27 #include "../../utils.hpp"
28 #include "../../bfloat.hpp"
35 void a64_hybrid_fp32bf16fp32_mmla_6x16 (
36 unsigned int num_strings,
const unsigned int *string_lengths, IndirectInputArg<float> A_arg,
37 size_t M,
size_t N,
const bfloat16 *B_ptr, IndirectOutputArg<float> output_arg,
42 float maxval =
static_cast<float>(std::numeric_limits<float>::infinity());
43 float minval = -
static_cast<float>(std::numeric_limits<float>::infinity());
44 unsigned int num_strings = {};
45 const unsigned int *string_lengths = {};
48 size_t output_offset = {};
49 size_t input_initial_col = {};
50 size_t input_offset = {};
53 unsigned long flags=0;
57 if (output_arg.is_indirect) {
58 output_ptr=(
void *)(output_arg.indirect.ptr);
59 ka.output_offset=output_arg.indirect.offset;
62 output_ptr=(
void *)(output_arg.direct.base);
63 ka.output_offset=output_arg.direct.stride;
66 if (A_arg.is_indirect) {
67 input_ptr=(
void *)(A_arg.indirect.ptr);
68 ka.input_offset=A_arg.indirect.start_row;
69 ka.input_initial_col=A_arg.indirect.start_col;
72 assert(num_strings==1);
73 input_ptr=(
void *)(A_arg.direct.base);
74 ka.input_offset=A_arg.direct.stride;
79 ka.num_strings = num_strings;
80 ka.string_lengths = string_lengths;
88 ka.maxval =
static_cast<float>(act.param1);
105 "mov x12, %x[bias]\n"
106 "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
107 "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
108 "mov x9, %x[output_ptr]\n"
111 "ldr q8, [x12, #0x0]\n"
112 "ldr q9, [x12, #0x10]\n"
113 "zip2 v12.2d, v8.2d, v8.2d\n"
114 "zip1 v8.2d, v8.2d, v8.2d\n"
115 "ldr q10, [x12, #0x20]\n"
116 "ldr q11, [x12, #0x30]\n"
117 "zip2 v13.2d, v9.2d, v9.2d\n"
118 "zip1 v9.2d, v9.2d, v9.2d\n"
119 "zip2 v14.2d, v10.2d, v10.2d\n"
120 "zip1 v10.2d, v10.2d, v10.2d\n"
121 "add x12, x12, #0x40\n"
122 "zip2 v15.2d, v11.2d, v11.2d\n"
123 "zip1 v11.2d, v11.2d, v11.2d\n"
126 "tbz %x[flags], #0, 14f\n"
130 "ld1 { v9.4s }, [x9], #0x10\n"
131 "ld1 { v10.4s }, [x9], #0x10\n"
133 "ld1 { v11.4s }, [x9], #0x10\n"
135 "ldr d16, [x9], #0x8\n"
138 "ld1 { v16.s }[2], [x9]\n"
143 "ldr s16, [x9, #0x0]\n"
147 "ldr d11, [x9], #0x8\n"
150 "ld1 { v11.s }[2], [x9]\n"
155 "ldr s11, [x9, #0x0]\n"
159 "ld1 { v9.4s }, [x9], #0x10\n"
161 "ldr d10, [x9], #0x8\n"
164 "ld1 { v10.s }[2], [x9]\n"
169 "ldr s10, [x9, #0x0]\n"
173 "ldr d9, [x9], #0x8\n"
176 "ld1 { v9.s }[2], [x9]\n"
179 "ldr s9, [x9, #0x0]\n"
185 "ldr q9, [x9, #0x0]\n"
186 "ldr q10, [x9, #0x10]\n"
187 "ldr q11, [x9, #0x20]\n"
188 "ldr q16, [x9, #0x30]\n"
190 "zip1 v8.2d, v9.2d, v12.2d\n"
191 "zip2 v12.2d, v9.2d, v12.2d\n"
192 "zip1 v9.2d, v10.2d, v13.2d\n"
193 "zip2 v13.2d, v10.2d, v13.2d\n"
194 "zip1 v10.2d, v11.2d, v14.2d\n"
195 "zip2 v14.2d, v11.2d, v14.2d\n"
196 "zip1 v11.2d, v16.2d, v15.2d\n"
197 "zip2 v15.2d, v16.2d, v15.2d\n"
200 "movi v8.16b, #0x0\n"
201 "movi v9.16b, #0x0\n"
202 "movi v10.16b, #0x0\n"
203 "movi v11.16b, #0x0\n"
204 "movi v12.16b, #0x0\n"
205 "movi v13.16b, #0x0\n"
206 "movi v14.16b, #0x0\n"
207 "movi v15.16b, #0x0\n"
211 "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
212 "ldr w27, [x20, x28, LSL #0x2]\n"
213 "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
214 "tbz %x[flags], #3, 17f\n"
215 "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
216 "add x20, x20, x21, LSL #3\n"
217 "ldr x26, [x20, #0x0]\n"
219 "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
220 "add x26, x26, x20, LSL #2\n"
223 "mov x26, %x[input_ptr]\n"
227 "ld1 { v0.4s }, [x26], #0x10\n"
228 "ldr q6, [x10, #0x0]\n"
230 "ldr q7, [x10, #0x10]\n"
233 ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n"
234 ".inst 0x6e46ec08 // bfmmla v8.4s, v0.8h, v6.8h\n"
235 "ldr q18, [x10, #0x20]\n"
236 ".inst 0x6e47ec0c // bfmmla v12.4s, v0.8h, v7.8h\n"
237 "ldr q17, [x10, #0x30]\n"
238 ".inst 0x6e52ec09 // bfmmla v9.4s, v0.8h, v18.8h\n"
239 "ldr q18, [x10, #0x40]\n"
240 ".inst 0x6e51ec0d // bfmmla v13.4s, v0.8h, v17.8h\n"
241 "ldr q17, [x10, #0x50]\n"
242 ".inst 0x6e52ec0a // bfmmla v10.4s, v0.8h, v18.8h\n"
243 "ldr q18, [x10, #0x60]\n"
244 ".inst 0x6e51ec0e // bfmmla v14.4s, v0.8h, v17.8h\n"
245 "ldr q17, [x10, #0x70]\n"
246 "sub x27, x27, #0x4\n"
248 ".inst 0x6e52ec0b // bfmmla v11.4s, v0.8h, v18.8h\n"
249 "add x10, x10, #0x80\n"
250 "ldr q6, [x10, #0x0]\n"
251 ".inst 0x6e51ec0f // bfmmla v15.4s, v0.8h, v17.8h\n"
252 "ldr q7, [x10, #0x10]\n"
253 "prfm pldl1keep, [x26, #0x80]\n"
254 "ld1 { v0.4s }, [x26], #0x10\n"
257 ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n"
258 ".inst 0x6e46ec08 // bfmmla v8.4s, v0.8h, v6.8h\n"
259 "ldr q18, [x10, #0x20]\n"
260 ".inst 0x6e47ec0c // bfmmla v12.4s, v0.8h, v7.8h\n"
261 "ldr q17, [x10, #0x30]\n"
262 ".inst 0x6e52ec09 // bfmmla v9.4s, v0.8h, v18.8h\n"
263 "ldr q18, [x10, #0x40]\n"
264 ".inst 0x6e51ec0d // bfmmla v13.4s, v0.8h, v17.8h\n"
265 "ldr q17, [x10, #0x50]\n"
266 ".inst 0x6e52ec0a // bfmmla v10.4s, v0.8h, v18.8h\n"
267 "ldr q18, [x10, #0x60]\n"
268 ".inst 0x6e51ec0e // bfmmla v14.4s, v0.8h, v17.8h\n"
269 "ldr q17, [x10, #0x70]\n"
270 "sub x27, x27, #0x4\n"
271 ".inst 0x6e52ec0b // bfmmla v11.4s, v0.8h, v18.8h\n"
272 ".inst 0x6e51ec0f // bfmmla v15.4s, v0.8h, v17.8h\n"
273 "prfm pldl1keep, [x26, #0x80]\n"
274 "add x10, x10, #0x80\n"
279 "ldr d0, [x26], #0x8\n"
281 "ld1 { v0.s }[2], [x26]\n"
284 "ldr s0, [x26, #0x0]\n"
286 "ldr q18, [x10, #0x0]\n"
287 "ldr q17, [x10, #0x10]\n"
288 ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n"
289 ".inst 0x6e52ec08 // bfmmla v8.4s, v0.8h, v18.8h\n"
290 "ldr q18, [x10, #0x20]\n"
291 ".inst 0x6e51ec0c // bfmmla v12.4s, v0.8h, v17.8h\n"
292 "ldr q17, [x10, #0x30]\n"
293 ".inst 0x6e52ec09 // bfmmla v9.4s, v0.8h, v18.8h\n"
294 "ldr q18, [x10, #0x40]\n"
295 ".inst 0x6e51ec0d // bfmmla v13.4s, v0.8h, v17.8h\n"
296 "ldr q17, [x10, #0x50]\n"
297 ".inst 0x6e52ec0a // bfmmla v10.4s, v0.8h, v18.8h\n"
298 "ldr q18, [x10, #0x60]\n"
299 ".inst 0x6e51ec0e // bfmmla v14.4s, v0.8h, v17.8h\n"
300 "ldr q17, [x10, #0x70]\n"
301 ".inst 0x6e52ec0b // bfmmla v11.4s, v0.8h, v18.8h\n"
302 ".inst 0x6e51ec0f // bfmmla v15.4s, v0.8h, v17.8h\n"
303 "add x10, x10, #0x80\n"
305 "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
306 "add x28, x28, #0x1\n"
309 "uzp1 v8.2d, v8.2d, v12.2d\n"
310 "uzp1 v9.2d, v9.2d, v13.2d\n"
311 "prfm pstl1keep, [x9, #0x0]\n"
312 "uzp1 v10.2d, v10.2d, v14.2d\n"
313 "uzp1 v11.2d, v11.2d, v15.2d\n"
314 "tbz %x[flags], #1, 25f\n"
315 "add x20, %x[args_ptr], %[offset_max]\n"
316 "ld1r { v18.4s }, [x20]\n"
317 "add x20, %x[args_ptr], %[offset_min]\n"
318 "ld1r { v17.4s }, [x20]\n"
319 "fmin v8.4s, v8.4s, v18.4s\n"
320 "fmin v9.4s, v9.4s, v18.4s\n"
321 "fmin v10.4s, v10.4s, v18.4s\n"
322 "fmin v11.4s, v11.4s, v18.4s\n"
323 "fmax v8.4s, v8.4s, v17.4s\n"
324 "fmax v9.4s, v9.4s, v17.4s\n"
325 "fmax v10.4s, v10.4s, v17.4s\n"
326 "fmax v11.4s, v11.4s, v17.4s\n"
331 "st1 { v8.4s }, [x9], #0x10\n"
332 "st1 { v9.4s }, [x9], #0x10\n"
334 "st1 { v10.4s }, [x9], #0x10\n"
336 "str d11, [x9], #0x8\n"
338 "st1 { v11.s }[2], [x9]\n"
342 "str s11, [x9, #0x0]\n"
346 "str d10, [x9], #0x8\n"
348 "st1 { v10.s }[2], [x9]\n"
352 "str s10, [x9, #0x0]\n"
356 "st1 { v8.4s }, [x9], #0x10\n"
358 "str d9, [x9], #0x8\n"
360 "st1 { v9.s }[2], [x9]\n"
364 "str s9, [x9, #0x0]\n"
368 "str d8, [x9], #0x8\n"
370 "st1 { v8.s }[2], [x9]\n"
373 "str s8, [x9, #0x0]\n"
377 "str q8, [x9, #0x0]\n"
378 "str q9, [x9, #0x10]\n"
379 "str q10, [x9, #0x20]\n"
380 "str q11, [x9, #0x30]\n"
381 "add x9, x9, #0x40\n"
383 "subs x11, x11, #0x10\n"
387 "mov x12, %x[bias]\n"
388 "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
389 "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
390 "mov x9, %x[output_ptr]\n"
393 "ldr q8, [x12, #0x0]\n"
394 "ldr q9, [x12, #0x10]\n"
395 "zip2 v12.2d, v8.2d, v8.2d\n"
396 "zip1 v8.2d, v8.2d, v8.2d\n"
397 "ldr q10, [x12, #0x20]\n"
398 "ldr q11, [x12, #0x30]\n"
399 "zip2 v13.2d, v9.2d, v9.2d\n"
400 "zip1 v9.2d, v9.2d, v9.2d\n"
401 "zip2 v14.2d, v10.2d, v10.2d\n"
402 "zip1 v10.2d, v10.2d, v10.2d\n"
403 "add x12, x12, #0x40\n"
404 "zip2 v15.2d, v11.2d, v11.2d\n"
405 "zip1 v11.2d, v11.2d, v11.2d\n"
408 "tbz %x[flags], #0, 49f\n"
409 "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
411 "add x25, x9, x20, LSL #2\n"
414 "ld1 { v9.4s }, [x9], #0x10\n"
415 "ld1 { v12.4s }, [x25], #0x10\n"
416 "ld1 { v10.4s }, [x9], #0x10\n"
417 "ld1 { v13.4s }, [x25], #0x10\n"
419 "ld1 { v11.4s }, [x9], #0x10\n"
420 "ld1 { v14.4s }, [x25], #0x10\n"
422 "ldr d16, [x9], #0x8\n"
423 "ldr d15, [x25], #0x8\n"
426 "ld1 { v16.s }[2], [x9]\n"
427 "ld1 { v15.s }[2], [x25]\n"
432 "ldr s16, [x9, #0x0]\n"
433 "ldr s15, [x25, #0x0]\n"
437 "ldr d11, [x9], #0x8\n"
438 "ldr d14, [x25], #0x8\n"
441 "ld1 { v11.s }[2], [x9]\n"
442 "ld1 { v14.s }[2], [x25]\n"
447 "ldr s11, [x9, #0x0]\n"
448 "ldr s14, [x25, #0x0]\n"
452 "ld1 { v9.4s }, [x9], #0x10\n"
453 "ld1 { v12.4s }, [x25], #0x10\n"
455 "ldr d10, [x9], #0x8\n"
456 "ldr d13, [x25], #0x8\n"
459 "ld1 { v10.s }[2], [x9]\n"
460 "ld1 { v13.s }[2], [x25]\n"
465 "ldr s10, [x9, #0x0]\n"
466 "ldr s13, [x25, #0x0]\n"
470 "ldr d9, [x9], #0x8\n"
471 "ldr d12, [x25], #0x8\n"
474 "ld1 { v9.s }[2], [x9]\n"
475 "ld1 { v12.s }[2], [x25]\n"
478 "ldr s9, [x9, #0x0]\n"
479 "ldr s12, [x25, #0x0]\n"
485 "ldr q9, [x9, #0x0]\n"
486 "ldr q10, [x9, #0x10]\n"
487 "ldr q11, [x9, #0x20]\n"
488 "ldr q16, [x9, #0x30]\n"
489 "ldr q12, [x25, #0x0]\n"
490 "ldr q13, [x25, #0x10]\n"
491 "ldr q14, [x25, #0x20]\n"
492 "ldr q15, [x25, #0x30]\n"
494 "zip1 v8.2d, v9.2d, v12.2d\n"
495 "zip2 v12.2d, v9.2d, v12.2d\n"
496 "zip1 v9.2d, v10.2d, v13.2d\n"
497 "zip2 v13.2d, v10.2d, v13.2d\n"
498 "zip1 v10.2d, v11.2d, v14.2d\n"
499 "zip2 v14.2d, v11.2d, v14.2d\n"
500 "zip1 v11.2d, v16.2d, v15.2d\n"
501 "zip2 v15.2d, v16.2d, v15.2d\n"
504 "movi v8.16b, #0x0\n"
505 "movi v9.16b, #0x0\n"
506 "movi v10.16b, #0x0\n"
507 "movi v11.16b, #0x0\n"
508 "movi v12.16b, #0x0\n"
509 "movi v13.16b, #0x0\n"
510 "movi v14.16b, #0x0\n"
511 "movi v15.16b, #0x0\n"
515 "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
516 "ldr w27, [x20, x28, LSL #0x2]\n"
517 "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
518 "tbz %x[flags], #3, 52f\n"
519 "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
520 "add x20, x20, x21, LSL #3\n"
521 "ldr x26, [x20, #0x0]\n"
522 "ldr x25, [x20, #0x8]\n"
524 "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
525 "add x26, x26, x20, LSL #2\n"
526 "add x25, x25, x20, LSL #2\n"
529 "mov x26, %x[input_ptr]\n"
530 "add x25, x26, x21, LSL #2\n"
534 "ld1 { v0.4s }, [x26], #0x10\n"
535 "ld1 { v1.4s }, [x25], #0x10\n"
537 "ldr q6, [x10, #0x0]\n"
538 "ldr q7, [x10, #0x10]\n"
541 ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n"
542 ".inst 0x4ea16820 // bfcvtn2 v0.8h, v1.4s\n"
543 ".inst 0x6e46ec08 // bfmmla v8.4s, v0.8h, v6.8h\n"
544 "ldr q18, [x10, #0x20]\n"
545 ".inst 0x6e47ec0c // bfmmla v12.4s, v0.8h, v7.8h\n"
546 "ldr q17, [x10, #0x30]\n"
547 ".inst 0x6e52ec09 // bfmmla v9.4s, v0.8h, v18.8h\n"
548 "ldr q18, [x10, #0x40]\n"
549 ".inst 0x6e51ec0d // bfmmla v13.4s, v0.8h, v17.8h\n"
550 "ldr q17, [x10, #0x50]\n"
551 ".inst 0x6e52ec0a // bfmmla v10.4s, v0.8h, v18.8h\n"
552 "ldr q18, [x10, #0x60]\n"
553 ".inst 0x6e51ec0e // bfmmla v14.4s, v0.8h, v17.8h\n"
554 "ldr q17, [x10, #0x70]\n"
555 "sub x27, x27, #0x4\n"
557 "add x10, x10, #0x80\n"
558 ".inst 0x6e52ec0b // bfmmla v11.4s, v0.8h, v18.8h\n"
559 "ldr q6, [x10, #0x0]\n"
