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
|
/*
* Copyright (c) 2017 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to
* deal in the Software without restriction, including without limitation the
* rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
* sell copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
#include "helpers.h"
/** This OpenCL kernel computes the "vector" 1x4 transposition of input matrix
*
* @param[in] src_ptr Pointer to the source matrix. Supported data types: F32
* @param[in] src_stride_x Stride of the source matrix in X dimension (in bytes)
* @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] src_stride_y Stride of the source matrix in Y dimension (in bytes)
* @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] src_offset_first_element_in_bytes The offset of the first element in the source matrix
* @param[out] dst_ptr Pointer to the destination matrix Supported data types: F32
* @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes)
* @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes)
* @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix
*/
__kernel void gemm_transpose1x4_f32(IMAGE_DECLARATION(src),
IMAGE_DECLARATION(dst))
{
uint x = get_global_id(0);
uint y = get_global_id(1);
/* Compute address for Matrix B - source */
Image src = CONVERT_TO_IMAGE_STRUCT(src);
/* Compute address for Matrix B transposed - destination. X and Y are swapped */
uint dst_addr_in_bytes = y * 16 + ((x * dst_stride_y + dst_offset_first_element_in_bytes));
float4 b0 = vload4(0, (__global float *)src.ptr);
vstore4(b0, 0, (__global float *)(dst_ptr + dst_addr_in_bytes));
}
/** This OpenCL kernel computes the "vector" 1x8 transposition of input matrix
*
* @param[in] src_ptr Pointer to the source matrix. Supported data types: F16
* @param[in] src_stride_x Stride of the source matrix in X dimension (in bytes)
* @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] src_stride_y Stride of the source matrix in Y dimension (in bytes)
* @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] src_offset_first_element_in_bytes The offset of the first element in the source matrix
* @param[out] dst_ptr Pointer to the destination matrix Supported data types: F16
* @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes)
* @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes)
* @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix
*/
__kernel void gemm_transpose1x8_f16(IMAGE_DECLARATION(src),
IMAGE_DECLARATION(dst))
{
uint x = get_global_id(0);
uint y = get_global_id(1);
/* Compute address for Matrix B - source */
Image src = CONVERT_TO_IMAGE_STRUCT(src);
/* Compute address for Matrix B transposed - destination. X and Y are swapped */
uint dst_addr_in_bytes = y * 16 + ((x * dst_stride_y + dst_offset_first_element_in_bytes));
half8 b0 = vload8(0, (__global half *)src.ptr);
vstore8(b0, 0, (__global half *)(dst_ptr + dst_addr_in_bytes));
}
/** This OpenCL kernel computes the "vector" 1x16 transposition of input matrix
*
* @param[in] src_ptr Pointer to the source matrix. Supported data types: U8
* @param[in] src_stride_x Stride of the source matrix in X dimension (in bytes)
* @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] src_stride_y Stride of the source matrix in Y dimension (in bytes)
* @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] src_offset_first_element_in_bytes The offset of the first element in the source matrix
* @param[out] dst_ptr Pointer to the destination matrix Supported data types: U8
* @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes)
* @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes)
* @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix
*/
__kernel void gemm_transpose1x16_u8(IMAGE_DECLARATION(src),
IMAGE_DECLARATION(dst))
{
uint x = get_global_id(0);
uint y = get_global_id(1);
/* Compute address for Matrix B - source */
Image src = CONVERT_TO_IMAGE_STRUCT(src);
/* Compute address for Matrix B transposed - destination. X and Y are swapped */
uint dst_addr_in_bytes = y * 16 + ((x * dst_stride_y + dst_offset_first_element_in_bytes));
uchar16 b0 = vload16(0, (__global uchar *)src.ptr);
vstore16(b0, 0, (__global uchar *)(dst_ptr + dst_addr_in_bytes));
}
/** This OpenCL kernel reshapes the input matrix transposing each 4x4 block and interleaving the values
*
* @param[in] src_ptr Pointer to the source matrix. Supported data types: U32/S32/F32
* @param[in] src_stride_x Stride of the source matrix in X dimension (in bytes)
* @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] src_stride_y Stride of the source matrix in Y dimension (in bytes)
* @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] src_offset_first_element_in_bytes The offset of the first element in the source matrix
* @param[out] dst_ptr Pointer to the destination matrix Supported data types: U32/S32/F32
* @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes)
* @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes)
* @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix
*/
__kernel void gemm_interleave4x4_32bit(IMAGE_DECLARATION(src),
IMAGE_DECLARATION(dst))
{
/* Compute source and destination addresses */
Image src = CONVERT_TO_IMAGE_STRUCT(src);
Image dst = CONVERT_TO_IMAGE_STRUCT(dst);
/* Load values from Matrix A */
float4 a0 = vload4(0, (__global float *)(offset(&src, 0, 0)));
