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
|
/*
* Copyright (c) 2022 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 "../../bfloat.hpp"
#include "../../asmlib.hpp"
#include "../../utils.hpp"
namespace arm_gemm {
void sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL(const bfloat16 *const A, const bfloat16 *const B, float *const C, int ldc, const int M, const int N, const int K, const float *const bias, const Activation act, bool accumulate, float *const accumulator_buffer)
{
struct KernelArgs
{
KernelArgs(
const bfloat16 *const A,
const bfloat16 *const B,
float *const C, const int ldc,
const int M, const int N, const int K,
const float *const bias,
const Activation act,
bool accumulate,
float *const accumulator_buffer
) : A(A),
B(B), kstride_bytes(roundup(K, 2) * sizeof(bfloat16)),
C(C), ldcb(ldc * sizeof(float)),
M(M), N(N), K(K),
n_loops(((K / 2) - 1) / 2), n_tail_iters(((K / 2) - 1) % 2),
min(-std::numeric_limits<float>::infinity()),
max(std::numeric_limits<float>::infinity()),
bias(bias),
accumulator_buffer(accumulator_buffer),
flags(0x0)
{
if (accumulate)
{
flags |= 1 << 0; // FILL_ACCUMULATORS_FROM_BUFFER
}
if (C == nullptr)
{
flags |= 1 << 1; // STORE_ACCUMULATORS_TO_BUFFER
}
if (act.type == Activation::Type::None)
{
flags |= 1 << 2; // SKIP_ACTIVATION
}
// Initialise the activation values
switch (act.type)
{
default:
case Activation::Type::None:
break;
case Activation::Type::BoundedReLU:
this->max = static_cast<float>(act.param1);
/* fall through */
case Activation::Type::ReLU:
this->min = static_cast<float>(0);
break;
}
}
const bfloat16 *const A;
const bfloat16 *const B;
const long kstride_bytes;
float *const C;
const long ldcb;
const long M, N, K, n_loops, n_tail_iters;
float min = -std::numeric_limits<float>::infinity();
float max = std::numeric_limits<float>::infinity();
const float *const bias;
float *const accumulator_buffer;
uint64_t flags;
};
// Construct arguments for this kernel
KernelArgs args(A, B, C, ldc, M, N, K, bias, act, accumulate, accumulator_buffer);
__asm__ __volatile__(
"ldr x15, [%x[args], %[offsetof_flags]]\n"
".inst 0xd503477f // SMSTART ZA\n"
"ptrue p1.b\n"
".inst 0x25207810 // ptrue pn8.b\n"
"ldr x14, [%x[args], %[offsetof_accumulator_buffer]]\n"
"ldr x13, [%x[args], %[offsetof_accumulator_buffer]]\n"
"tbz x15, #0, 2f\n"
"mov x12, #0x0\n"
"cntw x19\n"
"1:" // Initial accumulator load from buffer: Loop
".inst 0xa040c1c4 // ld1w { z4.s-z7.s }, pn8.b/Z, [x14]\n"
".inst 0xc0840480 // mova za0h.s[x12], { z4.s-z7.s }\n"
".inst 0xa041c1d8 // ld1w { z24.s-z27.s }, pn8.b/Z, [x14, #0x4, MUL VL]\n"
".inst 0xc0840701 // mova za1h.s[x12], { z24.s-z27.s }\n"
".inst 0xa042c1c0 // ld1w { z0.s-z3.s }, pn8.b/Z, [x14, #0x8, MUL VL]\n"
".inst 0xc0840402 // mova za2h.s[x12], { z0.s-z3.s }\n"
".inst 0xa043c1c4 // ld1w { z4.s-z7.s }, pn8.b/Z, [x14, #0xc, MUL VL]\n"
".inst 0xc0840483 // mova za3h.s[x12], { z4.s-z7.s }\n"
"add x12, x12, #0x4\n"
"cmp x12, x19\n"
"addvl x14, x14, #16\n"
"blt 1b\n"
"2:" // Initial accumulator load from buffer: End
