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
|
/*
* Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved.
*
* SPDX-License-Identifier: Apache-2.0
*
* Licensed under the Apache License, Version 2.0 (the License); you may
* not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS, WITHOUT
* WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
/* ----------------------------------------------------------------------
* Project: CMSIS NN Library
* Title: arm_pool_q7_HWC.c
* Description: Pooling function implementations
*
* $Date: 17. January 2018
* $Revision: V.1.0.0
*
* Target Processor: Cortex-M cores
*
* -------------------------------------------------------------------- */
#include "arm_math.h"
#include "arm_nnfunctions.h"
#if defined (ARM_MATH_DSP)
/**
* @brief A few utility functions used by pooling functions
*
*
*/
static void buffer_scale_back_q15_to_q7(q15_t * buffer, q7_t * target, uint16_t length, uint16_t scale)
{
int i;
for (i = 0; i < length; i++)
{
target[i] = (q7_t) (buffer[i] / scale);
}
}
static void compare_and_replace_if_larger_q7(q7_t * base, // base data
const q7_t * target, // compare target
const uint16_t length // data length
)
{
q7_t *pIn = base;
const q7_t *pCom = target;
union arm_nnword in;
union arm_nnword com;
uint16_t cnt = length >> 2;
while (cnt > 0u)
{
in.word = *__SIMD32(pIn);
com.word = *__SIMD32(pCom)++;
// if version
if (com.bytes[0] > in.bytes[0])
in.bytes[0] = com.bytes[0];
if (com.bytes[1] > in.bytes[1])
in.bytes[1] = com.bytes[1];
if (com.bytes[2] > in.bytes[2])
in.bytes[2] = com.bytes[2];
if (com.bytes[3] > in.bytes[3])
in.bytes[3] = com.bytes[3];
*__SIMD32(pIn)++ = in.word;
cnt--;
}
cnt = length & 0x3;
while (cnt > 0u)
{
if (*pCom > *pIn)
{
*pIn = *pCom;
}
pIn++;
pCom++;
cnt--;
}
}
static void accumulate_q7_to_q15(q15_t * base, q7_t * target, const uint16_t length)
{
q15_t *pCnt = base;
q7_t *pV = target;
q31_t v1, v2, vo1, vo2;
uint16_t cnt = length >> 2;
q31_t in;
while (cnt > 0u)
{
q31_t value = *__SIMD32(pV)++;
v1 = __SXTB16(__ROR(value, 8));
v2 = __SXTB16(value);
#ifndef ARM_MATH_BIG_ENDIAN
vo2 = __PKHTB(v1, v2, 16);
vo1 = __PKHBT(v2, v1, 16);
#else
vo1 = __PKHTB(v1, v2, 16);
vo2 = __PKHBT(v2, v1, 16);
#endif
in = *__SIMD32(pCnt);
*__SIMD32(pCnt)++ = __QADD16(vo1, in);
in = *__SIMD32(pCnt);
*__SIMD32(pCnt)++ = __QADD16(vo2, in);
cnt--;
}
cnt = length & 0x3;
while (cnt > 0u)
{
*pCnt++ += *pV++;
cnt--;
}
}
#endif // ARM_MATH_DSP
/**
* @ingroup groupNN
*/
/**
* @addtogroup Pooling
* @{
*/
/**
* @brief Q7 max pooling function
* @param[in, out] Im_in pointer to input tensor
* @param[in] dim_im_in input tensor dimention
* @param[in] ch_im_in number of input tensor channels
* @param[in] dim_kernel filter kernel size
* @param[in] padding padding sizes
* @param[in] stride convolution stride
* @param[in] dim_im_out output tensor dimension
* @param[in,out] bufferA pointer to buffer space for input
* @param[in,out] Im_out pointer to output tensor
* @return none.
*
* @details
*
* <b>Buffer size:</b>
*
* bufferA size: 0
*
* The pooling function is implemented as split x-pooling then
* y-pooling.
*
* This pooling function is input-destructive. Input data is undefined
* after calling this function.
