forked from Sunyouteng/darknet_prune_yoloV3
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathgemm.c
2870 lines (2476 loc) · 103 KB
/
gemm.c
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
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
#include "gemm.h"
#include "utils.h"
#include "im2col.h"
#include "dark_cuda.h"
#include <stdlib.h>
#include <stdio.h>
#include <math.h>
#include <float.h>
#include <string.h>
#include <stdint.h>
#if defined(_OPENMP)
#include <omp.h>
#endif
#if defined(_MSC_VER)
#if defined(_M_ARM) || defined(_M_ARM64)
static inline uint32_t popcnt(uint32_t v) {
v = v - ((v >> 1) & 0x55555555);
v = (v & 0x33333333) + ((v >> 2) & 0x33333333);
return ((v + (v >> 4) & 0xF0F0F0F) * 0x1010101) >> 24;
}
#define POPCNT(x) popcnt((x))
#define POPCNT64(x) (popcnt((unsigned)(x)) + popcnt((unsigned)((uint64_t)(x) >> 32)))
#else
#include <intrin.h>
#ifdef _WIN64
#define POPCNT(x) __popcnt(x)
#define POPCNT64(x) __popcnt64(x)
#else
static inline int popcnt_64(uint64_t val64) {
int tmp_count = __popcnt(val64);
tmp_count += __popcnt(val64 >> 32);
return tmp_count;
}
#define POPCNT(x) __popcnt(x)
#define POPCNT64(x) popcnt_64(x)
#endif
#endif
#elif defined(__GNUC__)
#define POPCNT(x) __builtin_popcount(x)
#define POPCNT64(x) __builtin_popcountll(x)
#endif
#define TILE_M 4 // 4 ops
#define TILE_N 16 // AVX2 = 2 ops * 8 floats
#define TILE_K 16 // loop
#ifdef __cplusplus
#define PUT_IN_REGISTER
#else
#define PUT_IN_REGISTER register
#endif
void gemm_bin(int M, int N, int K, float ALPHA,
char *A, int lda,
float *B, int ldb,
float *C, int ldc)
{
int i,j,k;
for(i = 0; i < M; ++i){
for(k = 0; k < K; ++k){
char A_PART = A[i*lda+k];
if(A_PART){
for(j = 0; j < N; ++j){
C[i*ldc+j] += B[k*ldb+j];
}
} else {
for(j = 0; j < N; ++j){
C[i*ldc+j] -= B[k*ldb+j];
}
}
}
}
}
float *random_matrix(int rows, int cols)
{
int i;
float* m = (float*)xcalloc(rows * cols, sizeof(float));
for(i = 0; i < rows*cols; ++i){
m[i] = (float)rand()/RAND_MAX;
}
return m;
}
void time_random_matrix(int TA, int TB, int m, int k, int n)
{
float *a;
if(!TA) a = random_matrix(m,k);
else a = random_matrix(k,m);
int lda = (!TA)?k:m;
float *b;
if(!TB) b = random_matrix(k,n);
else b = random_matrix(n,k);
int ldb = (!TB)?n:k;
float *c = random_matrix(m,n);
int i;
clock_t start = clock(), end;
for(i = 0; i<10; ++i){
gemm_cpu(TA,TB,m,n,k,1,a,lda,b,ldb,1,c,n);
}
end = clock();
printf("Matrix Multiplication %dx%d * %dx%d, TA=%d, TB=%d: %lf ms\n",m,k,k,n, TA, TB, (float)(end-start)/CLOCKS_PER_SEC);
free(a);
free(b);
free(c);
}
void gemm(int TA, int TB, int M, int N, int K, float ALPHA,
float *A, int lda,
float *B, int ldb,
float BETA,
float *C, int ldc)
{
gemm_cpu( TA, TB, M, N, K, ALPHA,A,lda, B, ldb,BETA,C,ldc);
}
//--------------------------------------------
// XNOR bitwise GEMM for binary neural network
//--------------------------------------------
static inline unsigned char xnor(unsigned char a, unsigned char b) {
//return a == b;
return !(a^b);
}
// INT-32
static inline uint32_t get_bit_int32(uint32_t const*const src, size_t index) {
size_t src_i = index / 32;
int src_shift = index % 32;
unsigned char val = (src[src_i] & (1 << src_shift)) > 0;
return val;
}
static inline uint32_t xnor_int32(uint32_t a, uint32_t b) {
return ~(a^b);
}
static inline uint64_t xnor_int64(uint64_t a, uint64_t b) {
return ~(a^b);
}
static inline uint32_t fill_bit_int32(char src) {
if (src == 0) return 0x00000000;
else return 0xFFFFFFFF;
}
static inline uint64_t fill_bit_int64(char src) {
if (src == 0) return 0x0000000000000000;
