File size: 7,855 Bytes
c310e19
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
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
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
// Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
#include <ATen/ATen.h>
#include <ATen/cuda/CUDAContext.h>

#include <THC/THC.h>
#include <THC/THCAtomics.cuh>
#include <THC/THCDeviceUtils.cuh>


// TODO make it in a common file
#define CUDA_1D_KERNEL_LOOP(i, n)                            \
  for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < n; \
       i += blockDim.x * gridDim.x)


template <typename T>
__global__ void RoIPoolFForward(const int nthreads, const T* bottom_data,
    const T spatial_scale, const int channels, const int height,
    const int width, const int pooled_height, const int pooled_width,
    const T* bottom_rois, T* top_data, int* argmax_data) {
  CUDA_1D_KERNEL_LOOP(index, nthreads) {
    // (n, c, ph, pw) is an element in the pooled output
    int pw = index % pooled_width;
    int ph = (index / pooled_width) % pooled_height;
    int c = (index / pooled_width / pooled_height) % channels;
    int n = index / pooled_width / pooled_height / channels;

    const T* offset_bottom_rois = bottom_rois + n * 5;
    int roi_batch_ind = offset_bottom_rois[0];
    int roi_start_w = round(offset_bottom_rois[1] * spatial_scale);
    int roi_start_h = round(offset_bottom_rois[2] * spatial_scale);
    int roi_end_w = round(offset_bottom_rois[3] * spatial_scale);
    int roi_end_h = round(offset_bottom_rois[4] * spatial_scale);

    // Force malformed ROIs to be 1x1
    int roi_width = max(roi_end_w - roi_start_w + 1, 1);
    int roi_height = max(roi_end_h - roi_start_h + 1, 1);
    T bin_size_h = static_cast<T>(roi_height)
                       / static_cast<T>(pooled_height);
    T bin_size_w = static_cast<T>(roi_width)
                       / static_cast<T>(pooled_width);

    int hstart = static_cast<int>(floor(static_cast<T>(ph)
                                        * bin_size_h));
    int wstart = static_cast<int>(floor(static_cast<T>(pw)
                                        * bin_size_w));
    int hend = static_cast<int>(ceil(static_cast<T>(ph + 1)
                                     * bin_size_h));
    int wend = static_cast<int>(ceil(static_cast<T>(pw + 1)
                                     * bin_size_w));

    // Add roi offsets and clip to input boundaries
    hstart = min(max(hstart + roi_start_h, 0), height);
    hend = min(max(hend + roi_start_h, 0), height);
    wstart = min(max(wstart + roi_start_w, 0), width);
    wend = min(max(wend + roi_start_w, 0), width);
    bool is_empty = (hend <= hstart) || (wend <= wstart);

    // Define an empty pooling region to be zero
    T maxval = is_empty ? 0 : -FLT_MAX;
    // If nothing is pooled, argmax = -1 causes nothing to be backprop'd
    int maxidx = -1;
    const T* offset_bottom_data =
        bottom_data + (roi_batch_ind * channels + c) * height * width;
    for (int h = hstart; h < hend; ++h) {
      for (int w = wstart; w < wend; ++w) {
        int bottom_index = h * width + w;
        if (offset_bottom_data[bottom_index] > maxval) {
          maxval = offset_bottom_data[bottom_index];
          maxidx = bottom_index;
        }
      }
    }
    top_data[index] = maxval;
    argmax_data[index] = maxidx;
  }
}

template <typename T>
__global__ void RoIPoolFBackward(const int nthreads, const T* top_diff,
    const int* argmax_data, const int num_rois, const T spatial_scale,
    const int channels, const int height, const int width,
    const int pooled_height, const int pooled_width, T* bottom_diff,
    const T* bottom_rois) {
  CUDA_1D_KERNEL_LOOP(index, nthreads) {
    // (n, c, ph, pw) is an element in the pooled output
    int pw = index % pooled_width;
    int ph = (index / pooled_width) % pooled_height;
    int c = (index / pooled_width / pooled_height) % channels;
    int n = index / pooled_width / pooled_height / channels;

