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
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
|
/*
* Copyright (c) 2022-2023 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.
*/
#if defined(ARM_COMPUTE_ENABLE_SME2)
#include <algorithm>
#include <cstddef>
#include "arm_gemm.hpp"
using arm_gemm::Requantize32;
namespace arm_conv {
namespace depthwise {
void sme2_u8s8u8q_planar_3x3_s1_4rows_dot_za_impl(
const uint8_t *inptr,
size_t ld_in_row,
size_t ld_in_col,
size_t ld_in_vl,
unsigned int pad_top,
unsigned int valid_input_rows,
unsigned int pad_left,
unsigned int valid_input_cols,
const int8_t *weights,
uint8_t **outptrs,
const size_t *outlds,
const size_t *outvllds,
unsigned int output_cols,
unsigned int start_channel,
unsigned int valid_channels,
const arm_gemm::Requantize32 &qp
)
{
struct Args
{
const uint8_t *inptr;
size_t ld_in_vl;
long unsigned int pad_top, pad_bottom, pad_left;
const int8_t *weights;
long unsigned int input_cols, output_cols;
uint8_t **outptrs;
const size_t *ld_out_cols;
const size_t *ld_out_vls;
long unsigned int current_channel, n_channels;
};
Args args = { inptr, ld_in_vl, pad_top, 6u - std::min(6u, pad_top + valid_input_rows), pad_left, weights, valid_input_cols, output_cols, outptrs, outlds, outvllds, start_channel, valid_channels };
__asm__ __volatile__(
".inst 0xd503477f // SMSTART ZA\n"
"ldr x6, [%x[args], %[offsetof_Args_pad_bottom]]\n"
"ptrue p2.b\n"
"mov x20, #0x6\n"
"ldr x7, [%x[args], %[offsetof_Args_pad_top]]\n"
"ld1rh { z21.h }, p2/Z, [%x[qp], %[offsetof_Requantize32_a_offset]]\n"
"sub x20, x20, x6\n"
".inst 0x25207812 // ptrue pn10.b\n"
"ldr x17, [%x[args], %[offsetof_Args_n_channels]]\n"
"whilelt p1.s, XZR, x17\n"
"whilelt p9.s, XZR, x20\n"
"ld1rw { z15.s }, p2/Z, [%x[qp], %[offsetof_Requantize32_c_offset]]\n"
"whilelt p8.s, XZR, x7\n"
"addvl SP, SP, #-12\n"
"ldr x16, [%x[args], %[offsetof_Args_current_channel]]\n"
"neg z21.h, p2/M, z21.h\n"
"eor p8.b, p2/Z, p8.b, p9.b\n"
"ld1rw { z14.s }, p2/Z, [%x[qp], %[offsetof_Requantize32_per_layer_mul]]\n"
"ld1rw { z12.s }, p2/Z, [%x[qp], %[offsetof_Requantize32_per_layer_right_shift]]\n"
"ld1rw { z29.s }, p2/Z, [%x[qp], %[offsetof_Requantize32_minval]]\n"
"ld1rw { z28.s }, p2/Z, [%x[qp], %[offsetof_Requantize32_maxval]]\n"
"1:" // Channel loop
"ldr x20, [%x[qp], %[offsetof_Requantize32_bias]]\n"
"mov z30.s, #0x0\n"
"cbz x20, 2f\n"
"ld1w { z30.s }, p1/Z, [x20, x16, LSL #2]\n"
"2:" // Load bias: Done
"ldr x22, [%x[args], %[offsetof_Args_weights]]\n"
"mov x20, x22\n"
"ld1sb { z10.s }, p2/Z, [x20]\n"
"incw x20, ALL, MUL #3\n"
"ld1rh { z31.h }, p2/Z, [%x[qp], %[offsetof_Requantize32_b_offset]]\n"
"mov z7.h, #0x0\n"
"sub z10.h, z10.h, z31.h\n"
"incw x22\n"
"ld1sb { z16.s }, p2/Z, [x20]\n"
"incw x20, ALL, MUL #3\n"
"sub z16.h, z16.h, z31.h\n"
"trn1 z20.h, z7.h, z10.h\n"
"ld1sb { z11.s }, p2/Z, [x20]\n"
"sub z11.h, z11.h, z31.h\n"
"mov x20, x22\n"
"trn1 z19.h, z10.h, z16.h\n"
"ld1sb { z24.s }, p2/Z, [x20]\n"
"incw x20, ALL, MUL #3\n"
"trn1 z26.h, z16.h, z11.h\n"
"trn1 z13.h, z11.h, z7.h\n"
"ld1sb { z11.s }, p2/Z, [x20]\n"
"incw x20, ALL, MUL #3\n"
"sub z24.h, z24.h, z31.h\n"
"sub z11.h, z11.h, z31.h\n"
"ld1sb { z2.s }, p2/Z, [x20]\n"
"sub z2.h, z2.h, z31.h\n"
"addvl x21, SP, #12\n"
"incw x22\n"
