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
|
/*
* Copyright (c) 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 "pooling.hpp"
#include <cstdint>
#include <cstring>
#include <cmath>
#if defined(__ARM_FEATURE_SVE) && defined(SVE2)
namespace arm_conv {
namespace pooling {
namespace {
struct RescaleParams
{
int32_t multiplier, shift;
};
constexpr RescaleParams rescale_params[8] = {
{0x40000000, -0}, // 1/2
{0x55555556, -1}, // 1/3
{0x40000000, -1}, // 1/4
{0x66666666, -2}, // 1/5
{0x55555556, -2}, // 1/6
{0x49249249, -2}, // 1/7
{0x40000000, -2}, // 1/8
{0x71c71c72, -3}, // 1/9
};
}
void sve_s8q_nhwc_avg_generic_depthfirst_impl(
const uint64_t window_cells,
const uint64_t n_valid_cells,
uint64_t n_channels,
const int8_t *const *const inptrs,
int8_t *outptr,
const Requantize32 &qp
)
{
if (n_valid_cells == 1 && window_cells == 1)
{
// In this case, simply copy from the input to the output
std::memcpy(outptr, *inptrs, n_channels);
return;
}
// Compute (or look up) the rescale values
int32_t shift_value = 0, rescale_value = 0;
if (2 <= window_cells && window_cells <= 9)
{
auto ¶ms = rescale_params[window_cells - 2];
rescale_value = params.multiplier;
shift_value = params.shift;
}
else
{
auto f_rescale_value = 1.0f / static_cast<float>(window_cells);
shift_value = 0;
while (f_rescale_value < 0.5f)
{
shift_value--;
f_rescale_value *= 2.0f;
}
rescale_value = static_cast<int32_t>(round(f_rescale_value * static_cast<float>(1ll << 31)));
if (static_cast<int64_t>(rescale_value) == (1ll << 31))
{
shift_value++;
rescale_value >>= 1;
}
}
// Combine together the rescale value for the requantization and the scaling
// factor for the average pool.
const int32_t shift = qp.per_layer_left_shift - qp.per_layer_right_shift + shift_value;
const int32_t left_shift = shift > 0 ? shift : 0;
const int32_t right_shift = shift <= 0 ? shift : 0;
int32_t combined_rescale_value = 0;
__asm__ __volatile__ (
"mov v16.s[0], %w[per_layer_mul]\n"
"mov v17.s[0], %w[rescale_value]\n"
"sqrdmulh s18, s16, s17\n"
"mov %w[combined_rescale_value], v18.s[0]\n"
: [combined_rescale_value] "=r" (combined_rescale_value)
: [per_layer_mul] "r" (qp.per_layer_mul), [rescale_value] "r" (rescale_value)
: "v16", "v17", "v18"
);
__asm__ __volatile__(
"ptrue p4.b\n"
"mov x26, #0x0\n"
"cntb x25\n"
"cntb x24, ALL, MUL #2\n"
"cntb x23, ALL, MUL #3\n"
"whilelt p3.b, x26, %x[n_channels]\n"
"whilelt p2.b, x25, %x[n_channels]\n"
"whilelt p1.b, x24, %x[n_channels]\n"
"whilelt p0.b, x23, %x[n_channels]\n"
"b.none 7f\n"
"1:" // 4-vectors of channels
"mov z15.s, #0x0\n"
"mov x19, %x[inptrs]\n"
"mov z14.s, #0x0\n"
"lsr x22, %x[n_valid_cells], #0x1\n"
"mov z13.s, #0x0\n"
"mov z12.s, #0x0\n"
