Lines Matching refs:src0
44 __m256i src0, src1, src2, src3, src4, out;
53 DUP2_ARG2(__lasx_xvld, chroma_mask_arr, 0, src, 0, mask, src0);
57 src0 = __lasx_xvshuf_b(src0, src0, mask);
59 DUP2_ARG2(__lasx_xvdp2_h_bu, src0, coeff_hz_vec, src1, coeff_hz_vec, res_hz0, res_hz1);
82 __m256i src0, src1, src2, src3, src4, src5, src6, src7, src8;
93 DUP2_ARG2(__lasx_xvld, chroma_mask_arr, 0, src, 0, mask, src0);
101 src0 = __lasx_xvshuf_b(src0, src0, mask);
104 DUP4_ARG2(__lasx_xvdp2_h_bu, src0, coeff_hz_vec, src1, coeff_hz_vec, src3,
135 __m256i src0, src1, src2, src3, out;
143 DUP2_ARG2(__lasx_xvld, chroma_mask_arr, 0, src, 0, mask, src0);
146 DUP2_ARG3(__lasx_xvpermi_q, src1, src0, 0x20, src3, src2, 0x20, src0, src2);
147 DUP2_ARG3(__lasx_xvshuf_b, src0, src0, mask, src2, src2, mask, src0, src2);
148 DUP2_ARG2(__lasx_xvdp2_h_bu, src0, coeff_vec, src2, coeff_vec, res0, res1);
163 __m256i src0, src1, src2, src3, src4, src5, src6, src7;
172 DUP2_ARG2(__lasx_xvld, chroma_mask_arr, 0, src, 0, mask, src0);
178 DUP4_ARG3(__lasx_xvpermi_q, src1, src0, 0x20, src3, src2, 0x20, src5, src4, 0x20,
179 src7, src6, 0x20, src0, src2, src4, src6);
180 DUP4_ARG3(__lasx_xvshuf_b, src0, src0, mask, src2, src2, mask, src4, src4, mask,
181 src6, src6, mask, src0, src2, src4, src6);
182 DUP4_ARG2(__lasx_xvdp2_h_bu, src0, coeff_vec, src2, coeff_vec, src4, coeff_vec, src6,
204 __m256i src0, src1, src2, src3, out;
216 src0, src1, src2, src3);
218 DUP2_ARG3(__lasx_xvpermi_q, src1, src0, 0x20, src3, src2, 0x20, src0, src2);
219 DUP2_ARG3(__lasx_xvshuf_b, src0, src0, mask, src2, src2, mask, src0, src2);
220 DUP2_ARG2(__lasx_xvdp2_h_bu, src0, coeff_vec, src2, coeff_vec, res0, res1);
230 src0 = __lasx_xvld(src, 0);
232 src1 = __lasx_xvpermi_q(src1, src0, 0x20);
233 src0 = __lasx_xvshuf_b(src1, src1, mask);
234 res0 = __lasx_xvdp2_h_bu(src0, coeff_vec);
247 __m256i src0, src1, src2, src3, src4, out;
254 src0 = __lasx_xvld(src, 0);
258 DUP4_ARG3(__lasx_xvpermi_q, src1, src0, 0x20, src2, src1, 0x20, src3, src2, 0x20,
259 src4, src3, 0x20, src0, src1, src2, src3);
260 DUP2_ARG2(__lasx_xvilvl_b, src1, src0, src3, src2, src0, src2);
261 DUP2_ARG2(__lasx_xvdp2_h_bu, src0, coeff_vec, src2, coeff_vec, res0, res1);
275 __m256i src0, src1, src2, src3, src4, src5, src6, src7, src8;
283 src0 = __lasx_xvld(src, 0);
290 DUP4_ARG3(__lasx_xvpermi_q, src1, src0, 0x20, src2, src1, 0x20, src3, src2, 0x20,
291 src4, src3, 0x20, src0, src1, src2, src3);
294 DUP4_ARG2(__lasx_xvilvl_b, src1, src0, src3, src2, src5, src4, src7, src6,
295 src0, src2, src4, src6);
296 DUP4_ARG2(__lasx_xvdp2_h_bu, src0, coeff_vec, src2, coeff_vec, src4, coeff_vec,