560 ".inst 0x6e51ec0f // bfmmla v15.4s, v0.8h, v17.8h\n"
561 "ldr q7, [x10, #0x10]\n"
562 "prfm pldl1keep, [x26, #0x80]\n"
563 "ld1 { v0.4s }, [x26], #0x10\n"
564 "prfm pldl1keep, [x25, #0x80]\n"
565 "ld1 { v1.4s }, [x25], #0x10\n"
568 ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n"
569 ".inst 0x4ea16820 // bfcvtn2 v0.8h, v1.4s\n"
570 ".inst 0x6e46ec08 // bfmmla v8.4s, v0.8h, v6.8h\n"
571 "ldr q18, [x10, #0x20]\n"
572 ".inst 0x6e47ec0c // bfmmla v12.4s, v0.8h, v7.8h\n"
573 "ldr q17, [x10, #0x30]\n"
574 ".inst 0x6e52ec09 // bfmmla v9.4s, v0.8h, v18.8h\n"
575 "ldr q18, [x10, #0x40]\n"
576 ".inst 0x6e51ec0d // bfmmla v13.4s, v0.8h, v17.8h\n"
577 "ldr q17, [x10, #0x50]\n"
578 ".inst 0x6e52ec0a // bfmmla v10.4s, v0.8h, v18.8h\n"
579 "ldr q18, [x10, #0x60]\n"
580 ".inst 0x6e51ec0e // bfmmla v14.4s, v0.8h, v17.8h\n"
581 "ldr q17, [x10, #0x70]\n"
582 "sub x27, x27, #0x4\n"
583 ".inst 0x6e52ec0b // bfmmla v11.4s, v0.8h, v18.8h\n"
584 ".inst 0x6e51ec0f // bfmmla v15.4s, v0.8h, v17.8h\n"
585 "prfm pldl1keep, [x26, #0x80]\n"
586 "prfm pldl1keep, [x25, #0x80]\n"
587 "add x10, x10, #0x80\n"
592 "ldr d0, [x26], #0x8\n"
593 "ldr d1, [x25], #0x8\n"
595 "ld1 { v0.s }[2], [x26]\n"
596 "ld1 { v1.s }[2], [x25]\n"
599 "ldr s0, [x26, #0x0]\n"
600 "ldr s1, [x25, #0x0]\n"
602 "ldr q18, [x10, #0x0]\n"
603 "ldr q17, [x10, #0x10]\n"
604 ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n"
605 ".inst 0x4ea16820 // bfcvtn2 v0.8h, v1.4s\n"
606 ".inst 0x6e52ec08 // bfmmla v8.4s, v0.8h, v18.8h\n"
607 "ldr q18, [x10, #0x20]\n"
608 ".inst 0x6e51ec0c // bfmmla v12.4s, v0.8h, v17.8h\n"
609 "ldr q17, [x10, #0x30]\n"
610 ".inst 0x6e52ec09 // bfmmla v9.4s, v0.8h, v18.8h\n"
611 "ldr q18, [x10, #0x40]\n"
612 ".inst 0x6e51ec0d // bfmmla v13.4s, v0.8h, v17.8h\n"
613 "ldr q17, [x10, #0x50]\n"
614 ".inst 0x6e52ec0a // bfmmla v10.4s, v0.8h, v18.8h\n"
615 "ldr q18, [x10, #0x60]\n"
616 ".inst 0x6e51ec0e // bfmmla v14.4s, v0.8h, v17.8h\n"
617 "ldr q17, [x10, #0x70]\n"
618 ".inst 0x6e52ec0b // bfmmla v11.4s, v0.8h, v18.8h\n"
619 ".inst 0x6e51ec0f // bfmmla v15.4s, v0.8h, v17.8h\n"
620 "add x10, x10, #0x80\n"
622 "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
623 "add x28, x28, #0x1\n"
626 "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
627 "add x25, x9, x20, LSL #2\n"
628 "uzp1 v6.2d, v8.2d, v12.2d\n"
629 "uzp2 v8.2d, v8.2d, v12.2d\n"
630 "uzp1 v12.2d, v9.2d, v13.2d\n"
631 "uzp2 v9.2d, v9.2d, v13.2d\n"
632 "prfm pstl1keep, [x9, #0x0]\n"
633 "prfm pstl1keep, [x25, #0x0]\n"
634 "uzp1 v13.2d, v10.2d, v14.2d\n"
635 "uzp2 v10.2d, v10.2d, v14.2d\n"
636 "uzp1 v14.2d, v11.2d, v15.2d\n"
637 "uzp2 v11.2d, v11.2d, v15.2d\n"
638 "tbz %x[flags], #1, 60f\n"
639 "add x20, %x[args_ptr], %[offset_max]\n"
640 "ld1r { v18.4s }, [x20]\n"
641 "add x20, %x[args_ptr], %[offset_min]\n"
642 "ld1r { v17.4s }, [x20]\n"
643 "fmin v6.4s, v6.4s, v18.4s\n"
644 "fmin v12.4s, v12.4s, v18.4s\n"
645 "fmin v13.4s, v13.4s, v18.4s\n"
646 "fmin v14.4s, v14.4s, v18.4s\n"
647 "fmin v8.4s, v8.4s, v18.4s\n"
648 "fmin v9.4s, v9.4s, v18.4s\n"
649 "fmin v10.4s, v10.4s, v18.4s\n"
650 "fmin v11.4s, v11.4s, v18.4s\n"
651 "fmax v6.4s, v6.4s, v17.4s\n"
652 "fmax v12.4s, v12.4s, v17.4s\n"
653 "fmax v13.4s, v13.4s, v17.4s\n"
654 "fmax v14.4s, v14.4s, v17.4s\n"
655 "fmax v8.4s, v8.4s, v17.4s\n"
656 "fmax v9.4s, v9.4s, v17.4s\n"
657 "fmax v10.4s, v10.4s, v17.4s\n"
658 "fmax v11.4s, v11.4s, v17.4s\n"
663 "st1 { v6.4s }, [x9], #0x10\n"
664 "st1 { v12.4s }, [x9], #0x10\n"
665 "st1 { v8.4s }, [x25], #0x10\n"
666 "st1 { v9.4s }, [x25], #0x10\n"
668 "st1 { v13.4s }, [x9], #0x10\n"
669 "st1 { v10.4s }, [x25], #0x10\n"
671 "str d14, [x9], #0x8\n"
672 "str d11, [x25], #0x8\n"
674 "st1 { v14.s }[2], [x9]\n"
675 "st1 { v11.s }[2], [x25]\n"
679 "str s14, [x9, #0x0]\n"
680 "str s11, [x25, #0x0]\n"
684 "str d13, [x9], #0x8\n"
685 "str d10, [x25], #0x8\n"
687 "st1 { v13.s }[2], [x9]\n"
688 "st1 { v10.s }[2], [x25]\n"
692 "str s13, [x9, #0x0]\n"
693 "str s10, [x25, #0x0]\n"
697 "st1 { v6.4s }, [x9], #0x10\n"
698 "st1 { v8.4s }, [x25], #0x10\n"
700 "str d12, [x9], #0x8\n"
701 "str d9, [x25], #0x8\n"
703 "st1 { v12.s }[2], [x9]\n"
704 "st1 { v9.s }[2], [x25]\n"
708 "str s12, [x9, #0x0]\n"
709 "str s9, [x25, #0x0]\n"
713 "str d6, [x9], #0x8\n"
714 "str d8, [x25], #0x8\n"
716 "st1 { v6.s }[2], [x9]\n"
717 "st1 { v8.s }[2], [x25]\n"
720 "str s6, [x9, #0x0]\n"
721 "str s8, [x25, #0x0]\n"
725 "str q6, [x9, #0x0]\n"
726 "str q12, [x9, #0x10]\n"
727 "str q13, [x9, #0x20]\n"
728 "str q14, [x9, #0x30]\n"
729 "add x9, x9, #0x40\n"
730 "str q8, [x25, #0x0]\n"
731 "str q9, [x25, #0x10]\n"
732 "str q10, [x25, #0x20]\n"
733 "str q11, [x25, #0x30]\n"
735 "subs x11, x11, #0x10\n"
739 "mov x12, %x[bias]\n"
740 "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
741 "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
742 "mov x9, %x[output_ptr]\n"
745 "ldr q8, [x12, #0x0]\n"
746 "ldr q9, [x12, #0x10]\n"
747 "zip2 v12.2d, v8.2d, v8.2d\n"
748 "zip1 v8.2d, v8.2d, v8.2d\n"
749 "ldr q10, [x12, #0x20]\n"
750 "ldr q11, [x12, #0x30]\n"
751 "zip2 v13.2d, v9.2d, v9.2d\n"
752 "zip1 v9.2d, v9.2d, v9.2d\n"
753 "zip2 v14.2d, v10.2d, v10.2d\n"
754 "zip1 v10.2d, v10.2d, v10.2d\n"
755 "add x12, x12, #0x40\n"
756 "zip2 v15.2d, v11.2d, v11.2d\n"
757 "zip1 v11.2d, v11.2d, v11.2d\n"
758 "mov v16.16b, v8.16b\n"
759 "mov v20.16b, v12.16b\n"
760 "mov v17.16b, v9.16b\n"
761 "mov v21.16b, v13.16b\n"
762 "mov v18.16b, v10.16b\n"
763 "mov v22.16b, v14.16b\n"
764 "mov v19.16b, v11.16b\n"
765 "mov v23.16b, v15.16b\n"
768 "tbz %x[flags], #0, 84f\n"
769 "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
770 "add x25, x9, x20, LSL #2\n"
772 "add x24, x25, x20, LSL #2\n"
775 "ld1 { v9.4s }, [x9], #0x10\n"
776 "ld1 { v12.4s }, [x25], #0x10\n"
777 "ld1 { v17.4s }, [x24], #0x10\n"
778 "ld1 { v10.4s }, [x9], #0x10\n"
779 "ld1 { v13.4s }, [x25], #0x10\n"
780 "ld1 { v18.4s }, [x24], #0x10\n"
782 "ld1 { v11.4s }, [x9], #0x10\n"
783 "ld1 { v14.4s }, [x25], #0x10\n"
784 "ld1 { v19.4s }, [x24], #0x10\n"
786 "ldr d16, [x9], #0x8\n"
787 "ldr d15, [x25], #0x8\n"
789 "ldr d24, [x24], #0x8\n"
791 "ld1 { v16.s }[2], [x9]\n"
792 "ld1 { v15.s }[2], [x25]\n"
793 "ld1 { v24.s }[2], [x24]\n"
798 "ldr s16, [x9, #0x0]\n"
799 "ldr s15, [x25, #0x0]\n"
800 "ldr s24, [x24, #0x0]\n"
804 "ldr d11, [x9], #0x8\n"
805 "ldr d14, [x25], #0x8\n"
807 "ldr d19, [x24], #0x8\n"
809 "ld1 { v11.s }[2], [x9]\n"
810 "ld1 { v14.s }[2], [x25]\n"
811 "ld1 { v19.s }[2], [x24]\n"
816 "ldr s11, [x9, #0x0]\n"
817 "ldr s14, [x25, #0x0]\n"
818 "ldr s19, [x24, #0x0]\n"
822 "ld1 { v9.4s }, [x9], #0x10\n"
823 "ld1 { v12.4s }, [x25], #0x10\n"
824 "ld1 { v17.4s }, [x24], #0x10\n"
826 "ldr d10, [x9], #0x8\n"
827 "ldr d13, [x25], #0x8\n"
829 "ldr d18, [x24], #0x8\n"
831 "ld1 { v10.s }[2], [x9]\n"
832 "ld1 { v13.s }[2], [x25]\n"
833 "ld1 { v18.s }[2], [x24]\n"
838 "ldr s10, [x9, #0x0]\n"
839 "ldr s13, [x25, #0x0]\n"
840 "ldr s18, [x24, #0x0]\n"
844 "ldr d9, [x9], #0x8\n"
845 "ldr d12, [x25], #0x8\n"
847 "ldr d17, [x24], #0x8\n"
849 "ld1 { v9.s }[2], [x9]\n"
850 "ld1 { v12.s }[2], [x25]\n"
851 "ld1 { v17.s }[2], [x24]\n"
854 "ldr s9, [x9, #0x0]\n"
855 "ldr s12, [x25, #0x0]\n"
857 "ldr s17, [x24, #0x0]\n"
862 "ldr q9, [x9, #0x0]\n"
863 "ldr q10, [x9, #0x10]\n"
864 "ldr q11, [x9, #0x20]\n"
865 "ldr q16, [x9, #0x30]\n"
866 "ldr q12, [x25, #0x0]\n"
867 "ldr q13, [x25, #0x10]\n"
868 "ldr q14, [x25, #0x20]\n"
869 "ldr q15, [x25, #0x30]\n"
870 "ldr q17, [x24, #0x0]\n"
871 "ldr q18, [x24, #0x10]\n"
872 "ldr q19, [x24, #0x20]\n"
873 "ldr q24, [x24, #0x30]\n"
875 "zip1 v8.2d, v9.2d, v12.2d\n"
876 "zip2 v12.2d, v9.2d, v12.2d\n"
877 "zip1 v9.2d, v10.2d, v13.2d\n"
878 "zip2 v13.2d, v10.2d, v13.2d\n"
879 "zip1 v10.2d, v11.2d, v14.2d\n"
880 "zip2 v14.2d, v11.2d, v14.2d\n"
881 "zip1 v11.2d, v16.2d, v15.2d\n"
882 "zip2 v15.2d, v16.2d, v15.2d\n"
883 "zip1 v16.2d, v17.2d, v20.2d\n"
884 "zip2 v20.2d, v17.2d, v20.2d\n"
885 "zip1 v17.2d, v18.2d, v21.2d\n"
886 "zip2 v21.2d, v18.2d, v21.2d\n"
887 "zip1 v18.2d, v19.2d, v22.2d\n"
888 "zip2 v22.2d, v19.2d, v22.2d\n"
889 "zip1 v19.2d, v24.2d, v23.2d\n"
890 "zip2 v23.2d, v24.2d, v23.2d\n"
893 "movi v8.16b, #0x0\n"
894 "movi v9.16b, #0x0\n"
895 "movi v10.16b, #0x0\n"
896 "movi v11.16b, #0x0\n"
897 "movi v12.16b, #0x0\n"
898 "movi v13.16b, #0x0\n"
899 "movi v14.16b, #0x0\n"
900 "movi v15.16b, #0x0\n"
901 "movi v16.16b, #0x0\n"
902 "movi v17.16b, #0x0\n"
903 "movi v18.16b, #0x0\n"
904 "movi v19.16b, #0x0\n"
905 "movi v20.16b, #0x0\n"
906 "movi v21.16b, #0x0\n"
907 "movi v22.16b, #0x0\n"
908 "movi v23.16b, #0x0\n"
912 "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
913 "ldr w27, [x20, x28, LSL #0x2]\n"
914 "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
915 "tbz %x[flags], #3, 87f\n"
916 "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
917 "add x20, x20, x21, LSL #3\n"
918 "ldr x26, [x20, #0x0]\n"
919 "ldr x25, [x20, #0x8]\n"
920 "ldr x24, [x20, #0x10]\n"
922 "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
923 "add x26, x26, x20, LSL #2\n"
924 "add x25, x25, x20, LSL #2\n"
925 "add x24, x24, x20, LSL #2\n"
928 "mov x26, %x[input_ptr]\n"
929 "add x25, x26, x21, LSL #2\n"
930 "add x24, x25, x21, LSL #2\n"
934 "ld1 { v0.4s }, [x26], #0x10\n"
935 "ld1 { v1.4s }, [x25], #0x10\n"
937 "ld1 { v2.4s }, [x24], #0x10\n"
938 "ldr q6, [x10, #0x0]\n"
939 "ldr q7, [x10, #0x10]\n"
942 ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n"
943 ".inst 0x4ea16820 // bfcvtn2 v0.8h, v1.4s\n"
944 ".inst 0x6e46ec08 // bfmmla v8.4s, v0.8h, v6.8h\n"
945 "sub x27, x27, #0x4\n"
946 ".inst 0x0ea16842 // bfcvtn v2.4h, v2.4s\n"
947 ".inst 0x6e46ec50 // bfmmla v16.4s, v2.8h, v6.8h\n"
948 "ldr q26, [x10, #0x20]\n"
949 ".inst 0x6e47ec0c // bfmmla v12.4s, v0.8h, v7.8h\n"
950 ".inst 0x6e47ec54 // bfmmla v20.4s, v2.8h, v7.8h\n"
951 "ldr q25, [x10, #0x30]\n"
952 ".inst 0x6e5aec09 // bfmmla v9.4s, v0.8h, v26.8h\n"
954 ".inst 0x6e5aec51 // bfmmla v17.4s, v2.8h, v26.8h\n"
955 "ldr q26, [x10, #0x40]\n"
956 ".inst 0x6e59ec0d // bfmmla v13.4s, v0.8h, v25.8h\n"
957 "prfm pldl1keep, [x26, #0x80]\n"
958 ".inst 0x6e59ec55 // bfmmla v21.4s, v2.8h, v25.8h\n"
959 "ldr q25, [x10, #0x50]\n"
960 ".inst 0x6e5aec0a // bfmmla v10.4s, v0.8h, v26.8h\n"
961 "prfm pldl1keep, [x25, #0x80]\n"
962 "ld1 { v1.4s }, [x25], #0x10\n"
963 ".inst 0x6e5aec52 // bfmmla v18.4s, v2.8h, v26.8h\n"
964 "ldr q26, [x10, #0x60]\n"
965 ".inst 0x6e59ec0e // bfmmla v14.4s, v0.8h, v25.8h\n"
966 ".inst 0x6e59ec56 // bfmmla v22.4s, v2.8h, v25.8h\n"
967 "ldr q25, [x10, #0x70]\n"
968 "add x10, x10, #0x80\n"
969 "prfm pldl1keep, [x24, #0x80]\n"
970 ".inst 0x6e5aec0b // bfmmla v11.4s, v0.8h, v26.8h\n"
971 ".inst 0x6e5aec53 // bfmmla v19.4s, v2.8h, v26.8h\n"
972 "ldr q6, [x10, #0x0]\n"
973 ".inst 0x6e59ec0f // bfmmla v15.4s, v0.8h, v25.8h\n"
974 "ld1 { v0.4s }, [x26], #0x10\n"
975 ".inst 0x6e59ec57 // bfmmla v23.4s, v2.8h, v25.8h\n"
976 "ld1 { v2.4s }, [x24], #0x10\n"
977 "ldr q7, [x10, #0x10]\n"
980 ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n"
981 ".inst 0x4ea16820 // bfcvtn2 v0.8h, v1.4s\n"
982 ".inst 0x6e46ec08 // bfmmla v8.4s, v0.8h, v6.8h\n"
983 "sub x27, x27, #0x4\n"
984 ".inst 0x0ea16842 // bfcvtn v2.4h, v2.4s\n"
985 ".inst 0x6e46ec50 // bfmmla v16.4s, v2.8h, v6.8h\n"
986 "ldr q26, [x10, #0x20]\n"
987 ".inst 0x6e47ec0c // bfmmla v12.4s, v0.8h, v7.8h\n"
988 ".inst 0x6e47ec54 // bfmmla v20.4s, v2.8h, v7.8h\n"
989 "ldr q25, [x10, #0x30]\n"
