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
|
/*******************************************************************************
* 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 <math.h>
#include "mkldnn_thread.hpp"
#include "utils.hpp"
namespace mkldnn {
namespace impl {
namespace cpu {
namespace gemm_utils {
#define BM_NOCOPY_AVX 64
#define BN_NOCOPY_AVX 48
#define BK_NOCOPY_AVX 384
#define BN_LARGE_NOCOPY_AVX 192
#define BM_SMALL_NOCOPY_AVX 16
#define BN_SMALL_NOCOPY_AVX 1
#define BK_SMALL_NOCOPY_AVX 4
// Determine number of threads for each dimension of a 3-D partitioning
// algorithm based on input parameters
// m/n/k - First/second/third parameter for GEMM
// nthrs - total available number of threads
// nthrs_m/nthrs_n/nthrs_k - number of threads to use in each dimension
// BM/BN/BK - blocking values
void calc_nthr_nocopy_avx(int m, int n, int k,
int nthrs, int *nthrs_m, int *nthrs_n, int *nthrs_k, int *BM, int *BN,
int *BK)
{
int nthr, nthr_m, nthr_n, nthr_k;
int MB, NB, KB;
nthr = nthrs;
nthr_m = (m + BM_NOCOPY_AVX - 1) / BM_NOCOPY_AVX;
nthr_n = (n + BN_NOCOPY_AVX - 1) / BN_NOCOPY_AVX;
nthr_k = 1;
// Partition along K dimension
// - if threading allows having barriers (e.g. OMP)
// - if there is not enough parallelism along M or N
if (mkldnn_thr_syncable()) {
int nthr_other = nthr_k = 1;
while ((nthr_m * nthr_n * nthr_other < nthr)
&& (k / (nthr_other + 1) > BK_NOCOPY_AVX)) {
nthr_other++;
if ((nthr / nthr_other) * nthr_other > 0.9 * nthr)
nthr_k = nthr_other;
}
}
nthr /= nthr_k;
if (nthr_m == 1)
nthr_n = nthr;
if (nthr_n == 1)
nthr_m = nthr;
// Simple partition reduction
while (nthr_m * nthr_n > nthr)
if (nthr_m > nthr_n)
nthr_m--;
else
nthr_n--;
while (nthr_m * nthr_n < nthr)
if (nthr_m < nthr_n)
nthr_m++;
else
nthr_n++;
if ((nthr_m * nthr_n > nthr) && (nthr_m > 1) && (nthr_n > 1)) {
if (nthr_m <= nthr_n) {
nthr_m = (int)sqrt((double)nthr);
if (nthr_m > (m + BM_SMALL_NOCOPY_AVX - 1) / BM_SMALL_NOCOPY_AVX)
nthr_m = (m + BM_SMALL_NOCOPY_AVX - 1) / BM_SMALL_NOCOPY_AVX;
nthr_n = nthr / nthr_m;
while ((nthr_m > 1) && (nthr_m * nthr_n != nthr)) {
nthr_m--;
nthr_n = nthr / nthr_m;
}
} else {
nthr_n = (int)sqrt((double)nthr);
if (nthr_n > (n + BN_SMALL_NOCOPY_AVX - 1) / BN_SMALL_NOCOPY_AVX)
nthr_n = (n + BN_SMALL_NOCOPY_AVX - 1) / BN_SMALL_NOCOPY_AVX;
nthr_m = nthr / nthr_n;
while ((nthr_n > 1) && (nthr_m * nthr_n != nthr)) {
nthr_n--;
nthr_m = nthr / nthr_n;
}
}
}
MB = (m + nthr_m - 1) / nthr_m + BM_SMALL_NOCOPY_AVX - 1;
MB -= MB % BM_SMALL_NOCOPY_AVX;
NB = (n + nthr_n - 1) / nthr_n + BN_SMALL_NOCOPY_AVX - 1;
NB -= NB % BN_SMALL_NOCOPY_AVX;
KB = (k + nthr_k - 1) / nthr_k + BK_SMALL_NOCOPY_AVX - 1;
KB -= KB % BK_SMALL_NOCOPY_AVX;
if (MB * nthr_m > m)
nthr_m = (m + MB - 1) / MB;
