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
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
|
// Copyright (c) 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.
// --------------------------------------------------------------------------------------------------------------------------------
// L3_SIMD_4x8
// Input matrices dimensions: M x K x N
// Output matrix dimensions: M x N
// --------------------------------------------------------------------------------------------------------------------------------
#include "include/data_types.cl"
#include "include/activation_functions.cl"
#define DOT8i_0( _result, _A, _B, i) \
{ \
_result = mad(_A.s0, sub_group_broadcast( _B.s0, i), _result); \
}
#define DOT8i_1( _result, _A, _B, i) \
{ \
_result = mad(_A.s1, sub_group_broadcast( _B.s1, i), _result); \
}
#define DOT8i_2( _result, _A, _B, i) \
{ \
_result = mad(_A.s2, sub_group_broadcast( _B.s2, i), _result); \
}
#define DOT8i_3( _result, _A, _B, i) \
{ \
_result = mad(_A.s3, sub_group_broadcast( _B.s3, i), _result); \
}
#define DOT8i_4( _result, _A, _B, i) \
{ \
_result = mad(_A.s4, sub_group_broadcast( _B.s4, i), _result); \
}
#define DOT8i_5( _result, _A, _B, i) \
{ \
_result = mad(_A.s5, sub_group_broadcast( _B.s5, i), _result); \
}
#define DOT8i_6( _result, _A, _B, i) \
{ \
_result = mad(_A.s6, sub_group_broadcast( _B.s6, i), _result); \
}
#define DOT8i_7( _result, _A, _B, i) \
{ \
_result = mad(_A.s7, sub_group_broadcast( _B.s7, i), _result); \
}
#define DOT8i_( _result, _A, _B, i) \
{ \
_result = mad(_A.s0, sub_group_broadcast( _B.s0, i), _result); \
_result = mad(_A.s1, sub_group_broadcast( _B.s1, i), _result); \
_result = mad(_A.s2, sub_group_broadcast( _B.s2, i), _result); \
_result = mad(_A.s3, sub_group_broadcast( _B.s3, i), _result); \
_result = mad(_A.s4, sub_group_broadcast( _B.s4, i), _result); \
_result = mad(_A.s5, sub_group_broadcast( _B.s5, i), _result); \
_result = mad(_A.s6, sub_group_broadcast( _B.s6, i), _result); \
_result = mad(_A.s7, sub_group_broadcast( _B.s7, i), _result); \
}
#define UNIT_TYPE_2 CAT(UNIT_TYPE, 2)
#define UNIT_TYPE_4 CAT(UNIT_TYPE, 4)
#define UNIT_TYPE_8 CAT(UNIT_TYPE, 8)
__attribute__((reqd_work_group_size(16, 1, 8)))
__attribute__((intel_reqd_sub_group_size(16)))
KERNEL(convolution_gpu_winograd_6x3_s1_fused)
(
__global INPUT0_TYPE* I,
__global OUTPUT_TYPE* O,
#if FILTER_LAYOUT_IMAGE_2D_WEIGHTS_WINOGRAD_6x3_S1_FBXYB || FILTER_LAYOUT_IMAGE_2D_WEIGHTS_WINOGRAD_6x3_S1_XFBYB
__read_only image2d_t U,
#else
__global FILTER_TYPE* U,
#endif
#if BIAS_TERM
const __global UNIT_TYPE * bias,
#endif
uint split_idx)
{
// (DxC2)x(UxWx8c)
const uint slmSize = (2 * 8)*(16 * 4);
__local UNIT_TYPE_4 V[slmSize]; // 8 KB
/* These constants are defined as precompiler macros during compilation. */
const uint WC = W*INPUT0_FEATURE_NUM;
const uint HW = H*W;
const uint HWC = H*WC;
const uint WC4 = WC >> 2;
const uint K16 = FILTER_OFM_NUM >> 4;
const uint C4 = INPUT0_FEATURE_NUM >> 2;
const uint K2 = FILTER_OFM_NUM >> 1;
const uint QK2 = Q*K2;
const uint QK = Q*FILTER_OFM_NUM;
const uint PQK = P*QK;
const UNIT_TYPE sc = 0.1h;
const UNIT_TYPE scl = 1.0h/sc;
const UNIT_TYPE_4 scl_vec = (UNIT_TYPE_4)(sc, sc, sc, sc);
uint gx = get_group_id(0);
uint gy = get_group_id(1);
uint gz = get_group_id(2);
uint gk = gz % K16;
uint gn = gz / K16;
#define lx get_local_id(0)
#define lz get_local_id(2)
uint lxd8 = lx >> 3;
uint lxm8 = lx % 8;
// Load 16x8 input tile, with 2 pixel overlap in X and y.
// Compute 14x6 output tile.
// Load 32 filters.
// 8 threads total
int x = gx * 14 + lz * 2 + lxd8 - px;
int y = gy * 6 - py;
uint k = gk * 16;
uint c0 = lxm8 * 4;
// # x->
// # M0 M1 M2 M3 M4 M5 M6
// # +------------------------------------------
// # u | 0 1 | 0 1 | 0 1 | 0 1 | 0 1 | 0 1 | 0 1 |
// # | | 2 3 | 2 3 | 2 3 | 2 3 | 2 3 | 2 3 | 2 3 |
// # v
// #
UNIT_TYPE_2 M0 = (UNIT_TYPE_2)(UNIT_VAL_ZERO, UNIT_VAL_ZERO);
UNIT_TYPE_2 M1 = (UNIT_TYPE_2)(UNIT_VAL_ZERO, UNIT_VAL_ZERO);
UNIT_TYPE_2 M2 = (UNIT_TYPE_2)(UNIT_VAL_ZERO, UNIT_VAL_ZERO);
UNIT_TYPE_2 M3 = (UNIT_TYPE_2)(UNIT_VAL_ZERO, UNIT_VAL_ZERO);
UNIT_TYPE_2 M4 = (UNIT_TYPE_2)(UNIT_VAL_ZERO, UNIT_VAL_ZERO);
UNIT_TYPE_2 M5 = (UNIT_TYPE_2)(UNIT_VAL_ZERO, UNIT_VAL_ZERO);
UNIT_TYPE_2 M6 = (UNIT_TYPE_2)(UNIT_VAL_ZERO, UNIT_VAL_ZERO);
/*if (gy == 0) {
y = 0;
}*/
uint lxm4 = lx % 4;
uint lxb2 = (lx & 4) / 4;
#if INPUT0_LAYOUT_BYXF
uint adr = gn*HWC + ((uint)y)*WC + ((uint)x)*INPUT0_FEATURE_NUM + c0;
const __global UNIT_TYPE_4 *I_load = ((const __global UNIT_TYPE_4*)&(I[adr]));
#else
uint adr = gn*HWC + c0*HW + ((uint)y)*W + ((uint)x);
const __global UNIT_TYPE *I_load = (const __global UNIT_TYPE*)&I[adr];
#endif
// c, Kdsk
uint2 coordU0;
#if FILTER_LAYOUT_IMAGE_2D_WEIGHTS_WINOGRAD_6x3_S1_FBXYB
coordU0.x = (lz * 48 + k * 24);
coordU0.y = 0;
#else // FILTER_LAYOUT_IMAGE_2D_WEIGHTS_WINOGRAD_6x3_S1_XFBYB
coordU0.x = (k * 3);
coordU0.y = lz*C_;
int last_coord_y = lz*C_ + C_;
#endif
#if FILTER_LAYOUT_IMAGE_2D_WEIGHTS_WINOGRAD_6x3_S1_FBXYB || FILTER_LAYOUT_IMAGE_2D_WEIGHTS_WINOGRAD_6x3_S1_XFBYB
coordU0.x *= sizeof(UNIT_TYPE);
#endif
__attribute__((opencl_unroll_hint(1)))
for (uint c = 0; c < C4; c += 8) {
__local UNIT_TYPE_4 *V_write = &V[lxb2 * 512 + lz * 8 + lxd8 * 4 + lxm4];
__local const UNIT_TYPE_8 *V_read = (__local const UNIT_TYPE_8 *)&V[lz * 64 + lx * 2];
// 2*14 * 3 * 16 = 1344 MADs
// Transform HxW x C -> DxUxW x C
// 6x16x16 inputs -> 4x2x16x16 winograd components.
