pool_cudnn_op.cu.cc 6.0 KB
Newer Older
C
chengduoZH 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39
/* 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 "paddle/operators/pool_cudnn_op.h"
#include "paddle/platform/cudnn_helper.h"

namespace paddle {
namespace operators {

using Tensor = framework::Tensor;
using ScopedTensorDescriptor = platform::ScopedTensorDescriptor;
using ScopedPoolingDescriptor = platform::ScopedPoolingDescriptor;
using DataLayout = platform::DataLayout;
using PoolingMode = platform::PoolingMode;

template <typename T>
class PoolCudnnOpKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext &ctx) const override {
    PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
                   "It must use GPUPlace.");

    const Tensor *input = ctx.Input<Tensor>("X");
    Tensor *output = ctx.Output<Tensor>("Out");

    const T *input_data = input->data<T>();
    T *output_data = output->mutable_data<T>(ctx.GetPlace());

C
chengduoZH 已提交
40
    std::string pooling_type = ctx.Attr<std::string>("pooling_type");
C
chengduoZH 已提交
41 42 43
    std::vector<int> ksize = ctx.Attr<std::vector<int>>("ksize");
    std::vector<int> strides = ctx.Attr<std::vector<int>>("strides");
    std::vector<int> paddings = ctx.Attr<std::vector<int>>("paddings");
C
chengduoZH 已提交
44
    if (ctx.Attr<bool>("global_pooling")) {
C
chengduoZH 已提交
45
      for (size_t i = 0; i < ksize.size(); ++i) {
C
fix bug  
chengduoZH 已提交
46
        paddings[i] = 0;
C
chengduoZH 已提交
47 48 49 50 51 52 53 54
        ksize[i] = static_cast<int>(input->dims()[i + 2]);
      }
    }

    // ------------------- cudnn descriptors ---------------------
    ScopedTensorDescriptor input_desc;
    ScopedTensorDescriptor output_desc;
    ScopedPoolingDescriptor pool_desc;
C
chengduoZH 已提交
55 56 57 58 59 60 61
    DataLayout layout;

    if (strides.size() == 2U) {
      layout = DataLayout::kNCHW;
    } else {
      layout = DataLayout::kNCDHW;
    }
C
chengduoZH 已提交
62

C
chengduoZH 已提交
63 64 65 66
    cudnnTensorDescriptor_t cudnn_input_desc = input_desc.descriptor<T>(
        layout, framework::vectorize2int(input->dims()));
    cudnnTensorDescriptor_t cudnn_output_desc = output_desc.descriptor<T>(
        layout, framework::vectorize2int(output->dims()));
C
chengduoZH 已提交
67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100

    PoolingMode pooling_mode;
    if (pooling_type == "max") {
      pooling_mode = PoolingMode::kMaximum;
    } else {
      pooling_mode = PoolingMode::kAverage;
    }

    cudnnPoolingDescriptor_t cudnn_pool_desc =
        pool_desc.descriptor(pooling_mode, ksize, paddings, strides);

    // ------------------- cudnn pool algorithm ---------------------
    auto handle = ctx.cuda_device_context().cudnn_handle();
    T alpha = 1.0f, beta = 0.0f;

    PADDLE_ENFORCE(platform::dynload::cudnnPoolingForward(
        handle, cudnn_pool_desc, &alpha, cudnn_input_desc, input_data, &beta,
        cudnn_output_desc, output_data));
  }
};

template <typename T>
class PoolCudnnGradOpKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext &ctx) const override {
    PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
                   "It must use GPUPlace.");

    const Tensor *input = ctx.Input<Tensor>("X");
    const Tensor *output = ctx.Input<Tensor>("Out");
    const Tensor *output_grad =
        ctx.Input<Tensor>(framework::GradVarName("Out"));
    Tensor *input_grad = ctx.Output<Tensor>(framework::GradVarName("X"));

C
chengduoZH 已提交
101
    std::string pooling_type = ctx.Attr<std::string>("pooling_type");
C
chengduoZH 已提交
102 103 104 105
    std::vector<int> ksize = ctx.Attr<std::vector<int>>("ksize");
    std::vector<int> strides = ctx.Attr<std::vector<int>>("strides");
    std::vector<int> paddings = ctx.Attr<std::vector<int>>("paddings");

C
chengduoZH 已提交
106
    if (ctx.Attr<bool>("global_pooling")) {
C
fix bug  
chengduoZH 已提交
107 108
      for (size_t i = 0; i < ksize.size(); ++i) {
        paddings[i] = 0;
C
chengduoZH 已提交
109
        ksize[i] = static_cast<int>(input->dims()[i + 2]);
C
fix bug  
chengduoZH 已提交
110
      }
C
chengduoZH 已提交
111 112 113 114 115 116 117 118 119 120
    }

    const T *input_data = input->data<T>();
    const T *output_data = output->data<T>();
    const T *output_grad_data = output_grad->data<T>();

    // ------------------- cudnn descriptors ---------------------
    ScopedTensorDescriptor input_desc;
    ScopedTensorDescriptor output_desc;
    ScopedPoolingDescriptor pool_desc;
C
chengduoZH 已提交
121 122 123 124 125 126 127
    DataLayout layout;

    if (strides.size() == 2U) {
      layout = DataLayout::kNCHW;
    } else {
      layout = DataLayout::kNCDHW;
    }
C
chengduoZH 已提交
128

C
chengduoZH 已提交
129 130 131 132
    cudnnTensorDescriptor_t cudnn_input_desc = input_desc.descriptor<T>(
        layout, framework::vectorize2int(input->dims()));
    cudnnTensorDescriptor_t cudnn_output_desc = output_desc.descriptor<T>(
        layout, framework::vectorize2int(output->dims()));
C
chengduoZH 已提交
133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149

    PoolingMode pooling_mode;
    if (pooling_type == "max") {
      pooling_mode = PoolingMode::kMaximum;
    } else {
      pooling_mode = PoolingMode::kAverage;
    }

    cudnnPoolingDescriptor_t cudnn_pool_desc =
        pool_desc.descriptor(pooling_mode, ksize, paddings, strides);

    // ------------------- cudnn pool algorithm ---------------------
    auto handle = ctx.cuda_device_context().cudnn_handle();
    T alpha = 1.0f, beta = 0.0f;

    if (input_grad) {
      T *input_grad_data = input_grad->mutable_data<T>(ctx.GetPlace());
C
chengduoZH 已提交
150
      // Because beta is zero, it is unnecessary to reset input_grad.
C
chengduoZH 已提交
151 152 153

      PADDLE_ENFORCE(platform::dynload::cudnnPoolingBackward(
          handle, cudnn_pool_desc, &alpha, cudnn_output_desc, output_data,
154 155
          cudnn_output_desc, output_grad_data, cudnn_input_desc, input_data,
          &beta, cudnn_input_desc, input_grad_data));
C
chengduoZH 已提交
156 157 158 159 160 161 162 163 164 165 166
    }
  }
};

}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;

REGISTER_OP_GPU_KERNEL(pool2d_cudnn, ops::PoolCudnnOpKernel<float>);
REGISTER_OP_GPU_KERNEL(pool2d_cudnn_grad, ops::PoolCudnnGradOpKernel<float>);
C
chengduoZH 已提交
167 168

REGISTER_OP_GPU_KERNEL(pool3d_cudnn, ops::PoolCudnnOpKernel<float>);
C
chengduoZH 已提交
169
REGISTER_OP_GPU_KERNEL(pool3d_cudnn_grad, ops::PoolCudnnGradOpKernel<float>);
反馈
建议
客服 返回
顶部