27 #include "../../utils.hpp"
34 void a64_hybrid_fp32_mla_8x4_a55 (
35 unsigned int num_strings,
const unsigned int *string_lengths, IndirectInputArg<float> A_arg,
36 size_t M,
size_t N,
const float *B_ptr, IndirectOutputArg<float> output_arg,
41 float maxval =
static_cast<float>(std::numeric_limits<float>::infinity());
42 float minval = -
static_cast<float>(std::numeric_limits<float>::infinity());
43 unsigned int num_strings = {};
44 const unsigned int *string_lengths = {};
46 const float *B_ptr = {};
47 size_t output_offset = {};
48 size_t input_initial_col = {};
49 size_t input_offset = {};
52 unsigned long flags=0;
56 if (output_arg.is_indirect) {
57 output_ptr=(
void *)(output_arg.indirect.ptr);
58 ka.output_offset=output_arg.indirect.offset;
61 output_ptr=(
void *)(output_arg.direct.base);
62 ka.output_offset=output_arg.direct.stride;
65 if (A_arg.is_indirect) {
66 input_ptr=(
void *)(A_arg.indirect.ptr);
67 ka.input_offset=A_arg.indirect.start_row;
68 ka.input_initial_col=A_arg.indirect.start_col;
71 assert(num_strings==1);
72 input_ptr=(
void *)(A_arg.direct.base);
73 ka.input_offset=A_arg.direct.stride;
78 ka.num_strings = num_strings;
79 ka.string_lengths = string_lengths;
87 ka.maxval =
static_cast<float>(act.param1);
108 "ldr x4, [%x[args_ptr], %[offsetof_N]]\n"
109 "ldr x5, [%x[args_ptr], %[offsetof_B_ptr]]\n"
110 "mov x6, %x[output_ptr]\n"
113 "ldr q24, [x3, #0x0]\n"
114 "add x3, x3, #0x10\n"
117 "tbz %x[flags], #0, 7f\n"
121 "ldr d24, [x6], #0x8\n"
124 "ld1 { v24.s }[2], [x6]\n"
127 "ldr s24, [x6, #0x0]\n"
133 "ldr q24, [x6, #0x0]\n"
136 "movi v24.16b, #0x0\n"
140 "ldr x26, [%x[args_ptr], %[offsetof_string_lengths]]\n"
141 "ldr w8, [x26, x7, LSL #0x2]\n"
142 "ldr x27, [%x[args_ptr], %[offsetof_input_offset]]\n"
143 "tbz %x[flags], #3, 10f\n"
144 "ldr x26, [%x[input_ptr], x7, LSL #0x3]\n"
145 "add x26, x26, x27, LSL #3\n"
146 "ldr x17, [x26, #0x0]\n"
148 "ldr x26, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
149 "add x17, x17, x26, LSL #2\n"
152 "mov x17, %x[input_ptr]\n"
156 "ldr q0, [x17, #0x0]\n"
158 "ldr q8, [x5, #0x0]\n"
159 "ldr q9, [x5, #0x10]\n"
160 "ldr q10, [x5, #0x20]\n"
161 "ldr q11, [x5, #0x30]\n"
164 "fmla v24.4s, v8.4s, v0.s[0]\n"
165 "add x17, x17, #0x10\n"
166 "fmla v24.4s, v9.4s, v0.s[1]\n"
167 "add x5, x5, #0x40\n"
168 "ldr d8, [x5, #0x0]\n"
169 "fmla v24.4s, v10.4s, v0.s[2]\n"
170 "ldr d9, [x5, #0x10]\n"
171 "fmla v24.4s, v11.4s, v0.s[3]\n"
172 "ldr d0, [x17, #0x0]\n"
174 "ldr d10, [x5, #0x20]\n"
176 "ldr d11, [x5, #0x30]\n"
177 "ldr x26, [x5, #0x8]\n"
179 "ldr x26, [x5, #0x18]\n"
181 "ldr x26, [x17, #0x8]\n"
183 "ldr x26, [x5, #0x28]\n"
184 "mov v10.d[1], x26\n"
185 "ldr x26, [x5, #0x38]\n"
186 "mov v11.d[1], x26\n"
187 "prfm pldl1keep, [x17, #0x80]\n"
190 "fmla v24.4s, v8.4s, v0.s[0]\n"
191 "add x17, x17, #0x10\n"
192 "fmla v24.4s, v9.4s, v0.s[1]\n"
194 "fmla v24.4s, v10.4s, v0.s[2]\n"
195 "prfm pldl1keep, [x17, #0x80]\n"
196 "fmla v24.4s, v11.4s, v0.s[3]\n"
197 "add x5, x5, #0x40\n"
201 "ldr s17, [x17], #0x4\n"
203 "ldr q16, [x5, #0x0]\n"
204 "fmla v24.4s, v16.4s, v17.s[0]\n"
205 "add x5, x5, #0x10\n"
208 "ldr w26, [%x[args_ptr], %[offsetof_num_strings]]\n"
212 "prfm pstl1keep, [x6, #0x0]\n"
213 "tbz %x[flags], #1, 17f\n"
214 "add x26, %x[args_ptr], %[offset_max]\n"
215 "ld1r { v16.4s }, [x26]\n"
216 "fmin v24.4s, v24.4s, v16.4s\n"
217 "add x26, %x[args_ptr], %[offset_min]\n"
218 "ld1r { v16.4s }, [x26]\n"
219 "fmax v24.4s, v24.4s, v16.4s\n"
224 "str d24, [x6], #0x8\n"
226 "st1 { v24.s }[2], [x6]\n"
229 "str s24, [x6, #0x0]\n"
233 "str q24, [x6, #0x0]\n"
234 "add x6, x6, #0x10\n"
236 "subs x4, x4, #0x4\n"
241 "ldr x4, [%x[args_ptr], %[offsetof_N]]\n"
242 "ldr x5, [%x[args_ptr], %[offsetof_B_ptr]]\n"
243 "mov x6, %x[output_ptr]\n"
246 "ldr q24, [x3, #0x0]\n"
247 "mov v25.16b, v24.16b\n"
248 "add x3, x3, #0x10\n"
251 "tbz %x[flags], #0, 28f\n"
252 "ldr x26, [%x[args_ptr], %[offsetof_output_offset]]\n"
254 "add x13, x6, x26, LSL #2\n"
257 "ldr d24, [x6], #0x8\n"
259 "ldr d25, [x13], #0x8\n"
261 "ld1 { v24.s }[2], [x6]\n"
262 "ld1 { v25.s }[2], [x13]\n"
265 "ldr s24, [x6, #0x0]\n"
267 "ldr s25, [x13, #0x0]\n"
272 "ldr q24, [x6, #0x0]\n"
273 "ldr q25, [x13, #0x0]\n"
276 "movi v24.16b, #0x0\n"
277 "movi v25.16b, #0x0\n"
281 "ldr x26, [%x[args_ptr], %[offsetof_string_lengths]]\n"
282 "ldr w8, [x26, x7, LSL #0x2]\n"
283 "ldr x27, [%x[args_ptr], %[offsetof_input_offset]]\n"
284 "tbz %x[flags], #3, 31f\n"
285 "ldr x26, [%x[input_ptr], x7, LSL #0x3]\n"
286 "add x26, x26, x27, LSL #3\n"
287 "ldr x17, [x26, #0x0]\n"
288 "ldr x16, [x26, #0x8]\n"
290 "ldr x26, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
291 "add x17, x17, x26, LSL #2\n"
292 "add x16, x16, x26, LSL #2\n"
295 "mov x17, %x[input_ptr]\n"
296 "add x16, x17, x27, LSL #2\n"
300 "ldr q0, [x17, #0x0]\n"
302 "ldr q1, [x16, #0x0]\n"
303 "ldr q8, [x5, #0x0]\n"
304 "ldr q9, [x5, #0x10]\n"
305 "ldr q10, [x5, #0x20]\n"
306 "ldr q11, [x5, #0x30]\n"
309 "fmla v24.4s, v8.4s, v0.s[0]\n"
310 "add x17, x17, #0x10\n"
311 "fmla v25.4s, v8.4s, v1.s[0]\n"
312 "add x16, x16, #0x10\n"
313 "fmla v24.4s, v9.4s, v0.s[1]\n"
314 "add x5, x5, #0x40\n"
315 "ldr d8, [x5, #0x0]\n"
316 "fmla v25.4s, v9.4s, v1.s[1]\n"
317 "ldr d9, [x5, #0x10]\n"
318 "fmla v24.4s, v10.4s, v0.s[2]\n"
319 "fmla v25.4s, v10.4s, v1.s[2]\n"
320 "ldr d10, [x5, #0x20]\n"
321 "ldr x27, [x5, #0x8]\n"
322 "fmla v24.4s, v11.4s, v0.s[3]\n"
323 "ldr d0, [x17, #0x0]\n"
324 "fmla v25.4s, v11.4s, v1.s[3]\n"
325 "ldr d1, [x16, #0x0]\n"
327 "ldr d11, [x5, #0x30]\n"
329 "ldr x26, [x5, #0x18]\n"
331 "ldr x27, [x17, #0x8]\n"
333 "ldr x26, [x16, #0x8]\n"
335 "ldr x27, [x5, #0x28]\n"
337 "ldr x26, [x5, #0x38]\n"
338 "mov v10.d[1], x27\n"
339 "mov v11.d[1], x26\n"
340 "prfm pldl1keep, [x17, #0x80]\n"
341 "prfm pldl1keep, [x16, #0x80]\n"
344 "fmla v24.4s, v8.4s, v0.s[0]\n"
345 "add x17, x17, #0x10\n"
346 "fmla v25.4s, v8.4s, v1.s[0]\n"
347 "add x16, x16, #0x10\n"
348 "fmla v24.4s, v9.4s, v0.s[1]\n"
350 "fmla v25.4s, v9.4s, v1.s[1]\n"
351 "prfm pldl1keep, [x17, #0x80]\n"
352 "fmla v24.4s, v10.4s, v0.s[2]\n"
353 "prfm pldl1keep, [x16, #0x80]\n"
354 "fmla v25.4s, v10.4s, v1.s[2]\n"
355 "add x5, x5, #0x40\n"
356 "fmla v24.4s, v11.4s, v0.s[3]\n"
357 "fmla v25.4s, v11.4s, v1.s[3]\n"
361 "ldr s18, [x17], #0x4\n"
363 "ldr s17, [x16], #0x4\n"
364 "ldr q16, [x5, #0x0]\n"
365 "fmla v24.4s, v16.4s, v18.s[0]\n"
366 "fmla v25.4s, v16.4s, v17.s[0]\n"
367 "add x5, x5, #0x10\n"
370 "ldr w26, [%x[args_ptr], %[offsetof_num_strings]]\n"
374 "ldr x26, [%x[args_ptr], %[offsetof_output_offset]]\n"
375 "add x13, x6, x26, LSL #2\n"
376 "prfm pstl1keep, [x6, #0x0]\n"
377 "prfm pstl1keep, [x13, #0x0]\n"
378 "tbz %x[flags], #1, 38f\n"
379 "add x26, %x[args_ptr], %[offset_max]\n"
380 "ld1r { v16.4s }, [x26]\n"
381 "fmin v24.4s, v24.4s, v16.4s\n"
382 "fmin v25.4s, v25.4s, v16.4s\n"
383 "add x26, %x[args_ptr], %[offset_min]\n"
384 "ld1r { v16.4s }, [x26]\n"
385 "fmax v24.4s, v24.4s, v16.4s\n"
386 "fmax v25.4s, v25.4s, v16.4s\n"
391 "str d24, [x6], #0x8\n"
392 "str d25, [x13], #0x8\n"
394 "st1 { v24.s }[2], [x6]\n"
395 "st1 { v25.s }[2], [x13]\n"
398 "str s24, [x6, #0x0]\n"