    const T* offset_bottom_rois = bottom_rois + n * 5;
    int roi_batch_ind = offset_bottom_rois[0];
    int bottom_offset = (roi_batch_ind * channels + c) * height * width;
    int top_offset    = (n * channels + c) * pooled_height * pooled_width;
    const T* offset_top_diff = top_diff + top_offset;
    T* offset_bottom_diff = bottom_diff + bottom_offset;
    const int* offset_argmax_data = argmax_data + top_offset;

    int argmax = offset_argmax_data[ph * pooled_width + pw];
    if (argmax != -1) {
      atomicAdd(
          offset_bottom_diff + argmax,
          static_cast<T>(offset_top_diff[ph * pooled_width + pw]));

    }
  }
}

std::tuple<at::Tensor, at::Tensor> ROIPool_forward_cuda(const at::Tensor& input,
                                const at::Tensor& rois,
                                const float spatial_scale,
                                const int pooled_height,
                                const int pooled_width) {
  AT_ASSERTM(input.type().is_cuda(), "input must be a CUDA tensor");
  AT_ASSERTM(rois.type().is_cuda(), "rois must be a CUDA tensor");

  auto num_rois = rois.size(0);
  auto channels = input.size(1);
  auto height = input.size(2);
  auto width = input.size(3);

  auto output = at::empty({num_rois, channels, pooled_height, pooled_width}, input.options());
  auto output_size = num_rois * pooled_height * pooled_width * channels;
  auto argmax = at::zeros({num_rois, channels, pooled_height, pooled_width}, input.options().dtype(at::kInt));

  cudaStream_t stream = at::cuda::getCurrentCUDAStream();

  dim3 grid(std::min(THCCeilDiv((long)output_size, 512L), 4096L));
  dim3 block(512);

  if (output.numel() == 0) {
    THCudaCheck(cudaGetLastError());
    return std::make_tuple(output, argmax);
  }

  AT_DISPATCH_FLOATING_TYPES(input.type(), "ROIPool_forward", [&] {
    RoIPoolFForward<scalar_t><<<grid, block, 0, stream>>>(
         output_size,
         input.contiguous().data<scalar_t>(),
         spatial_scale,
         channels,
         height,
         width,
         pooled_height,
         pooled_width,
         rois.contiguous().data<scalar_t>(),
         output.data<scalar_t>(),
         argmax.data<int>());
  });
  THCudaCheck(cudaGetLastError());
  return std::make_tuple(output, argmax);
}

// TODO remove the dependency on input and use instead its sizes -> save memory
at::Tensor ROIPool_backward_cuda(const at::Tensor& grad,
                                 const at::Tensor& input,
                                 const at::Tensor& rois,
                                 const at::Tensor& argmax,
                                 const float spatial_scale,
                                 const int pooled_height,
                                 const int pooled_width,
                                 const int batch_size,
                                 const int channels,
                                 const int height,
                                 const int width) {
  AT_ASSERTM(grad.type().is_cuda(), "grad must be a CUDA tensor");
  AT_ASSERTM(rois.type().is_cuda(), "rois must be a CUDA tensor");
  // TODO add more checks

  auto num_rois = rois.size(0);
  auto grad_input = at::zeros({batch_size, channels, height, width}, grad.options());

  cudaStream_t stream = at::cuda::getCurrentCUDAStream();

  dim3 grid(std::min(THCCeilDiv((long)grad.numel(), 512L), 4096L));
  dim3 block(512);

  // handle possibly empty gradients
  if (grad.numel() == 0) {
    THCudaCheck(cudaGetLastError());
    return grad_input;
  }

  AT_DISPATCH_FLOATING_TYPES(grad.type(), "ROIPool_backward", [&] {
    RoIPoolFBackward<scalar_t><<<grid, block, 0, stream>>>(
         grad.numel(),
         grad.contiguous().data<scalar_t>(),
         argmax.data<int>(),
         num_rois,
         spatial_scale,
         channels,
         height,
         width,
         pooled_height,
         pooled_width,
         grad_input.data<scalar_t>(),
         rois.contiguous().data<scalar_t>());
  });
  THCudaCheck(cudaGetLastError());
  return grad_input;
}