float4 a1 = vload4(0, (__global float *)(offset(&src, 0, 1)));
float4 a2 = vload4(0, (__global float *)(offset(&src, 0, 2)));
float4 a3 = vload4(0, (__global float *)(offset(&src, 0, 3)));
float4 val0 = (float4)(a0.s0, a1.s0, a2.s0, a3.s0);
vstore4(val0, 0, ((__global float *)dst.ptr) + 0);
val0 = (float4)(a0.s1, a1.s1, a2.s1, a3.s1);
vstore4(val0, 0, ((__global float *)dst.ptr) + 4);
val0 = (float4)(a0.s2, a1.s2, a2.s2, a3.s2);
vstore4(val0, 0, ((__global float *)dst.ptr) + 8);
val0 = (float4)(a0.s3, a1.s3, a2.s3, a3.s3);
vstore4(val0, 0, ((__global float *)dst.ptr) + 12);
}
/** This OpenCL kernel reshapes the input matrix transposing each 4x4 block and interleaving the values
*
* @param[in] src_ptr Pointer to the source matrix. Supported data types: U16/S16/F16
* @param[in] src_stride_x Stride of the source matrix in X dimension (in bytes)
* @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] src_stride_y Stride of the source matrix in Y dimension (in bytes)
* @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] src_offset_first_element_in_bytes The offset of the first element in the source matrix
* @param[out] dst_ptr Pointer to the destination matrix Supported data types: U16/S16/F16
* @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes)
* @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes)
* @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix
*/
__kernel void gemm_interleave4x4_16bit(IMAGE_DECLARATION(src),
IMAGE_DECLARATION(dst))
{
/* Compute source and destination addresses */
Image src = CONVERT_TO_IMAGE_STRUCT(src);
Image dst = CONVERT_TO_IMAGE_STRUCT(dst);
/* Load values from Matrix A */
half8 a0 = vload8(0, (__global half *)(offset(&src, 0, 0)));
half8 a1 = vload8(0, (__global half *)(offset(&src, 0, 1)));
half8 a2 = vload8(0, (__global half *)(offset(&src, 0, 2)));
half8 a3 = vload8(0, (__global half *)(offset(&src, 0, 3)));
half8 val0 = (half8)((half4)(a0.s0, a1.s0, a2.s0, a3.s0), (half4)(a0.s1, a1.s1, a2.s1, a3.s1));
vstore8(val0, 0, ((__global half *)dst.ptr) + 0);
val0 = (half8)((half4)(a0.s2, a1.s2, a2.s2, a3.s2), (half4)(a0.s3, a1.s3, a2.s3, a3.s3));
vstore8(val0, 0, ((__global half *)dst.ptr) + 8);
val0 = (half8)((half4)(a0.s4, a1.s4, a2.s4, a3.s4), (half4)(a0.s5, a1.s5, a2.s5, a3.s5));
vstore8(val0, 0, ((__global half *)dst.ptr) + 16);
val0 = (half8)((half4)(a0.s6, a1.s6, a2.s6, a3.s6), (half4)(a0.s7, a1.s7, a2.s7, a3.s7));
vstore8(val0, 0, ((__global half *)dst.ptr) + 24);
}
/** This OpenCL kernel reshapes the input matrix transposing each 4x4 block and interleaving the values
*
* @param[in] src_ptr Pointer to the source matrix. Supported data types: U8/S8
* @param[in] src_stride_x Stride of the source matrix in X dimension (in bytes)
* @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] src_stride_y Stride of the source matrix in Y dimension (in bytes)
* @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] src_offset_first_element_in_bytes The offset of the first element in the source matrix
* @param[out] dst_ptr Pointer to the destination matrix Supported data types: U8/S8
* @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes)
* @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes)
* @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix
*/
__kernel void gemm_interleave4x4_8bit(IMAGE_DECLARATION(src),
IMAGE_DECLARATION(dst))
{
/* Compute source and destination addresses */
Image src = CONVERT_TO_IMAGE_STRUCT(src);
Image dst = CONVERT_TO_IMAGE_STRUCT(dst);
/* Load values from Matrix A */
uchar16 a0 = vload16(0, (__global uchar *)(offset(&src, 0, 0)));
uchar16 a1 = vload16(0, (__global uchar *)(offset(&src, 0, 1)));
uchar16 a2 = vload16(0, (__global uchar *)(offset(&src, 0, 2)));
uchar16 a3 = vload16(0, (__global uchar *)(offset(&src, 0, 3)));
uchar16 val0 = (uchar16)((uchar4)(a0.s0, a1.s0, a2.s0, a3.s0), (uchar4)(a0.s1, a1.s1, a2.s1, a3.s1),
(uchar4)(a0.s2, a1.s2, a2.s2, a3.s2), (uchar4)(a0.s3, a1.s3, a2.s3, a3.s3));
vstore16(val0, 0, ((__global uchar *)dst.ptr) + 0);
val0 = (uchar16)((uchar4)(a0.s4, a1.s4, a2.s4, a3.s4), (uchar4)(a0.s5, a1.s5, a2.s5, a3.s5),
(uchar4)(a0.s6, a1.s6, a2.s6, a3.s6), (uchar4)(a0.s7, a1.s7, a2.s7, a3.s7));
vstore16(val0, 0, ((__global uchar *)dst.ptr) + 16);
val0 = (uchar16)((uchar4)(a0.s8, a1.s8, a2.s8, a3.s8), (uchar4)(a0.s9, a1.s9, a2.s9, a3.s9),
(uchar4)(a0.sA, a1.sA, a2.sA, a3.sA), (uchar4)(a0.sB, a1.sB, a2.sB, a3.sB));
vstore16(val0, 0, ((__global uchar *)dst.ptr) + 32);
val0 = (uchar16)((uchar4)(a0.sC, a1.sC, a2.sC, a3.sC), (uchar4)(a0.sD, a1.sD, a2.sD, a3.sD),
(uchar4)(a0.sE, a1.sE, a2.sE, a3.sE), (uchar4)(a0.sF, a1.sF, a2.sF, a3.sF));
vstore16(val0, 0, ((__global uchar *)dst.ptr) + 48);
}
/** This kernel accumulates each row with the biases vector
*
* @param[in, out] accum_ptr Pointer to the accumulate tensor. Supported data type: F32
* @param[in] accum_stride_x Stride of the accmulate tensor in X dimension (in bytes)
* @param[in] accum_step_x accum_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] accum_stride_y Stride of the accumlulate tensor in Y dimension (in bytes)
* @param[in] accum_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] accum_offset_first_element_in_bytes The offset of the first element in the accumulate tensor
* @param[in] biases_ptr Pointer to the biases vector. Same as input.