if (NB * nthr_n > n)
nthr_n = (n + NB - 1) / NB;
if (KB * nthr_k > k)
nthr_k = (k + KB - 1) / KB;
*nthrs_m = nthr_m;
*nthrs_n = nthr_n;
*nthrs_k = nthr_k;
*BM = MB;
*BN = NB;
*BK = KB;
}
#undef BM_NOCOPY_AVX
#undef BN_NOCOPY_AVX
#undef BK_NOCOPY_AVX
#undef BN_LARGE_NOCOPY_AVX
#undef BM_SMALL_NOCOPY_AVX
#undef BN_SMALL_NOCOPY_AVX
#undef BK_SMALL_NOCOPY_AVX
#define BM_NOCOPY_AVX512_COMMON 32
#define BN_NOCOPY_AVX512_COMMON 64
#define BK_NOCOPY_AVX512_COMMON 192
#define BN_LARGE_NOCOPY_AVX512_COMMON 192
#define BM_SMALL_NOCOPY_AVX512_COMMON 16
#define BN_SMALL_NOCOPY_AVX512_COMMON 1
#define BK_SMALL_NOCOPY_AVX512_COMMON 4
// Determine number of threads for each dimension of a 3-D partitioning
// algorithm based on input parameters
// m/n/k - First/second/third parameter for GEMM
// nthrs - total available number of threads
// nthrs_m/nthrs_n/nthrs_k - number of threads to use in each dimension
// BM/BN/BK - blocking values
void calc_nthr_nocopy_avx512_common(int m,
int n, int k, int nthrs, int *nthrs_m, int *nthrs_n, int *nthrs_k,
int *BM, int *BN, int *BK)
{
int nthr, nthr_m, nthr_n, nthr_k = 1;
int MB, NB, KB;
nthr = nthrs;
int counter = 0;
float ratio_float = 1.;
int ratio = 1;
nthr = nthrs;
int nthr_m_gt_n;
// Partition along K dimension
// - if threading allows having barriers (e.g. OMP)
// - if there is not enough parallelism along M or N
if (mkldnn_thr_syncable()) {
if (n <= 2 * BN_NOCOPY_AVX512_COMMON &&
m <= 2 * BM_NOCOPY_AVX512_COMMON * nthr) {
nthr_k = k / BK_NOCOPY_AVX512_COMMON;
if (nthr_k > nthr / 4)
nthr_k = nthr / 4;
if (nthr_k < 1)
nthr_k = 1;
while ((nthr_k > 1) && (nthr % nthr_k)) {
nthr_k--;
}
nthr /= nthr_k;
} else {
nthr_k = 1;
}
}
nthr_m = (m + BM_NOCOPY_AVX512_COMMON - 1) / BM_NOCOPY_AVX512_COMMON;
nthr_n = (n + BN_NOCOPY_AVX512_COMMON - 1) / BN_NOCOPY_AVX512_COMMON;
if (nthr_m < 1)
nthr_m = 1;
if (nthr_n < 1)
nthr_n = 1;
nthr_m_gt_n = nthr_m > nthr_n ? 1 : 0;
ratio_float = (float)nthr_m / nthr_n;
if (nthr_m_gt_n)
ratio = (int)ratio_float;
else
ratio = (int)(1. / ratio_float);
// scale down nthr_m and nthr_n if they are too large
while (nthr_m * nthr_n > 4 * nthr) {
nthr_m /= 2;
nthr_n /= 2;
}
if (nthr_m < 1)
nthr_m = 1;
if (nthr_n < 1)
nthr_n = 1;
// Simple partition reduction
counter = 0;
while (nthr_m * nthr_n > nthr) {
if (nthr_m > nthr_n) {
if (counter < ratio)
nthr_m--;
else {
nthr_n--;
counter = -1;
}
} else {
if (counter < ratio)
nthr_n--;
else {
nthr_m--;
counter = -1;
}
}
counter++;
}
// Simple partition increment
counter = 0;
while (nthr_m * nthr_n < 0.95 * nthr) {
if (nthr_m > nthr_n) {
if (counter < ratio)
nthr_m++;
else {
nthr_n++;
counter = -1;
}
} else {
if (counter < ratio)
nthr_n++;
else {
nthr_m++;
counter = -1;
}
}
counter++;
}
