27 #include "../../utils.hpp"
34 void a64_hybrid_fp32_mla_8x4 (
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);
107 "mov x14, %x[bias]\n"
108 "ldr x13, [%x[args_ptr], %[offsetof_N]]\n"
109 "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
110 "mov x11, %x[output_ptr]\n"
113 "ldr q24, [x14, #0x0]\n"
114 "add x14, x14, #0x10\n"
117 "tbz %x[flags], #0, 7f\n"
121 "ldr d24, [x11], #0x8\n"
124 "ld1 { v24.s }[2], [x11]\n"
127 "ldr s24, [x11, #0x0]\n"
130 "sub x11, x11, x20\n"
133 "ldr q24, [x11, #0x0]\n"
136 "movi v24.16b, #0x0\n"
140 "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
141 "ldr w9, [x20, x10, LSL #0x2]\n"
142 "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
143 "tbz %x[flags], #3, 10f\n"
144 "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n"
145 "add x20, x20, x21, LSL #3\n"
146 "ldr x28, [x20, #0x0]\n"
148 "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
149 "add x28, x28, x20, LSL #2\n"
152 "mov x28, %x[input_ptr]\n"
156 "ldr q0, [x28, #0x0]\n"
157 "ldr q8, [x12, #0x0]\n"
159 "ldr q9, [x12, #0x10]\n"
160 "ldr q10, [x12, #0x20]\n"
161 "ldr q11, [x12, #0x30]\n"
164 "fmla v24.4s, v8.4s, v0.s[0]\n"
165 "fmla v24.4s, v9.4s, v0.s[1]\n"
167 "add x28, x28, #0x10\n"
168 "fmla v24.4s, v10.4s, v0.s[2]\n"
170 "add x12, x12, #0x40\n"
171 "ldr q8, [x12, #0x0]\n"
172 "ldr q9, [x12, #0x10]\n"
173 "ldr q10, [x12, #0x20]\n"
174 "fmla v24.4s, v11.4s, v0.s[3]\n"
175 "ldr q0, [x28, #0x0]\n"
176 "ldr q11, [x12, #0x30]\n"
177 "prfm pldl1keep, [x28, #0x80]\n"
180 "fmla v24.4s, v8.4s, v0.s[0]\n"
181 "fmla v24.4s, v9.4s, v0.s[1]\n"
182 "add x28, x28, #0x10\n"
184 "fmla v24.4s, v10.4s, v0.s[2]\n"
185 "fmla v24.4s, v11.4s, v0.s[3]\n"
186 "prfm pldl1keep, [x28, #0x80]\n"
187 "add x12, x12, #0x40\n"
191 "ldr s17, [x28], #0x4\n"
192 "ldr q16, [x12, #0x0]\n"
194 "fmla v24.4s, v16.4s, v17.s[0]\n"
195 "add x12, x12, #0x10\n"
198 "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
199 "add x10, x10, #0x1\n"
202 "prfm pstl1keep, [x11, #0x0]\n"
203 "tbz %x[flags], #1, 17f\n"
204 "add x20, %x[args_ptr], %[offset_max]\n"
205 "ld1r { v17.4s }, [x20]\n"
206 "add x20, %x[args_ptr], %[offset_min]\n"
207 "ld1r { v16.4s }, [x20]\n"
208 "fmin v24.4s, v24.4s, v17.4s\n"
209 "fmax v24.4s, v24.4s, v16.4s\n"
214 "str d24, [x11], #0x8\n"
216 "st1 { v24.s }[2], [x11]\n"
219 "str s24, [x11, #0x0]\n"
223 "str q24, [x11, #0x0]\n"
224 "add x11, x11, #0x10\n"
226 "subs x13, x13, #0x4\n"
230 "mov x14, %x[bias]\n"
231 "ldr x13, [%x[args_ptr], %[offsetof_N]]\n"
232 "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
233 "mov x11, %x[output_ptr]\n"
236 "ldr q24, [x14, #0x0]\n"
237 "mov v25.16b, v24.16b\n"
238 "add x14, x14, #0x10\n"
241 "tbz %x[flags], #0, 28f\n"
242 "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
244 "add x27, x11, x20, LSL #2\n"
247 "ldr d24, [x11], #0x8\n"
248 "ldr d25, [x27], #0x8\n"
251 "ld1 { v24.s }[2], [x11]\n"
252 "ld1 { v25.s }[2], [x27]\n"
255 "ldr s24, [x11, #0x0]\n"
256 "ldr s25, [x27, #0x0]\n"
259 "sub x11, x11, x20\n"
262 "ldr q24, [x11, #0x0]\n"
263 "ldr q25, [x27, #0x0]\n"
266 "movi v24.16b, #0x0\n"
267 "movi v25.16b, #0x0\n"
271 "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
272 "ldr w9, [x20, x10, LSL #0x2]\n"
273 "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
274 "tbz %x[flags], #3, 31f\n"
275 "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n"
276 "add x20, x20, x21, LSL #3\n"
277 "ldr x28, [x20, #0x0]\n"
278 "ldr x27, [x20, #0x8]\n"
280 "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
281 "add x28, x28, x20, LSL #2\n"
282 "add x27, x27, x20, LSL #2\n"
285 "mov x28, %x[input_ptr]\n"
286 "add x27, x28, x21, LSL #2\n"
290 "ldr q0, [x28, #0x0]\n"
291 "ldr q1, [x27, #0x0]\n"
293 "ldr q8, [x12, #0x0]\n"
294 "ldr q9, [x12, #0x10]\n"
295 "ldr q10, [x12, #0x20]\n"
296 "ldr q11, [x12, #0x30]\n"
299 "fmla v24.4s, v8.4s, v0.s[0]\n"
300 "fmla v25.4s, v8.4s, v1.s[0]\n"
302 "add x28, x28, #0x10\n"
303 "fmla v24.4s, v9.4s, v0.s[1]\n"
304 "fmla v25.4s, v9.4s, v1.s[1]\n"
305 "add x27, x27, #0x10\n"
307 "fmla v24.4s, v10.4s, v0.s[2]\n"
308 "fmla v25.4s, v10.4s, v1.s[2]\n"
309 "add x12, x12, #0x40\n"
310 "ldr q8, [x12, #0x0]\n"
311 "ldr q9, [x12, #0x10]\n"
312 "ldr q10, [x12, #0x20]\n"
313 "fmla v24.4s, v11.4s, v0.s[3]\n"
314 "fmla v25.4s, v11.4s, v1.s[3]\n"
315 "ldr q0, [x28, #0x0]\n"
316 "ldr q1, [x27, #0x0]\n"
317 "ldr q11, [x12, #0x30]\n"
318 "prfm pldl1keep, [x28, #0x80]\n"
319 "prfm pldl1keep, [x27, #0x80]\n"
322 "fmla v24.4s, v8.4s, v0.s[0]\n"
323 "fmla v25.4s, v8.4s, v1.s[0]\n"
324 "add x28, x28, #0x10\n"
325 "add x27, x27, #0x10\n"
326 "fmla v24.4s, v9.4s, v0.s[1]\n"
327 "fmla v25.4s, v9.4s, v1.s[1]\n"
329 "prfm pldl1keep, [x28, #0x80]\n"
330 "fmla v24.4s, v10.4s, v0.s[2]\n"
331 "fmla v25.4s, v10.4s, v1.s[2]\n"
332 "prfm pldl1keep, [x27, #0x80]\n"
333 "add x12, x12, #0x40\n"
334 "fmla v24.4s, v11.4s, v0.s[3]\n"
335 "fmla v25.4s, v11.4s, v1.s[3]\n"
339 "ldr s18, [x28], #0x4\n"
340 "ldr s17, [x27], #0x4\n"
342 "ldr q16, [x12, #0x0]\n"
343 "fmla v24.4s, v16.4s, v18.s[0]\n"
344 "fmla v25.4s, v16.4s, v17.s[0]\n"
345 "add x12, x12, #0x10\n"
348 "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
349 "add x10, x10, #0x1\n"
352 "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
353 "add x27, x11, x20, LSL #2\n"
354 "prfm pstl1keep, [x11, #0x0]\n"
355 "prfm pstl1keep, [x27, #0x0]\n"
356 "tbz %x[flags], #1, 38f\n"
357 "add x20, %x[args_ptr], %[offset_max]\n"
358 "ld1r { v17.4s }, [x20]\n"
359 "add x20, %x[args_ptr], %[offset_min]\n"
360 "ld1r { v16.4s }, [x20]\n"
361 "fmin v24.4s, v24.4s, v17.4s\n"
362 "fmin v25.4s, v25.4s, v17.4s\n"
363 "fmax v24.4s, v24.4s, v16.4s\n"
364 "fmax v25.4s, v25.4s, v16.4s\n"
369 "str d24, [x11], #0x8\n"
370 "str d25, [x27], #0x8\n"
372 "st1 { v24.s }[2], [x11]\n"
373 "st1 { v25.s }[2], [x27]\n"
376 "str s24, [x11, #0x0]\n"
377 "str s25, [x27, #0x0]\n"
381 "str q24, [x11, #0x0]\n"
382 "add x11, x11, #0x10\n"
383 "str q25, [x27, #0x0]\n"
385 "subs x13, x13, #0x4\n"
389 "mov x14, %x[bias]\n"
390 "ldr x13, [%x[args_ptr], %[offsetof_N]]\n"
391 "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
392 "mov x11, %x[output_ptr]\n"
395 "ldr q24, [x14, #0x0]\n"
396 "mov v25.16b, v24.16b\n"
397 "mov v26.16b, v24.16b\n"
398 "add x14, x14, #0x10\n"
401 "tbz %x[flags], #0, 49f\n"
402 "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
403 "add x27, x11, x20, LSL #2\n"
405 "add x26, x27, x20, LSL #2\n"
408 "ldr d24, [x11], #0x8\n"
409 "ldr d25, [x27], #0x8\n"
411 "ldr d26, [x26], #0x8\n"
413 "ld1 { v24.s }[2], [x11]\n"
414 "ld1 { v25.s }[2], [x27]\n"
415 "ld1 { v26.s }[2], [x26]\n"
418 "ldr s24, [x11, #0x0]\n"
