Lines Matching refs:src0

36     v16i8 src0, src1;
46 LD_SB2(src, stride, src0, src1);
48 src0 = __msa_vshf_b(mask, src1, src0);
49 res_r = __msa_dotp_u_h((v16u8) src0, coeff_vec);
66 v16u8 src0, src1, src2, src3;
76 LD_UB4(src, stride, src0, src1, src2, src3);
78 VSHF_B2_UB(src0, src1, src2, src3, mask, mask, src0, src2);
80 src0 = (v16u8) __msa_ilvr_d((v2i64) src2, (v2i64) src0);
82 res_r = __msa_dotp_u_h(src0, coeff_vec);
105 v16i8 src0, src1;
115 LD_SB2(src, stride, src0, src1);
117 src0 = __msa_vshf_b(mask, src1, src0);
118 res_r = __msa_dotp_u_h((v16u8) src0, coeff_vec);
130 v16u8 src0, src1, src2, src3, out;
139 LD_UB4(src, stride, src0, src1, src2, src3);
140 VSHF_B2_UB(src0, src1, src2, src3, mask, mask, src0, src2);
141 DOTP_UB2_UH(src0, src2, coeff_vec, coeff_vec, res0_r, res1_r);
153 v16u8 src0, src1, src2, src3, src4, src5, src6, src7, out0, out1;
162 LD_UB8(src, stride, src0, src1, src2, src3, src4, src5, src6, src7);
163 VSHF_B2_UB(src0, src1, src2, src3, mask, mask, src0, src2);
165 DOTP_UB2_UH(src0, src2, coeff_vec, coeff_vec, res0, res1);
190 v16u8 src0, src1, src2, src3, out0, out1;
198 LD_UB4(src, stride, src0, src1, src2, src3);
199 VSHF_B2_UB(src0, src0, src1, src1, mask, mask, src0, src1);
201 DOTP_UB4_UH(src0, src1, src2, src3, coeff_vec, coeff_vec, coeff_vec,
213 v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
223 LD_UB8(src, stride, src0, src1, src2, src3, src4, src5, src6, src7);
224 VSHF_B2_UB(src0, src0, src1, src1, mask, mask, src0, src1);
228 DOTP_UB4_UH(src0, src1, src2, src3, coeff_vec, coeff_vec, coeff_vec,
248 v16u8 src0, src1, src2, src3, out0, out1;
258 LD_UB4(src, stride, src0, src1, src2, src3);
261 VSHF_B2_UB(src0, src0, src1, src1, mask, mask, src0, src1);
263 DOTP_UB4_UH(src0, src1, src2, src3, coeff_vec, coeff_vec, coeff_vec,
275 src0 = LD_UB(src);
278 src0 = (v16u8) __msa_vshf_b(mask, (v16i8) src0, (v16i8) src0);
280 res0 = __msa_dotp_u_h(src0, coeff_vec);
309 v16i8 src0, src1, src2;
317 LD_SB3(src, stride, src0, src1, src2);
319 ILVR_B2_UB(src1, src0, src2, src1, tmp0, tmp1);
340 v16u8 src0, src1, src2, src3, src4;
348 LD_UB5(src, stride, src0, src1, src2, src3, src4);
349 ILVR_B4_UB(src1, src0, src2, src1, src3, src2, src4, src3,
379 v16u8 src0, src1, src2;
387 LD_UB3(src, stride, src0, src1, src2);
388 ILVR_B2_UB(src1, src0, src2, src1, tmp0, tmp1);
403 v16u8 src0, src1, src2, src3, src4;
411 LD_UB5(src, stride, src0, src1, src2, src3, src4);
412 ILVR_B4_UB(src1, src0, src2, src1, src3, src2, src4, src3, tmp0, tmp1, tmp2,
427 v16u8 src0, src1, src2, src3, src4, src5, src6, src7, src8;
