Lines Matching refs:v3
301 ld1 {v2.8b, v3.8b, v4.8b}, [x0], x1 //only use 21(16+5); v2=src[-2]
302 trn1 v2.2d, v2.2d, v3.2d
349 …VEC4_UADDL_8BITS v18, v19, v22, v23, v26, v27, v30, v31, v1, v3, v5, v7 //v1/v3/v5/v7=src[-1]+src[…
350 …VEC4_MLS_16BITS v1, v9, v3, v9, v5, v9, v7, v9, v0, v2, v4, v6 //v0/v2/v4/v6 -= 5*(src[-1]+src[2])
365 …VEC4_UADDL_8BITS v18, v19, v22, v23, v26, v27, v30, v31, v1, v3, v5, v7 //v1/v3/v5/v7=src[0]+src[1]
366 … VEC4_MLA_16BITS v1, v8, v3, v8, v5, v8, v7, v8, v0, v2, v4, v6 //v0/v2/v4/v6+=20*(src[0]+src[1])
368 VEC4_SQRSHRUN_16BITS_SHIFT5 v0, v2, v4, v6, v1, v3, v5, v7
370 VEC4_ST1_8BITS_8ELEMENT x2, x3, v1, v3, v5, v7
389 ld1 {v3.16b}, [x0], x1 //only use 9(4+5); 2nd row src[-2:6]
392 zip1 v4.4s, v2.4s, v3.4s // v4=src[-2] 1st:2nd
396 ext v3.16b, v3.16b, v4.16b, #1 //2nd row src[-1:6]
397 zip1 v5.4s, v2.4s, v3.4s // v5=src[-1:2] 1st:2nd
401 ext v3.16b, v3.16b, v4.16b, #1 //2nd row src[0:6]
402 zip1 v6.4s, v2.4s, v3.4s // v6=src[0:3] 1st:2nd
405 ext v3.16b, v3.16b, v4.16b, #1 //2nd row src[1:6]
406 zip1 v16.4s, v2.4s, v3.4s // v16=src[1:4] 1st:2nd
426 ld1 {v2.8b, v3.8b, v4.8b}, [x0], x1 //only use 21(16+5); v2=src[-2]
427 trn1 v2.2d, v2.2d, v3.2d
475 …VEC4_UADDL_8BITS v18, v19, v22, v23, v26, v27, v30, v31, v1, v3, v5, v7 //v1/v3/v5/v7=src[-1]+src[…
476 …VEC4_MLS_16BITS v1, v9, v3, v9, v5, v9, v7, v9, v0, v2, v4, v6 //v0/v2/v4/v6 -= 5*(src[-1]+src[2])
491 …VEC4_UADDL_8BITS v18, v19, v22, v23, v26, v27, v30, v31, v1, v3, v5, v7 //v1/v3/v5/v7=src[0]+src[1]
492 … VEC4_MLA_16BITS v1, v8, v3, v8, v5, v8, v7, v8, v0, v2, v4, v6 //v0/v2/v4/v6+=20*(src[0]+src[1])
493 VEC4_SQRSHRUN_16BITS_SHIFT5 v0, v2, v4, v6, v1, v3, v5, v7
495 VEC4_UADDL_8BITS v1, v18, v3, v22, v5, v26, v7, v30, v0, v2, v4, v6 //average with arc[0]
496 VEC4_RSHRN_16BITS_SHIFT1 v0, v2, v4, v6, v1, v3, v5, v7
498 VEC4_ST1_8BITS_8ELEMENT x2, x3, v1, v3, v5, v7
517 ld1 {v3.16b}, [x0], x1 //only use 9(4+5); 2nd row src[-2:6]
520 zip1 v4.4s, v2.4s, v3.4s // v4=src[-2] 1st:2nd
524 ext v3.16b, v3.16b, v4.16b, #1 //2nd row src[-1:6]
525 zip1 v5.4s, v2.4s, v3.4s // v5=src[-1:2] 1st:2nd
529 ext v3.16b, v3.16b, v4.16b, #1 //2nd row src[0:6]
530 zip1 v6.4s, v2.4s, v3.4s // v6=src[0:3] 1st:2nd
533 ext v3.16b, v3.16b, v4.16b, #1 //2nd row src[1:6]
534 zip1 v16.4s, v2.4s, v3.4s // v16=src[1:4] 1st:2nd
