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
|
/*
* Copyright (c) 2018-2021 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"
#define VECTOR_N VEC_DATA_TYPE(DATA_TYPE, VECTOR_SIZE)
#define COND_N SIGNED_INT_VEC_DATA_TYPE(DATA_TYPE, VECTOR_SIZE)
#if defined(IM2COL_3X3) || defined(IM2COL_9X9)
/** Store a 1x9 row or a 3x3 block in a boundary-aware manner to avoid paddings in the channel dimension
* @name IM2COL1X9_NHWC_STORE
*
* @note To use this macro for a 3x3 block, @p ROW has to be 0
*
* @param[in] VECTOR_SIZE The non-boundary vector width of @p DATA. Supported: 1(scalar), 2, 3, 4, 8, 16
* @param[in] BOUNDARY_VECTOR_SIZE The boundary vector width of @p DATA. Supported: 1-16, but has to be <= @p size
* @param[in] DATA_TYPE Data type of @p DATA
* @param[in] SRC_DEPTH Input channel size / depth
* @param[in] DATA Value variable base name
* @param[in] ROW The row number to store. Supported: 0-8
* @param[in] OUTPUT_PTR Output pointer
* @{
*/
#if defined(VECTOR_SIZE) && defined(BOUNDARY_VECTOR_SIZE) && BOUNDARY_VECTOR_SIZE < VECTOR_SIZE
#define IM2COL1X9_NHWC_STORE(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE, DATA_TYPE, SRC_DEPTH, DATA, ROW, OUTPUT_PTR) \
const bool at_channel_boundary = get_global_id(0) == 0; \
if(at_channel_boundary) \
{ \
IM2COL1X9_NHWC_STORE_PARTIAL(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE, DATA_TYPE, SRC_DEPTH, DATA, ROW, OUTPUT_PTR) \
} \
else \
{ \
IM2COL1X9_NHWC_STORE_NONPARTIAL(VECTOR_SIZE, DATA_TYPE, SRC_DEPTH, DATA, ROW, OUTPUT_PTR) \
}
#else // defined(VECTOR_SIZE) && defined(BOUNDARY_VECTOR_SIZE) && BOUNDARY_VECTOR_SIZE < VECTOR_SIZE
#define IM2COL1X9_NHWC_STORE(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE, DATA_TYPE, SRC_DEPTH, DATA, ROW, OUTPUT_PTR) \
IM2COL1X9_NHWC_STORE_NONPARTIAL(VECTOR_SIZE, DATA_TYPE, SRC_DEPTH, DATA, ROW, OUTPUT_PTR)
#endif // defined(VECTOR_SIZE) && defined(BOUNDARY_VECTOR_SIZE) && BOUNDARY_VECTOR_SIZE < VECTOR_SIZE
#define IM2COL1X9_NHWC_STORE_NONPARTIAL(VECTOR_SIZE, DATA_TYPE, SRC_DEPTH, DATA, ROW, OUTPUT_PTR) \
VSTORE(VECTOR_SIZE) \
(DATA##0, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (0 + ROW * 9) * SRC_DEPTH); \
VSTORE(VECTOR_SIZE) \
(DATA##1, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (1 + ROW * 9) * SRC_DEPTH); \
VSTORE(VECTOR_SIZE) \
(DATA##2, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (2 + ROW * 9) * SRC_DEPTH); \
VSTORE(VECTOR_SIZE) \
(DATA##3, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (3 + ROW * 9) * SRC_DEPTH); \
VSTORE(VECTOR_SIZE) \
(DATA##4, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (4 + ROW * 9) * SRC_DEPTH); \
VSTORE(VECTOR_SIZE) \
(DATA##5, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (5 + ROW * 9) * SRC_DEPTH); \
VSTORE(VECTOR_SIZE) \
(DATA##6, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (6 + ROW * 9) * SRC_DEPTH); \
VSTORE(VECTOR_SIZE) \
(DATA##7, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (7 + ROW * 9) * SRC_DEPTH); \
VSTORE(VECTOR_SIZE) \
(DATA##8, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (8 + ROW * 9) * SRC_DEPTH);
#define IM2COL1X9_NHWC_STORE_PARTIAL(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE, DATA_TYPE, SRC_DEPTH, DATA, ROW, OUTPUT_PTR) \
VSTORE_PARTIAL(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE) \
(DATA##0, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (0 + ROW * 9) * SRC_DEPTH); \
VSTORE_PARTIAL(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE) \
(DATA##1, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (1 + ROW * 9) * SRC_DEPTH); \
VSTORE_PARTIAL(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE) \
(DATA##2, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (2 + ROW * 9) * SRC_DEPTH); \
VSTORE_PARTIAL(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE) \
(DATA##3, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (3 + ROW * 9) * SRC_DEPTH); \
VSTORE_PARTIAL(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE) \
