Home | History | Annotate | Download | only in simd

Lines Matching full:movdqa

131         movdqa  xmm0, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_JCOEF)]
132 movdqa xmm1, XMMWORD [XMMBLOCK(2,0,rsi,SIZEOF_JCOEF)]
147 movdqa xmm5, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_JCOEF)]
152 movdqa xmm4,xmm5 ; xmm5=in0=(00 01 02 03 04 05 06 07)
165 movdqa XMMWORD [wk(8)], xmm6 ; wk(8)=col1
166 movdqa XMMWORD [wk(9)], xmm5 ; wk(9)=col3
167 movdqa XMMWORD [wk(10)], xmm3 ; wk(10)=col5
168 movdqa XMMWORD [wk(11)], xmm4 ; wk(11)=col7
175 movdqa xmm0, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_JCOEF)]
176 movdqa xmm1, XMMWORD [XMMBLOCK(2,0,rsi,SIZEOF_JCOEF)]
179 movdqa xmm2, XMMWORD [XMMBLOCK(4,0,rsi,SIZEOF_JCOEF)]
180 movdqa xmm3, XMMWORD [XMMBLOCK(6,0,rsi,SIZEOF_JCOEF)]
193 movdqa xmm4,xmm1 ; xmm1=in2=z2
194 movdqa xmm5,xmm1
197 movdqa xmm1,xmm4
198 movdqa xmm3,xmm5
204 movdqa xmm6,xmm0
215 movdqa xmm0,xmm7
218 movdqa xmm4,xmm2
222 movdqa XMMWORD [wk(0)], xmm7 ; wk(0)=tmp10L
223 movdqa XMMWORD [wk(1)], xmm2 ; wk(1)=tmp10H
224 movdqa XMMWORD [wk(2)], xmm0 ; wk(2)=tmp13L
225 movdqa XMMWORD [wk(3)], xmm4 ; wk(3)=tmp13H
234 movdqa xmm2,xmm5
237 movdqa xmm0,xmm7
241 movdqa XMMWORD [wk(4)], xmm5 ; wk(4)=tmp11L
242 movdqa XMMWORD [wk(5)], xmm7 ; wk(5)=tmp11H
243 movdqa XMMWORD [wk(6)], xmm2 ; wk(6)=tmp12L
244 movdqa XMMWORD [wk(7)], xmm0 ; wk(7)=tmp12H
248 movdqa xmm4, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_JCOEF)]
249 movdqa xmm6, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_JCOEF)]
252 movdqa xmm1, XMMWORD [XMMBLOCK(5,0,rsi,SIZEOF_JCOEF)]
253 movdqa xmm3, XMMWORD [XMMBLOCK(7,0,rsi,SIZEOF_JCOEF)]
257 movdqa xmm5,xmm6
258 movdqa xmm7,xmm4
271 movdqa xmm2,xmm5
272 movdqa xmm0,xmm5
275 movdqa xmm5,xmm2
276 movdqa xmm7,xmm0
282 movdqa XMMWORD [wk(10)], xmm2 ; wk(10)=z3L
283 movdqa XMMWORD [wk(11)], xmm0 ; wk(11)=z3H
301 movdqa xmm2,xmm3
302 movdqa xmm0,xmm3
305 movdqa xmm3,xmm2
306 movdqa xmm4,xmm0
317 movdqa XMMWORD [wk(8)], xmm2 ; wk(8)=tmp0L
318 movdqa XMMWORD [wk(9)], xmm0 ; wk(9)=tmp0H
320 movdqa xmm2,xmm1
321 movdqa xmm0,xmm1
324 movdqa xmm1,xmm2
325 movdqa xmm6,xmm0
336 movdqa XMMWORD [wk(10)], xmm2 ; wk(10)=tmp1L
337 movdqa XMMWORD [wk(11)], xmm0 ; wk(11)=tmp1H
341 movdqa xmm5, XMMWORD [wk(0)] ; xmm5=tmp10L
342 movdqa xmm7, XMMWORD [wk(1)] ; xmm7=tmp10H
344 movdqa xmm2,xmm5
345 movdqa xmm0,xmm7
351 movdqa xmm3,[rel PD_DESCALE_P1] ; xmm3=[rel PD_DESCALE_P1]
365 movdqa xmm4, XMMWORD [wk(4)] ; xmm4=tmp11L
366 movdqa xmm3, XMMWORD [wk(5)] ; xmm3=tmp11H
368 movdqa xmm7,xmm4
369 movdqa xmm0,xmm3
375 movdqa xmm1,[rel PD_DESCALE_P1] ; xmm1=[rel PD_DESCALE_P1]
