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
|
/*
* 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.h"
#if defined(POOL_AVG) || defined(POOL_L2)
#define POOL_OP(x, y) ((x) + (y))
#else /* defined(POOL_AVG) || defined(POOL_L2) */
#define POOL_OP(x, y) (fmax((x), (y)))
#endif /* defined(POOL_AVG) || defined(POOL_L2) */
#if defined(POOL_L2)
#define POW2_OP(x, vec_size) ((x) * (x))
#else /* defined(POOL_L2) */
#define POW2_OP(x, vec_size) (x)
#endif /* defined(POOL_L2) */
#define DIV_OP(x, y) (x * (1.f / y))
#define SQRT_OP(x) sqrt((x))
#define DIV_OP_NHWC(x, y) (x * (VEC_DATA_TYPE(ACC_DATA_TYPE, 8))(1.f / y))
#if STRIDE_X == 1
#define POOLING3x3(res, input, output) POOLING3x3_STRIDE1(res, input, output)
#elif STRIDE_X == 2 /* STRIDE_X == 1 */
#define POOLING3x3(res, input, output) POOLING3x3_STRIDE2(res, input, output)
#elif STRIDE_X == 3 /* STRIDE_X not equals 1 or 2 */
#define POOLING3x3(res, input, output) POOLING3x3_STRIDE3(res, input, output)
#endif /* STRIDE_X == 3 */
#if defined(FP_MIXED_PRECISION)
#define CONVERT_TO_ACC_DATA_TYPE(x, n) CONVERT(x, VEC_DATA_TYPE(ACC_DATA_TYPE, n))
#define VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(n, offset, ptr) \
CONVERT_TO_ACC_DATA_TYPE(vload##n(offset, ptr), n)
#else /* defined(FP_MIXED_PRECISION) */
#define VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(n, offset, ptr) vload##n(offset, ptr)
#endif /* defined(FP_MIXED_PRECISION) */
#define POOLING3x3_STRIDE1(res, input, output) \
({ \
VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \
data00 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(4, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0)); \
VEC_DATA_TYPE(ACC_DATA_TYPE, 2) \
data01 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(2, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0) + 4); \
VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \
data10 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(4, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 1, 0)); \
VEC_DATA_TYPE(ACC_DATA_TYPE, 2) \
data11 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(2, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 1, 0) + 4); \
VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \
data20 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(4, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 2, 0)); \
VEC_DATA_TYPE(ACC_DATA_TYPE, 2) \
data21 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(2, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 2, 0) + 4); \
data00 = POW2_OP(data00, 4); \
data01 = POW2_OP(data01, 2); \
data10 = POW2_OP(data10, 4); \
data11 = POW2_OP(data11, 2); \
data20 = POW2_OP(data20, 4); \
data21 = POW2_OP(data21, 2); \
\
VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \
values00 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 8))(data00.s01212323); \
VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \
values01 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(data01.s0, data00.s3, data01.s01); \
VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \
values10 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 8))(data10.s01212323); \
VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \
values11 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(data11.s0, data10.s3, data11.s01); \
VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \
values20 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 8))(data20.s01212323); \
VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \
values21 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(data21.s0, data20.s3, data21.s01); \
\
values00 = POOL_OP(values00, values10); \
values01 = POOL_OP(values01, values11); \
values00 = POOL_OP(values00, values20); \
values01 = POOL_OP(values01, values21); \
\
res = POOL_OP((VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(values00.s036, values01.s1), (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(values00.s147, values01.s2)); \
res = POOL_OP(res, (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(values00.s25, values01.s03)); \
})
#define POOLING3x3_STRIDE2(res, input, output) \
({ \
VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \
data00 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(8, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0)); \
ACC_DATA_TYPE data01 = (ACC_DATA_TYPE)(*((__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0) + 8)); \
VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \
data10 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(8, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 1, 0)); \
ACC_DATA_TYPE data11 = (ACC_DATA_TYPE)(*((__global DATA_TYPE *)tensor3D_offset(&input, 0, 1, 0) + 8)); \
VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \
data20 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(8, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 2, 0)); \
ACC_DATA_TYPE data21 = (ACC_DATA_TYPE)(*((__global DATA_TYPE *)tensor3D_offset(&input, 0, 2, 0) + 8)); \
data00 = POW2_OP(data00, 8); \
data01 = POW2_OP(data01, 1); \
data10 = POW2_OP(data10, 8); \
data11 = POW2_OP(data11, 1); \
data20 = POW2_OP(data20, 8); \
data21 = POW2_OP(data21, 1); \
\
VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \
values00 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 8))(data00.s01223445); \
VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \
values01 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(data00.s667, data01); \
VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \
values10 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 8))(data10.s01223445); \
VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \
values11 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(data10.s667, data11); \
VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \
values20 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 8))(data20.s01223445); \
VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \
values21 = (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(data20.s667, data21); \