419 "ldr s25, [x27, #0x0]\n"
421 "ldr s26, [x26, #0x0]\n"
423 "sub x11, x11, x20\n"
426 "ldr q24, [x11, #0x0]\n"
427 "ldr q25, [x27, #0x0]\n"
428 "ldr q26, [x26, #0x0]\n"
431 "movi v24.16b, #0x0\n"
432 "movi v25.16b, #0x0\n"
433 "movi v26.16b, #0x0\n"
437 "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
438 "ldr w9, [x20, x10, LSL #0x2]\n"
439 "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
440 "tbz %x[flags], #3, 52f\n"
441 "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n"
442 "add x20, x20, x21, LSL #3\n"
443 "ldr x28, [x20, #0x0]\n"
444 "ldr x27, [x20, #0x8]\n"
445 "ldr x26, [x20, #0x10]\n"
447 "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
448 "add x28, x28, x20, LSL #2\n"
449 "add x27, x27, x20, LSL #2\n"
450 "add x26, x26, x20, LSL #2\n"
453 "mov x28, %x[input_ptr]\n"
454 "add x27, x28, x21, LSL #2\n"
455 "add x26, x27, x21, LSL #2\n"
459 "ldr q0, [x28, #0x0]\n"
460 "ldr q1, [x27, #0x0]\n"
462 "ldr q2, [x26, #0x0]\n"
463 "ldr q8, [x12, #0x0]\n"
464 "ldr q9, [x12, #0x10]\n"
465 "ldr q10, [x12, #0x20]\n"
466 "ldr q11, [x12, #0x30]\n"
469 "fmla v24.4s, v8.4s, v0.s[0]\n"
470 "fmla v25.4s, v8.4s, v1.s[0]\n"
472 "add x28, x28, #0x10\n"
473 "fmla v26.4s, v8.4s, v2.s[0]\n"
474 "fmla v24.4s, v9.4s, v0.s[1]\n"
475 "add x27, x27, #0x10\n"
476 "add x26, x26, #0x10\n"
477 "fmla v25.4s, v9.4s, v1.s[1]\n"
478 "fmla v26.4s, v9.4s, v2.s[1]\n"
480 "add x12, x12, #0x40\n"
481 "ldr q8, [x12, #0x0]\n"
482 "ldr q9, [x12, #0x10]\n"
483 "fmla v24.4s, v10.4s, v0.s[2]\n"
484 "fmla v25.4s, v10.4s, v1.s[2]\n"
485 "fmla v26.4s, v10.4s, v2.s[2]\n"
486 "ldr q10, [x12, #0x20]\n"
487 "fmla v24.4s, v11.4s, v0.s[3]\n"
488 "ldr q0, [x28, #0x0]\n"
489 "prfm pldl1keep, [x28, #0x80]\n"
490 "fmla v25.4s, v11.4s, v1.s[3]\n"
491 "ldr q1, [x27, #0x0]\n"
492 "fmla v26.4s, v11.4s, v2.s[3]\n"
493 "ldr q2, [x26, #0x0]\n"
494 "ldr q11, [x12, #0x30]\n"
495 "prfm pldl1keep, [x27, #0x80]\n"
496 "prfm pldl1keep, [x26, #0x80]\n"
499 "fmla v24.4s, v8.4s, v0.s[0]\n"
500 "fmla v25.4s, v8.4s, v1.s[0]\n"
501 "add x28, x28, #0x10\n"
502 "add x27, x27, #0x10\n"
503 "fmla v26.4s, v8.4s, v2.s[0]\n"
504 "fmla v24.4s, v9.4s, v0.s[1]\n"
505 "add x26, x26, #0x10\n"
507 "fmla v25.4s, v9.4s, v1.s[1]\n"
508 "fmla v26.4s, v9.4s, v2.s[1]\n"
509 "prfm pldl1keep, [x28, #0x80]\n"
510 "prfm pldl1keep, [x27, #0x80]\n"
511 "fmla v24.4s, v10.4s, v0.s[2]\n"
512 "fmla v25.4s, v10.4s, v1.s[2]\n"
513 "prfm pldl1keep, [x26, #0x80]\n"
514 "add x12, x12, #0x40\n"
515 "fmla v26.4s, v10.4s, v2.s[2]\n"
516 "fmla v24.4s, v11.4s, v0.s[3]\n"
517 "fmla v25.4s, v11.4s, v1.s[3]\n"
518 "fmla v26.4s, v11.4s, v2.s[3]\n"
522 "ldr s19, [x28], #0x4\n"
523 "ldr s18, [x27], #0x4\n"
525 "ldr s17, [x26], #0x4\n"
526 "ldr q16, [x12, #0x0]\n"
527 "fmla v24.4s, v16.4s, v19.s[0]\n"
528 "fmla v25.4s, v16.4s, v18.s[0]\n"
529 "fmla v26.4s, v16.4s, v17.s[0]\n"
530 "add x12, x12, #0x10\n"
533 "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
534 "add x10, x10, #0x1\n"
537 "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
538 "add x27, x11, x20, LSL #2\n"
539 "add x26, x27, x20, LSL #2\n"
540 "prfm pstl1keep, [x11, #0x0]\n"
541 "prfm pstl1keep, [x27, #0x0]\n"
542 "prfm pstl1keep, [x26, #0x0]\n"
543 "tbz %x[flags], #1, 59f\n"
544 "add x20, %x[args_ptr], %[offset_max]\n"
545 "ld1r { v17.4s }, [x20]\n"
546 "add x20, %x[args_ptr], %[offset_min]\n"
547 "ld1r { v16.4s }, [x20]\n"
548 "fmin v24.4s, v24.4s, v17.4s\n"
549 "fmin v25.4s, v25.4s, v17.4s\n"
550 "fmin v26.4s, v26.4s, v17.4s\n"
551 "fmax v24.4s, v24.4s, v16.4s\n"
552 "fmax v25.4s, v25.4s, v16.4s\n"
553 "fmax v26.4s, v26.4s, v16.4s\n"
558 "str d24, [x11], #0x8\n"
559 "str d25, [x27], #0x8\n"
560 "str d26, [x26], #0x8\n"
562 "st1 { v24.s }[2], [x11]\n"
563 "st1 { v25.s }[2], [x27]\n"
564 "st1 { v26.s }[2], [x26]\n"
567 "str s24, [x11, #0x0]\n"
568 "str s25, [x27, #0x0]\n"
569 "str s26, [x26, #0x0]\n"
573 "str q24, [x11, #0x0]\n"
574 "add x11, x11, #0x10\n"
575 "str q25, [x27, #0x0]\n"
576 "str q26, [x26, #0x0]\n"
578 "subs x13, x13, #0x4\n"
582 "mov x14, %x[bias]\n"
583 "ldr x13, [%x[args_ptr], %[offsetof_N]]\n"
584 "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
585 "mov x11, %x[output_ptr]\n"
588 "ldr q24, [x14, #0x0]\n"
589 "mov v25.16b, v24.16b\n"
590 "mov v26.16b, v24.16b\n"
591 "add x14, x14, #0x10\n"
592 "mov v27.16b, v24.16b\n"
595 "tbz %x[flags], #0, 70f\n"
596 "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
597 "add x27, x11, x20, LSL #2\n"
598 "add x26, x27, x20, LSL #2\n"
600 "add x25, x26, x20, LSL #2\n"
603 "ldr d24, [x11], #0x8\n"
604 "ldr d25, [x27], #0x8\n"
606 "ldr d26, [x26], #0x8\n"
607 "ldr d27, [x25], #0x8\n"
609 "ld1 { v24.s }[2], [x11]\n"
610 "ld1 { v25.s }[2], [x27]\n"
611 "ld1 { v26.s }[2], [x26]\n"
612 "ld1 { v27.s }[2], [x25]\n"
615 "ldr s24, [x11, #0x0]\n"
616 "ldr s25, [x27, #0x0]\n"
618 "ldr s26, [x26, #0x0]\n"
619 "ldr s27, [x25, #0x0]\n"
621 "sub x11, x11, x20\n"
624 "ldr q24, [x11, #0x0]\n"
625 "ldr q25, [x27, #0x0]\n"
626 "ldr q26, [x26, #0x0]\n"
627 "ldr q27, [x25, #0x0]\n"
630 "movi v24.16b, #0x0\n"
631 "movi v25.16b, #0x0\n"
632 "movi v26.16b, #0x0\n"
633 "movi v27.16b, #0x0\n"
637 "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
638 "ldr w9, [x20, x10, LSL #0x2]\n"
639 "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
640 "tbz %x[flags], #3, 73f\n"
641 "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n"
642 "add x20, x20, x21, LSL #3\n"
643 "ldr x28, [x20, #0x0]\n"
644 "ldr x27, [x20, #0x8]\n"
645 "ldr x26, [x20, #0x10]\n"
646 "ldr x25, [x20, #0x18]\n"
648 "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
649 "add x28, x28, x20, LSL #2\n"
650 "add x27, x27, x20, LSL #2\n"
651 "add x26, x26, x20, LSL #2\n"
652 "add x25, x25, x20, LSL #2\n"
655 "mov x28, %x[input_ptr]\n"
656 "add x27, x28, x21, LSL #2\n"
657 "add x26, x27, x21, LSL #2\n"
658 "add x25, x26, x21, LSL #2\n"
662 "ldr q0, [x28, #0x0]\n"
663 "ldr q1, [x27, #0x0]\n"
665 "ldr q2, [x26, #0x0]\n"
666 "ldr q3, [x25, #0x0]\n"
667 "ldr q8, [x12, #0x0]\n"
668 "ldr q9, [x12, #0x10]\n"
669 "ldr q10, [x12, #0x20]\n"
670 "ldr q11, [x12, #0x30]\n"
673 "fmla v24.4s, v8.4s, v0.s[0]\n"
674 "fmla v25.4s, v8.4s, v1.s[0]\n"
676 "add x28, x28, #0x10\n"
677 "fmla v26.4s, v8.4s, v2.s[0]\n"
678 "fmla v27.4s, v8.4s, v3.s[0]\n"
679 "add x27, x27, #0x10\n"
680 "add x26, x26, #0x10\n"
681 "fmla v24.4s, v9.4s, v0.s[1]\n"
682 "fmla v25.4s, v9.4s, v1.s[1]\n"
683 "add x25, x25, #0x10\n"
685 "fmla v26.4s, v9.4s, v2.s[1]\n"
686 "fmla v27.4s, v9.4s, v3.s[1]\n"
687 "add x12, x12, #0x40\n"
688 "ldr q8, [x12, #0x0]\n"
689 "ldr q9, [x12, #0x10]\n"
690 "fmla v24.4s, v10.4s, v0.s[2]\n"