394 __m256i src0, src1, src2;
404 DUP2_ARG2(__lasx_xvld, chroma_mask_arr, 32, src, 0, mask, src0);
406 DUP2_ARG3(__lasx_xvshuf_b, src1, src0, mask, src2, src1, mask, src0, src1);
407 src0 = __lasx_xvpermi_q(src0, src1, 0x02);
408 res_hz = __lasx_xvdp2_h_bu(src0, coeff_hz_vec);
424 __m256i src0, src1, src2, src3, src4;
433 DUP2_ARG2(__lasx_xvld, chroma_mask_arr, 32, src, 0, mask, src0);
436 DUP4_ARG3(__lasx_xvshuf_b, src1, src0, mask, src2, src1, mask, src3, src2, mask,
437 src4, src3, mask, src0, src1, src2, src3);
438 DUP2_ARG3(__lasx_xvpermi_q, src0, src2, 0x02, src1, src3, 0x02, src0, src1);
439 DUP2_ARG2(__lasx_xvdp2_h_bu, src0, coeff_hz_vec, src1, coeff_hz_vec, res_hz0, res_hz1);
456 __m256i src0, src1, src2, src3, src4, src5, src6, src7, src8;
466 DUP2_ARG2(__lasx_xvld, chroma_mask_arr, 32, src, 0, mask, src0);
472 DUP4_ARG3(__lasx_xvshuf_b, src1, src0, mask, src2, src1, mask, src3, src2, mask,
473 src4, src3, mask, src0, src1, src2, src3);
476 DUP4_ARG3(__lasx_xvpermi_q, src0, src2, 0x02, src1, src3, 0x02, src4, src6, 0x02,
477 src5, src7, 0x02, src0, src1, src4, src5);
478 DUP4_ARG2(__lasx_xvdp2_h_bu, src0, coeff_hz_vec, src1, coeff_hz_vec, src4, coeff_hz_vec,
515 __m256i src0, src1;
521 DUP2_ARG2(__lasx_xvld, chroma_mask_arr, 32, src, 0, mask, src0);
523 src0 = __lasx_xvshuf_b(src1, src0, mask);
524 res = __lasx_xvdp2_h_bu(src0, coeff_vec);
536 __m256i src0, src1, src2, src3;
542 DUP2_ARG2(__lasx_xvld, chroma_mask_arr, 32, src, 0, mask, src0);
545 DUP2_ARG3(__lasx_xvshuf_b, src1, src0, mask, src3, src2, mask, src0, src2);
546 src0 = __lasx_xvpermi_q(src0, src2, 0x02);
547 res = __lasx_xvdp2_h_bu(src0, coeff_vec);
562 __m256i src0, src1, src2, src3, src4, src5, src6, src7;
569 DUP2_ARG2(__lasx_xvld, chroma_mask_arr, 32, src, 0, mask, src0);
575 DUP4_ARG3(__lasx_xvshuf_b, src1, src0, mask, src3, src2, mask, src5, src4, mask,
576 src7, src6, mask, src0, src2, src4, src6);
577 DUP2_ARG3(__lasx_xvpermi_q, src0, src2, 0x02, src4, src6, 0x02, src0, src4);
578 DUP2_ARG2(__lasx_xvdp2_h_bu, src0, coeff_vec, src4, coeff_vec, res0, res1);
620 __m256i src0, src1, src2;
627 src0 = __lasx_xvld(src, 0);
629 DUP2_ARG2(__lasx_xvilvl_b, src1, src0, src2, src1, tmp0, tmp1);
644 __m256i src0, src1, src2, src3, src4;
651 src0 = __lasx_xvld(src, 0);
654 DUP4_ARG2(__lasx_xvilvl_b, src1, src0, src2, src1, src3, src2, src4, src3,
673 __m256i src0, src1, src2, src3, src4, src5, src6, src7, src8;
681 src0 = __lasx_xvld(src, 0);
687 DUP4_ARG2(__lasx_xvilvl_b, src1, src0, src2, src1, src3, src2, src4, src3,
853 __m256i src0, src1, src2, src3, src4, out;