"addvl x21, x21, #-4\n"
"mov x20, x22\n"
"st1h { z20.h }, p2, [x21]\n"
"trn1 z22.h, z7.h, z24.h\n"
"st1h { z19.h }, p2, [x21, #1, MUL VL]\n"
"trn1 z1.h, z24.h, z11.h\n"
"ld1sb { z16.s }, p2/Z, [x20]\n"
"incw x20, ALL, MUL #3\n"
"st1h { z26.h }, p2, [x21, #2, MUL VL]\n"
"trn1 z3.h, z11.h, z2.h\n"
"ld1sb { z0.s }, p2/Z, [x20]\n"
"incw x20, ALL, MUL #3\n"
"st1h { z13.h }, p2, [x21, #3, MUL VL]\n"
"trn1 z25.h, z2.h, z7.h\n"
"ld1sb { z4.s }, p2/Z, [x20]\n"
"ldr x20, [%x[qp], %[offsetof_Requantize32_per_channel_muls]]\n"
"sub z16.h, z16.h, z31.h\n"
"sub z0.h, z0.h, z31.h\n"
"addvl x21, x21, #-4\n"
"st1h { z22.h }, p2, [x21]\n"
"sub z4.h, z4.h, z31.h\n"
"st1h { z1.h }, p2, [x21, #1, MUL VL]\n"
"mov z31.d, z30.d\n"
"st1h { z3.h }, p2, [x21, #2, MUL VL]\n"
"trn1 z24.h, z7.h, z16.h\n"
"trn1 z18.h, z16.h, z0.h\n"
"st1h { z25.h }, p2, [x21, #3, MUL VL]\n"
"addvl x21, x21, #-4\n"
"trn1 z0.h, z0.h, z4.h\n"
"trn1 z1.h, z4.h, z7.h\n"
"st1h { z24.h }, p2, [x21]\n"
"st1h { z18.h }, p2, [x21, #1, MUL VL]\n"
"st1h { z0.h }, p2, [x21, #2, MUL VL]\n"
"st1h { z1.h }, p2, [x21, #3, MUL VL]\n"
"cbz x20, 3f\n"
"ld1w { z14.s }, p1/Z, [x20, x16, LSL #2]\n"
"3:" // Load mul: End
"ldr x20, [%x[qp], %[offsetof_Requantize32_per_channel_right_shifts]]\n"
"cbz x20, 4f\n"
"ld1w { z12.s }, p1/Z, [x20, x16, LSL #2]\n"
"4:" // Load right_shift: End
"ldr x15, [%x[args], %[offsetof_Args_input_cols]]\n"
"sub x20, x15, #0x1\n"
"orr x23, x20, %x[ld_in_col], LSL #16\n"
"ldr x14, [%x[args], %[offsetof_Args_inptr]]\n"
"orr x23, x17, x23, LSL #22\n"
"mov x22, #0x6\n"
"add x21, x7, x6\n"
"lsl x20, %x[ld_in_row], #0x0\n"
"ldr x13, [%x[args], %[offsetof_Args_output_cols]]\n"
"mov x8, #0x0\n"
"lsl x23, x23, #0x0\n"
"sub x22, x22, x21\n"
"madd x20, x20, x7, x14\n"
"5:" // Issue prefetches
"subs x22, x22, #0x1\n"
".inst 0xf8b74a9c // rprfm pldstrm, x23, [x20]\n"
"add x20, x20, %x[ld_in_col]\n"
"bgt 5b\n"
"ldr x23, [%x[args], %[offsetof_Args_outptrs]]\n"
"lsl x20, %x[ld_in_row], #0x0\n"
"msub x14, x7, x20, x14\n"
".inst 0xc0040bc0 // mova za.d[x8, #0], { z30.d-z31.d }\n"
"ldr x20, [%x[args], %[offsetof_Args_ld_out_cols]]\n"
".inst 0xc0040bc1 // mova za.d[x8, #1], { z30.d-z31.d }\n"
"mov x22, #0x2\n"
"ldp x11, x10, [x23], #0x10\n"
".inst 0xc0040bc2 // mova za.d[x8, #2], { z30.d-z31.d }\n"
"ldp x9, x28, [x20], #0x10\n"
".inst 0xc0040bc3 // mova za.d[x8, #3], { z30.d-z31.d }\n"
"ldr x21, [%x[args], %[offsetof_Args_pad_left]]\n"
".inst 0xc0040bc4 // mova za.d[x8, #4], { z30.d-z31.d }\n"
"ldp x27, x26, [x23], #0x10\n"
".inst 0xc0040bc5 // mova za.d[x8, #5], { z30.d-z31.d }\n"
"ldp x25, x24, [x20], #0x10\n"
"cbz x21, 7f\n"
"cmp x21, x22\n"
"csel x20, x21, x22, LT\n"
"sub x21, x21, x20\n"
"sub x22, x22, x20\n"
"cbz x21, 7f\n"
".inst 0xc0060818 // mova { z24.d-z25.d }, za.d[x8, #0]\n"
"sub x13, x13, x21\n"
".inst 0xc006083a // mova { z26.d-z27.d }, za.d[x8, #1]\n"
".inst 0xc1aeac18 // sqdmulh { z24.s-z27.s }, { z24.s-z27.s }, z14.s\n"
".inst 0xc1acaa38 // srshl { z24.s-z27.s }, { z24.s-z27.s }, z12.s\n"
".inst 0xc1afab18 // add { z24.s-z27.s }, { z24.s-z27.s }, z15.s\n"
".inst 0xc1bccfb8 // sclamp { z24.s-z27.s }, z29.s, z28.s\n"
"6:" // Left padding
"subs x21, x21, #0x1\n"
"st1b { z24.s }, p1, [x11]\n"
"add x11, x11, x9\n"
"st1b { z26.s }, p1, [x10]\n"
