Lines Matching refs:m3
142 mova m3, [aq+mmsize*3]
149 mova [dstq+strideq*0+48], m3
153 mova [dstq+strideq*1+48], m3
161 mova m3, [lq]
164 pshufw m0, m3, q3333
165 pshufw m1, m3, q2222
166 pshufw m2, m3, q1111
167 pshufw m3, m3, q0000
171 mova [dstq+stride3q ], m3
179 punpckhwd m3, m2, m2
180 pshufd m0, m3, q3333
181 pshufd m1, m3, q2222
184 pshufd m0, m3, q1111
185 pshufd m1, m3, q0000
205 movh m3, [lq+cntq*8]
206 punpcklwd m3, m3
207 pshufd m0, m3, q3333
208 pshufd m1, m3, q2222
209 pshufd m2, m3, q1111
210 pshufd m3, m3, q0000
217 mova [dstq+stride3q + 0], m3
218 mova [dstq+stride3q +16], m3
229 movh m3, [lq+cntq*8]
230 punpcklwd m3, m3
231 pshufd m0, m3, q3333
232 pshufd m1, m3, q2222
233 pshufd m2, m3, q1111
234 pshufd m3, m3, q0000
247 mova [dstq+stride3q + 0], m3
248 mova [dstq+stride3q +16], m3
249 mova [dstq+stride3q +32], m3
250 mova [dstq+stride3q +48], m3
481 mova m3, [lq]
487 pshufw m0, m3, q3333
488 pshufw m1, m3, q2222
489 pshufw m2, m3, q1111
490 pshufw m3, m3, q0000
494 paddw m3, m4
499 pmaxsw m3, m4
503 pminsw m3, m5
507 mova [dstq+stride3q ], m3
528 movh m3, [lq+cntq*8]
529 punpcklwd m3, m3
530 pshufd m0, m3, q3333
531 pshufd m1, m3, q2222
532 pshufd m2, m3, q1111
533 pshufd m3, m3, q0000
537 paddw m3, m5
541 pmaxsw m3, m6
545 pminsw m3, m4
549 mova [dstq+stride3q ], m3
574 movd m3, [lq+cntq*4]
575 punpcklwd m3, m3
576 pshufd m2, m3, q1111
577 pshufd m3, m3, q0000
580 paddw m1, m3, m4
581 paddw m3, m5
585 pmaxsw m3, m6
589 pminsw m3, m7
593 mova [dstq+strideq*1+16], m3
634 pinsrw m3, [lq+cntq*2], 0
635 punpcklwd m3, m3
636 pshufd m3, m3, q0000
637 paddw m0, m3, m4
638 paddw m1, m3, m5
639 paddw m2, m3, m6
640 paddw m3, m7
644 pmaxsw m3, reg_min
648 pminsw m3, reg_max
652 mova [dstq+strideq*0+48], m3
728 shufps m3, m0, m1, q2121 ; DEFGHhhh
741 mova [dstq+strideq*0], m3
751 mova m3, [aq+mmsize] ; ijklmnop
752 PALIGNR m1, m3, m0, 2, m4 ; bcdefghi
753 PALIGNR m2, m3, m0, 4, m4 ; cdefghij
758 SHIFT_RIGHTx2 m2, m1, m3, m4 ; jklmnopp/klmnoppp
773 PALIGNR m3, m1, m0, 2, m4
774 mova m0, m3
786 mova m3, [aq+mmsize*3] ; yz012345
793 PALIGNR m4, m3, m2, 2, m6
794 PALIGNR m5, m3, m2, 4, m6
799 SHIFT_RIGHTx2 m4, m5, m3, m6
811 mova [dstq+stride8q*0+48], m3
814 mova [dstq+stride8q*1+32], m3
817 mova [dstq+stride8q*2+16], m3
820 mova [dstq+stride24q + 0], m3
828 vpalignr m2, m3, m2, 2
834 PALIGNR m5, m3, m2, 2, m6
837 SHIFT_RIGHT m3, m3, m6
857 vpalignr m3, m2, m0, 2 ; bcdefghijklmnopp
867 vpalignr m3, m2, m0, 2
869 mova [dstq+strideq*1], m3
871 vpalignr m3, m2, m0, 6
873 mova [dstq+stride3q ], m3
876 vpalignr m3, m2, m0, 10
878 mova [dstq+strideq*1], m3
880 vpalignr m3, m2, m0, 14
881 mova [dstq+stride3q ], m3
896 vpalignr m3, m5, m0, 4 ; cdefghijklmnopqr
900 vpalignr m3, m5, m1, 4 ; stuvwxyz01234555
911 vpalignr m3, m5, m0, 2
913 mova [dstq+strideq*1 + 0], m3
915 vpalignr m3, m5, m0, 4
917 mova [dstq+strideq*2 + 0], m3
919 vpalignr m3, m5, m0, 6
921 mova [dstq+stride3q*1+ 0], m3
924 vpalignr m3, m5, m0, 8
926 mova [dstq+strideq*0 + 0], m3
928 vpalignr m3, m5, m0, 10
930 mova [dstq+strideq*1 + 0], m3
932 vpalignr m3, m5, m0, 12
934 mova [dstq+strideq*2+ 0], m3
936 vpalignr m3, m5, m0, 14
938 mova [dstq+stride3q+ 0], m3
940 vpalignr m3, m5, m0, 16
942 vperm2i128 m5, m3, m4, q0201
944 mova m0, m3
976 psrldq m3, m2, 2 ; bcdefgh.
