27 #include "../../utils.hpp"
34 void a64_hybrid_fp32_mla_6x16 (
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);
104 "mov x12, %x[bias]\n"
105 "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
106 "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
107 "mov x9, %x[output_ptr]\n"
110 "ldr q8, [x12, #0x0]\n"
111 "ldr q9, [x12, #0x10]\n"
112 "ldr q10, [x12, #0x20]\n"
113 "ldr q11, [x12, #0x30]\n"
114 "add x12, x12, #0x40\n"
117 "tbz %x[flags], #0, 13f\n"
121 "ld1 { v8.4s }, [x9], #0x10\n"
122 "ld1 { v9.4s }, [x9], #0x10\n"
124 "ld1 { v10.4s }, [x9], #0x10\n"
126 "ldr d11, [x9], #0x8\n"
129 "ld1 { v11.s }[2], [x9]\n"
134 "ldr s11, [x9, #0x0]\n"
138 "ldr d10, [x9], #0x8\n"
141 "ld1 { v10.s }[2], [x9]\n"
146 "ldr s10, [x9, #0x0]\n"
150 "ld1 { v8.4s }, [x9], #0x10\n"
152 "ldr d9, [x9], #0x8\n"
155 "ld1 { v9.s }[2], [x9]\n"
160 "ldr s9, [x9, #0x0]\n"
164 "ldr d8, [x9], #0x8\n"
167 "ld1 { v8.s }[2], [x9]\n"
170 "ldr s8, [x9, #0x0]\n"
176 "ldr q8, [x9, #0x0]\n"
177 "ldr q9, [x9, #0x10]\n"
178 "ldr q10, [x9, #0x20]\n"
179 "ldr q11, [x9, #0x30]\n"
182 "movi v8.16b, #0x0\n"
183 "movi v9.16b, #0x0\n"
184 "movi v10.16b, #0x0\n"
185 "movi v11.16b, #0x0\n"
189 "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
190 "ldr w27, [x20, x28, LSL #0x2]\n"
191 "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
192 "tbz %x[flags], #3, 16f\n"
193 "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
194 "add x20, x20, x21, LSL #3\n"
195 "ldr x26, [x20, #0x0]\n"
197 "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
198 "add x26, x26, x20, LSL #2\n"
201 "mov x26, %x[input_ptr]\n"
205 "ldr q0, [x26, #0x0]\n"
206 "ldr q6, [x10, #0x0]\n"
208 "ldr q7, [x10, #0x10]\n"
211 "fmla v8.4s, v6.4s, v0.s[0]\n"
212 "ldr q17, [x10, #0x20]\n"
213 "fmla v9.4s, v7.4s, v0.s[0]\n"
214 "ldr q16, [x10, #0x30]\n"
215 "fmla v10.4s, v17.4s, v0.s[0]\n"
216 "ldr q17, [x10, #0x40]\n"
217 "fmla v11.4s, v16.4s, v0.s[0]\n"
218 "ldr q16, [x10, #0x50]\n"
219 "fmla v8.4s, v17.4s, v0.s[1]\n"
220 "ldr q17, [x10, #0x60]\n"
221 "fmla v9.4s, v16.4s, v0.s[1]\n"
222 "ldr q16, [x10, #0x70]\n"
223 "fmla v10.4s, v17.4s, v0.s[1]\n"
224 "ldr q17, [x10, #0x80]\n"
225 "fmla v11.4s, v16.4s, v0.s[1]\n"
226 "ldr q16, [x10, #0x90]\n"
227 "fmla v8.4s, v17.4s, v0.s[2]\n"
228 "ldr q17, [x10, #0xa0]\n"
229 "fmla v9.4s, v16.4s, v0.s[2]\n"
230 "ldr q16, [x10, #0xb0]\n"
231 "fmla v10.4s, v17.4s, v0.s[2]\n"
232 "ldr q17, [x10, #0xc0]\n"
233 "fmla v11.4s, v16.4s, v0.s[2]\n"
234 "ldr q16, [x10, #0xd0]\n"
235 "fmla v8.4s, v17.4s, v0.s[3]\n"
236 "ldr q17, [x10, #0xe0]\n"
237 "fmla v9.4s, v16.4s, v0.s[3]\n"
238 "ldr q16, [x10, #0xf0]\n"
239 "sub x27, x27, #0x4\n"
240 "add x26, x26, #0x10\n"
241 "fmla v10.4s, v17.4s, v0.s[3]\n"
242 "fmla v11.4s, v16.4s, v0.s[3]\n"
243 "ldr q0, [x26, #0x0]\n"
245 "add x10, x10, #0x100\n"
246 "ldr q6, [x10, #0x0]\n"
247 "ldr q7, [x10, #0x10]\n"
248 "prfm pldl1keep, [x26, #0x80]\n"
251 "fmla v8.4s, v6.4s, v0.s[0]\n"
252 "ldr q17, [x10, #0x20]\n"
253 "fmla v9.4s, v7.4s, v0.s[0]\n"
254 "ldr q16, [x10, #0x30]\n"
255 "fmla v10.4s, v17.4s, v0.s[0]\n"
256 "ldr q17, [x10, #0x40]\n"
257 "fmla v11.4s, v16.4s, v0.s[0]\n"
258 "ldr q16, [x10, #0x50]\n"
259 "fmla v8.4s, v17.4s, v0.s[1]\n"
260 "ldr q17, [x10, #0x60]\n"
261 "fmla v9.4s, v16.4s, v0.s[1]\n"
262 "ldr q16, [x10, #0x70]\n"
263 "fmla v10.4s, v17.4s, v0.s[1]\n"
264 "ldr q17, [x10, #0x80]\n"
265 "fmla v11.4s, v16.4s, v0.s[1]\n"
266 "ldr q16, [x10, #0x90]\n"
267 "fmla v8.4s, v17.4s, v0.s[2]\n"
268 "ldr q17, [x10, #0xa0]\n"
269 "fmla v9.4s, v16.4s, v0.s[2]\n"
270 "ldr q16, [x10, #0xb0]\n"
271 "fmla v10.4s, v17.4s, v0.s[2]\n"
272 "ldr q17, [x10, #0xc0]\n"
273 "fmla v11.4s, v16.4s, v0.s[2]\n"
274 "ldr q16, [x10, #0xd0]\n"
275 "fmla v8.4s, v17.4s, v0.s[3]\n"
276 "ldr q17, [x10, #0xe0]\n"
277 "fmla v9.4s, v16.4s, v0.s[3]\n"
278 "ldr q16, [x10, #0xf0]\n"
279 "add x26, x26, #0x10\n"
280 "sub x27, x27, #0x4\n"
281 "fmla v10.4s, v17.4s, v0.s[3]\n"
282 "fmla v11.4s, v16.4s, v0.s[3]\n"
283 "prfm pldl1keep, [x26, #0x80]\n"
284 "add x10, x10, #0x100\n"
288 "ldr s18, [x26], #0x4\n"
289 "ldr q16, [x10, #0x0]\n"
290 "fmla v8.4s, v16.4s, v18.s[0]\n"
291 "sub x27, x27, #0x1\n"
292 "ldr q17, [x10, #0x10]\n"
293 "ldr q16, [x10, #0x20]\n"
294 "fmla v9.4s, v17.4s, v18.s[0]\n"
295 "fmla v10.4s, v16.4s, v18.s[0]\n"
296 "ldr q16, [x10, #0x30]\n"
297 "fmla v11.4s, v16.4s, v18.s[0]\n"
298 "add x10, x10, #0x40\n"
301 "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
302 "add x28, x28, #0x1\n"
305 "prfm pstl1keep, [x9, #0x0]\n"
306 "tbz %x[flags], #1, 23f\n"
307 "add x20, %x[args_ptr], %[offset_max]\n"
308 "ld1r { v17.4s }, [x20]\n"
309 "add x20, %x[args_ptr], %[offset_min]\n"
310 "ld1r { v16.4s }, [x20]\n"
311 "fmin v8.4s, v8.4s, v17.4s\n"
312 "fmin v9.4s, v9.4s, v17.4s\n"
313 "fmin v10.4s, v10.4s, v17.4s\n"
314 "fmin v11.4s, v11.4s, v17.4s\n"
315 "fmax v8.4s, v8.4s, v16.4s\n"
316 "fmax v9.4s, v9.4s, v16.4s\n"
317 "fmax v10.4s, v10.4s, v16.4s\n"
318 "fmax v11.4s, v11.4s, v16.4s\n"
323 "st1 { v8.4s }, [x9], #0x10\n"
324 "st1 { v9.4s }, [x9], #0x10\n"
326 "st1 { v10.4s }, [x9], #0x10\n"
328 "str d11, [x9], #0x8\n"
330 "st1 { v11.s }[2], [x9]\n"
334 "str s11, [x9, #0x0]\n"
338 "str d10, [x9], #0x8\n"
340 "st1 { v10.s }[2], [x9]\n"
344 "str s10, [x9, #0x0]\n"
348 "st1 { v8.4s }, [x9], #0x10\n"
350 "str d9, [x9], #0x8\n"
352 "st1 { v9.s }[2], [x9]\n"
356 "str s9, [x9, #0x0]\n"
360 "str d8, [x9], #0x8\n"
362 "st1 { v8.s }[2], [x9]\n"
365 "str s8, [x9, #0x0]\n"
369 "str q8, [x9, #0x0]\n"
370 "str q9, [x9, #0x10]\n"
371 "str q10, [x9, #0x20]\n"
372 "str q11, [x9, #0x30]\n"
373 "add x9, x9, #0x40\n"
375 "subs x11, x11, #0x10\n"
379 "mov x12, %x[bias]\n"
380 "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
381 "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
382 "mov x9, %x[output_ptr]\n"
385 "ldr q8, [x12, #0x0]\n"
386 "ldr q9, [x12, #0x10]\n"
387 "mov v12.16b, v8.16b\n"
388 "mov v13.16b, v9.16b\n"
389 "ldr q10, [x12, #0x20]\n"
390 "ldr q11, [x12, #0x30]\n"
391 "mov v14.16b, v10.16b\n"
392 "mov v15.16b, v11.16b\n"
393 "add x12, x12, #0x40\n"
396 "tbz %x[flags], #0, 46f\n"
397 "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
399 "add x25, x9, x20, LSL #2\n"
402 "ld1 { v8.4s }, [x9], #0x10\n"
403 "ld1 { v12.4s }, [x25], #0x10\n"
404 "ld1 { v9.4s }, [x9], #0x10\n"
405 "ld1 { v13.4s }, [x25], #0x10\n"
407 "ld1 { v10.4s }, [x9], #0x10\n"
408 "ld1 { v14.4s }, [x25], #0x10\n"
410 "ldr d11, [x9], #0x8\n"
411 "ldr d15, [x25], #0x8\n"
414 "ld1 { v11.s }[2], [x9]\n"
415 "ld1 { v15.s }[2], [x25]\n"
420 "ldr s11, [x9, #0x0]\n"
421 "ldr s15, [x25, #0x0]\n"
425 "ldr d10, [x9], #0x8\n"
426 "ldr d14, [x25], #0x8\n"
429 "ld1 { v10.s }[2], [x9]\n"
430 "ld1 { v14.s }[2], [x25]\n"
435 "ldr s10, [x9, #0x0]\n"
436 "ldr s14, [x25, #0x0]\n"
440 "ld1 { v8.4s }, [x9], #0x10\n"
441 "ld1 { v12.4s }, [x25], #0x10\n"
443 "ldr d9, [x9], #0x8\n"
444 "ldr d13, [x25], #0x8\n"
447 "ld1 { v9.s }[2], [x9]\n"
448 "ld1 { v13.s }[2], [x25]\n"
453 "ldr s9, [x9, #0x0]\n"
454 "ldr s13, [x25, #0x0]\n"
458 "ldr d8, [x9], #0x8\n"
459 "ldr d12, [x25], #0x8\n"
462 "ld1 { v8.s }[2], [x9]\n"
463 "ld1 { v12.s }[2], [x25]\n"
466 "ldr s8, [x9, #0x0]\n"
467 "ldr s12, [x25, #0x0]\n"
473 "ldr q8, [x9, #0x0]\n"
474 "ldr q9, [x9, #0x10]\n"
475 "ldr q10, [x9, #0x20]\n"
476 "ldr q11, [x9, #0x30]\n"
477 "ldr q12, [x25, #0x0]\n"
478 "ldr q13, [x25, #0x10]\n"
479 "ldr q14, [x25, #0x20]\n"
480 "ldr q15, [x25, #0x30]\n"
483 "movi v8.16b, #0x0\n"
484 "movi v9.16b, #0x0\n"
485 "movi v10.16b, #0x0\n"
486 "movi v11.16b, #0x0\n"
487 "movi v12.16b, #0x0\n"
488 "movi v13.16b, #0x0\n"
489 "movi v14.16b, #0x0\n"
490 "movi v15.16b, #0x0\n"
494 "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
495 "ldr w27, [x20, x28, LSL #0x2]\n"
496 "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
497 "tbz %x[flags], #3, 49f\n"
498 "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
499 "add x20, x20, x21, LSL #3\n"
500 "ldr x26, [x20, #0x0]\n"
501 "ldr x25, [x20, #0x8]\n"
503 "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
504 "add x26, x26, x20, LSL #2\n"
505 "add x25, x25, x20, LSL #2\n"
508 "mov x26, %x[input_ptr]\n"
509 "add x25, x26, x21, LSL #2\n"
513 "ldr q0, [x26, #0x0]\n"
514 "ldr q1, [x25, #0x0]\n"
516 "ldr q6, [x10, #0x0]\n"
517 "ldr q7, [x10, #0x10]\n"
520 "fmla v8.4s, v6.4s, v0.s[0]\n"
521 "fmla v12.4s, v6.4s, v1.s[0]\n"
522 "ldr q17, [x10, #0x20]\n"
523 "sub x27, x27, #0x4\n"
524 "fmla v9.4s, v7.4s, v0.s[0]\n"
525 "fmla v13.4s, v7.4s, v1.s[0]\n"
526 "ldr q16, [x10, #0x30]\n"
527 "add x26, x26, #0x10\n"
528 "fmla v10.4s, v17.4s, v0.s[0]\n"
529 "fmla v14.4s, v17.4s, v1.s[0]\n"
530 "ldr q17, [x10, #0x40]\n"
531 "add x25, x25, #0x10\n"
532 "fmla v11.4s, v16.4s, v0.s[0]\n"
533 "fmla v15.4s, v16.4s, v1.s[0]\n"
534 "ldr q16, [x10, #0x50]\n"
536 "fmla v8.4s, v17.4s, v0.s[1]\n"
537 "fmla v12.4s, v17.4s, v1.s[1]\n"
538 "ldr q17, [x10, #0x60]\n"
539 "prfm pldl1keep, [x26, #0x80]\n"
540 "fmla v9.4s, v16.4s, v0.s[1]\n"
541 "fmla v13.4s, v16.4s, v1.s[1]\n"
542 "ldr q16, [x10, #0x70]\n"
543 "prfm pldl1keep, [x25, #0x80]\n"
544 "fmla v10.4s, v17.4s, v0.s[1]\n"
545 "fmla v14.4s, v17.4s, v1.s[1]\n"
546 "ldr q17, [x10, #0x80]\n"
547 "fmla v11.4s, v16.4s, v0.s[1]\n"
548 "fmla v15.4s, v16.4s, v1.s[1]\n"
549 "ldr q16, [x10, #0x90]\n"
550 "fmla v8.4s, v17.4s, v0.s[2]\n"
551 "fmla v12.4s, v17.4s, v1.s[2]\n"
552 "ldr q17, [x10, #0xa0]\n"
553 "fmla v9.4s, v16.4s, v0.s[2]\n"
554 "fmla v13.4s, v16.4s, v1.s[2]\n"
555 "ldr q16, [x10, #0xb0]\n"
556 "fmla v10.4s, v17.4s, v0.s[2]\n"
557 "fmla v14.4s, v17.4s, v1.s[2]\n"
558 "ldr q17, [x10, #0xc0]\n"
559 "fmla v11.4s, v16.4s, v0.s[2]\n"
560 "fmla v15.4s, v16.4s, v1.s[2]\n"
561 "ldr q16, [x10, #0xd0]\n"
562 "fmla v8.4s, v17.4s, v0.s[3]\n"
563 "fmla v12.4s, v17.4s, v1.s[3]\n"
564 "ldr q17, [x10, #0xe0]\n"
565 "fmla v9.4s, v16.4s, v0.s[3]\n"
566 "fmla v13.4s, v16.4s, v1.s[3]\n"
567 "ldr q16, [x10, #0xf0]\n"
568 "add x10, x10, #0x100\n"
569 "fmla v10.4s, v17.4s, v0.s[3]\n"
570 "fmla v14.4s, v17.4s, v1.s[3]\n"
571 "ldr q6, [x10, #0x0]\n"
572 "fmla v11.4s, v16.4s, v0.s[3]\n"
573 "ldr q0, [x26, #0x0]\n"
574 "fmla v15.4s, v16.4s, v1.s[3]\n"
575 "ldr q1, [x25, #0x0]\n"
576 "ldr q7, [x10, #0x10]\n"
579 "fmla v8.4s, v6.4s, v0.s[0]\n"
580 "fmla v12.4s, v6.4s, v1.s[0]\n"
581 "ldr q17, [x10, #0x20]\n"
582 "add x26, x26, #0x10\n"
583 "fmla v9.4s, v7.4s, v0.s[0]\n"
584 "fmla v13.4s, v7.4s, v1.s[0]\n"
585 "ldr q16, [x10, #0x30]\n"
586 "add x25, x25, #0x10\n"
587 "fmla v10.4s, v17.4s, v0.s[0]\n"
588 "fmla v14.4s, v17.4s, v1.s[0]\n"
589 "ldr q17, [x10, #0x40]\n"
590 "sub x27, x27, #0x4\n"
591 "fmla v11.4s, v16.4s, v0.s[0]\n"
592 "fmla v15.4s, v16.4s, v1.s[0]\n"
593 "ldr q16, [x10, #0x50]\n"
594 "prfm pldl1keep, [x26, #0x80]\n"
595 "fmla v8.4s, v17.4s, v0.s[1]\n"
596 "fmla v12.4s, v17.4s, v1.s[1]\n"
597 "ldr q17, [x10, #0x60]\n"
598 "prfm pldl1keep, [x25, #0x80]\n"
599 "fmla v9.4s, v16.4s, v0.s[1]\n"
600 "fmla v13.4s, v16.4s, v1.s[1]\n"
601 "ldr q16, [x10, #0x70]\n"
602 "fmla v10.4s, v17.4s, v0.s[1]\n"
603 "fmla v14.4s, v17.4s, v1.s[1]\n"
604 "ldr q17, [x10, #0x80]\n"
605 "fmla v11.4s, v16.4s, v0.s[1]\n"
606 "fmla v15.4s, v16.4s, v1.s[1]\n"
607 "ldr q16, [x10, #0x90]\n"
608 "fmla v8.4s, v17.4s, v0.s[2]\n"
609 "fmla v12.4s, v17.4s, v1.s[2]\n"
610 "ldr q17, [x10, #0xa0]\n"
611 "fmla v9.4s, v16.4s, v0.s[2]\n"
612 "fmla v13.4s, v16.4s, v1.s[2]\n"
613 "ldr q16, [x10, #0xb0]\n"
614 "fmla v10.4s, v17.4s, v0.s[2]\n"
615 "fmla v14.4s, v17.4s, v1.s[2]\n"
616 "ldr q17, [x10, #0xc0]\n"
