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
|
/*
* Copyright (c) 2022-2023 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to
* deal in the Software without restriction, including without limitation the
* rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
* sell copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
#if defined(ARM_COMPUTE_ENABLE_SME2)
#include <algorithm>
#include <cstddef>
namespace arm_conv {
namespace depthwise {
void sme2_fp32bf16fp32_planar_3x3_s1_4rows_dot_za_impl(
const float *inptr,
size_t ld_in_row,
size_t ld_in_col,
size_t ld_in_vl,
unsigned int pad_top,
unsigned int valid_input_rows,
unsigned int pad_left,
unsigned int valid_input_cols,
const float *weights,
const float *bias,
float **outptrs,
const size_t *outlds,
const size_t *outvllds,
unsigned int output_cols,
unsigned int start_channel,
unsigned int valid_channels,
float act_min,
float act_max
)
{
struct Args
{
const float *inptr;
size_t ld_in_vl;
long unsigned int pad_top, pad_bottom, pad_left;
const float *weights;
const float *bias;
long unsigned int input_cols, output_cols;
float **outptrs;
const size_t *ld_out_cols;
const size_t *ld_out_vls;
long unsigned int current_channel, n_channels;
float clamp_min, clamp_max;
};
Args args = { inptr, ld_in_vl, pad_top, 6u - std::min(6u, pad_top + valid_input_rows), pad_left, weights, bias, valid_input_cols, output_cols, outptrs, outlds, outvllds, start_channel, valid_channels, act_min, act_max };
__asm__ __volatile__(
"ldr x6, [%x[args], %[offsetof_Args_pad_bottom]]\n"
"mov x20, #0x6\n"
".inst 0xd503477f // SMSTART ZA\n"
"sub x20, x20, x6\n"
"ldr x7, [%x[args], %[offsetof_Args_pad_top]]\n"
"ptrue p2.b\n"
"ld1rw { z28.s }, p2/Z, [%x[args], %[offsetof_Args_clamp_min]]\n"
"ldr x17, [%x[args], %[offsetof_Args_n_channels]]\n"
"whilelt p1.s, XZR, x17\n"
"whilelt p9.s, XZR, x20\n"
"ld1rw { z29.s }, p2/Z, [%x[args], %[offsetof_Args_clamp_max]]\n"
"whilelt p8.s, XZR, x7\n"
"eor p8.b, p2/Z, p8.b, p9.b\n"
"ldr x16, [%x[args], %[offsetof_Args_current_channel]]\n"
"1:" // Channel loop
"ldr x20, [%x[args], %[offsetof_Args_bias]]\n"
"fmov z22.s, #0x0\n"
"cbz x20, 2f\n"
"ld1w { z22.s }, p1/Z, [x20, x16, LSL #2]\n"
"2:" // Load bias: Done
"ldr x20, [%x[args], %[offsetof_Args_weights]]\n"
"mov x21, x20\n"
"fmov z9.s, #0x0\n"
"ld1w { z25.s }, p2/Z, [x21]\n"
"incb x21, ALL, MUL #3\n"
"incb x20\n"
"ld1w { z27.s }, p2/Z, [x21]\n"
".inst 0x648aab29 // bfcvtnt z9.h, p2/M, z25.s\n"
"incb x21, ALL, MUL #3\n"
"ld1w { z21.s }, p2/Z, [x21]\n"
"mov x21, x20\n"
".inst 0x658aab28 // bfcvt z8.h, p2/M, z25.s\n"
"ld1w { z25.s }, p2/Z, [x21]\n"
".inst 0x658aab66 // bfcvt z6.h, p2/M, z27.s\n"
"fmov z2.s, #0x0\n"
"incb x21, ALL, MUL #3\n"
".inst 0x658aab21 // bfcvt z1.h, p2/M, z25.s\n"
".inst 0x648aab68 // bfcvtnt z8.h, p2/M, z27.s\n"