554 ld1 {v2.8b, v3.8b, v4.8b}, [x0], x1 //only use 21(16+5); v2=src[-2]
555 trn1 v2.2d, v2.2d, v3.2d
603 …VEC4_UADDL_8BITS v18, v19, v22, v23, v26, v27, v30, v31, v1, v3, v5, v7 //v1/v3/v5/v7=src[-1]+src[…
604 …VEC4_MLS_16BITS v1, v9, v3, v9, v5, v9, v7, v9, v0, v2, v4, v6 //v0/v2/v4/v6 -= 5*(src[-1]+src[2])
619 …VEC4_UADDL_8BITS v18, v19, v22, v23, v26, v27, v30, v31, v1, v3, v5, v7 //v1/v3/v5/v7=src[0]+src[1]
620 … VEC4_MLA_16BITS v1, v8, v3, v8, v5, v8, v7, v8, v0, v2, v4, v6 //v0/v2/v4/v6+=20*(src[0]+src[1])
621 VEC4_SQRSHRUN_16BITS_SHIFT5 v0, v2, v4, v6, v1, v3, v5, v7
623 VEC4_UADDL_8BITS v1, v19, v3, v23, v5, v27, v7, v31, v0, v2, v4, v6 //average with arc[0]
624 VEC4_RSHRN_16BITS_SHIFT1 v0, v2, v4, v6, v1, v3, v5, v7
626 VEC4_ST1_8BITS_8ELEMENT x2, x3, v1, v3, v5, v7
645 ld1 {v3.16b}, [x0], x1 //only use 9(4+5); 2nd row src[-2:6]
648 zip1 v4.4s, v2.4s, v3.4s // v4=src[-2] 1st:2nd
652 ext v3.16b, v3.16b, v4.16b, #1 //2nd row src[-1:6]
653 zip1 v5.4s, v2.4s, v3.4s // v5=src[-1:2] 1st:2nd
657 ext v3.16b, v3.16b, v4.16b, #1 //2nd row src[0:6]
658 zip1 v6.4s, v2.4s, v3.4s // v6=src[0:3] 1st:2nd
661 ext v3.16b, v3.16b, v4.16b, #1 //2nd row src[1:6]
662 zip1 v16.4s, v2.4s, v3.4s // v16=src[1:4] 1st:2nd
686 ld1 {v3.16b}, [x0], x1 // v3=src[-1*stride]
698 FILTER_6TAG_8BITS1_AVERAGE_WITH_0 v2, v3, v4, v5, v6, v7, v20, v0, v1
699 FILTER_6TAG_8BITS2_AVERAGE_WITH_0 v2, v3, v4, v5, v6, v7, v20, v0, v1
705 FILTER_6TAG_8BITS1_AVERAGE_WITH_0 v3, v4, v5, v6, v7, v2, v20, v0, v1
706 FILTER_6TAG_8BITS2_AVERAGE_WITH_0 v3, v4, v5, v6, v7, v2, v20, v0, v1
711 ld1 {v3.16b}, [x0], x1 // v3=src[3*stride]
712 FILTER_6TAG_8BITS1_AVERAGE_WITH_0 v4, v5, v6, v7, v2, v3, v20, v0, v1
713 FILTER_6TAG_8BITS2_AVERAGE_WITH_0 v4, v5, v6, v7, v2, v3, v20, v0, v1
719 FILTER_6TAG_8BITS1_AVERAGE_WITH_0 v5, v6, v7, v2, v3, v4, v20, v0, v1
720 FILTER_6TAG_8BITS2_AVERAGE_WITH_0 v5, v6, v7, v2, v3, v4, v20, v0, v1
726 FILTER_6TAG_8BITS1_AVERAGE_WITH_0 v6, v7, v2, v3, v4, v5, v20, v0, v1
727 FILTER_6TAG_8BITS2_AVERAGE_WITH_0 v6, v7, v2, v3, v4, v5, v20, v0, v1
733 FILTER_6TAG_8BITS1_AVERAGE_WITH_0 v7, v2, v3, v4, v5, v6, v20, v0, v1
734 FILTER_6TAG_8BITS2_AVERAGE_WITH_0 v7, v2, v3, v4, v5, v6, v20, v0, v1
739 FILTER_6TAG_8BITS1_AVERAGE_WITH_0 v2, v3, v4, v5, v6, v7, v20, v0, v1