617 "fmla v11.4s, v16.4s, v0.s[2]\n"
618 "fmla v15.4s, v16.4s, v1.s[2]\n"
619 "ldr q16, [x10, #0xd0]\n"
620 "fmla v8.4s, v17.4s, v0.s[3]\n"
621 "fmla v12.4s, v17.4s, v1.s[3]\n"
622 "ldr q17, [x10, #0xe0]\n"
623 "fmla v9.4s, v16.4s, v0.s[3]\n"
624 "fmla v13.4s, v16.4s, v1.s[3]\n"
625 "ldr q16, [x10, #0xf0]\n"
626 "add x10, x10, #0x100\n"
627 "fmla v10.4s, v17.4s, v0.s[3]\n"
628 "fmla v14.4s, v17.4s, v1.s[3]\n"
629 "fmla v11.4s, v16.4s, v0.s[3]\n"
630 "fmla v15.4s, v16.4s, v1.s[3]\n"
634 "ldr s19, [x26], #0x4\n"
635 "ldr s18, [x25], #0x4\n"
636 "sub x27, x27, #0x1\n"
637 "ldr q17, [x10, #0x0]\n"
638 "ldr q16, [x10, #0x10]\n"
639 "fmla v8.4s, v17.4s, v19.s[0]\n"
640 "fmla v12.4s, v17.4s, v18.s[0]\n"
641 "ldr q17, [x10, #0x20]\n"
642 "fmla v9.4s, v16.4s, v19.s[0]\n"
643 "fmla v13.4s, v16.4s, v18.s[0]\n"
644 "ldr q16, [x10, #0x30]\n"
645 "fmla v10.4s, v17.4s, v19.s[0]\n"
646 "fmla v14.4s, v17.4s, v18.s[0]\n"
647 "add x10, x10, #0x40\n"
648 "fmla v11.4s, v16.4s, v19.s[0]\n"
649 "fmla v15.4s, v16.4s, v18.s[0]\n"
652 "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
653 "add x28, x28, #0x1\n"
656 "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
657 "add x25, x9, x20, LSL #2\n"
658 "prfm pstl1keep, [x9, #0x0]\n"
659 "prfm pstl1keep, [x25, #0x0]\n"
660 "tbz %x[flags], #1, 56f\n"
661 "add x20, %x[args_ptr], %[offset_max]\n"
662 "ld1r { v17.4s }, [x20]\n"
663 "add x20, %x[args_ptr], %[offset_min]\n"
664 "ld1r { v16.4s }, [x20]\n"
665 "fmin v8.4s, v8.4s, v17.4s\n"
666 "fmin v9.4s, v9.4s, v17.4s\n"
667 "fmin v10.4s, v10.4s, v17.4s\n"
668 "fmin v11.4s, v11.4s, v17.4s\n"
669 "fmin v12.4s, v12.4s, v17.4s\n"
670 "fmin v13.4s, v13.4s, v17.4s\n"
671 "fmin v14.4s, v14.4s, v17.4s\n"
672 "fmin v15.4s, v15.4s, v17.4s\n"
673 "fmax v8.4s, v8.4s, v16.4s\n"
674 "fmax v9.4s, v9.4s, v16.4s\n"
675 "fmax v10.4s, v10.4s, v16.4s\n"
676 "fmax v11.4s, v11.4s, v16.4s\n"
677 "fmax v12.4s, v12.4s, v16.4s\n"
678 "fmax v13.4s, v13.4s, v16.4s\n"
679 "fmax v14.4s, v14.4s, v16.4s\n"
680 "fmax v15.4s, v15.4s, v16.4s\n"
685 "st1 { v8.4s }, [x9], #0x10\n"
686 "st1 { v9.4s }, [x9], #0x10\n"
687 "st1 { v12.4s }, [x25], #0x10\n"
688 "st1 { v13.4s }, [x25], #0x10\n"
690 "st1 { v10.4s }, [x9], #0x10\n"
691 "st1 { v14.4s }, [x25], #0x10\n"
693 "str d11, [x9], #0x8\n"
694 "str d15, [x25], #0x8\n"
696 "st1 { v11.s }[2], [x9]\n"
697 "st1 { v15.s }[2], [x25]\n"
701 "str s11, [x9, #0x0]\n"
702 "str s15, [x25, #0x0]\n"
706 "str d10, [x9], #0x8\n"
707 "str d14, [x25], #0x8\n"
709 "st1 { v10.s }[2], [x9]\n"
710 "st1 { v14.s }[2], [x25]\n"
714 "str s10, [x9, #0x0]\n"
715 "str s14, [x25, #0x0]\n"
719 "st1 { v8.4s }, [x9], #0x10\n"
720 "st1 { v12.4s }, [x25], #0x10\n"
722 "str d9, [x9], #0x8\n"
723 "str d13, [x25], #0x8\n"
725 "st1 { v9.s }[2], [x9]\n"
726 "st1 { v13.s }[2], [x25]\n"
730 "str s9, [x9, #0x0]\n"
731 "str s13, [x25, #0x0]\n"
735 "str d8, [x9], #0x8\n"
736 "str d12, [x25], #0x8\n"
738 "st1 { v8.s }[2], [x9]\n"
739 "st1 { v12.s }[2], [x25]\n"
742 "str s8, [x9, #0x0]\n"
743 "str s12, [x25, #0x0]\n"
747 "str q8, [x9, #0x0]\n"
748 "str q9, [x9, #0x10]\n"
749 "str q10, [x9, #0x20]\n"
750 "str q11, [x9, #0x30]\n"
751 "add x9, x9, #0x40\n"
752 "str q12, [x25, #0x0]\n"
753 "str q13, [x25, #0x10]\n"
754 "str q14, [x25, #0x20]\n"
755 "str q15, [x25, #0x30]\n"
757 "subs x11, x11, #0x10\n"
761 "mov x12, %x[bias]\n"
762 "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
763 "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
764 "mov x9, %x[output_ptr]\n"
767 "ldr q8, [x12, #0x0]\n"
768 "ldr q9, [x12, #0x10]\n"
769 "mov v12.16b, v8.16b\n"
770 "mov v13.16b, v9.16b\n"
771 "ldr q10, [x12, #0x20]\n"
772 "ldr q11, [x12, #0x30]\n"
773 "mov v14.16b, v10.16b\n"
774 "mov v15.16b, v11.16b\n"
775 "mov v16.16b, v8.16b\n"
776 "mov v17.16b, v9.16b\n"
777 "add x12, x12, #0x40\n"
778 "mov v18.16b, v10.16b\n"
779 "mov v19.16b, v11.16b\n"
782 "tbz %x[flags], #0, 79f\n"
783 "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
784 "add x25, x9, x20, LSL #2\n"
786 "add x24, x25, x20, LSL #2\n"
789 "ld1 { v8.4s }, [x9], #0x10\n"
790 "ld1 { v12.4s }, [x25], #0x10\n"
791 "ld1 { v16.4s }, [x24], #0x10\n"
792 "ld1 { v9.4s }, [x9], #0x10\n"
793 "ld1 { v13.4s }, [x25], #0x10\n"
794 "ld1 { v17.4s }, [x24], #0x10\n"
796 "ld1 { v10.4s }, [x9], #0x10\n"
797 "ld1 { v14.4s }, [x25], #0x10\n"
798 "ld1 { v18.4s }, [x24], #0x10\n"
800 "ldr d11, [x9], #0x8\n"
801 "ldr d15, [x25], #0x8\n"
803 "ldr d19, [x24], #0x8\n"
805 "ld1 { v11.s }[2], [x9]\n"
806 "ld1 { v15.s }[2], [x25]\n"
807 "ld1 { v19.s }[2], [x24]\n"
812 "ldr s11, [x9, #0x0]\n"
813 "ldr s15, [x25, #0x0]\n"
814 "ldr s19, [x24, #0x0]\n"
818 "ldr d10, [x9], #0x8\n"
819 "ldr d14, [x25], #0x8\n"
821 "ldr d18, [x24], #0x8\n"
823 "ld1 { v10.s }[2], [x9]\n"
824 "ld1 { v14.s }[2], [x25]\n"
825 "ld1 { v18.s }[2], [x24]\n"
830 "ldr s10, [x9, #0x0]\n"
831 "ldr s14, [x25, #0x0]\n"
832 "ldr s18, [x24, #0x0]\n"
836 "ld1 { v8.4s }, [x9], #0x10\n"
837 "ld1 { v12.4s }, [x25], #0x10\n"
838 "ld1 { v16.4s }, [x24], #0x10\n"
840 "ldr d9, [x9], #0x8\n"
841 "ldr d13, [x25], #0x8\n"
843 "ldr d17, [x24], #0x8\n"
845 "ld1 { v9.s }[2], [x9]\n"
846 "ld1 { v13.s }[2], [x25]\n"
847 "ld1 { v17.s }[2], [x24]\n"
852 "ldr s9, [x9, #0x0]\n"
853 "ldr s13, [x25, #0x0]\n"
854 "ldr s17, [x24, #0x0]\n"
858 "ldr d8, [x9], #0x8\n"
859 "ldr d12, [x25], #0x8\n"
861 "ldr d16, [x24], #0x8\n"
863 "ld1 { v8.s }[2], [x9]\n"
864 "ld1 { v12.s }[2], [x25]\n"
865 "ld1 { v16.s }[2], [x24]\n"
868 "ldr s8, [x9, #0x0]\n"
869 "ldr s12, [x25, #0x0]\n"
871 "ldr s16, [x24, #0x0]\n"
876 "ldr q8, [x9, #0x0]\n"
877 "ldr q9, [x9, #0x10]\n"
878 "ldr q10, [x9, #0x20]\n"
879 "ldr q11, [x9, #0x30]\n"
880 "ldr q12, [x25, #0x0]\n"
881 "ldr q13, [x25, #0x10]\n"
882 "ldr q14, [x25, #0x20]\n"
883 "ldr q15, [x25, #0x30]\n"
884 "ldr q16, [x24, #0x0]\n"
885 "ldr q17, [x24, #0x10]\n"
886 "ldr q18, [x24, #0x20]\n"
887 "ldr q19, [x24, #0x30]\n"
890 "movi v8.16b, #0x0\n"
891 "movi v9.16b, #0x0\n"
892 "movi v10.16b, #0x0\n"
893 "movi v11.16b, #0x0\n"
894 "movi v12.16b, #0x0\n"
895 "movi v13.16b, #0x0\n"
896 "movi v14.16b, #0x0\n"
897 "movi v15.16b, #0x0\n"
898 "movi v16.16b, #0x0\n"
899 "movi v17.16b, #0x0\n"
900 "movi v18.16b, #0x0\n"
901 "movi v19.16b, #0x0\n"
905 "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
906 "ldr w27, [x20, x28, LSL #0x2]\n"
907 "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
908 "tbz %x[flags], #3, 82f\n"
909 "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
910 "add x20, x20, x21, LSL #3\n"
911 "ldr x26, [x20, #0x0]\n"
912 "ldr x25, [x20, #0x8]\n"
913 "ldr x24, [x20, #0x10]\n"
915 "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
916 "add x26, x26, x20, LSL #2\n"
917 "add x25, x25, x20, LSL #2\n"
918 "add x24, x24, x20, LSL #2\n"
921 "mov x26, %x[input_ptr]\n"
922 "add x25, x26, x21, LSL #2\n"
923 "add x24, x25, x21, LSL #2\n"
927 "ldr q0, [x26, #0x0]\n"
928 "ldr q1, [x25, #0x0]\n"
930 "ldr q2, [x24, #0x0]\n"
931 "ldr q6, [x10, #0x0]\n"
932 "ldr q7, [x10, #0x10]\n"
935 "fmla v8.4s, v6.4s, v0.s[0]\n"
936 "fmla v12.4s, v6.4s, v1.s[0]\n"
937 "sub x27, x27, #0x4\n"
938 "add x26, x26, #0x10\n"
939 "fmla v16.4s, v6.4s, v2.s[0]\n"
940 "ldr q21, [x10, #0x20]\n"
941 "fmla v9.4s, v7.4s, v0.s[0]\n"
942 "add x25, x25, #0x10\n"
943 "fmla v13.4s, v7.4s, v1.s[0]\n"
944 "fmla v17.4s, v7.4s, v2.s[0]\n"
945 "ldr q20, [x10, #0x30]\n"
946 "add x24, x24, #0x10\n"
947 "fmla v10.4s, v21.4s, v0.s[0]\n"
948 "fmla v14.4s, v21.4s, v1.s[0]\n"
950 "prfm pldl1keep, [x26, #0x80]\n"
951 "fmla v18.4s, v21.4s, v2.s[0]\n"
952 "ldr q21, [x10, #0x40]\n"
953 "fmla v11.4s, v20.4s, v0.s[0]\n"
954 "prfm pldl1keep, [x25, #0x80]\n"
955 "fmla v15.4s, v20.4s, v1.s[0]\n"
956 "fmla v19.4s, v20.4s, v2.s[0]\n"
957 "ldr q20, [x10, #0x50]\n"
958 "prfm pldl1keep, [x24, #0x80]\n"
959 "fmla v8.4s, v21.4s, v0.s[1]\n"
960 "fmla v12.4s, v21.4s, v1.s[1]\n"
961 "fmla v16.4s, v21.4s, v2.s[1]\n"
962 "ldr q21, [x10, #0x60]\n"
963 "fmla v9.4s, v20.4s, v0.s[1]\n"
964 "fmla v13.4s, v20.4s, v1.s[1]\n"
965 "fmla v17.4s, v20.4s, v2.s[1]\n"
966 "ldr q20, [x10, #0x70]\n"
967 "fmla v10.4s, v21.4s, v0.s[1]\n"
968 "fmla v14.4s, v21.4s, v1.s[1]\n"
969 "fmla v18.4s, v21.4s, v2.s[1]\n"
970 "ldr q21, [x10, #0x80]\n"
971 "fmla v11.4s, v20.4s, v0.s[1]\n"
972 "fmla v15.4s, v20.4s, v1.s[1]\n"
973 "fmla v19.4s, v20.4s, v2.s[1]\n"
974 "ldr q20, [x10, #0x90]\n"
975 "fmla v8.4s, v21.4s, v0.s[2]\n"
976 "fmla v12.4s, v21.4s, v1.s[2]\n"
977 "fmla v16.4s, v21.4s, v2.s[2]\n"
978 "ldr q21, [x10, #0xa0]\n"
979 "fmla v9.4s, v20.4s, v0.s[2]\n"
980 "fmla v13.4s, v20.4s, v1.s[2]\n"
981 "fmla v17.4s, v20.4s, v2.s[2]\n"
982 "ldr q20, [x10, #0xb0]\n"
983 "fmla v10.4s, v21.4s, v0.s[2]\n"
984 "fmla v14.4s, v21.4s, v1.s[2]\n"
985 "fmla v18.4s, v21.4s, v2.s[2]\n"
986 "ldr q21, [x10, #0xc0]\n"
987 "fmla v11.4s, v20.4s, v0.s[2]\n"
988 "fmla v15.4s, v20.4s, v1.s[2]\n"
989 "fmla v19.4s, v20.4s, v2.s[2]\n"
990 "ldr q20, [x10, #0xd0]\n"
991 "fmla v8.4s, v21.4s, v0.s[3]\n"
992 "fmla v12.4s, v21.4s, v1.s[3]\n"
993 "fmla v16.4s, v21.4s, v2.s[3]\n"
994 "ldr q21, [x10, #0xe0]\n"
995 "fmla v9.4s, v20.4s, v0.s[3]\n"
996 "fmla v13.4s, v20.4s, v1.s[3]\n"
997 "fmla v17.4s, v20.4s, v2.s[3]\n"
998 "ldr q20, [x10, #0xf0]\n"
999 "add x10, x10, #0x100\n"
1000 "fmla v10.4s, v21.4s, v0.s[3]\n"
1001 "fmla v14.4s, v21.4s, v1.s[3]\n"
1002 "fmla v18.4s, v21.4s, v2.s[3]\n"
1003 "ldr q6, [x10, #0x0]\n"
1004 "fmla v11.4s, v20.4s, v0.s[3]\n"
1005 "ldr q0, [x26, #0x0]\n"
1006 "fmla v15.4s, v20.4s, v1.s[3]\n"
1007 "ldr q1, [x25, #0x0]\n"
1008 "fmla v19.4s, v20.4s, v2.s[3]\n"
1009 "ldr q2, [x24, #0x0]\n"
1010 "ldr q7, [x10, #0x10]\n"
1013 "fmla v8.4s, v6.4s, v0.s[0]\n"
1014 "fmla v12.4s, v6.4s, v1.s[0]\n"
1015 "add x26, x26, #0x10\n"
1016 "add x25, x25, #0x10\n"
1017 "fmla v16.4s, v6.4s, v2.s[0]\n"
1018 "ldr q21, [x10, #0x20]\n"
1019 "fmla v9.4s, v7.4s, v0.s[0]\n"
1020 "add x24, x24, #0x10\n"
1021 "fmla v13.4s, v7.4s, v1.s[0]\n"
1022 "fmla v17.4s, v7.4s, v2.s[0]\n"
1023 "ldr q20, [x10, #0x30]\n"
1024 "sub x27, x27, #0x4\n"
1025 "fmla v10.4s, v21.4s, v0.s[0]\n"
1026 "fmla v14.4s, v21.4s, v1.s[0]\n"
1027 "prfm pldl1keep, [x26, #0x80]\n"
1028 "prfm pldl1keep, [x25, #0x80]\n"
1029 "fmla v18.4s, v21.4s, v2.s[0]\n"
1030 "ldr q21, [x10, #0x40]\n"
1031 "fmla v11.4s, v20.4s, v0.s[0]\n"
1032 "prfm pldl1keep, [x24, #0x80]\n"
1033 "fmla v15.4s, v20.4s, v1.s[0]\n"
1034 "fmla v19.4s, v20.4s, v2.s[0]\n"
1035 "ldr q20, [x10, #0x50]\n"
1036 "fmla v8.4s, v21.4s, v0.s[1]\n"
1037 "fmla v12.4s, v21.4s, v1.s[1]\n"
1038 "fmla v16.4s, v21.4s, v2.s[1]\n"
1039 "ldr q21, [x10, #0x60]\n"
1040 "fmla v9.4s, v20.4s, v0.s[1]\n"
1041 "fmla v13.4s, v20.4s, v1.s[1]\n"
1042 "fmla v17.4s, v20.4s, v2.s[1]\n"
1043 "ldr q20, [x10, #0x70]\n"
1044 "fmla v10.4s, v21.4s, v0.s[1]\n"
1045 "fmla v14.4s, v21.4s, v1.s[1]\n"
1046 "fmla v18.4s, v21.4s, v2.s[1]\n"
1047 "ldr q21, [x10, #0x80]\n"
1048 "fmla v11.4s, v20.4s, v0.s[1]\n"
1049 "fmla v15.4s, v20.4s, v1.s[1]\n"
1050 "fmla v19.4s, v20.4s, v2.s[1]\n"
1051 "ldr q20, [x10, #0x90]\n"
1052 "fmla v8.4s, v21.4s, v0.s[2]\n"
1053 "fmla v12.4s, v21.4s, v1.s[2]\n"
1054 "fmla v16.4s, v21.4s, v2.s[2]\n"
1055 "ldr q21, [x10, #0xa0]\n"
1056 "fmla v9.4s, v20.4s, v0.s[2]\n"
1057 "fmla v13.4s, v20.4s, v1.s[2]\n"
1058 "fmla v17.4s, v20.4s, v2.s[2]\n"
1059 "ldr q20, [x10, #0xb0]\n"
1060 "fmla v10.4s, v21.4s, v0.s[2]\n"
1061 "fmla v14.4s, v21.4s, v1.s[2]\n"
1062 "fmla v18.4s, v21.4s, v2.s[2]\n"
1063 "ldr q21, [x10, #0xc0]\n"
1064 "fmla v11.4s, v20.4s, v0.s[2]\n"
1065 "fmla v15.4s, v20.4s, v1.s[2]\n"
1066 "fmla v19.4s, v20.4s, v2.s[2]\n"
