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
|
/*
* 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_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_2VLx2VL(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 x16, [%x[args], %[offsetof_flags]]\n"
".inst 0xd503477f // SMSTART ZA\n"
"ptrue p0.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 0xa040c5e4 // ld1w { z4.s-z7.s }, pn9.b/Z, [x15]\n"
".inst 0xc0840480 // mova za0h.s[x12], { z4.s-z7.s }\n"
".inst 0xa041c5e8 // ld1w { z8.s-z11.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n"
".inst 0xc0840501 // mova za1h.s[x12], { z8.s-z11.s }\n"
".inst 0xa042c5f4 // ld1w { z20.s-z23.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n"
".inst 0xc0840682 // mova za2h.s[x12], { z20.s-z23.s }\n"
".inst 0xa043c5e8 // ld1w { z8.s-z11.s }, pn9.b/Z, [x15, #0xc, MUL VL]\n"
".inst 0xc0840503 // mova za3h.s[x12], { z8.s-z11.s }\n"
"add x12, x12, #0x4\n"
"cmp x12, x20\n"
"addvl x15, x15, #16\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"
".inst 0x25a94550 // whilelt pn8.s, x10, x9, VLx2\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"
"fmov z12.s, #1.0\n"
".inst 0xa10a4289 // ldnt1w { z1.s, z9.s }, p8/Z, [x20, x10, LSL #2]\n"
".inst 0x80810180 // fmopa za0.s, p0/M, p0/M, z12.s, z1.s\n"
".inst 0x80890181 // fmopa za1.s, p0/M, p0/M, z12.s, z9.s\n"
".inst 0x80810182 // fmopa za2.s, p0/M, p0/M, z12.s, z1.s\n"
".inst 0x80890183 // fmopa za3.s, p0/M, p0/M, z12.s, z9.s\n"
"4:" // Prepare accumulators: Test for last block
"mov x20, x10\n"
"mov x21, x11\n"
"incw x20, ALL, MUL #2\n"
"incw x21, ALL, MUL #2\n"
"cmp x20, x9\n"
"csel x21, x11, x21, LT\n"
"mov x20, x16\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"
"add x20, x20, #0x1\n"
"lsr x20, x20, #0x1\n"
"ldr x23, [%x[args], %[offsetof_B]]\n"
"lsr x22, x20, #0x2\n"
"and x21, x20, #0x3\n"
"ldr x20, [%x[args], %[offsetof_kstride_bytes]]\n"
"madd x23, x10, x20, x23\n" // bptr = B + n * kstride_bytes
"cbz x22, 8f\n"
"subs x22, x22, #0x1\n"
".inst 0xa0402772 // ld1h { z18.h-z19.h }, pn9.b/Z, [x27]\n"
".inst 0xa04026e3 // ldnt1h { z2.h-z3.h }, pn9.b/Z, [x23]\n"
".inst 0xa0412764 // ld1h { z4.h-z5.h }, pn9.b/Z, [x27, #0x2, MUL VL]\n"
".inst 0xa04126fb // ldnt1h { z26.h-z27.h }, pn9.b/Z, [x23, #0x2, MUL VL]\n"
".inst 0xa042276a // ld1h { z10.h-z11.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
".inst 0xa04226f5 // ldnt1h { z20.h-z21.h }, pn9.b/Z, [x23, #0x4, MUL VL]\n"