// if nothing works out, then this should work
if ((nthr_m * nthr_n > nthr)) {
if (nthr_m <= nthr_n) {
nthr_m = (int)sqrt((double)nthr);
if (nthr_m > (m + BM_SMALL_NOCOPY_AVX512_COMMON - 1)
/ BM_SMALL_NOCOPY_AVX512_COMMON)
nthr_m = (m + BM_SMALL_NOCOPY_AVX512_COMMON - 1)
/ BM_SMALL_NOCOPY_AVX512_COMMON;
nthr_n = nthr / nthr_m;
while ((nthr_m > 1) && (nthr_m * nthr_n != nthr)) {
nthr_m--;
nthr_n = nthr / nthr_m;
}
} else {
nthr_n = (int)sqrt((double)nthr);
if (nthr_n > (n + BN_SMALL_NOCOPY_AVX512_COMMON - 1)
/ BN_SMALL_NOCOPY_AVX512_COMMON)
nthr_n = (n + BN_SMALL_NOCOPY_AVX512_COMMON - 1)
/ BN_SMALL_NOCOPY_AVX512_COMMON;
nthr_m = nthr / nthr_n;
while ((nthr_n > 1) && (nthr_m * nthr_n != nthr)) {
nthr_n--;
nthr_m = nthr / nthr_n;
}
}
}
MB = (m + nthr_m - 1) / nthr_m + BM_SMALL_NOCOPY_AVX512_COMMON - 1;
MB -= MB % BM_SMALL_NOCOPY_AVX512_COMMON;
NB = (n + nthr_n - 1) / nthr_n + BN_SMALL_NOCOPY_AVX512_COMMON - 1;
NB -= NB % BN_SMALL_NOCOPY_AVX512_COMMON;
KB = (k + nthr_k - 1) / nthr_k + BK_SMALL_NOCOPY_AVX512_COMMON - 1;
KB -= KB % BK_SMALL_NOCOPY_AVX512_COMMON;
if (MB * nthr_m > m)
nthr_m = (m + MB - 1) / MB;
if (NB * nthr_n > n)
nthr_n = (n + NB - 1) / NB;
if (KB * nthr_k > k)
nthr_k = (k + KB - 1) / KB;
*nthrs_m = nthr_m;
*nthrs_n = nthr_n;
*nthrs_k = nthr_k;
*BM = MB;
*BN = NB;
*BK = KB;
}
#undef BM_NOCOPY_AVX512_COMMON
#undef BN_NOCOPY_AVX512_COMMON
#undef BK_NOCOPY_AVX512_COMMON
#undef BN_LARGE_NOCOPY_AVX512_COMMON
#undef BM_SMALL_NOCOPY_AVX512_COMMON
#undef BN_SMALL_NOCOPY_AVX512_COMMON
#undef BK_SMALL_NOCOPY_AVX512_COMMON
// Partition n values as equally as possible among nthr threads
// and set the offset (t_offset) and number of values (t_block) for ithr
// Assumption: 0 <= ithr < nthr
void partition_unit_diff(
int ithr, int nthr, int n, int *t_offset, int *t_block)
{
int band = n / nthr;
if (band == 0)
band = 1;
int tail = n - band * nthr;
if (tail < 0)
tail = 0;
if (ithr < tail) {
band++;
*t_offset = band * ithr;
*t_block = band;
} else {
*t_offset = band * ithr + tail;
*t_block = band;
}
if (*t_offset >= n) {
*t_offset = 0;
*t_block = 0;
}
if (*t_offset + *t_block > n) {
*t_block = n - *t_offset;
}
}
// Sum the m*n values from p_src into p_dst, assuming the two-dimensional
// arrays have leading dimensions ld_src and ld_dst, respectively
template<typename data_t>
void sum_two_matrices(
int m, int n, data_t *p_src, int ld_src, data_t *p_dst, int ld_dst)
{
int i, j;
for (j = 0; j < n; j++) {
for (i = 0; i < m; i++) {
p_dst[i + j * ld_dst] += p_src[i + j * ld_src];
}
}
}
template void sum_two_matrices<float>(
int m, int n, float *p_src, int ld_src, float *p_dst, int ld_dst);
template void sum_two_matrices<double>(
int m, int n, double *p_src, int ld_src, double *p_dst, int ld_dst);
}
}
}
}
|
