Lines Matching refs:src0
136 __m128i src0, src1, src2, src3;
148 LSX_LD_4(src, src_stride, src0, src1, src2, src3);
149 DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128,
150 src0, src1, src2, src3);
151 HORIZ_8TAP_4WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2,
171 __m128i src0, src1, src2, src3;
183 src0 = __lsx_vld(_src, 0);
187 DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128,
188 src0, src1, src2, src3);
189 HORIZ_8TAP_4WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2,
191 src0 = __lsx_vld(_src, 0);
194 DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128,
195 src0, src1, src2, src3);
196 HORIZ_8TAP_4WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2,
232 __m128i src0, src1, src2, src3;
244 LSX_LD_4(src, src_stride, src0, src1, src2, src3);
245 DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128,
246 src0, src1, src2, src3);
247 HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2,
268 __m128i src0, src1, src2, src3;
281 src0 = __lsx_vld(_src, 0);
285 DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128,
286 src0, src1, src2, src3);
287 HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2,
320 __m128i src0, src1, src2, src3;
334 DUP2_ARG2(__lsx_vld, src, 0, _src, 0, src0, src2);
336 DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128,
337 src0, src1, src2, src3);
338 HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2,
355 __m128i src0, src1, src2, src3;
369 DUP2_ARG2(__lsx_vld, src, 0, src, 16, src0, src2);
371 src1 = __lsx_vshuf_b(src2, src0, shuff);
373 DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128,
374 src0, src1, src2, src3);
375 HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2,
382 DUP2_ARG2(__lsx_vld, src, 0, src, 16, src0, src2);
384 src1 = __lsx_vshuf_b(src2, src0, shuff);
388 DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128,
389 src0, src1, src2, src3);
390 HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2,
405 __m128i src0, src1, src2, src3;
419 DUP2_ARG2(__lsx_vld, src, 0, src, 16, src0, src2);
421 src1 = __lsx_vshuf_b(src2, src0, shuff);
422 DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128,
423 src0, src1, src2, src3);
424 HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2,
431 DUP2_ARG2(__lsx_vld, src, 32, src, 48, src0, src2);
433 src1 = __lsx_vshuf_b(src2, src0, shuff);
434 DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128,
435 src0, src1, src2, src3);
436 HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2,
455 __m128i src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
464 src0 = __lsx_vld(_src, 0);
471 DUP4_ARG2(__lsx_vilvl_b, src1, src0, src3, src2, src5, src4, src2, src1, tmp0,
515 __m128i src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
528 src0 = __lsx_vld(_src, 0);
536 DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128,
537 src0, src1, src2, src3);
540 DUP4_ARG2(__lsx_vilvl_b, src1, src0, src3, src2, src5, src4, src2, src1,
587 __m128i src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
599 src0 = __lsx_vld(_src, 0);
606 DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128,
607 src0, src1, src2, src3);
610 DUP4_ARG2(__lsx_vilvl_b, src1, src0, src3, src2, src5, src4, src2, src1,
613 DUP4_ARG2(__lsx_vilvh_b, src1, src0, src3, src2, src5, src4, src2, src1,
625 src0, src1, src2, src3);
628 tmp0 = FILT_8TAP_DPADD_S_H(reg0, reg1, reg2, src0, filter0, filter1,
642 tmp0 = FILT_8TAP_DPADD_S_H(reg1, reg2, src0, src2, filter0, filter1,
658 reg1 = src0;
681 __m128i src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
702 src0 = __lsx_vld(src_tmp, 0);
712 DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128,
713 src0, src1, src2, src3);