(DATA##4, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (4 + ROW * 9) * SRC_DEPTH); \
VSTORE_PARTIAL(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE) \
(DATA##5, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (5 + ROW * 9) * SRC_DEPTH); \
VSTORE_PARTIAL(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE) \
(DATA##6, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (6 + ROW * 9) * SRC_DEPTH); \
VSTORE_PARTIAL(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE) \
(DATA##7, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (7 + ROW * 9) * SRC_DEPTH); \
VSTORE_PARTIAL(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE) \
(DATA##8, 0, (__global DATA_TYPE *)(OUTPUT_PTR) + (8 + ROW * 9) * SRC_DEPTH);
/** @}*/
#endif // defined(IM2COL_3X3) || defined(IM2COL_9X9)
#if defined(IM2COL_3X3)
/** This kernel performs im2col when the kernel size is 3x3 and the data layout is NHWC
*
* @note This kernel computes VECTOR_SIZE elements
* @note This kernel stores VECTOR_SIZE or BOUNDARY_VECTOR_SIZE (if at boundary) elements
* @note The vector size must be passed at compile time using -DVECTOR_SIZE: e.g. -DVECTOR_SIZE=2
* @note The boundary vector size must be passed at compile time using -DBOUNDARY_VECTOR_SIZE: e.g. -DBOUNDARY_VECTOR_SIZE=1
* @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
* @note The width of output tensor after matrix multiplication must be passed at compile time using -DCONVOLVED_WIDTH: e.g. -DCONVOLVED_WIDTH=34
* @note The kernel depth must be passed at compile time using -DSRC_DEPTH: e.g. -DSRC_DEPTH=3
* @note The stride along the Y direction must be passed at compile time using -DSTRIDE_Y: e.g. -DSTRIDE_Y=1
* @note In case biases will be added to the convolution -DHAS_BIAS has to be passed to append the final matrix with 1 in each row.
*
* @param[in] src_ptr Pointer to the source tensor. Supported data types: QASYMM8_SIGNED/QASYMM8/F16/F32
* @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 Y 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] src_stride_w Stride of the source tensor in W dimension (in bytes).
* @param[in] dst_stride_w Stride of the destination tensor in W dimension (in bytes).
*/
__kernel void im2col3x3_nhwc(
TENSOR3D_DECLARATION(src),
IMAGE_DECLARATION(dst),
uint src_stride_w,
uint dst_stride_w)
{
// input feature map, boundary-corrected (shift all non-boundary vectors by shift_amount) to avoid padding
const int shift_amount = (int)VECTOR_SIZE - (int)BOUNDARY_VECTOR_SIZE;
const int ch = max((int)(get_global_id(0) * VECTOR_SIZE) - shift_amount, 0);
const int yo = get_global_id(1);
const int batch = get_global_id(2); // batch size
// Calculate input indices
const int xi = (get_global_id(1) % CONVOLVED_WIDTH) * STRIDE_X;
const int yi = (get_global_id(1) / (int)CONVOLVED_WIDTH) * STRIDE_Y;
// Get input and output address
__global uchar *input_ptr = src_ptr + src_offset_first_element_in_bytes + ch * sizeof(DATA_TYPE) + batch * (int)src_stride_w;
__global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + ch * sizeof(DATA_TYPE) + yo * (int)dst_stride_y + batch * (int)dst_stride_w;
int yi_coord = 0;
int3 offset = 0;
// Clamp xi
int3 xi_offset = ((int3)xi + (int3)(0, 1, 2) * DILATION_X - (int3)PAD_LEFT);
#if PAD_LEFT != 0 || PAD_RIGHT != 0
#define CLAMP(x, min_val, max_val) min(max(x, min_val), max_val)
xi_offset = CLAMP(xi_offset, (int3)0, (int3)(SRC_WIDTH - 1));
#endif // PAD_LEFT != 0 || PAD_RIGHT != 0
// Multiply by src_stride_y as the width (X) dimension here is the second (y) dimension in src NHWC tensor
xi_offset *= (int3)src_stride_y;
// Out-of-bound condition for X
int3 x_cond = (((int3)xi + (int3)(0, 1, 2) * DILATION_X - (int3)PAD_LEFT) < (int3)0) || (((int3)xi + (int3)(0, 1, 2) * DILATION_X - (int3)PAD_LEFT) >= (int3)SRC_WIDTH);
// yi == 0
// Clamp yi
// yi_coord is casted to unsigned int in order to use just a min() operation
// A "-1" 32 bit signed variable converted to unsigned gives 4294967295
// This is a trick so that the values loaded in the padding areas are always from the last row (SRC_HEIGHT - 1),
