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
|
/*
* 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.
*/
#ifdef __ARM_FEATURE_SVE
#ifdef ARM_COMPUTE_ENABLE_SME2
#include "arm_gemm.hpp"
#include "../../utils.hpp"
#include <cassert>
#include <limits>
namespace arm_gemm {
void sme2_gemv_fp32_mla_16VL (
const float *A_ptr, const float *B_ptr, float *output_ptr,
size_t N, size_t K,
const float *bias, Activation act, bool
)
{
struct KernelArgs {
float maxval = static_cast<float>(std::numeric_limits<float>::infinity());
float minval = - static_cast<float>(std::numeric_limits<float>::infinity());
const float *B_ptr = {};
size_t output_offset = {};
unsigned int input_initial_col = {};
} ka;
unsigned long flags=0;
ka.B_ptr = B_ptr;
switch(act.type) {
default:
case Activation::Type::None:
break;
case Activation::Type::BoundedReLU:
ka.maxval = static_cast<float>(act.param1);
/* fall through */
case Activation::Type::ReLU:
ka.minval = 0;
flags |= 0x2;
break;
}
__asm__ __volatile__(
"ptrue p1.b\n"
".inst 0xd503477f // SMSTART ZA\n"
"cntw x28, ALL, MUL #4\n"
"add x27, %x[N], x28\n"
"sub x27, x27, #0x1\n"
"udiv x27, x27, x28\n"
"add x22, x27, #0x3\n"
"and x22, x22, #0xfffffffffffffffc\n"
"mul x22, x22, x28\n"
"mul x22, x22, %x[K]\n"
"mov x9, #0x0\n"
"mov x26, %x[B_ptr]\n"
"mov x25, %x[output_ptr]\n"
"ptrue p1.b\n"
".inst 0x25207811 // ptrue pn9.b\n"
"lsl x22, x22, #0x2\n"
"mov x21, #0x1\n"
"1:" // RHS size check loop
"cmp x22, #0x200000\n"
"blt 2f\n"
"tbnz x22, #0, 3f\n"
"lsr x22, x22, #0x1\n"
"lsl x21, x21, #0x1\n"
"b 1b\n"
"2:" // RHS do prefetch
"lsl x20, x22, #0x26\n"
"sub x21, x21, #0x1\n"
"lsl x21, x21, #0x16\n"
"orr x22, x22, x20\n"
"orr x22, x22, x21\n"
".inst 0xf8b64b5a // rprfm pldonce, x22, [x26]\n"
"3:" // RHS prefetch exit
"mov x24, %x[bias]\n"
"4:" // Column loop
"cmp x27, #0x4\n"
"bge 28f\n"
"cmp x27, #0x2\n"
"bgt 20f\n"
"beq 12f\n"
"mov x23, %x[A_ptr]\n"
"lsl x22, %x[K], #0x2\n"
"mov x20, %x[N]\n"
"mov x21, %x[K]\n"
".inst 0xf8b64af8 // rprfm pldmany, x22, [x23]\n"
".inst 0x25b467f0 // whilelt p8.s, XZR, x20, VLx4\n"
"cbz x24, 5f\n"
".inst 0xa040c700 // ld1w { z0.s-z3.s }, pn9.b/Z, [x24]\n"
".inst 0xc0042c00 // mova za.d[x9, #0], { z0.d-z3.d }\n"
"b 6f\n"
"5:" // Width 1: no bias
".inst 0xc00800ff // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
"6:" // Width 1: setup done
"cmp x21, #0x4\n"
"ble 8f\n"
"7:" // Width 1: Multiply loop: Main loop head
"whilelt p0.s, XZR, x21\n"
"ld1rqw { z10.s }, p0/Z, [x23]\n"
"sub x21, x21, #0x4\n"
".inst 0xa040c741 // ldnt1w { z0.s-z3.s }, pn9.b/Z, [x26]\n"
".inst 0xc15aa000 // fmla za.s[x9, 0], { z0.s-z3.s }, z10.s[0]\n"
"addvl x26, x26, #16\n"
"cmp x21, #0x4\n"
".inst 0xa040c759 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x26]\n"
".inst 0xc15aa700 // fmla za.s[x9, 0], { z24.s-z27.s }, z10.s[1]\n"
"addvl x26, x26, #16\n"
"add x23, x23, #0x10\n"
".inst 0xa040c74d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x26]\n"