\
values00 = POOL_OP(values00, values10); \
values01 = POOL_OP(values01, values11); \
values00 = POOL_OP(values00, values20); \
values01 = POOL_OP(values01, values21); \
\
res = POOL_OP((VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(values00.s036, values01.s1), (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(values00.s147, values01.s2)); \
res = POOL_OP(res, (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(values00.s25, values01.s03)); \
})
#define POOLING3x3_STRIDE3(res, input, output) \
({ \
VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \
data00 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(8, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0)); \
VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \
data01 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(4, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0) + 8); \
VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \
data10 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(8, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 1, 0)); \
VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \
data11 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(4, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 1, 0) + 8); \
VEC_DATA_TYPE(ACC_DATA_TYPE, 8) \
data20 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(8, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 2, 0)); \
VEC_DATA_TYPE(ACC_DATA_TYPE, 4) \
data21 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(4, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 2, 0) + 8); \
data00 = POW2_OP(data00, 8); \
data01 = POW2_OP(data01, 4); \
data10 = POW2_OP(data10, 8); \
data11 = POW2_OP(data11, 4); \
data20 = POW2_OP(data20, 8); \
data21 = POW2_OP(data21, 4); \
\
data00 = POOL_OP(data00, data10); \
data01 = POOL_OP(data01, data11); \
data00 = POOL_OP(data00, data20); \
data01 = POOL_OP(data01, data21); \
\
res = POOL_OP((VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(data00.s036, data01.s1), (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(data00.s147, data01.s2)); \
res = POOL_OP(res, (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(data00.s25, data01.s03)); \
})
ACC_DATA_TYPE calculate_avg_scale(const int pool_size_x, const int pool_size_y, const int upper_bound_w, const int upper_bound_h,
const int pad_x, const int pad_y, const int stride_x, const int stride_y)
{
int start_x = get_global_id(0) * stride_x - pad_x;
int start_y = get_global_id(1) * stride_y - pad_y;
const int end_x = min(start_x + pool_size_x, upper_bound_w);
const int end_y = min(start_y + pool_size_y, upper_bound_h);
#if defined(EXCLUDE_PADDING)
start_x = max(0, start_x);
start_y = max(0, start_y);
#endif /* defined(EXCLUDE_PADDING) */
return ((end_y - start_y) * (end_x - start_x));
}
/** Performs a pooling function of pool size equal to 2.
*
* @note Datatype must be passed using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types are F16/F32;
* @note In case of average pooling the following information must be passed at compile time:
* -DPOOL_AVG or -DPOOL_L2 must be provided otherwise max pooling will be performed.
* -DMAX_WIDTH and -DMAX_HEIGHT which are the maximum accessible indeces in x and y dimensions (width + pad)
* -DSTRIDE_X and -DSTRIDE_Y which are the steps of the window along the x and y directions
* -DPAD_X and -DPAD_Y which are the pooling paddings in x and y dimension
*
* @param[in] input_ptr Pointer to the source image. Supported data types: F16/F32
* @param[in] input_stride_x Stride of the source image in X dimension (in bytes)
* @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] input_stride_y Stride of the source image in Y dimension (in bytes)
* @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes)
* @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] input_offset_first_element_in_bytes The offset of the first element in the source image
* @param[out] output_ptr Pointer to the destination image. Supported data types: same as @p input_ptr
* @param[in] output_stride_x Stride of the destination image in X dimension (in bytes)
* @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] output_stride_y Stride of the destination image in Y dimension (in bytes)
* @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] output_stride_z Stride of the source tensor in Z dimension (in bytes)
* @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination image
*/
__kernel void pooling_layer_2(
TENSOR3D_DECLARATION(input),
TENSOR3D_DECLARATION(output))
{
// Get pixels pointer
Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input);
Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output);
// Load data
VEC_DATA_TYPE(ACC_DATA_TYPE, 2)
data0 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(2, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0));
VEC_DATA_TYPE(ACC_DATA_TYPE, 2)
data1 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(2, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 1, 0));
#if defined(POOL_L2)
// Raise to power of 2 for L2 Pooling
data0 = POW2_OP(data0, 2);
data1 = POW2_OP(data1, 2);
#endif /* defined(POOL_L2) */
// Perform calculations
data0 = POOL_OP(data0, data1);
ACC_DATA_TYPE res = POOL_OP(data0.s0, data0.s1);
#if defined(POOL_AVG) || defined(POOL_L2)
// Divide by pool region in case of average or l2 pooling
res = DIV_OP(res, calculate_avg_scale(2, 2, MAX_WIDTH, MAX_HEIGHT, PAD_X, PAD_Y, STRIDE_X, STRIDE_Y));
#endif /* defined(POOL_AVG) || defined(POOL_L2) */
#if defined(POOL_L2)
// Take square root of the result in L2 pooling
res = SQRT_OP(res);
#endif /* defined(POOL_L2) */
// Store result
*(__global DATA_TYPE *)output.ptr = (DATA_TYPE)res;
}
/** Performs a pooling function of pool size equal to 3
*
* @note Datatype must be passed using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types are F16/F32;
* @note In case of average pooling the following information must be passed at compile time:
* -DPOOL_AVG or -DPOOL_L2 must be provided otherwise max pooling will be performed.