// because of the negative yi_coord wrap-around, but it gets overwritten by PAD_VALUE immediately as the wrap-around
// also causes y_cond (y padding condition) to be satisfied
yi_coord = yi - (int)PAD_TOP;
// Clamp only if PAD_TOP or PAD_BOTTOM is not equal to 0
#if PAD_TOP != 0 || PAD_BOTTOM != 0
yi_coord = min((uint)yi_coord, (uint)(SRC_HEIGHT - 1));
#endif // PAD_TOP != 0 || PAD_BOTTOM != 0
// Compute offset
offset = xi_offset + (yi_coord * (int)src_stride_z);
// Load input values
VECTOR_N values0 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset.s0));
VECTOR_N values1 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset.s1));
VECTOR_N values2 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset.s2));
#if PAD_TOP != 0 || PAD_LEFT != 0 || PAD_BOTTOM != 0 || PAD_RIGHT != 0
// Replace invalid values with PAD_VALUE
int y_cond = (int)((uint)(yi - (int)PAD_TOP) >= (uint)(SRC_HEIGHT));
values0 = select(values0, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond.s0)));
values1 = select(values1, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond.s1)));
values2 = select(values2, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond.s2)));
#endif // PAD_TOP != 0 || PAD_LEFT != 0 || PAD_BOTTOM != 0 || PAD_RIGHT != 0
// yi == 1
// Clamp yi_coord (it can be negative if PAD_TOP > 1)
yi_coord = yi - (int)PAD_TOP + 1 * DILATION_Y;
// Clamp only if PAD_TOP or PAD_BOTTOM is not equal to 0
#if PAD_TOP != 0 || PAD_BOTTOM != 0
yi_coord = min((uint)yi_coord, (uint)(SRC_HEIGHT - 1));
#endif // PAD_TOP != 0 || PAD_BOTTOM != 0
// Compute offset
offset = xi_offset + (yi_coord * (int)src_stride_z);
// Load input values
VECTOR_N values3 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset.s0));
VECTOR_N values4 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset.s1));
VECTOR_N values5 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset.s2));
#if PAD_TOP != 0 || PAD_LEFT != 0 || PAD_BOTTOM != 0 || PAD_RIGHT != 0
// Replace invalid values with zeros
y_cond = (int)((uint)(yi - (int)PAD_TOP + 1 * DILATION_Y) >= (uint)(SRC_HEIGHT));
values3 = select(values3, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond.s0)));
values4 = select(values4, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond.s1)));
values5 = select(values5, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond.s2)));
#endif // PAD_TOP != 0 || PAD_LEFT != 0 || PAD_BOTTOM != 0 || PAD_RIGHT != 0
// yi == 2
// Clamp yi_coord
yi_coord = yi - (int)PAD_TOP + 2 * DILATION_Y;
// Clamp only if PAD_TOP or PAD_BOTTOM is not equal to 0
#if PAD_TOP != 0 || PAD_BOTTOM != 0
yi_coord = min((uint)yi_coord, (uint)(SRC_HEIGHT - 1));
#endif // PAD_TOP != 0 || PAD_BOTTOM != 0
// Compute offset
offset = xi_offset + (yi_coord * (int)src_stride_z);
// Load input values
VECTOR_N values6 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset.s0));
VECTOR_N values7 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset.s1));
VECTOR_N values8 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset.s2));
#if PAD_TOP != 0 || PAD_LEFT != 0 || PAD_BOTTOM != 0 || PAD_RIGHT != 0
// Replace invalid values with PAD_VALUE
y_cond = (int)((uint)(yi - (int)PAD_TOP + 2 * DILATION_Y) >= (uint)(SRC_HEIGHT));
values6 = select(values6, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond.s0)));
values7 = select(values7, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond.s1)));
values8 = select(values8, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond.s2)));
#endif // PAD_TOP != 0 || PAD_LEFT != 0 || PAD_BOTTOM != 0 || PAD_RIGHT != 0
// Store in a boundary-aware way to avoid padding
IM2COL1X9_NHWC_STORE(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE, DATA_TYPE, SRC_DEPTH, values, 0, output_ptr)
#ifdef HAS_BIAS
// We can use VECTOR_SIZE instead of BOUNDARY_VECTOR_SIZE even if it's at the boundary. This is because the bias is
// added at the end of the channel, while the boundary vec is at the beginning of the channel.
