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
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
|
// TODO: reduce the apparent redundancy of all the code below.
#include <cfloat>
#include "caffe2/core/context_gpu.h"
#include "caffe2/operators/pool_op.h"
namespace caffe2 {
namespace {
struct LpPoolFunctor {
explicit LpPoolFunctor(const OperatorBase& /* op */) {}
};
} // namespace
namespace {
using c10::cuda::compat::abs;
using c10::cuda::compat::pow;
template <typename T>
__global__ void LpPoolForwardNCHW(
const int nthreads,
const T *const bottom_data,
const int channels,
const int height,
const int width,
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 top_data,
const T p) {
CUDA_1D_KERNEL_LOOP(index, nthreads) {
int n = index;
int pw = n % pooled_width;
n /= pooled_width;
int ph = n % pooled_height;
n /= pooled_height;
int c = n % channels;
n /= channels;
int hstart = ph * stride_h - pad_t;
int wstart = pw * stride_w - pad_l;
int hend = min(hstart + kernel_h, height);
int wend = min(wstart + kernel_w, width);
hstart = max(hstart, 0);
wstart = max(wstart, 0);
top_data[index] = 0;
int bottom_offset = (n * channels + c) * height * width;
for (int h = hstart; h < hend; ++h) {
for (int w = wstart; w < wend; ++w) {
top_data[index] +=
pow(abs(bottom_data[bottom_offset + h * width + w]), p);
}
}
top_data[index] = pow(top_data[index], static_cast<T>(1.0) / p);
}
}
template <typename T>
__global__ void LpPoolForwardNHWC(
const int nthreads,
const T *const bottom_data,
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 top_data,
const T p) {
CUDA_1D_KERNEL_LOOP(index, nthreads) {
int c = index % channels;
int pw = (index / channels) % pooled_width;
int ph = (index / channels / pooled_width) % pooled_height;
int n = index / channels / pooled_width / pooled_height;
int hstart = ph * stride_h - pad_t;
int wstart = pw * stride_w - pad_l;
int hend = min(hstart + kernel_h, height);
int wend = min(wstart + kernel_w, width);
hstart = max(hstart, 0);
wstart = max(wstart, 0);
T output = 0;
int bottom_offset = n * height * width * channels + c;
for (int h = hstart; h < hend; ++h) {
for (int w = wstart; w < wend; ++w) {
output += pow(
abs(bottom_data[bottom_offset + (h * width + w) * channels]), p);
}
}
top_data[index] = pow(output, static_cast<T>(1.0) / p);
}
}
template <typename T>
__global__ void LpPoolBackwardNCHW(
const int nthreads,
const T* const top_diff,
const T* const top_data,
const T* const bottom_data,
const int channels,
const int height,
const int width,
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 int p) {
CUDA_1D_KERNEL_LOOP(index, nthreads) {
// find out the local index
// find out the local offset
const int w = index % width + pad_l;
const int h = (index / width) % height + pad_t;
const int c = (index / width / height) % channels;
const int n = index / width / height / channels;
const int phstart = (h < kernel_h) ? 0 : (h - kernel_h) / stride_h + 1;
const int phend = min(h / stride_h + 1, pooled_height);
const int pwstart = (w < kernel_w) ? 0 : (w - kernel_w) / stride_w + 1;
const int pwend = min(w / stride_w + 1, pooled_width);
T gradient = 0;
const T* const top_diff_slice =
top_diff + (n * channels + c) * pooled_height * pooled_width;
const T* const top_data_slice =
top_data + (n * channels + c) * pooled_height * pooled_width;
for (int ph = phstart; ph < phend; ++ph) {
for (int pw = pwstart; pw < pwend; ++pw) {
// figure out the pooling size
int hstart = ph * stride_h - pad_t;
int wstart = pw * stride_w - pad_l;
hstart = max(hstart, 0);
wstart = max(wstart, 0);
gradient += top_diff_slice[ph * pooled_width + pw] *
bottom_data[index] * pow(abs(bottom_data[index]), p - 2) /
pow(top_data_slice[ph * pooled_width + pw], p - 1);
}
}
bottom_diff[index] = gradient;
}
}
template <typename T>
__global__ void LpPoolBackwardNHWC(
const int nthreads,
const T* const top_diff,
const T* const top_data,
const T* const bottom_data,
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 T p) {
CUDA_1D_KERNEL_LOOP(index, nthreads) {