"mov z11.s, #0x0\n"
"mov z10.s, #0x0\n"
"mov z9.s, #0x0\n"
"mov z8.s, #0x0\n"
"mov z7.s, #0x0\n"
"mov z6.s, #0x0\n"
"mov z5.s, #0x0\n"
"mov z4.s, #0x0\n"
"mov z3.s, #0x0\n"
"mov z2.s, #0x0\n"
"mov z1.s, #0x0\n"
"mov z0.s, #0x0\n"
"cbz x22, 4f\n"
"ldp x21, x20, [x19, #0x0]\n"
"ld1b { z31.b }, p3/Z, [x21, x26]\n"
"add x19, x19, #0x10\n"
"ld1b { z30.b }, p3/Z, [x20, x26]\n"
"subs x22, x22, #0x1\n"
"ld1b { z29.b }, p2/Z, [x21, x25]\n"
"ld1b { z28.b }, p2/Z, [x20, x25]\n"
"ld1b { z27.b }, p1/Z, [x21, x24]\n"
"ld1b { z26.b }, p1/Z, [x20, x24]\n"
"ld1b { z25.b }, p0/Z, [x21, x23]\n"
"ld1b { z24.b }, p0/Z, [x20, x23]\n"
"beq 3f\n"
"2:" // 4-vectors of channels: 2 inputs loop
".inst 0x455e03f7 // saddlb z23.h, z31.b, z30.b\n"
"ldp x21, x20, [x19, #0x0]\n"
"add x19, x19, #0x10\n"
".inst 0x455e07f6 // saddlt z22.h, z31.b, z30.b\n"
"ld1b { z31.b }, p3/Z, [x21, x26]\n"
".inst 0x455c03b5 // saddlb z21.h, z29.b, z28.b\n"
"subs x22, x22, #0x1\n"
".inst 0x455c07b4 // saddlt z20.h, z29.b, z28.b\n"
"ld1b { z30.b }, p3/Z, [x20, x26]\n"
".inst 0x455a0373 // saddlb z19.h, z27.b, z26.b\n"
"ld1b { z29.b }, p2/Z, [x21, x25]\n"
".inst 0x455a0772 // saddlt z18.h, z27.b, z26.b\n"
"ld1b { z28.b }, p2/Z, [x20, x25]\n"
".inst 0x45580331 // saddlb z17.h, z25.b, z24.b\n"
"ld1b { z27.b }, p1/Z, [x21, x24]\n"
".inst 0x45580730 // saddlt z16.h, z25.b, z24.b\n"
"ld1b { z26.b }, p1/Z, [x20, x24]\n"
".inst 0x459741ef // saddwb z15.s, z15.s, z23.h\n"
"ld1b { z25.b }, p0/Z, [x21, x23]\n"
".inst 0x459745ce // saddwt z14.s, z14.s, z23.h\n"
"ld1b { z24.b }, p0/Z, [x20, x23]\n"
".inst 0x459641ad // saddwb z13.s, z13.s, z22.h\n"
".inst 0x4596458c // saddwt z12.s, z12.s, z22.h\n"
".inst 0x4595416b // saddwb z11.s, z11.s, z21.h\n"
".inst 0x4595454a // saddwt z10.s, z10.s, z21.h\n"
".inst 0x45944129 // saddwb z9.s, z9.s, z20.h\n"
".inst 0x45944508 // saddwt z8.s, z8.s, z20.h\n"
".inst 0x459340e7 // saddwb z7.s, z7.s, z19.h\n"
".inst 0x459344c6 // saddwt z6.s, z6.s, z19.h\n"
".inst 0x459240a5 // saddwb z5.s, z5.s, z18.h\n"
".inst 0x45924484 // saddwt z4.s, z4.s, z18.h\n"
".inst 0x45914063 // saddwb z3.s, z3.s, z17.h\n"
".inst 0x45914442 // saddwt z2.s, z2.s, z17.h\n"
".inst 0x45904021 // saddwb z1.s, z1.s, z16.h\n"
".inst 0x45904400 // saddwt z0.s, z0.s, z16.h\n"
"bgt 2b\n"
"3:" // 4-vectors of channels: 2 inputs tail
".inst 0x455e03f7 // saddlb z23.h, z31.b, z30.b\n"
".inst 0x455e07f6 // saddlt z22.h, z31.b, z30.b\n"
".inst 0x455c03b5 // saddlb z21.h, z29.b, z28.b\n"
".inst 0x455c07b4 // saddlt z20.h, z29.b, z28.b\n"
".inst 0x455a0373 // saddlb z19.h, z27.b, z26.b\n"
".inst 0x455a0772 // saddlt z18.h, z27.b, z26.b\n"
".inst 0x45580331 // saddlb z17.h, z25.b, z24.b\n"