"add x10, x10, x28\n"
"st1b { z25.s }, p1, [x27]\n"
"add x27, x27, x25\n"
"st1b { z27.s }, p1, [x26]\n"
"add x26, x26, x24\n"
"bgt 6b\n"
"7:" // Left padding: End
"adds XZR, x7, x6\n"
"bne 12f\n"
"cbz x22, 10f\n"
"cmp x22, #0x1\n"
"sub x15, x15, x22\n"
"beq 9f\n"
"8:" // Unpadded: 2 priming loads
"add x21, x14, %x[ld_in_row]\n"
"ld1b { z20.s }, p1/Z, [x14]\n"
"addvl x20, SP, #8\n"
"ld1b { z16.s }, p1/Z, [x21]\n"
"add x21, x21, %x[ld_in_row]\n"
"trn1 z4.h, z20.h, z16.h\n"
"add z4.h, z4.h, z21.h\n"
"ld1b { z23.s }, p1/Z, [x21]\n"
"add x21, x21, %x[ld_in_row]\n"
"add x14, x14, %x[ld_in_col]\n"
"ld1b { z22.s }, p1/Z, [x21]\n"
"add x21, x21, %x[ld_in_row]\n"
"trn1 z5.h, z23.h, z22.h\n"
"add z5.h, z5.h, z21.h\n"
"ld1b { z17.s }, p1/Z, [x21]\n"
"add x21, x21, %x[ld_in_row]\n"
"ld1b { z16.s }, p1/Z, [x21]\n"
"trn1 z6.h, z17.h, z16.h\n"
"add z6.h, z6.h, z21.h\n"
".inst 0xa1402a83 // ld1h { z3.h, z11.h }, pn10.b/Z, [x20]\n"
".inst 0xc16b1488 // sdot za.s[x8, 0], { z4.h-z5.h }, z11.h\n"
".inst 0xc1631489 // sdot za.s[x8, 1], { z4.h-z5.h }, z3.h\n"
".inst 0xa1412a80 // ld1h { z0.h, z8.h }, pn10.b/Z, [x20, #0x2, MUL VL]\n"
".inst 0xc16814a8 // sdot za.s[x8, 0], { z5.h-z6.h }, z8.h\n"
".inst 0xc16014a9 // sdot za.s[x8, 1], { z5.h-z6.h }, z0.h\n"
"9:" // Unpadded: 1 priming loads
"add x22, x14, %x[ld_in_row]\n"
"ld1b { z25.s }, p1/Z, [x14]\n"
"addvl x21, SP, #4\n"
"ld1b { z6.s }, p1/Z, [x22]\n"
"add x22, x22, %x[ld_in_row]\n"
"trn1 z3.h, z25.h, z6.h\n"
"add z3.h, z3.h, z21.h\n"
"ld1b { z18.s }, p1/Z, [x22]\n"
"add x22, x22, %x[ld_in_row]\n"
"addvl x20, SP, #8\n"
"ld1b { z26.s }, p1/Z, [x22]\n"
"add x22, x22, %x[ld_in_row]\n"
"trn1 z4.h, z18.h, z26.h\n"
"add z4.h, z4.h, z21.h\n"
"ld1b { z2.s }, p1/Z, [x22]\n"
"add x22, x22, %x[ld_in_row]\n"
"add x14, x14, %x[ld_in_col]\n"
"ld1b { z5.s }, p1/Z, [x22]\n"
"trn1 z5.h, z2.h, z5.h\n"
"add z5.h, z5.h, z21.h\n"
".inst 0xa0402aa0 // ld1h { z0.h-z1.h }, pn10.b/Z, [x21]\n"
".inst 0xc1611468 // sdot za.s[x8, 0], { z3.h-z4.h }, z1.h\n"
".inst 0xc1601469 // sdot za.s[x8, 1], { z3.h-z4.h }, z0.h\n"
".inst 0xa1402a82 // ld1h { z2.h, z10.h }, pn10.b/Z, [x20]\n"
".inst 0xa0412aa8 // ld1h { z8.h-z9.h }, pn10.b/Z, [x21, #0x2, MUL VL]\n"
".inst 0xc16a146a // sdot za.s[x8, 2], { z3.h-z4.h }, z10.h\n"
".inst 0xc162146b // sdot za.s[x8, 3], { z3.h-z4.h }, z2.h\n"
".inst 0xc1691488 // sdot za.s[x8, 0], { z4.h-z5.h }, z9.h\n"
".inst 0xc1681489 // sdot za.s[x8, 1], { z4.h-z5.h }, z8.h\n"
".inst 0xa1412a82 // ld1h { z2.h, z10.h }, pn10.b/Z, [x20, #0x2, MUL VL]\n"
".inst 0xc16a148a // sdot za.s[x8, 2], { z4.h-z5.h }, z10.h\n"
".inst 0xc162148b // sdot za.s[x8, 3], { z4.h-z5.h }, z2.h\n"
"10:" // Unpadded: 0 priming loads
".inst 0xa1402be5 // ld1h { z5.h, z13.h }, pn10.b/Z, [SP]\n"
".inst 0xa0412bea // ld1h { z10.h-z11.h }, pn10.b/Z, [SP, #0x2, MUL VL]\n"
"cbz x15, 18f\n"
"add x20, x14, %x[ld_in_row]\n"
"ld1b { z17.s }, p1/Z, [x14]\n"
"sub x15, x15, #0x1\n"
"ld1b { z9.s }, p1/Z, [x20]\n"
"add x20, x20, %x[ld_in_row]\n"
"trn1 z6.h, z17.h, z9.h\n"
"sub x13, x13, #0x1\n"
"ld1b { z17.s }, p1/Z, [x20]\n"
"add x20, x20, %x[ld_in_row]\n"
"cmp x15, x13\n"
"add z6.h, z6.h, z21.h\n"
"ld1b { z7.s }, p1/Z, [x20]\n"
"add x20, x20, %x[ld_in_row]\n"
"trn1 z7.h, z17.h, z7.h\n"
