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
|
/*******************************************************************************
* Copyright 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 "c_types_map.hpp"
#include "type_helpers.hpp"
#include "mkldnn_thread.hpp"
#include "mkldnn_traits.hpp"
#include "math_utils.hpp"
#include "ref_deconvolution.hpp"
namespace mkldnn {
namespace impl {
namespace cpu {
void ref_deconvolution_fwd_t::compute_fwd_bias(const data_t *bias,
data_t *dst) const {
const memory_desc_wrapper dst_d(pd()->dst_md());
const int G = pd()->G();
const int MB = pd()->MB();
const int OH = pd()->OH();
const int OW = pd()->OW();
const int OD = pd()->OD();
const int OC = pd()->OC() / G;
const int ndims = pd()->desc()->src_desc.ndims;
parallel_nd(MB, G, OC, OD, OH, OW,
[&](int mb, int g, int oc, int od, int oh, int ow) {
auto b = bias[g * OC + oc];
switch (ndims) {
case 5: dst[dst_d.off(mb, g * OC + oc, od, oh, ow)] += b; break;
case 4: dst[dst_d.off(mb, g * OC + oc, oh, ow)] += b; break;
case 3: dst[dst_d.off(mb, g * OC + oc, ow)] += b; break;
default: assert(!"invalid dimension size");
}
});
}
void ref_deconvolution_fwd_t::compute_fwd_bias_ncdhw(const data_t *bias,
data_t *dst) const {
const memory_desc_wrapper dst_d(pd()->dst_md());
const int MB = pd()->MB();
const int OC = pd()->OC();
const int SP = pd()->OW()*pd()->OH()*pd()->OD();
parallel_nd(MB, OC, [&](int mb, int oc) {
PRAGMA_OMP_SIMD()
for (int sp = 0; sp < SP; ++sp) {
auto offset = (size_t)(mb * OC + oc) * SP + sp;
dst[offset] += bias[oc];
}
});
}
template <int blksize>
void ref_deconvolution_fwd_t::compute_fwd_bias_nCdhwXc(const data_t *bias,
data_t *dst) const {
const memory_desc_wrapper dst_d(pd()->dst_md());
const int MB = pd()->MB();
const int OC = pd()->OC();
const int SP = pd()->OW() * pd()->OH() * pd()->OD();
const ptrdiff_t stride_mb = dst_d.blocking_desc().strides[0];
parallel_nd(MB, utils::div_up(OC, blksize), SP,
[&](int mb, int oc_blk, int sp) {
int oc = oc_blk * blksize;
auto offset = mb * stride_mb + oc * SP + sp * blksize;
const int blk = nstl::min(blksize, OC - oc);
PRAGMA_OMP_SIMD()
for (int i = 0; i < blk; ++i)
dst[offset + i] += bias[oc + i];
});
}
void ref_deconvolution_bwd_weights_t::compute_bwd_bias(const data_t *diff_dst,
data_t *diff_bias) const {
const memory_desc_wrapper diff_dst_d(pd()->diff_dst_md());
const int G = pd()->G();
const int MB = pd()->MB();
const int OH = pd()->OH();
const int OW = pd()->OW();
const int OC = pd()->OC() / G;
const int OD = pd()->OD();
const int ndims = pd()->desc()->src_desc.ndims;
parallel_nd(G, OC, [&](int g, int oc) {
data_t db = 0;
for (int mb = 0; mb < MB; ++mb) {
for (int od = 0; od < OD; ++od) {
for (int oh = 0; oh < OH; ++oh) {
for (int ow = 0; ow < OW; ++ow) {
switch (ndims) {
case 5:
db += diff_dst[diff_dst_d.off(
mb, g * OC + oc, od, oh, ow)];
break;
case 4:
db += diff_dst[diff_dst_d.off(
mb, g * OC + oc, oh, ow)];
break;
case 3:
db += diff_dst[diff_dst_d.off(mb, g * OC + oc, ow)];
break;
default: assert(!"invalid dimension size");
}
}
}
}
}
diff_bias[g * OC + oc] = db;
});
}
void ref_deconvolution_bwd_weights_t::compute_bwd_bias_ncdhw(
const data_t *diff_dst, data_t *diff_bias) const {
const memory_desc_wrapper diff_dst_d(pd()->diff_dst_md());
const int OC = pd()->OC();
const int MB = pd()->MB();
const int SP = pd()->OH()*pd()->OW()*pd()->OD();
parallel_nd(OC, [&](int oc) {
data_t db = 0;
for (int mb = 0; mb < MB; ++mb) {
PRAGMA_OMP_SIMD()
for (int sp = 0; sp < SP; ++sp) {
auto offset = (size_t)(mb * OC + oc) * SP + sp;
db += diff_dst[offset];
}
}
diff_bias[oc] = db;
});
}
template <int blksize>
void ref_deconvolution_bwd_weights_t::compute_bwd_bias_nCdhwXc(
const data_t *diff_dst, data_t *diff_bias) const {
const memory_desc_wrapper diff_dst_d(pd()->diff_dst_md());
const int OC = pd()->OC();
const int MB = pd()->MB();
const int SP = pd()->OH() * pd()->OW() * pd()->OD();
const ptrdiff_t stride_mb = diff_dst_d.blocking_desc().strides[0];
parallel_nd(utils::div_up(OC, blksize), [&](int ocb) {
data_t db[blksize] = {0};
for (int mb = 0; mb < MB; ++mb) {
for (int sp = 0; sp < SP; ++sp) {
auto offset = mb * stride_mb + (ocb * SP + sp) * blksize;
PRAGMA_OMP_SIMD()
for (int i = 0; i < blksize; ++i)
db[i] += diff_dst[offset+i];
}
}
const int blk = nstl::min(blksize, OC - ocb * blksize);
PRAGMA_OMP_SIMD()
for (int i = 0; i < blk; ++i)
diff_bias[ocb * blksize + i] = db[i];
});
}
template void ref_deconvolution_fwd_t::compute_fwd_bias_nCdhwXc<8>(
const data_t *diff_dst, data_t *diff_bias) const;
template void ref_deconvolution_fwd_t::compute_fwd_bias_nCdhwXc<16>(
const data_t *diff_dst, data_t *diff_bias) const;
template void ref_deconvolution_bwd_weights_t::compute_bwd_bias_nCdhwXc<8>(
const data_t *diff_dst, data_t *diff_bias) const;
template void ref_deconvolution_bwd_weights_t::compute_bwd_bias_nCdhwXc<16>(
const data_t *diff_dst, data_t *diff_bias) const;
}
}
}
// vim: et ts=4 sw=4 cindent cino^=l0,\:0,N-s
|