Lines Matching refs:tmp0
1013 __m256i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
1023 tmp0 = __lasx_xvreplgr2vr_b(weight_src);
1025 wgt = __lasx_xvilvh_b(tmp1, tmp0);
1030 src, stride_3x, tmp0, tmp1, tmp2, tmp3);
1035 DUP4_ARG3(__lasx_xvpermi_q, tmp1, tmp0, 0x20, tmp3, tmp2, 0x20, tmp5, tmp4,
1038 dst, stride_3x, tmp0, tmp1, tmp2, tmp3);
1043 DUP4_ARG3(__lasx_xvpermi_q, tmp1, tmp0, 0x20, tmp3, tmp2, 0x20, tmp5, tmp4,
1056 offset, wgt, vec2, offset, wgt, vec3, tmp0, tmp1, tmp2, tmp3);
1060 tmp0 = __lasx_xvsra_h(tmp0, denom);
1069 DUP4_ARG1(__lasx_xvclip255_h, tmp0, tmp1, tmp2, tmp3,
1070 tmp0, tmp1, tmp2, tmp3);
1073 DUP4_ARG2(__lasx_xvpickev_b, tmp1, tmp0, tmp3, tmp2, tmp5, tmp4, tmp7, tmp6,
1102 src, stride_3x, tmp0, tmp1, tmp2, tmp3);
1107 DUP4_ARG3(__lasx_xvpermi_q, tmp1, tmp0, 0x20, tmp3, tmp2, 0x20, tmp5,
1110 dst, stride_3x, tmp0, tmp1, tmp2, tmp3);
1115 DUP4_ARG3(__lasx_xvpermi_q, tmp1, tmp0, 0x20, tmp3, tmp2, 0x20, tmp5,
1128 offset, wgt, vec2, offset, wgt, vec3, tmp0, tmp1, tmp2, tmp3);
1132 tmp0 = __lasx_xvsra_h(tmp0, denom);
1141 DUP4_ARG1(__lasx_xvclip255_h, tmp0, tmp1, tmp2, tmp3,
1142 tmp0, tmp1, tmp2, tmp3);
1145 DUP4_ARG2(__lasx_xvpickev_b, tmp1, tmp0, tmp3, tmp2, tmp5, tmp4, tmp7,
1179 __m256i tmp0, tmp1, tmp2, tmp3, denom, offset;
1187 tmp0 = __lasx_xvreplgr2vr_b(weight_src);
1189 wgt = __lasx_xvilvh_b(tmp1, tmp0);
1194 src, stride_3x, tmp0, tmp1, tmp2, tmp3);
1195 DUP2_ARG2(__lasx_xvilvl_d, tmp1, tmp0, tmp3, tmp2, tmp0, tmp1);
1196 src0 = __lasx_xvpermi_q(tmp1, tmp0, 0x20);
1198 dst, stride_3x, tmp0, tmp1, tmp2, tmp3);
1199 DUP2_ARG2(__lasx_xvilvl_d, tmp1, tmp0, tmp3, tmp2, tmp0, tmp1);
1200 dst0 = __lasx_xvpermi_q(tmp1, tmp0, 0x20);
1205 tmp0, tmp1);
1206 tmp0 = __lasx_xvsra_h(tmp0, denom);
1208 DUP2_ARG1(__lasx_xvclip255_h, tmp0, tmp1, tmp0, tmp1);
1209 dst0 = __lasx_xvpickev_b(tmp1, tmp0);
1222 __m256i tmp0, tmp1, tmp2, tmp3, denom, offset;
1232 tmp0 = __lasx_xvreplgr2vr_b(weight_src);
1234 wgt = __lasx_xvilvh_b(tmp1, tmp0);
1239 src, stride_3x, tmp0, tmp1, tmp2, tmp3);
1241 DUP2_ARG2(__lasx_xvilvl_d, tmp1, tmp0, tmp3, tmp2, tmp0, tmp1);
1242 src0 = __lasx_xvpermi_q(tmp1, tmp0, 0x20);
1244 src, stride_3x, tmp0, tmp1, tmp2, tmp3);
1245 DUP2_ARG2(__lasx_xvilvl_d, tmp1, tmp0, tmp3, tmp2, tmp0, tmp1);
1246 src1 = __lasx_xvpermi_q(tmp1, tmp0, 0x20);
1247 tmp0 = __lasx_xvld(dst_tmp, 0);
1251 DUP2_ARG2(__lasx_xvilvl_d, tmp1, tmp0, tmp3, tmp2, tmp0, tmp1);
1252 dst0 = __lasx_xvpermi_q(tmp1, tmp0, 0x20);
1254 dst_tmp, stride_3x, tmp0, tmp1, tmp2, tmp3);