"csel x23, x15, x13, LT\n"
"ld1b { z17.s }, p1/Z, [x20]\n"
"add x20, x20, %x[ld_in_row]\n"
"add z7.h, z7.h, z21.h\n"
"add x14, x14, %x[ld_in_col]\n"
"ld1b { z1.s }, p1/Z, [x20]\n"
"trn1 z8.h, z17.h, z1.h\n"
"add z8.h, z8.h, z21.h\n"
"sub x13, x13, x23\n"
"cbz x23, 17f\n"
"11:" // Unpadded: Main loop
".inst 0xc16d14c8 // sdot za.s[x8, 0], { z6.h-z7.h }, z13.h\n"
"addvl x22, SP, #4\n"
"addvl x21, SP, #8\n"
"ld1b { z2.s }, p1/Z, [x14]\n"
".inst 0xc16514c9 // sdot za.s[x8, 1], { z6.h-z7.h }, z5.h\n"
".inst 0xa1402ac5 // ld1h { z5.h, z13.h }, pn10.b/Z, [x22]\n"
"add x20, x14, %x[ld_in_row]\n"
"subs x23, x23, #0x1\n"
".inst 0xc16b14e8 // sdot za.s[x8, 0], { z7.h-z8.h }, z11.h\n"
"ld1b { z19.s }, p1/Z, [x20]\n"
"add x20, x20, %x[ld_in_row]\n"
"add x14, x14, %x[ld_in_col]\n"
".inst 0xc16a14e9 // sdot za.s[x8, 1], { z7.h-z8.h }, z10.h\n"
".inst 0xa1412ac3 // ld1h { z3.h, z11.h }, pn10.b/Z, [x22, #0x2, MUL VL]\n"
".inst 0xc0060818 // mova { z24.d-z25.d }, za.d[x8, #0]\n"
"ld1b { z23.s }, p1/Z, [x20]\n"
"add x20, x20, %x[ld_in_row]\n"
".inst 0xc006083a // mova { z26.d-z27.d }, za.d[x8, #1]\n"
".inst 0xc1aeac18 // sqdmulh { z24.s-z27.s }, { z24.s-z27.s }, z14.s\n"
".inst 0xc16d14ca // sdot za.s[x8, 2], { z6.h-z7.h }, z13.h\n"
"ld1b { z18.s }, p1/Z, [x20]\n"
"add x20, x20, %x[ld_in_row]\n"
".inst 0xc16514cb // sdot za.s[x8, 3], { z6.h-z7.h }, z5.h\n"
".inst 0xa1402aa1 // ld1h { z1.h, z9.h }, pn10.b/Z, [x21]\n"
".inst 0xc1acaa38 // srshl { z24.s-z27.s }, { z24.s-z27.s }, z12.s\n"
".inst 0xc16914cc // sdot za.s[x8, 4], { z6.h-z7.h }, z9.h\n"
"ld1b { z17.s }, p1/Z, [x20]\n"
"add x20, x20, %x[ld_in_row]\n"
".inst 0xc16114cd // sdot za.s[x8, 5], { z6.h-z7.h }, z1.h\n"
"ld1b { z16.s }, p1/Z, [x20]\n"
".inst 0xc1afab18 // add { z24.s-z27.s }, { z24.s-z27.s }, z15.s\n"
".inst 0xc16b14ea // sdot za.s[x8, 2], { z7.h-z8.h }, z11.h\n"
"trn1 z6.h, z2.h, z19.h\n"
".inst 0xa1402be5 // ld1h { z5.h, z13.h }, pn10.b/Z, [SP]\n"
".inst 0xc16314eb // sdot za.s[x8, 3], { z7.h-z8.h }, z3.h\n"
".inst 0xa1412aa1 // ld1h { z1.h, z9.h }, pn10.b/Z, [x21, #0x2, MUL VL]\n"
".inst 0xc1bccfb8 // sclamp { z24.s-z27.s }, z29.s, z28.s\n"
".inst 0xc16914ec // sdot za.s[x8, 4], { z7.h-z8.h }, z9.h\n"
"st1b { z24.s }, p1, [x11]\n"
"add x11, x11, x9\n"
"add z6.h, z6.h, z21.h\n"
".inst 0xc16114ed // sdot za.s[x8, 5], { z7.h-z8.h }, z1.h\n"
"trn1 z7.h, z23.h, z18.h\n"
"trn1 z8.h, z17.h, z16.h\n"
"add x8, x8, #0x2\n"
".inst 0xa0412bea // ld1h { z10.h-z11.h }, pn10.b/Z, [SP, #0x2, MUL VL]\n"
"st1b { z26.s }, p1, [x10]\n"
"add x10, x10, x28\n"
".inst 0xc0040bc4 // mova za.d[x8, #4], { z30.d-z31.d }\n"
"st1b { z25.s }, p1, [x27]\n"
"add x27, x27, x25\n"
".inst 0xc0040bc5 // mova za.d[x8, #5], { z30.d-z31.d }\n"
"add z7.h, z7.h, z21.h\n"
"st1b { z27.s }, p1, [x26]\n"
"add x26, x26, x24\n"
"add z8.h, z8.h, z21.h\n"
"bgt 11b\n"
"b 17f\n"
"12:" // Padded
"cbz x22, 15f\n"
"cmp x22, #0x1\n"
"sub x15, x15, x22\n"
"beq 14f\n"
"13:" // Padded: 2 priming loads
"mov x12, #0x0\n"
".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n"
"ld1b { z19.s }, p0/Z, [x14]\n"
"add z19.h, p0/M, z19.h, z21.h\n"
"add x20, x14, %x[ld_in_row]\n"
".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n"
"ld1b { z18.s }, p0/Z, [x20]\n"
"add z18.h, p0/M, z18.h, z21.h\n"