else return 0xFFFFFFFFFFFFFFFF;
}
void binary_int32_printf(uint32_t src) {
int i;
for (i = 0; i < 32; ++i) {
if (src & 1) printf("1");
else printf("0");
src = src >> 1;
}
printf("\n");
}
void binary_int64_printf(uint64_t src) {
int i;
for (i = 0; i < 64; ++i) {
if (src & 1) printf("1");
else printf("0");
src = src >> 1;
}
printf("\n");
}
/*
void gemm_nn_custom_bin_mean(int M, int N, int K, float ALPHA_UNUSED,
unsigned char *A, int lda,
unsigned char *B, int ldb,
float *C, int ldc, float *mean_arr)
{
int *count_arr = xcalloc(M*N, sizeof(int));
int i, j, k;
for (i = 0; i < M; ++i) { // l.n - filters [16 - 55 - 1024]
for (k = 0; k < K; ++k) { // l.size*l.size*l.c - one filter size [27 - 9216]
char a_bit = get_bit(A, i*lda + k);
for (j = 0; j < N; ++j) { // out_h*out_w - one channel output size [169 - 173056]
char b_bit = get_bit(B, k*ldb + j);
count_arr[i*ldc + j] += xnor(a_bit, b_bit);
}
}
}
for (i = 0; i < M; ++i) {
float mean_val = mean_arr[i];
for (j = 0; j < N; ++j) {
C[i*ldc + j] = (2 * count_arr[i*ldc + j] - K) * mean_val;
}
}
free(count_arr);
}
*/
/*
void gemm_nn_custom_bin_mean_transposed(int M, int N, int K, float ALPHA_UNUSED,
unsigned char *A, int lda,
unsigned char *B, int ldb,
float *C, int ldc, float *mean_arr)
{
int *count_arr = xcalloc(M*N, sizeof(int));
int i, j, k;
for (i = 0; i < M; ++i) { // l.n - filters [16 - 55 - 1024]
for (j = 0; j < N; ++j) { // out_h*out_w - one channel output size [169 - 173056]
for (k = 0; k < K; ++k) { // l.size*l.size*l.c - one filter size [27 - 9216]
char a_bit = get_bit(A, i*lda + k);
char b_bit = get_bit(B, j*ldb + k);
count_arr[i*ldc + j] += xnor(a_bit, b_bit);
}
}
}
for (i = 0; i < M; ++i) {
float mean_val = mean_arr[i];
for (j = 0; j < N; ++j) {
C[i*ldc + j] = (2 * count_arr[i*ldc + j] - K) * mean_val;
}
}
free(count_arr);
}
*/
/*
void gemm_nn_custom_bin_mean(int M, int N, int K, float ALPHA_UNUSED,
unsigned char *A, int lda,
unsigned char *B, int ldb,
float *C, int ldc, float *mean_arr)
{
int *count_arr = xcalloc(M*N, sizeof(int));
int i;
#pragma omp parallel for
for (i = 0; i < M; ++i) { // l.n - filters [16 - 55 - 1024]
int j, k, h;
for (k = 0; k < K; ++k) { // l.size*l.size*l.c - one filter size [27 - 9216]
const char a_bit = get_bit(A, i*lda + k);
uint64_t a_bit64 = fill_bit_int64(a_bit);
int k_ldb = k*ldb;
for (j = 0; j < N; j += 64) { // out_h*out_w - one channel output size [169 - 173056]
if ((N - j > 64) && (k_ldb % 8 == 0)) {
uint64_t b_bit64 = *((uint64_t *)(B + (k_ldb + j) / 8));
uint64_t c_bit64 = xnor_int64(a_bit64, b_bit64);
//printf("\n %d \n",__builtin_popcountll(c_bit64)); // gcc
printf("\n %d \n", POPCNT64(c_bit64)); // msvs
int h;
for (h = 0; h < 64; ++h)
if ((c_bit64 >> h) & 1) count_arr[i*ldc + j + h] += 1;
//binary_int64_printf(a_bit64);
//binary_int64_printf(b_bit64);
//binary_int64_printf(c_bit64);
}
else {
for (; j < N; ++j) { // out_h*out_w - one channel output size [169 - 173056]
char b_bit = get_bit(B, k_ldb + j);
if (xnor(a_bit, b_bit)) count_arr[i*ldc + j] += 1;
}
}
}
}
}
if (mean_arr) {
//int K_2 = K / 2;
for (i = 0; i < M; ++i) {
float mean_val = mean_arr[i];
//float mean_val2 = 2 * mean_val;
for (j = 0; j < N; ++j) {
C[i*ldc + j] = (2 * count_arr[i*ldc + j] - K) * mean_val;
//C[i*ldc + j] = (count_arr[i*ldc + j] - K_2) *mean_val2;
}
}
}
else {
for (i = 0; i < M; ++i) {
for (j = 0; j < N; ++j) {
C[i*ldc + j] = count_arr[i*ldc + j] - K / 2;
}
}
}
free(count_arr);
//getchar();
}
*/
/*
void gemm_nn_custom_bin_mean_transposed(int M, int N, int K, float ALPHA_UNUSED,
