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
|
/*
* Copyright (c) 2016-2020 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"
#include "helpers_asymm.h"
#if defined(FLOAT_DATA_TYPE)
#define ISGREATER(x, y) isgreater(x, y)
#define ISLESS(x, y) isless(x, y)
#else // !FLOAT_DATA_TYPE
#if defined(WIDTH)
#define ISGREATER(x, y) (x > y) ? 1 : 0
#define ISLESS(x, y) (x < y) ? 1 : 0
#else // !defined(WIDTH)
#define ISGREATER(x, y) select((int16)0, (int16)-1, x > y)
#define ISLESS(x, y) select((int16)0, (int16)-1, x < y)
#endif // defined(WIDTH)
#endif // defined(FLOAT_DATA_TYPE)
/** Calculate square sum of a vector
*
* @param[in] input Pointer to the first pixel.
*
* @return square sum of vector.
*/
inline DATA_TYPE square_sum(__global const DATA_TYPE *input)
{
VEC_DATA_TYPE(DATA_TYPE, 16)
in = vload16(0, input);
in *= in;
in.s01234567 += in.s89ABCDEF;
in.s0123 += in.s4567;
in.s01 += in.s23;
return (in.s0 + in.s1);
}
/** Calculate sum of a vector
*
* @param[in] input Pointer to the first pixel.
*
* @return sum of vector.
*/
inline DATA_TYPE sum(__global const DATA_TYPE *input)
{
VEC_DATA_TYPE(DATA_TYPE, 16)
in = vload16(0, input);
in.s01234567 += in.s89ABCDEF;
in.s0123 += in.s4567;
in.s01 += in.s23;
return (in.s0 + in.s1);
}
/** Calculate product of a vector
*
* @param[in] input Pointer to the first pixel.
*
* @return product of vector.
*/
inline DATA_TYPE product(__global const DATA_TYPE *input)
{
VEC_DATA_TYPE(DATA_TYPE, 16)
in = vload16(0, input);
in.s01234567 *= in.s89ABCDEF;
in.s0123 *= in.s4567;
in.s01 *= in.s23;
return (in.s0 * in.s1);
}
#if defined(OPERATION)
/** This kernel performs parallel reduction given an operation on x-axis.
*
* @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
* @note The operation we want to perform must be passed at compile time using -DOPERATION e.g. -DOPERATION=square_sum
* @note The mean flag must be passed at compile time using -DMEAN if we want to compute the mean value
* @note The product flag must be passed at compile time using -DPROD if we want to compute the product, otherwise sum will be used
* @note The width size must be passed at compile time using -DWIDTH e.g. -DWIDTH=128 if we want to compute the mean value
*
* @param[in] src_ptr Pointer to the source tensor. Supported data types: 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_offset_first_element_in_bytes The offset of the first element in the source tensor
* @param[in] partial_res_ptr The local buffer to hold partial result values. Supported data types: same as @p src_ptr
* @param[in] partial_res_stride_x Stride of the output tensor in X dimension (in bytes)
* @param[in] partial_res_step_x partial_res_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] partial_res_stride_y Stride of the output tensor in Y dimension (in bytes)
* @param[in] partial_res_step_y partial_res_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] partial_res_offset_first_element_in_bytes The offset of the first element in the source tensor
* @param[in] local_results Local buffer for storing the partial result
*/
__kernel void reduction_operation_x(
IMAGE_DECLARATION(src),
IMAGE_DECLARATION(partial_res),
__local DATA_TYPE *local_results)
{
Image src = CONVERT_TO_IMAGE_STRUCT(src);
Image partial_res = CONVERT_TO_IMAGE_STRUCT(partial_res);
unsigned int lsize = get_local_size(0);
unsigned int lid = get_local_id(0);
for(unsigned int y = 0; y < get_local_size(1); ++y)
{
local_results[lid] = OPERATION((__global DATA_TYPE *)offset(&src, 0, y));
