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
|
/*******************************************************************************
* Copyright 2017-2018 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 "mkldnn_types.h"
#include "c_types_map.hpp"
#include "jit_uni_pooling.hpp"
#include "type_helpers.hpp"
#include "nstl.hpp"
namespace mkldnn {
namespace impl {
namespace cpu {
template <cpu_isa_t isa>
void jit_uni_pooling_fwd_t<isa>::execute_forward() {
auto src = reinterpret_cast<const data_t *>(this->input_memory(0));
auto dst = reinterpret_cast<data_t*>(this->memory(0));
auto indices = conf_.desc()->alg_kind == alg_kind::pooling_max ?
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 indices_d(conf_.workspace_pd());
const size_t ind_dt_size = indices
? types::data_type_size(indices_d.data_type()) : 0;
const auto &jpp = conf_.jpp_;
int mb = conf_.MB();
auto ker = [&](int n, int b_c, int oh) {
auto arg = jit_pool_call_s();
const int ij = oh * jpp.stride_h;
const int i_t_overflow = nstl::max(0, jpp.t_pad-ij);
const int i_b_overflow = nstl::max(jpp.ih, ij+jpp.kh-jpp.t_pad)-jpp.ih;
const int ih = nstl::max(ij - jpp.t_pad, 0);
arg.src = &src[src_d.blk_off(n, b_c, ih)];
arg.dst = &dst[dst_d.blk_off(n, b_c, oh)];
if (indices) {
const size_t ind_off = indices_d.blk_off(n, b_c, oh);
arg.indices = &indices[ind_off * ind_dt_size];
}
arg.oh = oh == 0;
arg.kh_padding = jpp.kh - i_t_overflow - i_b_overflow;
arg.kh_padding_shift = i_t_overflow*jpp.kw;
arg.kw_padding = 0;
arg.ker_area_h = conf_.desc()->alg_kind == alg_kind::pooling_avg_exclude_padding
? (float)(jpp.kh - nstl::max(0, oh*jpp.stride_h - jpp.t_pad + jpp.kh - jpp.ih) -
nstl::max(0, jpp.t_pad - oh*jpp.stride_h))
: (float)(jpp.kh - nstl::max(0, oh*jpp.stride_h - jpp.t_pad + jpp.kh - jpp.ih - jpp.b_pad));
(*kernel_)(&arg);
};
# pragma omp parallel for collapse(3) schedule(static)
for (int n = 0; n < mb; ++n) {
for (int b_c = 0; b_c < jpp.nb_c; ++b_c) {
for (int oh = 0; oh < jpp.oh; ++oh) {
ker(n, b_c, oh);
}
}
}
}
template <cpu_isa_t isa>
void jit_uni_pooling_fwd_t<isa>::execute_forward_3d() {
auto src = reinterpret_cast<const data_t *>(this->input_memory(0));
auto dst = reinterpret_cast<data_t*>(this->memory(0));
auto indices = conf_.desc()->alg_kind == alg_kind::pooling_max ?
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 indices_d(conf_.workspace_pd());
const size_t ind_dt_size = indices
? types::data_type_size(indices_d.data_type()) : 0;
const auto &jpp = conf_.jpp_;
int mb = conf_.MB();
auto ker = [&](int n, int b_c, int od, int oh, int id, int d_t_overflow,
int d_b_overflow) {
auto arg = jit_pool_call_s();
const int ij = oh * jpp.stride_h;
const int i_t_overflow = nstl::max(0, jpp.t_pad-ij);
const int i_b_overflow = nstl::max(jpp.ih, ij+jpp.kh-jpp.t_pad)-jpp.ih;
const int ih = nstl::max(ij - jpp.t_pad, 0);
arg.src = &src[src_d.blk_off(n, b_c, id, ih)];
arg.dst = &dst[dst_d.blk_off(n, b_c, od, oh)];
if (indices) {
const size_t ind_off = indices_d.blk_off(n, b_c, od, oh);
