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
|
/*
* Copyright (c) 2017-2020 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_asymm.h"
#undef CONVERT_SAT_STR
#undef CONVERT_SAT
#if defined(DATA_TYPE) && defined(STRIDE_X) && defined(WEIGHTS_DEPTH) && defined(OUTPUT_MULTIPLIER) && defined(OUTPUT_SHIFT)
#define CONVERT_SAT_STR(x, type) (convert_##type##8_sat((x)))
#define CONVERT_SAT(x, type) CONVERT_SAT_STR(x, type)
#if defined(DATA_LAYOUT_NHWC)
#if KERNEL_SIZE == 5
#if STRIDE_X == 1
#define CONVOLUTION1x5(acc, src_ptr, weights_ptr) CONVOLUTION1x5_STRIDE1(acc, src_ptr, weights_ptr)
#elif STRIDE_X == 2
#define CONVOLUTION1x5(acc, src_ptr, weights_ptr) CONVOLUTION1x5_STRIDE2(acc, src_ptr, weights_ptr)
#else /* STRIDE_X not equals 1 or 2 */
#error "STRIDE_X larger than 2 is not supported"
#endif /* STRIDE_X */
#define CONVOLUTION1x5_STRIDE1(acc, src_ptr, weights_ptr) \
({ \
int4 weights_values0 = 0; \
int weights_value1 = 0; \
weights_values0.s0 = convert_int(*(weights_ptr + 0 * weights_stride_y)); \
weights_values0.s1 = convert_int(*(weights_ptr + 1 * weights_stride_y)); \
weights_values0.s2 = convert_int(*(weights_ptr + 2 * weights_stride_y)); \
weights_values0.s3 = convert_int(*(weights_ptr + 3 * weights_stride_y)); \
weights_value1 = convert_int(*(weights_ptr + 4 * weights_stride_y)); \
\
int8 src0 = 0; \
int4 src1 = 0; \
src0.s0 = convert_int(*(src_ptr + 0 * weights_stride_y)); \
src0.s1 = convert_int(*(src_ptr + 1 * weights_stride_y)); \
src0.s2 = convert_int(*(src_ptr + 2 * weights_stride_y)); \
src0.s3 = convert_int(*(src_ptr + 3 * weights_stride_y)); \
src0.s4 = convert_int(*(src_ptr + 4 * weights_stride_y)); \
src0.s5 = convert_int(*(src_ptr + 5 * weights_stride_y)); \
src0.s6 = convert_int(*(src_ptr + 6 * weights_stride_y)); \
src0.s7 = convert_int(*(src_ptr + 7 * weights_stride_y)); \
src1.s0 = convert_int(*(src_ptr + 8 * weights_stride_y)); \
src1.s1 = convert_int(*(src_ptr + 9 * weights_stride_y)); \
src1.s2 = convert_int(*(src_ptr + 10 * weights_stride_y)); \
src1.s3 = convert_int(*(src_ptr + 11 * weights_stride_y)); \
\
acc += (src0 + input_offset) * ((int8)weights_values0.s0 + weight_offset); \
acc += ((int8)(src0.s1234, src0.s567, src1.s0) + input_offset) * ((int8)weights_values0.s1 + weight_offset); \
acc += ((int8)(src0.s234, src0.s567, src1.s01) + input_offset) * ((int8)weights_values0.s2 + weight_offset); \
acc += ((int8)(src0.s345, src0.s67, src1.s012) + input_offset) * ((int8)weights_values0.s3 + weight_offset); \
acc += ((int8)(src0.s45, src0.s67, src1.s0123) + input_offset) * ((int8)weights_value1 + weight_offset); \
})
#define CONVOLUTION1x5_STRIDE2(acc, src_ptr, weights_ptr) \
({ \
int4 weights_values0 = 0; \
int weights_value1 = 0; \
weights_values0.s0 = convert_int(*(weights_ptr + 0 * weights_stride_y)); \
weights_values0.s1 = convert_int(*(weights_ptr + 1 * weights_stride_y)); \
weights_values0.s2 = convert_int(*(weights_ptr + 2 * weights_stride_y)); \
weights_values0.s3 = convert_int(*(weights_ptr + 3 * weights_stride_y)); \
weights_value1 = convert_int(*(weights_ptr + 4 * weights_stride_y)); \
\
int16 src0 = 0; \
int4 src1 = 0; \
src0.s0 = convert_int(*(src_ptr + 0 * weights_stride_y)); \
src0.s1 = convert_int(*(src_ptr + 1 * weights_stride_y)); \
src0.s2 = convert_int(*(src_ptr + 2 * weights_stride_y)); \
src0.s3 = convert_int(*(src_ptr + 3 * weights_stride_y)); \
src0.s4 = convert_int(*(src_ptr + 4 * weights_stride_y)); \
src0.s5 = convert_int(*(src_ptr + 5 * weights_stride_y)); \