"ldr w11, [%x[args], %[offsetof_M]]\n"
"mov x10, #0x0\n"
"mov x9, #0x0\n"
"ldr w28, [%x[args], %[offsetof_N]]\n"
"ldr x27, [%x[args], %[offsetof_A]]\n"
"3:" // M and N loop
"mov x26, x27\n"
"whilelt p0.s, x9, x28\n"
"tbnz x15, #0, 4f\n"
"ldr x19, [%x[args], %[offsetof_bias]]\n"
".inst 0xc00800ff // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
"cbz x19, 5f\n"
"fmov z8.s, #1.0\n"
"ldnt1w { z27.s }, p0/Z, [x19, x9, LSL #2]\n"
".inst 0x809b2500 // fmopa za0.s, p1/M, p1/M, z8.s, z27.s\n"
".inst 0x809b2501 // fmopa za1.s, p1/M, p1/M, z8.s, z27.s\n"
".inst 0x809b2502 // fmopa za2.s, p1/M, p1/M, z8.s, z27.s\n"
".inst 0x809b2503 // fmopa za3.s, p1/M, p1/M, z8.s, z27.s\n"
"4:" // Prepare accumulators: Test for last block
"mov x19, x9\n"
"mov x20, x10\n"
"incw x19\n"
"incw x20, ALL, MUL #4\n"
"cmp x19, x28\n"
"csel x20, x10, x20, LT\n"
"mov x19, x15\n"
"bfm x15, XZR, #0x0, #0x0 // bfc x15, #0x0, #0x1\n"
"cmp x20, x11\n"
"csel x15, x19, x15, LT\n"
"5:" // Prepare accumulators: End
"ldr x19, [%x[args], %[offsetof_K]]\n"
"add x19, x19, #0x1\n"
"lsr x19, x19, #0x1\n"
"ldr x22, [%x[args], %[offsetof_B]]\n"
"lsr x21, x19, #0x2\n"
"and x20, x19, #0x3\n"
"ldr x19, [%x[args], %[offsetof_kstride_bytes]]\n"
"madd x22, x9, x19, x22\n" // bptr = B + n * kstride_bytes
"cbz x21, 8f\n"
"subs x21, x21, #0x1\n"
".inst 0xa040a344 // ld1h { z4.h-z7.h }, pn8.b/Z, [x26]\n"
"ldnt1h { z29.h }, p1/Z, [x22]\n"
".inst 0xa041a34c // ld1h { z12.h-z15.h }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
"ldnt1h { z23.h }, p1/Z, [x22, #1, MUL VL]\n"
".inst 0xa042a340 // ld1h { z0.h-z3.h }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
"ldnt1h { z21.h }, p1/Z, [x22, #2, MUL VL]\n"
".inst 0xa143a352 // ld1h { z18.h, z22.h, z26.h, z30.h }, pn8.b/Z, [x26, #0xc, MUL VL]\n"
"addvl x26, x26, #16\n"
"ldnt1h { z27.h }, p1/Z, [x22, #3, MUL VL]\n"
"addvl x22, x22, #4\n"
"ble 7f\n"
"6:" // K loop
".inst 0x819d2480 // bfmopa za0.s, p1/M, p1/M, z4.h, z29.h\n"
"subs x21, x21, #0x1\n"
".inst 0x819d24a1 // bfmopa za1.s, p1/M, p1/M, z5.h, z29.h\n"
".inst 0x819d24c2 // bfmopa za2.s, p1/M, p1/M, z6.h, z29.h\n"
".inst 0x819d24e3 // bfmopa za3.s, p1/M, p1/M, z7.h, z29.h\n"
".inst 0xa040a344 // ld1h { z4.h-z7.h }, pn8.b/Z, [x26]\n"
".inst 0x81972580 // bfmopa za0.s, p1/M, p1/M, z12.h, z23.h\n"
"ldnt1h { z29.h }, p1/Z, [x22]\n"
".inst 0x819725a1 // bfmopa za1.s, p1/M, p1/M, z13.h, z23.h\n"
".inst 0x819725c2 // bfmopa za2.s, p1/M, p1/M, z14.h, z23.h\n"
".inst 0x819725e3 // bfmopa za3.s, p1/M, p1/M, z15.h, z23.h\n"
".inst 0xa041a34c // ld1h { z12.h-z15.h }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
".inst 0x81952400 // bfmopa za0.s, p1/M, p1/M, z0.h, z21.h\n"
"ldnt1h { z23.h }, p1/Z, [x22, #1, MUL VL]\n"
".inst 0x81952421 // bfmopa za1.s, p1/M, p1/M, z1.h, z21.h\n"
".inst 0x81952442 // bfmopa za2.s, p1/M, p1/M, z2.h, z21.h\n"
".inst 0x81952463 // bfmopa za3.s, p1/M, p1/M, z3.h, z21.h\n"
".inst 0xa042a340 // ld1h { z0.h-z3.h }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
"ldnt1h { z21.h }, p1/Z, [x22, #2, MUL VL]\n"