862 DUP2_ARG2(__lasx_xvld, chroma_mask_arr, 0, src, 0, mask, src0);
866 src0 = __lasx_xvshuf_b(src0, src0, mask);
868 DUP2_ARG2(__lasx_xvdp2_h_bu, src0, coeff_hz_vec, src1, coeff_hz_vec, res_hz0, res_hz1);
897 __m256i src0, src1, src2, src3, src4, src5, src6, src7, src8;
908 DUP2_ARG2(__lasx_xvld, chroma_mask_arr, 0, src, 0, mask, src0);
917 src0 = __lasx_xvshuf_b(src0, src0, mask);
920 DUP4_ARG2(__lasx_xvdp2_h_bu, src0, coeff_hz_vec, src1, coeff_hz_vec, src3,
967 __m256i src0, src1, src2, src3, out;
977 src0, src1, src2, src3);
978 DUP2_ARG3(__lasx_xvpermi_q, src1, src0, 0x20, src3, src2, 0x20, src0, src2);
979 DUP2_ARG3(__lasx_xvshuf_b, src0, src0, mask, src2, src2, mask, src0, src2);
980 DUP2_ARG2(__lasx_xvdp2_h_bu, src0, coeff_vec, src2, coeff_vec, res0, res1);
1001 __m256i src0, src1, src2, src3, src4, src5, src6, src7;
1012 src0, src1, src2, src3);
1016 DUP4_ARG3(__lasx_xvpermi_q, src1, src0, 0x20, src3, src2, 0x20, src5, src4, 0x20,
1017 src7, src6, 0x20, src0, src2, src4, src6);
1018 DUP4_ARG3(__lasx_xvshuf_b, src0, src0, mask, src2, src2, mask, src4, src4,
1019 mask, src6, src6, mask, src0, src2, src4, src6);
1020 DUP4_ARG2(__lasx_xvdp2_h_bu, src0, coeff_vec, src2, coeff_vec, src4, coeff_vec, src6,
1054 __m256i src0, src1, src2, src3, src4, out;
1061 src0 = __lasx_xvld(src, 0);
1064 DUP4_ARG3(__lasx_xvpermi_q, src1, src0, 0x20, src2, src1, 0x20, src3, src2, 0x20,
1065 src4, src3, 0x20, src0, src1, src2, src3);
1066 DUP2_ARG2(__lasx_xvilvl_b, src1, src0, src3, src2, src0, src2);
1067 DUP2_ARG2(__lasx_xvdp2_h_bu, src0, coeff_vec, src2, coeff_vec, res0, res1);
1088 __m256i src0, src1, src2, src3, src4, src5, src6, src7, src8;
1096 src0 = __lasx_xvld(src, 0);
1103 DUP4_ARG3(__lasx_xvpermi_q, src1, src0, 0x20, src2, src1, 0x20, src3, src2, 0x20,
1104 src4, src3, 0x20, src0, src1, src2, src3);
1107 DUP4_ARG2(__lasx_xvilvl_b, src1, src0, src3, src2, src5, src4, src7, src6,
1108 src0, src2, src4, src6);
1109 DUP4_ARG2(__lasx_xvdp2_h_bu, src0, coeff_vec, src2, coeff_vec, src4, coeff_vec, src6,
1138 __m256i src0, src1, src2, src3;
1144 src0 = __lasx_xvldrepl_d(src, 0);
1152 src0 = __lasx_xvpackev_d(src1,src0);
1154 src0 = __lasx_xvpermi_q(src0, src2, 0x02);
1158 dst0 = __lasx_xvavgr_bu(src0, dst0);
1166 src0 = __lasx_xvldrepl_d(src, 0);
1174 src0 = __lasx_xvpackev_d(src1,src0);
1176 src0 = __lasx_xvpermi_q(src0, src2, 0x02);
1180 dst0 = __lasx_xvavgr_bu(src0, dst0);
1190 __m256i src0, src1, src2, src3;
1195 src0 = __lasx_xvldrepl_d(src, 0);
1203 src0 = __lasx_xvpackev_d(src1,src0);
1205 src0 = __lasx_xvpermi_q(src0, src2, 0x02);
1209 dst0 = __lasx_xvavgr_bu(src0, dst0);