986 movh [dstq+ stride3q +8], m3
988 PALIGNR m1, m3, m2, 2, m0
989 psrldq m3, 2
991 movh [dstq+ strideq*2+8], m3
993 PALIGNR m2, m3, m1, 2, m0
994 psrldq m3, 2
996 movh [dstq+ strideq*1+8], m3
998 PALIGNR m1, m3, m2, 2, m0
999 psrldq m3, 2
1001 movh [dstq+ strideq*0+8], m3
1009 movu m3, [aq+mmsize-2] ; hijklmno
1014 PALIGNR m5, m4, 2, m3 ; bcdefghi
1016 PALIGNR m2, m1, 2, m3 ; tuvwxyz*
1017 PALIGNR m4, m2, 2, m3 ; uvwxyz*a
1019 PALIGNR m1, m0, 2, m3 ; lmnopqrs
1020 PALIGNR m2, m1, 2, m3 ; mnopqrst
1056 movu m3, [aq+mmsize*2-2] ; a[15-22]
1059 mova m3, [aq+mmsize*1] ; a[8-15]
1061 PALIGNR m1, m3, 2, m5 ; a[9-16]
1065 PALIGNR m3, m4, 2, m6 ; a[1-8]
1084 mova m3, [lq+mmsize*0] ; l[0-7]
1085 PALIGNR m1, m3, 2, m0 ; l[1-8]
1111 mova [dst8q+stride8q*0+16], m3
1116 mova [dst8q+stride8q*1+32], m3
1121 mova [dst8q+stride8q*2+48], m3
1130 vpalignr m4, m3, m4, 2
1131 vpalignr m3, m1, m3, 2
1145 PALIGNR m2, m3, m4, 2, m0
1147 PALIGNR m2, m1, m3, 2, m0
1148 mova m3, m2
1153 PALIGNR m2, m0, m1, 2, m3
1157 PALIGNR m1, m2, m0, 2, m3
1181 vperm2i128 m3, m0, m1, q0201 ; stuvwxyz*abcdefg
1183 vpalignr m4, m3, m0, 2 ; lmnopqrstuvwxyz*
1184 vpalignr m5, m3, m0, 4 ; mnopqrstuvwxyz*a
1192 vpalignr m3, m5, m0, 2
1194 mova [dst3q+stride5q*2], m3 ; 14
1196 vpalignr m3, m5, m0, 4
1199 mova [dst3q+stride5q*2], m3 ; 13
1202 vpalignr m3, m5, m0, 6
1204 mova [dstq+stride3q*4], m3 ; 12
1206 vpalignr m3, m5, m0, 8
1208 mova [dst3q+strideq*8], m3 ; 11
1210 vpalignr m3, m5, m0, 10
1212 mova [dstq+stride5q*2], m3 ; 10
1214 vpalignr m3, m5, m0, 12
1216 mova [dst3q+stride3q*2], m3 ; 9
1218 vpalignr m3, m5, m0, 14
1220 mova [dstq+strideq*8], m3 ; 8
1230 vpalignr m3, m2, m0, 2 ; bcdefghijklmnopp
1231 vperm2i128 m4, m3, m1, q0201 ; jklmnopppppppppp
1236 pavgw m3, m0 ; abcdefghijklmnop
1243 mova [dstq+strideq*0], m3 ; 0 abcdefghijklmnop
1245 vpalignr m0, m4, m3, 2
1249 vpalignr m0, m4, m3, 4
1253 vpalignr m0, m4, m3, 6
1257 vpalignr m0, m4, m3, 8
1261 vpalignr m0, m4, m3, 10
1263 vpalignr m0, m4, m3, 12
1265 vpalignr m0, m4, m3, 14
1280 vpalignr m3, m2, m1, 2 ; lmnopqrstuvwxyz*
1283 pavgw m3, m1 ; klmnopqrstuvwxyz
1287 vpunpcklwd m0, m3, m4 ; kLlMmNnOsTtUuVvW
