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
|
/*
* Copyright (c) 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
#include "arm_gemm.hpp"
#include "../../asmlib.hpp"
#include "../../utils.hpp"
namespace arm_gemm {
void sme2_interleaved_nomerge_fp16fp32fp16_mopa_4VLx1VL(const __fp16 *const A, const __fp16 *const B, __fp16 *const C, int ldc, const int M, const int N, const int K, const __fp16 *const bias, const Activation act, bool accumulate, float *const accumulator_buffer)
{
struct KernelArgs
{
KernelArgs(
const __fp16 *const A,
const __fp16 *const B,
__fp16 *const C, const int ldc,
const int M, const int N, const int K,
const __fp16 *const bias,
const Activation act,
bool accumulate,
float *const accumulator_buffer
) : A(A),
B(B), kstride_bytes(roundup(K, 2) * sizeof(__fp16)),
C(C), ldcb(ldc * sizeof(__fp16)),
M(M), N(N), K(K),
min(-static_cast<__fp16>(std::numeric_limits<float>::infinity())),
max(static_cast<__fp16>(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
}
// Initialise the activation values
switch (act.type)
{
default:
case Activation::Type::None:
break;
case Activation::Type::BoundedReLU:
this->max = static_cast<__fp16>(act.param1);
/* fall through */
case Activation::Type::ReLU:
this->min = static_cast<__fp16>(0);
break;
}
}
const __fp16 *const A;
const __fp16 *const B;
const long kstride_bytes;
__fp16 *const C;
const long ldcb;
const long M, N, K;
__fp16 min = -static_cast<__fp16>(std::numeric_limits<float>::infinity());
__fp16 max = static_cast<__fp16>(std::numeric_limits<float>::infinity());
const __fp16 *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 x16, [%x[args], %[offsetof_flags]]\n"
".inst 0xd503477f // SMSTART ZA\n"
"ptrue p1.b\n"
".inst 0x25207811 // ptrue pn9.b\n"
"ldr x15, [%x[args], %[offsetof_accumulator_buffer]]\n"
"ldr x14, [%x[args], %[offsetof_accumulator_buffer]]\n"
"tbz x16, #0, 2f\n"
"mov x12, #0x0\n"
"cntw x20\n"
"1:" // Initial accumulator load from buffer: Loop
".inst 0xa040c5f8 // ld1w { z24.s-z27.s }, pn9.b/Z, [x15]\n"
".inst 0xa041c5e4 // ld1w { z4.s-z7.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n"
".inst 0xa042c5e8 // ld1w { z8.s-z11.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n"
".inst 0xa043c5f4 // ld1w { z20.s-z23.s }, pn9.b/Z, [x15, #0xc, MUL VL]\n"
".inst 0xc0840700 // mova za0h.s[x12], { z24.s-z27.s }\n"
"addvl x15, x15, #16\n"
".inst 0xc0840481 // mova za1h.s[x12], { z4.s-z7.s }\n"
".inst 0xc0840502 // mova za2h.s[x12], { z8.s-z11.s }\n"
".inst 0xc0840683 // mova za3h.s[x12], { z20.s-z23.s }\n"
"add x12, x12, #0x4\n"
"cmp x12, x20\n"
"blt 1b\n"
"2:" // Initial accumulator load from buffer: End
"ldr w13, [%x[args], %[offsetof_M]]\n"
"mov x11, #0x0\n"
"mov x10, #0x0\n"
"ldr w9, [%x[args], %[offsetof_N]]\n"
