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
|
/*
* Copyright (c) 2017 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to
* deal in the Software without restriction, including without limitation the
* rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
* sell copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
#ifndef __ARM_COMPUTE_NEDIRECTCONVOLUTIONDETAIL_H__
#define __ARM_COMPUTE_NEDIRECTCONVOLUTIONDETAIL_H__
#include "arm_compute/core/AccessWindowStatic.h"
#include "arm_compute/core/NEON/NEFixedPoint.h"
#include <arm_neon.h>
namespace arm_compute
{
namespace detail
{
/** Loads a 3x3 matrix as a row (float).
*
* @param[in] ptr Pointer to a float 3x3 matrix.
*
* @return The loaded matrix.
*/
inline float32x4x3_t load_matrix_row(const float *ptr)
{
const float32x4x3_t r =
{
{
vld1q_dup_f32(ptr),
vld1q_dup_f32(1 + ptr),
vld1q_dup_f32(2 + ptr)
}
};
return r;
}
/** Loads a 3x3 matrix as a row (qint8_t).
*
* @param[in] ptr Pointer to a qint8 3x3 matrix.
*
* @return The loaded matrix.
*/
inline qint8x8x3_t load_matrix_row(const qint8_t *ptr)
{
/* ptr is a pointer to a row in a 3x3 matrix, the function returns 3 vectors holding exactly the same value in all lanes:
r.val[0] contains the first element, r.val[1] the second element and r.val[2] the third element (in all lanes) */
const qint8x8x3_t r =
{
{
vld1_dup_qs8(ptr),
vld1_dup_qs8(1 + ptr),
vld1_dup_qs8(2 + ptr)
}
};
return r;
}
/** Perform a convolve3x3 on float32.
*
* @param[in] in_top Pointer to the first row of the input.
* @param[in] in_mid Pointer to the second row of the input.
* @param[in] in_low Pointer to the third row of the input.
* @param[in] m0 First row of the filter.
* @param[in] m1 Second row of the filter.
* @param[in] m2 Third row of the filter.
* @param[in] fixed_point_position (Optional) Fixed point position.
*
*/
template <unsigned int stridex>
float32x4x2_t convolve_3x3(const float *in_top, const float *in_mid, const float *in_low, const float32x4x3_t &m0, const float32x4x3_t &m1, const float32x4x3_t &m2, int fixed_point_position);
template <>
inline float32x4x2_t convolve_3x3<1>(const float *in_top, const float *in_mid, const float *in_low, const float32x4x3_t &m0, const float32x4x3_t &m1, const float32x4x3_t &m2, int fixed_point_position)
{
ARM_COMPUTE_UNUSED(fixed_point_position);
const float32x4x3_t vtop =
{
{
vld1q_f32(in_top),
vld1q_f32(in_top + 4),
vld1q_f32(in_top + 8)
}
};
const float32x4x3_t vmid =
{
{
vld1q_f32(in_mid),
vld1q_f32(in_mid + 4),
vld1q_f32(in_mid + 8)
}
};
const float32x4x3_t vlow =
{
{
vld1q_f32(in_low),
vld1q_f32(in_low + 4),
vld1q_f32(in_low + 8)
}
};
float32x4x2_t out =
{
{
vmulq_f32(vtop.val[0], m0.val[0]),
vmulq_f32(vtop.val[1], m0.val[0])
}
};
out.val[0] = vmlaq_f32(out.val[0], vextq_f32(vtop.val[0], vtop.val[1], 1), m0.val[1]);
out.val[0] = vmlaq_f32(out.val[0], vextq_f32(vtop.val[0], vtop.val[1], 2), m0.val[2]);
out.val[0] = vmlaq_f32(out.val[0], vmid.val[0], m1.val[0]);
out.val[0] = vmlaq_f32(out.val[0], vextq_f32(vmid.val[0], vmid.val[1], 1), m1.val[1]);
out.val[0] = vmlaq_f32(out.val[0], vextq_f32(vmid.val[0], vmid.val[1], 2), m1.val[2]);
out.val[0] = vmlaq_f32(out.val[0], vlow.val[0], m2.val[0]);
out.val[0] = vmlaq_f32(out.val[0], vextq_f32(vlow.val[0], vlow.val[1], 1), m2.val[1]);
out.val[0] = vmlaq_f32(out.val[0], vextq_f32(vlow.val[0], vlow.val[1], 2), m2.val[2]);
