fix dynamic_loader more safe and error message on windows (#28144)

CMakeLists.txt

@@ -193,10 +193,19 @@ if(WITH_BRPC_RDMA)
     endif()
 endif()
 
-# lite subgraph compilation depends on CUDNN_ROOT,
-# so include(cudnn) needs to be in front of include(third_party/lite)
-include(cudnn)              # set cudnn libraries, must before configure
-include(third_party)        # download, build, install third_party
+if(WITH_GPU)
+    include(cuda)
+    # lite subgraph compilation depends on CUDNN_ROOT,
+    # so include(cudnn) needs to be in front of include(third_party/lite)
+    include(cudnn)              # set cudnn libraries, must before configure
+    include(tensorrt)
+    # there is no official support of nccl, cupti in windows
+    if(NOT WIN32)
+        include(cupti)
+    endif()
+endif()
+
+include(third_party)  # download, build, install third_party, Contains about 20+ dependencies
 
 if(WITH_DISTRIBUTE)
     if(WITH_GRPC)
@@ -209,18 +218,8 @@ if(WITH_DISTRIBUTE)
     endif()
 endif()
 
-# there is no official support of nccl, cupti in windows
-if(NOT WIN32)
-    include(cupti)
-endif()
-
 include(flags)              # set paddle compile flags
 
-if(WITH_GPU)
-    include(cuda)
-    include(tensorrt)
-endif()
-
 if(WITH_PROFILER)
     find_package(Gperftools REQUIRED)
     include_directories(${GPERFTOOLS_INCLUDE_DIR})

cmake/cuda.cmake

@@ -198,7 +198,9 @@ elseif (${CMAKE_CUDA_COMPILER_VERSION} LESS 12.0) # CUDA 11.x
   set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -D__STRICT_ANSI__")
 endif()
 
-add_definitions("-DPADDLE_CUDA_BINVER=\"${CUDA_VERSION_MAJOR}${CUDA_VERSION_MINOR}\"")
+add_definitions("-DCUDA_VERSION_MAJOR=\"${CUDA_VERSION_MAJOR}\"")
+add_definitions("-DCUDA_VERSION_MINOR=\"${CUDA_VERSION_MINOR}\"")
+add_definitions("-DCUDA_TOOLKIT_ROOT_DIR=\"${CUDA_TOOLKIT_ROOT_DIR}\"")
 
 # setting nvcc arch flags
 select_nvcc_arch_flags(NVCC_FLAGS_EXTRA)
@@ -249,3 +251,4 @@ endif()
 
 mark_as_advanced(CUDA_BUILD_CUBIN CUDA_BUILD_EMULATION CUDA_VERBOSE_BUILD)
 mark_as_advanced(CUDA_SDK_ROOT_DIR CUDA_SEPARABLE_COMPILATION)
+

cmake/cudnn.cmake

@@ -35,17 +35,18 @@ list(APPEND CUDNN_CHECK_LIBRARY_DIRS
 	${CUDA_TOOLKIT_ROOT_DIR}/lib/x64
 	)
 set(CUDNN_LIB_NAME "")
+
 if (LINUX)
-set(CUDNN_LIB_NAME "libcudnn.so")
+    set(CUDNN_LIB_NAME "libcudnn.so")
 endif(LINUX)
 
 if(WIN32)
-# only support cudnn7
-set(CUDNN_LIB_NAME "cudnn.lib" "cudnn64_7.dll")
+    # only support cudnn7
+    set(CUDNN_LIB_NAME "cudnn.lib" "cudnn64_7.dll")
 endif(WIN32)
 
 if(APPLE)
-set(CUDNN_LIB_NAME "libcudnn.dylib" "libcudnn.so")
+    set(CUDNN_LIB_NAME "libcudnn.dylib" "libcudnn.so")
 endif(APPLE)
 