716 DUP4_ARG2(__lsx_vilvl_b, src1, src0, src3, src2, src5, src4, src2, src1,
719 DUP4_ARG2(__lsx_vilvh_b, src1, src0, src3, src2, src5, src4, src2, src1,
732 src10, src9, src0, src1, src2, src3);
735 tmp0 = FILT_8TAP_DPADD_S_H(reg0, reg1, reg2, src0, filter0,
748 tmp0 = FILT_8TAP_DPADD_S_H(reg1, reg2, src0, src2, filter0,
764 reg1 = src0;
804 __m128i src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
822 src0 = __lsx_vld(_src, 0);
829 DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128,
830 src0, src1, src2, src3);
834 tmp0 = HORIZ_8TAP_FILT(src0, src1, mask0, mask1, mask2, mask3, filt_hz0,
863 src0 = __lsx_vshuf_b(src1, tmp3, shuff);
864 src0 = __lsx_vpackev_b(src1, src0);
865 out1 = FILT_8TAP_DPADD_S_H(tmp1, tmp2, tmp4, src0, filt_vt0, filt_vt1,
881 tmp2 = src0;
892 __m128i src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
909 src0 = __lsx_vld(_src, 0);
916 DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128,
917 src0, src1, src2, src3);
921 src0 = HORIZ_8TAP_FILT(src0, src0, mask0, mask1, mask2, mask3, filt_hz0,
938 DUP4_ARG2(__lsx_vpackev_b, src1, src0, src3, src2, src5, src4,
957 src0 = __lsx_vpackev_b(src8, src7);
958 out1 = FILT_8TAP_DPADD_S_H(tmp4, tmp5, tmp6, src0, filt_vt0, filt_vt1,
968 src4 = FILT_8TAP_DPADD_S_H(tmp5, tmp6, src0, src2, filt_vt0, filt_vt1,
986 tmp5 = src0;
1044 __m128i src0, src1, src2, src3;
1047 src0 = __lsx_vldrepl_d(src, 0);
1055 __lsx_vstelm_d(src0, dst, 0, 0);
1071 __m128i src0, src1, src2, src3;
1081 src0 = __lsx_vld(_src, 0);
1085 __lsx_vst(src0, dst, 0);
1108 __m128i src0, src1, src2, src3, src4, src5, src6, src7;
1111 src0 = __lsx_vld(src_tmp1, 0);
1123 __lsx_vst(src0, dst_tmp1, 0);
1141 __m128i src0, src1, src2, src3, src4, src5, src6, src7;
1146 src0, src1, src2, src3);
1157 __lsx_vst(src0, dst, 0);
1186 __m128i src0, src1, src2, src3;
1198 LSX_LD_4(src, src_stride, src0, src1, src2, src3);
1199 DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128,
1200 src0, src1, src2, src3);
1201 HORIZ_8TAP_4WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2, mask3,
1231 __m128i src0, src1, src2, src3, filter0, filter1, filter2, filter3;
1242 LSX_LD_4(src, src_stride, src0, src1, src2, src3);
1244 DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128,
1245 src0, src1, src2, src3);
1268 HORIZ_8TAP_4WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2, mask3,
1270 LSX_LD_4(src, src_stride, src0, src1, src2, src3);
1271 DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128,
1272 src0, src1, src2, src3);
1273 HORIZ_8TAP_4WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2, mask3,
1318 __m128i src0, src1, src2, src3, filter0, filter1, filter2, filter3;
1334 src0 = __lsx_vld(_src, 0);
1338 DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128,
1339 src0, src1, src2, src3);
1340 HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2,
1374 __m128i src0, src1, src2, src3, filter0, filter1, filter2, filter3;
1387 DUP2_ARG2(__lsx_vld, src, 0, src, 8, src0, src1);
1394 DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128,
1395 src0, src1, src2, src3);
1396 DUP4_ARG3(__lsx_vshuf_b, src0, src0, mask0, src1, src1, mask0, src2, src2,
1398 DUP4_ARG3(__lsx_vshuf_b, src0, src0, mask1, src1, src1, mask1, src2, src2,
1400 DUP4_ARG3(__lsx_vshuf_b, src0, src0, mask2, src1, src1, mask2, src2, src2,
1402 DUP4_ARG3(__lsx_vshuf_b, src0, src0, mask3, src1, src1, mask3, src2, src2,
1431 __m128i src0, src1, src2, src3, filter0, filter1, filter2, filter3;
1445 DUP2_ARG2(__lsx_vld, src, 0, src, 16, src0, src2);
1447 src1 = __lsx_vshuf_b(src2, src0, shuff);
1451 DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128,
1452 src0, src1, src2, src3);
1453 DUP4_ARG3(__lsx_vshuf_b, src0, src0, mask0, src1, src1, mask0, src2,