"ldr x28, [%x[args], %[offsetof_A]]\n"
"3:" // M and N loop
"mov x27, x28\n"
"whilelt p8.s, x10, x9\n"
"tbnz x16, #0, 4f\n"
"ldr x20, [%x[args], %[offsetof_bias]]\n"
".inst 0xc00800ff // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
"cbz x20, 5f\n"
"whilelt p0.h, x10, x9\n"
"fmov z5.h, #0.0\n"
"fmov z18.h, #1.0\n"
"ld1h { z31.h }, p0/Z, [x20, x10, LSL #1]\n"
"zip1 z15.h, z31.h, z5.h\n"
".inst 0x81af2640 // fmopa za0.s, p1/M, p1/M, z18.h, z15.h\n"
".inst 0x81af2641 // fmopa za1.s, p1/M, p1/M, z18.h, z15.h\n"
".inst 0x81af2642 // fmopa za2.s, p1/M, p1/M, z18.h, z15.h\n"
".inst 0x81af2643 // fmopa za3.s, p1/M, p1/M, z18.h, z15.h\n"
"4:" // Prepare accumulators: Test for last block
"mov x20, x10\n"
"mov x21, x11\n"
"incw x20\n"
"incw x21, ALL, MUL #4\n"
"cmp x20, x9\n"
"mov x20, x16\n"
"csel x21, x11, x21, LT\n"
"bfm x16, XZR, #0x0, #0x0 // bfc x16, #0x0, #0x1\n"
"cmp x21, x13\n"
"csel x16, x20, x16, LT\n"
"5:" // Prepare accumulators: End
"ldr x20, [%x[args], %[offsetof_K]]\n"
"ldr x23, [%x[args], %[offsetof_B]]\n"
"ldr x22, [%x[args], %[offsetof_kstride_bytes]]\n"
"add x20, x20, #0x1\n"
"lsr x20, x20, #0x1\n"
"lsr x21, x20, #0x2\n"
"and x20, x20, #0x3\n"
"madd x23, x10, x22, x23\n" // bptr = B + n * kstride_bytes
"cbz x21, 8f\n"
"subs x21, x21, #0x1\n"
".inst 0xa140a773 // ld1h { z19.h, z23.h, z27.h, z31.h }, pn9.b/Z, [x27]\n"
"ldnt1h { z17.h }, p1/Z, [x23]\n"
".inst 0xa041a76c // ld1h { z12.h-z15.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
"ldnt1h { z26.h }, p1/Z, [x23, #1, MUL VL]\n"
".inst 0xa042a760 // ld1h { z0.h-z3.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
"ldnt1h { z30.h }, p1/Z, [x23, #2, MUL VL]\n"
".inst 0xa143a770 // ld1h { z16.h, z20.h, z24.h, z28.h }, pn9.b/Z, [x27, #0xc, MUL VL]\n"
"addvl x27, x27, #16\n"
"ldnt1h { z18.h }, p1/Z, [x23, #3, MUL VL]\n"
"addvl x23, x23, #4\n"
"ble 7f\n"
"6:" // K loop
".inst 0x81b12660 // fmopa za0.s, p1/M, p1/M, z19.h, z17.h\n"
"subs x21, x21, #0x1\n"
".inst 0x81b126e1 // fmopa za1.s, p1/M, p1/M, z23.h, z17.h\n"
".inst 0x81b12762 // fmopa za2.s, p1/M, p1/M, z27.h, z17.h\n"
".inst 0x81b127e3 // fmopa za3.s, p1/M, p1/M, z31.h, z17.h\n"
".inst 0xa140a773 // ld1h { z19.h, z23.h, z27.h, z31.h }, pn9.b/Z, [x27]\n"
".inst 0x81ba2580 // fmopa za0.s, p1/M, p1/M, z12.h, z26.h\n"
"ld1h { z17.h }, p1/Z, [x23]\n"
".inst 0x81ba25a1 // fmopa za1.s, p1/M, p1/M, z13.h, z26.h\n"
".inst 0x81ba25c2 // fmopa za2.s, p1/M, p1/M, z14.h, z26.h\n"
".inst 0x81ba25e3 // fmopa za3.s, p1/M, p1/M, z15.h, z26.h\n"
".inst 0xa041a76c // ld1h { z12.h-z15.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
".inst 0x81be2400 // fmopa za0.s, p1/M, p1/M, z0.h, z30.h\n"
"ld1h { z26.h }, p1/Z, [x23, #1, MUL VL]\n"