990 ".inst 0x6e5aec09 // bfmmla v9.4s, v0.8h, v26.8h\n"
991 "prfm pldl1keep, [x26, #0x80]\n"
992 ".inst 0x6e5aec51 // bfmmla v17.4s, v2.8h, v26.8h\n"
993 "ldr q26, [x10, #0x40]\n"
994 ".inst 0x6e59ec0d // bfmmla v13.4s, v0.8h, v25.8h\n"
995 "prfm pldl1keep, [x25, #0x80]\n"
996 ".inst 0x6e59ec55 // bfmmla v21.4s, v2.8h, v25.8h\n"
997 "ldr q25, [x10, #0x50]\n"
998 ".inst 0x6e5aec0a // bfmmla v10.4s, v0.8h, v26.8h\n"
999 "prfm pldl1keep, [x24, #0x80]\n"
1000 ".inst 0x6e5aec52 // bfmmla v18.4s, v2.8h, v26.8h\n"
1001 "ldr q26, [x10, #0x60]\n"
1002 ".inst 0x6e59ec0e // bfmmla v14.4s, v0.8h, v25.8h\n"
1003 ".inst 0x6e59ec56 // bfmmla v22.4s, v2.8h, v25.8h\n"
1004 "ldr q25, [x10, #0x70]\n"
1005 "add x10, x10, #0x80\n"
1006 ".inst 0x6e5aec0b // bfmmla v11.4s, v0.8h, v26.8h\n"
1007 ".inst 0x6e5aec53 // bfmmla v19.4s, v2.8h, v26.8h\n"
1008 ".inst 0x6e59ec0f // bfmmla v15.4s, v0.8h, v25.8h\n"
1009 ".inst 0x6e59ec57 // bfmmla v23.4s, v2.8h, v25.8h\n"
1013 "tbz x27, #1, 92f\n"
1014 "ldr d0, [x26], #0x8\n"
1015 "ldr d1, [x25], #0x8\n"
1016 "ldr d2, [x24], #0x8\n"
1017 "tbz x27, #0, 93f\n"
1018 "ld1 { v0.s }[2], [x26]\n"
1019 "ld1 { v1.s }[2], [x25]\n"
1020 "ld1 { v2.s }[2], [x24]\n"
1023 "ldr s0, [x26, #0x0]\n"
1024 "ldr s1, [x25, #0x0]\n"
1025 "ldr s2, [x24, #0x0]\n"
1027 "ldr q26, [x10, #0x0]\n"
1028 "ldr q25, [x10, #0x10]\n"
1029 ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n"
1030 ".inst 0x4ea16820 // bfcvtn2 v0.8h, v1.4s\n"
1031 ".inst 0x0ea16842 // bfcvtn v2.4h, v2.4s\n"
1032 ".inst 0x6e5aec08 // bfmmla v8.4s, v0.8h, v26.8h\n"
1033 ".inst 0x6e5aec50 // bfmmla v16.4s, v2.8h, v26.8h\n"
1034 "ldr q26, [x10, #0x20]\n"
1035 ".inst 0x6e59ec0c // bfmmla v12.4s, v0.8h, v25.8h\n"
1036 ".inst 0x6e59ec54 // bfmmla v20.4s, v2.8h, v25.8h\n"
1037 "ldr q25, [x10, #0x30]\n"
1038 ".inst 0x6e5aec09 // bfmmla v9.4s, v0.8h, v26.8h\n"
1039 ".inst 0x6e5aec51 // bfmmla v17.4s, v2.8h, v26.8h\n"
1040 "ldr q26, [x10, #0x40]\n"
1041 ".inst 0x6e59ec0d // bfmmla v13.4s, v0.8h, v25.8h\n"
1042 ".inst 0x6e59ec55 // bfmmla v21.4s, v2.8h, v25.8h\n"
1043 "ldr q25, [x10, #0x50]\n"
1044 ".inst 0x6e5aec0a // bfmmla v10.4s, v0.8h, v26.8h\n"
1045 ".inst 0x6e5aec52 // bfmmla v18.4s, v2.8h, v26.8h\n"
1046 "ldr q26, [x10, #0x60]\n"
1047 ".inst 0x6e59ec0e // bfmmla v14.4s, v0.8h, v25.8h\n"
1048 ".inst 0x6e59ec56 // bfmmla v22.4s, v2.8h, v25.8h\n"
1049 "ldr q25, [x10, #0x70]\n"
1050 "add x10, x10, #0x80\n"
1051 ".inst 0x6e5aec0b // bfmmla v11.4s, v0.8h, v26.8h\n"
1052 ".inst 0x6e5aec53 // bfmmla v19.4s, v2.8h, v26.8h\n"
1053 ".inst 0x6e59ec0f // bfmmla v15.4s, v0.8h, v25.8h\n"
1054 ".inst 0x6e59ec57 // bfmmla v23.4s, v2.8h, v25.8h\n"
1056 "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
1057 "add x28, x28, #0x1\n"
1060 "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
1061 "add x25, x9, x20, LSL #2\n"
1062 "add x24, x25, x20, LSL #2\n"
1063 "uzp1 v6.2d, v8.2d, v12.2d\n"
1064 "uzp2 v8.2d, v8.2d, v12.2d\n"
1065 "uzp1 v12.2d, v9.2d, v13.2d\n"
1066 "prfm pstl1keep, [x9, #0x0]\n"
1067 "prfm pstl1keep, [x25, #0x0]\n"
1068 "uzp2 v9.2d, v9.2d, v13.2d\n"
1069 "uzp1 v13.2d, v10.2d, v14.2d\n"
1070 "prfm pstl1keep, [x24, #0x0]\n"
1071 "uzp2 v10.2d, v10.2d, v14.2d\n"
1072 "uzp1 v14.2d, v11.2d, v15.2d\n"
1073 "uzp2 v11.2d, v11.2d, v15.2d\n"
1074 "uzp1 v16.2d, v16.2d, v20.2d\n"
1075 "uzp1 v17.2d, v17.2d, v21.2d\n"
1076 "uzp1 v18.2d, v18.2d, v22.2d\n"
1077 "uzp1 v19.2d, v19.2d, v23.2d\n"
1078 "tbz %x[flags], #1, 95f\n"
1079 "add x20, %x[args_ptr], %[offset_max]\n"
1080 "ld1r { v26.4s }, [x20]\n"
1081 "add x20, %x[args_ptr], %[offset_min]\n"
1082 "ld1r { v25.4s }, [x20]\n"
1083 "fmin v6.4s, v6.4s, v26.4s\n"
1084 "fmin v12.4s, v12.4s, v26.4s\n"
1085 "fmin v13.4s, v13.4s, v26.4s\n"
1086 "fmin v14.4s, v14.4s, v26.4s\n"
1087 "fmin v8.4s, v8.4s, v26.4s\n"
1088 "fmin v9.4s, v9.4s, v26.4s\n"
1089 "fmin v10.4s, v10.4s, v26.4s\n"
1090 "fmin v11.4s, v11.4s, v26.4s\n"
1091 "fmin v16.4s, v16.4s, v26.4s\n"
1092 "fmin v17.4s, v17.4s, v26.4s\n"
1093 "fmin v18.4s, v18.4s, v26.4s\n"
1094 "fmin v19.4s, v19.4s, v26.4s\n"
1095 "fmax v6.4s, v6.4s, v25.4s\n"
1096 "fmax v12.4s, v12.4s, v25.4s\n"
1097 "fmax v13.4s, v13.4s, v25.4s\n"
1098 "fmax v14.4s, v14.4s, v25.4s\n"
1099 "fmax v8.4s, v8.4s, v25.4s\n"
1100 "fmax v9.4s, v9.4s, v25.4s\n"
1101 "fmax v10.4s, v10.4s, v25.4s\n"
1102 "fmax v11.4s, v11.4s, v25.4s\n"
1103 "fmax v16.4s, v16.4s, v25.4s\n"
1104 "fmax v17.4s, v17.4s, v25.4s\n"
1105 "fmax v18.4s, v18.4s, v25.4s\n"
1106 "fmax v19.4s, v19.4s, v25.4s\n"
1110 "tbz x11, #3, 99f\n"
1111 "st1 { v6.4s }, [x9], #0x10\n"
1112 "st1 { v12.4s }, [x9], #0x10\n"
1113 "st1 { v8.4s }, [x25], #0x10\n"
1114 "st1 { v9.4s }, [x25], #0x10\n"
1115 "st1 { v16.4s }, [x24], #0x10\n"
1116 "st1 { v17.4s }, [x24], #0x10\n"
1117 "tbz x11, #2, 97f\n"
1118 "st1 { v13.4s }, [x9], #0x10\n"
1119 "st1 { v10.4s }, [x25], #0x10\n"
1120 "st1 { v18.4s }, [x24], #0x10\n"
1121 "tbz x11, #1, 96f\n"
1122 "str d14, [x9], #0x8\n"
1123 "str d11, [x25], #0x8\n"
1124 "str d19, [x24], #0x8\n"
1125 "tbz x11, #0, 103f\n"
1126 "st1 { v14.s }[2], [x9]\n"
1127 "st1 { v11.s }[2], [x25]\n"
1128 "st1 { v19.s }[2], [x24]\n"
1131 "tbz x11, #0, 103f\n"
1132 "str s14, [x9, #0x0]\n"
1133 "str s11, [x25, #0x0]\n"
1134 "str s19, [x24, #0x0]\n"
1137 "tbz x11, #1, 98f\n"
1138 "str d13, [x9], #0x8\n"
1139 "str d10, [x25], #0x8\n"
1140 "str d18, [x24], #0x8\n"
1141 "tbz x11, #0, 103f\n"
1142 "st1 { v13.s }[2], [x9]\n"
1143 "st1 { v10.s }[2], [x25]\n"
1144 "st1 { v18.s }[2], [x24]\n"
1147 "tbz x11, #0, 103f\n"
1148 "str s13, [x9, #0x0]\n"
1149 "str s10, [x25, #0x0]\n"
1150 "str s18, [x24, #0x0]\n"
1153 "tbz x11, #2, 101f\n"
1154 "st1 { v6.4s }, [x9], #0x10\n"
1155 "st1 { v8.4s }, [x25], #0x10\n"
1156 "st1 { v16.4s }, [x24], #0x10\n"
1157 "tbz x11, #1, 100f\n"
1158 "str d12, [x9], #0x8\n"
1159 "str d9, [x25], #0x8\n"
1160 "str d17, [x24], #0x8\n"
1161 "tbz x11, #0, 103f\n"
1162 "st1 { v12.s }[2], [x9]\n"
1163 "st1 { v9.s }[2], [x25]\n"
1164 "st1 { v17.s }[2], [x24]\n"
1167 "tbz x11, #0, 103f\n"
1168 "str s12, [x9, #0x0]\n"
1169 "str s9, [x25, #0x0]\n"
1170 "str s17, [x24, #0x0]\n"
1173 "tbz x11, #1, 102f\n"
1174 "str d6, [x9], #0x8\n"
1175 "str d8, [x25], #0x8\n"
1176 "str d16, [x24], #0x8\n"
1177 "tbz x11, #0, 103f\n"
1178 "st1 { v6.s }[2], [x9]\n"
1179 "st1 { v8.s }[2], [x25]\n"
1180 "st1 { v16.s }[2], [x24]\n"
1183 "str s6, [x9, #0x0]\n"
1184 "str s8, [x25, #0x0]\n"
1185 "str s16, [x24, #0x0]\n"
1189 "str q6, [x9, #0x0]\n"
1190 "str q12, [x9, #0x10]\n"
1191 "str q13, [x9, #0x20]\n"
1192 "str q14, [x9, #0x30]\n"
1193 "add x9, x9, #0x40\n"
1194 "str q8, [x25, #0x0]\n"
1195 "str q9, [x25, #0x10]\n"
1196 "str q10, [x25, #0x20]\n"
1197 "str q11, [x25, #0x30]\n"
1198 "str q16, [x24, #0x0]\n"
1199 "str q17, [x24, #0x10]\n"
1200 "str q18, [x24, #0x20]\n"
1201 "str q19, [x24, #0x30]\n"
1203 "subs x11, x11, #0x10\n"
1207 "mov x12, %x[bias]\n"
1208 "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
1209 "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
1210 "mov x9, %x[output_ptr]\n"
1213 "ldr q8, [x12, #0x0]\n"
1214 "ldr q9, [x12, #0x10]\n"
1215 "zip2 v12.2d, v8.2d, v8.2d\n"
1216 "zip1 v8.2d, v8.2d, v8.2d\n"
1217 "ldr q10, [x12, #0x20]\n"
1218 "ldr q11, [x12, #0x30]\n"
1219 "zip2 v13.2d, v9.2d, v9.2d\n"
1220 "zip1 v9.2d, v9.2d, v9.2d\n"
1221 "zip2 v14.2d, v10.2d, v10.2d\n"
1222 "zip1 v10.2d, v10.2d, v10.2d\n"
1223 "add x12, x12, #0x40\n"
1224 "zip2 v15.2d, v11.2d, v11.2d\n"
1225 "zip1 v11.2d, v11.2d, v11.2d\n"
1226 "mov v16.16b, v8.16b\n"
1227 "mov v20.16b, v12.16b\n"
1228 "mov v17.16b, v9.16b\n"
1229 "mov v21.16b, v13.16b\n"
1230 "mov v18.16b, v10.16b\n"
1231 "mov v22.16b, v14.16b\n"
1232 "mov v19.16b, v11.16b\n"
1233 "mov v23.16b, v15.16b\n"
1236 "tbz %x[flags], #0, 119f\n"
1237 "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
1238 "add x25, x9, x20, LSL #2\n"
1239 "add x24, x25, x20, LSL #2\n"
1241 "add x23, x24, x20, LSL #2\n"
1243 "tbz x11, #3, 112f\n"
1244 "ld1 { v9.4s }, [x9], #0x10\n"
1245 "ld1 { v12.4s }, [x25], #0x10\n"
1246 "ld1 { v17.4s }, [x24], #0x10\n"
1247 "ld1 { v20.4s }, [x23], #0x10\n"
1248 "ld1 { v10.4s }, [x9], #0x10\n"
1249 "ld1 { v13.4s }, [x25], #0x10\n"
1250 "ld1 { v18.4s }, [x24], #0x10\n"
1251 "ld1 { v21.4s }, [x23], #0x10\n"
1252 "tbz x11, #2, 110f\n"
1253 "ld1 { v11.4s }, [x9], #0x10\n"
1254 "ld1 { v14.4s }, [x25], #0x10\n"
1255 "ld1 { v19.4s }, [x24], #0x10\n"
1256 "ld1 { v22.4s }, [x23], #0x10\n"
1257 "tbz x11, #1, 109f\n"
1258 "ldr d16, [x9], #0x8\n"
1259 "ldr d15, [x25], #0x8\n"
1261 "ldr d24, [x24], #0x8\n"
1262 "ldr d23, [x23], #0x8\n"
1263 "tbz x11, #0, 116f\n"
1264 "ld1 { v16.s }[2], [x9]\n"
1265 "ld1 { v15.s }[2], [x25]\n"
1266 "ld1 { v24.s }[2], [x24]\n"
1267 "ld1 { v23.s }[2], [x23]\n"
1271 "tbz x11, #0, 116f\n"
1272 "ldr s16, [x9, #0x0]\n"
1273 "ldr s15, [x25, #0x0]\n"
1274 "ldr s24, [x24, #0x0]\n"
1275 "ldr s23, [x23, #0x0]\n"
1278 "tbz x11, #1, 111f\n"
1279 "ldr d11, [x9], #0x8\n"
1280 "ldr d14, [x25], #0x8\n"
1282 "ldr d19, [x24], #0x8\n"
1283 "ldr d22, [x23], #0x8\n"
1284 "tbz x11, #0, 116f\n"
1285 "ld1 { v11.s }[2], [x9]\n"
1286 "ld1 { v14.s }[2], [x25]\n"
1287 "ld1 { v19.s }[2], [x24]\n"
1288 "ld1 { v22.s }[2], [x23]\n"
1292 "tbz x11, #0, 116f\n"
1293 "ldr s11, [x9, #0x0]\n"
1294 "ldr s14, [x25, #0x0]\n"
1295 "ldr s19, [x24, #0x0]\n"
1296 "ldr s22, [x23, #0x0]\n"
1299 "tbz x11, #2, 114f\n"
1300 "ld1 { v9.4s }, [x9], #0x10\n"
1301 "ld1 { v12.4s }, [x25], #0x10\n"
1302 "ld1 { v17.4s }, [x24], #0x10\n"
1303 "ld1 { v20.4s }, [x23], #0x10\n"
1304 "tbz x11, #1, 113f\n"
1305 "ldr d10, [x9], #0x8\n"
1306 "ldr d13, [x25], #0x8\n"
1308 "ldr d18, [x24], #0x8\n"
1309 "ldr d21, [x23], #0x8\n"
1310 "tbz x11, #0, 116f\n"
1311 "ld1 { v10.s }[2], [x9]\n"
1312 "ld1 { v13.s }[2], [x25]\n"
1313 "ld1 { v18.s }[2], [x24]\n"
1314 "ld1 { v21.s }[2], [x23]\n"
1318 "tbz x11, #0, 116f\n"
1319 "ldr s10, [x9, #0x0]\n"
1320 "ldr s13, [x25, #0x0]\n"
1321 "ldr s18, [x24, #0x0]\n"
1322 "ldr s21, [x23, #0x0]\n"
1325 "tbz x11, #1, 115f\n"
1326 "ldr d9, [x9], #0x8\n"
1327 "ldr d12, [x25], #0x8\n"
1329 "ldr d17, [x24], #0x8\n"
1330 "ldr d20, [x23], #0x8\n"
1331 "tbz x11, #0, 116f\n"
1332 "ld1 { v9.s }[2], [x9]\n"
1333 "ld1 { v12.s }[2], [x25]\n"
1334 "ld1 { v17.s }[2], [x24]\n"
1335 "ld1 { v20.s }[2], [x23]\n"
1338 "ldr s9, [x9, #0x0]\n"
1339 "ldr s12, [x25, #0x0]\n"
1341 "ldr s17, [x24, #0x0]\n"
1342 "ldr s20, [x23, #0x0]\n"
1347 "ldr q9, [x9, #0x0]\n"
1348 "ldr q10, [x9, #0x10]\n"
1349 "ldr q11, [x9, #0x20]\n"
1350 "ldr q16, [x9, #0x30]\n"
1351 "ldr q12, [x25, #0x0]\n"
1352 "ldr q13, [x25, #0x10]\n"
1353 "ldr q14, [x25, #0x20]\n"
1354 "ldr q15, [x25, #0x30]\n"
1355 "ldr q17, [x24, #0x0]\n"
1356 "ldr q18, [x24, #0x10]\n"
1357 "ldr q19, [x24, #0x20]\n"
1358 "ldr q24, [x24, #0x30]\n"
1359 "ldr q20, [x23, #0x0]\n"
1360 "ldr q21, [x23, #0x10]\n"
1361 "ldr q22, [x23, #0x20]\n"
1362 "ldr q23, [x23, #0x30]\n"
1364 "zip1 v8.2d, v9.2d, v12.2d\n"
1365 "zip2 v12.2d, v9.2d, v12.2d\n"
1366 "zip1 v9.2d, v10.2d, v13.2d\n"
1367 "zip2 v13.2d, v10.2d, v13.2d\n"
1368 "zip1 v10.2d, v11.2d, v14.2d\n"
1369 "zip2 v14.2d, v11.2d, v14.2d\n"
1370 "zip1 v11.2d, v16.2d, v15.2d\n"
1371 "zip2 v15.2d, v16.2d, v15.2d\n"
1372 "zip1 v16.2d, v17.2d, v20.2d\n"
1373 "zip2 v20.2d, v17.2d, v20.2d\n"
1374 "zip1 v17.2d, v18.2d, v21.2d\n"
1375 "zip2 v21.2d, v18.2d, v21.2d\n"
1376 "zip1 v18.2d, v19.2d, v22.2d\n"
1377 "zip2 v22.2d, v19.2d, v22.2d\n"
1378 "zip1 v19.2d, v24.2d, v23.2d\n"
