24 #if defined(__aarch64__) && (defined(FP16_KERNELS) || defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC))
27 #include "../../utils.hpp"
34 void a64_hybrid_fp16_mla_6x32_a55 (
35 unsigned int num_strings,
const unsigned int *string_lengths, IndirectInputArg<__fp16> A_arg,
36 size_t M,
size_t N,
const __fp16 *B_ptr, IndirectOutputArg<__fp16> output_arg,
41 __fp16 maxval =
static_cast<__fp16
>(std::numeric_limits<float>::infinity());
42 __fp16 minval = -
static_cast<__fp16
>(std::numeric_limits<float>::infinity());
43 unsigned int num_strings = {};
44 const unsigned int *string_lengths = {};
46 const __fp16 *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<__fp16
>(act.param1);
105 "ldr x8, [%x[args_ptr], %[offsetof_N]]\n"
106 "ldr x17, [%x[args_ptr], %[offsetof_B_ptr]]\n"
107 "mov x16, %x[output_ptr]\n"
110 "ldr q8, [x7, #0x0]\n"
111 "ldr q9, [x7, #0x10]\n"
112 "ldr q10, [x7, #0x20]\n"
113 "ldr q11, [x7, #0x30]\n"
114 "add x7, x7, #0x40\n"
117 "tbz %x[flags], #0, 21f\n"
121 "ld1 { v8.8h }, [x16], #0x10\n"
122 "ld1 { v9.8h }, [x16], #0x10\n"
124 "ld1 { v10.8h }, [x16], #0x10\n"
126 "ldr d11, [x16], #0x8\n"
128 "ld1 { v11.s }[2], [x16], #0x4\n"
131 "ld1 { v11.h }[6], [x16]\n"
136 "ld1 { v11.h }[4], [x16]\n"
140 "ldr s11, [x16], #0x4\n"
143 "ld1 { v11.h }[2], [x16]\n"
148 "ldr h11, [x16, #0x0]\n"
152 "ldr d10, [x16], #0x8\n"
154 "ld1 { v10.s }[2], [x16], #0x4\n"
157 "ld1 { v10.h }[6], [x16]\n"
162 "ld1 { v10.h }[4], [x16]\n"
166 "ldr s10, [x16], #0x4\n"
169 "ld1 { v10.h }[2], [x16]\n"
174 "ldr h10, [x16, #0x0]\n"
178 "ld1 { v8.8h }, [x16], #0x10\n"
180 "ldr d9, [x16], #0x8\n"
182 "ld1 { v9.s }[2], [x16], #0x4\n"
185 "ld1 { v9.h }[6], [x16]\n"
190 "ld1 { v9.h }[4], [x16]\n"
194 "ldr s9, [x16], #0x4\n"
197 "ld1 { v9.h }[2], [x16]\n"
202 "ldr h9, [x16, #0x0]\n"
206 "ldr d8, [x16], #0x8\n"
208 "ld1 { v8.s }[2], [x16], #0x4\n"
211 "ld1 { v8.h }[6], [x16]\n"
216 "ld1 { v8.h }[4], [x16]\n"
220 "ldr s8, [x16], #0x4\n"
223 "ld1 { v8.h }[2], [x16]\n"
226 "ldr h8, [x16, #0x0]\n"
229 "sub x16, x16, x20\n"
232 "ldr q8, [x16, #0x0]\n"
233 "ldr q9, [x16, #0x10]\n"
234 "ldr q10, [x16, #0x20]\n"
235 "ldr q11, [x16, #0x30]\n"
238 "movi v8.16b, #0x0\n"
239 "movi v9.16b, #0x0\n"
240 "movi v10.16b, #0x0\n"
241 "movi v11.16b, #0x0\n"
245 "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
246 "ldr w14, [x20, x15, LSL #0x2]\n"
247 "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
248 "tbz %x[flags], #3, 24f\n"
249 "ldr x20, [%x[input_ptr], x15, LSL #0x3]\n"
250 "add x20, x20, x21, LSL #3\n"
251 "ldr x13, [x20, #0x0]\n"
253 "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
254 "add x13, x13, x20, LSL #1\n"
257 "mov x13, %x[input_ptr]\n"
261 "ldr q0, [x13, #0x0]\n"
263 "ldr q6, [x17, #0x0]\n"
264 "ldr q7, [x17, #0x10]\n"
267 "fmla v8.8h, v6.8h, v0.h[0]\n"
268 "ldr d17, [x17, #0x20]\n"
269 "ldr x20, [x17, #0x28]\n"
270 "fmla v9.8h, v7.8h, v0.h[0]\n"
271 "ldr d16, [x17, #0x30]\n"
272 "mov v17.d[1], x20\n"
273 "ldr x20, [x17, #0x38]\n"
274 "mov v16.d[1], x20\n"
275 "fmla v10.8h, v17.8h, v0.h[0]\n"
276 "ldr d17, [x17, #0x40]\n"
277 "ldr x20, [x17, #0x48]\n"
278 "fmla v11.8h, v16.8h, v0.h[0]\n"
279 "ldr d16, [x17, #0x50]\n"
280 "mov v17.d[1], x20\n"
281 "ldr x20, [x17, #0x58]\n"
282 "mov v16.d[1], x20\n"
283 "fmla v8.8h, v17.8h, v0.h[1]\n"
284 "ldr d17, [x17, #0x60]\n"
285 "ldr x20, [x17, #0x68]\n"
286 "fmla v9.8h, v16.8h, v0.h[1]\n"
287 "ldr d16, [x17, #0x70]\n"
288 "mov v17.d[1], x20\n"
289 "ldr x20, [x17, #0x78]\n"
290 "mov v16.d[1], x20\n"
291 "fmla v10.8h, v17.8h, v0.h[1]\n"
292 "ldr d17, [x17, #0x80]\n"
293 "ldr x20, [x17, #0x88]\n"
294 "fmla v11.8h, v16.8h, v0.h[1]\n"
295 "ldr d16, [x17, #0x90]\n"
296 "mov v17.d[1], x20\n"
297 "ldr x20, [x17, #0x98]\n"
298 "mov v16.d[1], x20\n"
299 "fmla v8.8h, v17.8h, v0.h[2]\n"
300 "ldr d17, [x17, #0xa0]\n"
301 "ldr x20, [x17, #0xa8]\n"
302 "fmla v9.8h, v16.8h, v0.h[2]\n"
303 "ldr d16, [x17, #0xb0]\n"
304 "mov v17.d[1], x20\n"
305 "ldr x20, [x17, #0xb8]\n"
306 "mov v16.d[1], x20\n"
307 "fmla v10.8h, v17.8h, v0.h[2]\n"
308 "ldr d17, [x17, #0xc0]\n"
309 "ldr x20, [x17, #0xc8]\n"
310 "fmla v11.8h, v16.8h, v0.h[2]\n"
311 "ldr d16, [x17, #0xd0]\n"
312 "mov v17.d[1], x20\n"
313 "ldr x20, [x17, #0xd8]\n"
314 "mov v16.d[1], x20\n"
315 "fmla v8.8h, v17.8h, v0.h[3]\n"
316 "ldr d17, [x17, #0xe0]\n"
317 "ldr x20, [x17, #0xe8]\n"
318 "fmla v9.8h, v16.8h, v0.h[3]\n"
319 "ldr d16, [x17, #0xf0]\n"
320 "mov v17.d[1], x20\n"
321 "ldr x20, [x17, #0xf8]\n"
322 "mov v16.d[1], x20\n"
323 "fmla v10.8h, v17.8h, v0.h[3]\n"
324 "ldr d17, [x17, #0x100]\n"
325 "ldr x20, [x17, #0x108]\n"
326 "fmla v11.8h, v16.8h, v0.h[3]\n"
327 "ldr d16, [x17, #0x110]\n"
328 "mov v17.d[1], x20\n"
329 "ldr x20, [x17, #0x118]\n"
330 "mov v16.d[1], x20\n"
331 "fmla v8.8h, v17.8h, v0.h[4]\n"
332 "ldr d17, [x17, #0x120]\n"
333 "ldr x20, [x17, #0x128]\n"
334 "fmla v9.8h, v16.8h, v0.h[4]\n"
335 "ldr d16, [x17, #0x130]\n"
336 "mov v17.d[1], x20\n"
337 "ldr x20, [x17, #0x138]\n"
338 "mov v16.d[1], x20\n"
339 "fmla v10.8h, v17.8h, v0.h[4]\n"
340 "ldr d17, [x17, #0x140]\n"
341 "ldr x20, [x17, #0x148]\n"
342 "fmla v11.8h, v16.8h, v0.h[4]\n"
343 "ldr d16, [x17, #0x150]\n"
344 "mov v17.d[1], x20\n"
345 "ldr x20, [x17, #0x158]\n"
346 "mov v16.d[1], x20\n"
347 "fmla v8.8h, v17.8h, v0.h[5]\n"
348 "ldr d17, [x17, #0x160]\n"
349 "ldr x20, [x17, #0x168]\n"
350 "fmla v9.8h, v16.8h, v0.h[5]\n"
351 "ldr d16, [x17, #0x170]\n"
352 "mov v17.d[1], x20\n"
353 "ldr x20, [x17, #0x178]\n"
354 "mov v16.d[1], x20\n"
355 "fmla v10.8h, v17.8h, v0.h[5]\n"
356 "ldr d17, [x17, #0x180]\n"
357 "ldr x20, [x17, #0x188]\n"
358 "fmla v11.8h, v16.8h, v0.h[5]\n"
359 "ldr d16, [x17, #0x190]\n"
360 "mov v17.d[1], x20\n"
361 "ldr x20, [x17, #0x198]\n"
362 "mov v16.d[1], x20\n"
363 "fmla v8.8h, v17.8h, v0.h[6]\n"
364 "ldr d17, [x17, #0x1a0]\n"
365 "ldr x20, [x17, #0x1a8]\n"
366 "fmla v9.8h, v16.8h, v0.h[6]\n"
367 "ldr d16, [x17, #0x1b0]\n"
368 "mov v17.d[1], x20\n"
369 "ldr x20, [x17, #0x1b8]\n"
370 "mov v16.d[1], x20\n"
371 "fmla v10.8h, v17.8h, v0.h[6]\n"
372 "ldr d17, [x17, #0x1c0]\n"
373 "ldr x20, [x17, #0x1c8]\n"
374 "fmla v11.8h, v16.8h, v0.h[6]\n"
375 "ldr d16, [x17, #0x1d0]\n"
376 "mov v17.d[1], x20\n"
377 "ldr x20, [x17, #0x1d8]\n"
378 "mov v16.d[1], x20\n"
379 "fmla v8.8h, v17.8h, v0.h[7]\n"
380 "ldr d17, [x17, #0x1e0]\n"
381 "ldr x20, [x17, #0x1e8]\n"
382 "fmla v9.8h, v16.8h, v0.h[7]\n"
383 "ldr d16, [x17, #0x1f0]\n"
384 "mov v17.d[1], x20\n"
385 "ldr x20, [x17, #0x1f8]\n"
386 "mov v16.d[1], x20\n"
387 "add x13, x13, #0x10\n"
388 "add x17, x17, #0x200\n"
389 "fmla v10.8h, v17.8h, v0.h[7]\n"
390 "ldr d6, [x17, #0x0]\n"
391 "ldr x20, [x17, #0x8]\n"
392 "fmla v11.8h, v16.8h, v0.h[7]\n"
393 "ldr d0, [x13, #0x0]\n"
394 "sub x14, x14, #0x8\n"
395 "ldr d7, [x17, #0x10]\n"
397 "ldr x21, [x13, #0x8]\n"
399 "ldr x20, [x17, #0x18]\n"
402 "prfm pldl1keep, [x13, #0x80]\n"
405 "fmla v8.8h, v6.8h, v0.h[0]\n"
406 "ldr q17, [x17, #0x20]\n"
407 "fmla v9.8h, v7.8h, v0.h[0]\n"
408 "ldr q16, [x17, #0x30]\n"
409 "fmla v10.8h, v17.8h, v0.h[0]\n"
410 "ldr q17, [x17, #0x40]\n"
411 "fmla v11.8h, v16.8h, v0.h[0]\n"
412 "ldr q16, [x17, #0x50]\n"
413 "fmla v8.8h, v17.8h, v0.h[1]\n"
414 "ldr q17, [x17, #0x60]\n"
415 "fmla v9.8h, v16.8h, v0.h[1]\n"
416 "ldr q16, [x17, #0x70]\n"
417 "fmla v10.8h, v17.8h, v0.h[1]\n"
418 "ldr q17, [x17, #0x80]\n"
419 "fmla v11.8h, v16.8h, v0.h[1]\n"
420 "ldr q16, [x17, #0x90]\n"
421 "fmla v8.8h, v17.8h, v0.h[2]\n"
422 "ldr q17, [x17, #0xa0]\n"
423 "fmla v9.8h, v16.8h, v0.h[2]\n"
424 "ldr q16, [x17, #0xb0]\n"
425 "fmla v10.8h, v17.8h, v0.h[2]\n"
426 "ldr q17, [x17, #0xc0]\n"
427 "fmla v11.8h, v16.8h, v0.h[2]\n"
428 "ldr q16, [x17, #0xd0]\n"
429 "fmla v8.8h, v17.8h, v0.h[3]\n"
430 "ldr q17, [x17, #0xe0]\n"
431 "fmla v9.8h, v16.8h, v0.h[3]\n"
432 "ldr q16, [x17, #0xf0]\n"
433 "fmla v10.8h, v17.8h, v0.h[3]\n"
434 "ldr q17, [x17, #0x100]\n"
435 "fmla v11.8h, v16.8h, v0.h[3]\n"
436 "ldr q16, [x17, #0x110]\n"
437 "fmla v8.8h, v17.8h, v0.h[4]\n"
438 "ldr q17, [x17, #0x120]\n"
439 "fmla v9.8h, v16.8h, v0.h[4]\n"
440 "ldr q16, [x17, #0x130]\n"
441 "fmla v10.8h, v17.8h, v0.h[4]\n"
442 "ldr q17, [x17, #0x140]\n"
443 "fmla v11.8h, v16.8h, v0.h[4]\n"
444 "ldr q16, [x17, #0x150]\n"
445 "fmla v8.8h, v17.8h, v0.h[5]\n"
446 "ldr q17, [x17, #0x160]\n"
447 "fmla v9.8h, v16.8h, v0.h[5]\n"
448 "ldr q16, [x17, #0x170]\n"
449 "fmla v10.8h, v17.8h, v0.h[5]\n"
450 "ldr q17, [x17, #0x180]\n"
451 "fmla v11.8h, v16.8h, v0.h[5]\n"
452 "ldr q16, [x17, #0x190]\n"
453 "fmla v8.8h, v17.8h, v0.h[6]\n"
454 "ldr q17, [x17, #0x1a0]\n"
455 "fmla v9.8h, v16.8h, v0.h[6]\n"
456 "ldr q16, [x17, #0x1b0]\n"
457 "fmla v10.8h, v17.8h, v0.h[6]\n"
458 "ldr q17, [x17, #0x1c0]\n"
459 "fmla v11.8h, v16.8h, v0.h[6]\n"
460 "ldr q16, [x17, #0x1d0]\n"
461 "fmla v8.8h, v17.8h, v0.h[7]\n"
462 "ldr q17, [x17, #0x1e0]\n"
463 "fmla v9.8h, v16.8h, v0.h[7]\n"
464 "ldr q16, [x17, #0x1f0]\n"
465 "add x13, x13, #0x10\n"
466 "sub x14, x14, #0x8\n"
467 "fmla v10.8h, v17.8h, v0.h[7]\n"
468 "prfm pldl1keep, [x13, #0x80]\n"
469 "fmla v11.8h, v16.8h, v0.h[7]\n"
470 "add x17, x17, #0x200\n"
474 "ldr h0, [x13], #0x2\n"
475 "sub x14, x14, #0x1\n"
476 "ldr q16, [x17, #0x0]\n"
477 "fmla v8.8h, v16.8h, v0.h[0]\n"
478 "ldr q16, [x17, #0x10]\n"
479 "fmla v9.8h, v16.8h, v0.h[0]\n"
480 "ldr q16, [x17, #0x20]\n"
481 "fmla v10.8h, v16.8h, v0.h[0]\n"
482 "ldr q16, [x17, #0x30]\n"
483 "fmla v11.8h, v16.8h, v0.h[0]\n"
484 "add x17, x17, #0x40\n"
487 "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
488 "add x15, x15, #0x1\n"
491 "prfm pstl1keep, [x16, #0x0]\n"
492 "tbz %x[flags], #1, 31f\n"
493 "add x20, %x[args_ptr], %[offset_max]\n"
494 "ld1r { v16.8h }, [x20]\n"
495 "fmin v8.8h, v8.8h, v16.8h\n"
496 "fmin v9.8h, v9.8h, v16.8h\n"
497 "fmin v10.8h, v10.8h, v16.8h\n"
498 "fmin v11.8h, v11.8h, v16.8h\n"
499 "add x20, %x[args_ptr], %[offset_min]\n"
500 "ld1r { v16.8h }, [x20]\n"
501 "fmax v8.8h, v8.8h, v16.8h\n"
502 "fmax v9.8h, v9.8h, v16.8h\n"
503 "fmax v10.8h, v10.8h, v16.8h\n"
504 "fmax v11.8h, v11.8h, v16.8h\n"
509 "st1 { v8.8h }, [x16], #0x10\n"
510 "st1 { v9.8h }, [x16], #0x10\n"
512 "st1 { v10.8h }, [x16], #0x10\n"
514 "str d11, [x16], #0x8\n"
516 "st1 { v11.s }[2], [x16], #0x4\n"
518 "st1 { v11.h }[6], [x16]\n"
522 "st1 { v11.h }[4], [x16]\n"
526 "str s11, [x16], #0x4\n"
528 "st1 { v11.h }[2], [x16]\n"
532 "str h11, [x16, #0x0]\n"
536 "str d10, [x16], #0x8\n"
538 "st1 { v10.s }[2], [x16], #0x4\n"
540 "st1 { v10.h }[6], [x16]\n"
544 "st1 { v10.h }[4], [x16]\n"
548 "str s10, [x16], #0x4\n"
550 "st1 { v10.h }[2], [x16]\n"
554 "str h10, [x16, #0x0]\n"
558 "st1 { v8.8h }, [x16], #0x10\n"
560 "str d9, [x16], #0x8\n"
562 "st1 { v9.s }[2], [x16], #0x4\n"
564 "st1 { v9.h }[6], [x16]\n"
568 "st1 { v9.h }[4], [x16]\n"
572 "str s9, [x16], #0x4\n"
574 "st1 { v9.h }[2], [x16]\n"
578 "str h9, [x16, #0x0]\n"
582 "str d8, [x16], #0x8\n"
584 "st1 { v8.s }[2], [x16], #0x4\n"
586 "st1 { v8.h }[6], [x16]\n"
590 "st1 { v8.h }[4], [x16]\n"
594 "str s8, [x16], #0x4\n"
596 "st1 { v8.h }[2], [x16]\n"
599 "str h8, [x16, #0x0]\n"
603 "str q8, [x16, #0x0]\n"
604 "str q9, [x16, #0x10]\n"
605 "str q10, [x16, #0x20]\n"
606 "str q11, [x16, #0x30]\n"
607 "add x16, x16, #0x40\n"
609 "subs x8, x8, #0x20\n"
614 "ldr x8, [%x[args_ptr], %[offsetof_N]]\n"
615 "ldr x17, [%x[args_ptr], %[offsetof_B_ptr]]\n"
616 "mov x16, %x[output_ptr]\n"
619 "ldr q8, [x7, #0x0]\n"
620 "mov v12.16b, v8.16b\n"
621 "ldr q9, [x7, #0x10]\n"
622 "mov v13.16b, v9.16b\n"
623 "ldr q10, [x7, #0x20]\n"
624 "mov v14.16b, v10.16b\n"
625 "ldr q11, [x7, #0x30]\n"
626 "mov v15.16b, v11.16b\n"
627 "add x7, x7, #0x40\n"
630 "tbz %x[flags], #0, 70f\n"
631 "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
633 "add x25, x16, x20, LSL #1\n"
636 "ld1 { v8.8h }, [x16], #0x10\n"
637 "ld1 { v12.8h }, [x25], #0x10\n"
638 "ld1 { v9.8h }, [x16], #0x10\n"
639 "ld1 { v13.8h }, [x25], #0x10\n"
641 "ld1 { v10.8h }, [x16], #0x10\n"
642 "ld1 { v14.8h }, [x25], #0x10\n"
644 "ldr d11, [x16], #0x8\n"
645 "ldr d15, [x25], #0x8\n"
647 "ld1 { v11.s }[2], [x16], #0x4\n"
649 "ld1 { v15.s }[2], [x25], #0x4\n"
651 "ld1 { v11.h }[6], [x16]\n"
652 "ld1 { v15.h }[6], [x25]\n"
657 "ld1 { v11.h }[4], [x16]\n"
658 "ld1 { v15.h }[4], [x25]\n"
662 "ldr s11, [x16], #0x4\n"
664 "ldr s15, [x25], #0x4\n"
666 "ld1 { v11.h }[2], [x16]\n"
667 "ld1 { v15.h }[2], [x25]\n"
672 "ldr h11, [x16, #0x0]\n"
673 "ldr h15, [x25, #0x0]\n"
677 "ldr d10, [x16], #0x8\n"
678 "ldr d14, [x25], #0x8\n"
680 "ld1 { v10.s }[2], [x16], #0x4\n"
682 "ld1 { v14.s }[2], [x25], #0x4\n"
684 "ld1 { v10.h }[6], [x16]\n"
685 "ld1 { v14.h }[6], [x25]\n"
690 "ld1 { v10.h }[4], [x16]\n"
691 "ld1 { v14.h }[4], [x25]\n"
695 "ldr s10, [x16], #0x4\n"
697 "ldr s14, [x25], #0x4\n"
699 "ld1 { v10.h }[2], [x16]\n"
700 "ld1 { v14.h }[2], [x25]\n"
705 "ldr h10, [x16, #0x0]\n"
706 "ldr h14, [x25, #0x0]\n"
710 "ld1 { v8.8h }, [x16], #0x10\n"
711 "ld1 { v12.8h }, [x25], #0x10\n"
713 "ldr d9, [x16], #0x8\n"
714 "ldr d13, [x25], #0x8\n"
716 "ld1 { v9.s }[2], [x16], #0x4\n"
718 "ld1 { v13.s }[2], [x25], #0x4\n"
720 "ld1 { v9.h }[6], [x16]\n"
721 "ld1 { v13.h }[6], [x25]\n"
726 "ld1 { v9.h }[4], [x16]\n"
727 "ld1 { v13.h }[4], [x25]\n"
731 "ldr s9, [x16], #0x4\n"
733 "ldr s13, [x25], #0x4\n"
735 "ld1 { v9.h }[2], [x16]\n"
736 "ld1 { v13.h }[2], [x25]\n"
741 "ldr h9, [x16, #0x0]\n"
742 "ldr h13, [x25, #0x0]\n"
746 "ldr d8, [x16], #0x8\n"
747 "ldr d12, [x25], #0x8\n"
749 "ld1 { v8.s }[2], [x16], #0x4\n"
751 "ld1 { v12.s }[2], [x25], #0x4\n"
753 "ld1 { v8.h }[6], [x16]\n"
754 "ld1 { v12.h }[6], [x25]\n"
759 "ld1 { v8.h }[4], [x16]\n"
760 "ld1 { v12.h }[4], [x25]\n"
764 "ldr s8, [x16], #0x4\n"
766 "ldr s12, [x25], #0x4\n"
768 "ld1 { v8.h }[2], [x16]\n"
769 "ld1 { v12.h }[2], [x25]\n"
772 "ldr h8, [x16, #0x0]\n"
774 "ldr h12, [x25, #0x0]\n"
776 "sub x16, x16, x20\n"
779 "ldr q8, [x16, #0x0]\n"
780 "ldr q9, [x16, #0x10]\n"
781 "ldr q10, [x16, #0x20]\n"
782 "ldr q11, [x16, #0x30]\n"
783 "ldr q12, [x25, #0x0]\n"
784 "ldr q13, [x25, #0x10]\n"
785 "ldr q14, [x25, #0x20]\n"
786 "ldr q15, [x25, #0x30]\n"
789 "movi v8.16b, #0x0\n"
790 "movi v9.16b, #0x0\n"
791 "movi v10.16b, #0x0\n"
792 "movi v11.16b, #0x0\n"
793 "movi v12.16b, #0x0\n"
794 "movi v13.16b, #0x0\n"
795 "movi v14.16b, #0x0\n"
796 "movi v15.16b, #0x0\n"
800 "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
801 "ldr w14, [x20, x15, LSL #0x2]\n"
802 "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
803 "tbz %x[flags], #3, 73f\n"
804 "ldr x20, [%x[input_ptr], x15, LSL #0x3]\n"
805 "add x20, x20, x21, LSL #3\n"
806 "ldr x13, [x20, #0x0]\n"
807 "ldr x12, [x20, #0x8]\n"
809 "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
810 "add x13, x13, x20, LSL #1\n"
811 "add x12, x12, x20, LSL #1\n"
814 "mov x13, %x[input_ptr]\n"
815 "add x12, x13, x21, LSL #1\n"
819 "ldr q0, [x13, #0x0]\n"
821 "ldr q1, [x12, #0x0]\n"
822 "ldr q6, [x17, #0x0]\n"
823 "ldr q7, [x17, #0x10]\n"
826 "fmla v8.8h, v6.8h, v0.h[0]\n"
827 "ldr x21, [x17, #0x28]\n"
828 "fmla v12.8h, v6.8h, v1.h[0]\n"
829 "ldr d17, [x17, #0x20]\n"
830 "fmla v9.8h, v7.8h, v0.h[0]\n"
831 "ldr x20, [x17, #0x38]\n"
832 "fmla v13.8h, v7.8h, v1.h[0]\n"
833 "ldr d16, [x17, #0x30]\n"
834 "mov v17.d[1], x21\n"
835 "fmla v10.8h, v17.8h, v0.h[0]\n"
836 "mov v16.d[1], x20\n"
837 "fmla v14.8h, v17.8h, v1.h[0]\n"
838 "ldr d17, [x17, #0x40]\n"
839 "fmla v11.8h, v16.8h, v0.h[0]\n"
840 "ldr x20, [x17, #0x48]\n"
841 "fmla v15.8h, v16.8h, v1.h[0]\n"
842 "ldr d16, [x17, #0x50]\n"
843 "mov v17.d[1], x20\n"
844 "ldr x20, [x17, #0x58]\n"
845 "mov v16.d[1], x20\n"
846 "fmla v8.8h, v17.8h, v0.h[1]\n"
847 "ldr x21, [x17, #0x68]\n"
848 "fmla v12.8h, v17.8h, v1.h[1]\n"
849 "ldr d17, [x17, #0x60]\n"
850 "fmla v9.8h, v16.8h, v0.h[1]\n"
851 "ldr x20, [x17, #0x78]\n"
852 "fmla v13.8h, v16.8h, v1.h[1]\n"
853 "ldr d16, [x17, #0x70]\n"
854 "mov v17.d[1], x21\n"
855 "fmla v10.8h, v17.8h, v0.h[1]\n"
856 "mov v16.d[1], x20\n"
857 "fmla v14.8h, v17.8h, v1.h[1]\n"
858 "ldr d17, [x17, #0x80]\n"
859 "fmla v11.8h, v16.8h, v0.h[1]\n"
860 "ldr x20, [x17, #0x88]\n"
861 "fmla v15.8h, v16.8h, v1.h[1]\n"
862 "ldr d16, [x17, #0x90]\n"
863 "mov v17.d[1], x20\n"
864 "ldr x20, [x17, #0x98]\n"
865 "mov v16.d[1], x20\n"
866 "fmla v8.8h, v17.8h, v0.h[2]\n"
867 "ldr x21, [x17, #0xa8]\n"
868 "fmla v12.8h, v17.8h, v1.h[2]\n"
869 "ldr d17, [x17, #0xa0]\n"
870 "fmla v9.8h, v16.8h, v0.h[2]\n"
871 "ldr x20, [x17, #0xb8]\n"
872 "fmla v13.8h, v16.8h, v1.h[2]\n"
873 "ldr d16, [x17, #0xb0]\n"
874 "mov v17.d[1], x21\n"
875 "fmla v10.8h, v17.8h, v0.h[2]\n"
876 "mov v16.d[1], x20\n"
877 "fmla v14.8h, v17.8h, v1.h[2]\n"
878 "ldr d17, [x17, #0xc0]\n"
879 "fmla v11.8h, v16.8h, v0.h[2]\n"
880 "ldr x20, [x17, #0xc8]\n"
881 "fmla v15.8h, v16.8h, v1.h[2]\n"
882 "ldr d16, [x17, #0xd0]\n"
883 "mov v17.d[1], x20\n"
884 "ldr x20, [x17, #0xd8]\n"
885 "mov v16.d[1], x20\n"
886 "fmla v8.8h, v17.8h, v0.h[3]\n"
887 "ldr x21, [x17, #0xe8]\n"
888 "fmla v12.8h, v17.8h, v1.h[3]\n"
889 "ldr d17, [x17, #0xe0]\n"
890 "fmla v9.8h, v16.8h, v0.h[3]\n"
891 "ldr x20, [x17, #0xf8]\n"
892 "fmla v13.8h, v16.8h, v1.h[3]\n"
893 "ldr d16, [x17, #0xf0]\n"
894 "mov v17.d[1], x21\n"
895 "fmla v10.8h, v17.8h, v0.h[3]\n"
896 "mov v16.d[1], x20\n"
897 "fmla v14.8h, v17.8h, v1.h[3]\n"
898 "ldr d17, [x17, #0x100]\n"
899 "fmla v11.8h, v16.8h, v0.h[3]\n"
900 "ldr x20, [x17, #0x108]\n"
901 "fmla v15.8h, v16.8h, v1.h[3]\n"
902 "ldr d16, [x17, #0x110]\n"
903 "mov v17.d[1], x20\n"
904 "ldr x20, [x17, #0x118]\n"
905 "mov v16.d[1], x20\n"
906 "fmla v8.8h, v17.8h, v0.h[4]\n"
907 "ldr x21, [x17, #0x128]\n"
908 "fmla v12.8h, v17.8h, v1.h[4]\n"
909 "ldr d17, [x17, #0x120]\n"
910 "fmla v9.8h, v16.8h, v0.h[4]\n"
911 "ldr x20, [x17, #0x138]\n"
912 "fmla v13.8h, v16.8h, v1.h[4]\n"
913 "ldr d16, [x17, #0x130]\n"
914 "mov v17.d[1], x21\n"
915 "fmla v10.8h, v17.8h, v0.h[4]\n"
916 "mov v16.d[1], x20\n"
917 "fmla v14.8h, v17.8h, v1.h[4]\n"
918 "ldr d17, [x17, #0x140]\n"
919 "fmla v11.8h, v16.8h, v0.h[4]\n"
920 "ldr x20, [x17, #0x148]\n"
921 "fmla v15.8h, v16.8h, v1.h[4]\n"
922 "ldr d16, [x17, #0x150]\n"
923 "mov v17.d[1], x20\n"
924 "ldr x20, [x17, #0x158]\n"
925 "mov v16.d[1], x20\n"
926 "fmla v8.8h, v17.8h, v0.h[5]\n"
927 "ldr x21, [x17, #0x168]\n"
928 "fmla v12.8h, v17.8h, v1.h[5]\n"
929 "ldr d17, [x17, #0x160]\n"
930 "fmla v9.8h, v16.8h, v0.h[5]\n"
931 "ldr x20, [x17, #0x178]\n"
932 "fmla v13.8h, v16.8h, v1.h[5]\n"
933 "ldr d16, [x17, #0x170]\n"
934 "mov v17.d[1], x21\n"
935 "fmla v10.8h, v17.8h, v0.h[5]\n"
936 "mov v16.d[1], x20\n"
937 "fmla v14.8h, v17.8h, v1.h[5]\n"
938 "ldr d17, [x17, #0x180]\n"
939 "fmla v11.8h, v16.8h, v0.h[5]\n"
940 "ldr x20, [x17, #0x188]\n"
941 "fmla v15.8h, v16.8h, v1.h[5]\n"
942 "ldr d16, [x17, #0x190]\n"
943 "mov v17.d[1], x20\n"
944 "ldr x20, [x17, #0x198]\n"
945 "mov v16.d[1], x20\n"
946 "fmla v8.8h, v17.8h, v0.h[6]\n"
947 "ldr x21, [x17, #0x1a8]\n"
948 "fmla v12.8h, v17.8h, v1.h[6]\n"
949 "ldr d17, [x17, #0x1a0]\n"
950 "fmla v9.8h, v16.8h, v0.h[6]\n"
951 "ldr x20, [x17, #0x1b8]\n"
952 "fmla v13.8h, v16.8h, v1.h[6]\n"
953 "ldr d16, [x17, #0x1b0]\n"
954 "mov v17.d[1], x21\n"
955 "fmla v10.8h, v17.8h, v0.h[6]\n"
956 "mov v16.d[1], x20\n"
957 "fmla v14.8h, v17.8h, v1.h[6]\n"
958 "ldr d17, [x17, #0x1c0]\n"
959 "fmla v11.8h, v16.8h, v0.h[6]\n"
960 "ldr x20, [x17, #0x1c8]\n"
961 "fmla v15.8h, v16.8h, v1.h[6]\n"
962 "ldr d16, [x17, #0x1d0]\n"
963 "mov v17.d[1], x20\n"
964 "ldr x20, [x17, #0x1d8]\n"
965 "mov v16.d[1], x20\n"
966 "fmla v8.8h, v17.8h, v0.h[7]\n"
967 "ldr x21, [x17, #0x1e8]\n"
968 "fmla v12.8h, v17.8h, v1.h[7]\n"
969 "ldr d17, [x17, #0x1e0]\n"
970 "fmla v9.8h, v16.8h, v0.h[7]\n"
971 "ldr x20, [x17, #0x1f8]\n"
972 "fmla v13.8h, v16.8h, v1.h[7]\n"
973 "ldr d16, [x17, #0x1f0]\n"
974 "mov v17.d[1], x21\n"
975 "add x13, x13, #0x10\n"
976 "mov v16.d[1], x20\n"
977 "add x12, x12, #0x10\n"
978 "add x17, x17, #0x200\n"
979 "fmla v10.8h, v17.8h, v0.h[7]\n"
980 "fmla v14.8h, v17.8h, v1.h[7]\n"
981 "ldr d6, [x17, #0x0]\n"
982 "ldr x21, [x17, #0x8]\n"
983 "fmla v11.8h, v16.8h, v0.h[7]\n"
984 "ldr d0, [x13, #0x0]\n"
985 "fmla v15.8h, v16.8h, v1.h[7]\n"
986 "ldr d1, [x12, #0x0]\n"
987 "sub x14, x14, #0x8\n"
988 "ldr d7, [x17, #0x10]\n"
990 "ldr x20, [x13, #0x8]\n"
992 "ldr x21, [x12, #0x8]\n"
994 "ldr x20, [x17, #0x18]\n"
996 "prfm pldl1keep, [x13, #0x80]\n"
998 "prfm pldl1keep, [x12, #0x80]\n"
1001 "fmla v8.8h, v6.8h, v0.h[0]\n"