".inst 0x81be2421 // fmopa za1.s, p1/M, p1/M, z1.h, z30.h\n"
".inst 0x81be2442 // fmopa za2.s, p1/M, p1/M, z2.h, z30.h\n"
".inst 0x81be2463 // fmopa za3.s, p1/M, p1/M, z3.h, z30.h\n"
".inst 0xa042a760 // ld1h { z0.h-z3.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n"
"ld1h { z30.h }, p1/Z, [x23, #2, MUL VL]\n"
".inst 0x81b22600 // fmopa za0.s, p1/M, p1/M, z16.h, z18.h\n"
".inst 0x81b22681 // fmopa za1.s, p1/M, p1/M, z20.h, z18.h\n"
".inst 0x81b22702 // fmopa za2.s, p1/M, p1/M, z24.h, z18.h\n"
".inst 0x81b22783 // fmopa za3.s, p1/M, p1/M, z28.h, z18.h\n"
".inst 0xa143a770 // ld1h { z16.h, z20.h, z24.h, z28.h }, pn9.b/Z, [x27, #0xc, MUL VL]\n"
"addvl x27, x27, #16\n"
"ld1h { z18.h }, p1/Z, [x23, #3, MUL VL]\n"
"addvl x23, x23, #4\n"
"bgt 6b\n"
"7:" // K loop tail
".inst 0x81b12660 // fmopa za0.s, p1/M, p1/M, z19.h, z17.h\n"
".inst 0x81b126e1 // fmopa za1.s, p1/M, p1/M, z23.h, z17.h\n"
".inst 0x81b12762 // fmopa za2.s, p1/M, p1/M, z27.h, z17.h\n"
".inst 0x81b127e3 // fmopa za3.s, p1/M, p1/M, z31.h, z17.h\n"
".inst 0x81ba2580 // fmopa za0.s, p1/M, p1/M, z12.h, z26.h\n"
".inst 0x81ba25a1 // fmopa za1.s, p1/M, p1/M, z13.h, z26.h\n"
".inst 0x81ba25c2 // fmopa za2.s, p1/M, p1/M, z14.h, z26.h\n"
".inst 0x81ba25e3 // fmopa za3.s, p1/M, p1/M, z15.h, z26.h\n"
".inst 0x81be2400 // fmopa za0.s, p1/M, p1/M, z0.h, z30.h\n"
".inst 0x81be2421 // fmopa za1.s, p1/M, p1/M, z1.h, z30.h\n"
".inst 0x81be2442 // fmopa za2.s, p1/M, p1/M, z2.h, z30.h\n"
".inst 0x81be2463 // fmopa za3.s, p1/M, p1/M, z3.h, z30.h\n"
".inst 0x81b22600 // fmopa za0.s, p1/M, p1/M, z16.h, z18.h\n"
".inst 0x81b22681 // fmopa za1.s, p1/M, p1/M, z20.h, z18.h\n"
".inst 0x81b22702 // fmopa za2.s, p1/M, p1/M, z24.h, z18.h\n"
".inst 0x81b22783 // fmopa za3.s, p1/M, p1/M, z28.h, z18.h\n"
"8:" // K oddments
"cbz x20, 10f\n"
"9:" // K oddments: Loop
".inst 0xa140a773 // ld1h { z19.h, z23.h, z27.h, z31.h }, pn9.b/Z, [x27]\n"
"subs x20, x20, #0x1\n"
"addvl x27, x27, #4\n"
"ld1h { z17.h }, p1/Z, [x23]\n"
"addvl x23, x23, #1\n"
".inst 0x81b12660 // fmopa za0.s, p1/M, p1/M, z19.h, z17.h\n"
".inst 0x81b126e1 // fmopa za1.s, p1/M, p1/M, z23.h, z17.h\n"
".inst 0x81b12762 // fmopa za2.s, p1/M, p1/M, z27.h, z17.h\n"
".inst 0x81b127e3 // fmopa za3.s, p1/M, p1/M, z31.h, z17.h\n"
"bgt 9b\n"
"10:" // K oddments: End
"tbz x16, #1, 14f\n"
"tbz x16, #0, 12f\n"
"mov x12, #0x0\n"
"cntw x20\n"
"11:" // Store to partial result buffer: Store and refill: Loop
".inst 0xa040c5e0 // ld1w { z0.s-z3.s }, pn9.b/Z, [x15]\n"
".inst 0xc0860404 // mova { z4.s-z7.s }, za0h.s[x12]\n"
".inst 0xc0860434 // mova { z20.s-z23.s }, za1h.s[x12]\n"
".inst 0xa041c5ec // ld1w { z12.s-z15.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n"
".inst 0xc086045c // mova { z28.s-z31.s }, za2h.s[x12]\n"