unsigned char *A, int lda,
unsigned char *B, int ldb,
float *C, int ldc, float *mean_arr)
{
int i;
#pragma omp parallel for
for (i = 0; i < M; ++i) { // l.n - filters [16 - 55 - 1024]
int j, k, h;
float mean_val = mean_arr[i];
for (j = 0; j < N; ++j) { // out_h*out_w - one channel output size [169 - 173056]
int count = 0;
for (k = 0; k < K; k += 64) { // l.size*l.size*l.c - one filter size [27 - 9216]
uint64_t a_bit64 = *((uint64_t *)(A + (i*lda + k) / 8));
uint64_t b_bit64 = *((uint64_t *)(B + (j*ldb + k) / 8));
uint64_t c_bit64 = xnor_int64(a_bit64, b_bit64);
int tmp_count = POPCNT64(c_bit64);
if (K - k < 64) tmp_count = tmp_count - (64 - (K - k)); // remove extra bits
count += tmp_count;
//binary_int64_printf(c_bit64);
//printf(", count = %d \n\n", tmp_count);
}
C[i*ldc + j] = (2 * count - K) * mean_val;
}
}
}
*/
//----------------------------
// is not used
/*
void transpose_32x32_bits_my(uint32_t *A, uint32_t *B, int lda, int ldb)
{
unsigned int x, y;
for (y = 0; y < 32; ++y) {
for (x = 0; x < 32; ++x) {
if (A[y * lda] & ((uint32_t)1 << x)) B[x * ldb] |= (uint32_t)1 << y;
}
}
}
*/
#ifndef GPU
uint8_t reverse_8_bit(uint8_t a) {
return ((a * 0x0802LU & 0x22110LU) | (a * 0x8020LU & 0x88440LU)) * 0x10101LU >> 16;
}
uint32_t reverse_32_bit(uint32_t a)
{
// unsigned int __rbit(unsigned int val) // for ARM //__asm__("rbit %0, %1\n" : "=r"(output) : "r"(input));
return (reverse_8_bit(a >> 24) << 0) |
(reverse_8_bit(a >> 16) << 8) |
(reverse_8_bit(a >> 8) << 16) |
(reverse_8_bit(a >> 0) << 24);
}
#define swap(a0, a1, j, m) t = (a0 ^ (a1 >>j)) & m; a0 = a0 ^ t; a1 = a1 ^ (t << j);
void transpose32_optimized(uint32_t A[32]) {
int j, k;
unsigned m, t;
//m = 0x0000FFFF;
//for (j = 16; j != 0; j = j >> 1, m = m ^ (m << j)) {
// for (k = 0; k < 32; k = (k + j + 1) & ~j) {
// t = (A[k] ^ (A[k + j] >> j)) & m;
// A[k] = A[k] ^ t;
// A[k + j] = A[k + j] ^ (t << j);
// }
//}
j = 16;
m = 0x0000FFFF;
for (k = 0; k < 32; k = (k + j + 1) & ~j) { swap(A[k], A[k + j], j, m); }
j = 8;
m = 0x00ff00ff;
for (k = 0; k < 32; k = (k + j + 1) & ~j) { swap(A[k], A[k + j], j, m); }
j = 4;
m = 0x0f0f0f0f;
for (k = 0; k < 32; k = (k + j + 1) & ~j) { swap(A[k], A[k + j], j, m); }
j = 2;
m = 0x33333333;
for (k = 0; k < 32; k = (k + j + 1) & ~j) { swap(A[k], A[k + j], j, m); }
j = 1;
m = 0x55555555;
for (k = 0; k < 32; k = (k + j + 1) & ~j) { swap(A[k], A[k + j], j, m); }
// reverse Y
for (j = 0; j < 16; ++j) {
uint32_t tmp = A[j];
A[j] = reverse_32_bit(A[31 - j]);
A[31 - j] = reverse_32_bit(tmp);
}
}
void transpose_32x32_bits_reversed_diagonale(uint32_t *A, uint32_t *B, int m, int n)
{
unsigned A_tmp[32];
int i;
#pragma unroll
for (i = 0; i < 32; ++i) A_tmp[i] = A[i * m];
transpose32_optimized(A_tmp);
#pragma unroll
for (i = 0; i < 32; ++i) B[i*n] = A_tmp[i];
}
void transpose_8x8_bits_my(unsigned char *A, unsigned char *B, int lda, int ldb)
{
unsigned x, y;
for (y = 0; y < 8; ++y) {
for (x = 0; x < 8; ++x) {
if (A[y * lda] & (1 << x)) B[x * ldb] |= 1 << y;
}
}
}
unsigned char reverse_byte_1(char a)
{
return ((a & 0x1) << 7) | ((a & 0x2) << 5) |
((a & 0x4) << 3) | ((a & 0x8) << 1) |
((a & 0x10) >> 1) | ((a & 0x20) >> 3) |
((a & 0x40) >> 5) | ((a & 0x80) >> 7);
}
unsigned char reverse_byte(unsigned char a)
{
return ((a * 0x0802LU & 0x22110LU) | (a * 0x8020LU & 0x88440LU)) * 0x10101LU >> 16;
}
static unsigned char lookup[16] = {
0x0, 0x8, 0x4, 0xc, 0x2, 0xa, 0x6, 0xe,
0x1, 0x9, 0x5, 0xd, 0x3, 0xb, 0x7, 0xf, };
unsigned char reverse_byte_3(unsigned char n) {
// Reverse the top and bottom nibble then swap them.
