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
|
#include <vector>
#include "caffe/layer.hpp"
#include "caffe/util/math_functions.hpp"
#include "caffe/vision_layers.hpp"
namespace caffe {
template <typename Dtype>
__global__ void LRNFillScale(const int nthreads, const Dtype* in,
const int num, const int channels, const int height,
const int width, const int size, const Dtype alpha_over_size,
Dtype* scale) {
CUDA_KERNEL_LOOP(index, nthreads) {
// find out the local offset
int w = index % width;
int h = (index / width) % height;
int n = index / width / height;
int offset = (n * channels * height + h) * width + w;
int step = height * width;
in += offset;
scale += offset;
int head = 0;
int pre_pad = (size - 1) / 2;
int post_pad = size - pre_pad - 1;
Dtype accum_scale = 0;
// fill the scale at [n, :, h, w]
// accumulate values
while (head < post_pad) {
accum_scale += in[head * step] * in[head * step];
++head;
}
// until we reach size, nothing needs to be subtracted
while (head < size) {
accum_scale += in[head * step] * in[head * step];
scale[(head - post_pad) * step] = 1. + accum_scale * alpha_over_size;
++head;
}
// both add and subtract
while (head < channels) {
accum_scale += in[head * step] * in[head * step];
accum_scale -= in[(head - size) * step] * in[(head - size) * step];
scale[(head - post_pad) * step] = 1. + accum_scale * alpha_over_size;
++head;
}
// subtract only
while (head < channels + post_pad) {
accum_scale -= in[(head - size) * step] * in[(head - size) * step];
scale[(head - post_pad) * step] = 1. + accum_scale * alpha_over_size;
++head;
}
}
}
template <typename Dtype>
Dtype LRNLayer<Dtype>::Forward_gpu(const vector<Blob<Dtype>*>& bottom,
vector<Blob<Dtype>*>* top) {
switch (this->layer_param_.lrn_param().norm_region()) {
case LRNParameter_NormRegion_ACROSS_CHANNELS:
return CrossChannelForward_gpu(bottom, top);
case LRNParameter_NormRegion_WITHIN_CHANNEL:
return WithinChannelForward(bottom, top);
default:
LOG(FATAL) << "Unknown normalization region.";
return Dtype(0);
}
}
// TODO: check if it would be faster to just put it into the previous kernel.
template <typename Dtype>
__global__ void LRNComputeOutput(const int nthreads, const Dtype* in,
const Dtype* scale, const Dtype negative_beta, Dtype* out) {
CUDA_KERNEL_LOOP(index, nthreads) {
out[index] = in[index] * pow(scale[index], negative_beta);
}
}
template <typename Dtype>
Dtype LRNLayer<Dtype>::CrossChannelForward_gpu(
const vector<Blob<Dtype>*>& bottom, vector<Blob<Dtype>*>* top) {
// First, compute scale
const Dtype* bottom_data = bottom[0]->gpu_data();
Dtype* top_data = (*top)[0]->mutable_gpu_data();
Dtype* scale_data = scale_.mutable_gpu_data();
// We will launch one kernel for each pixel location, and have the kernel
// go through all the channels.