// The only case where the boundary vec is at the end of the channel is when there's only a single boundary vec in
// the whole channel dimension, but in that case VECTOR_SIZE is also equal to BOUNDARY_VECTOR_SIZE
// See the value of num_elems_processed_per_iteration in configure_opencl_kernel method in CLIm2ColKernel.cpp
if((ch + VECTOR_SIZE) >= SRC_DEPTH)
{
*((__global DATA_TYPE *)(output_ptr) - ch + SRC_DEPTH * 9) = 1.0f;
}
#endif // HAS_BIAS
}
#endif // defined(IM2COL_3X3)
#if defined(IM2COL_9X9)
#if PAD_TOP != 0 || PAD_LEFT != 0 || PAD_BOTTOM != 0 || PAD_RIGHT != 0
#define IM2COL1x9(i) \
({ \
yi_coord = yi - (int)PAD_TOP + i * DILATION_Y; \
yi_coord = min((uint)yi_coord, (uint)(SRC_HEIGHT - 1)); \
\
offset0 = xi_offset0 + (yi_coord * (int)src_stride_z); \
offset1 = xi_offset1 + (yi_coord * (int)src_stride_z); \
\
VECTOR_N values0 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s0)); \
VECTOR_N values1 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s1)); \
VECTOR_N values2 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s2)); \
VECTOR_N values3 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s3)); \
VECTOR_N values4 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s4)); \
VECTOR_N values5 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s5)); \
VECTOR_N values6 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s6)); \
VECTOR_N values7 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s7)); \
VECTOR_N values8 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset1)); \
\
int y_cond = (int)((uint)(yi - (int)PAD_TOP + i * DILATION_Y) >= (uint)(SRC_HEIGHT)); \
values0 = select(values0, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond0.s0))); \
values1 = select(values1, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond0.s1))); \
values2 = select(values2, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond0.s2))); \
values3 = select(values3, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond0.s3))); \
values4 = select(values4, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond0.s4))); \
values5 = select(values5, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond0.s5))); \
values6 = select(values6, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond0.s6))); \
values7 = select(values7, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond0.s7))); \
values8 = select(values8, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)y_cond || (COND_N)(x_cond1))); \
\
IM2COL1X9_NHWC_STORE(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE, DATA_TYPE, SRC_DEPTH, values, i, output_ptr) \
})
#else // PAD_TOP != 0 || PAD_LEFT != 0 || PAD_BOTTOM != 0 || PAD_RIGHT != 0
#define IM2COL1x9(i) \
({ \
yi_coord = yi - (int)PAD_TOP + i * DILATION_Y; \
yi_coord = min((uint)yi_coord, (uint)(SRC_HEIGHT - 1)); \
\
offset0 = xi_offset0 + (yi_coord * (int)src_stride_z); \
offset1 = xi_offset1 + (yi_coord * (int)src_stride_z); \
\
VECTOR_N values0 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s0)); \
VECTOR_N values1 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s1)); \
VECTOR_N values2 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s2)); \
VECTOR_N values3 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s3)); \
VECTOR_N values4 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s4)); \
VECTOR_N values5 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s5)); \
VECTOR_N values6 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s6)); \
VECTOR_N values7 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset0.s7)); \
VECTOR_N values8 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset1)); \
\
IM2COL1X9_NHWC_STORE(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE, DATA_TYPE, SRC_DEPTH, values, i, output_ptr) \