"incb x20\n"
"ld1w { z27.s }, p2/Z, [x21]\n"
"incb x21, ALL, MUL #3\n"
".inst 0x648aaaa6 // bfcvtnt z6.h, p2/M, z21.s\n"
".inst 0x658aaaa5 // bfcvt z5.h, p2/M, z21.s\n"
"ld1w { z21.s }, p2/Z, [x21]\n"
"mov x21, x20\n"
".inst 0x648aab22 // bfcvtnt z2.h, p2/M, z25.s\n"
"ld1w { z25.s }, p2/Z, [x21]\n"
"incb x21, ALL, MUL #3\n"
".inst 0x648aab61 // bfcvtnt z1.h, p2/M, z27.s\n"
".inst 0x658aab6c // bfcvt z12.h, p2/M, z27.s\n"
"ld1w { z27.s }, p2/Z, [x21]\n"
"ldr x15, [%x[args], %[offsetof_Args_input_cols]]\n"
"incb x21, ALL, MUL #3\n"
"fmov z7.s, #0x0\n"
".inst 0x658aab24 // bfcvt z4.h, p2/M, z25.s\n"
"ldr x14, [%x[args], %[offsetof_Args_inptr]]\n"
".inst 0x658aab60 // bfcvt z0.h, p2/M, z27.s\n"
".inst 0x648aaaac // bfcvtnt z12.h, p2/M, z21.s\n"
"sub x20, x15, #0x1\n"
"orr x23, x20, %x[ld_in_col], LSL #18\n"
".inst 0x658aaaaa // bfcvt z10.h, p2/M, z21.s\n"
"ld1w { z21.s }, p2/Z, [x21]\n"
"orr x23, x17, x23, LSL #20\n"
"mov x22, #0x6\n"
"add x21, x7, x6\n"
"lsl x20, %x[ld_in_row], #0x2\n"
"mov z23.d, z22.d\n"
".inst 0x648aab27 // bfcvtnt z7.h, p2/M, z25.s\n"
".inst 0x648aab64 // bfcvtnt z4.h, p2/M, z27.s\n"
".inst 0x648aaaa0 // bfcvtnt z0.h, p2/M, z21.s\n"
"mov x8, #0x0\n"
"ldr x13, [%x[args], %[offsetof_Args_output_cols]]\n"
".inst 0x658aaaa3 // bfcvt z3.h, p2/M, z21.s\n"
"lsl x23, x23, #0x2\n"
"sub x22, x22, x21\n"
"madd x20, x20, x7, x14\n"
"3:" // Issue prefetches
"subs x22, x22, #0x1\n"
".inst 0xf8b74a9c // rprfm pldstrm, x23, [x20]\n"
"add x20, x20, %x[ld_in_col], LSL #2\n"
"bgt 3b\n"
"ldr x11, [%x[args], %[offsetof_Args_outptrs]]\n"
"lsl x20, %x[ld_in_row], #0x2\n"
"msub x14, x7, x20, x14\n"
".inst 0xc0040ac0 // mova za.d[x8, #0], { z22.d-z23.d }\n"
"ldr x20, [%x[args], %[offsetof_Args_ld_out_cols]]\n"
".inst 0xc0040ac1 // mova za.d[x8, #1], { z22.d-z23.d }\n"
"mov x10, #0x2\n"
"ldp x9, x28, [x11], #0x10\n"
".inst 0xc0040ac2 // mova za.d[x8, #2], { z22.d-z23.d }\n"
"ldp x27, x26, [x20], #0x10\n"
".inst 0xc0040ac3 // mova za.d[x8, #3], { z22.d-z23.d }\n"
"ldr x21, [%x[args], %[offsetof_Args_pad_left]]\n"
".inst 0xc0040ac4 // mova za.d[x8, #4], { z22.d-z23.d }\n"
"ldp x25, x24, [x11], #0x10\n"
".inst 0xc0040ac5 // mova za.d[x8, #5], { z22.d-z23.d }\n"
"ldp x23, x22, [x20], #0x10\n"
"cbz x21, 5f\n"
"cmp x21, x10\n"
"csel x20, x21, x10, LT\n"
"sub x21, x21, x20\n"
"sub x10, x10, x20\n"
"cbz x21, 5f\n"
".inst 0xc0060818 // mova { z24.d-z25.d }, za.d[x8, #0]\n"
"sub x13, x13, x21\n"
".inst 0xc006083a // mova { z26.d-z27.d }, za.d[x8, #1]\n"
".inst 0xc1bdcb98 // fclamp { z24.s-z27.s }, z28.s, z29.s\n"
"4:" // Left padding
"subs x21, x21, #0x1\n"
"st1w { z24.s }, p1, [x9]\n"
"add x9, x9, x27, LSL #2\n"
"st1w { z26.s }, p1, [x28]\n"
"add x28, x28, x26, LSL #2\n"
"st1w { z25.s }, p1, [x25]\n"