src0.s6 = convert_int(*(src_ptr + 6 * weights_stride_y)); \
src0.s7 = convert_int(*(src_ptr + 7 * weights_stride_y)); \
src0.s8 = convert_int(*(src_ptr + 8 * weights_stride_y)); \
src0.s9 = convert_int(*(src_ptr + 9 * weights_stride_y)); \
src0.sa = convert_int(*(src_ptr + 10 * weights_stride_y)); \
src0.sb = convert_int(*(src_ptr + 11 * weights_stride_y)); \
src0.sc = convert_int(*(src_ptr + 12 * weights_stride_y)); \
src0.sd = convert_int(*(src_ptr + 13 * weights_stride_y)); \
src0.se = convert_int(*(src_ptr + 14 * weights_stride_y)); \
src0.sf = convert_int(*(src_ptr + 15 * weights_stride_y)); \
src1.s0 = convert_int(*(src_ptr + 16 * weights_stride_y)); \
src1.s1 = convert_int(*(src_ptr + 17 * weights_stride_y)); \
src1.s2 = convert_int(*(src_ptr + 18 * weights_stride_y)); \
src1.s3 = convert_int(*(src_ptr + 19 * weights_stride_y)); \
\
acc += (src0.even + input_offset) * ((int8)weights_values0.s0 + weight_offset); \
acc += ((int8)(src0.s1357, src0.s9BDF) + input_offset) * ((int8)weights_values0.s1 + weight_offset); \
acc += ((int8)(src0.s2468, src0.sACE, src1.s0) + input_offset) * ((int8)weights_values0.s2 + weight_offset); \
acc += ((int8)(src0.s3579, src0.sBDF, src1.s1) + input_offset) * ((int8)weights_values0.s3 + weight_offset); \
acc += ((int8)(src0.s468a, src0.sCE, src1.s02) + input_offset) * ((int8)weights_value1 + weight_offset); \
})
#elif KERNEL_SIZE == 3
#if STRIDE_X == 1
#define CONVOLUTION1x3(acc, src_ptr, weights_ptr) CONVOLUTION1x3_STRIDE1(acc, src_ptr, weights_ptr)
#elif STRIDE_X == 2
#define CONVOLUTION1x3(acc, src_ptr, weights_ptr) CONVOLUTION1x3_STRIDE2(acc, src_ptr, weights_ptr)
#else /* STRIDE_X not equals 1 or 2 */
#error "STRIDE_X larger than 2 is not supported"
#endif /* STRIDE_X */
#define CONVOLUTION1x3_STRIDE1(acc, src_ptr, weights_ptr) \
({ \
int3 weights_values0 = 0; \
weights_values0.s0 = convert_int(*(weights_ptr + 0 * weights_stride_y)); \
weights_values0.s1 = convert_int(*(weights_ptr + 1 * weights_stride_y)); \
weights_values0.s2 = convert_int(*(weights_ptr + 2 * weights_stride_y)); \
\
int8 src0 = 0; \
int2 src1 = 0; \
src0.s0 = convert_int(*(src_ptr + 0 * weights_stride_y)); \
src0.s1 = convert_int(*(src_ptr + 1 * weights_stride_y)); \
src0.s2 = convert_int(*(src_ptr + 2 * weights_stride_y)); \
src0.s3 = convert_int(*(src_ptr + 3 * weights_stride_y)); \
src0.s4 = convert_int(*(src_ptr + 4 * weights_stride_y)); \
src0.s5 = convert_int(*(src_ptr + 5 * weights_stride_y)); \
src0.s6 = convert_int(*(src_ptr + 6 * weights_stride_y)); \
src0.s7 = convert_int(*(src_ptr + 7 * weights_stride_y)); \
src1.s0 = convert_int(*(src_ptr + 8 * weights_stride_y)); \
src1.s1 = convert_int(*(src_ptr + 9 * weights_stride_y)); \
\
acc += (src0 + input_offset) * ((int8)weights_values0.s0 + weight_offset); \
acc += ((int8)(src0.s1234, src0.s567, src1.s0) + input_offset) * ((int8)weights_values0.s1 + weight_offset); \
acc += ((int8)(src0.s234, src0.s567, src1.s01) + input_offset) * ((int8)weights_values0.s2 + weight_offset); \
})
#define CONVOLUTION1x3_STRIDE2(acc, src_ptr, weights_ptr) \
({ \
int3 weights_values0 = 0; \
weights_values0.s0 = convert_int(*(weights_ptr + 0 * weights_stride_y)); \
weights_values0.s1 = convert_int(*(weights_ptr + 1 * weights_stride_y)); \
weights_values0.s2 = convert_int(*(weights_ptr + 2 * weights_stride_y)); \
\
int16 src0 = 0; \
int src1 = 0; \
src0.s0 = convert_int(*(src_ptr + 0 * src_stride_y)); \
src0.s1 = convert_int(*(src_ptr + 1 * src_stride_y)); \
src0.s2 = convert_int(*(src_ptr + 2 * src_stride_y)); \
