Lines Matching full:mm0
69 movq mm0, mm4 ; q0
71 psubusb mm3, mm0 ; q1-=q0
158 pxor mm0, [GLOBAL(t80)] ; offset to convert to signed values
159 movq mm3, mm0 ; q0
160 psubsb mm0, mm6 ; q0 - p0
161 paddsb mm2, mm0 ; 1 * (q0 - p0) + hvm(p1 - q1)
162 paddsb mm2, mm0 ; 2 * (q0 - p0) + hvm(p1 - q1)
163 paddsb mm2, mm0 ; 3 * (q0 - p0) + hvm(p1 - q1)
169 pxor mm0, mm0 ;
171 punpcklbw mm0, mm2 ;
173 psraw mm0, 11 ;
175 packsswb mm0, mm5
176 movq mm2, mm0 ; (3* (q0 - p0) + hvm(p1 - q1) + 3) >> 3;
178 pxor mm0, mm0 ; 0
180 punpcklbw mm0, mm1 ; e0f0g0h0
181 psraw mm0, 11 ; sign extended shift right by 3
185 movq mm5, mm0 ; save results
187 packsswb mm0, mm1 ; (3* (q0 - p0) + hvm(p1 - q1) + 4) >>3
205 psubsb mm3, mm0 ; q0-= q0 add
297 movq mm0, mm7 ; 17 07 16 06 15 05 14 04
300 punpcklwd mm0, mm6 ; 35 25 15 05 34 24 14 04
313 movq mm5, mm0 ; 35 25 15 05 34 24 14 04
316 punpckldq mm0, mm3 ; 74 64 54 44 34 24 15 04 = q0
326 movq [rdx+16], mm0 ; save q0
331 movq mm0, mm3 ; 13 03 12 02 11 01 10 00
332 punpcklwd mm0, mm1 ; 31 21 11 01 30 20 10 00
335 movq mm1, mm0 ; 31 21 11 01 30 20 10 00
337 punpckldq mm0, mm2 ; 70 60 50 40 30 20 10 00 =p3
341 psubusb mm2, mm0 ; p2-p3
343 psubusb mm0, mm1 ; p3-p2
344 por mm0, mm2 ; mm0=abs(p3-p2)
364 psubusb mm0, mm4
370 por mm0, mm1
371 por mm0, mm7 ; abs(q3-q2) > limit || abs(p3-p2) > limit ||abs(p2-p1) > limit || abs(q2-q1) > limit
385 por mm0, mm5 ; mm0=mask
400 por mm0, mm7 ; mask
426 por mm1, mm0; ; mask
428 pxor mm0, mm0
429 pcmpeqb mm1, mm0
442 pcmpeqb mm4, mm0
444 pcmpeqb mm0, mm0
445 pxor mm4, mm0
456 movq mm0, [rdx+16] ; q0
465 pxor mm0, [GLOBAL(t80)] ; offset to convert to signed values
467 movq mm3, mm0 ; q0
468 psubsb mm0, mm6 ; q0 - p0
470 paddsb mm2, mm0 ; 1 * (q0 - p0) + hvm(p1 - q1)
471 paddsb mm2, mm0 ; 2 * (q0 - p0) + hvm(p1 - q1)
473 paddsb mm2, mm0 ; 3 * (q0 - p0) + hvm(p1 - q1)
480 pxor mm0, mm0 ;
483 punpcklbw mm0, mm2 ;
486 psraw mm0, 11 ;
489 packsswb mm0, mm5
491 movq mm2, mm0 ; (3* (q0 - p0) + hvm(p1 - q1) + 3) >> 3;
493 pxor mm0, mm0 ; 0
496 punpcklbw mm0, mm1 ; e0f0g0h0
497 psraw mm0, 11 ; sign extended shift right by 3
503 movq mm5, mm0 ; save results
505 packsswb mm0, mm1 ; (3* (q0 - p0) + hvm(p1 - q1) + 4) >>3
527 psubsb mm3, mm0 ; q0-= q0 add
660 movq mm0, mm4 ; q0
662 psubusb mm3, mm0 ; q1-=q0
669 ; mm1 = mask, mm0=q0, mm7 = limit, t0 = abs(q0-q1)
681 ; mm1 = mask, mm0=q0, mm7 = limit, t0 = abs(q0-q1)
694 ; mm1 = mask, mm0=q0, mm7 = limit, t0 = abs(q0-q1)
704 ; mm1 = mask, mm0=q0, mm7 = limit, t0 = abs(q0-q1) t1 = abs(p1-p0)
715 movq mm3, mm0 ; q0
732 ; mm1 = mask, mm0=q0, mm7 = flimit, t0 = abs(q0-q1) t1 = abs(p1-p0)
751 ; mm1 = mask, mm0=q0, mm7 = thresh, t0 = abs(q0-q1) t1 = abs(p1-p0)
761 pxor mm0, [GLOBAL(t80)] ; offset to convert to signed values
762 movq mm3, mm0 ; q0
763 psubsb mm0, mm6 ; q0 - p0
764 paddsb mm2, mm0 ; 1 * (q0 - p0) + (p1 - q1)
