device_context.h 5.0 KB
Newer Older
Q
QI JUN 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13
/* 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. */

#pragma once

L
liaogang 已提交
14
#include "paddle/platform/enforce.h"
L
liaogang 已提交
15 16
#include "paddle/platform/place.h"

Q
QI JUN 已提交
17 18 19 20
#ifndef PADDLE_ONLY_CPU
#include "paddle/platform/dynload/cublas.h"
#include "paddle/platform/dynload/cudnn.h"
#include "paddle/platform/dynload/curand.h"
L
liaogang 已提交
21
#include "paddle/platform/gpu_info.h"
Q
QI JUN 已提交
22 23
#define EIGEN_USE_GPU
#endif
Q
qijun 已提交
24
#include <memory>
Q
qijun 已提交
25 26
#include "paddle/platform/place.h"
#include "unsupported/Eigen/CXX11/Tensor"
Q
QI JUN 已提交
27 28 29 30 31 32 33

namespace paddle {
namespace platform {

class DeviceContext {
 public:
  virtual ~DeviceContext() {}
Q
qijun 已提交
34
  virtual Place GetPlace() const = 0;
Q
QI JUN 已提交
35

Q
qijun 已提交
36
  template <typename DeviceType>
Q
qijun 已提交
37
  DeviceType* get_eigen_device() const;
Q
QI JUN 已提交
38 39
};

Q
qijun 已提交
40 41
class CPUDeviceContext : public DeviceContext {
 public:
Q
qijun 已提交
42 43 44
  CPUDeviceContext() { eigen_device_.reset(new Eigen::DefaultDevice()); }

  Eigen::DefaultDevice* eigen_device() const { return eigen_device_.get(); }
Q
qijun 已提交
45

Y
Yu Yang 已提交
46
  Place GetPlace() const override {
Q
qijun 已提交
47
    Place retv = CPUPlace();
Y
Yu Yang 已提交
48 49 50
    return retv;
  }

Q
qijun 已提交
51
 private:
Q
qijun 已提交
52
  std::unique_ptr<Eigen::DefaultDevice> eigen_device_;
Q
QI JUN 已提交
53 54 55
};

#ifndef PADDLE_ONLY_CPU
D
dongzhihong 已提交
56

Q
QI JUN 已提交
57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74
class GPUPlaceGuard {
 public:
  explicit GPUPlaceGuard(GPUPlace new_place) : previous_(GetCurrentDeviceId()) {
    if (previous_ != new_place) {
      paddle::platform::SetDeviceId(new_place.device);
    }
  }

  ~GPUPlaceGuard() { paddle::platform::SetDeviceId(previous_.device); }

 private:
  GPUPlace previous_;
};

class CUDADeviceContext : public DeviceContext {
 public:
  explicit CUDADeviceContext(const GPUPlace gpu_place) : gpu_place_(gpu_place) {
    GPUPlaceGuard guard(gpu_place_);
L
liaogang 已提交
75
    PADDLE_ENFORCE(cudaStreamCreate(&stream_), "cudaStreamCreate failed");
Q
qijun 已提交
76 77
    eigen_stream_.reset(new Eigen::CudaStreamDevice(&stream_));
    eigen_device_.reset(new Eigen::GpuDevice(eigen_stream_.get()));
Q
qijun 已提交
78 79 80 81 82
  }

  Place GetPlace() const override {
    Place retv = GPUPlace();
    return retv;
Q
QI JUN 已提交
83 84 85
  }

  void Wait() {
L
liaogang 已提交
86 87
    PADDLE_ENFORCE(cudaStreamSynchronize(stream_),
                   "cudaStreamSynchronize failed");
Q
QI JUN 已提交
88 89 90 91
  }

  cudaStream_t stream() { return stream_; }

Q
qijun 已提交
92
  Eigen::GpuDevice* eigen_device() const { return eigen_device_.get(); }
Q
QI JUN 已提交
93 94 95 96