".inst 0x45580730 // saddlt z16.h, z25.b, z24.b\n"
".inst 0x459741ef // saddwb z15.s, z15.s, z23.h\n"
".inst 0x459745ce // saddwt z14.s, z14.s, z23.h\n"
".inst 0x459641ad // saddwb z13.s, z13.s, z22.h\n"
".inst 0x4596458c // saddwt z12.s, z12.s, z22.h\n"
".inst 0x4595416b // saddwb z11.s, z11.s, z21.h\n"
".inst 0x4595454a // saddwt z10.s, z10.s, z21.h\n"
".inst 0x45944129 // saddwb z9.s, z9.s, z20.h\n"
".inst 0x45944508 // saddwt z8.s, z8.s, z20.h\n"
".inst 0x459340e7 // saddwb z7.s, z7.s, z19.h\n"
".inst 0x459344c6 // saddwt z6.s, z6.s, z19.h\n"
".inst 0x459240a5 // saddwb z5.s, z5.s, z18.h\n"
".inst 0x45924484 // saddwt z4.s, z4.s, z18.h\n"
".inst 0x45914063 // saddwb z3.s, z3.s, z17.h\n"
".inst 0x45914442 // saddwt z2.s, z2.s, z17.h\n"
".inst 0x45904021 // saddwb z1.s, z1.s, z16.h\n"
".inst 0x45904400 // saddwt z0.s, z0.s, z16.h\n"
"4:" // 4-vectors of channels: After loop
"ands x20, %x[n_valid_cells], #0x1\n"
"beq 6f\n"
"5:" // 4-vectors of channels: Single input loop
"ldr x21, [x19], #0x8\n"
"subs x20, x20, #0x1\n"
"ld1b { z31.b }, p3/Z, [x21, x26]\n"
".inst 0x4508a3f1 // sshllb z17.h, z31.b, #0x0\n"
"ld1b { z29.b }, p2/Z, [x21, x25]\n"
".inst 0x4508a7f0 // sshllt z16.h, z31.b, #0x0\n"
"ld1b { z27.b }, p1/Z, [x21, x24]\n"
".inst 0x459141ef // saddwb z15.s, z15.s, z17.h\n"
"ld1b { z25.b }, p0/Z, [x21, x23]\n"
".inst 0x459145ce // saddwt z14.s, z14.s, z17.h\n"
".inst 0x459041ad // saddwb z13.s, z13.s, z16.h\n"
".inst 0x4590458c // saddwt z12.s, z12.s, z16.h\n"
".inst 0x4508a3b0 // sshllb z16.h, z29.b, #0x0\n"
".inst 0x4590416b // saddwb z11.s, z11.s, z16.h\n"
".inst 0x4590454a // saddwt z10.s, z10.s, z16.h\n"
".inst 0x4508a7b0 // sshllt z16.h, z29.b, #0x0\n"
".inst 0x45904129 // saddwb z9.s, z9.s, z16.h\n"
".inst 0x45904508 // saddwt z8.s, z8.s, z16.h\n"
".inst 0x4508a370 // sshllb z16.h, z27.b, #0x0\n"
".inst 0x459040e7 // saddwb z7.s, z7.s, z16.h\n"
".inst 0x459044c6 // saddwt z6.s, z6.s, z16.h\n"
".inst 0x4508a770 // sshllt z16.h, z27.b, #0x0\n"
".inst 0x459040a5 // saddwb z5.s, z5.s, z16.h\n"
".inst 0x45904484 // saddwt z4.s, z4.s, z16.h\n"
".inst 0x4508a330 // sshllb z16.h, z25.b, #0x0\n"
".inst 0x45904063 // saddwb z3.s, z3.s, z16.h\n"
".inst 0x45904442 // saddwt z2.s, z2.s, z16.h\n"
".inst 0x4508a730 // sshllt z16.h, z25.b, #0x0\n"
".inst 0x45904021 // saddwb z1.s, z1.s, z16.h\n"
".inst 0x45904400 // saddwt z0.s, z0.s, z16.h\n"
"bgt 5b\n"
"6:" // 4-vectors of channels: Single input loop: End
"mov z20.s, #0x7f\n"
"ld1rw { z18.s }, p4/Z, [%x[combined_rescale_value]]\n"
"ld1rw { z17.s }, p4/Z, [%x[left_shift]]\n"
"not z19.s, p4/M, z20.s\n"
"ld1rw { z16.s }, p4/Z, [%x[right_shift]]\n"
".inst 0x4482922f // srshl z15.s, p4/M, z15.s, z17.s\n"
