Merge pull request #15069 from wopeizl/windows/dsosupport

add cuda dso support for windows

cmake/cuda.cmake

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

cmake/cudnn.cmake

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

cmake/external/cub.cmake

@@ -32,4 +32,4 @@ endif()
 
 add_dependencies(cub extern_cub)
 
-LIST(APPEND externl_project_dependencies cub)
+LIST(APPEND external_project_dependencies cub)

cmake/external/dlpack.cmake

@@ -28,4 +28,4 @@ endif()
 
 add_dependencies(dlpack extern_dlpack)
 
-LIST(APPEND externl_project_dependencies dlpack)
+LIST(APPEND external_project_dependencies dlpack)

paddle/fluid/framework/details/all_reduce_op_handle.cc

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

paddle/fluid/platform/dynload/cudnn.cc

@@ -38,6 +38,10 @@ CUDNN_DNN_ROUTINE_EACH_AFTER_R4(DEFINE_WRAP);
 CUDNN_DNN_ROUTINE_EACH_R5(DEFINE_WRAP);
 #endif
 
+#ifdef CUDNN_DNN_ROUTINE_EACH_R6
+CUDNN_DNN_ROUTINE_EACH_R6(DEFINE_WRAP);
+#endif
+
 #ifdef CUDNN_DNN_ROUTINE_EACH_R7
 CUDNN_DNN_ROUTINE_EACH_R7(DEFINE_WRAP);
 #endif

paddle/fluid/platform/dynload/dynamic_loader.cc

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