399 "str s25, [x13, #0x0]\n"
403 "str q24, [x6, #0x0]\n"
404 "add x6, x6, #0x10\n"
405 "str q25, [x13, #0x0]\n"
407 "subs x4, x4, #0x4\n"
412 "ldr x4, [%x[args_ptr], %[offsetof_N]]\n"
413 "ldr x5, [%x[args_ptr], %[offsetof_B_ptr]]\n"
414 "mov x6, %x[output_ptr]\n"
417 "ldr q24, [x3, #0x0]\n"
418 "mov v25.16b, v24.16b\n"
419 "mov v26.16b, v24.16b\n"
420 "add x3, x3, #0x10\n"
423 "tbz %x[flags], #0, 49f\n"
424 "ldr x26, [%x[args_ptr], %[offsetof_output_offset]]\n"
425 "add x13, x6, x26, LSL #2\n"
427 "add x12, x13, x26, LSL #2\n"
430 "ldr d24, [x6], #0x8\n"
432 "ldr d25, [x13], #0x8\n"
433 "ldr d26, [x12], #0x8\n"
435 "ld1 { v24.s }[2], [x6]\n"
436 "ld1 { v25.s }[2], [x13]\n"
437 "ld1 { v26.s }[2], [x12]\n"
440 "ldr s24, [x6, #0x0]\n"
442 "ldr s25, [x13, #0x0]\n"
443 "ldr s26, [x12, #0x0]\n"
448 "ldr q24, [x6, #0x0]\n"
449 "ldr q25, [x13, #0x0]\n"
450 "ldr q26, [x12, #0x0]\n"
453 "movi v24.16b, #0x0\n"
454 "movi v25.16b, #0x0\n"
455 "movi v26.16b, #0x0\n"
459 "ldr x26, [%x[args_ptr], %[offsetof_string_lengths]]\n"
460 "ldr w8, [x26, x7, LSL #0x2]\n"
461 "ldr x27, [%x[args_ptr], %[offsetof_input_offset]]\n"
462 "tbz %x[flags], #3, 52f\n"
463 "ldr x26, [%x[input_ptr], x7, LSL #0x3]\n"
464 "add x26, x26, x27, LSL #3\n"
465 "ldr x17, [x26, #0x0]\n"
466 "ldr x16, [x26, #0x8]\n"
467 "ldr x15, [x26, #0x10]\n"
469 "ldr x26, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
470 "add x17, x17, x26, LSL #2\n"
471 "add x16, x16, x26, LSL #2\n"
472 "add x15, x15, x26, LSL #2\n"
475 "mov x17, %x[input_ptr]\n"
476 "add x16, x17, x27, LSL #2\n"
477 "add x15, x16, x27, LSL #2\n"
481 "ldr q0, [x17, #0x0]\n"
483 "ldr q1, [x16, #0x0]\n"
484 "ldr q2, [x15, #0x0]\n"
485 "ldr q8, [x5, #0x0]\n"
486 "ldr q9, [x5, #0x10]\n"
487 "ldr q10, [x5, #0x20]\n"
488 "ldr q11, [x5, #0x30]\n"
491 "fmla v24.4s, v8.4s, v0.s[0]\n"
492 "add x17, x17, #0x10\n"
493 "fmla v25.4s, v8.4s, v1.s[0]\n"
494 "add x16, x16, #0x10\n"
495 "fmla v26.4s, v8.4s, v2.s[0]\n"
496 "add x15, x15, #0x10\n"
497 "fmla v24.4s, v9.4s, v0.s[1]\n"
498 "add x5, x5, #0x40\n"
499 "ldr d8, [x5, #0x0]\n"
500 "fmla v25.4s, v9.4s, v1.s[1]\n"
501 "fmla v26.4s, v9.4s, v2.s[1]\n"
502 "ldr d9, [x5, #0x10]\n"
503 "fmla v24.4s, v10.4s, v0.s[2]\n"
504 "ldr x28, [x5, #0x8]\n"
505 "fmla v25.4s, v10.4s, v1.s[2]\n"
506 "ldr x27, [x5, #0x18]\n"
507 "fmla v26.4s, v10.4s, v2.s[2]\n"
508 "ldr d10, [x5, #0x20]\n"
509 "ldr x26, [x5, #0x28]\n"
510 "fmla v24.4s, v11.4s, v0.s[3]\n"
511 "ldr d0, [x17, #0x0]\n"
512 "fmla v25.4s, v11.4s, v1.s[3]\n"
513 "ldr d1, [x16, #0x0]\n"
514 "fmla v26.4s, v11.4s, v2.s[3]\n"
515 "ldr d2, [x15, #0x0]\n"
517 "ldr d11, [x5, #0x30]\n"
519 "ldr x9, [x17, #0x8]\n"
521 "ldr x28, [x16, #0x8]\n"
523 "ldr x27, [x15, #0x8]\n"
524 "mov v10.d[1], x26\n"
525 "ldr x26, [x5, #0x38]\n"
528 "prfm pldl1keep, [x17, #0x80]\n"
530 "prfm pldl1keep, [x16, #0x80]\n"
531 "mov v11.d[1], x26\n"
532 "prfm pldl1keep, [x15, #0x80]\n"
535 "fmla v24.4s, v8.4s, v0.s[0]\n"
536 "add x17, x17, #0x10\n"
537 "fmla v25.4s, v8.4s, v1.s[0]\n"
538 "add x16, x16, #0x10\n"
539 "fmla v26.4s, v8.4s, v2.s[0]\n"
540 "add x15, x15, #0x10\n"
541 "fmla v24.4s, v9.4s, v0.s[1]\n"
543 "fmla v25.4s, v9.4s, v1.s[1]\n"
544 "prfm pldl1keep, [x17, #0x80]\n"
545 "fmla v26.4s, v9.4s, v2.s[1]\n"
546 "prfm pldl1keep, [x16, #0x80]\n"
547 "fmla v24.4s, v10.4s, v0.s[2]\n"
548 "prfm pldl1keep, [x15, #0x80]\n"
549 "fmla v25.4s, v10.4s, v1.s[2]\n"
550 "add x5, x5, #0x40\n"
551 "fmla v26.4s, v10.4s, v2.s[2]\n"
552 "fmla v24.4s, v11.4s, v0.s[3]\n"
553 "fmla v25.4s, v11.4s, v1.s[3]\n"
554 "fmla v26.4s, v11.4s, v2.s[3]\n"
558 "ldr s19, [x17], #0x4\n"
560 "ldr s18, [x16], #0x4\n"
561 "ldr s17, [x15], #0x4\n"
562 "ldr q16, [x5, #0x0]\n"
563 "fmla v24.4s, v16.4s, v19.s[0]\n"
564 "fmla v25.4s, v16.4s, v18.s[0]\n"
565 "add x5, x5, #0x10\n"
566 "fmla v26.4s, v16.4s, v17.s[0]\n"
569 "ldr w26, [%x[args_ptr], %[offsetof_num_strings]]\n"
573 "ldr x26, [%x[args_ptr], %[offsetof_output_offset]]\n"
574 "add x13, x6, x26, LSL #2\n"
575 "add x12, x13, x26, LSL #2\n"
576 "prfm pstl1keep, [x6, #0x0]\n"
577 "prfm pstl1keep, [x13, #0x0]\n"
578 "prfm pstl1keep, [x12, #0x0]\n"
579 "tbz %x[flags], #1, 59f\n"
580 "add x26, %x[args_ptr], %[offset_max]\n"
581 "ld1r { v16.4s }, [x26]\n"
582 "fmin v24.4s, v24.4s, v16.4s\n"
583 "fmin v25.4s, v25.4s, v16.4s\n"
584 "fmin v26.4s, v26.4s, v16.4s\n"
585 "add x26, %x[args_ptr], %[offset_min]\n"
586 "ld1r { v16.4s }, [x26]\n"
587 "fmax v24.4s, v24.4s, v16.4s\n"
588 "fmax v25.4s, v25.4s, v16.4s\n"
589 "fmax v26.4s, v26.4s, v16.4s\n"
594 "str d24, [x6], #0x8\n"
595 "str d25, [x13], #0x8\n"
596 "str d26, [x12], #0x8\n"
598 "st1 { v24.s }[2], [x6]\n"
599 "st1 { v25.s }[2], [x13]\n"
600 "st1 { v26.s }[2], [x12]\n"
603 "str s24, [x6, #0x0]\n"
604 "str s25, [x13, #0x0]\n"
605 "str s26, [x12, #0x0]\n"
609 "str q24, [x6, #0x0]\n"
610 "add x6, x6, #0x10\n"
611 "str q25, [x13, #0x0]\n"
612 "str q26, [x12, #0x0]\n"
614 "subs x4, x4, #0x4\n"
619 "ldr x4, [%x[args_ptr], %[offsetof_N]]\n"
620 "ldr x5, [%x[args_ptr], %[offsetof_B_ptr]]\n"
621 "mov x6, %x[output_ptr]\n"
624 "ldr q24, [x3, #0x0]\n"
625 "mov v25.16b, v24.16b\n"
626 "mov v26.16b, v24.16b\n"
627 "add x3, x3, #0x10\n"
628 "mov v27.16b, v24.16b\n"
631 "tbz %x[flags], #0, 70f\n"
632 "ldr x26, [%x[args_ptr], %[offsetof_output_offset]]\n"
633 "add x13, x6, x26, LSL #2\n"
634 "add x12, x13, x26, LSL #2\n"
636 "add x11, x12, x26, LSL #2\n"
639 "ldr d24, [x6], #0x8\n"
641 "ldr d25, [x13], #0x8\n"
642 "ldr d26, [x12], #0x8\n"
643 "ldr d27, [x11], #0x8\n"
645 "ld1 { v24.s }[2], [x6]\n"
646 "ld1 { v25.s }[2], [x13]\n"
647 "ld1 { v26.s }[2], [x12]\n"
648 "ld1 { v27.s }[2], [x11]\n"
651 "ldr s24, [x6, #0x0]\n"
653 "ldr s25, [x13, #0x0]\n"
654 "ldr s26, [x12, #0x0]\n"
655 "ldr s27, [x11, #0x0]\n"
660 "ldr q24, [x6, #0x0]\n"
661 "ldr q25, [x13, #0x0]\n"
662 "ldr q26, [x12, #0x0]\n"
663 "ldr q27, [x11, #0x0]\n"
666 "movi v24.16b, #0x0\n"
667 "movi v25.16b, #0x0\n"
668 "movi v26.16b, #0x0\n"
669 "movi v27.16b, #0x0\n"
673 "ldr x26, [%x[args_ptr], %[offsetof_string_lengths]]\n"
674 "ldr w8, [x26, x7, LSL #0x2]\n"
675 "ldr x27, [%x[args_ptr], %[offsetof_input_offset]]\n"
676 "tbz %x[flags], #3, 73f\n"
677 "ldr x26, [%x[input_ptr], x7, LSL #0x3]\n"
678 "add x26, x26, x27, LSL #3\n"
679 "ldr x17, [x26, #0x0]\n"
680 "ldr x16, [x26, #0x8]\n"
681 "ldr x15, [x26, #0x10]\n"
682 "ldr x14, [x26, #0x18]\n"
684 "ldr x26, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
685 "add x17, x17, x26, LSL #2\n"
686 "add x16, x16, x26, LSL #2\n"
687 "add x15, x15, x26, LSL #2\n"
688 "add x14, x14, x26, LSL #2\n"
691 "mov x17, %x[input_ptr]\n"
692 "add x16, x17, x27, LSL #2\n"
693 "add x15, x16, x27, LSL #2\n"
694 "add x14, x15, x27, LSL #2\n"
698 "ldr q0, [x17, #0x0]\n"
700 "ldr q1, [x16, #0x0]\n"
701 "ldr q2, [x15, #0x0]\n"
702 "ldr q3, [x14, #0x0]\n"
703 "ldr q8, [x5, #0x0]\n"
704 "ldr q9, [x5, #0x10]\n"
705 "ldr q10, [x5, #0x20]\n"
706 "ldr q11, [x5, #0x30]\n"
709 "fmla v24.4s, v8.4s, v0.s[0]\n"
710 "add x17, x17, #0x10\n"
711 "fmla v25.4s, v8.4s, v1.s[0]\n"
712 "add x16, x16, #0x10\n"
713 "fmla v26.4s, v8.4s, v2.s[0]\n"
714 "add x15, x15, #0x10\n"
715 "fmla v27.4s, v8.4s, v3.s[0]\n"