389 movdqa xmm6,xmm5 ; transpose coefficients(phase 1)
392 movdqa xmm1,xmm7 ; transpose coefficients(phase 1)
396 movdqa xmm3, XMMWORD [wk(6)] ; xmm3=tmp12L
397 movdqa xmm0, XMMWORD [wk(7)] ; xmm0=tmp12H
398 movdqa xmm4, XMMWORD [wk(10)] ; xmm4=tmp1L
399 movdqa xmm2, XMMWORD [wk(11)] ; xmm2=tmp1H
401 movdqa XMMWORD [wk(0)], xmm5 ; wk(0)=(00 10 01 11 02 12 03 13)
402 movdqa XMMWORD [wk(1)], xmm6 ; wk(1)=(04 14 05 15 06 16 07 17)
403 movdqa
404 movdqa XMMWORD [wk(5)], xmm1 ; wk(5)=(64 74 65 75 66 76 67 77)
406 movdqa xmm5,xmm3
407 movdqa xmm6,xmm0
413 movdqa xmm7,[rel PD_DESCALE_P1] ; xmm7=[rel PD_DESCALE_P1]
427 movdqa xmm1, XMMWORD [wk(2)] ; xmm1=tmp13L
428 movdqa xmm4, XMMWORD [wk(3)] ; xmm4=tmp13H
429 movdqa xmm2, XMMWORD [wk(8)] ; xmm2=tmp0L
430 movdqa xmm7, XMMWORD [wk(9)] ; xmm7=tmp0H
432 movdqa xmm0,xmm1
433 movdqa xmm6,xmm4
439 movdqa xmm2,[rel PD_DESCALE_P1] ; xmm2=[rel PD_DESCALE_P1]
453 movdqa xmm7, XMMWORD [wk(0)] ; xmm7=(00 10 01 11 02 12 03 13)
454 movdqa xmm2, XMMWORD [wk(1)] ; xmm2=(04 14 05 15 06 16 07 17)
456 movdqa xmm4,xmm3 ; transpose coefficients(phase 1)
459 movdqa xmm6,xmm0 ; transpose coefficients(phase 1)
463 movdqa xmm1,xmm7 ; transpose coefficients(phase 2)
466 movdqa xmm5,xmm2 ; transpose coefficients(phase 2)
470 movdqa xmm3, XMMWORD [wk(4)] ; xmm3=(60 70 61 71 62 72 63 73)
471 movdqa xmm4, XMMWORD [wk(5)] ; xmm4=(64 74 65 75 66 76 67 77)
473 movdqa XMMWORD [wk(6)], xmm2 ; wk(6)=(04 14 24 34 05 15 25 35)
474 movdqa XMMWORD [wk(7)], xmm5 ; wk(7)=(06 16 26 36 07 17 27 37)
476 movdqa xmm2,xmm0 ; transpose coefficients(phase 2)
479 movdqa xmm5,xmm6 ; transpose coefficients(phase 2)
483 movdqa xmm3,xmm7 ; transpose coefficients(phase 3)
486 movdqa xmm4,xmm1 ; transpose coefficients(phase 3)
490 movdqa xmm0, XMMWORD [wk(6)] ; xmm0=(04 14 24 34 05 15 25 35)
491 movdqa xmm2, XMMWORD [wk(7)] ; xmm2=(06 16 26 36 07 17 27 37)
493 movdqa XMMWORD [wk(8)], xmm3 ; wk(8)=col1
494 movdqa XMMWORD [wk(9)], xmm4 ; wk(9)=col3
496 movdqa xmm3,xmm0 ; transpose coefficients(phase 3)
499 movdqa xmm4,xmm2 ; transpose coefficients(phase 3)
503 movdqa XMMWORD [wk(10)], xmm3 ; wk(10)=col5
504 movdqa XMMWORD [wk(11)], xmm4 ; wk(11)=col7
533 movdqa xmm6,xmm1 ; xmm1=in2=z2
534 movdqa xmm5,xmm1
537 movdqa xmm1,xmm6
538 movdqa xmm2,xmm5
544 movdqa xmm3,xmm7
555 movdqa xmm7,xmm4
558 movdqa xmm6,xmm0
562 movdqa XMMWORD [wk(0)], xmm4 ; wk(0)=tmp10L
563 movdqa XMMWORD [wk(1)], xmm0 ; wk(1)=tmp10H
564 movdqa XMMWORD [wk(2)], xmm7 ; wk(2)=tmp13L
565 movdqa XMMWORD [wk(3)], xmm6 ; wk(3)=tmp13H
574 movdqa xmm0,xmm5
577 movdqa xmm7,xmm4
581 movdqa XMMWORD [wk(4)], xmm5 ; wk(4)=tmp11L