* @param[in] biases_stride_x Stride of the destination tensor in X dimension (in bytes)
* @param[in] biases_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] biases_offset_first_element_in_bytes The offset of the first element in the destination tensor
*/
__kernel void gemm_accumulate_biases_f32(
IMAGE_DECLARATION(accum),
VECTOR_DECLARATION(biases))
{
Image accum = CONVERT_TO_IMAGE_STRUCT(accum);
Vector biases = CONVERT_TO_VECTOR_STRUCT(biases);
float4 accum_value = vload4(0, (__global float *)accum.ptr);
float4 biases_value = vload4(0, (__global float *)biases.ptr);
accum_value = biases_value + accum_value;
// Store result in the accummulate buffer
vstore4(accum_value, 0, (__global float *)accum.ptr);
}
/** This kernel accumulates each row with the biases vector
*
* @param[in, out] accum_ptr Pointer to the accumulate tensor. Supported data type: F16
* @param[in] accum_stride_x Stride of the accumulate tensor in X dimension (in bytes)
* @param[in] accum_step_x accum_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] accum_stride_y Stride of the accumlulate tensor in Y dimension (in bytes)
* @param[in] accum_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] accum_offset_first_element_in_bytes The offset of the first element in the accumulate tensor
* @param[in] biases_ptr Pointer to the biases vector. Same as input.
* @param[in] biases_stride_x Stride of the destination tensor in X dimension (in bytes)
* @param[in] biases_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] biases_offset_first_element_in_bytes The offset of the first element in the destination tensor
*/
__kernel void gemm_accumulate_biases_f16(
IMAGE_DECLARATION(accum),
VECTOR_DECLARATION(biases))
{
Image accum = CONVERT_TO_IMAGE_STRUCT(accum);
Vector biases = CONVERT_TO_VECTOR_STRUCT(biases);
half8 accum_value = vload8(0, (__global half *)accum.ptr);
half8 biases_value = vload8(0, (__global half *)biases.ptr);
accum_value = biases_value + accum_value;
// Store result in the accummulate buffer
vstore8(accum_value, 0, (__global half *)accum.ptr);
}
#if(defined WIDTH_MATRIX_B)
/** This OpenCL kernel computes the matrix multiplication between matrix A (src0) and matrix B (src1)
* Matrix A and matrix B must be reshaped respectively with @ref gemm_interleave4x4_u8 and @ref gemm_transpose1x16_u8 before running the matrix multiplication
*
* @attention The width of matrix B and the alpha's value need to be passed at compile time using -DWIDTH_MATRIX_B
*
* @param[in] src0_ptr Pointer to the source matrix. Supported formats: U8
* @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes)
* @param[in] src0_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] src0_stride_y Stride of the source matrix in Y dimension (in bytes)
* @param[in] src0_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] src0_offset_first_element_in_bytes The offset of the first element in the source matrix
* @param[in] src1_ptr Pointer to the source matrix. Supported formats: U8
* @param[in] src1_stride_x Stride of the source matrix in X dimension (in bytes)
* @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] src1_stride_y Stride of the source matrix in Y dimension (in bytes)
* @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source matrix
* @param[out] dst_ptr Pointer to the destination matrix Supported formats: U8
* @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes)
* @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes)
* @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix
* @param[in] a_offset Offset to be added to each element of the matrix A
* @param[in] b_offset Offset to be added to each element of the matrix B.
* @param[in] c_offset Offset to be added to each element of the matrix C.
* @param[in] c_mult_int Multiplied with each element of the matrix C.
* @param[in] shift Number of bits to shift right the result.
*/
__kernel void gemm_mm_u8(IMAGE_DECLARATION(src0),
IMAGE_DECLARATION(src1),
IMAGE_DECLARATION(dst),
int a_offset,
int b_offset,
int c_offset,
int c_mult_int,
int shift)
{
/* src_addr.s0 = address of matrix A */
/* src_addr.s1 = address of matrix B */
/* Compute address for matrix A and B */
int2 src_addr = (int2)(get_global_id(1), get_global_id(0)) * (int2)((src0_stride_y),
(src1_stride_y));
/* Add offset_first_element_in_bytes */
src_addr = src_addr + ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes));
/* Compute end row address for matrix B */
int end_row_mtx_b = src_addr.s1 + WIDTH_MATRIX_B;
/* Reset accumulators */
int16 c00 = 0.0f;
int16 c10 = 0.0f;
int16 c20 = 0.0f;
int16 c30 = 0.0f;
for(; src_addr.s1 <= (end_row_mtx_b - 8); src_addr += (int2)(8, 32))
{
/* Load values from matrix A (interleaved) and matrix B (transposed) */
int8 a0 = (int8)a_offset + convert_int8(vload8(0, ((__global uchar *)src0_ptr) + src_addr.s0));
int16 b0 = (int16)b_offset + convert_int16(vload16(0, ((__global uchar *)src1_ptr) + src_addr.s1));
c00 += (int16)a0.s0 * b0;
c10 += (int16)a0.s1 * b0;
c20 += (int16)a0.s2 * b0;
c30 += (int16)a0.s3 * b0;
int16 b1 = (int16)b_offset + convert_int16(vload16(0, ((__global uchar *)src1_ptr) + src_addr.s1 + 16));
c00 += (int16)a0.s4 * b1;
c10 += (int16)a0.s5 * b1;
c20 += (int16)a0.s6 * b1;
c30 += (int16)a0.s7 * b1;
}
for(; src_addr.s1 < end_row_mtx_b; src_addr += (int2)(4, 16))
{