  cublasHandle_t cublas_handle() {
    if (!blas_handle_) {
      GPUPlaceGuard guard(gpu_place_);
L
liaogang 已提交
97
      PADDLE_ENFORCE(paddle::platform::dynload::cublasCreate(&blas_handle_),
Q
QI JUN 已提交
98
                     "cublasCreate failed");
L
liaogang 已提交
99 100 101
      PADDLE_ENFORCE(
          paddle::platform::dynload::cublasSetStream(blas_handle_, stream_),
          "cublasSetStream failed");
Q
QI JUN 已提交
102 103 104 105 106 107 108
    }
    return blas_handle_;
  }

  cudnnHandle_t cudnn_handle() {
    if (!dnn_handle_) {
      GPUPlaceGuard guard(gpu_place_);
L
liaogang 已提交
109
      PADDLE_ENFORCE(paddle::platform::dynload::cudnnCreate(&dnn_handle_),
Q
QI JUN 已提交
110
                     "cudnnCreate failed");
L
liaogang 已提交
111 112 113
      PADDLE_ENFORCE(
          paddle::platform::dynload::cudnnSetStream(dnn_handle_, stream_),
          "cudnnSetStream failed");
Q
QI JUN 已提交
114 115 116 117 118 119 120 121
    }
    return dnn_handle_;
  }

  curandGenerator_t curand_generator() {
    if (!rand_generator_) {
      GPUPlaceGuard guard(gpu_place_);
      PADDLE_ENFORCE(paddle::platform::dynload::curandCreateGenerator(
L
liaogang 已提交
122
                         &rand_generator_, CURAND_RNG_PSEUDO_DEFAULT),
Q
QI JUN 已提交
123 124 125
                     "curandCreateGenerator failed");
      PADDLE_ENFORCE(
          paddle::platform::dynload::curandSetPseudoRandomGeneratorSeed(
L
liaogang 已提交
126
              rand_generator_, random_seed_),
Q
QI JUN 已提交
127
          "curandSetPseudoRandomGeneratorSeed failed");
L
liaogang 已提交
128 129 130
      PADDLE_ENFORCE(
          paddle::platform::dynload::curandSetStream(rand_generator_, stream_),
          "curandSetStream failed");
Q
QI JUN 已提交
131 132 133 134 135 136 137
    }
    return rand_generator_;
  }

  ~CUDADeviceContext() {
    Wait();
    if (blas_handle_) {
L
liaogang 已提交
138
      PADDLE_ENFORCE(paddle::platform::dynload::cublasDestroy(blas_handle_),
Q
QI JUN 已提交
139 140 141 142
                     "cublasDestroy failed");
    }

    if (dnn_handle_) {
L
liaogang 已提交
143
      PADDLE_ENFORCE(paddle::platform::dynload::cudnnDestroy(dnn_handle_),
Q
QI JUN 已提交
144 145 146 147
                     "cudnnDestroy failed");
    }

    if (rand_generator_) {
L
liaogang 已提交
148 149 150
      PADDLE_ENFORCE(
          paddle::platform::dynload::curandDestroyGenerator(rand_generator_),
          "curandDestroyGenerator failed");
Q
QI JUN 已提交
151
    }
Q
qijun 已提交
152 153
    eigen_stream_.reset();
    eigen_device_.reset();
L
liaogang 已提交
154
    PADDLE_ENFORCE(cudaStreamDestroy(stream_), "cudaStreamDestroy failed");
Q
QI JUN 已提交
155 156 157 158 159 160
  }

 private:
  GPUPlace gpu_place_;
  cudaStream_t stream_;

Q
qijun 已提交
161 162
  std::unique_ptr<Eigen::CudaStreamDevice> eigen_stream_;
  std::unique_ptr<Eigen::GpuDevice> eigen_device_;
Q
QI JUN 已提交
163 164 165 166 167 168 169 170

  cublasHandle_t blas_handle_{nullptr};

  cudnnHandle_t dnn_handle_{nullptr};

  int random_seed_;
  curandGenerator_t rand_generator_{nullptr};
};
Q
qijun 已提交
171

Q
QI JUN 已提交
172
#endif
Q
qijun 已提交
173

Q
QI JUN 已提交
174 175
}  // namespace platform
}  // namespace paddle