{
bool x_in = 0 <= x && x < W;
bool y0_in = 0 <= (y + 0) && (y + 0) < H && x_in;
bool y1_in = 0 <= (y + 1) && (y + 1) < H && x_in;
bool y2_in = 0 <= (y + 2) && (y + 2) < H && x_in;
bool y3_in = 0 <= (y + 3) && (y + 3) < H && x_in;
bool y4_in = 0 <= (y + 4) && (y + 4) < H && x_in;
bool y5_in = 0 <= (y + 5) && (y + 5) < H && x_in;
bool y6_in = 0 <= (y + 6) && (y + 6) < H && x_in;
bool y7_in = 0 <= (y + 7) && (y + 7) < H && x_in;
#if INPUT0_LAYOUT_BYXF
UNIT_TYPE_4 I0 = y0_in ? *((const __global UNIT_TYPE_4*)(I + adr + (0 * WC4 + c) * 4)) : (UNIT_TYPE_4)(UNIT_VAL_ZERO);
UNIT_TYPE_4 I1 = y1_in ? *((const __global UNIT_TYPE_4*)(I + adr + (1 * WC4 + c) * 4)) : (UNIT_TYPE_4)(UNIT_VAL_ZERO);
UNIT_TYPE_4 I2 = y2_in ? *((const __global UNIT_TYPE_4*)(I + adr + (2 * WC4 + c) * 4)) : (UNIT_TYPE_4)(UNIT_VAL_ZERO);
UNIT_TYPE_4 I3 = y3_in ? *((const __global UNIT_TYPE_4*)(I + adr + (3 * WC4 + c) * 4)) : (UNIT_TYPE_4)(UNIT_VAL_ZERO);
UNIT_TYPE_4 I4 = y4_in ? *((const __global UNIT_TYPE_4*)(I + adr + (4 * WC4 + c) * 4)) : (UNIT_TYPE_4)(UNIT_VAL_ZERO);
UNIT_TYPE_4 I5 = y5_in ? *((const __global UNIT_TYPE_4*)(I + adr + (5 * WC4 + c) * 4)) : (UNIT_TYPE_4)(UNIT_VAL_ZERO);
UNIT_TYPE_4 I6 = y6_in ? *((const __global UNIT_TYPE_4*)(I + adr + (6 * WC4 + c) * 4)) : (UNIT_TYPE_4)(UNIT_VAL_ZERO);
UNIT_TYPE_4 I7 = y7_in ? *((const __global UNIT_TYPE_4*)(I + adr + (7 * WC4 + c) * 4)) : (UNIT_TYPE_4)(UNIT_VAL_ZERO);
#else
const __global UNIT_TYPE *I_load_0 = &I_load[0 * W];
const __global UNIT_TYPE *I_load_1 = &I_load[1 * W];
const __global UNIT_TYPE *I_load_2 = &I_load[2 * W];
const __global UNIT_TYPE *I_load_3 = &I_load[3 * W];
const __global UNIT_TYPE *I_load_4 = &I_load[4 * W];
const __global UNIT_TYPE *I_load_5 = &I_load[5 * W];
const __global UNIT_TYPE *I_load_6 = &I_load[6 * W];
const __global UNIT_TYPE *I_load_7 = &I_load[7 * W];
UNIT_TYPE_4 I0 = y0_in ? (UNIT_TYPE_4)(I_load_0[c*HW * 4], I_load_0[c*HW * 4 + HW], I_load_0[c*HW * 4 + HW * 2], I_load_0[c*HW * 4 + HW * 3]) : (UNIT_TYPE_4)(UNIT_VAL_ZERO, UNIT_VAL_ZERO, UNIT_VAL_ZERO, UNIT_VAL_ZERO);
UNIT_TYPE_4 I1 = y1_in ? (UNIT_TYPE_4)(I_load_1[c*HW * 4], I_load_1[c*HW * 4 + HW], I_load_1[c*HW * 4 + HW * 2], I_load_1[c*HW * 4 + HW * 3]) : (UNIT_TYPE_4)(UNIT_VAL_ZERO, UNIT_VAL_ZERO, UNIT_VAL_ZERO, UNIT_VAL_ZERO);
UNIT_TYPE_4 I2 = y2_in ? (UNIT_TYPE_4)(I_load_2[c*HW * 4], I_load_2[c*HW * 4 + HW], I_load_2[c*HW * 4 + HW * 2], I_load_2[c*HW * 4 + HW * 3]) : (UNIT_TYPE_4)(UNIT_VAL_ZERO, UNIT_VAL_ZERO, UNIT_VAL_ZERO, UNIT_VAL_ZERO);
UNIT_TYPE_4 I3 = y3_in ? (UNIT_TYPE_4)(I_load_3[c*HW * 4], I_load_3[c*HW * 4 + HW], I_load_3[c*HW * 4 + HW * 2], I_load_3[c*HW * 4 + HW * 3]) : (UNIT_TYPE_4)(UNIT_VAL_ZERO, UNIT_VAL_ZERO, UNIT_VAL_ZERO, UNIT_VAL_ZERO);
UNIT_TYPE_4 I4 = y4_in ? (UNIT_TYPE_4)(I_load_4[c*HW * 4], I_load_4[c*HW * 4 + HW], I_load_4[c*HW * 4 + HW * 2], I_load_4[c*HW * 4 + HW * 3]) : (UNIT_TYPE_4)(UNIT_VAL_ZERO, UNIT_VAL_ZERO, UNIT_VAL_ZERO, UNIT_VAL_ZERO);
UNIT_TYPE_4 I5 = y5_in ? (UNIT_TYPE_4)(I_load_5[c*HW * 4], I_load_5[c*HW * 4 + HW], I_load_5[c*HW * 4 + HW * 2], I_load_5[c*HW * 4 + HW * 3]) : (UNIT_TYPE_4)(UNIT_VAL_ZERO, UNIT_VAL_ZERO, UNIT_VAL_ZERO, UNIT_VAL_ZERO);
UNIT_TYPE_4 I6 = y6_in ? (UNIT_TYPE_4)(I_load_6[c*HW * 4], I_load_6[c*HW * 4 + HW], I_load_6[c*HW * 4 + HW * 2], I_load_6[c*HW * 4 + HW * 3]) : (UNIT_TYPE_4)(UNIT_VAL_ZERO, UNIT_VAL_ZERO, UNIT_VAL_ZERO, UNIT_VAL_ZERO);
UNIT_TYPE_4 I7 = y7_in ? (UNIT_TYPE_4)(I_load_7[c*HW * 4], I_load_7[c*HW * 4 + HW], I_load_7[c*HW * 4 + HW * 2], I_load_7[c*HW * 4 + HW * 3]) : (UNIT_TYPE_4)(UNIT_VAL_ZERO, UNIT_VAL_ZERO, UNIT_VAL_ZERO, UNIT_VAL_ZERO);
#endif
//For winograd 6x3 the WA to scale input needed to be added, as the intermediate computations overflow in some cases
//Later on the output is adjusted with the same scale factor before adding bias and ACTIVATION
I0 = I0*scl_vec;
I1 = I1*scl_vec;
I2 = I2*scl_vec;
I3 = I3*scl_vec;
I4 = I4*scl_vec;
I5 = I5*scl_vec;
I6 = I6*scl_vec;
I7 = I7*scl_vec;
// Compute Winograd f6x3 data transform and store components in SLM.