*
*/
void
arm_maxpool_q7_HWC(q7_t * Im_in,
const uint16_t dim_im_in,
const uint16_t ch_im_in,
const uint16_t dim_kernel,
const uint16_t padding,
const uint16_t stride, const uint16_t dim_im_out, q7_t * bufferA, q7_t * Im_out)
{
#if defined (ARM_MATH_DSP)
/* Run the following code for Cortex-M4 and Cortex-M7 */
int16_t i_x, i_y;
/* first does the pooling along x axis */
for (i_y = 0; i_y < dim_im_in; i_y++)
{
for (i_x = 0; i_x < dim_im_out; i_x++)
{
/* for each output pixel */
q7_t *target = Im_in + (i_y * dim_im_in + i_x) * ch_im_in;
q7_t *win_start;
q7_t *win_stop;
if (i_x * stride - padding < 0)
{
win_start = target;
} else
{
win_start = Im_in + (i_y * dim_im_in + i_x * stride - padding) * ch_im_in;
}
if (i_x * stride - padding + dim_kernel >= dim_im_in)
{
win_stop = Im_in + (i_y * dim_im_in + dim_im_in) * ch_im_in;
} else
{
win_stop = Im_in + (i_y * dim_im_in + i_x * stride - padding + dim_kernel) * ch_im_in;
}
/* first step is to copy over initial data */
/* arm_copy_q7(win_start, target, ch_im_in); */
memmove(target, win_start, ch_im_in);
/* start the max operation from the second part */
win_start += ch_im_in;
for (; win_start < win_stop; win_start += ch_im_in)
{
compare_and_replace_if_larger_q7(target, win_start, ch_im_in);
}
}
}
/* then does the pooling along y axis */
for (i_y = 0; i_y < dim_im_out; i_y++)
{
/* for each output row */
q7_t *target = Im_out + i_y * dim_im_out * ch_im_in;
q7_t *row_start;
q7_t *row_end;
/* setting the starting row */
if (i_y * stride - padding < 0)
{
row_start = Im_in;
} else
{
row_start = Im_in + (i_y * stride - padding) * dim_im_in * ch_im_in;
}
/* setting the stopping row */
if (i_y * stride - padding + dim_kernel >= dim_im_in)
{
row_end = Im_in + dim_im_in * dim_im_in * ch_im_in;
} else
{
row_end = Im_in + (i_y * stride - padding + dim_kernel) * dim_im_in * ch_im_in;
}
/* copy over the first row */
/* arm_copy_q7(row_start, target, dim_im_out * ch_im_in); */
memmove(target, row_start, dim_im_out * ch_im_in);
/* move over to next row */
row_start += ch_im_in * dim_im_in;
for (; row_start < row_end; row_start += dim_im_in * ch_im_in)
{
compare_and_replace_if_larger_q7(target, row_start, dim_im_out * ch_im_in);
}
}
#else
/* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */
int16_t i_ch_in, i_x, i_y;
int16_t k_x, k_y;
for (i_ch_in = 0; i_ch_in < ch_im_in; i_ch_in++)
{
for (i_y = 0; i_y < dim_im_out; i_y++)
{
for (i_x = 0; i_x < dim_im_out; i_x++)
{
int max = -129;
for (k_y = i_y * stride - padding; k_y < i_y * stride - padding + dim_kernel; k_y++)
{
for (k_x = i_x * stride - padding; k_x < i_x * stride - padding + dim_kernel; k_x++)
{
if (k_y >= 0 && k_x >= 0 && k_y < dim_im_in && k_x < dim_im_in)
{
if (Im_in[i_ch_in + ch_im_in * (k_x + k_y * dim_im_in)] > max)
{
max = Im_in[i_ch_in + ch_im_in * (k_x + k_y * dim_im_in)];
}
}
}
}
Im_out[i_ch_in + ch_im_in * (i_x + i_y * dim_im_out)] = max;
}
}
}
#endif /* ARM_MATH_DSP */
}
/**
* @brief Q7 average pooling function
* @param[in,out] Im_in pointer to input tensor
* @param[in] dim_im_in input tensor dimention
* @param[in] ch_im_in number of input tensor channels
* @param[in] dim_kernel filter kernel size
* @param[in] padding padding sizes
* @param[in] stride convolution stride
* @param[in] dim_im_out output tensor dimension
* @param[in,out] bufferA pointer to buffer space for input
* @param[in,out] Im_out pointer to output tensor
* @return none.
*
* @details
*
* <b>Buffer size:</b>
*
* bufferA size: 2*dim_im_out*ch_im_in
*
* The pooling function is implemented as split x-pooling then
* y-pooling.