 find_library(CUDNN_LIBRARY NAMES ${CUDNN_LIB_NAME} # libcudnn_static.a
@@ -88,7 +89,7 @@ macro(find_cudnn_version cudnn_header_file)
         if(NOT CUDNN_MAJOR_VERSION)
             set(CUDNN_VERSION "???")
         else()
-            add_definitions("-DPADDLE_CUDNN_BINVER=\"${CUDNN_MAJOR_VERSION}\"")
+            add_definitions("-DCUDNN_MAJOR_VERSION=\"${CUDNN_MAJOR_VERSION}\"")
             math(EXPR CUDNN_VERSION
                 "${CUDNN_MAJOR_VERSION} * 1000 +
                  ${CUDNN_MINOR_VERSION} * 100 + ${CUDNN_PATCHLEVEL_VERSION}")

paddle/fluid/platform/dynload/dynamic_loader.cc

@@ -57,17 +57,26 @@ struct PathNode {
 
 static constexpr char cupti_lib_path[] = CUPTI_LIB_PATH;
 
-// NOTE: In order to adapt to the default installation path of cuda on linux
-static constexpr char linux_cudnn_lib_path[] = "/usr/local/cuda/lib64";
+// NOTE: In order to adapt to the default installation path of cuda
+#if defined(_WIN32) && defined(PADDLE_WITH_CUDA)
+static constexpr char cuda_lib_path[] = CUDA_TOOLKIT_ROOT_DIR "/bin";
+#else
+static constexpr char cuda_lib_path[] = "/usr/local/cuda/lib64";
+#endif
 
 static PathNode s_py_site_pkg_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";
+static constexpr char* win_cublas_lib =
+    "cublas64_" CUDA_VERSION_MAJOR CUDA_VERSION_MINOR
+    ".dll;cublas64_" CUDA_VERSION_MAJOR ".dll";
+static constexpr char* win_curand_lib =
+    "curand64_" CUDA_VERSION_MAJOR CUDA_VERSION_MINOR
+    ".dll;curand64_" CUDA_VERSION_MAJOR ".dll";
+static constexpr char* win_cudnn_lib = "cudnn64_" CUDNN_MAJOR_VERSION ".dll";
 static constexpr char* win_cusolver_lib =
-    "cusolver64_" PADDLE_CUDA_BINVER ".dll";
+    "cusolver64_" CUDA_VERSION_MAJOR CUDA_VERSION_MINOR
+    ".dll;cusolver64_" CUDA_VERSION_MAJOR ".dll";
 #endif
 
 static inline std::string join(const std::string& part1,
@@ -87,6 +96,24 @@ static inline std::string join(const std::string& part1,
   return ret;
 }
 
+static inline std::vector<std::string> split(
+    const std::string& str, const std::string separator = " ") {
+  std::vector<std::string> str_list;
+  std::string::size_type firstPos;
+  firstPos = str.find_first_not_of(separator, 0);
+  std::string::size_type lastPos;
+  lastPos = str.find_first_of(separator, firstPos);
+  while (std::string::npos != firstPos && std::string::npos != lastPos) {
+    str_list.push_back(str.substr(firstPos, lastPos - firstPos));
+    firstPos = str.find_first_not_of(separator, lastPos);
+    lastPos = str.find_first_of(separator, firstPos);
+  }
+  if (std::string::npos == lastPos) {
+    str_list.push_back(str.substr(firstPos, lastPos - firstPos));
+  }
+  return str_list;
+}
+
 void SetPaddleLibPath(const std::string& py_site_pkg_path) {
   s_py_site_pkg_path.path = py_site_pkg_path;
   VLOG(3) << "Set paddle lib path : " << py_site_pkg_path;
@@ -147,26 +174,31 @@ static inline void* GetDsoHandleFromSearchPath(
 #else
   int dynload_flags = 0;
 #endif  // !_WIN32
+  std::vector<std::string> dso_names = split(dso_name, ";");
+  void* dso_handle = nullptr;
+  for (auto dso : dso_names) {
     // 1. search in user config path by FLAGS
-  void* dso_handle =
-      GetDsoHandleFromSpecificPath(config_path, dso_name, dynload_flags);
-  // 2. search in system default path
+    dso_handle = GetDsoHandleFromSpecificPath(config_path, dso, dynload_flags);
+    // 2. search in extra paths
     if (nullptr == dso_handle) {
-    dso_handle = GetDsoHandleFromDefaultPath(dso_name, dynload_flags);
+      for (auto path : extra_paths) {
+        VLOG(3) << "extra_paths: " << path;
+        dso_handle = GetDsoHandleFromSpecificPath(path, dso, dynload_flags);
       }
-  // 3. search in extra paths
+    }
+    // 3. search in system default path
     if (nullptr == dso_handle) {
-    for (auto path : extra_paths) {
-      dso_handle = GetDsoHandleFromSpecificPath(path, dso_name, dynload_flags);
+      dso_handle = GetDsoHandleFromDefaultPath(dso, dynload_flags);
     }
+    if (nullptr != dso_handle) break;
   }
 