arg.indices = &indices[ind_off * ind_dt_size];
}
arg.oh = (oh + od == 0);
arg.kd_padding = jpp.kd - d_t_overflow - d_b_overflow;
arg.kh_padding = jpp.kh - i_t_overflow - i_b_overflow;
arg.kh_padding_shift = i_t_overflow*jpp.kw + d_t_overflow*jpp.kw*jpp.kh;
arg.kd_padding_shift = (i_t_overflow + i_b_overflow)*jpp.kw;
arg.kw_padding = 0;
arg.ker_area_h = (float)(jpp.kh -
nstl::max(0, oh*jpp.stride_h - jpp.t_pad + jpp.kh - jpp.ih) -
nstl::max(0, jpp.t_pad - oh*jpp.stride_h)) * (jpp.kd -
nstl::max(0, od*jpp.stride_d - jpp.f_pad + jpp.kd - jpp.id) -
nstl::max(0, jpp.f_pad - od*jpp.stride_d));
(*kernel_)(&arg);
};
# pragma omp parallel for collapse(3) schedule(static)
for (int n = 0; n < mb; ++n) {
for (int b_c = 0; b_c < jpp.nb_c; ++b_c) {
for (int od = 0; od < jpp.od; ++od) {
const int ik = od * jpp.stride_d;
const int d_t_overflow = nstl::max(0, jpp.f_pad-ik);
const int d_b_overflow = nstl::max(jpp.id, ik+jpp.kd-jpp.f_pad)
-jpp.id;
const int id = nstl::max(ik - jpp.f_pad, 0);
for (int oh = 0; oh < jpp.oh; ++oh) {
ker(n, b_c, od, oh, id, d_t_overflow, d_b_overflow);
}
}
}
}
}
template <cpu_isa_t isa>
void jit_uni_pooling_bwd_t<isa>::execute_backward() {
auto diff_dst = reinterpret_cast<const data_t *>(this->input_memory(0));
auto diff_src = reinterpret_cast<data_t*>(this->memory(0));
auto indices = conf_.desc()->alg_kind == alg_kind::pooling_max ?
reinterpret_cast<const char*>(this->input_memory(1)) : nullptr;
const memory_desc_wrapper diff_src_d(conf_.diff_src_pd());
const memory_desc_wrapper diff_dst_d(conf_.diff_dst_pd());
const memory_desc_wrapper indices_d(conf_.workspace_pd());
const size_t ind_dt_size = indices
? types::data_type_size(indices_d.data_type()) : 0;
const auto &jpp = conf_.jpp_;
int mb = conf_.MB();
auto ker = [&](int n, int b_c, int oh) {
auto arg = jit_pool_call_s();
const int ij = oh * jpp.stride_h;
const int i_t_overflow = nstl::max(0, jpp.t_pad-ij);
const int i_b_overflow = nstl::max(jpp.ih, ij+jpp.kh-jpp.t_pad)-jpp.ih;
const int ih = nstl::max(ij - jpp.t_pad, 0);
arg.src = &diff_src[diff_src_d.blk_off(n, b_c, ih)];
arg.dst = &diff_dst[diff_dst_d.blk_off(n, b_c, oh)];
if (indices) {
const size_t ind_off = indices_d.blk_off(n, b_c, oh);
arg.indices = &indices[ind_off * ind_dt_size];
}
arg.oh = (oh == 0);
arg.kh_padding = jpp.kh - i_t_overflow - i_b_overflow;
arg.kh_padding_shift = i_t_overflow*jpp.kw;
arg.kw_padding = 0;
arg.ker_area_h = (float)(jpp.kh -
nstl::max(0, oh*jpp.stride_h - jpp.t_pad + jpp.kh - jpp.ih) -
nstl::max(0, jpp.t_pad - oh*jpp.stride_h));
(*kernel_)(&arg);
};
# pragma omp parallel for collapse(2) schedule(static)
for (int n = 0; n < mb; ++n) {
for (int b_c = 0; b_c < jpp.nb_c; ++b_c) {
for (int oh = 0; oh < jpp.oh; ++oh) {
ker(n, b_c, oh);
}
}
}
}
template <cpu_isa_t isa>
void jit_uni_pooling_bwd_t<isa>::execute_backward_3d() {
auto diff_dst = reinterpret_cast<const data_t *>(this->input_memory(0));
auto diff_src = reinterpret_cast<data_t*>(this->memory(0));
auto indices = conf_.desc()->alg_kind == alg_kind::pooling_max ?