".inst 0x819b2640 // bfmopa za0.s, p1/M, p1/M, z18.h, z27.h\n"
".inst 0x819b26c1 // bfmopa za1.s, p1/M, p1/M, z22.h, z27.h\n"
".inst 0x819b2742 // bfmopa za2.s, p1/M, p1/M, z26.h, z27.h\n"
".inst 0x819b27c3 // bfmopa za3.s, p1/M, p1/M, z30.h, z27.h\n"
".inst 0xa143a352 // ld1h { z18.h, z22.h, z26.h, z30.h }, pn8.b/Z, [x26, #0xc, MUL VL]\n"
"addvl x26, x26, #16\n"
"ldnt1h { z27.h }, p1/Z, [x22, #3, MUL VL]\n"
"addvl x22, x22, #4\n"
"bgt 6b\n"
"7:" // K loop tail
".inst 0x819d2480 // bfmopa za0.s, p1/M, p1/M, z4.h, z29.h\n"
".inst 0x819d24a1 // bfmopa za1.s, p1/M, p1/M, z5.h, z29.h\n"
".inst 0x819d24c2 // bfmopa za2.s, p1/M, p1/M, z6.h, z29.h\n"
".inst 0x819d24e3 // bfmopa za3.s, p1/M, p1/M, z7.h, z29.h\n"
".inst 0x81972580 // bfmopa za0.s, p1/M, p1/M, z12.h, z23.h\n"
".inst 0x819725a1 // bfmopa za1.s, p1/M, p1/M, z13.h, z23.h\n"
".inst 0x819725c2 // bfmopa za2.s, p1/M, p1/M, z14.h, z23.h\n"
".inst 0x819725e3 // bfmopa za3.s, p1/M, p1/M, z15.h, z23.h\n"
".inst 0x81952400 // bfmopa za0.s, p1/M, p1/M, z0.h, z21.h\n"
".inst 0x81952421 // bfmopa za1.s, p1/M, p1/M, z1.h, z21.h\n"
".inst 0x81952442 // bfmopa za2.s, p1/M, p1/M, z2.h, z21.h\n"
".inst 0x81952463 // bfmopa za3.s, p1/M, p1/M, z3.h, z21.h\n"
".inst 0x819b2640 // bfmopa za0.s, p1/M, p1/M, z18.h, z27.h\n"
".inst 0x819b26c1 // bfmopa za1.s, p1/M, p1/M, z22.h, z27.h\n"
".inst 0x819b2742 // bfmopa za2.s, p1/M, p1/M, z26.h, z27.h\n"
".inst 0x819b27c3 // bfmopa za3.s, p1/M, p1/M, z30.h, z27.h\n"
"8:" // K oddments
"cbz x20, 10f\n"
"9:" // K oddments: Loop
".inst 0xa040a344 // ld1h { z4.h-z7.h }, pn8.b/Z, [x26]\n"
"subs x20, x20, #0x1\n"
"addvl x26, x26, #4\n"
"ld1h { z29.h }, p1/Z, [x22]\n"
"addvl x22, x22, #1\n"
".inst 0x819d2480 // bfmopa za0.s, p1/M, p1/M, z4.h, z29.h\n"
".inst 0x819d24a1 // bfmopa za1.s, p1/M, p1/M, z5.h, z29.h\n"
".inst 0x819d24c2 // bfmopa za2.s, p1/M, p1/M, z6.h, z29.h\n"
".inst 0x819d24e3 // bfmopa za3.s, p1/M, p1/M, z7.h, z29.h\n"
"bgt 9b\n"
"10:" // K oddments: End
"tbz x15, #1, 14f\n"
"tbz x15, #0, 12f\n"
"mov x12, #0x0\n"
"cntw x19\n"
"11:" // Store to partial result buffer: Store and refill: Loop
".inst 0xa040c1c8 // ld1w { z8.s-z11.s }, pn8.b/Z, [x14]\n"
".inst 0xc0860418 // mova { z24.s-z27.s }, za0h.s[x12]\n"
".inst 0xc0840500 // mova za0h.s[x12], { z8.s-z11.s }\n"
".inst 0xc0860424 // mova { z4.s-z7.s }, za1h.s[x12]\n"
".inst 0xa041c1cc // ld1w { z12.s-z15.s }, pn8.b/Z, [x14, #0x4, MUL VL]\n"
".inst 0xc0840581 // mova za1h.s[x12], { z12.s-z15.s }\n"
".inst 0xc086044c // mova { z12.s-z15.s }, za2h.s[x12]\n"
".inst 0xc0860460 // mova { z0.s-z3.s }, za3h.s[x12]\n"
".inst 0xa042c1c8 // ld1w { z8.s-z11.s }, pn8.b/Z, [x14, #0x8, MUL VL]\n"
".inst 0xc0840502 // mova za2h.s[x12], { z8.s-z11.s }\n"
".inst 0xa043c1dc // ld1w { z28.s-z31.s }, pn8.b/Z, [x14, #0xc, MUL VL]\n"
".inst 0xc0840783 // mova za3h.s[x12], { z28.s-z31.s }\n"
"add x12, x12, #0x4\n"
"cmp x12, x19\n"
".inst 0xa060c1b8 // st1w { z24.s-z27.s }, pn8.b, [x13]\n"
"addvl x14, x14, #16\n"
".inst 0xa061c1a4 // st1w { z4.s-z7.s }, pn8.b, [x13, #0x4, MUL VL]\n"