"add x20, x20, %x[ld_in_row]\n"
".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n"
"ld1b { z17.s }, p0/Z, [x20]\n"
"add z17.h, p0/M, z17.h, z21.h\n"
"add x20, x20, %x[ld_in_row]\n"
".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n"
"ld1b { z16.s }, p0/Z, [x20]\n"
"add z16.h, p0/M, z16.h, z21.h\n"
"mov x12, #0x4\n"
"add x20, x20, %x[ld_in_row]\n"
"trn1 z7.h, z19.h, z18.h\n"
"trn1 z8.h, z17.h, z16.h\n"
".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n"
"ld1b { z17.s }, p0/Z, [x20]\n"
"add z17.h, p0/M, z17.h, z21.h\n"
"add x20, x20, %x[ld_in_row]\n"
".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n"
"ld1b { z16.s }, p0/Z, [x20]\n"
"addvl x20, SP, #8\n"
"add z16.h, p0/M, z16.h, z21.h\n"
".inst 0xa1402a82 // ld1h { z2.h, z10.h }, pn10.b/Z, [x20]\n"
"trn1 z9.h, z17.h, z16.h\n"
".inst 0xc16a14e8 // sdot za.s[x8, 0], { z7.h-z8.h }, z10.h\n"
"add x14, x14, %x[ld_in_col]\n"
".inst 0xc16214e9 // sdot za.s[x8, 1], { z7.h-z8.h }, z2.h\n"
".inst 0xa1412a85 // ld1h { z5.h, z13.h }, pn10.b/Z, [x20, #0x2, MUL VL]\n"
".inst 0xc16d1508 // sdot za.s[x8, 0], { z8.h-z9.h }, z13.h\n"
".inst 0xc1651509 // sdot za.s[x8, 1], { z8.h-z9.h }, z5.h\n"
"14:" // Padded: 1 priming loads
"mov x12, #0x0\n"
".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n"
"ld1b { z19.s }, p0/Z, [x14]\n"
"add z19.h, p0/M, z19.h, z21.h\n"
"add x20, x14, %x[ld_in_row]\n"
".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n"
"ld1b { z18.s }, p0/Z, [x20]\n"
"add z18.h, p0/M, z18.h, z21.h\n"
"add x20, x20, %x[ld_in_row]\n"
".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n"
"ld1b { z17.s }, p0/Z, [x20]\n"
"add z17.h, p0/M, z17.h, z21.h\n"
"add x20, x20, %x[ld_in_row]\n"
".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n"
"ld1b { z16.s }, p0/Z, [x20]\n"
"add z16.h, p0/M, z16.h, z21.h\n"
"mov x12, #0x4\n"
"add x20, x20, %x[ld_in_row]\n"
"trn1 z22.h, z19.h, z18.h\n"
"trn1 z23.h, z17.h, z16.h\n"
".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n"
"ld1b { z17.s }, p0/Z, [x20]\n"
"add z17.h, p0/M, z17.h, z21.h\n"
"add x20, x20, %x[ld_in_row]\n"
".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n"
"ld1b { z16.s }, p0/Z, [x20]\n"
"addvl x21, SP, #4\n"
"add z16.h, p0/M, z16.h, z21.h\n"
".inst 0xa0402aa0 // ld1h { z0.h-z1.h }, pn10.b/Z, [x21]\n"
"addvl x20, SP, #8\n"
"trn1 z24.h, z17.h, z16.h\n"
".inst 0xc16116c8 // sdot za.s[x8, 0], { z22.h-z23.h }, z1.h\n"
".inst 0xc16016c9 // sdot za.s[x8, 1], { z22.h-z23.h }, z0.h\n"
".inst 0xa1402a85 // ld1h { z5.h, z13.h }, pn10.b/Z, [x20]\n"
"add x14, x14, %x[ld_in_col]\n"
".inst 0xa0412aa0 // ld1h { z0.h-z1.h }, pn10.b/Z, [x21, #0x2, MUL VL]\n"
".inst 0xc16d16ca // sdot za.s[x8, 2], { z22.h-z23.h }, z13.h\n"
".inst 0xc16516cb // sdot za.s[x8, 3], { z22.h-z23.h }, z5.h\n"
".inst 0xc16116e8 // sdot za.s[x8, 0], { z23.h-z24.h }, z1.h\n"
".inst 0xc16016e9 // sdot za.s[x8, 1], { z23.h-z24.h }, z0.h\n"
".inst 0xa0412a80 // ld1h { z0.h-z1.h }, pn10.b/Z, [x20, #0x2, MUL VL]\n"
".inst 0xc16116ea // sdot za.s[x8, 2], { z23.h-z24.h }, z1.h\n"
".inst 0xc16016eb // sdot za.s[x8, 3], { z23.h-z24.h }, z0.h\n"
"15:" // Padded: 0 priming loads
".inst 0xa1402be5 // ld1h { z5.h, z13.h }, pn10.b/Z, [SP]\n"
".inst 0xa0412bea // ld1h { z10.h-z11.h }, pn10.b/Z, [SP, #0x2, MUL VL]\n"