740 FILTER_6TAG_8BITS2_AVERAGE_WITH_0 v2, v3, v4, v5, v6, v7, v20, v0, v1
745 FILTER_6TAG_8BITS1_AVERAGE_WITH_0 v3, v4, v5, v6, v7, v2, v20, v0, v1
746 FILTER_6TAG_8BITS2_AVERAGE_WITH_0 v3, v4, v5, v6, v7, v2, v20, v0, v1
749 mov v3.16b, v5.16b
782 …VEC4_UADDL_8BITS v17, v20, v18, v21, v19, v22, v20, v23, v1, v3, v5, v7 //v1/v3/v5/v7 =src[-1]+src…
783 …VEC4_MLS_16BITS v1, v31, v3, v31, v5, v31, v7, v31, v0, v2, v4, v6 //v0/v2/v4/v6 -=5*(src[-1]+src…
784 …VEC4_UADDL_8BITS v18, v19, v19, v20, v20, v21, v21, v22, v1, v3, v5, v7 //v1/v3/v5/v7 =src[0]+src[…
785 …VEC4_MLA_16BITS v1, v30, v3, v30, v5, v30, v7, v30, v0, v2, v4, v6 //v0/v2/v4/v6 += 20*(src[0]+sr…
786 VEC4_SQRSHRUN_16BITS_SHIFT5 v0, v2, v4, v6, v1, v3, v5, v7
788 …VEC4_UADDL_8BITS v1, v18, v3, v19, v5, v20, v7, v21, v0, v2, v4, v6 //v0/v2/v4/v6 = average with s…
789 VEC4_RSHRN_16BITS_SHIFT1 v0, v2, v4, v6, v1, v3, v5, v7
791 VEC4_ST1_8BITS_8ELEMENT x2, x3, v1, v3, v5, v7 //store 8bytes*4row
817 ld1 {v3.s}[0], [x0], x1 // v3=src[-1*stride]
818 mov v2.s[1], v3.s[0]
821 mov v3.s[1], v4.s[0]
835 FILTER_6TAG_8BITS1_AVERAGE_WITH_0 v2, v3, v4, v5, v6, v7, v20, v0, v1
843 ld1 {v3.s}[1], [x0], x1 // v2=src[6*stride]
844 mov v3.s[0], v2.s[1]
845 FILTER_6TAG_8BITS1_AVERAGE_WITH_0 v4, v5, v6, v7, v2, v3, v20, v0, v1
848 mov v4.s[0], v3.s[1]
854 mov v21.8b, v3.8b
855 mov v3.8b, v7.8b
877 ld1 {v3.16b}, [x0], x1 // v3=src[-1*stride]
889 FILTER_6TAG_8BITS1_AVERAGE_WITH_1 v2, v3, v4, v5, v6, v7, v20, v0, v1
890 FILTER_6TAG_8BITS2_AVERAGE_WITH_1 v2, v3, v4, v5, v6, v7, v20, v0, v1
896 FILTER_6TAG_8BITS1_AVERAGE_WITH_1 v3, v4, v5, v6, v7, v2, v20, v0, v1
897 FILTER_6TAG_8BITS2_AVERAGE_WITH_1 v3, v4, v5, v6, v7, v2, v20, v0, v1
902 ld1 {v3.16b}, [x0], x1 // v3=src[3*stride]
903 FILTER_6TAG_8BITS1_AVERAGE_WITH_1 v4, v5, v6, v7, v2, v3, v20, v0, v1
904 FILTER_6TAG_8BITS2_AVERAGE_WITH_1 v4, v5, v6, v7, v2, v3, v20, v0, v1
910 FILTER_6TAG_8BITS1_AVERAGE_WITH_1 v5, v6, v7, v2, v3, v4, v20, v0, v1
911 FILTER_6TAG_8BITS2_AVERAGE_WITH_1 v5, v6, v7, v2, v3, v4, v20, v0, v1
917 FILTER_6TAG_8BITS1_AVERAGE_WITH_1 v6, v7, v2, v3, v4, v5, v20, v0, v1
918 FILTER_6TAG_8BITS2_AVERAGE_WITH_1 v6, v7, v2, v3, v4, v5, v20, v0, v1
924 FILTER_6TAG_8BITS1_AVERAGE_WITH_1 v7, v2, v3, v4, v5, v6, v20, v0, v1
925 FILTER_6TAG_8BITS2_AVERAGE_WITH_1 v7, v2, v3, v4, v5, v6, v20, v0, v1