1288 vpunpckhwd m1, m3, m4 ; oPpQqRrSwXxYyZz#
1289 vperm2i128 m3, m1, m0, q0002 ; kLlMmNnOoPpQqRrS
1292 vperm2i128 m1, m3, m4, q0201 ; oPpQqRrSsTtUuVvW
1298 mova [dst5q+stride5q*2], m3 ; 15 kLlMmNnOoPpQqRrS
1314 vpalignr m5, m1, m3, 12
1323 vpalignr m5, m1, m3, 4
1326 vpalignr m5, m1, m3, 8
1335 mova m3, [aq+mmsize*0+0] ; abcdefghijklmnop
1345 vperm2i128 m5, m3, m4, q0201 ; ijklmnopqrstuvwx
1346 vpalignr m6, m5, m3, 2 ; bcdefghijklmnopq
1348 movu m3, [aq+mmsize*1-2] ; pqrstuvwxyz01234
1354 vperm2i128 m8, m2, m3, q0201 ; IJKLMNOPQRSTUVWX
1422 mova m8, m3
1456 mova m3, [pb_2to15_14_15]
1458 SHIFT_RIGHTx2 m1, m2, m0, m3 ; bcdefghh/cdefghhh
1466 SHIFT_RIGHT m1, m1, m3
1467 SHIFT_RIGHT m2, m2, m3
1471 SHIFT_RIGHT m1, m1, m3
1472 SHIFT_RIGHT m2, m2, m3
1475 SHIFT_RIGHT m1, m1, m3
1476 SHIFT_RIGHT m2, m2, m3
1485 PALIGNR m2, m1, m0, 2, m3
1486 PALIGNR m3, m1, m0, 4, m4
1501 mova [dstq+strideq*1+ 0], m3
1506 vpalignr m3, m0, m3, 2
1510 PALIGNR m5, m0, m3, 2, m4
1511 mova m3, m5
1535 PALIGNR m3, m0, m2, 4, m6
1567 mova [dstq+strideq*1+32], m3
1576 mova [dstq+stride17q+16], m3
1587 vpalignr m5, m3, m5, 2
1588 vpalignr m3, m1, m3, 2
1594 PALIGNR m3, m4, m6, 2, m1
1595 mova m6, m3
1596 PALIGNR m3, m2, m4, 2, m1
1597 mova m4, m3
1598 PALIGNR m3, m0, m2, 2, m1
1599 mova m2, m3
1600 PALIGNR m3, m5, m7, 2, m1
1601 mova m7, m3
1608 PALIGNR m6, m3, m5, 2, m7
1610 PALIGNR m6, m1, m3, 2, m7
1611 mova m3, m6
1676 PALIGNR m3, m1, m2, 14, m4 ; z*abcdef
1686 mova [dstq+strideq*1], m3
1689 PALIGNR m3, m4, 14, m1
1692 mova [dstq+stride3q ], m3
1696 PALIGNR m3, m4, 14, m1
1699 mova [dstq+strideq*1], m3
1702 PALIGNR m3, m4, 14, m4
1704 mova [dstq+stride3q ], m3
1710 mova m3, [aq] ; abcdefgh
1718 pavgw m3, m1
1736 mova [dstq+strideq*0+ 0], m3
1741 PALIGNR m2, m3, 14, m4
1742 PALIGNR m3, m7, 14, m4
1755 movu m3, [aq+mmsize*3-2] ; a[23-30]
1760 pavgw m3, m4
1819 mova [dstq+strideq*0+48], m3
1837 PALIGNR m3, m2, 14, m5
1865 mova [dstq+strideq*0+16], m3
1870 mova [dstq+stride16q+32], m3
1874 PALIGNR m2, m3, 14, m6
1875 PALIGNR m3, m4, 14, m6
1919 mova m3, [pb_2to15_14_15]
1921 SHIFT_RIGHTx2 m1, m2, m0, m3
1926 pshufd m3, m2, q3332