1067 "ldr q20, [x10, #0xd0]\n"
1068 "fmla v8.4s, v21.4s, v0.s[3]\n"
1069 "fmla v12.4s, v21.4s, v1.s[3]\n"
1070 "fmla v16.4s, v21.4s, v2.s[3]\n"
1071 "ldr q21, [x10, #0xe0]\n"
1072 "fmla v9.4s, v20.4s, v0.s[3]\n"
1073 "fmla v13.4s, v20.4s, v1.s[3]\n"
1074 "fmla v17.4s, v20.4s, v2.s[3]\n"
1075 "ldr q20, [x10, #0xf0]\n"
1076 "add x10, x10, #0x100\n"
1077 "fmla v10.4s, v21.4s, v0.s[3]\n"
1078 "fmla v14.4s, v21.4s, v1.s[3]\n"
1079 "fmla v18.4s, v21.4s, v2.s[3]\n"
1080 "fmla v11.4s, v20.4s, v0.s[3]\n"
1081 "fmla v15.4s, v20.4s, v1.s[3]\n"
1082 "fmla v19.4s, v20.4s, v2.s[3]\n"
1086 "ldr s24, [x26], #0x4\n"
1087 "ldr s23, [x25], #0x4\n"
1088 "sub x27, x27, #0x1\n"
1089 "ldr s22, [x24], #0x4\n"
1090 "ldr q21, [x10, #0x0]\n"
1091 "fmla v8.4s, v21.4s, v24.s[0]\n"
1092 "fmla v12.4s, v21.4s, v23.s[0]\n"
1093 "ldr q20, [x10, #0x10]\n"
1094 "fmla v16.4s, v21.4s, v22.s[0]\n"
1095 "ldr q21, [x10, #0x20]\n"
1096 "fmla v9.4s, v20.4s, v24.s[0]\n"
1097 "fmla v13.4s, v20.4s, v23.s[0]\n"
1098 "fmla v17.4s, v20.4s, v22.s[0]\n"
1099 "ldr q20, [x10, #0x30]\n"
1100 "add x10, x10, #0x40\n"
1101 "fmla v10.4s, v21.4s, v24.s[0]\n"
1102 "fmla v14.4s, v21.4s, v23.s[0]\n"
1103 "fmla v18.4s, v21.4s, v22.s[0]\n"
1104 "fmla v11.4s, v20.4s, v24.s[0]\n"
1105 "fmla v15.4s, v20.4s, v23.s[0]\n"
1106 "fmla v19.4s, v20.4s, v22.s[0]\n"
1109 "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
1110 "add x28, x28, #0x1\n"
1113 "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
1114 "add x25, x9, x20, LSL #2\n"
1115 "add x24, x25, x20, LSL #2\n"
1116 "prfm pstl1keep, [x9, #0x0]\n"
1117 "prfm pstl1keep, [x25, #0x0]\n"
1118 "prfm pstl1keep, [x24, #0x0]\n"
1119 "tbz %x[flags], #1, 89f\n"
1120 "add x20, %x[args_ptr], %[offset_max]\n"
1121 "ld1r { v21.4s }, [x20]\n"
1122 "add x20, %x[args_ptr], %[offset_min]\n"
1123 "ld1r { v20.4s }, [x20]\n"
1124 "fmin v8.4s, v8.4s, v21.4s\n"
1125 "fmin v9.4s, v9.4s, v21.4s\n"
1126 "fmin v10.4s, v10.4s, v21.4s\n"
1127 "fmin v11.4s, v11.4s, v21.4s\n"
1128 "fmin v12.4s, v12.4s, v21.4s\n"
1129 "fmin v13.4s, v13.4s, v21.4s\n"
1130 "fmin v14.4s, v14.4s, v21.4s\n"
1131 "fmin v15.4s, v15.4s, v21.4s\n"
1132 "fmin v16.4s, v16.4s, v21.4s\n"
1133 "fmin v17.4s, v17.4s, v21.4s\n"
1134 "fmin v18.4s, v18.4s, v21.4s\n"
1135 "fmin v19.4s, v19.4s, v21.4s\n"
1136 "fmax v8.4s, v8.4s, v20.4s\n"
1137 "fmax v9.4s, v9.4s, v20.4s\n"
1138 "fmax v10.4s, v10.4s, v20.4s\n"
1139 "fmax v11.4s, v11.4s, v20.4s\n"
1140 "fmax v12.4s, v12.4s, v20.4s\n"
1141 "fmax v13.4s, v13.4s, v20.4s\n"
1142 "fmax v14.4s, v14.4s, v20.4s\n"
1143 "fmax v15.4s, v15.4s, v20.4s\n"
1144 "fmax v16.4s, v16.4s, v20.4s\n"
1145 "fmax v17.4s, v17.4s, v20.4s\n"
1146 "fmax v18.4s, v18.4s, v20.4s\n"
1147 "fmax v19.4s, v19.4s, v20.4s\n"
1151 "tbz x11, #3, 93f\n"
1152 "st1 { v8.4s }, [x9], #0x10\n"
1153 "st1 { v9.4s }, [x9], #0x10\n"
1154 "st1 { v12.4s }, [x25], #0x10\n"
1155 "st1 { v13.4s }, [x25], #0x10\n"
1156 "st1 { v16.4s }, [x24], #0x10\n"
1157 "st1 { v17.4s }, [x24], #0x10\n"
1158 "tbz x11, #2, 91f\n"
1159 "st1 { v10.4s }, [x9], #0x10\n"
1160 "st1 { v14.4s }, [x25], #0x10\n"
1161 "st1 { v18.4s }, [x24], #0x10\n"
1162 "tbz x11, #1, 90f\n"
1163 "str d11, [x9], #0x8\n"
1164 "str d15, [x25], #0x8\n"
1165 "str d19, [x24], #0x8\n"
1166 "tbz x11, #0, 97f\n"
1167 "st1 { v11.s }[2], [x9]\n"
1168 "st1 { v15.s }[2], [x25]\n"
1169 "st1 { v19.s }[2], [x24]\n"
1172 "tbz x11, #0, 97f\n"
1173 "str s11, [x9, #0x0]\n"
1174 "str s15, [x25, #0x0]\n"
1175 "str s19, [x24, #0x0]\n"
1178 "tbz x11, #1, 92f\n"
1179 "str d10, [x9], #0x8\n"
1180 "str d14, [x25], #0x8\n"
1181 "str d18, [x24], #0x8\n"
1182 "tbz x11, #0, 97f\n"
1183 "st1 { v10.s }[2], [x9]\n"
1184 "st1 { v14.s }[2], [x25]\n"
1185 "st1 { v18.s }[2], [x24]\n"
1188 "tbz x11, #0, 97f\n"
1189 "str s10, [x9, #0x0]\n"
1190 "str s14, [x25, #0x0]\n"
1191 "str s18, [x24, #0x0]\n"
1194 "tbz x11, #2, 95f\n"
1195 "st1 { v8.4s }, [x9], #0x10\n"
1196 "st1 { v12.4s }, [x25], #0x10\n"
1197 "st1 { v16.4s }, [x24], #0x10\n"
1198 "tbz x11, #1, 94f\n"
1199 "str d9, [x9], #0x8\n"
1200 "str d13, [x25], #0x8\n"
1201 "str d17, [x24], #0x8\n"
1202 "tbz x11, #0, 97f\n"
1203 "st1 { v9.s }[2], [x9]\n"
1204 "st1 { v13.s }[2], [x25]\n"
1205 "st1 { v17.s }[2], [x24]\n"
1208 "tbz x11, #0, 97f\n"
1209 "str s9, [x9, #0x0]\n"
1210 "str s13, [x25, #0x0]\n"
1211 "str s17, [x24, #0x0]\n"
1214 "tbz x11, #1, 96f\n"
1215 "str d8, [x9], #0x8\n"
1216 "str d12, [x25], #0x8\n"
1217 "str d16, [x24], #0x8\n"
1218 "tbz x11, #0, 97f\n"
1219 "st1 { v8.s }[2], [x9]\n"
1220 "st1 { v12.s }[2], [x25]\n"
1221 "st1 { v16.s }[2], [x24]\n"
1224 "str s8, [x9, #0x0]\n"
1225 "str s12, [x25, #0x0]\n"
1226 "str s16, [x24, #0x0]\n"
1230 "str q8, [x9, #0x0]\n"
1231 "str q9, [x9, #0x10]\n"
1232 "str q10, [x9, #0x20]\n"
1233 "str q11, [x9, #0x30]\n"
1234 "add x9, x9, #0x40\n"
1235 "str q12, [x25, #0x0]\n"
1236 "str q13, [x25, #0x10]\n"
1237 "str q14, [x25, #0x20]\n"
1238 "str q15, [x25, #0x30]\n"
1239 "str q16, [x24, #0x0]\n"
1240 "str q17, [x24, #0x10]\n"
1241 "str q18, [x24, #0x20]\n"
1242 "str q19, [x24, #0x30]\n"
1244 "subs x11, x11, #0x10\n"
1248 "mov x12, %x[bias]\n"
1249 "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
1250 "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
1251 "mov x9, %x[output_ptr]\n"
1254 "ldr q8, [x12, #0x0]\n"
1255 "ldr q9, [x12, #0x10]\n"
1256 "mov v12.16b, v8.16b\n"
1257 "mov v13.16b, v9.16b\n"
1258 "ldr q10, [x12, #0x20]\n"
1259 "ldr q11, [x12, #0x30]\n"
1260 "mov v14.16b, v10.16b\n"
1261 "mov v15.16b, v11.16b\n"
1262 "mov v16.16b, v8.16b\n"
1263 "mov v17.16b, v9.16b\n"
1264 "add x12, x12, #0x40\n"
1265 "mov v18.16b, v10.16b\n"
1266 "mov v19.16b, v11.16b\n"
1267 "mov v20.16b, v8.16b\n"
1268 "mov v21.16b, v9.16b\n"
1269 "mov v22.16b, v10.16b\n"
1270 "mov v23.16b, v11.16b\n"
1273 "tbz %x[flags], #0, 112f\n"
1274 "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
1275 "add x25, x9, x20, LSL #2\n"
1276 "add x24, x25, x20, LSL #2\n"
1278 "add x23, x24, x20, LSL #2\n"
1280 "tbz x11, #3, 106f\n"
1281 "ld1 { v8.4s }, [x9], #0x10\n"
1282 "ld1 { v12.4s }, [x25], #0x10\n"
1283 "ld1 { v16.4s }, [x24], #0x10\n"
1284 "ld1 { v20.4s }, [x23], #0x10\n"
1285 "ld1 { v9.4s }, [x9], #0x10\n"
1286 "ld1 { v13.4s }, [x25], #0x10\n"
1287 "ld1 { v17.4s }, [x24], #0x10\n"
1288 "ld1 { v21.4s }, [x23], #0x10\n"
1289 "tbz x11, #2, 104f\n"
1290 "ld1 { v10.4s }, [x9], #0x10\n"
1291 "ld1 { v14.4s }, [x25], #0x10\n"
1292 "ld1 { v18.4s }, [x24], #0x10\n"
1293 "ld1 { v22.4s }, [x23], #0x10\n"
1294 "tbz x11, #1, 103f\n"
1295 "ldr d11, [x9], #0x8\n"
1296 "ldr d15, [x25], #0x8\n"
1298 "ldr d19, [x24], #0x8\n"
1299 "ldr d23, [x23], #0x8\n"
1300 "tbz x11, #0, 110f\n"
1301 "ld1 { v11.s }[2], [x9]\n"
1302 "ld1 { v15.s }[2], [x25]\n"
1303 "ld1 { v19.s }[2], [x24]\n"
1304 "ld1 { v23.s }[2], [x23]\n"
1308 "tbz x11, #0, 110f\n"
1309 "ldr s11, [x9, #0x0]\n"
1310 "ldr s15, [x25, #0x0]\n"
1311 "ldr s19, [x24, #0x0]\n"
1312 "ldr s23, [x23, #0x0]\n"
1315 "tbz x11, #1, 105f\n"
1316 "ldr d10, [x9], #0x8\n"
1317 "ldr d14, [x25], #0x8\n"
1319 "ldr d18, [x24], #0x8\n"
1320 "ldr d22, [x23], #0x8\n"
1321 "tbz x11, #0, 110f\n"
1322 "ld1 { v10.s }[2], [x9]\n"
1323 "ld1 { v14.s }[2], [x25]\n"
1324 "ld1 { v18.s }[2], [x24]\n"
1325 "ld1 { v22.s }[2], [x23]\n"
1329 "tbz x11, #0, 110f\n"
1330 "ldr s10, [x9, #0x0]\n"
1331 "ldr s14, [x25, #0x0]\n"
1332 "ldr s18, [x24, #0x0]\n"
1333 "ldr s22, [x23, #0x0]\n"
1336 "tbz x11, #2, 108f\n"
1337 "ld1 { v8.4s }, [x9], #0x10\n"
1338 "ld1 { v12.4s }, [x25], #0x10\n"
1339 "ld1 { v16.4s }, [x24], #0x10\n"
1340 "ld1 { v20.4s }, [x23], #0x10\n"
1341 "tbz x11, #1, 107f\n"
1342 "ldr d9, [x9], #0x8\n"
1343 "ldr d13, [x25], #0x8\n"
1345 "ldr d17, [x24], #0x8\n"
1346 "ldr d21, [x23], #0x8\n"
1347 "tbz x11, #0, 110f\n"
1348 "ld1 { v9.s }[2], [x9]\n"
1349 "ld1 { v13.s }[2], [x25]\n"
1350 "ld1 { v17.s }[2], [x24]\n"
1351 "ld1 { v21.s }[2], [x23]\n"
1355 "tbz x11, #0, 110f\n"
1356 "ldr s9, [x9, #0x0]\n"
1357 "ldr s13, [x25, #0x0]\n"
1358 "ldr s17, [x24, #0x0]\n"
1359 "ldr s21, [x23, #0x0]\n"
1362 "tbz x11, #1, 109f\n"
1363 "ldr d8, [x9], #0x8\n"
1364 "ldr d12, [x25], #0x8\n"
1366 "ldr d16, [x24], #0x8\n"
1367 "ldr d20, [x23], #0x8\n"
1368 "tbz x11, #0, 110f\n"
1369 "ld1 { v8.s }[2], [x9]\n"
1370 "ld1 { v12.s }[2], [x25]\n"
1371 "ld1 { v16.s }[2], [x24]\n"
1372 "ld1 { v20.s }[2], [x23]\n"
1375 "ldr s8, [x9, #0x0]\n"
1376 "ldr s12, [x25, #0x0]\n"
1378 "ldr s16, [x24, #0x0]\n"
1379 "ldr s20, [x23, #0x0]\n"
1384 "ldr q8, [x9, #0x0]\n"
1385 "ldr q9, [x9, #0x10]\n"
1386 "ldr q10, [x9, #0x20]\n"
1387 "ldr q11, [x9, #0x30]\n"
1388 "ldr q12, [x25, #0x0]\n"
1389 "ldr q13, [x25, #0x10]\n"
1390 "ldr q14, [x25, #0x20]\n"
1391 "ldr q15, [x25, #0x30]\n"
1392 "ldr q16, [x24, #0x0]\n"
1393 "ldr q17, [x24, #0x10]\n"
1394 "ldr q18, [x24, #0x20]\n"
1395 "ldr q19, [x24, #0x30]\n"
1396 "ldr q20, [x23, #0x0]\n"
1397 "ldr q21, [x23, #0x10]\n"
1398 "ldr q22, [x23, #0x20]\n"
1399 "ldr q23, [x23, #0x30]\n"
1402 "movi v8.16b, #0x0\n"
1403 "movi v9.16b, #0x0\n"
1404 "movi v10.16b, #0x0\n"
1405 "movi v11.16b, #0x0\n"
1406 "movi v12.16b, #0x0\n"
1407 "movi v13.16b, #0x0\n"
1408 "movi v14.16b, #0x0\n"
1409 "movi v15.16b, #0x0\n"
1410 "movi v16.16b, #0x0\n"
1411 "movi v17.16b, #0x0\n"
1412 "movi v18.16b, #0x0\n"
1413 "movi v19.16b, #0x0\n"
1414 "movi v20.16b, #0x0\n"
1415 "movi v21.16b, #0x0\n"
1416 "movi v22.16b, #0x0\n"
1417 "movi v23.16b, #0x0\n"
1421 "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
1422 "ldr w27, [x20, x28, LSL #0x2]\n"
1423 "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
1424 "tbz %x[flags], #3, 115f\n"
1425 "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
1426 "add x20, x20, x21, LSL #3\n"
1427 "ldr x26, [x20, #0x0]\n"
1428 "ldr x25, [x20, #0x8]\n"
1429 "ldr x24, [x20, #0x10]\n"
1430 "ldr x23, [x20, #0x18]\n"
1432 "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
1433 "add x26, x26, x20, LSL #2\n"
1434 "add x25, x25, x20, LSL #2\n"
1435 "add x24, x24, x20, LSL #2\n"
1436 "add x23, x23, x20, LSL #2\n"
1439 "mov x26, %x[input_ptr]\n"
1440 "add x25, x26, x21, LSL #2\n"
1441 "add x24, x25, x21, LSL #2\n"
1442 "add x23, x24, x21, LSL #2\n"
1446 "ldr q0, [x26, #0x0]\n"
1447 "ldr q1, [x25, #0x0]\n"
1449 "ldr q2, [x24, #0x0]\n"
1450 "ldr q3, [x23, #0x0]\n"
1451 "ldr q6, [x10, #0x0]\n"
1452 "ldr q7, [x10, #0x10]\n"
1455 "fmla v8.4s, v6.4s, v0.s[0]\n"
1456 "fmla v12.4s, v6.4s, v1.s[0]\n"
1457 "sub x27, x27, #0x4\n"
1458 "add x26, x26, #0x10\n"
1459 "fmla v16.4s, v6.4s, v2.s[0]\n"
1460 "fmla v20.4s, v6.4s, v3.s[0]\n"
1461 "ldr q25, [x10, #0x20]\n"
1462 "add x25, x25, #0x10\n"
1463 "fmla v9.4s, v7.4s, v0.s[0]\n"
1464 "fmla v13.4s, v7.4s, v1.s[0]\n"
1465 "add x24, x24, #0x10\n"
1466 "add x23, x23, #0x10\n"
1467 "fmla v17.4s, v7.4s, v2.s[0]\n"
1468 "fmla v21.4s, v7.4s, v3.s[0]\n"
1469 "ldr q24, [x10, #0x30]\n"
1471 "fmla v10.4s, v25.4s, v0.s[0]\n"
1472 "fmla v14.4s, v25.4s, v1.s[0]\n"
1473 "prfm pldl1keep, [x26, #0x80]\n"
1474 "prfm pldl1keep, [x25, #0x80]\n"
1475 "fmla v18.4s, v25.4s, v2.s[0]\n"
1476 "fmla v22.4s, v25.4s, v3.s[0]\n"
1477 "ldr q25, [x10, #0x40]\n"
1478 "prfm pldl1keep, [x24, #0x80]\n"
1479 "fmla v11.4s, v24.4s, v0.s[0]\n"
1480 "fmla v15.4s, v24.4s, v1.s[0]\n"
1481 "prfm pldl1keep, [x23, #0x80]\n"
1482 "fmla v19.4s, v24.4s, v2.s[0]\n"
1483 "fmla v23.4s, v24.4s, v3.s[0]\n"
1484 "ldr q24, [x10, #0x50]\n"
1485 "fmla v8.4s, v25.4s, v0.s[1]\n"
1486 "fmla v12.4s, v25.4s, v1.s[1]\n"
1487 "fmla v16.4s, v25.4s, v2.s[1]\n"
1488 "fmla v20.4s, v25.4s, v3.s[1]\n"
1489 "ldr q25, [x10, #0x60]\n"