out.val[1] = vmlaq_f32(out.val[1], vextq_f32(vtop.val[1], vtop.val[2], 1), m0.val[1]);
out.val[1] = vmlaq_f32(out.val[1], vextq_f32(vtop.val[1], vtop.val[2], 2), m0.val[2]);
out.val[1] = vmlaq_f32(out.val[1], vmid.val[1], m1.val[0]);
out.val[1] = vmlaq_f32(out.val[1], vextq_f32(vmid.val[1], vmid.val[2], 1), m1.val[1]);
out.val[1] = vmlaq_f32(out.val[1], vextq_f32(vmid.val[1], vmid.val[2], 2), m1.val[2]);
out.val[1] = vmlaq_f32(out.val[1], vlow.val[1], m2.val[0]);
out.val[1] = vmlaq_f32(out.val[1], vextq_f32(vlow.val[1], vlow.val[2], 1), m2.val[1]);
out.val[1] = vmlaq_f32(out.val[1], vextq_f32(vlow.val[1], vlow.val[2], 2), m2.val[2]);
return out;
}
template <>
inline float32x4x2_t convolve_3x3<2>(const float *in_top, const float *in_mid, const float *in_low, const float32x4x3_t &m0, const float32x4x3_t &m1, const float32x4x3_t &m2, int fixed_point_position)
{
float32x4x2_t out = convolve_3x3<1>(in_top, in_mid, in_low, m0, m1, m2, fixed_point_position);
out.val[0] = vsetq_lane_f32(vgetq_lane_f32(out.val[0], 2), out.val[0], 1);
out.val[0] = vsetq_lane_f32(vgetq_lane_f32(out.val[1], 0), out.val[0], 2);
out.val[0] = vsetq_lane_f32(vgetq_lane_f32(out.val[1], 2), out.val[0], 3);
return out;
}
template <>
inline float32x4x2_t convolve_3x3<3>(const float *in_top, const float *in_mid, const float *in_low, const float32x4x3_t &m0, const float32x4x3_t &m1, const float32x4x3_t &m2, int fixed_point_position)
{
float32x4x2_t out = convolve_3x3<1>(in_top, in_mid, in_low, m0, m1, m2, fixed_point_position);
out.val[0] = vsetq_lane_f32(vgetq_lane_f32(out.val[0], 3), out.val[0], 1);
return out;
}
/** Perform a convolve3x3 on qint16.
*
* @param[in] in_top Pointer to the first row of the input.
* @param[in] in_mid Pointer to the second row of the input.
* @param[in] in_low Pointer to the third row of the input.
* @param[in] m0 First row of the filter.
* @param[in] m1 Second row of the filter.
* @param[in] m2 Third row of the filter.
* @param[in] fixed_point_position (Optional) Fixed point position.
*
*/
template <unsigned int stridex>
qint16x8x2_t convolve_3x3(const qint8_t *in_top, const qint8_t *in_mid, const qint8_t *in_low, const qint8x8x3_t &m0, const qint8x8x3_t &m1, const qint8x8x3_t &m2, int fixed_point_position);
template <>
inline qint16x8x2_t convolve_3x3<1>(const qint8_t *in_top, const qint8_t *in_mid, const qint8_t *in_low, const qint8x8x3_t &m0, const qint8x8x3_t &m1, const qint8x8x3_t &m2, int fixed_point_position)
{
ARM_COMPUTE_UNUSED(fixed_point_position);
const qint8x8x3_t vtop =
{
{
vld1_qs8(in_top),
vld1_qs8(in_top + 8),
vld1_qs8(in_top + 16)
}
};
const qint8x8x3_t vmid =
{
{
vld1_qs8(in_mid),
vld1_qs8(in_mid + 8),
vld1_qs8(in_mid + 16)
}
};
const qint8x8x3_t vlow =
{
{
vld1_qs8(in_low),
vld1_qs8(in_low + 8),
vld1_qs8(in_low + 16)
}
};
qint16x8x2_t out =
{
{
vmull_qs8(vtop.val[0], m0.val[0], fixed_point_position),
vmull_qs8(vtop.val[1], m0.val[0], fixed_point_position)
}
};
out.val[0] = vqmlal_qs8(out.val[0], vext_s8(vtop.val[0], vtop.val[1], 1), m0.val[1], fixed_point_position);
out.val[0] = vqmlal_qs8(out.val[0], vext_s8(vtop.val[0], vtop.val[1], 2), m0.val[2], fixed_point_position);
out.val[0] = vqmlal_qs8(out.val[0], vmid.val[0], m1.val[0], fixed_point_position);
out.val[0] = vqmlal_qs8(out.val[0], vext_s8(vmid.val[0], vmid.val[1], 1), m1.val[1], fixed_point_position);