1002 "add x13, x13, #0x10\n"
1003 "fmla v12.8h, v6.8h, v1.h[0]\n"
1004 "ldr q17, [x17, #0x20]\n"
1005 "fmla v9.8h, v7.8h, v0.h[0]\n"
1006 "add x12, x12, #0x10\n"
1007 "fmla v13.8h, v7.8h, v1.h[0]\n"
1008 "ldr q16, [x17, #0x30]\n"
1009 "fmla v10.8h, v17.8h, v0.h[0]\n"
1010 "sub x14, x14, #0x8\n"
1011 "fmla v14.8h, v17.8h, v1.h[0]\n"
1012 "ldr q17, [x17, #0x40]\n"
1013 "fmla v11.8h, v16.8h, v0.h[0]\n"
1014 "prfm pldl1keep, [x13, #0x80]\n"
1015 "fmla v15.8h, v16.8h, v1.h[0]\n"
1016 "ldr q16, [x17, #0x50]\n"
1017 "fmla v8.8h, v17.8h, v0.h[1]\n"
1018 "prfm pldl1keep, [x12, #0x80]\n"
1019 "fmla v12.8h, v17.8h, v1.h[1]\n"
1020 "ldr q17, [x17, #0x60]\n"
1021 "fmla v9.8h, v16.8h, v0.h[1]\n"
1022 "fmla v13.8h, v16.8h, v1.h[1]\n"
1023 "ldr q16, [x17, #0x70]\n"
1024 "fmla v10.8h, v17.8h, v0.h[1]\n"
1025 "fmla v14.8h, v17.8h, v1.h[1]\n"
1026 "ldr q17, [x17, #0x80]\n"
1027 "fmla v11.8h, v16.8h, v0.h[1]\n"
1028 "fmla v15.8h, v16.8h, v1.h[1]\n"
1029 "ldr q16, [x17, #0x90]\n"
1030 "fmla v8.8h, v17.8h, v0.h[2]\n"
1031 "fmla v12.8h, v17.8h, v1.h[2]\n"
1032 "ldr q17, [x17, #0xa0]\n"
1033 "fmla v9.8h, v16.8h, v0.h[2]\n"
1034 "fmla v13.8h, v16.8h, v1.h[2]\n"
1035 "ldr q16, [x17, #0xb0]\n"
1036 "fmla v10.8h, v17.8h, v0.h[2]\n"
1037 "fmla v14.8h, v17.8h, v1.h[2]\n"
1038 "ldr q17, [x17, #0xc0]\n"
1039 "fmla v11.8h, v16.8h, v0.h[2]\n"
1040 "fmla v15.8h, v16.8h, v1.h[2]\n"
1041 "ldr q16, [x17, #0xd0]\n"
1042 "fmla v8.8h, v17.8h, v0.h[3]\n"
1043 "fmla v12.8h, v17.8h, v1.h[3]\n"
1044 "ldr q17, [x17, #0xe0]\n"
1045 "fmla v9.8h, v16.8h, v0.h[3]\n"
1046 "fmla v13.8h, v16.8h, v1.h[3]\n"
1047 "ldr q16, [x17, #0xf0]\n"
1048 "fmla v10.8h, v17.8h, v0.h[3]\n"
1049 "fmla v14.8h, v17.8h, v1.h[3]\n"
1050 "ldr q17, [x17, #0x100]\n"
1051 "fmla v11.8h, v16.8h, v0.h[3]\n"
1052 "fmla v15.8h, v16.8h, v1.h[3]\n"
1053 "ldr q16, [x17, #0x110]\n"
1054 "fmla v8.8h, v17.8h, v0.h[4]\n"
1055 "fmla v12.8h, v17.8h, v1.h[4]\n"
1056 "ldr q17, [x17, #0x120]\n"
1057 "fmla v9.8h, v16.8h, v0.h[4]\n"
1058 "fmla v13.8h, v16.8h, v1.h[4]\n"
1059 "ldr q16, [x17, #0x130]\n"
1060 "fmla v10.8h, v17.8h, v0.h[4]\n"
1061 "fmla v14.8h, v17.8h, v1.h[4]\n"
1062 "ldr q17, [x17, #0x140]\n"
1063 "fmla v11.8h, v16.8h, v0.h[4]\n"
1064 "fmla v15.8h, v16.8h, v1.h[4]\n"
1065 "ldr q16, [x17, #0x150]\n"
1066 "fmla v8.8h, v17.8h, v0.h[5]\n"
1067 "fmla v12.8h, v17.8h, v1.h[5]\n"
1068 "ldr q17, [x17, #0x160]\n"
1069 "fmla v9.8h, v16.8h, v0.h[5]\n"
1070 "fmla v13.8h, v16.8h, v1.h[5]\n"
1071 "ldr q16, [x17, #0x170]\n"
1072 "fmla v10.8h, v17.8h, v0.h[5]\n"
1073 "fmla v14.8h, v17.8h, v1.h[5]\n"
1074 "ldr q17, [x17, #0x180]\n"
1075 "fmla v11.8h, v16.8h, v0.h[5]\n"
1076 "fmla v15.8h, v16.8h, v1.h[5]\n"
1077 "ldr q16, [x17, #0x190]\n"
1078 "fmla v8.8h, v17.8h, v0.h[6]\n"
1079 "fmla v12.8h, v17.8h, v1.h[6]\n"
1080 "ldr q17, [x17, #0x1a0]\n"
1081 "fmla v9.8h, v16.8h, v0.h[6]\n"
1082 "fmla v13.8h, v16.8h, v1.h[6]\n"
1083 "ldr q16, [x17, #0x1b0]\n"
1084 "fmla v10.8h, v17.8h, v0.h[6]\n"
1085 "fmla v14.8h, v17.8h, v1.h[6]\n"
1086 "ldr q17, [x17, #0x1c0]\n"
1087 "fmla v11.8h, v16.8h, v0.h[6]\n"
1088 "fmla v15.8h, v16.8h, v1.h[6]\n"
1089 "ldr q16, [x17, #0x1d0]\n"
1090 "fmla v8.8h, v17.8h, v0.h[7]\n"
1091 "fmla v12.8h, v17.8h, v1.h[7]\n"
1092 "ldr q17, [x17, #0x1e0]\n"
1093 "fmla v9.8h, v16.8h, v0.h[7]\n"
1094 "fmla v13.8h, v16.8h, v1.h[7]\n"
1095 "ldr q16, [x17, #0x1f0]\n"
1096 "fmla v10.8h, v17.8h, v0.h[7]\n"
1097 "add x17, x17, #0x200\n"
1098 "fmla v14.8h, v17.8h, v1.h[7]\n"
1099 "fmla v11.8h, v16.8h, v0.h[7]\n"
1100 "fmla v15.8h, v16.8h, v1.h[7]\n"
1104 "ldr h1, [x13], #0x2\n"
1105 "sub x14, x14, #0x1\n"
1106 "ldr h0, [x12], #0x2\n"
1107 "ldr q17, [x17, #0x0]\n"
1108 "fmla v8.8h, v17.8h, v1.h[0]\n"
1109 "ldr q16, [x17, #0x10]\n"
1110 "fmla v12.8h, v17.8h, v0.h[0]\n"
1111 "ldr q17, [x17, #0x20]\n"
1112 "fmla v9.8h, v16.8h, v1.h[0]\n"
1113 "fmla v13.8h, v16.8h, v0.h[0]\n"
1114 "ldr q16, [x17, #0x30]\n"
1115 "fmla v10.8h, v17.8h, v1.h[0]\n"
1116 "add x17, x17, #0x40\n"
1117 "fmla v14.8h, v17.8h, v0.h[0]\n"
1118 "fmla v11.8h, v16.8h, v1.h[0]\n"
1119 "fmla v15.8h, v16.8h, v0.h[0]\n"
1122 "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
1123 "add x15, x15, #0x1\n"
1126 "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
1127 "add x25, x16, x20, LSL #1\n"
1128 "prfm pstl1keep, [x16, #0x0]\n"
1129 "prfm pstl1keep, [x25, #0x0]\n"
1130 "tbz %x[flags], #1, 80f\n"
1131 "add x20, %x[args_ptr], %[offset_max]\n"
1132 "ld1r { v16.8h }, [x20]\n"
1133 "fmin v8.8h, v8.8h, v16.8h\n"
1134 "fmin v9.8h, v9.8h, v16.8h\n"
1135 "fmin v10.8h, v10.8h, v16.8h\n"
1136 "fmin v11.8h, v11.8h, v16.8h\n"
1137 "fmin v12.8h, v12.8h, v16.8h\n"
1138 "fmin v13.8h, v13.8h, v16.8h\n"
1139 "fmin v14.8h, v14.8h, v16.8h\n"
1140 "fmin v15.8h, v15.8h, v16.8h\n"
1141 "add x20, %x[args_ptr], %[offset_min]\n"
1142 "ld1r { v16.8h }, [x20]\n"
1143 "fmax v8.8h, v8.8h, v16.8h\n"
1144 "fmax v9.8h, v9.8h, v16.8h\n"
1145 "fmax v10.8h, v10.8h, v16.8h\n"
1146 "fmax v11.8h, v11.8h, v16.8h\n"
1147 "fmax v12.8h, v12.8h, v16.8h\n"
1148 "fmax v13.8h, v13.8h, v16.8h\n"
1149 "fmax v14.8h, v14.8h, v16.8h\n"
1150 "fmax v15.8h, v15.8h, v16.8h\n"
1155 "st1 { v8.8h }, [x16], #0x10\n"
1156 "st1 { v9.8h }, [x16], #0x10\n"
1157 "st1 { v12.8h }, [x25], #0x10\n"
1158 "st1 { v13.8h }, [x25], #0x10\n"
1160 "st1 { v10.8h }, [x16], #0x10\n"
1161 "st1 { v14.8h }, [x25], #0x10\n"
1163 "str d11, [x16], #0x8\n"
1164 "str d15, [x25], #0x8\n"
1166 "st1 { v11.s }[2], [x16], #0x4\n"
1167 "st1 { v15.s }[2], [x25], #0x4\n"
1169 "st1 { v11.h }[6], [x16]\n"
1170 "st1 { v15.h }[6], [x25]\n"
1174 "st1 { v11.h }[4], [x16]\n"
1175 "st1 { v15.h }[4], [x25]\n"
1179 "str s11, [x16], #0x4\n"
1180 "str s15, [x25], #0x4\n"
1182 "st1 { v11.h }[2], [x16]\n"
1183 "st1 { v15.h }[2], [x25]\n"
1187 "str h11, [x16, #0x0]\n"
1188 "str h15, [x25, #0x0]\n"
1192 "str d10, [x16], #0x8\n"
1193 "str d14, [x25], #0x8\n"
1195 "st1 { v10.s }[2], [x16], #0x4\n"
1196 "st1 { v14.s }[2], [x25], #0x4\n"
1198 "st1 { v10.h }[6], [x16]\n"
1199 "st1 { v14.h }[6], [x25]\n"
1203 "st1 { v10.h }[4], [x16]\n"
1204 "st1 { v14.h }[4], [x25]\n"
1208 "str s10, [x16], #0x4\n"
1209 "str s14, [x25], #0x4\n"
1211 "st1 { v10.h }[2], [x16]\n"
1212 "st1 { v14.h }[2], [x25]\n"
1216 "str h10, [x16, #0x0]\n"
1217 "str h14, [x25, #0x0]\n"
1221 "st1 { v8.8h }, [x16], #0x10\n"
1222 "st1 { v12.8h }, [x25], #0x10\n"
1224 "str d9, [x16], #0x8\n"
1225 "str d13, [x25], #0x8\n"
1227 "st1 { v9.s }[2], [x16], #0x4\n"
1228 "st1 { v13.s }[2], [x25], #0x4\n"
1230 "st1 { v9.h }[6], [x16]\n"
1231 "st1 { v13.h }[6], [x25]\n"
1235 "st1 { v9.h }[4], [x16]\n"
1236 "st1 { v13.h }[4], [x25]\n"
1240 "str s9, [x16], #0x4\n"
1241 "str s13, [x25], #0x4\n"
1243 "st1 { v9.h }[2], [x16]\n"
1244 "st1 { v13.h }[2], [x25]\n"
1248 "str h9, [x16, #0x0]\n"
1249 "str h13, [x25, #0x0]\n"
1253 "str d8, [x16], #0x8\n"
1254 "str d12, [x25], #0x8\n"
1256 "st1 { v8.s }[2], [x16], #0x4\n"
1257 "st1 { v12.s }[2], [x25], #0x4\n"
1259 "st1 { v8.h }[6], [x16]\n"
1260 "st1 { v12.h }[6], [x25]\n"
1264 "st1 { v8.h }[4], [x16]\n"
1265 "st1 { v12.h }[4], [x25]\n"
1269 "str s8, [x16], #0x4\n"
1270 "str s12, [x25], #0x4\n"
1272 "st1 { v8.h }[2], [x16]\n"
1273 "st1 { v12.h }[2], [x25]\n"
1276 "str h8, [x16, #0x0]\n"
1277 "str h12, [x25, #0x0]\n"
1281 "str q8, [x16, #0x0]\n"
1282 "str q9, [x16, #0x10]\n"
1283 "str q10, [x16, #0x20]\n"
1284 "str q11, [x16, #0x30]\n"
1285 "add x16, x16, #0x40\n"
1286 "str q12, [x25, #0x0]\n"
1287 "str q13, [x25, #0x10]\n"
1288 "str q14, [x25, #0x20]\n"
1289 "str q15, [x25, #0x30]\n"
1291 "subs x8, x8, #0x20\n"
1295 "mov x7, %x[bias]\n"
1296 "ldr x8, [%x[args_ptr], %[offsetof_N]]\n"
1297 "ldr x17, [%x[args_ptr], %[offsetof_B_ptr]]\n"
1298 "mov x16, %x[output_ptr]\n"
1301 "ldr q8, [x7, #0x0]\n"
1302 "mov v12.16b, v8.16b\n"
1303 "ldr q9, [x7, #0x10]\n"
1304 "mov v13.16b, v9.16b\n"
1305 "ldr q10, [x7, #0x20]\n"
1306 "mov v14.16b, v10.16b\n"
1307 "ldr q11, [x7, #0x30]\n"
1308 "mov v15.16b, v11.16b\n"
1309 "mov v16.16b, v8.16b\n"
1310 "add x7, x7, #0x40\n"
1311 "mov v17.16b, v9.16b\n"
1312 "mov v18.16b, v10.16b\n"
1313 "mov v19.16b, v11.16b\n"
1316 "tbz %x[flags], #0, 119f\n"
1317 "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
1318 "add x25, x16, x20, LSL #1\n"
1320 "add x24, x25, x20, LSL #1\n"
1322 "tbz x8, #4, 109f\n"
1323 "ld1 { v8.8h }, [x16], #0x10\n"
1324 "ld1 { v12.8h }, [x25], #0x10\n"
1325 "ld1 { v16.8h }, [x24], #0x10\n"
1326 "ld1 { v9.8h }, [x16], #0x10\n"
1327 "ld1 { v13.8h }, [x25], #0x10\n"
1328 "ld1 { v17.8h }, [x24], #0x10\n"
1329 "tbz x8, #3, 105f\n"
1330 "ld1 { v10.8h }, [x16], #0x10\n"
1331 "ld1 { v14.8h }, [x25], #0x10\n"
1332 "ld1 { v18.8h }, [x24], #0x10\n"
1333 "tbz x8, #2, 103f\n"
1334 "ldr d11, [x16], #0x8\n"
1335 "ldr d15, [x25], #0x8\n"
1336 "ldr d19, [x24], #0x8\n"
1337 "tbz x8, #1, 102f\n"
1338 "ld1 { v11.s }[2], [x16], #0x4\n"
1340 "ld1 { v15.s }[2], [x25], #0x4\n"
1341 "ld1 { v19.s }[2], [x24], #0x4\n"
1342 "tbz x8, #0, 117f\n"
1343 "ld1 { v11.h }[6], [x16]\n"
1344 "ld1 { v15.h }[6], [x25]\n"
1345 "ld1 { v19.h }[6], [x24]\n"
1349 "tbz x8, #0, 117f\n"
1350 "ld1 { v11.h }[4], [x16]\n"
1351 "ld1 { v15.h }[4], [x25]\n"
1352 "ld1 { v19.h }[4], [x24]\n"
1355 "tbz x8, #1, 104f\n"
1356 "ldr s11, [x16], #0x4\n"
1358 "ldr s15, [x25], #0x4\n"
1359 "ldr s19, [x24], #0x4\n"
1360 "tbz x8, #0, 117f\n"
1361 "ld1 { v11.h }[2], [x16]\n"
1362 "ld1 { v15.h }[2], [x25]\n"
1363 "ld1 { v19.h }[2], [x24]\n"
1367 "tbz x8, #0, 117f\n"
1368 "ldr h11, [x16, #0x0]\n"
1369 "ldr h15, [x25, #0x0]\n"
1370 "ldr h19, [x24, #0x0]\n"
1373 "tbz x8, #2, 107f\n"
1374 "ldr d10, [x16], #0x8\n"
1375 "ldr d14, [x25], #0x8\n"
1376 "ldr d18, [x24], #0x8\n"
1377 "tbz x8, #1, 106f\n"
1378 "ld1 { v10.s }[2], [x16], #0x4\n"
1380 "ld1 { v14.s }[2], [x25], #0x4\n"
1381 "ld1 { v18.s }[2], [x24], #0x4\n"
1382 "tbz x8, #0, 117f\n"
1383 "ld1 { v10.h }[6], [x16]\n"
1384 "ld1 { v14.h }[6], [x25]\n"
1385 "ld1 { v18.h }[6], [x24]\n"
1389 "tbz x8, #0, 117f\n"
1390 "ld1 { v10.h }[4], [x16]\n"
1391 "ld1 { v14.h }[4], [x25]\n"
1392 "ld1 { v18.h }[4], [x24]\n"
1395 "tbz x8, #1, 108f\n"
1396 "ldr s10, [x16], #0x4\n"
1398 "ldr s14, [x25], #0x4\n"
1399 "ldr s18, [x24], #0x4\n"
1400 "tbz x8, #0, 117f\n"
1401 "ld1 { v10.h }[2], [x16]\n"
1402 "ld1 { v14.h }[2], [x25]\n"
1403 "ld1 { v18.h }[2], [x24]\n"
1407 "tbz x8, #0, 117f\n"
1408 "ldr h10, [x16, #0x0]\n"
1409 "ldr h14, [x25, #0x0]\n"
1410 "ldr h18, [x24, #0x0]\n"
1413 "tbz x8, #3, 113f\n"
1414 "ld1 { v8.8h }, [x16], #0x10\n"
1415 "ld1 { v12.8h }, [x25], #0x10\n"
1416 "ld1 { v16.8h }, [x24], #0x10\n"
1417 "tbz x8, #2, 111f\n"
1418 "ldr d9, [x16], #0x8\n"
1419 "ldr d13, [x25], #0x8\n"
1420 "ldr d17, [x24], #0x8\n"
1421 "tbz x8, #1, 110f\n"
1422 "ld1 { v9.s }[2], [x16], #0x4\n"
1424 "ld1 { v13.s }[2], [x25], #0x4\n"
1425 "ld1 { v17.s }[2], [x24], #0x4\n"
1426 "tbz x8, #0, 117f\n"
1427 "ld1 { v9.h }[6], [x16]\n"
1428 "ld1 { v13.h }[6], [x25]\n"
1429 "ld1 { v17.h }[6], [x24]\n"
1433 "tbz x8, #0, 117f\n"
1434 "ld1 { v9.h }[4], [x16]\n"
1435 "ld1 { v13.h }[4], [x25]\n"
1436 "ld1 { v17.h }[4], [x24]\n"
1439 "tbz x8, #1, 112f\n"
1440 "ldr s9, [x16], #0x4\n"
1442 "ldr s13, [x25], #0x4\n"
1443 "ldr s17, [x24], #0x4\n"
1444 "tbz x8, #0, 117f\n"
1445 "ld1 { v9.h }[2], [x16]\n"
1446 "ld1 { v13.h }[2], [x25]\n"
1447 "ld1 { v17.h }[2], [x24]\n"
1451 "tbz x8, #0, 117f\n"
1452 "ldr h9, [x16, #0x0]\n"
1453 "ldr h13, [x25, #0x0]\n"
1454 "ldr h17, [x24, #0x0]\n"
1457 "tbz x8, #2, 115f\n"
1458 "ldr d8, [x16], #0x8\n"
1459 "ldr d12, [x25], #0x8\n"
1460 "ldr d16, [x24], #0x8\n"
1461 "tbz x8, #1, 114f\n"
1462 "ld1 { v8.s }[2], [x16], #0x4\n"
1464 "ld1 { v12.s }[2], [x25], #0x4\n"
1465 "ld1 { v16.s }[2], [x24], #0x4\n"
1466 "tbz x8, #0, 117f\n"
1467 "ld1 { v8.h }[6], [x16]\n"
1468 "ld1 { v12.h }[6], [x25]\n"
1469 "ld1 { v16.h }[6], [x24]\n"
1473 "tbz x8, #0, 117f\n"
1474 "ld1 { v8.h }[4], [x16]\n"
1475 "ld1 { v12.h }[4], [x25]\n"
1476 "ld1 { v16.h }[4], [x24]\n"
1479 "tbz x8, #1, 116f\n"
1480 "ldr s8, [x16], #0x4\n"
1482 "ldr s12, [x25], #0x4\n"
1483 "ldr s16, [x24], #0x4\n"
1484 "tbz x8, #0, 117f\n"
1485 "ld1 { v8.h }[2], [x16]\n"
1486 "ld1 { v12.h }[2], [x25]\n"
1487 "ld1 { v16.h }[2], [x24]\n"
1490 "ldr h8, [x16, #0x0]\n"
1492 "ldr h12, [x25, #0x0]\n"
1493 "ldr h16, [x24, #0x0]\n"
1495 "sub x16, x16, x20\n"
1498 "ldr q8, [x16, #0x0]\n"
1499 "ldr q9, [x16, #0x10]\n"
1500 "ldr q10, [x16, #0x20]\n"
1501 "ldr q11, [x16, #0x30]\n"
1502 "ldr q12, [x25, #0x0]\n"
1503 "ldr q13, [x25, #0x10]\n"
1504 "ldr q14, [x25, #0x20]\n"
1505 "ldr q15, [x25, #0x30]\n"
1506 "ldr q16, [x24, #0x0]\n"
1507 "ldr q17, [x24, #0x10]\n"
1508 "ldr q18, [x24, #0x20]\n"
1509 "ldr q19, [x24, #0x30]\n"
1512 "movi v8.16b, #0x0\n"
1513 "movi v9.16b, #0x0\n"
1514 "movi v10.16b, #0x0\n"
1515 "movi v11.16b, #0x0\n"
1516 "movi v12.16b, #0x0\n"
1517 "movi v13.16b, #0x0\n"
1518 "movi v14.16b, #0x0\n"
1519 "movi v15.16b, #0x0\n"
1520 "movi v16.16b, #0x0\n"
1521 "movi v17.16b, #0x0\n"
1522 "movi v18.16b, #0x0\n"
1523 "movi v19.16b, #0x0\n"
1527 "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
1528 "ldr w14, [x20, x15, LSL #0x2]\n"
1529 "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
1530 "tbz %x[flags], #3, 122f\n"
1531 "ldr x20, [%x[input_ptr], x15, LSL #0x3]\n"
1532 "add x20, x20, x21, LSL #3\n"
1533 "ldr x13, [x20, #0x0]\n"
1534 "ldr x12, [x20, #0x8]\n"
1535 "ldr x11, [x20, #0x10]\n"
1537 "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
1538 "add x13, x13, x20, LSL #1\n"
1539 "add x12, x12, x20, LSL #1\n"
1540 "add x11, x11, x20, LSL #1\n"
1543 "mov x13, %x[input_ptr]\n"
1544 "add x12, x13, x21, LSL #1\n"
1545 "add x11, x12, x21, LSL #1\n"
1549 "ldr q0, [x13, #0x0]\n"
1551 "ldr q1, [x12, #0x0]\n"
1552 "ldr q2, [x11, #0x0]\n"
1553 "ldr q6, [x17, #0x0]\n"
1554 "ldr q7, [x17, #0x10]\n"
1557 "fmla v8.8h, v6.8h, v0.h[0]\n"
1558 "ldr x21, [x17, #0x28]\n"
1559 "fmla v12.8h, v6.8h, v1.h[0]\n"
1560 "ldr x20, [x17, #0x38]\n"
1561 "fmla v16.8h, v6.8h, v2.h[0]\n"
1562 "ldr d21, [x17, #0x20]\n"
1563 "fmla v9.8h, v7.8h, v0.h[0]\n"
1564 "mov v21.d[1], x21\n"
1565 "fmla v13.8h, v7.8h, v1.h[0]\n"
1566 "ldr x21, [x17, #0x48]\n"
1567 "fmla v17.8h, v7.8h, v2.h[0]\n"
1568 "ldr d20, [x17, #0x30]\n"
1569 "mov v20.d[1], x20\n"
1570 "fmla v10.8h, v21.8h, v0.h[0]\n"
1571 "fmla v14.8h, v21.8h, v1.h[0]\n"
1572 "ldr x20, [x17, #0x58]\n"
1573 "fmla v18.8h, v21.8h, v2.h[0]\n"
1574 "ldr d21, [x17, #0x40]\n"
1575 "fmla v11.8h, v20.8h, v0.h[0]\n"
1576 "mov v21.d[1], x21\n"
1577 "fmla v15.8h, v20.8h, v1.h[0]\n"
1578 "ldr x21, [x17, #0x68]\n"
1579 "fmla v19.8h, v20.8h, v2.h[0]\n"
1580 "ldr d20, [x17, #0x50]\n"
1581 "mov v20.d[1], x20\n"
1582 "fmla v8.8h, v21.8h, v0.h[1]\n"
1583 "fmla v12.8h, v21.8h, v1.h[1]\n"
1584 "ldr x20, [x17, #0x78]\n"
1585 "fmla v16.8h, v21.8h, v2.h[1]\n"
1586 "ldr d21, [x17, #0x60]\n"
1587 "fmla v9.8h, v20.8h, v0.h[1]\n"
1588 "mov v21.d[1], x21\n"
1589 "fmla v13.8h, v20.8h, v1.h[1]\n"
1590 "ldr x21, [x17, #0x88]\n"
1591 "fmla v17.8h, v20.8h, v2.h[1]\n"
1592 "ldr d20, [x17, #0x70]\n"
1593 "mov v20.d[1], x20\n"
1594 "fmla v10.8h, v21.8h, v0.h[1]\n"
1595 "fmla v14.8h, v21.8h, v1.h[1]\n"
1596 "ldr x20, [x17, #0x98]\n"
1597 "fmla v18.8h, v21.8h, v2.h[1]\n"
1598 "ldr d21, [x17, #0x80]\n"
1599 "fmla v11.8h, v20.8h, v0.h[1]\n"
1600 "mov v21.d[1], x21\n"
1601 "fmla v15.8h, v20.8h, v1.h[1]\n"
1602 "ldr x21, [x17, #0xa8]\n"
1603 "fmla v19.8h, v20.8h, v2.h[1]\n"
1604 "ldr d20, [x17, #0x90]\n"
1605 "mov v20.d[1], x20\n"
1606 "fmla v8.8h, v21.8h, v0.h[2]\n"
1607 "fmla v12.8h, v21.8h, v1.h[2]\n"
1608 "ldr x20, [x17, #0xb8]\n"
1609 "fmla v16.8h, v21.8h, v2.h[2]\n"
1610 "ldr d21, [x17, #0xa0]\n"
1611 "fmla v9.8h, v20.8h, v0.h[2]\n"
1612 "mov v21.d[1], x21\n"
1613 "fmla v13.8h, v20.8h, v1.h[2]\n"
1614 "ldr x21, [x17, #0xc8]\n"
1615 "fmla v17.8h, v20.8h, v2.h[2]\n"
1616 "ldr d20, [x17, #0xb0]\n"
1617 "mov v20.d[1], x20\n"
1618 "fmla v10.8h, v21.8h, v0.h[2]\n"
1619 "fmla v14.8h, v21.8h, v1.h[2]\n"
1620 "ldr x20, [x17, #0xd8]\n"
1621 "fmla v18.8h, v21.8h, v2.h[2]\n"
1622 "ldr d21, [x17, #0xc0]\n"
1623 "fmla v11.8h, v20.8h, v0.h[2]\n"
1624 "mov v21.d[1], x21\n"
1625 "fmla v15.8h, v20.8h, v1.h[2]\n"
1626 "ldr x21, [x17, #0xe8]\n"
1627 "fmla v19.8h, v20.8h, v2.h[2]\n"
1628 "ldr d20, [x17, #0xd0]\n"
1629 "mov v20.d[1], x20\n"
1630 "fmla v8.8h, v21.8h, v0.h[3]\n"
1631 "fmla v12.8h, v21.8h, v1.h[3]\n"
1632 "ldr x20, [x17, #0xf8]\n"
1633 "fmla v16.8h, v21.8h, v2.h[3]\n"
1634 "ldr d21, [x17, #0xe0]\n"
1635 "fmla v9.8h, v20.8h, v0.h[3]\n"
1636 "mov v21.d[1], x21\n"
1637 "fmla v13.8h, v20.8h, v1.h[3]\n"
1638 "ldr x21, [x17, #0x108]\n"
1639 "fmla v17.8h, v20.8h, v2.h[3]\n"
1640 "ldr d20, [x17, #0xf0]\n"
1641 "mov v20.d[1], x20\n"
1642 "fmla v10.8h, v21.8h, v0.h[3]\n"
1643 "fmla v14.8h, v21.8h, v1.h[3]\n"
1644 "ldr x20, [x17, #0x118]\n"
1645 "fmla v18.8h, v21.8h, v2.h[3]\n"
1646 "ldr d21, [x17, #0x100]\n"
1647 "fmla v11.8h, v20.8h, v0.h[3]\n"
1648 "mov v21.d[1], x21\n"
1649 "fmla v15.8h, v20.8h, v1.h[3]\n"
1650 "ldr x21, [x17, #0x128]\n"
1651 "fmla v19.8h, v20.8h, v2.h[3]\n"
1652 "ldr d20, [x17, #0x110]\n"
1653 "mov v20.d[1], x20\n"
1654 "fmla v8.8h, v21.8h, v0.h[4]\n"
1655 "fmla v12.8h, v21.8h, v1.h[4]\n"
1656 "ldr x20, [x17, #0x138]\n"
1657 "fmla v16.8h, v21.8h, v2.h[4]\n"
1658 "ldr d21, [x17, #0x120]\n"
1659 "fmla v9.8h, v20.8h, v0.h[4]\n"
1660 "mov v21.d[1], x21\n"
1661 "fmla v13.8h, v20.8h, v1.h[4]\n"
1662 "ldr x21, [x17, #0x148]\n"
1663 "fmla v17.8h, v20.8h, v2.h[4]\n"
1664 "ldr d20, [x17, #0x130]\n"
1665 "mov v20.d[1], x20\n"
1666 "fmla v10.8h, v21.8h, v0.h[4]\n"
1667 "fmla v14.8h, v21.8h, v1.h[4]\n"
1668 "ldr x20, [x17, #0x158]\n"
1669 "fmla v18.8h, v21.8h, v2.h[4]\n"
1670 "ldr d21, [x17, #0x140]\n"
1671 "fmla v11.8h, v20.8h, v0.h[4]\n"
1672 "mov v21.d[1], x21\n"
1673 "fmla v15.8h, v20.8h, v1.h[4]\n"
1674 "ldr x21, [x17, #0x168]\n"
1675 "fmla v19.8h, v20.8h, v2.h[4]\n"
1676 "ldr d20, [x17, #0x150]\n"
1677 "mov v20.d[1], x20\n"
1678 "fmla v8.8h, v21.8h, v0.h[5]\n"
1679 "fmla v12.8h, v21.8h, v1.h[5]\n"
1680 "ldr x20, [x17, #0x178]\n"
1681 "fmla v16.8h, v21.8h, v2.h[5]\n"
1682 "ldr d21, [x17, #0x160]\n"
1683 "fmla v9.8h, v20.8h, v0.h[5]\n"
1684 "mov v21.d[1], x21\n"
1685 "fmla v13.8h, v20.8h, v1.h[5]\n"
1686 "ldr x21, [x17, #0x188]\n"
1687 "fmla v17.8h, v20.8h, v2.h[5]\n"
1688 "ldr d20, [x17, #0x170]\n"
1689 "mov v20.d[1], x20\n"
1690 "fmla v10.8h, v21.8h, v0.h[5]\n"
1691 "fmla v14.8h, v21.8h, v1.h[5]\n"
1692 "ldr x20, [x17, #0x198]\n"
1693 "fmla v18.8h, v21.8h, v2.h[5]\n"
1694 "ldr d21, [x17, #0x180]\n"
1695 "fmla v11.8h, v20.8h, v0.h[5]\n"
1696 "mov v21.d[1], x21\n"
1697 "fmla v15.8h, v20.8h, v1.h[5]\n"
1698 "ldr x21, [x17, #0x1a8]\n"
1699 "fmla v19.8h, v20.8h, v2.h[5]\n"
1700 "ldr d20, [x17, #0x190]\n"
1701 "mov v20.d[1], x20\n"
1702 "fmla v8.8h, v21.8h, v0.h[6]\n"
1703 "fmla v12.8h, v21.8h, v1.h[6]\n"
1704 "ldr x20, [x17, #0x1b8]\n"
1705 "fmla v16.8h, v21.8h, v2.h[6]\n"
1706 "ldr d21, [x17, #0x1a0]\n"
1707 "fmla v9.8h, v20.8h, v0.h[6]\n"
1708 "mov v21.d[1], x21\n"
1709 "fmla v13.8h, v20.8h, v1.h[6]\n"
1710 "ldr x21, [x17, #0x1c8]\n"
1711 "fmla v17.8h, v20.8h, v2.h[6]\n"
1712 "ldr d20, [x17, #0x1b0]\n"
1713 "mov v20.d[1], x20\n"
1714 "fmla v10.8h, v21.8h, v0.h[6]\n"
1715 "fmla v14.8h, v21.8h, v1.h[6]\n"
1716 "ldr x20, [x17, #0x1d8]\n"
1717 "fmla v18.8h, v21.8h, v2.h[6]\n"
1718 "ldr d21, [x17, #0x1c0]\n"
1719 "fmla v11.8h, v20.8h, v0.h[6]\n"
1720 "mov v21.d[1], x21\n"
1721 "fmla v15.8h, v20.8h, v1.h[6]\n"
1722 "ldr x21, [x17, #0x1e8]\n"
1723 "fmla v19.8h, v20.8h, v2.h[6]\n"
1724 "ldr d20, [x17, #0x1d0]\n"
1725 "mov v20.d[1], x20\n"
1726 "fmla v8.8h, v21.8h, v0.h[7]\n"
1727 "fmla v12.8h, v21.8h, v1.h[7]\n"
1728 "ldr x20, [x17, #0x1f8]\n"
1729 "fmla v16.8h, v21.8h, v2.h[7]\n"
1730 "ldr d21, [x17, #0x1e0]\n"
1731 "fmla v9.8h, v20.8h, v0.h[7]\n"
1732 "mov v21.d[1], x21\n"
1733 "fmla v13.8h, v20.8h, v1.h[7]\n"
1734 "add x13, x13, #0x10\n"
1735 "fmla v17.8h, v20.8h, v2.h[7]\n"
1736 "ldr d20, [x17, #0x1f0]\n"
1737 "mov v20.d[1], x20\n"
1738 "add x12, x12, #0x10\n"
1739 "add x11, x11, #0x10\n"
1740 "add x17, x17, #0x200\n"
1741 "fmla v10.8h, v21.8h, v0.h[7]\n"
1742 "ldr x20, [x17, #0x8]\n"
1743 "fmla v14.8h, v21.8h, v1.h[7]\n"
1744 "ldr x23, [x13, #0x8]\n"
1745 "fmla v18.8h, v21.8h, v2.h[7]\n"
1746 "ldr d6, [x17, #0x0]\n"
1747 "fmla v11.8h, v20.8h, v0.h[7]\n"
1748 "ldr d0, [x13, #0x0]\n"
1749 "fmla v15.8h, v20.8h, v1.h[7]\n"
1750 "ldr d1, [x12, #0x0]\n"