/* Load values from matrix A (interleaved) and matrix B (transposed) */
int4 a0 = (int4)a_offset + convert_int4(vload4(0, ((__global uchar *)src0_ptr) + src_addr.s0));
int16 b0 = (int16)b_offset + convert_int16(vload16(0, ((__global uchar *)src1_ptr) + src_addr.s1));
c00 += (int16)a0.s0 * b0;
c10 += (int16)a0.s1 * b0;
c20 += (int16)a0.s2 * b0;
c30 += (int16)a0.s3 * b0;
}
/* Compute destination address */
Image dst = CONVERT_TO_IMAGE_STRUCT(dst);
/* Multiply by the weight of matrix product */
c00 = (((int16)c_offset + c00) * (int16)c_mult_int) >> shift;
c10 = (((int16)c_offset + c10) * (int16)c_mult_int) >> shift;
c20 = (((int16)c_offset + c20) * (int16)c_mult_int) >> shift;
c30 = (((int16)c_offset + c30) * (int16)c_mult_int) >> shift;
/* Store 4x16 block */
vstore16(convert_uchar16_sat(c00), 0, (__global uchar *)(offset(&dst, 0, 0)));
vstore16(convert_uchar16_sat(c10), 0, (__global uchar *)(offset(&dst, 0, 1)));
vstore16(convert_uchar16_sat(c20), 0, (__global uchar *)(offset(&dst, 0, 2)));
vstore16(convert_uchar16_sat(c30), 0, (__global uchar *)(offset(&dst, 0, 3)));
}
#endif
#if(defined WIDTH_MATRIX_B && defined ALPHA)
/** This OpenCL kernel is optimised for Midgard. It computes the matrix multiplication between matrix A (src0) and matrix B (src1)
* Matrix A and matrix B must be reshaped respectively with @ref gemm_interleave4x4_f32 and @ref gemm_transpose1x4_f32 before running the matrix multiplication
*
* @attention The width of matrix B and the alpha's value need to be passed at compile time using -DWIDTH_MATRIX_B and -DALPHA
*
* @param[in] src0_ptr Pointer to the source matrix. Supported data types: F32
* @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes)
* @param[in] src0_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] src0_stride_y Stride of the source matrix in Y dimension (in bytes)
* @param[in] src0_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] src0_offset_first_element_in_bytes The offset of the first element in the source matrix
* @param[in] src1_ptr Pointer to the source matrix. Supported data types: F32
* @param[in] src1_stride_x Stride of the source matrix in X dimension (in bytes)
* @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] src1_stride_y Stride of the source matrix in Y dimension (in bytes)
* @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source matrix
* @param[out] dst_ptr Pointer to the destination matrix Supported data types: F32
* @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes)
* @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes)
* @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix
*/
__kernel void gemm_mm_f32_midgard(IMAGE_DECLARATION(src0),
IMAGE_DECLARATION(src1),
IMAGE_DECLARATION(dst))
{
/* src_addr.s0 = address of matrix A */
/* src_addr.s1 = address of matrix B */
/* Compute address for matrix A and B */
int2 src_addr = (int2)(get_global_id(1), get_global_id(0)) * (int2)((src0_stride_y),
(src1_stride_y));
/* Add offset_first_element_in_bytes */
src_addr = src_addr + ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes));
/* Divide by 4 in order to get the src_addr in unit of float */
src_addr = src_addr >> 2;
/* Compute end row address for matrix B */
int end_row_mtx_b = src_addr.s1 + WIDTH_MATRIX_B;
/* Reset accumulators */
float4 c00 = 0.0f;
float4 c10 = 0.0f;
float4 c20 = 0.0f;
float4 c30 = 0.0f;
for(; src_addr.s1 <= (end_row_mtx_b - 8); src_addr += (int2)(8, 8))
{
/* Load values from matrix A (interleaved) and matrix B (transposed) */
float4 a0 = vload4(0, ((__global float *)src0_ptr) + src_addr.s0);
float4 b0 = vload4(0, ((__global float *)src1_ptr) + src_addr.s1);
c00 += (float4)a0.s0 * b0;
c10 += (float4)a0.s1 * b0;
c20 += (float4)a0.s2 * b0;
c30 += (float4)a0.s3 * b0;
/* Load values from matrix A (interleaved) and matrix B (transposed) */
a0 = vload4(0, ((__global float *)src0_ptr) + src_addr.s0 + 4);
b0 = vload4(0, ((__global float *)src1_ptr) + src_addr.s1 + 4);
c00 += (float4)a0.s0 * b0;
c10 += (float4)a0.s1 * b0;
c20 += (float4)a0.s2 * b0;
c30 += (float4)a0.s3 * b0;
}
for(; src_addr.s1 < end_row_mtx_b; src_addr += (int2)(4, 4))
{
/* Load values from matrix A (interleaved) and matrix B (transposed) */
float4 a0 = vload4(0, ((__global float *)src0_ptr) + src_addr.s0);
float4 b0 = vload4(0, ((__global float *)src1_ptr) + src_addr.s1);
c00 += (float4)a0.s0 * b0;
c10 += (float4)a0.s1 * b0;
c20 += (float4)a0.s2 * b0;
c30 += (float4)a0.s3 * b0;
}
/* Compute destination address */
Image dst = CONVERT_TO_IMAGE_STRUCT(dst);
/* Multiply by the weight of matrix product */
c00 = c00 * (float4)ALPHA;
c10 = c10 * (float4)ALPHA;
c20 = c20 * (float4)ALPHA;
c30 = c30 * (float4)ALPHA;
/* Store 4x4 block */
vstore4(c00, 0, (__global float *)(offset(&dst, 0, 0)));
vstore4(c10, 0, (__global float *)(offset(&dst, 0, 1)));
vstore4(c20, 0, (__global float *)(offset(&dst, 0, 2)));
vstore4(c30, 0, (__global float *)(offset(&dst, 0, 3)));
}
/** This OpenCL kernel is optimised for Bifrost. It computes the matrix multiplication between matrix A (src0) and matrix B (src1)
* Matrix A and matrix B must be reshaped respectively with @ref gemm_interleave4x4_f32 and @ref gemm_transpose1x4_f32 before running the matrix multiplication
*
* @attention The width of matrix B and the alpha's value need to be passed at compile time using -DWIDTH_MATRIX_B and -DALPHA