".inst 0xc0860478 // mova { z24.s-z27.s }, za3h.s[x12]\n"
".inst 0xa042c5e8 // ld1w { z8.s-z11.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n"
".inst 0xa043c5f0 // ld1w { z16.s-z19.s }, pn9.b/Z, [x15, #0xc, MUL VL]\n"
".inst 0xc0840400 // mova za0h.s[x12], { z0.s-z3.s }\n"
"addvl x15, x15, #16\n"
".inst 0xc0840581 // mova za1h.s[x12], { z12.s-z15.s }\n"
".inst 0xa060c5c4 // st1w { z4.s-z7.s }, pn9.b, [x14]\n"
".inst 0xc0840502 // mova za2h.s[x12], { z8.s-z11.s }\n"
".inst 0xa061c5d4 // st1w { z20.s-z23.s }, pn9.b, [x14, #0x4, MUL VL]\n"
".inst 0xc0840603 // mova za3h.s[x12], { z16.s-z19.s }\n"
"add x12, x12, #0x4\n"
".inst 0xa062c5dc // st1w { z28.s-z31.s }, pn9.b, [x14, #0x8, MUL VL]\n"
"cmp x12, x20\n"
".inst 0xa063c5d8 // st1w { z24.s-z27.s }, pn9.b, [x14, #0xc, MUL VL]\n"
"addvl x14, x14, #16\n"
"blt 11b\n"
"b 29f\n"
"12:" // Store to partial result buffer: Store only
"mov x12, #0x0\n"
"cntw x20\n"
"13:" // Store to partial result buffer: Store only: Loop
".inst 0xc086041c // mova { z28.s-z31.s }, za0h.s[x12]\n"
".inst 0xc0860438 // mova { z24.s-z27.s }, za1h.s[x12]\n"
".inst 0xc086044c // mova { z12.s-z15.s }, za2h.s[x12]\n"
".inst 0xc0860470 // mova { z16.s-z19.s }, za3h.s[x12]\n"
".inst 0xa060c5dc // st1w { z28.s-z31.s }, pn9.b, [x14]\n"
"add x12, x12, #0x4\n"
".inst 0xa061c5d8 // st1w { z24.s-z27.s }, pn9.b, [x14, #0x4, MUL VL]\n"
"cmp x12, x20\n"
".inst 0xa062c5cc // st1w { z12.s-z15.s }, pn9.b, [x14, #0x8, MUL VL]\n"
".inst 0xa063c5d0 // st1w { z16.s-z19.s }, pn9.b, [x14, #0xc, MUL VL]\n"
"addvl x14, x14, #16\n"
"blt 13b\n"
"b 29f\n"
"14:" // Store to output array
"ldr x26, [%x[args], %[offsetof_C]]\n"
"sub x25, x13, x11\n"
"cntw x24\n"
"ld1rh { z29.h }, p1/Z, [%x[args], %[offsetof_KernelArgs_min]]\n"
"ldr x23, [%x[args], %[offsetof_ldcb]]\n"
"whilelt p0.s, x10, x9\n"
"cmp x25, x24\n"
"ld1rh { z28.h }, p1/Z, [%x[args], %[offsetof_KernelArgs_max]]\n"
"csel x22, x25, x24, LT\n"
"mov x12, #0x0\n"
"add x26, x26, x10, LSL #1\n" // C += n
"lsr x21, x22, #0x2\n"
"madd x26, x11, x23, x26\n" // C += m * ldc
"and x20, x22, #0x3\n"
"cbz x21, 16f\n"
"15:" // Store to output array: Accumulator row 0 loop
".inst 0xc0860400 // mova { z0.s-z3.s }, za0h.s[x12]\n"
"add x12, x12, #0x4\n"
"fcvt z0.h, p1/m, z0.s\n"
"fcvt z1.h, p1/m, z1.s\n"
"fcvt z2.h, p1/m, z2.s\n"
"fcvt z3.h, p1/m, z3.s\n"
"cmp x12, x21, LSL #2\n"
".inst 0xc17ccba0 // fclamp { z0.h-z3.h }, z29.h, z28.h\n"
"st1h { z0.s }, p0, [x26]\n"
"add x26, x26, x23\n"
"st1h { z1.s }, p0, [x26]\n"
"add x26, x26, x23\n"
"st1h { z2.s }, p0, [x26]\n"
"add x26, x26, x23\n"
"st1h { z3.s }, p0, [x26]\n"
"add x26, x26, x23\n"
"blt 15b\n"
"16:" // Store to output array: Accumulator row 0 oddments
"cbz x20, 17f\n"
".inst 0xc0860410 // mova { z16.s-z19.s }, za0h.s[x12]\n"