"add x25, x25, x23, LSL #2\n"
"st1w { z27.s }, p1, [x24]\n"
"add x24, x24, x22, LSL #2\n"
"bgt 4b\n"
"5:" // Left padding: End
"adds XZR, x7, x6\n"
"bne 10f\n"
"cbz x10, 8f\n"
"cmp x10, #0x1\n"
"sub x15, x15, x10\n"
"beq 7f\n"
"6:" // Unpadded: 2 priming loads
"add x20, x14, %x[ld_in_row], LSL #2\n"
"ld1w { z16.s }, p1/Z, [x14]\n"
".inst 0x658aaa0d // bfcvt z13.h, p2/M, z16.s\n"
"add x14, x14, %x[ld_in_col], LSL #2\n"
"ld1w { z16.s }, p1/Z, [x20]\n"
"add x20, x20, %x[ld_in_row], LSL #2\n"
".inst 0x648aaa0d // bfcvtnt z13.h, p2/M, z16.s\n"
"ld1w { z16.s }, p1/Z, [x20]\n"
"add x20, x20, %x[ld_in_row], LSL #2\n"
".inst 0x658aaa0e // bfcvt z14.h, p2/M, z16.s\n"
"ld1w { z16.s }, p1/Z, [x20]\n"
"add x20, x20, %x[ld_in_row], LSL #2\n"
".inst 0x648aaa0e // bfcvtnt z14.h, p2/M, z16.s\n"
".inst 0xc12811b0 // bfdot za.s[x8, 0], { z13.h-z14.h }, z8.h\n"
"ld1w { z16.s }, p1/Z, [x20]\n"
"add x20, x20, %x[ld_in_row], LSL #2\n"
".inst 0x658aaa0f // bfcvt z15.h, p2/M, z16.s\n"
".inst 0xc12911b1 // bfdot za.s[x8, 1], { z13.h-z14.h }, z9.h\n"
"ld1w { z16.s }, p1/Z, [x20]\n"
".inst 0x648aaa0f // bfcvtnt z15.h, p2/M, z16.s\n"
".inst 0xc12511d0 // bfdot za.s[x8, 0], { z14.h-z15.h }, z5.h\n"
".inst 0xc12611d1 // bfdot za.s[x8, 1], { z14.h-z15.h }, z6.h\n"
"7:" // Unpadded: 1 priming loads
"add x20, x14, %x[ld_in_row], LSL #2\n"
"ld1w { z16.s }, p1/Z, [x14]\n"
".inst 0x658aaa0d // bfcvt z13.h, p2/M, z16.s\n"
"add x14, x14, %x[ld_in_col], LSL #2\n"
"ld1w { z16.s }, p1/Z, [x20]\n"
"add x20, x20, %x[ld_in_row], LSL #2\n"
".inst 0x648aaa0d // bfcvtnt z13.h, p2/M, z16.s\n"
"ld1w { z16.s }, p1/Z, [x20]\n"
"add x20, x20, %x[ld_in_row], LSL #2\n"
".inst 0x658aaa0e // bfcvt z14.h, p2/M, z16.s\n"
"ld1w { z16.s }, p1/Z, [x20]\n"
"add x20, x20, %x[ld_in_row], LSL #2\n"
".inst 0x648aaa0e // bfcvtnt z14.h, p2/M, z16.s\n"
".inst 0xc12111b0 // bfdot za.s[x8, 0], { z13.h-z14.h }, z1.h\n"
"ld1w { z16.s }, p1/Z, [x20]\n"
"add x20, x20, %x[ld_in_row], LSL #2\n"
".inst 0x658aaa0f // bfcvt z15.h, p2/M, z16.s\n"
".inst 0xc12211b1 // bfdot za.s[x8, 1], { z13.h-z14.h }, z2.h\n"
"ld1w { z16.s }, p1/Z, [x20]\n"
".inst 0x648aaa0f // bfcvtnt z15.h, p2/M, z16.s\n"
".inst 0xc12811b2 // bfdot za.s[x8, 2], { z13.h-z14.h }, z8.h\n"
".inst 0xc12911b3 // bfdot za.s[x8, 3], { z13.h-z14.h }, z9.h\n"
".inst 0xc12a11d0 // bfdot za.s[x8, 0], { z14.h-z15.h }, z10.h\n"
".inst 0xc12c11d1 // bfdot za.s[x8, 1], { z14.h-z15.h }, z12.h\n"
".inst 0xc12511d2 // bfdot za.s[x8, 2], { z14.h-z15.h }, z5.h\n"
".inst 0xc12611d3 // bfdot za.s[x8, 3], { z14.h-z15.h }, z6.h\n"
"8:" // Unpadded: 0 priming loads
"cbz x15, 16f\n"
"add x20, x14, %x[ld_in_row], LSL #2\n"
"ld1w { z16.s }, p1/Z, [x14]\n"
".inst 0x658aaa0d // bfcvt z13.h, p2/M, z16.s\n"
"sub x15, x15, #0x1\n"
"ld1w { z16.s }, p1/Z, [x20]\n"