src0.s3 = convert_int(*(src_ptr + 3 * src_stride_y)); \
src0.s4 = convert_int(*(src_ptr + 4 * src_stride_y)); \
src0.s5 = convert_int(*(src_ptr + 5 * src_stride_y)); \
src0.s6 = convert_int(*(src_ptr + 6 * src_stride_y)); \
src0.s7 = convert_int(*(src_ptr + 7 * src_stride_y)); \
src0.s8 = convert_int(*(src_ptr + 8 * src_stride_y)); \
src0.s9 = convert_int(*(src_ptr + 9 * src_stride_y)); \
src0.sa = convert_int(*(src_ptr + 10 * src_stride_y)); \
src0.sb = convert_int(*(src_ptr + 11 * src_stride_y)); \
src0.sc = convert_int(*(src_ptr + 12 * src_stride_y)); \
src0.sd = convert_int(*(src_ptr + 13 * src_stride_y)); \
src0.se = convert_int(*(src_ptr + 14 * src_stride_y)); \
src0.sf = convert_int(*(src_ptr + 15 * src_stride_y)); \
src1 = convert_int(*(src_ptr + 16 * src_stride_y)); \
acc += (src0.even + input_offset) * ((int8)weights_values0.s0 + weight_offset); \
acc += ((int8)(src0.s1357, src0.s9BDF) + input_offset) * ((int8)weights_values0.s1 + weight_offset); \
acc += ((int8)(src0.s2468, src0.sACE, src1) + input_offset) * ((int8)weights_values0.s2 + weight_offset); \
})
#elif KERNEL_SIZE == 1
#if STRIDE_X == 3
#define INPUT_VALUE extract_input_stride3
#elif STRIDE_X == 2
#define INPUT_VALUE extract_input_stride2
#elif STRIDE_X == 1
#define INPUT_VALUE extract_input_stride1
#else /* STRIDE_X not equals 1, 2 or 3 */
#error "Only support strides 1, 2 and 3"
#endif /* STRIDE_X */
#endif // KERNEL_SIZE == 1
/** Extracts a 1D horizontal vector from the input tensor with stride as 1.
*
* @param[in] input_value Pointer to the first value.
*
* @return extracted input values.
*/
inline VEC_DATA_TYPE(DATA_TYPE, 8) extract_input_stride1(__global const DATA_TYPE *input_value, const uchar stride_y)
{
VEC_DATA_TYPE(DATA_TYPE, 8)
vals;
vals.s0 = *(input_value + 0 * stride_y);
vals.s1 = *(input_value + 1 * stride_y);
vals.s2 = *(input_value + 2 * stride_y);
vals.s3 = *(input_value + 3 * stride_y);
vals.s4 = *(input_value + 4 * stride_y);
vals.s5 = *(input_value + 5 * stride_y);
vals.s6 = *(input_value + 6 * stride_y);
vals.s7 = *(input_value + 7 * stride_y);
return vals;
}
/** Extracts a 1D horizontal vector from the input tensor with stride as 2.
*
* @param[in] input_value Pointer to the first value.
*
* @return extracted input values.
*/
inline VEC_DATA_TYPE(DATA_TYPE, 8) extract_input_stride2(__global const DATA_TYPE *input_value, const uchar stride_y)
{
VEC_DATA_TYPE(DATA_TYPE, 8)
vals;
vals.s0 = *(input_value + 0 * stride_y);
vals.s1 = *(input_value + 2 * stride_y);
vals.s2 = *(input_value + 4 * stride_y);
vals.s3 = *(input_value + 6 * stride_y);
vals.s4 = *(input_value + 8 * stride_y);
vals.s5 = *(input_value + 10 * stride_y);
vals.s6 = *(input_value + 12 * stride_y);
vals.s7 = *(input_value + 14 * stride_y);
return vals;
}
/** Extracts a 1D horizontal vector from the input tensor with stride as 3 and 8-bit data size.
*
* @param[in] input_value Pointer to the first value.
*
* @return extracted input values.
*/
inline VEC_DATA_TYPE(DATA_TYPE, 8) extract_input_stride3(__global const DATA_TYPE *input_value, const uchar stride_y)
{
VEC_DATA_TYPE(DATA_TYPE, 8)
vals;
vals.s0 = *(input_value + 0 * stride_y);
vals.s1 = *(input_value + 3 * stride_y);
vals.s2 = *(input_value + 6 * stride_y);
vals.s3 = *(input_value + 9 * stride_y);
vals.s4 = *(input_value + 12 * stride_y);
vals.s5 = *(input_value + 15 * stride_y);
vals.s6 = *(input_value + 18 * stride_y);
vals.s7 = *(input_value + 21 * stride_y);
return vals;
}
/** This kernel performs a direct convolution to convolve the low three dimensions.