* -DMAX_WIDTH and -DMAX_HEIGHT which are the maximum accessible indeces in x and y dimensions (width + pad)
* -DSTRIDE_X and -DSTRIDE_Y which are the steps of the window along the x and y directions
* -DPAD_X and -DPAD_Y which are the pooling paddings in x and y dimension
*
* @param[in] input_ptr Pointer to the source image. Supported data types: F16/F32
* @param[in] input_stride_x Stride of the source image in X dimension (in bytes)
* @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] input_stride_y Stride of the source image in Y dimension (in bytes)
* @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes)
* @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] input_offset_first_element_in_bytes The offset of the first element in the source image
* @param[out] output_ptr Pointer to the destination image. Supported data types: same as @p input_ptr
* @param[in] output_stride_x Stride of the destination image in X dimension (in bytes)
* @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] output_stride_y Stride of the destination image in Y dimension (in bytes)
* @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] output_stride_z Stride of the source tensor in Z dimension (in bytes)
* @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination image
*/
__kernel void pooling_layer_3(
TENSOR3D_DECLARATION(input),
TENSOR3D_DECLARATION(output))
{
// Get pixels pointer
Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input);
Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output);
// Load data
VEC_DATA_TYPE(ACC_DATA_TYPE, 3)
data0 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(3, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0));
VEC_DATA_TYPE(ACC_DATA_TYPE, 3)
data1 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(3, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 1, 0));
VEC_DATA_TYPE(ACC_DATA_TYPE, 3)
data2 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(3, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 2, 0));
#if defined(POOL_L2)
// Raise to power of 2 for L2 Pooling
data0 = POW2_OP(data0, 3);
data1 = POW2_OP(data1, 3);
data2 = POW2_OP(data2, 3);
#endif /* defined(POOL_L2) */
// Perform calculations
data0 = POOL_OP(data0, data1);
data0 = POOL_OP(data0, data2);
ACC_DATA_TYPE res = POOL_OP(POOL_OP(data0.s0, data0.s1), data0.s2);
#if defined(POOL_AVG) || defined(POOL_L2)
// Divide by pool region in case of average pooling
res = DIV_OP(res, calculate_avg_scale(3, 3, MAX_WIDTH, MAX_HEIGHT, PAD_X, PAD_Y, STRIDE_X, STRIDE_Y));
#endif /* defined(POOL_AVG) || defined(POOL_L2) */
#if defined(POOL_L2)
// Take square root of the result in L2 pooling
res = SQRT_OP(res);
#endif /* defined(POOL_L2) */
// Store result
*(__global DATA_TYPE *)output.ptr = (DATA_TYPE)res;
}
#if defined(POOLING3x3)
#define CONVERT_OP(data_type) convert_##data_type##4
#define CONVERT_VECTOR4(data_type) CONVERT_OP(data_type)
VEC_DATA_TYPE(ACC_DATA_TYPE, 4)
calculate_avg_scale4(const int pool_size, const int upper_bound_w, const int upper_bound_h,
const int pad_x, const int pad_y, const int stride_x, const int stride_y)
{
int4 start_x = ((int4)get_global_id(0) * 4 + (int4)(0, 1, 2, 3)) * (int4)stride_x - (int4)pad_x;
int start_y = get_global_id(1) * stride_y - pad_y;
const int4 end_x = min(start_x + (int4)pool_size, (int4)upper_bound_w);
const int end_y = min(start_y + pool_size, upper_bound_h);
#if defined(EXCLUDE_PADDING)
start_x = max((int4)0, start_x);
start_y = max(0, start_y);
#endif /* defined(EXCLUDE_PADDING) */
return (VEC_DATA_TYPE(ACC_DATA_TYPE, 4))(1.f) / CONVERT_VECTOR4(ACC_DATA_TYPE)(((int4)(end_y - start_y)) * (end_x - start_x));
}
/** Performs an optimized pooling function of pool size equal to 3 when the stride_x is less equal than 3
*
* @note Datatype must be passed using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types are F16/F32;
* @note In case of average pooling the following information must be passed at compile time:
* -DPOOL_AVG or -DPOOL_L2 must be provided otherwise max pooling will be performed.