691 "fmla v25.4s, v10.4s, v1.s[2]\n"
692 "prfm pldl1keep, [x28, #0x80]\n"
693 "fmla v26.4s, v10.4s, v2.s[2]\n"
694 "fmla v27.4s, v10.4s, v3.s[2]\n"
695 "ldr q10, [x12, #0x20]\n"
696 "prfm pldl1keep, [x27, #0x80]\n"
697 "prfm pldl1keep, [x26, #0x80]\n"
698 "fmla v24.4s, v11.4s, v0.s[3]\n"
699 "ldr q0, [x28, #0x0]\n"
700 "fmla v25.4s, v11.4s, v1.s[3]\n"
701 "ldr q1, [x27, #0x0]\n"
702 "fmla v26.4s, v11.4s, v2.s[3]\n"
703 "ldr q2, [x26, #0x0]\n"
704 "fmla v27.4s, v11.4s, v3.s[3]\n"
705 "ldr q3, [x25, #0x0]\n"
706 "ldr q11, [x12, #0x30]\n"
707 "prfm pldl1keep, [x25, #0x80]\n"
710 "fmla v24.4s, v8.4s, v0.s[0]\n"
711 "fmla v25.4s, v8.4s, v1.s[0]\n"
712 "add x28, x28, #0x10\n"
713 "add x27, x27, #0x10\n"
714 "fmla v26.4s, v8.4s, v2.s[0]\n"
715 "fmla v27.4s, v8.4s, v3.s[0]\n"
716 "add x26, x26, #0x10\n"
717 "add x25, x25, #0x10\n"
718 "fmla v24.4s, v9.4s, v0.s[1]\n"
719 "fmla v25.4s, v9.4s, v1.s[1]\n"
721 "prfm pldl1keep, [x28, #0x80]\n"
722 "fmla v26.4s, v9.4s, v2.s[1]\n"
723 "fmla v27.4s, v9.4s, v3.s[1]\n"
724 "prfm pldl1keep, [x27, #0x80]\n"
725 "prfm pldl1keep, [x26, #0x80]\n"
726 "fmla v24.4s, v10.4s, v0.s[2]\n"
727 "fmla v25.4s, v10.4s, v1.s[2]\n"
728 "prfm pldl1keep, [x25, #0x80]\n"
729 "add x12, x12, #0x40\n"
730 "fmla v26.4s, v10.4s, v2.s[2]\n"
731 "fmla v27.4s, v10.4s, v3.s[2]\n"
732 "fmla v24.4s, v11.4s, v0.s[3]\n"
733 "fmla v25.4s, v11.4s, v1.s[3]\n"
734 "fmla v26.4s, v11.4s, v2.s[3]\n"
735 "fmla v27.4s, v11.4s, v3.s[3]\n"
739 "ldr s20, [x28], #0x4\n"
740 "ldr s19, [x27], #0x4\n"
742 "ldr s18, [x26], #0x4\n"
743 "ldr s17, [x25], #0x4\n"
744 "ldr q16, [x12, #0x0]\n"
745 "fmla v24.4s, v16.4s, v20.s[0]\n"
746 "fmla v25.4s, v16.4s, v19.s[0]\n"
747 "add x12, x12, #0x10\n"
748 "fmla v26.4s, v16.4s, v18.s[0]\n"
749 "fmla v27.4s, v16.4s, v17.s[0]\n"
752 "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
753 "add x10, x10, #0x1\n"
756 "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
757 "add x27, x11, x20, LSL #2\n"
758 "add x26, x27, x20, LSL #2\n"
759 "prfm pstl1keep, [x11, #0x0]\n"
760 "add x25, x26, x20, LSL #2\n"
761 "prfm pstl1keep, [x27, #0x0]\n"
762 "prfm pstl1keep, [x26, #0x0]\n"
763 "prfm pstl1keep, [x25, #0x0]\n"
764 "tbz %x[flags], #1, 80f\n"
765 "add x20, %x[args_ptr], %[offset_max]\n"
766 "ld1r { v17.4s }, [x20]\n"
767 "add x20, %x[args_ptr], %[offset_min]\n"
768 "ld1r { v16.4s }, [x20]\n"
769 "fmin v24.4s, v24.4s, v17.4s\n"
770 "fmin v25.4s, v25.4s, v17.4s\n"
771 "fmin v26.4s, v26.4s, v17.4s\n"
772 "fmin v27.4s, v27.4s, v17.4s\n"
773 "fmax v24.4s, v24.4s, v16.4s\n"
774 "fmax v25.4s, v25.4s, v16.4s\n"
775 "fmax v26.4s, v26.4s, v16.4s\n"
776 "fmax v27.4s, v27.4s, v16.4s\n"
781 "str d24, [x11], #0x8\n"
782 "str d25, [x27], #0x8\n"
783 "str d26, [x26], #0x8\n"
784 "str d27, [x25], #0x8\n"
786 "st1 { v24.s }[2], [x11]\n"
787 "st1 { v25.s }[2], [x27]\n"
788 "st1 { v26.s }[2], [x26]\n"
789 "st1 { v27.s }[2], [x25]\n"
792 "str s24, [x11, #0x0]\n"
793 "str s25, [x27, #0x0]\n"
794 "str s26, [x26, #0x0]\n"
795 "str s27, [x25, #0x0]\n"
799 "str q24, [x11, #0x0]\n"
800 "add x11, x11, #0x10\n"
801 "str q25, [x27, #0x0]\n"
802 "str q26, [x26, #0x0]\n"
803 "str q27, [x25, #0x0]\n"
805 "subs x13, x13, #0x4\n"
809 "mov x14, %x[bias]\n"
810 "ldr x13, [%x[args_ptr], %[offsetof_N]]\n"
811 "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
812 "mov x11, %x[output_ptr]\n"
815 "ldr q24, [x14, #0x0]\n"
816 "mov v25.16b, v24.16b\n"
817 "mov v26.16b, v24.16b\n"
818 "add x14, x14, #0x10\n"
819 "mov v27.16b, v24.16b\n"
820 "mov v28.16b, v24.16b\n"
823 "tbz %x[flags], #0, 91f\n"
824 "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
825 "add x27, x11, x20, LSL #2\n"
826 "add x26, x27, x20, LSL #2\n"
827 "add x25, x26, x20, LSL #2\n"
829 "add x24, x25, x20, LSL #2\n"
832 "ldr d24, [x11], #0x8\n"
833 "ldr d25, [x27], #0x8\n"
835 "ldr d26, [x26], #0x8\n"
836 "ldr d27, [x25], #0x8\n"
837 "ldr d28, [x24], #0x8\n"
839 "ld1 { v24.s }[2], [x11]\n"
840 "ld1 { v25.s }[2], [x27]\n"
841 "ld1 { v26.s }[2], [x26]\n"
842 "ld1 { v27.s }[2], [x25]\n"
843 "ld1 { v28.s }[2], [x24]\n"
846 "ldr s24, [x11, #0x0]\n"
847 "ldr s25, [x27, #0x0]\n"
849 "ldr s26, [x26, #0x0]\n"
850 "ldr s27, [x25, #0x0]\n"
851 "ldr s28, [x24, #0x0]\n"
853 "sub x11, x11, x20\n"
856 "ldr q24, [x11, #0x0]\n"
857 "ldr q25, [x27, #0x0]\n"
858 "ldr q26, [x26, #0x0]\n"
859 "ldr q27, [x25, #0x0]\n"
860 "ldr q28, [x24, #0x0]\n"
863 "movi v24.16b, #0x0\n"
864 "movi v25.16b, #0x0\n"
865 "movi v26.16b, #0x0\n"
866 "movi v27.16b, #0x0\n"
867 "movi v28.16b, #0x0\n"
871 "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
872 "ldr w9, [x20, x10, LSL #0x2]\n"
873 "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
874 "tbz %x[flags], #3, 94f\n"
875 "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n"
876 "add x20, x20, x21, LSL #3\n"
877 "ldr x28, [x20, #0x0]\n"
878 "ldr x27, [x20, #0x8]\n"
879 "ldr x26, [x20, #0x10]\n"
880 "ldr x25, [x20, #0x18]\n"
881 "ldr x24, [x20, #0x20]\n"
883 "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
884 "add x28, x28, x20, LSL #2\n"
885 "add x27, x27, x20, LSL #2\n"
886 "add x26, x26, x20, LSL #2\n"
887 "add x25, x25, x20, LSL #2\n"
888 "add x24, x24, x20, LSL #2\n"
891 "mov x28, %x[input_ptr]\n"
892 "add x27, x28, x21, LSL #2\n"
893 "add x26, x27, x21, LSL #2\n"
894 "add x25, x26, x21, LSL #2\n"
895 "add x24, x25, x21, LSL #2\n"
899 "ldr q0, [x28, #0x0]\n"
900 "ldr q1, [x27, #0x0]\n"
902 "ldr q2, [x26, #0x0]\n"
903 "ldr q3, [x25, #0x0]\n"
904 "ldr q4, [x24, #0x0]\n"
905 "ldr q8, [x12, #0x0]\n"
906 "ldr q9, [x12, #0x10]\n"
907 "ldr q10, [x12, #0x20]\n"
908 "ldr q11, [x12, #0x30]\n"
911 "fmla v24.4s, v8.4s, v0.s[0]\n"
912 "fmla v25.4s, v8.4s, v1.s[0]\n"
914 "add x28, x28, #0x10\n"
915 "fmla v26.4s, v8.4s, v2.s[0]\n"
916 "fmla v27.4s, v8.4s, v3.s[0]\n"
917 "add x27, x27, #0x10\n"
918 "add x26, x26, #0x10\n"
919 "fmla v28.4s, v8.4s, v4.s[0]\n"
920 "fmla v24.4s, v9.4s, v0.s[1]\n"
921 "add x25, x25, #0x10\n"
922 "add x24, x24, #0x10\n"
923 "fmla v25.4s, v9.4s, v1.s[1]\n"
924 "fmla v26.4s, v9.4s, v2.s[1]\n"
926 "add x12, x12, #0x40\n"
927 "ldr q8, [x12, #0x0]\n"
928 "fmla v27.4s, v9.4s, v3.s[1]\n"
929 "fmla v28.4s, v9.4s, v4.s[1]\n"
930 "ldr q9, [x12, #0x10]\n"
931 "fmla v24.4s, v10.4s, v0.s[2]\n"
932 "fmla v25.4s, v10.4s, v1.s[2]\n"
933 "prfm pldl1keep, [x28, #0x80]\n"
934 "prfm pldl1keep, [x27, #0x80]\n"
935 "fmla v26.4s, v10.4s, v2.s[2]\n"
936 "fmla v27.4s, v10.4s, v3.s[2]\n"
937 "prfm pldl1keep, [x26, #0x80]\n"
938 "prfm pldl1keep, [x25, #0x80]\n"
939 "fmla v28.4s, v10.4s, v4.s[2]\n"
940 "ldr q10, [x12, #0x20]\n"
941 "fmla v24.4s, v11.4s, v0.s[3]\n"
942 "ldr q0, [x28, #0x0]\n"
