Lines Matching refs:xmm0
281 movdqa xmm0, [eax] in ScaleRowDown2_SSE2()
284 pand xmm0, xmm7 in ScaleRowDown2_SSE2()
286 packuswb xmm0, xmm1 in ScaleRowDown2_SSE2()
287 movdqa [edx], xmm0 in ScaleRowDown2_SSE2()
310 movdqa xmm0, [eax] in ScaleRowDown2Int_SSE2()
315 pavgb xmm0, xmm2 // average rows in ScaleRowDown2Int_SSE2()
318 movdqa xmm2, xmm0 // average columns (32 to 16 pixels) in ScaleRowDown2Int_SSE2()
319 psrlw xmm0, 8 in ScaleRowDown2Int_SSE2()
324 pavgw xmm0, xmm2 in ScaleRowDown2Int_SSE2()
326 packuswb xmm0, xmm1 in ScaleRowDown2Int_SSE2()
328 movdqa [edx], xmm0 in ScaleRowDown2Int_SSE2()
354 movdqa xmm0, [esi] in ScaleRowDown4_SSE2()
357 pand xmm0, xmm7 in ScaleRowDown4_SSE2()
359 packuswb xmm0, xmm1 in ScaleRowDown4_SSE2()
360 packuswb xmm0, xmm0 in ScaleRowDown4_SSE2()
361 movq qword ptr [edi], xmm0 in ScaleRowDown4_SSE2()
387 movdqa xmm0, [esi] in ScaleRowDown4Int_SSE2()
391 pavgb xmm0, xmm2 // average rows in ScaleRowDown4Int_SSE2()
400 pavgb xmm0, xmm2 in ScaleRowDown4Int_SSE2()
403 movdqa xmm2, xmm0 // average columns (32 to 16 pixels) in ScaleRowDown4Int_SSE2()
404 psrlw xmm0, 8 in ScaleRowDown4Int_SSE2()
409 pavgw xmm0, xmm2 in ScaleRowDown4Int_SSE2()
411 packuswb xmm0, xmm1 in ScaleRowDown4Int_SSE2()
413 movdqa xmm2, xmm0 // average columns (16 to 8 pixels) in ScaleRowDown4Int_SSE2()
414 psrlw xmm0, 8 in ScaleRowDown4Int_SSE2()
416 pavgw xmm0, xmm2 in ScaleRowDown4Int_SSE2()
417 packuswb xmm0, xmm0 in ScaleRowDown4Int_SSE2()
419 movq qword ptr [edi], xmm0 in ScaleRowDown4Int_SSE2()
445 movdqa xmm0, [esi] in ScaleRowDown8_SSE2()
448 pand xmm0, xmm7 in ScaleRowDown8_SSE2()
450 packuswb xmm0, xmm1 // 32->16 in ScaleRowDown8_SSE2()
451 packuswb xmm0, xmm0 // 16->8 in ScaleRowDown8_SSE2()
452 packuswb xmm0, xmm0 // 8->4 in ScaleRowDown8_SSE2()
453 movd dword ptr [edi], xmm0 in ScaleRowDown8_SSE2()
478 movdqa xmm0, [esi] // average 8 rows to 1 in ScaleRowDown8Int_SSE2()
482 pavgb xmm0, xmm2 in ScaleRowDown8Int_SSE2()
492 pavgb xmm0, xmm2 in ScaleRowDown8Int_SSE2()
509 pavgb xmm0, xmm2 in ScaleRowDown8Int_SSE2()
512 psadbw xmm0, xmm7 // average 32 pixels to 4 in ScaleRowDown8Int_SSE2()
514 pshufd xmm0, xmm0, 0xd8 // x1x0 -> xx01 in ScaleRowDown8Int_SSE2()
516 por xmm0, xmm1 // -> 3201 in ScaleRowDown8Int_SSE2()
517 psrlw xmm0, 3 in ScaleRowDown8Int_SSE2()
518 packuswb xmm0, xmm0 in ScaleRowDown8Int_SSE2()
519 packuswb xmm0, xmm0 in ScaleRowDown8Int_SSE2()
