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
|
/*
* 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_fully_connected_mat_q7_vec_q15_opt.c
* Description: Mixed Q15-Q7 opt fully-connected layer function
*
* $Date: 17. January 2018
* $Revision: V.1.0.0
*
* Target Processor: Cortex-M cores
*
* -------------------------------------------------------------------- */
#include "arm_math.h"
#include "arm_nnfunctions.h"
/**
* @ingroup groupNN
*/
/**
* @addtogroup FC
* @{
*/
/**
* @brief Mixed Q15-Q7 opt fully-connected layer function
* @param[in] pV pointer to input vector
* @param[in] pM pointer to matrix weights
* @param[in] dim_vec length of the vector
* @param[in] num_of_rows number of rows in weight matrix
* @param[in] bias_shift amount of left-shift for bias
* @param[in] out_shift amount of right-shift for output
* @param[in] bias pointer to bias
* @param[in,out] pOut pointer to output vector
* @param[in,out] vec_buffer pointer to buffer space for input
* @return The function returns <code>ARM_MATH_SUCCESS</code>
*
* @details
*
* <b>Buffer size:</b>
*
* vec_buffer size: 0
*
* Q7_Q15 version of the fully connected layer
*
* Weights are in q7_t and Activations are in q15_t
*
* Limitation: x4 version requires weight reordering to work
*
* Here we use only one pointer to read 4 rows in the weight
* matrix. So if the original q7_t matrix looks like this:
*
* | a11 | a12 | a13 | a14 | a15 | a16 | a17 |
*
* | a21 | a22 | a23 | a24 | a25 | a26 | a27 |
*
* | a31 | a32 | a33 | a34 | a35 | a36 | a37 |
*
* | a41 | a42 | a43 | a44 | a45 | a46 | a47 |
*
* | a51 | a52 | a53 | a54 | a55 | a56 | a57 |
*
* | a61 | a62 | a63 | a64 | a65 | a66 | a67 |
*
* We operates on multiple-of-4 rows, so the first four rows becomes
*
* | a11 | a21 | a12 | a22 | a31 | a41 | a32 | a42 |
*
* | a13 | a23 | a14 | a24 | a33 | a43 | a34 | a44 |
*
* | a15 | a25 | a16 | a26 | a35 | a45 | a36 | a46 |
*
* The column left over will be in-order.
* which is:
* | a17 | a27 | a37 | a47 |
*
* For the left-over rows, we do 1x1 computation, so the data remains
* as its original order.
*
* So the stored weight matrix looks like this:
*
* | a11 | a21 | a12 | a22 | a31 | a41 |
*
* | a32 | a42 | a13 | a23 | a14 | a24 |
*
* | a33 | a43 | a34 | a44 | a15 | a25 |
*
* | a16 | a26 | a35 | a45 | a36 | a46 |
*
* | a17 | a27 | a37 | a47 | a51 | a52 |
*
* | a53 | a54 | a55 | a56 | a57 | a61 |
*
* | a62 | a63 | a64 | a65 | a66 | a67 |
*
*/
arm_status
arm_fully_connected_mat_q7_vec_q15_opt(const q15_t * pV,
const q7_t * pM,
const uint16_t dim_vec,
const uint16_t num_of_rows,
const uint16_t bias_shift,
const uint16_t out_shift, const q7_t * bias, q15_t * pOut, q15_t * vec_buffer)
{
#if defined (ARM_MATH_DSP)
/* Run the following code for Cortex-M4 and Cortex-M7 */
