Lines Matching refs:m0
88 mova m0, [aq]
91 mova [dstq+strideq*0], m0
92 mova [dstq+strideq*1], m0
93 mova [dstq+strideq*2], m0
94 mova [dstq+stride3q ], m0
100 mova m0, [aq]
103 mova [dstq+strideq*0], m0
104 mova [dstq+strideq*1], m0
105 mova [dstq+strideq*2], m0
106 mova [dstq+stride3q ], m0
108 mova [dstq+strideq*0], m0
109 mova [dstq+strideq*1], m0
110 mova [dstq+strideq*2], m0
111 mova [dstq+stride3q ], m0
117 mova m0, [aq]
123 mova [dstq+strideq*0+ 0], m0
125 mova [dstq+strideq*1+ 0], m0
127 mova [dstq+strideq*2+ 0], m0
129 mova [dstq+stride3q + 0], m0
139 mova m0, [aq+mmsize*0]
146 mova [dstq+strideq*0+ 0], m0
150 mova [dstq+strideq*1+ 0], m0
164 pshufw m0, m3, q3333
168 mova [dstq+strideq*0], m0
180 pshufd m0, m3, q3333
182 mova [dstq+strideq*0], m0
184 pshufd m0, m3, q1111
186 mova [dstq+strideq*2], m0
190 pshufd m0, m2, q3333
192 mova [dstq+strideq*0], m0
194 pshufd m0, m2, q1111
196 mova [dstq+strideq*2], m0
207 pshufd m0, m3, q3333
211 mova [dstq+strideq*0+ 0], m0
212 mova [dstq+strideq*0+16], m0
231 pshufd m0, m3, q3333
235 mova [dstq+strideq*0+ 0], m0
236 mova [dstq+strideq*0+16], m0
237 mova [dstq+strideq*0+32], m0
238 mova [dstq+strideq*0+48], m0
258 mova m0, [lq]
259 paddw m0, [aq]
262 pmaddwd m0, [pw_1]
263 pshufw m1, m0, q3232
264 paddd m0, [pd_4]
265 paddd m0, m1
266 psrad m0, 3
267 pshufw m0, m0, q0000
268 mova [dstq+strideq*0], m0
269 mova [dstq+strideq*1], m0
270 mova [dstq+strideq*2], m0
271 mova [dstq+stride3q ], m0
276 mova m0, [lq]
277 paddw m0, [aq]
280 pmaddwd m0, [pw_1]
281 pshufd m1, m0, q3232
282 paddd m0, m1
283 pshufd m1, m0, q1111
284 paddd m0, [pd_8]
285 paddd m0, m1
286 psrad m0, 4
287 pshuflw m0, m0, q0000
288 punpcklqdq m0, m0
289 mova [dstq+strideq*0], m0
290 mova [dstq+strideq*1], m0
291 mova [dstq+strideq*2], m0
292 mova [dstq+stride3q ], m0
294 mova [dstq+strideq*0], m0
295 mova [dstq+strideq*1], m0
296 mova [dstq+strideq*2], m0
297 mova [dstq+stride3q ], m0
302 mova m0, [lq]
303 paddw m0, [lq+mmsize]
304 paddw m0, [aq]
305 paddw m0, [aq+mmsize]
309 pmaddwd m0, [pw_1]
310 pshufd m1, m0, q3232
311 paddd m0, m1
312 pshufd m1, m0, q1111
313 paddd m0, [pd_16]
314 paddd m0, m1
315 psrad m0, 5
316 pshuflw m0, m0, q0000
317 punpcklqdq m0, m0
319 mova [dstq+strideq*0+ 0], m0
320 mova [dstq+strideq*0+16], m0
321 mova [dstq+strideq*1+ 0], m0
322 mova [dstq+strideq*1+16], m0
323 mova [dstq+strideq*2+ 0], m0
324 mova [dstq+strideq*2+16], m0
325 mova [dstq+stride3q + 0], m0
326 mova [dstq+stride3q +16], m0
334 mova m0, [lq+mmsize*0]
335 paddw m0, [lq+mmsize*1]
336 paddw m0, [lq+mmsize*2]
