add dynamic_load

paddle/platform/cublas.h

+#include <cublas_v2.h>
+#include "paddle/platform/dynamic_loader.h"
+
+namespace paddle {
+namespace dyload {
+namespace dynload {
+
+std::once_flag cublas_dso_flag;
+void *cublas_dso_handle = nullptr;
+
+/**
+ * The following macro definition can generate structs
+ * (for each function) to dynamic load cublas routine
+ * via operator overloading.
+ *
+ * note: default dynamic linked libs
+ */
+#ifdef PADDLE_USE_DSO
+#define DYNAMIC_LOAD_CUBLAS_WRAP(__name)                                       \
+  struct DynLoad__##__name {                                                   \
+    template <typename... Args>                                                \
+    cublasStatus_t operator()(Args... args) {                                  \
+      typedef cublasStatus_t (*cublasFunc)(Args...);                           \
+      std::call_once(cublas_dso_flag, GetCublasDsoHandle, &cublas_dso_handle); \
+      void *p_##__name = dlsym(cublas_dso_handle, #__name);                    \
+      return reinterpret_cast<cublasFunc>(p_##__name)(args...);                \
+    }                                                                          \
+  } __name;  // struct DynLoad__##__name
+#else
+#define DYNAMIC_LOAD_CUBLAS_WRAP(__name)      \
+  struct DynLoad__##__name {                  \
+    template <typename... Args>               \
+    cublasStatus_t operator()(Args... args) { \
+      return __name(args...);                 \
+    }                                         \
+  } __name;  // struct DynLoad__##__name
+#endif
+
+#define DYNAMIC_LOAD_CUBLAS_V2_WRAP(__name) DYNAMIC_LOAD_CUBLAS_WRAP(__name)
+
+// include all needed cublas functions in HPPL
+// clang-format off
+#define CUBLAS_BLAS_ROUTINE_EACH(__macro) \
+  __macro(cublasSgemv)                    \
+  __macro(cublasDgemv)                    \
+  __macro(cublasSgemm)                    \
+  __macro(cublasDgemm)                    \
+  __macro(cublasSgeam)                    \
+  __macro(cublasDgeam)                    \
+
+DYNAMIC_LOAD_CUBLAS_V2_WRAP(cublasCreate)
+DYNAMIC_LOAD_CUBLAS_V2_WRAP(cublasDestroy)
+DYNAMIC_LOAD_CUBLAS_V2_WRAP(cublasSetStream)
+DYNAMIC_LOAD_CUBLAS_V2_WRAP(cublasSetPointerMode)
+DYNAMIC_LOAD_CUBLAS_V2_WRAP(cublasGetPointerMode)
+DYNAMIC_LOAD_CUBLAS_WRAP(cublasSgemmBatched)
+DYNAMIC_LOAD_CUBLAS_WRAP(cublasDgemmBatched)
+DYNAMIC_LOAD_CUBLAS_WRAP(cublasCgemmBatched)
+DYNAMIC_LOAD_CUBLAS_WRAP(cublasZgemmBatched)
+DYNAMIC_LOAD_CUBLAS_WRAP(cublasSgetrfBatched)
+DYNAMIC_LOAD_CUBLAS_WRAP(cublasSgetriBatched)
+DYNAMIC_LOAD_CUBLAS_WRAP(cublasDgetrfBatched)
+DYNAMIC_LOAD_CUBLAS_WRAP(cublasDgetriBatched)
+CUBLAS_BLAS_ROUTINE_EACH(DYNAMIC_LOAD_CUBLAS_V2_WRAP)
+
+#undef DYNAMIC_LOAD_CUBLAS_WRAP
+#undef DYNAMIC_LOAD_CUBLAS_V2_WRAP
+#undef CUBLAS_BLAS_ROUTINE_EACH
+
+} /* namespace dynload */
+
+// clang-format on
+#ifndef PADDLE_TYPE_DOUBLE
+#define CUBLAS_GEAM dynload::cublasSgeam
+#define CUBLAS_GEMV dynload::cublasSgemv
+#define CUBLAS_GEMM dynload::cublasSgemm
+#define CUBLAS_GETRF dynload::cublasSgetrfBatched
+#define CUBLAS_GETRI dynload::cublasSgetriBatched
+#else
+#define CUBLAS_GEAM dynload::cublasDgeam
+#define CUBLAS_GEMV dynload::cublasDgemv
+#define CUBLAS_GEMM dynload::cublasDgemm
+#define CUBLAS_GETRF dynload::cublasDgetrfBatched
+#define CUBLAS_GETRI dynload::cublasDgetriBatched
+#endif
+}  // namespace dyload
+}  // namespace paddle