".inst 0xc15aa980 // fmla za.s[x9, 0], { z12.s-z15.s }, z10.s[2]\n"
"addvl x26, x26, #16\n"
".inst 0xa040c751 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26]\n"
".inst 0xc15aae00 // fmla za.s[x9, 0], { z16.s-z19.s }, z10.s[3]\n"
"addvl x26, x26, #16\n"
"bgt 7b\n"
"8:" // Width 1: Multiply loop: Single iteration only
"whilelt p0.s, XZR, x21\n"
"ld1rqw { z10.s }, p0/Z, [x23]\n"
"subs x21, x21, #0x1\n"
".inst 0xa040c741 // ldnt1w { z0.s-z3.s }, pn9.b/Z, [x26]\n"
"add x23, x23, #0x10\n"
".inst 0xc15aa000 // fmla za.s[x9, 0], { z0.s-z3.s }, z10.s[0]\n"
"addvl x26, x26, #16\n"
"ble 9f\n"
".inst 0xa040c759 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x26]\n"
"subs x21, x21, #0x1\n"
".inst 0xc15aa700 // fmla za.s[x9, 0], { z24.s-z27.s }, z10.s[1]\n"
"addvl x26, x26, #16\n"
"ble 9f\n"
".inst 0xa040c74d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x26]\n"
"subs x21, x21, #0x1\n"
".inst 0xc15aa980 // fmla za.s[x9, 0], { z12.s-z15.s }, z10.s[2]\n"
"addvl x26, x26, #16\n"
"ble 9f\n"
".inst 0xa040c751 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26]\n"
".inst 0xc15aae00 // fmla za.s[x9, 0], { z16.s-z19.s }, z10.s[3]\n"
"addvl x26, x26, #16\n"
"9:" // Width 1: Multiply loop: multiply skip
"tbz %x[flags], #1, 10f\n"
"add x21, %x[args_ptr], %[offset_min]\n"
"add x20, %x[args_ptr], %[offset_max]\n"
".inst 0xc0062c08 // mova { z8.d-z11.d }, za.d[x9, #0]\n"
"ld1rw { z0.s }, p1/Z, [x21]\n"
"ld1rw { z6.s }, p1/Z, [x20]\n"
".inst 0xc1a6c808 // fclamp { z8.s-z11.s }, z0.s, z6.s\n"
".inst 0xa060c328 // st1w { z8.s-z11.s }, p8, [x25]\n"
"addvl x25, x25, #4\n"
"b 11f\n"
"10:" // Width 1: No activation
".inst 0xc0062c08 // mova { z8.d-z11.d }, za.d[x9, #0]\n"
".inst 0xa060c328 // st1w { z8.s-z11.s }, p8, [x25]\n"
"addvl x25, x25, #4\n"
"11:" // Width 1: Output done
"b 36f\n"
"12:" // Width 2
"mov x23, %x[A_ptr]\n"
"lsl x22, %x[K], #0x2\n"
"sub x20, %x[N], x28\n"
"mov x21, %x[K]\n"
".inst 0xf8b64af8 // rprfm pldmany, x22, [x23]\n"
".inst 0x25b467f0 // whilelt p8.s, XZR, x20, VLx4\n"
"cbz x24, 13f\n"
".inst 0xa040c700 // ld1w { z0.s-z3.s }, pn9.b/Z, [x24]\n"
".inst 0xc0042c00 // mova za.d[x9, #0], { z0.d-z3.d }\n"
".inst 0xa041c708 // ld1w { z8.s-z11.s }, pn9.b/Z, [x24, #0x4, MUL VL]\n"
".inst 0xc0042d01 // mova za.d[x9, #1], { z8.d-z11.d }\n"
"b 14f\n"
"13:" // Width 2: no bias
".inst 0xc00800ff // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
"14:" // Width 2: setup done
"cmp x21, #0x4\n"
"ble 16f\n"
"15:" // Width 2: Multiply loop: Main loop head
"whilelt p0.s, XZR, x21\n"
"ld1rqw { z10.s }, p0/Z, [x23]\n"
"sub x21, x21, #0x4\n"
".inst 0xa040c741 // ldnt1w { z0.s-z3.s }, pn9.b/Z, [x26]\n"
".inst 0xc15aa000 // fmla za.s[x9, 0], { z0.s-z3.s }, z10.s[0]\n"
"cmp x21, #0x4\n"
"add x23, x23, #0x10\n"
".inst 0xa041c745 // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
".inst 0xc15aa081 // fmla za.s[x9, 1], { z4.s-z7.s }, z10.s[0]\n"
"addvl x26, x26, #16\n"
".inst 0xa040c759 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x26]\n"