-  // 4. [If Failed] logging warning if exists
+  // 4. [If Failed for All dso_names] logging warning if exists
   if (nullptr == dso_handle && !warning_msg.empty()) {
     LOG(WARNING) << warning_msg;
   }
 
-  // 5. [If Failed] logging or throw error info
+  // 5. [If Failed for All dso_names] logging or throw error info
   if (nullptr == dso_handle) {
     auto error_msg =
         "The third-party dynamic library (%s) that Paddle depends on is not "
@@ -203,7 +235,8 @@ 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);
+  return GetDsoHandleFromSearchPath(FLAGS_cuda_dir, win_cublas_lib, true,
+                                    {cuda_lib_path});
 #else
   return GetDsoHandleFromSearchPath(FLAGS_cuda_dir, "libcublas.so");
 #endif
@@ -220,10 +253,19 @@ void* GetCUDNNDsoHandle() {
   return GetDsoHandleFromSearchPath(FLAGS_cudnn_dir, "libcudnn.dylib", false,
                                     {}, mac_warn_meg);
 #elif defined(_WIN32) && defined(PADDLE_WITH_CUDA)
-  return GetDsoHandleFromSearchPath(FLAGS_cudnn_dir, win_cudnn_lib);
+  std::string win_warn_meg(
+      "Note: [Recommend] copy cudnn into CUDA installation directory. \n "
+      "For instance, download cudnn-10.0-windows10-x64-v7.6.5.32.zip from "
+      "NVIDIA's official website, \n"
+      "then, unzip it and copy it into C:\\Program Files\\NVIDIA GPU Computing "
+      "Toolkit\\CUDA/v10.0\n"
+      "You should do this according to your CUDA installation directory and "
+      "CUDNN version.");
+  return GetDsoHandleFromSearchPath(FLAGS_cudnn_dir, win_cudnn_lib, true,
+                                    {cuda_lib_path}, win_warn_meg);
 #else
   return GetDsoHandleFromSearchPath(FLAGS_cudnn_dir, "libcudnn.so", false,
-                                    {linux_cudnn_lib_path});
+                                    {cuda_lib_path});
 #endif
 }
 
@@ -241,7 +283,8 @@ 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);
+  return GetDsoHandleFromSearchPath(FLAGS_cuda_dir, win_curand_lib, true,
+                                    {cuda_lib_path});
 #else
   return GetDsoHandleFromSearchPath(FLAGS_cuda_dir, "libcurand.so");
 #endif
@@ -251,7 +294,8 @@ void* GetCusolverDsoHandle() {
 #if defined(__APPLE__) || defined(__OSX__)
   return GetDsoHandleFromSearchPath(FLAGS_cuda_dir, "libcusolver.dylib");
 #elif defined(_WIN32) && defined(PADDLE_WITH_CUDA)
-  return GetDsoHandleFromSearchPath(FLAGS_cuda_dir, win_cusolver_lib);
+  return GetDsoHandleFromSearchPath(FLAGS_cuda_dir, win_cusolver_lib, true,
+                                    {cuda_lib_path});
 #else
   return GetDsoHandleFromSearchPath(FLAGS_cuda_dir, "libcusolver.so");
 #endif

paddle/fluid/platform/port.h

@@ -56,7 +56,11 @@ static void *dlopen(const char *filename, int flag) {
   std::string file_name(filename);
   HMODULE hModule = LoadLibrary(file_name.c_str());
   if (!hModule) {
+    if (flag) {
       throw std::runtime_error(file_name + " not found.");
+    } else {
+      return nullptr;
+    }
   }
   return reinterpret_cast<void *>(hModule);
 }