1490 "fmla v9.4s, v24.4s, v0.s[1]\n"
1491 "fmla v13.4s, v24.4s, v1.s[1]\n"
1492 "fmla v17.4s, v24.4s, v2.s[1]\n"
1493 "fmla v21.4s, v24.4s, v3.s[1]\n"
1494 "ldr q24, [x10, #0x70]\n"
1495 "fmla v10.4s, v25.4s, v0.s[1]\n"
1496 "fmla v14.4s, v25.4s, v1.s[1]\n"
1497 "fmla v18.4s, v25.4s, v2.s[1]\n"
1498 "fmla v22.4s, v25.4s, v3.s[1]\n"
1499 "ldr q25, [x10, #0x80]\n"
1500 "fmla v11.4s, v24.4s, v0.s[1]\n"
1501 "fmla v15.4s, v24.4s, v1.s[1]\n"
1502 "fmla v19.4s, v24.4s, v2.s[1]\n"
1503 "fmla v23.4s, v24.4s, v3.s[1]\n"
1504 "ldr q24, [x10, #0x90]\n"
1505 "fmla v8.4s, v25.4s, v0.s[2]\n"
1506 "fmla v12.4s, v25.4s, v1.s[2]\n"
1507 "fmla v16.4s, v25.4s, v2.s[2]\n"
1508 "fmla v20.4s, v25.4s, v3.s[2]\n"
1509 "ldr q25, [x10, #0xa0]\n"
1510 "fmla v9.4s, v24.4s, v0.s[2]\n"
1511 "fmla v13.4s, v24.4s, v1.s[2]\n"
1512 "fmla v17.4s, v24.4s, v2.s[2]\n"
1513 "fmla v21.4s, v24.4s, v3.s[2]\n"
1514 "ldr q24, [x10, #0xb0]\n"
1515 "fmla v10.4s, v25.4s, v0.s[2]\n"
1516 "fmla v14.4s, v25.4s, v1.s[2]\n"
1517 "fmla v18.4s, v25.4s, v2.s[2]\n"
1518 "fmla v22.4s, v25.4s, v3.s[2]\n"
1519 "ldr q25, [x10, #0xc0]\n"
1520 "fmla v11.4s, v24.4s, v0.s[2]\n"
1521 "fmla v15.4s, v24.4s, v1.s[2]\n"
1522 "fmla v19.4s, v24.4s, v2.s[2]\n"
1523 "fmla v23.4s, v24.4s, v3.s[2]\n"
1524 "ldr q24, [x10, #0xd0]\n"
1525 "fmla v8.4s, v25.4s, v0.s[3]\n"
1526 "fmla v12.4s, v25.4s, v1.s[3]\n"
1527 "fmla v16.4s, v25.4s, v2.s[3]\n"
1528 "fmla v20.4s, v25.4s, v3.s[3]\n"
1529 "ldr q25, [x10, #0xe0]\n"
1530 "fmla v9.4s, v24.4s, v0.s[3]\n"
1531 "fmla v13.4s, v24.4s, v1.s[3]\n"
1532 "fmla v17.4s, v24.4s, v2.s[3]\n"
1533 "fmla v21.4s, v24.4s, v3.s[3]\n"
1534 "ldr q24, [x10, #0xf0]\n"
1535 "add x10, x10, #0x100\n"
1536 "fmla v10.4s, v25.4s, v0.s[3]\n"
1537 "fmla v14.4s, v25.4s, v1.s[3]\n"
1538 "fmla v18.4s, v25.4s, v2.s[3]\n"
1539 "fmla v22.4s, v25.4s, v3.s[3]\n"
1540 "ldr q6, [x10, #0x0]\n"
1541 "fmla v11.4s, v24.4s, v0.s[3]\n"
1542 "ldr q0, [x26, #0x0]\n"
1543 "fmla v15.4s, v24.4s, v1.s[3]\n"
1544 "ldr q1, [x25, #0x0]\n"
1545 "fmla v19.4s, v24.4s, v2.s[3]\n"
1546 "ldr q2, [x24, #0x0]\n"
1547 "fmla v23.4s, v24.4s, v3.s[3]\n"
1548 "ldr q3, [x23, #0x0]\n"
1549 "ldr q7, [x10, #0x10]\n"
1552 "fmla v8.4s, v6.4s, v0.s[0]\n"
1553 "fmla v12.4s, v6.4s, v1.s[0]\n"
1554 "add x26, x26, #0x10\n"
1555 "add x25, x25, #0x10\n"
1556 "fmla v16.4s, v6.4s, v2.s[0]\n"
1557 "fmla v20.4s, v6.4s, v3.s[0]\n"
1558 "ldr q25, [x10, #0x20]\n"
1559 "add x24, x24, #0x10\n"
1560 "fmla v9.4s, v7.4s, v0.s[0]\n"
1561 "fmla v13.4s, v7.4s, v1.s[0]\n"
1562 "add x23, x23, #0x10\n"
1563 "sub x27, x27, #0x4\n"
1564 "fmla v17.4s, v7.4s, v2.s[0]\n"
1565 "fmla v21.4s, v7.4s, v3.s[0]\n"
1566 "ldr q24, [x10, #0x30]\n"
1567 "prfm pldl1keep, [x26, #0x80]\n"
1568 "fmla v10.4s, v25.4s, v0.s[0]\n"
1569 "fmla v14.4s, v25.4s, v1.s[0]\n"
1570 "prfm pldl1keep, [x25, #0x80]\n"
1571 "prfm pldl1keep, [x24, #0x80]\n"
1572 "fmla v18.4s, v25.4s, v2.s[0]\n"
1573 "fmla v22.4s, v25.4s, v3.s[0]\n"
1574 "ldr q25, [x10, #0x40]\n"
1575 "prfm pldl1keep, [x23, #0x80]\n"
1576 "fmla v11.4s, v24.4s, v0.s[0]\n"
1577 "fmla v15.4s, v24.4s, v1.s[0]\n"
1578 "fmla v19.4s, v24.4s, v2.s[0]\n"
1579 "fmla v23.4s, v24.4s, v3.s[0]\n"
1580 "ldr q24, [x10, #0x50]\n"
1581 "fmla v8.4s, v25.4s, v0.s[1]\n"
1582 "fmla v12.4s, v25.4s, v1.s[1]\n"
1583 "fmla v16.4s, v25.4s, v2.s[1]\n"
1584 "fmla v20.4s, v25.4s, v3.s[1]\n"
1585 "ldr q25, [x10, #0x60]\n"
1586 "fmla v9.4s, v24.4s, v0.s[1]\n"
1587 "fmla v13.4s, v24.4s, v1.s[1]\n"
1588 "fmla v17.4s, v24.4s, v2.s[1]\n"
1589 "fmla v21.4s, v24.4s, v3.s[1]\n"
1590 "ldr q24, [x10, #0x70]\n"
1591 "fmla v10.4s, v25.4s, v0.s[1]\n"
1592 "fmla v14.4s, v25.4s, v1.s[1]\n"
1593 "fmla v18.4s, v25.4s, v2.s[1]\n"
1594 "fmla v22.4s, v25.4s, v3.s[1]\n"
1595 "ldr q25, [x10, #0x80]\n"
1596 "fmla v11.4s, v24.4s, v0.s[1]\n"
1597 "fmla v15.4s, v24.4s, v1.s[1]\n"
1598 "fmla v19.4s, v24.4s, v2.s[1]\n"
1599 "fmla v23.4s, v24.4s, v3.s[1]\n"
1600 "ldr q24, [x10, #0x90]\n"
1601 "fmla v8.4s, v25.4s, v0.s[2]\n"
1602 "fmla v12.4s, v25.4s, v1.s[2]\n"
1603 "fmla v16.4s, v25.4s, v2.s[2]\n"
1604 "fmla v20.4s, v25.4s, v3.s[2]\n"
1605 "ldr q25, [x10, #0xa0]\n"
1606 "fmla v9.4s, v24.4s, v0.s[2]\n"
1607 "fmla v13.4s, v24.4s, v1.s[2]\n"
1608 "fmla v17.4s, v24.4s, v2.s[2]\n"
1609 "fmla v21.4s, v24.4s, v3.s[2]\n"
1610 "ldr q24, [x10, #0xb0]\n"
1611 "fmla v10.4s, v25.4s, v0.s[2]\n"
1612 "fmla v14.4s, v25.4s, v1.s[2]\n"
1613 "fmla v18.4s, v25.4s, v2.s[2]\n"
1614 "fmla v22.4s, v25.4s, v3.s[2]\n"
1615 "ldr q25, [x10, #0xc0]\n"
1616 "fmla v11.4s, v24.4s, v0.s[2]\n"
1617 "fmla v15.4s, v24.4s, v1.s[2]\n"
1618 "fmla v19.4s, v24.4s, v2.s[2]\n"
1619 "fmla v23.4s, v24.4s, v3.s[2]\n"
1620 "ldr q24, [x10, #0xd0]\n"
1621 "fmla v8.4s, v25.4s, v0.s[3]\n"
1622 "fmla v12.4s, v25.4s, v1.s[3]\n"
1623 "fmla v16.4s, v25.4s, v2.s[3]\n"
1624 "fmla v20.4s, v25.4s, v3.s[3]\n"
1625 "ldr q25, [x10, #0xe0]\n"
1626 "fmla v9.4s, v24.4s, v0.s[3]\n"
1627 "fmla v13.4s, v24.4s, v1.s[3]\n"
1628 "fmla v17.4s, v24.4s, v2.s[3]\n"
1629 "fmla v21.4s, v24.4s, v3.s[3]\n"
1630 "ldr q24, [x10, #0xf0]\n"
1631 "add x10, x10, #0x100\n"
1632 "fmla v10.4s, v25.4s, v0.s[3]\n"
1633 "fmla v14.4s, v25.4s, v1.s[3]\n"
1634 "fmla v18.4s, v25.4s, v2.s[3]\n"
1635 "fmla v22.4s, v25.4s, v3.s[3]\n"
1636 "fmla v11.4s, v24.4s, v0.s[3]\n"
1637 "fmla v15.4s, v24.4s, v1.s[3]\n"
1638 "fmla v19.4s, v24.4s, v2.s[3]\n"
1639 "fmla v23.4s, v24.4s, v3.s[3]\n"
1643 "ldr s29, [x26], #0x4\n"
1644 "ldr s28, [x25], #0x4\n"
1645 "sub x27, x27, #0x1\n"
1646 "ldr s27, [x24], #0x4\n"
1647 "ldr s26, [x23], #0x4\n"
1648 "ldr q25, [x10, #0x0]\n"
1649 "ldr q24, [x10, #0x10]\n"
1650 "fmla v8.4s, v25.4s, v29.s[0]\n"
1651 "fmla v12.4s, v25.4s, v28.s[0]\n"
1652 "fmla v16.4s, v25.4s, v27.s[0]\n"
1653 "fmla v20.4s, v25.4s, v26.s[0]\n"
1654 "ldr q25, [x10, #0x20]\n"
1655 "fmla v9.4s, v24.4s, v29.s[0]\n"
1656 "fmla v13.4s, v24.4s, v28.s[0]\n"
1657 "fmla v17.4s, v24.4s, v27.s[0]\n"
1658 "fmla v21.4s, v24.4s, v26.s[0]\n"
1659 "ldr q24, [x10, #0x30]\n"
1660 "add x10, x10, #0x40\n"
1661 "fmla v10.4s, v25.4s, v29.s[0]\n"
1662 "fmla v14.4s, v25.4s, v28.s[0]\n"
1663 "fmla v18.4s, v25.4s, v27.s[0]\n"
1664 "fmla v22.4s, v25.4s, v26.s[0]\n"
1665 "fmla v11.4s, v24.4s, v29.s[0]\n"
1666 "fmla v15.4s, v24.4s, v28.s[0]\n"
1667 "fmla v19.4s, v24.4s, v27.s[0]\n"
1668 "fmla v23.4s, v24.4s, v26.s[0]\n"
1671 "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
1672 "add x28, x28, #0x1\n"
1675 "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
1676 "add x25, x9, x20, LSL #2\n"
1677 "add x24, x25, x20, LSL #2\n"
1678 "prfm pstl1keep, [x9, #0x0]\n"
1679 "add x23, x24, x20, LSL #2\n"
1680 "prfm pstl1keep, [x25, #0x0]\n"
1681 "prfm pstl1keep, [x24, #0x0]\n"
1682 "prfm pstl1keep, [x23, #0x0]\n"
1683 "tbz %x[flags], #1, 122f\n"
1684 "add x20, %x[args_ptr], %[offset_max]\n"
1685 "ld1r { v25.4s }, [x20]\n"
1686 "add x20, %x[args_ptr], %[offset_min]\n"
1687 "ld1r { v24.4s }, [x20]\n"
1688 "fmin v8.4s, v8.4s, v25.4s\n"
1689 "fmin v9.4s, v9.4s, v25.4s\n"
1690 "fmin v10.4s, v10.4s, v25.4s\n"
1691 "fmin v11.4s, v11.4s, v25.4s\n"
1692 "fmin v12.4s, v12.4s, v25.4s\n"
1693 "fmin v13.4s, v13.4s, v25.4s\n"
1694 "fmin v14.4s, v14.4s, v25.4s\n"
1695 "fmin v15.4s, v15.4s, v25.4s\n"
1696 "fmin v16.4s, v16.4s, v25.4s\n"
1697 "fmin v17.4s, v17.4s, v25.4s\n"
1698 "fmin v18.4s, v18.4s, v25.4s\n"
1699 "fmin v19.4s, v19.4s, v25.4s\n"
1700 "fmin v20.4s, v20.4s, v25.4s\n"
1701 "fmin v21.4s, v21.4s, v25.4s\n"
1702 "fmin v22.4s, v22.4s, v25.4s\n"
1703 "fmin v23.4s, v23.4s, v25.4s\n"
1704 "fmax v8.4s, v8.4s, v24.4s\n"
1705 "fmax v9.4s, v9.4s, v24.4s\n"
1706 "fmax v10.4s, v10.4s, v24.4s\n"
1707 "fmax v11.4s, v11.4s, v24.4s\n"
1708 "fmax v12.4s, v12.4s, v24.4s\n"
1709 "fmax v13.4s, v13.4s, v24.4s\n"
1710 "fmax v14.4s, v14.4s, v24.4s\n"
1711 "fmax v15.4s, v15.4s, v24.4s\n"
1712 "fmax v16.4s, v16.4s, v24.4s\n"
1713 "fmax v17.4s, v17.4s, v24.4s\n"
1714 "fmax v18.4s, v18.4s, v24.4s\n"
1715 "fmax v19.4s, v19.4s, v24.4s\n"
1716 "fmax v20.4s, v20.4s, v24.4s\n"
1717 "fmax v21.4s, v21.4s, v24.4s\n"
1718 "fmax v22.4s, v22.4s, v24.4s\n"
1719 "fmax v23.4s, v23.4s, v24.4s\n"
1723 "tbz x11, #3, 126f\n"
1724 "st1 { v8.4s }, [x9], #0x10\n"
1725 "st1 { v9.4s }, [x9], #0x10\n"
1726 "st1 { v12.4s }, [x25], #0x10\n"
1727 "st1 { v13.4s }, [x25], #0x10\n"
1728 "st1 { v16.4s }, [x24], #0x10\n"
1729 "st1 { v17.4s }, [x24], #0x10\n"
1730 "st1 { v20.4s }, [x23], #0x10\n"
1731 "st1 { v21.4s }, [x23], #0x10\n"
1732 "tbz x11, #2, 124f\n"
1733 "st1 { v10.4s }, [x9], #0x10\n"
1734 "st1 { v14.4s }, [x25], #0x10\n"
1735 "st1 { v18.4s }, [x24], #0x10\n"
1736 "st1 { v22.4s }, [x23], #0x10\n"
1737 "tbz x11, #1, 123f\n"
1738 "str d11, [x9], #0x8\n"
1739 "str d15, [x25], #0x8\n"
1740 "str d19, [x24], #0x8\n"
1741 "str d23, [x23], #0x8\n"
1742 "tbz x11, #0, 130f\n"
1743 "st1 { v11.s }[2], [x9]\n"
1744 "st1 { v15.s }[2], [x25]\n"
1745 "st1 { v19.s }[2], [x24]\n"
1746 "st1 { v23.s }[2], [x23]\n"
1749 "tbz x11, #0, 130f\n"
1750 "str s11, [x9, #0x0]\n"
1751 "str s15, [x25, #0x0]\n"
1752 "str s19, [x24, #0x0]\n"
1753 "str s23, [x23, #0x0]\n"
1756 "tbz x11, #1, 125f\n"
1757 "str d10, [x9], #0x8\n"
1758 "str d14, [x25], #0x8\n"
1759 "str d18, [x24], #0x8\n"
1760 "str d22, [x23], #0x8\n"
1761 "tbz x11, #0, 130f\n"
1762 "st1 { v10.s }[2], [x9]\n"
1763 "st1 { v14.s }[2], [x25]\n"
1764 "st1 { v18.s }[2], [x24]\n"
1765 "st1 { v22.s }[2], [x23]\n"
1768 "tbz x11, #0, 130f\n"
1769 "str s10, [x9, #0x0]\n"
1770 "str s14, [x25, #0x0]\n"
1771 "str s18, [x24, #0x0]\n"
1772 "str s22, [x23, #0x0]\n"
1775 "tbz x11, #2, 128f\n"
1776 "st1 { v8.4s }, [x9], #0x10\n"
1777 "st1 { v12.4s }, [x25], #0x10\n"
1778 "st1 { v16.4s }, [x24], #0x10\n"
1779 "st1 { v20.4s }, [x23], #0x10\n"
1780 "tbz x11, #1, 127f\n"
1781 "str d9, [x9], #0x8\n"
1782 "str d13, [x25], #0x8\n"
1783 "str d17, [x24], #0x8\n"
1784 "str d21, [x23], #0x8\n"
1785 "tbz x11, #0, 130f\n"
1786 "st1 { v9.s }[2], [x9]\n"
1787 "st1 { v13.s }[2], [x25]\n"
1788 "st1 { v17.s }[2], [x24]\n"
1789 "st1 { v21.s }[2], [x23]\n"
1792 "tbz x11, #0, 130f\n"
1793 "str s9, [x9, #0x0]\n"
1794 "str s13, [x25, #0x0]\n"
1795 "str s17, [x24, #0x0]\n"
1796 "str s21, [x23, #0x0]\n"
1799 "tbz x11, #1, 129f\n"
1800 "str d8, [x9], #0x8\n"
1801 "str d12, [x25], #0x8\n"
1802 "str d16, [x24], #0x8\n"
1803 "str d20, [x23], #0x8\n"
1804 "tbz x11, #0, 130f\n"
1805 "st1 { v8.s }[2], [x9]\n"
1806 "st1 { v12.s }[2], [x25]\n"
1807 "st1 { v16.s }[2], [x24]\n"
1808 "st1 { v20.s }[2], [x23]\n"
1811 "str s8, [x9, #0x0]\n"
1812 "str s12, [x25, #0x0]\n"
1813 "str s16, [x24, #0x0]\n"
1814 "str s20, [x23, #0x0]\n"
1818 "str q8, [x9, #0x0]\n"
1819 "str q9, [x9, #0x10]\n"
1820 "str q10, [x9, #0x20]\n"
1821 "str q11, [x9, #0x30]\n"
1822 "add x9, x9, #0x40\n"
1823 "str q12, [x25, #0x0]\n"
1824 "str q13, [x25, #0x10]\n"
1825 "str q14, [x25, #0x20]\n"
1826 "str q15, [x25, #0x30]\n"
1827 "str q16, [x24, #0x0]\n"
1828 "str q17, [x24, #0x10]\n"
1829 "str q18, [x24, #0x20]\n"
1830 "str q19, [x24, #0x30]\n"
1831 "str q20, [x23, #0x0]\n"
1832 "str q21, [x23, #0x10]\n"
1833 "str q22, [x23, #0x20]\n"
1834 "str q23, [x23, #0x30]\n"
1836 "subs x11, x11, #0x10\n"
1840 "mov x12, %x[bias]\n"
1841 "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
1842 "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
1843 "mov x9, %x[output_ptr]\n"
1846 "ldr q8, [x12, #0x0]\n"
1847 "ldr q9, [x12, #0x10]\n"
1848 "mov v12.16b, v8.16b\n"
1849 "mov v13.16b, v9.16b\n"
1850 "ldr q10, [x12, #0x20]\n"
1851 "ldr q11, [x12, #0x30]\n"
1852 "mov v14.16b, v10.16b\n"
1853 "mov v15.16b, v11.16b\n"
1854 "mov v16.16b, v8.16b\n"