".inst 0xc15aa700 // fmla za.s[x9, 0], { z24.s-z27.s }, z10.s[1]\n"
".inst 0xa041c751 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
".inst 0xc15aa601 // fmla za.s[x9, 1], { z16.s-z19.s }, z10.s[1]\n"
"addvl x26, x26, #16\n"
".inst 0xa040c74d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x26]\n"
".inst 0xc15aa980 // fmla za.s[x9, 0], { z12.s-z15.s }, z10.s[2]\n"
".inst 0xa041c751 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
".inst 0xc15aaa01 // fmla za.s[x9, 1], { z16.s-z19.s }, z10.s[2]\n"
"addvl x26, x26, #16\n"
".inst 0xa040c751 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26]\n"
".inst 0xc15aae00 // fmla za.s[x9, 0], { z16.s-z19.s }, z10.s[3]\n"
".inst 0xa041c759 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
".inst 0xc15aaf01 // fmla za.s[x9, 1], { z24.s-z27.s }, z10.s[3]\n"
"addvl x26, x26, #16\n"
"bgt 15b\n"
"16:" // Width 2: Multiply loop: Single iteration only
"whilelt p0.s, XZR, x21\n"
"ld1rqw { z10.s }, p0/Z, [x23]\n"
"subs x21, x21, #0x1\n"
".inst 0xa040c741 // ldnt1w { z0.s-z3.s }, pn9.b/Z, [x26]\n"
"add x23, x23, #0x10\n"
".inst 0xc15aa000 // fmla za.s[x9, 0], { z0.s-z3.s }, z10.s[0]\n"
".inst 0xa041c745 // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
".inst 0xc15aa081 // fmla za.s[x9, 1], { z4.s-z7.s }, z10.s[0]\n"
"addvl x26, x26, #16\n"
"ble 17f\n"
".inst 0xa040c759 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x26]\n"
"subs x21, x21, #0x1\n"
".inst 0xc15aa700 // fmla za.s[x9, 0], { z24.s-z27.s }, z10.s[1]\n"
".inst 0xa041c751 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
".inst 0xc15aa601 // fmla za.s[x9, 1], { z16.s-z19.s }, z10.s[1]\n"
"addvl x26, x26, #16\n"
"ble 17f\n"
".inst 0xa040c74d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x26]\n"
"subs x21, x21, #0x1\n"
".inst 0xc15aa980 // fmla za.s[x9, 0], { z12.s-z15.s }, z10.s[2]\n"
".inst 0xa041c751 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
".inst 0xc15aaa01 // fmla za.s[x9, 1], { z16.s-z19.s }, z10.s[2]\n"
"addvl x26, x26, #16\n"
"ble 17f\n"
".inst 0xa040c751 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26]\n"
".inst 0xc15aae00 // fmla za.s[x9, 0], { z16.s-z19.s }, z10.s[3]\n"
".inst 0xa041c759 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
".inst 0xc15aaf01 // fmla za.s[x9, 1], { z24.s-z27.s }, z10.s[3]\n"
"addvl x26, x26, #16\n"
"17:" // Width 2: Multiply loop: multiply skip
"tbz %x[flags], #1, 18f\n"
"add x21, %x[args_ptr], %[offset_min]\n"
"add x20, %x[args_ptr], %[offset_max]\n"
".inst 0xc0062c08 // mova { z8.d-z11.d }, za.d[x9, #0]\n"
"ld1rw { z0.s }, p1/Z, [x21]\n"
".inst 0xc0062c34 // mova { z20.d-z23.d }, za.d[x9, #1]\n"
"ld1rw { z6.s }, p1/Z, [x20]\n"
".inst 0xc1a6c808 // fclamp { z8.s-z11.s }, z0.s, z6.s\n"
".inst 0xa060c728 // st1w { z8.s-z11.s }, pn9.b, [x25]\n"
".inst 0xc1a6c814 // fclamp { z20.s-z23.s }, z0.s, z6.s\n"
".inst 0xa061c334 // st1w { z20.s-z23.s }, p8, [x25, #0x4, MUL VL]\n"
"addvl x25, x25, #8\n"
"b 19f\n"