*
* @note The convolution stride x must be passed at compile time using -DSTRIDE_X e.g. -DSTRIDE_X=1
* @note The third dimensions of the weights tensors must be passed at compile time using -DWEIGHTS_DEPTH
* @note If biases are used then -DHAS_BIAS has to be passed at compile time
* @note The output quantization multiplier must be passed at compile time using -DOUTPUT_MULTIPLIER e.g. -DOUTPUT_MULTIPLIER=1234
* @note The output quantization shift must be passed at compile time using -DOUTPUT_SHIFT e.g. -DOUTPUT_SHIFT=4
*
* @param[in] src_ptr Pointer to the source tensor. Supported data types: QASYMM8/QASYMM8_SIGNED
* @param[in] src_stride_x Stride of the source tensor 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 tensor 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_stride_z Stride of the source tensor in Z dimension (in bytes)
* @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor
* @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr
* @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)
* @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes)
* @param[in] dst_step_y dst_stride_y * number of elements along Z processed per workitem(in bytes)
* @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes)
* @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
* @param[in] weights_ptr Pointer to the weights tensor. Supported data types: same as @p src_ptr
* @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes)
* @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes)
* @param[in] weights_step_y weights_stride_y * number of elements along y processed per workitem(in bytes)
* @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes)
* @param[in] weights_step_z weights_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor
* @param[in] biases_ptr Pointer to the biases tensor. Supported data types: S32
* @param[in] biases_stride_x Stride of the biases tensor in X dimension (in bytes)
* @param[in] biases_step_x biases_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 biases tensor
* @param[in] weights_stride_w Stride of the weights tensor in the 4th dimension
* @param[in] input_offset Input offset quantization parameter
* @param[in] weight_offset Weights offset quantization parameter
* @param[in] output_offset Output offset quantization parameter
*/
__kernel void direct_convolution_quantized(
TENSOR3D_DECLARATION(src),
TENSOR3D_DECLARATION(dst),
TENSOR3D_DECLARATION(weights),
#ifdef HAS_BIAS
VECTOR_DECLARATION(biases),
#endif /* defined(HAS_BIAS) */
unsigned int weights_stride_w,
int input_offset,
int weight_offset,
int output_offset)
{
Image src = CONVERT_TO_IMAGE_STRUCT(src);
Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(weights);
Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst);
int8 values0 = 0;
const int id0 = get_global_id(0);
const int y_coord = (get_global_id(2) * STRIDE_Y) - PAD_TOP;
__global DATA_TYPE *weights_addr = (__global DATA_TYPE *)tensor3D_offset(&weights, 0, 0, 0);
__global DATA_TYPE *src_addr = (__global DATA_TYPE *)offset(&src, 0, 0) - src_stride_x * id0 + y_coord * (int)src_stride_z;
weights_addr += id0 * weights_stride_w;
for(volatile int d = 0; d < WEIGHTS_DEPTH; ++d)
{
#if KERNEL_SIZE == 5
#if(PAD_TOP == 1) || (PAD_BOTTM == 1)
if(y_coord < 0) // special case Z = -1 doesn't exists
{
CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 1 * weights_stride_z));
CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 2 * weights_stride_z));
CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 3 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 3 * weights_stride_z));
CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 4 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 4 * weights_stride_z));
}
else if(get_global_id(2) == (DST_HEIGHT - 1))
{
CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 0 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 0 * weights_stride_z));
CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 1 * weights_stride_z));
CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 2 * weights_stride_z));
CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 3 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 3 * weights_stride_z));
}
else
{
CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 0 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 0 * weights_stride_z));
CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 1 * weights_stride_z));
CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 2 * weights_stride_z));
CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 3 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 3 * weights_stride_z));
CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 4 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 4 * weights_stride_z));
}
#elif(PAD_TOP == 2) || (PAD_BOTTM == 2)
if(y_coord < -1)
{
CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 2 * weights_stride_z));
CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 3 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 3 * weights_stride_z));
CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 4 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 4 * weights_stride_z));
}
else if(y_coord == -1)
{
CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 1 * weights_stride_z));
CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 2 * weights_stride_z));
CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 3 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 3 * weights_stride_z));
CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 4 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 4 * weights_stride_z));
}
else if(y_coord == (SRC_HEIGHT - 3))
{
CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 0 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 0 * weights_stride_z));
CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 1 * weights_stride_z));
CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 2 * weights_stride_z));
}
else if(y_coord >= (SRC_HEIGHT - 4))
{
CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 0 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 0 * weights_stride_z));
CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 1 * weights_stride_z));
CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 2 * weights_stride_z));
CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 3 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 3 * weights_stride_z));
}
else
{
CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 0 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 0 * weights_stride_z));
CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 1 * weights_stride_z));
CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 2 * weights_stride_z));
CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 3 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 3 * weights_stride_z));
CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 4 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 4 * weights_stride_z));
}
#else /* PAD_TOP == 2 || || PAD_BOTTM == 2 */
CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 0 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 0 * weights_stride_z));
CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 1 * weights_stride_z));
CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 2 * weights_stride_z));
CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 3 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 3 * weights_stride_z));
CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 4 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 4 * weights_stride_z));
#endif /* PAD_TOP == 1 || || PAD_BOTTM == 1 */
#elif KERNEL_SIZE == 3
#if(PAD_TOP > 0) || (PAD_BOTTOM > 0)
if(y_coord < 0) // special case Z = -1 doesn't exists
{
//skip first row and load the two next ones
CONVOLUTION1x3(values0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 1 * weights_stride_z));
CONVOLUTION1x3(values0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 2 * weights_stride_z));
}
else if(y_coord == (SRC_HEIGHT - PAD_BOTTOM - 1))
{
// special case when computing the last row of the output we must read the last three rows from the input buffer (including padding) but the
// Z axis has no padding at all.