1751 "ldr x22, [x12, #0x8]\n"
1752 "fmla v19.8h, v20.8h, v2.h[7]\n"
1753 "ldr d2, [x11, #0x0]\n"
1754 "sub x14, x14, #0x8\n"
1755 "ldr d7, [x17, #0x10]\n"
1757 "ldr x21, [x11, #0x8]\n"
1758 "mov v6.d[1], x20\n"
1759 "ldr x20, [x17, #0x18]\n"
1760 "mov v0.d[1], x23\n"
1761 "prfm pldl1keep, [x13, #0x80]\n"
1762 "mov v1.d[1], x22\n"
1763 "prfm pldl1keep, [x12, #0x80]\n"
1764 "mov v2.d[1], x21\n"
1765 "prfm pldl1keep, [x11, #0x80]\n"
1766 "mov v7.d[1], x20\n"
1769 "fmla v8.8h, v6.8h, v0.h[0]\n"
1770 "add x13, x13, #0x10\n"
1771 "fmla v12.8h, v6.8h, v1.h[0]\n"
1772 "add x12, x12, #0x10\n"
1773 "fmla v16.8h, v6.8h, v2.h[0]\n"
1774 "ldr q21, [x17, #0x20]\n"
1775 "fmla v9.8h, v7.8h, v0.h[0]\n"
1776 "add x11, x11, #0x10\n"
1777 "fmla v13.8h, v7.8h, v1.h[0]\n"
1778 "sub x14, x14, #0x8\n"
1779 "fmla v17.8h, v7.8h, v2.h[0]\n"
1780 "ldr q20, [x17, #0x30]\n"
1781 "fmla v10.8h, v21.8h, v0.h[0]\n"
1782 "prfm pldl1keep, [x13, #0x80]\n"
1783 "fmla v14.8h, v21.8h, v1.h[0]\n"
1784 "prfm pldl1keep, [x12, #0x80]\n"
1785 "fmla v18.8h, v21.8h, v2.h[0]\n"
1786 "ldr q21, [x17, #0x40]\n"
1787 "fmla v11.8h, v20.8h, v0.h[0]\n"
1788 "prfm pldl1keep, [x11, #0x80]\n"
1789 "fmla v15.8h, v20.8h, v1.h[0]\n"
1790 "fmla v19.8h, v20.8h, v2.h[0]\n"
1791 "ldr q20, [x17, #0x50]\n"
1792 "fmla v8.8h, v21.8h, v0.h[1]\n"
1793 "fmla v12.8h, v21.8h, v1.h[1]\n"
1794 "fmla v16.8h, v21.8h, v2.h[1]\n"
1795 "ldr q21, [x17, #0x60]\n"
1796 "fmla v9.8h, v20.8h, v0.h[1]\n"
1797 "fmla v13.8h, v20.8h, v1.h[1]\n"
1798 "fmla v17.8h, v20.8h, v2.h[1]\n"
1799 "ldr q20, [x17, #0x70]\n"
1800 "fmla v10.8h, v21.8h, v0.h[1]\n"
1801 "fmla v14.8h, v21.8h, v1.h[1]\n"
1802 "fmla v18.8h, v21.8h, v2.h[1]\n"
1803 "ldr q21, [x17, #0x80]\n"
1804 "fmla v11.8h, v20.8h, v0.h[1]\n"
1805 "fmla v15.8h, v20.8h, v1.h[1]\n"
1806 "fmla v19.8h, v20.8h, v2.h[1]\n"
1807 "ldr q20, [x17, #0x90]\n"
1808 "fmla v8.8h, v21.8h, v0.h[2]\n"
1809 "fmla v12.8h, v21.8h, v1.h[2]\n"
1810 "fmla v16.8h, v21.8h, v2.h[2]\n"
1811 "ldr q21, [x17, #0xa0]\n"
1812 "fmla v9.8h, v20.8h, v0.h[2]\n"
1813 "fmla v13.8h, v20.8h, v1.h[2]\n"
1814 "fmla v17.8h, v20.8h, v2.h[2]\n"
1815 "ldr q20, [x17, #0xb0]\n"
1816 "fmla v10.8h, v21.8h, v0.h[2]\n"
1817 "fmla v14.8h, v21.8h, v1.h[2]\n"
1818 "fmla v18.8h, v21.8h, v2.h[2]\n"
1819 "ldr q21, [x17, #0xc0]\n"
1820 "fmla v11.8h, v20.8h, v0.h[2]\n"
1821 "fmla v15.8h, v20.8h, v1.h[2]\n"
1822 "fmla v19.8h, v20.8h, v2.h[2]\n"
1823 "ldr q20, [x17, #0xd0]\n"
1824 "fmla v8.8h, v21.8h, v0.h[3]\n"
1825 "fmla v12.8h, v21.8h, v1.h[3]\n"
1826 "fmla v16.8h, v21.8h, v2.h[3]\n"
1827 "ldr q21, [x17, #0xe0]\n"
1828 "fmla v9.8h, v20.8h, v0.h[3]\n"
1829 "fmla v13.8h, v20.8h, v1.h[3]\n"
1830 "fmla v17.8h, v20.8h, v2.h[3]\n"
1831 "ldr q20, [x17, #0xf0]\n"
1832 "fmla v10.8h, v21.8h, v0.h[3]\n"
1833 "fmla v14.8h, v21.8h, v1.h[3]\n"
1834 "fmla v18.8h, v21.8h, v2.h[3]\n"
1835 "ldr q21, [x17, #0x100]\n"
1836 "fmla v11.8h, v20.8h, v0.h[3]\n"
1837 "fmla v15.8h, v20.8h, v1.h[3]\n"
1838 "fmla v19.8h, v20.8h, v2.h[3]\n"
1839 "ldr q20, [x17, #0x110]\n"
1840 "fmla v8.8h, v21.8h, v0.h[4]\n"
1841 "fmla v12.8h, v21.8h, v1.h[4]\n"
1842 "fmla v16.8h, v21.8h, v2.h[4]\n"
1843 "ldr q21, [x17, #0x120]\n"
1844 "fmla v9.8h, v20.8h, v0.h[4]\n"
1845 "fmla v13.8h, v20.8h, v1.h[4]\n"
1846 "fmla v17.8h, v20.8h, v2.h[4]\n"
1847 "ldr q20, [x17, #0x130]\n"
1848 "fmla v10.8h, v21.8h, v0.h[4]\n"
1849 "fmla v14.8h, v21.8h, v1.h[4]\n"
1850 "fmla v18.8h, v21.8h, v2.h[4]\n"
1851 "ldr q21, [x17, #0x140]\n"
1852 "fmla v11.8h, v20.8h, v0.h[4]\n"
1853 "fmla v15.8h, v20.8h, v1.h[4]\n"
1854 "fmla v19.8h, v20.8h, v2.h[4]\n"
1855 "ldr q20, [x17, #0x150]\n"
1856 "fmla v8.8h, v21.8h, v0.h[5]\n"
1857 "fmla v12.8h, v21.8h, v1.h[5]\n"
1858 "fmla v16.8h, v21.8h, v2.h[5]\n"
1859 "ldr q21, [x17, #0x160]\n"
1860 "fmla v9.8h, v20.8h, v0.h[5]\n"
1861 "fmla v13.8h, v20.8h, v1.h[5]\n"
1862 "fmla v17.8h, v20.8h, v2.h[5]\n"
1863 "ldr q20, [x17, #0x170]\n"
1864 "fmla v10.8h, v21.8h, v0.h[5]\n"
1865 "fmla v14.8h, v21.8h, v1.h[5]\n"
1866 "fmla v18.8h, v21.8h, v2.h[5]\n"
1867 "ldr q21, [x17, #0x180]\n"
1868 "fmla v11.8h, v20.8h, v0.h[5]\n"
1869 "fmla v15.8h, v20.8h, v1.h[5]\n"
1870 "fmla v19.8h, v20.8h, v2.h[5]\n"
1871 "ldr q20, [x17, #0x190]\n"
1872 "fmla v8.8h, v21.8h, v0.h[6]\n"
1873 "fmla v12.8h, v21.8h, v1.h[6]\n"
1874 "fmla v16.8h, v21.8h, v2.h[6]\n"
1875 "ldr q21, [x17, #0x1a0]\n"
1876 "fmla v9.8h, v20.8h, v0.h[6]\n"
1877 "fmla v13.8h, v20.8h, v1.h[6]\n"
1878 "fmla v17.8h, v20.8h, v2.h[6]\n"
1879 "ldr q20, [x17, #0x1b0]\n"
1880 "fmla v10.8h, v21.8h, v0.h[6]\n"
1881 "fmla v14.8h, v21.8h, v1.h[6]\n"
1882 "fmla v18.8h, v21.8h, v2.h[6]\n"
1883 "ldr q21, [x17, #0x1c0]\n"
1884 "fmla v11.8h, v20.8h, v0.h[6]\n"
1885 "fmla v15.8h, v20.8h, v1.h[6]\n"
1886 "fmla v19.8h, v20.8h, v2.h[6]\n"
1887 "ldr q20, [x17, #0x1d0]\n"
1888 "fmla v8.8h, v21.8h, v0.h[7]\n"
1889 "fmla v12.8h, v21.8h, v1.h[7]\n"
1890 "fmla v16.8h, v21.8h, v2.h[7]\n"
1891 "ldr q21, [x17, #0x1e0]\n"
1892 "fmla v9.8h, v20.8h, v0.h[7]\n"
1893 "fmla v13.8h, v20.8h, v1.h[7]\n"
1894 "fmla v17.8h, v20.8h, v2.h[7]\n"
1895 "ldr q20, [x17, #0x1f0]\n"
1896 "fmla v10.8h, v21.8h, v0.h[7]\n"
1897 "add x17, x17, #0x200\n"
1898 "fmla v14.8h, v21.8h, v1.h[7]\n"
1899 "fmla v18.8h, v21.8h, v2.h[7]\n"
1900 "fmla v11.8h, v20.8h, v0.h[7]\n"
1901 "fmla v15.8h, v20.8h, v1.h[7]\n"
1902 "fmla v19.8h, v20.8h, v2.h[7]\n"
1906 "ldr h2, [x13], #0x2\n"
1907 "sub x14, x14, #0x1\n"
1908 "ldr h1, [x12], #0x2\n"
1909 "ldr h0, [x11], #0x2\n"
1910 "ldr q21, [x17, #0x0]\n"
1911 "fmla v8.8h, v21.8h, v2.h[0]\n"
1912 "ldr q20, [x17, #0x10]\n"
1913 "fmla v12.8h, v21.8h, v1.h[0]\n"
1914 "fmla v16.8h, v21.8h, v0.h[0]\n"
1915 "ldr q21, [x17, #0x20]\n"
1916 "fmla v9.8h, v20.8h, v2.h[0]\n"
1917 "fmla v13.8h, v20.8h, v1.h[0]\n"
1918 "fmla v17.8h, v20.8h, v0.h[0]\n"
1919 "ldr q20, [x17, #0x30]\n"
1920 "fmla v10.8h, v21.8h, v2.h[0]\n"
1921 "add x17, x17, #0x40\n"
1922 "fmla v14.8h, v21.8h, v1.h[0]\n"
1923 "fmla v18.8h, v21.8h, v0.h[0]\n"
1924 "fmla v11.8h, v20.8h, v2.h[0]\n"
1925 "fmla v15.8h, v20.8h, v1.h[0]\n"
1926 "fmla v19.8h, v20.8h, v0.h[0]\n"
1929 "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
1930 "add x15, x15, #0x1\n"
1933 "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
1934 "add x25, x16, x20, LSL #1\n"
1935 "add x24, x25, x20, LSL #1\n"
1936 "prfm pstl1keep, [x16, #0x0]\n"
1937 "prfm pstl1keep, [x25, #0x0]\n"
1938 "prfm pstl1keep, [x24, #0x0]\n"
1939 "tbz %x[flags], #1, 129f\n"
1940 "add x20, %x[args_ptr], %[offset_max]\n"
1941 "ld1r { v20.8h }, [x20]\n"
1942 "fmin v8.8h, v8.8h, v20.8h\n"
1943 "fmin v9.8h, v9.8h, v20.8h\n"
1944 "fmin v10.8h, v10.8h, v20.8h\n"
1945 "fmin v11.8h, v11.8h, v20.8h\n"
1946 "fmin v12.8h, v12.8h, v20.8h\n"
1947 "fmin v13.8h, v13.8h, v20.8h\n"
1948 "fmin v14.8h, v14.8h, v20.8h\n"
1949 "fmin v15.8h, v15.8h, v20.8h\n"
1950 "fmin v16.8h, v16.8h, v20.8h\n"
1951 "fmin v17.8h, v17.8h, v20.8h\n"
1952 "fmin v18.8h, v18.8h, v20.8h\n"
1953 "fmin v19.8h, v19.8h, v20.8h\n"
1954 "add x20, %x[args_ptr], %[offset_min]\n"
1955 "ld1r { v20.8h }, [x20]\n"
1956 "fmax v8.8h, v8.8h, v20.8h\n"
1957 "fmax v9.8h, v9.8h, v20.8h\n"
1958 "fmax v10.8h, v10.8h, v20.8h\n"
1959 "fmax v11.8h, v11.8h, v20.8h\n"
1960 "fmax v12.8h, v12.8h, v20.8h\n"
1961 "fmax v13.8h, v13.8h, v20.8h\n"
1962 "fmax v14.8h, v14.8h, v20.8h\n"
1963 "fmax v15.8h, v15.8h, v20.8h\n"
1964 "fmax v16.8h, v16.8h, v20.8h\n"
1965 "fmax v17.8h, v17.8h, v20.8h\n"
1966 "fmax v18.8h, v18.8h, v20.8h\n"
1967 "fmax v19.8h, v19.8h, v20.8h\n"
1971 "tbz x8, #4, 137f\n"
1972 "st1 { v8.8h }, [x16], #0x10\n"
1973 "st1 { v9.8h }, [x16], #0x10\n"
1974 "st1 { v12.8h }, [x25], #0x10\n"
1975 "st1 { v13.8h }, [x25], #0x10\n"
1976 "st1 { v16.8h }, [x24], #0x10\n"
1977 "st1 { v17.8h }, [x24], #0x10\n"
1978 "tbz x8, #3, 133f\n"
1979 "st1 { v10.8h }, [x16], #0x10\n"
1980 "st1 { v14.8h }, [x25], #0x10\n"
1981 "st1 { v18.8h }, [x24], #0x10\n"
1982 "tbz x8, #2, 131f\n"
1983 "str d11, [x16], #0x8\n"
1984 "str d15, [x25], #0x8\n"
1985 "str d19, [x24], #0x8\n"
1986 "tbz x8, #1, 130f\n"
1987 "st1 { v11.s }[2], [x16], #0x4\n"
1988 "st1 { v15.s }[2], [x25], #0x4\n"
1989 "st1 { v19.s }[2], [x24], #0x4\n"
1990 "tbz x8, #0, 145f\n"
1991 "st1 { v11.h }[6], [x16]\n"
1992 "st1 { v15.h }[6], [x25]\n"
1993 "st1 { v19.h }[6], [x24]\n"
1996 "tbz x8, #0, 145f\n"
1997 "st1 { v11.h }[4], [x16]\n"
1998 "st1 { v15.h }[4], [x25]\n"
1999 "st1 { v19.h }[4], [x24]\n"
2002 "tbz x8, #1, 132f\n"
2003 "str s11, [x16], #0x4\n"
2004 "str s15, [x25], #0x4\n"
2005 "str s19, [x24], #0x4\n"
2006 "tbz x8, #0, 145f\n"
2007 "st1 { v11.h }[2], [x16]\n"
2008 "st1 { v15.h }[2], [x25]\n"
2009 "st1 { v19.h }[2], [x24]\n"
2012 "tbz x8, #0, 145f\n"
2013 "str h11, [x16, #0x0]\n"
2014 "str h15, [x25, #0x0]\n"
2015 "str h19, [x24, #0x0]\n"
2018 "tbz x8, #2, 135f\n"
2019 "str d10, [x16], #0x8\n"
2020 "str d14, [x25], #0x8\n"
2021 "str d18, [x24], #0x8\n"
2022 "tbz x8, #1, 134f\n"
2023 "st1 { v10.s }[2], [x16], #0x4\n"
2024 "st1 { v14.s }[2], [x25], #0x4\n"
2025 "st1 { v18.s }[2], [x24], #0x4\n"
2026 "tbz x8, #0, 145f\n"
2027 "st1 { v10.h }[6], [x16]\n"
2028 "st1 { v14.h }[6], [x25]\n"
2029 "st1 { v18.h }[6], [x24]\n"
2032 "tbz x8, #0, 145f\n"
2033 "st1 { v10.h }[4], [x16]\n"
2034 "st1 { v14.h }[4], [x25]\n"
2035 "st1 { v18.h }[4], [x24]\n"
2038 "tbz x8, #1, 136f\n"
2039 "str s10, [x16], #0x4\n"
2040 "str s14, [x25], #0x4\n"
2041 "str s18, [x24], #0x4\n"
2042 "tbz x8, #0, 145f\n"
2043 "st1 { v10.h }[2], [x16]\n"
2044 "st1 { v14.h }[2], [x25]\n"
2045 "st1 { v18.h }[2], [x24]\n"
2048 "tbz x8, #0, 145f\n"
2049 "str h10, [x16, #0x0]\n"
2050 "str h14, [x25, #0x0]\n"
2051 "str h18, [x24, #0x0]\n"
2054 "tbz x8, #3, 141f\n"
2055 "st1 { v8.8h }, [x16], #0x10\n"
2056 "st1 { v12.8h }, [x25], #0x10\n"
2057 "st1 { v16.8h }, [x24], #0x10\n"
2058 "tbz x8, #2, 139f\n"
2059 "str d9, [x16], #0x8\n"
2060 "str d13, [x25], #0x8\n"
2061 "str d17, [x24], #0x8\n"
2062 "tbz x8, #1, 138f\n"
2063 "st1 { v9.s }[2], [x16], #0x4\n"
2064 "st1 { v13.s }[2], [x25], #0x4\n"
2065 "st1 { v17.s }[2], [x24], #0x4\n"
2066 "tbz x8, #0, 145f\n"
2067 "st1 { v9.h }[6], [x16]\n"
2068 "st1 { v13.h }[6], [x25]\n"
2069 "st1 { v17.h }[6], [x24]\n"
2072 "tbz x8, #0, 145f\n"
2073 "st1 { v9.h }[4], [x16]\n"
2074 "st1 { v13.h }[4], [x25]\n"
2075 "st1 { v17.h }[4], [x24]\n"
2078 "tbz x8, #1, 140f\n"
2079 "str s9, [x16], #0x4\n"
2080 "str s13, [x25], #0x4\n"
2081 "str s17, [x24], #0x4\n"
2082 "tbz x8, #0, 145f\n"
2083 "st1 { v9.h }[2], [x16]\n"
2084 "st1 { v13.h }[2], [x25]\n"
2085 "st1 { v17.h }[2], [x24]\n"
2088 "tbz x8, #0, 145f\n"
2089 "str h9, [x16, #0x0]\n"
2090 "str h13, [x25, #0x0]\n"
2091 "str h17, [x24, #0x0]\n"
2094 "tbz x8, #2, 143f\n"
2095 "str d8, [x16], #0x8\n"
2096 "str d12, [x25], #0x8\n"
2097 "str d16, [x24], #0x8\n"
2098 "tbz x8, #1, 142f\n"
2099 "st1 { v8.s }[2], [x16], #0x4\n"
2100 "st1 { v12.s }[2], [x25], #0x4\n"
2101 "st1 { v16.s }[2], [x24], #0x4\n"
2102 "tbz x8, #0, 145f\n"
2103 "st1 { v8.h }[6], [x16]\n"
2104 "st1 { v12.h }[6], [x25]\n"
2105 "st1 { v16.h }[6], [x24]\n"
2108 "tbz x8, #0, 145f\n"
2109 "st1 { v8.h }[4], [x16]\n"
2110 "st1 { v12.h }[4], [x25]\n"
2111 "st1 { v16.h }[4], [x24]\n"
2114 "tbz x8, #1, 144f\n"
2115 "str s8, [x16], #0x4\n"
2116 "str s12, [x25], #0x4\n"
2117 "str s16, [x24], #0x4\n"
2118 "tbz x8, #0, 145f\n"
2119 "st1 { v8.h }[2], [x16]\n"
2120 "st1 { v12.h }[2], [x25]\n"
2121 "st1 { v16.h }[2], [x24]\n"
2124 "str h8, [x16, #0x0]\n"
2125 "str h12, [x25, #0x0]\n"
2126 "str h16, [x24, #0x0]\n"
2130 "str q8, [x16, #0x0]\n"
2131 "str q9, [x16, #0x10]\n"
2132 "str q10, [x16, #0x20]\n"
2133 "str q11, [x16, #0x30]\n"
2134 "add x16, x16, #0x40\n"
2135 "str q12, [x25, #0x0]\n"
2136 "str q13, [x25, #0x10]\n"
2137 "str q14, [x25, #0x20]\n"
2138 "str q15, [x25, #0x30]\n"
2139 "str q16, [x24, #0x0]\n"
2140 "str q17, [x24, #0x10]\n"
2141 "str q18, [x24, #0x20]\n"
2142 "str q19, [x24, #0x30]\n"
2144 "subs x8, x8, #0x20\n"
2148 "mov x7, %x[bias]\n"
2149 "ldr x8, [%x[args_ptr], %[offsetof_N]]\n"
2150 "ldr x17, [%x[args_ptr], %[offsetof_B_ptr]]\n"
2151 "mov x16, %x[output_ptr]\n"
2154 "ldr q8, [x7, #0x0]\n"
2155 "mov v12.16b, v8.16b\n"
2156 "ldr q9, [x7, #0x10]\n"
2157 "mov v13.16b, v9.16b\n"
2158 "ldr q10, [x7, #0x20]\n"
2159 "mov v14.16b, v10.16b\n"
2160 "ldr q11, [x7, #0x30]\n"
2161 "mov v15.16b, v11.16b\n"
2162 "mov v16.16b, v8.16b\n"
2163 "add x7, x7, #0x40\n"
2164 "mov v17.16b, v9.16b\n"
2165 "mov v18.16b, v10.16b\n"
2166 "mov v19.16b, v11.16b\n"
2167 "mov v20.16b, v8.16b\n"
2168 "mov v21.16b, v9.16b\n"
2169 "mov v22.16b, v10.16b\n"
2170 "mov v23.16b, v11.16b\n"
2173 "tbz %x[flags], #0, 168f\n"
2174 "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
2175 "add x25, x16, x20, LSL #1\n"
2176 "add x24, x25, x20, LSL #1\n"
2178 "add x23, x24, x20, LSL #1\n"
2180 "tbz x8, #4, 158f\n"
2181 "ld1 { v8.8h }, [x16], #0x10\n"
2182 "ld1 { v12.8h }, [x25], #0x10\n"
2183 "ld1 { v16.8h }, [x24], #0x10\n"
2184 "ld1 { v20.8h }, [x23], #0x10\n"
2185 "ld1 { v9.8h }, [x16], #0x10\n"
2186 "ld1 { v13.8h }, [x25], #0x10\n"
2187 "ld1 { v17.8h }, [x24], #0x10\n"
2188 "ld1 { v21.8h }, [x23], #0x10\n"
2189 "tbz x8, #3, 154f\n"
2190 "ld1 { v10.8h }, [x16], #0x10\n"
2191 "ld1 { v14.8h }, [x25], #0x10\n"
2192 "ld1 { v18.8h }, [x24], #0x10\n"
2193 "ld1 { v22.8h }, [x23], #0x10\n"
2194 "tbz x8, #2, 152f\n"
2195 "ldr d11, [x16], #0x8\n"
2196 "ldr d15, [x25], #0x8\n"
2197 "ldr d19, [x24], #0x8\n"
2198 "ldr d23, [x23], #0x8\n"
2199 "tbz x8, #1, 151f\n"
2200 "ld1 { v11.s }[2], [x16], #0x4\n"
2202 "ld1 { v15.s }[2], [x25], #0x4\n"
2203 "ld1 { v19.s }[2], [x24], #0x4\n"
2204 "ld1 { v23.s }[2], [x23], #0x4\n"
2205 "tbz x8, #0, 166f\n"
2206 "ld1 { v11.h }[6], [x16]\n"
2207 "ld1 { v15.h }[6], [x25]\n"
2208 "ld1 { v19.h }[6], [x24]\n"
2209 "ld1 { v23.h }[6], [x23]\n"
2213 "tbz x8, #0, 166f\n"
2214 "ld1 { v11.h }[4], [x16]\n"
2215 "ld1 { v15.h }[4], [x25]\n"
2216 "ld1 { v19.h }[4], [x24]\n"
2217 "ld1 { v23.h }[4], [x23]\n"
2220 "tbz x8, #1, 153f\n"
2221 "ldr s11, [x16], #0x4\n"
2223 "ldr s15, [x25], #0x4\n"
2224 "ldr s19, [x24], #0x4\n"
2225 "ldr s23, [x23], #0x4\n"
2226 "tbz x8, #0, 166f\n"
2227 "ld1 { v11.h }[2], [x16]\n"
2228 "ld1 { v15.h }[2], [x25]\n"
2229 "ld1 { v19.h }[2], [x24]\n"
2230 "ld1 { v23.h }[2], [x23]\n"
2234 "tbz x8, #0, 166f\n"
2235 "ldr h11, [x16, #0x0]\n"
2236 "ldr h15, [x25, #0x0]\n"
2237 "ldr h19, [x24, #0x0]\n"
2238 "ldr h23, [x23, #0x0]\n"
2241 "tbz x8, #2, 156f\n"
2242 "ldr d10, [x16], #0x8\n"
2243 "ldr d14, [x25], #0x8\n"
2244 "ldr d18, [x24], #0x8\n"
2245 "ldr d22, [x23], #0x8\n"
2246 "tbz x8, #1, 155f\n"
2247 "ld1 { v10.s }[2], [x16], #0x4\n"
2249 "ld1 { v14.s }[2], [x25], #0x4\n"
2250 "ld1 { v18.s }[2], [x24], #0x4\n"
2251 "ld1 { v22.s }[2], [x23], #0x4\n"
2252 "tbz x8, #0, 166f\n"
2253 "ld1 { v10.h }[6], [x16]\n"
2254 "ld1 { v14.h }[6], [x25]\n"
2255 "ld1 { v18.h }[6], [x24]\n"
2256 "ld1 { v22.h }[6], [x23]\n"
2260 "tbz x8, #0, 166f\n"
2261 "ld1 { v10.h }[4], [x16]\n"
2262 "ld1 { v14.h }[4], [x25]\n"
2263 "ld1 { v18.h }[4], [x24]\n"
2264 "ld1 { v22.h }[4], [x23]\n"
2267 "tbz x8, #1, 157f\n"
2268 "ldr s10, [x16], #0x4\n"
2270 "ldr s14, [x25], #0x4\n"
2271 "ldr s18, [x24], #0x4\n"
2272 "ldr s22, [x23], #0x4\n"
2273 "tbz x8, #0, 166f\n"
2274 "ld1 { v10.h }[2], [x16]\n"
2275 "ld1 { v14.h }[2], [x25]\n"
2276 "ld1 { v18.h }[2], [x24]\n"
2277 "ld1 { v22.h }[2], [x23]\n"
2281 "tbz x8, #0, 166f\n"
2282 "ldr h10, [x16, #0x0]\n"
2283 "ldr h14, [x25, #0x0]\n"
2284 "ldr h18, [x24, #0x0]\n"
2285 "ldr h22, [x23, #0x0]\n"
2288 "tbz x8, #3, 162f\n"
2289 "ld1 { v8.8h }, [x16], #0x10\n"
2290 "ld1 { v12.8h }, [x25], #0x10\n"
2291 "ld1 { v16.8h }, [x24], #0x10\n"
2292 "ld1 { v20.8h }, [x23], #0x10\n"
2293 "tbz x8, #2, 160f\n"
2294 "ldr d9, [x16], #0x8\n"
2295 "ldr d13, [x25], #0x8\n"
2296 "ldr d17, [x24], #0x8\n"
2297 "ldr d21, [x23], #0x8\n"
2298 "tbz x8, #1, 159f\n"
2299 "ld1 { v9.s }[2], [x16], #0x4\n"
2301 "ld1 { v13.s }[2], [x25], #0x4\n"
2302 "ld1 { v17.s }[2], [x24], #0x4\n"
2303 "ld1 { v21.s }[2], [x23], #0x4\n"
2304 "tbz x8, #0, 166f\n"
2305 "ld1 { v9.h }[6], [x16]\n"
2306 "ld1 { v13.h }[6], [x25]\n"
2307 "ld1 { v17.h }[6], [x24]\n"
2308 "ld1 { v21.h }[6], [x23]\n"
2312 "tbz x8, #0, 166f\n"
2313 "ld1 { v9.h }[4], [x16]\n"
2314 "ld1 { v13.h }[4], [x25]\n"
2315 "ld1 { v17.h }[4], [x24]\n"
2316 "ld1 { v21.h }[4], [x23]\n"
2319 "tbz x8, #1, 161f\n"
2320 "ldr s9, [x16], #0x4\n"
2322 "ldr s13, [x25], #0x4\n"
2323 "ldr s17, [x24], #0x4\n"
2324 "ldr s21, [x23], #0x4\n"
2325 "tbz x8, #0, 166f\n"
2326 "ld1 { v9.h }[2], [x16]\n"
2327 "ld1 { v13.h }[2], [x25]\n"
2328 "ld1 { v17.h }[2], [x24]\n"
2329 "ld1 { v21.h }[2], [x23]\n"
2333 "tbz x8, #0, 166f\n"
2334 "ldr h9, [x16, #0x0]\n"
2335 "ldr h13, [x25, #0x0]\n"
2336 "ldr h17, [x24, #0x0]\n"
2337 "ldr h21, [x23, #0x0]\n"
2340 "tbz x8, #2, 164f\n"
2341 "ldr d8, [x16], #0x8\n"
2342 "ldr d12, [x25], #0x8\n"
2343 "ldr d16, [x24], #0x8\n"
2344 "ldr d20, [x23], #0x8\n"
2345 "tbz x8, #1, 163f\n"
2346 "ld1 { v8.s }[2], [x16], #0x4\n"
2348 "ld1 { v12.s }[2], [x25], #0x4\n"
2349 "ld1 { v16.s }[2], [x24], #0x4\n"
2350 "ld1 { v20.s }[2], [x23], #0x4\n"
2351 "tbz x8, #0, 166f\n"
2352 "ld1 { v8.h }[6], [x16]\n"
2353 "ld1 { v12.h }[6], [x25]\n"
2354 "ld1 { v16.h }[6], [x24]\n"
2355 "ld1 { v20.h }[6], [x23]\n"
2359 "tbz x8, #0, 166f\n"
2360 "ld1 { v8.h }[4], [x16]\n"
2361 "ld1 { v12.h }[4], [x25]\n"
2362 "ld1 { v16.h }[4], [x24]\n"
2363 "ld1 { v20.h }[4], [x23]\n"
2366 "tbz x8, #1, 165f\n"
2367 "ldr s8, [x16], #0x4\n"
2369 "ldr s12, [x25], #0x4\n"
2370 "ldr s16, [x24], #0x4\n"
2371 "ldr s20, [x23], #0x4\n"
2372 "tbz x8, #0, 166f\n"
2373 "ld1 { v8.h }[2], [x16]\n"
2374 "ld1 { v12.h }[2], [x25]\n"
2375 "ld1 { v16.h }[2], [x24]\n"
2376 "ld1 { v20.h }[2], [x23]\n"
2379 "ldr h8, [x16, #0x0]\n"
2381 "ldr h12, [x25, #0x0]\n"
2382 "ldr h16, [x24, #0x0]\n"
2383 "ldr h20, [x23, #0x0]\n"
2385 "sub x16, x16, x20\n"
2388 "ldr q8, [x16, #0x0]\n"
2389 "ldr q9, [x16, #0x10]\n"
2390 "ldr q10, [x16, #0x20]\n"
2391 "ldr q11, [x16, #0x30]\n"
2392 "ldr q12, [x25, #0x0]\n"
2393 "ldr q13, [x25, #0x10]\n"
2394 "ldr q14, [x25, #0x20]\n"
2395 "ldr q15, [x25, #0x30]\n"
2396 "ldr q16, [x24, #0x0]\n"
2397 "ldr q17, [x24, #0x10]\n"
2398 "ldr q18, [x24, #0x20]\n"
2399 "ldr q19, [x24, #0x30]\n"
2400 "ldr q20, [x23, #0x0]\n"
2401 "ldr q21, [x23, #0x10]\n"
2402 "ldr q22, [x23, #0x20]\n"
2403 "ldr q23, [x23, #0x30]\n"
2406 "movi v8.16b, #0x0\n"
2407 "movi v9.16b, #0x0\n"
2408 "movi v10.16b, #0x0\n"
2409 "movi v11.16b, #0x0\n"
2410 "movi v12.16b, #0x0\n"
2411 "movi v13.16b, #0x0\n"
2412 "movi v14.16b, #0x0\n"
2413 "movi v15.16b, #0x0\n"
2414 "movi v16.16b, #0x0\n"
2415 "movi v17.16b, #0x0\n"
2416 "movi v18.16b, #0x0\n"
2417 "movi v19.16b, #0x0\n"
2418 "movi v20.16b, #0x0\n"
2419 "movi v21.16b, #0x0\n"
2420 "movi v22.16b, #0x0\n"
2421 "movi v23.16b, #0x0\n"
2425 "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
2426 "ldr w14, [x20, x15, LSL #0x2]\n"
2427 "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
2428 "tbz %x[flags], #3, 171f\n"
2429 "ldr x20, [%x[input_ptr], x15, LSL #0x3]\n"
2430 "add x20, x20, x21, LSL #3\n"
2431 "ldr x13, [x20, #0x0]\n"
2432 "ldr x12, [x20, #0x8]\n"
2433 "ldr x11, [x20, #0x10]\n"
2434 "ldr x10, [x20, #0x18]\n"
2436 "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
2437 "add x13, x13, x20, LSL #1\n"
2438 "add x12, x12, x20, LSL #1\n"
2439 "add x11, x11, x20, LSL #1\n"
2440 "add x10, x10, x20, LSL #1\n"
2443 "mov x13, %x[input_ptr]\n"
2444 "add x12, x13, x21, LSL #1\n"
2445 "add x11, x12, x21, LSL #1\n"
2446 "add x10, x11, x21, LSL #1\n"
2450 "ldr q0, [x13, #0x0]\n"
2452 "ldr q1, [x12, #0x0]\n"
2453 "ldr q2, [x11, #0x0]\n"
2454 "ldr q3, [x10, #0x0]\n"
2455 "ldr q6, [x17, #0x0]\n"
2456 "ldr q7, [x17, #0x10]\n"
2459 "fmla v8.8h, v6.8h, v0.h[0]\n"
2460 "ldr x21, [x17, #0x28]\n"
2461 "fmla v12.8h, v6.8h, v1.h[0]\n"
2462 "ldr x20, [x17, #0x38]\n"
2463 "fmla v16.8h, v6.8h, v2.h[0]\n"
2464 "add x13, x13, #0x10\n"
2465 "fmla v20.8h, v6.8h, v3.h[0]\n"
2466 "ldr d25, [x17, #0x20]\n"
2467 "fmla v9.8h, v7.8h, v0.h[0]\n"
2468 "mov v25.d[1], x21\n"
2469 "fmla v13.8h, v7.8h, v1.h[0]\n"
2470 "ldr x21, [x17, #0x48]\n"
2471 "fmla v17.8h, v7.8h, v2.h[0]\n"
2472 "add x12, x12, #0x10\n"
2473 "fmla v21.8h, v7.8h, v3.h[0]\n"
2474 "ldr d24, [x17, #0x30]\n"
2475 "mov v24.d[1], x20\n"
2476 "fmla v10.8h, v25.8h, v0.h[0]\n"
2477 "fmla v14.8h, v25.8h, v1.h[0]\n"
2478 "ldr x20, [x17, #0x58]\n"