"18:" // Width 2: No activation
".inst 0xc0062c08 // mova { z8.d-z11.d }, za.d[x9, #0]\n"
".inst 0xa060c728 // st1w { z8.s-z11.s }, pn9.b, [x25]\n"
".inst 0xc0062c34 // mova { z20.d-z23.d }, za.d[x9, #1]\n"
".inst 0xa061c334 // st1w { z20.s-z23.s }, p8, [x25, #0x4, MUL VL]\n"
"addvl x25, x25, #8\n"
"19:" // Width 2: Output done
"b 36f\n"
"20:" // Width 3
"mov x20, #0x2\n"
"mov x23, %x[A_ptr]\n"
"lsl x22, %x[K], #0x2\n"
"msub x20, x28, x20, %x[N]\n"
"mov x21, %x[K]\n"
".inst 0xf8b64af8 // rprfm pldmany, x22, [x23]\n"
".inst 0x25b467f0 // whilelt p8.s, XZR, x20, VLx4\n"
"cbz x24, 21f\n"
".inst 0xa040c700 // ld1w { z0.s-z3.s }, pn9.b/Z, [x24]\n"
".inst 0xc0042c00 // mova za.d[x9, #0], { z0.d-z3.d }\n"
".inst 0xa041c708 // ld1w { z8.s-z11.s }, pn9.b/Z, [x24, #0x4, MUL VL]\n"
".inst 0xc0042d01 // mova za.d[x9, #1], { z8.d-z11.d }\n"
".inst 0xa042c704 // ld1w { z4.s-z7.s }, pn9.b/Z, [x24, #0x8, MUL VL]\n"
".inst 0xc0042c82 // mova za.d[x9, #2], { z4.d-z7.d }\n"
"b 22f\n"
"21:" // Width 3: no bias
".inst 0xc00800ff // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
"22:" // Width 3: setup done
"cmp x21, #0x4\n"
"ble 24f\n"
"23:" // Width 3: Multiply loop: Main loop head
"whilelt p0.s, XZR, x21\n"
"ld1rqw { z10.s }, p0/Z, [x23]\n"
"sub x21, x21, #0x4\n"
".inst 0xa040c741 // ldnt1w { z0.s-z3.s }, pn9.b/Z, [x26]\n"
".inst 0xc15aa000 // fmla za.s[x9, 0], { z0.s-z3.s }, z10.s[0]\n"
"cmp x21, #0x4\n"
"add x23, x23, #0x10\n"
".inst 0xa041c745 // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
".inst 0xc15aa081 // fmla za.s[x9, 1], { z4.s-z7.s }, z10.s[0]\n"
".inst 0xa042c755 // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
".inst 0xc15aa282 // fmla za.s[x9, 2], { z20.s-z23.s }, z10.s[0]\n"
"addvl x26, x26, #16\n"
".inst 0xa040c759 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x26]\n"
".inst 0xc15aa700 // fmla za.s[x9, 0], { z24.s-z27.s }, z10.s[1]\n"
".inst 0xa041c751 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
".inst 0xc15aa601 // fmla za.s[x9, 1], { z16.s-z19.s }, z10.s[1]\n"
".inst 0xa042c759 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
".inst 0xc15aa702 // fmla za.s[x9, 2], { z24.s-z27.s }, z10.s[1]\n"
"addvl x26, x26, #16\n"
".inst 0xa040c74d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x26]\n"
".inst 0xc15aa980 // fmla za.s[x9, 0], { z12.s-z15.s }, z10.s[2]\n"
".inst 0xa041c751 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
".inst 0xc15aaa01 // fmla za.s[x9, 1], { z16.s-z19.s }, z10.s[2]\n"
".inst 0xa042c75d // ldnt1w { z28.s-z31.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
".inst 0xc15aab82 // fmla za.s[x9, 2], { z28.s-z31.s }, z10.s[2]\n"
"addvl x26, x26, #16\n"
".inst 0xa040c751 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26]\n"
".inst 0xc15aae00 // fmla za.s[x9, 0], { z16.s-z19.s }, z10.s[3]\n"
".inst 0xa041c759 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
".inst 0xc15aaf01 // fmla za.s[x9, 1], { z24.s-z27.s }, z10.s[3]\n"