943 "prfm pldl1keep, [x24, #0x80]\n"
944 "fmla v25.4s, v11.4s, v1.s[3]\n"
945 "ldr q1, [x27, #0x0]\n"
946 "fmla v26.4s, v11.4s, v2.s[3]\n"
947 "ldr q2, [x26, #0x0]\n"
948 "fmla v27.4s, v11.4s, v3.s[3]\n"
949 "ldr q3, [x25, #0x0]\n"
950 "fmla v28.4s, v11.4s, v4.s[3]\n"
951 "ldr q4, [x24, #0x0]\n"
952 "ldr q11, [x12, #0x30]\n"
955 "fmla v24.4s, v8.4s, v0.s[0]\n"
956 "fmla v25.4s, v8.4s, v1.s[0]\n"
957 "add x28, x28, #0x10\n"
958 "add x27, x27, #0x10\n"
959 "fmla v26.4s, v8.4s, v2.s[0]\n"
960 "fmla v27.4s, v8.4s, v3.s[0]\n"
961 "add x26, x26, #0x10\n"
962 "add x25, x25, #0x10\n"
963 "fmla v28.4s, v8.4s, v4.s[0]\n"
964 "fmla v24.4s, v9.4s, v0.s[1]\n"
965 "add x24, x24, #0x10\n"
967 "fmla v25.4s, v9.4s, v1.s[1]\n"
968 "fmla v26.4s, v9.4s, v2.s[1]\n"
969 "prfm pldl1keep, [x28, #0x80]\n"
970 "prfm pldl1keep, [x27, #0x80]\n"
971 "fmla v27.4s, v9.4s, v3.s[1]\n"
972 "fmla v28.4s, v9.4s, v4.s[1]\n"
973 "prfm pldl1keep, [x26, #0x80]\n"
974 "prfm pldl1keep, [x25, #0x80]\n"
975 "fmla v24.4s, v10.4s, v0.s[2]\n"
976 "fmla v25.4s, v10.4s, v1.s[2]\n"
977 "prfm pldl1keep, [x24, #0x80]\n"
978 "add x12, x12, #0x40\n"
979 "fmla v26.4s, v10.4s, v2.s[2]\n"
980 "fmla v27.4s, v10.4s, v3.s[2]\n"
981 "fmla v28.4s, v10.4s, v4.s[2]\n"
982 "fmla v24.4s, v11.4s, v0.s[3]\n"
983 "fmla v25.4s, v11.4s, v1.s[3]\n"
984 "fmla v26.4s, v11.4s, v2.s[3]\n"
985 "fmla v27.4s, v11.4s, v3.s[3]\n"
986 "fmla v28.4s, v11.4s, v4.s[3]\n"
990 "ldr s21, [x28], #0x4\n"
991 "ldr s20, [x27], #0x4\n"
993 "ldr s19, [x26], #0x4\n"
994 "ldr s18, [x25], #0x4\n"
995 "ldr s17, [x24], #0x4\n"
996 "ldr q16, [x12, #0x0]\n"
997 "fmla v24.4s, v16.4s, v21.s[0]\n"
998 "fmla v25.4s, v16.4s, v20.s[0]\n"
999 "fmla v26.4s, v16.4s, v19.s[0]\n"
1000 "fmla v27.4s, v16.4s, v18.s[0]\n"
1001 "add x12, x12, #0x10\n"
1002 "fmla v28.4s, v16.4s, v17.s[0]\n"
1005 "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
1006 "add x10, x10, #0x1\n"
1009 "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
1010 "add x27, x11, x20, LSL #2\n"
1011 "add x26, x27, x20, LSL #2\n"
1012 "prfm pstl1keep, [x11, #0x0]\n"
1013 "add x25, x26, x20, LSL #2\n"
1014 "add x24, x25, x20, LSL #2\n"
1015 "prfm pstl1keep, [x27, #0x0]\n"
1016 "prfm pstl1keep, [x26, #0x0]\n"
1017 "prfm pstl1keep, [x25, #0x0]\n"
1018 "prfm pstl1keep, [x24, #0x0]\n"
1019 "tbz %x[flags], #1, 101f\n"
1020 "add x20, %x[args_ptr], %[offset_max]\n"
1021 "ld1r { v17.4s }, [x20]\n"
1022 "add x20, %x[args_ptr], %[offset_min]\n"
1023 "ld1r { v16.4s }, [x20]\n"
1024 "fmin v24.4s, v24.4s, v17.4s\n"
1025 "fmin v25.4s, v25.4s, v17.4s\n"
1026 "fmin v26.4s, v26.4s, v17.4s\n"
1027 "fmin v27.4s, v27.4s, v17.4s\n"
1028 "fmin v28.4s, v28.4s, v17.4s\n"
1029 "fmax v24.4s, v24.4s, v16.4s\n"
1030 "fmax v25.4s, v25.4s, v16.4s\n"
1031 "fmax v26.4s, v26.4s, v16.4s\n"
1032 "fmax v27.4s, v27.4s, v16.4s\n"
1033 "fmax v28.4s, v28.4s, v16.4s\n"
1037 "tbz x13, #1, 102f\n"
1038 "str d24, [x11], #0x8\n"
1039 "str d25, [x27], #0x8\n"
1040 "str d26, [x26], #0x8\n"
1041 "str d27, [x25], #0x8\n"
1042 "str d28, [x24], #0x8\n"
1043 "tbz x13, #0, 103f\n"
1044 "st1 { v24.s }[2], [x11]\n"
1045 "st1 { v25.s }[2], [x27]\n"
1046 "st1 { v26.s }[2], [x26]\n"
1047 "st1 { v27.s }[2], [x25]\n"
1048 "st1 { v28.s }[2], [x24]\n"
1051 "str s24, [x11, #0x0]\n"
1052 "str s25, [x27, #0x0]\n"
1053 "str s26, [x26, #0x0]\n"
1054 "str s27, [x25, #0x0]\n"
1055 "str s28, [x24, #0x0]\n"
1059 "str q24, [x11, #0x0]\n"
1060 "add x11, x11, #0x10\n"
1061 "str q25, [x27, #0x0]\n"
1062 "str q26, [x26, #0x0]\n"
1063 "str q27, [x25, #0x0]\n"
1064 "str q28, [x24, #0x0]\n"
1066 "subs x13, x13, #0x4\n"
1070 "mov x14, %x[bias]\n"
1071 "ldr x13, [%x[args_ptr], %[offsetof_N]]\n"
1072 "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
1073 "mov x11, %x[output_ptr]\n"
1076 "ldr q24, [x14, #0x0]\n"
1077 "mov v25.16b, v24.16b\n"
1078 "mov v26.16b, v24.16b\n"
1079 "add x14, x14, #0x10\n"
1080 "mov v27.16b, v24.16b\n"
1081 "mov v28.16b, v24.16b\n"
1082 "mov v29.16b, v24.16b\n"
1085 "tbz %x[flags], #0, 112f\n"
1086 "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
1087 "add x27, x11, x20, LSL #2\n"
1088 "add x26, x27, x20, LSL #2\n"
1089 "add x25, x26, x20, LSL #2\n"
1090 "add x24, x25, x20, LSL #2\n"
1092 "add x23, x24, x20, LSL #2\n"
1094 "tbz x13, #1, 109f\n"
1095 "ldr d24, [x11], #0x8\n"
1096 "ldr d25, [x27], #0x8\n"
1098 "ldr d26, [x26], #0x8\n"
1099 "ldr d27, [x25], #0x8\n"
1100 "ldr d28, [x24], #0x8\n"
1101 "ldr d29, [x23], #0x8\n"
1102 "tbz x13, #0, 110f\n"
1103 "ld1 { v24.s }[2], [x11]\n"
1104 "ld1 { v25.s }[2], [x27]\n"
1105 "ld1 { v26.s }[2], [x26]\n"
1106 "ld1 { v27.s }[2], [x25]\n"
1107 "ld1 { v28.s }[2], [x24]\n"
1108 "ld1 { v29.s }[2], [x23]\n"
1111 "ldr s24, [x11, #0x0]\n"
1112 "ldr s25, [x27, #0x0]\n"
1114 "ldr s26, [x26, #0x0]\n"
1115 "ldr s27, [x25, #0x0]\n"
1116 "ldr s28, [x24, #0x0]\n"
1117 "ldr s29, [x23, #0x0]\n"
1119 "sub x11, x11, x20\n"
1122 "ldr q24, [x11, #0x0]\n"
1123 "ldr q25, [x27, #0x0]\n"
1124 "ldr q26, [x26, #0x0]\n"
1125 "ldr q27, [x25, #0x0]\n"
1126 "ldr q28, [x24, #0x0]\n"
1127 "ldr q29, [x23, #0x0]\n"
1130 "movi v24.16b, #0x0\n"
1131 "movi v25.16b, #0x0\n"
1132 "movi v26.16b, #0x0\n"
1133 "movi v27.16b, #0x0\n"
1134 "movi v28.16b, #0x0\n"
1135 "movi v29.16b, #0x0\n"
1139 "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
1140 "ldr w9, [x20, x10, LSL #0x2]\n"
1141 "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
1142 "tbz %x[flags], #3, 115f\n"
1143 "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n"
1144 "add x20, x20, x21, LSL #3\n"
1145 "ldr x28, [x20, #0x0]\n"
1146 "ldr x27, [x20, #0x8]\n"
1147 "ldr x26, [x20, #0x10]\n"
1148 "ldr x25, [x20, #0x18]\n"
1149 "ldr x24, [x20, #0x20]\n"
1150 "ldr x23, [x20, #0x28]\n"
1152 "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
1153 "add x28, x28, x20, LSL #2\n"
1154 "add x27, x27, x20, LSL #2\n"
1155 "add x26, x26, x20, LSL #2\n"
1156 "add x25, x25, x20, LSL #2\n"
1157 "add x24, x24, x20, LSL #2\n"
1158 "add x23, x23, x20, LSL #2\n"
1161 "mov x28, %x[input_ptr]\n"
1162 "add x27, x28, x21, LSL #2\n"
1163 "add x26, x27, x21, LSL #2\n"
1164 "add x25, x26, x21, LSL #2\n"
1165 "add x24, x25, x21, LSL #2\n"
1166 "add x23, x24, x21, LSL #2\n"
1170 "ldr q0, [x28, #0x0]\n"
1171 "ldr q1, [x27, #0x0]\n"
1173 "ldr q2, [x26, #0x0]\n"
1174 "ldr q3, [x25, #0x0]\n"
1175 "ldr q4, [x24, #0x0]\n"
1176 "ldr q5, [x23, #0x0]\n"
1177 "ldr q8, [x12, #0x0]\n"
1178 "ldr q9, [x12, #0x10]\n"
1179 "ldr q10, [x12, #0x20]\n"
1180 "ldr q11, [x12, #0x30]\n"
1183 "fmla v24.4s, v8.4s, v0.s[0]\n"
1184 "fmla v25.4s, v8.4s, v1.s[0]\n"
1185 "sub x9, x9, #0x4\n"
1186 "add x28, x28, #0x10\n"
1187 "fmla v26.4s, v8.4s, v2.s[0]\n"