1379 "zip2 v23.2d, v24.2d, v23.2d\n"
1382 "movi v8.16b, #0x0\n"
1383 "movi v9.16b, #0x0\n"
1384 "movi v10.16b, #0x0\n"
1385 "movi v11.16b, #0x0\n"
1386 "movi v12.16b, #0x0\n"
1387 "movi v13.16b, #0x0\n"
1388 "movi v14.16b, #0x0\n"
1389 "movi v15.16b, #0x0\n"
1390 "movi v16.16b, #0x0\n"
1391 "movi v17.16b, #0x0\n"
1392 "movi v18.16b, #0x0\n"
1393 "movi v19.16b, #0x0\n"
1394 "movi v20.16b, #0x0\n"
1395 "movi v21.16b, #0x0\n"
1396 "movi v22.16b, #0x0\n"
1397 "movi v23.16b, #0x0\n"
1401 "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
1402 "ldr w27, [x20, x28, LSL #0x2]\n"
1403 "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
1404 "tbz %x[flags], #3, 122f\n"
1405 "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
1406 "add x20, x20, x21, LSL #3\n"
1407 "ldr x26, [x20, #0x0]\n"
1408 "ldr x25, [x20, #0x8]\n"
1409 "ldr x24, [x20, #0x10]\n"
1410 "ldr x23, [x20, #0x18]\n"
1412 "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
1413 "add x26, x26, x20, LSL #2\n"
1414 "add x25, x25, x20, LSL #2\n"
1415 "add x24, x24, x20, LSL #2\n"
1416 "add x23, x23, x20, LSL #2\n"
1419 "mov x26, %x[input_ptr]\n"
1420 "add x25, x26, x21, LSL #2\n"
1421 "add x24, x25, x21, LSL #2\n"
1422 "add x23, x24, x21, LSL #2\n"
1426 "ld1 { v0.4s }, [x26], #0x10\n"
1427 "ld1 { v2.4s }, [x24], #0x10\n"
1429 "ld1 { v1.4s }, [x25], #0x10\n"
1430 "ld1 { v3.4s }, [x23], #0x10\n"
1431 "ldr q6, [x10, #0x0]\n"
1432 "ldr q7, [x10, #0x10]\n"
1435 ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n"
1436 ".inst 0x0ea16842 // bfcvtn v2.4h, v2.4s\n"
1437 "sub x27, x27, #0x4\n"
1439 ".inst 0x4ea16820 // bfcvtn2 v0.8h, v1.4s\n"
1440 ".inst 0x4ea16862 // bfcvtn2 v2.8h, v3.4s\n"
1441 ".inst 0x6e46ec08 // bfmmla v8.4s, v0.8h, v6.8h\n"
1442 "prfm pldl1keep, [x26, #0x80]\n"
1443 ".inst 0x6e46ec50 // bfmmla v16.4s, v2.8h, v6.8h\n"
1444 "ldr q26, [x10, #0x20]\n"
1445 ".inst 0x6e47ec0c // bfmmla v12.4s, v0.8h, v7.8h\n"
1446 "prfm pldl1keep, [x25, #0x80]\n"
1447 "ld1 { v1.4s }, [x25], #0x10\n"
1448 ".inst 0x6e47ec54 // bfmmla v20.4s, v2.8h, v7.8h\n"
1449 "ldr q25, [x10, #0x30]\n"
1450 ".inst 0x6e5aec09 // bfmmla v9.4s, v0.8h, v26.8h\n"
1451 ".inst 0x6e5aec51 // bfmmla v17.4s, v2.8h, v26.8h\n"
1452 "ldr q26, [x10, #0x40]\n"
1453 ".inst 0x6e59ec0d // bfmmla v13.4s, v0.8h, v25.8h\n"
1454 "prfm pldl1keep, [x24, #0x80]\n"
1455 ".inst 0x6e59ec55 // bfmmla v21.4s, v2.8h, v25.8h\n"
1456 "ldr q25, [x10, #0x50]\n"
1457 ".inst 0x6e5aec0a // bfmmla v10.4s, v0.8h, v26.8h\n"
1458 "prfm pldl1keep, [x23, #0x80]\n"
1459 "ld1 { v3.4s }, [x23], #0x10\n"
1460 ".inst 0x6e5aec52 // bfmmla v18.4s, v2.8h, v26.8h\n"
1461 "ldr q26, [x10, #0x60]\n"
1462 ".inst 0x6e59ec0e // bfmmla v14.4s, v0.8h, v25.8h\n"
1463 ".inst 0x6e59ec56 // bfmmla v22.4s, v2.8h, v25.8h\n"
1464 "ldr q25, [x10, #0x70]\n"
1465 "add x10, x10, #0x80\n"
1466 ".inst 0x6e5aec0b // bfmmla v11.4s, v0.8h, v26.8h\n"
1467 ".inst 0x6e5aec53 // bfmmla v19.4s, v2.8h, v26.8h\n"
1468 "ldr q6, [x10, #0x0]\n"
1469 ".inst 0x6e59ec0f // bfmmla v15.4s, v0.8h, v25.8h\n"
1470 "ld1 { v0.4s }, [x26], #0x10\n"
1471 ".inst 0x6e59ec57 // bfmmla v23.4s, v2.8h, v25.8h\n"
1472 "ld1 { v2.4s }, [x24], #0x10\n"
1473 "ldr q7, [x10, #0x10]\n"
1476 ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n"
1477 ".inst 0x0ea16842 // bfcvtn v2.4h, v2.4s\n"
1478 "sub x27, x27, #0x4\n"
1479 "prfm pldl1keep, [x26, #0x80]\n"
1480 ".inst 0x4ea16820 // bfcvtn2 v0.8h, v1.4s\n"
1481 ".inst 0x4ea16862 // bfcvtn2 v2.8h, v3.4s\n"
1482 ".inst 0x6e46ec08 // bfmmla v8.4s, v0.8h, v6.8h\n"
1483 "prfm pldl1keep, [x25, #0x80]\n"
1484 ".inst 0x6e46ec50 // bfmmla v16.4s, v2.8h, v6.8h\n"
1485 "ldr q26, [x10, #0x20]\n"
1486 ".inst 0x6e47ec0c // bfmmla v12.4s, v0.8h, v7.8h\n"
1487 "prfm pldl1keep, [x24, #0x80]\n"
1488 ".inst 0x6e47ec54 // bfmmla v20.4s, v2.8h, v7.8h\n"
1489 "ldr q25, [x10, #0x30]\n"
1490 ".inst 0x6e5aec09 // bfmmla v9.4s, v0.8h, v26.8h\n"
1491 "prfm pldl1keep, [x23, #0x80]\n"
1492 ".inst 0x6e5aec51 // bfmmla v17.4s, v2.8h, v26.8h\n"
1493 "ldr q26, [x10, #0x40]\n"
1494 ".inst 0x6e59ec0d // bfmmla v13.4s, v0.8h, v25.8h\n"
1495 ".inst 0x6e59ec55 // bfmmla v21.4s, v2.8h, v25.8h\n"
1496 "ldr q25, [x10, #0x50]\n"
1497 ".inst 0x6e5aec0a // bfmmla v10.4s, v0.8h, v26.8h\n"
1498 ".inst 0x6e5aec52 // bfmmla v18.4s, v2.8h, v26.8h\n"
1499 "ldr q26, [x10, #0x60]\n"
1500 ".inst 0x6e59ec0e // bfmmla v14.4s, v0.8h, v25.8h\n"
1501 ".inst 0x6e59ec56 // bfmmla v22.4s, v2.8h, v25.8h\n"
1502 "ldr q25, [x10, #0x70]\n"
1503 "add x10, x10, #0x80\n"
1504 ".inst 0x6e5aec0b // bfmmla v11.4s, v0.8h, v26.8h\n"
1505 ".inst 0x6e5aec53 // bfmmla v19.4s, v2.8h, v26.8h\n"
1506 ".inst 0x6e59ec0f // bfmmla v15.4s, v0.8h, v25.8h\n"
1507 ".inst 0x6e59ec57 // bfmmla v23.4s, v2.8h, v25.8h\n"
1511 "tbz x27, #1, 127f\n"
1512 "ldr d0, [x26], #0x8\n"
1513 "ldr d1, [x25], #0x8\n"
1514 "ldr d2, [x24], #0x8\n"
1515 "ldr d3, [x23], #0x8\n"
1516 "tbz x27, #0, 128f\n"
1517 "ld1 { v0.s }[2], [x26]\n"
1518 "ld1 { v1.s }[2], [x25]\n"
1519 "ld1 { v2.s }[2], [x24]\n"
1520 "ld1 { v3.s }[2], [x23]\n"
1523 "ldr s0, [x26, #0x0]\n"
1524 "ldr s1, [x25, #0x0]\n"
1525 "ldr s2, [x24, #0x0]\n"
1526 "ldr s3, [x23, #0x0]\n"
1528 "ldr q26, [x10, #0x0]\n"
1529 "ldr q25, [x10, #0x10]\n"
1530 ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n"
1531 ".inst 0x0ea16842 // bfcvtn v2.4h, v2.4s\n"
1532 ".inst 0x4ea16820 // bfcvtn2 v0.8h, v1.4s\n"
1533 ".inst 0x4ea16862 // bfcvtn2 v2.8h, v3.4s\n"
1534 ".inst 0x6e5aec08 // bfmmla v8.4s, v0.8h, v26.8h\n"
1535 ".inst 0x6e5aec50 // bfmmla v16.4s, v2.8h, v26.8h\n"
1536 "ldr q26, [x10, #0x20]\n"
1537 ".inst 0x6e59ec0c // bfmmla v12.4s, v0.8h, v25.8h\n"
1538 ".inst 0x6e59ec54 // bfmmla v20.4s, v2.8h, v25.8h\n"
1539 "ldr q25, [x10, #0x30]\n"
1540 ".inst 0x6e5aec09 // bfmmla v9.4s, v0.8h, v26.8h\n"
1541 ".inst 0x6e5aec51 // bfmmla v17.4s, v2.8h, v26.8h\n"
1542 "ldr q26, [x10, #0x40]\n"
1543 ".inst 0x6e59ec0d // bfmmla v13.4s, v0.8h, v25.8h\n"
1544 ".inst 0x6e59ec55 // bfmmla v21.4s, v2.8h, v25.8h\n"
1545 "ldr q25, [x10, #0x50]\n"
1546 ".inst 0x6e5aec0a // bfmmla v10.4s, v0.8h, v26.8h\n"
1547 ".inst 0x6e5aec52 // bfmmla v18.4s, v2.8h, v26.8h\n"
1548 "ldr q26, [x10, #0x60]\n"
1549 ".inst 0x6e59ec0e // bfmmla v14.4s, v0.8h, v25.8h\n"
1550 ".inst 0x6e59ec56 // bfmmla v22.4s, v2.8h, v25.8h\n"
1551 "ldr q25, [x10, #0x70]\n"
1552 ".inst 0x6e5aec0b // bfmmla v11.4s, v0.8h, v26.8h\n"
1553 "add x10, x10, #0x80\n"
1554 ".inst 0x6e5aec53 // bfmmla v19.4s, v2.8h, v26.8h\n"
1555 ".inst 0x6e59ec0f // bfmmla v15.4s, v0.8h, v25.8h\n"
1556 ".inst 0x6e59ec57 // bfmmla v23.4s, v2.8h, v25.8h\n"
1558 "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
1559 "add x28, x28, #0x1\n"
1562 "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
1563 "add x25, x9, x20, LSL #2\n"
1564 "add x24, x25, x20, LSL #2\n"
1565 "uzp1 v6.2d, v8.2d, v12.2d\n"
1566 "add x23, x24, x20, LSL #2\n"
1567 "uzp2 v8.2d, v8.2d, v12.2d\n"
1568 "uzp1 v12.2d, v9.2d, v13.2d\n"
1569 "prfm pstl1keep, [x9, #0x0]\n"
1570 "uzp2 v9.2d, v9.2d, v13.2d\n"
1571 "uzp1 v13.2d, v10.2d, v14.2d\n"
1572 "prfm pstl1keep, [x25, #0x0]\n"
1573 "prfm pstl1keep, [x24, #0x0]\n"
1574 "uzp2 v10.2d, v10.2d, v14.2d\n"
1575 "uzp1 v14.2d, v11.2d, v15.2d\n"
1576 "prfm pstl1keep, [x23, #0x0]\n"
1577 "uzp2 v11.2d, v11.2d, v15.2d\n"
1578 "uzp1 v15.2d, v16.2d, v20.2d\n"
1579 "uzp2 v16.2d, v16.2d, v20.2d\n"
1580 "uzp1 v20.2d, v17.2d, v21.2d\n"
1581 "uzp2 v17.2d, v17.2d, v21.2d\n"
1582 "uzp1 v21.2d, v18.2d, v22.2d\n"
1583 "uzp2 v18.2d, v18.2d, v22.2d\n"
1584 "uzp1 v22.2d, v19.2d, v23.2d\n"
1585 "uzp2 v19.2d, v19.2d, v23.2d\n"
1586 "tbz %x[flags], #1, 130f\n"
1587 "add x20, %x[args_ptr], %[offset_max]\n"
1588 "ld1r { v26.4s }, [x20]\n"
1589 "add x20, %x[args_ptr], %[offset_min]\n"
1590 "ld1r { v25.4s }, [x20]\n"
1591 "fmin v6.4s, v6.4s, v26.4s\n"
1592 "fmin v12.4s, v12.4s, v26.4s\n"
1593 "fmin v13.4s, v13.4s, v26.4s\n"
1594 "fmin v14.4s, v14.4s, v26.4s\n"
1595 "fmin v8.4s, v8.4s, v26.4s\n"
1596 "fmin v9.4s, v9.4s, v26.4s\n"
1597 "fmin v10.4s, v10.4s, v26.4s\n"
1598 "fmin v11.4s, v11.4s, v26.4s\n"
1599 "fmin v15.4s, v15.4s, v26.4s\n"
1600 "fmin v20.4s, v20.4s, v26.4s\n"
1601 "fmin v21.4s, v21.4s, v26.4s\n"
1602 "fmin v22.4s, v22.4s, v26.4s\n"
1603 "fmin v16.4s, v16.4s, v26.4s\n"
1604 "fmin v17.4s, v17.4s, v26.4s\n"
1605 "fmin v18.4s, v18.4s, v26.4s\n"
1606 "fmin v19.4s, v19.4s, v26.4s\n"
1607 "fmax v6.4s, v6.4s, v25.4s\n"
1608 "fmax v12.4s, v12.4s, v25.4s\n"
1609 "fmax v13.4s, v13.4s, v25.4s\n"
1610 "fmax v14.4s, v14.4s, v25.4s\n"
1611 "fmax v8.4s, v8.4s, v25.4s\n"
1612 "fmax v9.4s, v9.4s, v25.4s\n"
1613 "fmax v10.4s, v10.4s, v25.4s\n"
1614 "fmax v11.4s, v11.4s, v25.4s\n"
1615 "fmax v15.4s, v15.4s, v25.4s\n"
1616 "fmax v20.4s, v20.4s, v25.4s\n"
1617 "fmax v21.4s, v21.4s, v25.4s\n"
1618 "fmax v22.4s, v22.4s, v25.4s\n"
1619 "fmax v16.4s, v16.4s, v25.4s\n"
1620 "fmax v17.4s, v17.4s, v25.4s\n"
1621 "fmax v18.4s, v18.4s, v25.4s\n"
1622 "fmax v19.4s, v19.4s, v25.4s\n"
1626 "tbz x11, #3, 134f\n"
1627 "st1 { v6.4s }, [x9], #0x10\n"
1628 "st1 { v12.4s }, [x9], #0x10\n"
1629 "st1 { v8.4s }, [x25], #0x10\n"
1630 "st1 { v9.4s }, [x25], #0x10\n"
1631 "st1 { v15.4s }, [x24], #0x10\n"
1632 "st1 { v20.4s }, [x24], #0x10\n"
1633 "st1 { v16.4s }, [x23], #0x10\n"
1634 "st1 { v17.4s }, [x23], #0x10\n"
1635 "tbz x11, #2, 132f\n"
1636 "st1 { v13.4s }, [x9], #0x10\n"
1637 "st1 { v10.4s }, [x25], #0x10\n"
1638 "st1 { v21.4s }, [x24], #0x10\n"
1639 "st1 { v18.4s }, [x23], #0x10\n"
1640 "tbz x11, #1, 131f\n"
1641 "str d14, [x9], #0x8\n"
1642 "str d11, [x25], #0x8\n"
1643 "str d22, [x24], #0x8\n"
1644 "str d19, [x23], #0x8\n"
1645 "tbz x11, #0, 138f\n"
1646 "st1 { v14.s }[2], [x9]\n"
1647 "st1 { v11.s }[2], [x25]\n"
1648 "st1 { v22.s }[2], [x24]\n"
1649 "st1 { v19.s }[2], [x23]\n"
1652 "tbz x11, #0, 138f\n"
1653 "str s14, [x9, #0x0]\n"
1654 "str s11, [x25, #0x0]\n"
1655 "str s22, [x24, #0x0]\n"
1656 "str s19, [x23, #0x0]\n"
1659 "tbz x11, #1, 133f\n"
1660 "str d13, [x9], #0x8\n"
1661 "str d10, [x25], #0x8\n"
1662 "str d21, [x24], #0x8\n"
1663 "str d18, [x23], #0x8\n"
1664 "tbz x11, #0, 138f\n"
1665 "st1 { v13.s }[2], [x9]\n"
1666 "st1 { v10.s }[2], [x25]\n"
1667 "st1 { v21.s }[2], [x24]\n"
1668 "st1 { v18.s }[2], [x23]\n"
1671 "tbz x11, #0, 138f\n"
1672 "str s13, [x9, #0x0]\n"
1673 "str s10, [x25, #0x0]\n"
1674 "str s21, [x24, #0x0]\n"
1675 "str s18, [x23, #0x0]\n"
1678 "tbz x11, #2, 136f\n"
1679 "st1 { v6.4s }, [x9], #0x10\n"
1680 "st1 { v8.4s }, [x25], #0x10\n"
1681 "st1 { v15.4s }, [x24], #0x10\n"
1682 "st1 { v16.4s }, [x23], #0x10\n"
1683 "tbz x11, #1, 135f\n"
1684 "str d12, [x9], #0x8\n"
1685 "str d9, [x25], #0x8\n"
1686 "str d20, [x24], #0x8\n"
1687 "str d17, [x23], #0x8\n"
1688 "tbz x11, #0, 138f\n"
1689 "st1 { v12.s }[2], [x9]\n"
1690 "st1 { v9.s }[2], [x25]\n"
1691 "st1 { v20.s }[2], [x24]\n"
1692 "st1 { v17.s }[2], [x23]\n"
1695 "tbz x11, #0, 138f\n"
1696 "str s12, [x9, #0x0]\n"
1697 "str s9, [x25, #0x0]\n"
1698 "str s20, [x24, #0x0]\n"
1699 "str s17, [x23, #0x0]\n"
1702 "tbz x11, #1, 137f\n"
1703 "str d6, [x9], #0x8\n"
1704 "str d8, [x25], #0x8\n"
1705 "str d15, [x24], #0x8\n"
1706 "str d16, [x23], #0x8\n"
1707 "tbz x11, #0, 138f\n"
1708 "st1 { v6.s }[2], [x9]\n"
1709 "st1 { v8.s }[2], [x25]\n"
1710 "st1 { v15.s }[2], [x24]\n"
1711 "st1 { v16.s }[2], [x23]\n"
1714 "str s6, [x9, #0x0]\n"
1715 "str s8, [x25, #0x0]\n"
1716 "str s15, [x24, #0x0]\n"
1717 "str s16, [x23, #0x0]\n"