int n_threads = num_ * height_ * width_;
// NOLINT_NEXT_LINE(whitespace/operators)
LRNFillScale<<<CAFFE_GET_BLOCKS(n_threads), CAFFE_CUDA_NUM_THREADS>>>(
n_threads, bottom_data, num_, channels_, height_, width_, size_,
alpha_ / size_, scale_data);
CUDA_POST_KERNEL_CHECK;
n_threads = bottom[0]->count();
// NOLINT_NEXT_LINE(whitespace/operators)
LRNComputeOutput<<<CAFFE_GET_BLOCKS(n_threads), CAFFE_CUDA_NUM_THREADS>>>(
n_threads, bottom_data, scale_data, -beta_, top_data);
CUDA_POST_KERNEL_CHECK;
return Dtype(0.);
}
template <typename Dtype>
void LRNLayer<Dtype>::Backward_gpu(const vector<Blob<Dtype>*>& top,
const vector<bool>& propagate_down, vector<Blob<Dtype>*>* bottom) {
switch (this->layer_param_.lrn_param().norm_region()) {
case LRNParameter_NormRegion_ACROSS_CHANNELS:
CrossChannelBackward_gpu(top, propagate_down, bottom);
break;
case LRNParameter_NormRegion_WITHIN_CHANNEL:
WithinChannelBackward(top, propagate_down, bottom);
break;
default:
LOG(FATAL) << "Unknown normalization region.";
}
}
template <typename Dtype>
__global__ void LRNComputeDiff(const int nthreads, const Dtype* bottom_data,
const Dtype* top_data, const Dtype* scale, const Dtype* top_diff,
const int num, const int channels, const int height,
const int width, const int size, const Dtype negative_beta,
const Dtype cache_ratio,
Dtype* bottom_diff) {
CUDA_KERNEL_LOOP(index, nthreads) {
// find out the local offset
int w = index % width;
int h = (index / width) % height;
int n = index / width / height;
int offset = (n * channels * height + h) * width + w;
int step = height * width;
bottom_data += offset;
top_data += offset;
scale += offset;
top_diff += offset;
bottom_diff += offset;
int head = 0;
int pre_pad = size - (size + 1) / 2;
int post_pad = size - pre_pad - 1;
Dtype accum_ratio = 0;
// accumulate values
while (head < post_pad) {
accum_ratio += top_diff[head * step] * top_data[head * step] /
scale[head * step];
++head;
}
// until we reach size, nothing needs to be subtracted
while (head < size) {
accum_ratio += top_diff[head * step] * top_data[head * step] /
scale[head * step];
bottom_diff[(head - post_pad) * step] = top_diff[(head - post_pad) * step]
* pow(scale[(head - post_pad) * step], negative_beta) - cache_ratio *
bottom_data[(head - post_pad) * step] * accum_ratio;
++head;
}
// both add and subtract
while (head < channels) {
accum_ratio += top_diff[head * step] * top_data[head * step] /
scale[head * step];
accum_ratio -= top_diff[(head - size) * step] *
top_data[(head - size) * step] / scale[(head - size) * step];
bottom_diff[(head - post_pad) * step] = top_diff[(head - post_pad) * step]
* pow(scale[(head - post_pad) * step], negative_beta) - cache_ratio *
bottom_data[(head - post_pad) * step] * accum_ratio;
++head;
}
// subtract only
while (head < channels + post_pad) {
accum_ratio -= top_diff[(head - size) * step] *
top_data[(head - size) * step] / scale[(head - size) * step];
bottom_diff[(head - post_pad) * step] = top_diff[(head - post_pad) * step]
* pow(scale[(head - post_pad) * step], negative_beta) - cache_ratio *
bottom_data[(head - post_pad) * step] * accum_ratio;
++head;
}
}
}
template <typename Dtype>
void LRNLayer<Dtype>::CrossChannelBackward_gpu(
const vector<Blob<Dtype>*>& top, const vector<bool>& propagate_down,
vector<Blob<Dtype>*>* bottom) {
int n_threads = num_ * height_ * width_;
// NOLINT_NEXT_LINE(whitespace/operators)
LRNComputeDiff<<<CAFFE_GET_BLOCKS(n_threads), CAFFE_CUDA_NUM_THREADS>>>(
n_threads, (*bottom)[0]->gpu_data(), top[0]->gpu_data(),
scale_.gpu_data(), top[0]->gpu_diff(), num_, channels_, height_, width_,
size_, -beta_, Dtype(2. * alpha_ * beta_ / size_),
(*bottom)[0]->mutable_gpu_diff());
}
INSTANTIATE_CLASS(LRNLayer);
} // namespace caffe
|