520 movd dword ptr [edi], xmm0 in ScaleRowDown8Int_SSE2()
552 movdqa xmm0, [esi] in ScaleRowDown34_SSSE3()
556 palignr xmm1, xmm0, 8 in ScaleRowDown34_SSSE3()
557 pshufb xmm0, xmm3 in ScaleRowDown34_SSSE3()
560 movq qword ptr [edi], xmm0 in ScaleRowDown34_SSSE3()
605 movdqa xmm0, [esi] // pixels 0..7 in ScaleRowDown34_1_Int_SSSE3()
607 pavgb xmm0, xmm1 in ScaleRowDown34_1_Int_SSSE3()
608 pshufb xmm0, xmm2 in ScaleRowDown34_1_Int_SSSE3()
609 pmaddubsw xmm0, xmm5 in ScaleRowDown34_1_Int_SSSE3()
610 paddsw xmm0, xmm7 in ScaleRowDown34_1_Int_SSSE3()
611 psrlw xmm0, 2 in ScaleRowDown34_1_Int_SSSE3()
612 packuswb xmm0, xmm0 in ScaleRowDown34_1_Int_SSSE3()
613 movq qword ptr [edi], xmm0 in ScaleRowDown34_1_Int_SSSE3()
614 movdqu xmm0, [esi+8] // pixels 8..15 in ScaleRowDown34_1_Int_SSSE3()
616 pavgb xmm0, xmm1 in ScaleRowDown34_1_Int_SSSE3()
617 pshufb xmm0, xmm3 in ScaleRowDown34_1_Int_SSSE3()
618 pmaddubsw xmm0, xmm6 in ScaleRowDown34_1_Int_SSSE3()
619 paddsw xmm0, xmm7 in ScaleRowDown34_1_Int_SSSE3()
620 psrlw xmm0, 2 in ScaleRowDown34_1_Int_SSSE3()
621 packuswb xmm0, xmm0 in ScaleRowDown34_1_Int_SSSE3()
622 movq qword ptr [edi+8], xmm0 in ScaleRowDown34_1_Int_SSSE3()
623 movdqa xmm0, [esi+16] // pixels 16..23 in ScaleRowDown34_1_Int_SSSE3()
626 pavgb xmm0, xmm1 in ScaleRowDown34_1_Int_SSSE3()
627 pshufb xmm0, xmm4 in ScaleRowDown34_1_Int_SSSE3()
629 pmaddubsw xmm0, xmm1 in ScaleRowDown34_1_Int_SSSE3()
630 paddsw xmm0, xmm7 in ScaleRowDown34_1_Int_SSSE3()
631 psrlw xmm0, 2 in ScaleRowDown34_1_Int_SSSE3()
632 packuswb xmm0, xmm0 in ScaleRowDown34_1_Int_SSSE3()
633 movq qword ptr [edi+16], xmm0 in ScaleRowDown34_1_Int_SSSE3()
662 movdqa xmm0, [esi] // pixels 0..7 in ScaleRowDown34_0_Int_SSSE3()
664 pavgb xmm1, xmm0 in ScaleRowDown34_0_Int_SSSE3()
665 pavgb xmm0, xmm1 in ScaleRowDown34_0_Int_SSSE3()
666 pshufb xmm0, xmm2 in ScaleRowDown34_0_Int_SSSE3()
667 pmaddubsw xmm0, xmm5 in ScaleRowDown34_0_Int_SSSE3()
668 paddsw xmm0, xmm7 in ScaleRowDown34_0_Int_SSSE3()
669 psrlw xmm0, 2 in ScaleRowDown34_0_Int_SSSE3()
670 packuswb xmm0, xmm0 in ScaleRowDown34_0_Int_SSSE3()
671 movq qword ptr [edi], xmm0 in ScaleRowDown34_0_Int_SSSE3()
672 movdqu xmm0, [esi+8] // pixels 8..15 in ScaleRowDown34_0_Int_SSSE3()
674 pavgb xmm1, xmm0 in ScaleRowDown34_0_Int_SSSE3()
675 pavgb xmm0, xmm1 in ScaleRowDown34_0_Int_SSSE3()
676 pshufb xmm0, xmm3 in ScaleRowDown34_0_Int_SSSE3()
677 pmaddubsw xmm0, xmm6 in ScaleRowDown34_0_Int_SSSE3()