CONVOLUTION1x3(values0, (__global DATA_TYPE *)(src_addr + 0 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 0 * weights_stride_z));
CONVOLUTION1x3(values0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 1 * weights_stride_z));
}
else
{
CONVOLUTION1x3(values0, (__global DATA_TYPE *)(src_addr + 0 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 0 * weights_stride_z));
CONVOLUTION1x3(values0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 1 * weights_stride_z));
CONVOLUTION1x3(values0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 2 * weights_stride_z));
}
#else // PAD_TOP > 0 || PAD_BOTTOM > 0
CONVOLUTION1x3(values0, (__global DATA_TYPE *)(src_addr + 0 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 0 * weights_stride_z));
CONVOLUTION1x3(values0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 1 * weights_stride_z));
CONVOLUTION1x3(values0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_z), (__global DATA_TYPE *)(weights_addr + 2 * weights_stride_z));
#endif // PAD_TOP > 0 || PAD_BOTTOM > 0
#elif KERNEL_SIZE == 1
int weight = convert_int(*(__global DATA_TYPE *)weights_addr);
int8 input_value = convert_int8(INPUT_VALUE((__global DATA_TYPE *)src_addr, src_stride_y));
values0 += (input_value + input_offset) * ((int8)weight + weight_offset);
#endif /* (KERNEL_SIZE == 1) || (KERNEL_SIZE == 3) || (KERNEL_SIZE == 5) */
src_addr += src_stride_x;
weights_addr += weights_stride_x;
}
#ifdef HAS_BIAS
Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases);
__global int *bias_addr = ((__global int *)(vector_offset(&biases, id0)));
values0 += (int8)(*bias_addr);
#endif /* defined(HAS_BIAS) */
#if OUTPUT_SHIFT < 0
values0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(values0, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, 8);
#else // OUTPUT_SHIFT < 0
values0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(values0, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, 8);
#endif // OUTPUT_SHIFT < 0
values0 = values0 + output_offset;
VEC_DATA_TYPE(DATA_TYPE, 8)
values = CONVERT_SAT(values0, DATA_TYPE);
*(dst.ptr + 0 * dst_stride_y) = values.s0;
*(dst.ptr + 1 * dst_stride_y) = values.s1;
*(dst.ptr + 2 * dst_stride_y) = values.s2;
*(dst.ptr + 3 * dst_stride_y) = values.s3;
*(dst.ptr + 4 * dst_stride_y) = values.s4;
*(dst.ptr + 5 * dst_stride_y) = values.s5;
*(dst.ptr + 6 * dst_stride_y) = values.s6;
*(dst.ptr + 7 * dst_stride_y) = values.s7;
}
#else // defined(DATA_LAYOUT_NHWC)
#if KERNEL_SIZE == 9
#if STRIDE_X == 1
#define CONVOLUTION1x9(acc, src_row_ptr, weights_row_ptr) CONVOLUTION1x9_STRIDE1(acc, src_row_ptr, weights_row_ptr)
#elif STRIDE_X == 2
#define CONVOLUTION1x9(acc, src_row_ptr, weights_row_ptr) CONVOLUTION1x9_STRIDE2(acc, src_row_ptr, weights_row_ptr)
#else /* STRIDE_X not equals 1 or 2 */
#error "STRIDE_X larger than 2 is not supported"
#endif /* STRIDE_X */
#define CONVOLUTION1x9_STRIDE1(acc, src_row_ptr, weights_row_ptr) \
({ \
int8 weights_values0 = convert_int8(vload8(0, weights_row_ptr)); \
int weights_value1 = convert_int(*(weights_row_ptr + 8)); \
int16 src0 = convert_int16(vload16(0, src_row_ptr)); \
acc += (src0.lo + input_offset) * ((int8)weights_values0.s0 + weight_offset); \
acc += ((int8)(src0.s1234, src0.s5678) + input_offset) * ((int8)weights_values0.s1 + weight_offset); \
acc += ((int8)(src0.s2345, src0.s6789) + input_offset) * ((int8)weights_values0.s2 + weight_offset); \
acc += ((int8)(src0.s3456, src0.s789A) + input_offset) * ((int8)weights_values0.s3 + weight_offset); \