return (lookup[n & 0b1111] << 4) | lookup[n >> 4];
}
void transpose8rS32_reversed_diagonale(unsigned char* A, unsigned char* B, int m, int n)
{
unsigned x, y, t;
x = y = 0;
// Load the array and pack it into x and y.
//x = (A[0] << 24) | (A[m] << 16) | (A[2 * m] << 8) | A[3 * m];
//y = (A[4 * m] << 24) | (A[5 * m] << 16) | (A[6 * m] << 8) | A[7 * m];
t = (x ^ (x >> 7)) & 0x00AA00AA; x = x ^ t ^ (t << 7);
t = (y ^ (y >> 7)) & 0x00AA00AA; y = y ^ t ^ (t << 7);
t = (x ^ (x >> 14)) & 0x0000CCCC; x = x ^ t ^ (t << 14);
t = (y ^ (y >> 14)) & 0x0000CCCC; y = y ^ t ^ (t << 14);
t = (x & 0xF0F0F0F0) | ((y >> 4) & 0x0F0F0F0F);
y = ((x << 4) & 0xF0F0F0F0) | (y & 0x0F0F0F0F);
x = t;
B[7 * n] = reverse_byte(x >> 24); B[6 * n] = reverse_byte(x >> 16); B[5 * n] = reverse_byte(x >> 8); B[4 * n] = reverse_byte(x);
B[3 * n] = reverse_byte(y >> 24); B[2 * n] = reverse_byte(y >> 16); B[1 * n] = reverse_byte(y >> 8); B[0 * n] = reverse_byte(y);
}
/*
// transpose by 8-bit
void transpose_bin(char *A, char *B, const int n, const int m,
const int lda, const int ldb, const int block_size)
{
//printf("\n n = %d, ldb = %d \t\t m = %d, lda = %d \n", n, ldb, m, lda);
int i;
#pragma omp parallel for
for (i = 0; i < n; i += 8) {
int j;
for (j = 0; j < m; j += 8) {
int a_index = i*lda + j;
int b_index = j*ldb + i;
//transpose_8x8_bits_my(&A[a_index/8], &B[b_index/8], lda/8, ldb/8);
transpose8rS32_reversed_diagonale(&A[a_index / 8], &B[b_index / 8], lda / 8, ldb / 8);
}
for (; j < m; ++j) {
if (get_bit(A, i*lda + j)) set_bit(B, j*ldb + i);
}
}
}
*/
#endif
// transpose by 32-bit
void transpose_bin(uint32_t *A, uint32_t *B, const int n, const int m,
const int lda, const int ldb, const int block_size)
{
//printf("\n n = %d (n mod 32 = %d), m = %d (m mod 32 = %d) \n", n, n % 32, m, m % 32);
//printf("\n lda = %d (lda mod 32 = %d), ldb = %d (ldb mod 32 = %d) \n", lda, lda % 32, ldb, ldb % 32);
int i;
#pragma omp parallel for
for (i = 0; i < n; i += 32) {
int j;
for (j = 0; j < m; j += 32) {
int a_index = i*lda + j;
int b_index = j*ldb + i;
transpose_32x32_bits_reversed_diagonale(&A[a_index / 32], &B[b_index / 32], lda / 32, ldb / 32);
//transpose_32x32_bits_my(&A[a_index/32], &B[b_index/32], lda/32, ldb/32);
}
for (; j < m; ++j) {
if (get_bit((const unsigned char* const)A, i * lda + j)) set_bit((unsigned char* const)B, j * ldb + i);
}
}
}
#if (defined(__AVX__) && defined(__x86_64__)) || (defined(_WIN64) && !defined(__MINGW32__) && !defined(_M_ARM64))
#if (defined(_WIN64) && !defined(__MINGW64__))
#include <intrin.h>
#include <ammintrin.h>
#include <immintrin.h>
#include <smmintrin.h>
#if defined(_MSC_VER) && _MSC_VER <= 1900
static inline __int32 _mm256_extract_epi64(__m256i a, const int index) {
return a.m256i_i64[index];
}
static inline __int32 _mm256_extract_epi32(__m256i a, const int index) {
return a.m256i_i32[index];
}
#endif
static inline float _dn_castu32_f32(uint32_t a) {
return *((float *)&a);
}
static inline float _mm256_extract_float32(__m256 a, const int index) {
return a.m256_f32[index];
}
#else // Linux GCC/Clang
#include <x86intrin.h>
#include <ammintrin.h>
#include <immintrin.h>