*
* @param[in] src0_ptr Pointer to the source matrix. Supported data types: F32
* @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes)
* @param[in] src0_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] src0_stride_y Stride of the source matrix in Y dimension (in bytes)
* @param[in] src0_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] src0_offset_first_element_in_bytes The offset of the first element in the source matrix
* @param[in] src1_ptr Pointer to the source matrix. Supported data types: F32
* @param[in] src1_stride_x Stride of the source matrix in X dimension (in bytes)
* @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] src1_stride_y Stride of the source matrix in Y dimension (in bytes)
* @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source matrix
* @param[out] dst_ptr Pointer to the destination matrix Supported data types: F32
* @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes)
* @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes)
* @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix
*/
__kernel void gemm_mm_f32_bifrost(IMAGE_DECLARATION(src0),
IMAGE_DECLARATION(src1),
IMAGE_DECLARATION(dst))
{
// src_addr_a = address of matrix A
// src_addr_b = address of matrix B
__global float *src_addr_a = (__global float *)(src0_ptr + get_global_id(1) * src0_stride_y + src0_offset_first_element_in_bytes);
__global float *src_addr_b = (__global float *)(src1_ptr + get_global_id(0) * src1_stride_y + src1_offset_first_element_in_bytes);
// Compute end row address for matrix B
__global float *src_end_addr_b = src_addr_b + WIDTH_MATRIX_B;
// Reset accumulators
float c00 = 0.0f;
float c01 = 0.0f;
float c02 = 0.0f;
float c03 = 0.0f;
float c10 = 0.0f;
float c11 = 0.0f;
float c12 = 0.0f;
float c13 = 0.0f;
float c20 = 0.0f;
float c21 = 0.0f;
float c22 = 0.0f;
float c23 = 0.0f;
float c30 = 0.0f;
float c31 = 0.0f;
float c32 = 0.0f;
float c33 = 0.0f;
for(; src_addr_b <= (src_end_addr_b - 16); src_addr_a += 16, src_addr_b += 16)
{
// Load values from matrix A (interleaved) and matrix B (transposed)
float4 a0 = vload4(0, src_addr_a);
float4 b0 = vload4(0, src_addr_b);
c00 = fma(a0.s0, b0.s0, c00);
c01 = fma(a0.s0, b0.s1, c01);
c02 = fma(a0.s0, b0.s2, c02);
c03 = fma(a0.s0, b0.s3, c03);
c10 = fma(a0.s1, b0.s0, c10);
c11 = fma(a0.s1, b0.s1, c11);
c12 = fma(a0.s1, b0.s2, c12);
c13 = fma(a0.s1, b0.s3, c13);
c20 = fma(a0.s2, b0.s0, c20);
c21 = fma(a0.s2, b0.s1, c21);
c22 = fma(a0.s2, b0.s2, c22);
c23 = fma(a0.s2, b0.s3, c23);
c30 = fma(a0.s3, b0.s0, c30);
c31 = fma(a0.s3, b0.s1, c31);
c32 = fma(a0.s3, b0.s2, c32);
c33 = fma(a0.s3, b0.s3, c33);
// Load values from matrix A (interleaved) and matrix B (transposed)
a0 = vload4(0, src_addr_a + 4);
b0 = vload4(0, src_addr_b + 4);
c00 = fma(a0.s0, b0.s0, c00);
c01 = fma(a0.s0, b0.s1, c01);
c02 = fma(a0.s0, b0.s2, c02);
c03 = fma(a0.s0, b0.s3, c03);
c10 = fma(a0.s1, b0.s0, c10);
c11 = fma(a0.s1, b0.s1, c11);
c12 = fma(a0.s1, b0.s2, c12);
c13 = fma(a0.s1, b0.s3, c13);
c20 = fma(a0.s2, b0.s0, c20);
c21 = fma(a0.s2, b0.s1, c21);
c22 = fma(a0.s2, b0.s2, c22);
c23 = fma(a0.s2, b0.s3, c23);
c30 = fma(a0.s3, b0.s0, c30);
c31 = fma(a0.s3, b0.s1, c31);
c32 = fma(a0.s3, b0.s2, c32);
c33 = fma(a0.s3, b0.s3, c33);
// Load values from matrix A (interleaved) and matrix B (transposed)
a0 = vload4(0, src_addr_a + 8);
b0 = vload4(0, src_addr_b + 8);
c00 = fma(a0.s0, b0.s0, c00);
c01 = fma(a0.s0, b0.s1, c01);
c02 = fma(a0.s0, b0.s2, c02);
c03 = fma(a0.s0, b0.s3, c03);
c10 = fma(a0.s1, b0.s0, c10);
c11 = fma(a0.s1, b0.s1, c11);
c12 = fma(a0.s1, b0.s2, c12);
c13 = fma(a0.s1, b0.s3, c13);
c20 = fma(a0.s2, b0.s0, c20);
c21 = fma(a0.s2, b0.s1, c21);
c22 = fma(a0.s2, b0.s2, c22);
c23 = fma(a0.s2, b0.s3, c23);
c30 = fma(a0.s3, b0.s0, c30);
c31 = fma(a0.s3, b0.s1, c31);
c32 = fma(a0.s3, b0.s2, c32);
c33 = fma(a0.s3, b0.s3, c33);
// Load values from matrix A (interleaved) and matrix B (transposed)
a0 = vload4(0, src_addr_a + 12);
b0 = vload4(0, src_addr_b + 12);
c00 = fma(a0.s0, b0.s0, c00);
c01 = fma(a0.s0, b0.s1, c01);
c02 = fma(a0.s0, b0.s2, c02);
c03 = fma(a0.s0, b0.s3, c03);
c10 = fma(a0.s1, b0.s0, c10);
c11 = fma(a0.s1, b0.s1, c11);
c12 = fma(a0.s1, b0.s2, c12);
c13 = fma(a0.s1, b0.s3, c13);
c20 = fma(a0.s2, b0.s0, c20);
c21 = fma(a0.s2, b0.s1, c21);
c22 = fma(a0.s2, b0.s2, c22);
c23 = fma(a0.s2, b0.s3, c23);
c30 = fma(a0.s3, b0.s0, c30);
c31 = fma(a0.s3, b0.s1, c31);
c32 = fma(a0.s3, b0.s2, c32);
c33 = fma(a0.s3, b0.s3, c33);
}
for(; src_addr_b < src_end_addr_b; src_addr_a += 4, src_addr_b += 4)
{
// Load values from matrix A (interleaved) and matrix B (transposed)
float4 a0 = vload4(0, src_addr_a);
float4 b0 = vload4(0, src_addr_b);
c00 = fma(a0.s0, b0.s0, c00);
c01 = fma(a0.s0, b0.s1, c01);
c02 = fma(a0.s0, b0.s2, c02);
c03 = fma(a0.s0, b0.s3, c03);
c10 = fma(a0.s1, b0.s0, c10);
c11 = fma(a0.s1, b0.s1, c11);
c12 = fma(a0.s1, b0.s2, c12);
c13 = fma(a0.s1, b0.s3, c13);
c20 = fma(a0.s2, b0.s0, c20);
c21 = fma(a0.s2, b0.s1, c21);
c22 = fma(a0.s2, b0.s2, c22);
c23 = fma(a0.s2, b0.s3, c23);
c30 = fma(a0.s3, b0.s0, c30);
c31 = fma(a0.s3, b0.s1, c31);
c32 = fma(a0.s3, b0.s2, c32);
c33 = fma(a0.s3, b0.s3, c33);
}
// Compute destination address
Image dst = CONVERT_TO_IMAGE_STRUCT(dst);
// Multiply by the weight of matrix product