"add x20, x20, %x[ld_in_row], LSL #2\n"
"sub x13, x13, #0x1\n"
".inst 0x648aaa0d // bfcvtnt z13.h, p2/M, z16.s\n"
"ld1w { z16.s }, p1/Z, [x20]\n"
"add x20, x20, %x[ld_in_row], LSL #2\n"
".inst 0x658aaa0e // bfcvt z14.h, p2/M, z16.s\n"
"cmp x15, x13\n"
"ld1w { z16.s }, p1/Z, [x20]\n"
"add x20, x20, %x[ld_in_row], LSL #2\n"
"csel x21, x15, x13, LT\n"
".inst 0x648aaa0e // bfcvtnt z14.h, p2/M, z16.s\n"
"ld1w { z16.s }, p1/Z, [x20]\n"
"add x20, x20, %x[ld_in_row], LSL #2\n"
".inst 0x658aaa0f // bfcvt z15.h, p2/M, z16.s\n"
"add x14, x14, %x[ld_in_col], LSL #2\n"
"ld1w { z16.s }, p1/Z, [x20]\n"
".inst 0x648aaa0f // bfcvtnt z15.h, p2/M, z16.s\n"
"sub x13, x13, x21\n"
"cbz x21, 15f\n"
"9:" // Unpadded: Main loop
".inst 0xc12411b0 // bfdot za.s[x8, 0], { z13.h-z14.h }, z4.h\n"
"add x20, x14, %x[ld_in_row], LSL #2\n"
"ld1w { z21.s }, p1/Z, [x14]\n"
"subs x21, x21, #0x1\n"
".inst 0xc12711b1 // bfdot za.s[x8, 1], { z13.h-z14.h }, z7.h\n"
"ld1w { z20.s }, p1/Z, [x20]\n"
"add x20, x20, %x[ld_in_row], LSL #2\n"
"add x14, x14, %x[ld_in_col], LSL #2\n"
"ld1w { z19.s }, p1/Z, [x20]\n"
"add x20, x20, %x[ld_in_row], LSL #2\n"
".inst 0xc12311d0 // bfdot za.s[x8, 0], { z14.h-z15.h }, z3.h\n"
".inst 0xc12011d1 // bfdot za.s[x8, 1], { z14.h-z15.h }, z0.h\n"
"ld1w { z18.s }, p1/Z, [x20]\n"
"add x20, x20, %x[ld_in_row], LSL #2\n"
".inst 0xc12111b2 // bfdot za.s[x8, 2], { z13.h-z14.h }, z1.h\n"
"ld1w { z17.s }, p1/Z, [x20]\n"
"add x20, x20, %x[ld_in_row], LSL #2\n"
".inst 0xc12211b3 // bfdot za.s[x8, 3], { z13.h-z14.h }, z2.h\n"
"ld1w { z16.s }, p1/Z, [x20]\n"
".inst 0xc12811b4 // bfdot za.s[x8, 4], { z13.h-z14.h }, z8.h\n"
".inst 0xc12911b5 // bfdot za.s[x8, 5], { z13.h-z14.h }, z9.h\n"
".inst 0x658aaaad // bfcvt z13.h, p2/M, z21.s\n"
".inst 0x648aaa8d // bfcvtnt z13.h, p2/M, z20.s\n"
".inst 0xc12a11d2 // bfdot za.s[x8, 2], { z14.h-z15.h }, z10.h\n"
".inst 0xc12c11d3 // bfdot za.s[x8, 3], { z14.h-z15.h }, z12.h\n"
".inst 0xc12511d4 // bfdot za.s[x8, 4], { z14.h-z15.h }, z5.h\n"
".inst 0xc12611d5 // bfdot za.s[x8, 5], { z14.h-z15.h }, z6.h\n"
".inst 0x658aaa6e // bfcvt z14.h, p2/M, z19.s\n"
".inst 0x658aaa2f // bfcvt z15.h, p2/M, z17.s\n"
".inst 0xc0060818 // mova { z24.d-z25.d }, za.d[x8, #0]\n"
".inst 0x648aaa4e // bfcvtnt z14.h, p2/M, z18.s\n"
".inst 0x648aaa0f // bfcvtnt z15.h, p2/M, z16.s\n"
".inst 0xc006083a // mova { z26.d-z27.d }, za.d[x8, #1]\n"
"add x8, x8, #0x2\n"
".inst 0xc1bdcb98 // fclamp { z24.s-z27.s }, z28.s, z29.s\n"
"st1w { z24.s }, p1, [x9]\n"
"add x9, x9, x27, LSL #2\n"
"st1w { z26.s }, p1, [x28]\n"
"add x28, x28, x26, LSL #2\n"
".inst 0xc0040ac4 // mova za.d[x8, #4], { z22.d-z23.d }\n"
"st1w { z25.s }, p1, [x25]\n"
"add x25, x25, x23, LSL #2\n"
".inst 0xc0040ac5 // mova za.d[x8, #5], { z22.d-z23.d }\n"
"st1w { z27.s }, p1, [x24]\n"