1255 DUP2_ARG2(__lasx_xvilvl_d, tmp1, tmp0, tmp3, tmp2, tmp0, tmp1);
1256 dst1 = __lasx_xvpermi_q(tmp1, tmp0, 0x20);
1263 offset, wgt, vec2, offset, wgt, vec3, tmp0, tmp1, tmp2, tmp3);
1264 tmp0 = __lasx_xvsra_h(tmp0, denom);
1268 DUP4_ARG1(__lasx_xvclip255_h, tmp0, tmp1, tmp2, tmp3,
1269 tmp0, tmp1, tmp2, tmp3);
1270 DUP2_ARG2(__lasx_xvpickev_b, tmp1, tmp0, tmp3, tmp2, dst0, dst1);
1288 __m256i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, denom, offset;
1298 tmp0 = __lasx_xvreplgr2vr_b(weight_src);
1300 wgt = __lasx_xvilvh_b(tmp1, tmp0);
1305 src, stride_3x, tmp0, tmp1, tmp2, tmp3);
1307 DUP2_ARG2(__lasx_xvilvl_d, tmp1, tmp0, tmp3, tmp2, tmp0, tmp1);
1308 src0 = __lasx_xvpermi_q(tmp1, tmp0, 0x20);
1310 src, stride_3x, tmp0, tmp1, tmp2, tmp3);
1312 DUP2_ARG2(__lasx_xvilvl_d, tmp1, tmp0, tmp3, tmp2, tmp0, tmp1);
1313 src1 = __lasx_xvpermi_q(tmp1, tmp0, 0x20);
1315 src, stride_3x, tmp0, tmp1, tmp2, tmp3);
1317 DUP2_ARG2(__lasx_xvilvl_d, tmp1, tmp0, tmp3, tmp2, tmp0, tmp1);
1318 src2 = __lasx_xvpermi_q(tmp1, tmp0, 0x20);
1320 src, stride_3x, tmp0, tmp1, tmp2, tmp3);
1321 DUP2_ARG2(__lasx_xvilvl_d, tmp1, tmp0, tmp3, tmp2, tmp0, tmp1);
1322 src3 = __lasx_xvpermi_q(tmp1, tmp0, 0x20);
1325 dst_tmp, stride_3x, tmp0, tmp1, tmp2, tmp3);
1327 DUP2_ARG2(__lasx_xvilvl_d, tmp1, tmp0, tmp3, tmp2, tmp0, tmp1);
1328 dst0 = __lasx_xvpermi_q(tmp1, tmp0, 0x20);
1330 dst_tmp, stride_3x, tmp0, tmp1, tmp2, tmp3);
1332 DUP2_ARG2(__lasx_xvilvl_d, tmp1, tmp0, tmp3, tmp2, tmp0, tmp1);
1333 dst1 = __lasx_xvpermi_q(tmp1, tmp0, 0x20);
1335 dst_tmp, stride_3x, tmp0, tmp1, tmp2, tmp3);
1337 DUP2_ARG2(__lasx_xvilvl_d, tmp1, tmp0, tmp3, tmp2, tmp0, tmp1);
1338 dst2 = __lasx_xvpermi_q(tmp1, tmp0, 0x20);
1340 dst_tmp, stride_3x, tmp0, tmp1, tmp2, tmp3);
1341 DUP2_ARG2(__lasx_xvilvl_d, tmp1, tmp0, tmp3, tmp2, tmp0, tmp1);
1342 dst3 = __lasx_xvpermi_q(tmp1, tmp0, 0x20);
1353 offset, wgt, vec2, offset, wgt, vec3, tmp0, tmp1, tmp2, tmp3);
1356 tmp0 = __lasx_xvsra_h(tmp0, denom);
1364 DUP4_ARG1(__lasx_xvclip255_h, tmp0, tmp1, tmp2, tmp3,
1365 tmp0, tmp1, tmp2, tmp3);
1368 DUP4_ARG2(__lasx_xvpickev_b, tmp1, tmp0, tmp3, tmp2, tmp5, tmp4, tmp7, tmp6,
1414 __m256i tmp0, tmp1, denom, offset;
1420 tmp0 = __lasx_xvreplgr2vr_b(weight_src);
1422 wgt = __lasx_xvilvh_b(tmp1, tmp0);
1426 DUP2_ARG2(__lasx_xvldx, src, 0, src, stride, tmp0, tmp1);
1427 src0 = __lasx_xvilvl_w(tmp1, tmp0);
1428 DUP2_ARG2(__lasx_xvldx, dst, 0, dst, stride, tmp0, tmp1);
1429 dst0 = __lasx_xvilvl_w(tmp1, tmp0);
1432 tmp0 = __lasx_xvdp2add_h_b(offset, wgt, vec0);