".inst 0xa0432766 // ld1h { z6.h-z7.h }, pn9.b/Z, [x27, #0x6, MUL VL]\n"
"addvl x27, x27, #8\n"
".inst 0xa04326e9 // ldnt1h { z8.h-z9.h }, pn9.b/Z, [x23, #0x6, MUL VL]\n"
"addvl x23, x23, #8\n"
"ble 7f\n"
"6:" // K loop
".inst 0x81820240 // bfmopa za0.s, p0/M, p0/M, z18.h, z2.h\n"
"subs x22, x22, #0x1\n"
".inst 0x81830241 // bfmopa za1.s, p0/M, p0/M, z18.h, z3.h\n"
".inst 0x81820262 // bfmopa za2.s, p0/M, p0/M, z19.h, z2.h\n"
".inst 0x81830263 // bfmopa za3.s, p0/M, p0/M, z19.h, z3.h\n"
".inst 0xa0402772 // ld1h { z18.h-z19.h }, pn9.b/Z, [x27]\n"
".inst 0x819a0080 // bfmopa za0.s, p0/M, p0/M, z4.h, z26.h\n"
".inst 0xa04026e3 // ldnt1h { z2.h-z3.h }, pn9.b/Z, [x23]\n"
".inst 0x819b0081 // bfmopa za1.s, p0/M, p0/M, z4.h, z27.h\n"
".inst 0x819a00a2 // bfmopa za2.s, p0/M, p0/M, z5.h, z26.h\n"
".inst 0x819b00a3 // bfmopa za3.s, p0/M, p0/M, z5.h, z27.h\n"
".inst 0xa0412764 // ld1h { z4.h-z5.h }, pn9.b/Z, [x27, #0x2, MUL VL]\n"
".inst 0x81940140 // bfmopa za0.s, p0/M, p0/M, z10.h, z20.h\n"
".inst 0xa04126fb // ldnt1h { z26.h-z27.h }, pn9.b/Z, [x23, #0x2, MUL VL]\n"
".inst 0x81950141 // bfmopa za1.s, p0/M, p0/M, z10.h, z21.h\n"
".inst 0x81940162 // bfmopa za2.s, p0/M, p0/M, z11.h, z20.h\n"
".inst 0x81950163 // bfmopa za3.s, p0/M, p0/M, z11.h, z21.h\n"
".inst 0xa042276a // ld1h { z10.h-z11.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n"
".inst 0xa04226f5 // ldnt1h { z20.h-z21.h }, pn9.b/Z, [x23, #0x4, MUL VL]\n"
".inst 0x818800c0 // bfmopa za0.s, p0/M, p0/M, z6.h, z8.h\n"
".inst 0x818900c1 // bfmopa za1.s, p0/M, p0/M, z6.h, z9.h\n"
".inst 0x818800e2 // bfmopa za2.s, p0/M, p0/M, z7.h, z8.h\n"
".inst 0x818900e3 // bfmopa za3.s, p0/M, p0/M, z7.h, z9.h\n"
".inst 0xa0432766 // ld1h { z6.h-z7.h }, pn9.b/Z, [x27, #0x6, MUL VL]\n"
"addvl x27, x27, #8\n"
".inst 0xa04326e9 // ldnt1h { z8.h-z9.h }, pn9.b/Z, [x23, #0x6, MUL VL]\n"
"addvl x23, x23, #8\n"
"bgt 6b\n"
"7:" // K loop tail
".inst 0x81820240 // bfmopa za0.s, p0/M, p0/M, z18.h, z2.h\n"
".inst 0x81830241 // bfmopa za1.s, p0/M, p0/M, z18.h, z3.h\n"
".inst 0x81820262 // bfmopa za2.s, p0/M, p0/M, z19.h, z2.h\n"
".inst 0x81830263 // bfmopa za3.s, p0/M, p0/M, z19.h, z3.h\n"
".inst 0x819a0080 // bfmopa za0.s, p0/M, p0/M, z4.h, z26.h\n"
".inst 0x819b0081 // bfmopa za1.s, p0/M, p0/M, z4.h, z27.h\n"
".inst 0x819a00a2 // bfmopa za2.s, p0/M, p0/M, z5.h, z26.h\n"
".inst 0x819b00a3 // bfmopa za3.s, p0/M, p0/M, z5.h, z27.h\n"
".inst 0x81940140 // bfmopa za0.s, p0/M, p0/M, z10.h, z20.h\n"