1188 "fmla v27.4s, v8.4s, v3.s[0]\n"
1189 "add x27, x27, #0x10\n"
1190 "add x26, x26, #0x10\n"
1191 "fmla v28.4s, v8.4s, v4.s[0]\n"
1192 "fmla v29.4s, v8.4s, v5.s[0]\n"
1193 "add x25, x25, #0x10\n"
1194 "add x24, x24, #0x10\n"
1195 "fmla v24.4s, v9.4s, v0.s[1]\n"
1196 "fmla v25.4s, v9.4s, v1.s[1]\n"
1197 "add x23, x23, #0x10\n"
1199 "fmla v26.4s, v9.4s, v2.s[1]\n"
1200 "fmla v27.4s, v9.4s, v3.s[1]\n"
1201 "add x12, x12, #0x40\n"
1202 "ldr q8, [x12, #0x0]\n"
1203 "fmla v28.4s, v9.4s, v4.s[1]\n"
1204 "fmla v29.4s, v9.4s, v5.s[1]\n"
1205 "ldr q9, [x12, #0x10]\n"
1206 "prfm pldl1keep, [x28, #0x80]\n"
1207 "fmla v24.4s, v10.4s, v0.s[2]\n"
1208 "fmla v25.4s, v10.4s, v1.s[2]\n"
1209 "prfm pldl1keep, [x27, #0x80]\n"
1210 "prfm pldl1keep, [x26, #0x80]\n"
1211 "fmla v26.4s, v10.4s, v2.s[2]\n"
1212 "fmla v27.4s, v10.4s, v3.s[2]\n"
1213 "prfm pldl1keep, [x25, #0x80]\n"
1214 "prfm pldl1keep, [x24, #0x80]\n"
1215 "fmla v28.4s, v10.4s, v4.s[2]\n"
1216 "fmla v29.4s, v10.4s, v5.s[2]\n"
1217 "ldr q10, [x12, #0x20]\n"
1218 "prfm pldl1keep, [x23, #0x80]\n"
1219 "fmla v24.4s, v11.4s, v0.s[3]\n"
1220 "ldr q0, [x28, #0x0]\n"
1221 "fmla v25.4s, v11.4s, v1.s[3]\n"
1222 "ldr q1, [x27, #0x0]\n"
1223 "fmla v26.4s, v11.4s, v2.s[3]\n"
1224 "ldr q2, [x26, #0x0]\n"
1225 "fmla v27.4s, v11.4s, v3.s[3]\n"
1226 "ldr q3, [x25, #0x0]\n"
1227 "fmla v28.4s, v11.4s, v4.s[3]\n"
1228 "ldr q4, [x24, #0x0]\n"
1229 "fmla v29.4s, v11.4s, v5.s[3]\n"
1230 "ldr q5, [x23, #0x0]\n"
1231 "ldr q11, [x12, #0x30]\n"
1234 "fmla v24.4s, v8.4s, v0.s[0]\n"
1235 "fmla v25.4s, v8.4s, v1.s[0]\n"
1236 "add x28, x28, #0x10\n"
1237 "add x27, x27, #0x10\n"
1238 "fmla v26.4s, v8.4s, v2.s[0]\n"
1239 "fmla v27.4s, v8.4s, v3.s[0]\n"
1240 "add x26, x26, #0x10\n"
1241 "add x25, x25, #0x10\n"
1242 "fmla v28.4s, v8.4s, v4.s[0]\n"
1243 "fmla v29.4s, v8.4s, v5.s[0]\n"
1244 "add x24, x24, #0x10\n"
1245 "add x23, x23, #0x10\n"
1246 "fmla v24.4s, v9.4s, v0.s[1]\n"
1247 "fmla v25.4s, v9.4s, v1.s[1]\n"
1248 "sub x9, x9, #0x4\n"
1249 "prfm pldl1keep, [x28, #0x80]\n"
1250 "fmla v26.4s, v9.4s, v2.s[1]\n"
1251 "fmla v27.4s, v9.4s, v3.s[1]\n"
1252 "prfm pldl1keep, [x27, #0x80]\n"
1253 "prfm pldl1keep, [x26, #0x80]\n"
1254 "fmla v28.4s, v9.4s, v4.s[1]\n"
1255 "fmla v29.4s, v9.4s, v5.s[1]\n"
1256 "prfm pldl1keep, [x25, #0x80]\n"
1257 "prfm pldl1keep, [x24, #0x80]\n"
1258 "fmla v24.4s, v10.4s, v0.s[2]\n"
1259 "fmla v25.4s, v10.4s, v1.s[2]\n"
1260 "prfm pldl1keep, [x23, #0x80]\n"
1261 "add x12, x12, #0x40\n"
1262 "fmla v26.4s, v10.4s, v2.s[2]\n"
1263 "fmla v27.4s, v10.4s, v3.s[2]\n"
1264 "fmla v28.4s, v10.4s, v4.s[2]\n"
1265 "fmla v29.4s, v10.4s, v5.s[2]\n"
1266 "fmla v24.4s, v11.4s, v0.s[3]\n"
1267 "fmla v25.4s, v11.4s, v1.s[3]\n"
1268 "fmla v26.4s, v11.4s, v2.s[3]\n"
1269 "fmla v27.4s, v11.4s, v3.s[3]\n"
1270 "fmla v28.4s, v11.4s, v4.s[3]\n"
1271 "fmla v29.4s, v11.4s, v5.s[3]\n"
1275 "ldr s22, [x28], #0x4\n"
1276 "ldr s21, [x27], #0x4\n"
1277 "sub x9, x9, #0x1\n"
1278 "ldr s20, [x26], #0x4\n"
1279 "ldr s19, [x25], #0x4\n"
1280 "ldr s18, [x24], #0x4\n"
1281 "ldr s17, [x23], #0x4\n"
1282 "ldr q16, [x12, #0x0]\n"
1283 "fmla v24.4s, v16.4s, v22.s[0]\n"
1284 "fmla v25.4s, v16.4s, v21.s[0]\n"
1285 "add x12, x12, #0x10\n"
1286 "fmla v26.4s, v16.4s, v20.s[0]\n"
1287 "fmla v27.4s, v16.4s, v19.s[0]\n"
1288 "fmla v28.4s, v16.4s, v18.s[0]\n"
1289 "fmla v29.4s, v16.4s, v17.s[0]\n"
1292 "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
1293 "add x10, x10, #0x1\n"
1296 "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
1297 "add x27, x11, x20, LSL #2\n"
1298 "add x26, x27, x20, LSL #2\n"
1299 "prfm pstl1keep, [x11, #0x0]\n"
1300 "add x25, x26, x20, LSL #2\n"
1301 "add x24, x25, x20, LSL #2\n"
1302 "prfm pstl1keep, [x27, #0x0]\n"
1303 "prfm pstl1keep, [x26, #0x0]\n"
1304 "add x23, x24, x20, LSL #2\n"
1305 "prfm pstl1keep, [x25, #0x0]\n"
1306 "prfm pstl1keep, [x24, #0x0]\n"
1307 "prfm pstl1keep, [x23, #0x0]\n"
1308 "tbz %x[flags], #1, 122f\n"
1309 "add x20, %x[args_ptr], %[offset_max]\n"
1310 "ld1r { v17.4s }, [x20]\n"
1311 "add x20, %x[args_ptr], %[offset_min]\n"
1312 "ld1r { v16.4s }, [x20]\n"
1313 "fmin v24.4s, v24.4s, v17.4s\n"
1314 "fmin v25.4s, v25.4s, v17.4s\n"
1315 "fmin v26.4s, v26.4s, v17.4s\n"
1316 "fmin v27.4s, v27.4s, v17.4s\n"
1317 "fmin v28.4s, v28.4s, v17.4s\n"
1318 "fmin v29.4s, v29.4s, v17.4s\n"
1319 "fmax v24.4s, v24.4s, v16.4s\n"
1320 "fmax v25.4s, v25.4s, v16.4s\n"
1321 "fmax v26.4s, v26.4s, v16.4s\n"
1322 "fmax v27.4s, v27.4s, v16.4s\n"
1323 "fmax v28.4s, v28.4s, v16.4s\n"
1324 "fmax v29.4s, v29.4s, v16.4s\n"
1328 "tbz x13, #1, 123f\n"
1329 "str d24, [x11], #0x8\n"
1330 "str d25, [x27], #0x8\n"
1331 "str d26, [x26], #0x8\n"
1332 "str d27, [x25], #0x8\n"
1333 "str d28, [x24], #0x8\n"
1334 "str d29, [x23], #0x8\n"
1335 "tbz x13, #0, 124f\n"
1336 "st1 { v24.s }[2], [x11]\n"
1337 "st1 { v25.s }[2], [x27]\n"
1338 "st1 { v26.s }[2], [x26]\n"
1339 "st1 { v27.s }[2], [x25]\n"
1340 "st1 { v28.s }[2], [x24]\n"
1341 "st1 { v29.s }[2], [x23]\n"
1344 "str s24, [x11, #0x0]\n"
1345 "str s25, [x27, #0x0]\n"
1346 "str s26, [x26, #0x0]\n"
1347 "str s27, [x25, #0x0]\n"
1348 "str s28, [x24, #0x0]\n"
1349 "str s29, [x23, #0x0]\n"
1353 "str q24, [x11, #0x0]\n"
1354 "add x11, x11, #0x10\n"
1355 "str q25, [x27, #0x0]\n"
1356 "str q26, [x26, #0x0]\n"
1357 "str q27, [x25, #0x0]\n"
1358 "str q28, [x24, #0x0]\n"
1359 "str q29, [x23, #0x0]\n"
1361 "subs x13, x13, #0x4\n"
1365 "mov x14, %x[bias]\n"
1366 "ldr x13, [%x[args_ptr], %[offsetof_N]]\n"
1367 "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
1368 "mov x11, %x[output_ptr]\n"
1371 "ldr q24, [x14, #0x0]\n"
1372 "mov v25.16b, v24.16b\n"
1373 "mov v26.16b, v24.16b\n"
1374 "add x14, x14, #0x10\n"
1375 "mov v27.16b, v24.16b\n"
1376 "mov v28.16b, v24.16b\n"
1377 "mov v29.16b, v24.16b\n"
1378 "mov v30.16b, v24.16b\n"
1381 "tbz %x[flags], #0, 133f\n"
1382 "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
1383 "add x27, x11, x20, LSL #2\n"
1384 "add x26, x27, x20, LSL #2\n"
1385 "add x25, x26, x20, LSL #2\n"
1386 "add x24, x25, x20, LSL #2\n"
1387 "add x23, x24, x20, LSL #2\n"
1389 "add x22, x23, x20, LSL #2\n"
1391 "tbz x13, #1, 130f\n"
1392 "ldr d24, [x11], #0x8\n"
1393 "ldr d25, [x27], #0x8\n"
1395 "ldr d26, [x26], #0x8\n"
1396 "ldr d27, [x25], #0x8\n"
1397 "ldr d28, [x24], #0x8\n"
1398 "ldr d29, [x23], #0x8\n"
1399 "ldr d30, [x22], #0x8\n"
1400 "tbz x13, #0, 131f\n"
1401 "ld1 { v24.s }[2], [x11]\n"
1402 "ld1 { v25.s }[2], [x27]\n"
1403 "ld1 { v26.s }[2], [x26]\n"
1404 "ld1 { v27.s }[2], [x25]\n"
1405 "ld1 { v28.s }[2], [x24]\n"
1406 "ld1 { v29.s }[2], [x23]\n"
1407 "ld1 { v30.s }[2], [x22]\n"
1410 "ldr s24, [x11, #0x0]\n"