V_write[0 * 64] = I0 - 5.25h*I2 + 5.25h*I4 - I6;
UNIT_TYPE_4 x0 = I1 - 4.25h*I3 + I5;
UNIT_TYPE_4 x1 = I2 - 4.25h*I4 + I6;
V_write[1 * 64] = x1 + x0;
V_write[2 * 64] = x1 - x0;
UNIT_TYPE_4 x2 = -5.h*I3 + I1;
UNIT_TYPE_4 x3 = 4.h*I5 + x2;
UNIT_TYPE_4 x4 = 0.25h*I2 + I6;
UNIT_TYPE_4 x5 = -1.25h*I4 + x4;
V_write[3 * 64] = +0.5h * x3 + x5;
V_write[4 * 64] = -0.5h * x3 + x5;
UNIT_TYPE_4 x6 = 4.h*I1 + I5;
UNIT_TYPE_4 x7 = -5.h*I3 + x6;
UNIT_TYPE_4 x8 = 4.h*I2 + I6;
UNIT_TYPE_4 x9 = -5.h*I4 + x8;
V_write[5 * 64] = +0.5h*x7 + x9;
V_write[6 * 64] = -0.5h*x7 + x9;
V_write[7 * 64] = -I1 + 5.25h*I3 - 5.25h*I5 + I7;
}
barrier(CLK_LOCAL_MEM_FENCE);
__local const UNIT_TYPE_8 *V_read_c16 = V_read;
__attribute__((opencl_unroll_hint(1)))
for (uint c16 = 0; c16 < 2
#ifndef FILTER_LAYOUT_IMAGE_2D_WEIGHTS_WINOGRAD_6x3_S1_FBXYB
&& coordU0.y < last_coord_y
#endif
; ++c16) {
// 2*14 * 3 * 8 = 672 MADs
// Fetch 16 channels of Winograd input components, spread across subgroup.
UNIT_TYPE_8 V0 = V_read_c16[0 * 16 + c16 * 256];
UNIT_TYPE_8 V8 = V_read_c16[1 * 16 + c16 * 256];
__attribute__((opencl_unroll_hint(2)))
for (int c8 = 0; c8 < 2 ; ++c8) {
// filter 0
// 2*14 * 3 * 4 = 336 MADs
const uint2 coordU = coordU0;
#if FILTER_LAYOUT_IMAGE_2D_WEIGHTS_WINOGRAD_6x3_S1_FBXYB
const uint WEIGHTWIDTH = FILTER_OFM_NUM*KCOLSW*KROWSW;
#else
const uint WEIGHTWIDTH = FILTER_OFM_NUM*KROWSW;
#endif
const uint flatA = coordU0.y*WEIGHTWIDTH + coordU0.x;
// Fetch 8 channels of Winograd components from f(k,s)
#if FILTER_LAYOUT_IMAGE_2D_WEIGHTS_WINOGRAD_6x3_S1_FBXYB || FILTER_LAYOUT_IMAGE_2D_WEIGHTS_WINOGRAD_6x3_S1_XFBYB
const UNIT_TYPE_8 f00 = as_half8(intel_sub_group_block_read_us8(U, (int2)(coordU0.x, coordU0.y)));
#else
const UNIT_TYPE_8 f00 = (UNIT_TYPE_8)(
as_half(intel_sub_group_block_read_us((__global unsigned short *)&U[flatA + 0 * WEIGHTWIDTH])),
as_half(intel_sub_group_block_read_us((__global unsigned short *)&U[flatA + 1 * WEIGHTWIDTH])),
as_half(intel_sub_group_block_read_us((__global unsigned short *)&U[flatA + 2 * WEIGHTWIDTH])),
as_half(intel_sub_group_block_read_us((__global unsigned short *)&U[flatA + 3 * WEIGHTWIDTH])),
as_half(intel_sub_group_block_read_us((__global unsigned short *)&U[flatA + 4 * WEIGHTWIDTH])),
as_half(intel_sub_group_block_read_us((__global unsigned short *)&U[flatA + 5 * WEIGHTWIDTH])),
as_half(intel_sub_group_block_read_us((__global unsigned short *)&U[flatA + 6 * WEIGHTWIDTH])),
as_half(intel_sub_group_block_read_us((__global unsigned short *)&U[flatA + 7 * WEIGHTWIDTH])));
#endif
// f0 x v[0 .. 14]
DOT8i_0(M0.s0, f00, V0, 0 + c8);
DOT8i_0(M0.s1, f00, V0, 2 + c8);
DOT8i_0(M1.s0, f00, V0, 4 + c8);
DOT8i_0(M1.s1, f00, V0, 6 + c8);
DOT8i_0(M2.s0, f00, V0, 8 + c8);
DOT8i_0(M2.s1, f00, V0, 10 + c8);
DOT8i_0(M3.s0, f00, V0, 12 + c8);
DOT8i_0(M3.s1, f00, V0, 14 + c8);
DOT8i_0(M4.s0, f00, V8, 0 + c8);
DOT8i_0(M4.s1, f00, V8, 2 + c8);
DOT8i_0(M5.s0, f00, V8, 4 + c8);
DOT8i_0(M5.s1, f00, V8, 6 + c8);
DOT8i_0(M6.s0, f00, V8, 8 + c8);
DOT8i_0(M6.s1, f00, V8, 10 + c8);
// f0 x v[0 .. 14]
DOT8i_1(M0.s0, f00, V0, 0 + c8);
DOT8i_1(M0.s1, f00, V0, 2 + c8);
DOT8i_1(M1.s0, f00, V0, 4 + c8);
DOT8i_1(M1.s1, f00, V0, 6 + c8);
DOT8i_1(M2.s0, f00, V0, 8 + c8);
DOT8i_1(M2.s1, f00, V0, 10 + c8);
DOT8i_1(M3.s0, f00, V0, 12 + c8);
DOT8i_1(M3.s1, f00, V0, 14 + c8);
DOT8i_1(M4.s0, f00, V8, 0 + c8);
DOT8i_1(M4.s1, f00, V8, 2 + c8);
DOT8i_1(M5.s0, f00, V8, 4 + c8);
DOT8i_1(M5.s1, f00, V8, 6 + c8);
DOT8i_1(M6.s0, f00, V8, 8 + c8);
DOT8i_1(M6.s1, f00, V8, 10 + c8);
// f0 x v[0 .. 14]
DOT8i_2(M0.s0, f00, V0, 0 + c8);
DOT8i_2(M0.s1, f00, V0, 2 + c8);
DOT8i_2(M1.s0, f00, V0, 4 + c8);
DOT8i_2(M1.s1, f00, V0, 6 + c8);
DOT8i_2(M2.s0, f00, V0, 8 + c8);
DOT8i_2(M2.s1, f00, V0, 10 + c8);
DOT8i_2(M3.s0, f00, V0, 12 + c8);
DOT8i_2(M3.s1, f00, V0, 14 + c8);
DOT8i_2(M4.s0, f00, V8, 0 + c8);
DOT8i_2(M4.s1, f00, V8, 2 + c8);
DOT8i_2(M5.s0, f00, V8, 4 + c8);
DOT8i_2(M5.s1, f00, V8, 6 + c8);
DOT8i_2(M6.s0, f00, V8, 8 + c8);
DOT8i_2(M6.s1, f00, V8, 10 + c8);
// f0 x v[0 .. 14]
DOT8i_3(M0.s0, f00, V0, 0 + c8);
DOT8i_3(M0.s1, f00, V0, 2 + c8);
DOT8i_3(M1.s0, f00, V0, 4 + c8);
DOT8i_3(M1.s1, f00, V0, 6 + c8);