".inst 0xa062c1ac // st1w { z12.s-z15.s }, pn8.b, [x13, #0x8, MUL VL]\n"
".inst 0xa063c1a0 // st1w { z0.s-z3.s }, pn8.b, [x13, #0xc, MUL VL]\n"
"addvl x13, x13, #16\n"
"blt 11b\n"
"b 42f\n"
"12:" // Store to partial result buffer: Store only
"mov x12, #0x0\n"
"cntw x19\n"
"13:" // Store to partial result buffer: Store only: Loop
".inst 0xc086040c // mova { z12.s-z15.s }, za0h.s[x12]\n"
".inst 0xc0860438 // mova { z24.s-z27.s }, za1h.s[x12]\n"
".inst 0xa060c1ac // st1w { z12.s-z15.s }, pn8.b, [x13]\n"
".inst 0xc0860440 // mova { z0.s-z3.s }, za2h.s[x12]\n"
".inst 0xc0860468 // mova { z8.s-z11.s }, za3h.s[x12]\n"
".inst 0xa061c1b8 // st1w { z24.s-z27.s }, pn8.b, [x13, #0x4, MUL VL]\n"
"add x12, x12, #0x4\n"
"cmp x12, x19\n"
".inst 0xa062c1a0 // st1w { z0.s-z3.s }, pn8.b, [x13, #0x8, MUL VL]\n"
".inst 0xa063c1a8 // st1w { z8.s-z11.s }, pn8.b, [x13, #0xc, MUL VL]\n"
"addvl x13, x13, #16\n"
"blt 13b\n"
"b 42f\n"
"14:" // Store to output array
"ldr x25, [%x[args], %[offsetof_C]]\n"
"add x25, x25, x9, LSL #2\n" // C += n
"sub x24, x11, x10\n"
"ldr x23, [%x[args], %[offsetof_ldcb]]\n"
"madd x25, x10, x23, x25\n" // C += m * ldc
"tbz x15, #2, 27f\n"
"cntw x22\n"
"cmp x24, x22\n"
"csel x21, x24, x22, LT\n"
"lsr x20, x21, #0x2\n"
"mov x12, #0x0\n"
"and x19, x21, #0x3\n"
"cbz x20, 16f\n"
"15:" // Store to output array: Skip activation: Accumulator row 0 loop
".inst 0xc0860410 // mova { z16.s-z19.s }, za0h.s[x12]\n"
"st1w { z16.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"st1w { z17.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"add x12, x12, #0x4\n"
"st1w { z18.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"cmp x12, x20, LSL #2\n"
"st1w { z19.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"blt 15b\n"
"16:" // Store to output array: Skip activation: Accumulator row 0 oddments
"cbz x19, 17f\n"
"subs x19, x19, #0x1\n"
".inst 0xc0860404 // mova { z4.s-z7.s }, za0h.s[x12]\n"
"st1w { z4.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"beq 17f\n"
"subs x19, x19, #0x1\n"
"st1w { z5.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"beq 17f\n"
"st1w { z6.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"17:" // Store to output array: Skip activation: Accumulator row 0 oddments: End
"subs x24, x24, x21\n"
"beq 27f\n"
"cmp x24, x22\n"
"csel x21, x24, x22, LT\n"
"lsr x20, x21, #0x2\n"
"mov x12, #0x0\n"
"and x19, x21, #0x3\n"
"cbz x20, 19f\n"
"18:" // Store to output array: Skip activation: Accumulator row 1 loop
".inst 0xc0860430 // mova { z16.s-z19.s }, za1h.s[x12]\n"
"st1w { z16.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"st1w { z17.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"add x12, x12, #0x4\n"
"st1w { z18.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"cmp x12, x20, LSL #2\n"
"st1w { z19.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"blt 18b\n"
"19:" // Store to output array: Skip activation: Accumulator row 1 oddments
"cbz x19, 20f\n"
"subs x19, x19, #0x1\n"
".inst 0xc0860424 // mova { z4.s-z7.s }, za1h.s[x12]\n"