acc += ((int8)(src0.s4567, src0.s89AB) + input_offset) * ((int8)weights_values0.s4 + weight_offset); \
acc += ((int8)(src0.s5678, src0.s9ABC) + input_offset) * ((int8)weights_values0.s5 + weight_offset); \
acc += ((int8)(src0.s6789, src0.sABCD) + input_offset) * ((int8)weights_values0.s6 + weight_offset); \
acc += ((int8)(src0.s789A, src0.sBCDE) + input_offset) * ((int8)weights_values0.s7 + weight_offset); \
acc += ((int8)(src0.s89AB, src0.sCDEF) + input_offset) * ((int8)weights_value1 + weight_offset); \
})
#define CONVOLUTION1x9_STRIDE2(acc, src_row_ptr, weights_row_ptr) \
({ \
int8 weights_values0 = convert_int8(vload8(0, weights_row_ptr)); \
int weights_value1 = convert_int(*(weights_row_ptr + 8)); \
int16 src0 = convert_int16(vload16(0, src_row_ptr)); \
int8 src1 = convert_int8(vload8(0, src_row_ptr + 16)); \
acc += (src0.even + input_offset) * ((int8)weights_values0.s0 + weight_offset); \
acc += ((int8)(src0.s1357, src0.s9BDF) + input_offset) * ((int8)weights_values0.s1 + weight_offset); \
acc += ((int8)(src0.s2468, src0.sACE, src1.s0) + input_offset) * ((int8)weights_values0.s2 + weight_offset); \
acc += ((int8)(src0.s3579, src0.sBDF, src1.s1) + input_offset) * ((int8)weights_values0.s3 + weight_offset); \
acc += ((int8)(src0.s468A, src0.sCE, src1.s02) + input_offset) * ((int8)weights_values0.s4 + weight_offset); \
acc += ((int8)(src0.s579B, src0.sDF, src1.s13) + input_offset) * ((int8)weights_values0.s5 + weight_offset); \
acc += ((int8)(src0.s68AC, src0.sE, src1.s024) + input_offset) * ((int8)weights_values0.s6 + weight_offset); \
acc += ((int8)(src0.s79BD, src0.sF, src1.s135) + input_offset) * ((int8)weights_values0.s7 + weight_offset); \
acc += ((int8)(src0.s8ACE, src1.s0246) + input_offset) * ((int8)weights_value1 + weight_offset); \
})
#elif KERNEL_SIZE == 5
#if STRIDE_X == 1
#define CONVOLUTION1x5(acc, src_row_ptr, weights_row_ptr) CONVOLUTION1x5_STRIDE1(acc, src_row_ptr, weights_row_ptr)
#elif STRIDE_X == 2
#define CONVOLUTION1x5(acc, src_row_ptr, weights_row_ptr) CONVOLUTION1x5_STRIDE2(acc, src_row_ptr, weights_row_ptr)
#else /* STRIDE_X not equals 1 or 2 */
#error "STRIDE_X larger than 2 is not supported"
#endif /* STRIDE_X */
#define CONVOLUTION1x5_STRIDE1(acc, src_row_ptr, weights_row_ptr) \
({ \
int4 weights_values0 = convert_int4(vload4(0, weights_row_ptr)); \
int weights_value1 = convert_int(*(weights_row_ptr + 4)); \
int8 src0 = convert_int8(vload8(0, src_row_ptr)); \
int4 src1 = convert_int4(vload4(0, src_row_ptr + 8)); \
acc += (src0 + input_offset) * ((int8)weights_values0.s0 + weight_offset); \
acc += ((int8)(src0.s1234, src0.s567, src1.s0) + input_offset) * ((int8)weights_values0.s1 + weight_offset); \
acc += ((int8)(src0.s234, src0.s567, src1.s01) + input_offset) * ((int8)weights_values0.s2 + weight_offset); \
acc += ((int8)(src0.s345, src0.s67, src1.s012) + input_offset) * ((int8)weights_values0.s3 + weight_offset); \
acc += ((int8)(src0.s45, src0.s67, src1.s0123) + input_offset) * ((int8)weights_value1 + weight_offset); \
})
#define CONVOLUTION1x5_STRIDE2(acc, src_row_ptr, weights_row_ptr) \
({ \
int4 weights_values0 = convert_int4(vload4(0, weights_row_ptr)); \
int weights_value1 = convert_int(*(weights_row_ptr + 4)); \
int16 src0 = convert_int16(vload16(0, src_row_ptr)); \
int4 src1 = convert_int4(vload4(0, src_row_ptr + 16)); \
acc += (src0.even + input_offset) * ((int8)weights_values0.s0 + weight_offset); \
acc += ((int8)(src0.s1357, src0.s9BDF) + input_offset) * ((int8)weights_values0.s1 + weight_offset); \
