• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /* Copyright 2015 The TensorFlow Authors. All Rights Reserved.
2 
3 Licensed under the Apache License, Version 2.0 (the "License");
4 you may not use this file except in compliance with the License.
5 You may obtain a copy of the License at
6 
7     http://www.apache.org/licenses/LICENSE-2.0
8 
9 Unless required by applicable law or agreed to in writing, software
10 distributed under the License is distributed on an "AS IS" BASIS,
11 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 See the License for the specific language governing permissions and
13 limitations under the License.
14 ==============================================================================*/
15 
16 #if GOOGLE_CUDA
17 
18 #define EIGEN_USE_GPU
19 
20 #include <stdio.h>
21 #include <iostream>
22 
23 #include "tensorflow/core/kernels/avgpooling_op.h"
24 
25 #include "tensorflow/core/framework/register_types.h"
26 #include "tensorflow/core/framework/tensor_types.h"
27 #include "tensorflow/core/util/cuda_kernel_helper.h"
28 
29 namespace tensorflow {
30 
31 typedef Eigen::GpuDevice GPUDevice;
32 
33 #define DEFINE_GPU_KERNELS(T) \
34   template struct functor::SpatialAvgPooling<GPUDevice, T>;
35 
36 DEFINE_GPU_KERNELS(Eigen::half)
DEFINE_GPU_KERNELS(float)37 DEFINE_GPU_KERNELS(float)
38 DEFINE_GPU_KERNELS(double)
39 
40 #undef DEFINE_GPU_KERNELS
41 
42 template <typename dtype>
43 __global__ void AvePoolBackwardNHWC(const int nthreads,
44                                     const dtype* const top_diff, const int num,
45                                     const int height, const int width,
46                                     const int channels, const int pooled_height,
47                                     const int pooled_width, const int kernel_h,
48                                     const int kernel_w, const int stride_h,
49                                     const int stride_w, const int pad_t,
50                                     const int pad_l, dtype* const bottom_diff) {
51   CUDA_1D_KERNEL_LOOP(index, nthreads) {
52     // find out the local index
53     // find out the local offset
54     const int c = index % channels;
55     const int w = index / channels % width + pad_l;
56     const int h = (index / channels / width) % height + pad_t;
57     const int n = index / channels / width / height;
58     const int phstart = (h < kernel_h) ? 0 : (h - kernel_h) / stride_h + 1;
59     const int phend = min(h / stride_h + 1, pooled_height);
60     const int pwstart = (w < kernel_w) ? 0 : (w - kernel_w) / stride_w + 1;
61     const int pwend = min(w / stride_w + 1, pooled_width);
62     dtype gradient(0);
63     const dtype* const top_diff_slice =
64         top_diff + n * pooled_height * pooled_width * channels + c;
65     for (int ph = phstart; ph < phend; ++ph) {
66       for (int pw = pwstart; pw < pwend; ++pw) {
67         // figure out the pooling size
68         int hstart = ph * stride_h - pad_t;
69         int wstart = pw * stride_w - pad_l;
70         int hend = min(hstart + kernel_h, height);
71         int wend = min(wstart + kernel_w, width);
72         hstart = max(hstart, 0);
73         wstart = max(wstart, 0);
74         int pool_size = (hend - hstart) * (wend - wstart);
75         gradient += top_diff_slice[(ph * pooled_width + pw) * channels] /
76                     dtype(pool_size);
77       }
78     }
79     bottom_diff[index] = gradient;
80   }
81 }
82 
83 template <typename T>
RunAvePoolBackwardNHWC(const T * const top_diff,const int num,const int height,const int width,const int channels,const int pooled_height,const int pooled_width,const int kernel_h,const int kernel_w,const int stride_h,const int stride_w,const int pad_t,const int pad_l,T * const bottom_diff,const GPUDevice & d)84 bool RunAvePoolBackwardNHWC(const T* const top_diff, const int num,
85                             const int height, const int width,
86                             const int channels, const int pooled_height,
87                             const int pooled_width, const int kernel_h,
88                             const int kernel_w, const int stride_h,
89                             const int stride_w, const int pad_t,
90                             const int pad_l, T* const bottom_diff,
91                             const GPUDevice& d) {
92   int x_size = num * height * width * channels;
93   CudaLaunchConfig config = GetCudaLaunchConfig(x_size, d);
94   TF_CHECK_OK(CudaLaunchKernel(
95       AvePoolBackwardNHWC<T>, config.block_count, config.thread_per_block, 0,
96       d.stream(), config.virtual_thread_count, top_diff, num, height, width,
97       channels, pooled_height, pooled_width, kernel_h, kernel_w, stride_h,
98       stride_w, pad_t, pad_t, bottom_diff));
99 
100   return d.ok();
101 }
102 
103 template bool RunAvePoolBackwardNHWC(
104     const double* const top_diff, const int num, const int height,
105     const int width, const int channels, const int pooled_height,
106     const int pooled_width, const int kernel_h, const int kernel_w,
107     const int stride_h, const int stride_w, const int pad_t, const int pad_l,
108     double* const bottom_diff, const GPUDevice& d);
109 template bool RunAvePoolBackwardNHWC(
110     const float* const top_diff, const int num, const int height,
111     const int width, const int channels, const int pooled_height,
112     const int pooled_width, const int kernel_h, const int kernel_w,
113     const int stride_h, const int stride_w, const int pad_t, const int pad_l,
114     float* const bottom_diff, const GPUDevice& d);
115 template bool RunAvePoolBackwardNHWC(
116     const Eigen::half* const top_diff, const int num, const int height,
117     const int width, const int channels, const int pooled_height,
118     const int pooled_width, const int kernel_h, const int kernel_w,
119     const int stride_h, const int stride_w, const int pad_t, const int pad_l,
120     Eigen::half* const bottom_diff, const GPUDevice& d);
121 
122 }  // end namespace tensorflow
123 
124 #endif  // GOOGLE_CUDA
125