434 LD_UB5(src, stride, src0, src1, src2, src3, src4);
437 ILVR_B4_UB(src1, src0, src2, src1, src3, src2, src4, src3, tmp0, tmp1, tmp2,
468 v16u8 src0, src1, src2, src3, src4, out0, out1;
474 LD_UB5(src, stride, src0, src1, src2, src3, src4);
475 ILVR_B4_UB(src1, src0, src2, src1, src3, src2, src4, src3, src0, src1, src2,
477 DOTP_UB4_UH(src0, src1, src2, src3, coeff_vec, coeff_vec, coeff_vec,
489 v16u8 src0, src1, src2, src3, src4, src5, src6, src7, src8;
496 LD_UB5(src, stride, src0, src1, src2, src3, src4);
499 ILVR_B4_UB(src1, src0, src2, src1, src3, src2, src4, src3, src0, src1, src2,
503 DOTP_UB4_UH(src0, src1, src2, src3, coeff_vec, coeff_vec, coeff_vec,
534 v16u8 src0, src1, src2;
546 LD_UB3(src, stride, src0, src1, src2);
547 VSHF_B2_UB(src0, src1, src1, src2, mask, mask, src0, src1);
548 DOTP_UB2_UH(src0, src1, coeff_hz_vec, coeff_hz_vec, res_hz0, res_hz1);
568 v16u8 src0, src1, src2, src3, src4;
581 LD_UB5(src, stride, src0, src1, src2, src3, src4);
583 VSHF_B2_UB(src0, src1, src2, src3, mask, mask, tmp0, tmp1);
585 ILVR_D2_UB(tmp1, tmp0, tmp3, tmp2, src0, src1);
586 DOTP_UB2_UH(src0, src1, coeff_hz_vec, coeff_hz_vec, res_hz0, res_hz1);
616 v16u8 src0, src1, src2;
627 LD_UB3(src, stride, src0, src1, src2);
628 VSHF_B2_UB(src0, src1, src1, src2, mask, mask, src0, src1);
629 DOTP_UB2_UH(src0, src1, coeff_hz_vec, coeff_hz_vec, res_hz0, res_hz1);
644 v16u8 src0, src1, src2, src3, src4;
657 LD_UB5(src, stride, src0, src1, src2, src3, src4);
658 VSHF_B2_UB(src0, src1, src1, src2, mask, mask, src0, src1);
660 DOTP_UB4_UH(src0, src1, src2, src3, coeff_hz_vec, coeff_hz_vec,
677 v16u8 src0, src1, src2, src3, src4, src5, src6, src7, src8, res0, res1;
689 LD_UB5(src, stride, src0, src1, src2, src3, src4);
693 VSHF_B2_UB(src0, src1, src1, src2, mask, mask, src0, src1);
697 DOTP_UB4_UH(src0, src1, src2, src3, coeff_hz_vec, coeff_hz_vec,
734 v16u8 src0, src1, src2, src3, src4, out0, out1;
746 src0 = LD_UB(src);
749 src0 = (v16u8) __msa_vshf_b(mask, (v16i8) src0, (v16i8) src0);
750 res_hz0 = __msa_dotp_u_h(src0, coeff_hz_vec);
777 v16u8 src0, src1, src2, src3, src4, src5, src6, src7, src8;
792 LD_UB5(src, stride, src0, src1, src2, src3, src4);
795 src0 = (v16u8) __msa_vshf_b(mask, (v16i8) src0, (v16i8) src0);
800 res_hz0 = __msa_dotp_u_h(src0, coeff_hz_vec);
848 v16i8 src0, src1;
859 LD_SB2(src, stride, src0, src1);
867 src0 = __msa_vshf_b(mask, src1, src0);
869 res_r = __msa_dotp_u_h((v16u8) src0, coeff_vec);
890 v16u8 src0, src1, src2, src3;
900 LD_UB4(src, stride, src0, src1, src2, src3);
910 VSHF_B2_UB(src0, src1, src2, src3, mask, mask, src0, src2);
