Lines Matching refs:ACC
306 my @ACC=map("q$_",(6..13));
342 vmull.u32 @ACC[0],$Bi,${A0}[0]
343 vmull.u32 @ACC[1],$Bi,${A0}[1]
344 vmull.u32 @ACC[2],$Bi,${A1}[0]
345 vshl.i64 $Ni,@ACC[0]#hi,#16
346 vmull.u32 @ACC[3],$Bi,${A1}[1]
348 vadd.u64 $Ni,$Ni,@ACC[0]#lo
352 vmull.u32 @ACC[4],$Bi,${A2}[0]
354 vmull.u32 @ACC[5],$Bi,${A2}[1]
355 vmull.u32 @ACC[6],$Bi,${A3}[0]
357 vmull.u32 @ACC[7],$Bi,${A3}[1]
359 vmlal.u32 @ACC[0],$Ni,${N0}[0]
361 vmlal.u32 @ACC[1],$Ni,${N0}[1]
362 vmlal.u32 @ACC[2],$Ni,${N1}[0]
363 vmlal.u32 @ACC[3],$Ni,${N1}[1]
365 vmlal.u32 @ACC[4],$Ni,${N2}[0]
366 vmov $Temp,@ACC[0]
367 vmlal.u32 @ACC[5],$Ni,${N2}[1]
368 vmov @ACC[0],@ACC[1]
369 vmlal.u32 @ACC[6],$Ni,${N3}[0]
370 vmov @ACC[1],@ACC[2]
371 vmlal.u32 @ACC[7],$Ni,${N3}[1]
372 vmov @ACC[2],@ACC[3]
373 vmov @ACC[3],@ACC[4]
375 vmov @ACC[4],@ACC[5]
376 vmov @ACC[5],@ACC[6]
378 vmov @ACC[6],@ACC[7]
379 veor @ACC[7],@ACC[7]
389 vadd.u64 @ACC[0]#lo,@ACC[0]#lo,$temp
391 vmlal.u32 @ACC[0],$Bi,${A0}[0]
392 vmlal.u32 @ACC[1],$Bi,${A0}[1]
393 vmlal.u32 @ACC[2],$Bi,${A1}[0]
394 vshl.i64 $Ni,@ACC[0]#hi,#16
395 vmlal.u32 @ACC[3],$Bi,${A1}[1]
397 vadd.u64 $Ni,$Ni,@ACC[0]#lo
402 vmlal.u32 @ACC[4],$Bi,${A2}[0]
403 vmlal.u32 @ACC[5],$Bi,${A2}[1]
404 vmlal.u32 @ACC[6],$Bi,${A3}[0]
406 vmlal.u32 @ACC[7],$Bi,${A3}[1]
408 vmlal.u32 @ACC[0],$Ni,${N0}[0]
409 vmlal.u32 @ACC[1],$Ni,${N0}[1]
410 vmlal.u32 @ACC[2],$Ni,${N1}[0]
411 vmlal.u32 @ACC[3],$Ni,${N1}[1]
413 vmlal.u32 @ACC[4],$Ni,${N2}[0]
414 vmov $Temp,@ACC[0]
415 vmlal.u32 @ACC[5],$Ni,${N2}[1]
416 vmov @ACC[0],@ACC[1]
417 vmlal.u32 @ACC[6],$Ni,${N3}[0]
418 vmov @ACC[1],@ACC[2]
419 vmlal.u32 @ACC[7],$Ni,${N3}[1]
420 vmov @ACC[2],@ACC[3]
421 vmov @ACC[3],@ACC[4]
423 vmov @ACC[4],@ACC[5]
424 vmov @ACC[5],@ACC[6]
426 vmov @ACC[6],@ACC[7]
427 veor @ACC[7],@ACC[7]
432 vadd.u64 @ACC[0]#lo,@ACC[0]#lo,$temp
434 vshr.u64 $temp,@ACC[0]#lo,#16
436 vadd.u64 @ACC[0]#hi,@ACC[0]#hi,$temp
438 vshr.u64 $temp,@ACC[0]#hi,#16
439 vzip.16 @ACC[0]#lo,@ACC[0]#hi
445 veor @ACC[0],@ACC[0],@ACC[0]
447 veor @ACC[1],@ACC[1],@ACC[1]
449 veor @ACC[2],@ACC[2],@ACC[2]
451 veor @ACC[3],@ACC[3],@ACC[3]
453 veor @ACC[4],@ACC[4],@ACC[4]
455 veor @ACC[5],@ACC[5],@ACC[5]
457 veor @ACC[6],@ACC[6],@ACC[6]
458 veor @ACC[7],@ACC[7],@ACC[7]
461 vst1.64 {@ACC[0]-@ACC[1]},[$toutptr,:256]!