"st1w { z4.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"beq 20f\n"
"subs x19, x19, #0x1\n"
"st1w { z5.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"beq 20f\n"
"st1w { z6.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"20:" // Store to output array: Skip activation: Accumulator row 1 oddments: End
"subs x24, x24, x21\n"
"beq 27f\n"
"cmp x24, x22\n"
"csel x21, x24, x22, LT\n"
"lsr x20, x21, #0x2\n"
"mov x12, #0x0\n"
"and x19, x21, #0x3\n"
"cbz x20, 22f\n"
"21:" // Store to output array: Skip activation: Accumulator row 2 loop
".inst 0xc0860450 // mova { z16.s-z19.s }, za2h.s[x12]\n"
"st1w { z16.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"st1w { z17.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"add x12, x12, #0x4\n"
"st1w { z18.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"cmp x12, x20, LSL #2\n"
"st1w { z19.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"blt 21b\n"
"22:" // Store to output array: Skip activation: Accumulator row 2 oddments
"cbz x19, 23f\n"
"subs x19, x19, #0x1\n"
".inst 0xc0860454 // mova { z20.s-z23.s }, za2h.s[x12]\n"
"st1w { z20.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"beq 23f\n"
"subs x19, x19, #0x1\n"
"st1w { z21.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"beq 23f\n"
"st1w { z22.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"23:" // Store to output array: Skip activation: Accumulator row 2 oddments: End
"subs x24, x24, x21\n"
"beq 27f\n"
"cmp x24, x22\n"
"csel x21, x24, x22, LT\n"
"lsr x20, x21, #0x2\n"
"mov x12, #0x0\n"
"and x19, x21, #0x3\n"
"cbz x20, 25f\n"
"24:" // Store to output array: Skip activation: Accumulator row 3 loop
".inst 0xc0860464 // mova { z4.s-z7.s }, za3h.s[x12]\n"
"st1w { z4.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"st1w { z5.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"add x12, x12, #0x4\n"
"st1w { z6.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"cmp x12, x20, LSL #2\n"
"st1w { z7.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"blt 24b\n"
"25:" // Store to output array: Skip activation: Accumulator row 3 oddments
"cbz x19, 26f\n"
"subs x19, x19, #0x1\n"
".inst 0xc086046c // mova { z12.s-z15.s }, za3h.s[x12]\n"
"st1w { z12.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"beq 26f\n"
"subs x19, x19, #0x1\n"
"st1w { z13.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"beq 26f\n"
"st1w { z14.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"26:" // Store to output array: Skip activation: Accumulator row 3 oddments: End
"subs x24, x24, x21\n"
"beq 27f\n"
"b 40f\n"
"27:" // Store to output array: Skip activation: End
"cntw x22\n"
"cmp x24, x22\n"
"ld1rw { z25.s }, p1/Z, [%x[args], %[offsetof_KernelArgs_min]]\n"
"csel x21, x24, x22, LT\n"
"lsr x20, x21, #0x2\n"
"ld1rw { z24.s }, p1/Z, [%x[args], %[offsetof_KernelArgs_max]]\n"
"mov x12, #0x0\n"
"and x19, x21, #0x3\n"
"cbz x20, 29f\n"
"28:" // Store to output array: Accumulator row 0 loop
".inst 0xc0860414 // mova { z20.s-z23.s }, za0h.s[x12]\n"
".inst 0xc1b8cb34 // fclamp { z20.s-z23.s }, z25.s, z24.s\n"