".inst 0xa042c751 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
".inst 0xc15aae02 // fmla za.s[x9, 2], { z16.s-z19.s }, z10.s[3]\n"
"addvl x26, x26, #16\n"
"bgt 23b\n"
"24:" // Width 3: Multiply loop: Single iteration only
"whilelt p0.s, XZR, x21\n"
"ld1rqw { z10.s }, p0/Z, [x23]\n"
"subs x21, x21, #0x1\n"
".inst 0xa040c741 // ldnt1w { z0.s-z3.s }, pn9.b/Z, [x26]\n"
"add x23, x23, #0x10\n"
".inst 0xc15aa000 // fmla za.s[x9, 0], { z0.s-z3.s }, z10.s[0]\n"
".inst 0xa041c745 // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
".inst 0xc15aa081 // fmla za.s[x9, 1], { z4.s-z7.s }, z10.s[0]\n"
".inst 0xa042c755 // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
".inst 0xc15aa282 // fmla za.s[x9, 2], { z20.s-z23.s }, z10.s[0]\n"
"addvl x26, x26, #16\n"
"ble 25f\n"
".inst 0xa040c759 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x26]\n"
"subs x21, x21, #0x1\n"
".inst 0xc15aa700 // fmla za.s[x9, 0], { z24.s-z27.s }, z10.s[1]\n"
".inst 0xa041c751 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
".inst 0xc15aa601 // fmla za.s[x9, 1], { z16.s-z19.s }, z10.s[1]\n"
".inst 0xa042c759 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
".inst 0xc15aa702 // fmla za.s[x9, 2], { z24.s-z27.s }, z10.s[1]\n"
"addvl x26, x26, #16\n"
"ble 25f\n"
".inst 0xa040c74d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x26]\n"
"subs x21, x21, #0x1\n"
".inst 0xc15aa980 // fmla za.s[x9, 0], { z12.s-z15.s }, z10.s[2]\n"
".inst 0xa041c751 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
".inst 0xc15aaa01 // fmla za.s[x9, 1], { z16.s-z19.s }, z10.s[2]\n"
".inst 0xa042c75d // ldnt1w { z28.s-z31.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
".inst 0xc15aab82 // fmla za.s[x9, 2], { z28.s-z31.s }, z10.s[2]\n"
"addvl x26, x26, #16\n"
"ble 25f\n"
".inst 0xa040c751 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26]\n"
".inst 0xc15aae00 // fmla za.s[x9, 0], { z16.s-z19.s }, z10.s[3]\n"
".inst 0xa041c759 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
".inst 0xc15aaf01 // fmla za.s[x9, 1], { z24.s-z27.s }, z10.s[3]\n"
".inst 0xa042c751 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
".inst 0xc15aae02 // fmla za.s[x9, 2], { z16.s-z19.s }, z10.s[3]\n"
"addvl x26, x26, #16\n"
"25:" // Width 3: Multiply loop: multiply skip
"tbz %x[flags], #1, 26f\n"
"add x21, %x[args_ptr], %[offset_min]\n"
"add x20, %x[args_ptr], %[offset_max]\n"
".inst 0xc0062c08 // mova { z8.d-z11.d }, za.d[x9, #0]\n"
"ld1rw { z0.s }, p1/Z, [x21]\n"
".inst 0xc0062c34 // mova { z20.d-z23.d }, za.d[x9, #1]\n"
"ld1rw { z6.s }, p1/Z, [x20]\n"
".inst 0xc1a6c808 // fclamp { z8.s-z11.s }, z0.s, z6.s\n"
".inst 0xc0062c50 // mova { z16.d-z19.d }, za.d[x9, #2]\n"
".inst 0xa060c728 // st1w { z8.s-z11.s }, pn9.b, [x25]\n"
".inst 0xc1a6c814 // fclamp { z20.s-z23.s }, z0.s, z6.s\n"
".inst 0xa061c734 // st1w { z20.s-z23.s }, pn9.b, [x25, #0x4, MUL VL]\n"
".inst 0xc1a6c810 // fclamp { z16.s-z19.s }, z0.s, z6.s\n"