paddle/platform/cudnn.h

+#include <cudnn.h>
+#include "paddle/platform/dynamic_loader.h"
+
+namespace paddle {
+namespace dyload {
+
+std::once_flag cudnn_dso_flag;
+void* cudnn_dso_handle = nullptr;
+
+#ifdef PADDLE_USE_DSO
+
+#define DYNAMIC_LOAD_CUDNN_WRAP(__name)                                     \
+  struct DynLoad__##__name {                                                \
+    template <typename... Args>                                             \
+    auto operator()(Args... args) -> decltype(__name(args...)) {            \
+      using cudnn_func = decltype(__name(args...)) (*)(Args...);            \
+      std::call_once(cudnn_dso_flag, GetCudnnDsoHandle, &cudnn_dso_handle); \
+      void* p_##__name = dlsym(cudnn_dso_handle, #__name);                  \
+      return reinterpret_cast<cudnn_func>(p_##__name)(args...);             \
+    }                                                                       \
+  } __name; /* struct DynLoad__##__name */
+
+#else
+
+#define DYNAMIC_LOAD_CUDNN_WRAP(__name)                          \
+  struct DynLoad__##__name {                                     \
+    template <typename... Args>                                  \
+    auto operator()(Args... args) -> decltype(__name(args...)) { \
+      return __name(args...);                                    \
+    }                                                            \
+  } __name; /* struct DynLoad__##__name */
+
+#endif
+
+/**
+ * include all needed cudnn functions in HPPL
+ * different cudnn version has different interfaces
+ **/
+// clang-format off
+#define CUDNN_DNN_ROUTINE_EACH(__macro)                   \
+  __macro(cudnnSetTensor4dDescriptor)                     \
+  __macro(cudnnSetTensor4dDescriptorEx)                   \
+  __macro(cudnnGetConvolutionNdForwardOutputDim)          \
+  __macro(cudnnGetConvolutionForwardAlgorithm)            \
+  __macro(cudnnCreateTensorDescriptor)                    \
+  __macro(cudnnDestroyTensorDescriptor)                   \
+  __macro(cudnnCreateFilterDescriptor)                    \
+  __macro(cudnnSetFilter4dDescriptor)                     \
+  __macro(cudnnSetPooling2dDescriptor)                    \
+  __macro(cudnnDestroyFilterDescriptor)                   \
+  __macro(cudnnCreateConvolutionDescriptor)               \
+  __macro(cudnnCreatePoolingDescriptor)                   \
+  __macro(cudnnDestroyPoolingDescriptor)                  \
+  __macro(cudnnSetConvolution2dDescriptor)                \
+  __macro(cudnnDestroyConvolutionDescriptor)              \
+  __macro(cudnnCreate)                                    \
+  __macro(cudnnDestroy)                                   \
+  __macro(cudnnSetStream)                                 \
+  __macro(cudnnActivationForward)                         \
+  __macro(cudnnConvolutionForward)                        \
+  __macro(cudnnConvolutionBackwardBias)                   \
+  __macro(cudnnGetConvolutionForwardWorkspaceSize)        \
+  __macro(cudnnTransformTensor)                           \
+  __macro(cudnnPoolingForward)                            \
+  __macro(cudnnPoolingBackward)                           \
+  __macro(cudnnSoftmaxBackward)                           \
+  __macro(cudnnSoftmaxForward)                            \
+  __macro(cudnnGetVersion)                                \
+  __macro(cudnnGetErrorString)