c00 = c00 * ALPHA;
c01 = c01 * ALPHA;
c02 = c02 * ALPHA;
c03 = c03 * ALPHA;
c10 = c10 * ALPHA;
c11 = c11 * ALPHA;
c12 = c12 * ALPHA;
c13 = c13 * ALPHA;
c20 = c20 * ALPHA;
c21 = c21 * ALPHA;
c22 = c22 * ALPHA;
c23 = c23 * ALPHA;
c30 = c30 * ALPHA;
c31 = c31 * ALPHA;
c32 = c32 * ALPHA;
c33 = c33 * ALPHA;
barrier(CLK_GLOBAL_MEM_FENCE);
// Store 4x4 block
vstore4((float4)(c00, c01, c02, c03), 0, (__global float *)(offset(&dst, 0, 0)));
vstore4((float4)(c10, c11, c12, c13), 0, (__global float *)(offset(&dst, 0, 1)));
vstore4((float4)(c20, c21, c22, c23), 0, (__global float *)(offset(&dst, 0, 2)));
vstore4((float4)(c30, c31, c32, c33), 0, (__global float *)(offset(&dst, 0, 3)));
}
/** This OpenCL kernel computes the matrix multiplication between matrix A (src0) and matrix B (src1)
* Matrix A and matrix B must be reshaped respectively with @ref gemm_interleave4x4_f16 and @ref gemm_transpose1x8_f16 before running the matrix multiplication
*
* @attention The width of matrix B and the alpha's value need to be passed at compile time using -DWIDTH_MATRIX_B and -DALPHA
*
* @param[in] src0_ptr Pointer to the source matrix. Supported data types: F16
* @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes)
* @param[in] src0_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] src0_stride_y Stride of the source matrix in Y dimension (in bytes)
* @param[in] src0_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] src0_offset_first_element_in_bytes The offset of the first element in the source matrix
* @param[in] src1_ptr Pointer to the source matrix. Supported data types: F16
* @param[in] src1_stride_x Stride of the source matrix in X dimension (in bytes)
* @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] src1_stride_y Stride of the source matrix in Y dimension (in bytes)
* @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source matrix
* @param[out] dst_ptr Pointer to the destination matrix Supported data types: F16
* @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes)
* @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes)
* @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix
*/
__kernel void gemm_mm_f16(IMAGE_DECLARATION(src0),
IMAGE_DECLARATION(src1),
IMAGE_DECLARATION(dst))
{
/* src_addr.s0 = address of matrix A */
/* src_addr.s1 = address of matrix B */
/* Compute address for matrix A and B */
int2 src_addr = (int2)(get_global_id(1), get_global_id(0)) * (int2)((src0_stride_y),
(src1_stride_y));
/* Add offset_first_element_in_bytes */
src_addr = src_addr + ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes));
/* Divide by 2 in order to get the src_addr in unit of half */
src_addr = src_addr >> 1;
/* Compute end row address for matrix B */
int end_row_mtx_b = src_addr.s1 + WIDTH_MATRIX_B;
/* Reset accumulators */
half8 c00 = 0.0f;
half8 c10 = 0.0f;
half8 c20 = 0.0f;
half8 c30 = 0.0f;
for(; src_addr.s1 <= (end_row_mtx_b - 8); src_addr += (int2)(8, 16))
{
/* Load values from matrix A (interleaved) and matrix B (transposed) */
half4 a0 = vload4(0, ((__global half *)src0_ptr) + src_addr.s0);
half8 b0 = vload8(0, ((__global half *)src1_ptr) + src_addr.s1);
c00 += (half8)a0.s0 * b0;
c10 += (half8)a0.s1 * b0;
c20 += (half8)a0.s2 * b0;
c30 += (half8)a0.s3 * b0;
/* Load values from matrix A (interleaved) and matrix B (transposed) */
a0 = vload4(0, ((__global half *)src0_ptr) + src_addr.s0 + 4);
b0 = vload8(0, ((__global half *)src1_ptr) + src_addr.s1 + 8);
c00 += (half8)a0.s0 * b0;
c10 += (half8)a0.s1 * b0;
c20 += (half8)a0.s2 * b0;
c30 += (half8)a0.s3 * b0;
}
for(; src_addr.s1 < end_row_mtx_b; src_addr += (int2)(4, 8))
{
/* Load values from matrix A (interleaved) and matrix B (transposed) */
half4 a0 = vload4(0, ((__global half *)src0_ptr) + src_addr.s0);
half8 b0 = vload8(0, ((__global half *)src1_ptr) + src_addr.s1);
c00 += (half8)a0.s0 * b0;
c10 += (half8)a0.s1 * b0;
c20 += (half8)a0.s2 * b0;
c30 += (half8)a0.s3 * b0;
}
/* Compute destination address */
Image dst = CONVERT_TO_IMAGE_STRUCT(dst);
/* Multiply by the weight of matrix product */
c00 = c00 * (half8)ALPHA;
c10 = c10 * (half8)ALPHA;
c20 = c20 * (half8)ALPHA;
c30 = c30 * (half8)ALPHA;
/* Store 4x8 block */
vstore8(c00, 0, (__global half *)(offset(&dst, 0, 0)));
vstore8(c10, 0, (__global half *)(offset(&dst, 0, 1)));
vstore8(c20, 0, (__global half *)(offset(&dst, 0, 2)));
vstore8(c30, 0, (__global half *)(offset(&dst, 0, 3)));
}
#if(defined WIDTH_VECTOR_A)
/** This OpenCL kernel computes the vector by matrix multiplication between the vector A (src0) and matrix B (src1)
*
* @attention The width of vector A, the width of matrix B and the alpha's value need to be passed at compile time using -DWIDTH_VECTOR_A -DWIDTH_MATRIX_B and -DALPHA
*
* @attention The input vector A and matrix B must not be reshaped
*
* @param[in] src0_ptr Pointer to the source matrix. Supported data types: F32
* @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes)
* @param[in] src0_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] src0_stride_y Stride of the source matrix in Y dimension (in bytes)
* @param[in] src0_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] src0_offset_first_element_in_bytes The offset of the first element in the source matrix
* @param[in] src1_ptr Pointer to the source matrix. Supported data types: F32
