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
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
|
/*******************************************************************************
* Copyright 2016-2018 Intel Corporation
*
* 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 <assert.h>
#include <math.h>
#include "c_types_map.hpp"
#include "math_utils.hpp"
#include "mkldnn_thread.hpp"
#include "nstl.hpp"
#include "type_helpers.hpp"
#include "ref_pooling.hpp"
namespace mkldnn {
namespace impl {
namespace cpu {
template <data_type_t data_type, data_type_t acc_type>
void ref_pooling_fwd_t<data_type, acc_type>::execute_forward() {
using namespace alg_kind;
using namespace prop_kind;
auto alg = conf_.desc()->alg_kind;
auto src = reinterpret_cast<const data_t *>(this->input_memory(0));
auto dst = reinterpret_cast<data_t *>(this->memory(0));
auto ws = alg == pooling_max && conf_.desc()->prop_kind == forward_training
? reinterpret_cast<unsigned char *>(this->memory(1)) : nullptr;
const memory_desc_wrapper src_d(conf_.src_pd());
const memory_desc_wrapper dst_d(conf_.dst_pd());
const memory_desc_wrapper ws_d(conf_.workspace_pd());
const data_type_t ws_dt = ws ? ws_d.data_type() : data_type::undef;
const int ID = conf_.ID();
const int IH = conf_.IH();
const int IW = conf_.IW();
const int KD = conf_.KD();
const int KH = conf_.KH();
const int KW = conf_.KW();
const int SD = conf_.KSD();
const int SH = conf_.KSH();
const int SW = conf_.KSW();
const int padF = conf_.padFront();
const int padT = conf_.padT();
const int padL = conf_.padL();
const int padB = conf_.padB();
const int padR = conf_.padR();
const bool is_3d = conf_.desc()->src_desc.ndims == 5;
auto apply_offset = [=](int index, int offset) {
return (index > offset) ? index - offset : 0;
};
auto set_ws = [=](int mb, int oc, int od, int oh, int ow, int value) {
if (ws) {
assert(ws_dt == data_type::u8 || ws_dt == data_type::s32);
size_t offset = is_3d
? ws_d.off(mb, oc, od, oh, ow) : ws_d.off(mb, oc, oh, ow);;
if (ws_dt == data_type::u8) {
assert(0 <= value && value <= 255);
ws[offset] = value;
} else
reinterpret_cast<int *>(ws)[offset] = value;
}
};
auto ker_max = [=](data_t *d, int mb, int oc, int oh, int ow) {
for (int kh = 0; kh < KH; ++kh) {
for (int kw = 0; kw < KW; ++kw) {
const int ih = oh * SH - padT + kh;
const int iw = ow * SW - padL + kw;
if (ih < 0 || ih >= IH) continue;
if (iw < 0 || iw >= IW) continue;
auto s = src[src_d.off(mb, oc, ih, iw)];
if (s > d[0]) {
d[0] = s;
set_ws(mb, oc, 1, oh, ow, kh*KW + kw);
}
}
}
};
auto ker_avg = [=](data_t *d, int mb, int oc, int oh, int ow) {
auto ih_start = oh*SH - padT;
auto iw_start = ow*SW - padL;
auto ih_end = nstl::min(oh*SH - padT + KH, IH + padB);
auto iw_end = nstl::min(ow*SW - padL + KW, IW + padR);
// case alg == pooling_avg_include_padding
auto num_summands = (ih_end - ih_start)*(iw_end - iw_start);
ih_start = nstl::max(ih_start, 0);
iw_start = nstl::max(iw_start, 0);
ih_end = nstl::min(ih_end, IH);
iw_end = nstl::min(iw_end, IW);
if (alg == pooling_avg_exclude_padding)
num_summands = (ih_end - ih_start)*(iw_end - iw_start);
acc_data_t dst = 0;
for (int ih = ih_start; ih < ih_end; ++ih) {
for (int iw = iw_start; iw < iw_end; ++iw) {
dst += src[src_d.off(mb, oc, ih, iw)];
}
}
d[0] = math::out_round<data_t>((float)dst / num_summands);
};
auto ker_max_3d = [=](data_t *d, int mb, int oc, int od, int oh, int ow) {
for (int kd = 0; kd < KD; ++kd) {
for (int kh = 0; kh < KH; ++kh) {
for (int kw = 0; kw < KW; ++kw) {
const int id = od * SD - padF + kd;
const int ih = oh * SH - padT + kh;
const int iw = ow * SW - padL + kw;
if (id < 0 || id >= ID) continue;
if (ih < 0 || ih >= IH) continue;