".inst 0x81950141 // bfmopa za1.s, p0/M, p0/M, z10.h, z21.h\n"
".inst 0x81940162 // bfmopa za2.s, p0/M, p0/M, z11.h, z20.h\n"
".inst 0x81950163 // bfmopa za3.s, p0/M, p0/M, z11.h, z21.h\n"
".inst 0x818800c0 // bfmopa za0.s, p0/M, p0/M, z6.h, z8.h\n"
".inst 0x818900c1 // bfmopa za1.s, p0/M, p0/M, z6.h, z9.h\n"
".inst 0x818800e2 // bfmopa za2.s, p0/M, p0/M, z7.h, z8.h\n"
".inst 0x818900e3 // bfmopa za3.s, p0/M, p0/M, z7.h, z9.h\n"
"8:" // K oddments
"cbz x21, 10f\n"
"9:" // K oddments: Loop
".inst 0xa040277e // ld1h { z30.h-z31.h }, pn9.b/Z, [x27]\n"
"subs x21, x21, #0x1\n"
"addvl x27, x27, #2\n"
".inst 0xa14026e5 // ld1h { z5.h, z13.h }, pn9.b/Z, [x23]\n"
"addvl x23, x23, #2\n"
".inst 0x818503c0 // bfmopa za0.s, p0/M, p0/M, z30.h, z5.h\n"
".inst 0x818d03c1 // bfmopa za1.s, p0/M, p0/M, z30.h, z13.h\n"
".inst 0x818503e2 // bfmopa za2.s, p0/M, p0/M, z31.h, z5.h\n"
".inst 0x818d03e3 // bfmopa za3.s, p0/M, p0/M, z31.h, z13.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 0xa040c5ec // ld1w { z12.s-z15.s }, pn9.b/Z, [x15]\n"
".inst 0xc0860400 // mova { z0.s-z3.s }, za0h.s[x12]\n"
".inst 0xc0840580 // mova za0h.s[x12], { z12.s-z15.s }\n"
".inst 0xc0860424 // mova { z4.s-z7.s }, za1h.s[x12]\n"
".inst 0xa041c5f0 // ld1w { z16.s-z19.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n"
".inst 0xc0840601 // mova za1h.s[x12], { z16.s-z19.s }\n"
".inst 0xc086045c // mova { z28.s-z31.s }, za2h.s[x12]\n"
".inst 0xc086046c // mova { z12.s-z15.s }, za3h.s[x12]\n"
".inst 0xa042c5f0 // ld1w { z16.s-z19.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n"
".inst 0xc0840602 // mova za2h.s[x12], { z16.s-z19.s }\n"
".inst 0xa043c5e8 // ld1w { z8.s-z11.s }, pn9.b/Z, [x15, #0xc, MUL VL]\n"
".inst 0xc0840503 // mova za3h.s[x12], { z8.s-z11.s }\n"
"add x12, x12, #0x4\n"
"cmp x12, x20\n"
".inst 0xa060c5c0 // st1w { z0.s-z3.s }, pn9.b, [x14]\n"
"addvl x15, x15, #16\n"
".inst 0xa061c5c4 // st1w { z4.s-z7.s }, pn9.b, [x14, #0x4, MUL VL]\n"
".inst 0xa062c5dc // st1w { z28.s-z31.s }, pn9.b, [x14, #0x8, MUL VL]\n"
".inst 0xa063c5cc // st1w { z12.s-z15.s }, pn9.b, [x14, #0xc, MUL VL]\n"
"addvl x14, x14, #16\n"
"blt 11b\n"
"b 30f\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 0xc0860410 // mova { z16.s-z19.s }, za0h.s[x12]\n"
".inst 0xc0860420 // mova { z0.s-z3.s }, za1h.s[x12]\n"
".inst 0xa060c5d0 // st1w { z16.s-z19.s }, pn9.b, [x14]\n"
".inst 0xc0860458 // mova { z24.s-z27.s }, za2h.s[x12]\n"