1411 "ldr s25, [x27, #0x0]\n"
1413 "ldr s26, [x26, #0x0]\n"
1414 "ldr s27, [x25, #0x0]\n"
1415 "ldr s28, [x24, #0x0]\n"
1416 "ldr s29, [x23, #0x0]\n"
1417 "ldr s30, [x22, #0x0]\n"
1419 "sub x11, x11, x20\n"
1422 "ldr q24, [x11, #0x0]\n"
1423 "ldr q25, [x27, #0x0]\n"
1424 "ldr q26, [x26, #0x0]\n"
1425 "ldr q27, [x25, #0x0]\n"
1426 "ldr q28, [x24, #0x0]\n"
1427 "ldr q29, [x23, #0x0]\n"
1428 "ldr q30, [x22, #0x0]\n"
1431 "movi v24.16b, #0x0\n"
1432 "movi v25.16b, #0x0\n"
1433 "movi v26.16b, #0x0\n"
1434 "movi v27.16b, #0x0\n"
1435 "movi v28.16b, #0x0\n"
1436 "movi v29.16b, #0x0\n"
1437 "movi v30.16b, #0x0\n"
1441 "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
1442 "ldr w9, [x20, x10, LSL #0x2]\n"
1443 "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
1444 "tbz %x[flags], #3, 136f\n"
1445 "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n"
1446 "add x20, x20, x21, LSL #3\n"
1447 "ldr x28, [x20, #0x0]\n"
1448 "ldr x27, [x20, #0x8]\n"
1449 "ldr x26, [x20, #0x10]\n"
1450 "ldr x25, [x20, #0x18]\n"
1451 "ldr x24, [x20, #0x20]\n"
1452 "ldr x23, [x20, #0x28]\n"
1453 "ldr x22, [x20, #0x30]\n"
1455 "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
1456 "add x28, x28, x20, LSL #2\n"
1457 "add x27, x27, x20, LSL #2\n"
1458 "add x26, x26, x20, LSL #2\n"
1459 "add x25, x25, x20, LSL #2\n"
1460 "add x24, x24, x20, LSL #2\n"
1461 "add x23, x23, x20, LSL #2\n"
1462 "add x22, x22, x20, LSL #2\n"
1465 "mov x28, %x[input_ptr]\n"
1466 "add x27, x28, x21, LSL #2\n"
1467 "add x26, x27, x21, LSL #2\n"
1468 "add x25, x26, x21, LSL #2\n"
1469 "add x24, x25, x21, LSL #2\n"
1470 "add x23, x24, x21, LSL #2\n"
1471 "add x22, x23, x21, LSL #2\n"
1475 "ldr q0, [x28, #0x0]\n"
1476 "ldr q1, [x27, #0x0]\n"
1478 "ldr q2, [x26, #0x0]\n"
1479 "ldr q3, [x25, #0x0]\n"
1480 "ldr q4, [x24, #0x0]\n"
1481 "ldr q5, [x23, #0x0]\n"
1482 "ldr q6, [x22, #0x0]\n"
1483 "ldr q8, [x12, #0x0]\n"
1484 "ldr q9, [x12, #0x10]\n"
1485 "ldr q10, [x12, #0x20]\n"
1486 "ldr q11, [x12, #0x30]\n"
1489 "fmla v24.4s, v8.4s, v0.s[0]\n"
1490 "fmla v25.4s, v8.4s, v1.s[0]\n"
1491 "sub x9, x9, #0x4\n"
1492 "add x28, x28, #0x10\n"
1493 "fmla v26.4s, v8.4s, v2.s[0]\n"
1494 "fmla v27.4s, v8.4s, v3.s[0]\n"
1495 "add x27, x27, #0x10\n"
1496 "add x26, x26, #0x10\n"
1497 "fmla v28.4s, v8.4s, v4.s[0]\n"
1498 "fmla v29.4s, v8.4s, v5.s[0]\n"
1499 "add x25, x25, #0x10\n"
1500 "add x24, x24, #0x10\n"
1501 "fmla v30.4s, v8.4s, v6.s[0]\n"
1502 "fmla v24.4s, v9.4s, v0.s[1]\n"
1503 "add x23, x23, #0x10\n"
1504 "add x22, x22, #0x10\n"
1505 "fmla v25.4s, v9.4s, v1.s[1]\n"
1506 "fmla v26.4s, v9.4s, v2.s[1]\n"
1508 "add x12, x12, #0x40\n"
1509 "ldr q8, [x12, #0x0]\n"
1510 "fmla v27.4s, v9.4s, v3.s[1]\n"
1511 "fmla v28.4s, v9.4s, v4.s[1]\n"
1512 "prfm pldl1keep, [x28, #0x80]\n"
1513 "fmla v29.4s, v9.4s, v5.s[1]\n"
1514 "fmla v30.4s, v9.4s, v6.s[1]\n"
1515 "ldr q9, [x12, #0x10]\n"
1516 "prfm pldl1keep, [x27, #0x80]\n"
1517 "fmla v24.4s, v10.4s, v0.s[2]\n"
1518 "fmla v25.4s, v10.4s, v1.s[2]\n"
1519 "prfm pldl1keep, [x26, #0x80]\n"
1520 "prfm pldl1keep, [x25, #0x80]\n"
1521 "fmla v26.4s, v10.4s, v2.s[2]\n"
1522 "fmla v27.4s, v10.4s, v3.s[2]\n"
1523 "prfm pldl1keep, [x24, #0x80]\n"
1524 "prfm pldl1keep, [x23, #0x80]\n"
1525 "fmla v28.4s, v10.4s, v4.s[2]\n"
1526 "fmla v29.4s, v10.4s, v5.s[2]\n"
1527 "prfm pldl1keep, [x22, #0x80]\n"
1528 "fmla v30.4s, v10.4s, v6.s[2]\n"
1529 "ldr q10, [x12, #0x20]\n"
1530 "fmla v24.4s, v11.4s, v0.s[3]\n"
1531 "ldr q0, [x28, #0x0]\n"
1532 "fmla v25.4s, v11.4s, v1.s[3]\n"
1533 "ldr q1, [x27, #0x0]\n"
1534 "fmla v26.4s, v11.4s, v2.s[3]\n"
1535 "ldr q2, [x26, #0x0]\n"
1536 "fmla v27.4s, v11.4s, v3.s[3]\n"
1537 "ldr q3, [x25, #0x0]\n"
1538 "fmla v28.4s, v11.4s, v4.s[3]\n"
1539 "ldr q4, [x24, #0x0]\n"
1540 "fmla v29.4s, v11.4s, v5.s[3]\n"
1541 "ldr q5, [x23, #0x0]\n"
1542 "fmla v30.4s, v11.4s, v6.s[3]\n"
1543 "ldr q6, [x22, #0x0]\n"
1544 "ldr q11, [x12, #0x30]\n"
1547 "fmla v24.4s, v8.4s, v0.s[0]\n"
1548 "fmla v25.4s, v8.4s, v1.s[0]\n"
1549 "add x28, x28, #0x10\n"
1550 "add x27, x27, #0x10\n"
1551 "fmla v26.4s, v8.4s, v2.s[0]\n"
1552 "fmla v27.4s, v8.4s, v3.s[0]\n"
1553 "add x26, x26, #0x10\n"
1554 "add x25, x25, #0x10\n"
1555 "fmla v28.4s, v8.4s, v4.s[0]\n"
1556 "fmla v29.4s, v8.4s, v5.s[0]\n"
1557 "add x24, x24, #0x10\n"
1558 "add x23, x23, #0x10\n"
1559 "fmla v30.4s, v8.4s, v6.s[0]\n"
1560 "fmla v24.4s, v9.4s, v0.s[1]\n"
1561 "add x22, x22, #0x10\n"
1562 "sub x9, x9, #0x4\n"
1563 "fmla v25.4s, v9.4s, v1.s[1]\n"
1564 "fmla v26.4s, v9.4s, v2.s[1]\n"
1565 "prfm pldl1keep, [x28, #0x80]\n"
1566 "prfm pldl1keep, [x27, #0x80]\n"
1567 "fmla v27.4s, v9.4s, v3.s[1]\n"
1568 "fmla v28.4s, v9.4s, v4.s[1]\n"
1569 "prfm pldl1keep, [x26, #0x80]\n"
1570 "prfm pldl1keep, [x25, #0x80]\n"
1571 "fmla v29.4s, v9.4s, v5.s[1]\n"
1572 "fmla v30.4s, v9.4s, v6.s[1]\n"
1573 "prfm pldl1keep, [x24, #0x80]\n"
1574 "prfm pldl1keep, [x23, #0x80]\n"
1575 "fmla v24.4s, v10.4s, v0.s[2]\n"
1576 "fmla v25.4s, v10.4s, v1.s[2]\n"
1577 "prfm pldl1keep, [x22, #0x80]\n"
1578 "add x12, x12, #0x40\n"
1579 "fmla v26.4s, v10.4s, v2.s[2]\n"
1580 "fmla v27.4s, v10.4s, v3.s[2]\n"
1581 "fmla v28.4s, v10.4s, v4.s[2]\n"
1582 "fmla v29.4s, v10.4s, v5.s[2]\n"
1583 "fmla v30.4s, v10.4s, v6.s[2]\n"
1584 "fmla v24.4s, v11.4s, v0.s[3]\n"
1585 "fmla v25.4s, v11.4s, v1.s[3]\n"
1586 "fmla v26.4s, v11.4s, v2.s[3]\n"
1587 "fmla v27.4s, v11.4s, v3.s[3]\n"
1588 "fmla v28.4s, v11.4s, v4.s[3]\n"
1589 "fmla v29.4s, v11.4s, v5.s[3]\n"
1590 "fmla v30.4s, v11.4s, v6.s[3]\n"
1594 "ldr s23, [x28], #0x4\n"
1595 "ldr s22, [x27], #0x4\n"
1596 "sub x9, x9, #0x1\n"
1597 "ldr s21, [x26], #0x4\n"
1598 "ldr s20, [x25], #0x4\n"
1599 "ldr s19, [x24], #0x4\n"
1600 "ldr s18, [x23], #0x4\n"
1601 "ldr s17, [x22], #0x4\n"
1602 "ldr q16, [x12, #0x0]\n"
1603 "fmla v24.4s, v16.4s, v23.s[0]\n"
1604 "fmla v25.4s, v16.4s, v22.s[0]\n"
1605 "fmla v26.4s, v16.4s, v21.s[0]\n"
1606 "fmla v27.4s, v16.4s, v20.s[0]\n"
1607 "add x12, x12, #0x10\n"
1608 "fmla v28.4s, v16.4s, v19.s[0]\n"
1609 "fmla v29.4s, v16.4s, v18.s[0]\n"
1610 "fmla v30.4s, v16.4s, v17.s[0]\n"
1613 "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
1614 "add x10, x10, #0x1\n"
1617 "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
1618 "add x27, x11, x20, LSL #2\n"
1619 "add x26, x27, x20, LSL #2\n"
1620 "prfm pstl1keep, [x11, #0x0]\n"
1621 "add x25, x26, x20, LSL #2\n"
1622 "add x24, x25, x20, LSL #2\n"
1623 "prfm pstl1keep, [x27, #0x0]\n"
1624 "prfm pstl1keep, [x26, #0x0]\n"
1625 "add x23, x24, x20, LSL #2\n"
1626 "add x22, x23, x20, LSL #2\n"
1627 "prfm pstl1keep, [x25, #0x0]\n"
1628 "prfm pstl1keep, [x24, #0x0]\n"