463 vst1.64 {@ACC[2]-@ACC[3]},[$toutptr,:256]!
464 vst1.64 {@ACC[4]-@ACC[5]},[$toutptr,:256]!
465 vst1.64 {@ACC[6]-@ACC[7]},[$toutptr,:256]!
483 vmlal.u32 @ACC[0],$Bi,${A0}[0]
484 vmlal.u32 @ACC[1],$Bi,${A0}[1]
486 vmlal.u32 @ACC[2],$Bi,${A1}[0]
487 vshl.i64 $Ni,@ACC[0]#hi,#16
488 vmlal.u32 @ACC[3],$Bi,${A1}[1]
489 vadd.u64 $Ni,$Ni,@ACC[0]#lo
490 vmlal.u32 @ACC[4],$Bi,${A2}[0]
492 vmlal.u32 @ACC[5],$Bi,${A2}[1]
494 vmlal.u32 @ACC[6],$Bi,${A3}[0]
496 vmlal.u32 @ACC[7],$Bi,${A3}[1]
501 vmlal.u32 @ACC[0],$Ni,${N0}[0]
503 vmlal.u32 @ACC[1],$Ni,${N0}[1]
505 vmlal.u32 @ACC[2],$Ni,${N1}[0]
506 vshr.u64 @ACC[0]#lo,@ACC[0]#lo,#16
507 vmlal.u32 @ACC[3],$Ni,${N1}[1]
508 vmlal.u32 @ACC[4],$Ni,${N2}[0]
509 vadd.u64 @ACC[0]#lo,@ACC[0]#lo,@ACC[0]#hi
510 vmlal.u32 @ACC[5],$Ni,${N2}[1]
511 vshr.u64 @ACC[0]#lo,@ACC[0]#lo,#16
512 vmlal.u32 @ACC[6],$Ni,${N3}[0]
513 vmlal.u32 @ACC[7],$Ni,${N3}[1]
514 vadd.u64 @ACC[1]#lo,@ACC[1]#lo,@ACC[0]#lo
517 push(@ACC,shift(@ACC)); $i++;
519 vmlal.u32 @ACC[0],$Bi,${A0}[0]
520 vld1.64 {@ACC[7]},[$tinptr,:128]!
521 vmlal.u32 @ACC[1],$Bi,${A0}[1]
523 vmlal.u32 @ACC[2],$Bi,${A1}[0]
524 vshl.i64 $Ni,@ACC[0]#hi,#16
525 vmlal.u32 @ACC[3],$Bi,${A1}[1]
526 vadd.u64 $Ni,$Ni,@ACC[0]#lo
527 vmlal.u32 @ACC[4],$Bi,${A2}[0]
529 vmlal.u32 @ACC[5],$Bi,${A2}[1]
531 vmlal.u32 @ACC[6],$Bi,${A3}[0]
533 vmlal.u32 @ACC[7],$Bi,${A3}[1]
538 vmlal.u32 @ACC[0],$Ni,${N0}[0]
540 vmlal.u32 @ACC[1],$Ni,${N0}[1]
541 vmlal.u32 @ACC[2],$Ni,${N1}[0]
542 vshr.u64 @ACC[0]#lo,@ACC[0]#lo,#16
543 vmlal.u32 @ACC[3],$Ni,${N1}[1]
544 vmlal.u32 @ACC[4],$Ni,${N2}[0]
545 vadd.u64 @ACC[0]#lo,@ACC[0]#lo,@ACC[0]#hi
546 vmlal.u32 @ACC[5],$Ni,${N2}[1]
547 vshr.u64 @ACC[0]#lo,@ACC[0]#lo,#16
548 vmlal.u32 @ACC[6],$Ni,${N3}[0]
549 vmlal.u32 @ACC[7],$Ni,${N3}[1]
550 vadd.u64 @ACC[1]#lo,@ACC[1]#lo,@ACC[0]#lo
554 push(@ACC,shift(@ACC));
562 vmlal.u32 @ACC[0],$Bi,${A0}[0]
563 vld1.64 {@ACC[7]},[$tinptr,:128]
564 vmlal.u32 @ACC[1],$Bi,${A0}[1]
566 vmlal.u32 @ACC[2],$Bi,${A1}[0]
568 vmlal.u32 @ACC[3],$Bi,${A1}[1]
571 vmlal.u32 @ACC[4],$Bi,${A2}[0]
572 vmlal.u32 @ACC[5],$Bi,${A2}[1]
573 vmlal.u32 @ACC[6],$Bi,${A3}[0]
574 vmlal.u32 @ACC[7],$Bi,${A3}[1]
579 vmlal.u32 @ACC[0],$Ni,${N0}[0]
580 vmlal.u32 @ACC[1],$Ni,${N0}[1]
581 vmlal.u32 @ACC[2],$Ni,${N1}[0]
582 vmlal.u32 @ACC[3],$Ni,${N1}[1]
583 vmlal.u32 @ACC[4],$Ni,${N2}[0]
584 vmlal.u32 @ACC[5],$Ni,${N2}[1]
585 vmlal.u32 @ACC[6],$Ni,${N3}[0]
586 vmlal.u32 @ACC[7],$Ni,${N3}[1]
587 vst1.64 {@ACC[0]},[$toutptr,:128]!