1721 "str q6, [x9, #0x0]\n"
1722 "str q12, [x9, #0x10]\n"
1723 "str q13, [x9, #0x20]\n"
1724 "str q14, [x9, #0x30]\n"
1725 "add x9, x9, #0x40\n"
1726 "str q8, [x25, #0x0]\n"
1727 "str q9, [x25, #0x10]\n"
1728 "str q10, [x25, #0x20]\n"
1729 "str q11, [x25, #0x30]\n"
1730 "str q15, [x24, #0x0]\n"
1731 "str q20, [x24, #0x10]\n"
1732 "str q21, [x24, #0x20]\n"
1733 "str q22, [x24, #0x30]\n"
1734 "str q16, [x23, #0x0]\n"
1735 "str q17, [x23, #0x10]\n"
1736 "str q18, [x23, #0x20]\n"
1737 "str q19, [x23, #0x30]\n"
1739 "subs x11, x11, #0x10\n"
1743 "mov x12, %x[bias]\n"
1744 "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
1745 "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
1746 "mov x9, %x[output_ptr]\n"
1749 "ldr q8, [x12, #0x0]\n"
1750 "ldr q9, [x12, #0x10]\n"
1751 "zip2 v12.2d, v8.2d, v8.2d\n"
1752 "zip1 v8.2d, v8.2d, v8.2d\n"
1753 "ldr q10, [x12, #0x20]\n"
1754 "ldr q11, [x12, #0x30]\n"
1755 "zip2 v13.2d, v9.2d, v9.2d\n"
1756 "zip1 v9.2d, v9.2d, v9.2d\n"
1757 "zip2 v14.2d, v10.2d, v10.2d\n"
1758 "zip1 v10.2d, v10.2d, v10.2d\n"
1759 "add x12, x12, #0x40\n"
1760 "zip2 v15.2d, v11.2d, v11.2d\n"
1761 "zip1 v11.2d, v11.2d, v11.2d\n"
1762 "mov v16.16b, v8.16b\n"
1763 "mov v20.16b, v12.16b\n"
1764 "mov v17.16b, v9.16b\n"
1765 "mov v21.16b, v13.16b\n"
1766 "mov v18.16b, v10.16b\n"
1767 "mov v22.16b, v14.16b\n"
1768 "mov v19.16b, v11.16b\n"
1769 "mov v23.16b, v15.16b\n"
1770 "mov v24.16b, v8.16b\n"
1771 "mov v28.16b, v12.16b\n"
1772 "mov v25.16b, v9.16b\n"
1773 "mov v29.16b, v13.16b\n"
1774 "mov v26.16b, v10.16b\n"
1775 "mov v30.16b, v14.16b\n"
1776 "mov v27.16b, v11.16b\n"
1777 "mov v31.16b, v15.16b\n"
1780 "tbz %x[flags], #0, 154f\n"
1781 "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
1782 "add x25, x9, x20, LSL #2\n"
1783 "add x24, x25, x20, LSL #2\n"
1784 "add x23, x24, x20, LSL #2\n"
1786 "add x22, x23, x20, LSL #2\n"
1788 "tbz x11, #3, 147f\n"
1789 "ld1 { v9.4s }, [x9], #0x10\n"
1790 "ld1 { v12.4s }, [x25], #0x10\n"
1791 "ld1 { v17.4s }, [x24], #0x10\n"
1792 "ld1 { v20.4s }, [x23], #0x10\n"
1793 "ld1 { v25.4s }, [x22], #0x10\n"
1794 "ld1 { v10.4s }, [x9], #0x10\n"
1795 "ld1 { v13.4s }, [x25], #0x10\n"
1796 "ld1 { v18.4s }, [x24], #0x10\n"
1797 "ld1 { v21.4s }, [x23], #0x10\n"
1798 "ld1 { v26.4s }, [x22], #0x10\n"
1799 "tbz x11, #2, 145f\n"
1800 "ld1 { v11.4s }, [x9], #0x10\n"
1801 "ld1 { v14.4s }, [x25], #0x10\n"
1802 "ld1 { v19.4s }, [x24], #0x10\n"
1803 "ld1 { v22.4s }, [x23], #0x10\n"
1804 "ld1 { v27.4s }, [x22], #0x10\n"
1805 "tbz x11, #1, 144f\n"
1806 "ldr d16, [x9], #0x8\n"
1807 "ldr d15, [x25], #0x8\n"
1809 "ldr d24, [x24], #0x8\n"
1810 "ldr d23, [x23], #0x8\n"
1811 "ldr d6, [x22], #0x8\n"
1812 "tbz x11, #0, 151f\n"
1813 "ld1 { v16.s }[2], [x9]\n"
1814 "ld1 { v15.s }[2], [x25]\n"
1815 "ld1 { v24.s }[2], [x24]\n"
1816 "ld1 { v23.s }[2], [x23]\n"
1817 "ld1 { v6.s }[2], [x22]\n"
1821 "tbz x11, #0, 151f\n"
1822 "ldr s16, [x9, #0x0]\n"
1823 "ldr s15, [x25, #0x0]\n"
1824 "ldr s24, [x24, #0x0]\n"
1825 "ldr s23, [x23, #0x0]\n"
1826 "ldr s6, [x22, #0x0]\n"
1829 "tbz x11, #1, 146f\n"
1830 "ldr d11, [x9], #0x8\n"
1831 "ldr d14, [x25], #0x8\n"
1833 "ldr d19, [x24], #0x8\n"
1834 "ldr d22, [x23], #0x8\n"
1835 "ldr d27, [x22], #0x8\n"
1836 "tbz x11, #0, 151f\n"
1837 "ld1 { v11.s }[2], [x9]\n"
1838 "ld1 { v14.s }[2], [x25]\n"
1839 "ld1 { v19.s }[2], [x24]\n"
1840 "ld1 { v22.s }[2], [x23]\n"
1841 "ld1 { v27.s }[2], [x22]\n"
1845 "tbz x11, #0, 151f\n"
1846 "ldr s11, [x9, #0x0]\n"
1847 "ldr s14, [x25, #0x0]\n"
1848 "ldr s19, [x24, #0x0]\n"
1849 "ldr s22, [x23, #0x0]\n"
1850 "ldr s27, [x22, #0x0]\n"
1853 "tbz x11, #2, 149f\n"
1854 "ld1 { v9.4s }, [x9], #0x10\n"
1855 "ld1 { v12.4s }, [x25], #0x10\n"
1856 "ld1 { v17.4s }, [x24], #0x10\n"
1857 "ld1 { v20.4s }, [x23], #0x10\n"
1858 "ld1 { v25.4s }, [x22], #0x10\n"
1859 "tbz x11, #1, 148f\n"
1860 "ldr d10, [x9], #0x8\n"
1861 "ldr d13, [x25], #0x8\n"
1863 "ldr d18, [x24], #0x8\n"
1864 "ldr d21, [x23], #0x8\n"
1865 "ldr d26, [x22], #0x8\n"
1866 "tbz x11, #0, 151f\n"
1867 "ld1 { v10.s }[2], [x9]\n"
1868 "ld1 { v13.s }[2], [x25]\n"
1869 "ld1 { v18.s }[2], [x24]\n"
1870 "ld1 { v21.s }[2], [x23]\n"
1871 "ld1 { v26.s }[2], [x22]\n"
1875 "tbz x11, #0, 151f\n"
1876 "ldr s10, [x9, #0x0]\n"
1877 "ldr s13, [x25, #0x0]\n"
1878 "ldr s18, [x24, #0x0]\n"
1879 "ldr s21, [x23, #0x0]\n"
1880 "ldr s26, [x22, #0x0]\n"
1883 "tbz x11, #1, 150f\n"
1884 "ldr d9, [x9], #0x8\n"
1885 "ldr d12, [x25], #0x8\n"
1887 "ldr d17, [x24], #0x8\n"
1888 "ldr d20, [x23], #0x8\n"
1889 "ldr d25, [x22], #0x8\n"
1890 "tbz x11, #0, 151f\n"
1891 "ld1 { v9.s }[2], [x9]\n"
1892 "ld1 { v12.s }[2], [x25]\n"
1893 "ld1 { v17.s }[2], [x24]\n"
1894 "ld1 { v20.s }[2], [x23]\n"
1895 "ld1 { v25.s }[2], [x22]\n"
1898 "ldr s9, [x9, #0x0]\n"
1899 "ldr s12, [x25, #0x0]\n"
1901 "ldr s17, [x24, #0x0]\n"
1902 "ldr s20, [x23, #0x0]\n"
1903 "ldr s25, [x22, #0x0]\n"
1908 "ldr q9, [x9, #0x0]\n"
1909 "ldr q10, [x9, #0x10]\n"
1910 "ldr q11, [x9, #0x20]\n"
1911 "ldr q16, [x9, #0x30]\n"
1912 "ldr q12, [x25, #0x0]\n"
1913 "ldr q13, [x25, #0x10]\n"
1914 "ldr q14, [x25, #0x20]\n"
1915 "ldr q15, [x25, #0x30]\n"
1916 "ldr q17, [x24, #0x0]\n"
1917 "ldr q18, [x24, #0x10]\n"
1918 "ldr q19, [x24, #0x20]\n"
1919 "ldr q24, [x24, #0x30]\n"
1920 "ldr q20, [x23, #0x0]\n"
1921 "ldr q21, [x23, #0x10]\n"
1922 "ldr q22, [x23, #0x20]\n"
1923 "ldr q23, [x23, #0x30]\n"
1924 "ldr q25, [x22, #0x0]\n"
1925 "ldr q26, [x22, #0x10]\n"
1926 "ldr q27, [x22, #0x20]\n"
1927 "ldr q6, [x22, #0x30]\n"
1929 "zip1 v8.2d, v9.2d, v12.2d\n"
1930 "zip2 v12.2d, v9.2d, v12.2d\n"
1931 "zip1 v9.2d, v10.2d, v13.2d\n"
1932 "zip2 v13.2d, v10.2d, v13.2d\n"
1933 "zip1 v10.2d, v11.2d, v14.2d\n"
1934 "zip2 v14.2d, v11.2d, v14.2d\n"
1935 "zip1 v11.2d, v16.2d, v15.2d\n"
1936 "zip2 v15.2d, v16.2d, v15.2d\n"
1937 "zip1 v16.2d, v17.2d, v20.2d\n"
1938 "zip2 v20.2d, v17.2d, v20.2d\n"
1939 "zip1 v17.2d, v18.2d, v21.2d\n"
1940 "zip2 v21.2d, v18.2d, v21.2d\n"
1941 "zip1 v18.2d, v19.2d, v22.2d\n"
1942 "zip2 v22.2d, v19.2d, v22.2d\n"
1943 "zip1 v19.2d, v24.2d, v23.2d\n"
1944 "zip2 v23.2d, v24.2d, v23.2d\n"
1945 "zip1 v24.2d, v25.2d, v28.2d\n"
1946 "zip2 v28.2d, v25.2d, v28.2d\n"
1947 "zip1 v25.2d, v26.2d, v29.2d\n"
1948 "zip2 v29.2d, v26.2d, v29.2d\n"
1949 "zip1 v26.2d, v27.2d, v30.2d\n"
1950 "zip2 v30.2d, v27.2d, v30.2d\n"
1951 "zip1 v27.2d, v6.2d, v31.2d\n"
1952 "zip2 v31.2d, v6.2d, v31.2d\n"
1955 "movi v8.16b, #0x0\n"
1956 "movi v9.16b, #0x0\n"
1957 "movi v10.16b, #0x0\n"
1958 "movi v11.16b, #0x0\n"
1959 "movi v12.16b, #0x0\n"
1960 "movi v13.16b, #0x0\n"
1961 "movi v14.16b, #0x0\n"
1962 "movi v15.16b, #0x0\n"
1963 "movi v16.16b, #0x0\n"
1964 "movi v17.16b, #0x0\n"
1965 "movi v18.16b, #0x0\n"
1966 "movi v19.16b, #0x0\n"
1967 "movi v20.16b, #0x0\n"
1968 "movi v21.16b, #0x0\n"
1969 "movi v22.16b, #0x0\n"
1970 "movi v23.16b, #0x0\n"
1971 "movi v24.16b, #0x0\n"
1972 "movi v25.16b, #0x0\n"
1973 "movi v26.16b, #0x0\n"
1974 "movi v27.16b, #0x0\n"
1975 "movi v28.16b, #0x0\n"
1976 "movi v29.16b, #0x0\n"
1977 "movi v30.16b, #0x0\n"
1978 "movi v31.16b, #0x0\n"
1982 "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
1983 "ldr w27, [x20, x28, LSL #0x2]\n"
1984 "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
1985 "tbz %x[flags], #3, 157f\n"
1986 "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
1987 "add x20, x20, x21, LSL #3\n"
1988 "ldr x26, [x20, #0x0]\n"
1989 "ldr x25, [x20, #0x8]\n"
1990 "ldr x24, [x20, #0x10]\n"
1991 "ldr x23, [x20, #0x18]\n"
1992 "ldr x22, [x20, #0x20]\n"
1994 "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
1995 "add x26, x26, x20, LSL #2\n"
1996 "add x25, x25, x20, LSL #2\n"
1997 "add x24, x24, x20, LSL #2\n"
1998 "add x23, x23, x20, LSL #2\n"
1999 "add x22, x22, x20, LSL #2\n"
2002 "mov x26, %x[input_ptr]\n"
2003 "add x25, x26, x21, LSL #2\n"
2004 "add x24, x25, x21, LSL #2\n"
2005 "add x23, x24, x21, LSL #2\n"
2006 "add x22, x23, x21, LSL #2\n"
2010 "ld1 { v0.4s }, [x26], #0x10\n"
2011 "ld1 { v2.4s }, [x24], #0x10\n"
2013 "ld1 { v1.4s }, [x25], #0x10\n"
2014 "ld1 { v3.4s }, [x23], #0x10\n"
2015 "ld1 { v4.4s }, [x22], #0x10\n"
2016 "ldr q6, [x10, #0x0]\n"
2017 "ldr q7, [x10, #0x10]\n"
2020 ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n"
2021 ".inst 0x0ea16842 // bfcvtn v2.4h, v2.4s\n"
2022 "sub x27, x27, #0x4\n"
2024 ".inst 0x4ea16820 // bfcvtn2 v0.8h, v1.4s\n"
2025 ".inst 0x4ea16862 // bfcvtn2 v2.8h, v3.4s\n"
2026 ".inst 0x6e46ec08 // bfmmla v8.4s, v0.8h, v6.8h\n"
2027 "prfm pldl1keep, [x26, #0x80]\n"
2028 ".inst 0x0ea16884 // bfcvtn v4.4h, v4.4s\n"
2029 ".inst 0x6e46ec50 // bfmmla v16.4s, v2.8h, v6.8h\n"
2030 ".inst 0x6e46ec98 // bfmmla v24.4s, v4.8h, v6.8h\n"
2031 "ldr q3, [x10, #0x20]\n"
2032 ".inst 0x6e47ec0c // bfmmla v12.4s, v0.8h, v7.8h\n"
2033 ".inst 0x6e47ec54 // bfmmla v20.4s, v2.8h, v7.8h\n"
2034 "prfm pldl1keep, [x25, #0x80]\n"
2035 "ld1 { v1.4s }, [x25], #0x10\n"
2036 ".inst 0x6e47ec9c // bfmmla v28.4s, v4.8h, v7.8h\n"
2037 "ldr q5, [x10, #0x30]\n"
2038 ".inst 0x6e43ec09 // bfmmla v9.4s, v0.8h, v3.8h\n"
2039 "prfm pldl1keep, [x24, #0x80]\n"
2040 ".inst 0x6e43ec51 // bfmmla v17.4s, v2.8h, v3.8h\n"
2041 ".inst 0x6e43ec99 // bfmmla v25.4s, v4.8h, v3.8h\n"
2042 "ldr q6, [x10, #0x40]\n"
2043 "prfm pldl1keep, [x23, #0x80]\n"
2044 "ld1 { v3.4s }, [x23], #0x10\n"
2045 ".inst 0x6e45ec0d // bfmmla v13.4s, v0.8h, v5.8h\n"
2046 ".inst 0x6e45ec55 // bfmmla v21.4s, v2.8h, v5.8h\n"
2047 "prfm pldl1keep, [x22, #0x80]\n"
2048 ".inst 0x6e45ec9d // bfmmla v29.4s, v4.8h, v5.8h\n"
2049 "ldr q5, [x10, #0x50]\n"
2050 ".inst 0x6e46ec0a // bfmmla v10.4s, v0.8h, v6.8h\n"
2051 ".inst 0x6e46ec52 // bfmmla v18.4s, v2.8h, v6.8h\n"
2052 ".inst 0x6e46ec9a // bfmmla v26.4s, v4.8h, v6.8h\n"
2053 "ldr q6, [x10, #0x60]\n"
2054 ".inst 0x6e45ec0e // bfmmla v14.4s, v0.8h, v5.8h\n"
2055 ".inst 0x6e45ec56 // bfmmla v22.4s, v2.8h, v5.8h\n"
2056 ".inst 0x6e45ec9e // bfmmla v30.4s, v4.8h, v5.8h\n"
2057 "ldr q5, [x10, #0x70]\n"
2058 "add x10, x10, #0x80\n"
2059 ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n"
2060 ".inst 0x6e46ec53 // bfmmla v19.4s, v2.8h, v6.8h\n"
2061 ".inst 0x6e46ec9b // bfmmla v27.4s, v4.8h, v6.8h\n"
2062 "ldr q6, [x10, #0x0]\n"
2063 ".inst 0x6e45ec0f // bfmmla v15.4s, v0.8h, v5.8h\n"
2064 "ld1 { v0.4s }, [x26], #0x10\n"
2065 ".inst 0x6e45ec57 // bfmmla v23.4s, v2.8h, v5.8h\n"
2066 "ld1 { v2.4s }, [x24], #0x10\n"
2067 ".inst 0x6e45ec9f // bfmmla v31.4s, v4.8h, v5.8h\n"
2068 "ld1 { v4.4s }, [x22], #0x10\n"
2069 "ldr q7, [x10, #0x10]\n"
2072 ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n"
2073 ".inst 0x0ea16842 // bfcvtn v2.4h, v2.4s\n"
2074 "sub x27, x27, #0x4\n"
2075 "prfm pldl1keep, [x26, #0x80]\n"
2076 ".inst 0x4ea16820 // bfcvtn2 v0.8h, v1.4s\n"
2077 ".inst 0x4ea16862 // bfcvtn2 v2.8h, v3.4s\n"
2078 ".inst 0x6e46ec08 // bfmmla v8.4s, v0.8h, v6.8h\n"
2079 "prfm pldl1keep, [x25, #0x80]\n"
2080 ".inst 0x0ea16884 // bfcvtn v4.4h, v4.4s\n"
2081 ".inst 0x6e46ec50 // bfmmla v16.4s, v2.8h, v6.8h\n"
2082 ".inst 0x6e46ec98 // bfmmla v24.4s, v4.8h, v6.8h\n"
2083 "ldr q3, [x10, #0x20]\n"
2084 ".inst 0x6e47ec0c // bfmmla v12.4s, v0.8h, v7.8h\n"
2085 ".inst 0x6e47ec54 // bfmmla v20.4s, v2.8h, v7.8h\n"
2086 "prfm pldl1keep, [x24, #0x80]\n"
2087 "prfm pldl1keep, [x23, #0x80]\n"
2088 ".inst 0x6e47ec9c // bfmmla v28.4s, v4.8h, v7.8h\n"
2089 "ldr q1, [x10, #0x30]\n"
2090 ".inst 0x6e43ec09 // bfmmla v9.4s, v0.8h, v3.8h\n"
2091 "prfm pldl1keep, [x22, #0x80]\n"