716 "add x14, x14, #0x10\n"
717 "fmla v24.4s, v9.4s, v0.s[1]\n"
718 "add x5, x5, #0x40\n"
719 "ldr d8, [x5, #0x0]\n"
720 "fmla v25.4s, v9.4s, v1.s[1]\n"
721 "fmla v26.4s, v9.4s, v2.s[1]\n"
722 "ldr x27, [x5, #0x8]\n"
723 "fmla v27.4s, v9.4s, v3.s[1]\n"
724 "ldr d9, [x5, #0x10]\n"
725 "fmla v24.4s, v10.4s, v0.s[2]\n"
726 "ldr x26, [x5, #0x18]\n"
727 "fmla v25.4s, v10.4s, v1.s[2]\n"
728 "ldr x11, [x5, #0x28]\n"
729 "fmla v26.4s, v10.4s, v2.s[2]\n"
730 "ldr x10, [x17, #0x8]\n"
731 "fmla v27.4s, v10.4s, v3.s[2]\n"
732 "ldr d10, [x5, #0x20]\n"
733 "fmla v24.4s, v11.4s, v0.s[3]\n"
734 "ldr d0, [x17, #0x0]\n"
735 "fmla v25.4s, v11.4s, v1.s[3]\n"
736 "ldr d1, [x16, #0x0]\n"
737 "ldr x9, [x16, #0x8]\n"
738 "fmla v26.4s, v11.4s, v2.s[3]\n"
739 "ldr d2, [x15, #0x0]\n"
740 "fmla v27.4s, v11.4s, v3.s[3]\n"
741 "ldr d3, [x14, #0x0]\n"
743 "ldr d11, [x5, #0x30]\n"
745 "ldr x28, [x15, #0x8]\n"
747 "ldr x27, [x14, #0x8]\n"
749 "ldr x26, [x5, #0x38]\n"
750 "mov v10.d[1], x11\n"
751 "prfm pldl1keep, [x17, #0x80]\n"
753 "prfm pldl1keep, [x16, #0x80]\n"
756 "prfm pldl1keep, [x15, #0x80]\n"
758 "prfm pldl1keep, [x14, #0x80]\n"
759 "mov v11.d[1], x26\n"
762 "fmla v24.4s, v8.4s, v0.s[0]\n"
763 "add x17, x17, #0x10\n"
764 "fmla v25.4s, v8.4s, v1.s[0]\n"
765 "add x16, x16, #0x10\n"
766 "fmla v26.4s, v8.4s, v2.s[0]\n"
767 "add x15, x15, #0x10\n"
768 "fmla v27.4s, v8.4s, v3.s[0]\n"
769 "add x14, x14, #0x10\n"
770 "fmla v24.4s, v9.4s, v0.s[1]\n"
772 "fmla v25.4s, v9.4s, v1.s[1]\n"
773 "prfm pldl1keep, [x17, #0x80]\n"
774 "fmla v26.4s, v9.4s, v2.s[1]\n"
775 "prfm pldl1keep, [x16, #0x80]\n"
776 "fmla v27.4s, v9.4s, v3.s[1]\n"
777 "prfm pldl1keep, [x15, #0x80]\n"
778 "fmla v24.4s, v10.4s, v0.s[2]\n"
779 "prfm pldl1keep, [x14, #0x80]\n"
780 "fmla v25.4s, v10.4s, v1.s[2]\n"
781 "add x5, x5, #0x40\n"
782 "fmla v26.4s, v10.4s, v2.s[2]\n"
783 "fmla v27.4s, v10.4s, v3.s[2]\n"
784 "fmla v24.4s, v11.4s, v0.s[3]\n"
785 "fmla v25.4s, v11.4s, v1.s[3]\n"
786 "fmla v26.4s, v11.4s, v2.s[3]\n"
787 "fmla v27.4s, v11.4s, v3.s[3]\n"
791 "ldr s20, [x17], #0x4\n"
793 "ldr s19, [x16], #0x4\n"
794 "ldr s18, [x15], #0x4\n"
795 "ldr s17, [x14], #0x4\n"
796 "ldr q16, [x5, #0x0]\n"
797 "fmla v24.4s, v16.4s, v20.s[0]\n"
798 "fmla v25.4s, v16.4s, v19.s[0]\n"
799 "add x5, x5, #0x10\n"
800 "fmla v26.4s, v16.4s, v18.s[0]\n"
801 "fmla v27.4s, v16.4s, v17.s[0]\n"
804 "ldr w26, [%x[args_ptr], %[offsetof_num_strings]]\n"
808 "ldr x26, [%x[args_ptr], %[offsetof_output_offset]]\n"
809 "add x13, x6, x26, LSL #2\n"
810 "add x12, x13, x26, LSL #2\n"
811 "add x11, x12, x26, LSL #2\n"
812 "prfm pstl1keep, [x6, #0x0]\n"
813 "prfm pstl1keep, [x13, #0x0]\n"
814 "prfm pstl1keep, [x12, #0x0]\n"
815 "prfm pstl1keep, [x11, #0x0]\n"
816 "tbz %x[flags], #1, 80f\n"
817 "add x26, %x[args_ptr], %[offset_max]\n"
818 "ld1r { v16.4s }, [x26]\n"
819 "fmin v24.4s, v24.4s, v16.4s\n"
820 "fmin v25.4s, v25.4s, v16.4s\n"
821 "fmin v26.4s, v26.4s, v16.4s\n"
822 "fmin v27.4s, v27.4s, v16.4s\n"
823 "add x26, %x[args_ptr], %[offset_min]\n"
824 "ld1r { v16.4s }, [x26]\n"
825 "fmax v24.4s, v24.4s, v16.4s\n"
826 "fmax v25.4s, v25.4s, v16.4s\n"
827 "fmax v26.4s, v26.4s, v16.4s\n"
828 "fmax v27.4s, v27.4s, v16.4s\n"
833 "str d24, [x6], #0x8\n"
834 "str d25, [x13], #0x8\n"
835 "str d26, [x12], #0x8\n"
836 "str d27, [x11], #0x8\n"
838 "st1 { v24.s }[2], [x6]\n"
839 "st1 { v25.s }[2], [x13]\n"
840 "st1 { v26.s }[2], [x12]\n"
841 "st1 { v27.s }[2], [x11]\n"
844 "str s24, [x6, #0x0]\n"
845 "str s25, [x13, #0x0]\n"
846 "str s26, [x12, #0x0]\n"
847 "str s27, [x11, #0x0]\n"
851 "str q24, [x6, #0x0]\n"
852 "add x6, x6, #0x10\n"
853 "str q25, [x13, #0x0]\n"
854 "str q26, [x12, #0x0]\n"
855 "str q27, [x11, #0x0]\n"
857 "subs x4, x4, #0x4\n"
862 "ldr x4, [%x[args_ptr], %[offsetof_N]]\n"
863 "ldr x5, [%x[args_ptr], %[offsetof_B_ptr]]\n"
864 "mov x6, %x[output_ptr]\n"
867 "ldr q24, [x3, #0x0]\n"
868 "mov v25.16b, v24.16b\n"
869 "mov v26.16b, v24.16b\n"
870 "add x3, x3, #0x10\n"
871 "mov v27.16b, v24.16b\n"
872 "mov v28.16b, v24.16b\n"
875 "tbz %x[flags], #0, 91f\n"
876 "ldr x26, [%x[args_ptr], %[offsetof_output_offset]]\n"
877 "add x13, x6, x26, LSL #2\n"
878 "add x12, x13, x26, LSL #2\n"
879 "add x11, x12, x26, LSL #2\n"
881 "add x10, x11, x26, LSL #2\n"
884 "ldr d24, [x6], #0x8\n"
886 "ldr d25, [x13], #0x8\n"
887 "ldr d26, [x12], #0x8\n"
888 "ldr d27, [x11], #0x8\n"
889 "ldr d28, [x10], #0x8\n"
891 "ld1 { v24.s }[2], [x6]\n"
892 "ld1 { v25.s }[2], [x13]\n"
893 "ld1 { v26.s }[2], [x12]\n"
894 "ld1 { v27.s }[2], [x11]\n"
895 "ld1 { v28.s }[2], [x10]\n"
898 "ldr s24, [x6, #0x0]\n"
900 "ldr s25, [x13, #0x0]\n"
901 "ldr s26, [x12, #0x0]\n"
902 "ldr s27, [x11, #0x0]\n"
903 "ldr s28, [x10, #0x0]\n"
908 "ldr q24, [x6, #0x0]\n"
909 "ldr q25, [x13, #0x0]\n"
910 "ldr q26, [x12, #0x0]\n"
911 "ldr q27, [x11, #0x0]\n"
912 "ldr q28, [x10, #0x0]\n"
915 "movi v24.16b, #0x0\n"
916 "movi v25.16b, #0x0\n"
917 "movi v26.16b, #0x0\n"
918 "movi v27.16b, #0x0\n"
919 "movi v28.16b, #0x0\n"
923 "ldr x26, [%x[args_ptr], %[offsetof_string_lengths]]\n"
924 "ldr w8, [x26, x7, LSL #0x2]\n"
925 "ldr x27, [%x[args_ptr], %[offsetof_input_offset]]\n"
926 "tbz %x[flags], #3, 94f\n"
927 "ldr x26, [%x[input_ptr], x7, LSL #0x3]\n"
928 "add x26, x26, x27, LSL #3\n"
929 "ldr x17, [x26, #0x0]\n"
930 "ldr x16, [x26, #0x8]\n"
931 "ldr x15, [x26, #0x10]\n"
932 "ldr x14, [x26, #0x18]\n"
933 "ldr x13, [x26, #0x20]\n"
935 "ldr x26, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
936 "add x17, x17, x26, LSL #2\n"
937 "add x16, x16, x26, LSL #2\n"
938 "add x15, x15, x26, LSL #2\n"
939 "add x14, x14, x26, LSL #2\n"
940 "add x13, x13, x26, LSL #2\n"
943 "mov x17, %x[input_ptr]\n"
944 "add x16, x17, x27, LSL #2\n"
945 "add x15, x16, x27, LSL #2\n"
946 "add x14, x15, x27, LSL #2\n"
947 "add x13, x14, x27, LSL #2\n"
951 "ldr q0, [x17, #0x0]\n"
953 "ldr q1, [x16, #0x0]\n"
954 "ldr q2, [x15, #0x0]\n"
955 "ldr q3, [x14, #0x0]\n"
956 "ldr q4, [x13, #0x0]\n"
957 "ldr q8, [x5, #0x0]\n"
958 "ldr q9, [x5, #0x10]\n"
959 "ldr q10, [x5, #0x20]\n"
960 "ldr q11, [x5, #0x30]\n"
963 "fmla v24.4s, v8.4s, v0.s[0]\n"
964 "add x17, x17, #0x10\n"
965 "fmla v25.4s, v8.4s, v1.s[0]\n"
966 "add x16, x16, #0x10\n"
967 "fmla v26.4s, v8.4s, v2.s[0]\n"
968 "add x15, x15, #0x10\n"
969 "fmla v27.4s, v8.4s, v3.s[0]\n"
970 "add x14, x14, #0x10\n"
971 "fmla v28.4s, v8.4s, v4.s[0]\n"
972 "add x13, x13, #0x10\n"
973 "fmla v24.4s, v9.4s, v0.s[1]\n"
974 "add x5, x5, #0x40\n"
975 "ldr d8, [x5, #0x0]\n"
976 "fmla v25.4s, v9.4s, v1.s[1]\n"
977 "fmla v26.4s, v9.4s, v2.s[1]\n"
978 "ldr x27, [x5, #0x8]\n"
979 "fmla v27.4s, v9.4s, v3.s[1]\n"
980 "ldr x26, [x5, #0x18]\n"
981 "fmla v28.4s, v9.4s, v4.s[1]\n"
982 "ldr d9, [x5, #0x10]\n"
983 "fmla v24.4s, v10.4s, v0.s[2]\n"
984 "ldr x12, [x5, #0x28]\n"
985 "fmla v25.4s, v10.4s, v1.s[2]\n"
986 "ldr x11, [x17, #0x8]\n"
987 "fmla v26.4s, v10.4s, v2.s[2]\n"
988 "ldr x10, [x16, #0x8]\n"
989 "fmla v27.4s, v10.4s, v3.s[2]\n"
990 "ldr x9, [x15, #0x8]\n"
991 "fmla v28.4s, v10.4s, v4.s[2]\n"
992 "ldr d10, [x5, #0x20]\n"
993 "fmla v24.4s, v11.4s, v0.s[3]\n"
994 "ldr d0, [x17, #0x0]\n"
995 "fmla v25.4s, v11.4s, v1.s[3]\n"
996 "ldr d1, [x16, #0x0]\n"