678 paddsw xmm0, xmm7 in ScaleRowDown34_0_Int_SSSE3()
679 psrlw xmm0, 2 in ScaleRowDown34_0_Int_SSSE3()
680 packuswb xmm0, xmm0 in ScaleRowDown34_0_Int_SSSE3()
681 movq qword ptr [edi+8], xmm0 in ScaleRowDown34_0_Int_SSSE3()
682 movdqa xmm0, [esi+16] // pixels 16..23 in ScaleRowDown34_0_Int_SSSE3()
685 pavgb xmm1, xmm0 in ScaleRowDown34_0_Int_SSSE3()
686 pavgb xmm0, xmm1 in ScaleRowDown34_0_Int_SSSE3()
687 pshufb xmm0, xmm4 in ScaleRowDown34_0_Int_SSSE3()
689 pmaddubsw xmm0, xmm1 in ScaleRowDown34_0_Int_SSSE3()
690 paddsw xmm0, xmm7 in ScaleRowDown34_0_Int_SSSE3()
691 psrlw xmm0, 2 in ScaleRowDown34_0_Int_SSSE3()
692 packuswb xmm0, xmm0 in ScaleRowDown34_0_Int_SSSE3()
693 movq qword ptr [edi+16], xmm0 in ScaleRowDown34_0_Int_SSSE3()
721 movdqa xmm0, [esi] // 16 pixels -> 0,1,2,3,4,5 in ScaleRowDown38_SSSE3()
724 pshufb xmm0, xmm5 in ScaleRowDown38_SSSE3()
726 paddusb xmm0, xmm1 in ScaleRowDown38_SSSE3()
728 movq qword ptr [edi], xmm0 // write 12 pixels in ScaleRowDown38_SSSE3()
729 movhlps xmm1, xmm0 in ScaleRowDown38_SSSE3()
756 movdqa xmm0, [esi] // sum up 3 rows into xmm0/1 in ScaleRowDown38_3_Int_SSSE3()
758 movhlps xmm1, xmm0 in ScaleRowDown38_3_Int_SSSE3()
760 punpcklbw xmm0, xmm7 in ScaleRowDown38_3_Int_SSSE3()
764 paddusw xmm0, xmm2 in ScaleRowDown38_3_Int_SSSE3()
771 paddusw xmm0, xmm2 in ScaleRowDown38_3_Int_SSSE3()
774 movdqa xmm2, xmm0 // 8 pixels -> 0,1,2 of xmm2 in ScaleRowDown38_3_Int_SSSE3()
775 psrldq xmm0, 2 in ScaleRowDown38_3_Int_SSSE3()
776 paddusw xmm2, xmm0 in ScaleRowDown38_3_Int_SSSE3()
777 psrldq xmm0, 2 in ScaleRowDown38_3_Int_SSSE3()
778 paddusw xmm2, xmm0 in ScaleRowDown38_3_Int_SSSE3()
824 movdqa xmm0, xmm2 // 16 pixels -> 0,1,2,3,4,5 of xmm0 in ScaleRowDown38_2_Int_SSSE3()
825 pshufb xmm0, xmm4 in ScaleRowDown38_2_Int_SSSE3()
828 paddusw xmm0, xmm1 in ScaleRowDown38_2_Int_SSSE3()
830 paddusw xmm0, xmm2 in ScaleRowDown38_2_Int_SSSE3()
832 pmulhuw xmm0, xmm7 // divide by 3,3,2, 3,3,2 in ScaleRowDown38_2_Int_SSSE3()
833 packuswb xmm0, xmm0 in ScaleRowDown38_2_Int_SSSE3()
835 movd [edi], xmm0 // write 6 pixels in ScaleRowDown38_2_Int_SSSE3()
836 pextrw eax, xmm0, 2 in ScaleRowDown38_2_Int_SSSE3()
875 movdqa xmm0, [eax] // read 16 pixels in ScaleAddRows_SSE2()
877 movhlps xmm1, xmm0 in ScaleAddRows_SSE2()
878 punpcklbw xmm0, xmm7 in ScaleAddRows_SSE2()
880 paddusw xmm2, xmm0 // sum 16 words in ScaleAddRows_SSE2()
928 movdqa xmm0, [esi] in ScaleFilterRows_SSE2()