1933 mova [dstq+stride3q*2], m3
1938 PALIGNR m0, m2, m1, 4, m3
1943 pshufd m3, m2, q3332
1947 mova [dstq+stride3q*2], m3
1952 mova m3, [lq+mmsize]
1961 SHIFT_RIGHTx2 m0, m4, m3, m5
1963 pavgw m3, m0
1974 mova [dstq+strideq *4+16], m3
1975 mova [dstq+strideq *8+ 0], m3
1982 vpalignr m2, m3, m2, 4
1983 vpalignr m3, m4, m3, 4
1988 PALIGNR m5, m3, m2, 4, m6
1990 PALIGNR m5, m4, m3, 4, m6
1991 mova m3, m5
2009 movu m3, [lq+mmsize*1+2]
2012 pavgw m3, m4
2048 mova [dstq+strideq *0+32], m3
2051 mova [dstq+stride4q*1+16], m3
2054 mova [dstq+stride4q*2+ 0], m3
2060 vpalignr m0, m3, m0, 4
2061 vpalignr m3, m2, m3, 4
2069 PALIGNR m6, m3, m0, 4, m7
2071 PALIGNR m6, m2, m3, 4, m7
2072 mova m3, m6
2113 PALIGNR m0, m5, m2, 4, m3
2115 PALIGNR m0, m4, m5, 4, m3
2117 PALIGNR m0, m7, m4, 4, m3
2119 PALIGNR m0, m6, m7, 4, m3
2121 PALIGNR m0, m1, m6, 4, m3
2164 PALIGNR m2, m1, m0, 2, m3
2165 PALIGNR m3, m1, m0, 4, m4
2180 mova [dst8q+mstrideq*4], m3
2182 vpalignr m2, m3, m2, 4
2183 vpalignr m3, m1, m3, 4
2185 PALIGNR m0, m3, m2, 4, m4
2187 PALIGNR m0, m1, m3, 4, m4
2188 mova m3, m0
2203 PALIGNR m3, m5, m4, 2, m6
2206 pavgw m3, m4
2227 mova [dstq+mstrideq *8+ 0], m3
2230 mova [dstq+mstrideq *4+16], m3
2235 vpalignr m0, m3, m0, 4
2236 vpalignr m3, m2, m3, 4
2242 PALIGNR m6, m3, m0, 4, m7
2244 PALIGNR m6, m2, m3, 4, m7
2245 mova m3, m6
2265 movu m3, [lq+mmsize*1+2]
2268 pavgw m3, m4
2292 movu m3, [aq+mmsize*1+2]
2297 mova m3, [aq+mmsize*2+0]
2300 movu m3, [aq+mmsize*3-2]
2301 psrldq m6, m3, 2
2302 psrldq m7, m3, 4
2374 vpalignr m2, m3, m2, 4
2390 PALIGNR m3, m4, m5, 4, m2
2391 mova m5, m3
2392 PALIGNR m3, m7, m4, 4, m2
2393 mova m4, m3
2394 PALIGNR m3, m6, m7, 4, m2
2395 mova m7, m3
2396 PALIGNR m3, m0, m6, 4, m2
2397 mova m6, m3
2398 PALIGNR m3, m1, m0, 4, m2
2399 mova m0, m3
2404 PALIGNR m3, m2, m1, 4, m0
2405 mova m1, m3
2406 PALIGNR m3, reg_sh, m2, 4, m0
2407 mova m2, m3
2413 psrldq m3, 4
2434 mova [dstq+stride3q*4+32], m3
2437 mova [dstq+strideq* 8+16], m3
2440 mova [dstq+strideq* 4+ 0], m3
2451 vpalignr m0, m3, m0, 4
2452 vpalignr m3, m2, m3, 4
2464 PALIGNR m6, m3, m0, 4, m7
2466 PALIGNR m6, m2, m3, 4, m7
2467 mova m3, m6