1629 "prfm pstl1keep, [x23, #0x0]\n"
1630 "prfm pstl1keep, [x22, #0x0]\n"
1631 "tbz %x[flags], #1, 143f\n"
1632 "add x20, %x[args_ptr], %[offset_max]\n"
1633 "ld1r { v17.4s }, [x20]\n"
1634 "add x20, %x[args_ptr], %[offset_min]\n"
1635 "ld1r { v16.4s }, [x20]\n"
1636 "fmin v24.4s, v24.4s, v17.4s\n"
1637 "fmin v25.4s, v25.4s, v17.4s\n"
1638 "fmin v26.4s, v26.4s, v17.4s\n"
1639 "fmin v27.4s, v27.4s, v17.4s\n"
1640 "fmin v28.4s, v28.4s, v17.4s\n"
1641 "fmin v29.4s, v29.4s, v17.4s\n"
1642 "fmin v30.4s, v30.4s, v17.4s\n"
1643 "fmax v24.4s, v24.4s, v16.4s\n"
1644 "fmax v25.4s, v25.4s, v16.4s\n"
1645 "fmax v26.4s, v26.4s, v16.4s\n"
1646 "fmax v27.4s, v27.4s, v16.4s\n"
1647 "fmax v28.4s, v28.4s, v16.4s\n"
1648 "fmax v29.4s, v29.4s, v16.4s\n"
1649 "fmax v30.4s, v30.4s, v16.4s\n"
1653 "tbz x13, #1, 144f\n"
1654 "str d24, [x11], #0x8\n"
1655 "str d25, [x27], #0x8\n"
1656 "str d26, [x26], #0x8\n"
1657 "str d27, [x25], #0x8\n"
1658 "str d28, [x24], #0x8\n"
1659 "str d29, [x23], #0x8\n"
1660 "str d30, [x22], #0x8\n"
1661 "tbz x13, #0, 145f\n"
1662 "st1 { v24.s }[2], [x11]\n"
1663 "st1 { v25.s }[2], [x27]\n"
1664 "st1 { v26.s }[2], [x26]\n"
1665 "st1 { v27.s }[2], [x25]\n"
1666 "st1 { v28.s }[2], [x24]\n"
1667 "st1 { v29.s }[2], [x23]\n"
1668 "st1 { v30.s }[2], [x22]\n"
1671 "str s24, [x11, #0x0]\n"
1672 "str s25, [x27, #0x0]\n"
1673 "str s26, [x26, #0x0]\n"
1674 "str s27, [x25, #0x0]\n"
1675 "str s28, [x24, #0x0]\n"
1676 "str s29, [x23, #0x0]\n"
1677 "str s30, [x22, #0x0]\n"
1681 "str q24, [x11, #0x0]\n"
1682 "add x11, x11, #0x10\n"
1683 "str q25, [x27, #0x0]\n"
1684 "str q26, [x26, #0x0]\n"
1685 "str q27, [x25, #0x0]\n"
1686 "str q28, [x24, #0x0]\n"
1687 "str q29, [x23, #0x0]\n"
1688 "str q30, [x22, #0x0]\n"
1690 "subs x13, x13, #0x4\n"
1694 "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
1696 "mov x14, %x[bias]\n"
1697 "ldr x13, [%x[args_ptr], %[offsetof_N]]\n"
1698 "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
1699 "mov x11, %x[output_ptr]\n"
1700 "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
1703 "ldr q24, [x14, #0x0]\n"
1704 "mov v25.16b, v24.16b\n"
1705 "mov v26.16b, v24.16b\n"
1706 "add x14, x14, #0x10\n"
1707 "mov v27.16b, v24.16b\n"
1708 "mov v28.16b, v24.16b\n"
1709 "mov v29.16b, v24.16b\n"
1710 "mov v30.16b, v24.16b\n"
1711 "mov v31.16b, v24.16b\n"
1714 "tbz %x[flags], #0, 154f\n"
1715 "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
1716 "add x27, x11, x20, LSL #2\n"
1717 "add x26, x27, x20, LSL #2\n"
1718 "add x25, x26, x20, LSL #2\n"
1719 "add x24, x25, x20, LSL #2\n"
1720 "add x23, x24, x20, LSL #2\n"
1721 "add x22, x23, x20, LSL #2\n"
1723 "add x21, x22, x20, LSL #2\n"
1725 "tbz x13, #1, 151f\n"
1726 "ldr d24, [x11], #0x8\n"
1727 "ldr d25, [x27], #0x8\n"
1729 "ldr d26, [x26], #0x8\n"
1730 "ldr d27, [x25], #0x8\n"
1731 "ldr d28, [x24], #0x8\n"
1732 "ldr d29, [x23], #0x8\n"
1733 "ldr d30, [x22], #0x8\n"
1734 "ldr d31, [x21], #0x8\n"
1735 "tbz x13, #0, 152f\n"
1736 "ld1 { v24.s }[2], [x11]\n"
1737 "ld1 { v25.s }[2], [x27]\n"
1738 "ld1 { v26.s }[2], [x26]\n"
1739 "ld1 { v27.s }[2], [x25]\n"
1740 "ld1 { v28.s }[2], [x24]\n"
1741 "ld1 { v29.s }[2], [x23]\n"
1742 "ld1 { v30.s }[2], [x22]\n"
1743 "ld1 { v31.s }[2], [x21]\n"
1746 "ldr s24, [x11, #0x0]\n"
1747 "ldr s25, [x27, #0x0]\n"
1749 "ldr s26, [x26, #0x0]\n"
1750 "ldr s27, [x25, #0x0]\n"
1751 "ldr s28, [x24, #0x0]\n"
1752 "ldr s29, [x23, #0x0]\n"
1753 "ldr s30, [x22, #0x0]\n"
1754 "ldr s31, [x21, #0x0]\n"
1756 "sub x11, x11, x20\n"
1759 "ldr q24, [x11, #0x0]\n"
1760 "ldr q25, [x27, #0x0]\n"
1761 "ldr q26, [x26, #0x0]\n"
1762 "ldr q27, [x25, #0x0]\n"
1763 "ldr q28, [x24, #0x0]\n"
1764 "ldr q29, [x23, #0x0]\n"
1765 "ldr q30, [x22, #0x0]\n"
1766 "ldr q31, [x21, #0x0]\n"
1769 "movi v24.16b, #0x0\n"
1770 "movi v25.16b, #0x0\n"
1771 "movi v26.16b, #0x0\n"
1772 "movi v27.16b, #0x0\n"
1773 "movi v28.16b, #0x0\n"
1774 "movi v29.16b, #0x0\n"
1775 "movi v30.16b, #0x0\n"
1776 "movi v31.16b, #0x0\n"
1780 "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
1781 "ldr w9, [x20, x10, LSL #0x2]\n"
1782 "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
1783 "tbz %x[flags], #3, 157f\n"
1784 "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n"
1785 "add x20, x20, x21, LSL #3\n"
1786 "ldr x28, [x20, #0x0]\n"
1787 "ldr x27, [x20, #0x8]\n"
1788 "ldr x26, [x20, #0x10]\n"
1789 "ldr x25, [x20, #0x18]\n"
1790 "ldr x24, [x20, #0x20]\n"
1791 "ldr x23, [x20, #0x28]\n"
1792 "ldr x22, [x20, #0x30]\n"
1793 "ldr x21, [x20, #0x38]\n"
1795 "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
1796 "add x28, x28, x20, LSL #2\n"
1797 "add x27, x27, x20, LSL #2\n"
1798 "add x26, x26, x20, LSL #2\n"
1799 "add x25, x25, x20, LSL #2\n"
1800 "add x24, x24, x20, LSL #2\n"
1801 "add x23, x23, x20, LSL #2\n"
1802 "add x22, x22, x20, LSL #2\n"
1803 "add x21, x21, x20, LSL #2\n"
1806 "mov x28, %x[input_ptr]\n"
1807 "add x27, x28, x21, LSL #2\n"
1808 "add x26, x27, x21, LSL #2\n"
1809 "add x25, x26, x21, LSL #2\n"
1810 "add x24, x25, x21, LSL #2\n"
1811 "add x23, x24, x21, LSL #2\n"
1812 "add x22, x23, x21, LSL #2\n"
1813 "add x21, x22, x21, LSL #2\n"
1817 "ldr q0, [x28, #0x0]\n"
1818 "ldr q1, [x27, #0x0]\n"
1820 "ldr q2, [x26, #0x0]\n"
1821 "ldr q3, [x25, #0x0]\n"
1822 "ldr q4, [x24, #0x0]\n"
1823 "ldr q5, [x23, #0x0]\n"
1824 "ldr q6, [x22, #0x0]\n"
1825 "ldr q7, [x21, #0x0]\n"
1826 "ldr q8, [x12, #0x0]\n"
1827 "ldr q9, [x12, #0x10]\n"
1828 "ldr q10, [x12, #0x20]\n"
1829 "ldr q11, [x12, #0x30]\n"
1832 "fmla v24.4s, v8.4s, v0.s[0]\n"
1833 "fmla v25.4s, v8.4s, v1.s[0]\n"
1834 "sub x9, x9, #0x4\n"
1835 "add x28, x28, #0x10\n"
1836 "fmla v26.4s, v8.4s, v2.s[0]\n"
1837 "fmla v27.4s, v8.4s, v3.s[0]\n"
1838 "add x27, x27, #0x10\n"
1839 "add x26, x26, #0x10\n"
1840 "fmla v28.4s, v8.4s, v4.s[0]\n"
1841 "fmla v29.4s, v8.4s, v5.s[0]\n"
1842 "add x25, x25, #0x10\n"
1843 "add x24, x24, #0x10\n"
1844 "fmla v30.4s, v8.4s, v6.s[0]\n"
1845 "fmla v31.4s, v8.4s, v7.s[0]\n"
1846 "add x23, x23, #0x10\n"
1847 "add x22, x22, #0x10\n"
1848 "fmla v24.4s, v9.4s, v0.s[1]\n"
1849 "fmla v25.4s, v9.4s, v1.s[1]\n"
1850 "add x21, x21, #0x10\n"
1852 "fmla v26.4s, v9.4s, v2.s[1]\n"
1853 "fmla v27.4s, v9.4s, v3.s[1]\n"
1854 "add x12, x12, #0x40\n"
1855 "ldr q8, [x12, #0x0]\n"
1856 "fmla v28.4s, v9.4s, v4.s[1]\n"
1857 "fmla v29.4s, v9.4s, v5.s[1]\n"
1858 "prfm pldl1keep, [x28, #0x80]\n"
1859 "prfm pldl1keep, [x27, #0x80]\n"
1860 "fmla v30.4s, v9.4s, v6.s[1]\n"
1861 "fmla v31.4s, v9.4s, v7.s[1]\n"
1862 "ldr q9, [x12, #0x10]\n"
1863 "prfm pldl1keep, [x26, #0x80]\n"
1864 "fmla v24.4s, v10.4s, v0.s[2]\n"
1865 "fmla v25.4s, v10.4s, v1.s[2]\n"
1866 "prfm pldl1keep, [x25, #0x80]\n"