".inst 0xc0860464 // mova { z4.s-z7.s }, za3h.s[x12]\n"
".inst 0xa061c5c0 // st1w { z0.s-z3.s }, pn9.b, [x14, #0x4, MUL VL]\n"
"add x12, x12, #0x4\n"
"cmp x12, x20\n"
".inst 0xa062c5d8 // st1w { z24.s-z27.s }, pn9.b, [x14, #0x8, MUL VL]\n"
".inst 0xa063c5c4 // st1w { z4.s-z7.s }, pn9.b, [x14, #0xc, MUL VL]\n"
"addvl x14, x14, #16\n"
"blt 13b\n"
"b 30f\n"
"14:" // Store to output array
"ldr x26, [%x[args], %[offsetof_C]]\n"
"add x26, x26, x10, LSL #2\n" // C += n
"sub x25, x13, x11\n"
"ldr x24, [%x[args], %[offsetof_ldcb]]\n"
"madd x26, x11, x24, x26\n" // C += m * ldc
"tbz x16, #2, 21f\n"
"cntw x23\n"
"cmp x25, x23\n"
"csel x22, x25, x23, LT\n"
"lsr x21, x22, #0x2\n"
"mov x12, #0x0\n"
"and x20, x22, #0x3\n"
"cbz x21, 16f\n"
"15:" // Store to output array: Skip activation: Accumulator row 0 loop
".inst 0xc0860404 // mova { z4.s-z7.s }, za0h.s[x12]\n"
".inst 0xc086042c // mova { z12.s-z15.s }, za1h.s[x12]\n"
".inst 0xa1604344 // st1w { z4.s, z12.s }, p8, [x26]\n"
"add x26, x26, x24\n"
".inst 0xa1604345 // st1w { z5.s, z13.s }, p8, [x26]\n"
"add x26, x26, x24\n"
"add x12, x12, #0x4\n"
".inst 0xa1604346 // st1w { z6.s, z14.s }, p8, [x26]\n"
"add x26, x26, x24\n"
"cmp x12, x21, LSL #2\n"
".inst 0xa1604347 // st1w { z7.s, z15.s }, p8, [x26]\n"
"add x26, x26, x24\n"
"blt 15b\n"
"16:" // Store to output array: Skip activation: Accumulator row 0 oddments
"cbz x20, 17f\n"
"subs x20, x20, #0x1\n"
".inst 0xc0860414 // mova { z20.s-z23.s }, za0h.s[x12]\n"
".inst 0xc086043c // mova { z28.s-z31.s }, za1h.s[x12]\n"
".inst 0xa1604354 // st1w { z20.s, z28.s }, p8, [x26]\n"
"add x26, x26, x24\n"
"beq 17f\n"
"subs x20, x20, #0x1\n"
".inst 0xa1604355 // st1w { z21.s, z29.s }, p8, [x26]\n"
"add x26, x26, x24\n"
"beq 17f\n"
".inst 0xa1604356 // st1w { z22.s, z30.s }, p8, [x26]\n"
"add x26, x26, x24\n"
"17:" // Store to output array: Skip activation: Accumulator row 0 oddments: End
"subs x25, x25, x22\n"
"beq 21f\n"
"cmp x25, x23\n"
"csel x22, x25, x23, LT\n"
"lsr x21, x22, #0x2\n"
"mov x12, #0x0\n"
"and x20, x22, #0x3\n"
"cbz x21, 19f\n"
"18:" // Store to output array: Skip activation: Accumulator row 1 loop
".inst 0xc0860444 // mova { z4.s-z7.s }, za2h.s[x12]\n"
".inst 0xc086046c // mova { z12.s-z15.s }, za3h.s[x12]\n"
".inst 0xa1604344 // st1w { z4.s, z12.s }, p8, [x26]\n"
"add x26, x26, x24\n"
".inst 0xa1604345 // st1w { z5.s, z13.s }, p8, [x26]\n"
"add x26, x26, x24\n"
"add x12, x12, #0x4\n"
".inst 0xa1604346 // st1w { z6.s, z14.s }, p8, [x26]\n"