out.val[0] = vqmlal_qs8(out.val[0], vext_s8(vmid.val[0], vmid.val[1], 2), m1.val[2], fixed_point_position);
out.val[0] = vqmlal_qs8(out.val[0], vlow.val[0], m2.val[0], fixed_point_position);
out.val[0] = vqmlal_qs8(out.val[0], vext_s8(vlow.val[0], vlow.val[1], 1), m2.val[1], fixed_point_position);
out.val[0] = vqmlal_qs8(out.val[0], vext_s8(vlow.val[0], vlow.val[1], 2), m2.val[2], fixed_point_position);
out.val[1] = vqmlal_qs8(out.val[1], vext_s8(vtop.val[1], vtop.val[2], 1), m0.val[1], fixed_point_position);
out.val[1] = vqmlal_qs8(out.val[1], vext_s8(vtop.val[1], vtop.val[2], 2), m0.val[2], fixed_point_position);
out.val[1] = vqmlal_qs8(out.val[1], vmid.val[1], m1.val[0], fixed_point_position);
out.val[1] = vqmlal_qs8(out.val[1], vext_s8(vmid.val[1], vmid.val[2], 1), m1.val[1], fixed_point_position);
out.val[1] = vqmlal_qs8(out.val[1], vext_s8(vmid.val[1], vmid.val[2], 2), m1.val[2], fixed_point_position);
out.val[1] = vqmlal_qs8(out.val[1], vlow.val[1], m2.val[0], fixed_point_position);
out.val[1] = vqmlal_qs8(out.val[1], vext_s8(vlow.val[1], vlow.val[2], 1), m2.val[1], fixed_point_position);
out.val[1] = vqmlal_qs8(out.val[1], vext_s8(vlow.val[1], vlow.val[2], 2), m2.val[2], fixed_point_position);
return out;
}
template <>
inline qint16x8x2_t convolve_3x3<2>(const qint8_t *in_top, const qint8_t *in_mid, const qint8_t *in_low, const qint8x8x3_t &m0, const qint8x8x3_t &m1, const qint8x8x3_t &m2, int fixed_point_position)
{
qint16x8x2_t out = convolve_3x3<1>(in_top, in_mid, in_low, m0, m1, m2, fixed_point_position);
out.val[0] = vsetq_lane_s16(vgetq_lane_s16(out.val[0], 2), out.val[0], 1);
out.val[0] = vsetq_lane_s16(vgetq_lane_s16(out.val[0], 4), out.val[0], 2);
out.val[0] = vsetq_lane_s16(vgetq_lane_s16(out.val[0], 6), out.val[0], 3);
out.val[0] = vsetq_lane_s16(vgetq_lane_s16(out.val[1], 0), out.val[0], 4);
out.val[0] = vsetq_lane_s16(vgetq_lane_s16(out.val[1], 2), out.val[0], 5);
out.val[0] = vsetq_lane_s16(vgetq_lane_s16(out.val[1], 4), out.val[0], 6);
out.val[0] = vsetq_lane_s16(vgetq_lane_s16(out.val[1], 6), out.val[0], 7);
return out;
}
template <>
inline qint16x8x2_t convolve_3x3<3>(const qint8_t *in_top, const qint8_t *in_mid, const qint8_t *in_low, const qint8x8x3_t &m0, const qint8x8x3_t &m1, const qint8x8x3_t &m2, int fixed_point_position)
{
qint16x8x2_t out = convolve_3x3<1>(in_top, in_mid, in_low, m0, m1, m2, fixed_point_position);
out.val[0] = vsetq_lane_s16(vgetq_lane_s16(out.val[0], 3), out.val[0], 1);
out.val[0] = vsetq_lane_s16(vgetq_lane_s16(out.val[0], 6), out.val[0], 2);
out.val[0] = vsetq_lane_s16(vgetq_lane_s16(out.val[1], 1), out.val[0], 3);
return out;
}
/** Stores a float32x4x2_t array into a memory location.
*
* @param[in] buffer Pointer to the memory location where the values will be stored.
* @param[in] values Values that will be stored.
*
*/
template <unsigned int stridex>
void store_results(float *buffer, const float32x4x2_t &values);
template <>
inline void store_results<1>(float *buffer, const float32x4x2_t &values)
{
vst1q_f32(buffer, values.val[0]);
vst1q_f32(buffer + 4, values.val[1]);
}
template <>
inline void store_results<2>(float *buffer, const float32x4x2_t &values)
{
vst1q_f32(buffer, values.val[0]);
}
template <>
inline void store_results<3>(float *buffer, const float32x4x2_t &values)
{
vst1_f32(buffer, vget_low_f32(values.val[0]));
}
/** Stores a qint16_t array into a memory location.