reinterpret_cast<const char*>(this->input_memory(1)) : nullptr;
const memory_desc_wrapper diff_src_d(conf_.diff_src_pd());
const memory_desc_wrapper diff_dst_d(conf_.diff_dst_pd());
const memory_desc_wrapper indices_d(conf_.workspace_pd());
const size_t ind_dt_size = indices
? types::data_type_size(indices_d.data_type()) : 0;
const auto &jpp = conf_.jpp_;
int mb = conf_.MB();
auto ker = [&](int n, int b_c, int od, int oh, int id, int d_t_overflow,
int d_b_overflow, int zero_size, int kd) {
auto arg = jit_pool_call_s();
const int ij = oh * jpp.stride_h;
const int i_t_overflow = nstl::max(0, jpp.t_pad-ij);
const int i_b_overflow = nstl::max(jpp.ih, ij+jpp.kh-jpp.t_pad)-jpp.ih;
const int ih = nstl::max(ij - jpp.t_pad, 0);
arg.src = &diff_src[diff_src_d.blk_off(n, b_c, id + kd, ih)];
arg.dst = &diff_dst[diff_dst_d.blk_off(n, b_c, od, oh)];
if (indices) {
const size_t ind_off = indices_d.blk_off(n, b_c, od, oh);
arg.indices = &indices[ind_off * ind_dt_size];
}
arg.oh = zero_size;
arg.kd_padding = jpp.kd - d_t_overflow - d_b_overflow;
arg.kh_padding = jpp.kh - i_t_overflow - i_b_overflow;
arg.kh_padding_shift = i_t_overflow*jpp.kw + d_t_overflow*jpp.kw*jpp.kh
+ kd * jpp.kw * jpp.kh;
arg.kd_padding_shift = (i_t_overflow + i_b_overflow)*jpp.kw;
arg.kw_padding = 0;
arg.ker_area_h = (float)(jpp.kh -
nstl::max(0, oh*jpp.stride_h - jpp.t_pad + jpp.kh - jpp.ih) -
nstl::max(0, jpp.t_pad - oh*jpp.stride_h)) * (jpp.kd -
nstl::max(0, od*jpp.stride_d - jpp.f_pad + jpp.kd - jpp.id) -
nstl::max(0, jpp.f_pad - od*jpp.stride_d));
(*kernel_)(&arg);
};
if (jpp.simple_alg) {
# pragma omp parallel for collapse(3) schedule(static)
for (int n = 0; n < mb; ++n) {
for (int b_c = 0; b_c < jpp.nb_c; ++b_c) {
for (int od = 0; od < jpp.od; ++od) {
const int ik = od * jpp.stride_d;
const int d_t_overflow = nstl::max(0, jpp.f_pad - ik);
const int d_b_overflow = nstl::max(jpp.id, ik + jpp.kd
- jpp.f_pad) - jpp.id;
const int id = nstl::max(ik - jpp.f_pad, 0);
int zero_s = jpp.stride_d - d_t_overflow - (nstl::max(
jpp.id, ik + jpp.stride_d - jpp.f_pad) - jpp.id);
for (int oh = 0; oh < jpp.oh; ++oh) {
ker(n, b_c, od, oh, id, d_t_overflow, d_b_overflow,
(oh == 0) ? zero_s : 0, 0);
}
}
}
}
} else {
ptrdiff_t nelems = (ptrdiff_t)mb * (ptrdiff_t)jpp.c
* (ptrdiff_t)jpp.id * (ptrdiff_t)jpp.ih * (ptrdiff_t)jpp.iw;
# pragma omp parallel for
for (ptrdiff_t i = 0; i < nelems; ++i)
diff_src[i] = 0.;
for (int kd = 0; kd < jpp.kd; ++kd) {
# pragma omp parallel for collapse(2) schedule(static)
for (int n = 0; n < mb; ++n) {
for (int b_c = 0; b_c < jpp.nb_c; ++b_c) {
for (int od = 0; od < jpp.od; ++od) {
const int ik = od * jpp.stride_d;
const int d_t_overflow = nstl::max(0, jpp.f_pad-ik);
const int d_b_overflow = nstl::max(jpp.id, ik + jpp.kd
- jpp.f_pad) - jpp.id;
if (kd >= jpp.kd - d_t_overflow - d_b_overflow)
continue;
const int id = nstl::max(ik - jpp.f_pad, 0);
for (int oh = 0; oh < jpp.oh; ++oh) {
ker(n, b_c, od, oh, id, d_t_overflow, d_b_overflow,
0, kd);
}
}
}
}
}
}
}
template struct jit_uni_pooling_fwd_t<sse42>;
template struct jit_uni_pooling_bwd_t<sse42>;
template struct jit_uni_pooling_fwd_t<avx2>;
template struct jit_uni_pooling_bwd_t<avx2>;
template struct jit_uni_pooling_fwd_t<avx512_common>;
template struct jit_uni_pooling_bwd_t<avx512_common>;
}
}
}
// vim: et ts=4 sw=4 cindent cino^=l0,\:0,N-s
|
