 
+#if !defined(_WIN32)
 struct RecordEvent {
   RecordEvent(const std::string& name, const DeviceContext* dev_ctx);
 
@@ -94,6 +95,15 @@ struct RecordBlock {
   std::string name_;
   uint64_t start_ns_;
 };
+#else
+// windows do not support profiler temporarily.
+struct RecordEvent {
+  RecordEvent(const std::string& name, const DeviceContext* dev_ctx) {}
+};
+struct RecordBlock {
+  explicit RecordBlock(int block_id) {}
+};
+#endif
 
 // Return the event list of all threads. Assumed the returned value calls
 // event_lists, event_lists[i][j] represents the j-th Event of i-th thread.

paddle/fluid/pybind/CMakeLists.txt

-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)

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(

python/paddle/fluid/tests/unittests/test_fusion_lstm_op.py

@@ -43,13 +43,13 @@ def fusion_lstm(
         act_cell, act_cand)
 
 
-class TestLstmOp(OpTest):
-    def set_argument(self):
-        self.lod = [[2, 3, 2]]
+class TestFusionLSTMOp(OpTest):
+    def set_conf(self):
+        pass
 
     def setUp(self):
         self.op_type = 'fusion_lstm'
-        self.lod = [[2, 3, 2]]
+        self.lod = [[2, 3, 5, 4]]
         self.M = 8
         self.D = 16
         self.has_initial_state = False
@@ -58,33 +58,33 @@ class TestLstmOp(OpTest):
         self.act_cell = 'tanh'
         self.act_cand = 'tanh'
         self.use_peepholes = False
-        self.set_argument()
+        self.set_conf()
 
         T = sum(self.lod[0])
         bs = len(self.lod[0])
 
-        x = np.random.normal(size=(T, self.M)).astype('float64')
+        x = np.random.normal(size=(T, self.M)).astype('float32')
         if self.has_initial_state:
-            h0 = np.random.normal(size=(bs, self.D)).astype('float64')
-            c0 = np.random.normal(size=(bs, self.D)).astype('float64')
+            h0 = np.random.normal(size=(bs, self.D)).astype('float32')
+            c0 = np.random.normal(size=(bs, self.D)).astype('float32')
         else:
-            h0 = np.zeros((bs, self.D)).astype('float64')
-            c0 = np.zeros((bs, self.D)).astype('float64')
+            h0 = np.zeros((bs, self.D)).astype('float32')
+            c0 = np.zeros((bs, self.D)).astype('float32')
 
-        wh = np.random.normal(size=(self.D, 4 * self.D)).astype('float64')
+        wh = np.random.normal(size=(self.D, 4 * self.D)).astype('float32')
 
         if self.use_peepholes:
-            b = np.random.normal(size=(1, 7 * self.D)).astype('float64')
+            b = np.random.normal(size=(1, 7 * self.D)).astype('float32')
         else:
-            b = np.random.normal(size=(1, 4 * self.D)).astype('float64')
+            b = np.random.normal(size=(1, 4 * self.D)).astype('float32')
         w_b = np.copy(b[:, 0:4 * self.D])
         w_c = b[:, 4 * self.D:] if self.use_peepholes else None
 
         # this is the weight of fc
-        wx = np.random.normal(size=(self.M, 4 * self.D)).astype('float64')
+        wx = np.random.normal(size=(self.M, 4 * self.D)).astype('float32')
         # this is the bias of fc
         # and it should be manually added into the bias of this fusion LSTM
-        bx = np.random.normal(size=(1, 4 * self.D)).astype('float64')
+        bx = np.random.normal(size=(1, 4 * self.D)).astype('float32')
         b[0, 0:4 * self.D] += bx[0, :]
         h, c = fusion_lstm(x, self.lod, wx, bx, h0, c0, wh, w_b, w_c,
                            self.is_reverse, ACTIVATION[self.act_gate],
@@ -114,35 +114,45 @@ class TestLstmOp(OpTest):
         }
 
     def test_check_output(self):
-        self.check_output(atol=1e-8)
+        self.check_output()
 
 
-class TestLstmOpInitReverse(TestLstmOp):
-    def set_argument(self):
+class TestFusionLSTMOpInit(TestFusionLSTMOp):
+    def set_conf(self):
+        self.has_initial_state = True
+
+
+class TestFusionLSTMOpReverse(TestFusionLSTMOp):
+    def set_conf(self):
+        self.is_reverse = True
+
+
+class TestFusionLSTMOpInitReverse(TestFusionLSTMOp):
+    def set_conf(self):
         self.has_initial_state = True
         self.is_reverse = True
 
 
-class TestLstmOpMD1(TestLstmOp):
-    def set_argument(self):
+class TestFusionLSTMOpMD1(TestFusionLSTMOp):
+    def set_conf(self):
         self.M = 36
         self.D = 8
 
 
-class TestLstmOpMD2(TestLstmOp):
-    def set_argument(self):
+class TestFusionLSTMOpMD2(TestFusionLSTMOp):
+    def set_conf(self):
         self.M = 8
         self.D = 8
 
 
-class TestLstmOpMD3(TestLstmOp):
-    def set_argument(self):
+class TestFusionLSTMOpMD3(TestFusionLSTMOp):
+    def set_conf(self):
         self.M = 15
         self.D = 3
 
 
-class TestLstmOpBS1(TestLstmOp):
-    def set_argument(self):
+class TestFusionLSTMOpBS1(TestFusionLSTMOp):
+    def set_conf(self):
         self.lod = [[3]]
         self.D = 16