2479 "fmla v18.8h, v25.8h, v2.h[0]\n"
2480 "add x11, x11, #0x10\n"
2481 "fmla v22.8h, v25.8h, v3.h[0]\n"
2482 "ldr d25, [x17, #0x40]\n"
2483 "fmla v11.8h, v24.8h, v0.h[0]\n"
2484 "mov v25.d[1], x21\n"
2485 "fmla v15.8h, v24.8h, v1.h[0]\n"
2486 "ldr x21, [x17, #0x68]\n"
2487 "fmla v19.8h, v24.8h, v2.h[0]\n"
2488 "add x10, x10, #0x10\n"
2489 "fmla v23.8h, v24.8h, v3.h[0]\n"
2490 "ldr d24, [x17, #0x50]\n"
2491 "mov v24.d[1], x20\n"
2492 "fmla v8.8h, v25.8h, v0.h[1]\n"
2493 "fmla v12.8h, v25.8h, v1.h[1]\n"
2494 "ldr x20, [x17, #0x78]\n"
2495 "fmla v16.8h, v25.8h, v2.h[1]\n"
2496 "ldr x25, [x13, #0x8]\n"
2497 "fmla v20.8h, v25.8h, v3.h[1]\n"
2498 "ldr d25, [x17, #0x60]\n"
2499 "fmla v9.8h, v24.8h, v0.h[1]\n"
2500 "mov v25.d[1], x21\n"
2501 "fmla v13.8h, v24.8h, v1.h[1]\n"
2502 "ldr x21, [x17, #0x88]\n"
2503 "fmla v17.8h, v24.8h, v2.h[1]\n"
2504 "ldr x24, [x12, #0x8]\n"
2505 "fmla v21.8h, v24.8h, v3.h[1]\n"
2506 "ldr d24, [x17, #0x70]\n"
2507 "mov v24.d[1], x20\n"
2508 "fmla v10.8h, v25.8h, v0.h[1]\n"
2509 "fmla v14.8h, v25.8h, v1.h[1]\n"
2510 "ldr x20, [x17, #0x98]\n"
2511 "fmla v18.8h, v25.8h, v2.h[1]\n"
2512 "ldr x23, [x11, #0x8]\n"
2513 "fmla v22.8h, v25.8h, v3.h[1]\n"
2514 "ldr d25, [x17, #0x80]\n"
2515 "fmla v11.8h, v24.8h, v0.h[1]\n"
2516 "mov v25.d[1], x21\n"
2517 "fmla v15.8h, v24.8h, v1.h[1]\n"
2518 "ldr x21, [x17, #0xa8]\n"
2519 "fmla v19.8h, v24.8h, v2.h[1]\n"
2520 "ldr x22, [x10, #0x8]\n"
2521 "fmla v23.8h, v24.8h, v3.h[1]\n"
2522 "ldr d24, [x17, #0x90]\n"
2523 "mov v24.d[1], x20\n"
2524 "fmla v8.8h, v25.8h, v0.h[2]\n"
2525 "fmla v12.8h, v25.8h, v1.h[2]\n"
2526 "ldr x20, [x17, #0xb8]\n"
2527 "fmla v16.8h, v25.8h, v2.h[2]\n"
2528 "sub x14, x14, #0x8\n"
2529 "fmla v20.8h, v25.8h, v3.h[2]\n"
2530 "ldr d25, [x17, #0xa0]\n"
2531 "fmla v9.8h, v24.8h, v0.h[2]\n"
2532 "mov v25.d[1], x21\n"
2533 "fmla v13.8h, v24.8h, v1.h[2]\n"
2534 "ldr x21, [x17, #0xc8]\n"
2535 "fmla v17.8h, v24.8h, v2.h[2]\n"
2537 "fmla v21.8h, v24.8h, v3.h[2]\n"
2538 "ldr d24, [x17, #0xb0]\n"
2539 "mov v24.d[1], x20\n"
2540 "fmla v10.8h, v25.8h, v0.h[2]\n"
2541 "fmla v14.8h, v25.8h, v1.h[2]\n"
2542 "ldr x20, [x17, #0xd8]\n"
2543 "fmla v18.8h, v25.8h, v2.h[2]\n"
2544 "prfm pldl1keep, [x13, #0x80]\n"
2545 "fmla v22.8h, v25.8h, v3.h[2]\n"
2546 "ldr d25, [x17, #0xc0]\n"
2547 "fmla v11.8h, v24.8h, v0.h[2]\n"
2548 "mov v25.d[1], x21\n"
2549 "fmla v15.8h, v24.8h, v1.h[2]\n"
2550 "ldr x21, [x17, #0xe8]\n"
2551 "fmla v19.8h, v24.8h, v2.h[2]\n"
2552 "prfm pldl1keep, [x12, #0x80]\n"
2553 "fmla v23.8h, v24.8h, v3.h[2]\n"
2554 "ldr d24, [x17, #0xd0]\n"
2555 "mov v24.d[1], x20\n"
2556 "fmla v8.8h, v25.8h, v0.h[3]\n"
2557 "fmla v12.8h, v25.8h, v1.h[3]\n"
2558 "ldr x20, [x17, #0xf8]\n"
2559 "fmla v16.8h, v25.8h, v2.h[3]\n"
2560 "prfm pldl1keep, [x11, #0x80]\n"
2561 "fmla v20.8h, v25.8h, v3.h[3]\n"
2562 "ldr d25, [x17, #0xe0]\n"
2563 "fmla v9.8h, v24.8h, v0.h[3]\n"
2564 "mov v25.d[1], x21\n"
2565 "fmla v13.8h, v24.8h, v1.h[3]\n"
2566 "ldr x21, [x17, #0x108]\n"
2567 "fmla v17.8h, v24.8h, v2.h[3]\n"
2568 "prfm pldl1keep, [x10, #0x80]\n"
2569 "fmla v21.8h, v24.8h, v3.h[3]\n"
2570 "ldr d24, [x17, #0xf0]\n"
2571 "mov v24.d[1], x20\n"
2572 "fmla v10.8h, v25.8h, v0.h[3]\n"
2573 "fmla v14.8h, v25.8h, v1.h[3]\n"
2574 "ldr x20, [x17, #0x118]\n"
2575 "fmla v18.8h, v25.8h, v2.h[3]\n"
2576 "fmla v22.8h, v25.8h, v3.h[3]\n"
2577 "ldr d25, [x17, #0x100]\n"
2578 "fmla v11.8h, v24.8h, v0.h[3]\n"
2579 "mov v25.d[1], x21\n"
2580 "fmla v15.8h, v24.8h, v1.h[3]\n"
2581 "ldr x21, [x17, #0x128]\n"
2582 "fmla v19.8h, v24.8h, v2.h[3]\n"
2583 "fmla v23.8h, v24.8h, v3.h[3]\n"
2584 "ldr d24, [x17, #0x110]\n"
2585 "mov v24.d[1], x20\n"
2586 "fmla v8.8h, v25.8h, v0.h[4]\n"
2587 "fmla v12.8h, v25.8h, v1.h[4]\n"
2588 "ldr x20, [x17, #0x138]\n"
2589 "fmla v16.8h, v25.8h, v2.h[4]\n"
2590 "fmla v20.8h, v25.8h, v3.h[4]\n"
2591 "ldr d25, [x17, #0x120]\n"
2592 "fmla v9.8h, v24.8h, v0.h[4]\n"
2593 "mov v25.d[1], x21\n"
2594 "fmla v13.8h, v24.8h, v1.h[4]\n"
2595 "ldr x21, [x17, #0x148]\n"
2596 "fmla v17.8h, v24.8h, v2.h[4]\n"
2597 "fmla v21.8h, v24.8h, v3.h[4]\n"
2598 "ldr d24, [x17, #0x130]\n"
2599 "mov v24.d[1], x20\n"
2600 "fmla v10.8h, v25.8h, v0.h[4]\n"
2601 "fmla v14.8h, v25.8h, v1.h[4]\n"
2602 "ldr x20, [x17, #0x158]\n"
2603 "fmla v18.8h, v25.8h, v2.h[4]\n"
2604 "fmla v22.8h, v25.8h, v3.h[4]\n"
2605 "ldr d25, [x17, #0x140]\n"
2606 "fmla v11.8h, v24.8h, v0.h[4]\n"
2607 "mov v25.d[1], x21\n"
2608 "fmla v15.8h, v24.8h, v1.h[4]\n"
2609 "ldr x21, [x17, #0x168]\n"
2610 "fmla v19.8h, v24.8h, v2.h[4]\n"
2611 "fmla v23.8h, v24.8h, v3.h[4]\n"
2612 "ldr d24, [x17, #0x150]\n"
2613 "mov v24.d[1], x20\n"
2614 "fmla v8.8h, v25.8h, v0.h[5]\n"
2615 "fmla v12.8h, v25.8h, v1.h[5]\n"
2616 "ldr x20, [x17, #0x178]\n"
2617 "fmla v16.8h, v25.8h, v2.h[5]\n"
2618 "fmla v20.8h, v25.8h, v3.h[5]\n"
2619 "ldr d25, [x17, #0x160]\n"
2620 "fmla v9.8h, v24.8h, v0.h[5]\n"
2621 "mov v25.d[1], x21\n"
2622 "fmla v13.8h, v24.8h, v1.h[5]\n"
2623 "ldr x21, [x17, #0x188]\n"
2624 "fmla v17.8h, v24.8h, v2.h[5]\n"
2625 "fmla v21.8h, v24.8h, v3.h[5]\n"
2626 "ldr d24, [x17, #0x170]\n"
2627 "mov v24.d[1], x20\n"
2628 "fmla v10.8h, v25.8h, v0.h[5]\n"
2629 "fmla v14.8h, v25.8h, v1.h[5]\n"
2630 "ldr x20, [x17, #0x198]\n"
2631 "fmla v18.8h, v25.8h, v2.h[5]\n"
2632 "fmla v22.8h, v25.8h, v3.h[5]\n"
2633 "ldr d25, [x17, #0x180]\n"
2634 "fmla v11.8h, v24.8h, v0.h[5]\n"
2635 "mov v25.d[1], x21\n"
2636 "fmla v15.8h, v24.8h, v1.h[5]\n"
2637 "ldr x21, [x17, #0x1a8]\n"
2638 "fmla v19.8h, v24.8h, v2.h[5]\n"
2639 "fmla v23.8h, v24.8h, v3.h[5]\n"
2640 "ldr d24, [x17, #0x190]\n"
2641 "mov v24.d[1], x20\n"
2642 "fmla v8.8h, v25.8h, v0.h[6]\n"
2643 "fmla v12.8h, v25.8h, v1.h[6]\n"
2644 "ldr x20, [x17, #0x1b8]\n"
2645 "fmla v16.8h, v25.8h, v2.h[6]\n"
2646 "fmla v20.8h, v25.8h, v3.h[6]\n"
2647 "ldr d25, [x17, #0x1a0]\n"
2648 "fmla v9.8h, v24.8h, v0.h[6]\n"
2649 "mov v25.d[1], x21\n"
2650 "fmla v13.8h, v24.8h, v1.h[6]\n"
2651 "ldr x21, [x17, #0x1c8]\n"
2652 "fmla v17.8h, v24.8h, v2.h[6]\n"
2653 "fmla v21.8h, v24.8h, v3.h[6]\n"
2654 "ldr d24, [x17, #0x1b0]\n"
2655 "mov v24.d[1], x20\n"
2656 "fmla v10.8h, v25.8h, v0.h[6]\n"
2657 "fmla v14.8h, v25.8h, v1.h[6]\n"
2658 "ldr x20, [x17, #0x1d8]\n"
2659 "fmla v18.8h, v25.8h, v2.h[6]\n"
2660 "fmla v22.8h, v25.8h, v3.h[6]\n"
2661 "ldr d25, [x17, #0x1c0]\n"
2662 "fmla v11.8h, v24.8h, v0.h[6]\n"
2663 "mov v25.d[1], x21\n"
2664 "fmla v15.8h, v24.8h, v1.h[6]\n"
2665 "ldr x21, [x17, #0x1e8]\n"
2666 "fmla v19.8h, v24.8h, v2.h[6]\n"
2667 "fmla v23.8h, v24.8h, v3.h[6]\n"
2668 "ldr d24, [x17, #0x1d0]\n"
2669 "mov v24.d[1], x20\n"
2670 "fmla v8.8h, v25.8h, v0.h[7]\n"
2671 "fmla v12.8h, v25.8h, v1.h[7]\n"
2672 "ldr x20, [x17, #0x1f8]\n"
2673 "fmla v16.8h, v25.8h, v2.h[7]\n"
2674 "fmla v20.8h, v25.8h, v3.h[7]\n"
2675 "ldr d25, [x17, #0x1e0]\n"
2676 "fmla v9.8h, v24.8h, v0.h[7]\n"
2677 "mov v25.d[1], x21\n"
2678 "fmla v13.8h, v24.8h, v1.h[7]\n"
2679 "fmla v17.8h, v24.8h, v2.h[7]\n"
2680 "fmla v21.8h, v24.8h, v3.h[7]\n"
2681 "ldr d24, [x17, #0x1f0]\n"
2682 "mov v24.d[1], x20\n"
2683 "add x17, x17, #0x200\n"
2684 "fmla v10.8h, v25.8h, v0.h[7]\n"
2685 "ldr x21, [x17, #0x8]\n"
2686 "fmla v14.8h, v25.8h, v1.h[7]\n"
2687 "ldr x20, [x17, #0x18]\n"
2688 "fmla v18.8h, v25.8h, v2.h[7]\n"
2689 "fmla v22.8h, v25.8h, v3.h[7]\n"
2690 "ldr d6, [x17, #0x0]\n"
2691 "fmla v11.8h, v24.8h, v0.h[7]\n"
2692 "ldr d0, [x13, #0x0]\n"
2693 "fmla v15.8h, v24.8h, v1.h[7]\n"
2694 "ldr d1, [x12, #0x0]\n"
2695 "fmla v19.8h, v24.8h, v2.h[7]\n"
2696 "ldr d2, [x11, #0x0]\n"
2697 "fmla v23.8h, v24.8h, v3.h[7]\n"
2698 "ldr d3, [x10, #0x0]\n"
2699 "ldr d7, [x17, #0x10]\n"
2700 "mov v6.d[1], x21\n"
2701 "mov v0.d[1], x25\n"
2702 "mov v1.d[1], x24\n"
2703 "mov v2.d[1], x23\n"
2704 "mov v3.d[1], x22\n"
2705 "mov v7.d[1], x20\n"
2708 "fmla v8.8h, v6.8h, v0.h[0]\n"
2709 "add x13, x13, #0x10\n"
2710 "fmla v12.8h, v6.8h, v1.h[0]\n"
2711 "add x12, x12, #0x10\n"
2712 "fmla v16.8h, v6.8h, v2.h[0]\n"
2713 "add x11, x11, #0x10\n"
2714 "fmla v20.8h, v6.8h, v3.h[0]\n"
2715 "ldr q25, [x17, #0x20]\n"
2716 "fmla v9.8h, v7.8h, v0.h[0]\n"
2717 "add x10, x10, #0x10\n"
2718 "fmla v13.8h, v7.8h, v1.h[0]\n"
2719 "sub x14, x14, #0x8\n"
2720 "fmla v17.8h, v7.8h, v2.h[0]\n"
2721 "prfm pldl1keep, [x13, #0x80]\n"
2722 "fmla v21.8h, v7.8h, v3.h[0]\n"
2723 "ldr q24, [x17, #0x30]\n"
2724 "fmla v10.8h, v25.8h, v0.h[0]\n"
2725 "prfm pldl1keep, [x12, #0x80]\n"
2726 "fmla v14.8h, v25.8h, v1.h[0]\n"
2727 "prfm pldl1keep, [x11, #0x80]\n"
2728 "fmla v18.8h, v25.8h, v2.h[0]\n"
2729 "prfm pldl1keep, [x10, #0x80]\n"
2730 "fmla v22.8h, v25.8h, v3.h[0]\n"
2731 "ldr q25, [x17, #0x40]\n"
2732 "fmla v11.8h, v24.8h, v0.h[0]\n"
2733 "fmla v15.8h, v24.8h, v1.h[0]\n"
2734 "fmla v19.8h, v24.8h, v2.h[0]\n"
2735 "fmla v23.8h, v24.8h, v3.h[0]\n"
2736 "ldr q24, [x17, #0x50]\n"
2737 "fmla v8.8h, v25.8h, v0.h[1]\n"
2738 "fmla v12.8h, v25.8h, v1.h[1]\n"
2739 "fmla v16.8h, v25.8h, v2.h[1]\n"
2740 "fmla v20.8h, v25.8h, v3.h[1]\n"
2741 "ldr q25, [x17, #0x60]\n"
2742 "fmla v9.8h, v24.8h, v0.h[1]\n"
2743 "fmla v13.8h, v24.8h, v1.h[1]\n"
2744 "fmla v17.8h, v24.8h, v2.h[1]\n"
2745 "fmla v21.8h, v24.8h, v3.h[1]\n"
2746 "ldr q24, [x17, #0x70]\n"
2747 "fmla v10.8h, v25.8h, v0.h[1]\n"
2748 "fmla v14.8h, v25.8h, v1.h[1]\n"
2749 "fmla v18.8h, v25.8h, v2.h[1]\n"
2750 "fmla v22.8h, v25.8h, v3.h[1]\n"
2751 "ldr q25, [x17, #0x80]\n"
2752 "fmla v11.8h, v24.8h, v0.h[1]\n"
2753 "fmla v15.8h, v24.8h, v1.h[1]\n"
2754 "fmla v19.8h, v24.8h, v2.h[1]\n"
2755 "fmla v23.8h, v24.8h, v3.h[1]\n"
2756 "ldr q24, [x17, #0x90]\n"
2757 "fmla v8.8h, v25.8h, v0.h[2]\n"
2758 "fmla v12.8h, v25.8h, v1.h[2]\n"
2759 "fmla v16.8h, v25.8h, v2.h[2]\n"
2760 "fmla v20.8h, v25.8h, v3.h[2]\n"
2761 "ldr q25, [x17, #0xa0]\n"
2762 "fmla v9.8h, v24.8h, v0.h[2]\n"
2763 "fmla v13.8h, v24.8h, v1.h[2]\n"
2764 "fmla v17.8h, v24.8h, v2.h[2]\n"
2765 "fmla v21.8h, v24.8h, v3.h[2]\n"
2766 "ldr q24, [x17, #0xb0]\n"
2767 "fmla v10.8h, v25.8h, v0.h[2]\n"
2768 "fmla v14.8h, v25.8h, v1.h[2]\n"
2769 "fmla v18.8h, v25.8h, v2.h[2]\n"
2770 "fmla v22.8h, v25.8h, v3.h[2]\n"
2771 "ldr q25, [x17, #0xc0]\n"
2772 "fmla v11.8h, v24.8h, v0.h[2]\n"
2773 "fmla v15.8h, v24.8h, v1.h[2]\n"
2774 "fmla v19.8h, v24.8h, v2.h[2]\n"
2775 "fmla v23.8h, v24.8h, v3.h[2]\n"
2776 "ldr q24, [x17, #0xd0]\n"
2777 "fmla v8.8h, v25.8h, v0.h[3]\n"
2778 "fmla v12.8h, v25.8h, v1.h[3]\n"
2779 "fmla v16.8h, v25.8h, v2.h[3]\n"
2780 "fmla v20.8h, v25.8h, v3.h[3]\n"
2781 "ldr q25, [x17, #0xe0]\n"
2782 "fmla v9.8h, v24.8h, v0.h[3]\n"
2783 "fmla v13.8h, v24.8h, v1.h[3]\n"
2784 "fmla v17.8h, v24.8h, v2.h[3]\n"
2785 "fmla v21.8h, v24.8h, v3.h[3]\n"
2786 "ldr q24, [x17, #0xf0]\n"
2787 "fmla v10.8h, v25.8h, v0.h[3]\n"
2788 "fmla v14.8h, v25.8h, v1.h[3]\n"
2789 "fmla v18.8h, v25.8h, v2.h[3]\n"
2790 "fmla v22.8h, v25.8h, v3.h[3]\n"
2791 "ldr q25, [x17, #0x100]\n"
2792 "fmla v11.8h, v24.8h, v0.h[3]\n"
2793 "fmla v15.8h, v24.8h, v1.h[3]\n"
2794 "fmla v19.8h, v24.8h, v2.h[3]\n"
2795 "fmla v23.8h, v24.8h, v3.h[3]\n"
2796 "ldr q24, [x17, #0x110]\n"
2797 "fmla v8.8h, v25.8h, v0.h[4]\n"
2798 "fmla v12.8h, v25.8h, v1.h[4]\n"
2799 "fmla v16.8h, v25.8h, v2.h[4]\n"
2800 "fmla v20.8h, v25.8h, v3.h[4]\n"
2801 "ldr q25, [x17, #0x120]\n"
2802 "fmla v9.8h, v24.8h, v0.h[4]\n"
2803 "fmla v13.8h, v24.8h, v1.h[4]\n"
2804 "fmla v17.8h, v24.8h, v2.h[4]\n"
2805 "fmla v21.8h, v24.8h, v3.h[4]\n"
2806 "ldr q24, [x17, #0x130]\n"
2807 "fmla v10.8h, v25.8h, v0.h[4]\n"
2808 "fmla v14.8h, v25.8h, v1.h[4]\n"
2809 "fmla v18.8h, v25.8h, v2.h[4]\n"
2810 "fmla v22.8h, v25.8h, v3.h[4]\n"
2811 "ldr q25, [x17, #0x140]\n"
2812 "fmla v11.8h, v24.8h, v0.h[4]\n"
2813 "fmla v15.8h, v24.8h, v1.h[4]\n"
2814 "fmla v19.8h, v24.8h, v2.h[4]\n"
2815 "fmla v23.8h, v24.8h, v3.h[4]\n"
2816 "ldr q24, [x17, #0x150]\n"
2817 "fmla v8.8h, v25.8h, v0.h[5]\n"
2818 "fmla v12.8h, v25.8h, v1.h[5]\n"
2819 "fmla v16.8h, v25.8h, v2.h[5]\n"
2820 "fmla v20.8h, v25.8h, v3.h[5]\n"
2821 "ldr q25, [x17, #0x160]\n"
2822 "fmla v9.8h, v24.8h, v0.h[5]\n"
2823 "fmla v13.8h, v24.8h, v1.h[5]\n"
2824 "fmla v17.8h, v24.8h, v2.h[5]\n"
2825 "fmla v21.8h, v24.8h, v3.h[5]\n"
2826 "ldr q24, [x17, #0x170]\n"
2827 "fmla v10.8h, v25.8h, v0.h[5]\n"
2828 "fmla v14.8h, v25.8h, v1.h[5]\n"
2829 "fmla v18.8h, v25.8h, v2.h[5]\n"
2830 "fmla v22.8h, v25.8h, v3.h[5]\n"
2831 "ldr q25, [x17, #0x180]\n"
2832 "fmla v11.8h, v24.8h, v0.h[5]\n"
2833 "fmla v15.8h, v24.8h, v1.h[5]\n"
2834 "fmla v19.8h, v24.8h, v2.h[5]\n"
2835 "fmla v23.8h, v24.8h, v3.h[5]\n"
2836 "ldr q24, [x17, #0x190]\n"
2837 "fmla v8.8h, v25.8h, v0.h[6]\n"
2838 "fmla v12.8h, v25.8h, v1.h[6]\n"
2839 "fmla v16.8h, v25.8h, v2.h[6]\n"
2840 "fmla v20.8h, v25.8h, v3.h[6]\n"
2841 "ldr q25, [x17, #0x1a0]\n"
2842 "fmla v9.8h, v24.8h, v0.h[6]\n"
2843 "fmla v13.8h, v24.8h, v1.h[6]\n"
2844 "fmla v17.8h, v24.8h, v2.h[6]\n"
2845 "fmla v21.8h, v24.8h, v3.h[6]\n"
2846 "ldr q24, [x17, #0x1b0]\n"
2847 "fmla v10.8h, v25.8h, v0.h[6]\n"
2848 "fmla v14.8h, v25.8h, v1.h[6]\n"
2849 "fmla v18.8h, v25.8h, v2.h[6]\n"
2850 "fmla v22.8h, v25.8h, v3.h[6]\n"
2851 "ldr q25, [x17, #0x1c0]\n"
2852 "fmla v11.8h, v24.8h, v0.h[6]\n"
2853 "fmla v15.8h, v24.8h, v1.h[6]\n"
2854 "fmla v19.8h, v24.8h, v2.h[6]\n"
2855 "fmla v23.8h, v24.8h, v3.h[6]\n"
2856 "ldr q24, [x17, #0x1d0]\n"
2857 "fmla v8.8h, v25.8h, v0.h[7]\n"
2858 "fmla v12.8h, v25.8h, v1.h[7]\n"
2859 "fmla v16.8h, v25.8h, v2.h[7]\n"
2860 "fmla v20.8h, v25.8h, v3.h[7]\n"
2861 "ldr q25, [x17, #0x1e0]\n"
2862 "fmla v9.8h, v24.8h, v0.h[7]\n"
2863 "fmla v13.8h, v24.8h, v1.h[7]\n"
2864 "fmla v17.8h, v24.8h, v2.h[7]\n"
2865 "fmla v21.8h, v24.8h, v3.h[7]\n"
2866 "ldr q24, [x17, #0x1f0]\n"
2867 "fmla v10.8h, v25.8h, v0.h[7]\n"
2868 "add x17, x17, #0x200\n"
2869 "fmla v14.8h, v25.8h, v1.h[7]\n"
2870 "fmla v18.8h, v25.8h, v2.h[7]\n"
2871 "fmla v22.8h, v25.8h, v3.h[7]\n"
2872 "fmla v11.8h, v24.8h, v0.h[7]\n"
2873 "fmla v15.8h, v24.8h, v1.h[7]\n"
2874 "fmla v19.8h, v24.8h, v2.h[7]\n"
2875 "fmla v23.8h, v24.8h, v3.h[7]\n"
2879 "ldr h3, [x13], #0x2\n"
2880 "sub x14, x14, #0x1\n"
2881 "ldr h2, [x12], #0x2\n"
2882 "ldr h1, [x11], #0x2\n"
2883 "ldr h0, [x10], #0x2\n"
2884 "ldr q25, [x17, #0x0]\n"
2885 "fmla v8.8h, v25.8h, v3.h[0]\n"
2886 "ldr q24, [x17, #0x10]\n"
2887 "fmla v12.8h, v25.8h, v2.h[0]\n"
2888 "fmla v16.8h, v25.8h, v1.h[0]\n"
2889 "fmla v20.8h, v25.8h, v0.h[0]\n"
2890 "ldr q25, [x17, #0x20]\n"
2891 "fmla v9.8h, v24.8h, v3.h[0]\n"
2892 "fmla v13.8h, v24.8h, v2.h[0]\n"
2893 "fmla v17.8h, v24.8h, v1.h[0]\n"
2894 "fmla v21.8h, v24.8h, v0.h[0]\n"
2895 "ldr q24, [x17, #0x30]\n"
2896 "fmla v10.8h, v25.8h, v3.h[0]\n"
2897 "add x17, x17, #0x40\n"
2898 "fmla v14.8h, v25.8h, v2.h[0]\n"
2899 "fmla v18.8h, v25.8h, v1.h[0]\n"
2900 "fmla v22.8h, v25.8h, v0.h[0]\n"
2901 "fmla v11.8h, v24.8h, v3.h[0]\n"
2902 "fmla v15.8h, v24.8h, v2.h[0]\n"
2903 "fmla v19.8h, v24.8h, v1.h[0]\n"
2904 "fmla v23.8h, v24.8h, v0.h[0]\n"
2907 "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
2908 "add x15, x15, #0x1\n"
2911 "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
2912 "add x25, x16, x20, LSL #1\n"
2913 "add x24, x25, x20, LSL #1\n"
2914 "add x23, x24, x20, LSL #1\n"
2915 "prfm pstl1keep, [x16, #0x0]\n"
2916 "prfm pstl1keep, [x25, #0x0]\n"
2917 "prfm pstl1keep, [x24, #0x0]\n"
2918 "prfm pstl1keep, [x23, #0x0]\n"
2919 "tbz %x[flags], #1, 178f\n"
2920 "add x20, %x[args_ptr], %[offset_max]\n"
2921 "ld1r { v24.8h }, [x20]\n"
2922 "fmin v8.8h, v8.8h, v24.8h\n"
2923 "fmin v9.8h, v9.8h, v24.8h\n"
2924 "fmin v10.8h, v10.8h, v24.8h\n"
2925 "fmin v11.8h, v11.8h, v24.8h\n"
2926 "fmin v12.8h, v12.8h, v24.8h\n"
2927 "fmin v13.8h, v13.8h, v24.8h\n"
2928 "fmin v14.8h, v14.8h, v24.8h\n"
2929 "fmin v15.8h, v15.8h, v24.8h\n"
2930 "fmin v16.8h, v16.8h, v24.8h\n"
2931 "fmin v17.8h, v17.8h, v24.8h\n"
2932 "fmin v18.8h, v18.8h, v24.8h\n"
2933 "fmin v19.8h, v19.8h, v24.8h\n"
2934 "fmin v20.8h, v20.8h, v24.8h\n"
2935 "fmin v21.8h, v21.8h, v24.8h\n"
2936 "fmin v22.8h, v22.8h, v24.8h\n"
2937 "fmin v23.8h, v23.8h, v24.8h\n"
2938 "add x20, %x[args_ptr], %[offset_min]\n"
2939 "ld1r { v24.8h }, [x20]\n"
2940 "fmax v8.8h, v8.8h, v24.8h\n"
2941 "fmax v9.8h, v9.8h, v24.8h\n"
2942 "fmax v10.8h, v10.8h, v24.8h\n"
2943 "fmax v11.8h, v11.8h, v24.8h\n"
2944 "fmax v12.8h, v12.8h, v24.8h\n"
2945 "fmax v13.8h, v13.8h, v24.8h\n"
2946 "fmax v14.8h, v14.8h, v24.8h\n"
2947 "fmax v15.8h, v15.8h, v24.8h\n"
2948 "fmax v16.8h, v16.8h, v24.8h\n"
2949 "fmax v17.8h, v17.8h, v24.8h\n"
2950 "fmax v18.8h, v18.8h, v24.8h\n"
2951 "fmax v19.8h, v19.8h, v24.8h\n"
2952 "fmax v20.8h, v20.8h, v24.8h\n"
2953 "fmax v21.8h, v21.8h, v24.8h\n"
2954 "fmax v22.8h, v22.8h, v24.8h\n"
2955 "fmax v23.8h, v23.8h, v24.8h\n"
2959 "tbz x8, #4, 186f\n"
2960 "st1 { v8.8h }, [x16], #0x10\n"
2961 "st1 { v9.8h }, [x16], #0x10\n"
2962 "st1 { v12.8h }, [x25], #0x10\n"
2963 "st1 { v13.8h }, [x25], #0x10\n"
2964 "st1 { v16.8h }, [x24], #0x10\n"
2965 "st1 { v17.8h }, [x24], #0x10\n"
2966 "st1 { v20.8h }, [x23], #0x10\n"
2967 "st1 { v21.8h }, [x23], #0x10\n"
2968 "tbz x8, #3, 182f\n"
2969 "st1 { v10.8h }, [x16], #0x10\n"
2970 "st1 { v14.8h }, [x25], #0x10\n"
2971 "st1 { v18.8h }, [x24], #0x10\n"
2972 "st1 { v22.8h }, [x23], #0x10\n"
2973 "tbz x8, #2, 180f\n"
2974 "str d11, [x16], #0x8\n"
2975 "str d15, [x25], #0x8\n"
2976 "str d19, [x24], #0x8\n"
2977 "str d23, [x23], #0x8\n"
2978 "tbz x8, #1, 179f\n"
2979 "st1 { v11.s }[2], [x16], #0x4\n"
2980 "st1 { v15.s }[2], [x25], #0x4\n"
2981 "st1 { v19.s }[2], [x24], #0x4\n"
2982 "st1 { v23.s }[2], [x23], #0x4\n"
2983 "tbz x8, #0, 194f\n"
2984 "st1 { v11.h }[6], [x16]\n"
2985 "st1 { v15.h }[6], [x25]\n"
2986 "st1 { v19.h }[6], [x24]\n"
2987 "st1 { v23.h }[6], [x23]\n"
2990 "tbz x8, #0, 194f\n"
2991 "st1 { v11.h }[4], [x16]\n"
2992 "st1 { v15.h }[4], [x25]\n"
2993 "st1 { v19.h }[4], [x24]\n"
2994 "st1 { v23.h }[4], [x23]\n"
2997 "tbz x8, #1, 181f\n"
2998 "str s11, [x16], #0x4\n"
2999 "str s15, [x25], #0x4\n"
3000 "str s19, [x24], #0x4\n"
3001 "str s23, [x23], #0x4\n"
3002 "tbz x8, #0, 194f\n"
3003 "st1 { v11.h }[2], [x16]\n"
3004 "st1 { v15.h }[2], [x25]\n"
3005 "st1 { v19.h }[2], [x24]\n"
3006 "st1 { v23.h }[2], [x23]\n"
3009 "tbz x8, #0, 194f\n"
3010 "str h11, [x16, #0x0]\n"
3011 "str h15, [x25, #0x0]\n"
3012 "str h19, [x24, #0x0]\n"
3013 "str h23, [x23, #0x0]\n"
3016 "tbz x8, #2, 184f\n"
3017 "str d10, [x16], #0x8\n"
3018 "str d14, [x25], #0x8\n"
3019 "str d18, [x24], #0x8\n"
3020 "str d22, [x23], #0x8\n"
3021 "tbz x8, #1, 183f\n"
3022 "st1 { v10.s }[2], [x16], #0x4\n"
3023 "st1 { v14.s }[2], [x25], #0x4\n"
3024 "st1 { v18.s }[2], [x24], #0x4\n"
3025 "st1 { v22.s }[2], [x23], #0x4\n"
3026 "tbz x8, #0, 194f\n"
3027 "st1 { v10.h }[6], [x16]\n"
3028 "st1 { v14.h }[6], [x25]\n"
3029 "st1 { v18.h }[6], [x24]\n"
3030 "st1 { v22.h }[6], [x23]\n"
3033 "tbz x8, #0, 194f\n"
3034 "st1 { v10.h }[4], [x16]\n"
3035 "st1 { v14.h }[4], [x25]\n"
3036 "st1 { v18.h }[4], [x24]\n"
3037 "st1 { v22.h }[4], [x23]\n"
3040 "tbz x8, #1, 185f\n"
3041 "str s10, [x16], #0x4\n"
3042 "str s14, [x25], #0x4\n"
3043 "str s18, [x24], #0x4\n"
3044 "str s22, [x23], #0x4\n"
3045 "tbz x8, #0, 194f\n"
3046 "st1 { v10.h }[2], [x16]\n"
3047 "st1 { v14.h }[2], [x25]\n"
3048 "st1 { v18.h }[2], [x24]\n"
3049 "st1 { v22.h }[2], [x23]\n"
3052 "tbz x8, #0, 194f\n"
3053 "str h10, [x16, #0x0]\n"
3054 "str h14, [x25, #0x0]\n"
3055 "str h18, [x24, #0x0]\n"
3056 "str h22, [x23, #0x0]\n"
3059 "tbz x8, #3, 190f\n"
3060 "st1 { v8.8h }, [x16], #0x10\n"
3061 "st1 { v12.8h }, [x25], #0x10\n"
3062 "st1 { v16.8h }, [x24], #0x10\n"
3063 "st1 { v20.8h }, [x23], #0x10\n"
3064 "tbz x8, #2, 188f\n"
3065 "str d9, [x16], #0x8\n"
3066 "str d13, [x25], #0x8\n"
3067 "str d17, [x24], #0x8\n"
3068 "str d21, [x23], #0x8\n"
3069 "tbz x8, #1, 187f\n"
3070 "st1 { v9.s }[2], [x16], #0x4\n"
3071 "st1 { v13.s }[2], [x25], #0x4\n"
3072 "st1 { v17.s }[2], [x24], #0x4\n"
3073 "st1 { v21.s }[2], [x23], #0x4\n"
3074 "tbz x8, #0, 194f\n"
3075 "st1 { v9.h }[6], [x16]\n"
3076 "st1 { v13.h }[6], [x25]\n"
3077 "st1 { v17.h }[6], [x24]\n"
3078 "st1 { v21.h }[6], [x23]\n"
3081 "tbz x8, #0, 194f\n"
3082 "st1 { v9.h }[4], [x16]\n"
3083 "st1 { v13.h }[4], [x25]\n"
3084 "st1 { v17.h }[4], [x24]\n"
3085 "st1 { v21.h }[4], [x23]\n"
3088 "tbz x8, #1, 189f\n"
3089 "str s9, [x16], #0x4\n"
3090 "str s13, [x25], #0x4\n"
3091 "str s17, [x24], #0x4\n"
3092 "str s21, [x23], #0x4\n"
3093 "tbz x8, #0, 194f\n"
3094 "st1 { v9.h }[2], [x16]\n"
3095 "st1 { v13.h }[2], [x25]\n"