1433 tmp0 = __lasx_xvsra_h(tmp0, denom);
1434 tmp0 = __lasx_xvclip255_h(tmp0);
1435 tmp0 = __lasx_xvpickev_b(tmp0, tmp0);
1436 __lasx_xvstelm_w(tmp0, dst, 0, 0);
1437 __lasx_xvstelm_w(tmp0, dst + stride, 0, 1);
1446 __m256i tmp0, tmp1, tmp2, tmp3, denom, offset;
1454 tmp0 = __lasx_xvreplgr2vr_b(weight_src);
1456 wgt = __lasx_xvilvh_b(tmp1, tmp0);
1461 src, stride_3x, tmp0, tmp1, tmp2, tmp3);
1462 DUP2_ARG2(__lasx_xvilvl_w, tmp2, tmp0, tmp3, tmp1, tmp0, tmp1);
1463 src0 = __lasx_xvilvl_w(tmp1, tmp0);
1465 dst, stride_3x, tmp0, tmp1, tmp2, tmp3);
1466 DUP2_ARG2(__lasx_xvilvl_w, tmp2, tmp0, tmp3, tmp1, tmp0, tmp1);
1467 dst0 = __lasx_xvilvl_w(tmp1, tmp0);
1472 tmp0 = __lasx_xvdp2add_h_b(offset, wgt, vec0);
1473 tmp0 = __lasx_xvsra_h(tmp0, denom);
1474 tmp0 = __lasx_xvclip255_h(tmp0);
1475 tmp0 = __lasx_xvpickev_b(tmp0, tmp0);
1476 __lasx_xvstelm_w(tmp0, dst, 0, 0);
1477 __lasx_xvstelm_w(tmp0, dst + stride, 0, 1);
1478 __lasx_xvstelm_w(tmp0, dst + stride_2x, 0, 4);
1479 __lasx_xvstelm_w(tmp0, dst + stride_3x, 0, 5);
1488 __m256i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, denom, offset;
1497 tmp0 = __lasx_xvreplgr2vr_b(weight_src);
1499 wgt = __lasx_xvilvh_b(tmp1, tmp0);
1504 src, stride_3x, tmp0, tmp1, tmp2, tmp3);
1508 DUP4_ARG2(__lasx_xvilvl_w, tmp2, tmp0, tmp3, tmp1, tmp6, tmp4, tmp7, tmp5,
1509 tmp0, tmp1, tmp2, tmp3);
1510 DUP2_ARG2(__lasx_xvilvl_w, tmp1, tmp0, tmp3, tmp2, tmp0, tmp1);
1511 src0 = __lasx_xvpermi_q(tmp1, tmp0, 0x20);
1513 dst, stride_3x, tmp0, tmp1, tmp2, tmp3);
1518 DUP4_ARG2(__lasx_xvilvl_w, tmp2, tmp0, tmp3, tmp1, tmp6, tmp4, tmp7, tmp5,
1519 tmp0, tmp1, tmp2, tmp3);
1520 DUP2_ARG2(__lasx_xvilvl_w, tmp1, tmp0, tmp3, tmp2, tmp0, tmp1);
1521 dst0 = __lasx_xvpermi_q(tmp1, tmp0, 0x20);
1526 tmp0, tmp1);
1527 tmp0 = __lasx_xvsra_h(tmp0, denom);
1529 DUP2_ARG1(__lasx_xvclip255_h, tmp0, tmp1, tmp0, tmp1);
1530 tmp0 = __lasx_xvpickev_b(tmp1, tmp0);
1531 __lasx_xvstelm_w(tmp0, dst, 0, 0);
1532 __lasx_xvstelm_w(tmp0, dst + stride, 0, 1);
1533 __lasx_xvstelm_w(tmp0, dst + stride_2x, 0, 2);
1534 __lasx_xvstelm_w(tmp0, dst + stride_3x, 0, 3);
1536 __lasx_xvstelm_w(tmp0, dst, 0, 4);
1537 __lasx_xvstelm_w(tmp0, dst + stride, 0, 5);
1538 __lasx_xvstelm_w(tmp0, dst + stride_2x, 0, 6);
1539 __lasx_xvstelm_w(tmp0, dst + stride_3x, 0, 7);
1570 __m256i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
1580 src, stride_3x, tmp0, tmp1, tmp2, tmp3);
1585 DUP4_ARG3(__lasx_xvpermi_q, tmp1, tmp0, 0x20, tmp3, tmp2, 0x20, tmp5, tmp4,
1646 src, stride_3x, tmp0, tmp1, tmp2, tmp3);