"add x24, x24, x22, LSL #2\n"
"bgt 9b\n"
"b 15f\n"
"10:" // Padded
"cbz x10, 13f\n"
"cmp x10, #0x1\n"
"sub x15, x15, x10\n"
"beq 12f\n"
"11:" // Padded: 2 priming loads
"mov x12, #0x0\n"
".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n"
"ld1w { z16.s }, p0/Z, [x14]\n"
".inst 0x658aaa0d // bfcvt z13.h, p2/M, z16.s\n"
"add x20, x14, %x[ld_in_row], LSL #2\n"
".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n"
"ld1w { z16.s }, p0/Z, [x20]\n"
".inst 0x648aaa0d // bfcvtnt z13.h, p2/M, z16.s\n"
"add x20, x20, %x[ld_in_row], LSL #2\n"
".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n"
"ld1w { z16.s }, p0/Z, [x20]\n"
".inst 0x658aaa0e // bfcvt z14.h, p2/M, z16.s\n"
"add x20, x20, %x[ld_in_row], LSL #2\n"
".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n"
"ld1w { z16.s }, p0/Z, [x20]\n"
".inst 0x648aaa0e // bfcvtnt z14.h, p2/M, z16.s\n"
"mov x12, #0x4\n"
"add x20, x20, %x[ld_in_row], LSL #2\n"
".inst 0xc12811b0 // bfdot za.s[x8, 0], { z13.h-z14.h }, z8.h\n"
".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n"
"ld1w { z16.s }, p0/Z, [x20]\n"
"add x20, x20, %x[ld_in_row], LSL #2\n"
".inst 0x658aaa0f // bfcvt z15.h, p2/M, z16.s\n"
".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n"
"ld1w { z16.s }, p0/Z, [x20]\n"
".inst 0x648aaa0f // bfcvtnt z15.h, p2/M, z16.s\n"
".inst 0xc12911b1 // bfdot za.s[x8, 1], { z13.h-z14.h }, z9.h\n"
"add x14, x14, %x[ld_in_col], LSL #2\n"
".inst 0xc12511d0 // bfdot za.s[x8, 0], { z14.h-z15.h }, z5.h\n"
".inst 0xc12611d1 // bfdot za.s[x8, 1], { z14.h-z15.h }, z6.h\n"
"12:" // Padded: 1 priming loads
"mov x12, #0x0\n"
".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n"
"ld1w { z16.s }, p0/Z, [x14]\n"
".inst 0x658aaa0d // bfcvt z13.h, p2/M, z16.s\n"
"add x20, x14, %x[ld_in_row], LSL #2\n"
".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n"
"ld1w { z16.s }, p0/Z, [x20]\n"
".inst 0x648aaa0d // bfcvtnt z13.h, p2/M, z16.s\n"
"add x20, x20, %x[ld_in_row], LSL #2\n"
".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n"
"ld1w { z16.s }, p0/Z, [x20]\n"
".inst 0x658aaa0e // bfcvt z14.h, p2/M, z16.s\n"
"add x20, x20, %x[ld_in_row], LSL #2\n"
".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n"
"ld1w { z16.s }, p0/Z, [x20]\n"
".inst 0x648aaa0e // bfcvtnt z14.h, p2/M, z16.s\n"
"mov x12, #0x4\n"
"add x20, x20, %x[ld_in_row], LSL #2\n"
".inst 0xc12111b0 // bfdot za.s[x8, 0], { z13.h-z14.h }, z1.h\n"
".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n"
"ld1w { z16.s }, p0/Z, [x20]\n"
"add x20, x20, %x[ld_in_row], LSL #2\n"
".inst 0x658aaa0f // bfcvt z15.h, p2/M, z16.s\n"
".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n"
"ld1w { z16.s }, p0/Z, [x20]\n"
".inst 0x648aaa0f // bfcvtnt z15.h, p2/M, z16.s\n"
".inst 0xc12211b1 // bfdot za.s[x8, 1], { z13.h-z14.h }, z2.h\n"