765 paddsb mm2, mm0 ; 2 * (q0 - p0)
766 paddsb mm2, mm0 ; 3 * (q0 - p0) + (p1 - q1)
777 pxor mm0, mm0 ; 0
780 punpcklbw mm0, mm5 ; e0f0g0h0
781 psraw mm0, 11 ; sign extended shift right by 3
784 packsswb mm0, mm7 ; Filter2 >>=3;
786 movq mm5, mm0 ; Filter2
789 pxor mm0, mm0 ; 0
792 punpcklbw mm0, mm2 ; e0f0g0h0
793 psraw mm0, 11 ; sign extended shift right by 3
796 packsswb mm0, mm7 ; Filter2 >>=3;
798 ; mm0= filter2 mm1 = vp8_filter, mm3 =qs0 mm5=s mm4 =hev mm6=ps0
799 psubsb mm3, mm0 ; qs0 =qs0 - filter1
815 pxor mm0, mm0
956 movq mm0, [rdi+2*rax] ; 77 76 75 74 73 72 71 70
960 punpckhbw mm7, mm0 ; 77 67 76 66 75 65 74 64
962 punpcklbw mm6, mm0 ; 73 63 72 62 71 61 70 60
963 movq mm0, [rsi+rax] ; 57 56 55 54 53 52 51 50
968 punpckhbw mm5, mm0 ; 57 47 56 46 55 45 54 44
969 punpcklbw mm4, mm0 ; 53 43 52 42 51 41 50 40
993 movq mm0, mm7 ; 17 07 16 06 15 05 14 04
996 punpcklwd mm0, mm6 ; 35 25 15 05 34 24 14 04
1013 movq mm5, mm0 ; 35 25 15 05 34 24 14 04
1016 punpckldq mm0, mm3 ; 74 64 54 44 34 24 15 04 = q0
1025 movq [rdx+32], mm0 ; save q0
1030 movq mm0, mm3 ; 13 03 12 02 11 01 10 00
1031 punpcklwd mm0, mm1 ; 31 21 11 01 30 20 10 00
1034 movq mm1, mm0 ; 31 21 11 01 30 20 10 00
1036 punpckldq mm0, mm2 ; 70 60 50 40 30 20 10 00 =p3
1039 movq [rdx], mm0 ; save p3
1043 psubusb mm2, mm0 ; p2-p3
1045 psubusb mm0, mm1 ; p3-p2
1046 por mm0, mm2 ; mm0=abs(p3-p2)
1066 psubusb mm0, mm4 ; abs(p3-p2) > limit
1072 por mm0, mm1 ;
1073 por mm0, mm7 ; abs(q3-q2) > limit || abs(p3-p2) > limit ||abs(p2-p1) > limit || abs(q2-q1) > limit
1087 por mm0, mm5 ; mm0=mask
1102 por mm0, mm7 ; mask
1128 por mm1, mm0; ; mask
1130 pxor mm0, mm0
1131 pcmpeqb mm1, mm0
1144 pcmpeqb mm4, mm0
1146 pcmpeqb mm0, mm0
1147 pxor mm4, mm0
1163 movq mm0, [rdx+32] ; q0
1165 pxor mm0, [GLOBAL(t80)] ; offset to convert to signed values
1167 movq mm3, mm0 ; q0
1168 psubsb mm0, mm6 ; q0 - p0
1169 paddsb mm2, mm0 ; 1 * (q0 - p0) + (p1 - q1)
1170 paddsb mm2, mm0 ; 2 * (q0 - p0)
1171 paddsb mm2, mm0 ; 3 * (q0 - p0) + (p1 - q1)
1181 pxor mm0, mm0 ; 0
1184 punpcklbw mm0, mm5 ; e0f0g0h0
1185 psraw mm0, 11 ; sign extended shift right by 3
1188 packsswb mm0, mm7 ; Filter2 >>=3;
1190 movq mm5, mm0 ; Filter2
1193 pxor mm0, mm0 ; 0
1196 punpcklbw mm0, mm2 ; e0f0g0h0
1197 psraw mm0, 11 ; sign extended shift right by 3
1200 packsswb mm0, mm7 ; Filter2 >>=3;
1202 ; mm0= filter2 mm1 = vp8_filter, mm3 =qs0 mm5=s mm4 =hev mm6=ps0
1203 psubsb mm3, mm0 ; qs0 =qs0 - filter1
1219 pxor mm0, mm0
1303 movq mm0, [rdx] ; mm0 = 70 60 50 40 30 20 10 00
1304 movq mm1, mm0 ; mm0 = 70 60 50 40 30 20 10 00
1306 punpcklbw mm0, mm6 ; mm0 = 31 30 21 20 11 10 01 00
1315 movq mm5, mm0 ; mm5 = 31 30 21 20 11 10 01 00
1316 punpcklwd mm0, mm2 ; mm0 = 13 12 11 10 03 02 01 00
1334 movq mm6, mm0 ; mm6 = 13 12 11 10 03 02 01 00
1336 punpckldq mm0, mm2 ; mm0 = 07 06 05 04 03 02 01 00