"subs x20, x20, #0x1\n"
"fcvt z16.h, p1/m, z16.s\n"
"fcvt z17.h, p1/m, z17.s\n"
"fcvt z18.h, p1/m, z18.s\n"
"fcvt z19.h, p1/m, z19.s\n"
".inst 0xc17ccbb0 // fclamp { z16.h-z19.h }, z29.h, z28.h\n"
"st1h { z16.s }, p0, [x26]\n"
"add x26, x26, x23\n"
"beq 17f\n"
"subs x20, x20, #0x1\n"
"st1h { z17.s }, p0, [x26]\n"
"add x26, x26, x23\n"
"beq 17f\n"
"st1h { z18.s }, p0, [x26]\n"
"add x26, x26, x23\n"
"17:" // Store to output array: Accumulator row 0 oddments: End
"subs x25, x25, x22\n"
"beq 27f\n"
"whilelt p0.s, x10, x9\n"
"cmp x25, x24\n"
"csel x22, x25, x24, LT\n"
"mov x12, #0x0\n"
"lsr x21, x22, #0x2\n"
"and x20, x22, #0x3\n"
"cbz x21, 19f\n"
"18:" // Store to output array: Accumulator row 1 loop
".inst 0xc0860438 // mova { z24.s-z27.s }, za1h.s[x12]\n"
"add x12, x12, #0x4\n"
"fcvt z24.h, p1/m, z24.s\n"
"fcvt z25.h, p1/m, z25.s\n"
"fcvt z26.h, p1/m, z26.s\n"
"fcvt z27.h, p1/m, z27.s\n"
"cmp x12, x21, LSL #2\n"
".inst 0xc17ccbb8 // fclamp { z24.h-z27.h }, z29.h, z28.h\n"
"st1h { z24.s }, p0, [x26]\n"
"add x26, x26, x23\n"
"st1h { z25.s }, p0, [x26]\n"
"add x26, x26, x23\n"
"st1h { z26.s }, p0, [x26]\n"
"add x26, x26, x23\n"
"st1h { z27.s }, p0, [x26]\n"
"add x26, x26, x23\n"
"blt 18b\n"
"19:" // Store to output array: Accumulator row 1 oddments
"cbz x20, 20f\n"
".inst 0xc0860420 // mova { z0.s-z3.s }, za1h.s[x12]\n"
"subs x20, x20, #0x1\n"
"fcvt z0.h, p1/m, z0.s\n"
"fcvt z1.h, p1/m, z1.s\n"
"fcvt z2.h, p1/m, z2.s\n"
"fcvt z3.h, p1/m, z3.s\n"
".inst 0xc17ccba0 // fclamp { z0.h-z3.h }, z29.h, z28.h\n"
"st1h { z0.s }, p0, [x26]\n"
"add x26, x26, x23\n"
"beq 20f\n"
"subs x20, x20, #0x1\n"
"st1h { z1.s }, p0, [x26]\n"
"add x26, x26, x23\n"
"beq 20f\n"
"st1h { z2.s }, p0, [x26]\n"
"add x26, x26, x23\n"
"20:" // Store to output array: Accumulator row 1 oddments: End
"subs x25, x25, x22\n"
"beq 27f\n"
"whilelt p0.s, x10, x9\n"
"cmp x25, x24\n"
"csel x22, x25, x24, LT\n"
"mov x12, #0x0\n"
"lsr x21, x22, #0x2\n"
"and x20, x22, #0x3\n"
"cbz x21, 22f\n"
"21:" // Store to output array: Accumulator row 2 loop
".inst 0xc0860454 // mova { z20.s-z23.s }, za2h.s[x12]\n"
"add x12, x12, #0x4\n"
"fcvt z20.h, p1/m, z20.s\n"
"fcvt z21.h, p1/m, z21.s\n"
"fcvt z22.h, p1/m, z22.s\n"
"fcvt z23.h, p1/m, z23.s\n"
"cmp x12, x21, LSL #2\n"
".inst 0xc17ccbb4 // fclamp { z20.h-z23.h }, z29.h, z28.h\n"
"st1h { z20.s }, p0, [x26]\n"
"add x26, x26, x23\n"
"st1h { z21.s }, p0, [x26]\n"
"add x26, x26, x23\n"
"st1h { z22.s }, p0, [x26]\n"
"add x26, x26, x23\n"
"st1h { z23.s }, p0, [x26]\n"
"add x26, x26, x23\n"
"blt 21b\n"
"22:" // Store to output array: Accumulator row 2 oddments
"cbz x20, 23f\n"
".inst 0xc086044c // mova { z12.s-z15.s }, za2h.s[x12]\n"
"subs x20, x20, #0x1\n"
"fcvt z12.h, p1/m, z12.s\n"