"cbz x15, 18f\n"
"mov x12, #0x0\n"
".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n"
"ld1b { z19.s }, p0/Z, [x14]\n"
"add z19.h, p0/M, z19.h, z21.h\n"
"add x20, x14, %x[ld_in_row]\n"
".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n"
"ld1b { z18.s }, p0/Z, [x20]\n"
"add z18.h, p0/M, z18.h, z21.h\n"
"add x20, x20, %x[ld_in_row]\n"
".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n"
"ld1b { z17.s }, p0/Z, [x20]\n"
"add z17.h, p0/M, z17.h, z21.h\n"
"add x20, x20, %x[ld_in_row]\n"
".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n"
"ld1b { z16.s }, p0/Z, [x20]\n"
"add z16.h, p0/M, z16.h, z21.h\n"
"mov x12, #0x4\n"
"add x20, x20, %x[ld_in_row]\n"
"trn1 z6.h, z19.h, z18.h\n"
"trn1 z7.h, z17.h, z16.h\n"
".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n"
"ld1b { z17.s }, p0/Z, [x20]\n"
"add z17.h, p0/M, z17.h, z21.h\n"
"add x20, x20, %x[ld_in_row]\n"
".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n"
"ld1b { z16.s }, p0/Z, [x20]\n"
"add z16.h, p0/M, z16.h, z21.h\n"
"sub x15, x15, #0x1\n"
"sub x13, x13, #0x1\n"
"cmp x15, x13\n"
"trn1 z8.h, z17.h, z16.h\n"
"csel x23, x15, x13, LT\n"
"add x14, x14, %x[ld_in_col]\n"
"sub x13, x13, x23\n"
"cbz x23, 17f\n"
"16:" // Padded: Main loop
"mov x12, #0x0\n"
".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n"
"ld1b { z9.s }, p0/Z, [x14]\n"
".inst 0xc16d14c8 // sdot za.s[x8, 0], { z6.h-z7.h }, z13.h\n"
".inst 0xc16514c9 // sdot za.s[x8, 1], { z6.h-z7.h }, z5.h\n"
"add z9.h, p0/M, z9.h, z21.h\n"
"add x22, x14, %x[ld_in_row]\n"
".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n"
"ld1b { z19.s }, p0/Z, [x22]\n"
".inst 0xc16b14e8 // sdot za.s[x8, 0], { z7.h-z8.h }, z11.h\n"
"add z19.h, p0/M, z19.h, z21.h\n"
"add x22, x22, %x[ld_in_row]\n"
".inst 0xc16a14e9 // sdot za.s[x8, 1], { z7.h-z8.h }, z10.h\n"
".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n"
"ld1b { z18.s }, p0/Z, [x22]\n"
"add z18.h, p0/M, z18.h, z21.h\n"
".inst 0xc0060818 // mova { z24.d-z25.d }, za.d[x8, #0]\n"
"add x22, x22, %x[ld_in_row]\n"
".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n"
"ld1b { z16.s }, p0/Z, [x22]\n"
".inst 0xc006083a // mova { z26.d-z27.d }, za.d[x8, #1]\n"
"mov x12, #0x4\n"
"addvl x21, SP, #4\n"
"add z16.h, p0/M, z16.h, z21.h\n"
".inst 0xc1aeac18 // sqdmulh { z24.s-z27.s }, { z24.s-z27.s }, z14.s\n"
"add x22, x22, %x[ld_in_row]\n"
".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n"
".inst 0xa1402aa3 // ld1h { z3.h, z11.h }, pn10.b/Z, [x21]\n"
"addvl x20, SP, #8\n"
".inst 0xc16b14ca // sdot za.s[x8, 2], { z6.h-z7.h }, z11.h\n"
"subs x23, x23, #0x1\n"
"ld1b { z17.s }, p0/Z, [x22]\n"
".inst 0xc16314cb // sdot za.s[x8, 3], { z6.h-z7.h }, z3.h\n"
".inst 0xc1acaa38 // srshl { z24.s-z27.s }, { z24.s-z27.s }, z12.s\n"
".inst 0xa1402a85 // ld1h { z5.h, z13.h }, pn10.b/Z, [x20]\n"
"add z17.h, p0/M, z17.h, z21.h\n"
"add x22, x22, %x[ld_in_row]\n"
".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n"
".inst 0xa0412aaa // ld1h { z10.h-z11.h }, pn10.b/Z, [x21, #0x2, MUL VL]\n"
".inst 0xc16d14cc // sdot za.s[x8, 4], { z6.h-z7.h }, z13.h\n"
".inst 0xc1afab18 // add { z24.s-z27.s }, { z24.s-z27.s }, z15.s\n"
"ld1b { z2.s }, p0/Z, [x22]\n"
".inst 0xc16514cd // sdot za.s[x8, 5], { z6.h-z7.h }, z5.h\n"