3096 "st1 { v17.h }[2], [x24]\n"
3097 "st1 { v21.h }[2], [x23]\n"
3100 "tbz x8, #0, 194f\n"
3101 "str h9, [x16, #0x0]\n"
3102 "str h13, [x25, #0x0]\n"
3103 "str h17, [x24, #0x0]\n"
3104 "str h21, [x23, #0x0]\n"
3107 "tbz x8, #2, 192f\n"
3108 "str d8, [x16], #0x8\n"
3109 "str d12, [x25], #0x8\n"
3110 "str d16, [x24], #0x8\n"
3111 "str d20, [x23], #0x8\n"
3112 "tbz x8, #1, 191f\n"
3113 "st1 { v8.s }[2], [x16], #0x4\n"
3114 "st1 { v12.s }[2], [x25], #0x4\n"
3115 "st1 { v16.s }[2], [x24], #0x4\n"
3116 "st1 { v20.s }[2], [x23], #0x4\n"
3117 "tbz x8, #0, 194f\n"
3118 "st1 { v8.h }[6], [x16]\n"
3119 "st1 { v12.h }[6], [x25]\n"
3120 "st1 { v16.h }[6], [x24]\n"
3121 "st1 { v20.h }[6], [x23]\n"
3124 "tbz x8, #0, 194f\n"
3125 "st1 { v8.h }[4], [x16]\n"
3126 "st1 { v12.h }[4], [x25]\n"
3127 "st1 { v16.h }[4], [x24]\n"
3128 "st1 { v20.h }[4], [x23]\n"
3131 "tbz x8, #1, 193f\n"
3132 "str s8, [x16], #0x4\n"
3133 "str s12, [x25], #0x4\n"
3134 "str s16, [x24], #0x4\n"
3135 "str s20, [x23], #0x4\n"
3136 "tbz x8, #0, 194f\n"
3137 "st1 { v8.h }[2], [x16]\n"
3138 "st1 { v12.h }[2], [x25]\n"
3139 "st1 { v16.h }[2], [x24]\n"
3140 "st1 { v20.h }[2], [x23]\n"
3143 "str h8, [x16, #0x0]\n"
3144 "str h12, [x25, #0x0]\n"
3145 "str h16, [x24, #0x0]\n"
3146 "str h20, [x23, #0x0]\n"
3150 "str q8, [x16, #0x0]\n"
3151 "str q9, [x16, #0x10]\n"
3152 "str q10, [x16, #0x20]\n"
3153 "str q11, [x16, #0x30]\n"
3154 "add x16, x16, #0x40\n"
3155 "str q12, [x25, #0x0]\n"
3156 "str q13, [x25, #0x10]\n"
3157 "str q14, [x25, #0x20]\n"
3158 "str q15, [x25, #0x30]\n"
3159 "str q16, [x24, #0x0]\n"
3160 "str q17, [x24, #0x10]\n"
3161 "str q18, [x24, #0x20]\n"
3162 "str q19, [x24, #0x30]\n"
3163 "str q20, [x23, #0x0]\n"
3164 "str q21, [x23, #0x10]\n"
3165 "str q22, [x23, #0x20]\n"
3166 "str q23, [x23, #0x30]\n"
3168 "subs x8, x8, #0x20\n"
3172 "mov x7, %x[bias]\n"
3173 "ldr x8, [%x[args_ptr], %[offsetof_N]]\n"
3174 "ldr x17, [%x[args_ptr], %[offsetof_B_ptr]]\n"
3175 "mov x16, %x[output_ptr]\n"
3178 "ldr q8, [x7, #0x0]\n"
3179 "mov v12.16b, v8.16b\n"
3180 "ldr q9, [x7, #0x10]\n"
3181 "mov v13.16b, v9.16b\n"
3182 "ldr q10, [x7, #0x20]\n"
3183 "mov v14.16b, v10.16b\n"
3184 "ldr q11, [x7, #0x30]\n"
3185 "mov v15.16b, v11.16b\n"
3186 "mov v16.16b, v8.16b\n"
3187 "add x7, x7, #0x40\n"
3188 "mov v17.16b, v9.16b\n"
3189 "mov v18.16b, v10.16b\n"
3190 "mov v19.16b, v11.16b\n"
3191 "mov v20.16b, v8.16b\n"
3192 "mov v21.16b, v9.16b\n"
3193 "mov v22.16b, v10.16b\n"
3194 "mov v23.16b, v11.16b\n"
3195 "mov v24.16b, v8.16b\n"
3196 "mov v25.16b, v9.16b\n"
3197 "mov v26.16b, v10.16b\n"
3198 "mov v27.16b, v11.16b\n"
3201 "tbz %x[flags], #0, 217f\n"
3202 "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
3203 "add x25, x16, x20, LSL #1\n"
3204 "add x24, x25, x20, LSL #1\n"
3205 "add x23, x24, x20, LSL #1\n"
3207 "add x22, x23, x20, LSL #1\n"
3209 "tbz x8, #4, 207f\n"
3210 "ld1 { v8.8h }, [x16], #0x10\n"
3211 "ld1 { v12.8h }, [x25], #0x10\n"
3212 "ld1 { v16.8h }, [x24], #0x10\n"
3213 "ld1 { v20.8h }, [x23], #0x10\n"
3214 "ld1 { v24.8h }, [x22], #0x10\n"
3215 "ld1 { v9.8h }, [x16], #0x10\n"
3216 "ld1 { v13.8h }, [x25], #0x10\n"
3217 "ld1 { v17.8h }, [x24], #0x10\n"
3218 "ld1 { v21.8h }, [x23], #0x10\n"
3219 "ld1 { v25.8h }, [x22], #0x10\n"
3220 "tbz x8, #3, 203f\n"
3221 "ld1 { v10.8h }, [x16], #0x10\n"
3222 "ld1 { v14.8h }, [x25], #0x10\n"
3223 "ld1 { v18.8h }, [x24], #0x10\n"
3224 "ld1 { v22.8h }, [x23], #0x10\n"
3225 "ld1 { v26.8h }, [x22], #0x10\n"
3226 "tbz x8, #2, 201f\n"
3227 "ldr d11, [x16], #0x8\n"
3228 "ldr d15, [x25], #0x8\n"
3229 "ldr d19, [x24], #0x8\n"
3230 "ldr d23, [x23], #0x8\n"
3231 "ldr d27, [x22], #0x8\n"
3232 "tbz x8, #1, 200f\n"
3233 "ld1 { v11.s }[2], [x16], #0x4\n"
3235 "ld1 { v15.s }[2], [x25], #0x4\n"
3236 "ld1 { v19.s }[2], [x24], #0x4\n"
3237 "ld1 { v23.s }[2], [x23], #0x4\n"
3238 "ld1 { v27.s }[2], [x22], #0x4\n"
3239 "tbz x8, #0, 215f\n"
3240 "ld1 { v11.h }[6], [x16]\n"
3241 "ld1 { v15.h }[6], [x25]\n"
3242 "ld1 { v19.h }[6], [x24]\n"
3243 "ld1 { v23.h }[6], [x23]\n"
3244 "ld1 { v27.h }[6], [x22]\n"
3248 "tbz x8, #0, 215f\n"
3249 "ld1 { v11.h }[4], [x16]\n"
3250 "ld1 { v15.h }[4], [x25]\n"
3251 "ld1 { v19.h }[4], [x24]\n"
3252 "ld1 { v23.h }[4], [x23]\n"
3253 "ld1 { v27.h }[4], [x22]\n"
3256 "tbz x8, #1, 202f\n"
3257 "ldr s11, [x16], #0x4\n"
3259 "ldr s15, [x25], #0x4\n"
3260 "ldr s19, [x24], #0x4\n"
3261 "ldr s23, [x23], #0x4\n"
3262 "ldr s27, [x22], #0x4\n"
3263 "tbz x8, #0, 215f\n"
3264 "ld1 { v11.h }[2], [x16]\n"
3265 "ld1 { v15.h }[2], [x25]\n"
3266 "ld1 { v19.h }[2], [x24]\n"
3267 "ld1 { v23.h }[2], [x23]\n"
3268 "ld1 { v27.h }[2], [x22]\n"
3272 "tbz x8, #0, 215f\n"
3273 "ldr h11, [x16, #0x0]\n"
3274 "ldr h15, [x25, #0x0]\n"
3275 "ldr h19, [x24, #0x0]\n"
3276 "ldr h23, [x23, #0x0]\n"
3277 "ldr h27, [x22, #0x0]\n"
3280 "tbz x8, #2, 205f\n"
3281 "ldr d10, [x16], #0x8\n"
3282 "ldr d14, [x25], #0x8\n"
3283 "ldr d18, [x24], #0x8\n"
3284 "ldr d22, [x23], #0x8\n"
3285 "ldr d26, [x22], #0x8\n"
3286 "tbz x8, #1, 204f\n"
3287 "ld1 { v10.s }[2], [x16], #0x4\n"
3289 "ld1 { v14.s }[2], [x25], #0x4\n"
3290 "ld1 { v18.s }[2], [x24], #0x4\n"
3291 "ld1 { v22.s }[2], [x23], #0x4\n"
3292 "ld1 { v26.s }[2], [x22], #0x4\n"
3293 "tbz x8, #0, 215f\n"
3294 "ld1 { v10.h }[6], [x16]\n"
3295 "ld1 { v14.h }[6], [x25]\n"
3296 "ld1 { v18.h }[6], [x24]\n"
3297 "ld1 { v22.h }[6], [x23]\n"
3298 "ld1 { v26.h }[6], [x22]\n"
3302 "tbz x8, #0, 215f\n"
3303 "ld1 { v10.h }[4], [x16]\n"
3304 "ld1 { v14.h }[4], [x25]\n"
3305 "ld1 { v18.h }[4], [x24]\n"
3306 "ld1 { v22.h }[4], [x23]\n"
3307 "ld1 { v26.h }[4], [x22]\n"
3310 "tbz x8, #1, 206f\n"
3311 "ldr s10, [x16], #0x4\n"
3313 "ldr s14, [x25], #0x4\n"
3314 "ldr s18, [x24], #0x4\n"
3315 "ldr s22, [x23], #0x4\n"
3316 "ldr s26, [x22], #0x4\n"
3317 "tbz x8, #0, 215f\n"
3318 "ld1 { v10.h }[2], [x16]\n"
3319 "ld1 { v14.h }[2], [x25]\n"
3320 "ld1 { v18.h }[2], [x24]\n"
3321 "ld1 { v22.h }[2], [x23]\n"
3322 "ld1 { v26.h }[2], [x22]\n"
3326 "tbz x8, #0, 215f\n"
3327 "ldr h10, [x16, #0x0]\n"
3328 "ldr h14, [x25, #0x0]\n"
3329 "ldr h18, [x24, #0x0]\n"
3330 "ldr h22, [x23, #0x0]\n"
3331 "ldr h26, [x22, #0x0]\n"
3334 "tbz x8, #3, 211f\n"
3335 "ld1 { v8.8h }, [x16], #0x10\n"
3336 "ld1 { v12.8h }, [x25], #0x10\n"
3337 "ld1 { v16.8h }, [x24], #0x10\n"
3338 "ld1 { v20.8h }, [x23], #0x10\n"
3339 "ld1 { v24.8h }, [x22], #0x10\n"
3340 "tbz x8, #2, 209f\n"
3341 "ldr d9, [x16], #0x8\n"
3342 "ldr d13, [x25], #0x8\n"
3343 "ldr d17, [x24], #0x8\n"
3344 "ldr d21, [x23], #0x8\n"
3345 "ldr d25, [x22], #0x8\n"
3346 "tbz x8, #1, 208f\n"
3347 "ld1 { v9.s }[2], [x16], #0x4\n"
3349 "ld1 { v13.s }[2], [x25], #0x4\n"
3350 "ld1 { v17.s }[2], [x24], #0x4\n"
3351 "ld1 { v21.s }[2], [x23], #0x4\n"
3352 "ld1 { v25.s }[2], [x22], #0x4\n"
3353 "tbz x8, #0, 215f\n"
3354 "ld1 { v9.h }[6], [x16]\n"
3355 "ld1 { v13.h }[6], [x25]\n"
3356 "ld1 { v17.h }[6], [x24]\n"
3357 "ld1 { v21.h }[6], [x23]\n"
3358 "ld1 { v25.h }[6], [x22]\n"
3362 "tbz x8, #0, 215f\n"
3363 "ld1 { v9.h }[4], [x16]\n"
3364 "ld1 { v13.h }[4], [x25]\n"
3365 "ld1 { v17.h }[4], [x24]\n"
3366 "ld1 { v21.h }[4], [x23]\n"
3367 "ld1 { v25.h }[4], [x22]\n"
3370 "tbz x8, #1, 210f\n"
3371 "ldr s9, [x16], #0x4\n"
3373 "ldr s13, [x25], #0x4\n"
3374 "ldr s17, [x24], #0x4\n"
3375 "ldr s21, [x23], #0x4\n"
3376 "ldr s25, [x22], #0x4\n"
3377 "tbz x8, #0, 215f\n"
3378 "ld1 { v9.h }[2], [x16]\n"
3379 "ld1 { v13.h }[2], [x25]\n"
3380 "ld1 { v17.h }[2], [x24]\n"
3381 "ld1 { v21.h }[2], [x23]\n"
3382 "ld1 { v25.h }[2], [x22]\n"
3386 "tbz x8, #0, 215f\n"
3387 "ldr h9, [x16, #0x0]\n"
3388 "ldr h13, [x25, #0x0]\n"
3389 "ldr h17, [x24, #0x0]\n"
3390 "ldr h21, [x23, #0x0]\n"
3391 "ldr h25, [x22, #0x0]\n"
3394 "tbz x8, #2, 213f\n"
3395 "ldr d8, [x16], #0x8\n"
3396 "ldr d12, [x25], #0x8\n"
3397 "ldr d16, [x24], #0x8\n"
3398 "ldr d20, [x23], #0x8\n"
3399 "ldr d24, [x22], #0x8\n"
3400 "tbz x8, #1, 212f\n"
3401 "ld1 { v8.s }[2], [x16], #0x4\n"
3403 "ld1 { v12.s }[2], [x25], #0x4\n"
3404 "ld1 { v16.s }[2], [x24], #0x4\n"
3405 "ld1 { v20.s }[2], [x23], #0x4\n"
3406 "ld1 { v24.s }[2], [x22], #0x4\n"
3407 "tbz x8, #0, 215f\n"
3408 "ld1 { v8.h }[6], [x16]\n"
3409 "ld1 { v12.h }[6], [x25]\n"
3410 "ld1 { v16.h }[6], [x24]\n"
3411 "ld1 { v20.h }[6], [x23]\n"
3412 "ld1 { v24.h }[6], [x22]\n"
3416 "tbz x8, #0, 215f\n"
3417 "ld1 { v8.h }[4], [x16]\n"
3418 "ld1 { v12.h }[4], [x25]\n"
3419 "ld1 { v16.h }[4], [x24]\n"
3420 "ld1 { v20.h }[4], [x23]\n"
3421 "ld1 { v24.h }[4], [x22]\n"
3424 "tbz x8, #1, 214f\n"
3425 "ldr s8, [x16], #0x4\n"
3427 "ldr s12, [x25], #0x4\n"
3428 "ldr s16, [x24], #0x4\n"
3429 "ldr s20, [x23], #0x4\n"
3430 "ldr s24, [x22], #0x4\n"
3431 "tbz x8, #0, 215f\n"
3432 "ld1 { v8.h }[2], [x16]\n"
3433 "ld1 { v12.h }[2], [x25]\n"
3434 "ld1 { v16.h }[2], [x24]\n"
3435 "ld1 { v20.h }[2], [x23]\n"
3436 "ld1 { v24.h }[2], [x22]\n"
3439 "ldr h8, [x16, #0x0]\n"
3441 "ldr h12, [x25, #0x0]\n"
3442 "ldr h16, [x24, #0x0]\n"
3443 "ldr h20, [x23, #0x0]\n"
3444 "ldr h24, [x22, #0x0]\n"
3446 "sub x16, x16, x20\n"
3449 "ldr q8, [x16, #0x0]\n"
3450 "ldr q9, [x16, #0x10]\n"
3451 "ldr q10, [x16, #0x20]\n"
3452 "ldr q11, [x16, #0x30]\n"
3453 "ldr q12, [x25, #0x0]\n"
3454 "ldr q13, [x25, #0x10]\n"
3455 "ldr q14, [x25, #0x20]\n"
3456 "ldr q15, [x25, #0x30]\n"
3457 "ldr q16, [x24, #0x0]\n"
3458 "ldr q17, [x24, #0x10]\n"
3459 "ldr q18, [x24, #0x20]\n"
3460 "ldr q19, [x24, #0x30]\n"
3461 "ldr q20, [x23, #0x0]\n"
3462 "ldr q21, [x23, #0x10]\n"
3463 "ldr q22, [x23, #0x20]\n"
3464 "ldr q23, [x23, #0x30]\n"
3465 "ldr q24, [x22, #0x0]\n"
3466 "ldr q25, [x22, #0x10]\n"
3467 "ldr q26, [x22, #0x20]\n"
3468 "ldr q27, [x22, #0x30]\n"
3471 "movi v8.16b, #0x0\n"
3472 "movi v9.16b, #0x0\n"
3473 "movi v10.16b, #0x0\n"
3474 "movi v11.16b, #0x0\n"
3475 "movi v12.16b, #0x0\n"
3476 "movi v13.16b, #0x0\n"
3477 "movi v14.16b, #0x0\n"
3478 "movi v15.16b, #0x0\n"
3479 "movi v16.16b, #0x0\n"
3480 "movi v17.16b, #0x0\n"
3481 "movi v18.16b, #0x0\n"
3482 "movi v19.16b, #0x0\n"
3483 "movi v20.16b, #0x0\n"
3484 "movi v21.16b, #0x0\n"
3485 "movi v22.16b, #0x0\n"
3486 "movi v23.16b, #0x0\n"
3487 "movi v24.16b, #0x0\n"
3488 "movi v25.16b, #0x0\n"
3489 "movi v26.16b, #0x0\n"
3490 "movi v27.16b, #0x0\n"
3494 "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
3495 "ldr w14, [x20, x15, LSL #0x2]\n"
3496 "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
3497 "tbz %x[flags], #3, 220f\n"
3498 "ldr x20, [%x[input_ptr], x15, LSL #0x3]\n"
3499 "add x20, x20, x21, LSL #3\n"
3500 "ldr x13, [x20, #0x0]\n"
3501 "ldr x12, [x20, #0x8]\n"
3502 "ldr x11, [x20, #0x10]\n"
3503 "ldr x10, [x20, #0x18]\n"
3504 "ldr x9, [x20, #0x20]\n"
3506 "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
3507 "add x13, x13, x20, LSL #1\n"
3508 "add x12, x12, x20, LSL #1\n"
3509 "add x11, x11, x20, LSL #1\n"
3510 "add x10, x10, x20, LSL #1\n"
3511 "add x9, x9, x20, LSL #1\n"
3514 "mov x13, %x[input_ptr]\n"
3515 "add x12, x13, x21, LSL #1\n"
3516 "add x11, x12, x21, LSL #1\n"
3517 "add x10, x11, x21, LSL #1\n"
3518 "add x9, x10, x21, LSL #1\n"
3522 "ldr q0, [x13, #0x0]\n"
3524 "ldr q1, [x12, #0x0]\n"
3525 "ldr q2, [x11, #0x0]\n"
3526 "ldr q3, [x10, #0x0]\n"
3527 "ldr q4, [x9, #0x0]\n"
3528 "ldr q6, [x17, #0x0]\n"
3529 "ldr q7, [x17, #0x10]\n"
3532 "fmla v8.8h, v6.8h, v0.h[0]\n"
3533 "ldr x21, [x17, #0x28]\n"
3534 "fmla v12.8h, v6.8h, v1.h[0]\n"
3535 "ldr x20, [x17, #0x38]\n"
3536 "fmla v16.8h, v6.8h, v2.h[0]\n"
3537 "add x13, x13, #0x10\n"
3538 "fmla v20.8h, v6.8h, v3.h[0]\n"
3539 "add x12, x12, #0x10\n"
3540 "fmla v24.8h, v6.8h, v4.h[0]\n"
3541 "ldr d29, [x17, #0x20]\n"
3542 "fmla v9.8h, v7.8h, v0.h[0]\n"
3543 "mov v29.d[1], x21\n"
3544 "fmla v13.8h, v7.8h, v1.h[0]\n"
3545 "ldr x21, [x17, #0x48]\n"
3546 "fmla v17.8h, v7.8h, v2.h[0]\n"
3547 "add x11, x11, #0x10\n"
3548 "fmla v21.8h, v7.8h, v3.h[0]\n"
3549 "add x10, x10, #0x10\n"
3550 "fmla v25.8h, v7.8h, v4.h[0]\n"
3551 "ldr d28, [x17, #0x30]\n"
3552 "mov v28.d[1], x20\n"
3553 "fmla v10.8h, v29.8h, v0.h[0]\n"
3554 "fmla v14.8h, v29.8h, v1.h[0]\n"
3555 "ldr x20, [x17, #0x58]\n"
3556 "fmla v18.8h, v29.8h, v2.h[0]\n"
3557 "add x9, x9, #0x10\n"
3558 "fmla v22.8h, v29.8h, v3.h[0]\n"
3559 "ldr x26, [x13, #0x8]\n"
3560 "fmla v26.8h, v29.8h, v4.h[0]\n"
3561 "ldr d29, [x17, #0x40]\n"
3562 "fmla v11.8h, v28.8h, v0.h[0]\n"
3563 "mov v29.d[1], x21\n"
3564 "fmla v15.8h, v28.8h, v1.h[0]\n"
3565 "ldr x21, [x17, #0x68]\n"
3566 "fmla v19.8h, v28.8h, v2.h[0]\n"
3567 "ldr x25, [x12, #0x8]\n"
3568 "fmla v23.8h, v28.8h, v3.h[0]\n"
3569 "ldr x24, [x11, #0x8]\n"
3570 "fmla v27.8h, v28.8h, v4.h[0]\n"
3571 "ldr d28, [x17, #0x50]\n"
3572 "mov v28.d[1], x20\n"
3573 "fmla v8.8h, v29.8h, v0.h[1]\n"
3574 "fmla v12.8h, v29.8h, v1.h[1]\n"
3575 "ldr x20, [x17, #0x78]\n"
3576 "fmla v16.8h, v29.8h, v2.h[1]\n"
3577 "ldr x23, [x10, #0x8]\n"
3578 "fmla v20.8h, v29.8h, v3.h[1]\n"
3579 "ldr x22, [x9, #0x8]\n"
3580 "fmla v24.8h, v29.8h, v4.h[1]\n"
3581 "ldr d29, [x17, #0x60]\n"
3582 "fmla v9.8h, v28.8h, v0.h[1]\n"
3583 "mov v29.d[1], x21\n"
3584 "fmla v13.8h, v28.8h, v1.h[1]\n"
3585 "ldr x21, [x17, #0x88]\n"
3586 "fmla v17.8h, v28.8h, v2.h[1]\n"
3587 "sub x14, x14, #0x8\n"
3588 "fmla v21.8h, v28.8h, v3.h[1]\n"
3590 "fmla v25.8h, v28.8h, v4.h[1]\n"
3591 "ldr d28, [x17, #0x70]\n"
3592 "mov v28.d[1], x20\n"
3593 "fmla v10.8h, v29.8h, v0.h[1]\n"
3594 "fmla v14.8h, v29.8h, v1.h[1]\n"
3595 "ldr x20, [x17, #0x98]\n"
3596 "fmla v18.8h, v29.8h, v2.h[1]\n"
3597 "prfm pldl1keep, [x13, #0x80]\n"
3598 "fmla v22.8h, v29.8h, v3.h[1]\n"
3599 "prfm pldl1keep, [x12, #0x80]\n"
3600 "fmla v26.8h, v29.8h, v4.h[1]\n"
3601 "ldr d29, [x17, #0x80]\n"
3602 "fmla v11.8h, v28.8h, v0.h[1]\n"
3603 "mov v29.d[1], x21\n"
3604 "fmla v15.8h, v28.8h, v1.h[1]\n"
3605 "ldr x21, [x17, #0xa8]\n"
3606 "fmla v19.8h, v28.8h, v2.h[1]\n"
3607 "prfm pldl1keep, [x11, #0x80]\n"
3608 "fmla v23.8h, v28.8h, v3.h[1]\n"
3609 "prfm pldl1keep, [x10, #0x80]\n"
3610 "fmla v27.8h, v28.8h, v4.h[1]\n"
3611 "ldr d28, [x17, #0x90]\n"
3612 "mov v28.d[1], x20\n"
3613 "fmla v8.8h, v29.8h, v0.h[2]\n"
3614 "fmla v12.8h, v29.8h, v1.h[2]\n"
3615 "ldr x20, [x17, #0xb8]\n"
3616 "fmla v16.8h, v29.8h, v2.h[2]\n"
3617 "prfm pldl1keep, [x9, #0x80]\n"
3618 "fmla v20.8h, v29.8h, v3.h[2]\n"
3619 "fmla v24.8h, v29.8h, v4.h[2]\n"
3620 "ldr d29, [x17, #0xa0]\n"
3621 "fmla v9.8h, v28.8h, v0.h[2]\n"
3622 "mov v29.d[1], x21\n"
3623 "fmla v13.8h, v28.8h, v1.h[2]\n"
3624 "ldr x21, [x17, #0xc8]\n"
3625 "fmla v17.8h, v28.8h, v2.h[2]\n"
3626 "fmla v21.8h, v28.8h, v3.h[2]\n"
3627 "fmla v25.8h, v28.8h, v4.h[2]\n"
3628 "ldr d28, [x17, #0xb0]\n"
3629 "mov v28.d[1], x20\n"
3630 "fmla v10.8h, v29.8h, v0.h[2]\n"
3631 "fmla v14.8h, v29.8h, v1.h[2]\n"
3632 "ldr x20, [x17, #0xd8]\n"
3633 "fmla v18.8h, v29.8h, v2.h[2]\n"
3634 "fmla v22.8h, v29.8h, v3.h[2]\n"
3635 "fmla v26.8h, v29.8h, v4.h[2]\n"
3636 "ldr d29, [x17, #0xc0]\n"
3637 "fmla v11.8h, v28.8h, v0.h[2]\n"
3638 "mov v29.d[1], x21\n"
3639 "fmla v15.8h, v28.8h, v1.h[2]\n"
3640 "ldr x21, [x17, #0xe8]\n"
3641 "fmla v19.8h, v28.8h, v2.h[2]\n"
3642 "fmla v23.8h, v28.8h, v3.h[2]\n"
3643 "fmla v27.8h, v28.8h, v4.h[2]\n"
3644 "ldr d28, [x17, #0xd0]\n"
3645 "mov v28.d[1], x20\n"
3646 "fmla v8.8h, v29.8h, v0.h[3]\n"
3647 "fmla v12.8h, v29.8h, v1.h[3]\n"
3648 "ldr x20, [x17, #0xf8]\n"
3649 "fmla v16.8h, v29.8h, v2.h[3]\n"
3650 "fmla v20.8h, v29.8h, v3.h[3]\n"
3651 "fmla v24.8h, v29.8h, v4.h[3]\n"
3652 "ldr d29, [x17, #0xe0]\n"
3653 "fmla v9.8h, v28.8h, v0.h[3]\n"
3654 "mov v29.d[1], x21\n"
3655 "fmla v13.8h, v28.8h, v1.h[3]\n"
3656 "ldr x21, [x17, #0x108]\n"
3657 "fmla v17.8h, v28.8h, v2.h[3]\n"
3658 "fmla v21.8h, v28.8h, v3.h[3]\n"
3659 "fmla v25.8h, v28.8h, v4.h[3]\n"
3660 "ldr d28, [x17, #0xf0]\n"
3661 "mov v28.d[1], x20\n"
3662 "fmla v10.8h, v29.8h, v0.h[3]\n"
3663 "fmla v14.8h, v29.8h, v1.h[3]\n"
3664 "ldr x20, [x17, #0x118]\n"
3665 "fmla v18.8h, v29.8h, v2.h[3]\n"
3666 "fmla v22.8h, v29.8h, v3.h[3]\n"
3667 "fmla v26.8h, v29.8h, v4.h[3]\n"
3668 "ldr d29, [x17, #0x100]\n"
3669 "fmla v11.8h, v28.8h, v0.h[3]\n"
3670 "mov v29.d[1], x21\n"
3671 "fmla v15.8h, v28.8h, v1.h[3]\n"
3672 "ldr x21, [x17, #0x128]\n"
3673 "fmla v19.8h, v28.8h, v2.h[3]\n"
3674 "fmla v23.8h, v28.8h, v3.h[3]\n"
3675 "fmla v27.8h, v28.8h, v4.h[3]\n"
3676 "ldr d28, [x17, #0x110]\n"
3677 "mov v28.d[1], x20\n"
3678 "fmla v8.8h, v29.8h, v0.h[4]\n"
3679 "fmla v12.8h, v29.8h, v1.h[4]\n"
3680 "ldr x20, [x17, #0x138]\n"
3681 "fmla v16.8h, v29.8h, v2.h[4]\n"
3682 "fmla v20.8h, v29.8h, v3.h[4]\n"
3683 "fmla v24.8h, v29.8h, v4.h[4]\n"
3684 "ldr d29, [x17, #0x120]\n"
3685 "fmla v9.8h, v28.8h, v0.h[4]\n"
3686 "mov v29.d[1], x21\n"
3687 "fmla v13.8h, v28.8h, v1.h[4]\n"
3688 "ldr x21, [x17, #0x148]\n"
3689 "fmla v17.8h, v28.8h, v2.h[4]\n"
3690 "fmla v21.8h, v28.8h, v3.h[4]\n"
3691 "fmla v25.8h, v28.8h, v4.h[4]\n"
3692 "ldr d28, [x17, #0x130]\n"
3693 "mov v28.d[1], x20\n"
3694 "fmla v10.8h, v29.8h, v0.h[4]\n"
3695 "fmla v14.8h, v29.8h, v1.h[4]\n"
3696 "ldr x20, [x17, #0x158]\n"
3697 "fmla v18.8h, v29.8h, v2.h[4]\n"
3698 "fmla v22.8h, v29.8h, v3.h[4]\n"
3699 "fmla v26.8h, v29.8h, v4.h[4]\n"
3700 "ldr d29, [x17, #0x140]\n"
3701 "fmla v11.8h, v28.8h, v0.h[4]\n"
3702 "mov v29.d[1], x21\n"
3703 "fmla v15.8h, v28.8h, v1.h[4]\n"
3704 "ldr x21, [x17, #0x168]\n"
3705 "fmla v19.8h, v28.8h, v2.h[4]\n"
3706 "fmla v23.8h, v28.8h, v3.h[4]\n"
3707 "fmla v27.8h, v28.8h, v4.h[4]\n"
3708 "ldr d28, [x17, #0x150]\n"
3709 "mov v28.d[1], x20\n"
3710 "fmla v8.8h, v29.8h, v0.h[5]\n"
3711 "fmla v12.8h, v29.8h, v1.h[5]\n"
3712 "ldr x20, [x17, #0x178]\n"
3713 "fmla v16.8h, v29.8h, v2.h[5]\n"
3714 "fmla v20.8h, v29.8h, v3.h[5]\n"
3715 "fmla v24.8h, v29.8h, v4.h[5]\n"
3716 "ldr d29, [x17, #0x160]\n"
3717 "fmla v9.8h, v28.8h, v0.h[5]\n"
3718 "mov v29.d[1], x21\n"
3719 "fmla v13.8h, v28.8h, v1.h[5]\n"
3720 "ldr x21, [x17, #0x188]\n"
3721 "fmla v17.8h, v28.8h, v2.h[5]\n"
3722 "fmla v21.8h, v28.8h, v3.h[5]\n"
3723 "fmla v25.8h, v28.8h, v4.h[5]\n"
3724 "ldr d28, [x17, #0x170]\n"
3725 "mov v28.d[1], x20\n"
3726 "fmla v10.8h, v29.8h, v0.h[5]\n"
3727 "fmla v14.8h, v29.8h, v1.h[5]\n"
3728 "ldr x20, [x17, #0x198]\n"
3729 "fmla v18.8h, v29.8h, v2.h[5]\n"
3730 "fmla v22.8h, v29.8h, v3.h[5]\n"
3731 "fmla v26.8h, v29.8h, v4.h[5]\n"
3732 "ldr d29, [x17, #0x180]\n"
3733 "fmla v11.8h, v28.8h, v0.h[5]\n"
3734 "mov v29.d[1], x21\n"
3735 "fmla v15.8h, v28.8h, v1.h[5]\n"
3736 "ldr x21, [x17, #0x1a8]\n"
3737 "fmla v19.8h, v28.8h, v2.h[5]\n"
3738 "fmla v23.8h, v28.8h, v3.h[5]\n"
3739 "fmla v27.8h, v28.8h, v4.h[5]\n"
3740 "ldr d28, [x17, #0x190]\n"
3741 "mov v28.d[1], x20\n"
3742 "fmla v8.8h, v29.8h, v0.h[6]\n"
3743 "fmla v12.8h, v29.8h, v1.h[6]\n"
3744 "ldr x20, [x17, #0x1b8]\n"
3745 "fmla v16.8h, v29.8h, v2.h[6]\n"
3746 "fmla v20.8h, v29.8h, v3.h[6]\n"
3747 "fmla v24.8h, v29.8h, v4.h[6]\n"
3748 "ldr d29, [x17, #0x1a0]\n"
3749 "fmla v9.8h, v28.8h, v0.h[6]\n"
3750 "mov v29.d[1], x21\n"
3751 "fmla v13.8h, v28.8h, v1.h[6]\n"
3752 "ldr x21, [x17, #0x1c8]\n"
3753 "fmla v17.8h, v28.8h, v2.h[6]\n"
3754 "fmla v21.8h, v28.8h, v3.h[6]\n"
3755 "fmla v25.8h, v28.8h, v4.h[6]\n"
3756 "ldr d28, [x17, #0x1b0]\n"
3757 "mov v28.d[1], x20\n"
3758 "fmla v10.8h, v29.8h, v0.h[6]\n"
3759 "fmla v14.8h, v29.8h, v1.h[6]\n"
3760 "ldr x20, [x17, #0x1d8]\n"
3761 "fmla v18.8h, v29.8h, v2.h[6]\n"
3762 "fmla v22.8h, v29.8h, v3.h[6]\n"
3763 "fmla v26.8h, v29.8h, v4.h[6]\n"
3764 "ldr d29, [x17, #0x1c0]\n"
3765 "fmla v11.8h, v28.8h, v0.h[6]\n"
3766 "mov v29.d[1], x21\n"
3767 "fmla v15.8h, v28.8h, v1.h[6]\n"
3768 "ldr x21, [x17, #0x1e8]\n"
3769 "fmla v19.8h, v28.8h, v2.h[6]\n"
3770 "fmla v23.8h, v28.8h, v3.h[6]\n"
3771 "fmla v27.8h, v28.8h, v4.h[6]\n"
3772 "ldr d28, [x17, #0x1d0]\n"
3773 "mov v28.d[1], x20\n"
3774 "fmla v8.8h, v29.8h, v0.h[7]\n"
3775 "fmla v12.8h, v29.8h, v1.h[7]\n"
3776 "ldr x20, [x17, #0x1f8]\n"
3777 "fmla v16.8h, v29.8h, v2.h[7]\n"
3778 "fmla v20.8h, v29.8h, v3.h[7]\n"
3779 "fmla v24.8h, v29.8h, v4.h[7]\n"
3780 "ldr d29, [x17, #0x1e0]\n"
3781 "fmla v9.8h, v28.8h, v0.h[7]\n"
3782 "mov v29.d[1], x21\n"
3783 "fmla v13.8h, v28.8h, v1.h[7]\n"
3784 "fmla v17.8h, v28.8h, v2.h[7]\n"
3785 "fmla v21.8h, v28.8h, v3.h[7]\n"
3786 "fmla v25.8h, v28.8h, v4.h[7]\n"
3787 "ldr d28, [x17, #0x1f0]\n"
3788 "mov v28.d[1], x20\n"
3789 "add x17, x17, #0x200\n"
3790 "fmla v10.8h, v29.8h, v0.h[7]\n"
3791 "ldr x21, [x17, #0x8]\n"
3792 "fmla v14.8h, v29.8h, v1.h[7]\n"
3793 "ldr x20, [x17, #0x18]\n"
3794 "fmla v18.8h, v29.8h, v2.h[7]\n"