"st1w { z20.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"st1w { z21.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"add x12, x12, #0x4\n"
"st1w { z22.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"cmp x12, x20, LSL #2\n"
"st1w { z23.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"blt 28b\n"
"29:" // Store to output array: Accumulator row 0 oddments
"cbz x19, 30f\n"
".inst 0xc0860408 // mova { z8.s-z11.s }, za0h.s[x12]\n"
"subs x19, x19, #0x1\n"
".inst 0xc1b8cb28 // fclamp { z8.s-z11.s }, z25.s, z24.s\n"
"st1w { z8.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"beq 30f\n"
"subs x19, x19, #0x1\n"
"st1w { z9.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"beq 30f\n"
"st1w { z10.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"30:" // Store to output array: Accumulator row 0 oddments: End
"subs x24, x24, x21\n"
"beq 40f\n"
"cmp x24, x22\n"
"csel x21, x24, x22, LT\n"
"lsr x20, x21, #0x2\n"
"mov x12, #0x0\n"
"and x19, x21, #0x3\n"
"cbz x20, 32f\n"
"31:" // Store to output array: Accumulator row 1 loop
".inst 0xc0860430 // mova { z16.s-z19.s }, za1h.s[x12]\n"
".inst 0xc1b8cb30 // fclamp { z16.s-z19.s }, z25.s, z24.s\n"
"st1w { z16.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"st1w { z17.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"add x12, x12, #0x4\n"
"st1w { z18.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"cmp x12, x20, LSL #2\n"
"st1w { z19.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"blt 31b\n"
"32:" // Store to output array: Accumulator row 1 oddments
"cbz x19, 33f\n"
".inst 0xc0860430 // mova { z16.s-z19.s }, za1h.s[x12]\n"
"subs x19, x19, #0x1\n"
".inst 0xc1b8cb30 // fclamp { z16.s-z19.s }, z25.s, z24.s\n"
"st1w { z16.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"beq 33f\n"
"subs x19, x19, #0x1\n"
"st1w { z17.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"beq 33f\n"
"st1w { z18.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"33:" // Store to output array: Accumulator row 1 oddments: End
"subs x24, x24, x21\n"
"beq 40f\n"
"cmp x24, x22\n"
"csel x21, x24, x22, LT\n"
"lsr x20, x21, #0x2\n"
"mov x12, #0x0\n"
"and x19, x21, #0x3\n"
"cbz x20, 35f\n"
"34:" // Store to output array: Accumulator row 2 loop
".inst 0xc0860450 // mova { z16.s-z19.s }, za2h.s[x12]\n"
".inst 0xc1b8cb30 // fclamp { z16.s-z19.s }, z25.s, z24.s\n"
"st1w { z16.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"st1w { z17.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"add x12, x12, #0x4\n"
"st1w { z18.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"cmp x12, x20, LSL #2\n"
"st1w { z19.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"blt 34b\n"
"35:" // Store to output array: Accumulator row 2 oddments
"cbz x19, 36f\n"
".inst 0xc0860450 // mova { z16.s-z19.s }, za2h.s[x12]\n"
"subs x19, x19, #0x1\n"
".inst 0xc1b8cb30 // fclamp { z16.s-z19.s }, z25.s, z24.s\n"
"st1w { z16.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"beq 36f\n"
"subs x19, x19, #0x1\n"
"st1w { z17.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"beq 36f\n"