*
* @param[in] buffer Pointer to the memory location where the values will be stored.
* @param[in] values Values that will be stored.
*
*/
template <unsigned int stridex>
void store_results(qint16_t *buffer, const qint16x8x2_t &values);
template <>
inline void store_results<1>(qint16_t *buffer, const qint16x8x2_t &values)
{
vst1q_qs16(buffer, values.val[0]);
vst1q_qs16(buffer + 8, values.val[1]);
}
template <>
inline void store_results<2>(qint16_t *buffer, const qint16x8x2_t &values)
{
vst1q_qs16(buffer, values.val[0]);
}
template <>
inline void store_results<3>(qint16_t *buffer, const qint16x8x2_t &values)
{
vst1_qs16(buffer, vget_low_s16(values.val[0]));
}
#ifdef ARM_COMPUTE_AARCH64_V8_2
/** Loads a 3x3 matrix as a row (float16_t).
*
* @param[in] ptr Pointer to a float 3x3 matrix.
*
* @return The loaded matrix.
*/
inline float16x8x3_t load_matrix_row(const float16_t *ptr)
{
/* ptr is a pointer to a row in a 3x3 matrix, the function returns 3 vectors holding exactly the same value in all lanes:
r.val[0] contains the first element, r.val[1] the second element and r.val[2] the third element (in all lanes) */
const float16x8x3_t r =
{
{
vld1q_dup_f16(ptr),
vld1q_dup_f16(1 + ptr),
vld1q_dup_f16(2 + ptr)
}
};
return r;
}
/** Perform a convolve3x3 on float16.
*
* @param[in] in_top Pointer to the first row of the input.
* @param[in] in_mid Pointer to the second row of the input.
* @param[in] in_low Pointer to the third row of the input.
* @param[in] m0 First row of the filter.
* @param[in] m1 Second row of the filter.
* @param[in] m2 Third row of the filter.
* @param[in] fixed_point_position (Optional) Fixed point position.
*
*/
template <unsigned int stridex>
float16x8x2_t convolve_3x3(const float16_t *in_top, const float16_t *in_mid, const float16_t *in_low, const float16x8x3_t &m0, const float16x8x3_t &m1, const float16x8x3_t &m2,
int fixed_point_position);
template <>
inline float16x8x2_t convolve_3x3<1>(const float16_t *in_top, const float16_t *in_mid, const float16_t *in_low, const float16x8x3_t &m0, const float16x8x3_t &m1, const float16x8x3_t &m2,
int fixed_point_position)
{
ARM_COMPUTE_UNUSED(fixed_point_position);
const float16x8x3_t vtop =
{
{
vld1q_f16(in_top),
vld1q_f16(in_top + 8),
vld1q_f16(in_top + 16)
}
};
const float16x8x3_t vmid =
{
{
vld1q_f16(in_mid),
vld1q_f16(in_mid + 8),
vld1q_f16(in_mid + 16)
}
};
const float16x8x3_t vlow =
{
{
vld1q_f16(in_low),
vld1q_f16(in_low + 8),
vld1q_f16(in_low + 16)
}
};
float16x8x2_t out =
{
{
vmulq_f16(vtop.val[0], m0.val[0]),
vmulq_f16(vtop.val[1], m0.val[0])
}
};
out.val[0] = vaddq_f16(out.val[0], vmulq_f16(vextq_f16(vtop.val[0], vtop.val[1], 1), m0.val[1]));
out.val[0] = vaddq_f16(out.val[0], vmulq_f16(vextq_f16(vtop.val[0], vtop.val[1], 2), m0.val[2]));
out.val[0] = vaddq_f16(out.val[0], vmulq_f16(vmid.val[0], m1.val[0]));
out.val[0] = vaddq_f16(out.val[0], vmulq_f16(vextq_f16(vmid.val[0], vmid.val[1], 1), m1.val[1]));
out.val[0] = vaddq_f16(out.val[0], vmulq_f16(vextq_f16(vmid.val[0], vmid.val[1], 2), m1.val[2]));
out.val[0] = vaddq_f16(out.val[0], vmulq_f16(vlow.val[0], m2.val[0]));
out.val[0] = vaddq_f16(out.val[0], vmulq_f16(vextq_f16(vlow.val[0], vlow.val[1], 1), m2.val[1]));
out.val[0] = vaddq_f16(out.val[0], vmulq_f16(vextq_f16(vlow.val[0], vlow.val[1], 2), m2.val[2]));