582 movdqa XMMWORD [wk(5)], xmm4 ; wk(5)=tmp11H
583 movdqa XMMWORD [wk(6)], xmm0 ; wk(6)=tmp12L
584 movdqa XMMWORD [wk(7)], xmm7 ; wk(7)=tmp12H
588 movdqa xmm6, XMMWORD [wk(9)] ; xmm6=col3
589 movdqa xmm3, XMMWORD [wk(8)] ; xmm3=col1
590 movdqa xmm1, XMMWORD [wk(11)] ; xmm1=col7
591 movdqa xmm2, XMMWORD [wk(10)] ; xmm2=col5
593 movdqa xmm5,xmm6
594 movdqa xmm4,xmm3
607 movdqa xmm0,xmm5
608 movdqa xmm7,xmm5
611 movdqa xmm5,xmm0
612 movdqa xmm4,xmm7
618 movdqa XMMWORD [wk(10)], xmm0 ; wk(10)=z3L
619 movdqa XMMWORD [wk(11)], xmm7 ; wk(11)=z3H
637 movdqa xmm0,xmm1
638 movdqa xmm7,xmm1
641 movdqa xmm1,xmm0
642 movdqa xmm3,xmm7
653 movdqa XMMWORD [wk(8)], xmm0 ; wk(8)=tmp0L
654 movdqa XMMWORD [wk(9)], xmm7 ; wk(9)=tmp0H
656 movdqa xmm0,xmm2
657 movdqa xmm7,xmm2
660 movdqa xmm2,xmm0
661 movdqa xmm6,xmm7
672 movdqa XMMWORD [wk(10)], xmm0 ; wk(10)=tmp1L
673 movdqa XMMWORD [wk(11)], xmm7 ; wk(11)=tmp1H
677 movdqa xmm5, XMMWORD [wk(0)] ; xmm5=tmp10L
678 movdqa xmm4, XMMWORD [wk(1)] ; xmm4=tmp10H
680 movdqa xmm0,xmm5
681 movdqa xmm7,xmm4
687 movdqa xmm1,[rel PD_DESCALE_P2] ; xmm1=[rel PD_DESCALE_P2]
701 movdqa xmm3, XMMWORD [wk(4)] ; xmm3=tmp11L
702 movdqa xmm1, XMMWORD [wk(5)] ; xmm1=tmp11H
704 movdqa xmm4,xmm3
705 movdqa xmm7,xmm1
711 movdqa xmm2,[rel PD_DESCALE_P2] ; xmm2=[rel PD_DESCALE_P2]
728 movdqa xmm6, XMMWORD [wk(6)] ; xmm6=tmp12L
729 movdqa xmm2, XMMWORD [wk(7)] ; xmm2=tmp12H
730 movdqa xmm1, XMMWORD [wk(10)] ; xmm1=tmp1L
731 movdqa xmm7, XMMWORD [wk(11)] ; xmm7=tmp1H
733 movdqa XMMWORD [wk(0)], xmm5 ; wk(0)=(00 10 20 30 40 50 60 70 06 16 26 36 46 56 66 76)
734 movdqa XMMWORD [wk(1)], xmm3 ; wk(1)=(01 11 21 31 41 51 61 71 07 17 27 37 47 57 67 77)
736 movdqa xmm4,xmm6
737 movdqa xmm0,xmm2
743 movdqa xmm5,[rel PD_DESCALE_P2] ; xmm5=[rel PD_DESCALE_P2]
757 movdqa xmm3, XMMWORD [wk(2)] ; xmm3=tmp13L
758 movdqa xmm1, XMMWORD [wk(3)] ; xmm1=tmp13H
759 movdqa xmm7, XMMWORD [wk(8)] ; xmm7=tmp0L
760 movdqa xmm5, XMMWORD [wk(9)] ; xmm5=tmp0H
762 movdqa xmm2,xmm3
763 movdqa xmm0,xmm1
769 movdqa xmm7,[rel PD_DESCALE_P2] ; xmm7=[rel PD_DESCALE_P2]
780 movdqa xmm5,[rel PB_CENTERJSAMP] ; xmm5=[rel PB_CENTERJSAMP]
785 movdqa xmm7, XMMWORD [wk(0)] ; xmm7=(00 10 20 30 40 50 60 70 06 16 26 36 46 56 66 76)
786 movdqa xmm1, XMMWORD [wk(1)] ; xmm1=(01 11 21 31 41 51 61 71 07 17 27 37 47 57 67 77)
796 movdqa xmm0,xmm7 ; transpose coefficients(phase 1)
799 movdqa xmm2,xmm6 ; transpose coefficients(phase 1)
803 movdqa xmm4,xmm7 ; transpose coefficients(phase 2)
806 movdqa xmm5,xmm2 ; transpose coefficients(phase 2)
810 movdqa xmm1,xmm7 ; transpose coefficients(phase 3)
813 movdqa xmm3,xmm4 ; transpose coefficients(phase 3)