DOT8i_3(M2.s0, f00, V0, 8 + c8);
DOT8i_3(M2.s1, f00, V0, 10 + c8);
DOT8i_3(M3.s0, f00, V0, 12 + c8);
DOT8i_3(M3.s1, f00, V0, 14 + c8);
DOT8i_3(M4.s0, f00, V8, 0 + c8);
DOT8i_3(M4.s1, f00, V8, 2 + c8);
DOT8i_3(M5.s0, f00, V8, 4 + c8);
DOT8i_3(M5.s1, f00, V8, 6 + c8);
DOT8i_3(M6.s0, f00, V8, 8 + c8);
DOT8i_3(M6.s1, f00, V8, 10 + c8);
// f0 x v[0 .. 14]
DOT8i_4(M0.s0, f00, V0, 0 + c8);
DOT8i_4(M0.s1, f00, V0, 2 + c8);
DOT8i_4(M1.s0, f00, V0, 4 + c8);
DOT8i_4(M1.s1, f00, V0, 6 + c8);
DOT8i_4(M2.s0, f00, V0, 8 + c8);
DOT8i_4(M2.s1, f00, V0, 10 + c8);
DOT8i_4(M3.s0, f00, V0, 12 + c8);
DOT8i_4(M3.s1, f00, V0, 14 + c8);
DOT8i_4(M4.s0, f00, V8, 0 + c8);
DOT8i_4(M4.s1, f00, V8, 2 + c8);
DOT8i_4(M5.s0, f00, V8, 4 + c8);
DOT8i_4(M5.s1, f00, V8, 6 + c8);
DOT8i_4(M6.s0, f00, V8, 8 + c8);
DOT8i_4(M6.s1, f00, V8, 10 + c8);
// f0 x v[0 .. 14]
DOT8i_5(M0.s0, f00, V0, 0 + c8);
DOT8i_5(M0.s1, f00, V0, 2 + c8);
DOT8i_5(M1.s0, f00, V0, 4 + c8);
DOT8i_5(M1.s1, f00, V0, 6 + c8);
DOT8i_5(M2.s0, f00, V0, 8 + c8);
DOT8i_5(M2.s1, f00, V0, 10 + c8);
DOT8i_5(M3.s0, f00, V0, 12 + c8);
DOT8i_5(M3.s1, f00, V0, 14 + c8);
DOT8i_5(M4.s0, f00, V8, 0 + c8);
DOT8i_5(M4.s1, f00, V8, 2 + c8);
DOT8i_5(M5.s0, f00, V8, 4 + c8);
DOT8i_5(M5.s1, f00, V8, 6 + c8);
DOT8i_5(M6.s0, f00, V8, 8 + c8);
DOT8i_5(M6.s1, f00, V8, 10 + c8);
// f0 x v[0 .. 14]
DOT8i_6(M0.s0, f00, V0, 0 + c8);
DOT8i_6(M0.s1, f00, V0, 2 + c8);
DOT8i_6(M1.s0, f00, V0, 4 + c8);
DOT8i_6(M1.s1, f00, V0, 6 + c8);
DOT8i_6(M2.s0, f00, V0, 8 + c8);
DOT8i_6(M2.s1, f00, V0, 10 + c8);
DOT8i_6(M3.s0, f00, V0, 12 + c8);
DOT8i_6(M3.s1, f00, V0, 14 + c8);
DOT8i_6(M4.s0, f00, V8, 0 + c8);
DOT8i_6(M4.s1, f00, V8, 2 + c8);
DOT8i_6(M5.s0, f00, V8, 4 + c8);
DOT8i_6(M5.s1, f00, V8, 6 + c8);
DOT8i_6(M6.s0, f00, V8, 8 + c8);
DOT8i_6(M6.s1, f00, V8, 10 + c8);
// f0 x v[0 .. 14]
DOT8i_7(M0.s0, f00, V0, 0 + c8);
DOT8i_7(M0.s1, f00, V0, 2 + c8);
DOT8i_7(M1.s0, f00, V0, 4 + c8);
DOT8i_7(M1.s1, f00, V0, 6 + c8);
DOT8i_7(M2.s0, f00, V0, 8 + c8);
DOT8i_7(M2.s1, f00, V0, 10 + c8);
DOT8i_7(M3.s0, f00, V0, 12 + c8);
DOT8i_7(M3.s1, f00, V0, 14 + c8);
DOT8i_7(M4.s0, f00, V8, 0 + c8);
DOT8i_7(M4.s1, f00, V8, 2 + c8);
DOT8i_7(M5.s0, f00, V8, 4 + c8);
DOT8i_7(M5.s1, f00, V8, 6 + c8);
DOT8i_7(M6.s0, f00, V8, 8 + c8);
DOT8i_7(M6.s1, f00, V8, 10 + c8);
#if FILTER_LAYOUT_IMAGE_2D_WEIGHTS_WINOGRAD_6x3_S1_FBXYB || FILTER_LAYOUT_IMAGE_2D_WEIGHTS_WINOGRAD_6x3_S1_XFBYB
const UNIT_TYPE_8 f01 = as_half8(intel_sub_group_block_read_us8(U, (int2)(coordU0.x + 16 * sizeof(UNIT_TYPE), coordU0.y)));
#else
const UNIT_TYPE_8 f01 = (UNIT_TYPE_8)(
as_half(intel_sub_group_block_read_us((__global unsigned short *)&U[flatA + 16 + 0 * WEIGHTWIDTH])),
as_half(intel_sub_group_block_read_us((__global unsigned short *)&U[flatA + 16 + 1 * WEIGHTWIDTH])),
as_half(intel_sub_group_block_read_us((__global unsigned short *)&U[flatA + 16 + 2 * WEIGHTWIDTH])),
as_half(intel_sub_group_block_read_us((__global unsigned short *)&U[flatA + 16 + 3 * WEIGHTWIDTH])),
as_half(intel_sub_group_block_read_us((__global unsigned short *)&U[flatA + 16 + 4 * WEIGHTWIDTH])),
as_half(intel_sub_group_block_read_us((__global unsigned short *)&U[flatA + 16 + 5 * WEIGHTWIDTH])),
as_half(intel_sub_group_block_read_us((__global unsigned short *)&U[flatA + 16 + 6 * WEIGHTWIDTH])),
as_half(intel_sub_group_block_read_us((__global unsigned short *)&U[flatA + 16 + 7 * WEIGHTWIDTH])));
#endif
// f1[c8] x v[1 .. 15]
DOT8i_0(M0.s0, f01, V0, 2 + c8);
DOT8i_0(M0.s1, f01, V0, 4 + c8);
DOT8i_0(M1.s0, f01, V0, 6 + c8);
DOT8i_0(M1.s1, f01, V0, 8 + c8);
DOT8i_0(M2.s0, f01, V0, 10 + c8);
DOT8i_0(M2.s1, f01, V0, 12 + c8);
DOT8i_0(M3.s0, f01, V0, 14 + c8);
DOT8i_0(M3.s1, f01, V8, 0 + c8);
DOT8i_0(M4.s0, f01, V8, 2 + c8);
DOT8i_0(M4.s1, f01, V8, 4 + c8);
DOT8i_0(M5.s0, f01, V8, 6 + c8);
DOT8i_0(M5.s1, f01, V8, 8 + c8);
DOT8i_0(M6.s0, f01, V8, 10 + c8);
DOT8i_0(M6.s1, f01, V8, 12 + c8);
// f1[c8] x v[1 .. 15]
DOT8i_1(M0.s0, f01, V0, 2 + c8);
DOT8i_1(M0.s1, f01, V0, 4 + c8);
DOT8i_1(M1.s0, f01, V0, 6 + c8);
DOT8i_1(M1.s1, f01, V0, 8 + c8);
DOT8i_1(M2.s0, f01, V0, 10 + c8);
DOT8i_1(M2.s1, f01, V0, 12 + c8);
DOT8i_1(M3.s0, f01, V0, 14 + c8);
DOT8i_1(M3.s1, f01, V8, 0 + c8);