".inst 0xa062c330 // st1w { z16.s-z19.s }, p8, [x25, #0x8, MUL VL]\n"
"addvl x25, x25, #12\n"
"b 27f\n"
"26:" // Width 3: No activation
".inst 0xc0062c08 // mova { z8.d-z11.d }, za.d[x9, #0]\n"
".inst 0xa060c728 // st1w { z8.s-z11.s }, pn9.b, [x25]\n"
".inst 0xc0062c34 // mova { z20.d-z23.d }, za.d[x9, #1]\n"
".inst 0xa061c734 // st1w { z20.s-z23.s }, pn9.b, [x25, #0x4, MUL VL]\n"
".inst 0xc0062c50 // mova { z16.d-z19.d }, za.d[x9, #2]\n"
".inst 0xa062c330 // st1w { z16.s-z19.s }, p8, [x25, #0x8, MUL VL]\n"
"addvl x25, x25, #12\n"
"27:" // Width 3: Output done
"b 36f\n"
"28:" // Width 4
"mov x20, #0x3\n"
"mov x23, %x[A_ptr]\n"
"lsl x22, %x[K], #0x2\n"
"msub x20, x28, x20, %x[N]\n"
"mov x21, %x[K]\n"
".inst 0xf8b64af8 // rprfm pldmany, x22, [x23]\n"
".inst 0x25b467f0 // whilelt p8.s, XZR, x20, VLx4\n"
"cbz x24, 29f\n"
".inst 0xa040c700 // ld1w { z0.s-z3.s }, pn9.b/Z, [x24]\n"
".inst 0xc0042c00 // mova za.d[x9, #0], { z0.d-z3.d }\n"
".inst 0xa041c708 // ld1w { z8.s-z11.s }, pn9.b/Z, [x24, #0x4, MUL VL]\n"
".inst 0xc0042d01 // mova za.d[x9, #1], { z8.d-z11.d }\n"
".inst 0xa042c704 // ld1w { z4.s-z7.s }, pn9.b/Z, [x24, #0x8, MUL VL]\n"
".inst 0xc0042c82 // mova za.d[x9, #2], { z4.d-z7.d }\n"
".inst 0xa043c710 // ld1w { z16.s-z19.s }, pn9.b/Z, [x24, #0xc, MUL VL]\n"
".inst 0xc0042e03 // mova za.d[x9, #3], { z16.d-z19.d }\n"
"addvl x24, x24, #16\n"
"b 30f\n"
"29:" // Width 4: no bias
".inst 0xc00800ff // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
"30:" // Width 4: setup done
"cmp x21, #0x4\n"
"ble 32f\n"
"31:" // Width 4: Multiply loop: Main loop head
"whilelt p0.s, XZR, x21\n"
"ld1rqw { z10.s }, p0/Z, [x23]\n"
"sub x21, x21, #0x4\n"
".inst 0xa040c741 // ldnt1w { z0.s-z3.s }, pn9.b/Z, [x26]\n"
".inst 0xc15aa000 // fmla za.s[x9, 0], { z0.s-z3.s }, z10.s[0]\n"
"cmp x21, #0x4\n"
"add x23, x23, #0x10\n"
".inst 0xa041c745 // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
".inst 0xc15aa081 // fmla za.s[x9, 1], { z4.s-z7.s }, z10.s[0]\n"
".inst 0xa042c755 // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
".inst 0xc15aa282 // fmla za.s[x9, 2], { z20.s-z23.s }, z10.s[0]\n"
".inst 0xa043c751 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26, #0xc, MUL VL]\n"
".inst 0xc15aa203 // fmla za.s[x9, 3], { z16.s-z19.s }, z10.s[0]\n"
"addvl x26, x26, #16\n"
".inst 0xa040c759 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x26]\n"
".inst 0xc15aa700 // fmla za.s[x9, 0], { z24.s-z27.s }, z10.s[1]\n"
".inst 0xa041c751 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
".inst 0xc15aa601 // fmla za.s[x9, 1], { z16.s-z19.s }, z10.s[1]\n"
".inst 0xa042c759 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
".inst 0xc15aa702 // fmla za.s[x9, 2], { z24.s-z27.s }, z10.s[1]\n"
".inst 0xa043c751 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26, #0xc, MUL VL]\n"
".inst 0xc15aa603 // fmla za.s[x9, 3], { z16.s-z19.s }, z10.s[1]\n"
"addvl x26, x26, #16\n"
".inst 0xa040c74d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x26]\n"