2092 ".inst 0x6e43ec51 // bfmmla v17.4s, v2.8h, v3.8h\n"
2093 ".inst 0x6e43ec99 // bfmmla v25.4s, v4.8h, v3.8h\n"
2094 "ldr q3, [x10, #0x40]\n"
2095 ".inst 0x6e41ec0d // bfmmla v13.4s, v0.8h, v1.8h\n"
2096 ".inst 0x6e41ec55 // bfmmla v21.4s, v2.8h, v1.8h\n"
2097 ".inst 0x6e41ec9d // bfmmla v29.4s, v4.8h, v1.8h\n"
2098 "ldr q1, [x10, #0x50]\n"
2099 ".inst 0x6e43ec0a // bfmmla v10.4s, v0.8h, v3.8h\n"
2100 ".inst 0x6e43ec52 // bfmmla v18.4s, v2.8h, v3.8h\n"
2101 ".inst 0x6e43ec9a // bfmmla v26.4s, v4.8h, v3.8h\n"
2102 "ldr q3, [x10, #0x60]\n"
2103 ".inst 0x6e41ec0e // bfmmla v14.4s, v0.8h, v1.8h\n"
2104 ".inst 0x6e41ec56 // bfmmla v22.4s, v2.8h, v1.8h\n"
2105 ".inst 0x6e41ec9e // bfmmla v30.4s, v4.8h, v1.8h\n"
2106 "ldr q1, [x10, #0x70]\n"
2107 "add x10, x10, #0x80\n"
2108 ".inst 0x6e43ec0b // bfmmla v11.4s, v0.8h, v3.8h\n"
2109 ".inst 0x6e43ec53 // bfmmla v19.4s, v2.8h, v3.8h\n"
2110 ".inst 0x6e43ec9b // bfmmla v27.4s, v4.8h, v3.8h\n"
2111 ".inst 0x6e41ec0f // bfmmla v15.4s, v0.8h, v1.8h\n"
2112 ".inst 0x6e41ec57 // bfmmla v23.4s, v2.8h, v1.8h\n"
2113 ".inst 0x6e41ec9f // bfmmla v31.4s, v4.8h, v1.8h\n"
2117 "tbz x27, #1, 162f\n"
2118 "ldr d0, [x26], #0x8\n"
2119 "ldr d1, [x25], #0x8\n"
2120 "ldr d2, [x24], #0x8\n"
2121 "ldr d3, [x23], #0x8\n"
2122 "ldr d4, [x22], #0x8\n"
2123 "tbz x27, #0, 163f\n"
2124 "ld1 { v0.s }[2], [x26]\n"
2125 "ld1 { v1.s }[2], [x25]\n"
2126 "ld1 { v2.s }[2], [x24]\n"
2127 "ld1 { v3.s }[2], [x23]\n"
2128 "ld1 { v4.s }[2], [x22]\n"
2131 "ldr s0, [x26, #0x0]\n"
2132 "ldr s1, [x25, #0x0]\n"
2133 "ldr s2, [x24, #0x0]\n"
2134 "ldr s3, [x23, #0x0]\n"
2135 "ldr s4, [x22, #0x0]\n"
2137 "ldr q6, [x10, #0x0]\n"
2138 "ldr q5, [x10, #0x10]\n"
2139 ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n"
2140 ".inst 0x0ea16842 // bfcvtn v2.4h, v2.4s\n"
2141 ".inst 0x4ea16820 // bfcvtn2 v0.8h, v1.4s\n"
2142 ".inst 0x4ea16862 // bfcvtn2 v2.8h, v3.4s\n"
2143 ".inst 0x6e46ec08 // bfmmla v8.4s, v0.8h, v6.8h\n"
2144 ".inst 0x0ea16884 // bfcvtn v4.4h, v4.4s\n"
2145 ".inst 0x6e46ec50 // bfmmla v16.4s, v2.8h, v6.8h\n"
2146 ".inst 0x6e46ec98 // bfmmla v24.4s, v4.8h, v6.8h\n"
2147 "ldr q3, [x10, #0x20]\n"
2148 ".inst 0x6e45ec0c // bfmmla v12.4s, v0.8h, v5.8h\n"
2149 ".inst 0x6e45ec54 // bfmmla v20.4s, v2.8h, v5.8h\n"
2150 ".inst 0x6e45ec9c // bfmmla v28.4s, v4.8h, v5.8h\n"
2151 "ldr q1, [x10, #0x30]\n"
2152 ".inst 0x6e43ec09 // bfmmla v9.4s, v0.8h, v3.8h\n"
2153 ".inst 0x6e43ec51 // bfmmla v17.4s, v2.8h, v3.8h\n"
2154 ".inst 0x6e43ec99 // bfmmla v25.4s, v4.8h, v3.8h\n"
2155 "ldr q3, [x10, #0x40]\n"
2156 ".inst 0x6e41ec0d // bfmmla v13.4s, v0.8h, v1.8h\n"
2157 ".inst 0x6e41ec55 // bfmmla v21.4s, v2.8h, v1.8h\n"
2158 ".inst 0x6e41ec9d // bfmmla v29.4s, v4.8h, v1.8h\n"
2159 "ldr q1, [x10, #0x50]\n"
2160 ".inst 0x6e43ec0a // bfmmla v10.4s, v0.8h, v3.8h\n"
2161 ".inst 0x6e43ec52 // bfmmla v18.4s, v2.8h, v3.8h\n"
2162 ".inst 0x6e43ec9a // bfmmla v26.4s, v4.8h, v3.8h\n"
2163 "ldr q3, [x10, #0x60]\n"
2164 ".inst 0x6e41ec0e // bfmmla v14.4s, v0.8h, v1.8h\n"
2165 ".inst 0x6e41ec56 // bfmmla v22.4s, v2.8h, v1.8h\n"
2166 ".inst 0x6e41ec9e // bfmmla v30.4s, v4.8h, v1.8h\n"
2167 "ldr q1, [x10, #0x70]\n"
2168 ".inst 0x6e43ec0b // bfmmla v11.4s, v0.8h, v3.8h\n"
2169 "add x10, x10, #0x80\n"
2170 ".inst 0x6e43ec53 // bfmmla v19.4s, v2.8h, v3.8h\n"
2171 ".inst 0x6e43ec9b // bfmmla v27.4s, v4.8h, v3.8h\n"
2172 ".inst 0x6e41ec0f // bfmmla v15.4s, v0.8h, v1.8h\n"
2173 ".inst 0x6e41ec57 // bfmmla v23.4s, v2.8h, v1.8h\n"
2174 ".inst 0x6e41ec9f // bfmmla v31.4s, v4.8h, v1.8h\n"
2176 "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
2177 "add x28, x28, #0x1\n"
2180 "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
2181 "add x25, x9, x20, LSL #2\n"
2182 "add x24, x25, x20, LSL #2\n"
2183 "uzp1 v6.2d, v8.2d, v12.2d\n"
2184 "add x23, x24, x20, LSL #2\n"
2185 "add x22, x23, x20, LSL #2\n"
2186 "uzp2 v8.2d, v8.2d, v12.2d\n"
2187 "uzp1 v12.2d, v9.2d, v13.2d\n"
2188 "uzp2 v9.2d, v9.2d, v13.2d\n"
2189 "uzp1 v13.2d, v10.2d, v14.2d\n"
2190 "prfm pstl1keep, [x9, #0x0]\n"
2191 "prfm pstl1keep, [x25, #0x0]\n"
2192 "uzp2 v10.2d, v10.2d, v14.2d\n"
2193 "uzp1 v14.2d, v11.2d, v15.2d\n"
2194 "prfm pstl1keep, [x24, #0x0]\n"
2195 "prfm pstl1keep, [x23, #0x0]\n"
2196 "uzp2 v11.2d, v11.2d, v15.2d\n"
2197 "uzp1 v15.2d, v16.2d, v20.2d\n"
2198 "prfm pstl1keep, [x22, #0x0]\n"
2199 "uzp2 v16.2d, v16.2d, v20.2d\n"
2200 "uzp1 v20.2d, v17.2d, v21.2d\n"
2201 "uzp2 v17.2d, v17.2d, v21.2d\n"
2202 "uzp1 v21.2d, v18.2d, v22.2d\n"
2203 "uzp2 v18.2d, v18.2d, v22.2d\n"
2204 "uzp1 v22.2d, v19.2d, v23.2d\n"
2205 "uzp2 v19.2d, v19.2d, v23.2d\n"
2206 "uzp1 v24.2d, v24.2d, v28.2d\n"
2207 "uzp1 v25.2d, v25.2d, v29.2d\n"
2208 "uzp1 v26.2d, v26.2d, v30.2d\n"
2209 "uzp1 v27.2d, v27.2d, v31.2d\n"
2210 "tbz %x[flags], #1, 165f\n"
2211 "add x20, %x[args_ptr], %[offset_max]\n"
2212 "ld1r { v1.4s }, [x20]\n"
2213 "add x20, %x[args_ptr], %[offset_min]\n"
2214 "ld1r { v0.4s }, [x20]\n"
2215 "fmin v6.4s, v6.4s, v1.4s\n"
2216 "fmin v12.4s, v12.4s, v1.4s\n"
2217 "fmin v13.4s, v13.4s, v1.4s\n"
2218 "fmin v14.4s, v14.4s, v1.4s\n"
2219 "fmin v8.4s, v8.4s, v1.4s\n"
2220 "fmin v9.4s, v9.4s, v1.4s\n"
2221 "fmin v10.4s, v10.4s, v1.4s\n"
2222 "fmin v11.4s, v11.4s, v1.4s\n"
2223 "fmin v15.4s, v15.4s, v1.4s\n"
2224 "fmin v20.4s, v20.4s, v1.4s\n"
2225 "fmin v21.4s, v21.4s, v1.4s\n"
2226 "fmin v22.4s, v22.4s, v1.4s\n"
2227 "fmin v16.4s, v16.4s, v1.4s\n"
2228 "fmin v17.4s, v17.4s, v1.4s\n"
2229 "fmin v18.4s, v18.4s, v1.4s\n"
2230 "fmin v19.4s, v19.4s, v1.4s\n"
2231 "fmin v24.4s, v24.4s, v1.4s\n"
2232 "fmin v25.4s, v25.4s, v1.4s\n"
2233 "fmin v26.4s, v26.4s, v1.4s\n"
2234 "fmin v27.4s, v27.4s, v1.4s\n"
2235 "fmax v6.4s, v6.4s, v0.4s\n"
2236 "fmax v12.4s, v12.4s, v0.4s\n"
2237 "fmax v13.4s, v13.4s, v0.4s\n"
2238 "fmax v14.4s, v14.4s, v0.4s\n"
2239 "fmax v8.4s, v8.4s, v0.4s\n"
2240 "fmax v9.4s, v9.4s, v0.4s\n"
2241 "fmax v10.4s, v10.4s, v0.4s\n"
2242 "fmax v11.4s, v11.4s, v0.4s\n"
2243 "fmax v15.4s, v15.4s, v0.4s\n"
2244 "fmax v20.4s, v20.4s, v0.4s\n"
2245 "fmax v21.4s, v21.4s, v0.4s\n"
2246 "fmax v22.4s, v22.4s, v0.4s\n"
2247 "fmax v16.4s, v16.4s, v0.4s\n"
2248 "fmax v17.4s, v17.4s, v0.4s\n"
2249 "fmax v18.4s, v18.4s, v0.4s\n"
2250 "fmax v19.4s, v19.4s, v0.4s\n"
2251 "fmax v24.4s, v24.4s, v0.4s\n"
2252 "fmax v25.4s, v25.4s, v0.4s\n"
2253 "fmax v26.4s, v26.4s, v0.4s\n"
2254 "fmax v27.4s, v27.4s, v0.4s\n"
2258 "tbz x11, #3, 169f\n"
2259 "st1 { v6.4s }, [x9], #0x10\n"
2260 "st1 { v12.4s }, [x9], #0x10\n"
2261 "st1 { v8.4s }, [x25], #0x10\n"
2262 "st1 { v9.4s }, [x25], #0x10\n"
2263 "st1 { v15.4s }, [x24], #0x10\n"
2264 "st1 { v20.4s }, [x24], #0x10\n"
2265 "st1 { v16.4s }, [x23], #0x10\n"
2266 "st1 { v17.4s }, [x23], #0x10\n"
2267 "st1 { v24.4s }, [x22], #0x10\n"
2268 "st1 { v25.4s }, [x22], #0x10\n"
2269 "tbz x11, #2, 167f\n"
2270 "st1 { v13.4s }, [x9], #0x10\n"
2271 "st1 { v10.4s }, [x25], #0x10\n"
2272 "st1 { v21.4s }, [x24], #0x10\n"
2273 "st1 { v18.4s }, [x23], #0x10\n"
2274 "st1 { v26.4s }, [x22], #0x10\n"
2275 "tbz x11, #1, 166f\n"
2276 "str d14, [x9], #0x8\n"
2277 "str d11, [x25], #0x8\n"
2278 "str d22, [x24], #0x8\n"
2279 "str d19, [x23], #0x8\n"
2280 "str d27, [x22], #0x8\n"
2281 "tbz x11, #0, 173f\n"
2282 "st1 { v14.s }[2], [x9]\n"
2283 "st1 { v11.s }[2], [x25]\n"
2284 "st1 { v22.s }[2], [x24]\n"
2285 "st1 { v19.s }[2], [x23]\n"
2286 "st1 { v27.s }[2], [x22]\n"
2289 "tbz x11, #0, 173f\n"
2290 "str s14, [x9, #0x0]\n"
2291 "str s11, [x25, #0x0]\n"
2292 "str s22, [x24, #0x0]\n"
2293 "str s19, [x23, #0x0]\n"
2294 "str s27, [x22, #0x0]\n"
2297 "tbz x11, #1, 168f\n"
2298 "str d13, [x9], #0x8\n"
2299 "str d10, [x25], #0x8\n"
2300 "str d21, [x24], #0x8\n"
2301 "str d18, [x23], #0x8\n"
2302 "str d26, [x22], #0x8\n"
2303 "tbz x11, #0, 173f\n"
2304 "st1 { v13.s }[2], [x9]\n"
2305 "st1 { v10.s }[2], [x25]\n"
2306 "st1 { v21.s }[2], [x24]\n"
2307 "st1 { v18.s }[2], [x23]\n"
2308 "st1 { v26.s }[2], [x22]\n"
2311 "tbz x11, #0, 173f\n"
2312 "str s13, [x9, #0x0]\n"
2313 "str s10, [x25, #0x0]\n"
2314 "str s21, [x24, #0x0]\n"
2315 "str s18, [x23, #0x0]\n"
2316 "str s26, [x22, #0x0]\n"
2319 "tbz x11, #2, 171f\n"
2320 "st1 { v6.4s }, [x9], #0x10\n"
2321 "st1 { v8.4s }, [x25], #0x10\n"
2322 "st1 { v15.4s }, [x24], #0x10\n"
2323 "st1 { v16.4s }, [x23], #0x10\n"
2324 "st1 { v24.4s }, [x22], #0x10\n"
2325 "tbz x11, #1, 170f\n"
2326 "str d12, [x9], #0x8\n"
2327 "str d9, [x25], #0x8\n"
2328 "str d20, [x24], #0x8\n"
2329 "str d17, [x23], #0x8\n"
2330 "str d25, [x22], #0x8\n"
2331 "tbz x11, #0, 173f\n"
2332 "st1 { v12.s }[2], [x9]\n"
2333 "st1 { v9.s }[2], [x25]\n"
2334 "st1 { v20.s }[2], [x24]\n"
2335 "st1 { v17.s }[2], [x23]\n"
2336 "st1 { v25.s }[2], [x22]\n"
2339 "tbz x11, #0, 173f\n"
2340 "str s12, [x9, #0x0]\n"
2341 "str s9, [x25, #0x0]\n"
2342 "str s20, [x24, #0x0]\n"
2343 "str s17, [x23, #0x0]\n"
2344 "str s25, [x22, #0x0]\n"
2347 "tbz x11, #1, 172f\n"
2348 "str d6, [x9], #0x8\n"
2349 "str d8, [x25], #0x8\n"
2350 "str d15, [x24], #0x8\n"
2351 "str d16, [x23], #0x8\n"
2352 "str d24, [x22], #0x8\n"
2353 "tbz x11, #0, 173f\n"
2354 "st1 { v6.s }[2], [x9]\n"
2355 "st1 { v8.s }[2], [x25]\n"
2356 "st1 { v15.s }[2], [x24]\n"
2357 "st1 { v16.s }[2], [x23]\n"
2358 "st1 { v24.s }[2], [x22]\n"
2361 "str s6, [x9, #0x0]\n"
2362 "str s8, [x25, #0x0]\n"
2363 "str s15, [x24, #0x0]\n"
2364 "str s16, [x23, #0x0]\n"
2365 "str s24, [x22, #0x0]\n"
2369 "str q6, [x9, #0x0]\n"
2370 "str q12, [x9, #0x10]\n"
2371 "str q13, [x9, #0x20]\n"
2372 "str q14, [x9, #0x30]\n"
2373 "add x9, x9, #0x40\n"
2374 "str q8, [x25, #0x0]\n"
2375 "str q9, [x25, #0x10]\n"
2376 "str q10, [x25, #0x20]\n"
2377 "str q11, [x25, #0x30]\n"
2378 "str q15, [x24, #0x0]\n"
2379 "str q20, [x24, #0x10]\n"
2380 "str q21, [x24, #0x20]\n"
2381 "str q22, [x24, #0x30]\n"
2382 "str q16, [x23, #0x0]\n"
2383 "str q17, [x23, #0x10]\n"
2384 "str q18, [x23, #0x20]\n"
2385 "str q19, [x23, #0x30]\n"
2386 "str q24, [x22, #0x0]\n"
2387 "str q25, [x22, #0x10]\n"
2388 "str q26, [x22, #0x20]\n"
2389 "str q27, [x22, #0x30]\n"
2391 "subs x11, x11, #0x10\n"
2395 "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
2397 "mov x12, %x[bias]\n"
2398 "ldr x11, [%x[args_ptr], %[offsetof_N]]\n"
2399 "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n"
2400 "mov x9, %x[output_ptr]\n"
2401 "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
2404 "ldr q8, [x12, #0x0]\n"
2405 "ldr q9, [x12, #0x10]\n"
2406 "zip2 v12.2d, v8.2d, v8.2d\n"
2407 "zip1 v8.2d, v8.2d, v8.2d\n"
2408 "ldr q10, [x12, #0x20]\n"
2409 "ldr q11, [x12, #0x30]\n"
2410 "zip2 v13.2d, v9.2d, v9.2d\n"
2411 "zip1 v9.2d, v9.2d, v9.2d\n"
2412 "zip2 v14.2d, v10.2d, v10.2d\n"
2413 "zip1 v10.2d, v10.2d, v10.2d\n"
2414 "add x12, x12, #0x40\n"
2415 "zip2 v15.2d, v11.2d, v11.2d\n"
2416 "zip1 v11.2d, v11.2d, v11.2d\n"
2417 "mov v16.16b, v8.16b\n"
2418 "mov v20.16b, v12.16b\n"
2419 "mov v17.16b, v9.16b\n"
2420 "mov v21.16b, v13.16b\n"
2421 "mov v18.16b, v10.16b\n"
2422 "mov v22.16b, v14.16b\n"
2423 "mov v19.16b, v11.16b\n"
2424 "mov v23.16b, v15.16b\n"
2425 "mov v24.16b, v8.16b\n"
2426 "mov v28.16b, v12.16b\n"
2427 "mov v25.16b, v9.16b\n"
2428 "mov v29.16b, v13.16b\n"
2429 "mov v26.16b, v10.16b\n"
2430 "mov v30.16b, v14.16b\n"
2431 "mov v27.16b, v11.16b\n"
2432 "mov v31.16b, v15.16b\n"
2435 "tbz %x[flags], #0, 189f\n"
2436 "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
2437 "add x25, x9, x20, LSL #2\n"