997 "fmla v26.4s, v11.4s, v2.s[3]\n"
998 "ldr d2, [x15, #0x0]\n"
999 "fmla v27.4s, v11.4s, v3.s[3]\n"
1000 "ldr d3, [x14, #0x0]\n"
1001 "ldr x28, [x14, #0x8]\n"
1002 "fmla v28.4s, v11.4s, v4.s[3]\n"
1003 "ldr d4, [x13, #0x0]\n"
1004 "sub x8, x8, #0x4\n"
1005 "ldr d11, [x5, #0x30]\n"
1007 "mov v8.d[1], x27\n"
1008 "ldr x27, [x13, #0x8]\n"
1009 "mov v9.d[1], x26\n"
1010 "ldr x26, [x5, #0x38]\n"
1011 "prfm pldl1keep, [x17, #0x80]\n"
1012 "mov v10.d[1], x12\n"
1013 "prfm pldl1keep, [x16, #0x80]\n"
1014 "mov v0.d[1], x11\n"
1015 "prfm pldl1keep, [x15, #0x80]\n"
1016 "mov v1.d[1], x10\n"
1017 "prfm pldl1keep, [x14, #0x80]\n"
1019 "mov v3.d[1], x28\n"
1020 "prfm pldl1keep, [x13, #0x80]\n"
1021 "mov v4.d[1], x27\n"
1022 "mov v11.d[1], x26\n"
1025 "fmla v24.4s, v8.4s, v0.s[0]\n"
1026 "add x17, x17, #0x10\n"
1027 "fmla v25.4s, v8.4s, v1.s[0]\n"
1028 "add x16, x16, #0x10\n"
1029 "fmla v26.4s, v8.4s, v2.s[0]\n"
1030 "add x15, x15, #0x10\n"
1031 "fmla v27.4s, v8.4s, v3.s[0]\n"
1032 "add x14, x14, #0x10\n"
1033 "fmla v28.4s, v8.4s, v4.s[0]\n"
1034 "add x13, x13, #0x10\n"
1035 "fmla v24.4s, v9.4s, v0.s[1]\n"
1036 "sub x8, x8, #0x4\n"
1037 "fmla v25.4s, v9.4s, v1.s[1]\n"
1038 "prfm pldl1keep, [x17, #0x80]\n"
1039 "fmla v26.4s, v9.4s, v2.s[1]\n"
1040 "prfm pldl1keep, [x16, #0x80]\n"
1041 "fmla v27.4s, v9.4s, v3.s[1]\n"
1042 "prfm pldl1keep, [x15, #0x80]\n"
1043 "fmla v28.4s, v9.4s, v4.s[1]\n"
1044 "prfm pldl1keep, [x14, #0x80]\n"
1045 "fmla v24.4s, v10.4s, v0.s[2]\n"
1046 "prfm pldl1keep, [x13, #0x80]\n"
1047 "fmla v25.4s, v10.4s, v1.s[2]\n"
1048 "add x5, x5, #0x40\n"
1049 "fmla v26.4s, v10.4s, v2.s[2]\n"
1050 "fmla v27.4s, v10.4s, v3.s[2]\n"
1051 "fmla v28.4s, v10.4s, v4.s[2]\n"
1052 "fmla v24.4s, v11.4s, v0.s[3]\n"
1053 "fmla v25.4s, v11.4s, v1.s[3]\n"
1054 "fmla v26.4s, v11.4s, v2.s[3]\n"
1055 "fmla v27.4s, v11.4s, v3.s[3]\n"
1056 "fmla v28.4s, v11.4s, v4.s[3]\n"
1060 "ldr s21, [x17], #0x4\n"
1061 "sub x8, x8, #0x1\n"
1062 "ldr s20, [x16], #0x4\n"
1063 "ldr s19, [x15], #0x4\n"
1064 "ldr s18, [x14], #0x4\n"
1065 "ldr s17, [x13], #0x4\n"
1066 "ldr q16, [x5, #0x0]\n"
1067 "fmla v24.4s, v16.4s, v21.s[0]\n"
1068 "fmla v25.4s, v16.4s, v20.s[0]\n"
1069 "add x5, x5, #0x10\n"
1070 "fmla v26.4s, v16.4s, v19.s[0]\n"
1071 "fmla v27.4s, v16.4s, v18.s[0]\n"
1072 "fmla v28.4s, v16.4s, v17.s[0]\n"
1075 "ldr w26, [%x[args_ptr], %[offsetof_num_strings]]\n"
1076 "add x7, x7, #0x1\n"
1079 "ldr x26, [%x[args_ptr], %[offsetof_output_offset]]\n"
1080 "add x13, x6, x26, LSL #2\n"
1081 "add x12, x13, x26, LSL #2\n"
1082 "add x11, x12, x26, LSL #2\n"
1083 "add x10, x11, x26, LSL #2\n"
1084 "prfm pstl1keep, [x6, #0x0]\n"
1085 "prfm pstl1keep, [x13, #0x0]\n"
1086 "prfm pstl1keep, [x12, #0x0]\n"
1087 "prfm pstl1keep, [x11, #0x0]\n"
1088 "prfm pstl1keep, [x10, #0x0]\n"
1089 "tbz %x[flags], #1, 101f\n"
1090 "add x26, %x[args_ptr], %[offset_max]\n"
1091 "ld1r { v16.4s }, [x26]\n"
1092 "fmin v24.4s, v24.4s, v16.4s\n"
1093 "fmin v25.4s, v25.4s, v16.4s\n"
1094 "fmin v26.4s, v26.4s, v16.4s\n"
1095 "fmin v27.4s, v27.4s, v16.4s\n"
1096 "fmin v28.4s, v28.4s, v16.4s\n"
1097 "add x26, %x[args_ptr], %[offset_min]\n"
1098 "ld1r { v16.4s }, [x26]\n"
1099 "fmax v24.4s, v24.4s, v16.4s\n"
1100 "fmax v25.4s, v25.4s, v16.4s\n"
1101 "fmax v26.4s, v26.4s, v16.4s\n"
1102 "fmax v27.4s, v27.4s, v16.4s\n"
1103 "fmax v28.4s, v28.4s, v16.4s\n"
1107 "tbz x4, #1, 102f\n"
1108 "str d24, [x6], #0x8\n"
1109 "str d25, [x13], #0x8\n"
1110 "str d26, [x12], #0x8\n"
1111 "str d27, [x11], #0x8\n"
1112 "str d28, [x10], #0x8\n"
1113 "tbz x4, #0, 103f\n"
1114 "st1 { v24.s }[2], [x6]\n"
1115 "st1 { v25.s }[2], [x13]\n"
1116 "st1 { v26.s }[2], [x12]\n"
1117 "st1 { v27.s }[2], [x11]\n"
1118 "st1 { v28.s }[2], [x10]\n"
1121 "str s24, [x6, #0x0]\n"
1122 "str s25, [x13, #0x0]\n"
1123 "str s26, [x12, #0x0]\n"
1124 "str s27, [x11, #0x0]\n"
1125 "str s28, [x10, #0x0]\n"
1129 "str q24, [x6, #0x0]\n"
1130 "add x6, x6, #0x10\n"
1131 "str q25, [x13, #0x0]\n"
1132 "str q26, [x12, #0x0]\n"
1133 "str q27, [x11, #0x0]\n"
1134 "str q28, [x10, #0x0]\n"
1136 "subs x4, x4, #0x4\n"
1140 "mov x3, %x[bias]\n"
1141 "ldr x4, [%x[args_ptr], %[offsetof_N]]\n"
1142 "ldr x5, [%x[args_ptr], %[offsetof_B_ptr]]\n"
1143 "mov x6, %x[output_ptr]\n"
1146 "ldr q24, [x3, #0x0]\n"
1147 "mov v25.16b, v24.16b\n"
1148 "mov v26.16b, v24.16b\n"
1149 "add x3, x3, #0x10\n"
1150 "mov v27.16b, v24.16b\n"
1151 "mov v28.16b, v24.16b\n"
1152 "mov v29.16b, v24.16b\n"
1155 "tbz %x[flags], #0, 112f\n"
1156 "ldr x26, [%x[args_ptr], %[offsetof_output_offset]]\n"
1157 "add x13, x6, x26, LSL #2\n"
1158 "add x12, x13, x26, LSL #2\n"
1159 "add x11, x12, x26, LSL #2\n"
1160 "add x10, x11, x26, LSL #2\n"
1162 "add x9, x10, x26, LSL #2\n"
1164 "tbz x4, #1, 109f\n"
1165 "ldr d24, [x6], #0x8\n"
1167 "ldr d25, [x13], #0x8\n"
1168 "ldr d26, [x12], #0x8\n"
1169 "ldr d27, [x11], #0x8\n"
1170 "ldr d28, [x10], #0x8\n"
1171 "ldr d29, [x9], #0x8\n"
1172 "tbz x4, #0, 110f\n"
1173 "ld1 { v24.s }[2], [x6]\n"
1174 "ld1 { v25.s }[2], [x13]\n"
1175 "ld1 { v26.s }[2], [x12]\n"
1176 "ld1 { v27.s }[2], [x11]\n"
1177 "ld1 { v28.s }[2], [x10]\n"
1178 "ld1 { v29.s }[2], [x9]\n"
1181 "ldr s24, [x6, #0x0]\n"
1183 "ldr s25, [x13, #0x0]\n"
1184 "ldr s26, [x12, #0x0]\n"
1185 "ldr s27, [x11, #0x0]\n"
1186 "ldr s28, [x10, #0x0]\n"
1187 "ldr s29, [x9, #0x0]\n"
1192 "ldr q24, [x6, #0x0]\n"
1193 "ldr q25, [x13, #0x0]\n"
1194 "ldr q26, [x12, #0x0]\n"
1195 "ldr q27, [x11, #0x0]\n"
1196 "ldr q28, [x10, #0x0]\n"
1197 "ldr q29, [x9, #0x0]\n"
1200 "movi v24.16b, #0x0\n"
1201 "movi v25.16b, #0x0\n"
1202 "movi v26.16b, #0x0\n"
1203 "movi v27.16b, #0x0\n"
1204 "movi v28.16b, #0x0\n"
1205 "movi v29.16b, #0x0\n"
1209 "ldr x26, [%x[args_ptr], %[offsetof_string_lengths]]\n"
1210 "ldr w8, [x26, x7, LSL #0x2]\n"
1211 "ldr x27, [%x[args_ptr], %[offsetof_input_offset]]\n"
1212 "tbz %x[flags], #3, 115f\n"
1213 "ldr x26, [%x[input_ptr], x7, LSL #0x3]\n"
1214 "add x26, x26, x27, LSL #3\n"
1215 "ldr x17, [x26, #0x0]\n"
1216 "ldr x16, [x26, #0x8]\n"
1217 "ldr x15, [x26, #0x10]\n"
1218 "ldr x14, [x26, #0x18]\n"
1219 "ldr x13, [x26, #0x20]\n"
1220 "ldr x12, [x26, #0x28]\n"
1222 "ldr x26, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
1223 "add x17, x17, x26, LSL #2\n"
1224 "add x16, x16, x26, LSL #2\n"
1225 "add x15, x15, x26, LSL #2\n"
1226 "add x14, x14, x26, LSL #2\n"
1227 "add x13, x13, x26, LSL #2\n"
1228 "add x12, x12, x26, LSL #2\n"
1231 "mov x17, %x[input_ptr]\n"
1232 "add x16, x17, x27, LSL #2\n"
1233 "add x15, x16, x27, LSL #2\n"
1234 "add x14, x15, x27, LSL #2\n"
1235 "add x13, x14, x27, LSL #2\n"
1236 "add x12, x13, x27, LSL #2\n"
1240 "ldr q0, [x17, #0x0]\n"
1242 "ldr q1, [x16, #0x0]\n"
1243 "ldr q2, [x15, #0x0]\n"
1244 "ldr q3, [x14, #0x0]\n"
1245 "ldr q4, [x13, #0x0]\n"
1246 "ldr q5, [x12, #0x0]\n"
1247 "ldr q8, [x5, #0x0]\n"
1248 "ldr q9, [x5, #0x10]\n"
1249 "ldr q10, [x5, #0x20]\n"
1250 "ldr q11, [x5, #0x30]\n"
1253 "fmla v24.4s, v8.4s, v0.s[0]\n"
1254 "add x17, x17, #0x10\n"
1255 "fmla v25.4s, v8.4s, v1.s[0]\n"
1256 "add x16, x16, #0x10\n"
1257 "fmla v26.4s, v8.4s, v2.s[0]\n"
1258 "add x15, x15, #0x10\n"
1259 "fmla v27.4s, v8.4s, v3.s[0]\n"
1260 "add x14, x14, #0x10\n"