930 FILTER_6TAG_8BITS1_AVERAGE_WITH_1 v2, v3, v4, v5, v6, v7, v20, v0, v1
931 FILTER_6TAG_8BITS2_AVERAGE_WITH_1 v2, v3, v4, v5, v6, v7, v20, v0, v1
936 FILTER_6TAG_8BITS1_AVERAGE_WITH_1 v3, v4, v5, v6, v7, v2, v20, v0, v1
937 FILTER_6TAG_8BITS2_AVERAGE_WITH_1 v3, v4, v5, v6, v7, v2, v20, v0, v1
940 mov v3.16b, v5.16b
973 …VEC4_UADDL_8BITS v17, v20, v18, v21, v19, v22, v20, v23, v1, v3, v5, v7 //v1/v3/v5/v7 =src[-1]+src…
974 …VEC4_MLS_16BITS v1, v31, v3, v31, v5, v31, v7, v31, v0, v2, v4, v6 //v0/v2/v4/v6 -=5*(src[-1]+src…
975 …VEC4_UADDL_8BITS v18, v19, v19, v20, v20, v21, v21, v22, v1, v3, v5, v7 //v1/v3/v5/v7 =src[0]+src[…
976 …VEC4_MLA_16BITS v1, v30, v3, v30, v5, v30, v7, v30, v0, v2, v4, v6 //v0/v2/v4/v6 += 20*(src[0]+sr…
977 VEC4_SQRSHRUN_16BITS_SHIFT5 v0, v2, v4, v6, v1, v3, v5, v7
979 …VEC4_UADDL_8BITS v1, v19, v3, v20, v5, v21, v7, v22, v0, v2, v4, v6 //v0/v2/v4/v6 = average with s…
980 VEC4_RSHRN_16BITS_SHIFT1 v0, v2, v4, v6, v1, v3, v5, v7
982 VEC4_ST1_8BITS_8ELEMENT x2, x3, v1, v3, v5, v7 //store 8bytes*4row
1008 ld1 {v3.s}[0], [x0], x1 // v3=src[-1*stride]
1009 mov v2.s[1], v3.s[0]
1012 mov v3.s[1], v4.s[0]
1026 FILTER_6TAG_8BITS1_AVERAGE_WITH_1 v2, v3, v4, v5, v6, v7, v20, v0, v1
1034 ld1 {v3.s}[1], [x0], x1 // v2=src[6*stride]
1035 mov v3.s[0], v2.s[1]
1036 FILTER_6TAG_8BITS1_AVERAGE_WITH_1 v4, v5, v6, v7, v2, v3, v20, v0, v1
1039 mov v4.s[0], v3.s[1]
1045 mov v21.8b, v3.8b
1046 mov v3.8b, v7.8b
1068 ld1 {v3.16b}, [x0], x1 // v3=src[-1*stride]
1080 FILTER_6TAG_8BITS1 v2, v3, v4, v5, v6, v7, v20, v0, v1
1081 FILTER_6TAG_8BITS2 v2, v3, v4, v5, v6, v7, v20, v0, v1
1087 FILTER_6TAG_8BITS1 v3, v4, v5, v6, v7, v2, v20, v0, v1
1088 FILTER_6TAG_8BITS2 v3, v4, v5, v6, v7, v2, v20, v0, v1
1093 ld1 {v3.16b}, [x0], x1 // v3=src[3*stride]
1094 FILTER_6TAG_8BITS1 v4, v5, v6, v7, v2, v3, v20, v0, v1
1095 FILTER_6TAG_8BITS2 v4, v5, v6, v7, v2, v3, v20, v0, v1
1101 FILTER_6TAG_8BITS1 v5, v6, v7, v2, v3, v4, v20, v0, v1
1102 FILTER_6TAG_8BITS2 v5, v6, v7, v2, v3, v4, v20, v0, v1
1108 FILTER_6TAG_8BITS1 v6, v7, v2, v3, v4, v5, v20, v0, v1
1109 FILTER_6TAG_8BITS2 v6, v7, v2, v3, v4, v5, v20, v0, v1
1115 FILTER_6TAG_8BITS1 v7, v2, v3, v4, v5, v6, v20, v0, v1
1116 FILTER_6TAG_8BITS2 v7, v2, v3, v4, v5, v6, v20, v0, v1