"add x26, x26, x24\n"
"cmp x12, x21, LSL #2\n"
".inst 0xa1604347 // st1w { z7.s, z15.s }, p8, [x26]\n"
"add x26, x26, x24\n"
"blt 18b\n"
"19:" // Store to output array: Skip activation: Accumulator row 1 oddments
"cbz x20, 20f\n"
"subs x20, x20, #0x1\n"
".inst 0xc0860450 // mova { z16.s-z19.s }, za2h.s[x12]\n"
".inst 0xc0860478 // mova { z24.s-z27.s }, za3h.s[x12]\n"
".inst 0xa1604350 // st1w { z16.s, z24.s }, p8, [x26]\n"
"add x26, x26, x24\n"
"beq 20f\n"
"subs x20, x20, #0x1\n"
".inst 0xa1604351 // st1w { z17.s, z25.s }, p8, [x26]\n"
"add x26, x26, x24\n"
"beq 20f\n"
".inst 0xa1604352 // st1w { z18.s, z26.s }, p8, [x26]\n"
"add x26, x26, x24\n"
"20:" // Store to output array: Skip activation: Accumulator row 1 oddments: End
"subs x25, x25, x22\n"
"beq 21f\n"
"b 28f\n"
"21:" // Store to output array: Skip activation: End
"cntw x23\n"
"cmp x25, x23\n"
"ld1rw { z1.s }, p0/Z, [%x[args], %[offsetof_KernelArgs_min]]\n"
"csel x22, x25, x23, LT\n"
"lsr x21, x22, #0x2\n"
"ld1rw { z0.s }, p0/Z, [%x[args], %[offsetof_KernelArgs_max]]\n"
"mov x12, #0x0\n"
"and x20, x22, #0x3\n"
"cbz x21, 23f\n"
"22:" // Store to output array: Accumulator row 0 loop
".inst 0xc0860414 // mova { z20.s-z23.s }, za0h.s[x12]\n"
".inst 0xc086043c // mova { z28.s-z31.s }, za1h.s[x12]\n"
".inst 0xc1a0c834 // fclamp { z20.s-z23.s }, z1.s, z0.s\n"
".inst 0xc1a0c83c // fclamp { z28.s-z31.s }, z1.s, z0.s\n"
".inst 0xa1604354 // st1w { z20.s, z28.s }, p8, [x26]\n"
"add x26, x26, x24\n"
"add x12, x12, #0x4\n"
".inst 0xa1604355 // st1w { z21.s, z29.s }, p8, [x26]\n"
"add x26, x26, x24\n"
"cmp x12, x21, LSL #2\n"
".inst 0xa1604356 // st1w { z22.s, z30.s }, p8, [x26]\n"
"add x26, x26, x24\n"
".inst 0xa1604357 // st1w { z23.s, z31.s }, p8, [x26]\n"
"add x26, x26, x24\n"
"blt 22b\n"
"23:" // Store to output array: Accumulator row 0 oddments
"cbz x20, 24f\n"
".inst 0xc0860410 // mova { z16.s-z19.s }, za0h.s[x12]\n"
".inst 0xc0860438 // mova { z24.s-z27.s }, za1h.s[x12]\n"
".inst 0xc1a0c830 // fclamp { z16.s-z19.s }, z1.s, z0.s\n"
".inst 0xc1a0c838 // fclamp { z24.s-z27.s }, z1.s, z0.s\n"
"subs x20, x20, #0x1\n"
".inst 0xa1604350 // st1w { z16.s, z24.s }, p8, [x26]\n"
"add x26, x26, x24\n"
"beq 24f\n"
"subs x20, x20, #0x1\n"
".inst 0xa1604351 // st1w { z17.s, z25.s }, p8, [x26]\n"
"add x26, x26, x24\n"
"beq 24f\n"
".inst 0xa1604352 // st1w { z18.s, z26.s }, p8, [x26]\n"
"add x26, x26, x24\n"
"24:" // Store to output array: Accumulator row 0 oddments: End