* -DMAX_WIDTH and -DMAX_HEIGHT which are the maximum accessible indeces in x and y dimensions (width + pad)
* -DSTRIDE_X and -DSTRIDE_Y which are the steps of the window along the x and y directions
* -DPAD_X and -DPAD_Y which are the pooling paddings in x and y dimension
*
* @param[in] input_ptr Pointer to the source image. Supported data types: F16/F32
* @param[in] input_stride_x Stride of the source image in X dimension (in bytes)
* @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] input_stride_y Stride of the source image in Y dimension (in bytes)
* @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes)
* @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] input_offset_first_element_in_bytes The offset of the first element in the source image
* @param[out] output_ptr Pointer to the destination image. Supported data types: same as @p input_ptr
* @param[in] output_stride_x Stride of the destination image in X dimension (in bytes)
* @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] output_stride_y Stride of the destination image in Y dimension (in bytes)
* @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] output_stride_z Stride of the source tensor in Z dimension (in bytes)
* @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination image
*/
__kernel void pooling_layer_optimized_3(
TENSOR3D_DECLARATION(input),
TENSOR3D_DECLARATION(output))
{
// Get pixels pointer
Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input);
Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output);
VEC_DATA_TYPE(ACC_DATA_TYPE, 4)
res;
// Perform pooling 3x3 for 4 output elements
POOLING3x3(res, input, output);
#if defined(POOL_AVG) || defined(POOL_L2)
// Divide by pool region in case of average pooling
res *= calculate_avg_scale4(3, MAX_WIDTH, MAX_HEIGHT, PAD_X, PAD_Y, STRIDE_X, STRIDE_Y);
#endif /* defined(POOL_AVG) || defined(POOL_L2) */
#if defined(POOL_L2)
// Take square root of the result in L2 pooling
res = SQRT_OP(res);
#endif /* defined(POOL_L2) */
vstore4(CONVERT(res, VEC_DATA_TYPE(DATA_TYPE, 4)), 0, (__global DATA_TYPE *)output.ptr);
}
#endif // defined(POOLING3x3)
#if defined(POOL_SIZE_X) && defined(POOL_SIZE_Y)
/** Performs a pooling function of pool size equal to N (NCHW)
*
* @note Datatype must be passed using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types are F16/F32;
* @note Pool sizes must be passed using -DPOOL_SIZE_X and -DPOOL_SIZE_Y e.g. -DPOOL_SIZE_X=13;
* @note In case of average pooling the following information must be passed at compile time:
* -DPOOL_AVG must be provided otherwise max pooling will be performed.