1455 DUP4_ARG3(__lsx_vshuf_b, src0, src0, mask1, src1, src1, mask1, src2,
1457 DUP4_ARG3(__lsx_vshuf_b, src0, src0, mask2, src1, src1, mask2, src2,
1459 DUP4_ARG3(__lsx_vshuf_b, src0, src0, mask3, src1, src1, mask3, src2,
1487 __m128i src0, src1, src2, src3;
1501 DUP2_ARG2(__lsx_vld, src, 0, src, 16, src0, src2);
1503 src1 = __lsx_vshuf_b(src2, src0, shuff);
1505 DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128,
1506 src0, src1, src2, src3);
1507 HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2,
1515 DUP2_ARG2(__lsx_vld, src, 32, src, 48, src0, src2);
1517 src1 = __lsx_vshuf_b(src2, src0, shuff);
1519 DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128,
1520 src0, src1, src2, src3);
1521 HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2,
1541 __m128i src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
1553 src0 = __lsx_vld(_src, 0);
1560 DUP4_ARG2(__lsx_vilvl_b, src1, src0, src3, src2, src5, src4, src2, src1,
1573 src0 = __lsx_vldrepl_w(dst_tmp, 0);
1581 DUP2_ARG2(__lsx_vilvl_w, src1, src0, src3, src2, src0, src1);
1582 src0 = __lsx_vilvl_d(src1, src0);
1593 out0 = __lsx_vavgr_bu(out0, src0);
1617 __m128i src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
1630 src0 = __lsx_vld(_src, 0);
1637 DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128,
1638 src0, src1, src2, src3);
1641 DUP4_ARG2(__lsx_vilvl_b, src1, src0, src3, src2, src5, src4, src2,
1650 src0 = __lsx_vldrepl_d(dst_tmp, 0);
1658 DUP2_ARG2(__lsx_vilvl_d, src1, src0, src3, src2, src0, src1);
1673 DUP2_ARG2(__lsx_vavgr_bu, out0, src0, out1, src1, out0, out1);
1703 __m128i src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
1723 src0 = __lsx_vld(src_tmp, 0);
1732 DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128,
1733 src0, src1, src2, src3);
1736 DUP4_ARG2(__lsx_vilvl_b, src1, src0, src3, src2, src5, src4, src2, src1,
1739 DUP4_ARG2(__lsx_vilvh_b, src1, src0, src3, src2, src5, src4, src2, src1,
1752 src10, src9, src0, src1, src2, src3);
1755 tmp0 = FILT_8TAP_DPADD_S_H(reg0, reg1, reg2, src0, filter0,
1771 tmp0 = FILT_8TAP_DPADD_S_H(reg1, reg2, src0, src2, filter0,
1790 reg1 = src0;
1848 __m128i src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
1866 src0 = __lsx_vld(_src, 0);
1874 DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128,
1875 src0, src1, src2, src3);
1879 tmp0 = HORIZ_8TAP_FILT(src0, src1, mask0, mask1, mask2, mask3, filt_hz0,
1918 src0 = __lsx_vshuf_b(src1, tmp3, shuff);
1919 src0 = __lsx_vpackev_b(src1, src0);
1920 out1 = FILT_8TAP_DPADD_S_H(tmp1, tmp2, tmp4, src0, filt_vt0, filt_vt1,
1937 tmp2 = src0;
1951 __m128i src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
1968 src0 = __lsx_vld(_src, 0);
1975 DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128,
1976 src0, src1, src2, src3);
1980 src0 = HORIZ_8TAP_FILT(src0, src0, mask0, mask1, mask2, mask3, filt_hz0,
1997 DUP4_ARG2(__lsx_vpackev_b, src1, src0, src3, src2, src5, src4,
2016 src0 = __lsx_vpackev_b(src8, src7);
2017 out1 = FILT_8TAP_DPADD_S_H(tmp4, tmp5, tmp6, src0, filt_vt0, filt_vt1,
2027 src4 = FILT_8TAP_DPADD_S_H(tmp5, tmp6, src0, src2, filt_vt0, filt_vt1,
2055 tmp5 = src0;
2126 __m128i src0, src1, dst0, dst1;
2138 DUP2_ARG2(__lsx_vilvl_d, tmp1, tmp0, tmp3, tmp2, src0, src1);
2148 DUP2_ARG2(__lsx_vavgr_bu, src0, dst0, src1, dst1, dst0, dst1);
2165 __m128i src0, src1, src2, src3;
2176 src0 = __lsx_vld(_src, 0);
2185 DUP4_ARG2(__lsx_vavgr_bu, src0, dst0, src1, dst1,
2203 __m128i src0, src1, src2, src3, src4, src5, src6, src7;
2215 src0 = __lsx_vld(src_tmp1, 0);
2236 DUP4_ARG2(__lsx_vavgr_bu, src0, dst0, src1, dst1,
2259 __m128i src0, src1, src2, src3, src4, src5, src6, src7;
2266 src0, src1, src2, src3);
2289 DUP4_ARG2(__lsx_vavgr_bu, src0, dst0, src1, dst1,