"add z2.h, p0/M, z2.h, z21.h\n"
"add x14, x14, %x[ld_in_col]\n"
".inst 0xc16b14ea // sdot za.s[x8, 2], { z7.h-z8.h }, z11.h\n"
".inst 0xa1402be5 // ld1h { z5.h, z13.h }, pn10.b/Z, [SP]\n"
".inst 0xc1bccfb8 // sclamp { z24.s-z27.s }, z29.s, z28.s\n"
".inst 0xc16a14eb // sdot za.s[x8, 3], { z7.h-z8.h }, z10.h\n"
".inst 0xa1412a83 // ld1h { z3.h, z11.h }, pn10.b/Z, [x20, #0x2, MUL VL]\n"
"st1b { z24.s }, p1, [x11]\n"
"add x11, x11, x9\n"
".inst 0xc16b14ec // sdot za.s[x8, 4], { z7.h-z8.h }, z11.h\n"
"st1b { z26.s }, p1, [x10]\n"
"add x10, x10, x28\n"
"trn1 z6.h, z9.h, z19.h\n"
".inst 0xc16314ed // sdot za.s[x8, 5], { z7.h-z8.h }, z3.h\n"
"add x8, x8, #0x2\n"
".inst 0xa0412bea // ld1h { z10.h-z11.h }, pn10.b/Z, [SP, #0x2, MUL VL]\n"
"st1b { z25.s }, p1, [x27]\n"
"add x27, x27, x25\n"
"st1b { z27.s }, p1, [x26]\n"
"add x26, x26, x24\n"
".inst 0xc0040bc4 // mova za.d[x8, #4], { z30.d-z31.d }\n"
".inst 0xc0040bc5 // mova za.d[x8, #5], { z30.d-z31.d }\n"
"trn1 z7.h, z18.h, z16.h\n"
"trn1 z8.h, z17.h, z2.h\n"
"bgt 16b\n"
"17:" // Main loop tail
".inst 0xc16d14c8 // sdot za.s[x8, 0], { z6.h-z7.h }, z13.h\n"
"addvl x21, SP, #4\n"
"addvl x20, SP, #8\n"
".inst 0xc16514c9 // sdot za.s[x8, 1], { z6.h-z7.h }, z5.h\n"
".inst 0xa0402aa0 // ld1h { z0.h-z1.h }, pn10.b/Z, [x21]\n"
".inst 0xc16b14e8 // sdot za.s[x8, 0], { z7.h-z8.h }, z11.h\n"
".inst 0xc16a14e9 // sdot za.s[x8, 1], { z7.h-z8.h }, z10.h\n"
".inst 0xa0412aa2 // ld1h { z2.h-z3.h }, pn10.b/Z, [x21, #0x2, MUL VL]\n"
".inst 0xc0060818 // mova { z24.d-z25.d }, za.d[x8, #0]\n"
".inst 0xc006083a // mova { z26.d-z27.d }, za.d[x8, #1]\n"
".inst 0xc1aeac18 // sqdmulh { z24.s-z27.s }, { z24.s-z27.s }, z14.s\n"
".inst 0xc16114ca // sdot za.s[x8, 2], { z6.h-z7.h }, z1.h\n"
".inst 0xc1acaa38 // srshl { z24.s-z27.s }, { z24.s-z27.s }, z12.s\n"
".inst 0xc16014cb // sdot za.s[x8, 3], { z6.h-z7.h }, z0.h\n"
".inst 0xa1402a81 // ld1h { z1.h, z9.h }, pn10.b/Z, [x20]\n"
".inst 0xc1afab18 // add { z24.s-z27.s }, { z24.s-z27.s }, z15.s\n"
".inst 0xc16914cc // sdot za.s[x8, 4], { z6.h-z7.h }, z9.h\n"
".inst 0xc1bccfb8 // sclamp { z24.s-z27.s }, z29.s, z28.s\n"
"st1b { z24.s }, p1, [x11]\n"
"add x11, x11, x9\n"
".inst 0xc16114cd // sdot za.s[x8, 5], { z6.h-z7.h }, z1.h\n"
"st1b { z26.s }, p1, [x10]\n"
"add x10, x10, x28\n"
".inst 0xc16314ea // sdot za.s[x8, 2], { z7.h-z8.h }, z3.h\n"
"st1b { z25.s }, p1, [x27]\n"
"add x27, x27, x25\n"
".inst 0xc16214eb // sdot za.s[x8, 3], { z7.h-z8.h }, z2.h\n"
".inst 0xa0412a80 // ld1h { z0.h-z1.h }, pn10.b/Z, [x20, #0x2, MUL VL]\n"
"st1b { z27.s }, p1, [x26]\n"
"add x26, x26, x24\n"
".inst 0xc16114ec // sdot za.s[x8, 4], { z7.h-z8.h }, z1.h\n"
".inst 0xc16014ed // sdot za.s[x8, 5], { z7.h-z8.h }, z0.h\n"
"add x8, x8, #0x2\n"
".inst 0xc0040bc4 // mova za.d[x8, #4], { z30.d-z31.d }\n"
".inst 0xc0040bc5 // mova za.d[x8, #5], { z30.d-z31.d }\n"
"18:" // Main loop skip tail
"cbz x13, 20f\n"
"19:" // Right padding loop
".inst 0xc0060804 // mova { z4.d-z5.d }, za.d[x8, #0]\n"
"subs x13, x13, #0x1\n"
".inst 0xc0060826 // mova { z6.d-z7.d }, za.d[x8, #1]\n"
".inst 0xc1aeac04 // sqdmulh { z4.s-z7.s }, { z4.s-z7.s }, z14.s\n"
"add x8, x8, #0x2\n"
".inst 0xc1acaa24 // srshl { z4.s-z7.s }, { z4.s-z7.s }, z12.s\n"