acc += ((int8)(src0.s2468, src0.sACE, src1.s0) + input_offset) * ((int8)weights_values0.s2 + weight_offset); \
acc += ((int8)(src0.s3579, src0.sBDF, src1.s1) + input_offset) * ((int8)weights_values0.s3 + weight_offset); \
acc += ((int8)(src0.s468a, src0.sCE, src1.s02) + input_offset) * ((int8)weights_value1 + weight_offset); \
})
#elif KERNEL_SIZE == 3
#if STRIDE_X == 1
#define CONVOLUTION1x3(acc, src_row_ptr, weights_row_ptr) CONVOLUTION1x3_STRIDE1(acc, src_row_ptr, weights_row_ptr)
#elif STRIDE_X == 2
#define CONVOLUTION1x3(acc, src_row_ptr, weights_row_ptr) CONVOLUTION1x3_STRIDE2(acc, src_row_ptr, weights_row_ptr)
#else /* STRIDE_X not equals 1 or 2 */
#error "STRIDE_X larger than 2 is not supported"
#endif /* STRIDE_X */
#define CONVOLUTION1x3_STRIDE1(acc, src_row_ptr, weights_row_ptr) \
({ \
int3 weights_values0 = convert_int3(vload3(0, weights_row_ptr)); \
int8 src0 = convert_int8(vload8(0, src_row_ptr)); \
int2 src1 = convert_int2(vload2(0, src_row_ptr + 8)); \
acc += (src0 + input_offset) * ((int8)weights_values0.s0 + weight_offset); \
acc += ((int8)(src0.s1234, src0.s567, src1.s0) + input_offset) * ((int8)weights_values0.s1 + weight_offset); \
acc += ((int8)(src0.s234, src0.s567, src1.s01) + input_offset) * ((int8)weights_values0.s2 + weight_offset); \
})
#define CONVOLUTION1x3_STRIDE2(acc, src_row_ptr, weights_row_ptr) \
({ \
int3 weights_values0 = convert_int3(vload3(0, weights_row_ptr)); \
int16 src0 = convert_int16(vload16(0, src_row_ptr)); \
int src1 = convert_int(*(src_row_ptr + 16)); \
acc += (src0.even + input_offset) * ((int8)weights_values0.s0 + weight_offset); \
acc += ((int8)(src0.s1357, src0.s9BDF) + input_offset) * ((int8)weights_values0.s1 + weight_offset); \
acc += ((int8)(src0.s2468, src0.sACE, src1) + input_offset) * ((int8)weights_values0.s2 + weight_offset); \
})
#elif KERNEL_SIZE == 1
#if STRIDE_X == 3
#define INPUT_VALUE extract_input_stride3
#elif STRIDE_X == 2
#define INPUT_VALUE extract_input_stride2
#elif STRIDE_X == 1
#define INPUT_VALUE extract_input_stride1
#else /* STRIDE_X not equals 1, 2 or 3 */
#error "Only support strides 1, 2 and 3"
#endif /* STRIDE_X */
/** Extracts a 1D horizontal vector from the input tensor with stride as 1.
*
* @param[in] input_value Pointer to the first value.
*
* @return extracted input values.
*/
inline VEC_DATA_TYPE(DATA_TYPE, 8) extract_input_stride1(__global const DATA_TYPE *input_value)
{
return vload8(0, input_value);
}
/** Extracts a 1D horizontal vector from the input tensor with stride as 2.
*
* @param[in] input_value Pointer to the first value.
*
* @return extracted input values.
*/
inline VEC_DATA_TYPE(DATA_TYPE, 8) extract_input_stride2(__global const DATA_TYPE *input_value)
{
VEC_DATA_TYPE(DATA_TYPE, 16)
temp = vload16(0, input_value);
return temp.s02468ace;
}
/** Extracts a 1D horizontal vector from the input tensor with stride as 3 and 8-bit data size.
*
* @param[in] input_value Pointer to the first value.
*
* @return extracted input values.
*/
inline VEC_DATA_TYPE(DATA_TYPE, 8) extract_input_stride3(__global const DATA_TYPE *input_value)
{
VEC_DATA_TYPE(DATA_TYPE, 16)
temp1 = vload16(0, input_value);
VEC_DATA_TYPE(DATA_TYPE, 16)
temp2 = vload16(0, input_value + 12);
return (VEC_DATA_TYPE(DATA_TYPE, 8))(temp1.s0369, temp2.s0369);
}
#else /* KERNEL_SIZE not equals 1, 3 , 5, 9 */
#error "Only kernel sizes 1, 3, 5 and 9 are supported"
#endif /* KERNEL_SIZE */
/** This kernel performs a direct convolution to convolve the low three dimensions.