1867 "prfm pldl1keep, [x24, #0x80]\n"
1868 "fmla v26.4s, v10.4s, v2.s[2]\n"
1869 "fmla v27.4s, v10.4s, v3.s[2]\n"
1870 "prfm pldl1keep, [x23, #0x80]\n"
1871 "prfm pldl1keep, [x22, #0x80]\n"
1872 "fmla v28.4s, v10.4s, v4.s[2]\n"
1873 "fmla v29.4s, v10.4s, v5.s[2]\n"
1874 "prfm pldl1keep, [x21, #0x80]\n"
1875 "fmla v30.4s, v10.4s, v6.s[2]\n"
1876 "fmla v31.4s, v10.4s, v7.s[2]\n"
1877 "ldr q10, [x12, #0x20]\n"
1878 "fmla v24.4s, v11.4s, v0.s[3]\n"
1879 "ldr q0, [x28, #0x0]\n"
1880 "fmla v25.4s, v11.4s, v1.s[3]\n"
1881 "ldr q1, [x27, #0x0]\n"
1882 "fmla v26.4s, v11.4s, v2.s[3]\n"
1883 "ldr q2, [x26, #0x0]\n"
1884 "fmla v27.4s, v11.4s, v3.s[3]\n"
1885 "ldr q3, [x25, #0x0]\n"
1886 "fmla v28.4s, v11.4s, v4.s[3]\n"
1887 "ldr q4, [x24, #0x0]\n"
1888 "fmla v29.4s, v11.4s, v5.s[3]\n"
1889 "ldr q5, [x23, #0x0]\n"
1890 "fmla v30.4s, v11.4s, v6.s[3]\n"
1891 "ldr q6, [x22, #0x0]\n"
1892 "fmla v31.4s, v11.4s, v7.s[3]\n"
1893 "ldr q7, [x21, #0x0]\n"
1894 "ldr q11, [x12, #0x30]\n"
1897 "fmla v24.4s, v8.4s, v0.s[0]\n"
1898 "fmla v25.4s, v8.4s, v1.s[0]\n"
1899 "add x28, x28, #0x10\n"
1900 "add x27, x27, #0x10\n"
1901 "fmla v26.4s, v8.4s, v2.s[0]\n"
1902 "fmla v27.4s, v8.4s, v3.s[0]\n"
1903 "add x26, x26, #0x10\n"
1904 "add x25, x25, #0x10\n"
1905 "fmla v28.4s, v8.4s, v4.s[0]\n"
1906 "fmla v29.4s, v8.4s, v5.s[0]\n"
1907 "add x24, x24, #0x10\n"
1908 "add x23, x23, #0x10\n"
1909 "fmla v30.4s, v8.4s, v6.s[0]\n"
1910 "fmla v31.4s, v8.4s, v7.s[0]\n"
1911 "add x22, x22, #0x10\n"
1912 "add x21, x21, #0x10\n"
1913 "fmla v24.4s, v9.4s, v0.s[1]\n"
1914 "fmla v25.4s, v9.4s, v1.s[1]\n"
1915 "sub x9, x9, #0x4\n"
1916 "prfm pldl1keep, [x28, #0x80]\n"
1917 "fmla v26.4s, v9.4s, v2.s[1]\n"
1918 "fmla v27.4s, v9.4s, v3.s[1]\n"
1919 "prfm pldl1keep, [x27, #0x80]\n"
1920 "prfm pldl1keep, [x26, #0x80]\n"
1921 "fmla v28.4s, v9.4s, v4.s[1]\n"
1922 "fmla v29.4s, v9.4s, v5.s[1]\n"
1923 "prfm pldl1keep, [x25, #0x80]\n"
1924 "prfm pldl1keep, [x24, #0x80]\n"
1925 "fmla v30.4s, v9.4s, v6.s[1]\n"
1926 "fmla v31.4s, v9.4s, v7.s[1]\n"
1927 "prfm pldl1keep, [x23, #0x80]\n"
1928 "prfm pldl1keep, [x22, #0x80]\n"
1929 "fmla v24.4s, v10.4s, v0.s[2]\n"
1930 "fmla v25.4s, v10.4s, v1.s[2]\n"
1931 "prfm pldl1keep, [x21, #0x80]\n"
1932 "add x12, x12, #0x40\n"
1933 "fmla v26.4s, v10.4s, v2.s[2]\n"
1934 "fmla v27.4s, v10.4s, v3.s[2]\n"
1935 "fmla v28.4s, v10.4s, v4.s[2]\n"
1936 "fmla v29.4s, v10.4s, v5.s[2]\n"
1937 "fmla v30.4s, v10.4s, v6.s[2]\n"
1938 "fmla v31.4s, v10.4s, v7.s[2]\n"
1939 "fmla v24.4s, v11.4s, v0.s[3]\n"
1940 "fmla v25.4s, v11.4s, v1.s[3]\n"
1941 "fmla v26.4s, v11.4s, v2.s[3]\n"
1942 "fmla v27.4s, v11.4s, v3.s[3]\n"
1943 "fmla v28.4s, v11.4s, v4.s[3]\n"
1944 "fmla v29.4s, v11.4s, v5.s[3]\n"
1945 "fmla v30.4s, v11.4s, v6.s[3]\n"
1946 "fmla v31.4s, v11.4s, v7.s[3]\n"
1950 "ldr s0, [x28], #0x4\n"
1951 "ldr s23, [x27], #0x4\n"
1952 "sub x9, x9, #0x1\n"
1953 "ldr s22, [x26], #0x4\n"
1954 "ldr s21, [x25], #0x4\n"
1955 "ldr s20, [x24], #0x4\n"
1956 "ldr s19, [x23], #0x4\n"
1957 "ldr s18, [x22], #0x4\n"
1958 "ldr s17, [x21], #0x4\n"
1959 "ldr q16, [x12, #0x0]\n"
1960 "fmla v24.4s, v16.4s, v0.s[0]\n"
1961 "fmla v25.4s, v16.4s, v23.s[0]\n"
1962 "add x12, x12, #0x10\n"
1963 "fmla v26.4s, v16.4s, v22.s[0]\n"
1964 "fmla v27.4s, v16.4s, v21.s[0]\n"
1965 "fmla v28.4s, v16.4s, v20.s[0]\n"
1966 "fmla v29.4s, v16.4s, v19.s[0]\n"
1967 "fmla v30.4s, v16.4s, v18.s[0]\n"
1968 "fmla v31.4s, v16.4s, v17.s[0]\n"
1971 "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
1972 "add x10, x10, #0x1\n"
1975 "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
1976 "add x27, x11, x20, LSL #2\n"
1977 "add x26, x27, x20, LSL #2\n"
1978 "prfm pstl1keep, [x11, #0x0]\n"
1979 "add x25, x26, x20, LSL #2\n"
1980 "add x24, x25, x20, LSL #2\n"
1981 "prfm pstl1keep, [x27, #0x0]\n"
1982 "prfm pstl1keep, [x26, #0x0]\n"
1983 "add x23, x24, x20, LSL #2\n"
1984 "add x22, x23, x20, LSL #2\n"
1985 "prfm pstl1keep, [x25, #0x0]\n"
1986 "prfm pstl1keep, [x24, #0x0]\n"
1987 "add x21, x22, x20, LSL #2\n"
1988 "prfm pstl1keep, [x23, #0x0]\n"
1989 "prfm pstl1keep, [x22, #0x0]\n"
1990 "prfm pstl1keep, [x21, #0x0]\n"
1991 "tbz %x[flags], #1, 164f\n"
1992 "add x20, %x[args_ptr], %[offset_max]\n"
1993 "ld1r { v17.4s }, [x20]\n"
1994 "add x20, %x[args_ptr], %[offset_min]\n"
1995 "ld1r { v16.4s }, [x20]\n"
1996 "fmin v24.4s, v24.4s, v17.4s\n"
1997 "fmin v25.4s, v25.4s, v17.4s\n"
1998 "fmin v26.4s, v26.4s, v17.4s\n"
1999 "fmin v27.4s, v27.4s, v17.4s\n"
2000 "fmin v28.4s, v28.4s, v17.4s\n"
2001 "fmin v29.4s, v29.4s, v17.4s\n"
2002 "fmin v30.4s, v30.4s, v17.4s\n"
2003 "fmin v31.4s, v31.4s, v17.4s\n"
2004 "fmax v24.4s, v24.4s, v16.4s\n"
2005 "fmax v25.4s, v25.4s, v16.4s\n"
2006 "fmax v26.4s, v26.4s, v16.4s\n"
2007 "fmax v27.4s, v27.4s, v16.4s\n"
2008 "fmax v28.4s, v28.4s, v16.4s\n"
2009 "fmax v29.4s, v29.4s, v16.4s\n"
2010 "fmax v30.4s, v30.4s, v16.4s\n"
2011 "fmax v31.4s, v31.4s, v16.4s\n"
2015 "tbz x13, #1, 165f\n"
2016 "str d24, [x11], #0x8\n"
2017 "str d25, [x27], #0x8\n"
2018 "str d26, [x26], #0x8\n"
2019 "str d27, [x25], #0x8\n"
2020 "str d28, [x24], #0x8\n"
2021 "str d29, [x23], #0x8\n"
2022 "str d30, [x22], #0x8\n"
2023 "str d31, [x21], #0x8\n"
2024 "tbz x13, #0, 166f\n"
2025 "st1 { v24.s }[2], [x11]\n"
2026 "st1 { v25.s }[2], [x27]\n"
2027 "st1 { v26.s }[2], [x26]\n"
2028 "st1 { v27.s }[2], [x25]\n"
2029 "st1 { v28.s }[2], [x24]\n"
2030 "st1 { v29.s }[2], [x23]\n"
2031 "st1 { v30.s }[2], [x22]\n"
2032 "st1 { v31.s }[2], [x21]\n"
2035 "str s24, [x11, #0x0]\n"
2036 "str s25, [x27, #0x0]\n"
2037 "str s26, [x26, #0x0]\n"
2038 "str s27, [x25, #0x0]\n"
2039 "str s28, [x24, #0x0]\n"
2040 "str s29, [x23, #0x0]\n"
2041 "str s30, [x22, #0x0]\n"
2042 "str s31, [x21, #0x0]\n"
2046 "str q24, [x11, #0x0]\n"
2047 "add x11, x11, #0x10\n"
2048 "str q25, [x27, #0x0]\n"
2049 "str q26, [x26, #0x0]\n"
2050 "str q27, [x25, #0x0]\n"
2051 "str q28, [x24, #0x0]\n"
2052 "str q29, [x23, #0x0]\n"
2053 "str q30, [x22, #0x0]\n"
2054 "str q31, [x21, #0x0]\n"
2056 "subs x13, x13, #0x4\n"
2058 "subs %x[M], %x[M], #0x8\n"
2060 "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
2061 "tbz %x[flags], #3, 169f\n"
2062 "add x21, x21, #0x8\n"
2063 "str x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
2067 "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
2070 : [
M]
"+&r" (
M), [input_ptr]
"+&r" (input_ptr), [output_ptr]
"+&r" (output_ptr)
2071 : [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))
2072 :
"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",
"x9",
"x10",
"x11",
"x12",
"x13",
"x14",
"x20",
"x21",
"x22",
"x23",
"x24",
"x25",
"x26",
"x27",
"x28"
2077 #endif // __aarch64__