912 src0 = (v16u8) __msa_ilvr_d((v2i64) src2, (v2i64) src0);
914 res_r = __msa_dotp_u_h(src0, coeff_vec);
941 v16i8 src0, src1;
951 LD_SB2(src, stride, src0, src1);
957 src0 = __msa_vshf_b(mask, src1, src0);
959 res_r = __msa_dotp_u_h((v16u8) src0, coeff_vec);
974 v16u8 src0, src1, src2, src3;
984 LD_UB4(src, stride, src0, src1, src2, src3);
987 VSHF_B2_UB(src0, src1, src2, src3, mask, mask, src0, src2);
988 DOTP_UB2_UH(src0, src2, coeff_vec, coeff_vec, res0_r, res1_r);
1003 v16u8 src0, src1, src2, src3, src4, src5, src6, src7, out0, out1;
1013 LD_UB8(src, stride, src0, src1, src2, src3, src4, src5, src6, src7);
1018 VSHF_B2_UB(src0, src1, src2, src3, mask, mask, src0, src2);
1020 DOTP_UB2_UH(src0, src2, coeff_vec, coeff_vec, res0, res1);
1048 v16u8 src0, src1, src2, src3, out0, out1;
1057 LD_UB4(src, stride, src0, src1, src2, src3);
1061 VSHF_B2_UB(src0, src0, src1, src1, mask, mask, src0, src1);
1063 DOTP_UB4_UH(src0, src1, src2, src3, coeff_vec, coeff_vec, coeff_vec,
1078 v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
1089 LD_UB8(src, stride, src0, src1, src2, src3, src4, src5, src6, src7);
1096 VSHF_B2_UB(src0, src0, src1, src1, mask, mask, src0, src1);
1100 DOTP_UB4_UH(src0, src1, src2, src3, coeff_vec, coeff_vec, coeff_vec,
1133 v16i8 src0, src1, src2, tmp0, tmp1, res;
1141 LD_SB3(src, stride, src0, src1, src2);
1148 ILVR_B2_SB(src1, src0, src2, src1, tmp0, tmp1);
1170 v16i8 src0, src1, src2, src3, src4;
1179 LD_SB5(src, stride, src0, src1, src2, src3, src4);
1190 ILVR_B4_UB(src1, src0, src2, src1, src3, src2, src4, src3,
1223 v16u8 src0, src1, src2, tmp0, tmp1;
1231 LD_UB3(src, stride, src0, src1, src2);
1236 ILVR_B2_UB(src1, src0, src2, src1, tmp0, tmp1);
1255 v16u8 src0, src1, src2, src3, src4;
1264 LD_UB5(src, stride, src0, src1, src2, src3, src4);
1267 ILVR_B4_UB(src1, src0, src2, src1, src3, src2, src4, src3, tmp0, tmp1, tmp2,
1285 v16u8 src0, src1, src2, src3, src4, src5, src6, src7, src8;
1293 LD_UB5(src, stride, src0, src1, src2, src3, src4);
1300 ILVR_B4_UB(src1, src0, src2, src1, src3, src2, src4, src3, tmp0, tmp1, tmp2,
1334 v16u8 src0, src1, src2, src3, src4;
1342 LD_UB5(src, stride, src0, src1, src2, src3, src4);
1346 ILVR_B4_UB(src1, src0, src2, src1, src3, src2, src4, src3,
1347 src0, src1, src2, src3);
1348 DOTP_UB4_UH(src0, src1, src2, src3, coeff_vec, coeff_vec, coeff_vec,
1363 v16u8 src0, src1, src2, src3, src4, src5, src6, src7, src8;
1371 LD_UB5(src, stride, src0, src1, src2, src3, src4);
1380 ILVR_B4_UB(src1, src0, src2, src1, src3, src2, src4, src3,
1381 src0, src1, src2, src3);