* @param[in] src1_stride_x Stride of the source matrix in X dimension (in bytes)
* @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] src1_stride_y Stride of the source matrix in Y dimension (in bytes)
* @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source matrix
* @param[out] dst_ptr Pointer to the destination matrix Supported data types: F32
* @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes)
* @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes)
* @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix
*/
__kernel void gemm_vm_f32(IMAGE_DECLARATION(src0),
IMAGE_DECLARATION(src1),
IMAGE_DECLARATION(dst))
{
int idx = get_global_id(0) * 4;
/* Compute the address for the vector A and matrix B */
int2 src_addr = ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes));
src_addr.s1 += idx * sizeof(float);
int end_row_vec_a = src_addr.s0 + (WIDTH_VECTOR_A * sizeof(float));
float4 acc = 0.0f;
for(; src_addr.s0 <= (end_row_vec_a - 2 * sizeof(float)); src_addr += (int2)(2 * sizeof(float), 2 * src1_stride_y))
{
float2 a0 = vload2(0, (__global float *)(src0_ptr + src_addr.s0));
float4 b0 = vload4(0, (__global float *)(src1_ptr + src_addr.s1));
float4 b1 = vload4(0, (__global float *)(src1_ptr + src_addr.s1 + src1_stride_y));
acc += b0 * (float4)a0.s0;
acc += b1 * (float4)a0.s1;
}
for(; src_addr.s0 < end_row_vec_a; src_addr += (int2)(sizeof(float), src1_stride_y))
{
float a0 = *((__global float *)(src0_ptr + src_addr.s0));
float4 b0 = vload4(0, (__global float *)(src1_ptr + src_addr.s1));
acc += b0 * (float4)a0;
}
/* Compute destination address */
Image dst = CONVERT_TO_IMAGE_STRUCT(dst);
/* Multiply by the weight of vector-matrix product */
acc = acc * (float4)ALPHA;
vstore4(acc, 0, (__global float *)(offset(&dst, 0, 0)));
}
/** This OpenCL kernel computes the vector by matrix multiplication between the vector A (src0) and matrix B (src1)
*
* @attention The width of vector A, the width of matrix B and the alpha's value need to be passed at compile time using -DWIDTH_VECTOR_A -DWIDTH_MATRIX_B and -DALPHA
*
* @attention The input vector A and matrix B must not be reshaped
*
* @param[in] src0_ptr Pointer to the source matrix. Supported data types: F16
* @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes)
* @param[in] src0_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] src0_stride_y Stride of the source matrix in Y dimension (in bytes)
* @param[in] src0_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] src0_offset_first_element_in_bytes The offset of the first element in the source matrix
* @param[in] src1_ptr Pointer to the source matrix. Supported data types: F16
* @param[in] src1_stride_x Stride of the source matrix in X dimension (in bytes)
* @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] src1_stride_y Stride of the source matrix in Y dimension (in bytes)
* @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source matrix
* @param[out] dst_ptr Pointer to the destination matrix Supported data types: F16
* @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes)
* @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes)
* @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix
*/
__kernel void gemm_vm_f16(IMAGE_DECLARATION(src0),
IMAGE_DECLARATION(src1),
IMAGE_DECLARATION(dst))
{
int idx = get_global_id(0) * 8;
/* Compute the address for the vector A and matrix B */
int2 src_addr = ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes));
src_addr.s1 += idx * sizeof(half);
int end_row_vec_a = src_addr.s0 + (WIDTH_VECTOR_A * sizeof(half));
half8 acc = 0.0f;
for(; src_addr.s0 <= (end_row_vec_a - 4 * sizeof(half)); src_addr += (int2)(4 * sizeof(half), 4 * src1_stride_y))
{
half4 a0 = vload4(0, (__global half *)(src0_ptr + src_addr.s0));
half8 b0 = vload8(0, (__global half *)(src1_ptr + src_addr.s1 + 0 * src1_stride_y));
half8 b1 = vload8(0, (__global half *)(src1_ptr + src_addr.s1 + 1 * src1_stride_y));
half8 b2 = vload8(0, (__global half *)(src1_ptr + src_addr.s1 + 2 * src1_stride_y));
half8 b3 = vload8(0, (__global half *)(src1_ptr + src_addr.s1 + 3 * src1_stride_y));
acc += b0 * (half8)a0.s0;
acc += b1 * (half8)a0.s1;
acc += b2 * (half8)a0.s2;
acc += b3 * (half8)a0.s3;
}
for(; src_addr.s0 < end_row_vec_a; src_addr += (int2)(sizeof(half), src1_stride_y))
{
half a0 = *((__global half *)(src0_ptr + src_addr.s0));
half8 b0 = vload8(0, (__global half *)(src1_ptr + src_addr.s1));
acc += b0 * (half8)a0;
}
/* Compute destination address */
Image dst = CONVERT_TO_IMAGE_STRUCT(dst);
/* Multiply by the weight of vector-matrix product */
acc = acc * (half8)ALPHA;
vstore8(acc, 0, (__global half *)(offset(&dst, 0, 0)));
}
#endif /* (defined WIDTH_VECTOR_A) */
#endif /* (defined WIDTH_MATRIX_B && defined ALPHA) */
#if(defined BETA)
/** This OpenCL kernel performs the in-place matrix addition between 2 matrices taking into account that the second matrix might be weighted by a scalar value beta:
*
* @attention The beta's value need to be passed at compile time using -DBETA
*
* @param[in] src_ptr Pointer to the source matrix. Supported data types: F32
* @param[in] src_stride_x Stride of the source matrix in X dimension (in bytes)
* @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] src_stride_y Stride of the source matrix in Y dimension (in bytes)