1261 "fmla v28.4s, v8.4s, v4.s[0]\n"
1262 "add x13, x13, #0x10\n"
1263 "fmla v29.4s, v8.4s, v5.s[0]\n"
1264 "add x12, x12, #0x10\n"
1265 "fmla v24.4s, v9.4s, v0.s[1]\n"
1266 "add x5, x5, #0x40\n"
1267 "ldr d8, [x5, #0x0]\n"
1268 "fmla v25.4s, v9.4s, v1.s[1]\n"
1269 "fmla v26.4s, v9.4s, v2.s[1]\n"
1270 "ldr x9, [x5, #0x8]\n"
1271 "fmla v27.4s, v9.4s, v3.s[1]\n"
1272 "ldr x28, [x5, #0x18]\n"
1273 "fmla v28.4s, v9.4s, v4.s[1]\n"
1274 "ldr x27, [x5, #0x28]\n"
1275 "fmla v29.4s, v9.4s, v5.s[1]\n"
1276 "ldr d9, [x5, #0x10]\n"
1277 "fmla v24.4s, v10.4s, v0.s[2]\n"
1278 "ldr x26, [x17, #0x8]\n"
1279 "fmla v25.4s, v10.4s, v1.s[2]\n"
1280 "ldr x11, [x16, #0x8]\n"
1281 "fmla v26.4s, v10.4s, v2.s[2]\n"
1282 "ldr x10, [x15, #0x8]\n"
1283 "fmla v27.4s, v10.4s, v3.s[2]\n"
1284 "sub x8, x8, #0x4\n"
1285 "fmla v28.4s, v10.4s, v4.s[2]\n"
1287 "fmla v29.4s, v10.4s, v5.s[2]\n"
1288 "ldr d10, [x5, #0x20]\n"
1289 "fmla v24.4s, v11.4s, v0.s[3]\n"
1290 "ldr d0, [x17, #0x0]\n"
1291 "fmla v25.4s, v11.4s, v1.s[3]\n"
1292 "ldr d1, [x16, #0x0]\n"
1293 "fmla v26.4s, v11.4s, v2.s[3]\n"
1294 "ldr d2, [x15, #0x0]\n"
1295 "fmla v27.4s, v11.4s, v3.s[3]\n"
1296 "ldr d3, [x14, #0x0]\n"
1297 "fmla v28.4s, v11.4s, v4.s[3]\n"
1298 "ldr d4, [x13, #0x0]\n"
1299 "fmla v29.4s, v11.4s, v5.s[3]\n"
1300 "ldr d5, [x12, #0x0]\n"
1301 "ldr d11, [x5, #0x30]\n"
1303 "ldr x9, [x14, #0x8]\n"
1304 "mov v9.d[1], x28\n"
1305 "ldr x28, [x13, #0x8]\n"
1306 "mov v10.d[1], x27\n"
1307 "ldr x27, [x12, #0x8]\n"
1308 "mov v0.d[1], x26\n"
1309 "ldr x26, [x5, #0x38]\n"
1310 "mov v1.d[1], x11\n"
1311 "prfm pldl1keep, [x17, #0x80]\n"
1312 "mov v2.d[1], x10\n"
1313 "prfm pldl1keep, [x16, #0x80]\n"
1315 "prfm pldl1keep, [x15, #0x80]\n"
1316 "mov v4.d[1], x28\n"
1317 "prfm pldl1keep, [x14, #0x80]\n"
1318 "mov v5.d[1], x27\n"
1319 "prfm pldl1keep, [x13, #0x80]\n"
1320 "mov v11.d[1], x26\n"
1321 "prfm pldl1keep, [x12, #0x80]\n"
1324 "fmla v24.4s, v8.4s, v0.s[0]\n"
1325 "add x17, x17, #0x10\n"
1326 "fmla v25.4s, v8.4s, v1.s[0]\n"
1327 "add x16, x16, #0x10\n"
1328 "fmla v26.4s, v8.4s, v2.s[0]\n"
1329 "add x15, x15, #0x10\n"
1330 "fmla v27.4s, v8.4s, v3.s[0]\n"
1331 "add x14, x14, #0x10\n"
1332 "fmla v28.4s, v8.4s, v4.s[0]\n"
1333 "add x13, x13, #0x10\n"
1334 "fmla v29.4s, v8.4s, v5.s[0]\n"
1335 "add x12, x12, #0x10\n"
1336 "fmla v24.4s, v9.4s, v0.s[1]\n"
1337 "sub x8, x8, #0x4\n"
1338 "fmla v25.4s, v9.4s, v1.s[1]\n"
1339 "prfm pldl1keep, [x17, #0x80]\n"
1340 "fmla v26.4s, v9.4s, v2.s[1]\n"
1341 "prfm pldl1keep, [x16, #0x80]\n"
1342 "fmla v27.4s, v9.4s, v3.s[1]\n"
1343 "prfm pldl1keep, [x15, #0x80]\n"
1344 "fmla v28.4s, v9.4s, v4.s[1]\n"
1345 "prfm pldl1keep, [x14, #0x80]\n"
1346 "fmla v29.4s, v9.4s, v5.s[1]\n"
1347 "prfm pldl1keep, [x13, #0x80]\n"
1348 "fmla v24.4s, v10.4s, v0.s[2]\n"
1349 "prfm pldl1keep, [x12, #0x80]\n"
1350 "fmla v25.4s, v10.4s, v1.s[2]\n"
1351 "add x5, x5, #0x40\n"
1352 "fmla v26.4s, v10.4s, v2.s[2]\n"
1353 "fmla v27.4s, v10.4s, v3.s[2]\n"
1354 "fmla v28.4s, v10.4s, v4.s[2]\n"
1355 "fmla v29.4s, v10.4s, v5.s[2]\n"
1356 "fmla v24.4s, v11.4s, v0.s[3]\n"
1357 "fmla v25.4s, v11.4s, v1.s[3]\n"
1358 "fmla v26.4s, v11.4s, v2.s[3]\n"
1359 "fmla v27.4s, v11.4s, v3.s[3]\n"
1360 "fmla v28.4s, v11.4s, v4.s[3]\n"
1361 "fmla v29.4s, v11.4s, v5.s[3]\n"
1365 "ldr s22, [x17], #0x4\n"
1366 "sub x8, x8, #0x1\n"
1367 "ldr s21, [x16], #0x4\n"
1368 "ldr s20, [x15], #0x4\n"
1369 "ldr s19, [x14], #0x4\n"
1370 "ldr s18, [x13], #0x4\n"
1371 "ldr s17, [x12], #0x4\n"
1372 "ldr q16, [x5, #0x0]\n"
1373 "fmla v24.4s, v16.4s, v22.s[0]\n"
1374 "fmla v25.4s, v16.4s, v21.s[0]\n"
1375 "add x5, x5, #0x10\n"
1376 "fmla v26.4s, v16.4s, v20.s[0]\n"
1377 "fmla v27.4s, v16.4s, v19.s[0]\n"
1378 "fmla v28.4s, v16.4s, v18.s[0]\n"
1379 "fmla v29.4s, v16.4s, v17.s[0]\n"
1382 "ldr w26, [%x[args_ptr], %[offsetof_num_strings]]\n"
1383 "add x7, x7, #0x1\n"
1386 "ldr x26, [%x[args_ptr], %[offsetof_output_offset]]\n"
1387 "add x13, x6, x26, LSL #2\n"
1388 "add x12, x13, x26, LSL #2\n"
1389 "add x11, x12, x26, LSL #2\n"
1390 "add x10, x11, x26, LSL #2\n"
1391 "add x9, x10, x26, LSL #2\n"
1392 "prfm pstl1keep, [x6, #0x0]\n"
1393 "prfm pstl1keep, [x13, #0x0]\n"
1394 "prfm pstl1keep, [x12, #0x0]\n"
1395 "prfm pstl1keep, [x11, #0x0]\n"
1396 "prfm pstl1keep, [x10, #0x0]\n"
1397 "prfm pstl1keep, [x9, #0x0]\n"
1398 "tbz %x[flags], #1, 122f\n"
1399 "add x26, %x[args_ptr], %[offset_max]\n"
1400 "ld1r { v16.4s }, [x26]\n"
1401 "fmin v24.4s, v24.4s, v16.4s\n"
1402 "fmin v25.4s, v25.4s, v16.4s\n"
1403 "fmin v26.4s, v26.4s, v16.4s\n"
1404 "fmin v27.4s, v27.4s, v16.4s\n"
1405 "fmin v28.4s, v28.4s, v16.4s\n"
1406 "fmin v29.4s, v29.4s, v16.4s\n"
1407 "add x26, %x[args_ptr], %[offset_min]\n"
1408 "ld1r { v16.4s }, [x26]\n"
1409 "fmax v24.4s, v24.4s, v16.4s\n"
1410 "fmax v25.4s, v25.4s, v16.4s\n"
1411 "fmax v26.4s, v26.4s, v16.4s\n"
1412 "fmax v27.4s, v27.4s, v16.4s\n"
1413 "fmax v28.4s, v28.4s, v16.4s\n"
1414 "fmax v29.4s, v29.4s, v16.4s\n"
1418 "tbz x4, #1, 123f\n"
1419 "str d24, [x6], #0x8\n"
1420 "str d25, [x13], #0x8\n"
1421 "str d26, [x12], #0x8\n"
1422 "str d27, [x11], #0x8\n"
1423 "str d28, [x10], #0x8\n"
1424 "str d29, [x9], #0x8\n"
1425 "tbz x4, #0, 124f\n"
1426 "st1 { v24.s }[2], [x6]\n"
1427 "st1 { v25.s }[2], [x13]\n"
1428 "st1 { v26.s }[2], [x12]\n"
1429 "st1 { v27.s }[2], [x11]\n"
1430 "st1 { v28.s }[2], [x10]\n"
1431 "st1 { v29.s }[2], [x9]\n"
1434 "str s24, [x6, #0x0]\n"
1435 "str s25, [x13, #0x0]\n"
1436 "str s26, [x12, #0x0]\n"
1437 "str s27, [x11, #0x0]\n"
1438 "str s28, [x10, #0x0]\n"
1439 "str s29, [x9, #0x0]\n"
1443 "str q24, [x6, #0x0]\n"
1444 "add x6, x6, #0x10\n"
1445 "str q25, [x13, #0x0]\n"
1446 "str q26, [x12, #0x0]\n"
1447 "str q27, [x11, #0x0]\n"
1448 "str q28, [x10, #0x0]\n"
1449 "str q29, [x9, #0x0]\n"
1451 "subs x4, x4, #0x4\n"
1455 "mov x3, %x[bias]\n"
1456 "ldr x4, [%x[args_ptr], %[offsetof_N]]\n"
1457 "ldr x5, [%x[args_ptr], %[offsetof_B_ptr]]\n"
1458 "mov x6, %x[output_ptr]\n"
1461 "ldr q24, [x3, #0x0]\n"
1462 "mov v25.16b, v24.16b\n"
1463 "mov v26.16b, v24.16b\n"
1464 "add x3, x3, #0x10\n"
1465 "mov v27.16b, v24.16b\n"
1466 "mov v28.16b, v24.16b\n"
1467 "mov v29.16b, v24.16b\n"
1468 "mov v30.16b, v24.16b\n"
1471 "tbz %x[flags], #0, 133f\n"
1472 "ldr x26, [%x[args_ptr], %[offsetof_output_offset]]\n"
1473 "add x13, x6, x26, LSL #2\n"
1474 "add x12, x13, x26, LSL #2\n"
1475 "add x11, x12, x26, LSL #2\n"
1476 "add x10, x11, x26, LSL #2\n"
1477 "add x9, x10, x26, LSL #2\n"
1479 "add x28, x9, x26, LSL #2\n"
1481 "tbz x4, #1, 130f\n"
1482 "ldr d24, [x6], #0x8\n"
1484 "ldr d25, [x13], #0x8\n"
1485 "ldr d26, [x12], #0x8\n"
1486 "ldr d27, [x11], #0x8\n"
1487 "ldr d28, [x10], #0x8\n"
1488 "ldr d29, [x9], #0x8\n"
1489 "ldr d30, [x28], #0x8\n"
1490 "tbz x4, #0, 131f\n"
1491 "ld1 { v24.s }[2], [x6]\n"
1492 "ld1 { v25.s }[2], [x13]\n"
1493 "ld1 { v26.s }[2], [x12]\n"
1494 "ld1 { v27.s }[2], [x11]\n"
1495 "ld1 { v28.s }[2], [x10]\n"
1496 "ld1 { v29.s }[2], [x9]\n"
1497 "ld1 { v30.s }[2], [x28]\n"
1500 "ldr s24, [x6, #0x0]\n"
1502 "ldr s25, [x13, #0x0]\n"
1503 "ldr s26, [x12, #0x0]\n"
1504 "ldr s27, [x11, #0x0]\n"
1505 "ldr s28, [x10, #0x0]\n"
1506 "ldr s29, [x9, #0x0]\n"
1507 "ldr s30, [x28, #0x0]\n"