})
#endif // PAD_TOP != 0 || PAD_LEFT != 0 || PAD_BOTTOM != 0 || PAD_RIGHT != 0
/** This kernel performs im2col when the kernel size is 9x9 and the data layout is NHWC
*
* @note This kernel computes VECTOR_SIZE elements
* @note This kernel stores VECTOR_SIZE or BOUNDARY_VECTOR_SIZE (if at boundary) elements
* @note The vector size must be passed at compile time using -DVECTOR_SIZE: e.g. -DVECTOR_SIZE=2
* @note The boundary vector size must be passed at compile time using -DBOUNDARY_VECTOR_SIZE: e.g. -DBOUNDARY_VECTOR_SIZE=1
* @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
* @note The width of output tensor after matrix multiplication must be passed at compile time using -DCONVOLVED_WIDTH: e.g. -DCONVOLVED_WIDTH=34
* @note The kernel depth must be passed at compile time using -DSRC_DEPTH: e.g. -DSRC_DEPTH=3
* @note The stride along the Y direction must be passed at compile time using -DSTRIDE_Y: e.g. -DSTRIDE_Y=1
* @note In case biases will be added to the convolution -DHAS_BIAS has to be passed to append the final matrix with 1 in each row.
*
* @param[in] src_ptr Pointer to the source tensor. Supported data types: QASYMM8_SIGNED/QASYMM8/F16/F32
* @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 Y 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] src_stride_w Stride of the source tensor in W dimension (in bytes).
* @param[in] dst_stride_w Stride of the destination tensor in W dimension (in bytes).
*/
__kernel void im2col9x9_nhwc(
TENSOR3D_DECLARATION(src),
IMAGE_DECLARATION(dst),
uint src_stride_w,
uint dst_stride_w)
{
// input feature map, boundary-corrected (shift all non-boundary vectors by shift_amount) to avoid padding
const int shift_amount = (int)VECTOR_SIZE - (int)BOUNDARY_VECTOR_SIZE;
const int ch = max((int)(get_global_id(0) * VECTOR_SIZE) - shift_amount, 0);
const int yo = get_global_id(1);
const int batch = get_global_id(2); // batch size
// Calculate input indices
const int xi = (get_global_id(1) % CONVOLVED_WIDTH) * STRIDE_X;
const int yi = (get_global_id(1) / (int)CONVOLVED_WIDTH) * STRIDE_Y;
// Get input and output address
__global uchar *input_ptr = src_ptr + src_offset_first_element_in_bytes + ch * sizeof(DATA_TYPE) + batch * (int)src_stride_w;
__global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + ch * sizeof(DATA_TYPE) + yo * (int)dst_stride_y + batch * (int)dst_stride_w;
int yi_coord = 0;
int8 offset0 = 0;
int offset1 = 0;
// Clamp xi
int8 xi_offset0 = ((int8)xi + (int8)(0, 1, 2, 3, 4, 5, 6, 7) * DILATION_X - (int8)PAD_LEFT);
int xi_offset1 = ((int)xi + (int)(8) * DILATION_X - (int)PAD_LEFT);
#if PAD_LEFT != 0 || PAD_RIGHT != 0
#define CLAMP(x, min_val, max_val) min(max(x, min_val), max_val)
xi_offset0 = CLAMP(xi_offset0, (int8)0, (int8)(SRC_WIDTH - 1));
xi_offset1 = CLAMP(xi_offset1, (int)0, (int)(SRC_WIDTH - 1));
#endif // PAD_LEFT != 0 || PAD_RIGHT != 0
xi_offset0 *= (int8)src_stride_y;
xi_offset1 *= (int)src_stride_y;
// Out-of-bound condition for X
int8 x_cond0 = (((int8)xi + (int8)(0, 1, 2, 3, 4, 5, 6, 7) * DILATION_X - (int8)PAD_LEFT) < (int8)0) || (((int8)xi + (int8)(0, 1, 2, 3, 4, 5, 6, 7) * DILATION_X - (int8)PAD_LEFT) >= (int8)SRC_WIDTH);
int x_cond1 = (((int)xi + (int)(8) * DILATION_X - (int)PAD_LEFT) < (int)0) || (((int)xi + (int)(8) * DILATION_X - (int)PAD_LEFT) >= (int)SRC_WIDTH);
IM2COL1x9(0);
IM2COL1x9(1);
IM2COL1x9(2);
IM2COL1x9(3);
IM2COL1x9(4);
IM2COL1x9(5);
IM2COL1x9(6);
IM2COL1x9(7);
IM2COL1x9(8);
#ifdef HAS_BIAS
// We can use VECTOR_SIZE instead of BOUNDARY_VECTOR_SIZE even if it's at the boundary. This is because the bias is
// added at the end of the channel, while the boundary vec is at the beginning of the channel.