1121 FILTER_6TAG_8BITS1 v2, v3, v4, v5, v6, v7, v20, v0, v1
1122 FILTER_6TAG_8BITS2 v2, v3, v4, v5, v6, v7, v20, v0, v1
1127 FILTER_6TAG_8BITS1 v3, v4, v5, v6, v7, v2, v20, v0, v1
1128 FILTER_6TAG_8BITS2 v3, v4, v5, v6, v7, v2, v20, v0, v1
1131 mov v3.16b, v5.16b
1164 …VEC4_UADDL_8BITS v17, v20, v18, v21, v19, v22, v20, v23, v1, v3, v5, v7 //v1/v3/v5/v7 =src[-1]+src…
1165 …VEC4_MLS_16BITS v1, v31, v3, v31, v5, v31, v7, v31, v0, v2, v4, v6 //v0/v2/v4/v6 -=5*(src[-1]+src…
1166 …VEC4_UADDL_8BITS v18, v19, v19, v20, v20, v21, v21, v22, v1, v3, v5, v7 //v1/v3/v5/v7 =src[0]+src[…
1167 …VEC4_MLA_16BITS v1, v30, v3, v30, v5, v30, v7, v30, v0, v2, v4, v6 //v0/v2/v4/v6 += 20*(src[0]+sr…
1168 VEC4_SQRSHRUN_16BITS_SHIFT5 v0, v2, v4, v6, v1, v3, v5, v7
1169 VEC4_ST1_8BITS_8ELEMENT x2, x3, v1, v3, v5, v7 //store 8bytes*4row
1195 ld1 {v3.s}[0], [x0], x1 // v3=src[-1*stride]
1196 mov v2.s[1], v3.s[0]
1199 mov v3.s[1], v4.s[0]
1213 FILTER_6TAG_8BITS1 v2, v3, v4, v5, v6, v7, v20, v0, v1
1221 ld1 {v3.s}[1], [x0], x1 // v2=src[6*stride]
1222 mov v3.s[0], v2.s[1]
1223 FILTER_6TAG_8BITS1 v4, v5, v6, v7, v2, v3, v20, v0, v1
1226 mov v4.s[0], v3.s[1]
1232 mov v21.8b, v3.8b
1233 mov v3.8b, v7.8b
1258 ld1 {v2.8b, v3.8b, v4.8b}, [x0], x1 // v2=src[-2*stride]
1273 FILTER_6TAG_8BITS_TO_16BITS1 v3, v6, v9, v12, v15, v18, v21, v0, v1
1284 ld1 {v2.8b, v3.8b, v4.8b}, [x0], x1 // v2=src[3*stride]
1287 FILTER_6TAG_8BITS_TO_16BITS1 v6, v9, v12, v15, v18, v3, v21, v0, v1
1301 FILTER_6TAG_8BITS_TO_16BITS1 v9, v12, v15, v18, v3, v6, v21, v0, v1
1315 FILTER_6TAG_8BITS_TO_16BITS1 v12, v15, v18, v3, v6, v9, v21, v0, v1
1329 FILTER_6TAG_8BITS_TO_16BITS1 v15, v18, v3, v6, v9, v12, v21, v0, v1
1343 FILTER_6TAG_8BITS_TO_16BITS1 v18, v3, v6, v9, v12, v15, v21, v0, v1
1357 FILTER_6TAG_8BITS_TO_16BITS1 v3, v6, v9, v12, v15, v18, v21, v0, v1
1368 ld1 {v2.8b, v3.8b, v4.8b}, [x0], x1 // v2=src[3*stride]
1371 FILTER_6TAG_8BITS_TO_16BITS1 v6, v9, v12, v15, v18, v3, v21, v0, v1
1390 mov v30.16b, v3.16b
1391 mov v3.16b, v9.16b
1426 ld1 {v3.16b}, [x0], x1 // v5=src[-1*stride]
1438 FILTER_6TAG_8BITS_TO_16BITS1 v2, v3, v4, v5, v6, v7, v20, v0, v1
1439 FILTER_6TAG_8BITS_TO_16BITS2 v2, v3, v4, v5, v6, v7, v21, v0, v1
1448 FILTER_6TAG_8BITS_TO_16BITS1 v3, v4, v5, v6, v7, v2, v20, v0, v1
1449 FILTER_6TAG_8BITS_TO_16BITS2 v3, v4, v5, v6, v7, v2, v21, v0, v1