out.val[1] = vaddq_f16(out.val[1], vmulq_f16(vextq_f16(vtop.val[1], vtop.val[2], 1), m0.val[1]));
out.val[1] = vaddq_f16(out.val[1], vmulq_f16(vextq_f16(vtop.val[1], vtop.val[2], 2), m0.val[2]));
out.val[1] = vaddq_f16(out.val[1], vmulq_f16(vmid.val[1], m1.val[0]));
out.val[1] = vaddq_f16(out.val[1], vmulq_f16(vextq_f16(vmid.val[1], vmid.val[2], 1), m1.val[1]));
out.val[1] = vaddq_f16(out.val[1], vmulq_f16(vextq_f16(vmid.val[1], vmid.val[2], 2), m1.val[2]));
out.val[1] = vaddq_f16(out.val[1], vmulq_f16(vlow.val[1], m2.val[0]));
out.val[1] = vaddq_f16(out.val[1], vmulq_f16(vextq_f16(vlow.val[1], vlow.val[2], 1), m2.val[1]));
out.val[1] = vaddq_f16(out.val[1], vmulq_f16(vextq_f16(vlow.val[1], vlow.val[2], 2), m2.val[2]));
return out;
}
template <>
inline float16x8x2_t convolve_3x3<2>(const float16_t *in_top, const float16_t *in_mid, const float16_t *in_low, const float16x8x3_t &m0, const float16x8x3_t &m1, const float16x8x3_t &m2,
int fixed_point_position)
{
float16x8x2_t out = convolve_3x3<1>(in_top, in_mid, in_low, m0, m1, m2, fixed_point_position);
out.val[0] = vsetq_lane_f16(vgetq_lane_f16(out.val[0], 2), out.val[0], 1);
out.val[0] = vsetq_lane_f16(vgetq_lane_f16(out.val[1], 0), out.val[0], 2);
out.val[0] = vsetq_lane_f16(vgetq_lane_f16(out.val[1], 2), out.val[0], 3);
return out;
}
template <>
inline float16x8x2_t convolve_3x3<3>(const float16_t *in_top, const float16_t *in_mid, const float16_t *in_low, const float16x8x3_t &m0, const float16x8x3_t &m1, const float16x8x3_t &m2,
int fixed_point_position)
{
float16x8x2_t out = convolve_3x3<1>(in_top, in_mid, in_low, m0, m1, m2, fixed_point_position);
out.val[0] = vsetq_lane_f16(vgetq_lane_f16(out.val[0], 3), out.val[0], 1);
return out;
}
/** Stores a float16x8x2_t array into a memory location.
*
* @param[in] buffer Pointer to the memory location where the values will be stored.
* @param[in] values Values that will be stored.
*
*/
template <unsigned int stridex>
void store_results(float16_t *buffer, const float16x8x2_t &values);
template <>
inline void store_results<1>(float16_t *buffer, const float16x8x2_t &values)
{
vst1q_f16(buffer, values.val[0]);
vst1q_f16(buffer + 8, values.val[1]);
}
template <>
inline void store_results<2>(float16_t *buffer, const float16x8x2_t &values)
{
vst1q_f16(buffer, values.val[0]);
}
template <>
inline void store_results<3>(float16_t *buffer, const float16x8x2_t &values)
{
vst1_f16(buffer, vget_low_f16(values.val[0]));
}
#endif /* ARM_COMPUTE_AARCH64_V8_2 */
/** Get the number of elements processed on 3x3 convolution.
*
* @param[in] num_elems_written_per_iteration Number of elements written per iteration on 3x3 convolution.
*
* @return The number of elements processed.
*/
template <unsigned int stridex>
int get_input_num_elems_processed(unsigned int num_elems_written_per_iteration);
template <>
inline int get_input_num_elems_processed<1>(unsigned int num_elems_written_per_iteration)
{
return num_elems_written_per_iteration;
}
template <>
inline int get_input_num_elems_processed<2>(unsigned int num_elems_written_per_iteration)
{
return num_elems_written_per_iteration << 1;
}
template <>
inline int get_input_num_elems_processed<3>(unsigned int num_elems_written_per_iteration)
{
return num_elems_written_per_iteration * 3;
}
}
} // namespace arm_compute
#endif /* __ARM_COMPUTE_NEDIRECTCONVOLUTIONDETAIL_H__ */
|