#include <smmintrin.h>
#include <cpuid.h>
static inline float _dn_castu32_f32(uint32_t a) {
return *((float *)&a);
}
static inline float _mm256_extract_float32(__m256 a, const int index) {
switch(index) {
case 0:
return _dn_castu32_f32(_mm256_extract_epi32(_mm256_castps_si256(a), 0));
case 1:
return _dn_castu32_f32(_mm256_extract_epi32(_mm256_castps_si256(a), 1));
case 2:
return _dn_castu32_f32(_mm256_extract_epi32(_mm256_castps_si256(a), 2));
case 3:
return _dn_castu32_f32(_mm256_extract_epi32(_mm256_castps_si256(a), 3));
case 4:
return _dn_castu32_f32(_mm256_extract_epi32(_mm256_castps_si256(a), 4));
case 5:
return _dn_castu32_f32(_mm256_extract_epi32(_mm256_castps_si256(a), 5));
case 6:
return _dn_castu32_f32(_mm256_extract_epi32(_mm256_castps_si256(a), 6));
case 7:
return _dn_castu32_f32(_mm256_extract_epi32(_mm256_castps_si256(a), 7));
default:
return _dn_castu32_f32(_mm256_extract_epi32(_mm256_castps_si256(a), 0));
}
}
void asm_cpuid(uint32_t* abcd, uint32_t eax)
{
uint32_t ebx = 0, edx = 0, ecx = 0;
// EBX is saved to EDI and later restored
__asm__("movl %%ebx, %%edi;"
"cpuid;"
"xchgl %%ebx, %%edi;"
: "=D"(ebx),
"+a"(eax), "+c"(ecx), "=d"(edx));
abcd[0] = eax;
abcd[1] = ebx;
abcd[2] = ecx;
abcd[3] = edx;
}
#endif
#ifdef _WIN32
// Windows
#define cpuid(info, x) __cpuidex(info, x, 0)
#else
// GCC Intrinsics
void cpuid(int info[4], int InfoType) {
__cpuid_count(InfoType, 0, info[0], info[1], info[2], info[3]);
}
#endif
// Misc.
static int HW_MMX, HW_x64, HW_RDRAND, HW_BMI1, HW_BMI2, HW_ADX, HW_PREFETCHWT1;
static int HW_ABM; // Advanced Bit Manipulation
// SIMD: 128-bit
static int HW_SSE, HW_SSE2, HW_SSE3, HW_SSSE3, HW_SSE41, HW_SSE42, HW_SSE4a, HW_AES, HW_SHA;
// SIMD: 256-bit
static int HW_AVX, HW_XOP, HW_FMA3, HW_FMA4, HW_AVX2;
// SIMD: 512-bit
static int HW_AVX512F; // AVX512 Foundation
static int HW_AVX512CD; // AVX512 Conflict Detection
static int HW_AVX512PF; // AVX512 Prefetch
static int HW_AVX512ER; // AVX512 Exponential + Reciprocal
static int HW_AVX512VL; // AVX512 Vector Length Extensions
static int HW_AVX512BW; // AVX512 Byte + Word
static int HW_AVX512DQ; // AVX512 Doubleword + Quadword
static int HW_AVX512IFMA; // AVX512 Integer 52-bit Fused Multiply-Add
static int HW_AVX512VBMI; // AVX512 Vector Byte Manipulation Instructions
// https://stackoverflow.com/questions/6121792/how-to-check-if-a-cpu-supports-the-sse3-instruction-set
void check_cpu_features(void) {
int info[4];
cpuid(info, 0);
int nIds = info[0];
cpuid(info, 0x80000000);
unsigned nExIds = info[0];
// Detect Features
if (nIds >= 0x00000001) {
cpuid(info, 0x00000001);
HW_MMX = (info[3] & ((uint32_t)1 << 23)) != 0;
HW_SSE = (info[3] & ((uint32_t)1 << 25)) != 0;
HW_SSE2 = (info[3] & ((uint32_t)1 << 26)) != 0;
HW_SSE3 = (info[2] & ((uint32_t)1 << 0)) != 0;
HW_SSSE3 = (info[2] & ((uint32_t)1 << 9)) != 0;
HW_SSE41 = (info[2] & ((uint32_t)1 << 19)) != 0;
HW_SSE42 = (info[2] & ((uint32_t)1 << 20)) != 0;
HW_AES = (info[2] & ((uint32_t)1 << 25)) != 0;
HW_AVX = (info[2] & ((uint32_t)1 << 28)) != 0;
HW_FMA3 = (info[2] & ((uint32_t)1 << 12)) != 0;
HW_RDRAND = (info[2] & ((uint32_t)1 << 30)) != 0;