+CUDNN_DNN_ROUTINE_EACH(DYNAMIC_LOAD_CUDNN_WRAP)
+
+#define CUDNN_DNN_ROUTINE_EACH_R2(__macro)                \
+  __macro(cudnnAddTensor)                                 \
+  __macro(cudnnConvolutionBackwardData)                   \
+  __macro(cudnnConvolutionBackwardFilter)
+CUDNN_DNN_ROUTINE_EACH_R2(DYNAMIC_LOAD_CUDNN_WRAP)
+
+// APIs available after R3:
+#if CUDNN_VERSION >= 3000
+#define CUDNN_DNN_ROUTINE_EACH_AFTER_R3(__macro)              \
+  __macro(cudnnGetConvolutionBackwardFilterWorkspaceSize)     \
+  __macro(cudnnGetConvolutionBackwardDataAlgorithm)           \
+  __macro(cudnnGetConvolutionBackwardFilterAlgorithm)         \
+  __macro(cudnnGetConvolutionBackwardDataWorkspaceSize)
+CUDNN_DNN_ROUTINE_EACH_AFTER_R3(DYNAMIC_LOAD_CUDNN_WRAP)
+#undef CUDNN_DNN_ROUTINE_EACH_AFTER_R3
+#endif
+
+
+// APIs available after R4:
+#if CUDNN_VERSION >= 4007
+#define CUDNN_DNN_ROUTINE_EACH_AFTER_R4(__macro)             \
+  __macro(cudnnBatchNormalizationForwardTraining)            \
+  __macro(cudnnBatchNormalizationForwardInference)           \
+  __macro(cudnnBatchNormalizationBackward)
+CUDNN_DNN_ROUTINE_EACH_AFTER_R4(DYNAMIC_LOAD_CUDNN_WRAP)
+#undef CUDNN_DNN_ROUTINE_EACH_AFTER_R4
+#endif
+
+// APIs in R5
+#if CUDNN_VERSION >= 5000
+#define CUDNN_DNN_ROUTINE_EACH_R5(__macro)                    \
+  __macro(cudnnCreateActivationDescriptor)                    \
+  __macro(cudnnSetActivationDescriptor)                       \
+  __macro(cudnnGetActivationDescriptor)                       \
+  __macro(cudnnDestroyActivationDescriptor)
+CUDNN_DNN_ROUTINE_EACH_R5(DYNAMIC_LOAD_CUDNN_WRAP)
+#undef CUDNN_DNN_ROUTINE_EACH_R5
+#endif
+
+#undef CUDNN_DNN_ROUTINE_EACH
+// clang-format on
+}  // namespace dyload
+}  // namespace paddle

paddle/platform/curand.h

+#include <curand.h>
+#include "paddle/platform/dynamic_loader.h"
+
+namespace paddle {
+namespace dyload {
+#ifdef PADDLE_USE_DSO
+#define DYNAMIC_LOAD_CURAND_WRAP(__name)                                       \
+  struct DynLoad__##__name {                                                   \
+    template <typename... Args>                                                \
+    curandStatus_t operator()(Args... args) {                                  \
+      typedef curandStatus_t (*curandFunc)(Args...);                           \
+      std::call_once(curand_dso_flag, GetCurandDsoHandle, &curand_dso_handle); \
+      void *p_##__name = dlsym(curand_dso_handle, #__name);                    \
+      return reinterpret_cast<curandFunc>(p_##__name)(args...);                \
+    }                                                                          \
+  } __name; /* struct DynLoad__##__name */
+#else
+#define DYNAMIC_LOAD_CURAND_WRAP(__name)      \
+  struct DynLoad__##__name {                  \
+    template <typename... Args>               \
+    curandStatus_t operator()(Args... args) { \
+      return __name(args...);                 \
+    }                                         \
+  } __name; /* struct DynLoad__##__name */
+#endif
+
+/* include all needed curand functions in HPPL */
+// clang-format off
+#define CURAND_RAND_ROUTINE_EACH(__macro)    \
+  __macro(curandCreateGenerator)             \
+  __macro(curandSetStream)                   \
+  __macro(curandSetPseudoRandomGeneratorSeed)\
+  __macro(curandGenerateUniform)             \
+  __macro(curandGenerateUniformDouble)
+// clang-format on
+
