Lines Matching refs:ACC
295 my @ACC=map("q$_",(6..13));
331 vmull.u32 @ACC[0],$Bi,${A0}[0]
332 vmull.u32 @ACC[1],$Bi,${A0}[1]
333 vmull.u32 @ACC[2],$Bi,${A1}[0]
334 vshl.i64 $Ni,@ACC[0]#hi,#16
335 vmull.u32 @ACC[3],$Bi,${A1}[1]
337 vadd.u64 $Ni,$Ni,@ACC[0]#lo
341 vmull.u32 @ACC[4],$Bi,${A2}[0]
343 vmull.u32 @ACC[5],$Bi,${A2}[1]
344 vmull.u32 @ACC[6],$Bi,${A3}[0]
346 vmull.u32 @ACC[7],$Bi,${A3}[1]
348 vmlal.u32 @ACC[0],$Ni,${N0}[0]
350 vmlal.u32 @ACC[1],$Ni,${N0}[1]
351 vmlal.u32 @ACC[2],$Ni,${N1}[0]
352 vmlal.u32 @ACC[3],$Ni,${N1}[1]
354 vmlal.u32 @ACC[4],$Ni,${N2}[0]
355 vmov $Temp,@ACC[0]
356 vmlal.u32 @ACC[5],$Ni,${N2}[1]
357 vmov @ACC[0],@ACC[1]
358 vmlal.u32 @ACC[6],$Ni,${N3}[0]
359 vmov @ACC[1],@ACC[2]
360 vmlal.u32 @ACC[7],$Ni,${N3}[1]
361 vmov @ACC[2],@ACC[3]
362 vmov @ACC[3],@ACC[4]
364 vmov @ACC[4],@ACC[5]
365 vmov @ACC[5],@ACC[6]
367 vmov @ACC[6],@ACC[7]
368 veor @ACC[7],@ACC[7]
378 vadd.u64 @ACC[0]#lo,@ACC[0]#lo,$temp
380 vmlal.u32 @ACC[0],$Bi,${A0}[0]
381 vmlal.u32 @ACC[1],$Bi,${A0}[1]
382 vmlal.u32 @ACC[2],$Bi,${A1}[0]
383 vshl.i64 $Ni,@ACC[0]#hi,#16
384 vmlal.u32 @ACC[3],$Bi,${A1}[1]
386 vadd.u64 $Ni,$Ni,@ACC[0]#lo
391 vmlal.u32 @ACC[4],$Bi,${A2}[0]
392 vmlal.u32 @ACC[5],$Bi,${A2}[1]
393 vmlal.u32 @ACC[6],$Bi,${A3}[0]
395 vmlal.u32 @ACC[7],$Bi,${A3}[1]
397 vmlal.u32 @ACC[0],$Ni,${N0}[0]
398 vmlal.u32 @ACC[1],$Ni,${N0}[1]
399 vmlal.u32 @ACC[2],$Ni,${N1}[0]
400 vmlal.u32 @ACC[3],$Ni,${N1}[1]
402 vmlal.u32 @ACC[4],$Ni,${N2}[0]
403 vmov $Temp,@ACC[0]
404 vmlal.u32 @ACC[5],$Ni,${N2}[1]
405 vmov @ACC[0],@ACC[1]
406 vmlal.u32 @ACC[6],$Ni,${N3}[0]
407 vmov @ACC[1],@ACC[2]
408 vmlal.u32 @ACC[7],$Ni,${N3}[1]
409 vmov @ACC[2],@ACC[3]
410 vmov @ACC[3],@ACC[4]
412 vmov @ACC[4],@ACC[5]
413 vmov @ACC[5],@ACC[6]
415 vmov @ACC[6],@ACC[7]
416 veor @ACC[7],@ACC[7]
421 vadd.u64 @ACC[0]#lo,@ACC[0]#lo,$temp
423 vshr.u64 $temp,@ACC[0]#lo,#16
425 vadd.u64 @ACC[0]#hi,@ACC[0]#hi,$temp
427 vshr.u64 $temp,@ACC[0]#hi,#16
428 vzip.16 @ACC[0]#lo,@ACC[0]#hi
434 veor @ACC[0],@ACC[0],@ACC[0]
436 veor @ACC[1],@ACC[1],@ACC[1]
438 veor @ACC[2],@ACC[2],@ACC[2]
440 veor @ACC[3],@ACC[3],@ACC[3]
442 veor @ACC[4],@ACC[4],@ACC[4]
444 veor @ACC[5],@ACC[5],@ACC[5]
446 veor @ACC[6],@ACC[6],@ACC[6]
447 veor @ACC[7],@ACC[7],@ACC[7]
450 vst1.64 {@ACC[0]-@ACC[1]},[$toutptr,:256]!