barrier(CLK_LOCAL_MEM_FENCE);
// Perform parallel reduction
for(unsigned int i = lsize >> 1; i > 0; i >>= 1)
{
if(lid < i)
{
#if defined(PROD)
local_results[lid] *= local_results[lid + i];
#else // !defined(PROD)
local_results[lid] += local_results[lid + i];
#endif // defined(PROD)
}
barrier(CLK_LOCAL_MEM_FENCE);
}
if(lid == 0)
{
#if defined(MEAN) && defined(WIDTH)
if(y == get_local_size(1) - 1)
{
local_results[0] /= WIDTH;
}
#endif // defined(MEAN) && defined(WIDTH)
((__global DATA_TYPE *)offset(&partial_res, get_group_id(0), y))[0] = local_results[0];
}
}
}
#endif // defined(OPERATION)
#if defined(WIDTH)
/** This kernel performs reduction on x-axis. (Non parallel)
*
* @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
* @note The width size must be passed at compile time using -DWIDTH e.g. -DWIDTH=128
* @note The product flag must be passed at compile time using -DPROD if we want to compute the product, otherwise sum will be used
* @note In case of MIN and MAX the condition data type must be passed at compile time using -DCOND_DATA_TYPE e.g. -DCOND_DATA_TYPE=short
*
* @param[in] src_ptr Pointer to the source tensor. Supported data types: S32/F16/F32 and QASYMM8 for operation MEAN
* @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_offset_first_element_in_bytes The offset of the first element in the source tensor
* @param[in] output_ptr The local buffer to hold sumed values. Supported data types: same as @p src_ptt
* @param[in] output_stride_x Stride of the output 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_offset_first_element_in_bytes The offset of the first element in the source tensor
*/
__kernel void reduction_operation_non_parallel_x(
VECTOR_DECLARATION(src),
VECTOR_DECLARATION(output))
{
Vector src = CONVERT_TO_VECTOR_STRUCT(src);
Vector output = CONVERT_TO_VECTOR_STRUCT(output);
DATA_TYPE_PROMOTED res = CONVERT(*((__global DATA_TYPE *)vector_offset(&src, 0)), DATA_TYPE_PROMOTED);
// Convert input into F32 in order to perform quantized multiplication
#if defined(PROD) && defined(OFFSET) && defined(SCALE)
float res_f = DEQUANTIZE(res, OFFSET, SCALE, DATA_TYPE_PROMOTED, 1);
#endif // defined(PROD) && defined(OFFSET) && defined(SCALE)
for(unsigned int x = 1; x < WIDTH; ++x)
{
DATA_TYPE_PROMOTED in = CONVERT(*((__global DATA_TYPE *)vector_offset(&src, x)), DATA_TYPE_PROMOTED);
#if defined(MIN)
res = select(res, in, CONVERT(ISLESS(in, res), COND_DATA_TYPE));
#elif defined(MAX)
res = select(res, in, CONVERT(ISGREATER(in, res), COND_DATA_TYPE));
#elif defined(PROD)
#if defined(OFFSET) && defined(SCALE)
res_f *= DEQUANTIZE(in, OFFSET, SCALE, DATA_TYPE_PROMOTED, 1);
#else // !(defined(OFFSET) && defined(SCALE))
res *= in;
#endif // defined(OFFSET) && defined(SCALE)
#else // defined(SUM))
res += in;
#endif // defined(MAX) || defined(MIN) || defined(PROD)
}
// Store result
#if defined(MEAN)
res /= WIDTH;
#endif // defined(MEAN)
// Subtract the offsets in case of quantized SUM
#if defined(SUM) && defined(OFFSET) && defined(SCALE)
res -= (WIDTH - 1) * OFFSET;
#endif // defined(OFFSET) && defined(OFFSET) && defined(SCALE)
// Re-quantize
#if defined(PROD) && defined(OFFSET) && defined(SCALE)
res = QUANTIZE(res_f, OFFSET, SCALE, DATA_TYPE_PROMOTED, 1);
#endif // defined(PROD) && defined(OFFSET) && defined(SCALE)
*((__global DATA_TYPE *)output.ptr) = CONVERT_SAT(res, DATA_TYPE);
}
#endif // defined(WIDTH)
#if defined(HEIGHT)
/** This kernel performs reduction on y-axis.