const q7_t *pB = pM;
q15_t *pO = pOut;
const q7_t *pBias = bias;
const q15_t *pA = pV;
uint16_t rowCnt = num_of_rows >> 2;
while (rowCnt)
{
q31_t sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
q31_t sum2 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
q31_t sum3 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
q31_t sum4 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
uint16_t colCnt = dim_vec >> 1;
pA = pV;
#ifdef USE_INTRINSIC
#ifndef ARM_MATH_BIG_ENDIAN
while (colCnt)
{
q31_t inM11, inM12, inM13, inM14;
q31_t inV;
inV = *__SIMD32(pA)++;
inM11 = *__SIMD32(pB)++;
inM12 = __SXTB16(__ROR(inM11, 8));
inM11 = __SXTB16(inM11);
sum = __SMLAD(inM11, inV, sum);
sum2 = __SMLAD(inM12, inV, sum2);
inM13 = *__SIMD32(pB)++;
inM14 = __SXTB16(__ROR(inM13, 8));
inM13 = __SXTB16(inM13);
sum3 = __SMLAD(inM13, inV, sum3);
sum4 = __SMLAD(inM14, inV, sum4);
colCnt--;
}
#else
while (colCnt)
{
q31_t inM11, inM12, inM13, inM14;
q31_t inV;
inV = *__SIMD32(pA)++;
inM11 = *__SIMD32(pB)++;
inM12 = __SXTB16(__ROR(inM11, 8));
inM11 = __SXTB16(inM11);
sum = __SMLAD(inM12, inV, sum);
sum2 = __SMLAD(inM11, inV, sum2);
inM13 = *__SIMD32(pB)++;
inM14 = __SXTB16(__ROR(inM13, 8));
inM13 = __SXTB16(inM13);
sum3 = __SMLAD(inM14, inV, sum3);
sum4 = __SMLAD(inM13, inV, sum4);
colCnt--;
}
#endif /* ARM_MATH_BIG_ENDIAN */
#else
/*
* register needed:
* loop counter: colCnt
* accumulators: sum, sum2, sum3, sum4
* pointers: pB, pA
* weight data: inM11, inM12, inM13, inM14
* activation data: inV
*/
#ifndef ARM_MATH_BIG_ENDIAN
asm volatile ("COL_LOOP_%=:\n"
"ldr.w r4, [%[pA]], #4\n"
"ldr.w r1, [%[pB]], #8\n"
"mov.w r0, r1, ror #8\n"
"sxtb16 r0, r0\n"
"sxtb16 r1, r1\n"
"smlad %[sum], r4, r1, %[sum]\n"
"smlad %[sum2], r4, r0, %[sum2]\n"
"ldr.w r3, [%[pB], #-4]\n"
"mov.w r2, r3, ror #8\n"
"sxtb16 r2, r2\n"
"sxtb16 r3, r3\n"
"smlad %[sum3], r4, r3, %[sum3]\n"
"smlad %[sum4], r4, r2, %[sum4]\n"
"subs %[colCnt], #1\n"
"bne COL_LOOP_%=\n":[sum] "+r"(sum),
[sum2] "+r"(sum2),[sum3] "+r"(sum3),
[sum4] "+r"(sum4),[pB] "+r"(pB),[pA] "+r"(pA):[colCnt] "r"(colCnt):"r0", "r1", "r2", "r3", "r4");
#else
asm volatile ("COL_LOOP_%=:\n"
"ldr.w r4, [%[pA]], #4\n"
"ldr.w r1, [%[pB]], #8\n"
"mov.w r0, r1, ror #8\n"
"sxtb16 r0, r0\n"
"sxtb16 r1, r1\n"
"smlad %[sum], r4, r0, %[sum]\n"
"smlad %[sum2], r4, r1, %[sum2]\n"
"ldr.w r3, [%[pB], #-4]\n"
"mov.w r2, r3, ror #8\n"
"sxtb16 r2, r2\n"
"sxtb16 r3, r3\n"
"smlad %[sum3], r4, r2, %[sum3]\n"
"smlad %[sum4], r4, r3, %[sum4]\n"
"subs %[colCnt], #1\n"
"bne COL_LOOP_%=\n":[sum] "+r"(sum),
[sum2] "+r"(sum2),[sum3] "+r"(sum3),
[sum4] "+r"(sum4),[pB] "+r"(pB),[pA] "+r"(pA):[colCnt] "r"(colCnt):"r0", "r1", "r2", "r3", "r4");
#endif /* ARM_MATH_BIG_ENDIAN */
#endif /* USE_INTRINSIC */
colCnt = dim_vec & 0x1;