337 paddw m0, [lq+mmsize*3]
338 paddw m0, [aq+mmsize*0]
339 paddw m0, [aq+mmsize*1]
340 paddw m0, [aq+mmsize*2]
341 paddw m0, [aq+mmsize*3]
345 pmaddwd m0, [pw_1]
346 pshufd m1, m0, q3232
347 paddd m0, m1
348 pshufd m1, m0, q1111
349 paddd m0, [pd_32]
350 paddd m0, m1
351 psrad m0, 6
352 pshuflw m0, m0, q0000
353 punpcklqdq m0, m0
355 mova [dstq+strideq*0+ 0], m0
356 mova [dstq+strideq*0+16], m0
357 mova [dstq+strideq*0+32], m0
358 mova [dstq+strideq*0+48], m0
359 mova [dstq+strideq*1+ 0], m0
360 mova [dstq+strideq*1+16], m0
361 mova [dstq+strideq*1+32], m0
362 mova [dstq+strideq*1+48], m0
371 mova m0, [%2]
374 pmaddwd m0, [pw_1]
375 pshufw m1, m0, q3232
376 paddd m0, [pd_2]
377 paddd m0, m1
378 psrad m0, 2
379 pshufw m0, m0, q0000
380 mova [dstq+strideq*0], m0
381 mova [dstq+strideq*1], m0
382 mova [dstq+strideq*2], m0
383 mova [dstq+stride3q ], m0
388 mova m0, [%2]
391 pmaddwd m0, [pw_1]
392 pshufd m1, m0, q3232
393 paddd m0, m1
394 pshufd m1, m0, q1111
395 paddd m0, [pd_4]
396 paddd m0, m1
397 psrad m0, 3
398 pshuflw m0, m0, q0000
399 punpcklqdq m0, m0
400 mova [dstq+strideq*0], m0
401 mova [dstq+strideq*1], m0
402 mova [dstq+strideq*2], m0
403 mova [dstq+stride3q ], m0
405 mova [dstq+strideq*0], m0
406 mova [dstq+strideq*1], m0
407 mova [dstq+strideq*2], m0
408 mova [dstq+stride3q ], m0
413 mova m0, [%2]
414 paddw m0, [%2+mmsize]
418 pmaddwd m0, [pw_1]
419 pshufd m1, m0, q3232
420 paddd m0, m1
421 pshufd m1, m0, q1111
422 paddd m0, [pd_8]
423 paddd m0, m1
424 psrad m0, 4
425 pshuflw m0, m0, q0000
426 punpcklqdq m0, m0
428 mova [dstq+strideq*0+ 0], m0
429 mova [dstq+strideq*0+16], m0
430 mova [dstq+strideq*1+ 0], m0
431 mova [dstq+strideq*1+16], m0
432 mova [dstq+strideq*2+ 0], m0
433 mova [dstq+strideq*2+16], m0
434 mova [dstq+stride3q + 0], m0
435 mova [dstq+stride3q +16], m0
443 mova m0, [%2+mmsize*0]
444 paddw m0, [%2+mmsize*1]
445 paddw m0, [%2+mmsize*2]
446 paddw m0, [%2+mmsize*3]
449 pmaddwd m0, [pw_1]
450 pshufd m1, m0, q3232
451 paddd m0, m1
452 pshufd m1, m0, q1111
453 paddd m0, [pd_16]
454 paddd m0, m1
455 psrad m0, 5
456 pshuflw m0, m0, q0000
457 punpcklqdq m0, m0
459 mova [dstq+strideq*0+ 0], m0
460 mova [dstq+strideq*0+16], m0
461 mova [dstq+strideq*0+32], m0
462 mova [dstq+strideq*0+48], m0
463 mova [dstq+strideq*1+ 0], m0
464 mova [dstq+strideq*1+16], m0
465 mova [dstq+strideq*1+32], m0
466 mova [dstq+strideq*1+48], m0
482 movd m0, [aq-4]
483 pshufw m0, m0, q1111
484 psubw m4, m0
487 pshufw m0, m3, q3333
491 paddw m0, m4
496 pmaxsw m0, m4
500 pminsw m0, m5
504 mova [dstq+strideq*0], m0
520 movd m0, [aq-4]