DOT8i_1(M4.s0, f01, V8, 2 + c8);
DOT8i_1(M4.s1, f01, V8, 4 + c8);
DOT8i_1(M5.s0, f01, V8, 6 + c8);
DOT8i_1(M5.s1, f01, V8, 8 + c8);
DOT8i_1(M6.s0, f01, V8, 10 + c8);
DOT8i_1(M6.s1, f01, V8, 12 + c8);
// f1[c8] x v[1 .. 15]
DOT8i_2(M0.s0, f01, V0, 2 + c8);
DOT8i_2(M0.s1, f01, V0, 4 + c8);
DOT8i_2(M1.s0, f01, V0, 6 + c8);
DOT8i_2(M1.s1, f01, V0, 8 + c8);
DOT8i_2(M2.s0, f01, V0, 10 + c8);
DOT8i_2(M2.s1, f01, V0, 12 + c8);
DOT8i_2(M3.s0, f01, V0, 14 + c8);
DOT8i_2(M3.s1, f01, V8, 0 + c8);
DOT8i_2(M4.s0, f01, V8, 2 + c8);
DOT8i_2(M4.s1, f01, V8, 4 + c8);
DOT8i_2(M5.s0, f01, V8, 6 + c8);
DOT8i_2(M5.s1, f01, V8, 8 + c8);
DOT8i_2(M6.s0, f01, V8, 10 + c8);
DOT8i_2(M6.s1, f01, V8, 12 + c8);
// f1[c8] x v[1 .. 15]
DOT8i_3(M0.s0, f01, V0, 2 + c8);
DOT8i_3(M0.s1, f01, V0, 4 + c8);
DOT8i_3(M1.s0, f01, V0, 6 + c8);
DOT8i_3(M1.s1, f01, V0, 8 + c8);
DOT8i_3(M2.s0, f01, V0, 10 + c8);
DOT8i_3(M2.s1, f01, V0, 12 + c8);
DOT8i_3(M3.s0, f01, V0, 14 + c8);
DOT8i_3(M3.s1, f01, V8, 0 + c8);
DOT8i_3(M4.s0, f01, V8, 2 + c8);
DOT8i_3(M4.s1, f01, V8, 4 + c8);
DOT8i_3(M5.s0, f01, V8, 6 + c8);
DOT8i_3(M5.s1, f01, V8, 8 + c8);
DOT8i_3(M6.s0, f01, V8, 10 + c8);
DOT8i_3(M6.s1, f01, V8, 12 + c8);
// f1[c8] x v[1 .. 15]
DOT8i_4(M0.s0, f01, V0, 2 + c8);
DOT8i_4(M0.s1, f01, V0, 4 + c8);
DOT8i_4(M1.s0, f01, V0, 6 + c8);
DOT8i_4(M1.s1, f01, V0, 8 + c8);
DOT8i_4(M2.s0, f01, V0, 10 + c8);
DOT8i_4(M2.s1, f01, V0, 12 + c8);
DOT8i_4(M3.s0, f01, V0, 14 + c8);
DOT8i_4(M3.s1, f01, V8, 0 + c8);
DOT8i_4(M4.s0, f01, V8, 2 + c8);
DOT8i_4(M4.s1, f01, V8, 4 + c8);
DOT8i_4(M5.s0, f01, V8, 6 + c8);
DOT8i_4(M5.s1, f01, V8, 8 + c8);
DOT8i_4(M6.s0, f01, V8, 10 + c8);
DOT8i_4(M6.s1, f01, V8, 12 + c8);
// f1[c8] x v[1 .. 15]
DOT8i_5(M0.s0, f01, V0, 2 + c8);
DOT8i_5(M0.s1, f01, V0, 4 + c8);
DOT8i_5(M1.s0, f01, V0, 6 + c8);
DOT8i_5(M1.s1, f01, V0, 8 + c8);
DOT8i_5(M2.s0, f01, V0, 10 + c8);
DOT8i_5(M2.s1, f01, V0, 12 + c8);
DOT8i_5(M3.s0, f01, V0, 14 + c8);
DOT8i_5(M3.s1, f01, V8, 0 + c8);
DOT8i_5(M4.s0, f01, V8, 2 + c8);
DOT8i_5(M4.s1, f01, V8, 4 + c8);
DOT8i_5(M5.s0, f01, V8, 6 + c8);
DOT8i_5(M5.s1, f01, V8, 8 + c8);
DOT8i_5(M6.s0, f01, V8, 10 + c8);
DOT8i_5(M6.s1, f01, V8, 12 + c8);
// f1[c8] x v[1 .. 15]
DOT8i_6(M0.s0, f01, V0, 2 + c8);
DOT8i_6(M0.s1, f01, V0, 4 + c8);
DOT8i_6(M1.s0, f01, V0, 6 + c8);
DOT8i_6(M1.s1, f01, V0, 8 + c8);
DOT8i_6(M2.s0, f01, V0, 10 + c8);
DOT8i_6(M2.s1, f01, V0, 12 + c8);
DOT8i_6(M3.s0, f01, V0, 14 + c8);
DOT8i_6(M3.s1, f01, V8, 0 + c8);
DOT8i_6(M4.s0, f01, V8, 2 + c8);
DOT8i_6(M4.s1, f01, V8, 4 + c8);
DOT8i_6(M5.s0, f01, V8, 6 + c8);
DOT8i_6(M5.s1, f01, V8, 8 + c8);
DOT8i_6(M6.s0, f01, V8, 10 + c8);
DOT8i_6(M6.s1, f01, V8, 12 + c8);
// f1[c8] x v[1 .. 15]
DOT8i_7(M0.s0, f01, V0, 2 + c8);
DOT8i_7(M0.s1, f01, V0, 4 + c8);
DOT8i_7(M1.s0, f01, V0, 6 + c8);
DOT8i_7(M1.s1, f01, V0, 8 + c8);
DOT8i_7(M2.s0, f01, V0, 10 + c8);
DOT8i_7(M2.s1, f01, V0, 12 + c8);
DOT8i_7(M3.s0, f01, V0, 14 + c8);
DOT8i_7(M3.s1, f01, V8, 0 + c8);
DOT8i_7(M4.s0, f01, V8, 2 + c8);
DOT8i_7(M4.s1, f01, V8, 4 + c8);
DOT8i_7(M5.s0, f01, V8, 6 + c8);
DOT8i_7(M5.s1, f01, V8, 8 + c8);
DOT8i_7(M6.s0, f01, V8, 10 + c8);
DOT8i_7(M6.s1, f01, V8, 12 + c8);
#if FILTER_LAYOUT_IMAGE_2D_WEIGHTS_WINOGRAD_6x3_S1_FBXYB || FILTER_LAYOUT_IMAGE_2D_WEIGHTS_WINOGRAD_6x3_S1_XFBYB
const UNIT_TYPE_8 f02 = as_half8(intel_sub_group_block_read_us8(U, (int2)(coordU0.x + 32 * sizeof(UNIT_TYPE), coordU0.y)));
#else
const UNIT_TYPE_8 f02 = (UNIT_TYPE_8)(
as_half(intel_sub_group_block_read_us((__global unsigned short *)&U[flatA + 32 + 0 * WEIGHTWIDTH])),
as_half(intel_sub_group_block_read_us((__global unsigned short *)&U[flatA + 32 + 1 * WEIGHTWIDTH])),
as_half(intel_sub_group_block_read_us((__global unsigned short *)&U[flatA + 32 + 2 * WEIGHTWIDTH])),
as_half(intel_sub_group_block_read_us((__global unsigned short *)&U[flatA + 32 + 3 * WEIGHTWIDTH])),
as_half(intel_sub_group_block_read_us((__global unsigned short *)&U[flatA + 32 + 4 * WEIGHTWIDTH])),
as_half(intel_sub_group_block_read_us((__global unsigned short *)&U[flatA + 32 + 5 * WEIGHTWIDTH])),
as_half(intel_sub_group_block_read_us((__global unsigned short *)&U[flatA + 32 + 6 * WEIGHTWIDTH])),