if (iw < 0 || iw >= IW) continue;
auto s = src[src_d.off(mb, oc, id, ih, iw)];
if (s > d[0]) {
d[0] = s;
set_ws(mb, oc, od, oh, ow, kd * KH * KW + kh*KW + kw);
}
}
}
}
};
auto ker_avg_3d = [=](data_t *d, int mb, int oc, int od, int oh, int ow) {
auto id_start = apply_offset(od*SD, padF);
auto ih_start = apply_offset(oh*SH, padT);
auto iw_start = apply_offset(ow*SW, padL);
auto id_end = nstl::min(od*SD - padF + KD, ID);
auto ih_end = nstl::min(oh*SH - padT + KH, IH);
auto iw_end = nstl::min(ow*SW - padL + KW, IW);
auto num_summands = (alg == pooling_avg_include_padding) ? KW*KH*KD
: (ih_end - ih_start)*(iw_end - iw_start)*(id_end - id_start);
acc_data_t dst = 0;
for (int id = id_start; id < id_end; ++id) {
for (int ih = ih_start; ih < ih_end; ++ih) {
for (int iw = iw_start; iw < iw_end; ++iw) {
dst += src[src_d.off(mb, oc, id, ih, iw)];
}
}
}
d[0] = math::out_round<data_t>((float)dst / num_summands);
};
const int MB = conf_.MB();
const int OC = conf_.C();
const int OD = conf_.OD();
const int OH = conf_.OH();
const int OW = conf_.OW();
if (alg == pooling_max) {
# pragma omp parallel for collapse(5) schedule(static)
for (int mb = 0; mb < MB; ++mb) {
for (int oc = 0; oc < OC; ++oc) {
for (int od = 0; od < OD; ++od)
for (int oh = 0; oh < OH; ++oh)
for (int ow = 0; ow < OW; ++ow) {
data_t *d = is_3d
? &dst[dst_d.off(mb, oc, od, oh, ow)]
: &dst[dst_d.off(mb, oc, oh, ow)];
d[0] = nstl::numeric_limits<data_t>::lowest();
set_ws(mb, oc, od, oh, ow, 0);
if (is_3d) ker_max_3d(d, mb, oc, od, oh, ow);
else ker_max(d, mb, oc, oh, ow);
}
}
}
} else {
# pragma omp parallel for collapse(5) schedule(static)
for (int mb = 0; mb < MB; ++mb) {
for (int oc = 0; oc < OC; ++oc) {
for (int od = 0; od < OD; ++od)
for (int oh = 0; oh < OH; ++oh)
for (int ow = 0; ow < OW; ++ow) {
data_t *d = is_3d
? &dst[dst_d.off(mb, oc, od, oh, ow)]
: &dst[dst_d.off(mb, oc, oh, ow)];
d[0] = 0;
if (is_3d) ker_avg_3d(d, mb, oc, od, oh, ow);
else ker_avg(d, mb, oc, oh, ow);
}
}
}
}
}
template <data_type_t data_type, data_type_t acc_type>
void ref_pooling_bwd_t<data_type, acc_type>::execute_backward() {
using namespace alg_kind;
auto diff_dst = reinterpret_cast<const data_t *>(this->input_memory(0));
auto ws = conf_.desc()->alg_kind != alg_kind::pooling_max ? nullptr
: reinterpret_cast<const unsigned char *>(this->input_memory(1));
auto diff_src = reinterpret_cast<data_t *>(this->memory(0));
const memory_desc_wrapper diff_dst_d(conf_.diff_dst_pd());
const memory_desc_wrapper ws_d(conf_.workspace_pd());
const memory_desc_wrapper diff_src_d(conf_.diff_src_pd());
const int ID = conf_.ID();
const int IH = conf_.IH();
const int IW = conf_.IW();
const int KD = conf_.KD();
const int KH = conf_.KH();
const int KW = conf_.KW();
const int SD = conf_.KSD();
const int SH = conf_.KSH();
const int SW = conf_.KSW();
const int padF = conf_.padFront();
const int padT = conf_.padT();
const int padL = conf_.padL();
const bool is_3d = conf_.desc()->diff_src_desc.ndims == 5;
auto alg = conf_.desc()->alg_kind;
auto apply_offset = [=](int index, int offset) {
return (index > offset) ? index - offset : 0;
};
auto ker_zero = [=](int _mb, int _oc) {
for (int ih = 0; ih < IH; ++ih) {
for (int iw = 0; iw < IW; ++iw) {
diff_src[diff_src_d.off(_mb, _oc, ih, iw)] = data_type_t(0);
}
}
};
auto ker_max = [=](const data_t *d, int mb, int oc, int oh, int ow) {
const size_t ws_off = ws_d.off(mb, oc, oh, ow);
const int index = ws_d.data_type() == data_type::u8
? (int)ws[ws_off] : ((int *)ws)[ws_off];
const int kw = index % KW;
const int kh = index / KW;
const int ih = oh * SH - padT + kh;
const int iw = ow * SW - padL + kw;
// If padding area could fit the kernel,
// then input displacement would be out of bounds.
