Merge pull request #14943 from wopeizl/windows/ctc

add ctc support for windows

CMakeLists.txt

@@ -208,10 +208,10 @@ include(external/xxhash)    # download xxhash
 include(external/dlpack)
 include(external/snappy)    # download snappy
 include(external/snappystream) # download snappystream
+include(external/warpctc)   # download, build, install warpctc
 
 if (NOT WIN32)
-# there is no official support of warpctc, nccl, cupti in windows
-include(external/warpctc)   # download, build, install warpctc
+# there is no official support of nccl, cupti in windows
 include(cupti)
 include(external/gzstream)
 endif (NOT WIN32)

cmake/external/warpctc.cmake

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

cmake/operators.cmake

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

paddle/fluid/operators/CMakeLists.txt

@@ -64,9 +64,7 @@ endif()
 set(COMMON_OP_DEPS ${OP_HEADER_DEPS})
 
 set(COMMON_OP_DEPS ${COMMON_OP_DEPS} selected_rows_functor selected_rows lod_tensor maxouting unpooling pooling lod_rank_table context_project sequence_pooling executor)
-if (NOT WIN32)
-    set(COMMON_OP_DEPS ${COMMON_OP_DEPS} dynload_warpctc)
-endif()
+set(COMMON_OP_DEPS ${COMMON_OP_DEPS} dynload_warpctc)
 set(COMMON_OP_DEPS ${COMMON_OP_DEPS} sequence_padding sequence_scale cos_sim_functor memory jit_kernel concat_and_split cross_entropy softmax vol2col im2col sampler)
 set(COMMON_OP_DEPS ${COMMON_OP_DEPS} sequence2batch lstm_compute matrix_bit_code gru_compute activation_functions)
 if (WITH_GPU)

paddle/fluid/platform/dynload/CMakeLists.txt

@@ -16,9 +16,7 @@ if (CUPTI_FOUND)
     list(APPEND CUDA_SRCS cupti.cc)
 endif(CUPTI_FOUND)
 nv_library(dynload_cuda SRCS ${CUDA_SRCS} DEPS dynamic_loader)
-if (NOT WIN32)
 cc_library(dynload_warpctc SRCS warpctc.cc DEPS dynamic_loader warpctc)
-endif(NOT WIN32)
 if (WITH_MKLML)
     cc_library(dynload_mklml SRCS mklml.cc DEPS dynamic_loader mklml)
 endif()

paddle/fluid/platform/dynload/cudnn.h

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

paddle/fluid/platform/dynload/dynamic_loader.cc

@@ -201,6 +201,8 @@ void* GetCurandDsoHandle() {
 void* GetWarpCTCDsoHandle() {
 #if defined(__APPLE__) || defined(__OSX__)
   return GetDsoHandleFromSearchPath(FLAGS_warpctc_dir, "libwarpctc.dylib");
+#elif defined(_WIN32)
+  return GetDsoHandleFromSearchPath(FLAGS_warpctc_dir, "warpctc.dll");
 #else
   return GetDsoHandleFromSearchPath(FLAGS_warpctc_dir, "libwarpctc.so");
 #endif

paddle/fluid/platform/dynload/dynamic_loader.h

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

paddle/fluid/platform/dynload/mklml.h

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

paddle/fluid/platform/dynload/tensorrt.h

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

paddle/fluid/platform/dynload/warpctc.h

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

paddle/fluid/platform/port.h

@@ -55,7 +55,6 @@ static void *dlsym(void *handle, const char *symbol_name) {
 
 static void *dlopen(const char *filename, int flag) {
   std::string file_name(filename);
-  file_name.replace(0, file_name.size() - 1, '/', '\\');
   HMODULE hModule = LoadLibrary(file_name.c_str());
   if (!hModule) {
     throw std::runtime_error(file_name + " not found.");

python/paddle/fluid/__init__.py

@@ -102,6 +102,13 @@ def __bootstrap__():
     import sys
     import os
     import platform
+
+    if os.name == 'nt':
+        third_lib_path = os.path.abspath(os.path.dirname(
+            __file__)) + os.sep + '..' + os.sep + 'libs'
+        os.environ['path'] += ';' + third_lib_path
+        sys.path.append(third_lib_path)
+
     from . import core
 
     in_test = 'unittest' in sys.modules
@@ -128,13 +135,12 @@ def __bootstrap__():
         'free_idle_memory', 'paddle_num_threads', "dist_threadpool_size",
         'eager_delete_tensor_gb', 'fast_eager_deletion_mode',
         'allocator_strategy', 'reader_queue_speed_test_mode',
-        'print_sub_graph_dir', 'pe_profile_fname'
+        'print_sub_graph_dir', 'pe_profile_fname', 'warpctc_dir'
     ]
     if 'Darwin' not in sysstr:
         read_env_flags.append('use_pinned_memory')
 
     if os.name != 'nt':
-        read_env_flags.append('warpctc_dir')
         read_env_flags.append('cpu_deterministic')
 
     if core.is_compiled_with_dist():

python/paddle/fluid/framework.py

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

python/setup.py.in

@@ -160,10 +160,11 @@ if '${WITH_FLUID_ONLY}'== 'OFF':
 
 # put all thirdparty libraries in paddle.libs
 libs_path='${PADDLE_BINARY_DIR}/python/paddle/libs'
-if os.name != 'nt':
-    package_data['paddle.libs']= []
-    package_data['paddle.libs']=['libwarpctc' + ext_name]
-    shutil.copy('${WARPCTC_LIBRARIES}', libs_path)
+
+package_data['paddle.libs']= []
+package_data['paddle.libs']=[('libwarpctc' if os.name != 'nt' else 'warpctc') + ext_name]
+shutil.copy('${WARPCTC_LIBRARIES}', libs_path)
+
 if '${WITH_MKL}' == 'ON':
     shutil.copy('${MKLML_LIB}', libs_path)
     shutil.copy('${MKLML_IOMP_LIB}', libs_path)