*
* This pooling function is input-destructive. Input data is undefined
* after calling this function.
*
*/
void
arm_avepool_q7_HWC(q7_t * Im_in,
const uint16_t dim_im_in,
const uint16_t ch_im_in,
const uint16_t dim_kernel,
const uint16_t padding,
const uint16_t stride, const uint16_t dim_im_out, q7_t * bufferA, q7_t * Im_out)
{
#if defined (ARM_MATH_DSP)
/* Run the following code for Cortex-M4 and Cortex-M7 */
q15_t *buffer = (q15_t *) bufferA;
int16_t i_x, i_y;
int16_t count = 0;
/* first does the pooling along x axis */
for (i_y = 0; i_y < dim_im_in; i_y++)
{
for (i_x = 0; i_x < dim_im_out; i_x++)
{
/* for each output pixel */
q7_t *target = Im_in + (i_y * dim_im_in + i_x) * ch_im_in;
q7_t *win_start;
q7_t *win_stop;
if (i_x * stride - padding < 0)
{
win_start = target;
} else
{
win_start = Im_in + (i_y * dim_im_in + i_x * stride - padding) * ch_im_in;
}
if (i_x * stride - padding + dim_kernel >= dim_im_in)
{
win_stop = Im_in + (i_y * dim_im_in + dim_im_in) * ch_im_in;
} else
{
win_stop = Im_in + (i_y * dim_im_in + i_x * stride - padding + dim_kernel) * ch_im_in;
}
/* first step is to copy over initial data */
arm_q7_to_q15_no_shift(win_start, buffer, ch_im_in);
count = 1;
/* start the max operation from the second part */
win_start += ch_im_in;
for (; win_start < win_stop; win_start += ch_im_in)
{
accumulate_q7_to_q15(buffer, win_start, ch_im_in);
count++;
}
buffer_scale_back_q15_to_q7(buffer, target, ch_im_in, count);
}
}
/* then does the pooling along y axis */
for (i_y = 0; i_y < dim_im_out; i_y++)
{
/* for each output row */
q7_t *target = Im_out + i_y * dim_im_out * ch_im_in;
q7_t *row_start;
q7_t *row_end;
/* setting the starting row */
if (i_y * stride - padding < 0)
{
row_start = Im_in;
} else
{
row_start = Im_in + (i_y * stride - padding) * dim_im_in * ch_im_in;
}
/* setting the stopping row */
if (i_y * stride - padding + dim_kernel >= dim_im_in)
{
row_end = Im_in + dim_im_in * dim_im_in * ch_im_in;
} else
{
row_end = Im_in + (i_y * stride - padding + dim_kernel) * dim_im_in * ch_im_in;
}
/* copy over the first row */
arm_q7_to_q15_no_shift(row_start, buffer, dim_im_out * ch_im_in);
count = 1;
/* move over to next row */
row_start += ch_im_in * dim_im_in;
for (; row_start < row_end; row_start += dim_im_in * ch_im_in)
{
accumulate_q7_to_q15(buffer, row_start, dim_im_out * ch_im_in);
count++;
}
buffer_scale_back_q15_to_q7(buffer, target, dim_im_out * ch_im_in, count);
}
#else
/* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */
int16_t i_ch_in, i_x, i_y;
int16_t k_x, k_y;
for (i_ch_in = 0; i_ch_in < ch_im_in; i_ch_in++)
{
for (i_y = 0; i_y < dim_im_out; i_y++)
{
for (i_x = 0; i_x < dim_im_out; i_x++)
{
int sum = 0;
int count = 0;
for (k_y = i_y * stride - padding; k_y < i_y * stride - padding + dim_kernel; k_y++)
{
for (k_x = i_x * stride - padding; k_x < i_x * stride - padding + dim_kernel; k_x++)
{
if (k_y >= 0 && k_x >= 0 && k_y < dim_im_in && k_x < dim_im_in)
{
sum += Im_in[i_ch_in + ch_im_in * (k_x + k_y * dim_im_in)];
count++;
}
}
}
Im_out[i_ch_in + ch_im_in * (i_x + i_y * dim_im_out)] = sum / count;
}
}
}
#endif /* ARM_MATH_DSP */
}
/**
* @} end of Pooling group
*/
|