1855 "mov v17.16b, v9.16b\n"
1856 "add x12, x12, #0x40\n"
1857 "mov v18.16b, v10.16b\n"
1858 "mov v19.16b, v11.16b\n"
1859 "mov v20.16b, v8.16b\n"
1860 "mov v21.16b, v9.16b\n"
1861 "mov v22.16b, v10.16b\n"
1862 "mov v23.16b, v11.16b\n"
1863 "mov v24.16b, v8.16b\n"
1864 "mov v25.16b, v9.16b\n"
1865 "mov v26.16b, v10.16b\n"
1866 "mov v27.16b, v11.16b\n"
1869 "tbz %x[flags], #0, 145f\n"
1870 "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
1871 "add x25, x9, x20, LSL #2\n"
1872 "add x24, x25, x20, LSL #2\n"
1873 "add x23, x24, x20, LSL #2\n"
1875 "add x22, x23, x20, LSL #2\n"
1877 "tbz x11, #3, 139f\n"
1878 "ld1 { v8.4s }, [x9], #0x10\n"
1879 "ld1 { v12.4s }, [x25], #0x10\n"
1880 "ld1 { v16.4s }, [x24], #0x10\n"
1881 "ld1 { v20.4s }, [x23], #0x10\n"
1882 "ld1 { v24.4s }, [x22], #0x10\n"
1883 "ld1 { v9.4s }, [x9], #0x10\n"
1884 "ld1 { v13.4s }, [x25], #0x10\n"
1885 "ld1 { v17.4s }, [x24], #0x10\n"
1886 "ld1 { v21.4s }, [x23], #0x10\n"
1887 "ld1 { v25.4s }, [x22], #0x10\n"
1888 "tbz x11, #2, 137f\n"
1889 "ld1 { v10.4s }, [x9], #0x10\n"
1890 "ld1 { v14.4s }, [x25], #0x10\n"
1891 "ld1 { v18.4s }, [x24], #0x10\n"
1892 "ld1 { v22.4s }, [x23], #0x10\n"
1893 "ld1 { v26.4s }, [x22], #0x10\n"
1894 "tbz x11, #1, 136f\n"
1895 "ldr d11, [x9], #0x8\n"
1896 "ldr d15, [x25], #0x8\n"
1898 "ldr d19, [x24], #0x8\n"
1899 "ldr d23, [x23], #0x8\n"
1900 "ldr d27, [x22], #0x8\n"
1901 "tbz x11, #0, 143f\n"
1902 "ld1 { v11.s }[2], [x9]\n"
1903 "ld1 { v15.s }[2], [x25]\n"
1904 "ld1 { v19.s }[2], [x24]\n"
1905 "ld1 { v23.s }[2], [x23]\n"
1906 "ld1 { v27.s }[2], [x22]\n"
1910 "tbz x11, #0, 143f\n"
1911 "ldr s11, [x9, #0x0]\n"
1912 "ldr s15, [x25, #0x0]\n"
1913 "ldr s19, [x24, #0x0]\n"
1914 "ldr s23, [x23, #0x0]\n"
1915 "ldr s27, [x22, #0x0]\n"
1918 "tbz x11, #1, 138f\n"
1919 "ldr d10, [x9], #0x8\n"
1920 "ldr d14, [x25], #0x8\n"
1922 "ldr d18, [x24], #0x8\n"
1923 "ldr d22, [x23], #0x8\n"
1924 "ldr d26, [x22], #0x8\n"
1925 "tbz x11, #0, 143f\n"
1926 "ld1 { v10.s }[2], [x9]\n"
1927 "ld1 { v14.s }[2], [x25]\n"
1928 "ld1 { v18.s }[2], [x24]\n"
1929 "ld1 { v22.s }[2], [x23]\n"
1930 "ld1 { v26.s }[2], [x22]\n"
1934 "tbz x11, #0, 143f\n"
1935 "ldr s10, [x9, #0x0]\n"
1936 "ldr s14, [x25, #0x0]\n"
1937 "ldr s18, [x24, #0x0]\n"
1938 "ldr s22, [x23, #0x0]\n"
1939 "ldr s26, [x22, #0x0]\n"
1942 "tbz x11, #2, 141f\n"
1943 "ld1 { v8.4s }, [x9], #0x10\n"
1944 "ld1 { v12.4s }, [x25], #0x10\n"
1945 "ld1 { v16.4s }, [x24], #0x10\n"
1946 "ld1 { v20.4s }, [x23], #0x10\n"
1947 "ld1 { v24.4s }, [x22], #0x10\n"
1948 "tbz x11, #1, 140f\n"
1949 "ldr d9, [x9], #0x8\n"
1950 "ldr d13, [x25], #0x8\n"
1952 "ldr d17, [x24], #0x8\n"
1953 "ldr d21, [x23], #0x8\n"
1954 "ldr d25, [x22], #0x8\n"
1955 "tbz x11, #0, 143f\n"
1956 "ld1 { v9.s }[2], [x9]\n"
1957 "ld1 { v13.s }[2], [x25]\n"
1958 "ld1 { v17.s }[2], [x24]\n"
1959 "ld1 { v21.s }[2], [x23]\n"
1960 "ld1 { v25.s }[2], [x22]\n"
1964 "tbz x11, #0, 143f\n"
1965 "ldr s9, [x9, #0x0]\n"
1966 "ldr s13, [x25, #0x0]\n"
1967 "ldr s17, [x24, #0x0]\n"
1968 "ldr s21, [x23, #0x0]\n"
1969 "ldr s25, [x22, #0x0]\n"
1972 "tbz x11, #1, 142f\n"
1973 "ldr d8, [x9], #0x8\n"
1974 "ldr d12, [x25], #0x8\n"
1976 "ldr d16, [x24], #0x8\n"
1977 "ldr d20, [x23], #0x8\n"
1978 "ldr d24, [x22], #0x8\n"
1979 "tbz x11, #0, 143f\n"
1980 "ld1 { v8.s }[2], [x9]\n"
1981 "ld1 { v12.s }[2], [x25]\n"
1982 "ld1 { v16.s }[2], [x24]\n"
1983 "ld1 { v20.s }[2], [x23]\n"
1984 "ld1 { v24.s }[2], [x22]\n"
1987 "ldr s8, [x9, #0x0]\n"
1988 "ldr s12, [x25, #0x0]\n"
1990 "ldr s16, [x24, #0x0]\n"
1991 "ldr s20, [x23, #0x0]\n"
1992 "ldr s24, [x22, #0x0]\n"
1997 "ldr q8, [x9, #0x0]\n"
1998 "ldr q9, [x9, #0x10]\n"
1999 "ldr q10, [x9, #0x20]\n"
2000 "ldr q11, [x9, #0x30]\n"
2001 "ldr q12, [x25, #0x0]\n"
2002 "ldr q13, [x25, #0x10]\n"
2003 "ldr q14, [x25, #0x20]\n"
2004 "ldr q15, [x25, #0x30]\n"
2005 "ldr q16, [x24, #0x0]\n"
2006 "ldr q17, [x24, #0x10]\n"
2007 "ldr q18, [x24, #0x20]\n"
2008 "ldr q19, [x24, #0x30]\n"
2009 "ldr q20, [x23, #0x0]\n"
2010 "ldr q21, [x23, #0x10]\n"
2011 "ldr q22, [x23, #0x20]\n"
2012 "ldr q23, [x23, #0x30]\n"
2013 "ldr q24, [x22, #0x0]\n"
2014 "ldr q25, [x22, #0x10]\n"
2015 "ldr q26, [x22, #0x20]\n"
2016 "ldr q27, [x22, #0x30]\n"
2019 "movi v8.16b, #0x0\n"
2020 "movi v9.16b, #0x0\n"
2021 "movi v10.16b, #0x0\n"
2022 "movi v11.16b, #0x0\n"
2023 "movi v12.16b, #0x0\n"
2024 "movi v13.16b, #0x0\n"
2025 "movi v14.16b, #0x0\n"
2026 "movi v15.16b, #0x0\n"
2027 "movi v16.16b, #0x0\n"
2028 "movi v17.16b, #0x0\n"
2029 "movi v18.16b, #0x0\n"
2030 "movi v19.16b, #0x0\n"
2031 "movi v20.16b, #0x0\n"
2032 "movi v21.16b, #0x0\n"
2033 "movi v22.16b, #0x0\n"
2034 "movi v23.16b, #0x0\n"
2035 "movi v24.16b, #0x0\n"
2036 "movi v25.16b, #0x0\n"
2037 "movi v26.16b, #0x0\n"
2038 "movi v27.16b, #0x0\n"
2042 "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
2043 "ldr w27, [x20, x28, LSL #0x2]\n"
2044 "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
2045 "tbz %x[flags], #3, 148f\n"
2046 "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
2047 "add x20, x20, x21, LSL #3\n"
2048 "ldr x26, [x20, #0x0]\n"
2049 "ldr x25, [x20, #0x8]\n"
2050 "ldr x24, [x20, #0x10]\n"
2051 "ldr x23, [x20, #0x18]\n"
2052 "ldr x22, [x20, #0x20]\n"
2054 "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
2055 "add x26, x26, x20, LSL #2\n"
2056 "add x25, x25, x20, LSL #2\n"
2057 "add x24, x24, x20, LSL #2\n"
2058 "add x23, x23, x20, LSL #2\n"
2059 "add x22, x22, x20, LSL #2\n"
2062 "mov x26, %x[input_ptr]\n"
2063 "add x25, x26, x21, LSL #2\n"
2064 "add x24, x25, x21, LSL #2\n"
2065 "add x23, x24, x21, LSL #2\n"
2066 "add x22, x23, x21, LSL #2\n"
2070 "ldr q0, [x26, #0x0]\n"
2071 "ldr q1, [x25, #0x0]\n"
2073 "ldr q2, [x24, #0x0]\n"
2074 "ldr q3, [x23, #0x0]\n"
2075 "ldr q4, [x22, #0x0]\n"
2076 "ldr q6, [x10, #0x0]\n"
2077 "ldr q7, [x10, #0x10]\n"
2080 "fmla v8.4s, v6.4s, v0.s[0]\n"
2081 "fmla v12.4s, v6.4s, v1.s[0]\n"
2082 "sub x27, x27, #0x4\n"
2083 "add x26, x26, #0x10\n"
2084 "fmla v16.4s, v6.4s, v2.s[0]\n"
2085 "fmla v20.4s, v6.4s, v3.s[0]\n"
2086 "add x25, x25, #0x10\n"
2087 "add x24, x24, #0x10\n"
2088 "fmla v24.4s, v6.4s, v4.s[0]\n"
2089 "ldr q29, [x10, #0x20]\n"
2090 "fmla v9.4s, v7.4s, v0.s[0]\n"
2091 "add x23, x23, #0x10\n"
2092 "fmla v13.4s, v7.4s, v1.s[0]\n"
2093 "fmla v17.4s, v7.4s, v2.s[0]\n"
2094 "add x22, x22, #0x10\n"
2096 "fmla v21.4s, v7.4s, v3.s[0]\n"
2097 "fmla v25.4s, v7.4s, v4.s[0]\n"
2098 "ldr q28, [x10, #0x30]\n"
2099 "prfm pldl1keep, [x26, #0x80]\n"
2100 "fmla v10.4s, v29.4s, v0.s[0]\n"
2101 "fmla v14.4s, v29.4s, v1.s[0]\n"
2102 "prfm pldl1keep, [x25, #0x80]\n"
2103 "prfm pldl1keep, [x24, #0x80]\n"
2104 "fmla v18.4s, v29.4s, v2.s[0]\n"
2105 "fmla v22.4s, v29.4s, v3.s[0]\n"
2106 "prfm pldl1keep, [x23, #0x80]\n"
2107 "prfm pldl1keep, [x22, #0x80]\n"
2108 "fmla v26.4s, v29.4s, v4.s[0]\n"
2109 "ldr q29, [x10, #0x40]\n"
2110 "fmla v11.4s, v28.4s, v0.s[0]\n"
2111 "fmla v15.4s, v28.4s, v1.s[0]\n"
2112 "fmla v19.4s, v28.4s, v2.s[0]\n"
2113 "fmla v23.4s, v28.4s, v3.s[0]\n"
2114 "fmla v27.4s, v28.4s, v4.s[0]\n"
2115 "ldr q28, [x10, #0x50]\n"
2116 "fmla v8.4s, v29.4s, v0.s[1]\n"
2117 "fmla v12.4s, v29.4s, v1.s[1]\n"
2118 "fmla v16.4s, v29.4s, v2.s[1]\n"
2119 "fmla v20.4s, v29.4s, v3.s[1]\n"
2120 "fmla v24.4s, v29.4s, v4.s[1]\n"
2121 "ldr q29, [x10, #0x60]\n"
2122 "fmla v9.4s, v28.4s, v0.s[1]\n"
2123 "fmla v13.4s, v28.4s, v1.s[1]\n"
2124 "fmla v17.4s, v28.4s, v2.s[1]\n"
2125 "fmla v21.4s, v28.4s, v3.s[1]\n"
2126 "fmla v25.4s, v28.4s, v4.s[1]\n"
2127 "ldr q28, [x10, #0x70]\n"
2128 "fmla v10.4s, v29.4s, v0.s[1]\n"
2129 "fmla v14.4s, v29.4s, v1.s[1]\n"
2130 "fmla v18.4s, v29.4s, v2.s[1]\n"
2131 "fmla v22.4s, v29.4s, v3.s[1]\n"
2132 "fmla v26.4s, v29.4s, v4.s[1]\n"
2133 "ldr q29, [x10, #0x80]\n"
2134 "fmla v11.4s, v28.4s, v0.s[1]\n"
2135 "fmla v15.4s, v28.4s, v1.s[1]\n"
2136 "fmla v19.4s, v28.4s, v2.s[1]\n"
2137 "fmla v23.4s, v28.4s, v3.s[1]\n"
2138 "fmla v27.4s, v28.4s, v4.s[1]\n"
2139 "ldr q28, [x10, #0x90]\n"
2140 "fmla v8.4s, v29.4s, v0.s[2]\n"
2141 "fmla v12.4s, v29.4s, v1.s[2]\n"
2142 "fmla v16.4s, v29.4s, v2.s[2]\n"
2143 "fmla v20.4s, v29.4s, v3.s[2]\n"
2144 "fmla v24.4s, v29.4s, v4.s[2]\n"
2145 "ldr q29, [x10, #0xa0]\n"
2146 "fmla v9.4s, v28.4s, v0.s[2]\n"
2147 "fmla v13.4s, v28.4s, v1.s[2]\n"
2148 "fmla v17.4s, v28.4s, v2.s[2]\n"
2149 "fmla v21.4s, v28.4s, v3.s[2]\n"
2150 "fmla v25.4s, v28.4s, v4.s[2]\n"
2151 "ldr q28, [x10, #0xb0]\n"
2152 "fmla v10.4s, v29.4s, v0.s[2]\n"
2153 "fmla v14.4s, v29.4s, v1.s[2]\n"
2154 "fmla v18.4s, v29.4s, v2.s[2]\n"
2155 "fmla v22.4s, v29.4s, v3.s[2]\n"
2156 "fmla v26.4s, v29.4s, v4.s[2]\n"
2157 "ldr q29, [x10, #0xc0]\n"
2158 "fmla v11.4s, v28.4s, v0.s[2]\n"
2159 "fmla v15.4s, v28.4s, v1.s[2]\n"
2160 "fmla v19.4s, v28.4s, v2.s[2]\n"
2161 "fmla v23.4s, v28.4s, v3.s[2]\n"
2162 "fmla v27.4s, v28.4s, v4.s[2]\n"
2163 "ldr q28, [x10, #0xd0]\n"
2164 "fmla v8.4s, v29.4s, v0.s[3]\n"
2165 "fmla v12.4s, v29.4s, v1.s[3]\n"
2166 "fmla v16.4s, v29.4s, v2.s[3]\n"
2167 "fmla v20.4s, v29.4s, v3.s[3]\n"
2168 "fmla v24.4s, v29.4s, v4.s[3]\n"
2169 "ldr q29, [x10, #0xe0]\n"
2170 "fmla v9.4s, v28.4s, v0.s[3]\n"
2171 "fmla v13.4s, v28.4s, v1.s[3]\n"
2172 "fmla v17.4s, v28.4s, v2.s[3]\n"
2173 "fmla v21.4s, v28.4s, v3.s[3]\n"
2174 "fmla v25.4s, v28.4s, v4.s[3]\n"
2175 "ldr q28, [x10, #0xf0]\n"
2176 "add x10, x10, #0x100\n"
2177 "fmla v10.4s, v29.4s, v0.s[3]\n"
2178 "fmla v14.4s, v29.4s, v1.s[3]\n"
2179 "fmla v18.4s, v29.4s, v2.s[3]\n"
2180 "fmla v22.4s, v29.4s, v3.s[3]\n"
2181 "fmla v26.4s, v29.4s, v4.s[3]\n"
2182 "ldr q6, [x10, #0x0]\n"
2183 "fmla v11.4s, v28.4s, v0.s[3]\n"
2184 "ldr q0, [x26, #0x0]\n"
2185 "fmla v15.4s, v28.4s, v1.s[3]\n"
2186 "ldr q1, [x25, #0x0]\n"
2187 "fmla v19.4s, v28.4s, v2.s[3]\n"
2188 "ldr q2, [x24, #0x0]\n"
2189 "fmla v23.4s, v28.4s, v3.s[3]\n"
2190 "ldr q3, [x23, #0x0]\n"
2191 "fmla v27.4s, v28.4s, v4.s[3]\n"
2192 "ldr q4, [x22, #0x0]\n"
2193 "ldr q7, [x10, #0x10]\n"
2196 "fmla v8.4s, v6.4s, v0.s[0]\n"
2197 "fmla v12.4s, v6.4s, v1.s[0]\n"
2198 "add x26, x26, #0x10\n"
2199 "add x25, x25, #0x10\n"
2200 "fmla v16.4s, v6.4s, v2.s[0]\n"
2201 "fmla v20.4s, v6.4s, v3.s[0]\n"
2202 "add x24, x24, #0x10\n"
2203 "add x23, x23, #0x10\n"
2204 "fmla v24.4s, v6.4s, v4.s[0]\n"
2205 "ldr q29, [x10, #0x20]\n"
2206 "fmla v9.4s, v7.4s, v0.s[0]\n"
2207 "add x22, x22, #0x10\n"
2208 "fmla v13.4s, v7.4s, v1.s[0]\n"
2209 "fmla v17.4s, v7.4s, v2.s[0]\n"
2210 "sub x27, x27, #0x4\n"
2211 "prfm pldl1keep, [x26, #0x80]\n"
2212 "fmla v21.4s, v7.4s, v3.s[0]\n"
2213 "fmla v25.4s, v7.4s, v4.s[0]\n"
2214 "ldr q28, [x10, #0x30]\n"
2215 "prfm pldl1keep, [x25, #0x80]\n"
2216 "fmla v10.4s, v29.4s, v0.s[0]\n"
2217 "fmla v14.4s, v29.4s, v1.s[0]\n"
2218 "prfm pldl1keep, [x24, #0x80]\n"
2219 "prfm pldl1keep, [x23, #0x80]\n"
2220 "fmla v18.4s, v29.4s, v2.s[0]\n"
2221 "fmla v22.4s, v29.4s, v3.s[0]\n"
2222 "prfm pldl1keep, [x22, #0x80]\n"
2223 "fmla v26.4s, v29.4s, v4.s[0]\n"
2224 "ldr q29, [x10, #0x40]\n"
2225 "fmla v11.4s, v28.4s, v0.s[0]\n"
2226 "fmla v15.4s, v28.4s, v1.s[0]\n"
2227 "fmla v19.4s, v28.4s, v2.s[0]\n"
2228 "fmla v23.4s, v28.4s, v3.s[0]\n"
2229 "fmla v27.4s, v28.4s, v4.s[0]\n"
2230 "ldr q28, [x10, #0x50]\n"
2231 "fmla v8.4s, v29.4s, v0.s[1]\n"
2232 "fmla v12.4s, v29.4s, v1.s[1]\n"
2233 "fmla v16.4s, v29.4s, v2.s[1]\n"
2234 "fmla v20.4s, v29.4s, v3.s[1]\n"
2235 "fmla v24.4s, v29.4s, v4.s[1]\n"
2236 "ldr q29, [x10, #0x60]\n"
2237 "fmla v9.4s, v28.4s, v0.s[1]\n"
2238 "fmla v13.4s, v28.4s, v1.s[1]\n"
2239 "fmla v17.4s, v28.4s, v2.s[1]\n"
2240 "fmla v21.4s, v28.4s, v3.s[1]\n"