+CURAND_RAND_ROUTINE_EACH(DYNAMIC_LOAD_CURAND_WRAP)
+
+#undef CURAND_RAND_ROUTINE_EACH
+#undef DYNAMIC_LOAD_CURAND_WRAP
+}
+}  // namespace paddle

paddle/platform/dynamic_loader.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+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 <gflags/gflags.h>
+#include "DynamicLoader.h"
+#include "Logging.h"
+
+DEFINE_string(cudnn_dir, "",
+              "Specify path for loading libcudnn.so. For instance, "
+              "/usr/local/cudnn/lib. If empty [default], dlopen "
+              "will search cudnn from LD_LIBRARY_PATH");
+
+DEFINE_string(cuda_dir, "",
+              "Specify path for loading cuda library, such as libcublas, "
+              "libcurand. For instance, /usr/local/cuda/lib64. If default, "
+              "dlopen will search cuda from LD_LIBRARY_PATH");
+
+DEFINE_string(warpctc_dir, "", "Specify path for loading libwarpctc.so.");
+
+DEFINE_string(lapack_dir, "", "Specify path for loading liblapack.so.");
+
+static inline std::string join(const std::string& part1,
+                               const std::string& part2) {
+  // directory separator
+  const char sep = '/';
+  if (!part2.empty() && part2.front() == sep) {
+    return part2;
+  }
+  std::string ret;
+  ret.reserve(part1.size() + part2.size() + 1);
+  ret = part1;
+  if (!ret.empty() && ret.back() != sep) {
+    ret += sep;
+  }
+  ret += part2;
+  return ret;
+}
+
+static inline void GetDsoHandleFromDefaultPath(std::string& dso_path,
+                                               void** dso_handle,
+                                               int dynload_flags) {
+  VLOG(3) << "Try to find library: " << dso_path
+          << " from default system path.";
+  // default search from LD_LIBRARY_PATH/DYLD_LIBRARY_PATH
+  *dso_handle = dlopen(dso_path.c_str(), dynload_flags);
+
+// DYLD_LIBRARY_PATH is disabled after Mac OS 10.11 to
+// bring System Integrity Projection (SIP), if dso_handle
+// is null, search from default package path in Mac OS.
+#if defined(__APPLE__) || defined(__OSX__)
+  if (nullptr == *dso_handle) {
+    dso_path = join("/usr/local/cuda/lib/", dso_path);
+    *dso_handle = dlopen(dso_path.c_str(), dynload_flags);
+    if (nullptr == *dso_handle) {
+      if (dso_path == "libcudnn.dylib") {
+        LOG(FATAL)
+            << "Note: [Recommend] copy cudnn into /usr/local/cuda/ \n"  // NOLINT
+            << "For instance, sudo tar -xzf "
+               "cudnn-7.5-osx-x64-v5.0-ga.tgz -C "  // NOLINT
+            << "/usr/local \n sudo chmod a+r "
+               "/usr/local/cuda/include/cudnn.h "  // NOLINT
+            << "/usr/local/cuda/lib/libcudnn*";
+      }
+    }
+  }
+#endif
+}
+
+static inline void GetDsoHandleFromSearchPath(const std::string& search_root,
+                                              const std::string& dso_name,
+                                              void** dso_handle) {
+  int dynload_flags = RTLD_LAZY | RTLD_LOCAL;
+  *dso_handle = nullptr;
+
+  std::string dlPath = dso_name;
+  if (search_root.empty()) {
+    GetDsoHandleFromDefaultPath(dlPath, dso_handle, dynload_flags);
+  } else {
+    // search xxx.so from custom path
+    dlPath = join(search_root, dso_name);
+    *dso_handle = dlopen(dlPath.c_str(), dynload_flags);
+    // if not found, search from default path
+    if (nullptr == *dso_handle) {
+      LOG(WARNING) << "Failed to find dynamic library: " << dlPath << " ("