".inst 0xc12811b2 // bfdot za.s[x8, 2], { z13.h-z14.h }, z8.h\n"
"add x14, x14, %x[ld_in_col], LSL #2\n"
".inst 0xc12911b3 // bfdot za.s[x8, 3], { z13.h-z14.h }, z9.h\n"
".inst 0xc12a11d0 // bfdot za.s[x8, 0], { z14.h-z15.h }, z10.h\n"
".inst 0xc12c11d1 // bfdot za.s[x8, 1], { z14.h-z15.h }, z12.h\n"
".inst 0xc12511d2 // bfdot za.s[x8, 2], { z14.h-z15.h }, z5.h\n"
".inst 0xc12611d3 // bfdot za.s[x8, 3], { z14.h-z15.h }, z6.h\n"
"13:" // Padded: 0 priming loads
"cbz x15, 16f\n"
"mov x12, #0x0\n"
".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n"
"ld1w { z16.s }, p0/Z, [x14]\n"
".inst 0x658aaa0d // bfcvt z13.h, p2/M, z16.s\n"
"add x20, x14, %x[ld_in_row], LSL #2\n"
".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n"
"ld1w { z16.s }, p0/Z, [x20]\n"
".inst 0x648aaa0d // bfcvtnt z13.h, p2/M, z16.s\n"
"add x20, x20, %x[ld_in_row], LSL #2\n"
".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n"
"ld1w { z16.s }, p0/Z, [x20]\n"
".inst 0x658aaa0e // bfcvt z14.h, p2/M, z16.s\n"
"add x20, x20, %x[ld_in_row], LSL #2\n"
".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n"
"ld1w { z16.s }, p0/Z, [x20]\n"
".inst 0x648aaa0e // bfcvtnt z14.h, p2/M, z16.s\n"
"mov x12, #0x4\n"
"add x20, x20, %x[ld_in_row], LSL #2\n"
".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n"
"ld1w { z16.s }, p0/Z, [x20]\n"
"add x20, x20, %x[ld_in_row], LSL #2\n"
".inst 0x658aaa0f // bfcvt z15.h, p2/M, z16.s\n"
".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n"
"ld1w { z16.s }, p0/Z, [x20]\n"
"sub x15, x15, #0x1\n"
".inst 0x648aaa0f // bfcvtnt z15.h, p2/M, z16.s\n"
"sub x13, x13, #0x1\n"
"cmp x15, x13\n"
"csel x21, x15, x13, LT\n"
"add x14, x14, %x[ld_in_col], LSL #2\n"
"sub x13, x13, x21\n"
"cbz x21, 15f\n"
"14:" // Padded: Main loop
"mov x12, #0x0\n"
".inst 0xc12411b0 // bfdot za.s[x8, 0], { z13.h-z14.h }, z4.h\n"
".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n"
"ld1w { z21.s }, p0/Z, [x14]\n"
".inst 0xc12711b1 // bfdot za.s[x8, 1], { z13.h-z14.h }, z7.h\n"
"add x20, x14, %x[ld_in_row], LSL #2\n"
".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n"
"ld1w { z20.s }, p0/Z, [x20]\n"
"add x20, x20, %x[ld_in_row], LSL #2\n"
".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n"
"ld1w { z19.s }, p0/Z, [x20]\n"
".inst 0xc12311d0 // bfdot za.s[x8, 0], { z14.h-z15.h }, z3.h\n"
"add x20, x20, %x[ld_in_row], LSL #2\n"
".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n"
".inst 0xc12011d1 // bfdot za.s[x8, 1], { z14.h-z15.h }, z0.h\n"
"ld1w { z18.s }, p0/Z, [x20]\n"
"mov x12, #0x4\n"
".inst 0xc12111b2 // bfdot za.s[x8, 2], { z13.h-z14.h }, z1.h\n"
"add x20, x20, %x[ld_in_row], LSL #2\n"
".inst 0xc12211b3 // bfdot za.s[x8, 3], { z13.h-z14.h }, z2.h\n"
".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n"