".inst 0xc15aa980 // fmla za.s[x9, 0], { z12.s-z15.s }, z10.s[2]\n"
".inst 0xa041c751 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
".inst 0xc15aaa01 // fmla za.s[x9, 1], { z16.s-z19.s }, z10.s[2]\n"
".inst 0xa042c75d // ldnt1w { z28.s-z31.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
".inst 0xc15aab82 // fmla za.s[x9, 2], { z28.s-z31.s }, z10.s[2]\n"
".inst 0xa043c755 // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x26, #0xc, MUL VL]\n"
".inst 0xc15aaa83 // fmla za.s[x9, 3], { z20.s-z23.s }, z10.s[2]\n"
"addvl x26, x26, #16\n"
".inst 0xa040c751 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26]\n"
".inst 0xc15aae00 // fmla za.s[x9, 0], { z16.s-z19.s }, z10.s[3]\n"
".inst 0xa041c759 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
".inst 0xc15aaf01 // fmla za.s[x9, 1], { z24.s-z27.s }, z10.s[3]\n"
".inst 0xa042c751 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
".inst 0xc15aae02 // fmla za.s[x9, 2], { z16.s-z19.s }, z10.s[3]\n"
".inst 0xa043c751 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26, #0xc, MUL VL]\n"
".inst 0xc15aae03 // fmla za.s[x9, 3], { z16.s-z19.s }, z10.s[3]\n"
"addvl x26, x26, #16\n"
"bgt 31b\n"
"32:" // Width 4: Multiply loop: Single iteration only
"whilelt p0.s, XZR, x21\n"
"ld1rqw { z10.s }, p0/Z, [x23]\n"
"subs x21, x21, #0x1\n"
".inst 0xa040c741 // ldnt1w { z0.s-z3.s }, pn9.b/Z, [x26]\n"
"add x23, x23, #0x10\n"
".inst 0xc15aa000 // fmla za.s[x9, 0], { z0.s-z3.s }, z10.s[0]\n"
".inst 0xa041c745 // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
".inst 0xc15aa081 // fmla za.s[x9, 1], { z4.s-z7.s }, z10.s[0]\n"
".inst 0xa042c755 // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
".inst 0xc15aa282 // fmla za.s[x9, 2], { z20.s-z23.s }, z10.s[0]\n"
".inst 0xa043c751 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26, #0xc, MUL VL]\n"
".inst 0xc15aa203 // fmla za.s[x9, 3], { z16.s-z19.s }, z10.s[0]\n"
"addvl x26, x26, #16\n"
"ble 33f\n"
".inst 0xa040c759 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x26]\n"
"subs x21, x21, #0x1\n"
".inst 0xc15aa700 // fmla za.s[x9, 0], { z24.s-z27.s }, z10.s[1]\n"
".inst 0xa041c751 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
".inst 0xc15aa601 // fmla za.s[x9, 1], { z16.s-z19.s }, z10.s[1]\n"
".inst 0xa042c759 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
".inst 0xc15aa702 // fmla za.s[x9, 2], { z24.s-z27.s }, z10.s[1]\n"
".inst 0xa043c751 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26, #0xc, MUL VL]\n"
".inst 0xc15aa603 // fmla za.s[x9, 3], { z16.s-z19.s }, z10.s[1]\n"
"addvl x26, x26, #16\n"
"ble 33f\n"
".inst 0xa040c74d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x26]\n"
"subs x21, x21, #0x1\n"
".inst 0xc15aa980 // fmla za.s[x9, 0], { z12.s-z15.s }, z10.s[2]\n"
".inst 0xa041c751 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
".inst 0xc15aaa01 // fmla za.s[x9, 1], { z16.s-z19.s }, z10.s[2]\n"
".inst 0xa042c75d // ldnt1w { z28.s-z31.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