"subs x25, x25, x22\n"
"beq 28f\n"
"cmp x25, x23\n"
"csel x20, x25, x23, LT\n"
"lsr x21, x20, #0x2\n"
"mov x12, #0x0\n"
"and x20, x20, #0x3\n"
"cbz x21, 26f\n"
"25:" // Store to output array: Accumulator row 1 loop
".inst 0xc0860450 // mova { z16.s-z19.s }, za2h.s[x12]\n"
".inst 0xc0860478 // mova { z24.s-z27.s }, za3h.s[x12]\n"
".inst 0xc1a0c830 // fclamp { z16.s-z19.s }, z1.s, z0.s\n"
".inst 0xc1a0c838 // fclamp { z24.s-z27.s }, z1.s, z0.s\n"
".inst 0xa1604350 // st1w { z16.s, z24.s }, p8, [x26]\n"
"add x26, x26, x24\n"
"add x12, x12, #0x4\n"
".inst 0xa1604351 // st1w { z17.s, z25.s }, p8, [x26]\n"
"add x26, x26, x24\n"
"cmp x12, x21, LSL #2\n"
".inst 0xa1604352 // st1w { z18.s, z26.s }, p8, [x26]\n"
"add x26, x26, x24\n"
".inst 0xa1604353 // st1w { z19.s, z27.s }, p8, [x26]\n"
"add x26, x26, x24\n"
"blt 25b\n"
"26:" // Store to output array: Accumulator row 1 oddments
"cbz x20, 27f\n"
".inst 0xc0860450 // mova { z16.s-z19.s }, za2h.s[x12]\n"
".inst 0xc0860478 // mova { z24.s-z27.s }, za3h.s[x12]\n"
".inst 0xc1a0c830 // fclamp { z16.s-z19.s }, z1.s, z0.s\n"
".inst 0xc1a0c838 // fclamp { z24.s-z27.s }, z1.s, z0.s\n"
"subs x20, x20, #0x1\n"
".inst 0xa1604350 // st1w { z16.s, z24.s }, p8, [x26]\n"
"add x26, x26, x24\n"
"beq 27f\n"
"subs x20, x20, #0x1\n"
".inst 0xa1604351 // st1w { z17.s, z25.s }, p8, [x26]\n"
"add x26, x26, x24\n"
"beq 27f\n"
".inst 0xa1604352 // st1w { z18.s, z26.s }, p8, [x26]\n"
"27:" // Store to output array: Accumulator row 1 oddments: End
"28:" // Store to output array: End
"tbz x16, #0, 30f\n"
"mov x12, #0x0\n"
"cntw x20\n"
"29:" // Store to output array: Refill accumulators: Loop
".inst 0xa040c5ec // ld1w { z12.s-z15.s }, pn9.b/Z, [x15]\n"
".inst 0xc0840580 // mova za0h.s[x12], { z12.s-z15.s }\n"
".inst 0xa041c5f0 // ld1w { z16.s-z19.s }, pn9.b/Z, [x15, #0x4, MUL VL]\n"
".inst 0xc0840601 // mova za1h.s[x12], { z16.s-z19.s }\n"
".inst 0xa042c5f0 // ld1w { z16.s-z19.s }, pn9.b/Z, [x15, #0x8, MUL VL]\n"
".inst 0xc0840602 // mova za2h.s[x12], { z16.s-z19.s }\n"
".inst 0xa043c5ec // ld1w { z12.s-z15.s }, pn9.b/Z, [x15, #0xc, MUL VL]\n"
".inst 0xc0840583 // mova za3h.s[x12], { z12.s-z15.s }\n"
"add x12, x12, #0x4\n"
"cmp x12, x20\n"
"addvl x15, x15, #16\n"
"blt 29b\n"
"30:" // End block
"incw x10, ALL, MUL #2\n"
"cmp x10, x9\n"
"blt 3b\n"
"incw x11, ALL, MUL #2\n"
"cmp x11, x13\n"
"mov x10, #0x0\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_COMPUTE_ENABLE_SME2
|