1456 ld1 {v3.16b}, [x0], x1 // v3=src[3*stride]
1458 FILTER_6TAG_8BITS_TO_16BITS1 v4, v5, v6, v7, v2, v3, v20, v0, v1
1459 FILTER_6TAG_8BITS_TO_16BITS2 v4, v5, v6, v7, v2, v3, v21, v0, v1
1468 FILTER_6TAG_8BITS_TO_16BITS1 v5, v6, v7, v2, v3, v4, v20, v0, v1
1469 FILTER_6TAG_8BITS_TO_16BITS2 v5, v6, v7, v2, v3, v4, v21, v0, v1
1476 mov v5.16b, v3.16b
1477 mov v3.16b, v7.16b
1502 ld1 {v3.16b}, [x0], x1 // v3=src[-1*stride]
1514 FILTER_6TAG_8BITS_TO_16BITS1 v2, v3, v4, v5, v6, v7, v20, v0, v1
1515 FILTER_6TAG_8BITS_TO_16BITS2 v2, v3, v4, v5, v6, v7, v21, v0, v1
1519 FILTER_6TAG_8BITS_TO_16BITS1 v3, v4, v5, v6, v7, v2, v22, v0, v1
1520 FILTER_6TAG_8BITS_TO_16BITS2 v3, v4, v5, v6, v7, v2, v23, v0, v1
1533 ld1 {v3.16b}, [x0], x1 // v3=src[5*stride]
1535 FILTER_6TAG_8BITS_TO_16BITS1 v4, v5, v6, v7, v2, v3, v20, v0, v1
1536 FILTER_6TAG_8BITS_TO_16BITS2 v4, v5, v6, v7, v2, v3, v21, v0, v1
1539 FILTER_6TAG_8BITS_TO_16BITS1 v5, v6, v7, v2, v3, v4, v22, v0, v1
1540 FILTER_6TAG_8BITS_TO_16BITS2 v5, v6, v7, v2, v3, v4, v23, v0, v1
1551 mov v5.16b, v3.16b
1552 mov v3.16b, v7.16b
1627 ld1 {v3.16b}, [x4], x5 //read 16Byte : src1: 1 line
1637 AVERAGE_TWO_8BITS1 v16, v2, v3
1638 AVERAGE_TWO_8BITS2 v16, v2, v3
1671 ld1 {v3.8b}, [x4], x5 //read 8Byte : src1: 1 line
1683 AVERAGE_TWO_8BITS1 v16, v2, v3
1713 ld1 {v3.16b}, [x4], x5 //read 16Byte : src1: 1 line
1727 AVERAGE_TWO_8BITS1 v16, v2, v3
1728 AVERAGE_TWO_8BITS2 v16, v2, v3
1760 ld1 {v3.8b}, [x4], x5 //read 8Byte : src1: 1 line
1772 AVERAGE_TWO_8BITS1 v16, v2, v3
1856 rshrn v3.8b, v2.8h, #6
1857 st1 {v3.8b}, [x2], x3
1881 ext v3.8b, v2.8b, v2.8b, #1 // src[x+stride+1]
1886 zip1 v1.4s, v1.4s, v3.4s
1888 zip1 v3.4s, v3.4s, v19.4s
1893 umlal v16.8h, v3.8b, v7.8b
1917 ld1 {v2.16b, v3.16b}, [x0], x1 //only use 22(17+5); v2=src[-2]
1920 ext v5.16b, v2.16b, v3.16b, #1 //v5=src[-1]
1921 ext v6.16b, v2.16b, v3.16b, #2 //v6=src[0]
1922 ext v7.16b, v2.16b, v3.16b, #3 //v7=src[1]
1923 ext v16.16b, v2.16b, v3.16b, #4 //v16=src[2]
1924 ext v17.16b, v2.16b, v3.16b, #5 //v17=src[3]
1930 ext v21.8b, v3.8b, v3.8b, #7 // [0][1][2][3][4][5]XY-->O[0][1][2][3][4][5]X
1952 mov v3.d[0], v2.d[1]
1963 ext v21.8b, v3.8b, v3.8b, #7 // [0][1][2][3][4][5]XY-->O[0][1][2][3][4][5]X
2021 ld1 {v2.8b, v3.8b, v4.8b}, [x0], x1 // v2=src[-2*stride]