521 pshuflw m0, m0, q1111
522 punpcklqdq m0, m0
523 psubw m5, m0
530 pshufd m0, m3, q3333
534 paddw m0, m5
538 pmaxsw m0, m6
542 pminsw m0, m4
546 mova [dstq+strideq*0], m0
566 movd m0, [aq-4]
567 pshuflw m0, m0, q1111
568 punpcklqdq m0, m0
569 psubw m4, m0
570 psubw m5, m0
578 paddw m0, m2, m4
582 pmaxsw m0, m6
586 pminsw m0, m7
590 mova [dstq+strideq*0+ 0], m0
605 mova m0, [pw_1023]
614 mova [rsp+ 0], m0
624 movd m0, [aq-4]
625 pshuflw m0, m0, q1111
626 punpcklqdq m0, m0
627 psubw m4, m0
628 psubw m5, m0
629 psubw m6, m0
630 psubw m7, m0
637 paddw m0, m3, m4
641 pmaxsw m0, reg_min
645 pminsw m0, reg_max
649 mova [dstq+strideq*0+ 0], m0
659 mova m0, [pw_4095]
705 pshufhw m0, m1, q3310 ; abcdefhh
709 pshufd m1, m0, q3321 ; DEFGh...
710 movh [dstq+strideq*0], m0
713 psrldq m0, 2 ; CDEFGh..
715 movh [dstq+strideq*0], m0
721 mova m0, [aq] ; abcdefgh
725 SHIFT_RIGHTx2 m1, m2, m0, m4 ; bcdefghh/cdefghhh
727 shufps m1, m0, m2, q3332 ; FGHhhhhh
728 shufps m3, m0, m1, q2121 ; DEFGHhhh
732 mova [dstq+strideq*0], m0
734 SHIFT_RIGHT m0, m0, m4 ; CDEFGHhh
736 pshufd m2, m0, q3321 ; EFGHhhhh
737 mova [dstq+strideq*1], m0
750 mova m0, [aq] ; abcdefgh
752 PALIGNR m1, m3, m0, 2, m4 ; bcdefghi
753 PALIGNR m2, m3, m0, 4, m4 ; cdefghij
765 mova [dstq+strideq*0+ 0], m0
771 vpalignr m0, m1, m0, 2
773 PALIGNR m3, m1, m0, 2, m4
774 mova m0, m3
783 mova m0, [aq+mmsize*0] ; abcdefgh
787 PALIGNR m4, m1, m0, 2, m6
788 PALIGNR m5, m1, m0, 4, m6
808 mova [dstq+stride8q*0+ 0], m0
826 vpalignr m0, m1, m0, 2
830 PALIGNR m5, m1, m0, 2, m6
831 mova m0, m5
854 mova m0, [aq] ; abcdefghijklmnop
856 vperm2i128 m2, m0, m1, q0201 ; ijklmnoppppppppp
857 vpalignr m3, m2, m0, 2 ; bcdefghijklmnopp
858 vpalignr m4, m2, m0, 4 ; cdefghijklmnoppp
860 vperm2i128 m2, m0, m1, q0201 ; JKLMNOPppppppppp
866 mova [dstq+strideq*0], m0
867 vpalignr m3, m2, m0, 2
868 vpalignr m4, m2, m0, 4
871 vpalignr m3, m2, m0, 6
872 vpalignr m4, m2, m0, 8
876 vpalignr m3, m2, m0, 10
877 vpalignr m4, m2, m0, 12
880 vpalignr m3, m2, m0, 14
883 mova m0, m2
891 mova m0, [aq+mmsize*0+ 0] ; abcdefghijklmnop
894 vperm2i128 m5, m0, m1, q0201 ; ijklmnopqrstuvwx
895 vpalignr m2, m5, m0, 2 ; bcdefghijklmnopq
896 vpalignr m3, m5, m0, 4 ; cdefghijklmnopqr
903 vperm2i128 m5, m0, m1, q0201 ; JKLMNOPQRSTUVWXY
909 mova [dstq+strideq*0 + 0], m0
911 vpalignr m3, m5, m0, 2
915 vpalignr m3, m5, m0, 4
919 vpalignr m3, m5, m0, 6
924 vpalignr m3, m5, m0, 8
928 vpalignr m3, m5, m0, 10
932 vpalignr m3, m5, m0, 12
936 vpalignr m3, m5, m0, 14
940 vpalignr m3, m5, m0, 16
944 mova m0, m3
954 movh m0, [lq] ; wxyz....