".inst 0xc15aab82 // fmla za.s[x9, 2], { z28.s-z31.s }, z10.s[2]\n"
".inst 0xa043c755 // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x26, #0xc, MUL VL]\n"
".inst 0xc15aaa83 // fmla za.s[x9, 3], { z20.s-z23.s }, z10.s[2]\n"
"addvl x26, x26, #16\n"
"ble 33f\n"
".inst 0xa040c751 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26]\n"
".inst 0xc15aae00 // fmla za.s[x9, 0], { z16.s-z19.s }, z10.s[3]\n"
".inst 0xa041c759 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n"
".inst 0xc15aaf01 // fmla za.s[x9, 1], { z24.s-z27.s }, z10.s[3]\n"
".inst 0xa042c751 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n"
".inst 0xc15aae02 // fmla za.s[x9, 2], { z16.s-z19.s }, z10.s[3]\n"
".inst 0xa043c751 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26, #0xc, MUL VL]\n"
".inst 0xc15aae03 // fmla za.s[x9, 3], { z16.s-z19.s }, z10.s[3]\n"
"addvl x26, x26, #16\n"
"33:" // Width 4: Multiply loop: multiply skip
"tbz %x[flags], #1, 34f\n"
"add x21, %x[args_ptr], %[offset_min]\n"
"add x20, %x[args_ptr], %[offset_max]\n"
".inst 0xc0062c08 // mova { z8.d-z11.d }, za.d[x9, #0]\n"
"ld1rw { z0.s }, p1/Z, [x21]\n"
".inst 0xc0062c34 // mova { z20.d-z23.d }, za.d[x9, #1]\n"
"ld1rw { z6.s }, p1/Z, [x20]\n"
".inst 0xc1a6c808 // fclamp { z8.s-z11.s }, z0.s, z6.s\n"
".inst 0xc0062c50 // mova { z16.d-z19.d }, za.d[x9, #2]\n"
".inst 0xa060c728 // st1w { z8.s-z11.s }, pn9.b, [x25]\n"
".inst 0xc1a6c814 // fclamp { z20.s-z23.s }, z0.s, z6.s\n"
".inst 0xc0062c78 // mova { z24.d-z27.d }, za.d[x9, #3]\n"
".inst 0xa061c734 // st1w { z20.s-z23.s }, pn9.b, [x25, #0x4, MUL VL]\n"
".inst 0xc1a6c810 // fclamp { z16.s-z19.s }, z0.s, z6.s\n"
".inst 0xa062c730 // st1w { z16.s-z19.s }, pn9.b, [x25, #0x8, MUL VL]\n"
".inst 0xc1a6c818 // fclamp { z24.s-z27.s }, z0.s, z6.s\n"
".inst 0xa063c338 // st1w { z24.s-z27.s }, p8, [x25, #0xc, MUL VL]\n"
"addvl x25, x25, #16\n"
"b 35f\n"
"34:" // Width 4: No activation
".inst 0xc0062c08 // mova { z8.d-z11.d }, za.d[x9, #0]\n"
".inst 0xa060c728 // st1w { z8.s-z11.s }, pn9.b, [x25]\n"
".inst 0xc0062c34 // mova { z20.d-z23.d }, za.d[x9, #1]\n"
".inst 0xa061c734 // st1w { z20.s-z23.s }, pn9.b, [x25, #0x4, MUL VL]\n"
".inst 0xc0062c50 // mova { z16.d-z19.d }, za.d[x9, #2]\n"
".inst 0xa062c730 // st1w { z16.s-z19.s }, pn9.b, [x25, #0x8, MUL VL]\n"
".inst 0xc0062c78 // mova { z24.d-z27.d }, za.d[x9, #3]\n"
".inst 0xa063c338 // st1w { z24.s-z27.s }, p8, [x25, #0xc, MUL VL]\n"
"addvl x25, x25, #16\n"
"35:" // Width 4: Output done
"subs x27, x27, #0x4\n"
"sub %x[N], %x[N], x28, LSL #2\n"
"bgt 4b\n"
"36:" // Exit
".inst 0xd503467f // SMSTOP\n"
"ptrue p1.b\n"
: [N] "+&r" (N)
: [A_ptr] "r" (A_ptr), [B_ptr] "r" (B_ptr), [K] "r" (K), [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [output_ptr] "r" (output_ptr)
: "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "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 arm_gemm
#endif // ARM_COMPUTE_ENABLE_SME2
#endif
|