// The only case where the boundary vec is at the end of the channel is when there's only a single boundary vec in
// the whole channel dimension, but in that case VECTOR_SIZE is also equal to BOUNDARY_VECTOR_SIZE
// See the value of num_elems_processed_per_iteration in configure_opencl_kernel method in CLIm2ColKernel.cpp
if((ch + VECTOR_SIZE) >= SRC_DEPTH)
{
*((__global DATA_TYPE *)(output_ptr) - ch + SRC_DEPTH * 81) = 1.0f;
}
#endif // HAS_BIAS
}
#endif // defined(IM2COL_9X9)
#if defined(IM2COL_GENERIC)
/** This opencl kernel performs a generic im2col implementation when the data layout is NHWC
*
* @note This kernel computes VECTOR_SIZE elements
* @note This kernel stores VECTOR_SIZE or BOUNDARY_VECTOR_SIZE (if at boundary) elements
* @note The vector size must be passed at compile time using -DVECTOR_SIZE: e.g. -DVECTOR_SIZE=2
* @note The boundary vector size must be passed at compile time using -DBOUNDARY_VECTOR_SIZE: e.g. -DBOUNDARY_VECTOR_SIZE=1
* @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
* @note The width and height of the input tensor must be passed at compile time using -DSRC_WIDTH and -DSRC_HEIGHT: e.g. -DSRC_WIDTH=128 and -DSRC_HEIGHT=128
* @note The width of output tensor after matrix multiplication must be passed at compile time using -DCONVOLVED_WIDTH: e.g. -DCONVOLVED_WIDTH=34
* @note The kernel width, height and depth must be passed at compile time using -DKERNEL_WIDTH, -DKERNEL_HEIGHT and -DSRC_DEPTH: e.g. -DKERNEL_WIDTH=3, -DKERNEL_HEIGHT=3 and -DSRC_DEPTH=64
* @note The pad_left, pad_right, pad_top and pad_bottom must be passed at compile time using -DPAD_LEFT, -DPAD_RIGHT, -DPAD_TOP and -DPAD_BOTTOM: e.g. -DPAD_LEFT=1, -DPAD_RIGHT=2, -DPAD_TOP=3 and -DPAD_BOTTOM=2
* @note The zero value to store in case we load values out-of-bounds must be passed at compile time using -DPAD_VALUE: e.g. -DPAD_VALUE=0.0
* @note The stride along the X and Y directions must be passed at compile time using -DSTRIDE_X and -DSTRIDE_Y: e.g. -DSTRIDE_X=1 and -DSTRIDE_Y=1
* @note The dilation_x and dilation_y must be passed at compile time using -DDILATION_X and -DDILATION_Y: e.g. -DDILATION_X=1, -DDILATION_Y=1
* @note In case biases will be added to the convolution -DHAS_BIAS has to be passed to append the final matrix with 1 in each row.
*
* @param[in] src_ptr Pointer to the source tensor. Supported data types: QASYMM8_SIGNED/QASYMM8/F16/F32
* @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 Y 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] src_stride_w Stride of the source tensor in W dimension (in bytes).
* @param[in] dst_stride_w Stride of the destination tensor in W dimension (in bytes).