}
if (nIds >= 0x00000007) {
cpuid(info, 0x00000007);
HW_AVX2 = (info[1] & ((uint32_t)1 << 5)) != 0;
HW_BMI1 = (info[1] & ((uint32_t)1 << 3)) != 0;
HW_BMI2 = (info[1] & ((uint32_t)1 << 8)) != 0;
HW_ADX = (info[1] & ((uint32_t)1 << 19)) != 0;
HW_SHA = (info[1] & ((uint32_t)1 << 29)) != 0;
HW_PREFETCHWT1 = (info[2] & ((uint32_t)1 << 0)) != 0;
HW_AVX512F = (info[1] & ((uint32_t)1 << 16)) != 0;
HW_AVX512CD = (info[1] & ((uint32_t)1 << 28)) != 0;
HW_AVX512PF = (info[1] & ((uint32_t)1 << 26)) != 0;
HW_AVX512ER = (info[1] & ((uint32_t)1 << 27)) != 0;
HW_AVX512VL = (info[1] & ((uint32_t)1 << 31)) != 0;
HW_AVX512BW = (info[1] & ((uint32_t)1 << 30)) != 0;
HW_AVX512DQ = (info[1] & ((uint32_t)1 << 17)) != 0;
HW_AVX512IFMA = (info[1] & ((uint32_t)1 << 21)) != 0;
HW_AVX512VBMI = (info[2] & ((uint32_t)1 << 1)) != 0;
}
if (nExIds >= 0x80000001) {
cpuid(info, 0x80000001);
HW_x64 = (info[3] & ((uint32_t)1 << 29)) != 0;
HW_ABM = (info[2] & ((uint32_t)1 << 5)) != 0;
HW_SSE4a = (info[2] & ((uint32_t)1 << 6)) != 0;
HW_FMA4 = (info[2] & ((uint32_t)1 << 16)) != 0;
HW_XOP = (info[2] & ((uint32_t)1 << 11)) != 0;
}
}
int is_avx() {
static int result = -1;
if (result == -1) {
check_cpu_features();
result = HW_AVX;
if (result == 1) printf(" Used AVX \n");
else printf(" Not used AVX \n");
}
return result;
}
int is_fma_avx2() {
static int result = -1;
if (result == -1) {
check_cpu_features();
result = HW_FMA3 && HW_AVX2;
if (result == 1) printf(" Used FMA & AVX2 \n");
else printf(" Not used FMA & AVX2 \n");
}
return result;
}
// syt
/*
void gemm_nn(int m, int n, int k, float alpha,
float *a, int lda,
float *b, int ldb,
float *c, int ldc)
{
int i, j, k;
for (i = 0; i < m; ++i) {
for (k = 0; k < k; ++k) {
put_in_register float a_part = alpha * a[i * lda + k];
for (j = 0; j < n; ++j) {
c[i*ldc + j] += a_part * b[k*ldb + j];
}
}
}
}
*/
// https://software.intel.com/sites/landingpage/IntrinsicsGuide
void gemm_nn(int M, int N, int K, float ALPHA,
float *A, int lda,
float *B, int ldb,
float *C, int ldc)
{
int i, j, k;
if (is_avx() == 1) { // AVX
for (i = 0; i < M; ++i) {
for (k = 0; k < K; ++k) {
float A_PART = ALPHA*A[i*lda + k];
__m256 a256, b256, c256, result256; // AVX
a256 = _mm256_set1_ps(A_PART);
for (j = 0; j < N - 8; j += 8) {
b256 = _mm256_loadu_ps(&B[k*ldb + j]);
c256 = _mm256_loadu_ps(&C[i*ldc + j]);
// FMA - Intel Haswell (2013), AMD Piledriver (2012)
//result256 = _mm256_fmadd_ps(a256, b256, c256);
result256 = _mm256_mul_ps(a256, b256);
result256 = _mm256_add_ps(result256, c256);
_mm256_storeu_ps(&C[i*ldc + j], result256);
}
int prev_end = (N % 8 == 0) ? (N - 8) : (N / 8) * 8;
for (j = prev_end; j < N; ++j)
C[i*ldc + j] += A_PART*B[k*ldb + j];
}
}
}
else {
for (i = 0; i < M; ++i) {
for (k = 0; k < K; ++k) {
PUT_IN_REGISTER float A_PART = ALPHA * A[i * lda + k];
for (j = 0; j < N; ++j) {
C[i*ldc + j] += A_PART*B[k*ldb + j];
}
/* // SSE
__m128 a128, b128, c128, result128; // SSE
a128 = _mm_set1_ps(A_PART);
for (j = 0; j < N - 4; j += 4) {
b128 = _mm_loadu_ps(&B[k*ldb + j]);
c128 = _mm_loadu_ps(&C[i*ldc + j]);
//result128 = _mm_fmadd_ps(a128, b128, c128);
result128 = _mm_mul_ps(a128, b128);