955 movhps m0, [aq-2] ; wxyz*abc
957 PALIGNR m1, m0, 2, m2 ; xyz*abcd
963 movh [dstq+stride3q ], m0
964 psrldq m0, 2 ; YZ#ABC..
965 movh [dstq+strideq*2], m0
966 psrldq m0, 2 ; Z#ABC...
967 movh [dstq+strideq*1], m0
968 psrldq m0, 2 ; #ABC....
969 movh [dstq+strideq*0], m0
973 mova m0, [lq] ; stuvwxyz
978 PALIGNR m1, m0, 2, m4 ; tuvwxyz*
988 PALIGNR m1, m3, m2, 2, m0
993 PALIGNR m2, m3, m1, 2, m0
998 PALIGNR m1, m3, m2, 2, m0
1006 mova m0, [lq] ; klmnopqr
1019 PALIGNR m1, m0, 2, m3 ; lmnopqrs
1037 PALIGNR m0, m4, m2, 2, m1
1038 mova m2, m0
1039 PALIGNR m0, m5, m4, 2, m1
1040 mova m4, m0
1041 PALIGNR m0, m6, m5, 2, m1
1042 mova m5, m0
1051 mova m0, [aq+mmsize*3] ; a[24-31]
1053 psrldq m2, m0, 2 ; a[25-31].
1057 PALIGNR m0, m1, 2, m4 ; a[17-24]
1073 PALIGNR m5, m6, 2, m0 ; l[25-31]*
1074 PALIGNR m4, m5, 2, m0 ; l[26-31]*a
1077 PALIGNR m6, m7, 2, m0 ; l[17-24]
1078 PALIGNR m5, m6, 2, m0 ; l[18-25]
1081 PALIGNR m7, m1, 2, m0 ; l[9-16]
1082 PALIGNR m6, m7, 2, m0 ; l[10-17]
1085 PALIGNR m1, m3, 2, m0 ; l[1-8]
1086 PALIGNR m7, m1, 2, m0 ; l[2-9]
1113 mova [dst8q+stride8q*0+48], m0
1132 vpalignr m1, m0, m1, 2
1133 vpalignr m0, m2, m0, 2
1139 PALIGNR m2, m6, m7, 2, m0
1141 PALIGNR m2, m5, m6, 2, m0
1143 PALIGNR m2, m4, m5, 2, m0
1145 PALIGNR m2, m3, m4, 2, m0
1147 PALIGNR m2, m1, m3, 2, m0
1153 PALIGNR m2, m0, m1, 2, m3
1157 PALIGNR m1, m2, m0, 2, m3
1158 mova m0, m1
1176 mova m0, [lq] ; klmnopqrstuvwxyz
1181 vperm2i128 m3, m0, m1, q0201 ; stuvwxyz*abcdefg
1183 vpalignr m4, m3, m0, 2 ; lmnopqrstuvwxyz*
1184 vpalignr m5, m3, m0, 4 ; mnopqrstuvwxyz*a
1186 vperm2i128 m5, m0, m1, q0201 ; TUVWXYZ#ABCDEFGH
1192 vpalignr m3, m5, m0, 2
1196 vpalignr m3, m5, m0, 4
1201 mova [dst3q+stride3q*4], m0 ; 15
1202 vpalignr m3, m5, m0, 6
1206 vpalignr m3, m5, m0, 8
1210 vpalignr m3, m5, m0, 10
1214 vpalignr m3, m5, m0, 12
1218 vpalignr m3, m5, m0, 14
1227 mova m0, [aq] ; abcdefghijklmnop