".inst 0x4482922e // srshl z14.s, p4/M, z14.s, z17.s\n"
".inst 0x4482922d // srshl z13.s, p4/M, z13.s, z17.s\n"
".inst 0x4482922c // srshl z12.s, p4/M, z12.s, z17.s\n"
".inst 0x4482922b // srshl z11.s, p4/M, z11.s, z17.s\n"
".inst 0x04b275ef // sqrdmulh z15.s, z15.s, z18.s\n"
".inst 0x04b275ce // sqrdmulh z14.s, z14.s, z18.s\n"
".inst 0x04b275ad // sqrdmulh z13.s, z13.s, z18.s\n"
".inst 0x04b2758c // sqrdmulh z12.s, z12.s, z18.s\n"
".inst 0x04b2756b // sqrdmulh z11.s, z11.s, z18.s\n"
".inst 0x4482920f // srshl z15.s, p4/M, z15.s, z16.s\n"
".inst 0x4482920e // srshl z14.s, p4/M, z14.s, z16.s\n"
".inst 0x4482920d // srshl z13.s, p4/M, z13.s, z16.s\n"
".inst 0x4482920c // srshl z12.s, p4/M, z12.s, z16.s\n"
".inst 0x4482920b // srshl z11.s, p4/M, z11.s, z16.s\n"
".inst 0x4482922a // srshl z10.s, p4/M, z10.s, z17.s\n"
".inst 0x44829229 // srshl z9.s, p4/M, z9.s, z17.s\n"
".inst 0x44829228 // srshl z8.s, p4/M, z8.s, z17.s\n"
".inst 0x44829227 // srshl z7.s, p4/M, z7.s, z17.s\n"
".inst 0x04b2754a // sqrdmulh z10.s, z10.s, z18.s\n"
".inst 0x04b27529 // sqrdmulh z9.s, z9.s, z18.s\n"
".inst 0x04b27508 // sqrdmulh z8.s, z8.s, z18.s\n"
".inst 0x04b274e7 // sqrdmulh z7.s, z7.s, z18.s\n"
".inst 0x4482920a // srshl z10.s, p4/M, z10.s, z16.s\n"
".inst 0x44829209 // srshl z9.s, p4/M, z9.s, z16.s\n"
".inst 0x44829208 // srshl z8.s, p4/M, z8.s, z16.s\n"
".inst 0x44829207 // srshl z7.s, p4/M, z7.s, z16.s\n"
".inst 0x44829226 // srshl z6.s, p4/M, z6.s, z17.s\n"
".inst 0x44829225 // srshl z5.s, p4/M, z5.s, z17.s\n"
".inst 0x44829224 // srshl z4.s, p4/M, z4.s, z17.s\n"
".inst 0x44829223 // srshl z3.s, p4/M, z3.s, z17.s\n"
".inst 0x04b274c6 // sqrdmulh z6.s, z6.s, z18.s\n"
".inst 0x04b274a5 // sqrdmulh z5.s, z5.s, z18.s\n"
".inst 0x04b27484 // sqrdmulh z4.s, z4.s, z18.s\n"
".inst 0x04b27463 // sqrdmulh z3.s, z3.s, z18.s\n"
".inst 0x44829206 // srshl z6.s, p4/M, z6.s, z16.s\n"
".inst 0x44829205 // srshl z5.s, p4/M, z5.s, z16.s\n"
".inst 0x44829204 // srshl z4.s, p4/M, z4.s, z16.s\n"
".inst 0x44829203 // srshl z3.s, p4/M, z3.s, z16.s\n"
".inst 0x44829222 // srshl z2.s, p4/M, z2.s, z17.s\n"
".inst 0x44829221 // srshl z1.s, p4/M, z1.s, z17.s\n"
".inst 0x44829220 // srshl z0.s, p4/M, z0.s, z17.s\n"
"smax z15.s, p4/M, z15.s, z19.s\n"
".inst 0x04b27442 // sqrdmulh z2.s, z2.s, z18.s\n"
".inst 0x04b27421 // sqrdmulh z1.s, z1.s, z18.s\n"
".inst 0x04b27400 // sqrdmulh z0.s, z0.s, z18.s\n"
"smin z15.s, p4/M, z15.s, z20.s\n"
".inst 0x44829202 // srshl z2.s, p4/M, z2.s, z16.s\n"
".inst 0x44829201 // srshl z1.s, p4/M, z1.s, z16.s\n"
".inst 0x44829200 // srshl z0.s, p4/M, z0.s, z16.s\n"
"smax z14.s, p4/M, z14.s, z19.s\n"
"smax z13.s, p4/M, z13.s, z19.s\n"