result128 = _mm_add_ps(result128, c128);
_mm_storeu_ps(&C[i*ldc + j], result128);
}
int prev_end = (N % 4 == 0) ? (N - 4) : (N / 4) * 4;
for (j = prev_end; j < N; ++j){
C[i*ldc + j] += A_PART*B[k*ldb + j];
}
*/
}
}
}
}
void gemm_nn_fast(int M, int N, int K, float ALPHA,
float *A, int lda,
float *B, int ldb,
float *C, int ldc)
{
int i;
#pragma omp parallel for
for (i = 0; i < (M / TILE_M)*TILE_M; i += TILE_M)
{
int j, k;
int i_d, k_d;
for (k = 0; k < (K / TILE_K)*TILE_K; k += TILE_K)
{
for (j = 0; j < (N / TILE_N)*TILE_N; j += TILE_N)
{
// L1 - 6 bits tag [11:6] - cache size 32 KB, conflict for each 4 KB
// L2 - 9 bits tag [14:6] - cache size 256 KB, conflict for each 32 KB
// L3 - 13 bits tag [18:6] - cache size 8 MB, conflict for each 512 KB
__m256 result256;
__m256 a256_0, b256_0; // AVX
__m256 a256_1, b256_1; // AVX
__m256 a256_2;// , b256_2; // AVX
__m256 a256_3;// , b256_3; // AVX
__m256 c256_0, c256_1, c256_2, c256_3;
__m256 c256_4, c256_5, c256_6, c256_7;
c256_0 = _mm256_loadu_ps(&C[(0 + i)*ldc + (0 + j)]);
c256_1 = _mm256_loadu_ps(&C[(1 + i)*ldc + (0 + j)]);
c256_2 = _mm256_loadu_ps(&C[(0 + i)*ldc + (8 + j)]);
c256_3 = _mm256_loadu_ps(&C[(1 + i)*ldc + (8 + j)]);
c256_4 = _mm256_loadu_ps(&C[(2 + i)*ldc + (0 + j)]);
c256_5 = _mm256_loadu_ps(&C[(3 + i)*ldc + (0 + j)]);
c256_6 = _mm256_loadu_ps(&C[(2 + i)*ldc + (8 + j)]);
c256_7 = _mm256_loadu_ps(&C[(3 + i)*ldc + (8 + j)]);
for (k_d = 0; k_d < (TILE_K); ++k_d)
{
a256_0 = _mm256_set1_ps(ALPHA*A[(0 + i)*lda + (k_d + k)]);
a256_1 = _mm256_set1_ps(ALPHA*A[(1 + i)*lda + (k_d + k)]);
a256_2 = _mm256_set1_ps(ALPHA*A[(2 + i)*lda + (k_d + k)]);
a256_3 = _mm256_set1_ps(ALPHA*A[(3 + i)*lda + (k_d + k)]);
b256_0 = _mm256_loadu_ps(&B[(k_d + k)*ldb + (0 + j)]);
b256_1 = _mm256_loadu_ps(&B[(k_d + k)*ldb + (8 + j)]);
// FMA - Intel Haswell (2013), AMD Piledriver (2012)
//c256_0 = _mm256_fmadd_ps(a256_0, b256_0, c256_0);
//c256_1 = _mm256_fmadd_ps(a256_1, b256_0, c256_1);
//c256_2 = _mm256_fmadd_ps(a256_0, b256_1, c256_2);
//c256_3 = _mm256_fmadd_ps(a256_1, b256_1, c256_3);
//c256_4 = _mm256_fmadd_ps(a256_2, b256_0, c256_4);
//c256_5 = _mm256_fmadd_ps(a256_3, b256_0, c256_5);
//c256_6 = _mm256_fmadd_ps(a256_2, b256_1, c256_6);
//c256_7 = _mm256_fmadd_ps(a256_3, b256_1, c256_7);
result256 = _mm256_mul_ps(a256_0, b256_0);
c256_0 = _mm256_add_ps(result256, c256_0);
result256 = _mm256_mul_ps(a256_1, b256_0);
c256_1 = _mm256_add_ps(result256, c256_1);
result256 = _mm256_mul_ps(a256_0, b256_1);
c256_2 = _mm256_add_ps(result256, c256_2);
result256 = _mm256_mul_ps(a256_1, b256_1);
c256_3 = _mm256_add_ps(result256, c256_3);
result256 = _mm256_mul_ps(a256_2, b256_0);
c256_4 = _mm256_add_ps(result256, c256_4);
result256 = _mm256_mul_ps(a256_3, b256_0);
c256_5 = _mm256_add_ps(result256, c256_5);
result256 = _mm256_mul_ps(a256_2, b256_1);
c256_6 = _mm256_add_ps(result256, c256_6);
result256 = _mm256_mul_ps(a256_3, b256_1);
c256_7 = _mm256_add_ps(result256, c256_7);
}
_mm256_storeu_ps(&C[(0 + i)*ldc + (0 + j)], c256_0);
_mm256_storeu_ps(&C[(1 + i)*ldc + (0 + j)], c256_1);
_mm256_storeu_ps(&C[(0 + i)*ldc + (8 + j)], c256_2);