// No need to back propagate there as padding is
// virtual in pooling_max case.
if (ih < 0 || ih >= IH)
return;
if (iw < 0 || iw >= IW)
return;
diff_src[diff_src_d.off(mb, oc, ih, iw)] += d[0];
};
auto ker_avg = [=](const data_t *d, int mb, int oc, int oh, int ow) {
auto ih_start = apply_offset(oh*SH, padT);
auto iw_start = apply_offset(ow*SW, padL);
auto ih_end = nstl::min(oh*SH - padT + KH, IH);
auto iw_end = nstl::min(ow*SW - padL + KW, IW);
auto num_summands = (alg == pooling_avg_include_padding) ? KW*KH
: (ih_end - ih_start)*(iw_end - iw_start);
for (int ih = ih_start; ih < ih_end; ++ih) {
for (int iw = iw_start; iw < iw_end; ++iw) {
diff_src[diff_src_d.off(mb, oc, ih, iw)] += d[0] / num_summands;
}
}
};
auto ker_zero_3d = [=](int _mb, int _oc) {
for (int id = 0; id < ID; ++id) {
for (int ih = 0; ih < IH; ++ih) {
for (int iw = 0; iw < IW; ++iw) {
diff_src[diff_src_d.off(_mb, _oc, id, ih, iw)] =
data_type_t(0);
}
}
}
};
auto ker_max_3d = [=](const data_t *d, int mb, int oc, int od, int oh,
int ow) {
const size_t ws_off = ws_d.off(mb, oc, od, oh, ow);
const int index = ws_d.data_type() == data_type::u8
? (int)ws[ws_off] : ((int *)ws)[ws_off];
const int kw = index % KW;
const int kh = (index / KW) % KH;
const int kd = (index / KW) / KH;
const int id = od * SD - padF + kd;
const int ih = oh * SH - padT + kh;
const int iw = ow * SW - padL + kw;
// If padding area could fit the kernel,
// then input displacement would be out of bounds.
// No need to back propagate there as padding is
// virtual in pooling_max case.
if (id < 0 || id >= ID)
return;
if (ih < 0 || ih >= IH)
return;
if (iw < 0 || iw >= IW)
return;
diff_src[diff_src_d.off(mb, oc, id, ih, iw)] += d[0];
};
auto ker_avg_3d = [=](const data_t *d, int mb, int oc, int od, int oh,
int ow) {
auto id_start = apply_offset(od*SD, padF);
auto ih_start = apply_offset(oh*SH, padT);
auto iw_start = apply_offset(ow*SW, padL);
auto id_end = nstl::min(od*SD - padF + KD, ID);
auto ih_end = nstl::min(oh*SH - padT + KH, IH);
auto iw_end = nstl::min(ow*SW - padL + KW, IW);
auto num_summands = (alg == pooling_avg_include_padding) ? KW*KH*KD
: (ih_end - ih_start)*(iw_end - iw_start)*(id_end - id_start);
for (int id = id_start; id < id_end; ++id)
for (int ih = ih_start; ih < ih_end; ++ih)
for (int iw = iw_start; iw < iw_end; ++iw) {
diff_src[diff_src_d.off(mb, oc, id, ih, iw)] += d[0] / num_summands;
}
};
const int MB = conf_.MB();
const int OC = conf_.C();
const int OD = conf_.OD();
const int OH = conf_.OH();
const int OW = conf_.OW();
if (conf_.desc()->alg_kind == alg_kind::pooling_max) {
# pragma omp parallel for collapse(2) schedule(static)
for (int mb = 0; mb < MB; ++mb) {
for (int oc = 0; oc < OC; ++oc) {
if (is_3d) ker_zero_3d(mb, oc);
else ker_zero(mb, oc);
for (int od = 0; od < OD; ++od) {
for (int oh = 0; oh < OH; ++oh) {
for (int ow = 0; ow < OW; ++ow) {
const data_t *d = is_3d
? &diff_dst[diff_dst_d.off(mb, oc, od, oh, ow)]
: &diff_dst[diff_dst_d.off(mb, oc, oh, ow)];
if (is_3d) ker_max_3d(d, mb, oc, od, oh, ow);
else ker_max(d, mb, oc, oh, ow);
}
}
}
}
}
} else {
# pragma omp parallel for collapse(2) schedule(static)
for (int mb = 0; mb < MB; ++mb) {
for (int oc = 0; oc < OC; ++oc) {
if (is_3d) ker_zero_3d(mb, oc);
else ker_zero(mb, oc);
for (int od = 0; od < OD; ++od) {
for (int oh = 0; oh < OH; ++oh) {
for (int ow = 0; ow < OW; ++ow) {
const data_t *d = is_3d
? &diff_dst[diff_dst_d.off(mb, oc, od, oh, ow)]
: &diff_dst[diff_dst_d.off(mb, oc, oh, ow)];
if (is_3d) ker_avg_3d(d, mb, oc, od, oh, ow);
else ker_avg(d, mb, oc, oh, ow);
}
}
}
}
}
}
}
template struct ref_pooling_fwd_t<data_type::f32>;
template struct ref_pooling_fwd_t<data_type::s32>;
template struct ref_pooling_fwd_t<data_type::s16, data_type::s32>;
template struct ref_pooling_fwd_t<data_type::s8, data_type::s32>;
template struct ref_pooling_fwd_t<data_type::u8, data_type::s32>;
template struct ref_pooling_bwd_t<data_type::f32>;
template struct ref_pooling_bwd_t<data_type::s32>;
template struct ref_pooling_bwd_t<data_type::s16, data_type::s32>;
}
}
}
// vim: et ts=4 sw=4 cindent cino^=l0,\:0,N-s
|