*
* @note The input data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
* @note The height size must be passed at compile time using -DHEIGHT e.g. -DHEIGHT=128
*
* @param[in] src_ptr Pointer to the source tensor. Supported data types: QASYMM8/S32/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_offset_first_element_in_bytes The offset of the first element in the source tensor
* @param[in] output_ptr The local buffer to hold sumed values. Supported data types: same as @p src_ptt
* @param[in] output_stride_x Stride of the output 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 output 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_offset_first_element_in_bytes The offset of the first element in the source tensor
*/
__kernel void reduction_operation_y(
IMAGE_DECLARATION(src),
IMAGE_DECLARATION(output))
{
Image src = CONVERT_TO_IMAGE_STRUCT(src);
Image output = CONVERT_TO_IMAGE_STRUCT(output);
VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16)
res = CONVERT(vload16(0, (__global DATA_TYPE *)offset(&src, 0, 0)), VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16));
// Convert input into F32 in order to perform quantized multiplication
#if defined(PROD) && defined(OFFSET) && defined(SCALE)
float16 res_f = DEQUANTIZE(res, OFFSET, SCALE, DATA_TYPE_PROMOTED, 16);
#endif // defined(PROD) && defined(OFFSET) && defined(SCALE)
#if defined(SUM_SQUARE)
res *= res;
#endif // defined(SUM_SQUARE)
for(unsigned int y = 1; y < HEIGHT; ++y)
{
VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16)
in = CONVERT(vload16(0, (__global DATA_TYPE *)offset(&src, 0, y)), VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16));
#if defined(MIN)
res = select(res, in, ISLESS(in, res));
#elif defined(MAX)
res = select(res, in, ISGREATER(in, res));
#else // !(defined(MAX) || defined(MIN))
#if defined(SUM_SQUARE)
in *= in;
#endif // defined(SUM_SQUARE)
#if defined(PROD)
#if defined(OFFSET) && defined(SCALE)
res_f *= DEQUANTIZE(in, OFFSET, SCALE, DATA_TYPE_PROMOTED, 16);
#else // !(defined(OFFSET) && defined(SCALE))
res *= in;
#endif // defined(OFFSET) && defined(SCALE)
#else // !defined(PROD)
res += in;
#endif // defined(PROD)
#endif // defined(MAX) || defined(MIN)
}
#if defined(MEAN)
res /= HEIGHT;
#endif // defined(MEAN)
// Subtract the offsets in case of quantized SUM
#if defined(SUM) && defined(OFFSET) && defined(SCALE)
res -= (HEIGHT - 1) * OFFSET;
#endif // defined(OFFSET) && defined(OFFSET) && defined(SCALE)
// Re-quantize
#if defined(PROD) && defined(OFFSET) && defined(SCALE)
res = QUANTIZE(res_f, OFFSET, SCALE, DATA_TYPE_PROMOTED, 16);
#endif // defined(PROD) && defined(OFFSET) && defined(SCALE)
// Store result
vstore16(CONVERT_SAT(res, VEC_DATA_TYPE(DATA_TYPE, 16)), 0, (__global DATA_TYPE *)output.ptr);
}
#endif // defined(HEIGHT)
#if defined(DEPTH)
/** This kernel performs reduction on z-axis.
*
* @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
* @note The depth size must be passed at compile time using -DDEPTH e.g. -DDEPTH=128
*
* @param[in] input_ptr Pointer to the source tensor. Supported data types: QASYMM8/S32/F16/F32
* @param[in] input_stride_x Stride of the source tensor 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 tensor 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 tensor
* @param[in] output_ptr The local buffer to hold sumed values. Supported data types: same as @p input_ptt
* @param[in] output_stride_x Stride of the output 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 output 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 output 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 source tensor
*/
__kernel void reduction_operation_z(
TENSOR3D_DECLARATION(input),
TENSOR3D_DECLARATION(output))
{
Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input);
Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output);
VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16)
res = CONVERT(vload16(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0)), VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16));
// Convert input into F32 in order to perform quantized multiplication
#if defined(PROD) && defined(OFFSET) && defined(SCALE)
float16 res_f = DEQUANTIZE(res, OFFSET, SCALE, DATA_TYPE_PROMOTED, 16);
#endif // defined(PROD) && defined(OFFSET) && defined(SCALE)
#if defined(COMPLEX)
VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16)
res1 = CONVERT(vload16(0, (__global DATA_TYPE *)tensor3D_offset(&input, 8, 0, 0)), VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16));
#endif // defined(COMPLEX)
#if defined(SUM_SQUARE)
res *= res;
#endif // defined(SUM_SQUARE)
for(unsigned int z = 1; z < DEPTH; ++z)
{
VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16)
in = CONVERT(vload16(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, z)), VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16));
#if defined(COMPLEX)
VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16)
in1 = CONVERT(vload16(0, (__global DATA_TYPE *)tensor3D_offset(&input, 8, 0, z)), VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16));
#endif // defined(COMPLEX)
#if defined(MIN)
res = select(res, in, ISLESS(in, res));
#elif defined(MAX)
res = select(res, in, ISGREATER(in, res));
#else // !(defined(MAX) || defined(MIN))
#if defined(SUM_SQUARE)
in *= in;
#endif // defined(SUM_SQUARE)
#if defined(PROD)
#if defined(OFFSET) && defined(SCALE)
res_f *= DEQUANTIZE(in, OFFSET, SCALE, DATA_TYPE_PROMOTED, 16);
#else // !(defined(OFFSET) && defined(SCALE))
res *= in;
#endif // defined(OFFSET) && defined(SCALE)
#else // !defined(PROD)
res += in;
#if defined(COMPLEX)
res1 += in1;
#endif // defined(COMPLEX)
#endif // defined(PROD)
#endif // defined(MAX) || defined(MIN)
}
#if defined(MEAN)
res /= DEPTH;
#endif // defined(MEAN)
// Subtract the offsets in case of quantized SUM
#if defined(SUM) && defined(OFFSET) && defined(SCALE)
res -= (DEPTH - 1) * OFFSET;
#endif // defined(OFFSET) && defined(OFFSET) && defined(SCALE)
// Re-quantize
#if defined(PROD) && defined(OFFSET) && defined(SCALE)
res = QUANTIZE(res_f, OFFSET, SCALE, DATA_TYPE_PROMOTED, 16);
#endif // defined(PROD) && defined(OFFSET) && defined(SCALE)
// Store result
vstore16(CONVERT_SAT(res, VEC_DATA_TYPE(DATA_TYPE, 16)), 0, (__global DATA_TYPE *)output.ptr);
#if defined(COMPLEX)
vstore16(CONVERT(res1, VEC_DATA_TYPE(DATA_TYPE, 16)), 0, (__global DATA_TYPE *)tensor3D_offset(&output, 8, 0, 0));
#endif // defined(COMPLEX)
}
#endif /* defined(DEPTH) */
#if defined(BATCH) && defined(DEPTH)
/** This kernel performs reduction on w-axis.
