Lines Matching refs:src2

44     __m256i src0, src1, src2, src3, src4, out;
55 src1, src2, src3, src4);
56 DUP2_ARG3(__lasx_xvpermi_q, src2, src1, 0x20, src4, src3, 0x20, src1, src3);
82 __m256i src0, src1, src2, src3, src4, src5, src6, src7, src8;
95 src1, src2, src3, src4);
99 DUP4_ARG3(__lasx_xvpermi_q, src2, src1, 0x20, src4, src3, 0x20, src6, src5, 0x20,
135 __m256i src0, src1, src2, src3, out;
144 DUP2_ARG2(__lasx_xvldx, src, stride, src, stride_2x, src1, src2);
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;
174 src1, src2, src3, src4);
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);
247 __m256i src0, src1, src2, src3, src4, out;
257 src1, src2, src3, src4);
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;
286 src1, src2, src3, src4);
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;
405 DUP2_ARG2(__lasx_xvldx, src, stride, src, stride_2, src1, src2);
406 DUP2_ARG3(__lasx_xvshuf_b, src1, src0, mask, src2, src1, mask, src0, src1);
424 __m256i src0, src1, src2, src3, src4;
435 src, stride_4, src1, src2, src3, src4);
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);
456 __m256i src0, src1, src2, src3, src4, src5, src6, src7, src8;
468 src, stride_4, src1, src2, src3, src4);
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,
536 __m256i src0, src1, src2, src3;
543 DUP2_ARG2(__lasx_xvldx, src, stride, src, stride_2, src1, src2);
545 DUP2_ARG3(__lasx_xvshuf_b, src1, src0, mask, src3, src2, mask, src0, src2);
546 src0 = __lasx_xvpermi_q(src0, src2, 0x02);
562 __m256i src0, src1, src2, src3, src4, src5, src6, src7;
571 src, stride_4, src1, src2, src3, src4);
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);
620 __m256i src0, src1, src2;
628 DUP2_ARG2(__lasx_xvldx, src, stride, src, stride << 1, src1, src2);
629 DUP2_ARG2(__lasx_xvilvl_b, src1, src0, src2, src1, tmp0, tmp1);
644 __m256i src0, src1, src2, src3, src4;
653 src, stride_4, src1, src2, src3, src4);
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;
683 src, stride_4, src1, src2, src3, src4);
687 DUP4_ARG2(__lasx_xvilvl_b, src1, src0, src2, src1, src3, src2, src4, src3,
853 __m256i src0, src1, src2, src3, src4, out;
864 src1, src2, src3, src4);
865 DUP2_ARG3(__lasx_xvpermi_q, src2, src1, 0x20, src4, src3, 0x20, src1, src3);
897 __m256i src0, src1, src2, src3, src4, src5, src6, src7, src8;
911 src1, src2, src3, src4);
915 DUP4_ARG3(__lasx_xvpermi_q, src2, src1, 0x20, src4, src3, 0x20, src6, src5, 0x20,
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;
1063 src1, src2, src3, src4);
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;
1099 src1, src2, src3, src4);
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;
1146 src2 = __lasx_xvldrepl_d(src + stride_2x, 0);
1153 src2 = __lasx_xvpackev_d(src3,src2);
1154 src0 = __lasx_xvpermi_q(src0, src2, 0x02);
1168 src2 = __lasx_xvldrepl_d(src + stride_2x, 0);
1175 src2 = __lasx_xvpackev_d(src3,src2);
1176 src0 = __lasx_xvpermi_q(src0, src2, 0x02);
1190 __m256i src0, src1, src2, src3;
1197 src2 = __lasx_xvldrepl_d(src + stride_2x, 0);
1204 src2 = __lasx_xvpackev_d(src3,src2);
1205 src0 = __lasx_xvpermi_q(src0, src2, 0x02);