stream_gpu-inl.h

Go to the documentation of this file.

/*
 * Licensed to the Apache Software Foundation (ASF) under one
 * or more contributor license agreements.  See the NOTICE file
 * distributed with this work for additional information
 * regarding copyright ownership.  The ASF licenses this file
 * to you 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 MSHADOW_STREAM_GPU_INL_H_
#define MSHADOW_STREAM_GPU_INL_H_
#include <memory>
#include "./base.h"
#include "./tensor.h"
#include "./logging.h"

namespace mshadow {
#if MSHADOW_USE_CUDA == 1
// Stream alocation
// actual implementation of GPU stream in CUDA
template<>
struct Stream<gpu> {
  enum HandleState {
    NoHandle = 0,
    OwnHandle = 1,
  };
  cudaStream_t stream_;
  cublasHandle_t blas_handle_;
  #if MSHADOW_USE_CUSOLVER == 1
  cusolverDnHandle_t solver_handle_;
  #endif

  #if MSHADOW_USE_CUDNN == 1
  cudnnHandle_t dnn_handle_;
  #endif

  HandleState blas_handle_ownership_;
  HandleState solver_handle_ownership_;
  HandleState dnn_handle_ownership_;
  cudaDeviceProp prop;
  int dev_id;

  Stream(void)
    : stream_(0)
      , blas_handle_(0)
#if MSHADOW_USE_CUDNN == 1
      , dnn_handle_(0)
#endif
      , blas_handle_ownership_(NoHandle)
      , solver_handle_ownership_(NoHandle)
      , dnn_handle_ownership_(NoHandle) {}
  inline void Wait(void) {
    MSHADOW_CUDA_CALL(cudaStreamSynchronize(stream_));
  }
  inline bool CheckIdle(void) {
    cudaError_t err = cudaStreamQuery(stream_);
    if (err == cudaSuccess) return true;
    if (err == cudaErrorNotReady) return false;
    LOG(FATAL) << cudaGetErrorString(err);
    return false;
  }
  inline static cudaStream_t GetStream(Stream<gpu> *stream) {
    if (stream == NULL) {
#if MSHADOW_FORCE_STREAM
      LOG(FATAL) << "Default GPU stream was used when MSHADOW_FORCE_STREAM was on";
#endif
      return 0;
    } else {
      return stream->stream_;
    }
  }
  inline static cublasHandle_t GetBlasHandle(Stream<gpu> *stream) {
    if (stream == NULL) {
      return 0;
    } else {
      CHECK_NE(stream->blas_handle_ownership_, NoHandle)
        << "No handle exist in source stream";
      return stream->blas_handle_;
    }
  }
  inline void DestroyBlasHandle() {
    if (blas_handle_ownership_ == OwnHandle) {
      cublasStatus_t err = cublasDestroy(blas_handle_);
      blas_handle_ownership_ = NoHandle;
      CHECK_EQ(err, CUBLAS_STATUS_SUCCESS) << "Destory cublas handle failed";
    }
  }
  inline void CreateBlasHandle() {
    this->DestroyBlasHandle();
    cublasStatus_t err = cublasCreate(&blas_handle_);
    blas_handle_ownership_ = OwnHandle;
    CHECK_EQ(err, CUBLAS_STATUS_SUCCESS) << "Create cublas handle failed";
    err = cublasSetStream(blas_handle_, stream_);
    CHECK_EQ(err, CUBLAS_STATUS_SUCCESS) << "Setting cublas stream failed";
  }
#if MSHADOW_USE_CUSOLVER == 1
  inline static cusolverDnHandle_t GetSolverHandle(Stream<gpu> *stream) {
    if (stream == NULL) {
      return 0;
    } else {
      CHECK_NE(stream->solver_handle_ownership_, NoHandle) << "No handle exist in source stream";
