maxouting.cu 5.7 KB
Newer Older
W
wanghaox 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21
/* 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/math/maxouting.h"
#include "paddle/platform/cuda_helper.h"

namespace paddle {
namespace operators {
namespace math {

W
wanghaox 已提交
22
template <typename T>
W
wanghaox 已提交
23
__global__ void KernelMaxOut(const int nthreads, const T* input_data,
W
wanghaox 已提交
24
                            const int channels,
W
wanghaox 已提交
25
                             const int input_height, const int input_width,
W
wanghaox 已提交
26
                             int groups, T* output_data ) {
W
wanghaox 已提交
27 28
  const int size = input_height * input_width * channels / groups;
  const int feat_len = input_height * input_width;
W
wanghaox 已提交
29 30 31 32 33
  int index = blockIdx.x * blockDim.x + threadIdx.x;
  int offset = blockDim.x * gridDim.x;
  for (int i = index; i < nthreads; i += offset) {
    int batch_idx = i / size;
    int batch_offset = i % size;
W
wanghaox 已提交
34 35
    int channel_idx = batch_offset / feat_len;
    int feat_idx = batch_offset % feat_len;
W
wanghaox 已提交
36
    int data_idx =
W
wanghaox 已提交
37
      (batch_idx * size + channel_idx * feat_len) * groups + feat_idx;
W
wanghaox 已提交
38
    T ele = static_cast<T>(-FLT_MAX);
W
wanghaox 已提交
39
    for (int g = 0; g < groups; ++g) {
W
wanghaox 已提交
40
      T x = input_data[data_idx + g * feat_len];
W
wanghaox 已提交
41
      ele = ele > x ? ele : x;
W
wanghaox 已提交
42
    }
W
wanghaox 已提交
43
    output_data[i] = ele;
W
wanghaox 已提交
44 45 46 47 48 49 50
  }
}
template <typename T>
__global__ void KernelMaxoutGrad(
    const int nthreads, const T* input_data, const T* output_data,
    const T* output_grad, T* input_grad, const int channels,
    const int input_height, const int input_width, int groups) {
W
wanghaox 已提交
51 52
    const int size = input_height * input_width * channels / groups;
    const int feat_len = input_height * input_width;
W
wanghaox 已提交
53 54 55 56 57
    int index = blockIdx.x * blockDim.x + threadIdx.x;
    int offset = blockDim.x * gridDim.x;
    for (int i = index; i < nthreads; i += offset) {
      int batch_idx = i / size;
      int batch_offset = i % size;
W
wanghaox 已提交
58 59
      int channel_idx = batch_offset / feat_len;
      int feat_idx = batch_offset % feat_len;
W
wanghaox 已提交
60
      int data_idx =
W
wanghaox 已提交
61
        (batch_idx * size + channel_idx * feat_len) * groups + feat_idx;
W
wanghaox 已提交
62 63 64 65 66 67
      int max_index = -1;
      bool continue_match = true;
      for (int g = 0; g < groups && continue_match; ++g) {
        if (input_data[data_idx + g * feat_len] == output_data[i]) {
          max_index = data_idx + g * feat_len;
          continue_match = false;
S
sweetsky0901 已提交
68
          break;
W
wanghaox 已提交
69 70
        }
      }
W
wanghaox 已提交
71
      if (max_index != -1) {
W
wanghaox 已提交
72
        // atomic add
W
wanghaox 已提交
73
        platform::CudaAtomicAdd(input_grad + max_index, output_grad[index]);
W
wanghaox 已提交
74 75 76 77 78 79
      }
    }
}
/*
 * All tensors are in NCHW format.
 */
W
wanghaox 已提交
80 81
template <typename T>
class MaxOutFunctor<platform::GPUPlace, T> {
W
wanghaox 已提交
82 83
 public:
  void operator()(const platform::DeviceContext& context,
W
wanghaox 已提交
84
                  const framework::Tensor& input, framework::Tensor * output,
W
wanghaox 已提交
85
                  int groups) {
W
wanghaox 已提交
86 87 88 89
    const int batch_size = input.dims()[0];
    const int input_channels = input.dims()[1];
    const int input_height = input.dims()[2];
    const int input_width = input.dims()[3];
W
wanghaox 已提交
90 91 92
    const int output_channels = output->dims()[1];
    const int output_height = output->dims()[2];
    const int output_width = output->dims()[3];
W
wanghaox 已提交
93 94

    const T* input_data = input.data<T>();
W
wanghaox 已提交
95 96
    T* output_data = output->mutable_data<T>(context.GetPlace());
    int nthreads =  output->numel();
W
wanghaox 已提交
97 98 99 100 101 102 103
    int blocks = (nthreads + 1024 - 1) / 1024;
    dim3 threads(1024, 1);
    dim3 grid(blocks, 1);

    KernelMaxOut<
        T><<<grid, threads, 0,
             reinterpret_cast<const platform::CUDADeviceContext&>(context)
W
wanghaox 已提交
104
                 .stream()>>>(nthreads, input_data, input_channels,
W
wanghaox 已提交
105
                              input_height, input_width, groups,
W
wanghaox 已提交
106
                              output_data);
W
wanghaox 已提交
107 108 109 110 111 112 113 114 115
  }
};
/*
 * All tensors are in NCHW format.
 */
template <typename T>
class MaxOutGradFunctor<platform::GPUPlace, T> {
 public:
  void operator()(const platform::DeviceContext& context,
W
wanghaox 已提交
116 117
                  const framework::Tensor& input,
                  framework::Tensor * input_grad,
W
wanghaox 已提交
118 119
                  const framework::Tensor& output,
                  const framework::Tensor& output_grad,
W
wanghaox 已提交
120
                  int groups) {
W
wanghaox 已提交
121 122 123 124 125 126 127 128 129 130 131
    const int batch_size = input.dims()[0];
    const int input_channels = input.dims()[1];
    const int input_height = input.dims()[2];
    const int input_width = input.dims()[3];
    const int output_channels = output.dims()[1];
    const int output_height = output.dims()[2];
    const int output_width = output.dims()[3];

    const T* input_data = input.data<T>();
    const T* output_data = output.data<T>();
    const T* output_grad_data = output_grad.data<T>();
W
wanghaox 已提交
132
    T* input_grad_data = input_grad->mutable_data<T>(context.GetPlace());
W
wanghaox 已提交
133
    int nthreads =  output.numel();
W
wanghaox 已提交
134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149
    int blocks = (nthreads + 1024 - 1) / 1024;
    dim3 threads(1024, 1);
    dim3 grid(blocks, 1);

    KernelMaxoutGrad<
        T><<<grid, threads, 0,
             reinterpret_cast<const platform::CUDADeviceContext&>(context)
                 .stream()>>>(
        nthreads, input_data, output_data, output_grad_data, input_grad_data,
        input_channels, input_height, input_width, groups);
  }
};

template class MaxOutGradFunctor<platform::GPUPlace, float>;
template class MaxOutGradFunctor<platform::GPUPlace, double>;

W
wanghaox 已提交
150 151
template class MaxOutFunctor<platform::GPUPlace, float>;
template class MaxOutFunctor<platform::GPUPlace, double>;
W
wanghaox 已提交
152 153 154 155

}  // namespace math
}  // namespace operators
}  // namespace paddle