"fcvt z13.h, p1/m, z13.s\n"
"fcvt z14.h, p1/m, z14.s\n"
"fcvt z15.h, p1/m, z15.s\n"
".inst 0xc17ccbac // fclamp { z12.h-z15.h }, z29.h, z28.h\n"
"st1h { z12.s }, p0, [x26]\n"
"add x26, x26, x23\n"
"beq 23f\n"
"subs x20, x20, #0x1\n"
"st1h { z13.s }, p0, [x26]\n"
"add x26, x26, x23\n"
"beq 23f\n"
"st1h { z14.s }, p0, [x26]\n"
"add x26, x26, x23\n"
"23:" // Store to output array: Accumulator row 2 oddments: End
"subs x25, x25, x22\n"
"beq 27f\n"
"whilelt p0.s, x10, x9\n"
"cmp x25, x24\n"
"csel x20, x25, x24, LT\n"
"mov x12, #0x0\n"
"lsr x21, x20, #0x2\n"
"and x20, x20, #0x3\n"
"cbz x21, 25f\n"
"24:" // Store to output array: Accumulator row 3 loop
".inst 0xc0860464 // mova { z4.s-z7.s }, za3h.s[x12]\n"
"add x12, x12, #0x4\n"
"fcvt z4.h, p1/m, z4.s\n"
"fcvt z5.h, p1/m, z5.s\n"
"fcvt z6.h, p1/m, z6.s\n"
"fcvt z7.h, p1/m, z7.s\n"
"cmp x12, x21, LSL #2\n"
".inst 0xc17ccba4 // fclamp { z4.h-z7.h }, z29.h, z28.h\n"
"st1h { z4.s }, p0, [x26]\n"
"add x26, x26, x23\n"
"st1h { z5.s }, p0, [x26]\n"
"add x26, x26, x23\n"
"st1h { z6.s }, p0, [x26]\n"
"add x26, x26, x23\n"
"st1h { z7.s }, p0, [x26]\n"
"add x26, x26, x23\n"
"blt 24b\n"
"25:" // Store to output array: Accumulator row 3 oddments
"cbz x20, 26f\n"
".inst 0xc0860464 // mova { z4.s-z7.s }, za3h.s[x12]\n"
"subs x20, x20, #0x1\n"
"fcvt z4.h, p1/m, z4.s\n"
"fcvt z5.h, p1/m, z5.s\n"
"fcvt z6.h, p1/m, z6.s\n"
"fcvt z7.h, p1/m, z7.s\n"
".inst 0xc17ccba4 // fclamp { z4.h-z7.h }, z29.h, z28.h\n"
"st1h { z4.s }, p0, [x26]\n"
"add x26, x26, x23\n"
"beq 26f\n"
"subs x20, x20, #0x1\n"
"st1h { z5.s }, p0, [x26]\n"
"add x26, x26, x23\n"
"beq 26f\n"
"st1h { z6.s }, p0, [x26]\n"
"26:" // Store to output array: Accumulator row 3 oddments: End
"27:" // Store to output array: End
"tbz x16, #0, 29f\n"
"mov x12, #0x0\n"
"cntw x20\n"
"28:" // Store to output array: Refill accumulators: Loop
".inst 0xa040c5e0 // ld1w { z0.s-z3.s }, pn9.b/Z, [x15]\n"
".inst 0xa041c5ec // ld1w { z12.s-z15.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n"
".inst 0xa042c5e4 // ld1w { z4.s-z7.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n"
".inst 0xa043c5f4 // ld1w { z20.s-z23.s }, pn9.b/Z, [x15, #0xc, MUL VL]\n"
".inst 0xc0840400 // mova za0h.s[x12], { z0.s-z3.s }\n"
"addvl x15, x15, #16\n"
".inst 0xc0840581 // mova za1h.s[x12], { z12.s-z15.s }\n"
".inst 0xc0840482 // mova za2h.s[x12], { z4.s-z7.s }\n"
".inst 0xc0840683 // mova za3h.s[x12], { z20.s-z23.s }\n"
"add x12, x12, #0x4\n"
"cmp x12, x20\n"
"blt 28b\n"
"29:" // End block
"incw x10\n"
"cmp x10, x9\n"
"blt 3b\n"
"incw x11, ALL, MUL #4\n"
"mov x10, #0x0\n"
"cmp x11, x13\n"
"mov x28, x27\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", "x16", "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_FEATURE_SVE
|