"ld1w { z17.s }, p0/Z, [x20]\n"
"add x20, x20, %x[ld_in_row], LSL #2\n"
".inst 0xc12811b4 // bfdot za.s[x8, 4], { z13.h-z14.h }, z8.h\n"
".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n"
"ld1w { z16.s }, p0/Z, [x20]\n"
"subs x21, x21, #0x1\n"
".inst 0xc12911b5 // bfdot za.s[x8, 5], { z13.h-z14.h }, z9.h\n"
".inst 0x658aaaad // bfcvt z13.h, p2/M, z21.s\n"
".inst 0x648aaa8d // bfcvtnt z13.h, p2/M, z20.s\n"
"add x14, x14, %x[ld_in_col], LSL #2\n"
".inst 0xc12a11d2 // bfdot za.s[x8, 2], { z14.h-z15.h }, z10.h\n"
".inst 0xc12c11d3 // bfdot za.s[x8, 3], { z14.h-z15.h }, z12.h\n"
".inst 0xc12511d4 // bfdot za.s[x8, 4], { z14.h-z15.h }, z5.h\n"
".inst 0xc12611d5 // bfdot za.s[x8, 5], { z14.h-z15.h }, z6.h\n"
".inst 0x658aaa6e // bfcvt z14.h, p2/M, z19.s\n"
".inst 0x658aaa2f // bfcvt z15.h, p2/M, z17.s\n"
".inst 0xc0060818 // mova { z24.d-z25.d }, za.d[x8, #0]\n"
".inst 0x648aaa4e // bfcvtnt z14.h, p2/M, z18.s\n"
".inst 0x648aaa0f // bfcvtnt z15.h, p2/M, z16.s\n"
".inst 0xc006083a // mova { z26.d-z27.d }, za.d[x8, #1]\n"
"add x8, x8, #0x2\n"
".inst 0xc1bdcb98 // fclamp { z24.s-z27.s }, z28.s, z29.s\n"
"st1w { z24.s }, p1, [x9]\n"
"add x9, x9, x27, LSL #2\n"
"st1w { z26.s }, p1, [x28]\n"
"add x28, x28, x26, LSL #2\n"
".inst 0xc0040ac4 // mova za.d[x8, #4], { z22.d-z23.d }\n"
"st1w { z25.s }, p1, [x25]\n"
"add x25, x25, x23, LSL #2\n"
".inst 0xc0040ac5 // mova za.d[x8, #5], { z22.d-z23.d }\n"
"st1w { z27.s }, p1, [x24]\n"
"add x24, x24, x22, LSL #2\n"
"bgt 14b\n"
"15:" // Main loop tail
".inst 0xc12411b0 // bfdot za.s[x8, 0], { z13.h-z14.h }, z4.h\n"
".inst 0xc12711b1 // bfdot za.s[x8, 1], { z13.h-z14.h }, z7.h\n"
".inst 0xc12311d0 // bfdot za.s[x8, 0], { z14.h-z15.h }, z3.h\n"
".inst 0xc12011d1 // bfdot za.s[x8, 1], { z14.h-z15.h }, z0.h\n"
".inst 0xc12111b2 // bfdot za.s[x8, 2], { z13.h-z14.h }, z1.h\n"
".inst 0xc12211b3 // bfdot za.s[x8, 3], { z13.h-z14.h }, z2.h\n"
".inst 0xc12811b4 // bfdot za.s[x8, 4], { z13.h-z14.h }, z8.h\n"
".inst 0xc12911b5 // bfdot za.s[x8, 5], { z13.h-z14.h }, z9.h\n"
".inst 0xc0060818 // mova { z24.d-z25.d }, za.d[x8, #0]\n"
".inst 0xc006083a // mova { z26.d-z27.d }, za.d[x8, #1]\n"
".inst 0xc1bdcb98 // fclamp { z24.s-z27.s }, z28.s, z29.s\n"
"st1w { z24.s }, p1, [x9]\n"
"add x9, x9, x27, LSL #2\n"
".inst 0xc12a11d2 // bfdot za.s[x8, 2], { z14.h-z15.h }, z10.h\n"
"st1w { z26.s }, p1, [x28]\n"
"add x28, x28, x26, LSL #2\n"
".inst 0xc12c11d3 // bfdot za.s[x8, 3], { z14.h-z15.h }, z12.h\n"
"st1w { z25.s }, p1, [x25]\n"
"add x25, x25, x23, LSL #2\n"
".inst 0xc12511d4 // bfdot za.s[x8, 4], { z14.h-z15.h }, z5.h\n"
"st1w { z27.s }, p1, [x24]\n"
"add x24, x24, x22, LSL #2\n"
".inst 0xc12611d5 // bfdot za.s[x8, 5], { z14.h-z15.h }, z6.h\n"
"add x8, x8, #0x2\n"
".inst 0xc0040ac4 // mova za.d[x8, #4], { z22.d-z23.d }\n"