as_half(intel_sub_group_block_read_us((__global unsigned short *)&U[flatA + 32 + 7 * WEIGHTWIDTH])));
#endif
coordU0.y += 8;
// f2[c8] x v[2 .. 16]
DOT8i_0(M0.s0, f02, V0, 4 + c8);
DOT8i_0(M0.s1, f02, V0, 6 + c8);
DOT8i_0(M1.s0, f02, V0, 8 + c8);
DOT8i_0(M1.s1, f02, V0, 10 + c8);
DOT8i_0(M2.s0, f02, V0, 12 + c8);
DOT8i_0(M2.s1, f02, V0, 14 + c8);
DOT8i_0(M3.s0, f02, V8, 0 + c8);
DOT8i_0(M3.s1, f02, V8, 2 + c8);
DOT8i_0(M4.s0, f02, V8, 4 + c8);
DOT8i_0(M4.s1, f02, V8, 6 + c8);
DOT8i_0(M5.s0, f02, V8, 8 + c8);
DOT8i_0(M5.s1, f02, V8, 10 + c8);
DOT8i_0(M6.s0, f02, V8, 12 + c8);
DOT8i_0(M6.s1, f02, V8, 14 + c8);
// f2[c8] x v[2 .. 16]
DOT8i_1(M0.s0, f02, V0, 4 + c8);
DOT8i_1(M0.s1, f02, V0, 6 + c8);
DOT8i_1(M1.s0, f02, V0, 8 + c8);
DOT8i_1(M1.s1, f02, V0, 10 + c8);
DOT8i_1(M2.s0, f02, V0, 12 + c8);
DOT8i_1(M2.s1, f02, V0, 14 + c8);
DOT8i_1(M3.s0, f02, V8, 0 + c8);
DOT8i_1(M3.s1, f02, V8, 2 + c8);
DOT8i_1(M4.s0, f02, V8, 4 + c8);
DOT8i_1(M4.s1, f02, V8, 6 + c8);
DOT8i_1(M5.s0, f02, V8, 8 + c8);
DOT8i_1(M5.s1, f02, V8, 10 + c8);
DOT8i_1(M6.s0, f02, V8, 12 + c8);
DOT8i_1(M6.s1, f02, V8, 14 + c8);
// f2[c8] x v[2 .. 16]
DOT8i_2(M0.s0, f02, V0, 4 + c8);
DOT8i_2(M0.s1, f02, V0, 6 + c8);
DOT8i_2(M1.s0, f02, V0, 8 + c8);
DOT8i_2(M1.s1, f02, V0, 10 + c8);
DOT8i_2(M2.s0, f02, V0, 12 + c8);
DOT8i_2(M2.s1, f02, V0, 14 + c8);
DOT8i_2(M3.s0, f02, V8, 0 + c8);
DOT8i_2(M3.s1, f02, V8, 2 + c8);
DOT8i_2(M4.s0, f02, V8, 4 + c8);
DOT8i_2(M4.s1, f02, V8, 6 + c8);
DOT8i_2(M5.s0, f02, V8, 8 + c8);
DOT8i_2(M5.s1, f02, V8, 10 + c8);
DOT8i_2(M6.s0, f02, V8, 12 + c8);
DOT8i_2(M6.s1, f02, V8, 14 + c8);
// f2[c8] x v[2 .. 16]
DOT8i_3(M0.s0, f02, V0, 4 + c8);
DOT8i_3(M0.s1, f02, V0, 6 + c8);
DOT8i_3(M1.s0, f02, V0, 8 + c8);
DOT8i_3(M1.s1, f02, V0, 10 + c8);
DOT8i_3(M2.s0, f02, V0, 12 + c8);
DOT8i_3(M2.s1, f02, V0, 14 + c8);
DOT8i_3(M3.s0, f02, V8, 0 + c8);
DOT8i_3(M3.s1, f02, V8, 2 + c8);
DOT8i_3(M4.s0, f02, V8, 4 + c8);
DOT8i_3(M4.s1, f02, V8, 6 + c8);
DOT8i_3(M5.s0, f02, V8, 8 + c8);
DOT8i_3(M5.s1, f02, V8, 10 + c8);
DOT8i_3(M6.s0, f02, V8, 12 + c8);
DOT8i_3(M6.s1, f02, V8, 14 + c8);
// f2[c8] x v[2 .. 16]
DOT8i_4(M0.s0, f02, V0, 4 + c8);
DOT8i_4(M0.s1, f02, V0, 6 + c8);
DOT8i_4(M1.s0, f02, V0, 8 + c8);
DOT8i_4(M1.s1, f02, V0, 10 + c8);
DOT8i_4(M2.s0, f02, V0, 12 + c8);
DOT8i_4(M2.s1, f02, V0, 14 + c8);
DOT8i_4(M3.s0, f02, V8, 0 + c8);
DOT8i_4(M3.s1, f02, V8, 2 + c8);
DOT8i_4(M4.s0, f02, V8, 4 + c8);
DOT8i_4(M4.s1, f02, V8, 6 + c8);
DOT8i_4(M5.s0, f02, V8, 8 + c8);
DOT8i_4(M5.s1, f02, V8, 10 + c8);
DOT8i_4(M6.s0, f02, V8, 12 + c8);
DOT8i_4(M6.s1, f02, V8, 14 + c8);
// f2[c8] x v[2 .. 16]
DOT8i_5(M0.s0, f02, V0, 4 + c8);
DOT8i_5(M0.s1, f02, V0, 6 + c8);
DOT8i_5(M1.s0, f02, V0, 8 + c8);
DOT8i_5(M1.s1, f02, V0, 10 + c8);
DOT8i_5(M2.s0, f02, V0, 12 + c8);
DOT8i_5(M2.s1, f02, V0, 14 + c8);
DOT8i_5(M3.s0, f02, V8, 0 + c8);
DOT8i_5(M3.s1, f02, V8, 2 + c8);
DOT8i_5(M4.s0, f02, V8, 4 + c8);
DOT8i_5(M4.s1, f02, V8, 6 + c8);
DOT8i_5(M5.s0, f02, V8, 8 + c8);
DOT8i_5(M5.s1, f02, V8, 10 + c8);
DOT8i_5(M6.s0, f02, V8, 12 + c8);
DOT8i_5(M6.s1, f02, V8, 14 + c8);
// f2[c8] x v[2 .. 16]
DOT8i_6(M0.s0, f02, V0, 4 + c8);
DOT8i_6(M0.s1, f02, V0, 6 + c8);
DOT8i_6(M1.s0, f02, V0, 8 + c8);
DOT8i_6(M1.s1, f02, V0, 10 + c8);
DOT8i_6(M2.s0, f02, V0, 12 + c8);
DOT8i_6(M2.s1, f02, V0, 14 + c8);
DOT8i_6(M3.s0, f02, V8, 0 + c8);
DOT8i_6(M3.s1, f02, V8, 2 + c8);
DOT8i_6(M4.s0, f02, V8, 4 + c8);
DOT8i_6(M4.s1, f02, V8, 6 + c8);
DOT8i_6(M5.s0, f02, V8, 8 + c8);
DOT8i_6(M5.s1, f02, V8, 10 + c8);
DOT8i_6(M6.s0, f02, V8, 12 + c8);
DOT8i_6(M6.s1, f02, V8, 14 + c8);
// f2[c8] x v[2 .. 16]
DOT8i_7(M0.s0, f02, V0, 4 + c8);
DOT8i_7(M0.s1, f02, V0, 6 + c8);
DOT8i_7(M1.s0, f02, V0, 8 + c8);
DOT8i_7(M1.s1, f02, V0, 10 + c8);
DOT8i_7(M2.s0, f02, V0, 12 + c8);
DOT8i_7(M2.s1, f02, V0, 14 + c8);
DOT8i_7(M3.s0, f02, V8, 0 + c8);
DOT8i_7(M3.s1, f02, V8, 2 + c8);
DOT8i_7(M4.s0, f02, V8, 4 + c8);
DOT8i_7(M4.s1, f02, V8, 6 + c8);
DOT8i_7(M5.s0, f02, V8, 8 + c8);
DOT8i_7(M5.s1, f02, V8, 10 + c8);
DOT8i_7(M6.s0, f02, V8, 12 + c8);
DOT8i_7(M6.s1, f02, V8, 14 + c8);
}
}
barrier(CLK_LOCAL_MEM_FENCE);
}
//barrier(CLK_LOCAL_MEM_FENCE);
// Store multiplies in SLM.