1512 "ldr q24, [x6, #0x0]\n"
1513 "ldr q25, [x13, #0x0]\n"
1514 "ldr q26, [x12, #0x0]\n"
1515 "ldr q27, [x11, #0x0]\n"
1516 "ldr q28, [x10, #0x0]\n"
1517 "ldr q29, [x9, #0x0]\n"
1518 "ldr q30, [x28, #0x0]\n"
1521 "movi v24.16b, #0x0\n"
1522 "movi v25.16b, #0x0\n"
1523 "movi v26.16b, #0x0\n"
1524 "movi v27.16b, #0x0\n"
1525 "movi v28.16b, #0x0\n"
1526 "movi v29.16b, #0x0\n"
1527 "movi v30.16b, #0x0\n"
1531 "ldr x26, [%x[args_ptr], %[offsetof_string_lengths]]\n"
1532 "ldr w8, [x26, x7, LSL #0x2]\n"
1533 "ldr x27, [%x[args_ptr], %[offsetof_input_offset]]\n"
1534 "tbz %x[flags], #3, 136f\n"
1535 "ldr x26, [%x[input_ptr], x7, LSL #0x3]\n"
1536 "add x26, x26, x27, LSL #3\n"
1537 "ldr x17, [x26, #0x0]\n"
1538 "ldr x16, [x26, #0x8]\n"
1539 "ldr x15, [x26, #0x10]\n"
1540 "ldr x14, [x26, #0x18]\n"
1541 "ldr x13, [x26, #0x20]\n"
1542 "ldr x12, [x26, #0x28]\n"
1543 "ldr x11, [x26, #0x30]\n"
1545 "ldr x26, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
1546 "add x17, x17, x26, LSL #2\n"
1547 "add x16, x16, x26, LSL #2\n"
1548 "add x15, x15, x26, LSL #2\n"
1549 "add x14, x14, x26, LSL #2\n"
1550 "add x13, x13, x26, LSL #2\n"
1551 "add x12, x12, x26, LSL #2\n"
1552 "add x11, x11, x26, LSL #2\n"
1555 "mov x17, %x[input_ptr]\n"
1556 "add x16, x17, x27, LSL #2\n"
1557 "add x15, x16, x27, LSL #2\n"
1558 "add x14, x15, x27, LSL #2\n"
1559 "add x13, x14, x27, LSL #2\n"
1560 "add x12, x13, x27, LSL #2\n"
1561 "add x11, x12, x27, LSL #2\n"
1565 "ldr q0, [x17, #0x0]\n"
1567 "ldr q1, [x16, #0x0]\n"
1568 "ldr q2, [x15, #0x0]\n"
1569 "ldr q3, [x14, #0x0]\n"
1570 "ldr q4, [x13, #0x0]\n"
1571 "ldr q5, [x12, #0x0]\n"
1572 "ldr q6, [x11, #0x0]\n"
1573 "ldr q8, [x5, #0x0]\n"
1574 "ldr q9, [x5, #0x10]\n"
1575 "ldr q10, [x5, #0x20]\n"
1576 "ldr q11, [x5, #0x30]\n"
1579 "fmla v24.4s, v8.4s, v0.s[0]\n"
1580 "add x17, x17, #0x10\n"
1581 "fmla v25.4s, v8.4s, v1.s[0]\n"
1582 "add x16, x16, #0x10\n"
1583 "fmla v26.4s, v8.4s, v2.s[0]\n"
1584 "add x15, x15, #0x10\n"
1585 "fmla v27.4s, v8.4s, v3.s[0]\n"
1586 "add x14, x14, #0x10\n"
1587 "fmla v28.4s, v8.4s, v4.s[0]\n"
1588 "add x13, x13, #0x10\n"
1589 "fmla v29.4s, v8.4s, v5.s[0]\n"
1590 "add x12, x12, #0x10\n"
1591 "fmla v30.4s, v8.4s, v6.s[0]\n"
1592 "add x11, x11, #0x10\n"
1593 "fmla v24.4s, v9.4s, v0.s[1]\n"
1594 "add x5, x5, #0x40\n"
1595 "ldr d8, [x5, #0x0]\n"
1596 "fmla v25.4s, v9.4s, v1.s[1]\n"
1597 "fmla v26.4s, v9.4s, v2.s[1]\n"
1598 "ldr x26, [x5, #0x8]\n"
1599 "fmla v27.4s, v9.4s, v3.s[1]\n"
1600 "ldr x10, [x5, #0x18]\n"
1601 "fmla v28.4s, v9.4s, v4.s[1]\n"
1602 "ldr x9, [x5, #0x28]\n"
1603 "fmla v29.4s, v9.4s, v5.s[1]\n"
1604 "ldr x28, [x17, #0x8]\n"
1605 "fmla v30.4s, v9.4s, v6.s[1]\n"
1606 "ldr d9, [x5, #0x10]\n"
1607 "fmla v24.4s, v10.4s, v0.s[2]\n"
1608 "ldr x27, [x16, #0x8]\n"
1609 "fmla v25.4s, v10.4s, v1.s[2]\n"
1610 "sub x8, x8, #0x4\n"
1611 "fmla v26.4s, v10.4s, v2.s[2]\n"
1613 "fmla v27.4s, v10.4s, v3.s[2]\n"
1614 "mov v8.d[1], x26\n"
1615 "fmla v28.4s, v10.4s, v4.s[2]\n"
1616 "ldr x26, [x15, #0x8]\n"
1617 "fmla v29.4s, v10.4s, v5.s[2]\n"
1618 "prfm pldl1keep, [x17, #0x80]\n"
1619 "fmla v30.4s, v10.4s, v6.s[2]\n"
1620 "ldr d10, [x5, #0x20]\n"
1621 "fmla v24.4s, v11.4s, v0.s[3]\n"
1622 "ldr d0, [x17, #0x0]\n"
1623 "fmla v25.4s, v11.4s, v1.s[3]\n"
1624 "ldr d1, [x16, #0x0]\n"
1625 "fmla v26.4s, v11.4s, v2.s[3]\n"
1626 "ldr d2, [x15, #0x0]\n"
1627 "fmla v27.4s, v11.4s, v3.s[3]\n"
1628 "ldr d3, [x14, #0x0]\n"
1629 "fmla v28.4s, v11.4s, v4.s[3]\n"
1630 "ldr d4, [x13, #0x0]\n"
1631 "fmla v29.4s, v11.4s, v5.s[3]\n"
1632 "ldr d5, [x12, #0x0]\n"
1633 "fmla v30.4s, v11.4s, v6.s[3]\n"
1634 "ldr d6, [x11, #0x0]\n"
1635 "ldr d11, [x5, #0x30]\n"
1636 "mov v9.d[1], x10\n"
1637 "ldr x10, [x14, #0x8]\n"
1638 "mov v10.d[1], x9\n"
1639 "ldr x9, [x13, #0x8]\n"
1640 "mov v0.d[1], x28\n"
1641 "ldr x28, [x12, #0x8]\n"
1642 "mov v1.d[1], x27\n"
1643 "ldr x27, [x11, #0x8]\n"
1644 "mov v2.d[1], x26\n"
1645 "ldr x26, [x5, #0x38]\n"
1646 "mov v3.d[1], x10\n"
1647 "prfm pldl1keep, [x16, #0x80]\n"
1649 "prfm pldl1keep, [x15, #0x80]\n"
1650 "mov v5.d[1], x28\n"
1651 "prfm pldl1keep, [x14, #0x80]\n"
1652 "mov v6.d[1], x27\n"
1653 "prfm pldl1keep, [x13, #0x80]\n"
1654 "mov v11.d[1], x26\n"
1655 "prfm pldl1keep, [x12, #0x80]\n"
1656 "prfm pldl1keep, [x11, #0x80]\n"
1659 "fmla v24.4s, v8.4s, v0.s[0]\n"
1660 "add x17, x17, #0x10\n"
1661 "fmla v25.4s, v8.4s, v1.s[0]\n"
1662 "add x16, x16, #0x10\n"
1663 "fmla v26.4s, v8.4s, v2.s[0]\n"
1664 "add x15, x15, #0x10\n"
1665 "fmla v27.4s, v8.4s, v3.s[0]\n"
1666 "add x14, x14, #0x10\n"
1667 "fmla v28.4s, v8.4s, v4.s[0]\n"
1668 "add x13, x13, #0x10\n"
1669 "fmla v29.4s, v8.4s, v5.s[0]\n"
1670 "add x12, x12, #0x10\n"
1671 "fmla v30.4s, v8.4s, v6.s[0]\n"
1672 "add x11, x11, #0x10\n"
1673 "fmla v24.4s, v9.4s, v0.s[1]\n"
1674 "sub x8, x8, #0x4\n"
1675 "fmla v25.4s, v9.4s, v1.s[1]\n"
1676 "prfm pldl1keep, [x17, #0x80]\n"
1677 "fmla v26.4s, v9.4s, v2.s[1]\n"
1678 "prfm pldl1keep, [x16, #0x80]\n"
1679 "fmla v27.4s, v9.4s, v3.s[1]\n"
1680 "prfm pldl1keep, [x15, #0x80]\n"
1681 "fmla v28.4s, v9.4s, v4.s[1]\n"
1682 "prfm pldl1keep, [x14, #0x80]\n"
1683 "fmla v29.4s, v9.4s, v5.s[1]\n"
1684 "prfm pldl1keep, [x13, #0x80]\n"
1685 "fmla v30.4s, v9.4s, v6.s[1]\n"
1686 "prfm pldl1keep, [x12, #0x80]\n"
1687 "fmla v24.4s, v10.4s, v0.s[2]\n"
1688 "prfm pldl1keep, [x11, #0x80]\n"
1689 "fmla v25.4s, v10.4s, v1.s[2]\n"
1690 "add x5, x5, #0x40\n"
1691 "fmla v26.4s, v10.4s, v2.s[2]\n"
1692 "fmla v27.4s, v10.4s, v3.s[2]\n"
1693 "fmla v28.4s, v10.4s, v4.s[2]\n"
1694 "fmla v29.4s, v10.4s, v5.s[2]\n"
1695 "fmla v30.4s, v10.4s, v6.s[2]\n"
1696 "fmla v24.4s, v11.4s, v0.s[3]\n"
1697 "fmla v25.4s, v11.4s, v1.s[3]\n"
1698 "fmla v26.4s, v11.4s, v2.s[3]\n"
1699 "fmla v27.4s, v11.4s, v3.s[3]\n"
1700 "fmla v28.4s, v11.4s, v4.s[3]\n"
1701 "fmla v29.4s, v11.4s, v5.s[3]\n"
1702 "fmla v30.4s, v11.4s, v6.s[3]\n"
1706 "ldr s23, [x17], #0x4\n"
1707 "sub x8, x8, #0x1\n"
1708 "ldr s22, [x16], #0x4\n"
1709 "ldr s21, [x15], #0x4\n"
1710 "ldr s20, [x14], #0x4\n"
1711 "ldr s19, [x13], #0x4\n"
1712 "ldr s18, [x12], #0x4\n"
1713 "ldr s17, [x11], #0x4\n"
1714 "ldr q16, [x5, #0x0]\n"
1715 "fmla v24.4s, v16.4s, v23.s[0]\n"
1716 "fmla v25.4s, v16.4s, v22.s[0]\n"
1717 "add x5, x5, #0x10\n"
1718 "fmla v26.4s, v16.4s, v21.s[0]\n"
1719 "fmla v27.4s, v16.4s, v20.s[0]\n"
1720 "fmla v28.4s, v16.4s, v19.s[0]\n"
1721 "fmla v29.4s, v16.4s, v18.s[0]\n"
1722 "fmla v30.4s, v16.4s, v17.s[0]\n"
1725 "ldr w26, [%x[args_ptr], %[offsetof_num_strings]]\n"
1726 "add x7, x7, #0x1\n"
1729 "ldr x26, [%x[args_ptr], %[offsetof_output_offset]]\n"
1730 "add x13, x6, x26, LSL #2\n"
1731 "add x12, x13, x26, LSL #2\n"
1732 "add x11, x12, x26, LSL #2\n"
1733 "add x10, x11, x26, LSL #2\n"
1734 "add x9, x10, x26, LSL #2\n"
1735 "add x28, x9, x26, LSL #2\n"
1736 "prfm pstl1keep, [x6, #0x0]\n"
1737 "prfm pstl1keep, [x13, #0x0]\n"
1738 "prfm pstl1keep, [x12, #0x0]\n"
1739 "prfm pstl1keep, [x11, #0x0]\n"
1740 "prfm pstl1keep, [x10, #0x0]\n"
1741 "prfm pstl1keep, [x9, #0x0]\n"
1742 "prfm pstl1keep, [x28, #0x0]\n"
1743 "tbz %x[flags], #1, 143f\n"
1744 "add x26, %x[args_ptr], %[offset_max]\n"
1745 "ld1r { v16.4s }, [x26]\n"