".inst 0xc0040ac5 // mova za.d[x8, #5], { z22.d-z23.d }\n"
"16:" // Main loop skip tail
"cbz x13, 18f\n"
"17:" // Right padding loop
".inst 0xc0060818 // mova { z24.d-z25.d }, za.d[x8, #0]\n"
"subs x13, x13, #0x1\n"
".inst 0xc006083a // mova { z26.d-z27.d }, za.d[x8, #1]\n"
"add x8, x8, #0x2\n"
".inst 0xc1bdcb98 // fclamp { z24.s-z27.s }, z28.s, z29.s\n"
"st1w { z24.s }, p1, [x9]\n"
"add x9, x9, x27, LSL #2\n"
"st1w { z26.s }, p1, [x28]\n"
"add x28, x28, x26, LSL #2\n"
".inst 0xc0040ac4 // mova za.d[x8, #4], { z22.d-z23.d }\n"
"st1w { z25.s }, p1, [x25]\n"
"add x25, x25, x23, LSL #2\n"
".inst 0xc0040ac5 // mova za.d[x8, #5], { z22.d-z23.d }\n"
"st1w { z27.s }, p1, [x24]\n"
"add x24, x24, x22, LSL #2\n"
"bgt 17b\n"
"18:" // End
"ldr x20, [%x[args], %[offsetof_Args_weights]]\n"
"incb x20, ALL, MUL #9\n"
"str x20, [%x[args], %[offsetof_Args_weights]]\n"
"incw x16\n"
"ldr x20, [%x[args], %[offsetof_Args_ld_in_vl]]\n"
"whilelt p1.s, x16, x17\n"
"ldr x14, [%x[args], %[offsetof_Args_inptr]]\n"
"add x14, x14, x20, LSL #2\n"
"str x14, [%x[args], %[offsetof_Args_inptr]]\n"
"ldr x11, [%x[args], %[offsetof_Args_outptrs]]\n"
"ldr x24, [%x[args], %[offsetof_Args_ld_out_vls]]\n"
"ldp x23, x22, [x11, #0x0]\n"
"ldp x21, x20, [x24, #0x0]\n"
"add x23, x23, x21, LSL #2\n"
"add x22, x22, x20, LSL #2\n"
"stp x23, x22, [x11, #0x0]\n"
"ldp x23, x22, [x11, #0x10]\n"
"ldp x21, x20, [x24, #0x10]\n"
"add x23, x23, x21, LSL #2\n"
"add x22, x22, x20, LSL #2\n"
"stp x23, x22, [x11, #0x10]\n"
"b.any 1b\n"
".inst 0xd503467f // SMSTOP\n"
:
: [args] "r" (&args), [ld_in_col] "r" (ld_in_col), [ld_in_row] "r" (ld_in_row), [offsetof_Args_bias] "I" (offsetof(Args, bias)), [offsetof_Args_clamp_max] "I" (offsetof(Args, clamp_max)), [offsetof_Args_clamp_min] "I" (offsetof(Args, clamp_min)), [offsetof_Args_current_channel] "I" (offsetof(Args, current_channel)), [offsetof_Args_inptr] "I" (offsetof(Args, inptr)), [offsetof_Args_input_cols] "I" (offsetof(Args, input_cols)), [offsetof_Args_ld_in_vl] "I" (offsetof(Args, ld_in_vl)), [offsetof_Args_ld_out_cols] "I" (offsetof(Args, ld_out_cols)), [offsetof_Args_ld_out_vls] "I" (offsetof(Args, ld_out_vls)), [offsetof_Args_n_channels] "I" (offsetof(Args, n_channels)), [offsetof_Args_outptrs] "I" (offsetof(Args, outptrs)), [offsetof_Args_output_cols] "I" (offsetof(Args, output_cols)), [offsetof_Args_pad_bottom] "I" (offsetof(Args, pad_bottom)), [offsetof_Args_pad_left] "I" (offsetof(Args, pad_left)), [offsetof_Args_pad_top] "I" (offsetof(Args, pad_top)), [offsetof_Args_weights] "I" (offsetof(Args, weights))
: "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x6", "x7", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
);
}
} // namespace depthwise
} // namespace arm_conv
#endif // defined(ARM_COMPUTE_ENABLE_SME2)
|