1229 vperm2i128 m2, m0, m1, q0201 ; ijklmnoppppppppp
1230 vpalignr m3, m2, m0, 2 ; bcdefghijklmnopp
1232 vpalignr m5, m2, m0, 4 ; cdefghijklmnoppp
1236 pavgw m3, m0 ; abcdefghijklmnop
1245 vpalignr m0, m4, m3, 2
1247 mova [dstq+strideq*2 ], m0 ; 2 bcdefghijklmnopp
1249 vpalignr m0, m4, m3, 4
1251 mova [dst4q+strideq*0], m0 ; 4 cdefghijklmnoppp
1253 vpalignr m0, m4, m3, 6
1255 mova [ dstq+stride3q*2], m0 ; 6 defghijklmnopppp
1257 vpalignr m0, m4, m3, 8
1259 mova [ dstq+strideq*8], m0 ; 8 efghijklmnoppppp
1261 vpalignr m0, m4, m3, 10
1262 mova [dstq+stride5q*2], m0 ; 10 fghijklmnopppppp
1263 vpalignr m0, m4, m3, 12
1264 mova [dst4q+strideq*8], m0 ; 12 ghijklmnoppppppp
1265 vpalignr m0, m4, m3, 14
1266 mova [dst4q+stride5q*2], m0 ; 14 hijklmnopppppppp
1277 movu m0, [aq-2] ; *abcdefghijklmno
1279 vperm2i128 m2, m1, m0, q0201 ; stuvwxyz*abcdefg
1287 vpunpcklwd m0, m3, m4 ; kLlMmNnOsTtUuVvW
1289 vperm2i128 m3, m1, m0, q0002 ; kLlMmNnOoPpQqRrS
1290 vperm2i128 m4, m0, m1, q0301 ; sTtUuVvWwXxYyZz#
1291 vperm2i128 m0, m4, m2, q0201 ; wXxYyZz#ABCDEFGH
1301 mova [dstq+stride3q*1], m0 ; 3 wXxYyZz#ABCDEFGH
1304 vpalignr m5, m0, m4, 4
1305 vpalignr m6, m2, m0, 4
1310 vpalignr m5, m0, m4, 8
1311 vpalignr m6, m2, m0, 8
1318 vpalignr m5, m0, m4, 12
1319 vpalignr m6, m2, m0, 12
1332 mova m0, [lq+mmsize*0+0] ; l[0-15]
1337 vperm2i128 m5, m0, m1, q0201 ; lmnopqrstuvwxyz0
1338 vpalignr m6, m5, m0, 2 ; mnopqrstuvwxyz01
1339 vpalignr m7, m5, m0, 4 ; nopqrstuvwxyz012
1353 vperm2i128 m5, m0, m1, q0201 ; L[7-15]L[16-23]
1364 mova [dst24q+stride7q+0 ], m0 ; 31 23 15 7
1369 vpalignr m7, m5, m0, 2
1376 vpalignr m7, m5, m0, 4
1383 vpalignr m7, m5, m0, 6
1390 vpalignr m7, m5, m0, 8
1397 vpalignr m7, m5, m0, 10
1404 vpalignr m7, m5, m0, 12
1411 vpalignr m7, m5, m0, 14
1417 mova m0, m5
1436 movu m0, [aq] ; abcdefgh
1437 psrldq m1, m0, 2 ; bcdefgh.
1438 psrldq m2, m0, 4 ; cdefgh..
1440 pavgw m1, m0 ; ABCDEFG.