*
* @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
* @note The batch size must be passed at compile time using -DBATCH e.g. -DBATCH=128
* @note The depth size must be passed at compile time using -DBATCH e.g. -DDEPTH=128
*
* @param[in] input_ptr Pointer to the source tensor. Supported data types: QASYMM8/S32/F16/F32
* @param[in] input_stride_x Stride of the source tensor 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 tensor 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 tensor
* @param[in] output_ptr The local buffer to hold sumed values. Supported data types: same as @p input_ptt
* @param[in] output_stride_x Stride of the output 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 output 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 output 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 output 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 source tensor
*/
__kernel void reduction_operation_w(
TENSOR4D_DECLARATION(input),
TENSOR4D_DECLARATION(output))
{
Tensor4D input = CONVERT_TO_TENSOR4D_STRUCT(input, DEPTH);
Tensor4D output = CONVERT_TO_TENSOR4D_STRUCT(output, DEPTH);
VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16)
res = CONVERT(vload16(0, (__global DATA_TYPE *)tensor4D_offset(&input, 0, 0, 0, 0)), VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16));
// Convert input into F32 in order to perform quantized multiplication
#if defined(PROD) && defined(OFFSET) && defined(SCALE)
float16 res_f = DEQUANTIZE(res, OFFSET, SCALE, DATA_TYPE_PROMOTED, 16);
#endif // defined(PROD) && defined(OFFSET) && defined(SCALE)
#if defined(SUM_SQUARE)
res *= res;
#endif // defined(SUM_SQUARE)
for(unsigned int w = 1; w < BATCH; ++w)
{
VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16)
in = CONVERT(vload16(0, (__global DATA_TYPE *)tensor4D_offset(&input, 0, 0, 0, w)), VEC_DATA_TYPE(DATA_TYPE_PROMOTED, 16));
#if defined(MIN)
res = select(res, in, ISLESS(in, res));
#elif defined(MAX)
res = select(res, in, ISGREATER(in, res));
#else // !(defined(MAX) || defined(MIN))
#if defined(SUM_SQUARE)
in *= in;
#endif // defined(SUM_SQUARE)
#if defined(PROD)
#if defined(OFFSET) && defined(SCALE)
res_f *= DEQUANTIZE(in, OFFSET, SCALE, DATA_TYPE_PROMOTED, 16);
#else // !(defined(OFFSET) && defined(SCALE))
res *= in;
#endif // defined(OFFSET) && defined(SCALE)
#else // !defined(PROD)
res += in;
#endif //defined(PROD)
#endif // defined(MAX) || defined(MIN)
}
#if defined(MEAN)
res /= BATCH;
#endif // defined(MEAN)
// Subtract the offsets in case of quantized SUM
#if defined(SUM) && defined(OFFSET) && defined(SCALE)
res -= (BATCH - 1) * OFFSET;
#endif // defined(OFFSET) && defined(OFFSET) && defined(SCALE)
// Re-quantize
#if defined(PROD) && defined(OFFSET) && defined(SCALE)
res = QUANTIZE(res_f, OFFSET, SCALE, DATA_TYPE_PROMOTED, 16);
#endif // defined(PROD) && defined(OFFSET) && defined(SCALE)
// Store result
vstore16(CONVERT_SAT(res, VEC_DATA_TYPE(DATA_TYPE, 16)), 0, (__global DATA_TYPE *)output.ptr);
}
#endif /* defined(BATCH) && defined(DEPTH) */
|