3795 "fmla v22.8h, v29.8h, v3.h[7]\n"
3796 "fmla v26.8h, v29.8h, v4.h[7]\n"
3797 "ldr d6, [x17, #0x0]\n"
3798 "fmla v11.8h, v28.8h, v0.h[7]\n"
3799 "ldr d0, [x13, #0x0]\n"
3800 "fmla v15.8h, v28.8h, v1.h[7]\n"
3801 "ldr d1, [x12, #0x0]\n"
3802 "fmla v19.8h, v28.8h, v2.h[7]\n"
3803 "ldr d2, [x11, #0x0]\n"
3804 "fmla v23.8h, v28.8h, v3.h[7]\n"
3805 "ldr d3, [x10, #0x0]\n"
3806 "fmla v27.8h, v28.8h, v4.h[7]\n"
3807 "ldr d4, [x9, #0x0]\n"
3808 "ldr d7, [x17, #0x10]\n"
3809 "mov v6.d[1], x21\n"
3810 "mov v0.d[1], x26\n"
3811 "mov v1.d[1], x25\n"
3812 "mov v2.d[1], x24\n"
3813 "mov v3.d[1], x23\n"
3814 "mov v4.d[1], x22\n"
3815 "mov v7.d[1], x20\n"
3818 "fmla v8.8h, v6.8h, v0.h[0]\n"
3819 "add x13, x13, #0x10\n"
3820 "fmla v12.8h, v6.8h, v1.h[0]\n"
3821 "add x12, x12, #0x10\n"
3822 "fmla v16.8h, v6.8h, v2.h[0]\n"
3823 "add x11, x11, #0x10\n"
3824 "fmla v20.8h, v6.8h, v3.h[0]\n"
3825 "add x10, x10, #0x10\n"
3826 "fmla v24.8h, v6.8h, v4.h[0]\n"
3827 "ldr q29, [x17, #0x20]\n"
3828 "fmla v9.8h, v7.8h, v0.h[0]\n"
3829 "add x9, x9, #0x10\n"
3830 "fmla v13.8h, v7.8h, v1.h[0]\n"
3831 "sub x14, x14, #0x8\n"
3832 "fmla v17.8h, v7.8h, v2.h[0]\n"
3833 "prfm pldl1keep, [x13, #0x80]\n"
3834 "fmla v21.8h, v7.8h, v3.h[0]\n"
3835 "prfm pldl1keep, [x12, #0x80]\n"
3836 "fmla v25.8h, v7.8h, v4.h[0]\n"
3837 "ldr q28, [x17, #0x30]\n"
3838 "fmla v10.8h, v29.8h, v0.h[0]\n"
3839 "prfm pldl1keep, [x11, #0x80]\n"
3840 "fmla v14.8h, v29.8h, v1.h[0]\n"
3841 "prfm pldl1keep, [x10, #0x80]\n"
3842 "fmla v18.8h, v29.8h, v2.h[0]\n"
3843 "prfm pldl1keep, [x9, #0x80]\n"
3844 "fmla v22.8h, v29.8h, v3.h[0]\n"
3845 "fmla v26.8h, v29.8h, v4.h[0]\n"
3846 "ldr q29, [x17, #0x40]\n"
3847 "fmla v11.8h, v28.8h, v0.h[0]\n"
3848 "fmla v15.8h, v28.8h, v1.h[0]\n"
3849 "fmla v19.8h, v28.8h, v2.h[0]\n"
3850 "fmla v23.8h, v28.8h, v3.h[0]\n"
3851 "fmla v27.8h, v28.8h, v4.h[0]\n"
3852 "ldr q28, [x17, #0x50]\n"
3853 "fmla v8.8h, v29.8h, v0.h[1]\n"
3854 "fmla v12.8h, v29.8h, v1.h[1]\n"
3855 "fmla v16.8h, v29.8h, v2.h[1]\n"
3856 "fmla v20.8h, v29.8h, v3.h[1]\n"
3857 "fmla v24.8h, v29.8h, v4.h[1]\n"
3858 "ldr q29, [x17, #0x60]\n"
3859 "fmla v9.8h, v28.8h, v0.h[1]\n"
3860 "fmla v13.8h, v28.8h, v1.h[1]\n"
3861 "fmla v17.8h, v28.8h, v2.h[1]\n"
3862 "fmla v21.8h, v28.8h, v3.h[1]\n"
3863 "fmla v25.8h, v28.8h, v4.h[1]\n"
3864 "ldr q28, [x17, #0x70]\n"
3865 "fmla v10.8h, v29.8h, v0.h[1]\n"
3866 "fmla v14.8h, v29.8h, v1.h[1]\n"
3867 "fmla v18.8h, v29.8h, v2.h[1]\n"
3868 "fmla v22.8h, v29.8h, v3.h[1]\n"
3869 "fmla v26.8h, v29.8h, v4.h[1]\n"
3870 "ldr q29, [x17, #0x80]\n"
3871 "fmla v11.8h, v28.8h, v0.h[1]\n"
3872 "fmla v15.8h, v28.8h, v1.h[1]\n"
3873 "fmla v19.8h, v28.8h, v2.h[1]\n"
3874 "fmla v23.8h, v28.8h, v3.h[1]\n"
3875 "fmla v27.8h, v28.8h, v4.h[1]\n"
3876 "ldr q28, [x17, #0x90]\n"
3877 "fmla v8.8h, v29.8h, v0.h[2]\n"
3878 "fmla v12.8h, v29.8h, v1.h[2]\n"
3879 "fmla v16.8h, v29.8h, v2.h[2]\n"
3880 "fmla v20.8h, v29.8h, v3.h[2]\n"
3881 "fmla v24.8h, v29.8h, v4.h[2]\n"
3882 "ldr q29, [x17, #0xa0]\n"
3883 "fmla v9.8h, v28.8h, v0.h[2]\n"
3884 "fmla v13.8h, v28.8h, v1.h[2]\n"
3885 "fmla v17.8h, v28.8h, v2.h[2]\n"
3886 "fmla v21.8h, v28.8h, v3.h[2]\n"
3887 "fmla v25.8h, v28.8h, v4.h[2]\n"
3888 "ldr q28, [x17, #0xb0]\n"
3889 "fmla v10.8h, v29.8h, v0.h[2]\n"
3890 "fmla v14.8h, v29.8h, v1.h[2]\n"
3891 "fmla v18.8h, v29.8h, v2.h[2]\n"
3892 "fmla v22.8h, v29.8h, v3.h[2]\n"
3893 "fmla v26.8h, v29.8h, v4.h[2]\n"
3894 "ldr q29, [x17, #0xc0]\n"
3895 "fmla v11.8h, v28.8h, v0.h[2]\n"
3896 "fmla v15.8h, v28.8h, v1.h[2]\n"
3897 "fmla v19.8h, v28.8h, v2.h[2]\n"
3898 "fmla v23.8h, v28.8h, v3.h[2]\n"
3899 "fmla v27.8h, v28.8h, v4.h[2]\n"
3900 "ldr q28, [x17, #0xd0]\n"
3901 "fmla v8.8h, v29.8h, v0.h[3]\n"
3902 "fmla v12.8h, v29.8h, v1.h[3]\n"
3903 "fmla v16.8h, v29.8h, v2.h[3]\n"
3904 "fmla v20.8h, v29.8h, v3.h[3]\n"
3905 "fmla v24.8h, v29.8h, v4.h[3]\n"
3906 "ldr q29, [x17, #0xe0]\n"
3907 "fmla v9.8h, v28.8h, v0.h[3]\n"
3908 "fmla v13.8h, v28.8h, v1.h[3]\n"
3909 "fmla v17.8h, v28.8h, v2.h[3]\n"
3910 "fmla v21.8h, v28.8h, v3.h[3]\n"
3911 "fmla v25.8h, v28.8h, v4.h[3]\n"
3912 "ldr q28, [x17, #0xf0]\n"
3913 "fmla v10.8h, v29.8h, v0.h[3]\n"
3914 "fmla v14.8h, v29.8h, v1.h[3]\n"
3915 "fmla v18.8h, v29.8h, v2.h[3]\n"
3916 "fmla v22.8h, v29.8h, v3.h[3]\n"
3917 "fmla v26.8h, v29.8h, v4.h[3]\n"
3918 "ldr q29, [x17, #0x100]\n"
3919 "fmla v11.8h, v28.8h, v0.h[3]\n"
3920 "fmla v15.8h, v28.8h, v1.h[3]\n"
3921 "fmla v19.8h, v28.8h, v2.h[3]\n"
3922 "fmla v23.8h, v28.8h, v3.h[3]\n"
3923 "fmla v27.8h, v28.8h, v4.h[3]\n"
3924 "ldr q28, [x17, #0x110]\n"
3925 "fmla v8.8h, v29.8h, v0.h[4]\n"
3926 "fmla v12.8h, v29.8h, v1.h[4]\n"
3927 "fmla v16.8h, v29.8h, v2.h[4]\n"
3928 "fmla v20.8h, v29.8h, v3.h[4]\n"
3929 "fmla v24.8h, v29.8h, v4.h[4]\n"
3930 "ldr q29, [x17, #0x120]\n"
3931 "fmla v9.8h, v28.8h, v0.h[4]\n"
3932 "fmla v13.8h, v28.8h, v1.h[4]\n"
3933 "fmla v17.8h, v28.8h, v2.h[4]\n"
3934 "fmla v21.8h, v28.8h, v3.h[4]\n"
3935 "fmla v25.8h, v28.8h, v4.h[4]\n"
3936 "ldr q28, [x17, #0x130]\n"
3937 "fmla v10.8h, v29.8h, v0.h[4]\n"
3938 "fmla v14.8h, v29.8h, v1.h[4]\n"
3939 "fmla v18.8h, v29.8h, v2.h[4]\n"
3940 "fmla v22.8h, v29.8h, v3.h[4]\n"
3941 "fmla v26.8h, v29.8h, v4.h[4]\n"
3942 "ldr q29, [x17, #0x140]\n"
3943 "fmla v11.8h, v28.8h, v0.h[4]\n"
3944 "fmla v15.8h, v28.8h, v1.h[4]\n"
3945 "fmla v19.8h, v28.8h, v2.h[4]\n"
3946 "fmla v23.8h, v28.8h, v3.h[4]\n"
3947 "fmla v27.8h, v28.8h, v4.h[4]\n"
3948 "ldr q28, [x17, #0x150]\n"
3949 "fmla v8.8h, v29.8h, v0.h[5]\n"
3950 "fmla v12.8h, v29.8h, v1.h[5]\n"
3951 "fmla v16.8h, v29.8h, v2.h[5]\n"
3952 "fmla v20.8h, v29.8h, v3.h[5]\n"
3953 "fmla v24.8h, v29.8h, v4.h[5]\n"
3954 "ldr q29, [x17, #0x160]\n"
3955 "fmla v9.8h, v28.8h, v0.h[5]\n"
3956 "fmla v13.8h, v28.8h, v1.h[5]\n"
3957 "fmla v17.8h, v28.8h, v2.h[5]\n"
3958 "fmla v21.8h, v28.8h, v3.h[5]\n"
3959 "fmla v25.8h, v28.8h, v4.h[5]\n"
3960 "ldr q28, [x17, #0x170]\n"
3961 "fmla v10.8h, v29.8h, v0.h[5]\n"
3962 "fmla v14.8h, v29.8h, v1.h[5]\n"
3963 "fmla v18.8h, v29.8h, v2.h[5]\n"
3964 "fmla v22.8h, v29.8h, v3.h[5]\n"
3965 "fmla v26.8h, v29.8h, v4.h[5]\n"
3966 "ldr q29, [x17, #0x180]\n"
3967 "fmla v11.8h, v28.8h, v0.h[5]\n"
3968 "fmla v15.8h, v28.8h, v1.h[5]\n"
3969 "fmla v19.8h, v28.8h, v2.h[5]\n"
3970 "fmla v23.8h, v28.8h, v3.h[5]\n"
3971 "fmla v27.8h, v28.8h, v4.h[5]\n"
3972 "ldr q28, [x17, #0x190]\n"
3973 "fmla v8.8h, v29.8h, v0.h[6]\n"
3974 "fmla v12.8h, v29.8h, v1.h[6]\n"
3975 "fmla v16.8h, v29.8h, v2.h[6]\n"
3976 "fmla v20.8h, v29.8h, v3.h[6]\n"
3977 "fmla v24.8h, v29.8h, v4.h[6]\n"
3978 "ldr q29, [x17, #0x1a0]\n"
3979 "fmla v9.8h, v28.8h, v0.h[6]\n"
3980 "fmla v13.8h, v28.8h, v1.h[6]\n"
3981 "fmla v17.8h, v28.8h, v2.h[6]\n"
3982 "fmla v21.8h, v28.8h, v3.h[6]\n"
3983 "fmla v25.8h, v28.8h, v4.h[6]\n"
3984 "ldr q28, [x17, #0x1b0]\n"
3985 "fmla v10.8h, v29.8h, v0.h[6]\n"
3986 "fmla v14.8h, v29.8h, v1.h[6]\n"
3987 "fmla v18.8h, v29.8h, v2.h[6]\n"
3988 "fmla v22.8h, v29.8h, v3.h[6]\n"
3989 "fmla v26.8h, v29.8h, v4.h[6]\n"
3990 "ldr q29, [x17, #0x1c0]\n"
3991 "fmla v11.8h, v28.8h, v0.h[6]\n"
3992 "fmla v15.8h, v28.8h, v1.h[6]\n"
3993 "fmla v19.8h, v28.8h, v2.h[6]\n"
3994 "fmla v23.8h, v28.8h, v3.h[6]\n"
3995 "fmla v27.8h, v28.8h, v4.h[6]\n"
3996 "ldr q28, [x17, #0x1d0]\n"
3997 "fmla v8.8h, v29.8h, v0.h[7]\n"
3998 "fmla v12.8h, v29.8h, v1.h[7]\n"
3999 "fmla v16.8h, v29.8h, v2.h[7]\n"
4000 "fmla v20.8h, v29.8h, v3.h[7]\n"
4001 "fmla v24.8h, v29.8h, v4.h[7]\n"
4002 "ldr q29, [x17, #0x1e0]\n"
4003 "fmla v9.8h, v28.8h, v0.h[7]\n"
4004 "fmla v13.8h, v28.8h, v1.h[7]\n"
4005 "fmla v17.8h, v28.8h, v2.h[7]\n"
4006 "fmla v21.8h, v28.8h, v3.h[7]\n"
4007 "fmla v25.8h, v28.8h, v4.h[7]\n"
4008 "ldr q28, [x17, #0x1f0]\n"
4009 "fmla v10.8h, v29.8h, v0.h[7]\n"
4010 "add x17, x17, #0x200\n"
4011 "fmla v14.8h, v29.8h, v1.h[7]\n"
4012 "fmla v18.8h, v29.8h, v2.h[7]\n"
4013 "fmla v22.8h, v29.8h, v3.h[7]\n"
4014 "fmla v26.8h, v29.8h, v4.h[7]\n"
4015 "fmla v11.8h, v28.8h, v0.h[7]\n"
4016 "fmla v15.8h, v28.8h, v1.h[7]\n"
4017 "fmla v19.8h, v28.8h, v2.h[7]\n"
4018 "fmla v23.8h, v28.8h, v3.h[7]\n"
4019 "fmla v27.8h, v28.8h, v4.h[7]\n"
4023 "ldr h4, [x13], #0x2\n"
4024 "sub x14, x14, #0x1\n"
4025 "ldr h3, [x12], #0x2\n"
4026 "ldr h2, [x11], #0x2\n"
4027 "ldr h1, [x10], #0x2\n"
4028 "ldr h0, [x9], #0x2\n"
4029 "ldr q29, [x17, #0x0]\n"
4030 "fmla v8.8h, v29.8h, v4.h[0]\n"
4031 "ldr q28, [x17, #0x10]\n"
4032 "fmla v12.8h, v29.8h, v3.h[0]\n"
4033 "fmla v16.8h, v29.8h, v2.h[0]\n"
4034 "fmla v20.8h, v29.8h, v1.h[0]\n"
4035 "fmla v24.8h, v29.8h, v0.h[0]\n"
4036 "ldr q29, [x17, #0x20]\n"
4037 "fmla v9.8h, v28.8h, v4.h[0]\n"
4038 "fmla v13.8h, v28.8h, v3.h[0]\n"
4039 "fmla v17.8h, v28.8h, v2.h[0]\n"
4040 "fmla v21.8h, v28.8h, v1.h[0]\n"
4041 "fmla v25.8h, v28.8h, v0.h[0]\n"
4042 "ldr q28, [x17, #0x30]\n"
4043 "fmla v10.8h, v29.8h, v4.h[0]\n"
4044 "add x17, x17, #0x40\n"
4045 "fmla v14.8h, v29.8h, v3.h[0]\n"
4046 "fmla v18.8h, v29.8h, v2.h[0]\n"
4047 "fmla v22.8h, v29.8h, v1.h[0]\n"
4048 "fmla v26.8h, v29.8h, v0.h[0]\n"
4049 "fmla v11.8h, v28.8h, v4.h[0]\n"
4050 "fmla v15.8h, v28.8h, v3.h[0]\n"
4051 "fmla v19.8h, v28.8h, v2.h[0]\n"
4052 "fmla v23.8h, v28.8h, v1.h[0]\n"
4053 "fmla v27.8h, v28.8h, v0.h[0]\n"
4056 "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
4057 "add x15, x15, #0x1\n"
4060 "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
4061 "add x25, x16, x20, LSL #1\n"
4062 "add x24, x25, x20, LSL #1\n"
4063 "add x23, x24, x20, LSL #1\n"
4064 "add x22, x23, x20, LSL #1\n"
4065 "prfm pstl1keep, [x16, #0x0]\n"
4066 "prfm pstl1keep, [x25, #0x0]\n"
4067 "prfm pstl1keep, [x24, #0x0]\n"
4068 "prfm pstl1keep, [x23, #0x0]\n"
4069 "prfm pstl1keep, [x22, #0x0]\n"
4070 "tbz %x[flags], #1, 227f\n"
4071 "add x20, %x[args_ptr], %[offset_max]\n"
4072 "ld1r { v28.8h }, [x20]\n"
4073 "fmin v8.8h, v8.8h, v28.8h\n"
4074 "fmin v9.8h, v9.8h, v28.8h\n"
4075 "fmin v10.8h, v10.8h, v28.8h\n"
4076 "fmin v11.8h, v11.8h, v28.8h\n"
4077 "fmin v12.8h, v12.8h, v28.8h\n"
4078 "fmin v13.8h, v13.8h, v28.8h\n"
4079 "fmin v14.8h, v14.8h, v28.8h\n"
4080 "fmin v15.8h, v15.8h, v28.8h\n"
4081 "fmin v16.8h, v16.8h, v28.8h\n"
4082 "fmin v17.8h, v17.8h, v28.8h\n"
4083 "fmin v18.8h, v18.8h, v28.8h\n"
4084 "fmin v19.8h, v19.8h, v28.8h\n"
4085 "fmin v20.8h, v20.8h, v28.8h\n"
4086 "fmin v21.8h, v21.8h, v28.8h\n"
4087 "fmin v22.8h, v22.8h, v28.8h\n"
4088 "fmin v23.8h, v23.8h, v28.8h\n"
4089 "fmin v24.8h, v24.8h, v28.8h\n"
4090 "fmin v25.8h, v25.8h, v28.8h\n"
4091 "fmin v26.8h, v26.8h, v28.8h\n"
4092 "fmin v27.8h, v27.8h, v28.8h\n"
4093 "add x20, %x[args_ptr], %[offset_min]\n"
4094 "ld1r { v28.8h }, [x20]\n"
4095 "fmax v8.8h, v8.8h, v28.8h\n"
4096 "fmax v9.8h, v9.8h, v28.8h\n"
4097 "fmax v10.8h, v10.8h, v28.8h\n"
4098 "fmax v11.8h, v11.8h, v28.8h\n"
4099 "fmax v12.8h, v12.8h, v28.8h\n"
4100 "fmax v13.8h, v13.8h, v28.8h\n"
4101 "fmax v14.8h, v14.8h, v28.8h\n"
4102 "fmax v15.8h, v15.8h, v28.8h\n"
4103 "fmax v16.8h, v16.8h, v28.8h\n"
4104 "fmax v17.8h, v17.8h, v28.8h\n"
4105 "fmax v18.8h, v18.8h, v28.8h\n"
4106 "fmax v19.8h, v19.8h, v28.8h\n"
4107 "fmax v20.8h, v20.8h, v28.8h\n"
4108 "fmax v21.8h, v21.8h, v28.8h\n"
4109 "fmax v22.8h, v22.8h, v28.8h\n"
4110 "fmax v23.8h, v23.8h, v28.8h\n"
4111 "fmax v24.8h, v24.8h, v28.8h\n"
4112 "fmax v25.8h, v25.8h, v28.8h\n"
4113 "fmax v26.8h, v26.8h, v28.8h\n"
4114 "fmax v27.8h, v27.8h, v28.8h\n"
4118 "tbz x8, #4, 235f\n"
4119 "st1 { v8.8h }, [x16], #0x10\n"
4120 "st1 { v9.8h }, [x16], #0x10\n"
4121 "st1 { v12.8h }, [x25], #0x10\n"
4122 "st1 { v13.8h }, [x25], #0x10\n"
4123 "st1 { v16.8h }, [x24], #0x10\n"
4124 "st1 { v17.8h }, [x24], #0x10\n"
4125 "st1 { v20.8h }, [x23], #0x10\n"
4126 "st1 { v21.8h }, [x23], #0x10\n"
4127 "st1 { v24.8h }, [x22], #0x10\n"
4128 "st1 { v25.8h }, [x22], #0x10\n"
4129 "tbz x8, #3, 231f\n"
4130 "st1 { v10.8h }, [x16], #0x10\n"
4131 "st1 { v14.8h }, [x25], #0x10\n"
4132 "st1 { v18.8h }, [x24], #0x10\n"
4133 "st1 { v22.8h }, [x23], #0x10\n"
4134 "st1 { v26.8h }, [x22], #0x10\n"
4135 "tbz x8, #2, 229f\n"
4136 "str d11, [x16], #0x8\n"
4137 "str d15, [x25], #0x8\n"
4138 "str d19, [x24], #0x8\n"
4139 "str d23, [x23], #0x8\n"
4140 "str d27, [x22], #0x8\n"
4141 "tbz x8, #1, 228f\n"
4142 "st1 { v11.s }[2], [x16], #0x4\n"
4143 "st1 { v15.s }[2], [x25], #0x4\n"
4144 "st1 { v19.s }[2], [x24], #0x4\n"
4145 "st1 { v23.s }[2], [x23], #0x4\n"
4146 "st1 { v27.s }[2], [x22], #0x4\n"
4147 "tbz x8, #0, 243f\n"
4148 "st1 { v11.h }[6], [x16]\n"
4149 "st1 { v15.h }[6], [x25]\n"
4150 "st1 { v19.h }[6], [x24]\n"
4151 "st1 { v23.h }[6], [x23]\n"
4152 "st1 { v27.h }[6], [x22]\n"
4155 "tbz x8, #0, 243f\n"
4156 "st1 { v11.h }[4], [x16]\n"
4157 "st1 { v15.h }[4], [x25]\n"
4158 "st1 { v19.h }[4], [x24]\n"
4159 "st1 { v23.h }[4], [x23]\n"
4160 "st1 { v27.h }[4], [x22]\n"
4163 "tbz x8, #1, 230f\n"
4164 "str s11, [x16], #0x4\n"
4165 "str s15, [x25], #0x4\n"
4166 "str s19, [x24], #0x4\n"
4167 "str s23, [x23], #0x4\n"
4168 "str s27, [x22], #0x4\n"
4169 "tbz x8, #0, 243f\n"
4170 "st1 { v11.h }[2], [x16]\n"
4171 "st1 { v15.h }[2], [x25]\n"
4172 "st1 { v19.h }[2], [x24]\n"
4173 "st1 { v23.h }[2], [x23]\n"
4174 "st1 { v27.h }[2], [x22]\n"
4177 "tbz x8, #0, 243f\n"
4178 "str h11, [x16, #0x0]\n"
4179 "str h15, [x25, #0x0]\n"
4180 "str h19, [x24, #0x0]\n"
4181 "str h23, [x23, #0x0]\n"
4182 "str h27, [x22, #0x0]\n"
4185 "tbz x8, #2, 233f\n"
4186 "str d10, [x16], #0x8\n"
4187 "str d14, [x25], #0x8\n"
4188 "str d18, [x24], #0x8\n"
4189 "str d22, [x23], #0x8\n"
4190 "str d26, [x22], #0x8\n"
4191 "tbz x8, #1, 232f\n"
4192 "st1 { v10.s }[2], [x16], #0x4\n"
4193 "st1 { v14.s }[2], [x25], #0x4\n"
4194 "st1 { v18.s }[2], [x24], #0x4\n"
4195 "st1 { v22.s }[2], [x23], #0x4\n"
4196 "st1 { v26.s }[2], [x22], #0x4\n"
4197 "tbz x8, #0, 243f\n"
4198 "st1 { v10.h }[6], [x16]\n"
4199 "st1 { v14.h }[6], [x25]\n"
4200 "st1 { v18.h }[6], [x24]\n"
4201 "st1 { v22.h }[6], [x23]\n"
4202 "st1 { v26.h }[6], [x22]\n"
4205 "tbz x8, #0, 243f\n"
4206 "st1 { v10.h }[4], [x16]\n"
4207 "st1 { v14.h }[4], [x25]\n"
4208 "st1 { v18.h }[4], [x24]\n"
4209 "st1 { v22.h }[4], [x23]\n"
4210 "st1 { v26.h }[4], [x22]\n"
4213 "tbz x8, #1, 234f\n"
4214 "str s10, [x16], #0x4\n"
4215 "str s14, [x25], #0x4\n"
4216 "str s18, [x24], #0x4\n"
4217 "str s22, [x23], #0x4\n"
4218 "str s26, [x22], #0x4\n"
4219 "tbz x8, #0, 243f\n"
4220 "st1 { v10.h }[2], [x16]\n"
4221 "st1 { v14.h }[2], [x25]\n"
4222 "st1 { v18.h }[2], [x24]\n"
4223 "st1 { v22.h }[2], [x23]\n"
4224 "st1 { v26.h }[2], [x22]\n"
4227 "tbz x8, #0, 243f\n"
4228 "str h10, [x16, #0x0]\n"
4229 "str h14, [x25, #0x0]\n"
4230 "str h18, [x24, #0x0]\n"
4231 "str h22, [x23, #0x0]\n"
4232 "str h26, [x22, #0x0]\n"
4235 "tbz x8, #3, 239f\n"
4236 "st1 { v8.8h }, [x16], #0x10\n"
4237 "st1 { v12.8h }, [x25], #0x10\n"
4238 "st1 { v16.8h }, [x24], #0x10\n"
4239 "st1 { v20.8h }, [x23], #0x10\n"
4240 "st1 { v24.8h }, [x22], #0x10\n"
4241 "tbz x8, #2, 237f\n"
4242 "str d9, [x16], #0x8\n"
4243 "str d13, [x25], #0x8\n"
4244 "str d17, [x24], #0x8\n"
4245 "str d21, [x23], #0x8\n"
4246 "str d25, [x22], #0x8\n"
4247 "tbz x8, #1, 236f\n"
4248 "st1 { v9.s }[2], [x16], #0x4\n"
4249 "st1 { v13.s }[2], [x25], #0x4\n"
4250 "st1 { v17.s }[2], [x24], #0x4\n"
4251 "st1 { v21.s }[2], [x23], #0x4\n"
4252 "st1 { v25.s }[2], [x22], #0x4\n"
4253 "tbz x8, #0, 243f\n"
4254 "st1 { v9.h }[6], [x16]\n"
4255 "st1 { v13.h }[6], [x25]\n"
4256 "st1 { v17.h }[6], [x24]\n"
4257 "st1 { v21.h }[6], [x23]\n"
4258 "st1 { v25.h }[6], [x22]\n"
4261 "tbz x8, #0, 243f\n"
4262 "st1 { v9.h }[4], [x16]\n"
4263 "st1 { v13.h }[4], [x25]\n"
4264 "st1 { v17.h }[4], [x24]\n"
4265 "st1 { v21.h }[4], [x23]\n"
4266 "st1 { v25.h }[4], [x22]\n"
4269 "tbz x8, #1, 238f\n"
4270 "str s9, [x16], #0x4\n"
4271 "str s13, [x25], #0x4\n"
4272 "str s17, [x24], #0x4\n"
4273 "str s21, [x23], #0x4\n"
4274 "str s25, [x22], #0x4\n"
4275 "tbz x8, #0, 243f\n"
4276 "st1 { v9.h }[2], [x16]\n"
4277 "st1 { v13.h }[2], [x25]\n"
4278 "st1 { v17.h }[2], [x24]\n"
4279 "st1 { v21.h }[2], [x23]\n"
4280 "st1 { v25.h }[2], [x22]\n"
4283 "tbz x8, #0, 243f\n"
4284 "str h9, [x16, #0x0]\n"
4285 "str h13, [x25, #0x0]\n"
4286 "str h17, [x24, #0x0]\n"
4287 "str h21, [x23, #0x0]\n"
4288 "str h25, [x22, #0x0]\n"
4291 "tbz x8, #2, 241f\n"
4292 "str d8, [x16], #0x8\n"
4293 "str d12, [x25], #0x8\n"
4294 "str d16, [x24], #0x8\n"
4295 "str d20, [x23], #0x8\n"
4296 "str d24, [x22], #0x8\n"
4297 "tbz x8, #1, 240f\n"
4298 "st1 { v8.s }[2], [x16], #0x4\n"
4299 "st1 { v12.s }[2], [x25], #0x4\n"
4300 "st1 { v16.s }[2], [x24], #0x4\n"
4301 "st1 { v20.s }[2], [x23], #0x4\n"
4302 "st1 { v24.s }[2], [x22], #0x4\n"
4303 "tbz x8, #0, 243f\n"
4304 "st1 { v8.h }[6], [x16]\n"
4305 "st1 { v12.h }[6], [x25]\n"
4306 "st1 { v16.h }[6], [x24]\n"
4307 "st1 { v20.h }[6], [x23]\n"
4308 "st1 { v24.h }[6], [x22]\n"
4311 "tbz x8, #0, 243f\n"
4312 "st1 { v8.h }[4], [x16]\n"
4313 "st1 { v12.h }[4], [x25]\n"
4314 "st1 { v16.h }[4], [x24]\n"
4315 "st1 { v20.h }[4], [x23]\n"
4316 "st1 { v24.h }[4], [x22]\n"
4319 "tbz x8, #1, 242f\n"
4320 "str s8, [x16], #0x4\n"
4321 "str s12, [x25], #0x4\n"
4322 "str s16, [x24], #0x4\n"
4323 "str s20, [x23], #0x4\n"
4324 "str s24, [x22], #0x4\n"
4325 "tbz x8, #0, 243f\n"
4326 "st1 { v8.h }[2], [x16]\n"
4327 "st1 { v12.h }[2], [x25]\n"
4328 "st1 { v16.h }[2], [x24]\n"
4329 "st1 { v20.h }[2], [x23]\n"
4330 "st1 { v24.h }[2], [x22]\n"
4333 "str h8, [x16, #0x0]\n"
4334 "str h12, [x25, #0x0]\n"
4335 "str h16, [x24, #0x0]\n"
4336 "str h20, [x23, #0x0]\n"
4337 "str h24, [x22, #0x0]\n"
4341 "str q8, [x16, #0x0]\n"
4342 "str q9, [x16, #0x10]\n"
4343 "str q10, [x16, #0x20]\n"
4344 "str q11, [x16, #0x30]\n"
4345 "add x16, x16, #0x40\n"
4346 "str q12, [x25, #0x0]\n"
4347 "str q13, [x25, #0x10]\n"
4348 "str q14, [x25, #0x20]\n"
4349 "str q15, [x25, #0x30]\n"
4350 "str q16, [x24, #0x0]\n"
4351 "str q17, [x24, #0x10]\n"
4352 "str q18, [x24, #0x20]\n"
4353 "str q19, [x24, #0x30]\n"
4354 "str q20, [x23, #0x0]\n"
4355 "str q21, [x23, #0x10]\n"
4356 "str q22, [x23, #0x20]\n"
4357 "str q23, [x23, #0x30]\n"
4358 "str q24, [x22, #0x0]\n"
4359 "str q25, [x22, #0x10]\n"
4360 "str q26, [x22, #0x20]\n"
4361 "str q27, [x22, #0x30]\n"
4363 "subs x8, x8, #0x20\n"
4367 "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
4369 "mov x7, %x[bias]\n"
4370 "ldr x8, [%x[args_ptr], %[offsetof_N]]\n"
4371 "ldr x17, [%x[args_ptr], %[offsetof_B_ptr]]\n"
4372 "mov x16, %x[output_ptr]\n"
4373 "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
4376 "ldr q8, [x7, #0x0]\n"
4377 "mov v12.16b, v8.16b\n"
4378 "ldr q9, [x7, #0x10]\n"
4379 "mov v13.16b, v9.16b\n"
4380 "ldr q10, [x7, #0x20]\n"
4381 "mov v14.16b, v10.16b\n"
4382 "ldr q11, [x7, #0x30]\n"
4383 "mov v15.16b, v11.16b\n"
4384 "mov v16.16b, v8.16b\n"
4385 "add x7, x7, #0x40\n"
4386 "mov v17.16b, v9.16b\n"
4387 "mov v18.16b, v10.16b\n"
4388 "mov v19.16b, v11.16b\n"
4389 "mov v20.16b, v8.16b\n"
4390 "mov v21.16b, v9.16b\n"
4391 "mov v22.16b, v10.16b\n"
4392 "mov v23.16b, v11.16b\n"
4393 "mov v24.16b, v8.16b\n"
4394 "mov v25.16b, v9.16b\n"
4395 "mov v26.16b, v10.16b\n"
4396 "mov v27.16b, v11.16b\n"
4397 "mov v28.16b, v8.16b\n"
4398 "mov v29.16b, v9.16b\n"
4399 "mov v30.16b, v10.16b\n"
4400 "mov v31.16b, v11.16b\n"
4403 "tbz %x[flags], #0, 266f\n"
4404 "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
4405 "add x25, x16, x20, LSL #1\n"
4406 "add x24, x25, x20, LSL #1\n"
4407 "add x23, x24, x20, LSL #1\n"
4408 "add x22, x23, x20, LSL #1\n"
4410 "add x21, x22, x20, LSL #1\n"
4412 "tbz x8, #4, 256f\n"
4413 "ld1 { v8.8h }, [x16], #0x10\n"
4414 "ld1 { v12.8h }, [x25], #0x10\n"
4415 "ld1 { v16.8h }, [x24], #0x10\n"
4416 "ld1 { v20.8h }, [x23], #0x10\n"
4417 "ld1 { v24.8h }, [x22], #0x10\n"
4418 "ld1 { v28.8h }, [x21], #0x10\n"
4419 "ld1 { v9.8h }, [x16], #0x10\n"
4420 "ld1 { v13.8h }, [x25], #0x10\n"
4421 "ld1 { v17.8h }, [x24], #0x10\n"
4422 "ld1 { v21.8h }, [x23], #0x10\n"
4423 "ld1 { v25.8h }, [x22], #0x10\n"
4424 "ld1 { v29.8h }, [x21], #0x10\n"
4425 "tbz x8, #3, 252f\n"
4426 "ld1 { v10.8h }, [x16], #0x10\n"
4427 "ld1 { v14.8h }, [x25], #0x10\n"
4428 "ld1 { v18.8h }, [x24], #0x10\n"
4429 "ld1 { v22.8h }, [x23], #0x10\n"
4430 "ld1 { v26.8h }, [x22], #0x10\n"
4431 "ld1 { v30.8h }, [x21], #0x10\n"
4432 "tbz x8, #2, 250f\n"
4433 "ldr d11, [x16], #0x8\n"
4434 "ldr d15, [x25], #0x8\n"
4435 "ldr d19, [x24], #0x8\n"
4436 "ldr d23, [x23], #0x8\n"
4437 "ldr d27, [x22], #0x8\n"