".inst 0xc0040bc4 // mova za.d[x8, #4], { z30.d-z31.d }\n"
".inst 0xc1afab04 // add { z4.s-z7.s }, { z4.s-z7.s }, z15.s\n"
".inst 0xc0040bc5 // mova za.d[x8, #5], { z30.d-z31.d }\n"
".inst 0xc1bccfa4 // sclamp { z4.s-z7.s }, z29.s, z28.s\n"
"st1b { z4.s }, p1, [x11]\n"
"add x11, x11, x9\n"
"st1b { z6.s }, p1, [x10]\n"
"add x10, x10, x28\n"
"st1b { z5.s }, p1, [x27]\n"
"add x27, x27, x25\n"
"st1b { z7.s }, p1, [x26]\n"
"add x26, x26, x24\n"
"bgt 19b\n"
"20:" // End
"ldr x20, [%x[args], %[offsetof_Args_weights]]\n"
"incw x20, ALL, MUL #9\n"
"str x20, [%x[args], %[offsetof_Args_weights]]\n"
"incw x16\n"
"ldr x21, [%x[args], %[offsetof_Args_ld_in_vl]]\n"
"whilelt p1.s, x16, x17\n"
"ldr x20, [%x[args], %[offsetof_Args_inptr]]\n"
"add x20, x20, x21\n"
"str x20, [%x[args], %[offsetof_Args_inptr]]\n"
"ldr x25, [%x[args], %[offsetof_Args_outptrs]]\n"
"ldr x24, [%x[args], %[offsetof_Args_ld_out_vls]]\n"
"ldp x23, x22, [x25, #0x0]\n"
"ldp x21, x20, [x24, #0x0]\n"
"add x23, x23, x21\n"
"add x22, x22, x20\n"
"stp x23, x22, [x25, #0x0]\n"
"ldp x23, x22, [x25, #0x10]\n"
"ldp x21, x20, [x24, #0x10]\n"
"add x23, x23, x21\n"
"add x22, x22, x20\n"
"stp x23, x22, [x25, #0x10]\n"
"b.any 1b\n"
"addvl SP, SP, #12\n"
".inst 0xd503467f // SMSTOP\n"
:
: [args] "r" (&args), [ld_in_col] "r" (ld_in_col), [ld_in_row] "r" (ld_in_row), [offsetof_Args_current_channel] "I" (offsetof(Args, current_channel)), [offsetof_Args_inptr] "I" (offsetof(Args, inptr)), [offsetof_Args_input_cols] "I" (offsetof(Args, input_cols)), [offsetof_Args_ld_in_vl] "I" (offsetof(Args, ld_in_vl)), [offsetof_Args_ld_out_cols] "I" (offsetof(Args, ld_out_cols)), [offsetof_Args_ld_out_vls] "I" (offsetof(Args, ld_out_vls)), [offsetof_Args_n_channels] "I" (offsetof(Args, n_channels)), [offsetof_Args_outptrs] "I" (offsetof(Args, outptrs)), [offsetof_Args_output_cols] "I" (offsetof(Args, output_cols)), [offsetof_Args_pad_bottom] "I" (offsetof(Args, pad_bottom)), [offsetof_Args_pad_left] "I" (offsetof(Args, pad_left)), [offsetof_Args_pad_top] "I" (offsetof(Args, pad_top)), [offsetof_Args_weights] "I" (offsetof(Args, weights)), [offsetof_Requantize32_a_offset] "I" (offsetof(arm_gemm::Requantize32, a_offset)), [offsetof_Requantize32_b_offset] "I" (offsetof(arm_gemm::Requantize32, b_offset)), [offsetof_Requantize32_bias] "I" (offsetof(arm_gemm::Requantize32, bias)), [offsetof_Requantize32_c_offset] "I" (offsetof(arm_gemm::Requantize32, c_offset)), [offsetof_Requantize32_maxval] "I" (offsetof(arm_gemm::Requantize32, maxval)), [offsetof_Requantize32_minval] "I" (offsetof(arm_gemm::Requantize32, minval)), [offsetof_Requantize32_per_channel_muls] "I" (offsetof(arm_gemm::Requantize32, per_channel_muls)), [offsetof_Requantize32_per_channel_right_shifts] "I" (offsetof(arm_gemm::Requantize32, per_channel_right_shifts)), [offsetof_Requantize32_per_layer_mul] "I" (offsetof(arm_gemm::Requantize32, per_layer_mul)), [offsetof_Requantize32_per_layer_right_shift] "I" (offsetof(arm_gemm::Requantize32, per_layer_right_shift)), [qp] "r" (&qp)
: "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x6", "x7", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "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 depthwise
} // namespace arm_conv
#endif // defined(ARM_COMPUTE_ENABLE_SME2)
|