2241 "fmla v25.4s, v28.4s, v4.s[1]\n"
2242 "ldr q28, [x10, #0x70]\n"
2243 "fmla v10.4s, v29.4s, v0.s[1]\n"
2244 "fmla v14.4s, v29.4s, v1.s[1]\n"
2245 "fmla v18.4s, v29.4s, v2.s[1]\n"
2246 "fmla v22.4s, v29.4s, v3.s[1]\n"
2247 "fmla v26.4s, v29.4s, v4.s[1]\n"
2248 "ldr q29, [x10, #0x80]\n"
2249 "fmla v11.4s, v28.4s, v0.s[1]\n"
2250 "fmla v15.4s, v28.4s, v1.s[1]\n"
2251 "fmla v19.4s, v28.4s, v2.s[1]\n"
2252 "fmla v23.4s, v28.4s, v3.s[1]\n"
2253 "fmla v27.4s, v28.4s, v4.s[1]\n"
2254 "ldr q28, [x10, #0x90]\n"
2255 "fmla v8.4s, v29.4s, v0.s[2]\n"
2256 "fmla v12.4s, v29.4s, v1.s[2]\n"
2257 "fmla v16.4s, v29.4s, v2.s[2]\n"
2258 "fmla v20.4s, v29.4s, v3.s[2]\n"
2259 "fmla v24.4s, v29.4s, v4.s[2]\n"
2260 "ldr q29, [x10, #0xa0]\n"
2261 "fmla v9.4s, v28.4s, v0.s[2]\n"
2262 "fmla v13.4s, v28.4s, v1.s[2]\n"
2263 "fmla v17.4s, v28.4s, v2.s[2]\n"
2264 "fmla v21.4s, v28.4s, v3.s[2]\n"
2265 "fmla v25.4s, v28.4s, v4.s[2]\n"
2266 "ldr q28, [x10, #0xb0]\n"
2267 "fmla v10.4s, v29.4s, v0.s[2]\n"
2268 "fmla v14.4s, v29.4s, v1.s[2]\n"
2269 "fmla v18.4s, v29.4s, v2.s[2]\n"
2270 "fmla v22.4s, v29.4s, v3.s[2]\n"
2271 "fmla v26.4s, v29.4s, v4.s[2]\n"
2272 "ldr q29, [x10, #0xc0]\n"
2273 "fmla v11.4s, v28.4s, v0.s[2]\n"
2274 "fmla v15.4s, v28.4s, v1.s[2]\n"
2275 "fmla v19.4s, v28.4s, v2.s[2]\n"
2276 "fmla v23.4s, v28.4s, v3.s[2]\n"
2277 "fmla v27.4s, v28.4s, v4.s[2]\n"
2278 "ldr q28, [x10, #0xd0]\n"
2279 "fmla v8.4s, v29.4s, v0.s[3]\n"
2280 "fmla v12.4s, v29.4s, v1.s[3]\n"
2281 "fmla v16.4s, v29.4s, v2.s[3]\n"
2282 "fmla v20.4s, v29.4s, v3.s[3]\n"
2283 "fmla v24.4s, v29.4s, v4.s[3]\n"
2284 "ldr q29, [x10, #0xe0]\n"
2285 "fmla v9.4s, v28.4s, v0.s[3]\n"
2286 "fmla v13.4s, v28.4s, v1.s[3]\n"
2287 "fmla v17.4s, v28.4s, v2.s[3]\n"
2288 "fmla v21.4s, v28.4s, v3.s[3]\n"
2289 "fmla v25.4s, v28.4s, v4.s[3]\n"
2290 "ldr q28, [x10, #0xf0]\n"
2291 "add x10, x10, #0x100\n"
2292 "fmla v10.4s, v29.4s, v0.s[3]\n"
2293 "fmla v14.4s, v29.4s, v1.s[3]\n"
2294 "fmla v18.4s, v29.4s, v2.s[3]\n"
2295 "fmla v22.4s, v29.4s, v3.s[3]\n"
2296 "fmla v26.4s, v29.4s, v4.s[3]\n"
2297 "fmla v11.4s, v28.4s, v0.s[3]\n"
2298 "fmla v15.4s, v28.4s, v1.s[3]\n"
2299 "fmla v19.4s, v28.4s, v2.s[3]\n"
2300 "fmla v23.4s, v28.4s, v3.s[3]\n"
2301 "fmla v27.4s, v28.4s, v4.s[3]\n"
2305 "ldr s2, [x26], #0x4\n"
2306 "ldr s1, [x25], #0x4\n"
2307 "sub x27, x27, #0x1\n"
2308 "ldr s0, [x24], #0x4\n"
2309 "ldr s31, [x23], #0x4\n"
2310 "ldr s30, [x22], #0x4\n"
2311 "ldr q29, [x10, #0x0]\n"
2312 "fmla v8.4s, v29.4s, v2.s[0]\n"
2313 "fmla v12.4s, v29.4s, v1.s[0]\n"
2314 "ldr q28, [x10, #0x10]\n"
2315 "fmla v16.4s, v29.4s, v0.s[0]\n"
2316 "fmla v20.4s, v29.4s, v31.s[0]\n"
2317 "fmla v24.4s, v29.4s, v30.s[0]\n"
2318 "ldr q29, [x10, #0x20]\n"
2319 "fmla v9.4s, v28.4s, v2.s[0]\n"
2320 "fmla v13.4s, v28.4s, v1.s[0]\n"
2321 "fmla v17.4s, v28.4s, v0.s[0]\n"
2322 "fmla v21.4s, v28.4s, v31.s[0]\n"
2323 "fmla v25.4s, v28.4s, v30.s[0]\n"
2324 "ldr q28, [x10, #0x30]\n"
2325 "add x10, x10, #0x40\n"
2326 "fmla v10.4s, v29.4s, v2.s[0]\n"
2327 "fmla v14.4s, v29.4s, v1.s[0]\n"
2328 "fmla v18.4s, v29.4s, v0.s[0]\n"
2329 "fmla v22.4s, v29.4s, v31.s[0]\n"
2330 "fmla v26.4s, v29.4s, v30.s[0]\n"
2331 "fmla v11.4s, v28.4s, v2.s[0]\n"
2332 "fmla v15.4s, v28.4s, v1.s[0]\n"
2333 "fmla v19.4s, v28.4s, v0.s[0]\n"
2334 "fmla v23.4s, v28.4s, v31.s[0]\n"
2335 "fmla v27.4s, v28.4s, v30.s[0]\n"
2338 "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
2339 "add x28, x28, #0x1\n"
2342 "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
2343 "add x25, x9, x20, LSL #2\n"
2344 "add x24, x25, x20, LSL #2\n"
2345 "prfm pstl1keep, [x9, #0x0]\n"
2346 "add x23, x24, x20, LSL #2\n"
2347 "add x22, x23, x20, LSL #2\n"
2348 "prfm pstl1keep, [x25, #0x0]\n"
2349 "prfm pstl1keep, [x24, #0x0]\n"
2350 "prfm pstl1keep, [x23, #0x0]\n"
2351 "prfm pstl1keep, [x22, #0x0]\n"
2352 "tbz %x[flags], #1, 155f\n"
2353 "add x20, %x[args_ptr], %[offset_max]\n"
2354 "ld1r { v29.4s }, [x20]\n"
2355 "add x20, %x[args_ptr], %[offset_min]\n"
2356 "ld1r { v28.4s }, [x20]\n"
2357 "fmin v8.4s, v8.4s, v29.4s\n"
2358 "fmin v9.4s, v9.4s, v29.4s\n"
2359 "fmin v10.4s, v10.4s, v29.4s\n"
2360 "fmin v11.4s, v11.4s, v29.4s\n"
2361 "fmin v12.4s, v12.4s, v29.4s\n"
2362 "fmin v13.4s, v13.4s, v29.4s\n"
2363 "fmin v14.4s, v14.4s, v29.4s\n"
2364 "fmin v15.4s, v15.4s, v29.4s\n"
2365 "fmin v16.4s, v16.4s, v29.4s\n"
2366 "fmin v17.4s, v17.4s, v29.4s\n"
2367 "fmin v18.4s, v18.4s, v29.4s\n"
2368 "fmin v19.4s, v19.4s, v29.4s\n"
2369 "fmin v20.4s, v20.4s, v29.4s\n"
2370 "fmin v21.4s, v21.4s, v29.4s\n"
2371 "fmin v22.4s, v22.4s, v29.4s\n"
2372 "fmin v23.4s, v23.4s, v29.4s\n"
2373 "fmin v24.4s, v24.4s, v29.4s\n"
2374 "fmin v25.4s, v25.4s, v29.4s\n"
2375 "fmin v26.4s, v26.4s, v29.4s\n"
2376 "fmin v27.4s, v27.4s, v29.4s\n"
2377 "fmax v8.4s, v8.4s, v28.4s\n"
2378 "fmax v9.4s, v9.4s, v28.4s\n"
2379 "fmax v10.4s, v10.4s, v28.4s\n"
2380 "fmax v11.4s, v11.4s, v28.4s\n"
2381 "fmax v12.4s, v12.4s, v28.4s\n"
2382 "fmax v13.4s, v13.4s, v28.4s\n"
2383 "fmax v14.4s, v14.4s, v28.4s\n"
2384 "fmax v15.4s, v15.4s, v28.4s\n"
2385 "fmax v16.4s, v16.4s, v28.4s\n"
2386 "fmax v17.4s, v17.4s, v28.4s\n"
2387 "fmax v18.4s, v18.4s, v28.4s\n"
2388 "fmax v19.4s, v19.4s, v28.4s\n"
2389 "fmax v20.4s, v20.4s, v28.4s\n"
2390 "fmax v21.4s, v21.4s, v28.4s\n"
2391 "fmax v22.4s, v22.4s, v28.4s\n"
2392 "fmax v23.4s, v23.4s, v28.4s\n"
2393 "fmax v24.4s, v24.4s, v28.4s\n"
2394 "fmax v25.4s, v25.4s, v28.4s\n"
2395 "fmax v26.4s, v26.4s, v28.4s\n"
2396 "fmax v27.4s, v27.4s, v28.4s\n"
2400 "tbz x11, #3, 159f\n"
2401 "st1 { v8.4s }, [x9], #0x10\n"
2402 "st1 { v9.4s }, [x9], #0x10\n"
2403 "st1 { v12.4s }, [x25], #0x10\n"
2404 "st1 { v13.4s }, [x25], #0x10\n"
2405 "st1 { v16.4s }, [x24], #0x10\n"
2406 "st1 { v17.4s }, [x24], #0x10\n"
2407 "st1 { v20.4s }, [x23], #0x10\n"
2408 "st1 { v21.4s }, [x23], #0x10\n"
2409 "st1 { v24.4s }, [x22], #0x10\n"
2410 "st1 { v25.4s }, [x22], #0x10\n"
2411 "tbz x11, #2, 157f\n"
2412 "st1 { v10.4s }, [x9], #0x10\n"
2413 "st1 { v14.4s }, [x25], #0x10\n"
2414 "st1 { v18.4s }, [x24], #0x10\n"
2415 "st1 { v22.4s }, [x23], #0x10\n"
2416 "st1 { v26.4s }, [x22], #0x10\n"
2417 "tbz x11, #1, 156f\n"
2418 "str d11, [x9], #0x8\n"
2419 "str d15, [x25], #0x8\n"
2420 "str d19, [x24], #0x8\n"
2421 "str d23, [x23], #0x8\n"
2422 "str d27, [x22], #0x8\n"
2423 "tbz x11, #0, 163f\n"
2424 "st1 { v11.s }[2], [x9]\n"
2425 "st1 { v15.s }[2], [x25]\n"
2426 "st1 { v19.s }[2], [x24]\n"
2427 "st1 { v23.s }[2], [x23]\n"
2428 "st1 { v27.s }[2], [x22]\n"
2431 "tbz x11, #0, 163f\n"
2432 "str s11, [x9, #0x0]\n"
2433 "str s15, [x25, #0x0]\n"
2434 "str s19, [x24, #0x0]\n"
2435 "str s23, [x23, #0x0]\n"
2436 "str s27, [x22, #0x0]\n"
2439 "tbz x11, #1, 158f\n"
2440 "str d10, [x9], #0x8\n"
2441 "str d14, [x25], #0x8\n"
2442 "str d18, [x24], #0x8\n"
2443 "str d22, [x23], #0x8\n"
2444 "str d26, [x22], #0x8\n"
2445 "tbz x11, #0, 163f\n"
2446 "st1 { v10.s }[2], [x9]\n"
2447 "st1 { v14.s }[2], [x25]\n"
2448 "st1 { v18.s }[2], [x24]\n"
2449 "st1 { v22.s }[2], [x23]\n"
2450 "st1 { v26.s }[2], [x22]\n"
2453 "tbz x11, #0, 163f\n"
2454 "str s10, [x9, #0x0]\n"
2455 "str s14, [x25, #0x0]\n"
2456 "str s18, [x24, #0x0]\n"
2457 "str s22, [x23, #0x0]\n"
2458 "str s26, [x22, #0x0]\n"
2461 "tbz x11, #2, 161f\n"
2462 "st1 { v8.4s }, [x9], #0x10\n"
2463 "st1 { v12.4s }, [x25], #0x10\n"
2464 "st1 { v16.4s }, [x24], #0x10\n"
2465 "st1 { v20.4s }, [x23], #0x10\n"
2466 "st1 { v24.4s }, [x22], #0x10\n"
2467 "tbz x11, #1, 160f\n"
2468 "str d9, [x9], #0x8\n"
2469 "str d13, [x25], #0x8\n"
2470 "str d17, [x24], #0x8\n"
2471 "str d21, [x23], #0x8\n"
2472 "str d25, [x22], #0x8\n"
2473 "tbz x11, #0, 163f\n"
2474 "st1 { v9.s }[2], [x9]\n"
2475 "st1 { v13.s }[2], [x25]\n"
2476 "st1 { v17.s }[2], [x24]\n"
2477 "st1 { v21.s }[2], [x23]\n"
2478 "st1 { v25.s }[2], [x22]\n"
2481 "tbz x11, #0, 163f\n"
2482 "str s9, [x9, #0x0]\n"
2483 "str s13, [x25, #0x0]\n"
2484 "str s17, [x24, #0x0]\n"
2485 "str s21, [x23, #0x0]\n"
2486 "str s25, [x22, #0x0]\n"
2489 "tbz x11, #1, 162f\n"
2490 "str d8, [x9], #0x8\n"
2491 "str d12, [x25], #0x8\n"
2492 "str d16, [x24], #0x8\n"
2493 "str d20, [x23], #0x8\n"
2494 "str d24, [x22], #0x8\n"
2495 "tbz x11, #0, 163f\n"
2496 "st1 { v8.s }[2], [x9]\n"
2497 "st1 { v12.s }[2], [x25]\n"
2498 "st1 { v16.s }[2], [x24]\n"
2499 "st1 { v20.s }[2], [x23]\n"
2500 "st1 { v24.s }[2], [x22]\n"
2503 "str s8, [x9, #0x0]\n"
2504 "str s12, [x25, #0x0]\n"
2505 "str s16, [x24, #0x0]\n"
2506 "str s20, [x23, #0x0]\n"
2507 "str s24, [x22, #0x0]\n"
2511 "str q8, [x9, #0x0]\n"
2512 "str q9, [x9, #0x10]\n"
2513 "str q10, [x9, #0x20]\n"
2514 "str q11, [x9, #0x30]\n"
2515 "add x9, x9, #0x40\n"
2516 "str q12, [x25, #0x0]\n"
2517 "str q13, [x25, #0x10]\n"
2518 "str q14, [x25, #0x20]\n"
2519 "str q15, [x25, #0x30]\n"
2520 "str q16, [x24, #0x0]\n"
2521 "str q17, [x24, #0x10]\n"
2522 "str q18, [x24, #0x20]\n"
2523 "str q19, [x24, #0x30]\n"
2524 "str q20, [x23, #0x0]\n"
2525 "str q21, [x23, #0x10]\n"
2526 "str q22, [x23, #0x20]\n"
2527 "str q23, [x23, #0x30]\n"
2528 "str q24, [x22, #0x0]\n"
2529 "str q25, [x22, #0x10]\n"
2530 "str q26, [x22, #0x20]\n"
2531 "str q27, [x22, #0x30]\n"
2533 "subs x11, x11, #0x10\n"
2537 "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
2539 "mov x12, %x[bias]\n"
2540 "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
2541 "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
2542 "mov x9, %x[output_ptr]\n"
2543 "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
2546 "ldr q8, [x12, #0x0]\n"
2547 "ldr q9, [x12, #0x10]\n"
2548 "mov v12.16b, v8.16b\n"
2549 "mov v13.16b, v9.16b\n"
2550 "ldr q10, [x12, #0x20]\n"
2551 "ldr q11, [x12, #0x30]\n"
2552 "mov v14.16b, v10.16b\n"
2553 "mov v15.16b, v11.16b\n"
2554 "mov v16.16b, v8.16b\n"
2555 "mov v17.16b, v9.16b\n"
2556 "add x12, x12, #0x40\n"
2557 "mov v18.16b, v10.16b\n"
2558 "mov v19.16b, v11.16b\n"
2559 "mov v20.16b, v8.16b\n"
2560 "mov v21.16b, v9.16b\n"
2561 "mov v22.16b, v10.16b\n"
2562 "mov v23.16b, v11.16b\n"
2563 "mov v24.16b, v8.16b\n"
2564 "mov v25.16b, v9.16b\n"
2565 "mov v26.16b, v10.16b\n"
2566 "mov v27.16b, v11.16b\n"
2567 "mov v28.16b, v8.16b\n"
2568 "mov v29.16b, v9.16b\n"
2569 "mov v30.16b, v10.16b\n"
2570 "mov v31.16b, v11.16b\n"
2573 "tbz %x[flags], #0, 178f\n"
2574 "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
2575 "add x25, x9, x20, LSL #2\n"
2576 "add x24, x25, x20, LSL #2\n"
2577 "add x23, x24, x20, LSL #2\n"
2578 "add x22, x23, x20, LSL #2\n"
2580 "add x21, x22, x20, LSL #2\n"
2582 "tbz x11, #3, 172f\n"
2583 "ld1 { v8.4s }, [x9], #0x10\n"
2584 "ld1 { v12.4s }, [x25], #0x10\n"
2585 "ld1 { v16.4s }, [x24], #0x10\n"
2586 "ld1 { v20.4s }, [x23], #0x10\n"
2587 "ld1 { v24.4s }, [x22], #0x10\n"
2588 "ld1 { v28.4s }, [x21], #0x10\n"
2589 "ld1 { v9.4s }, [x9], #0x10\n"
2590 "ld1 { v13.4s }, [x25], #0x10\n"
2591 "ld1 { v17.4s }, [x24], #0x10\n"
2592 "ld1 { v21.4s }, [x23], #0x10\n"
2593 "ld1 { v25.4s }, [x22], #0x10\n"
2594 "ld1 { v29.4s }, [x21], #0x10\n"
2595 "tbz x11, #2, 170f\n"
2596 "ld1 { v10.4s }, [x9], #0x10\n"
2597 "ld1 { v14.4s }, [x25], #0x10\n"
2598 "ld1 { v18.4s }, [x24], #0x10\n"
2599 "ld1 { v22.4s }, [x23], #0x10\n"
2600 "ld1 { v26.4s }, [x22], #0x10\n"
2601 "ld1 { v30.4s }, [x21], #0x10\n"
2602 "tbz x11, #1, 169f\n"
2603 "ldr d11, [x9], #0x8\n"
2604 "ldr d15, [x25], #0x8\n"
2606 "ldr d19, [x24], #0x8\n"
2607 "ldr d23, [x23], #0x8\n"
2608 "ldr d27, [x22], #0x8\n"
2609 "ldr d31, [x21], #0x8\n"
2610 "tbz x11, #0, 176f\n"
2611 "ld1 { v11.s }[2], [x9]\n"
2612 "ld1 { v15.s }[2], [x25]\n"