2036 FILTER_6TAG_8BITS_TO_16BITS1 v3, v6, v9, v12, v15, v18, v21, v0, v1
2049 ld1 {v2.8b, v3.8b, v4.8b}, [x0], x1 // v2=src[4*stride]
2052 FILTER_6TAG_8BITS_TO_16BITS1 v6, v9, v12, v15, v18, v3, v21, v0, v1
2068 FILTER_6TAG_8BITS_TO_16BITS1 v9, v12, v15, v18, v3, v6, v21, v0, v1
2084 FILTER_6TAG_8BITS_TO_16BITS1 v12, v15, v18, v3, v6, v9, v21, v0, v1
2100 FILTER_6TAG_8BITS_TO_16BITS1 v15, v18, v3, v6, v9, v12, v21, v0, v1
2116 FILTER_6TAG_8BITS_TO_16BITS1 v18, v3, v6, v9, v12, v15, v21, v0, v1
2132 FILTER_6TAG_8BITS_TO_16BITS1 v3, v6, v9, v12, v15, v18, v21, v0, v1
2145 ld1 {v2.8b, v3.8b, v4.8b}, [x0], x1 // v2=src[10*stride]
2148 FILTER_6TAG_8BITS_TO_16BITS1 v6, v9, v12, v15, v18, v3, v21, v0, v1
2169 mov v30.16b, v3.16b
2170 mov v3.16b, v9.16b
2188 FILTER_6TAG_8BITS_TO_16BITS1 v3, v6, v9, v12, v15, v18, v21, v0, v1
2225 ld1 {v3.16b}, [x0], x1 // v5=src[-1*stride]
2237 FILTER_6TAG_8BITS_TO_16BITS1 v2, v3, v4, v5, v6, v7, v20, v0, v1
2238 FILTER_6TAG_8BITS_TO_16BITS2 v2, v3, v4, v5, v6, v7, v21, v0, v1
2249 FILTER_6TAG_8BITS_TO_16BITS1 v3, v4, v5, v6, v7, v2, v20, v0, v1
2250 FILTER_6TAG_8BITS_TO_16BITS2 v3, v4, v5, v6, v7, v2, v21, v0, v1
2259 ld1 {v3.16b}, [x0], x1 // v3=src[5*stride]
2261 FILTER_6TAG_8BITS_TO_16BITS1 v4, v5, v6, v7, v2, v3, v20, v0, v1
2262 FILTER_6TAG_8BITS_TO_16BITS2 v4, v5, v6, v7, v2, v3, v21, v0, v1
2273 FILTER_6TAG_8BITS_TO_16BITS1 v5, v6, v7, v2, v3, v4, v20, v0, v1
2274 FILTER_6TAG_8BITS_TO_16BITS2 v5, v6, v7, v2, v3, v4, v21, v0, v1
2283 mov v5.16b, v3.16b
2284 mov v3.16b, v7.16b
2296 FILTER_6TAG_8BITS_TO_16BITS1 v2, v3, v4, v5, v6, v7, v20, v0, v1
2297 FILTER_6TAG_8BITS_TO_16BITS2 v2, v3, v4, v5, v6, v7, v21, v0, v1
2325 ld1 {v3.16b}, [x0], x1 // v5=src[-1*stride]
2337 FILTER_6TAG_8BITS_TO_16BITS1 v2, v3, v4, v5, v6, v7, v20, v0, v1
2338 FILTER_6TAG_8BITS_TO_16BITS2 v2, v3, v4, v5, v6, v7, v21, v0, v1
2348 FILTER_6TAG_8BITS_TO_16BITS1 v3, v4, v5, v6, v7, v2, v20, v0, v1
2349 FILTER_6TAG_8BITS_TO_16BITS2 v3, v4, v5, v6, v7, v2, v21, v0, v1
2357 ld1 {v3.16b}, [x0], x1 // v3=src[5*stride]
2359 FILTER_6TAG_8BITS_TO_16BITS1 v4, v5, v6, v7, v2, v3, v20, v0, v1
2360 FILTER_6TAG_8BITS_TO_16BITS2 v4, v5, v6, v7, v2, v3, v21, v0, v1
2370 FILTER_6TAG_8BITS_TO_16BITS1 v5, v6, v7, v2, v3, v4, v20, v0, v1
2371 FILTER_6TAG_8BITS_TO_16BITS2 v5, v6, v7, v2, v3, v4, v21, v0, v1