_mm256_storeu_ps(&C[(1 + i)*ldc + (8 + j)], c256_3);
_mm256_storeu_ps(&C[(2 + i)*ldc + (0 + j)], c256_4);
_mm256_storeu_ps(&C[(3 + i)*ldc + (0 + j)], c256_5);
_mm256_storeu_ps(&C[(2 + i)*ldc + (8 + j)], c256_6);
_mm256_storeu_ps(&C[(3 + i)*ldc + (8 + j)], c256_7);
}
for (j = (N / TILE_N)*TILE_N; j < N; ++j) {
for (i_d = i; i_d < (i + TILE_M); ++i_d)
{
for (k_d = k; k_d < (k + TILE_K); ++k_d)
{
PUT_IN_REGISTER float A_PART = ALPHA*A[i_d*lda + k_d];
C[i_d*ldc + j] += A_PART*B[k_d*ldb + j];
}
}
}
}
for (k = (K / TILE_K)*TILE_K; k < K; ++k)
{
for (i_d = i; i_d < (i + TILE_M); ++i_d)
{
PUT_IN_REGISTER float A_PART = ALPHA*A[i_d*lda + k];
for (j = 0; j < N; ++j) {
C[i_d*ldc + j] += A_PART*B[k*ldb + j];
}
}
}
}
for (i = (M / TILE_M)*TILE_M; i < M; ++i) {
int j, k;
for (k = 0; k < K; ++k) {
PUT_IN_REGISTER float A_PART = ALPHA*A[i*lda + k];
for (j = 0; j < N; ++j) {
C[i*ldc + j] += A_PART*B[k*ldb + j];
}
}
}
}
void gemm_nn_bin_32bit_packed(int M, int N, int K, float ALPHA,
uint32_t *A, int lda,
uint32_t *B, int ldb,
float *C, int ldc, float *mean_arr)
{
int i;
#pragma omp parallel for
for (i = 0; i < M; ++i) { // l.n
int j, s;
float mean_val = mean_arr[i];
//printf(" l.mean_arr[i] = %d \n ", l.mean_arr[i]);
for (s = 0; s < K; ++s) // l.size*l.size*l.c/32 or (l.size*l.size*l.c)
{
PUT_IN_REGISTER uint32_t A_PART = A[i*lda + s];
__m256i a256 = _mm256_set1_epi32(A_PART);
for (j = 0; j < N - 8; j += 8)
{
__m256i b256 = *((__m256i*)&B[s*ldb + j]);
__m256i xor256 = _mm256_xor_si256(a256, b256); // xnor = xor(a,b)
__m256i all_1 = _mm256_set1_epi8((char)255);
__m256i xnor256 = _mm256_andnot_si256(xor256, all_1); // xnor = not(xor(a,b))
// waiting for - CPUID Flags: AVX512VPOPCNTDQ: __m512i _mm512_popcnt_epi32(__m512i a)
__m256 count = _mm256_setr_ps(
POPCNT(_mm256_extract_epi32(xnor256, 0)),
POPCNT(_mm256_extract_epi32(xnor256, 1)),
POPCNT(_mm256_extract_epi32(xnor256, 2)),
POPCNT(_mm256_extract_epi32(xnor256, 3)),
POPCNT(_mm256_extract_epi32(xnor256, 4)),
POPCNT(_mm256_extract_epi32(xnor256, 5)),
POPCNT(_mm256_extract_epi32(xnor256, 6)),
POPCNT(_mm256_extract_epi32(xnor256, 7)));
__m256 val2 = _mm256_set1_ps(2);
count = _mm256_mul_ps(count, val2); // count * 2
__m256 val32 = _mm256_set1_ps(32);
count = _mm256_sub_ps(count, val32); // count - 32
__m256 mean256 = _mm256_set1_ps(mean_val);
count = _mm256_mul_ps(count, mean256); // count * mean_val
__m256 c256 = *((__m256*)&C[i*ldc + j]);
count = _mm256_add_ps(count, c256); // c = c + count
*((__m256*)&C[i*ldc + j]) = count;
}
for (; j < N; ++j) // out_h*out_w;
{
PUT_IN_REGISTER uint32_t B_PART = B[s*ldb + j];
uint32_t xnor_result = ~(A_PART ^ B_PART);
int32_t count = POPCNT(xnor_result); // must be Signed int
C[i*ldc + j] += (2 * count - 32) * mean_val;
}
}
}
}
void convolution_2d_old(int w, int h, int ksize, int n, int c, int pad, int stride,
float *weights, float *input, float *output)
{
//const int out_h = (h + 2 * pad - ksize) / stride + 1; // output_height=input_height for stride=1 and pad=1
//const int out_w = (w + 2 * pad - ksize) / stride + 1; // output_width=input_width for stride=1 and pad=1
int fil;
// filter index