"smax z12.s, p4/M, z12.s, z19.s\n"
"smax z11.s, p4/M, z11.s, z19.s\n"
"smin z14.s, p4/M, z14.s, z20.s\n"
"smin z13.s, p4/M, z13.s, z20.s\n"
"smin z12.s, p4/M, z12.s, z20.s\n"
"smin z11.s, p4/M, z11.s, z20.s\n"
"trn1 z17.h, z15.h, z14.h\n"
"smax z10.s, p4/M, z10.s, z19.s\n"
"trn1 z16.h, z13.h, z12.h\n"
"smax z9.s, p4/M, z9.s, z19.s\n"
"trn1 z16.b, z17.b, z16.b\n"
"st1b { z16.b }, p3, [%x[outptr], x26]\n"
"smin z10.s, p4/M, z10.s, z20.s\n"
"incb x26, ALL, MUL #4\n"
"smin z9.s, p4/M, z9.s, z20.s\n"
"smax z8.s, p4/M, z8.s, z19.s\n"
"smax z7.s, p4/M, z7.s, z19.s\n"
"smax z6.s, p4/M, z6.s, z19.s\n"
"trn1 z18.h, z11.h, z10.h\n"
"smin z8.s, p4/M, z8.s, z20.s\n"
"smin z7.s, p4/M, z7.s, z20.s\n"
"smin z6.s, p4/M, z6.s, z20.s\n"
"smax z5.s, p4/M, z5.s, z19.s\n"
"trn1 z16.h, z9.h, z8.h\n"
"smax z4.s, p4/M, z4.s, z19.s\n"
"trn1 z17.h, z7.h, z6.h\n"
"trn1 z16.b, z18.b, z16.b\n"
"st1b { z16.b }, p2, [%x[outptr], x25]\n"
"smin z5.s, p4/M, z5.s, z20.s\n"
"incb x25, ALL, MUL #4\n"
"smin z4.s, p4/M, z4.s, z20.s\n"
"smax z3.s, p4/M, z3.s, z19.s\n"
"smax z2.s, p4/M, z2.s, z19.s\n"
"smax z1.s, p4/M, z1.s, z19.s\n"
"smax z0.s, p4/M, z0.s, z19.s\n"
"trn1 z16.h, z5.h, z4.h\n"
"smin z3.s, p4/M, z3.s, z20.s\n"
"trn1 z16.b, z17.b, z16.b\n"
"st1b { z16.b }, p1, [%x[outptr], x24]\n"
"smin z2.s, p4/M, z2.s, z20.s\n"
"incb x24, ALL, MUL #4\n"
"smin z1.s, p4/M, z1.s, z20.s\n"
"smin z0.s, p4/M, z0.s, z20.s\n"
"trn1 z17.h, z3.h, z2.h\n"
"trn1 z16.h, z1.h, z0.h\n"
"trn1 z16.b, z17.b, z16.b\n"
"st1b { z16.b }, p0, [%x[outptr], x23]\n"
"incb x23, ALL, MUL #4\n"
"whilelt p0.b, x23, %x[n_channels]\n"
"b.any 1b\n"
"7:" // Single vector of channels
"whilelt p3.b, x26, %x[n_channels]\n"
"b.none 14f\n"
"8:" // Single vector of channels: Loop
"mov z15.s, #0x0\n"
"mov x19, %x[inptrs]\n"
"mov z14.s, #0x0\n"
"lsr x22, %x[n_valid_cells], #0x1\n"
"mov z13.s, #0x0\n"
"mov z12.s, #0x0\n"
"cbz x22, 11f\n"
"ldp x21, x20, [x19, #0x0]\n"
"ld1b { z31.b }, p3/Z, [x21, x26]\n"
"add x19, x19, #0x10\n"
"ld1b { z30.b }, p3/Z, [x20, x26]\n"
"subs x22, x22, #0x1\n"
"beq 10f\n"
"9:" // Single vector of channels: Loop: 2 inputs loop
".inst 0x455e03f7 // saddlb z23.h, z31.b, z30.b\n"
"ldp x21, x20, [x19, #0x0]\n"
"add x19, x19, #0x10\n"
".inst 0x455e07f6 // saddlt z22.h, z31.b, z30.b\n"
"ld1b { z31.b }, p3/Z, [x21, x26]\n"
".inst 0x459741ef // saddwb z15.s, z15.s, z23.h\n"
"subs x22, x22, #0x1\n"
".inst 0x459745ce // saddwt z14.s, z14.s, z23.h\n"
"ld1b { z30.b }, p3/Z, [x20, x26]\n"
".inst 0x459641ad // saddwb z13.s, z13.s, z22.h\n"
".inst 0x4596458c // saddwt z12.s, z12.s, z22.h\n"
"bgt 9b\n"
"10:" // Single vector of channels: Loop: 2 inputs tail
".inst 0x455e03f7 // saddlb z23.h, z31.b, z30.b\n"