* -DMAX_WIDTH and -DMAX_HEIGHT which are the maximum accessible indeces in x and y dimensions (width + pad)
* -DSTRIDE_X and -DSTRIDE_Y which are the steps of the window along the x and y directions
* -DPAD_X and -DPAD_Y which are the pooling paddings in x and y dimension
* @note The initial value for the pooling operation must be passed at compile time using -DINITIAL_VALUE e.g. -DINITIAL_VALUE=0
*
* @param[in] input_ptr Pointer to the source image. Supported data types: F16/F32
* @param[in] input_stride_x Stride of the source image in X dimension (in bytes)
* @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] input_stride_y Stride of the source image in Y dimension (in bytes)
* @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes)
* @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] input_offset_first_element_in_bytes The offset of the first element in the source image
* @param[out] output_ptr Pointer to the destination image. Supported data types: same as @p input_ptr
* @param[in] output_stride_x Stride of the destination image in X dimension (in bytes)
* @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] output_stride_y Stride of the destination image in Y dimension (in bytes)
* @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] output_stride_z Stride of the source tensor in Z dimension (in bytes)
* @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination image
*/
__kernel void pooling_layer_MxN_nchw(
TENSOR3D_DECLARATION(input),
TENSOR3D_DECLARATION(output))
{
// Get pixels pointer
Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input);
Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output);
VEC_DATA_TYPE(ACC_DATA_TYPE, 8)
vdata = INITIAL_VALUE;
ACC_DATA_TYPE sdata = INITIAL_VALUE;
// Load data
for(int y = 0; y < POOL_SIZE_Y; y++)
{
int x = 0;
for(; x <= ((int)POOL_SIZE_X - 8); x += 8)
{
VEC_DATA_TYPE(ACC_DATA_TYPE, 8)
data0 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(8, 0, (__global DATA_TYPE *)tensor3D_offset(&input, x, y, 0));
#if defined(POOL_L2)
// Raise to power of 2 for L2 Pooling
data0 *= data0;
#endif /* defined(POOL_L2) */
vdata = POOL_OP(vdata, data0);
}
// Leftover
for(; x < (int)POOL_SIZE_X; ++x)
{
ACC_DATA_TYPE data0 = (ACC_DATA_TYPE)(*((__global DATA_TYPE *)tensor3D_offset(&input, x, y, 0)));
#if defined(POOL_L2)
// Raise to power of 2 for L2 Pooling
data0 *= data0;
#endif /* defined(POOL_L2) */
sdata = POOL_OP(sdata, data0);
}
}
// Reduce result
VEC_DATA_TYPE(ACC_DATA_TYPE, 4)
reduce4 = POOL_OP(vdata.s0123, vdata.s4567);
VEC_DATA_TYPE(ACC_DATA_TYPE, 2)
reduce2 = POOL_OP(reduce4.s01, reduce4.s23);
ACC_DATA_TYPE res = POOL_OP(reduce2.s0, reduce2.s1);
res = POOL_OP(res, sdata);
#if defined(POOL_AVG) || defined(POOL_L2)
// Divide by pool region in case of average pooling
res = DIV_OP(res, calculate_avg_scale(POOL_SIZE_X, POOL_SIZE_Y, MAX_WIDTH, MAX_HEIGHT, PAD_X, PAD_Y, STRIDE_X, STRIDE_Y));
#endif /* defined(POOL_AVG) || defined(POOL_L2) */
#if defined(POOL_L2)
// Take square root of the result in L2 pooling
res = SQRT_OP(res);
#endif /* defined(POOL_L2) */
// Store result
*(__global DATA_TYPE *)output.ptr = (DATA_TYPE)res;
}
#endif // defined(POOL_SIZE_X) && defined(POOL_SIZE_Y)
ACC_DATA_TYPE calculate_avg_scale_nhwc(const int pool_size_x, const int pool_size_y, int upper_bound_w, int upper_bound_h,
const int pad_x, const int pad_y, const int stride_x, const int stride_y)
{
int start_x = get_global_id(1) * stride_x - pad_x;
#if defined(DST_DEPTH)
int start_y = (get_global_id(2) % DST_DEPTH) * stride_y - pad_y;
#else /* defined(DST_DEPTH) */
int start_y = get_global_id(2) * stride_y - pad_y;
#endif /* defined(DST_DEPTH) */
#if !defined(EXCLUDE_PADDING)
upper_bound_w += pad_x;
upper_bound_h += pad_y;
#endif /* defined(EXCLUDE_PADDING) */
const int end_x = min(start_x + pool_size_x, upper_bound_w);
const int end_y = min(start_y + pool_size_y, upper_bound_h);
#if defined(EXCLUDE_PADDING)
start_x = max(0, start_x);
start_y = max(0, start_y);
#endif /* defined(EXCLUDE_PADDING) */
return ((end_y - start_y) * (end_x - start_x));
}
/** Performs a pooling function of pool size equal to N (NHWC)
*
* @note Datatype must be passed using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types are F16/F32
* @note Pool sizes must be passed using -DPOOL_SIZE_X and -DPOOL_SIZE_Y e.g. -DPOOL_SIZE_X=13;
* @note Tensors width and height must be passed at compile time using -DMAX_WIDTH and -DMAX_HEIGHT
* @note Strides must be passed at compile time using -DSTRIDE_X and -DSTRIDE_Y which are the steps of the window along the x and y directions
* @note Pad values must be passed at compile time using -DPAD_X and -DPAD_Y which are the pooling paddings in x and y dimension
* @note In case of average pooling the following information must be passed at compile time:
* -DPOOL_AVG must be provided otherwise max pooling will be performed.