2379 mov v5.16b, v3.16b
2380 mov v3.16b, v7.16b
2392 FILTER_6TAG_8BITS_TO_16BITS1 v2, v3, v4, v5, v6, v7, v20, v0, v1
2393 FILTER_6TAG_8BITS_TO_16BITS2 v2, v3, v4, v5, v6, v7, v21, v0, v1
2416 ld1 {v3.16b}, [x0], x1 // v3=src[-1*stride]
2428 FILTER_6TAG_8BITS1 v2, v3, v4, v5, v6, v7, v20, v0, v1
2429 FILTER_6TAG_8BITS2 v2, v3, v4, v5, v6, v7, v20, v0, v1
2435 FILTER_6TAG_8BITS1 v3, v4, v5, v6, v7, v2, v20, v0, v1
2436 FILTER_6TAG_8BITS2 v3, v4, v5, v6, v7, v2, v20, v0, v1
2441 ld1 {v3.16b}, [x0], x1 // v3=src[5*stride]
2442 FILTER_6TAG_8BITS1 v4, v5, v6, v7, v2, v3, v20, v0, v1
2443 FILTER_6TAG_8BITS2 v4, v5, v6, v7, v2, v3, v20, v0, v1
2449 FILTER_6TAG_8BITS1 v5, v6, v7, v2, v3, v4, v20, v0, v1
2450 FILTER_6TAG_8BITS2 v5, v6, v7, v2, v3, v4, v20, v0, v1
2456 FILTER_6TAG_8BITS1 v6, v7, v2, v3, v4, v5, v20, v0, v1
2457 FILTER_6TAG_8BITS2 v6, v7, v2, v3, v4, v5, v20, v0, v1
2463 FILTER_6TAG_8BITS1 v7, v2, v3, v4, v5, v6, v20, v0, v1
2464 FILTER_6TAG_8BITS2 v7, v2, v3, v4, v5, v6, v20, v0, v1
2469 FILTER_6TAG_8BITS1 v2, v3, v4, v5, v6, v7, v20, v0, v1
2470 FILTER_6TAG_8BITS2 v2, v3, v4, v5, v6, v7, v20, v0, v1
2475 FILTER_6TAG_8BITS1 v3, v4, v5, v6, v7, v2, v20, v0, v1
2476 FILTER_6TAG_8BITS2 v3, v4, v5, v6, v7, v2, v20, v0, v1
2479 mov v3.16b, v5.16b
2490 FILTER_6TAG_8BITS1 v2, v3, v4, v5, v6, v7, v20, v0, v1
2491 FILTER_6TAG_8BITS2 v2, v3, v4, v5, v6, v7, v20, v0, v1
2509 ld1 {v3.8b}, [x0], x1 // v3=src[-1*stride]
2520 FILTER_6TAG_8BITS1 v2, v3, v4, v5, v6, v7, v20, v0, v1
2525 FILTER_6TAG_8BITS1 v3, v4, v5, v6, v7, v2, v20, v0, v1
2529 ld1 {v3.8b}, [x0], x1 // v3=src[5*stride]
2530 FILTER_6TAG_8BITS1 v4, v5, v6, v7, v2, v3, v20, v0, v1
2535 FILTER_6TAG_8BITS1 v5, v6, v7, v2, v3, v4, v20, v0, v1
2538 mov v5.16b, v3.16b
2539 mov v3.16b, v7.16b
2549 FILTER_6TAG_8BITS1 v2, v3, v4, v5, v6, v7, v20, v0, v1
2568 ld1 {v3.8b}, [x0], x1 // v3=src[-1*stride]
2579 FILTER_6TAG_8BITS1 v2, v3, v4, v5, v6, v7, v20, v0, v1
2584 FILTER_6TAG_8BITS1 v3, v4, v5, v6, v7, v2, v20, v0, v1
2588 ld1 {v3.8b}, [x0], x1 // v3=src[5*stride]
2589 FILTER_6TAG_8BITS1 v4, v5, v6, v7, v2, v3, v20, v0, v1
2594 FILTER_6TAG_8BITS1 v5, v6, v7, v2, v3, v4, v20, v0, v1
2597 mov v5.16b, v3.16b
2598 mov v3.16b, v7.16b
2608 FILTER_6TAG_8BITS1 v2, v3, v4, v5, v6, v7, v20, v0, v1