2438 "add x24, x25, x20, LSL #2\n"
2439 "add x23, x24, x20, LSL #2\n"
2440 "add x22, x23, x20, LSL #2\n"
2442 "add x21, x22, x20, LSL #2\n"
2444 "tbz x11, #3, 182f\n"
2445 "ld1 { v9.4s }, [x9], #0x10\n"
2446 "ld1 { v12.4s }, [x25], #0x10\n"
2447 "ld1 { v17.4s }, [x24], #0x10\n"
2448 "ld1 { v20.4s }, [x23], #0x10\n"
2449 "ld1 { v25.4s }, [x22], #0x10\n"
2450 "ld1 { v28.4s }, [x21], #0x10\n"
2451 "ld1 { v10.4s }, [x9], #0x10\n"
2452 "ld1 { v13.4s }, [x25], #0x10\n"
2453 "ld1 { v18.4s }, [x24], #0x10\n"
2454 "ld1 { v21.4s }, [x23], #0x10\n"
2455 "ld1 { v26.4s }, [x22], #0x10\n"
2456 "ld1 { v29.4s }, [x21], #0x10\n"
2457 "tbz x11, #2, 180f\n"
2458 "ld1 { v11.4s }, [x9], #0x10\n"
2459 "ld1 { v14.4s }, [x25], #0x10\n"
2460 "ld1 { v19.4s }, [x24], #0x10\n"
2461 "ld1 { v22.4s }, [x23], #0x10\n"
2462 "ld1 { v27.4s }, [x22], #0x10\n"
2463 "ld1 { v30.4s }, [x21], #0x10\n"
2464 "tbz x11, #1, 179f\n"
2465 "ldr d16, [x9], #0x8\n"
2466 "ldr d15, [x25], #0x8\n"
2468 "ldr d24, [x24], #0x8\n"
2469 "ldr d23, [x23], #0x8\n"
2470 "ldr d6, [x22], #0x8\n"
2471 "ldr d31, [x21], #0x8\n"
2472 "tbz x11, #0, 186f\n"
2473 "ld1 { v16.s }[2], [x9]\n"
2474 "ld1 { v15.s }[2], [x25]\n"
2475 "ld1 { v24.s }[2], [x24]\n"
2476 "ld1 { v23.s }[2], [x23]\n"
2477 "ld1 { v6.s }[2], [x22]\n"
2478 "ld1 { v31.s }[2], [x21]\n"
2482 "tbz x11, #0, 186f\n"
2483 "ldr s16, [x9, #0x0]\n"
2484 "ldr s15, [x25, #0x0]\n"
2485 "ldr s24, [x24, #0x0]\n"
2486 "ldr s23, [x23, #0x0]\n"
2487 "ldr s6, [x22, #0x0]\n"
2488 "ldr s31, [x21, #0x0]\n"
2491 "tbz x11, #1, 181f\n"
2492 "ldr d11, [x9], #0x8\n"
2493 "ldr d14, [x25], #0x8\n"
2495 "ldr d19, [x24], #0x8\n"
2496 "ldr d22, [x23], #0x8\n"
2497 "ldr d27, [x22], #0x8\n"
2498 "ldr d30, [x21], #0x8\n"
2499 "tbz x11, #0, 186f\n"
2500 "ld1 { v11.s }[2], [x9]\n"
2501 "ld1 { v14.s }[2], [x25]\n"
2502 "ld1 { v19.s }[2], [x24]\n"
2503 "ld1 { v22.s }[2], [x23]\n"
2504 "ld1 { v27.s }[2], [x22]\n"
2505 "ld1 { v30.s }[2], [x21]\n"
2509 "tbz x11, #0, 186f\n"
2510 "ldr s11, [x9, #0x0]\n"
2511 "ldr s14, [x25, #0x0]\n"
2512 "ldr s19, [x24, #0x0]\n"
2513 "ldr s22, [x23, #0x0]\n"
2514 "ldr s27, [x22, #0x0]\n"
2515 "ldr s30, [x21, #0x0]\n"
2518 "tbz x11, #2, 184f\n"
2519 "ld1 { v9.4s }, [x9], #0x10\n"
2520 "ld1 { v12.4s }, [x25], #0x10\n"
2521 "ld1 { v17.4s }, [x24], #0x10\n"
2522 "ld1 { v20.4s }, [x23], #0x10\n"
2523 "ld1 { v25.4s }, [x22], #0x10\n"
2524 "ld1 { v28.4s }, [x21], #0x10\n"
2525 "tbz x11, #1, 183f\n"
2526 "ldr d10, [x9], #0x8\n"
2527 "ldr d13, [x25], #0x8\n"
2529 "ldr d18, [x24], #0x8\n"
2530 "ldr d21, [x23], #0x8\n"
2531 "ldr d26, [x22], #0x8\n"
2532 "ldr d29, [x21], #0x8\n"
2533 "tbz x11, #0, 186f\n"
2534 "ld1 { v10.s }[2], [x9]\n"
2535 "ld1 { v13.s }[2], [x25]\n"
2536 "ld1 { v18.s }[2], [x24]\n"
2537 "ld1 { v21.s }[2], [x23]\n"
2538 "ld1 { v26.s }[2], [x22]\n"
2539 "ld1 { v29.s }[2], [x21]\n"
2543 "tbz x11, #0, 186f\n"
2544 "ldr s10, [x9, #0x0]\n"
2545 "ldr s13, [x25, #0x0]\n"
2546 "ldr s18, [x24, #0x0]\n"
2547 "ldr s21, [x23, #0x0]\n"
2548 "ldr s26, [x22, #0x0]\n"
2549 "ldr s29, [x21, #0x0]\n"
2552 "tbz x11, #1, 185f\n"
2553 "ldr d9, [x9], #0x8\n"
2554 "ldr d12, [x25], #0x8\n"
2556 "ldr d17, [x24], #0x8\n"
2557 "ldr d20, [x23], #0x8\n"
2558 "ldr d25, [x22], #0x8\n"
2559 "ldr d28, [x21], #0x8\n"
2560 "tbz x11, #0, 186f\n"
2561 "ld1 { v9.s }[2], [x9]\n"
2562 "ld1 { v12.s }[2], [x25]\n"
2563 "ld1 { v17.s }[2], [x24]\n"
2564 "ld1 { v20.s }[2], [x23]\n"
2565 "ld1 { v25.s }[2], [x22]\n"
2566 "ld1 { v28.s }[2], [x21]\n"
2569 "ldr s9, [x9, #0x0]\n"
2570 "ldr s12, [x25, #0x0]\n"
2572 "ldr s17, [x24, #0x0]\n"
2573 "ldr s20, [x23, #0x0]\n"
2574 "ldr s25, [x22, #0x0]\n"
2575 "ldr s28, [x21, #0x0]\n"
2580 "ldr q9, [x9, #0x0]\n"
2581 "ldr q10, [x9, #0x10]\n"
2582 "ldr q11, [x9, #0x20]\n"
2583 "ldr q16, [x9, #0x30]\n"
2584 "ldr q12, [x25, #0x0]\n"
2585 "ldr q13, [x25, #0x10]\n"
2586 "ldr q14, [x25, #0x20]\n"
2587 "ldr q15, [x25, #0x30]\n"
2588 "ldr q17, [x24, #0x0]\n"
2589 "ldr q18, [x24, #0x10]\n"
2590 "ldr q19, [x24, #0x20]\n"
2591 "ldr q24, [x24, #0x30]\n"
2592 "ldr q20, [x23, #0x0]\n"
2593 "ldr q21, [x23, #0x10]\n"
2594 "ldr q22, [x23, #0x20]\n"
2595 "ldr q23, [x23, #0x30]\n"
2596 "ldr q25, [x22, #0x0]\n"
2597 "ldr q26, [x22, #0x10]\n"
2598 "ldr q27, [x22, #0x20]\n"
2599 "ldr q6, [x22, #0x30]\n"
2600 "ldr q28, [x21, #0x0]\n"
2601 "ldr q29, [x21, #0x10]\n"
2602 "ldr q30, [x21, #0x20]\n"
2603 "ldr q31, [x21, #0x30]\n"
2605 "zip1 v8.2d, v9.2d, v12.2d\n"
2606 "zip2 v12.2d, v9.2d, v12.2d\n"
2607 "zip1 v9.2d, v10.2d, v13.2d\n"
2608 "zip2 v13.2d, v10.2d, v13.2d\n"
2609 "zip1 v10.2d, v11.2d, v14.2d\n"
2610 "zip2 v14.2d, v11.2d, v14.2d\n"
2611 "zip1 v11.2d, v16.2d, v15.2d\n"
2612 "zip2 v15.2d, v16.2d, v15.2d\n"
2613 "zip1 v16.2d, v17.2d, v20.2d\n"
2614 "zip2 v20.2d, v17.2d, v20.2d\n"
2615 "zip1 v17.2d, v18.2d, v21.2d\n"
2616 "zip2 v21.2d, v18.2d, v21.2d\n"
2617 "zip1 v18.2d, v19.2d, v22.2d\n"
2618 "zip2 v22.2d, v19.2d, v22.2d\n"
2619 "zip1 v19.2d, v24.2d, v23.2d\n"
2620 "zip2 v23.2d, v24.2d, v23.2d\n"
2621 "zip1 v24.2d, v25.2d, v28.2d\n"
2622 "zip2 v28.2d, v25.2d, v28.2d\n"
2623 "zip1 v25.2d, v26.2d, v29.2d\n"
2624 "zip2 v29.2d, v26.2d, v29.2d\n"
2625 "zip1 v26.2d, v27.2d, v30.2d\n"
2626 "zip2 v30.2d, v27.2d, v30.2d\n"
2627 "zip1 v27.2d, v6.2d, v31.2d\n"
2628 "zip2 v31.2d, v6.2d, v31.2d\n"
2631 "movi v8.16b, #0x0\n"
2632 "movi v9.16b, #0x0\n"
2633 "movi v10.16b, #0x0\n"
2634 "movi v11.16b, #0x0\n"
2635 "movi v12.16b, #0x0\n"
2636 "movi v13.16b, #0x0\n"
2637 "movi v14.16b, #0x0\n"
2638 "movi v15.16b, #0x0\n"
2639 "movi v16.16b, #0x0\n"
2640 "movi v17.16b, #0x0\n"
2641 "movi v18.16b, #0x0\n"
2642 "movi v19.16b, #0x0\n"
2643 "movi v20.16b, #0x0\n"
2644 "movi v21.16b, #0x0\n"
2645 "movi v22.16b, #0x0\n"
2646 "movi v23.16b, #0x0\n"
2647 "movi v24.16b, #0x0\n"
2648 "movi v25.16b, #0x0\n"
2649 "movi v26.16b, #0x0\n"
2650 "movi v27.16b, #0x0\n"
2651 "movi v28.16b, #0x0\n"
2652 "movi v29.16b, #0x0\n"
2653 "movi v30.16b, #0x0\n"
2654 "movi v31.16b, #0x0\n"
2658 "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
2659 "ldr w27, [x20, x28, LSL #0x2]\n"
2660 "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
2661 "tbz %x[flags], #3, 192f\n"
2662 "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
2663 "add x20, x20, x21, LSL #3\n"
2664 "ldr x26, [x20, #0x0]\n"
2665 "ldr x25, [x20, #0x8]\n"
2666 "ldr x24, [x20, #0x10]\n"
2667 "ldr x23, [x20, #0x18]\n"
2668 "ldr x22, [x20, #0x20]\n"
2669 "ldr x21, [x20, #0x28]\n"
2671 "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
2672 "add x26, x26, x20, LSL #2\n"
2673 "add x25, x25, x20, LSL #2\n"
2674 "add x24, x24, x20, LSL #2\n"
2675 "add x23, x23, x20, LSL #2\n"
2676 "add x22, x22, x20, LSL #2\n"
2677 "add x21, x21, x20, LSL #2\n"
2680 "mov x26, %x[input_ptr]\n"
2681 "add x25, x26, x21, LSL #2\n"
2682 "add x24, x25, x21, LSL #2\n"
2683 "add x23, x24, x21, LSL #2\n"
2684 "add x22, x23, x21, LSL #2\n"
2685 "add x21, x22, x21, LSL #2\n"
2689 "ld1 { v0.4s }, [x26], #0x10\n"
2690 "ld1 { v2.4s }, [x24], #0x10\n"
2692 "ld1 { v4.4s }, [x22], #0x10\n"
2693 "ld1 { v1.4s }, [x25], #0x10\n"
2694 "ld1 { v3.4s }, [x23], #0x10\n"
2695 "ld1 { v5.4s }, [x21], #0x10\n"
2696 "ldr q6, [x10, #0x0]\n"
2697 "ldr q7, [x10, #0x10]\n"
2700 ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n"
2701 ".inst 0x0ea16842 // bfcvtn v2.4h, v2.4s\n"
2702 "sub x27, x27, #0x4\n"
2704 ".inst 0x0ea16884 // bfcvtn v4.4h, v4.4s\n"
2705 ".inst 0x4ea16820 // bfcvtn2 v0.8h, v1.4s\n"
2706 ".inst 0x6e46ec08 // bfmmla v8.4s, v0.8h, v6.8h\n"
2707 "prfm pldl1keep, [x26, #0x80]\n"
2708 ".inst 0x4ea16862 // bfcvtn2 v2.8h, v3.4s\n"
2709 ".inst 0x4ea168a4 // bfcvtn2 v4.8h, v5.4s\n"
2710 ".inst 0x6e46ec50 // bfmmla v16.4s, v2.8h, v6.8h\n"
2711 "prfm pldl1keep, [x25, #0x80]\n"
2712 "ld1 { v1.4s }, [x25], #0x10\n"
2713 ".inst 0x6e46ec98 // bfmmla v24.4s, v4.8h, v6.8h\n"
2714 "ldr q6, [x10, #0x20]\n"
2715 ".inst 0x6e47ec0c // bfmmla v12.4s, v0.8h, v7.8h\n"
2716 ".inst 0x6e47ec54 // bfmmla v20.4s, v2.8h, v7.8h\n"
2717 ".inst 0x6e47ec9c // bfmmla v28.4s, v4.8h, v7.8h\n"
2718 "ldr q5, [x10, #0x30]\n"
2719 "prfm pldl1keep, [x24, #0x80]\n"
2720 ".inst 0x6e46ec09 // bfmmla v9.4s, v0.8h, v6.8h\n"
2721 ".inst 0x6e46ec51 // bfmmla v17.4s, v2.8h, v6.8h\n"
2722 "prfm pldl1keep, [x23, #0x80]\n"
2723 "ld1 { v3.4s }, [x23], #0x10\n"
2724 ".inst 0x6e46ec99 // bfmmla v25.4s, v4.8h, v6.8h\n"
2725 "ldr q6, [x10, #0x40]\n"
2726 ".inst 0x6e45ec0d // bfmmla v13.4s, v0.8h, v5.8h\n"
2727 "prfm pldl1keep, [x22, #0x80]\n"
2728 ".inst 0x6e45ec55 // bfmmla v21.4s, v2.8h, v5.8h\n"
2729 ".inst 0x6e45ec9d // bfmmla v29.4s, v4.8h, v5.8h\n"
2730 "ldr q7, [x10, #0x50]\n"
2731 "prfm pldl1keep, [x21, #0x80]\n"
2732 "ld1 { v5.4s }, [x21], #0x10\n"
2733 ".inst 0x6e46ec0a // bfmmla v10.4s, v0.8h, v6.8h\n"
2734 ".inst 0x6e46ec52 // bfmmla v18.4s, v2.8h, v6.8h\n"
2735 ".inst 0x6e46ec9a // bfmmla v26.4s, v4.8h, v6.8h\n"
2736 "ldr q6, [x10, #0x60]\n"
2737 ".inst 0x6e47ec0e // bfmmla v14.4s, v0.8h, v7.8h\n"
2738 ".inst 0x6e47ec56 // bfmmla v22.4s, v2.8h, v7.8h\n"
2739 ".inst 0x6e47ec9e // bfmmla v30.4s, v4.8h, v7.8h\n"
2740 "ldr q7, [x10, #0x70]\n"
2741 "add x10, x10, #0x80\n"
2742 ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n"
2743 ".inst 0x6e46ec53 // bfmmla v19.4s, v2.8h, v6.8h\n"
2744 ".inst 0x6e46ec9b // bfmmla v27.4s, v4.8h, v6.8h\n"
2745 "ldr q6, [x10, #0x0]\n"
2746 ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n"
2747 "ld1 { v0.4s }, [x26], #0x10\n"
2748 ".inst 0x6e47ec57 // bfmmla v23.4s, v2.8h, v7.8h\n"
2749 "ld1 { v2.4s }, [x24], #0x10\n"
2750 ".inst 0x6e47ec9f // bfmmla v31.4s, v4.8h, v7.8h\n"
2751 "ld1 { v4.4s }, [x22], #0x10\n"
2752 "ldr q7, [x10, #0x10]\n"
2755 ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n"
2756 ".inst 0x0ea16842 // bfcvtn v2.4h, v2.4s\n"
2757 "sub x27, x27, #0x4\n"
2758 "prfm pldl1keep, [x26, #0x80]\n"
2759 ".inst 0x0ea16884 // bfcvtn v4.4h, v4.4s\n"
2760 ".inst 0x4ea16820 // bfcvtn2 v0.8h, v1.4s\n"
2761 ".inst 0x6e46ec08 // bfmmla v8.4s, v0.8h, v6.8h\n"
2762 "prfm pldl1keep, [x25, #0x80]\n"
2763 ".inst 0x4ea16862 // bfcvtn2 v2.8h, v3.4s\n"
2764 ".inst 0x4ea168a4 // bfcvtn2 v4.8h, v5.4s\n"
2765 ".inst 0x6e46ec50 // bfmmla v16.4s, v2.8h, v6.8h\n"
2766 "prfm pldl1keep, [x24, #0x80]\n"
2767 ".inst 0x6e46ec98 // bfmmla v24.4s, v4.8h, v6.8h\n"
2768 "ldr q3, [x10, #0x20]\n"
2769 ".inst 0x6e47ec0c // bfmmla v12.4s, v0.8h, v7.8h\n"
2770 "prfm pldl1keep, [x23, #0x80]\n"
2771 ".inst 0x6e47ec54 // bfmmla v20.4s, v2.8h, v7.8h\n"
2772 ".inst 0x6e47ec9c // bfmmla v28.4s, v4.8h, v7.8h\n"
2773 "ldr q1, [x10, #0x30]\n"
2774 "prfm pldl1keep, [x22, #0x80]\n"
2775 ".inst 0x6e43ec09 // bfmmla v9.4s, v0.8h, v3.8h\n"
2776 ".inst 0x6e43ec51 // bfmmla v17.4s, v2.8h, v3.8h\n"
2777 "prfm pldl1keep, [x21, #0x80]\n"
2778 ".inst 0x6e43ec99 // bfmmla v25.4s, v4.8h, v3.8h\n"
2779 "ldr q3, [x10, #0x40]\n"
2780 ".inst 0x6e41ec0d // bfmmla v13.4s, v0.8h, v1.8h\n"
2781 ".inst 0x6e41ec55 // bfmmla v21.4s, v2.8h, v1.8h\n"
2782 ".inst 0x6e41ec9d // bfmmla v29.4s, v4.8h, v1.8h\n"
2783 "ldr q1, [x10, #0x50]\n"
2784 ".inst 0x6e43ec0a // bfmmla v10.4s, v0.8h, v3.8h\n"
2785 ".inst 0x6e43ec52 // bfmmla v18.4s, v2.8h, v3.8h\n"
2786 ".inst 0x6e43ec9a // bfmmla v26.4s, v4.8h, v3.8h\n"
2787 "ldr q3, [x10, #0x60]\n"
2788 ".inst 0x6e41ec0e // bfmmla v14.4s, v0.8h, v1.8h\n"
2789 ".inst 0x6e41ec56 // bfmmla v22.4s, v2.8h, v1.8h\n"
2790 ".inst 0x6e41ec9e // bfmmla v30.4s, v4.8h, v1.8h\n"