452 vst1.64 {@ACC[2]-@ACC[3]},[$toutptr,:256]!
453 vst1.64 {@ACC[4]-@ACC[5]},[$toutptr,:256]!
454 vst1.64 {@ACC[6]-@ACC[7]},[$toutptr,:256]!
472 vmlal.u32 @ACC[0],$Bi,${A0}[0]
473 vmlal.u32 @ACC[1],$Bi,${A0}[1]
475 vmlal.u32 @ACC[2],$Bi,${A1}[0]
476 vshl.i64 $Ni,@ACC[0]#hi,#16
477 vmlal.u32 @ACC[3],$Bi,${A1}[1]
478 vadd.u64 $Ni,$Ni,@ACC[0]#lo
479 vmlal.u32 @ACC[4],$Bi,${A2}[0]
481 vmlal.u32 @ACC[5],$Bi,${A2}[1]
483 vmlal.u32 @ACC[6],$Bi,${A3}[0]
485 vmlal.u32 @ACC[7],$Bi,${A3}[1]
490 vmlal.u32 @ACC[0],$Ni,${N0}[0]
492 vmlal.u32 @ACC[1],$Ni,${N0}[1]
494 vmlal.u32 @ACC[2],$Ni,${N1}[0]
495 vshr.u64 @ACC[0]#lo,@ACC[0]#lo,#16
496 vmlal.u32 @ACC[3],$Ni,${N1}[1]
497 vmlal.u32 @ACC[4],$Ni,${N2}[0]
498 vadd.u64 @ACC[0]#lo,@ACC[0]#lo,@ACC[0]#hi
499 vmlal.u32 @ACC[5],$Ni,${N2}[1]
500 vshr.u64 @ACC[0]#lo,@ACC[0]#lo,#16
501 vmlal.u32 @ACC[6],$Ni,${N3}[0]
502 vmlal.u32 @ACC[7],$Ni,${N3}[1]
503 vadd.u64 @ACC[1]#lo,@ACC[1]#lo,@ACC[0]#lo
506 push(@ACC,shift(@ACC)); $i++;
508 vmlal.u32 @ACC[0],$Bi,${A0}[0]
509 vld1.64 {@ACC[7]},[$tinptr,:128]!
510 vmlal.u32 @ACC[1],$Bi,${A0}[1]
512 vmlal.u32 @ACC[2],$Bi,${A1}[0]
513 vshl.i64 $Ni,@ACC[0]#hi,#16
514 vmlal.u32 @ACC[3],$Bi,${A1}[1]
515 vadd.u64 $Ni,$Ni,@ACC[0]#lo
516 vmlal.u32 @ACC[4],$Bi,${A2}[0]
518 vmlal.u32 @ACC[5],$Bi,${A2}[1]
520 vmlal.u32 @ACC[6],$Bi,${A3}[0]
522 vmlal.u32 @ACC[7],$Bi,${A3}[1]
527 vmlal.u32 @ACC[0],$Ni,${N0}[0]
529 vmlal.u32 @ACC[1],$Ni,${N0}[1]
530 vmlal.u32 @ACC[2],$Ni,${N1}[0]
531 vshr.u64 @ACC[0]#lo,@ACC[0]#lo,#16
532 vmlal.u32 @ACC[3],$Ni,${N1}[1]
533 vmlal.u32 @ACC[4],$Ni,${N2}[0]
534 vadd.u64 @ACC[0]#lo,@ACC[0]#lo,@ACC[0]#hi
535 vmlal.u32 @ACC[5],$Ni,${N2}[1]
536 vshr.u64 @ACC[0]#lo,@ACC[0]#lo,#16
537 vmlal.u32 @ACC[6],$Ni,${N3}[0]
538 vmlal.u32 @ACC[7],$Ni,${N3}[1]
539 vadd.u64 @ACC[1]#lo,@ACC[1]#lo,@ACC[0]#lo
543 push(@ACC,shift(@ACC));
551 vmlal.u32 @ACC[0],$Bi,${A0}[0]
552 vld1.64 {@ACC[7]},[$tinptr,:128]
553 vmlal.u32 @ACC[1],$Bi,${A0}[1]
555 vmlal.u32 @ACC[2],$Bi,${A1}[0]
557 vmlal.u32 @ACC[3],$Bi,${A1}[1]
560 vmlal.u32 @ACC[4],$Bi,${A2}[0]
561 vmlal.u32 @ACC[5],$Bi,${A2}[1]
562 vmlal.u32 @ACC[6],$Bi,${A3}[0]
563 vmlal.u32 @ACC[7],$Bi,${A3}[1]
568 vmlal.u32 @ACC[0],$Ni,${N0}[0]
569 vmlal.u32 @ACC[1],$Ni,${N0}[1]
570 vmlal.u32 @ACC[2],$Ni,${N1}[0]
571 vmlal.u32 @ACC[3],$Ni,${N1}[1]
572 vmlal.u32 @ACC[4],$Ni,${N2}[0]
573 vmlal.u32 @ACC[5],$Ni,${N2}[1]
574 vmlal.u32 @ACC[6],$Ni,${N3}[0]
575 vmlal.u32 @ACC[7],$Ni,${N3}[1]
576 vst1.64 {@ACC[0]},[$toutptr,:128]!