1746 "fmin v24.4s, v24.4s, v16.4s\n"
1747 "fmin v25.4s, v25.4s, v16.4s\n"
1748 "fmin v26.4s, v26.4s, v16.4s\n"
1749 "fmin v27.4s, v27.4s, v16.4s\n"
1750 "fmin v28.4s, v28.4s, v16.4s\n"
1751 "fmin v29.4s, v29.4s, v16.4s\n"
1752 "fmin v30.4s, v30.4s, v16.4s\n"
1753 "add x26, %x[args_ptr], %[offset_min]\n"
1754 "ld1r { v16.4s }, [x26]\n"
1755 "fmax v24.4s, v24.4s, v16.4s\n"
1756 "fmax v25.4s, v25.4s, v16.4s\n"
1757 "fmax v26.4s, v26.4s, v16.4s\n"
1758 "fmax v27.4s, v27.4s, v16.4s\n"
1759 "fmax v28.4s, v28.4s, v16.4s\n"
1760 "fmax v29.4s, v29.4s, v16.4s\n"
1761 "fmax v30.4s, v30.4s, v16.4s\n"
1765 "tbz x4, #1, 144f\n"
1766 "str d24, [x6], #0x8\n"
1767 "str d25, [x13], #0x8\n"
1768 "str d26, [x12], #0x8\n"
1769 "str d27, [x11], #0x8\n"
1770 "str d28, [x10], #0x8\n"
1771 "str d29, [x9], #0x8\n"
1772 "str d30, [x28], #0x8\n"
1773 "tbz x4, #0, 145f\n"
1774 "st1 { v24.s }[2], [x6]\n"
1775 "st1 { v25.s }[2], [x13]\n"
1776 "st1 { v26.s }[2], [x12]\n"
1777 "st1 { v27.s }[2], [x11]\n"
1778 "st1 { v28.s }[2], [x10]\n"
1779 "st1 { v29.s }[2], [x9]\n"
1780 "st1 { v30.s }[2], [x28]\n"
1783 "str s24, [x6, #0x0]\n"
1784 "str s25, [x13, #0x0]\n"
1785 "str s26, [x12, #0x0]\n"
1786 "str s27, [x11, #0x0]\n"
1787 "str s28, [x10, #0x0]\n"
1788 "str s29, [x9, #0x0]\n"
1789 "str s30, [x28, #0x0]\n"
1793 "str q24, [x6, #0x0]\n"
1794 "add x6, x6, #0x10\n"
1795 "str q25, [x13, #0x0]\n"
1796 "str q26, [x12, #0x0]\n"
1797 "str q27, [x11, #0x0]\n"
1798 "str q28, [x10, #0x0]\n"
1799 "str q29, [x9, #0x0]\n"
1800 "str q30, [x28, #0x0]\n"
1802 "subs x4, x4, #0x4\n"
1806 "ldr x27, [%x[args_ptr], %[offsetof_output_offset]]\n"
1808 "mov x3, %x[bias]\n"
1809 "ldr x4, [%x[args_ptr], %[offsetof_N]]\n"
1810 "ldr x5, [%x[args_ptr], %[offsetof_B_ptr]]\n"
1811 "mov x6, %x[output_ptr]\n"
1812 "madd %x[output_ptr], x27, x26, %x[output_ptr]\n"
1815 "ldr q24, [x3, #0x0]\n"
1816 "mov v25.16b, v24.16b\n"
1817 "mov v26.16b, v24.16b\n"
1818 "add x3, x3, #0x10\n"
1819 "mov v27.16b, v24.16b\n"
1820 "mov v28.16b, v24.16b\n"
1821 "mov v29.16b, v24.16b\n"
1822 "mov v30.16b, v24.16b\n"
1823 "mov v31.16b, v24.16b\n"
1826 "tbz %x[flags], #0, 154f\n"
1827 "ldr x26, [%x[args_ptr], %[offsetof_output_offset]]\n"
1828 "add x13, x6, x26, LSL #2\n"
1829 "add x12, x13, x26, LSL #2\n"
1830 "add x11, x12, x26, LSL #2\n"
1831 "add x10, x11, x26, LSL #2\n"
1832 "add x9, x10, x26, LSL #2\n"
1833 "add x28, x9, x26, LSL #2\n"
1835 "add x27, x28, x26, LSL #2\n"
1837 "tbz x4, #1, 151f\n"
1838 "ldr d24, [x6], #0x8\n"
1840 "ldr d25, [x13], #0x8\n"
1841 "ldr d26, [x12], #0x8\n"
1842 "ldr d27, [x11], #0x8\n"
1843 "ldr d28, [x10], #0x8\n"
1844 "ldr d29, [x9], #0x8\n"
1845 "ldr d30, [x28], #0x8\n"
1846 "ldr d31, [x27], #0x8\n"
1847 "tbz x4, #0, 152f\n"
1848 "ld1 { v24.s }[2], [x6]\n"
1849 "ld1 { v25.s }[2], [x13]\n"
1850 "ld1 { v26.s }[2], [x12]\n"
1851 "ld1 { v27.s }[2], [x11]\n"
1852 "ld1 { v28.s }[2], [x10]\n"
1853 "ld1 { v29.s }[2], [x9]\n"
1854 "ld1 { v30.s }[2], [x28]\n"
1855 "ld1 { v31.s }[2], [x27]\n"
1858 "ldr s24, [x6, #0x0]\n"
1860 "ldr s25, [x13, #0x0]\n"
1861 "ldr s26, [x12, #0x0]\n"
1862 "ldr s27, [x11, #0x0]\n"
1863 "ldr s28, [x10, #0x0]\n"
1864 "ldr s29, [x9, #0x0]\n"
1865 "ldr s30, [x28, #0x0]\n"
1866 "ldr s31, [x27, #0x0]\n"
1871 "ldr q24, [x6, #0x0]\n"
1872 "ldr q25, [x13, #0x0]\n"
1873 "ldr q26, [x12, #0x0]\n"
1874 "ldr q27, [x11, #0x0]\n"
1875 "ldr q28, [x10, #0x0]\n"
1876 "ldr q29, [x9, #0x0]\n"
1877 "ldr q30, [x28, #0x0]\n"
1878 "ldr q31, [x27, #0x0]\n"
1881 "movi v24.16b, #0x0\n"
1882 "movi v25.16b, #0x0\n"
1883 "movi v26.16b, #0x0\n"
1884 "movi v27.16b, #0x0\n"
1885 "movi v28.16b, #0x0\n"
1886 "movi v29.16b, #0x0\n"
1887 "movi v30.16b, #0x0\n"
1888 "movi v31.16b, #0x0\n"
1892 "ldr x26, [%x[args_ptr], %[offsetof_string_lengths]]\n"
1893 "ldr w8, [x26, x7, LSL #0x2]\n"
1894 "ldr x27, [%x[args_ptr], %[offsetof_input_offset]]\n"
1895 "tbz %x[flags], #3, 157f\n"
1896 "ldr x26, [%x[input_ptr], x7, LSL #0x3]\n"
1897 "add x26, x26, x27, LSL #3\n"
1898 "ldr x17, [x26, #0x0]\n"
1899 "ldr x16, [x26, #0x8]\n"
1900 "ldr x15, [x26, #0x10]\n"
1901 "ldr x14, [x26, #0x18]\n"
1902 "ldr x13, [x26, #0x20]\n"
1903 "ldr x12, [x26, #0x28]\n"
1904 "ldr x11, [x26, #0x30]\n"
1905 "ldr x27, [x26, #0x38]\n"
1907 "ldr x26, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
1908 "add x17, x17, x26, LSL #2\n"
1909 "add x16, x16, x26, LSL #2\n"
1910 "add x15, x15, x26, LSL #2\n"
1911 "add x14, x14, x26, LSL #2\n"
1912 "add x13, x13, x26, LSL #2\n"
1913 "add x12, x12, x26, LSL #2\n"
1914 "add x11, x11, x26, LSL #2\n"
1915 "add x27, x27, x26, LSL #2\n"
1918 "mov x17, %x[input_ptr]\n"
1919 "add x16, x17, x27, LSL #2\n"
1920 "add x15, x16, x27, LSL #2\n"
1921 "add x14, x15, x27, LSL #2\n"
1922 "add x13, x14, x27, LSL #2\n"
1923 "add x12, x13, x27, LSL #2\n"
1924 "add x11, x12, x27, LSL #2\n"
1925 "add x27, x11, x27, LSL #2\n"
1929 "ldr q0, [x17, #0x0]\n"
1931 "ldr q1, [x16, #0x0]\n"
1932 "ldr q2, [x15, #0x0]\n"
1933 "ldr q3, [x14, #0x0]\n"
1934 "ldr q4, [x13, #0x0]\n"
1935 "ldr q5, [x12, #0x0]\n"
1936 "ldr q6, [x11, #0x0]\n"
1937 "ldr q7, [x27, #0x0]\n"
1938 "ldr q8, [x5, #0x0]\n"
1939 "ldr q9, [x5, #0x10]\n"
1940 "ldr q10, [x5, #0x20]\n"
1941 "ldr q11, [x5, #0x30]\n"
1944 "fmla v24.4s, v8.4s, v0.s[0]\n"
1945 "add x17, x17, #0x10\n"
1946 "fmla v25.4s, v8.4s, v1.s[0]\n"
1947 "add x16, x16, #0x10\n"
1948 "fmla v26.4s, v8.4s, v2.s[0]\n"
1949 "add x15, x15, #0x10\n"
1950 "fmla v27.4s, v8.4s, v3.s[0]\n"
1951 "add x14, x14, #0x10\n"
1952 "fmla v28.4s, v8.4s, v4.s[0]\n"
1953 "add x13, x13, #0x10\n"
1954 "fmla v29.4s, v8.4s, v5.s[0]\n"
1955 "add x12, x12, #0x10\n"
1956 "fmla v30.4s, v8.4s, v6.s[0]\n"
1957 "add x11, x11, #0x10\n"
1958 "fmla v31.4s, v8.4s, v7.s[0]\n"
1959 "add x27, x27, #0x10\n"
1960 "fmla v24.4s, v9.4s, v0.s[1]\n"
1961 "add x5, x5, #0x40\n"
1962 "ldr d8, [x5, #0x0]\n"
1963 "fmla v25.4s, v9.4s, v1.s[1]\n"
1964 "fmla v26.4s, v9.4s, v2.s[1]\n"
1965 "ldr x26, [x5, #0x8]\n"
1966 "fmla v27.4s, v9.4s, v3.s[1]\n"
1967 "sub x8, x8, #0x4\n"
1968 "fmla v28.4s, v9.4s, v4.s[1]\n"
1970 "fmla v29.4s, v9.4s, v5.s[1]\n"
1971 "mov v8.d[1], x26\n"
1972 "fmla v30.4s, v9.4s, v6.s[1]\n"
1973 "ldr x26, [x5, #0x18]\n"
1974 "fmla v31.4s, v9.4s, v7.s[1]\n"
1975 "ldr d9, [x5, #0x10]\n"
1976 "fmla v24.4s, v10.4s, v0.s[2]\n"
1977 "prfm pldl1keep, [x17, #0x80]\n"
1978 "fmla v25.4s, v10.4s, v1.s[2]\n"
1979 "prfm pldl1keep, [x16, #0x80]\n"
1980 "fmla v26.4s, v10.4s, v2.s[2]\n"
1981 "prfm pldl1keep, [x15, #0x80]\n"
1982 "fmla v27.4s, v10.4s, v3.s[2]\n"
1983 "mov v9.d[1], x26\n"
1984 "fmla v28.4s, v10.4s, v4.s[2]\n"
1985 "ldr x26, [x5, #0x28]\n"
1986 "fmla v29.4s, v10.4s, v5.s[2]\n"
1987 "prfm pldl1keep, [x14, #0x80]\n"
1988 "fmla v30.4s, v10.4s, v6.s[2]\n"
1989 "prfm pldl1keep, [x13, #0x80]\n"
1990 "fmla v31.4s, v10.4s, v7.s[2]\n"
1991 "ldr d10, [x5, #0x20]\n"
1992 "fmla v24.4s, v11.4s, v0.s[3]\n"
1993 "ldr d0, [x17, #0x0]\n"
1994 "fmla v25.4s, v11.4s, v1.s[3]\n"
1995 "ldr d1, [x16, #0x0]\n"
1996 "fmla v26.4s, v11.4s, v2.s[3]\n"
1997 "ldr d2, [x15, #0x0]\n"
1998 "fmla v27.4s, v11.4s, v3.s[3]\n"
1999 "ldr d3, [x14, #0x0]\n"
2000 "fmla v28.4s, v11.4s, v4.s[3]\n"
2001 "ldr d4, [x13, #0x0]\n"