"st1w { z18.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"36:" // Store to output array: Accumulator row 2 oddments: End
"subs x24, x24, x21\n"
"beq 40f\n"
"cmp x24, x22\n"
"csel x19, x24, x22, LT\n"
"lsr x20, x19, #0x2\n"
"mov x12, #0x0\n"
"and x19, x19, #0x3\n"
"cbz x20, 38f\n"
"37:" // Store to output array: Accumulator row 3 loop
".inst 0xc0860474 // mova { z20.s-z23.s }, za3h.s[x12]\n"
".inst 0xc1b8cb34 // fclamp { z20.s-z23.s }, z25.s, z24.s\n"
"st1w { z20.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"st1w { z21.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"add x12, x12, #0x4\n"
"st1w { z22.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"cmp x12, x20, LSL #2\n"
"st1w { z23.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"blt 37b\n"
"38:" // Store to output array: Accumulator row 3 oddments
"cbz x19, 39f\n"
".inst 0xc0860470 // mova { z16.s-z19.s }, za3h.s[x12]\n"
"subs x19, x19, #0x1\n"
".inst 0xc1b8cb30 // fclamp { z16.s-z19.s }, z25.s, z24.s\n"
"st1w { z16.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"beq 39f\n"
"subs x19, x19, #0x1\n"
"st1w { z17.s }, p0, [x25]\n"
"add x25, x25, x23\n"
"beq 39f\n"
"st1w { z18.s }, p0, [x25]\n"
"39:" // Store to output array: Accumulator row 3 oddments: End
"40:" // Store to output array: End
"tbz x15, #0, 42f\n"
"mov x12, #0x0\n"
"cntw x19\n"
"41:" // Store to output array: Refill accumulators: Loop
".inst 0xa040c1d0 // ld1w { z16.s-z19.s }, pn8.b/Z, [x14]\n"
".inst 0xc0840600 // mova za0h.s[x12], { z16.s-z19.s }\n"
".inst 0xa041c1d0 // ld1w { z16.s-z19.s }, pn8.b/Z, [x14, #0x4, MUL VL]\n"
".inst 0xc0840601 // mova za1h.s[x12], { z16.s-z19.s }\n"
".inst 0xa042c1d4 // ld1w { z20.s-z23.s }, pn8.b/Z, [x14, #0x8, MUL VL]\n"
".inst 0xc0840682 // mova za2h.s[x12], { z20.s-z23.s }\n"
".inst 0xa043c1c4 // ld1w { z4.s-z7.s }, pn8.b/Z, [x14, #0xc, MUL VL]\n"
".inst 0xc0840483 // mova za3h.s[x12], { z4.s-z7.s }\n"
"add x12, x12, #0x4\n"
"cmp x12, x19\n"
"addvl x14, x14, #16\n"
"blt 41b\n"
"42:" // End block
"incw x9\n"
"cmp x9, x28\n"
"blt 3b\n"
"incw x10, ALL, MUL #4\n"
"cmp x10, x11\n"
"mov x9, #0x0\n"
"mov x27, x26\n"
"blt 3b\n"
".inst 0xd503467f // SMSTOP\n"
:
: [args] "r" (&args), [offsetof_A] "I" (offsetof(KernelArgs, A)), [offsetof_B] "I" (offsetof(KernelArgs, B)), [offsetof_C] "I" (offsetof(KernelArgs, C)), [offsetof_K] "I" (offsetof(KernelArgs, K)), [offsetof_KernelArgs_max] "I" (offsetof(KernelArgs, max)), [offsetof_KernelArgs_min] "I" (offsetof(KernelArgs, min)), [offsetof_M] "I" (offsetof(KernelArgs, M)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_accumulator_buffer] "I" (offsetof(KernelArgs, accumulator_buffer)), [offsetof_bias] "I" (offsetof(KernelArgs, bias)), [offsetof_flags] "I" (offsetof(KernelArgs, flags)), [offsetof_kstride_bytes] "I" (offsetof(KernelArgs, kstride_bytes)), [offsetof_ldcb] "I" (offsetof(KernelArgs, ldcb))
: "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x19", "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 // __ARM_FEATURE_SVE
|