while (colCnt)
{
q15_t inV = *pA++;
q7_t inM = *pB++;
q7_t inM2 = *pB++;
q7_t inM3 = *pB++;
q7_t inM4 = *pB++;
sum += inV * inM;
sum2 += inV * inM2;
sum3 += inV * inM3;
sum4 += inV * inM4;
colCnt--;
} /* while over colCnt */
*pO++ = (q15_t) (__SSAT((sum >> out_shift), 16));
*pO++ = (q15_t) (__SSAT((sum2 >> out_shift), 16));
*pO++ = (q15_t) (__SSAT((sum3 >> out_shift), 16));
*pO++ = (q15_t) (__SSAT((sum4 >> out_shift), 16));
/* adjust the pointers and counters */
rowCnt--;
}
/* left-over part of the rows */
rowCnt = num_of_rows & 0x3;
while (rowCnt)
{
q31_t sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
uint16_t colCnt = dim_vec >> 2;
pA = pV;
while (colCnt)
{
q31_t inV1, inV2, inM11, inM12;
pB = (q7_t *) read_and_pad((void *)pB, &inM11, &inM12);
inV1 = *__SIMD32(pA)++;
sum = __SMLAD(inV1, inM11, sum);
inV2 = *__SIMD32(pA)++;
sum = __SMLAD(inV2, inM12, sum);
colCnt--;
}
/* left-over of the vector */
colCnt = dim_vec & 0x3;
while (colCnt)
{
q15_t inV = *pA++;
q7_t inM = *pB++;
sum += inV * inM;
colCnt--;
}
*pO++ = (q15_t) (__SSAT((sum >> out_shift), 16));
rowCnt--;
}
#else
/* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */
uint16_t rowCnt = num_of_rows >> 2;
const q7_t *pB = pM;
const q15_t *pA;
q15_t *pO = pOut;
const q7_t *pBias = bias;
while (rowCnt)
{
q31_t sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
q31_t sum2 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
q31_t sum3 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
q31_t sum4 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
uint16_t colCnt = dim_vec >> 1;
pA = pV;
while (colCnt)
{
q15_t inA1 = *pA++;
q15_t inA2 = *pA++;
q7_t inB1 = *pB++;
q7_t inB3 = *pB++;
q7_t inB2 = *pB++;
q7_t inB4 = *pB++;
sum += inA1 * inB1 + inA2 * inB2;
sum2 += inA1 * inB3 + inA2 * inB4;
inB1 = *pB++;
inB3 = *pB++;
inB2 = *pB++;
inB4 = *pB++;
sum3 += inA1 * inB1 + inA2 * inB2;
sum4 += inA1 * inB3 + inA2 * inB4;
colCnt--;
}
colCnt = dim_vec & 0x1;
while (colCnt)
{
q15_t inA = *pA++;
q7_t inB = *pB++;
sum += inA * inB;
inB = *pB++;
sum2 += inA * inB;
inB = *pB++;
sum3 += inA * inB;
inB = *pB++;
sum4 += inA * inB;
colCnt--;
}
*pO++ = (q15_t) __SSAT((sum >> out_shift), 16);
*pO++ = (q15_t) __SSAT((sum2 >> out_shift), 16);
*pO++ = (q15_t) __SSAT((sum3 >> out_shift), 16);
*pO++ = (q15_t) __SSAT((sum4 >> out_shift), 16);
rowCnt--;
}
rowCnt = num_of_rows & 0x3;
while (rowCnt)
{
int ip_out = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
int j;
pA = pV;
for (j = 0; j < dim_vec; j++)
{
q15_t inA = *pA++;
q7_t inB = *pB++;
ip_out += inA * inB;
}
*pO++ = (q15_t) __SSAT((ip_out >> out_shift), 16);
rowCnt--;
}
#endif /* ARM_MATH_DSP */
/* Return to ARM_MATH_SUCCESS */
return (ARM_MATH_SUCCESS);
}
/**
* @} end of FC group
*/
|