4438 "ldr d31, [x21], #0x8\n"
4439 "tbz x8, #1, 249f\n"
4440 "ld1 { v11.s }[2], [x16], #0x4\n"
4442 "ld1 { v15.s }[2], [x25], #0x4\n"
4443 "ld1 { v19.s }[2], [x24], #0x4\n"
4444 "ld1 { v23.s }[2], [x23], #0x4\n"
4445 "ld1 { v27.s }[2], [x22], #0x4\n"
4446 "ld1 { v31.s }[2], [x21], #0x4\n"
4447 "tbz x8, #0, 264f\n"
4448 "ld1 { v11.h }[6], [x16]\n"
4449 "ld1 { v15.h }[6], [x25]\n"
4450 "ld1 { v19.h }[6], [x24]\n"
4451 "ld1 { v23.h }[6], [x23]\n"
4452 "ld1 { v27.h }[6], [x22]\n"
4453 "ld1 { v31.h }[6], [x21]\n"
4457 "tbz x8, #0, 264f\n"
4458 "ld1 { v11.h }[4], [x16]\n"
4459 "ld1 { v15.h }[4], [x25]\n"
4460 "ld1 { v19.h }[4], [x24]\n"
4461 "ld1 { v23.h }[4], [x23]\n"
4462 "ld1 { v27.h }[4], [x22]\n"
4463 "ld1 { v31.h }[4], [x21]\n"
4466 "tbz x8, #1, 251f\n"
4467 "ldr s11, [x16], #0x4\n"
4469 "ldr s15, [x25], #0x4\n"
4470 "ldr s19, [x24], #0x4\n"
4471 "ldr s23, [x23], #0x4\n"
4472 "ldr s27, [x22], #0x4\n"
4473 "ldr s31, [x21], #0x4\n"
4474 "tbz x8, #0, 264f\n"
4475 "ld1 { v11.h }[2], [x16]\n"
4476 "ld1 { v15.h }[2], [x25]\n"
4477 "ld1 { v19.h }[2], [x24]\n"
4478 "ld1 { v23.h }[2], [x23]\n"
4479 "ld1 { v27.h }[2], [x22]\n"
4480 "ld1 { v31.h }[2], [x21]\n"
4484 "tbz x8, #0, 264f\n"
4485 "ldr h11, [x16, #0x0]\n"
4486 "ldr h15, [x25, #0x0]\n"
4487 "ldr h19, [x24, #0x0]\n"
4488 "ldr h23, [x23, #0x0]\n"
4489 "ldr h27, [x22, #0x0]\n"
4490 "ldr h31, [x21, #0x0]\n"
4493 "tbz x8, #2, 254f\n"
4494 "ldr d10, [x16], #0x8\n"
4495 "ldr d14, [x25], #0x8\n"
4496 "ldr d18, [x24], #0x8\n"
4497 "ldr d22, [x23], #0x8\n"
4498 "ldr d26, [x22], #0x8\n"
4499 "ldr d30, [x21], #0x8\n"
4500 "tbz x8, #1, 253f\n"
4501 "ld1 { v10.s }[2], [x16], #0x4\n"
4503 "ld1 { v14.s }[2], [x25], #0x4\n"
4504 "ld1 { v18.s }[2], [x24], #0x4\n"
4505 "ld1 { v22.s }[2], [x23], #0x4\n"
4506 "ld1 { v26.s }[2], [x22], #0x4\n"
4507 "ld1 { v30.s }[2], [x21], #0x4\n"
4508 "tbz x8, #0, 264f\n"
4509 "ld1 { v10.h }[6], [x16]\n"
4510 "ld1 { v14.h }[6], [x25]\n"
4511 "ld1 { v18.h }[6], [x24]\n"
4512 "ld1 { v22.h }[6], [x23]\n"
4513 "ld1 { v26.h }[6], [x22]\n"
4514 "ld1 { v30.h }[6], [x21]\n"
4518 "tbz x8, #0, 264f\n"
4519 "ld1 { v10.h }[4], [x16]\n"
4520 "ld1 { v14.h }[4], [x25]\n"
4521 "ld1 { v18.h }[4], [x24]\n"
4522 "ld1 { v22.h }[4], [x23]\n"
4523 "ld1 { v26.h }[4], [x22]\n"
4524 "ld1 { v30.h }[4], [x21]\n"
4527 "tbz x8, #1, 255f\n"
4528 "ldr s10, [x16], #0x4\n"
4530 "ldr s14, [x25], #0x4\n"
4531 "ldr s18, [x24], #0x4\n"
4532 "ldr s22, [x23], #0x4\n"
4533 "ldr s26, [x22], #0x4\n"
4534 "ldr s30, [x21], #0x4\n"
4535 "tbz x8, #0, 264f\n"
4536 "ld1 { v10.h }[2], [x16]\n"
4537 "ld1 { v14.h }[2], [x25]\n"
4538 "ld1 { v18.h }[2], [x24]\n"
4539 "ld1 { v22.h }[2], [x23]\n"
4540 "ld1 { v26.h }[2], [x22]\n"
4541 "ld1 { v30.h }[2], [x21]\n"
4545 "tbz x8, #0, 264f\n"
4546 "ldr h10, [x16, #0x0]\n"
4547 "ldr h14, [x25, #0x0]\n"
4548 "ldr h18, [x24, #0x0]\n"
4549 "ldr h22, [x23, #0x0]\n"
4550 "ldr h26, [x22, #0x0]\n"
4551 "ldr h30, [x21, #0x0]\n"
4554 "tbz x8, #3, 260f\n"
4555 "ld1 { v8.8h }, [x16], #0x10\n"
4556 "ld1 { v12.8h }, [x25], #0x10\n"
4557 "ld1 { v16.8h }, [x24], #0x10\n"
4558 "ld1 { v20.8h }, [x23], #0x10\n"
4559 "ld1 { v24.8h }, [x22], #0x10\n"
4560 "ld1 { v28.8h }, [x21], #0x10\n"
4561 "tbz x8, #2, 258f\n"
4562 "ldr d9, [x16], #0x8\n"
4563 "ldr d13, [x25], #0x8\n"
4564 "ldr d17, [x24], #0x8\n"
4565 "ldr d21, [x23], #0x8\n"
4566 "ldr d25, [x22], #0x8\n"
4567 "ldr d29, [x21], #0x8\n"
4568 "tbz x8, #1, 257f\n"
4569 "ld1 { v9.s }[2], [x16], #0x4\n"
4571 "ld1 { v13.s }[2], [x25], #0x4\n"
4572 "ld1 { v17.s }[2], [x24], #0x4\n"
4573 "ld1 { v21.s }[2], [x23], #0x4\n"
4574 "ld1 { v25.s }[2], [x22], #0x4\n"
4575 "ld1 { v29.s }[2], [x21], #0x4\n"
4576 "tbz x8, #0, 264f\n"
4577 "ld1 { v9.h }[6], [x16]\n"
4578 "ld1 { v13.h }[6], [x25]\n"
4579 "ld1 { v17.h }[6], [x24]\n"
4580 "ld1 { v21.h }[6], [x23]\n"
4581 "ld1 { v25.h }[6], [x22]\n"
4582 "ld1 { v29.h }[6], [x21]\n"
4586 "tbz x8, #0, 264f\n"
4587 "ld1 { v9.h }[4], [x16]\n"
4588 "ld1 { v13.h }[4], [x25]\n"
4589 "ld1 { v17.h }[4], [x24]\n"
4590 "ld1 { v21.h }[4], [x23]\n"
4591 "ld1 { v25.h }[4], [x22]\n"
4592 "ld1 { v29.h }[4], [x21]\n"
4595 "tbz x8, #1, 259f\n"
4596 "ldr s9, [x16], #0x4\n"
4598 "ldr s13, [x25], #0x4\n"
4599 "ldr s17, [x24], #0x4\n"
4600 "ldr s21, [x23], #0x4\n"
4601 "ldr s25, [x22], #0x4\n"
4602 "ldr s29, [x21], #0x4\n"
4603 "tbz x8, #0, 264f\n"
4604 "ld1 { v9.h }[2], [x16]\n"
4605 "ld1 { v13.h }[2], [x25]\n"
4606 "ld1 { v17.h }[2], [x24]\n"
4607 "ld1 { v21.h }[2], [x23]\n"
4608 "ld1 { v25.h }[2], [x22]\n"
4609 "ld1 { v29.h }[2], [x21]\n"
4613 "tbz x8, #0, 264f\n"
4614 "ldr h9, [x16, #0x0]\n"
4615 "ldr h13, [x25, #0x0]\n"
4616 "ldr h17, [x24, #0x0]\n"
4617 "ldr h21, [x23, #0x0]\n"
4618 "ldr h25, [x22, #0x0]\n"
4619 "ldr h29, [x21, #0x0]\n"
4622 "tbz x8, #2, 262f\n"
4623 "ldr d8, [x16], #0x8\n"
4624 "ldr d12, [x25], #0x8\n"
4625 "ldr d16, [x24], #0x8\n"
4626 "ldr d20, [x23], #0x8\n"
4627 "ldr d24, [x22], #0x8\n"
4628 "ldr d28, [x21], #0x8\n"
4629 "tbz x8, #1, 261f\n"
4630 "ld1 { v8.s }[2], [x16], #0x4\n"
4632 "ld1 { v12.s }[2], [x25], #0x4\n"
4633 "ld1 { v16.s }[2], [x24], #0x4\n"
4634 "ld1 { v20.s }[2], [x23], #0x4\n"
4635 "ld1 { v24.s }[2], [x22], #0x4\n"
4636 "ld1 { v28.s }[2], [x21], #0x4\n"
4637 "tbz x8, #0, 264f\n"
4638 "ld1 { v8.h }[6], [x16]\n"
4639 "ld1 { v12.h }[6], [x25]\n"
4640 "ld1 { v16.h }[6], [x24]\n"
4641 "ld1 { v20.h }[6], [x23]\n"
4642 "ld1 { v24.h }[6], [x22]\n"
4643 "ld1 { v28.h }[6], [x21]\n"
4647 "tbz x8, #0, 264f\n"
4648 "ld1 { v8.h }[4], [x16]\n"
4649 "ld1 { v12.h }[4], [x25]\n"
4650 "ld1 { v16.h }[4], [x24]\n"
4651 "ld1 { v20.h }[4], [x23]\n"
4652 "ld1 { v24.h }[4], [x22]\n"
4653 "ld1 { v28.h }[4], [x21]\n"
4656 "tbz x8, #1, 263f\n"
4657 "ldr s8, [x16], #0x4\n"
4659 "ldr s12, [x25], #0x4\n"
4660 "ldr s16, [x24], #0x4\n"
4661 "ldr s20, [x23], #0x4\n"
4662 "ldr s24, [x22], #0x4\n"
4663 "ldr s28, [x21], #0x4\n"
4664 "tbz x8, #0, 264f\n"
4665 "ld1 { v8.h }[2], [x16]\n"
4666 "ld1 { v12.h }[2], [x25]\n"
4667 "ld1 { v16.h }[2], [x24]\n"
4668 "ld1 { v20.h }[2], [x23]\n"
4669 "ld1 { v24.h }[2], [x22]\n"
4670 "ld1 { v28.h }[2], [x21]\n"
4673 "ldr h8, [x16, #0x0]\n"
4675 "ldr h12, [x25, #0x0]\n"
4676 "ldr h16, [x24, #0x0]\n"
4677 "ldr h20, [x23, #0x0]\n"
4678 "ldr h24, [x22, #0x0]\n"
4679 "ldr h28, [x21, #0x0]\n"
4681 "sub x16, x16, x20\n"
4684 "ldr q8, [x16, #0x0]\n"
4685 "ldr q9, [x16, #0x10]\n"
4686 "ldr q10, [x16, #0x20]\n"
4687 "ldr q11, [x16, #0x30]\n"
4688 "ldr q12, [x25, #0x0]\n"
4689 "ldr q13, [x25, #0x10]\n"
4690 "ldr q14, [x25, #0x20]\n"
4691 "ldr q15, [x25, #0x30]\n"
4692 "ldr q16, [x24, #0x0]\n"
4693 "ldr q17, [x24, #0x10]\n"
4694 "ldr q18, [x24, #0x20]\n"
4695 "ldr q19, [x24, #0x30]\n"
4696 "ldr q20, [x23, #0x0]\n"
4697 "ldr q21, [x23, #0x10]\n"
4698 "ldr q22, [x23, #0x20]\n"
4699 "ldr q23, [x23, #0x30]\n"
4700 "ldr q24, [x22, #0x0]\n"
4701 "ldr q25, [x22, #0x10]\n"
4702 "ldr q26, [x22, #0x20]\n"
4703 "ldr q27, [x22, #0x30]\n"
4704 "ldr q28, [x21, #0x0]\n"
4705 "ldr q29, [x21, #0x10]\n"
4706 "ldr q30, [x21, #0x20]\n"
4707 "ldr q31, [x21, #0x30]\n"
4710 "movi v8.16b, #0x0\n"
4711 "movi v9.16b, #0x0\n"
4712 "movi v10.16b, #0x0\n"
4713 "movi v11.16b, #0x0\n"
4714 "movi v12.16b, #0x0\n"
4715 "movi v13.16b, #0x0\n"
4716 "movi v14.16b, #0x0\n"
4717 "movi v15.16b, #0x0\n"
4718 "movi v16.16b, #0x0\n"
4719 "movi v17.16b, #0x0\n"
4720 "movi v18.16b, #0x0\n"
4721 "movi v19.16b, #0x0\n"
4722 "movi v20.16b, #0x0\n"
4723 "movi v21.16b, #0x0\n"
4724 "movi v22.16b, #0x0\n"
4725 "movi v23.16b, #0x0\n"
4726 "movi v24.16b, #0x0\n"
4727 "movi v25.16b, #0x0\n"
4728 "movi v26.16b, #0x0\n"
4729 "movi v27.16b, #0x0\n"
4730 "movi v28.16b, #0x0\n"
4731 "movi v29.16b, #0x0\n"
4732 "movi v30.16b, #0x0\n"
4733 "movi v31.16b, #0x0\n"
4737 "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
4738 "ldr w14, [x20, x15, LSL #0x2]\n"
4739 "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
4740 "tbz %x[flags], #3, 269f\n"
4741 "ldr x20, [%x[input_ptr], x15, LSL #0x3]\n"
4742 "add x20, x20, x21, LSL #3\n"
4743 "ldr x13, [x20, #0x0]\n"
4744 "ldr x12, [x20, #0x8]\n"
4745 "ldr x11, [x20, #0x10]\n"
4746 "ldr x10, [x20, #0x18]\n"
4747 "ldr x9, [x20, #0x20]\n"
4748 "ldr x28, [x20, #0x28]\n"
4750 "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
4751 "add x13, x13, x20, LSL #1\n"
4752 "add x12, x12, x20, LSL #1\n"
4753 "add x11, x11, x20, LSL #1\n"
4754 "add x10, x10, x20, LSL #1\n"
4755 "add x9, x9, x20, LSL #1\n"
4756 "add x28, x28, x20, LSL #1\n"
4759 "mov x13, %x[input_ptr]\n"
4760 "add x12, x13, x21, LSL #1\n"
4761 "add x11, x12, x21, LSL #1\n"
4762 "add x10, x11, x21, LSL #1\n"
4763 "add x9, x10, x21, LSL #1\n"
4764 "add x28, x9, x21, LSL #1\n"
4768 "ldr q0, [x13, #0x0]\n"
4770 "ldr q1, [x12, #0x0]\n"
4771 "ldr q2, [x11, #0x0]\n"
4772 "ldr q3, [x10, #0x0]\n"
4773 "ldr q4, [x9, #0x0]\n"
4774 "ldr q5, [x28, #0x0]\n"
4775 "ldr q6, [x17, #0x0]\n"
4776 "ldr q7, [x17, #0x10]\n"
4779 "fmla v8.8h, v6.8h, v0.h[0]\n"
4780 "ldr x21, [x17, #0x28]\n"
4781 "fmla v12.8h, v6.8h, v1.h[0]\n"
4782 "ldr x20, [x17, #0x38]\n"
4783 "fmla v16.8h, v6.8h, v2.h[0]\n"
4784 "add x13, x13, #0x10\n"
4785 "fmla v20.8h, v6.8h, v3.h[0]\n"
4786 "add x12, x12, #0x10\n"
4787 "fmla v24.8h, v6.8h, v4.h[0]\n"
4788 "add x11, x11, #0x10\n"
4789 "fmla v28.8h, v6.8h, v5.h[0]\n"
4790 "ldr d6, [x17, #0x20]\n"
4791 "fmla v9.8h, v7.8h, v0.h[0]\n"
4792 "mov v6.d[1], x21\n"
4793 "fmla v13.8h, v7.8h, v1.h[0]\n"
4794 "ldr x21, [x17, #0x48]\n"
4795 "fmla v17.8h, v7.8h, v2.h[0]\n"
4796 "add x10, x10, #0x10\n"
4797 "fmla v21.8h, v7.8h, v3.h[0]\n"
4798 "add x9, x9, #0x10\n"
4799 "fmla v25.8h, v7.8h, v4.h[0]\n"
4800 "add x28, x28, #0x10\n"
4801 "fmla v29.8h, v7.8h, v5.h[0]\n"
4802 "ldr d7, [x17, #0x30]\n"
4803 "mov v7.d[1], x20\n"
4804 "fmla v10.8h, v6.8h, v0.h[0]\n"
4805 "fmla v14.8h, v6.8h, v1.h[0]\n"
4806 "ldr x20, [x17, #0x58]\n"
4807 "fmla v18.8h, v6.8h, v2.h[0]\n"
4808 "ldr x27, [x13, #0x8]\n"
4809 "fmla v22.8h, v6.8h, v3.h[0]\n"
4810 "ldr x26, [x12, #0x8]\n"
4811 "fmla v26.8h, v6.8h, v4.h[0]\n"
4812 "ldr x25, [x11, #0x8]\n"
4813 "fmla v30.8h, v6.8h, v5.h[0]\n"
4814 "ldr d6, [x17, #0x40]\n"
4815 "fmla v11.8h, v7.8h, v0.h[0]\n"
4816 "mov v6.d[1], x21\n"
4817 "fmla v15.8h, v7.8h, v1.h[0]\n"
4818 "ldr x21, [x17, #0x68]\n"
4819 "fmla v19.8h, v7.8h, v2.h[0]\n"
4820 "ldr x24, [x10, #0x8]\n"
4821 "fmla v23.8h, v7.8h, v3.h[0]\n"
4822 "ldr x23, [x9, #0x8]\n"
4823 "fmla v27.8h, v7.8h, v4.h[0]\n"
4824 "ldr x22, [x28, #0x8]\n"
4825 "fmla v31.8h, v7.8h, v5.h[0]\n"
4826 "ldr d7, [x17, #0x50]\n"
4827 "mov v7.d[1], x20\n"
4828 "fmla v8.8h, v6.8h, v0.h[1]\n"
4829 "fmla v12.8h, v6.8h, v1.h[1]\n"
4830 "ldr x20, [x17, #0x78]\n"
4831 "fmla v16.8h, v6.8h, v2.h[1]\n"
4832 "sub x14, x14, #0x8\n"
4833 "fmla v20.8h, v6.8h, v3.h[1]\n"
4835 "fmla v24.8h, v6.8h, v4.h[1]\n"
4836 "prfm pldl1keep, [x13, #0x80]\n"
4837 "fmla v28.8h, v6.8h, v5.h[1]\n"
4838 "ldr d6, [x17, #0x60]\n"
4839 "fmla v9.8h, v7.8h, v0.h[1]\n"
4840 "mov v6.d[1], x21\n"
4841 "fmla v13.8h, v7.8h, v1.h[1]\n"
4842 "ldr x21, [x17, #0x88]\n"
4843 "fmla v17.8h, v7.8h, v2.h[1]\n"
4844 "prfm pldl1keep, [x12, #0x80]\n"
4845 "fmla v21.8h, v7.8h, v3.h[1]\n"
4846 "prfm pldl1keep, [x11, #0x80]\n"
4847 "fmla v25.8h, v7.8h, v4.h[1]\n"
4848 "prfm pldl1keep, [x10, #0x80]\n"
4849 "fmla v29.8h, v7.8h, v5.h[1]\n"
4850 "ldr d7, [x17, #0x70]\n"
4851 "mov v7.d[1], x20\n"
4852 "fmla v10.8h, v6.8h, v0.h[1]\n"
4853 "fmla v14.8h, v6.8h, v1.h[1]\n"
4854 "ldr x20, [x17, #0x98]\n"
4855 "fmla v18.8h, v6.8h, v2.h[1]\n"
4856 "prfm pldl1keep, [x9, #0x80]\n"
4857 "fmla v22.8h, v6.8h, v3.h[1]\n"
4858 "prfm pldl1keep, [x28, #0x80]\n"
4859 "fmla v26.8h, v6.8h, v4.h[1]\n"
4860 "fmla v30.8h, v6.8h, v5.h[1]\n"
4861 "ldr d6, [x17, #0x80]\n"
4862 "fmla v11.8h, v7.8h, v0.h[1]\n"
4863 "mov v6.d[1], x21\n"
4864 "fmla v15.8h, v7.8h, v1.h[1]\n"
4865 "ldr x21, [x17, #0xa8]\n"
4866 "fmla v19.8h, v7.8h, v2.h[1]\n"
4867 "fmla v23.8h, v7.8h, v3.h[1]\n"
4868 "fmla v27.8h, v7.8h, v4.h[1]\n"
4869 "fmla v31.8h, v7.8h, v5.h[1]\n"
4870 "ldr d7, [x17, #0x90]\n"
4871 "mov v7.d[1], x20\n"
4872 "fmla v8.8h, v6.8h, v0.h[2]\n"
4873 "fmla v12.8h, v6.8h, v1.h[2]\n"
4874 "ldr x20, [x17, #0xb8]\n"
4875 "fmla v16.8h, v6.8h, v2.h[2]\n"
4876 "fmla v20.8h, v6.8h, v3.h[2]\n"
4877 "fmla v24.8h, v6.8h, v4.h[2]\n"
4878 "fmla v28.8h, v6.8h, v5.h[2]\n"
4879 "ldr d6, [x17, #0xa0]\n"
4880 "fmla v9.8h, v7.8h, v0.h[2]\n"
4881 "mov v6.d[1], x21\n"
4882 "fmla v13.8h, v7.8h, v1.h[2]\n"
4883 "ldr x21, [x17, #0xc8]\n"
4884 "fmla v17.8h, v7.8h, v2.h[2]\n"
4885 "fmla v21.8h, v7.8h, v3.h[2]\n"
4886 "fmla v25.8h, v7.8h, v4.h[2]\n"
4887 "fmla v29.8h, v7.8h, v5.h[2]\n"
4888 "ldr d7, [x17, #0xb0]\n"
4889 "mov v7.d[1], x20\n"
4890 "fmla v10.8h, v6.8h, v0.h[2]\n"
4891 "fmla v14.8h, v6.8h, v1.h[2]\n"
4892 "ldr x20, [x17, #0xd8]\n"
4893 "fmla v18.8h, v6.8h, v2.h[2]\n"
4894 "fmla v22.8h, v6.8h, v3.h[2]\n"
4895 "fmla v26.8h, v6.8h, v4.h[2]\n"
4896 "fmla v30.8h, v6.8h, v5.h[2]\n"
4897 "ldr d6, [x17, #0xc0]\n"
4898 "fmla v11.8h, v7.8h, v0.h[2]\n"
4899 "mov v6.d[1], x21\n"
4900 "fmla v15.8h, v7.8h, v1.h[2]\n"
4901 "ldr x21, [x17, #0xe8]\n"
4902 "fmla v19.8h, v7.8h, v2.h[2]\n"
4903 "fmla v23.8h, v7.8h, v3.h[2]\n"
4904 "fmla v27.8h, v7.8h, v4.h[2]\n"
4905 "fmla v31.8h, v7.8h, v5.h[2]\n"
4906 "ldr d7, [x17, #0xd0]\n"
4907 "mov v7.d[1], x20\n"
4908 "fmla v8.8h, v6.8h, v0.h[3]\n"
4909 "fmla v12.8h, v6.8h, v1.h[3]\n"
4910 "ldr x20, [x17, #0xf8]\n"
4911 "fmla v16.8h, v6.8h, v2.h[3]\n"
4912 "fmla v20.8h, v6.8h, v3.h[3]\n"
4913 "fmla v24.8h, v6.8h, v4.h[3]\n"
4914 "fmla v28.8h, v6.8h, v5.h[3]\n"
4915 "ldr d6, [x17, #0xe0]\n"
4916 "fmla v9.8h, v7.8h, v0.h[3]\n"
4917 "mov v6.d[1], x21\n"
4918 "fmla v13.8h, v7.8h, v1.h[3]\n"
4919 "ldr x21, [x17, #0x108]\n"
4920 "fmla v17.8h, v7.8h, v2.h[3]\n"
4921 "fmla v21.8h, v7.8h, v3.h[3]\n"
4922 "fmla v25.8h, v7.8h, v4.h[3]\n"
4923 "fmla v29.8h, v7.8h, v5.h[3]\n"
4924 "ldr d7, [x17, #0xf0]\n"
4925 "mov v7.d[1], x20\n"
4926 "fmla v10.8h, v6.8h, v0.h[3]\n"
4927 "fmla v14.8h, v6.8h, v1.h[3]\n"
4928 "ldr x20, [x17, #0x118]\n"
4929 "fmla v18.8h, v6.8h, v2.h[3]\n"
4930 "fmla v22.8h, v6.8h, v3.h[3]\n"
4931 "fmla v26.8h, v6.8h, v4.h[3]\n"
4932 "fmla v30.8h, v6.8h, v5.h[3]\n"
4933 "ldr d6, [x17, #0x100]\n"
4934 "fmla v11.8h, v7.8h, v0.h[3]\n"
4935 "mov v6.d[1], x21\n"
4936 "fmla v15.8h, v7.8h, v1.h[3]\n"
4937 "ldr x21, [x17, #0x128]\n"
4938 "fmla v19.8h, v7.8h, v2.h[3]\n"
4939 "fmla v23.8h, v7.8h, v3.h[3]\n"
4940 "fmla v27.8h, v7.8h, v4.h[3]\n"
4941 "fmla v31.8h, v7.8h, v5.h[3]\n"
4942 "ldr d7, [x17, #0x110]\n"
4943 "mov v7.d[1], x20\n"
4944 "fmla v8.8h, v6.8h, v0.h[4]\n"
4945 "fmla v12.8h, v6.8h, v1.h[4]\n"
4946 "ldr x20, [x17, #0x138]\n"
4947 "fmla v16.8h, v6.8h, v2.h[4]\n"
4948 "fmla v20.8h, v6.8h, v3.h[4]\n"
4949 "fmla v24.8h, v6.8h, v4.h[4]\n"
4950 "fmla v28.8h, v6.8h, v5.h[4]\n"
4951 "ldr d6, [x17, #0x120]\n"
4952 "fmla v9.8h, v7.8h, v0.h[4]\n"
4953 "mov v6.d[1], x21\n"
4954 "fmla v13.8h, v7.8h, v1.h[4]\n"
4955 "ldr x21, [x17, #0x148]\n"
4956 "fmla v17.8h, v7.8h, v2.h[4]\n"
4957 "fmla v21.8h, v7.8h, v3.h[4]\n"
4958 "fmla v25.8h, v7.8h, v4.h[4]\n"
4959 "fmla v29.8h, v7.8h, v5.h[4]\n"
4960 "ldr d7, [x17, #0x130]\n"
4961 "mov v7.d[1], x20\n"
4962 "fmla v10.8h, v6.8h, v0.h[4]\n"
4963 "fmla v14.8h, v6.8h, v1.h[4]\n"
4964 "ldr x20, [x17, #0x158]\n"
4965 "fmla v18.8h, v6.8h, v2.h[4]\n"
4966 "fmla v22.8h, v6.8h, v3.h[4]\n"
4967 "fmla v26.8h, v6.8h, v4.h[4]\n"
4968 "fmla v30.8h, v6.8h, v5.h[4]\n"
4969 "ldr d6, [x17, #0x140]\n"
4970 "fmla v11.8h, v7.8h, v0.h[4]\n"
4971 "mov v6.d[1], x21\n"
4972 "fmla v15.8h, v7.8h, v1.h[4]\n"
4973 "ldr x21, [x17, #0x168]\n"
4974 "fmla v19.8h, v7.8h, v2.h[4]\n"
4975 "fmla v23.8h, v7.8h, v3.h[4]\n"
4976 "fmla v27.8h, v7.8h, v4.h[4]\n"
4977 "fmla v31.8h, v7.8h, v5.h[4]\n"
4978 "ldr d7, [x17, #0x150]\n"
4979 "mov v7.d[1], x20\n"
4980 "fmla v8.8h, v6.8h, v0.h[5]\n"
4981 "fmla v12.8h, v6.8h, v1.h[5]\n"
4982 "ldr x20, [x17, #0x178]\n"
4983 "fmla v16.8h, v6.8h, v2.h[5]\n"
4984 "fmla v20.8h, v6.8h, v3.h[5]\n"
4985 "fmla v24.8h, v6.8h, v4.h[5]\n"
4986 "fmla v28.8h, v6.8h, v5.h[5]\n"
4987 "ldr d6, [x17, #0x160]\n"
4988 "fmla v9.8h, v7.8h, v0.h[5]\n"
4989 "mov v6.d[1], x21\n"
4990 "fmla v13.8h, v7.8h, v1.h[5]\n"
4991 "ldr x21, [x17, #0x188]\n"
4992 "fmla v17.8h, v7.8h, v2.h[5]\n"
4993 "fmla v21.8h, v7.8h, v3.h[5]\n"
4994 "fmla v25.8h, v7.8h, v4.h[5]\n"
4995 "fmla v29.8h, v7.8h, v5.h[5]\n"
4996 "ldr d7, [x17, #0x170]\n"
4997 "mov v7.d[1], x20\n"
4998 "fmla v10.8h, v6.8h, v0.h[5]\n"
4999 "fmla v14.8h, v6.8h, v1.h[5]\n"
5000 "ldr x20, [x17, #0x198]\n"
5001 "fmla v18.8h, v6.8h, v2.h[5]\n"
5002 "fmla v22.8h, v6.8h, v3.h[5]\n"
5003 "fmla v26.8h, v6.8h, v4.h[5]\n"
5004 "fmla v30.8h, v6.8h, v5.h[5]\n"
5005 "ldr d6, [x17, #0x180]\n"
5006 "fmla v11.8h, v7.8h, v0.h[5]\n"
5007 "mov v6.d[1], x21\n"
5008 "fmla v15.8h, v7.8h, v1.h[5]\n"
5009 "ldr x21, [x17, #0x1a8]\n"
5010 "fmla v19.8h, v7.8h, v2.h[5]\n"
5011 "fmla v23.8h, v7.8h, v3.h[5]\n"
5012 "fmla v27.8h, v7.8h, v4.h[5]\n"
5013 "fmla v31.8h, v7.8h, v5.h[5]\n"
5014 "ldr d7, [x17, #0x190]\n"
5015 "mov v7.d[1], x20\n"
5016 "fmla v8.8h, v6.8h, v0.h[6]\n"
5017 "fmla v12.8h, v6.8h, v1.h[6]\n"
5018 "ldr x20, [x17, #0x1b8]\n"
5019 "fmla v16.8h, v6.8h, v2.h[6]\n"
5020 "fmla v20.8h, v6.8h, v3.h[6]\n"
5021 "fmla v24.8h, v6.8h, v4.h[6]\n"
5022 "fmla v28.8h, v6.8h, v5.h[6]\n"
5023 "ldr d6, [x17, #0x1a0]\n"
5024 "fmla v9.8h, v7.8h, v0.h[6]\n"
5025 "mov v6.d[1], x21\n"
5026 "fmla v13.8h, v7.8h, v1.h[6]\n"
5027 "ldr x21, [x17, #0x1c8]\n"
5028 "fmla v17.8h, v7.8h, v2.h[6]\n"
5029 "fmla v21.8h, v7.8h, v3.h[6]\n"
5030 "fmla v25.8h, v7.8h, v4.h[6]\n"
5031 "fmla v29.8h, v7.8h, v5.h[6]\n"
5032 "ldr d7, [x17, #0x1b0]\n"
5033 "mov v7.d[1], x20\n"
5034 "fmla v10.8h, v6.8h, v0.h[6]\n"
5035 "fmla v14.8h, v6.8h, v1.h[6]\n"
5036 "ldr x20, [x17, #0x1d8]\n"
5037 "fmla v18.8h, v6.8h, v2.h[6]\n"
5038 "fmla v22.8h, v6.8h, v3.h[6]\n"
5039 "fmla v26.8h, v6.8h, v4.h[6]\n"
5040 "fmla v30.8h, v6.8h, v5.h[6]\n"
5041 "ldr d6, [x17, #0x1c0]\n"
5042 "fmla v11.8h, v7.8h, v0.h[6]\n"
5043 "mov v6.d[1], x21\n"
5044 "fmla v15.8h, v7.8h, v1.h[6]\n"
5045 "ldr x21, [x17, #0x1e8]\n"
5046 "fmla v19.8h, v7.8h, v2.h[6]\n"
5047 "fmla v23.8h, v7.8h, v3.h[6]\n"
5048 "fmla v27.8h, v7.8h, v4.h[6]\n"
5049 "fmla v31.8h, v7.8h, v5.h[6]\n"
5050 "ldr d7, [x17, #0x1d0]\n"
5051 "mov v7.d[1], x20\n"
5052 "fmla v8.8h, v6.8h, v0.h[7]\n"
5053 "fmla v12.8h, v6.8h, v1.h[7]\n"
5054 "ldr x20, [x17, #0x1f8]\n"
5055 "fmla v16.8h, v6.8h, v2.h[7]\n"
5056 "fmla v20.8h, v6.8h, v3.h[7]\n"
5057 "fmla v24.8h, v6.8h, v4.h[7]\n"
5058 "fmla v28.8h, v6.8h, v5.h[7]\n"
5059 "ldr d6, [x17, #0x1e0]\n"
5060 "fmla v9.8h, v7.8h, v0.h[7]\n"
5061 "mov v6.d[1], x21\n"
5062 "fmla v13.8h, v7.8h, v1.h[7]\n"
5063 "fmla v17.8h, v7.8h, v2.h[7]\n"
5064 "fmla v21.8h, v7.8h, v3.h[7]\n"
5065 "fmla v25.8h, v7.8h, v4.h[7]\n"
5066 "fmla v29.8h, v7.8h, v5.h[7]\n"
5067 "ldr d7, [x17, #0x1f0]\n"
5068 "mov v7.d[1], x20\n"
5069 "add x17, x17, #0x200\n"
5070 "fmla v10.8h, v6.8h, v0.h[7]\n"
5071 "ldr x21, [x17, #0x8]\n"
5072 "fmla v14.8h, v6.8h, v1.h[7]\n"
5073 "ldr x20, [x17, #0x18]\n"
5074 "fmla v18.8h, v6.8h, v2.h[7]\n"
5075 "fmla v22.8h, v6.8h, v3.h[7]\n"
5076 "fmla v26.8h, v6.8h, v4.h[7]\n"
5077 "fmla v30.8h, v6.8h, v5.h[7]\n"
5078 "ldr d6, [x17, #0x0]\n"
5079 "fmla v11.8h, v7.8h, v0.h[7]\n"
5080 "ldr d0, [x13, #0x0]\n"
5081 "fmla v15.8h, v7.8h, v1.h[7]\n"
5082 "ldr d1, [x12, #0x0]\n"
5083 "fmla v19.8h, v7.8h, v2.h[7]\n"
5084 "ldr d2, [x11, #0x0]\n"
5085 "fmla v23.8h, v7.8h, v3.h[7]\n"
5086 "ldr d3, [x10, #0x0]\n"
5087 "fmla v27.8h, v7.8h, v4.h[7]\n"
5088 "ldr d4, [x9, #0x0]\n"
5089 "fmla v31.8h, v7.8h, v5.h[7]\n"
5090 "ldr d5, [x28, #0x0]\n"
5091 "ldr d7, [x17, #0x10]\n"
5092 "mov v6.d[1], x21\n"
5093 "mov v0.d[1], x27\n"
5094 "mov v1.d[1], x26\n"
5095 "mov v2.d[1], x25\n"
5096 "mov v3.d[1], x24\n"
5097 "mov v4.d[1], x23\n"
5098 "mov v5.d[1], x22\n"
5099 "mov v7.d[1], x20\n"
5102 "fmla v8.8h, v6.8h, v0.h[0]\n"
5103 "add x13, x13, #0x10\n"
5104 "fmla v12.8h, v6.8h, v1.h[0]\n"
5105 "add x12, x12, #0x10\n"
5106 "fmla v16.8h, v6.8h, v2.h[0]\n"
5107 "add x11, x11, #0x10\n"
5108 "fmla v20.8h, v6.8h, v3.h[0]\n"
5109 "add x10, x10, #0x10\n"
5110 "fmla v24.8h, v6.8h, v4.h[0]\n"
5111 "add x9, x9, #0x10\n"
5112 "fmla v28.8h, v6.8h, v5.h[0]\n"
5113 "ldr q6, [x17, #0x20]\n"
5114 "fmla v9.8h, v7.8h, v0.h[0]\n"
5115 "add x28, x28, #0x10\n"
5116 "fmla v13.8h, v7.8h, v1.h[0]\n"
5117 "sub x14, x14, #0x8\n"