{
//barrier(CLK_LOCAL_MEM_FENCE);
__local UNIT_TYPE_2 *M_write = (__local UNIT_TYPE_2 *)&V[lz * 7 * 8];
M_write += lx;
M_write[0 * 16] = M0;
M_write[1 * 16] = M1;
M_write[2 * 16] = M2;
M_write[3 * 16] = M3;
M_write[4 * 16] = M4;
M_write[5 * 16] = M5;
M_write[6 * 16] = M6;
barrier(CLK_LOCAL_MEM_FENCE);
}
//if ((gz) % 2) return;
if (lz < 7)
{
// Load multiplies from SLM.
__local const UNIT_TYPE_2 *M_read = (__local UNIT_TYPE_2*)&V[lz * 8 ];
M_read += lx;
UNIT_TYPE_2 M0 = M_read[0 * 112];
UNIT_TYPE_2 M1 = M_read[1 * 112];
UNIT_TYPE_2 M2 = M_read[2 * 112];
UNIT_TYPE_2 M3 = M_read[3 * 112];
UNIT_TYPE_2 M4 = M_read[4 * 112];
UNIT_TYPE_2 M5 = M_read[5 * 112];
UNIT_TYPE_2 M6 = M_read[6 * 112];
UNIT_TYPE_2 M7 = M_read[7 * 112];
// Inverse Transform.
UNIT_TYPE_2 x0 = M1 + M2;
UNIT_TYPE_2 x1 = M1 - M2;
UNIT_TYPE_2 x2 = M3 + M4;
UNIT_TYPE_2 x3 = M3 - M4;
UNIT_TYPE_2 x4 = M5 + M6;
UNIT_TYPE_2 x5 = M5 - M6;
UNIT_TYPE_2 S0 = M0 + x0 + x2 + x4;
UNIT_TYPE_2 S1 = x1 + ((UNIT_TYPE)2.f)*x3 + ((UNIT_TYPE)0.5f)*x5;
UNIT_TYPE_2 S2 = x0 + ((UNIT_TYPE)4.f)*x2 + ((UNIT_TYPE)0.25f)*x4;
UNIT_TYPE_2 S3 = x1 + ((UNIT_TYPE)8.f)*x3 + ((UNIT_TYPE)0.125f)*x5;
UNIT_TYPE_2 S4 = x0 + ((UNIT_TYPE)16.f)*x2 + ((UNIT_TYPE)0.0625f)*x4;
UNIT_TYPE_2 S5 = x1 + ((UNIT_TYPE)32.f)*x3 + ((UNIT_TYPE)0.03125f)*x5 + M7;
// Store output to global memory.
uint p = gy * 6 + OUTPUT_PAD_BEFORE_SIZE_Y;
uint q = gx * 14 + lz * 2 + OUTPUT_PAD_BEFORE_SIZE_X;
uint k = gk * 16 + lx;
// bias and activation
#if BIAS_TERM
#if BIAS_PER_OUTPUT
const unsigned bias_index0 = k*OUTPUT_SIZE_X*OUTPUT_SIZE_Y + trow*OUTPUT_SIZE_X + q;
const unsigned bias_index1 = bias_index0 + 1;
#else
const unsigned bias_index0 = k;
const unsigned bias_index1 = bias_index0 + 1;
#endif
#endif
#if OUTPUT_LAYOUT_BYXF
uint outindex = gn*PQK + p*Q*FILTER_OFM_NUM + q*FILTER_OFM_NUM + k;
__global UNIT_TYPE *O_write = (__global UNIT_TYPE *)&O[outindex];
#else
__global UNIT_TYPE *O_write_0 = (__global UNIT_TYPE *)(&O[gn*PQK + k*Q*P + (p + 0)*Q + q]);
__global UNIT_TYPE *O_write_1 = (__global UNIT_TYPE *)(&O[gn*PQK + k*Q*P + (p + 1)*Q + q]);
__global UNIT_TYPE *O_write_2 = (__global UNIT_TYPE *)(&O[gn*PQK + k*Q*P + (p + 2)*Q + q]);
__global UNIT_TYPE *O_write_3 = (__global UNIT_TYPE *)(&O[gn*PQK + k*Q*P + (p + 3)*Q + q]);
__global UNIT_TYPE *O_write_4 = (__global UNIT_TYPE *)(&O[gn*PQK + k*Q*P + (p + 4)*Q + q]);
__global UNIT_TYPE *O_write_5 = (__global UNIT_TYPE *)(&O[gn*PQK + k*Q*P + (p + 5)*Q + q]);
#endif
// TODO: clip output by P, Q
bool q0_in = q < Q - OUTPUT_PAD_AFTER_SIZE_X;
bool q1_in = q + 1 < Q - OUTPUT_PAD_AFTER_SIZE_X;
const uint K = FILTER_OFM_NUM;
if (k < FILTER_OFM_NUM) {
if (p < P - OUTPUT_PAD_AFTER_SIZE_Y) {
if (q0_in) {
#if OUTPUT_LAYOUT_BYXF
#if BIAS_TERM
O_write[0 * QK + 0 * K] = ACTIVATION(S0.s0 * scl + bias[bias_index0], NL_M, NL_N);
#else
O_write[0 * QK + 0 * K] = ACTIVATION(S0.s0 * scl, NL_M, NL_N);
#endif
#else
#if BIAS_TERM
O_write_0[0] = ACTIVATION(S0.s0 * scl + bias[bias_index0], NL_M, NL_N);
#else
O_write_0[0] = ACTIVATION(S0.s0 * scl, NL_M, NL_N);
#endif
#endif
}
if (q1_in) {
#if OUTPUT_LAYOUT_BYXF
#if BIAS_TERM
O_write[0 * QK + 1 * K] = ACTIVATION(S0.s1 * scl + bias[bias_index0], NL_M, NL_N);
#else
O_write[0 * QK + 1 * K] = ACTIVATION(S0.s1 * scl, NL_M, NL_N);
#endif
#else
#if BIAS_TERM
O_write_0[1] = ACTIVATION(S0.s1 * scl + bias[bias_index0], NL_M, NL_N);
#else
O_write_0[1] = ACTIVATION(S0.s1 * scl, NL_M, NL_N);
#endif
#endif
}
}
// row 1
if (p + 1 < P - OUTPUT_PAD_AFTER_SIZE_Y) {
if (q0_in) {
#if OUTPUT_LAYOUT_BYXF
#if BIAS_TERM
O_write[1 * QK + 0 * K] = ACTIVATION(S1.s0 * scl + bias[bias_index0], NL_M, NL_N);
#else
O_write[1 * QK + 0 * K] = ACTIVATION(S1.s0 * scl, NL_M, NL_N);
#endif
#else
#if BIAS_TERM
O_write_1[0] = ACTIVATION(S1.s0 * scl + bias[bias_index0], NL_M, NL_N);