// find out the local index
// find out the local offset
const int c = index % channels;
const int w = index / channels % width + pad_l;
const int h = (index / channels / width) % height + pad_t;
const int n = index / channels / width / height;
const int phstart = (h < kernel_h) ? 0 : (h - kernel_h) / stride_h + 1;
const int phend = min(h / stride_h + 1, pooled_height);
const int pwstart = (w < kernel_w) ? 0 : (w - kernel_w) / stride_w + 1;
const int pwend = min(w / stride_w + 1, pooled_width);
T gradient = 0;
const T* const top_diff_slice =
top_diff + n * pooled_height * pooled_width * channels + c;
const T* const top_data_slice =
top_data + n * pooled_height * pooled_width * channels + c;
for (int ph = phstart; ph < phend; ++ph) {
for (int pw = pwstart; pw < pwend; ++pw) {
// figure out the pooling size
gradient += top_diff_slice[(ph * pooled_width + pw) * channels] *
bottom_data[index] * pow(abs(bottom_data[index]), p - 2) /
pow(top_data_slice[(ph * pooled_width + pw) * channels], p - 1);
}
}
bottom_diff[index] = gradient;
}
}
} // namespace
template <>
bool PoolOp<float, CUDAContext, LpPoolFunctor>::RunOnDeviceWithOrderNCHW() {
auto& X = Input(0);
auto* Y = Output(0);
ConvPoolOpBase<CUDAContext>::SetOutputSize(X, Y, X.dim32(1));
int output_size = Y->numel();
LpPoolForwardNCHW<float>
<<<CAFFE_GET_BLOCKS(output_size),
CAFFE_CUDA_NUM_THREADS,
0,
context_.cuda_stream()>>>(
output_size,
X.data<float>(),
X.dim32(1),
X.dim32(2),
X.dim32(3),
Y->dim32(2),
Y->dim32(3),
kernel_h(),
kernel_w(),
stride_h(),
stride_w(),
pad_t(),
pad_l(),
Y->template mutable_data<float>(),
OperatorBase::GetSingleArgument<float>("p", 2.0));
C10_CUDA_KERNEL_LAUNCH_CHECK();
return true;
}
template <>
bool PoolOp<float, CUDAContext, LpPoolFunctor>::RunOnDeviceWithOrderNHWC() {
auto& X = Input(0);
auto* Y = Output(0);
ConvPoolOpBase<CUDAContext>::SetOutputSize(X, Y, X.dim32(3));
int output_size = Y->numel();
LpPoolForwardNHWC<float>
<<<CAFFE_GET_BLOCKS(output_size),
CAFFE_CUDA_NUM_THREADS,
0,
context_.cuda_stream()>>>(
output_size,
X.data<float>(),
X.dim32(1),
X.dim32(2),
X.dim32(3),
Y->dim32(1),
Y->dim32(2),
kernel_h(),
kernel_w(),
stride_h(),
stride_w(),
pad_t(),
pad_l(),
Y->template mutable_data<float>(),
OperatorBase::GetSingleArgument<float>("p", 2.0));
C10_CUDA_KERNEL_LAUNCH_CHECK();
return true;
}
template <>
bool PoolGradientOp<float, CUDAContext, LpPoolFunctor>::
RunOnDeviceWithOrderNCHW() {
auto& X = Input(0);
auto& Y = Input(1);
auto& dY = Input(2);
CAFFE_ENFORCE_EQ(dY.dim(), 4);
auto* dX = Output(0, X.sizes(), at::dtype<float>());
ConvPoolOpBase<CUDAContext>::ComputePads({X.dim32(2), X.dim32(3)});
LpPoolBackwardNCHW<float>
<<<CAFFE_GET_BLOCKS(X.numel()),
CAFFE_CUDA_NUM_THREADS,
0,
context_.cuda_stream()>>>(
X.numel(),
dY.data<float>(),
Y.data<float>(),
X.data<float>(),
X.dim32(1),
X.dim32(2),
X.dim32(3),
dY.dim32(2),
dY.dim32(3),
kernel_h(),
kernel_w(),
stride_h(),
stride_w(),
pad_t(),
pad_l(),
dX->template mutable_data<float>(),
OperatorBase::GetSingleArgument<float>("p", 2.0));
C10_CUDA_KERNEL_LAUNCH_CHECK();
return true;
}
template <>
bool PoolGradientOp<float, CUDAContext, LpPoolFunctor>::
RunOnDeviceWithOrderNHWC() {
auto& X = Input(0);
auto& Y = Input(1);
auto& dY = Input(2);
CAFFE_ENFORCE_EQ(dY.dim(), 4);
auto* dX = Output(0, X.sizes(), at::dtype<float>());
ConvPoolOpBase<CUDAContext>::ComputePads({X.dim32(1), X.dim32(2)});
LpPoolBackwardNHWC<float>
<<<CAFFE_GET_BLOCKS(X.numel()),
CAFFE_CUDA_NUM_THREADS,
0,
context_.cuda_stream()>>>(
X.numel(),
dY.data<float>(),
Y.data<float>(),
X.data<float>(),
X.dim32(1),
X.dim32(2),
X.dim32(3),
dY.dim32(1),
dY.dim32(2),
kernel_h(),
kernel_w(),
stride_h(),
stride_w(),
pad_t(),
pad_l(),
dX->template mutable_data<float>(),
OperatorBase::GetSingleArgument<float>("p", 2.0));
C10_CUDA_KERNEL_LAUNCH_CHECK();
return true;
}
REGISTER_CUDA_OPERATOR(LpPool, PoolOp<float, CUDAContext, LpPoolFunctor>);
REGISTER_CUDA_OPERATOR(
LpPoolGradient,
PoolGradientOp<float, CUDAContext, LpPoolFunctor>);
}
|