* @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] src_offset_first_element_in_bytes The offset of the first element in the source matrix
* @param[out] dst_ptr Pointer to the destination matrix Supported data types: F32
* @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes)
* @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes)
* @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix
*/
__kernel void gemm_ma_f32(IMAGE_DECLARATION(src),
IMAGE_DECLARATION(dst))
{
/* Compute source and destination addresses */
Image src = CONVERT_TO_IMAGE_STRUCT(src);
Image dst = CONVERT_TO_IMAGE_STRUCT(dst);
/* Load values from A x B */
float4 alpha_ab = vload4(0, (__global float *)dst.ptr);
/* Load values from Matrix C */
float4 c = vload4(0, (__global float *)src.ptr);
/* Computes alpha * axb + beta * c */
float4 out = alpha_ab + (float4)BETA * c;
/* Store final result in axb matrix */
vstore4(out, 0, (__global float *)dst.ptr);
}
/** This OpenCL kernel performs the in-place matrix addition between 2 matrices taking into account that the second matrix might be weighted by a scalar value beta:
*
* @param[in] src_ptr Pointer to the source matrix. Supported data types: F16
* @param[in] src_stride_x Stride of the source matrix in X dimension (in bytes)
* @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] src_stride_y Stride of the source matrix in Y dimension (in bytes)
* @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] src_offset_first_element_in_bytes The offset of the first element in the source matrix
* @param[out] dst_ptr Pointer to the destination matrix Supported data types: F16
* @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes)
* @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes)
* @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix
*/
__kernel void gemm_ma_f16(IMAGE_DECLARATION(src),
IMAGE_DECLARATION(dst))
{
/* Compute source and destination addresses */
Image src = CONVERT_TO_IMAGE_STRUCT(src);
Image dst = CONVERT_TO_IMAGE_STRUCT(dst);
/* Load values from A x B */
half8 alpha_ab = vload8(0, (__global half *)dst.ptr);
/* Load values from Matrix C */
half8 c = vload8(0, (__global half *)src.ptr);
/* Computes alpha * axb + beta * c */
half8 out = alpha_ab + (half8)BETA * c;
/* Store final result in axb matrix */
vstore8(out, 0, (__global half *)dst.ptr);
}
#endif /* (defined BETA) */
#if(defined WIDTH_VECTOR_A)
/** This OpenCL kernel computes the vector by matrix multiplication between each row of A (src0) and matrix B (src1) used for locally connected layer
*
* @attention The width of A need to be passed at compile time using -DWIDTH_VECTOR_A
*
* @attention The input A and matrix B must not be reshaped
*
* @param[in] src0_ptr Pointer to the source matrix. Supported data types: F32
* @param[in] src0_stride_x Stride of the source matrix in X dimension (in bytes)
* @param[in] src0_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] src0_stride_y Stride of the source matrix in Y dimension (in bytes)
* @param[in] src0_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] src0_offset_first_element_in_bytes The offset of the first element in the source matrix
* @param[in] src1_ptr Pointer to the source matrix. Supported data types: F32
* @param[in] src1_stride_x Stride of the source matrix in X dimension (in bytes)
* @param[in] src1_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] src1_stride_y Stride of the source matrix in Y dimension (in bytes)
* @param[in] src1_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] src1_stride_z Stride of the source matrix in Z dimension (in bytes)
* @param[in] src1_step_z src_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] src1_offset_first_element_in_bytes The offset of the first element in the source matrix
* @param[out] dst_ptr Pointer to the destination matrix Supported data types: F32
* @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes)
* @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes)
* @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix
*/
__kernel void gemm_lc_vm_f32(IMAGE_DECLARATION(src0),
TENSOR3D_DECLARATION(src1),
IMAGE_DECLARATION(dst))
{
int idx = get_global_id(0) * 4;
int idy = get_global_id(1);
/* Compute the address for the vector A and matrix B */
int2 src_addr = ((int2)(src0_offset_first_element_in_bytes + src0_stride_y * idy, src1_offset_first_element_in_bytes + src1_stride_z * idy));
src_addr.s1 += idx * sizeof(float);
int end_row_vec_a = src_addr.s0 + (WIDTH_VECTOR_A * sizeof(float));
float4 acc = 0.0f;
for(; src_addr.s0 <= (end_row_vec_a - 2 * sizeof(float)); src_addr += (int2)(2 * sizeof(float), 2 * src1_stride_y))
{
float2 a0 = vload2(0, (__global float *)(src0_ptr + src_addr.s0));
float4 b0 = vload4(0, (__global float *)(src1_ptr + src_addr.s1));
float4 b1 = vload4(0, (__global float *)(src1_ptr + src_addr.s1 + src1_stride_y));
acc += b0 * (float4)a0.s0;
acc += b1 * (float4)a0.s1;
}
for(; src_addr.s0 < end_row_vec_a; src_addr += (int2)(sizeof(float), src1_stride_y))
{
float a0 = *((__global float *)(src0_ptr + src_addr.s0));
float4 b0 = vload4(0, (__global float *)(src1_ptr + src_addr.s1));
acc += b0 * (float4)a0;
}
/* Compute destination address */
Image dst = CONVERT_TO_IMAGE_STRUCT(dst);
vstore4(acc, 0, (__global float *)(offset(&dst, 0, 0)));
}
#endif /* (defined WIDTH_VECTOR_A) */
|