931 movdqa xmm1, xmm0 in ScaleFilterRows_SSE2()
933 punpcklbw xmm0, xmm7 in ScaleFilterRows_SSE2()
937 pmullw xmm0, xmm5 // scale row 0 in ScaleFilterRows_SSE2()
941 paddusw xmm0, xmm2 // sum rows in ScaleFilterRows_SSE2()
943 psrlw xmm0, 8 in ScaleFilterRows_SSE2()
945 packuswb xmm0, xmm1 in ScaleFilterRows_SSE2()
946 movdqa [edi], xmm0 in ScaleFilterRows_SSE2()
958 movdqa xmm0, [esi] in ScaleFilterRows_SSE2()
960 movdqa [edi], xmm0 in ScaleFilterRows_SSE2()
972 movdqa xmm0, [esi] in ScaleFilterRows_SSE2()
975 pavgb xmm0, xmm2 in ScaleFilterRows_SSE2()
976 movdqa [edi], xmm0 in ScaleFilterRows_SSE2()
1017 movdqa xmm0, [esi] in ScaleFilterRows_SSSE3()
1020 movdqa xmm1, xmm0 in ScaleFilterRows_SSSE3()
1021 punpcklbw xmm0, xmm2 in ScaleFilterRows_SSSE3()
1023 pmaddubsw xmm0, xmm7 in ScaleFilterRows_SSSE3()
1025 psrlw xmm0, 7 in ScaleFilterRows_SSSE3()
1027 packuswb xmm0, xmm1 in ScaleFilterRows_SSSE3()
1028 movdqa [edi], xmm0 in ScaleFilterRows_SSSE3()
1040 movdqa xmm0, [esi] in ScaleFilterRows_SSSE3()
1042 movdqa [edi], xmm0 in ScaleFilterRows_SSSE3()
1054 movdqa xmm0, [esi] in ScaleFilterRows_SSSE3()
1057 pavgb xmm0, xmm2 in ScaleFilterRows_SSSE3()
1058 movdqa [edi], xmm0 in ScaleFilterRows_SSSE3()
1090 movdqa xmm0, [eax] // pixels 0..7 in ScaleFilterCols34_SSSE3()
1091 pshufb xmm0, xmm2 in ScaleFilterCols34_SSSE3()
1092 pmaddubsw xmm0, xmm5 in ScaleFilterCols34_SSSE3()
1093 paddsw xmm0, xmm1 in ScaleFilterCols34_SSSE3()
1094 psrlw xmm0, 2 in ScaleFilterCols34_SSSE3()
1095 packuswb xmm0, xmm0 in ScaleFilterCols34_SSSE3()
1096 movq qword ptr [edx], xmm0 in ScaleFilterCols34_SSSE3()
1097 movdqu xmm0, [eax+8] // pixels 8..15 in ScaleFilterCols34_SSSE3()
1098 pshufb xmm0, xmm3 in ScaleFilterCols34_SSSE3()
1099 pmaddubsw xmm0, xmm6 in ScaleFilterCols34_SSSE3()
1100 paddsw xmm0, xmm1 in ScaleFilterCols34_SSSE3()
1101 psrlw xmm0, 2 in ScaleFilterCols34_SSSE3()
1102 packuswb xmm0, xmm0 in ScaleFilterCols34_SSSE3()
1103 movq qword ptr [edx+8], xmm0 in ScaleFilterCols34_SSSE3()
1104 movdqa xmm0, [eax+16] // pixels 16..23 in ScaleFilterCols34_SSSE3()
1106 pshufb xmm0, xmm4 in ScaleFilterCols34_SSSE3()
1107 pmaddubsw xmm0, xmm7 in ScaleFilterCols34_SSSE3()
1108 paddsw xmm0, xmm1 in ScaleFilterCols34_SSSE3()
1109 psrlw xmm0, 2 in ScaleFilterCols34_SSSE3()
1110 packuswb xmm0, xmm0 in ScaleFilterCols34_SSSE3()
1111 movq qword ptr [edx+16], xmm0 in ScaleFilterCols34_SSSE3()