*/
__kernel void im2col_generic_nhwc(
TENSOR3D_DECLARATION(src),
IMAGE_DECLARATION(dst),
uint src_stride_w,
uint dst_stride_w)
{
// input feature map, boundary-corrected (shift all non-boundary vectors by shift_amount) to avoid padding
const int shift_amount = (int)VECTOR_SIZE - (int)BOUNDARY_VECTOR_SIZE;
const int ch = max((int)(get_global_id(0) * VECTOR_SIZE) - shift_amount, 0);
const int yo = get_global_id(1);
const int batch = get_global_id(2); // batch size
// Calculate input indices
const int xi = (yo % CONVOLVED_WIDTH) * STRIDE_X;
const int yi = (yo / (int)CONVOLVED_WIDTH) * STRIDE_Y;
// Get input and output address
const int stride_x = ch * sizeof(DATA_TYPE);
__global uchar *input_ptr = src_ptr + src_offset_first_element_in_bytes + stride_x + batch * (int)src_stride_w;
__global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + stride_x + yo * (int)dst_stride_y + batch * (int)dst_stride_w;
int i = 0;
for(int yk = 0; yk < KERNEL_HEIGHT; ++yk)
{
// Clamp yi_coord
int yi_coord = yi + yk * DILATION_Y - (int)PAD_TOP;
yi_coord = clamp(yi_coord, (int)0, (int)(SRC_HEIGHT - 1));
// Out-of-bound condition for Y
int y_border_condition = ((yi + yk * DILATION_Y - (int)PAD_TOP) < (int)0) || ((yi + yk * DILATION_Y - (int)PAD_TOP) >= (int)SRC_HEIGHT);
for(int xk = 0; xk < KERNEL_WIDTH; ++xk)
{
// Clamp xi_coord
int xi_coord = (xi + xk * DILATION_X - (int)PAD_LEFT);
xi_coord = clamp(xi_coord, (int)0, (int)(SRC_WIDTH - 1));
// Out-of-bound condition for X
int x_border_condition = ((xi + xk * DILATION_X - (int)PAD_LEFT) < (int)0) || ((xi + xk * DILATION_X - (int)PAD_LEFT) >= (int)SRC_WIDTH);
int offset = xi_coord * (int)src_stride_y + (yi_coord * (int)src_stride_z);
VECTOR_N values0 = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(input_ptr + offset));
#if PAD_LEFT != 0 || PAD_TOP != 0 || PAD_RIGHT != 0 || PAD_BOTTOM != 0
// Replace with PAD_VALUE if the value is out-of-bound
values0 = select(values0, (VECTOR_N)PAD_VALUE, (COND_N)((COND_N)x_border_condition || (COND_N)(y_border_condition)));
#endif // PAD_LEFT != 0 || PAD_TOP != 0 || PAD_RIGHT != 0 || PAD_BOTTOM != 0
// Store in a boundary-aware way to avoid padding
#if BOUNDARY_VECTOR_SIZE != VECTOR_SIZE
const bool at_channel_boundary = get_global_id(0) == 0;
if(at_channel_boundary)
{
VSTORE_PARTIAL(VECTOR_SIZE, BOUNDARY_VECTOR_SIZE)
(values0, 0, (__global DATA_TYPE *)(output_ptr) + i * (int)SRC_DEPTH);
}
else // at_channel_boundary
#endif // BOUNDARY_VECTOR_SIZE != VECTOR_SIZE
{
VSTORE(VECTOR_SIZE)
(values0, 0, (__global DATA_TYPE *)(output_ptr) + i * (int)SRC_DEPTH);
}
i++;
}
}
#ifdef HAS_BIAS
// We can use VECTOR_SIZE instead of BOUNDARY_VECTOR_SIZE even if it's at the boundary. This is because the bias is
// added at the end of the channel, while the boundary vec is at the beginning of the channel.
// The only case where the boundary vec is at the end of the channel is when there's only a single boundary vec in
// the whole channel dimension, but in that case VECTOR_SIZE is also equal to BOUNDARY_VECTOR_SIZE
// See the value of num_elems_processed_per_iteration in configure_opencl_kernel method in CLIm2ColKernel.cpp
if((ch + VECTOR_SIZE) >= SRC_DEPTH)
{
*((__global DATA_TYPE *)(output_ptr) - ch + SRC_DEPTH * KERNEL_WIDTH * KERNEL_HEIGHT) = 1.0f;
}
#endif // HAS_BIAS
}
#endif // defined(IM2COL_GENERIC)
|