589 push(@ACC,shift(@ACC));
591 vmlal.u32 @ACC[0],$Bi,${A0}[0]
592 vld1.64 {@ACC[7]},[$tinptr,:128]
593 vmlal.u32 @ACC[1],$Bi,${A0}[1]
595 vmlal.u32 @ACC[2],$Bi,${A1}[0]
598 vmlal.u32 @ACC[3],$Bi,${A1}[1]
599 vmlal.u32 @ACC[4],$Bi,${A2}[0]
600 vmlal.u32 @ACC[5],$Bi,${A2}[1]
601 vmlal.u32 @ACC[6],$Bi,${A3}[0]
602 vmlal.u32 @ACC[7],$Bi,${A3}[1]
608 vmlal.u32 @ACC[0],$Ni,${N0}[0]
610 vmlal.u32 @ACC[1],$Ni,${N0}[1]
612 vmlal.u32 @ACC[2],$Ni,${N1}[0]
614 vmlal.u32 @ACC[3],$Ni,${N1}[1]
615 vmlal.u32 @ACC[4],$Ni,${N2}[0]
616 vmlal.u32 @ACC[5],$Ni,${N2}[1]
617 vmlal.u32 @ACC[6],$Ni,${N3}[0]
618 vst1.64 {@ACC[0]},[$toutptr,:128]!
619 vmlal.u32 @ACC[7],$Ni,${N3}[1]
623 push(@ACC,shift(@ACC));
626 vst1.64 {@ACC[0]-@ACC[1]},[$toutptr,:256]!
628 vst1.64 {@ACC[2]-@ACC[3]},[$toutptr,:256]!
630 vst1.64 {@ACC[4]-@ACC[5]},[$toutptr,:256]!
631 vst1.64 {@ACC[6]},[$toutptr,:128]
634 vld1.64 {@ACC[0]-@ACC[1]},[$tinptr,:256]!
635 vld1.64 {@ACC[2]-@ACC[3]},[$tinptr,:256]!
636 vld1.64 {@ACC[4]-@ACC[5]},[$tinptr,:256]!
637 vld1.64 {@ACC[6]-@ACC[7]},[$tinptr,:256]!
645 vshr.u64 $temp,@ACC[0]#lo,#16
647 vadd.u64 @ACC[0]#hi,@ACC[0]#hi,$temp
649 vshr.u64 $temp,@ACC[0]#hi,#16
651 vzip.16 @ACC[0]#lo,@ACC[0]#hi
658 vadd.u64 @ACC[0]#lo,@ACC[0]#lo,$temp
659 vshr.u64 $temp,@ACC[0]#lo,#16
660 vld1.64 {@ACC[2]-@ACC[3]}, [$tinptr, :256]!
661 vadd.u64 @ACC[0]#hi,@ACC[0]#hi,$temp
662 vld1.64 {@ACC[4]-@ACC[5]}, [$tinptr, :256]!
663 vshr.u64 $temp,@ACC[0]#hi,#16
664 vld1.64 {@ACC[6]-@ACC[7]}, [$tinptr, :256]!
665 vzip.16 @ACC[0]#lo,@ACC[0]#hi
671 vadd.u64 @ACC[1]#lo,@ACC[1]#lo,$temp
672 vst1.32 {@ACC[0]#lo[0]}, [$toutptr, :32]!
673 vshr.u64 $temp,@ACC[1]#lo,#16
674 vadd.u64 @ACC[1]#hi,@ACC[1]#hi,$temp
675 vshr.u64 $temp,@ACC[1]#hi,#16
676 vzip.16 @ACC[1]#lo,@ACC[1]#hi
678 push(@ACC,shift(@ACC));
680 push(@ACC,shift(@ACC));
682 vld1.64 {@ACC[0]-@ACC[1]}, [$tinptr, :256]!
684 vst1.32 {@ACC[7]#lo[0]}, [$toutptr, :32]!