1651 DUP4_ARG3(__lasx_xvpermi_q, tmp1, tmp0, 0x20, tmp3, tmp2, 0x20, tmp5,
1720 __m256i tmp0, tmp1, tmp2, tmp3, denom, offset;
1729 src, stride_3x, tmp0, tmp1, tmp2, tmp3);
1730 DUP2_ARG2(__lasx_xvilvl_d, tmp1, tmp0, tmp3, tmp2, tmp0, tmp1);
1731 src0 = __lasx_xvpermi_q(tmp1, tmp0, 0x20);
1754 __m256i tmp0, tmp1, tmp2, tmp3, denom, offset, wgt;
1768 src_tmp, stride_3x, tmp0, tmp1, tmp2, tmp3);
1770 DUP2_ARG2(__lasx_xvilvl_d, tmp1, tmp0, tmp3, tmp2, tmp0, tmp1);
1771 src0 = __lasx_xvpermi_q(tmp1, tmp0, 0x20);
1773 src_tmp, stride_3x, tmp0, tmp1, tmp2, tmp3);
1774 DUP2_ARG2(__lasx_xvilvl_d, tmp1, tmp0, tmp3, tmp2, tmp0, tmp1);
1775 src1 = __lasx_xvpermi_q(tmp1, tmp0, 0x20);
1811 __m256i tmp0, tmp1, tmp2, tmp3, denom, offset, wgt;
1826 src_tmp, stride_3x, tmp0, tmp1, tmp2, tmp3);
1828 DUP2_ARG2(__lasx_xvilvl_d, tmp1, tmp0, tmp3, tmp2, tmp0, tmp1);
1829 src0 = __lasx_xvpermi_q(tmp1, tmp0, 0x20);
1831 src_tmp, stride_3x, tmp0, tmp1, tmp2, tmp3);
1833 DUP2_ARG2(__lasx_xvilvl_d, tmp1, tmp0, tmp3, tmp2, tmp0, tmp1);
1834 src1 = __lasx_xvpermi_q(tmp1, tmp0, 0x20);
1836 src_tmp, stride_3x, tmp0, tmp1, tmp2, tmp3);
1838 DUP2_ARG2(__lasx_xvilvl_d, tmp1, tmp0, tmp3, tmp2, tmp0, tmp1);
1839 src2 = __lasx_xvpermi_q(tmp1, tmp0, 0x20);
1841 src_tmp, stride_3x, tmp0, tmp1, tmp2, tmp3);
1842 DUP2_ARG2(__lasx_xvilvl_d, tmp1, tmp0, tmp3, tmp2, tmp0, tmp1);
1843 src3 = __lasx_xvpermi_q(tmp1, tmp0, 0x20);
1922 __m256i src0, tmp0, tmp1, denom, offset;
1930 DUP2_ARG2(__lasx_xvldx, src, 0, src, stride, tmp0, tmp1);
1931 src0 = __lasx_xvilvl_w(tmp1, tmp0);
1946 __m256i src0, tmp0, tmp1, tmp2, tmp3, denom, offset;
1958 src, stride_3x, tmp0, tmp1, tmp2, tmp3);
1959 DUP2_ARG2(__lasx_xvilvl_w, tmp2, tmp0, tmp3, tmp1, tmp0, tmp1);
1960 src0 = __lasx_xvilvl_w(tmp1, tmp0);
1977 __m256i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, denom, offset;
1991 src, stride_3x, tmp0, tmp1, tmp2, tmp3);
1996 DUP4_ARG2(__lasx_xvilvl_w, tmp2, tmp0, tmp3, tmp1, tmp6, tmp4, tmp7,
1997 tmp5, tmp0, tmp1, tmp2, tmp3);
1998 DUP2_ARG2(__lasx_xvilvl_w, tmp1, tmp0, tmp3, tmp2, tmp0, tmp1);
1999 src0 = __lasx_xvpermi_q(tmp1, tmp0, 0x20);
2037 __m256i tmp0, tmp1;
2047 tmp0 = __lasx_xvilvl_w(dst1, dst0);
2049 dst0 = __lasx_xvilvl_d(tmp1, tmp0);
2050 tmp0 = __lasx_vext2xv_hu_bu(dst0);
2052 tmp1 = __lasx_xvadd_h(src0, tmp0);
2065 __m256i tmp0, tmp1, tmp2, tmp3;
2087 tmp0 = __lasx_xvilvl_d(dst1, dst0);
2091 dst0 = __lasx_vext2xv_hu_bu(tmp0);
2096 tmp0 = __lasx_xvadd_h(src0, dst0);
2100 dst1 = __lasx_xvpickev_b(tmp1, tmp0);