2791 "ldr q1, [x10, #0x70]\n"
2792 "add x10, x10, #0x80\n"
2793 ".inst 0x6e43ec0b // bfmmla v11.4s, v0.8h, v3.8h\n"
2794 ".inst 0x6e43ec53 // bfmmla v19.4s, v2.8h, v3.8h\n"
2795 ".inst 0x6e43ec9b // bfmmla v27.4s, v4.8h, v3.8h\n"
2796 ".inst 0x6e41ec0f // bfmmla v15.4s, v0.8h, v1.8h\n"
2797 ".inst 0x6e41ec57 // bfmmla v23.4s, v2.8h, v1.8h\n"
2798 ".inst 0x6e41ec9f // bfmmla v31.4s, v4.8h, v1.8h\n"
2802 "tbz x27, #1, 197f\n"
2803 "ldr d0, [x26], #0x8\n"
2804 "ldr d1, [x25], #0x8\n"
2805 "ldr d2, [x24], #0x8\n"
2806 "ldr d3, [x23], #0x8\n"
2807 "ldr d4, [x22], #0x8\n"
2808 "ldr d5, [x21], #0x8\n"
2809 "tbz x27, #0, 198f\n"
2810 "ld1 { v0.s }[2], [x26]\n"
2811 "ld1 { v1.s }[2], [x25]\n"
2812 "ld1 { v2.s }[2], [x24]\n"
2813 "ld1 { v3.s }[2], [x23]\n"
2814 "ld1 { v4.s }[2], [x22]\n"
2815 "ld1 { v5.s }[2], [x21]\n"
2818 "ldr s0, [x26, #0x0]\n"
2819 "ldr s1, [x25, #0x0]\n"
2820 "ldr s2, [x24, #0x0]\n"
2821 "ldr s3, [x23, #0x0]\n"
2822 "ldr s4, [x22, #0x0]\n"
2823 "ldr s5, [x21, #0x0]\n"
2825 "ldr q7, [x10, #0x0]\n"
2826 "ldr q6, [x10, #0x10]\n"
2827 ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n"
2828 ".inst 0x0ea16842 // bfcvtn v2.4h, v2.4s\n"
2829 ".inst 0x0ea16884 // bfcvtn v4.4h, v4.4s\n"
2830 ".inst 0x4ea16820 // bfcvtn2 v0.8h, v1.4s\n"
2831 ".inst 0x6e47ec08 // bfmmla v8.4s, v0.8h, v7.8h\n"
2832 ".inst 0x4ea16862 // bfcvtn2 v2.8h, v3.4s\n"
2833 ".inst 0x4ea168a4 // bfcvtn2 v4.8h, v5.4s\n"
2834 ".inst 0x6e47ec50 // bfmmla v16.4s, v2.8h, v7.8h\n"
2835 ".inst 0x6e47ec98 // bfmmla v24.4s, v4.8h, v7.8h\n"
2836 "ldr q3, [x10, #0x20]\n"
2837 ".inst 0x6e46ec0c // bfmmla v12.4s, v0.8h, v6.8h\n"
2838 ".inst 0x6e46ec54 // bfmmla v20.4s, v2.8h, v6.8h\n"
2839 ".inst 0x6e46ec9c // bfmmla v28.4s, v4.8h, v6.8h\n"
2840 "ldr q1, [x10, #0x30]\n"
2841 ".inst 0x6e43ec09 // bfmmla v9.4s, v0.8h, v3.8h\n"
2842 ".inst 0x6e43ec51 // bfmmla v17.4s, v2.8h, v3.8h\n"
2843 ".inst 0x6e43ec99 // bfmmla v25.4s, v4.8h, v3.8h\n"
2844 "ldr q3, [x10, #0x40]\n"
2845 ".inst 0x6e41ec0d // bfmmla v13.4s, v0.8h, v1.8h\n"
2846 ".inst 0x6e41ec55 // bfmmla v21.4s, v2.8h, v1.8h\n"
2847 ".inst 0x6e41ec9d // bfmmla v29.4s, v4.8h, v1.8h\n"
2848 "ldr q1, [x10, #0x50]\n"
2849 ".inst 0x6e43ec0a // bfmmla v10.4s, v0.8h, v3.8h\n"
2850 ".inst 0x6e43ec52 // bfmmla v18.4s, v2.8h, v3.8h\n"
2851 ".inst 0x6e43ec9a // bfmmla v26.4s, v4.8h, v3.8h\n"
2852 "ldr q3, [x10, #0x60]\n"
2853 ".inst 0x6e41ec0e // bfmmla v14.4s, v0.8h, v1.8h\n"
2854 ".inst 0x6e41ec56 // bfmmla v22.4s, v2.8h, v1.8h\n"
2855 ".inst 0x6e41ec9e // bfmmla v30.4s, v4.8h, v1.8h\n"
2856 "ldr q1, [x10, #0x70]\n"
2857 "add x10, x10, #0x80\n"
2858 ".inst 0x6e43ec0b // bfmmla v11.4s, v0.8h, v3.8h\n"
2859 ".inst 0x6e43ec53 // bfmmla v19.4s, v2.8h, v3.8h\n"
2860 ".inst 0x6e43ec9b // bfmmla v27.4s, v4.8h, v3.8h\n"
2861 ".inst 0x6e41ec0f // bfmmla v15.4s, v0.8h, v1.8h\n"
2862 ".inst 0x6e41ec57 // bfmmla v23.4s, v2.8h, v1.8h\n"
2863 ".inst 0x6e41ec9f // bfmmla v31.4s, v4.8h, v1.8h\n"
2865 "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
2866 "add x28, x28, #0x1\n"
2869 "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
2870 "add x25, x9, x20, LSL #2\n"
2871 "add x24, x25, x20, LSL #2\n"
2872 "uzp1 v6.2d, v8.2d, v12.2d\n"
2873 "add x23, x24, x20, LSL #2\n"
2874 "add x22, x23, x20, LSL #2\n"
2875 "uzp2 v8.2d, v8.2d, v12.2d\n"
2876 "uzp1 v12.2d, v9.2d, v13.2d\n"
2877 "add x21, x22, x20, LSL #2\n"
2878 "uzp2 v9.2d, v9.2d, v13.2d\n"
2879 "uzp1 v13.2d, v10.2d, v14.2d\n"
2880 "prfm pstl1keep, [x9, #0x0]\n"
2881 "uzp2 v10.2d, v10.2d, v14.2d\n"
2882 "uzp1 v14.2d, v11.2d, v15.2d\n"
2883 "prfm pstl1keep, [x25, #0x0]\n"
2884 "prfm pstl1keep, [x24, #0x0]\n"
2885 "uzp2 v11.2d, v11.2d, v15.2d\n"
2886 "uzp1 v15.2d, v16.2d, v20.2d\n"
2887 "prfm pstl1keep, [x23, #0x0]\n"
2888 "prfm pstl1keep, [x22, #0x0]\n"
2889 "uzp2 v16.2d, v16.2d, v20.2d\n"
2890 "uzp1 v20.2d, v17.2d, v21.2d\n"
2891 "prfm pstl1keep, [x21, #0x0]\n"
2892 "uzp2 v17.2d, v17.2d, v21.2d\n"
2893 "uzp1 v21.2d, v18.2d, v22.2d\n"
2894 "uzp2 v18.2d, v18.2d, v22.2d\n"
2895 "uzp1 v22.2d, v19.2d, v23.2d\n"
2896 "uzp2 v19.2d, v19.2d, v23.2d\n"
2897 "uzp1 v23.2d, v24.2d, v28.2d\n"
2898 "uzp2 v24.2d, v24.2d, v28.2d\n"
2899 "uzp1 v28.2d, v25.2d, v29.2d\n"
2900 "uzp2 v25.2d, v25.2d, v29.2d\n"
2901 "uzp1 v29.2d, v26.2d, v30.2d\n"
2902 "uzp2 v26.2d, v26.2d, v30.2d\n"
2903 "uzp1 v30.2d, v27.2d, v31.2d\n"
2904 "uzp2 v27.2d, v27.2d, v31.2d\n"
2905 "tbz %x[flags], #1, 200f\n"
2906 "add x20, %x[args_ptr], %[offset_max]\n"
2907 "ld1r { v1.4s }, [x20]\n"
2908 "add x20, %x[args_ptr], %[offset_min]\n"
2909 "ld1r { v0.4s }, [x20]\n"
2910 "fmin v6.4s, v6.4s, v1.4s\n"
2911 "fmin v12.4s, v12.4s, v1.4s\n"
2912 "fmin v13.4s, v13.4s, v1.4s\n"
2913 "fmin v14.4s, v14.4s, v1.4s\n"
2914 "fmin v8.4s, v8.4s, v1.4s\n"
2915 "fmin v9.4s, v9.4s, v1.4s\n"
2916 "fmin v10.4s, v10.4s, v1.4s\n"
2917 "fmin v11.4s, v11.4s, v1.4s\n"
2918 "fmin v15.4s, v15.4s, v1.4s\n"
2919 "fmin v20.4s, v20.4s, v1.4s\n"
2920 "fmin v21.4s, v21.4s, v1.4s\n"
2921 "fmin v22.4s, v22.4s, v1.4s\n"
2922 "fmin v16.4s, v16.4s, v1.4s\n"
2923 "fmin v17.4s, v17.4s, v1.4s\n"
2924 "fmin v18.4s, v18.4s, v1.4s\n"
2925 "fmin v19.4s, v19.4s, v1.4s\n"
2926 "fmin v23.4s, v23.4s, v1.4s\n"
2927 "fmin v28.4s, v28.4s, v1.4s\n"
2928 "fmin v29.4s, v29.4s, v1.4s\n"
2929 "fmin v30.4s, v30.4s, v1.4s\n"
2930 "fmin v24.4s, v24.4s, v1.4s\n"
2931 "fmin v25.4s, v25.4s, v1.4s\n"
2932 "fmin v26.4s, v26.4s, v1.4s\n"
2933 "fmin v27.4s, v27.4s, v1.4s\n"
2934 "fmax v6.4s, v6.4s, v0.4s\n"
2935 "fmax v12.4s, v12.4s, v0.4s\n"
2936 "fmax v13.4s, v13.4s, v0.4s\n"
2937 "fmax v14.4s, v14.4s, v0.4s\n"
2938 "fmax v8.4s, v8.4s, v0.4s\n"
2939 "fmax v9.4s, v9.4s, v0.4s\n"
2940 "fmax v10.4s, v10.4s, v0.4s\n"
2941 "fmax v11.4s, v11.4s, v0.4s\n"
2942 "fmax v15.4s, v15.4s, v0.4s\n"
2943 "fmax v20.4s, v20.4s, v0.4s\n"
2944 "fmax v21.4s, v21.4s, v0.4s\n"
2945 "fmax v22.4s, v22.4s, v0.4s\n"
2946 "fmax v16.4s, v16.4s, v0.4s\n"
2947 "fmax v17.4s, v17.4s, v0.4s\n"
2948 "fmax v18.4s, v18.4s, v0.4s\n"
2949 "fmax v19.4s, v19.4s, v0.4s\n"
2950 "fmax v23.4s, v23.4s, v0.4s\n"
2951 "fmax v28.4s, v28.4s, v0.4s\n"
2952 "fmax v29.4s, v29.4s, v0.4s\n"
2953 "fmax v30.4s, v30.4s, v0.4s\n"
2954 "fmax v24.4s, v24.4s, v0.4s\n"
2955 "fmax v25.4s, v25.4s, v0.4s\n"
2956 "fmax v26.4s, v26.4s, v0.4s\n"
2957 "fmax v27.4s, v27.4s, v0.4s\n"
2961 "tbz x11, #3, 204f\n"
2962 "st1 { v6.4s }, [x9], #0x10\n"
2963 "st1 { v12.4s }, [x9], #0x10\n"
2964 "st1 { v8.4s }, [x25], #0x10\n"
2965 "st1 { v9.4s }, [x25], #0x10\n"
2966 "st1 { v15.4s }, [x24], #0x10\n"
2967 "st1 { v20.4s }, [x24], #0x10\n"
2968 "st1 { v16.4s }, [x23], #0x10\n"
2969 "st1 { v17.4s }, [x23], #0x10\n"
2970 "st1 { v23.4s }, [x22], #0x10\n"
2971 "st1 { v28.4s }, [x22], #0x10\n"
2972 "st1 { v24.4s }, [x21], #0x10\n"
2973 "st1 { v25.4s }, [x21], #0x10\n"
2974 "tbz x11, #2, 202f\n"
2975 "st1 { v13.4s }, [x9], #0x10\n"
2976 "st1 { v10.4s }, [x25], #0x10\n"
2977 "st1 { v21.4s }, [x24], #0x10\n"
2978 "st1 { v18.4s }, [x23], #0x10\n"
2979 "st1 { v29.4s }, [x22], #0x10\n"
2980 "st1 { v26.4s }, [x21], #0x10\n"
2981 "tbz x11, #1, 201f\n"
2982 "str d14, [x9], #0x8\n"
2983 "str d11, [x25], #0x8\n"
2984 "str d22, [x24], #0x8\n"
2985 "str d19, [x23], #0x8\n"
2986 "str d30, [x22], #0x8\n"
2987 "str d27, [x21], #0x8\n"
2988 "tbz x11, #0, 208f\n"
2989 "st1 { v14.s }[2], [x9]\n"
2990 "st1 { v11.s }[2], [x25]\n"
2991 "st1 { v22.s }[2], [x24]\n"
2992 "st1 { v19.s }[2], [x23]\n"
2993 "st1 { v30.s }[2], [x22]\n"
2994 "st1 { v27.s }[2], [x21]\n"
2997 "tbz x11, #0, 208f\n"
2998 "str s14, [x9, #0x0]\n"
2999 "str s11, [x25, #0x0]\n"
3000 "str s22, [x24, #0x0]\n"
3001 "str s19, [x23, #0x0]\n"
3002 "str s30, [x22, #0x0]\n"
3003 "str s27, [x21, #0x0]\n"
3006 "tbz x11, #1, 203f\n"
3007 "str d13, [x9], #0x8\n"
3008 "str d10, [x25], #0x8\n"
3009 "str d21, [x24], #0x8\n"
3010 "str d18, [x23], #0x8\n"
3011 "str d29, [x22], #0x8\n"
3012 "str d26, [x21], #0x8\n"
3013 "tbz x11, #0, 208f\n"
3014 "st1 { v13.s }[2], [x9]\n"
3015 "st1 { v10.s }[2], [x25]\n"
3016 "st1 { v21.s }[2], [x24]\n"
3017 "st1 { v18.s }[2], [x23]\n"
3018 "st1 { v29.s }[2], [x22]\n"
3019 "st1 { v26.s }[2], [x21]\n"
3022 "tbz x11, #0, 208f\n"
3023 "str s13, [x9, #0x0]\n"
3024 "str s10, [x25, #0x0]\n"
3025 "str s21, [x24, #0x0]\n"
3026 "str s18, [x23, #0x0]\n"
3027 "str s29, [x22, #0x0]\n"
3028 "str s26, [x21, #0x0]\n"
3031 "tbz x11, #2, 206f\n"
3032 "st1 { v6.4s }, [x9], #0x10\n"
3033 "st1 { v8.4s }, [x25], #0x10\n"
3034 "st1 { v15.4s }, [x24], #0x10\n"
3035 "st1 { v16.4s }, [x23], #0x10\n"
3036 "st1 { v23.4s }, [x22], #0x10\n"
3037 "st1 { v24.4s }, [x21], #0x10\n"
3038 "tbz x11, #1, 205f\n"
3039 "str d12, [x9], #0x8\n"
3040 "str d9, [x25], #0x8\n"
3041 "str d20, [x24], #0x8\n"
3042 "str d17, [x23], #0x8\n"
3043 "str d28, [x22], #0x8\n"
3044 "str d25, [x21], #0x8\n"
3045 "tbz x11, #0, 208f\n"
3046 "st1 { v12.s }[2], [x9]\n"
3047 "st1 { v9.s }[2], [x25]\n"
3048 "st1 { v20.s }[2], [x24]\n"
3049 "st1 { v17.s }[2], [x23]\n"
3050 "st1 { v28.s }[2], [x22]\n"
3051 "st1 { v25.s }[2], [x21]\n"
3054 "tbz x11, #0, 208f\n"
3055 "str s12, [x9, #0x0]\n"
3056 "str s9, [x25, #0x0]\n"
3057 "str s20, [x24, #0x0]\n"
3058 "str s17, [x23, #0x0]\n"
3059 "str s28, [x22, #0x0]\n"
3060 "str s25, [x21, #0x0]\n"
3063 "tbz x11, #1, 207f\n"
3064 "str d6, [x9], #0x8\n"
3065 "str d8, [x25], #0x8\n"
3066 "str d15, [x24], #0x8\n"
3067 "str d16, [x23], #0x8\n"
3068 "str d23, [x22], #0x8\n"
3069 "str d24, [x21], #0x8\n"
3070 "tbz x11, #0, 208f\n"
3071 "st1 { v6.s }[2], [x9]\n"
3072 "st1 { v8.s }[2], [x25]\n"
3073 "st1 { v15.s }[2], [x24]\n"
3074 "st1 { v16.s }[2], [x23]\n"
3075 "st1 { v23.s }[2], [x22]\n"
3076 "st1 { v24.s }[2], [x21]\n"
3079 "str s6, [x9, #0x0]\n"
3080 "str s8, [x25, #0x0]\n"
3081 "str s15, [x24, #0x0]\n"
3082 "str s16, [x23, #0x0]\n"
3083 "str s23, [x22, #0x0]\n"
3084 "str s24, [x21, #0x0]\n"
3088 "str q6, [x9, #0x0]\n"
3089 "str q12, [x9, #0x10]\n"
3090 "str q13, [x9, #0x20]\n"
3091 "str q14, [x9, #0x30]\n"
3092 "add x9, x9, #0x40\n"
3093 "str q8, [x25, #0x0]\n"
3094 "str q9, [x25, #0x10]\n"
3095 "str q10, [x25, #0x20]\n"
3096 "str q11, [x25, #0x30]\n"
3097 "str q15, [x24, #0x0]\n"
3098 "str q20, [x24, #0x10]\n"
3099 "str q21, [x24, #0x20]\n"
3100 "str q22, [x24, #0x30]\n"
3101 "str q16, [x23, #0x0]\n"
3102 "str q17, [x23, #0x10]\n"
3103 "str q18, [x23, #0x20]\n"
3104 "str q19, [x23, #0x30]\n"
3105 "str q23, [x22, #0x0]\n"
3106 "str q28, [x22, #0x10]\n"
3107 "str q29, [x22, #0x20]\n"
3108 "str q30, [x22, #0x30]\n"
3109 "str q24, [x21, #0x0]\n"
3110 "str q25, [x21, #0x10]\n"
3111 "str q26, [x21, #0x20]\n"
3112 "str q27, [x21, #0x30]\n"
3114 "subs x11, x11, #0x10\n"
3116 "subs %x[M], %x[M], #0x6\n"
3118 "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
3119 "tbz %x[flags], #3, 211f\n"
3120 "add x21, x21, #0x6\n"
3121 "str x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
3125 "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
3128 : [
M]
"+&r" (
M), [input_ptr]
"+&r" (input_ptr), [output_ptr]
"+&r" (output_ptr)
3129 : [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))
3130 :
"cc",
"memory",
"v0",
"v1",
"v2",
"v3",
"v4",
"v5",
"v6",
"v7",
"v8",
"v9",
"v10",
"v11",
"v12",
"v13",
"v14",
"v15",
"v16",
"v17",
"v18",
"v19",
"v20",
"v21",
"v22",
"v23",
"v24",
"v25",
"v26",
"v27",
"v28",
"v29",
"v30",
"v31",
"x9",
"x10",
"x11",
"x12",
"x20",
"x21",
"x22",
"x23",
"x24",
"x25",
"x26",
"x27",
"x28"
3135 #endif // __aarch64__