2002 "fmla v29.4s, v11.4s, v5.s[3]\n"
2003 "ldr d5, [x12, #0x0]\n"
2004 "fmla v30.4s, v11.4s, v6.s[3]\n"
2005 "ldr d6, [x11, #0x0]\n"
2006 "fmla v31.4s, v11.4s, v7.s[3]\n"
2007 "ldr d7, [x27, #0x0]\n"
2008 "ldr d11, [x5, #0x30]\n"
2009 "mov v10.d[1], x26\n"
2010 "ldr x26, [x17, #0x8]\n"
2011 "mov v0.d[1], x26\n"
2012 "ldr x26, [x16, #0x8]\n"
2013 "mov v1.d[1], x26\n"
2014 "ldr x26, [x15, #0x8]\n"
2015 "mov v2.d[1], x26\n"
2016 "ldr x26, [x14, #0x8]\n"
2017 "mov v3.d[1], x26\n"
2018 "ldr x26, [x13, #0x8]\n"
2019 "mov v4.d[1], x26\n"
2020 "ldr x26, [x12, #0x8]\n"
2021 "mov v5.d[1], x26\n"
2022 "ldr x26, [x11, #0x8]\n"
2023 "mov v6.d[1], x26\n"
2024 "ldr x26, [x27, #0x8]\n"
2025 "mov v7.d[1], x26\n"
2026 "ldr x26, [x5, #0x38]\n"
2027 "mov v11.d[1], x26\n"
2028 "prfm pldl1keep, [x12, #0x80]\n"
2029 "prfm pldl1keep, [x11, #0x80]\n"
2030 "prfm pldl1keep, [x27, #0x80]\n"
2033 "fmla v24.4s, v8.4s, v0.s[0]\n"
2034 "add x17, x17, #0x10\n"
2035 "fmla v25.4s, v8.4s, v1.s[0]\n"
2036 "add x16, x16, #0x10\n"
2037 "fmla v26.4s, v8.4s, v2.s[0]\n"
2038 "add x15, x15, #0x10\n"
2039 "fmla v27.4s, v8.4s, v3.s[0]\n"
2040 "add x14, x14, #0x10\n"
2041 "fmla v28.4s, v8.4s, v4.s[0]\n"
2042 "add x13, x13, #0x10\n"
2043 "fmla v29.4s, v8.4s, v5.s[0]\n"
2044 "add x12, x12, #0x10\n"
2045 "fmla v30.4s, v8.4s, v6.s[0]\n"
2046 "add x11, x11, #0x10\n"
2047 "fmla v31.4s, v8.4s, v7.s[0]\n"
2048 "add x27, x27, #0x10\n"
2049 "fmla v24.4s, v9.4s, v0.s[1]\n"
2050 "sub x8, x8, #0x4\n"
2051 "fmla v25.4s, v9.4s, v1.s[1]\n"
2052 "prfm pldl1keep, [x17, #0x80]\n"
2053 "fmla v26.4s, v9.4s, v2.s[1]\n"
2054 "prfm pldl1keep, [x16, #0x80]\n"
2055 "fmla v27.4s, v9.4s, v3.s[1]\n"
2056 "prfm pldl1keep, [x15, #0x80]\n"
2057 "fmla v28.4s, v9.4s, v4.s[1]\n"
2058 "prfm pldl1keep, [x14, #0x80]\n"
2059 "fmla v29.4s, v9.4s, v5.s[1]\n"
2060 "prfm pldl1keep, [x13, #0x80]\n"
2061 "fmla v30.4s, v9.4s, v6.s[1]\n"
2062 "prfm pldl1keep, [x12, #0x80]\n"
2063 "fmla v31.4s, v9.4s, v7.s[1]\n"
2064 "prfm pldl1keep, [x11, #0x80]\n"
2065 "fmla v24.4s, v10.4s, v0.s[2]\n"
2066 "prfm pldl1keep, [x27, #0x80]\n"
2067 "fmla v25.4s, v10.4s, v1.s[2]\n"
2068 "add x5, x5, #0x40\n"
2069 "fmla v26.4s, v10.4s, v2.s[2]\n"
2070 "fmla v27.4s, v10.4s, v3.s[2]\n"
2071 "fmla v28.4s, v10.4s, v4.s[2]\n"
2072 "fmla v29.4s, v10.4s, v5.s[2]\n"
2073 "fmla v30.4s, v10.4s, v6.s[2]\n"
2074 "fmla v31.4s, v10.4s, v7.s[2]\n"
2075 "fmla v24.4s, v11.4s, v0.s[3]\n"
2076 "fmla v25.4s, v11.4s, v1.s[3]\n"
2077 "fmla v26.4s, v11.4s, v2.s[3]\n"
2078 "fmla v27.4s, v11.4s, v3.s[3]\n"
2079 "fmla v28.4s, v11.4s, v4.s[3]\n"
2080 "fmla v29.4s, v11.4s, v5.s[3]\n"
2081 "fmla v30.4s, v11.4s, v6.s[3]\n"
2082 "fmla v31.4s, v11.4s, v7.s[3]\n"
2086 "ldr s0, [x17], #0x4\n"
2087 "sub x8, x8, #0x1\n"
2088 "ldr s23, [x16], #0x4\n"
2089 "ldr s22, [x15], #0x4\n"
2090 "ldr s21, [x14], #0x4\n"
2091 "ldr s20, [x13], #0x4\n"
2092 "ldr s19, [x12], #0x4\n"
2093 "ldr s18, [x11], #0x4\n"
2094 "ldr s17, [x27], #0x4\n"
2095 "ldr q16, [x5, #0x0]\n"
2096 "fmla v24.4s, v16.4s, v0.s[0]\n"
2097 "fmla v25.4s, v16.4s, v23.s[0]\n"
2098 "add x5, x5, #0x10\n"
2099 "fmla v26.4s, v16.4s, v22.s[0]\n"
2100 "fmla v27.4s, v16.4s, v21.s[0]\n"
2101 "fmla v28.4s, v16.4s, v20.s[0]\n"
2102 "fmla v29.4s, v16.4s, v19.s[0]\n"
2103 "fmla v30.4s, v16.4s, v18.s[0]\n"
2104 "fmla v31.4s, v16.4s, v17.s[0]\n"
2107 "ldr w26, [%x[args_ptr], %[offsetof_num_strings]]\n"
2108 "add x7, x7, #0x1\n"
2111 "ldr x26, [%x[args_ptr], %[offsetof_output_offset]]\n"
2112 "add x13, x6, x26, LSL #2\n"
2113 "add x12, x13, x26, LSL #2\n"
2114 "add x11, x12, x26, LSL #2\n"
2115 "add x10, x11, x26, LSL #2\n"
2116 "add x9, x10, x26, LSL #2\n"
2117 "add x28, x9, x26, LSL #2\n"
2118 "add x27, x28, x26, LSL #2\n"
2119 "prfm pstl1keep, [x6, #0x0]\n"
2120 "prfm pstl1keep, [x13, #0x0]\n"
2121 "prfm pstl1keep, [x12, #0x0]\n"
2122 "prfm pstl1keep, [x11, #0x0]\n"
2123 "prfm pstl1keep, [x10, #0x0]\n"
2124 "prfm pstl1keep, [x9, #0x0]\n"
2125 "prfm pstl1keep, [x28, #0x0]\n"
2126 "prfm pstl1keep, [x27, #0x0]\n"
2127 "tbz %x[flags], #1, 164f\n"
2128 "add x26, %x[args_ptr], %[offset_max]\n"
2129 "ld1r { v16.4s }, [x26]\n"
2130 "fmin v24.4s, v24.4s, v16.4s\n"
2131 "fmin v25.4s, v25.4s, v16.4s\n"
2132 "fmin v26.4s, v26.4s, v16.4s\n"
2133 "fmin v27.4s, v27.4s, v16.4s\n"
2134 "fmin v28.4s, v28.4s, v16.4s\n"
2135 "fmin v29.4s, v29.4s, v16.4s\n"
2136 "fmin v30.4s, v30.4s, v16.4s\n"
2137 "fmin v31.4s, v31.4s, v16.4s\n"
2138 "add x26, %x[args_ptr], %[offset_min]\n"
2139 "ld1r { v16.4s }, [x26]\n"
2140 "fmax v24.4s, v24.4s, v16.4s\n"
2141 "fmax v25.4s, v25.4s, v16.4s\n"
2142 "fmax v26.4s, v26.4s, v16.4s\n"
2143 "fmax v27.4s, v27.4s, v16.4s\n"
2144 "fmax v28.4s, v28.4s, v16.4s\n"
2145 "fmax v29.4s, v29.4s, v16.4s\n"
2146 "fmax v30.4s, v30.4s, v16.4s\n"
2147 "fmax v31.4s, v31.4s, v16.4s\n"
2151 "tbz x4, #1, 165f\n"
2152 "str d24, [x6], #0x8\n"
2153 "str d25, [x13], #0x8\n"
2154 "str d26, [x12], #0x8\n"
2155 "str d27, [x11], #0x8\n"
2156 "str d28, [x10], #0x8\n"
2157 "str d29, [x9], #0x8\n"
2158 "str d30, [x28], #0x8\n"
2159 "str d31, [x27], #0x8\n"
2160 "tbz x4, #0, 166f\n"
2161 "st1 { v24.s }[2], [x6]\n"
2162 "st1 { v25.s }[2], [x13]\n"
2163 "st1 { v26.s }[2], [x12]\n"
2164 "st1 { v27.s }[2], [x11]\n"
2165 "st1 { v28.s }[2], [x10]\n"
2166 "st1 { v29.s }[2], [x9]\n"
2167 "st1 { v30.s }[2], [x28]\n"
2168 "st1 { v31.s }[2], [x27]\n"
2171 "str s24, [x6, #0x0]\n"
2172 "str s25, [x13, #0x0]\n"
2173 "str s26, [x12, #0x0]\n"
2174 "str s27, [x11, #0x0]\n"
2175 "str s28, [x10, #0x0]\n"
2176 "str s29, [x9, #0x0]\n"
2177 "str s30, [x28, #0x0]\n"
2178 "str s31, [x27, #0x0]\n"
2182 "str q24, [x6, #0x0]\n"
2183 "add x6, x6, #0x10\n"
2184 "str q25, [x13, #0x0]\n"
2185 "str q26, [x12, #0x0]\n"
2186 "str q27, [x11, #0x0]\n"
2187 "str q28, [x10, #0x0]\n"
2188 "str q29, [x9, #0x0]\n"
2189 "str q30, [x28, #0x0]\n"
2190 "str q31, [x27, #0x0]\n"
2192 "subs x4, x4, #0x4\n"
2194 "subs %x[M], %x[M], #0x8\n"
2196 "ldr x27, [%x[args_ptr], %[offsetof_input_offset]]\n"
2197 "tbz %x[flags], #3, 169f\n"
2198 "add x27, x27, #0x8\n"
2199 "str x27, [%x[args_ptr], %[offsetof_input_offset]]\n"
2203 "madd %x[input_ptr], x26, x27, %x[input_ptr]\n"
2206 : [
M]
"+&r" (
M), [input_ptr]
"+&r" (input_ptr), [output_ptr]
"+&r" (output_ptr)
2207 : [args_ptr]
"r" (&ka), [
bias]
"r" (
bias), [flags]
"r" (flags), [offset_max]
"I" (offsetof(KernelArgs, maxval)), [offset_min]
"I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr]
"I" (offsetof(KernelArgs, B_ptr)), [offsetof_N]
"I" (offsetof(KernelArgs,
N)), [offsetof_input_initial_col]
"I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset]
"I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings]
"I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset]
"I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths]
"I" (offsetof(KernelArgs, string_lengths))
2208 :
"cc",
"memory",
"v0",
"v1",
"v2",
"v3",
"v4",
"v5",
"v6",
"v7",
"v8",
"v9",
"v10",
"v11",
"v16",
"v17",
"v18",
"v19",
"v20",
"v21",
"v22",
"v23",
"v24",
"v25",
"v26",
"v27",
"v28",
"v29",
"v30",
"v31",
"x3",
"x4",
"x5",
"x6",
"x7",
"x8",
"x9",
"x10",
"x11",
"x12",
"x13",
"x14",
"x15",
"x16",
"x17",
"x26",
"x27",
"x28"
2213 #endif // __aarch64__