* @note The initial value for the pooling operation must be passed at compile time using -DINITIAL_VALUE e.g. -DINITIAL_VALUE=0
*
* @param[in] input_ptr Pointer to the source image. Supported data types: F16/F32
* @param[in] input_stride_x Stride of the source image in X dimension (in bytes)
* @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] input_stride_y Stride of the source image in Y dimension (in bytes)
* @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes)
* @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] input_stride_w Stride of the source tensor in W dimension (in bytes)
* @param[in] input_step_w input_stride_w * number of elements along W processed per workitem(in bytes)
* @param[in] input_offset_first_element_in_bytes The offset of the first element in the source image
* @param[out] output_ptr Pointer to the destination image. Supported data types: same as @p input_ptr
* @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes)
* @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] output_stride_y Stride of the destination tensor in Y dimension (in bytes)
* @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] output_stride_z Stride of the destination tensor in Z dimension (in bytes)
* @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] output_stride_w Stride of the destination tensor in W dimension (in bytes)
* @param[in] output_step_w output_stride_w * number of elements along W processed per workitem(in bytes)
* @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination image
*/
__kernel void pooling_layer_MxN_nhwc(
TENSOR4D_DECLARATION(input),
TENSOR4D_DECLARATION(output))
{
// Get pixels pointer
#if defined(DST_DEPTH)
Tensor4D input = CONVERT_TO_TENSOR4D_STRUCT(input, DST_DEPTH);
Tensor4D output = CONVERT_TO_TENSOR4D_STRUCT(output, DST_DEPTH);
#else /* defined(DST_DEPTH) */
Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input);
Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output);
#endif /* defined(DST_DEPTH) */
VEC_DATA_TYPE(ACC_DATA_TYPE, 8)
vdata = INITIAL_VALUE;
const int idx_width = get_global_id(1) * STRIDE_X;
#if defined(DST_DEPTH)
const int idx_height = (get_global_id(2) % DST_DEPTH) * STRIDE_Y;
#else /* defined(DST_DEPTH) */
const int idx_height = get_global_id(2) * STRIDE_Y;
#endif /* defined(DST_DEPTH) */
for(int y = 0; y < POOL_SIZE_Y; ++y)
{
int y1 = select(y, PAD_Y - idx_height, y + idx_height - PAD_Y < 0 || y + idx_height - PAD_Y >= MAX_HEIGHT);
for(int x = 0; x < POOL_SIZE_X; ++x)
{
int x1 = select(x, PAD_X - idx_width - 1, x + idx_width - PAD_X < 0 || x + idx_width - PAD_X >= MAX_WIDTH);
x1 = select(x1, PAD_X - idx_width - 1, y != y1);
#if defined(DST_DEPTH)
VEC_DATA_TYPE(ACC_DATA_TYPE, 8)
data0 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(8, 0, (__global DATA_TYPE *)tensor4D_offset(&input, 0, x1 - PAD_X, y1 - PAD_Y, 0));
#else /* defined(DST_DEPTH) */
VEC_DATA_TYPE(ACC_DATA_TYPE, 8)
data0 = VLOAD_AND_CONVERT_TO_ACC_DATA_TYPE(8, 0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, x1 - PAD_X, y1 - PAD_Y));
#endif /* defined(DST_DEPTH) */
#if defined(POOL_L2)
// Raise to power of 2 for L2 Pooling
data0 *= data0;
#endif /* defined(POOL_L2) */
vdata = POOL_OP(vdata, CONVERT(data0, VEC_DATA_TYPE(ACC_DATA_TYPE, 8)));
}
}
#if defined(POOL_AVG) || defined(POOL_L2)
// Divide by pool region in case of average pooling
vdata = DIV_OP_NHWC(vdata, calculate_avg_scale_nhwc(POOL_SIZE_X, POOL_SIZE_Y, MAX_WIDTH, MAX_HEIGHT, PAD_X, PAD_Y, STRIDE_X, STRIDE_Y));
#endif /* defined(POOL_AVG) || defined(POOL_L2) */
#if defined(POOL_L2)
// Take square root of the result in L2 pooling
vdata = SQRT_OP(vdata);
#endif /* defined(POOL_L2) */
// Store result
vstore8(CONVERT(vdata, VEC_DATA_TYPE(DATA_TYPE, 8)), 0, (__global DATA_TYPE *)output.ptr);
}
|