1454 mova m0, [aq] ; abcdefgh
1458 SHIFT_RIGHTx2 m1, m2, m0, m3 ; bcdefghh/cdefghhh
1460 pavgw m1, m0 ; ABCDEFGh
1483 mova m0, [aq]
1485 PALIGNR m2, m1, m0, 2, m3
1486 PALIGNR m3, m1, m0, 4, m4
1488 pavgw m2, m0
1492 SHIFT_RIGHTx2 m5, m0, m1, m4
1502 mova [dstq+strideq*1+16], m0
1506 vpalignr m3, m0, m3, 2
1510 PALIGNR m5, m0, m3, 2, m4
1514 SHIFT_RIGHT m0, m0, m4
1521 mova m0, [aq+mmsize*0]
1524 PALIGNR m6, m1, m0, 2, m5
1525 PALIGNR m7, m1, m0, 4, m5
1527 pavgw m6, m0
1529 PALIGNR m4, m2, m1, 2, m0
1530 PALIGNR m5, m2, m1, 4, m0
1533 mova m0, [aq+mmsize*3]
1534 PALIGNR m1, m0, m2, 2, m6
1535 PALIGNR m3, m0, m2, 4, m6
1541 SHIFT_RIGHTx2 m6, m1, m0, reg_shuf
1543 pavgw m0, m6
1564 mova [dstq+strideq*0+48], m0
1571 mova [dstq+stride16q+32], m0
1585 vpalignr m2, m0, m2, 2
1598 PALIGNR m3, m0, m2, 2, m1
1617 SHIFT_RIGHT m0, m0, reg_shuf
1626 mova [dstq+strideq*0+48], m0
1627 mova [dstq+strideq*1+48], m0
1628 mova [dstq+strideq*2+48], m0
1629 mova [dstq+stride3q +48], m0
1648 movu m0, [aq-2]
1650 PALIGNR m0, m1, 10, m2 ; xyz*abcd
1651 pslldq m1, m0, 2 ; .xyz*abc
1652 pslldq m2, m0, 4 ; ..xyz*ab
1654 pavgw m1, m0 ; ....#ABC
1660 shufps m0, m2, m1, q3210
1667 psrldq m0, 6
1668 movh [dstq+strideq*2], m0
1675 mova m0, [aq] ; abcdefgh
1678 pavgw m0, m1
1685 mova [dstq+strideq*0], m0
1687 PALIGNR m0, m4, 14, m1
1691 mova [dstq+strideq*2], m0
1694 PALIGNR m0, m4, 14, m1
1698 mova [dstq+strideq*0], m0
1700 PALIGNR m0, m4, 14, m1
1703 mova [dstq+strideq*2], m0
1713 PALIGNR m0, m1, m5, 14, m6 ; z*abcdef
1738 mova [dstq+strideq*1+ 0], m0
1744 PALIGNR m6, m0, 14, m4
1745 PALIGNR m0, m5, 14, m4
1752 movu m0, [aq+mmsize*0-2] ; *a[0-6]
1773 PALIGNR m6, m0, m5, 14, m7 ; l[31]*a[0-5]
1776 pavgw m4, m0
1777 PALIGNR m0, m5, 2, m7 ; l[25-31]*
1784 psrld m7, m0, 16
1786 pand m0, [pd_65535]
1789 packssdw m5, m0
1793 mova m0, [lq+mmsize*1+0] ; l[8-15]
1796 movu m0, [lq+mmsize*0+2] ; l[1-8]
1797 pslldq m5, m0, 2 ; .l[1-7]
1798 pslldq m7, m0, 4 ; ..l[1-6]
1801 psrld m7, m0, 16
1803 pand m0, [pd_65535]
1805 packssdw m0, m6
1848 PALIGNR m12, m0, 14, m5
1849 pslldq m0, 2
1877 PALIGNR m5, m0, 14, m6
1878 pslldq m0, 2
1894 movh m0, [lq] ; abcd
1896 pshufb m0, [pb_0to7_67x4] ; abcddddd
1898 punpcklqdq m0, m0
1899 pshufhw m0, m0, q3333 ; abcddddd
1901 psrldq m1, m0, 2 ; bcddddd.
1902 psrldq m2, m0, 4 ; cddddd..