578 push(@ACC,shift(@ACC));
580 vmlal.u32 @ACC[0],$Bi,${A0}[0]
581 vld1.64 {@ACC[7]},[$tinptr,:128]
582 vmlal.u32 @ACC[1],$Bi,${A0}[1]
584 vmlal.u32 @ACC[2],$Bi,${A1}[0]
587 vmlal.u32 @ACC[3],$Bi,${A1}[1]
588 vmlal.u32 @ACC[4],$Bi,${A2}[0]
589 vmlal.u32 @ACC[5],$Bi,${A2}[1]
590 vmlal.u32 @ACC[6],$Bi,${A3}[0]
591 vmlal.u32 @ACC[7],$Bi,${A3}[1]
597 vmlal.u32 @ACC[0],$Ni,${N0}[0]
599 vmlal.u32 @ACC[1],$Ni,${N0}[1]
601 vmlal.u32 @ACC[2],$Ni,${N1}[0]
603 vmlal.u32 @ACC[3],$Ni,${N1}[1]
604 vmlal.u32 @ACC[4],$Ni,${N2}[0]
605 vmlal.u32 @ACC[5],$Ni,${N2}[1]
606 vmlal.u32 @ACC[6],$Ni,${N3}[0]
607 vst1.64 {@ACC[0]},[$toutptr,:128]!
608 vmlal.u32 @ACC[7],$Ni,${N3}[1]
612 push(@ACC,shift(@ACC));
615 vst1.64 {@ACC[0]-@ACC[1]},[$toutptr,:256]!
617 vst1.64 {@ACC[2]-@ACC[3]},[$toutptr,:256]!
619 vst1.64 {@ACC[4]-@ACC[5]},[$toutptr,:256]!
620 vst1.64 {@ACC[6]},[$toutptr,:128]
623 vld1.64 {@ACC[0]-@ACC[1]},[$tinptr,:256]!
624 vld1.64 {@ACC[2]-@ACC[3]},[$tinptr,:256]!
625 vld1.64 {@ACC[4]-@ACC[5]},[$tinptr,:256]!
626 vld1.64 {@ACC[6]-@ACC[7]},[$tinptr,:256]!
634 vshr.u64 $temp,@ACC[0]#lo,#16
636 vadd.u64 @ACC[0]#hi,@ACC[0]#hi,$temp
638 vshr.u64 $temp,@ACC[0]#hi,#16
640 vzip.16 @ACC[0]#lo,@ACC[0]#hi
647 vadd.u64 @ACC[0]#lo,@ACC[0]#lo,$temp
648 vshr.u64 $temp,@ACC[0]#lo,#16
649 vld1.64 {@ACC[2]-@ACC[3]}, [$tinptr, :256]!
650 vadd.u64 @ACC[0]#hi,@ACC[0]#hi,$temp
651 vld1.64 {@ACC[4]-@ACC[5]}, [$tinptr, :256]!
652 vshr.u64 $temp,@ACC[0]#hi,#16
653 vld1.64 {@ACC[6]-@ACC[7]}, [$tinptr, :256]!
654 vzip.16 @ACC[0]#lo,@ACC[0]#hi
660 vadd.u64 @ACC[1]#lo,@ACC[1]#lo,$temp
661 vst1.32 {@ACC[0]#lo[0]}, [$toutptr, :32]!
662 vshr.u64 $temp,@ACC[1]#lo,#16
663 vadd.u64 @ACC[1]#hi,@ACC[1]#hi,$temp
664 vshr.u64 $temp,@ACC[1]#hi,#16
665 vzip.16 @ACC[1]#lo,@ACC[1]#hi
667 push(@ACC,shift(@ACC));
669 push(@ACC,shift(@ACC));
671 vld1.64 {@ACC[0]-@ACC[1]}, [$tinptr, :256]!
673 vst1.32 {@ACC[7]#lo[0]}, [$toutptr, :32]!