5118 "fmla v17.8h, v7.8h, v2.h[0]\n"
5119 "prfm pldl1keep, [x13, #0x80]\n"
5120 "fmla v21.8h, v7.8h, v3.h[0]\n"
5121 "prfm pldl1keep, [x12, #0x80]\n"
5122 "fmla v25.8h, v7.8h, v4.h[0]\n"
5123 "prfm pldl1keep, [x11, #0x80]\n"
5124 "fmla v29.8h, v7.8h, v5.h[0]\n"
5125 "ldr q7, [x17, #0x30]\n"
5126 "fmla v10.8h, v6.8h, v0.h[0]\n"
5127 "prfm pldl1keep, [x10, #0x80]\n"
5128 "fmla v14.8h, v6.8h, v1.h[0]\n"
5129 "prfm pldl1keep, [x9, #0x80]\n"
5130 "fmla v18.8h, v6.8h, v2.h[0]\n"
5131 "prfm pldl1keep, [x28, #0x80]\n"
5132 "fmla v22.8h, v6.8h, v3.h[0]\n"
5133 "fmla v26.8h, v6.8h, v4.h[0]\n"
5134 "fmla v30.8h, v6.8h, v5.h[0]\n"
5135 "ldr q6, [x17, #0x40]\n"
5136 "fmla v11.8h, v7.8h, v0.h[0]\n"
5137 "fmla v15.8h, v7.8h, v1.h[0]\n"
5138 "fmla v19.8h, v7.8h, v2.h[0]\n"
5139 "fmla v23.8h, v7.8h, v3.h[0]\n"
5140 "fmla v27.8h, v7.8h, v4.h[0]\n"
5141 "fmla v31.8h, v7.8h, v5.h[0]\n"
5142 "ldr q7, [x17, #0x50]\n"
5143 "fmla v8.8h, v6.8h, v0.h[1]\n"
5144 "fmla v12.8h, v6.8h, v1.h[1]\n"
5145 "fmla v16.8h, v6.8h, v2.h[1]\n"
5146 "fmla v20.8h, v6.8h, v3.h[1]\n"
5147 "fmla v24.8h, v6.8h, v4.h[1]\n"
5148 "fmla v28.8h, v6.8h, v5.h[1]\n"
5149 "ldr q6, [x17, #0x60]\n"
5150 "fmla v9.8h, v7.8h, v0.h[1]\n"
5151 "fmla v13.8h, v7.8h, v1.h[1]\n"
5152 "fmla v17.8h, v7.8h, v2.h[1]\n"
5153 "fmla v21.8h, v7.8h, v3.h[1]\n"
5154 "fmla v25.8h, v7.8h, v4.h[1]\n"
5155 "fmla v29.8h, v7.8h, v5.h[1]\n"
5156 "ldr q7, [x17, #0x70]\n"
5157 "fmla v10.8h, v6.8h, v0.h[1]\n"
5158 "fmla v14.8h, v6.8h, v1.h[1]\n"
5159 "fmla v18.8h, v6.8h, v2.h[1]\n"
5160 "fmla v22.8h, v6.8h, v3.h[1]\n"
5161 "fmla v26.8h, v6.8h, v4.h[1]\n"
5162 "fmla v30.8h, v6.8h, v5.h[1]\n"
5163 "ldr q6, [x17, #0x80]\n"
5164 "fmla v11.8h, v7.8h, v0.h[1]\n"
5165 "fmla v15.8h, v7.8h, v1.h[1]\n"
5166 "fmla v19.8h, v7.8h, v2.h[1]\n"
5167 "fmla v23.8h, v7.8h, v3.h[1]\n"
5168 "fmla v27.8h, v7.8h, v4.h[1]\n"
5169 "fmla v31.8h, v7.8h, v5.h[1]\n"
5170 "ldr q7, [x17, #0x90]\n"
5171 "fmla v8.8h, v6.8h, v0.h[2]\n"
5172 "fmla v12.8h, v6.8h, v1.h[2]\n"
5173 "fmla v16.8h, v6.8h, v2.h[2]\n"
5174 "fmla v20.8h, v6.8h, v3.h[2]\n"
5175 "fmla v24.8h, v6.8h, v4.h[2]\n"
5176 "fmla v28.8h, v6.8h, v5.h[2]\n"
5177 "ldr q6, [x17, #0xa0]\n"
5178 "fmla v9.8h, v7.8h, v0.h[2]\n"
5179 "fmla v13.8h, v7.8h, v1.h[2]\n"
5180 "fmla v17.8h, v7.8h, v2.h[2]\n"
5181 "fmla v21.8h, v7.8h, v3.h[2]\n"
5182 "fmla v25.8h, v7.8h, v4.h[2]\n"
5183 "fmla v29.8h, v7.8h, v5.h[2]\n"
5184 "ldr q7, [x17, #0xb0]\n"
5185 "fmla v10.8h, v6.8h, v0.h[2]\n"
5186 "fmla v14.8h, v6.8h, v1.h[2]\n"
5187 "fmla v18.8h, v6.8h, v2.h[2]\n"
5188 "fmla v22.8h, v6.8h, v3.h[2]\n"
5189 "fmla v26.8h, v6.8h, v4.h[2]\n"
5190 "fmla v30.8h, v6.8h, v5.h[2]\n"
5191 "ldr q6, [x17, #0xc0]\n"
5192 "fmla v11.8h, v7.8h, v0.h[2]\n"
5193 "fmla v15.8h, v7.8h, v1.h[2]\n"
5194 "fmla v19.8h, v7.8h, v2.h[2]\n"
5195 "fmla v23.8h, v7.8h, v3.h[2]\n"
5196 "fmla v27.8h, v7.8h, v4.h[2]\n"
5197 "fmla v31.8h, v7.8h, v5.h[2]\n"
5198 "ldr q7, [x17, #0xd0]\n"
5199 "fmla v8.8h, v6.8h, v0.h[3]\n"
5200 "fmla v12.8h, v6.8h, v1.h[3]\n"
5201 "fmla v16.8h, v6.8h, v2.h[3]\n"
5202 "fmla v20.8h, v6.8h, v3.h[3]\n"
5203 "fmla v24.8h, v6.8h, v4.h[3]\n"
5204 "fmla v28.8h, v6.8h, v5.h[3]\n"
5205 "ldr q6, [x17, #0xe0]\n"
5206 "fmla v9.8h, v7.8h, v0.h[3]\n"
5207 "fmla v13.8h, v7.8h, v1.h[3]\n"
5208 "fmla v17.8h, v7.8h, v2.h[3]\n"
5209 "fmla v21.8h, v7.8h, v3.h[3]\n"
5210 "fmla v25.8h, v7.8h, v4.h[3]\n"
5211 "fmla v29.8h, v7.8h, v5.h[3]\n"
5212 "ldr q7, [x17, #0xf0]\n"
5213 "fmla v10.8h, v6.8h, v0.h[3]\n"
5214 "fmla v14.8h, v6.8h, v1.h[3]\n"
5215 "fmla v18.8h, v6.8h, v2.h[3]\n"
5216 "fmla v22.8h, v6.8h, v3.h[3]\n"
5217 "fmla v26.8h, v6.8h, v4.h[3]\n"
5218 "fmla v30.8h, v6.8h, v5.h[3]\n"
5219 "ldr q6, [x17, #0x100]\n"
5220 "fmla v11.8h, v7.8h, v0.h[3]\n"
5221 "fmla v15.8h, v7.8h, v1.h[3]\n"
5222 "fmla v19.8h, v7.8h, v2.h[3]\n"
5223 "fmla v23.8h, v7.8h, v3.h[3]\n"
5224 "fmla v27.8h, v7.8h, v4.h[3]\n"
5225 "fmla v31.8h, v7.8h, v5.h[3]\n"
5226 "ldr q7, [x17, #0x110]\n"
5227 "fmla v8.8h, v6.8h, v0.h[4]\n"
5228 "fmla v12.8h, v6.8h, v1.h[4]\n"
5229 "fmla v16.8h, v6.8h, v2.h[4]\n"
5230 "fmla v20.8h, v6.8h, v3.h[4]\n"
5231 "fmla v24.8h, v6.8h, v4.h[4]\n"
5232 "fmla v28.8h, v6.8h, v5.h[4]\n"
5233 "ldr q6, [x17, #0x120]\n"
5234 "fmla v9.8h, v7.8h, v0.h[4]\n"
5235 "fmla v13.8h, v7.8h, v1.h[4]\n"
5236 "fmla v17.8h, v7.8h, v2.h[4]\n"
5237 "fmla v21.8h, v7.8h, v3.h[4]\n"
5238 "fmla v25.8h, v7.8h, v4.h[4]\n"
5239 "fmla v29.8h, v7.8h, v5.h[4]\n"
5240 "ldr q7, [x17, #0x130]\n"
5241 "fmla v10.8h, v6.8h, v0.h[4]\n"
5242 "fmla v14.8h, v6.8h, v1.h[4]\n"
5243 "fmla v18.8h, v6.8h, v2.h[4]\n"
5244 "fmla v22.8h, v6.8h, v3.h[4]\n"
5245 "fmla v26.8h, v6.8h, v4.h[4]\n"
5246 "fmla v30.8h, v6.8h, v5.h[4]\n"
5247 "ldr q6, [x17, #0x140]\n"
5248 "fmla v11.8h, v7.8h, v0.h[4]\n"
5249 "fmla v15.8h, v7.8h, v1.h[4]\n"
5250 "fmla v19.8h, v7.8h, v2.h[4]\n"
5251 "fmla v23.8h, v7.8h, v3.h[4]\n"
5252 "fmla v27.8h, v7.8h, v4.h[4]\n"
5253 "fmla v31.8h, v7.8h, v5.h[4]\n"
5254 "ldr q7, [x17, #0x150]\n"
5255 "fmla v8.8h, v6.8h, v0.h[5]\n"
5256 "fmla v12.8h, v6.8h, v1.h[5]\n"
5257 "fmla v16.8h, v6.8h, v2.h[5]\n"
5258 "fmla v20.8h, v6.8h, v3.h[5]\n"
5259 "fmla v24.8h, v6.8h, v4.h[5]\n"
5260 "fmla v28.8h, v6.8h, v5.h[5]\n"
5261 "ldr q6, [x17, #0x160]\n"
5262 "fmla v9.8h, v7.8h, v0.h[5]\n"
5263 "fmla v13.8h, v7.8h, v1.h[5]\n"
5264 "fmla v17.8h, v7.8h, v2.h[5]\n"
5265 "fmla v21.8h, v7.8h, v3.h[5]\n"
5266 "fmla v25.8h, v7.8h, v4.h[5]\n"
5267 "fmla v29.8h, v7.8h, v5.h[5]\n"
5268 "ldr q7, [x17, #0x170]\n"
5269 "fmla v10.8h, v6.8h, v0.h[5]\n"
5270 "fmla v14.8h, v6.8h, v1.h[5]\n"
5271 "fmla v18.8h, v6.8h, v2.h[5]\n"
5272 "fmla v22.8h, v6.8h, v3.h[5]\n"
5273 "fmla v26.8h, v6.8h, v4.h[5]\n"
5274 "fmla v30.8h, v6.8h, v5.h[5]\n"
5275 "ldr q6, [x17, #0x180]\n"
5276 "fmla v11.8h, v7.8h, v0.h[5]\n"
5277 "fmla v15.8h, v7.8h, v1.h[5]\n"
5278 "fmla v19.8h, v7.8h, v2.h[5]\n"
5279 "fmla v23.8h, v7.8h, v3.h[5]\n"
5280 "fmla v27.8h, v7.8h, v4.h[5]\n"
5281 "fmla v31.8h, v7.8h, v5.h[5]\n"
5282 "ldr q7, [x17, #0x190]\n"
5283 "fmla v8.8h, v6.8h, v0.h[6]\n"
5284 "fmla v12.8h, v6.8h, v1.h[6]\n"
5285 "fmla v16.8h, v6.8h, v2.h[6]\n"
5286 "fmla v20.8h, v6.8h, v3.h[6]\n"
5287 "fmla v24.8h, v6.8h, v4.h[6]\n"
5288 "fmla v28.8h, v6.8h, v5.h[6]\n"
5289 "ldr q6, [x17, #0x1a0]\n"
5290 "fmla v9.8h, v7.8h, v0.h[6]\n"
5291 "fmla v13.8h, v7.8h, v1.h[6]\n"
5292 "fmla v17.8h, v7.8h, v2.h[6]\n"
5293 "fmla v21.8h, v7.8h, v3.h[6]\n"
5294 "fmla v25.8h, v7.8h, v4.h[6]\n"
5295 "fmla v29.8h, v7.8h, v5.h[6]\n"
5296 "ldr q7, [x17, #0x1b0]\n"
5297 "fmla v10.8h, v6.8h, v0.h[6]\n"
5298 "fmla v14.8h, v6.8h, v1.h[6]\n"
5299 "fmla v18.8h, v6.8h, v2.h[6]\n"
5300 "fmla v22.8h, v6.8h, v3.h[6]\n"
5301 "fmla v26.8h, v6.8h, v4.h[6]\n"
5302 "fmla v30.8h, v6.8h, v5.h[6]\n"
5303 "ldr q6, [x17, #0x1c0]\n"
5304 "fmla v11.8h, v7.8h, v0.h[6]\n"
5305 "fmla v15.8h, v7.8h, v1.h[6]\n"
5306 "fmla v19.8h, v7.8h, v2.h[6]\n"
5307 "fmla v23.8h, v7.8h, v3.h[6]\n"
5308 "fmla v27.8h, v7.8h, v4.h[6]\n"
5309 "fmla v31.8h, v7.8h, v5.h[6]\n"
5310 "ldr q7, [x17, #0x1d0]\n"
5311 "fmla v8.8h, v6.8h, v0.h[7]\n"
5312 "fmla v12.8h, v6.8h, v1.h[7]\n"
5313 "fmla v16.8h, v6.8h, v2.h[7]\n"
5314 "fmla v20.8h, v6.8h, v3.h[7]\n"
5315 "fmla v24.8h, v6.8h, v4.h[7]\n"
5316 "fmla v28.8h, v6.8h, v5.h[7]\n"
5317 "ldr q6, [x17, #0x1e0]\n"
5318 "fmla v9.8h, v7.8h, v0.h[7]\n"
5319 "fmla v13.8h, v7.8h, v1.h[7]\n"
5320 "fmla v17.8h, v7.8h, v2.h[7]\n"
5321 "fmla v21.8h, v7.8h, v3.h[7]\n"
5322 "fmla v25.8h, v7.8h, v4.h[7]\n"
5323 "fmla v29.8h, v7.8h, v5.h[7]\n"
5324 "ldr q7, [x17, #0x1f0]\n"
5325 "fmla v10.8h, v6.8h, v0.h[7]\n"
5326 "add x17, x17, #0x200\n"
5327 "fmla v14.8h, v6.8h, v1.h[7]\n"
5328 "fmla v18.8h, v6.8h, v2.h[7]\n"
5329 "fmla v22.8h, v6.8h, v3.h[7]\n"
5330 "fmla v26.8h, v6.8h, v4.h[7]\n"
5331 "fmla v30.8h, v6.8h, v5.h[7]\n"
5332 "fmla v11.8h, v7.8h, v0.h[7]\n"
5333 "fmla v15.8h, v7.8h, v1.h[7]\n"
5334 "fmla v19.8h, v7.8h, v2.h[7]\n"
5335 "fmla v23.8h, v7.8h, v3.h[7]\n"
5336 "fmla v27.8h, v7.8h, v4.h[7]\n"
5337 "fmla v31.8h, v7.8h, v5.h[7]\n"
5341 "ldr h7, [x13], #0x2\n"
5342 "sub x14, x14, #0x1\n"
5343 "ldr h6, [x12], #0x2\n"
5344 "ldr h5, [x11], #0x2\n"
5345 "ldr h4, [x10], #0x2\n"
5346 "ldr h3, [x9], #0x2\n"
5347 "ldr h2, [x28], #0x2\n"
5348 "ldr q1, [x17, #0x0]\n"
5349 "fmla v8.8h, v1.8h, v7.h[0]\n"
5350 "ldr q0, [x17, #0x10]\n"
5351 "fmla v12.8h, v1.8h, v6.h[0]\n"
5352 "fmla v16.8h, v1.8h, v5.h[0]\n"
5353 "fmla v20.8h, v1.8h, v4.h[0]\n"
5354 "fmla v24.8h, v1.8h, v3.h[0]\n"
5355 "fmla v28.8h, v1.8h, v2.h[0]\n"
5356 "ldr q1, [x17, #0x20]\n"
5357 "fmla v9.8h, v0.8h, v7.h[0]\n"
5358 "fmla v13.8h, v0.8h, v6.h[0]\n"
5359 "fmla v17.8h, v0.8h, v5.h[0]\n"
5360 "fmla v21.8h, v0.8h, v4.h[0]\n"
5361 "fmla v25.8h, v0.8h, v3.h[0]\n"
5362 "fmla v29.8h, v0.8h, v2.h[0]\n"
5363 "ldr q0, [x17, #0x30]\n"
5364 "fmla v10.8h, v1.8h, v7.h[0]\n"
5365 "add x17, x17, #0x40\n"
5366 "fmla v14.8h, v1.8h, v6.h[0]\n"
5367 "fmla v18.8h, v1.8h, v5.h[0]\n"
5368 "fmla v22.8h, v1.8h, v4.h[0]\n"
5369 "fmla v26.8h, v1.8h, v3.h[0]\n"
5370 "fmla v30.8h, v1.8h, v2.h[0]\n"
5371 "fmla v11.8h, v0.8h, v7.h[0]\n"
5372 "fmla v15.8h, v0.8h, v6.h[0]\n"
5373 "fmla v19.8h, v0.8h, v5.h[0]\n"
5374 "fmla v23.8h, v0.8h, v4.h[0]\n"
5375 "fmla v27.8h, v0.8h, v3.h[0]\n"
5376 "fmla v31.8h, v0.8h, v2.h[0]\n"
5379 "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
5380 "add x15, x15, #0x1\n"
5383 "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
5384 "add x25, x16, x20, LSL #1\n"
5385 "add x24, x25, x20, LSL #1\n"
5386 "add x23, x24, x20, LSL #1\n"
5387 "add x22, x23, x20, LSL #1\n"
5388 "add x21, x22, x20, LSL #1\n"
5389 "prfm pstl1keep, [x16, #0x0]\n"
5390 "prfm pstl1keep, [x25, #0x0]\n"
5391 "prfm pstl1keep, [x24, #0x0]\n"
5392 "prfm pstl1keep, [x23, #0x0]\n"
5393 "prfm pstl1keep, [x22, #0x0]\n"
5394 "prfm pstl1keep, [x21, #0x0]\n"
5395 "tbz %x[flags], #1, 276f\n"
5396 "add x20, %x[args_ptr], %[offset_max]\n"
5397 "ld1r { v0.8h }, [x20]\n"
5398 "fmin v8.8h, v8.8h, v0.8h\n"
5399 "fmin v9.8h, v9.8h, v0.8h\n"
5400 "fmin v10.8h, v10.8h, v0.8h\n"
5401 "fmin v11.8h, v11.8h, v0.8h\n"
5402 "fmin v12.8h, v12.8h, v0.8h\n"
5403 "fmin v13.8h, v13.8h, v0.8h\n"
5404 "fmin v14.8h, v14.8h, v0.8h\n"
5405 "fmin v15.8h, v15.8h, v0.8h\n"
5406 "fmin v16.8h, v16.8h, v0.8h\n"
5407 "fmin v17.8h, v17.8h, v0.8h\n"
5408 "fmin v18.8h, v18.8h, v0.8h\n"
5409 "fmin v19.8h, v19.8h, v0.8h\n"
5410 "fmin v20.8h, v20.8h, v0.8h\n"
5411 "fmin v21.8h, v21.8h, v0.8h\n"
5412 "fmin v22.8h, v22.8h, v0.8h\n"
5413 "fmin v23.8h, v23.8h, v0.8h\n"
5414 "fmin v24.8h, v24.8h, v0.8h\n"
5415 "fmin v25.8h, v25.8h, v0.8h\n"
5416 "fmin v26.8h, v26.8h, v0.8h\n"
5417 "fmin v27.8h, v27.8h, v0.8h\n"
5418 "fmin v28.8h, v28.8h, v0.8h\n"
5419 "fmin v29.8h, v29.8h, v0.8h\n"
5420 "fmin v30.8h, v30.8h, v0.8h\n"
5421 "fmin v31.8h, v31.8h, v0.8h\n"
5422 "add x20, %x[args_ptr], %[offset_min]\n"
5423 "ld1r { v0.8h }, [x20]\n"
5424 "fmax v8.8h, v8.8h, v0.8h\n"
5425 "fmax v9.8h, v9.8h, v0.8h\n"
5426 "fmax v10.8h, v10.8h, v0.8h\n"
5427 "fmax v11.8h, v11.8h, v0.8h\n"
5428 "fmax v12.8h, v12.8h, v0.8h\n"
5429 "fmax v13.8h, v13.8h, v0.8h\n"
5430 "fmax v14.8h, v14.8h, v0.8h\n"
5431 "fmax v15.8h, v15.8h, v0.8h\n"
5432 "fmax v16.8h, v16.8h, v0.8h\n"
5433 "fmax v17.8h, v17.8h, v0.8h\n"
5434 "fmax v18.8h, v18.8h, v0.8h\n"
5435 "fmax v19.8h, v19.8h, v0.8h\n"
5436 "fmax v20.8h, v20.8h, v0.8h\n"
5437 "fmax v21.8h, v21.8h, v0.8h\n"
5438 "fmax v22.8h, v22.8h, v0.8h\n"
5439 "fmax v23.8h, v23.8h, v0.8h\n"
5440 "fmax v24.8h, v24.8h, v0.8h\n"
5441 "fmax v25.8h, v25.8h, v0.8h\n"
5442 "fmax v26.8h, v26.8h, v0.8h\n"
5443 "fmax v27.8h, v27.8h, v0.8h\n"
5444 "fmax v28.8h, v28.8h, v0.8h\n"
5445 "fmax v29.8h, v29.8h, v0.8h\n"
5446 "fmax v30.8h, v30.8h, v0.8h\n"
5447 "fmax v31.8h, v31.8h, v0.8h\n"
5451 "tbz x8, #4, 284f\n"
5452 "st1 { v8.8h }, [x16], #0x10\n"
5453 "st1 { v9.8h }, [x16], #0x10\n"
5454 "st1 { v12.8h }, [x25], #0x10\n"
5455 "st1 { v13.8h }, [x25], #0x10\n"
5456 "st1 { v16.8h }, [x24], #0x10\n"
5457 "st1 { v17.8h }, [x24], #0x10\n"
5458 "st1 { v20.8h }, [x23], #0x10\n"
5459 "st1 { v21.8h }, [x23], #0x10\n"
5460 "st1 { v24.8h }, [x22], #0x10\n"
5461 "st1 { v25.8h }, [x22], #0x10\n"
5462 "st1 { v28.8h }, [x21], #0x10\n"
5463 "st1 { v29.8h }, [x21], #0x10\n"
5464 "tbz x8, #3, 280f\n"
5465 "st1 { v10.8h }, [x16], #0x10\n"
5466 "st1 { v14.8h }, [x25], #0x10\n"
5467 "st1 { v18.8h }, [x24], #0x10\n"
5468 "st1 { v22.8h }, [x23], #0x10\n"
5469 "st1 { v26.8h }, [x22], #0x10\n"
5470 "st1 { v30.8h }, [x21], #0x10\n"
5471 "tbz x8, #2, 278f\n"
5472 "str d11, [x16], #0x8\n"
5473 "str d15, [x25], #0x8\n"
5474 "str d19, [x24], #0x8\n"
5475 "str d23, [x23], #0x8\n"
5476 "str d27, [x22], #0x8\n"
5477 "str d31, [x21], #0x8\n"
5478 "tbz x8, #1, 277f\n"
5479 "st1 { v11.s }[2], [x16], #0x4\n"
5480 "st1 { v15.s }[2], [x25], #0x4\n"
5481 "st1 { v19.s }[2], [x24], #0x4\n"
5482 "st1 { v23.s }[2], [x23], #0x4\n"
5483 "st1 { v27.s }[2], [x22], #0x4\n"
5484 "st1 { v31.s }[2], [x21], #0x4\n"
5485 "tbz x8, #0, 292f\n"
5486 "st1 { v11.h }[6], [x16]\n"
5487 "st1 { v15.h }[6], [x25]\n"
5488 "st1 { v19.h }[6], [x24]\n"
5489 "st1 { v23.h }[6], [x23]\n"
5490 "st1 { v27.h }[6], [x22]\n"
5491 "st1 { v31.h }[6], [x21]\n"
5494 "tbz x8, #0, 292f\n"
5495 "st1 { v11.h }[4], [x16]\n"
5496 "st1 { v15.h }[4], [x25]\n"
5497 "st1 { v19.h }[4], [x24]\n"
5498 "st1 { v23.h }[4], [x23]\n"
5499 "st1 { v27.h }[4], [x22]\n"
5500 "st1 { v31.h }[4], [x21]\n"
5503 "tbz x8, #1, 279f\n"
5504 "str s11, [x16], #0x4\n"
5505 "str s15, [x25], #0x4\n"
5506 "str s19, [x24], #0x4\n"
5507 "str s23, [x23], #0x4\n"
5508 "str s27, [x22], #0x4\n"
5509 "str s31, [x21], #0x4\n"
5510 "tbz x8, #0, 292f\n"
5511 "st1 { v11.h }[2], [x16]\n"
5512 "st1 { v15.h }[2], [x25]\n"
5513 "st1 { v19.h }[2], [x24]\n"
5514 "st1 { v23.h }[2], [x23]\n"
5515 "st1 { v27.h }[2], [x22]\n"
5516 "st1 { v31.h }[2], [x21]\n"
5519 "tbz x8, #0, 292f\n"
5520 "str h11, [x16, #0x0]\n"
5521 "str h15, [x25, #0x0]\n"
5522 "str h19, [x24, #0x0]\n"
5523 "str h23, [x23, #0x0]\n"
5524 "str h27, [x22, #0x0]\n"
5525 "str h31, [x21, #0x0]\n"
5528 "tbz x8, #2, 282f\n"
5529 "str d10, [x16], #0x8\n"
5530 "str d14, [x25], #0x8\n"
5531 "str d18, [x24], #0x8\n"
5532 "str d22, [x23], #0x8\n"
5533 "str d26, [x22], #0x8\n"
5534 "str d30, [x21], #0x8\n"
5535 "tbz x8, #1, 281f\n"
5536 "st1 { v10.s }[2], [x16], #0x4\n"
5537 "st1 { v14.s }[2], [x25], #0x4\n"
5538 "st1 { v18.s }[2], [x24], #0x4\n"
5539 "st1 { v22.s }[2], [x23], #0x4\n"
5540 "st1 { v26.s }[2], [x22], #0x4\n"
5541 "st1 { v30.s }[2], [x21], #0x4\n"
5542 "tbz x8, #0, 292f\n"
5543 "st1 { v10.h }[6], [x16]\n"
5544 "st1 { v14.h }[6], [x25]\n"
5545 "st1 { v18.h }[6], [x24]\n"
5546 "st1 { v22.h }[6], [x23]\n"
5547 "st1 { v26.h }[6], [x22]\n"
5548 "st1 { v30.h }[6], [x21]\n"
5551 "tbz x8, #0, 292f\n"
5552 "st1 { v10.h }[4], [x16]\n"
5553 "st1 { v14.h }[4], [x25]\n"
5554 "st1 { v18.h }[4], [x24]\n"
5555 "st1 { v22.h }[4], [x23]\n"
5556 "st1 { v26.h }[4], [x22]\n"
5557 "st1 { v30.h }[4], [x21]\n"
5560 "tbz x8, #1, 283f\n"
5561 "str s10, [x16], #0x4\n"
5562 "str s14, [x25], #0x4\n"
5563 "str s18, [x24], #0x4\n"
5564 "str s22, [x23], #0x4\n"
5565 "str s26, [x22], #0x4\n"
5566 "str s30, [x21], #0x4\n"
5567 "tbz x8, #0, 292f\n"
5568 "st1 { v10.h }[2], [x16]\n"
5569 "st1 { v14.h }[2], [x25]\n"
5570 "st1 { v18.h }[2], [x24]\n"
5571 "st1 { v22.h }[2], [x23]\n"
5572 "st1 { v26.h }[2], [x22]\n"
5573 "st1 { v30.h }[2], [x21]\n"
5576 "tbz x8, #0, 292f\n"
5577 "str h10, [x16, #0x0]\n"
5578 "str h14, [x25, #0x0]\n"
5579 "str h18, [x24, #0x0]\n"
5580 "str h22, [x23, #0x0]\n"
5581 "str h26, [x22, #0x0]\n"
5582 "str h30, [x21, #0x0]\n"
5585 "tbz x8, #3, 288f\n"
5586 "st1 { v8.8h }, [x16], #0x10\n"
5587 "st1 { v12.8h }, [x25], #0x10\n"
5588 "st1 { v16.8h }, [x24], #0x10\n"
5589 "st1 { v20.8h }, [x23], #0x10\n"
5590 "st1 { v24.8h }, [x22], #0x10\n"
5591 "st1 { v28.8h }, [x21], #0x10\n"
5592 "tbz x8, #2, 286f\n"
5593 "str d9, [x16], #0x8\n"
5594 "str d13, [x25], #0x8\n"
5595 "str d17, [x24], #0x8\n"
5596 "str d21, [x23], #0x8\n"
5597 "str d25, [x22], #0x8\n"
5598 "str d29, [x21], #0x8\n"
5599 "tbz x8, #1, 285f\n"
5600 "st1 { v9.s }[2], [x16], #0x4\n"
5601 "st1 { v13.s }[2], [x25], #0x4\n"
5602 "st1 { v17.s }[2], [x24], #0x4\n"
5603 "st1 { v21.s }[2], [x23], #0x4\n"
5604 "st1 { v25.s }[2], [x22], #0x4\n"
5605 "st1 { v29.s }[2], [x21], #0x4\n"
5606 "tbz x8, #0, 292f\n"
5607 "st1 { v9.h }[6], [x16]\n"
5608 "st1 { v13.h }[6], [x25]\n"
5609 "st1 { v17.h }[6], [x24]\n"
5610 "st1 { v21.h }[6], [x23]\n"
5611 "st1 { v25.h }[6], [x22]\n"
5612 "st1 { v29.h }[6], [x21]\n"
5615 "tbz x8, #0, 292f\n"
5616 "st1 { v9.h }[4], [x16]\n"
5617 "st1 { v13.h }[4], [x25]\n"
5618 "st1 { v17.h }[4], [x24]\n"
5619 "st1 { v21.h }[4], [x23]\n"
5620 "st1 { v25.h }[4], [x22]\n"
5621 "st1 { v29.h }[4], [x21]\n"
5624 "tbz x8, #1, 287f\n"
5625 "str s9, [x16], #0x4\n"
5626 "str s13, [x25], #0x4\n"
5627 "str s17, [x24], #0x4\n"
5628 "str s21, [x23], #0x4\n"
5629 "str s25, [x22], #0x4\n"
5630 "str s29, [x21], #0x4\n"
5631 "tbz x8, #0, 292f\n"
5632 "st1 { v9.h }[2], [x16]\n"
5633 "st1 { v13.h }[2], [x25]\n"
5634 "st1 { v17.h }[2], [x24]\n"
5635 "st1 { v21.h }[2], [x23]\n"
5636 "st1 { v25.h }[2], [x22]\n"
5637 "st1 { v29.h }[2], [x21]\n"
5640 "tbz x8, #0, 292f\n"
5641 "str h9, [x16, #0x0]\n"
5642 "str h13, [x25, #0x0]\n"
5643 "str h17, [x24, #0x0]\n"
5644 "str h21, [x23, #0x0]\n"
5645 "str h25, [x22, #0x0]\n"
5646 "str h29, [x21, #0x0]\n"
5649 "tbz x8, #2, 290f\n"
5650 "str d8, [x16], #0x8\n"
5651 "str d12, [x25], #0x8\n"
5652 "str d16, [x24], #0x8\n"
5653 "str d20, [x23], #0x8\n"
5654 "str d24, [x22], #0x8\n"
5655 "str d28, [x21], #0x8\n"
5656 "tbz x8, #1, 289f\n"
5657 "st1 { v8.s }[2], [x16], #0x4\n"
5658 "st1 { v12.s }[2], [x25], #0x4\n"
5659 "st1 { v16.s }[2], [x24], #0x4\n"
5660 "st1 { v20.s }[2], [x23], #0x4\n"
5661 "st1 { v24.s }[2], [x22], #0x4\n"
5662 "st1 { v28.s }[2], [x21], #0x4\n"
5663 "tbz x8, #0, 292f\n"
5664 "st1 { v8.h }[6], [x16]\n"
5665 "st1 { v12.h }[6], [x25]\n"
5666 "st1 { v16.h }[6], [x24]\n"
5667 "st1 { v20.h }[6], [x23]\n"
5668 "st1 { v24.h }[6], [x22]\n"
5669 "st1 { v28.h }[6], [x21]\n"
5672 "tbz x8, #0, 292f\n"
5673 "st1 { v8.h }[4], [x16]\n"
5674 "st1 { v12.h }[4], [x25]\n"
5675 "st1 { v16.h }[4], [x24]\n"
5676 "st1 { v20.h }[4], [x23]\n"
5677 "st1 { v24.h }[4], [x22]\n"
5678 "st1 { v28.h }[4], [x21]\n"
5681 "tbz x8, #1, 291f\n"
5682 "str s8, [x16], #0x4\n"
5683 "str s12, [x25], #0x4\n"
5684 "str s16, [x24], #0x4\n"
5685 "str s20, [x23], #0x4\n"
5686 "str s24, [x22], #0x4\n"
5687 "str s28, [x21], #0x4\n"
5688 "tbz x8, #0, 292f\n"
5689 "st1 { v8.h }[2], [x16]\n"
5690 "st1 { v12.h }[2], [x25]\n"
5691 "st1 { v16.h }[2], [x24]\n"
5692 "st1 { v20.h }[2], [x23]\n"
5693 "st1 { v24.h }[2], [x22]\n"
5694 "st1 { v28.h }[2], [x21]\n"
5697 "str h8, [x16, #0x0]\n"
5698 "str h12, [x25, #0x0]\n"
5699 "str h16, [x24, #0x0]\n"
5700 "str h20, [x23, #0x0]\n"
5701 "str h24, [x22, #0x0]\n"
5702 "str h28, [x21, #0x0]\n"
5706 "str q8, [x16, #0x0]\n"
5707 "str q9, [x16, #0x10]\n"
5708 "str q10, [x16, #0x20]\n"
5709 "str q11, [x16, #0x30]\n"
5710 "add x16, x16, #0x40\n"
5711 "str q12, [x25, #0x0]\n"
5712 "str q13, [x25, #0x10]\n"
5713 "str q14, [x25, #0x20]\n"
5714 "str q15, [x25, #0x30]\n"
5715 "str q16, [x24, #0x0]\n"
5716 "str q17, [x24, #0x10]\n"
5717 "str q18, [x24, #0x20]\n"
5718 "str q19, [x24, #0x30]\n"
5719 "str q20, [x23, #0x0]\n"
5720 "str q21, [x23, #0x10]\n"
5721 "str q22, [x23, #0x20]\n"
5722 "str q23, [x23, #0x30]\n"
5723 "str q24, [x22, #0x0]\n"
5724 "str q25, [x22, #0x10]\n"
5725 "str q26, [x22, #0x20]\n"
5726 "str q27, [x22, #0x30]\n"
5727 "str q28, [x21, #0x0]\n"
5728 "str q29, [x21, #0x10]\n"
5729 "str q30, [x21, #0x20]\n"
5730 "str q31, [x21, #0x30]\n"
5732 "subs x8, x8, #0x20\n"
5734 "subs %x[M], %x[M], #0x6\n"
5736 "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
5737 "tbz %x[flags], #3, 295f\n"
5738 "add x21, x21, #0x6\n"
5739 "str x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
5743 "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
5746 : [
M]
"+&r" (
M), [input_ptr]
"+&r" (input_ptr), [output_ptr]
"+&r" (output_ptr)
5747 : [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))
5748 :
"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",
"x7",
"x8",
"x9",
"x10",
"x11",
"x12",
"x13",
"x14",
"x15",
"x16",
"x17",
"x20",
"x21",
"x22",
"x23",
"x24",
"x25",
"x26",
"x27",
"x28"
5753 #endif // __aarch64__