".inst 0x455e07f6 // saddlt z22.h, z31.b, z30.b\n"
".inst 0x459741ef // saddwb z15.s, z15.s, z23.h\n"
".inst 0x459745ce // saddwt z14.s, z14.s, z23.h\n"
".inst 0x459641ad // saddwb z13.s, z13.s, z22.h\n"
".inst 0x4596458c // saddwt z12.s, z12.s, z22.h\n"
"11:" // Single vector of channels: Loop: After loop
"ands x20, %x[n_valid_cells], #0x1\n"
"beq 13f\n"
"12:" // Single vector of channels: Loop: Single input loop
"ldr x21, [x19], #0x8\n"
"subs x20, x20, #0x1\n"
"ld1b { z31.b }, p3/Z, [x21, x26]\n"
".inst 0x4508a3f1 // sshllb z17.h, z31.b, #0x0\n"
".inst 0x4508a7f0 // sshllt z16.h, z31.b, #0x0\n"
".inst 0x459141ef // saddwb z15.s, z15.s, z17.h\n"
".inst 0x459145ce // saddwt z14.s, z14.s, z17.h\n"
".inst 0x459041ad // saddwb z13.s, z13.s, z16.h\n"
".inst 0x4590458c // saddwt z12.s, z12.s, z16.h\n"
"bgt 12b\n"
"13:" // Single vector of channels: Loop: Single input loop: End
"mov z20.s, #0x7f\n"
"ld1rw { z18.s }, p4/Z, [%x[combined_rescale_value]]\n"
"ld1rw { z17.s }, p4/Z, [%x[left_shift]]\n"
"not z19.s, p4/M, z20.s\n"
"ld1rw { z16.s }, p4/Z, [%x[right_shift]]\n"
".inst 0x4482922f // srshl z15.s, p4/M, z15.s, z17.s\n"
".inst 0x4482922e // srshl z14.s, p4/M, z14.s, z17.s\n"
".inst 0x4482922d // srshl z13.s, p4/M, z13.s, z17.s\n"
".inst 0x4482922c // srshl z12.s, p4/M, z12.s, z17.s\n"
".inst 0x04b275ef // sqrdmulh z15.s, z15.s, z18.s\n"
".inst 0x04b275ce // sqrdmulh z14.s, z14.s, z18.s\n"
".inst 0x04b275ad // sqrdmulh z13.s, z13.s, z18.s\n"
".inst 0x04b2758c // sqrdmulh z12.s, z12.s, z18.s\n"
".inst 0x4482920f // srshl z15.s, p4/M, z15.s, z16.s\n"
".inst 0x4482920e // srshl z14.s, p4/M, z14.s, z16.s\n"
".inst 0x4482920d // srshl z13.s, p4/M, z13.s, z16.s\n"
".inst 0x4482920c // srshl z12.s, p4/M, z12.s, z16.s\n"
"smax z15.s, p4/M, z15.s, z19.s\n"
"smax z14.s, p4/M, z14.s, z19.s\n"
"smax z13.s, p4/M, z13.s, z19.s\n"
"smax z12.s, p4/M, z12.s, z19.s\n"
"smin z15.s, p4/M, z15.s, z20.s\n"
"smin z14.s, p4/M, z14.s, z20.s\n"
"smin z13.s, p4/M, z13.s, z20.s\n"
"smin z12.s, p4/M, z12.s, z20.s\n"
"trn1 z17.h, z15.h, z14.h\n"
"trn1 z16.h, z13.h, z12.h\n"
"trn1 z16.b, z17.b, z16.b\n"
"st1b { z16.b }, p3, [%x[outptr], x26]\n"
"incb x26\n"
"whilelt p3.b, x26, %x[n_channels]\n"
"b.any 8b\n"
"14:" // End
:
: [combined_rescale_value] "r" (&combined_rescale_value), [inptrs] "r" (inptrs), [left_shift] "r" (&left_shift), [n_channels] "r" (n_channels), [n_valid_cells] "r" (n_valid_cells), [outptr] "r" (outptr), [right_shift] "r" (&right_shift)
: "cc", "memory", "p0", "p1", "p2", "p3", "p4", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
);
}
} // namespace pooling
} // namespace arm_conv
#endif // defined(__ARM_FEATURE_SVE) && defined(SVE2)
|