#else
O_write_1[0] = ACTIVATION(S1.s0 * scl, NL_M, NL_N);
#endif
#endif
}
if (q1_in) {
#if OUTPUT_LAYOUT_BYXF
#if BIAS_TERM
O_write[1 * QK + 1 * K] = ACTIVATION(S1.s1 * scl + bias[bias_index0], NL_M, NL_N);
#else
O_write[1 * QK + 1 * K] = ACTIVATION(S1.s1 * scl, NL_M, NL_N);
#endif
#else
#if BIAS_TERM
O_write_1[1] = ACTIVATION(S1.s1 * scl + bias[bias_index0], NL_M, NL_N);
#else
O_write_1[1] = ACTIVATION(S1.s1 * scl, NL_M, NL_N);
#endif
#endif
}
}
// row 2
if (p + 2 < P - OUTPUT_PAD_AFTER_SIZE_Y) {
if (q0_in) {
#if OUTPUT_LAYOUT_BYXF
#if BIAS_TERM
O_write[2 * QK + 0 * K] = ACTIVATION(S2.s0 * scl + bias[bias_index0], NL_M, NL_N);
#else
O_write[2 * QK + 0 * K] = ACTIVATION(S2.s0 * scl, NL_M, NL_N);
#endif
#else
#if BIAS_TERM
O_write_2[0] = ACTIVATION(S2.s0 * scl + bias[bias_index0], NL_M, NL_N);
#else
O_write_2[0] = ACTIVATION(S2.s0 * scl, NL_M, NL_N);
#endif
#endif
}
if (q1_in) {
#if OUTPUT_LAYOUT_BYXF
#if BIAS_TERM
O_write[2 * QK + 1 * K] = ACTIVATION(S2.s1 * scl + bias[bias_index0], NL_M, NL_N);
#else
O_write[2 * QK + 1 * K] = ACTIVATION(S2.s1 * scl, NL_M, NL_N);
#endif
#else
#if BIAS_TERM
O_write_2[1] = ACTIVATION(S2.s1 * scl + bias[bias_index0], NL_M, NL_N);
#else
O_write_2[1] = ACTIVATION(S2.s1 * scl, NL_M, NL_N);
#endif
#endif
}
}
// row 3
if (p + 3 < P - OUTPUT_PAD_AFTER_SIZE_Y) {
if (q0_in) {
#if OUTPUT_LAYOUT_BYXF
#if BIAS_TERM
O_write[3 * QK + 0 * K] = ACTIVATION(S3.s0 * scl + bias[bias_index0], NL_M, NL_N);
#else
O_write[3 * QK + 0 * K] = ACTIVATION(S3.s0 * scl, NL_M, NL_N);
#endif
#else
#if BIAS_TERM
O_write_3[0] = ACTIVATION(S3.s0 * scl + bias[bias_index0], NL_M, NL_N);
#else
O_write_3[0] = ACTIVATION(S3.s0 * scl, NL_M, NL_N);
#endif
#endif
}
if (q1_in) {
#if OUTPUT_LAYOUT_BYXF
#if BIAS_TERM
O_write[3 * QK + 1 * K] = ACTIVATION(S3.s1 * scl + bias[bias_index0], NL_M, NL_N);
#else
O_write[3 * QK + 1 * K] = ACTIVATION(S3.s1 * scl, NL_M, NL_N);
#endif
#else
#if BIAS_TERM
O_write_3[1] = ACTIVATION(S3.s1 * scl + bias[bias_index0], NL_M, NL_N);
#else
O_write_3[1] = ACTIVATION(S3.s1 * scl, NL_M, NL_N);
#endif
#endif
}
}
}
// row 4
if (p + 4 < P - OUTPUT_PAD_AFTER_SIZE_Y) {
if (q0_in) {
#if OUTPUT_LAYOUT_BYXF
#if BIAS_TERM
O_write[4 * QK + 0 * K] = ACTIVATION(S4.s0 * scl + bias[bias_index0], NL_M, NL_N);
#else
O_write[4 * QK + 0 * K] = ACTIVATION(S4.s0 * scl, NL_M, NL_N);
#endif
#else
#if BIAS_TERM
O_write_4[0] = ACTIVATION(S4.s0 * scl + bias[bias_index0], NL_M, NL_N);
#else
O_write_4[0] = ACTIVATION(S4.s0 * scl, NL_M, NL_N);
#endif
#endif
}
if (q1_in) {
#if OUTPUT_LAYOUT_BYXF
#if BIAS_TERM
O_write[4 * QK + 1 * K] = ACTIVATION(S4.s1 * scl + bias[bias_index0], NL_M, NL_N);
#else
O_write[4 * QK + 1 * K] = ACTIVATION(S4.s1 * scl, NL_M, NL_N);
#endif
#else
#if BIAS_TERM
O_write_4[1] = ACTIVATION(S4.s1 * scl + bias[bias_index0], NL_M, NL_N);
#else
O_write_4[1] = ACTIVATION(S4.s1 * scl, NL_M, NL_N);
#endif
#endif
}
}
// row 5
if (p + 5 < P - OUTPUT_PAD_AFTER_SIZE_Y) {
if (q0_in) {
#if OUTPUT_LAYOUT_BYXF
#if BIAS_TERM
O_write[5 * QK + 0 * K] = ACTIVATION(S5.s0 * scl + bias[bias_index0], NL_M, NL_N);
#else
O_write[5 * QK + 0 * K] = ACTIVATION(S5.s0 * scl, NL_M, NL_N);
#endif
#else
#if BIAS_TERM
O_write_5[0] = ACTIVATION(S5.s0 * scl + bias[bias_index0], NL_M, NL_N);
#else
O_write_5[0] = ACTIVATION(S5.s0 * scl, NL_M, NL_N);
#endif
#endif
}
if (q1_in) {
#if OUTPUT_LAYOUT_BYXF
#if BIAS_TERM
O_write[5 * QK + 1 * K] = ACTIVATION(S5.s1 * scl + bias[bias_index0], NL_M, NL_N);
#else
O_write[5 * QK + 1 * K] = ACTIVATION(S5.s1 * scl, NL_M, NL_N);
#endif
#else
#if BIAS_TERM
O_write_5[1] = ACTIVATION(S5.s1 * scl + bias[bias_index0], NL_M, NL_N);
#else
O_write_5[1] = ACTIVATION(S5.s1 * scl, NL_M, NL_N);
#endif
#endif
}
}
}
}
#undef UNIT_TYPE_2
#undef UNIT_TYPE_4
#undef UNIT_TYPE_8
|