      return stream->solver_handle_;
    }
  }
#endif
  inline void DestroySolverHandle() {
#if MSHADOW_USE_CUSOLVER == 1
    if (solver_handle_ownership_ == OwnHandle) {
      cusolverStatus_t err = cusolverDnDestroy(solver_handle_);
      CHECK_EQ(err, CUSOLVER_STATUS_SUCCESS) << "Destory cusolver handle failed";
    }
#endif
  }
  inline void CreateSolverHandle() {
#if MSHADOW_USE_CUSOLVER == 1
    this->DestroySolverHandle();
    cusolverStatus_t err = cusolverDnCreate(&solver_handle_);
    CHECK_EQ(err, CUSOLVER_STATUS_SUCCESS) << "Create cusolver handle failed";
    err = cusolverDnSetStream(solver_handle_, stream_);
    CHECK_EQ(err, CUSOLVER_STATUS_SUCCESS) << "Setting cusolver stream failed";
    this->solver_handle_ownership_ = OwnHandle;
#endif
  }
// #if MSHADOW_USE_CUDNN && defined(__CUDACC__)
#if MSHADOW_USE_CUDNN == 1
  inline static cudnnHandle_t GetDnnHandle(Stream<gpu> *stream) {
    if (stream == NULL) {
      return 0;
    } else {
      CHECK_NE(stream->dnn_handle_ownership_, NoHandle) << "No handle exist in source stream";
      return stream->dnn_handle_;
    }
  }
#endif
  inline void DestroyDnnHandle() {
// #if MSHADOW_USE_CUDNN && defined(__CUDACC__)
#if MSHADOW_USE_CUDNN == 1
    if (dnn_handle_ownership_ == OwnHandle) {
      cudnnStatus_t err = cudnnDestroy(dnn_handle_);
      this->dnn_handle_ownership_ = NoHandle;
      CHECK_EQ(err, CUDNN_STATUS_SUCCESS) << cudnnGetErrorString(err);
    }
#endif
  }
  inline void CreateDnnHandle() {
// #if MSHADOW_USE_CUDNN == 1 && defined(__CUDACC__)
#if MSHADOW_USE_CUDNN == 1
    this->DestroyDnnHandle();
    cudnnStatus_t err = cudnnCreate(&dnn_handle_);
    CHECK_EQ(err, CUDNN_STATUS_SUCCESS) << cudnnGetErrorString(err);
    // At this point, we have the resource which may need to be freed
    this->dnn_handle_ownership_ = OwnHandle;
    err = cudnnSetStream(dnn_handle_, stream_);
    CHECK_EQ(err, CUDNN_STATUS_SUCCESS) << cudnnGetErrorString(err);
#endif
  }
};
template<>
inline void DeleteStream<gpu>(Stream<gpu> *stream) {
  if (stream) {
    MSHADOW_CUDA_CALL(cudaStreamDestroy(stream->stream_));
    stream->DestroyBlasHandle();
    stream->DestroySolverHandle();
    stream->DestroyDnnHandle();
    delete stream;
  }
}
template<>
inline Stream<gpu> *NewStream<gpu>(bool create_blas_handle,
                                   bool create_dnn_handle,
                                   int dev_id) {
  // RAII on Cuda exception
  struct StreamDeleter { void operator()(Stream<gpu> *ptr) const { DeleteStream<gpu>(ptr); } };
  std::unique_ptr<Stream<gpu>, StreamDeleter> st(new Stream<gpu>());
  MSHADOW_CUDA_CALL(cudaStreamCreate(&st->stream_));
  if (create_blas_handle) {
    st->CreateBlasHandle();
    st->CreateSolverHandle();
  }
  if (create_dnn_handle) {
    st->CreateDnnHandle();
  }
  st->dev_id = dev_id;
  if (dev_id != -1) {
    MSHADOW_CUDA_CALL(cudaGetDeviceProperties(&st->prop, dev_id));
  }
  return st.release();
}
#endif
}  // namespace mshadow
#endif  // MSHADOW_STREAM_GPU_INL_H_