2613 "ld1 { v19.s }[2], [x24]\n"
2614 "ld1 { v23.s }[2], [x23]\n"
2615 "ld1 { v27.s }[2], [x22]\n"
2616 "ld1 { v31.s }[2], [x21]\n"
2620 "tbz x11, #0, 176f\n"
2621 "ldr s11, [x9, #0x0]\n"
2622 "ldr s15, [x25, #0x0]\n"
2623 "ldr s19, [x24, #0x0]\n"
2624 "ldr s23, [x23, #0x0]\n"
2625 "ldr s27, [x22, #0x0]\n"
2626 "ldr s31, [x21, #0x0]\n"
2629 "tbz x11, #1, 171f\n"
2630 "ldr d10, [x9], #0x8\n"
2631 "ldr d14, [x25], #0x8\n"
2633 "ldr d18, [x24], #0x8\n"
2634 "ldr d22, [x23], #0x8\n"
2635 "ldr d26, [x22], #0x8\n"
2636 "ldr d30, [x21], #0x8\n"
2637 "tbz x11, #0, 176f\n"
2638 "ld1 { v10.s }[2], [x9]\n"
2639 "ld1 { v14.s }[2], [x25]\n"
2640 "ld1 { v18.s }[2], [x24]\n"
2641 "ld1 { v22.s }[2], [x23]\n"
2642 "ld1 { v26.s }[2], [x22]\n"
2643 "ld1 { v30.s }[2], [x21]\n"
2647 "tbz x11, #0, 176f\n"
2648 "ldr s10, [x9, #0x0]\n"
2649 "ldr s14, [x25, #0x0]\n"
2650 "ldr s18, [x24, #0x0]\n"
2651 "ldr s22, [x23, #0x0]\n"
2652 "ldr s26, [x22, #0x0]\n"
2653 "ldr s30, [x21, #0x0]\n"
2656 "tbz x11, #2, 174f\n"
2657 "ld1 { v8.4s }, [x9], #0x10\n"
2658 "ld1 { v12.4s }, [x25], #0x10\n"
2659 "ld1 { v16.4s }, [x24], #0x10\n"
2660 "ld1 { v20.4s }, [x23], #0x10\n"
2661 "ld1 { v24.4s }, [x22], #0x10\n"
2662 "ld1 { v28.4s }, [x21], #0x10\n"
2663 "tbz x11, #1, 173f\n"
2664 "ldr d9, [x9], #0x8\n"
2665 "ldr d13, [x25], #0x8\n"
2667 "ldr d17, [x24], #0x8\n"
2668 "ldr d21, [x23], #0x8\n"
2669 "ldr d25, [x22], #0x8\n"
2670 "ldr d29, [x21], #0x8\n"
2671 "tbz x11, #0, 176f\n"
2672 "ld1 { v9.s }[2], [x9]\n"
2673 "ld1 { v13.s }[2], [x25]\n"
2674 "ld1 { v17.s }[2], [x24]\n"
2675 "ld1 { v21.s }[2], [x23]\n"
2676 "ld1 { v25.s }[2], [x22]\n"
2677 "ld1 { v29.s }[2], [x21]\n"
2681 "tbz x11, #0, 176f\n"
2682 "ldr s9, [x9, #0x0]\n"
2683 "ldr s13, [x25, #0x0]\n"
2684 "ldr s17, [x24, #0x0]\n"
2685 "ldr s21, [x23, #0x0]\n"
2686 "ldr s25, [x22, #0x0]\n"
2687 "ldr s29, [x21, #0x0]\n"
2690 "tbz x11, #1, 175f\n"
2691 "ldr d8, [x9], #0x8\n"
2692 "ldr d12, [x25], #0x8\n"
2694 "ldr d16, [x24], #0x8\n"
2695 "ldr d20, [x23], #0x8\n"
2696 "ldr d24, [x22], #0x8\n"
2697 "ldr d28, [x21], #0x8\n"
2698 "tbz x11, #0, 176f\n"
2699 "ld1 { v8.s }[2], [x9]\n"
2700 "ld1 { v12.s }[2], [x25]\n"
2701 "ld1 { v16.s }[2], [x24]\n"
2702 "ld1 { v20.s }[2], [x23]\n"
2703 "ld1 { v24.s }[2], [x22]\n"
2704 "ld1 { v28.s }[2], [x21]\n"
2707 "ldr s8, [x9, #0x0]\n"
2708 "ldr s12, [x25, #0x0]\n"
2710 "ldr s16, [x24, #0x0]\n"
2711 "ldr s20, [x23, #0x0]\n"
2712 "ldr s24, [x22, #0x0]\n"
2713 "ldr s28, [x21, #0x0]\n"
2718 "ldr q8, [x9, #0x0]\n"
2719 "ldr q9, [x9, #0x10]\n"
2720 "ldr q10, [x9, #0x20]\n"
2721 "ldr q11, [x9, #0x30]\n"
2722 "ldr q12, [x25, #0x0]\n"
2723 "ldr q13, [x25, #0x10]\n"
2724 "ldr q14, [x25, #0x20]\n"
2725 "ldr q15, [x25, #0x30]\n"
2726 "ldr q16, [x24, #0x0]\n"
2727 "ldr q17, [x24, #0x10]\n"
2728 "ldr q18, [x24, #0x20]\n"
2729 "ldr q19, [x24, #0x30]\n"
2730 "ldr q20, [x23, #0x0]\n"
2731 "ldr q21, [x23, #0x10]\n"
2732 "ldr q22, [x23, #0x20]\n"
2733 "ldr q23, [x23, #0x30]\n"
2734 "ldr q24, [x22, #0x0]\n"
2735 "ldr q25, [x22, #0x10]\n"
2736 "ldr q26, [x22, #0x20]\n"
2737 "ldr q27, [x22, #0x30]\n"
2738 "ldr q28, [x21, #0x0]\n"
2739 "ldr q29, [x21, #0x10]\n"
2740 "ldr q30, [x21, #0x20]\n"
2741 "ldr q31, [x21, #0x30]\n"
2744 "movi v8.16b, #0x0\n"
2745 "movi v9.16b, #0x0\n"
2746 "movi v10.16b, #0x0\n"
2747 "movi v11.16b, #0x0\n"
2748 "movi v12.16b, #0x0\n"
2749 "movi v13.16b, #0x0\n"
2750 "movi v14.16b, #0x0\n"
2751 "movi v15.16b, #0x0\n"
2752 "movi v16.16b, #0x0\n"
2753 "movi v17.16b, #0x0\n"
2754 "movi v18.16b, #0x0\n"
2755 "movi v19.16b, #0x0\n"
2756 "movi v20.16b, #0x0\n"
2757 "movi v21.16b, #0x0\n"
2758 "movi v22.16b, #0x0\n"
2759 "movi v23.16b, #0x0\n"
2760 "movi v24.16b, #0x0\n"
2761 "movi v25.16b, #0x0\n"
2762 "movi v26.16b, #0x0\n"
2763 "movi v27.16b, #0x0\n"
2764 "movi v28.16b, #0x0\n"
2765 "movi v29.16b, #0x0\n"
2766 "movi v30.16b, #0x0\n"
2767 "movi v31.16b, #0x0\n"
2771 "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
2772 "ldr w27, [x20, x28, LSL #0x2]\n"
2773 "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
2774 "tbz %x[flags], #3, 181f\n"
2775 "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
2776 "add x20, x20, x21, LSL #3\n"
2777 "ldr x26, [x20, #0x0]\n"
2778 "ldr x25, [x20, #0x8]\n"
2779 "ldr x24, [x20, #0x10]\n"
2780 "ldr x23, [x20, #0x18]\n"
2781 "ldr x22, [x20, #0x20]\n"
2782 "ldr x21, [x20, #0x28]\n"
2784 "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
2785 "add x26, x26, x20, LSL #2\n"
2786 "add x25, x25, x20, LSL #2\n"
2787 "add x24, x24, x20, LSL #2\n"
2788 "add x23, x23, x20, LSL #2\n"
2789 "add x22, x22, x20, LSL #2\n"
2790 "add x21, x21, x20, LSL #2\n"
2793 "mov x26, %x[input_ptr]\n"
2794 "add x25, x26, x21, LSL #2\n"
2795 "add x24, x25, x21, LSL #2\n"
2796 "add x23, x24, x21, LSL #2\n"
2797 "add x22, x23, x21, LSL #2\n"
2798 "add x21, x22, x21, LSL #2\n"
2802 "ldr q0, [x26, #0x0]\n"
2803 "ldr q1, [x25, #0x0]\n"
2805 "ldr q2, [x24, #0x0]\n"
2806 "ldr q3, [x23, #0x0]\n"
2807 "ldr q4, [x22, #0x0]\n"
2808 "ldr q5, [x21, #0x0]\n"
2809 "ldr q6, [x10, #0x0]\n"
2810 "ldr q7, [x10, #0x10]\n"
2813 "fmla v8.4s, v6.4s, v0.s[0]\n"
2814 "fmla v12.4s, v6.4s, v1.s[0]\n"
2815 "sub x27, x27, #0x4\n"
2816 "add x26, x26, #0x10\n"
2817 "fmla v16.4s, v6.4s, v2.s[0]\n"
2818 "fmla v20.4s, v6.4s, v3.s[0]\n"
2819 "add x25, x25, #0x10\n"
2820 "add x24, x24, #0x10\n"
2821 "fmla v24.4s, v6.4s, v4.s[0]\n"
2822 "fmla v28.4s, v6.4s, v5.s[0]\n"
2823 "ldr q6, [x10, #0x20]\n"
2824 "add x23, x23, #0x10\n"
2825 "fmla v9.4s, v7.4s, v0.s[0]\n"
2826 "fmla v13.4s, v7.4s, v1.s[0]\n"
2827 "add x22, x22, #0x10\n"
2828 "add x21, x21, #0x10\n"
2829 "fmla v17.4s, v7.4s, v2.s[0]\n"
2830 "fmla v21.4s, v7.4s, v3.s[0]\n"
2832 "prfm pldl1keep, [x26, #0x80]\n"
2833 "fmla v25.4s, v7.4s, v4.s[0]\n"
2834 "fmla v29.4s, v7.4s, v5.s[0]\n"
2835 "ldr q7, [x10, #0x30]\n"
2836 "prfm pldl1keep, [x25, #0x80]\n"
2837 "fmla v10.4s, v6.4s, v0.s[0]\n"
2838 "fmla v14.4s, v6.4s, v1.s[0]\n"
2839 "prfm pldl1keep, [x24, #0x80]\n"
2840 "prfm pldl1keep, [x23, #0x80]\n"
2841 "fmla v18.4s, v6.4s, v2.s[0]\n"
2842 "fmla v22.4s, v6.4s, v3.s[0]\n"
2843 "prfm pldl1keep, [x22, #0x80]\n"
2844 "prfm pldl1keep, [x21, #0x80]\n"
2845 "fmla v26.4s, v6.4s, v4.s[0]\n"
2846 "fmla v30.4s, v6.4s, v5.s[0]\n"
2847 "ldr q6, [x10, #0x40]\n"
2848 "fmla v11.4s, v7.4s, v0.s[0]\n"
2849 "fmla v15.4s, v7.4s, v1.s[0]\n"
2850 "fmla v19.4s, v7.4s, v2.s[0]\n"
2851 "fmla v23.4s, v7.4s, v3.s[0]\n"
2852 "fmla v27.4s, v7.4s, v4.s[0]\n"
2853 "fmla v31.4s, v7.4s, v5.s[0]\n"
2854 "ldr q7, [x10, #0x50]\n"
2855 "fmla v8.4s, v6.4s, v0.s[1]\n"
2856 "fmla v12.4s, v6.4s, v1.s[1]\n"
2857 "fmla v16.4s, v6.4s, v2.s[1]\n"
2858 "fmla v20.4s, v6.4s, v3.s[1]\n"
2859 "fmla v24.4s, v6.4s, v4.s[1]\n"
2860 "fmla v28.4s, v6.4s, v5.s[1]\n"
2861 "ldr q6, [x10, #0x60]\n"
2862 "fmla v9.4s, v7.4s, v0.s[1]\n"
2863 "fmla v13.4s, v7.4s, v1.s[1]\n"
2864 "fmla v17.4s, v7.4s, v2.s[1]\n"
2865 "fmla v21.4s, v7.4s, v3.s[1]\n"
2866 "fmla v25.4s, v7.4s, v4.s[1]\n"
2867 "fmla v29.4s, v7.4s, v5.s[1]\n"
2868 "ldr q7, [x10, #0x70]\n"
2869 "fmla v10.4s, v6.4s, v0.s[1]\n"
2870 "fmla v14.4s, v6.4s, v1.s[1]\n"
2871 "fmla v18.4s, v6.4s, v2.s[1]\n"
2872 "fmla v22.4s, v6.4s, v3.s[1]\n"
2873 "fmla v26.4s, v6.4s, v4.s[1]\n"
2874 "fmla v30.4s, v6.4s, v5.s[1]\n"
2875 "ldr q6, [x10, #0x80]\n"
2876 "fmla v11.4s, v7.4s, v0.s[1]\n"
2877 "fmla v15.4s, v7.4s, v1.s[1]\n"
2878 "fmla v19.4s, v7.4s, v2.s[1]\n"
2879 "fmla v23.4s, v7.4s, v3.s[1]\n"
2880 "fmla v27.4s, v7.4s, v4.s[1]\n"
2881 "fmla v31.4s, v7.4s, v5.s[1]\n"
2882 "ldr q7, [x10, #0x90]\n"
2883 "fmla v8.4s, v6.4s, v0.s[2]\n"
2884 "fmla v12.4s, v6.4s, v1.s[2]\n"
2885 "fmla v16.4s, v6.4s, v2.s[2]\n"
2886 "fmla v20.4s, v6.4s, v3.s[2]\n"
2887 "fmla v24.4s, v6.4s, v4.s[2]\n"
2888 "fmla v28.4s, v6.4s, v5.s[2]\n"
2889 "ldr q6, [x10, #0xa0]\n"
2890 "fmla v9.4s, v7.4s, v0.s[2]\n"
2891 "fmla v13.4s, v7.4s, v1.s[2]\n"
2892 "fmla v17.4s, v7.4s, v2.s[2]\n"
2893 "fmla v21.4s, v7.4s, v3.s[2]\n"
2894 "fmla v25.4s, v7.4s, v4.s[2]\n"
2895 "fmla v29.4s, v7.4s, v5.s[2]\n"
2896 "ldr q7, [x10, #0xb0]\n"
2897 "fmla v10.4s, v6.4s, v0.s[2]\n"
2898 "fmla v14.4s, v6.4s, v1.s[2]\n"
2899 "fmla v18.4s, v6.4s, v2.s[2]\n"
2900 "fmla v22.4s, v6.4s, v3.s[2]\n"
2901 "fmla v26.4s, v6.4s, v4.s[2]\n"
2902 "fmla v30.4s, v6.4s, v5.s[2]\n"
2903 "ldr q6, [x10, #0xc0]\n"
2904 "fmla v11.4s, v7.4s, v0.s[2]\n"
2905 "fmla v15.4s, v7.4s, v1.s[2]\n"
2906 "fmla v19.4s, v7.4s, v2.s[2]\n"
2907 "fmla v23.4s, v7.4s, v3.s[2]\n"
2908 "fmla v27.4s, v7.4s, v4.s[2]\n"
2909 "fmla v31.4s, v7.4s, v5.s[2]\n"
2910 "ldr q7, [x10, #0xd0]\n"
2911 "fmla v8.4s, v6.4s, v0.s[3]\n"
2912 "fmla v12.4s, v6.4s, v1.s[3]\n"
2913 "fmla v16.4s, v6.4s, v2.s[3]\n"
2914 "fmla v20.4s, v6.4s, v3.s[3]\n"
2915 "fmla v24.4s, v6.4s, v4.s[3]\n"
2916 "fmla v28.4s, v6.4s, v5.s[3]\n"
2917 "ldr q6, [x10, #0xe0]\n"
2918 "fmla v9.4s, v7.4s, v0.s[3]\n"
2919 "fmla v13.4s, v7.4s, v1.s[3]\n"
2920 "fmla v17.4s, v7.4s, v2.s[3]\n"
2921 "fmla v21.4s, v7.4s, v3.s[3]\n"
2922 "fmla v25.4s, v7.4s, v4.s[3]\n"
2923 "fmla v29.4s, v7.4s, v5.s[3]\n"
2924 "ldr q7, [x10, #0xf0]\n"
2925 "add x10, x10, #0x100\n"
2926 "fmla v10.4s, v6.4s, v0.s[3]\n"
2927 "fmla v14.4s, v6.4s, v1.s[3]\n"
2928 "fmla v18.4s, v6.4s, v2.s[3]\n"
2929 "fmla v22.4s, v6.4s, v3.s[3]\n"
2930 "fmla v26.4s, v6.4s, v4.s[3]\n"
2931 "fmla v30.4s, v6.4s, v5.s[3]\n"
2932 "ldr q6, [x10, #0x0]\n"
2933 "fmla v11.4s, v7.4s, v0.s[3]\n"
2934 "ldr q0, [x26, #0x0]\n"
2935 "fmla v15.4s, v7.4s, v1.s[3]\n"
2936 "ldr q1, [x25, #0x0]\n"
2937 "fmla v19.4s, v7.4s, v2.s[3]\n"
2938 "ldr q2, [x24, #0x0]\n"
2939 "fmla v23.4s, v7.4s, v3.s[3]\n"
2940 "ldr q3, [x23, #0x0]\n"
2941 "fmla v27.4s, v7.4s, v4.s[3]\n"
2942 "ldr q4, [x22, #0x0]\n"
2943 "fmla v31.4s, v7.4s, v5.s[3]\n"
2944 "ldr q5, [x21, #0x0]\n"
2945 "ldr q7, [x10, #0x10]\n"
2948 "fmla v8.4s, v6.4s, v0.s[0]\n"
2949 "fmla v12.4s, v6.4s, v1.s[0]\n"
2950 "add x26, x26, #0x10\n"
2951 "add x25, x25, #0x10\n"
2952 "fmla v16.4s, v6.4s, v2.s[0]\n"
2953 "fmla v20.4s, v6.4s, v3.s[0]\n"
2954 "add x24, x24, #0x10\n"
2955 "add x23, x23, #0x10\n"
2956 "fmla v24.4s, v6.4s, v4.s[0]\n"
2957 "fmla v28.4s, v6.4s, v5.s[0]\n"
2958 "ldr q6, [x10, #0x20]\n"
2959 "add x22, x22, #0x10\n"
2960 "fmla v9.4s, v7.4s, v0.s[0]\n"
2961 "fmla v13.4s, v7.4s, v1.s[0]\n"
2962 "add x21, x21, #0x10\n"
2963 "sub x27, x27, #0x4\n"
2964 "fmla v17.4s, v7.4s, v2.s[0]\n"
2965 "fmla v21.4s, v7.4s, v3.s[0]\n"
2966 "prfm pldl1keep, [x26, #0x80]\n"
2967 "prfm pldl1keep, [x25, #0x80]\n"
2968 "fmla v25.4s, v7.4s, v4.s[0]\n"
2969 "fmla v29.4s, v7.4s, v5.s[0]\n"
2970 "ldr q7, [x10, #0x30]\n"
2971 "prfm pldl1keep, [x24, #0x80]\n"
2972 "fmla v10.4s, v6.4s, v0.s[0]\n"
2973 "fmla v14.4s, v6.4s, v1.s[0]\n"
2974 "prfm pldl1keep, [x23, #0x80]\n"
2975 "prfm pldl1keep, [x22, #0x80]\n"
2976 "fmla v18.4s, v6.4s, v2.s[0]\n"
2977 "fmla v22.4s, v6.4s, v3.s[0]\n"
2978 "prfm pldl1keep, [x21, #0x80]\n"
2979 "fmla v26.4s, v6.4s, v4.s[0]\n"
2980 "fmla v30.4s, v6.4s, v5.s[0]\n"
2981 "ldr q6, [x10, #0x40]\n"
2982 "fmla v11.4s, v7.4s, v0.s[0]\n"
2983 "fmla v15.4s, v7.4s, v1.s[0]\n"
2984 "fmla v19.4s, v7.4s, v2.s[0]\n"
2985 "fmla v23.4s, v7.4s, v3.s[0]\n"
2986 "fmla v27.4s, v7.4s, v4.s[0]\n"
2987 "fmla v31.4s, v7.4s, v5.s[0]\n"
2988 "ldr q7, [x10, #0x50]\n"
2989 "fmla v8.4s, v6.4s, v0.s[1]\n"
2990 "fmla v12.4s, v6.4s, v1.s[1]\n"
2991 "fmla v16.4s, v6.4s, v2.s[1]\n"
2992 "fmla v20.4s, v6.4s, v3.s[1]\n"
2993 "fmla v24.4s, v6.4s, v4.s[1]\n"
2994 "fmla v28.4s, v6.4s, v5.s[1]\n"
2995 "ldr q6, [x10, #0x60]\n"
2996 "fmla v9.4s, v7.4s, v0.s[1]\n"