1339 movq [rsi+rax*4], mm0 ; write out
1342 movq mm0, mm5 ; mm0 = 33 32 31 30 23 22 21 20
1343 punpckldq mm0, mm7 ; mm0 = 27 26 25 24 23 22 20 20
1346 movq [rsi+rax*2], mm0 ; write out
1358 movq mm0, mm1 ; mm0= 53 52 51 50 43 42 41 40
1361 punpckldq mm0, mm2 ; mm0 = 47 46 45 44 43 42 41 40
1364 movq [rsi], mm0 ; write out
1428 movq mm0, [rdi] ; q1
1430 movq mm7, mm0
1431 movq mm4, mm0
1432 psubusb mm0, mm1 ; q1-=p1
1434 por mm1, mm0 ; abs(p1-q1)
1440 movq mm0, mm4 ; q0
1458 pxor mm0, [GLOBAL(t80)] ; offset to convert to signed values
1459 movq mm3, mm0 ; q0
1460 psubsb mm0, mm6 ; q0 - p0
1461 paddsb mm2, mm0 ; p1 - q1 + 1 * (q0 - p0)
1462 paddsb mm2, mm0 ; p1 - q1 + 2 * (q0 - p0)
1463 paddsb mm2, mm0 ; p1 - q1 + 3 * (q0 - p0)
1469 movq mm0, mm5 ; get a copy of filters
1470 psllw mm0, 8 ; shift left 8
1471 psraw mm0, 3 ; arithmetic shift right 11
1472 psrlw mm0, 8
1477 por mm0, mm1 ; put the two together to get result
1479 psubsb mm3, mm0 ; q0-= q0 add
1487 movq mm0, mm5 ; get a copy of filters
1488 psllw mm0, 8 ; shift left 8
1489 psraw mm0, 3 ; arithmetic shift right 11
1490 psrlw mm0, 8
1493 por mm0, mm5 ; put the two together to get result
1496 paddsb mm6, mm0 ; p0+= p0 add
1546 movd mm0, [rdi + rax * 2] ; xx xx xx xx 73 72 71 70
1549 punpcklbw mm6, mm0 ; 73 63 72 62 71 61 70 60
1551 movd mm0, [rsi + rax] ; xx xx xx xx 53 52 51 50
1554 punpcklbw mm4, mm0 ; 53 43 52 42 51 41 50 40
1568 movd mm0, [rsi + rax * 4] ; xx xx xx xx 03 02 01 00
1569 punpcklbw mm0, mm1 ; 13 03 12 02 11 01 10 00
1571 movq mm2, mm0 ; 13 03 12 02 11 01 10 00
1572 punpcklwd mm0, mm6 ; 31 21 11 01 30 20 10 00
1575 movq mm1, mm0 ; 13 03 12 02 11 01 10 00
1577 punpckldq mm0, mm4 ; 70 60 50 40 30 20 10 00 = p1
1587 movq mm6, mm0 ; p1
1617 mm0
1620 pxor mm0, [GLOBAL(t80)] ; p1 offset to convert to signed values
1623 psubsb mm0, mm3 ; p1 - q1
1633 paddsb mm0, mm7 ; p1 - q1 + 1 * (q0 - p0)
1635 paddsb mm0, mm7 ; p1 - q1 + 2 * (q0 - p0)
1636 paddsb mm0, mm7 ; p1 - q1 + 3 * (q0 - p0)
1638 pand mm5, mm0 ; mask filter values we don't care about
1642 movq mm0, mm5 ; get a copy of filters
1643 psllw mm0, 8 ; shift left 8
1644 psraw mm0, 3 ; arithmetic shift right 11
1645 psrlw mm0, 8
1651 por mm0, mm7 ; put the two together to get result
1653 psubsb mm3, mm0 ; q0-= q0sz add
1659 movq mm0, mm5 ; get a copy of filters
1660 psllw mm0, 8 ; shift left 8
1661 psraw mm0, 3 ; arithmetic shift right 11
1662 psrlw mm0, 8
1666 por mm0, mm5 ; put the two together to get result
1668 paddsb mm6, mm0 ; p0+= p0 add
1672 movq mm0, t0
1675 ; mm0 = 70 60 50 40 30 20 10 00
1681 movq mm1, mm0 ;
1682 punpcklbw mm0, mm6 ; 31 30 21 20 11 10 01 00
1691 movq mm6, mm0 ; 31 30 21 20 11 10 01 00
1693 punpcklwd mm0, mm2 ; 13 12 11 10 03 02 01 00
1696 movd [rsi+rax*4], mm0 ; write 03 02 01 00
1699 psrlq mm0, 32 ; xx xx xx xx 13 12 11 10
1702 movd [rdi+rax*4], mm0 ; write 13 12 11 10