1904 pavgw m1, m0 ; abcddddd
1906 PALIGNR m2, m1, 4, m0 ; bCcDdddd
1917 mova m0, [lq]
1921 SHIFT_RIGHTx2 m1, m2, m0, m3
1923 pavgw m1, m0
1925 shufps m0, m1, m2, q1032
1931 mova [dstq+strideq *2], m0
1938 PALIGNR m0, m2, m1, 4, m3
1939 mova m1, m0
1942 shufps m0, m1, m2, q1032
1945 mova [dstq+strideq *2], m0
1951 mova m0, [lq]
1956 pavgw m1, m0
1961 SHIFT_RIGHTx2 m0, m4, m3, m5
1963 pavgw m3, m0
1965 pshufd m0, m0, q3333
1978 mova [dstq+stride3q*4+16], m0
1984 vpalignr m4, m0, m4, 4
1992 PALIGNR m5, m0, m4, 4, m6
2003 movu m0, [lq+mmsize*0+4]
2023 mova m0, [pb_2to15_14_15]
2025 SHIFT_RIGHTx2 m1, m6, m7, m0
2047 mova [dstq+strideq *0+16], m0
2050 mova [dstq+stride4q*1+ 0], m0
2059 vpalignr m1, m0, m1, 4
2060 vpalignr m0, m3, m0, 4
2067 PALIGNR m6, m0, m1, 4, m7
2069 PALIGNR m6, m3, m0, 4, m7
2070 mova m0, m6
2113 PALIGNR m0, m5, m2, 4, m3
2114 mova m2, m0
2115 PALIGNR m0, m4, m5, 4, m3
2116 mova m5, m0
2117 PALIGNR m0, m7, m4, 4, m3
2118 mova m4, m0
2119 PALIGNR m0, m6, m7, 4, m3
2120 mova m7, m0
2121 PALIGNR m0, m1, m6, 4, m3
2122 mova m6, m0
2143 movh m0, [lq]
2144 movhps m0, [aq-2]
2145 psrldq m1, m0, 2
2146 psrldq m2, m0, 4
2148 pavgw m1, m0
2156 PALIGNR m2, m1, 4, m0
2162 mova m0, [lq]
2164 PALIGNR m2, m1, m0, 2, m3
2165 PALIGNR m3, m1, m0, 4, m4
2167 pavgw m2, m0
2169 psrldq m0, m1, 2
2185 PALIGNR m0, m3, m2, 4, m4
2186 mova m2, m0
2187 PALIGNR m0, m1, m3, 4, m4
2188 mova m3, m0
2198 movu m0, [lq+4]
2229 mova [dstq+mstrideq *4+ 0], m0
2232 mova [dstq+mstrideq *0+16], m0
2234 vpalignr m1, m0, m1, 4
2235 vpalignr m0, m3, m0, 4
2240 PALIGNR m6, m0, m1, 4, m7
2242 PALIGNR m6, m3, m0, 4, m7
2243 mova m0, m6
2260 movu m0, [lq+mmsize*0+4]
2280 mova m0, [lq+mmsize*3+0]
2282 PALIGNR m7, m1, m0, 2, m2
2283 PALIGNR m6, m1, m0, 4, m2
2285 pavgw m7, m0
2288 movu m0, [aq+mmsize*0+2]
2351 mova [dstq+strideq* 8+48], m0
2354 mova [dstq+strideq* 4+32], m0
2357 mova [dstq+strideq* 0+16], m0
2371 vpalignr m6, m0, m6, 4
2372 vpalignr m0, m1, m0, 4
2396 PALIGNR m3, m0, m6, 4, m2
2398 PALIGNR m3, m1, m0, 4, m2
2399 mova m0, m3
2404 PALIGNR m3, m2, m1, 4, m0
2406 PALIGNR m3, reg_sh, m2, 4, m0
2433 mova [dstq+stride3q*4+16], m0
2436 mova [dstq+strideq* 8+ 0], m0
2450 vpalignr m1, m0, m1, 4
2451 vpalignr m0, m3, m0, 4
2462 PALIGNR m6, m0, m1, 4, m7
2464 PALIGNR m6, m3, m0, 4, m7
2465 mova m0, m6