+                   << dlerror() << ")";
+      dlPath = dso_name;
+      GetDsoHandleFromDefaultPath(dlPath, dso_handle, dynload_flags);
+    }
+  }
+
+  CHECK(nullptr != *dso_handle) << "Failed to find dynamic library: " << dlPath
+                                << " (" << dlerror() << ") \n"
+                                << "Please specify its path correctly using "
+                                   "following ways: \n"
+
+                                << "Method. set environment variable "
+                                   "LD_LIBRARY_PATH on Linux or "
+                                << "DYLD_LIBRARY_PATH on Mac OS. \n"
+                                << "For instance, issue command: export "
+                                   "LD_LIBRARY_PATH=... \n"
+
+                                << "Note: After Mac OS 10.11, using the "
+                                   "DYLD_LIBRARY_PATH is impossible "
+                                << "unless System Integrity Protection (SIP) "
+                                   "is disabled.";
+}
+
+void GetCublasDsoHandle(void** dso_handle) {
+#if defined(__APPLE__) || defined(__OSX__)
+  GetDsoHandleFromSearchPath(FLAGS_cuda_dir, "libcublas.dylib", dso_handle);
+#else
+  GetDsoHandleFromSearchPath(FLAGS_cuda_dir, "libcublas.so", dso_handle);
+#endif
+}
+
+void GetCudnnDsoHandle(void** dso_handle) {
+#if defined(__APPLE__) || defined(__OSX__)
+  GetDsoHandleFromSearchPath(FLAGS_cudnn_dir, "libcudnn.dylib", dso_handle);
+#else
+  GetDsoHandleFromSearchPath(FLAGS_cudnn_dir, "libcudnn.so", dso_handle);
+#endif
+}
+
+void GetCurandDsoHandle(void** dso_handle) {
+#if defined(__APPLE__) || defined(__OSX__)
+  GetDsoHandleFromSearchPath(FLAGS_cuda_dir, "libcurand.dylib", dso_handle);
+#else
+  GetDsoHandleFromSearchPath(FLAGS_cuda_dir, "libcurand.so", dso_handle);
+#endif
+}
+
+void GetWarpCTCDsoHandle(void** dso_handle) {
+#if defined(__APPLE__) || defined(__OSX__)
+  GetDsoHandleFromSearchPath(FLAGS_warpctc_dir, "libwarpctc.dylib", dso_handle);
+#else
+  GetDsoHandleFromSearchPath(FLAGS_warpctc_dir, "libwarpctc.so", dso_handle);
+#endif
+}
+
+void GetLapackDsoHandle(void** dso_handle) {
+#if defined(__APPLE__) || defined(__OSX__)
+  GetDsoHandleFromSearchPath(FLAGS_lapack_dir, "liblapacke.dylib", dso_handle);
+#else
+  GetDsoHandleFromSearchPath(FLAGS_lapack_dir, "liblapacke.so", dso_handle);
+#endif
+}

paddle/platform/dynamic_loader.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+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. */
+
+#ifndef DYNAMIC_LOAD_H_
+#define DYNAMIC_LOAD_H_
+
+#include <dlfcn.h>
+#include <memory>
+#include <mutex>
+#include <string>
+
+/**
+ * @brief    load the DSO of CUBLAS
+ *
+ * @param    **dso_handle   dso handler
+ *
+ */
+void GetCublasDsoHandle(void** dso_handle);
+
+/**
+ * @brief    load the DSO of CUDNN
+ *
+ * @param    **dso_handle   dso handler
+ *
+ */
+void GetCudnnDsoHandle(void** dso_handle);
+
+/**
+ * @brief    load the DSO of CURAND
+ *
+ * @param    **dso_handle   dso handler
+ *
+ */
+void GetCurandDsoHandle(void** dso_handle);
+
+/**
+ * @brief    load the DSO of warp-ctc
+ *
+ * @param    **dso_handle   dso handler
+ *
+ */
+void GetWarpCTCDsoHandle(void** dso_handle);
+
+/**
+ * @brief    load the DSO of lapack
+ *
+ * @param    **dso_handle   dso handler
+ *
+ */
+void GetLapackDsoHandle(void** dso_handle);
+
+#endif  // DYNAMIC_LOAD_H_