*
* @note The convolution stride x must be passed at compile time using -DSTRIDE_X e.g. -DSTRIDE_X=1
* @note The third dimensions of the weights tensors must be passed at compile time using -DWEIGHTS_DEPTH
* @note If biases are used then -DHAS_BIAS has to be passed at compile time
* @note The output quantization multiplier must be passed at compile time using -DOUTPUT_MULTIPLIER e.g. -DOUTPUT_MULTIPLIER=1234
* @note The output quantization shift must be passed at compile time using -DOUTPUT_SHIFT e.g. -DOUTPUT_SHIFT=4
*
* @param[in] src_ptr Pointer to the source tensor. Supported data types: QASYMM8/QASYMM8_SIGNED
* @param[in] src_stride_x Stride of the source tensor 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 tensor 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_stride_z Stride of the source tensor in Z dimension (in bytes)
* @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor
* @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr
* @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)
* @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes)
* @param[in] dst_step_y dst_stride_y * number of elements along Z processed per workitem(in bytes)
* @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes)
* @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
* @param[in] weights_ptr Pointer to the weights tensor. Supported data types: same as @p src_ptr
* @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes)
* @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes)
* @param[in] weights_step_y weights_stride_y * number of elements along y processed per workitem(in bytes)
* @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes)
* @param[in] weights_step_z weights_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor
* @param[in] biases_ptr Pointer to the biases tensor. Supported data types: S32
* @param[in] biases_stride_x Stride of the biases tensor in X dimension (in bytes)
* @param[in] biases_step_x biases_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 biases tensor
* @param[in] weights_stride_w Stride of the weights tensor in the 4th dimension
* @param[in] input_offset Input offset quantization parameter
* @param[in] weight_offset Weights offset quantization parameter
* @param[in] output_offset Output offset quantization parameter
*/
__kernel void direct_convolution_quantized(
TENSOR3D_DECLARATION(src),
TENSOR3D_DECLARATION(dst),
TENSOR3D_DECLARATION(weights),
#ifdef HAS_BIAS
VECTOR_DECLARATION(biases),
#endif /* defined(HAS_BIAS) */
unsigned int weights_stride_w,
int input_offset,
int weight_offset,
int output_offset)
{
Image src = CONVERT_TO_IMAGE_STRUCT(src);
Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(weights);
Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst);
int8 values0 = 0;
__global DATA_TYPE *weights_addr = (__global DATA_TYPE *)tensor3D_offset(&weights, 0, 0, 0);
__global DATA_TYPE *src_addr = (__global DATA_TYPE *)offset(&src, 0, 0);
const int kernel_index = get_global_id(2);
weights_addr += kernel_index * weights_stride_w;
for(volatile int d = 0; d < WEIGHTS_DEPTH; ++d)
{
#if KERNEL_SIZE == 9
CONVOLUTION1x9(values0, (__global DATA_TYPE *)(src_addr + 0 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 0 * weights_stride_y));
CONVOLUTION1x9(values0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 1 * weights_stride_y));
CONVOLUTION1x9(values0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 2 * weights_stride_y));
CONVOLUTION1x9(values0, (__global DATA_TYPE *)(src_addr + 3 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 3 * weights_stride_y));
CONVOLUTION1x9(values0, (__global DATA_TYPE *)(src_addr + 4 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 4 * weights_stride_y));
CONVOLUTION1x9(values0, (__global DATA_TYPE *)(src_addr + 5 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 5 * weights_stride_y));
CONVOLUTION1x9(values0, (__global DATA_TYPE *)(src_addr + 6 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 6 * weights_stride_y));
CONVOLUTION1x9(values0, (__global DATA_TYPE *)(src_addr + 7 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 7 * weights_stride_y));
CONVOLUTION1x9(values0, (__global DATA_TYPE *)(src_addr + 8 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 8 * weights_stride_y));
#elif KERNEL_SIZE == 5
CONVOLUTION1x5(values0, (__global DATA_TYPE *)src_addr, (__global DATA_TYPE *)weights_addr);
CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 1 * weights_stride_y));
CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 2 * weights_stride_y));
CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 3 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 3 * weights_stride_y));
CONVOLUTION1x5(values0, (__global DATA_TYPE *)(src_addr + 4 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 4 * weights_stride_y));
#elif KERNEL_SIZE == 3
CONVOLUTION1x3(values0, (__global DATA_TYPE *)(src_addr + 0 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 0 * weights_stride_y));
CONVOLUTION1x3(values0, (__global DATA_TYPE *)(src_addr + 1 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 1 * weights_stride_y));
CONVOLUTION1x3(values0, (__global DATA_TYPE *)(src_addr + 2 * src_stride_y), (__global DATA_TYPE *)(weights_addr + 2 * weights_stride_y));
#elif KERNEL_SIZE == 1
int weight = convert_int(*(__global DATA_TYPE *)weights_addr);
int8 input_value = convert_int8(INPUT_VALUE((__global DATA_TYPE *)src_addr));
values0 += (input_value + input_offset) * ((int8)weight + weight_offset);
#endif /* (KERNEL_SIZE == 1) || (KERNEL_SIZE == 3) || (KERNEL_SIZE == 5) */
src_addr += src_stride_z;
weights_addr += weights_stride_z;
}
#ifdef HAS_BIAS
Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases);
__global int *bias_addr = ((__global int *)(vector_offset(&biases, kernel_index)));
values0 += (int8)(*bias_addr);
#endif /* defined(HAS_BIAS) */
#if OUTPUT_SHIFT < 0
values0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(values0, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, 8);
#else // OUTPUT_SHIFT < 0
values0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(values0, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, 8);
#endif // OUTPUT_SHIFT < 0
values0 = values0 + output_offset;
vstore8(CONVERT_SAT(values0, DATA_TYPE), 0, (__global DATA_TYPE *)dst.ptr);
}
#endif // defined(DATA_LAYOUT_NHWC)
#endif // defined(DATA_TYPE) && defined(STRIDE_X) && defined(WEIGHTS_DEPTH) && defined(OUTPUT_MULTIPLIER) && defined(OUTPUT_SHIFT)
|