2997 "fmla v13.4s, v7.4s, v1.s[1]\n"
2998 "fmla v17.4s, v7.4s, v2.s[1]\n"
2999 "fmla v21.4s, v7.4s, v3.s[1]\n"
3000 "fmla v25.4s, v7.4s, v4.s[1]\n"
3001 "fmla v29.4s, v7.4s, v5.s[1]\n"
3002 "ldr q7, [x10, #0x70]\n"
3003 "fmla v10.4s, v6.4s, v0.s[1]\n"
3004 "fmla v14.4s, v6.4s, v1.s[1]\n"
3005 "fmla v18.4s, v6.4s, v2.s[1]\n"
3006 "fmla v22.4s, v6.4s, v3.s[1]\n"
3007 "fmla v26.4s, v6.4s, v4.s[1]\n"
3008 "fmla v30.4s, v6.4s, v5.s[1]\n"
3009 "ldr q6, [x10, #0x80]\n"
3010 "fmla v11.4s, v7.4s, v0.s[1]\n"
3011 "fmla v15.4s, v7.4s, v1.s[1]\n"
3012 "fmla v19.4s, v7.4s, v2.s[1]\n"
3013 "fmla v23.4s, v7.4s, v3.s[1]\n"
3014 "fmla v27.4s, v7.4s, v4.s[1]\n"
3015 "fmla v31.4s, v7.4s, v5.s[1]\n"
3016 "ldr q7, [x10, #0x90]\n"
3017 "fmla v8.4s, v6.4s, v0.s[2]\n"
3018 "fmla v12.4s, v6.4s, v1.s[2]\n"
3019 "fmla v16.4s, v6.4s, v2.s[2]\n"
3020 "fmla v20.4s, v6.4s, v3.s[2]\n"
3021 "fmla v24.4s, v6.4s, v4.s[2]\n"
3022 "fmla v28.4s, v6.4s, v5.s[2]\n"
3023 "ldr q6, [x10, #0xa0]\n"
3024 "fmla v9.4s, v7.4s, v0.s[2]\n"
3025 "fmla v13.4s, v7.4s, v1.s[2]\n"
3026 "fmla v17.4s, v7.4s, v2.s[2]\n"
3027 "fmla v21.4s, v7.4s, v3.s[2]\n"
3028 "fmla v25.4s, v7.4s, v4.s[2]\n"
3029 "fmla v29.4s, v7.4s, v5.s[2]\n"
3030 "ldr q7, [x10, #0xb0]\n"
3031 "fmla v10.4s, v6.4s, v0.s[2]\n"
3032 "fmla v14.4s, v6.4s, v1.s[2]\n"
3033 "fmla v18.4s, v6.4s, v2.s[2]\n"
3034 "fmla v22.4s, v6.4s, v3.s[2]\n"
3035 "fmla v26.4s, v6.4s, v4.s[2]\n"
3036 "fmla v30.4s, v6.4s, v5.s[2]\n"
3037 "ldr q6, [x10, #0xc0]\n"
3038 "fmla v11.4s, v7.4s, v0.s[2]\n"
3039 "fmla v15.4s, v7.4s, v1.s[2]\n"
3040 "fmla v19.4s, v7.4s, v2.s[2]\n"
3041 "fmla v23.4s, v7.4s, v3.s[2]\n"
3042 "fmla v27.4s, v7.4s, v4.s[2]\n"
3043 "fmla v31.4s, v7.4s, v5.s[2]\n"
3044 "ldr q7, [x10, #0xd0]\n"
3045 "fmla v8.4s, v6.4s, v0.s[3]\n"
3046 "fmla v12.4s, v6.4s, v1.s[3]\n"
3047 "fmla v16.4s, v6.4s, v2.s[3]\n"
3048 "fmla v20.4s, v6.4s, v3.s[3]\n"
3049 "fmla v24.4s, v6.4s, v4.s[3]\n"
3050 "fmla v28.4s, v6.4s, v5.s[3]\n"
3051 "ldr q6, [x10, #0xe0]\n"
3052 "fmla v9.4s, v7.4s, v0.s[3]\n"
3053 "fmla v13.4s, v7.4s, v1.s[3]\n"
3054 "fmla v17.4s, v7.4s, v2.s[3]\n"
3055 "fmla v21.4s, v7.4s, v3.s[3]\n"
3056 "fmla v25.4s, v7.4s, v4.s[3]\n"
3057 "fmla v29.4s, v7.4s, v5.s[3]\n"
3058 "ldr q7, [x10, #0xf0]\n"
3059 "add x10, x10, #0x100\n"
3060 "fmla v10.4s, v6.4s, v0.s[3]\n"
3061 "fmla v14.4s, v6.4s, v1.s[3]\n"
3062 "fmla v18.4s, v6.4s, v2.s[3]\n"
3063 "fmla v22.4s, v6.4s, v3.s[3]\n"
3064 "fmla v26.4s, v6.4s, v4.s[3]\n"
3065 "fmla v30.4s, v6.4s, v5.s[3]\n"
3066 "fmla v11.4s, v7.4s, v0.s[3]\n"
3067 "fmla v15.4s, v7.4s, v1.s[3]\n"
3068 "fmla v19.4s, v7.4s, v2.s[3]\n"
3069 "fmla v23.4s, v7.4s, v3.s[3]\n"
3070 "fmla v27.4s, v7.4s, v4.s[3]\n"
3071 "fmla v31.4s, v7.4s, v5.s[3]\n"
3075 "ldr s7, [x26], #0x4\n"
3076 "ldr s6, [x25], #0x4\n"
3077 "sub x27, x27, #0x1\n"
3078 "ldr s5, [x24], #0x4\n"
3079 "ldr s4, [x23], #0x4\n"
3080 "ldr s3, [x22], #0x4\n"
3081 "ldr s2, [x21], #0x4\n"
3082 "ldr q1, [x10, #0x0]\n"
3083 "ldr q0, [x10, #0x10]\n"
3084 "fmla v8.4s, v1.4s, v7.s[0]\n"
3085 "fmla v12.4s, v1.4s, v6.s[0]\n"
3086 "fmla v16.4s, v1.4s, v5.s[0]\n"
3087 "fmla v20.4s, v1.4s, v4.s[0]\n"
3088 "fmla v24.4s, v1.4s, v3.s[0]\n"
3089 "fmla v28.4s, v1.4s, v2.s[0]\n"
3090 "ldr q1, [x10, #0x20]\n"
3091 "fmla v9.4s, v0.4s, v7.s[0]\n"
3092 "fmla v13.4s, v0.4s, v6.s[0]\n"
3093 "fmla v17.4s, v0.4s, v5.s[0]\n"
3094 "fmla v21.4s, v0.4s, v4.s[0]\n"
3095 "fmla v25.4s, v0.4s, v3.s[0]\n"
3096 "fmla v29.4s, v0.4s, v2.s[0]\n"
3097 "ldr q0, [x10, #0x30]\n"
3098 "add x10, x10, #0x40\n"
3099 "fmla v10.4s, v1.4s, v7.s[0]\n"
3100 "fmla v14.4s, v1.4s, v6.s[0]\n"
3101 "fmla v18.4s, v1.4s, v5.s[0]\n"
3102 "fmla v22.4s, v1.4s, v4.s[0]\n"
3103 "fmla v26.4s, v1.4s, v3.s[0]\n"
3104 "fmla v30.4s, v1.4s, v2.s[0]\n"
3105 "fmla v11.4s, v0.4s, v7.s[0]\n"
3106 "fmla v15.4s, v0.4s, v6.s[0]\n"
3107 "fmla v19.4s, v0.4s, v5.s[0]\n"
3108 "fmla v23.4s, v0.4s, v4.s[0]\n"
3109 "fmla v27.4s, v0.4s, v3.s[0]\n"
3110 "fmla v31.4s, v0.4s, v2.s[0]\n"
3113 "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
3114 "add x28, x28, #0x1\n"
3117 "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
3118 "add x25, x9, x20, LSL #2\n"
3119 "add x24, x25, x20, LSL #2\n"
3120 "prfm pstl1keep, [x9, #0x0]\n"
3121 "add x23, x24, x20, LSL #2\n"
3122 "add x22, x23, x20, LSL #2\n"
3123 "prfm pstl1keep, [x25, #0x0]\n"
3124 "prfm pstl1keep, [x24, #0x0]\n"
3125 "add x21, x22, x20, LSL #2\n"
3126 "prfm pstl1keep, [x23, #0x0]\n"
3127 "prfm pstl1keep, [x22, #0x0]\n"
3128 "prfm pstl1keep, [x21, #0x0]\n"
3129 "tbz %x[flags], #1, 188f\n"
3130 "add x20, %x[args_ptr], %[offset_max]\n"
3131 "ld1r { v1.4s }, [x20]\n"
3132 "add x20, %x[args_ptr], %[offset_min]\n"
3133 "ld1r { v0.4s }, [x20]\n"
3134 "fmin v8.4s, v8.4s, v1.4s\n"
3135 "fmin v9.4s, v9.4s, v1.4s\n"
3136 "fmin v10.4s, v10.4s, v1.4s\n"
3137 "fmin v11.4s, v11.4s, v1.4s\n"
3138 "fmin v12.4s, v12.4s, v1.4s\n"
3139 "fmin v13.4s, v13.4s, v1.4s\n"
3140 "fmin v14.4s, v14.4s, v1.4s\n"
3141 "fmin v15.4s, v15.4s, v1.4s\n"
3142 "fmin v16.4s, v16.4s, v1.4s\n"
3143 "fmin v17.4s, v17.4s, v1.4s\n"
3144 "fmin v18.4s, v18.4s, v1.4s\n"
3145 "fmin v19.4s, v19.4s, v1.4s\n"
3146 "fmin v20.4s, v20.4s, v1.4s\n"
3147 "fmin v21.4s, v21.4s, v1.4s\n"
3148 "fmin v22.4s, v22.4s, v1.4s\n"
3149 "fmin v23.4s, v23.4s, v1.4s\n"
3150 "fmin v24.4s, v24.4s, v1.4s\n"
3151 "fmin v25.4s, v25.4s, v1.4s\n"
3152 "fmin v26.4s, v26.4s, v1.4s\n"
3153 "fmin v27.4s, v27.4s, v1.4s\n"
3154 "fmin v28.4s, v28.4s, v1.4s\n"
3155 "fmin v29.4s, v29.4s, v1.4s\n"
3156 "fmin v30.4s, v30.4s, v1.4s\n"
3157 "fmin v31.4s, v31.4s, v1.4s\n"
3158 "fmax v8.4s, v8.4s, v0.4s\n"
3159 "fmax v9.4s, v9.4s, v0.4s\n"
3160 "fmax v10.4s, v10.4s, v0.4s\n"
3161 "fmax v11.4s, v11.4s, v0.4s\n"
3162 "fmax v12.4s, v12.4s, v0.4s\n"
3163 "fmax v13.4s, v13.4s, v0.4s\n"
3164 "fmax v14.4s, v14.4s, v0.4s\n"
3165 "fmax v15.4s, v15.4s, v0.4s\n"
3166 "fmax v16.4s, v16.4s, v0.4s\n"
3167 "fmax v17.4s, v17.4s, v0.4s\n"
3168 "fmax v18.4s, v18.4s, v0.4s\n"
3169 "fmax v19.4s, v19.4s, v0.4s\n"
3170 "fmax v20.4s, v20.4s, v0.4s\n"
3171 "fmax v21.4s, v21.4s, v0.4s\n"
3172 "fmax v22.4s, v22.4s, v0.4s\n"
3173 "fmax v23.4s, v23.4s, v0.4s\n"
3174 "fmax v24.4s, v24.4s, v0.4s\n"
3175 "fmax v25.4s, v25.4s, v0.4s\n"
3176 "fmax v26.4s, v26.4s, v0.4s\n"
3177 "fmax v27.4s, v27.4s, v0.4s\n"
3178 "fmax v28.4s, v28.4s, v0.4s\n"
3179 "fmax v29.4s, v29.4s, v0.4s\n"
3180 "fmax v30.4s, v30.4s, v0.4s\n"
3181 "fmax v31.4s, v31.4s, v0.4s\n"
3185 "tbz x11, #3, 192f\n"
3186 "st1 { v8.4s }, [x9], #0x10\n"
3187 "st1 { v9.4s }, [x9], #0x10\n"
3188 "st1 { v12.4s }, [x25], #0x10\n"
3189 "st1 { v13.4s }, [x25], #0x10\n"
3190 "st1 { v16.4s }, [x24], #0x10\n"
3191 "st1 { v17.4s }, [x24], #0x10\n"
3192 "st1 { v20.4s }, [x23], #0x10\n"
3193 "st1 { v21.4s }, [x23], #0x10\n"
3194 "st1 { v24.4s }, [x22], #0x10\n"
3195 "st1 { v25.4s }, [x22], #0x10\n"
3196 "st1 { v28.4s }, [x21], #0x10\n"
3197 "st1 { v29.4s }, [x21], #0x10\n"
3198 "tbz x11, #2, 190f\n"
3199 "st1 { v10.4s }, [x9], #0x10\n"
3200 "st1 { v14.4s }, [x25], #0x10\n"
3201 "st1 { v18.4s }, [x24], #0x10\n"
3202 "st1 { v22.4s }, [x23], #0x10\n"
3203 "st1 { v26.4s }, [x22], #0x10\n"
3204 "st1 { v30.4s }, [x21], #0x10\n"
3205 "tbz x11, #1, 189f\n"
3206 "str d11, [x9], #0x8\n"
3207 "str d15, [x25], #0x8\n"
3208 "str d19, [x24], #0x8\n"
3209 "str d23, [x23], #0x8\n"
3210 "str d27, [x22], #0x8\n"
3211 "str d31, [x21], #0x8\n"
3212 "tbz x11, #0, 196f\n"
3213 "st1 { v11.s }[2], [x9]\n"
3214 "st1 { v15.s }[2], [x25]\n"
3215 "st1 { v19.s }[2], [x24]\n"
3216 "st1 { v23.s }[2], [x23]\n"
3217 "st1 { v27.s }[2], [x22]\n"
3218 "st1 { v31.s }[2], [x21]\n"
3221 "tbz x11, #0, 196f\n"
3222 "str s11, [x9, #0x0]\n"
3223 "str s15, [x25, #0x0]\n"
3224 "str s19, [x24, #0x0]\n"
3225 "str s23, [x23, #0x0]\n"
3226 "str s27, [x22, #0x0]\n"
3227 "str s31, [x21, #0x0]\n"
3230 "tbz x11, #1, 191f\n"
3231 "str d10, [x9], #0x8\n"
3232 "str d14, [x25], #0x8\n"
3233 "str d18, [x24], #0x8\n"
3234 "str d22, [x23], #0x8\n"
3235 "str d26, [x22], #0x8\n"
3236 "str d30, [x21], #0x8\n"
3237 "tbz x11, #0, 196f\n"
3238 "st1 { v10.s }[2], [x9]\n"
3239 "st1 { v14.s }[2], [x25]\n"
3240 "st1 { v18.s }[2], [x24]\n"
3241 "st1 { v22.s }[2], [x23]\n"
3242 "st1 { v26.s }[2], [x22]\n"
3243 "st1 { v30.s }[2], [x21]\n"
3246 "tbz x11, #0, 196f\n"
3247 "str s10, [x9, #0x0]\n"
3248 "str s14, [x25, #0x0]\n"
3249 "str s18, [x24, #0x0]\n"
3250 "str s22, [x23, #0x0]\n"
3251 "str s26, [x22, #0x0]\n"
3252 "str s30, [x21, #0x0]\n"
3255 "tbz x11, #2, 194f\n"
3256 "st1 { v8.4s }, [x9], #0x10\n"
3257 "st1 { v12.4s }, [x25], #0x10\n"
3258 "st1 { v16.4s }, [x24], #0x10\n"
3259 "st1 { v20.4s }, [x23], #0x10\n"
3260 "st1 { v24.4s }, [x22], #0x10\n"
3261 "st1 { v28.4s }, [x21], #0x10\n"
3262 "tbz x11, #1, 193f\n"
3263 "str d9, [x9], #0x8\n"
3264 "str d13, [x25], #0x8\n"
3265 "str d17, [x24], #0x8\n"
3266 "str d21, [x23], #0x8\n"
3267 "str d25, [x22], #0x8\n"
3268 "str d29, [x21], #0x8\n"
3269 "tbz x11, #0, 196f\n"
3270 "st1 { v9.s }[2], [x9]\n"
3271 "st1 { v13.s }[2], [x25]\n"
3272 "st1 { v17.s }[2], [x24]\n"
3273 "st1 { v21.s }[2], [x23]\n"
3274 "st1 { v25.s }[2], [x22]\n"
3275 "st1 { v29.s }[2], [x21]\n"
3278 "tbz x11, #0, 196f\n"
3279 "str s9, [x9, #0x0]\n"
3280 "str s13, [x25, #0x0]\n"
3281 "str s17, [x24, #0x0]\n"
3282 "str s21, [x23, #0x0]\n"
3283 "str s25, [x22, #0x0]\n"
3284 "str s29, [x21, #0x0]\n"
3287 "tbz x11, #1, 195f\n"
3288 "str d8, [x9], #0x8\n"
3289 "str d12, [x25], #0x8\n"
3290 "str d16, [x24], #0x8\n"
3291 "str d20, [x23], #0x8\n"
3292 "str d24, [x22], #0x8\n"
3293 "str d28, [x21], #0x8\n"
3294 "tbz x11, #0, 196f\n"
3295 "st1 { v8.s }[2], [x9]\n"
3296 "st1 { v12.s }[2], [x25]\n"
3297 "st1 { v16.s }[2], [x24]\n"
3298 "st1 { v20.s }[2], [x23]\n"
3299 "st1 { v24.s }[2], [x22]\n"
3300 "st1 { v28.s }[2], [x21]\n"
3303 "str s8, [x9, #0x0]\n"
3304 "str s12, [x25, #0x0]\n"
3305 "str s16, [x24, #0x0]\n"
3306 "str s20, [x23, #0x0]\n"
3307 "str s24, [x22, #0x0]\n"
3308 "str s28, [x21, #0x0]\n"
3312 "str q8, [x9, #0x0]\n"
3313 "str q9, [x9, #0x10]\n"
3314 "str q10, [x9, #0x20]\n"
3315 "str q11, [x9, #0x30]\n"
3316 "add x9, x9, #0x40\n"
3317 "str q12, [x25, #0x0]\n"
3318 "str q13, [x25, #0x10]\n"
3319 "str q14, [x25, #0x20]\n"
3320 "str q15, [x25, #0x30]\n"
3321 "str q16, [x24, #0x0]\n"
3322 "str q17, [x24, #0x10]\n"
3323 "str q18, [x24, #0x20]\n"
3324 "str q19, [x24, #0x30]\n"
3325 "str q20, [x23, #0x0]\n"
3326 "str q21, [x23, #0x10]\n"
3327 "str q22, [x23, #0x20]\n"
3328 "str q23, [x23, #0x30]\n"
3329 "str q24, [x22, #0x0]\n"
3330 "str q25, [x22, #0x10]\n"
3331 "str q26, [x22, #0x20]\n"
3332 "str q27, [x22, #0x30]\n"
3333 "str q28, [x21, #0x0]\n"
3334 "str q29, [x21, #0x10]\n"
3335 "str q30, [x21, #0x20]\n"
3336 "str q31, [x21, #0x30]\n"
3338 "subs x11, x11, #0x10\n"
3340 "subs %x[M], %x[M], #0x6\n"
3342 "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
3343 "tbz %x[flags], #3, 199f\n"
3344 "add x21, x21, #0x6\n"
3345 "str x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
3349 "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
3352 : [
M]
"+&r" (
M), [input_ptr]
"+&r" (input_ptr), [output_ptr]
"+&r" (output_ptr)
3353 : [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))
3354 :
"cc",
"memory",
"v0",
"v1",
"v2",
"v3",
"v4",
"v5",
"v6",
"v7",
"v8",
"v9",
"v10",
"v11",
"v12",
"v13",
"v14",
"v15",
"v16",
"v17",
"v18",
"v19",
"v20",
"v21",
"v22",
"v23",
"v24",
"v25",
"v26",
"v27",
"v28",
"v29",
"v30",
"v31",
"x9",
"x10",
"x11",
"x12",
"x20",
"x21",
"x22",
"x23",
"x24",
"x25",
"x26",
"x27",
"x28"
3359 #endif // __aarch64__