1384 DOTP_UB4_UH(src0, src1, src2, src3, coeff_vec, coeff_vec, coeff_vec,
1421 v16u8 src0, src1, src2;
1432 LD_UB3(src, stride, src0, src1, src2);
1437 VSHF_B2_UB(src0, src1, src1, src2, mask, mask, src0, src1);
1438 DOTP_UB2_UH(src0, src1, coeff_hz_vec, coeff_hz_vec, res_hz0, res_hz1);
1462 v16u8 src0, src1, src2, src3, src4;
1475 LD_UB5(src, stride, src0, src1, src2, src3, src4);
1484 VSHF_B2_UB(src0, src1, src2, src3, mask, mask, tmp0, tmp1);
1486 ILVR_D2_UB(tmp1, tmp0, tmp3, tmp2, src0, src1);
1487 DOTP_UB2_UH(src0, src1, coeff_hz_vec, coeff_hz_vec, res_hz0, res_hz1);
1524 v16u8 src0, src1, src2;
1536 LD_UB3(src, stride, src0, src1, src2);
1539 VSHF_B2_UB(src0, src1, src1, src2, mask, mask, src0, src1);
1540 DOTP_UB2_UH(src0, src1, coeff_hz_vec, coeff_hz_vec, res_hz0, res_hz1);
1560 v16u8 src0, src1, src2, src3, src4;
1573 LD_UB5(src, stride, src0, src1, src2, src3, src4);
1576 VSHF_B2_UB(src0, src1, src1, src2, mask, mask, src0, src1);
1578 DOTP_UB4_UH(src0, src1, src2, src3, coeff_hz_vec, coeff_hz_vec,
1599 v16u8 src0, src1, src2, src3, src4, src5, src6, src7, src8, res0, res1;
1612 LD_UB5(src, stride, src0, src1, src2, src3, src4);
1619 VSHF_B2_UB(src0, src1, src1, src2, mask, mask, src0, src1);
1623 DOTP_UB4_UH(src0, src1, src2, src3, coeff_hz_vec, coeff_hz_vec,
1668 v16u8 src0, src1, src2, src3, src4, out0, out1;
1682 src0 = LD_UB(src);
1684 src0 = (v16u8) __msa_vshf_b(mask, (v16i8) src0, (v16i8) src0);
1685 res_hz0 = __msa_dotp_u_h(src0, coeff_hz_vec);
1716 v16u8 src0, src1, src2, src3, src4, src5, src6, src7, src8;
1732 LD_UB5(src, stride, src0, src1, src2, src3, src4);
1735 src0 = (v16u8) __msa_vshf_b(mask, (v16i8) src0, (v16i8) src0);
1740 res_hz0 = __msa_dotp_u_h(src0, coeff_hz_vec);
1820 uint64_t src0, src1, src2, src3, src4, src5, src6, src7;
1823 LD4(src, stride, src0, src1, src2, src3);
1826 SD4(src0, src1, src2, src3, dst, stride);
1830 LD4(src, stride, src0, src1, src2, src3);
1831 SD4(src0, src1, src2, src3, dst, stride);
1839 v16u8 src0 = { 0 }, src1 = { 0 }, dst0 = { 0 }, dst1 = { 0 };
1844 INSERT_W4_UB(tp0, tp1, tp2, tp3, src0);
1851 AVER_UB2_UB(src0, dst0, src1, dst1, dst0, dst1);
1855 INSERT_W4_UB(tp0, tp1, tp2, tp3, src0);
1858 dst0 = __msa_aver_u_b(src0, dst0);
1862 INSERT_W2_UB(tp0, tp1, src0);
1865 dst0 = __msa_aver_u_b(src0, dst0);
1874 v16u8 src0 = { 0 }, src1 = { 0 }, src2 = { 0 }, src3 = { 0 };
1881 INSERT_D2_UB(tp0, tp1, src0);
1891 AVER_UB4_UB(src0, dst0, src1, dst1, src2, dst2, src3, dst3, dst0, dst1,
1896 INSERT_D2_UB(tp0, tp1, src0);
1901 AVER_UB2_UB(src0, dst0, src1, dst1, dst0, dst1);