xref: /third_party/ffmpeg/libavcodec/loongarch/hevcdsp_lsx.c

613     __m128i vec0, vec1, vec2, vec3;
634                   src0, mask2, src1, src0, mask3, vec0, vec1, vec2, vec3);
635         dst0 = __lsx_vdp2_h_bu_b(vec0, filt0);
640                   src2, mask2, src3, src2, mask3, vec0, vec1, vec2, vec3);
641         dst1 = __lsx_vdp2_h_bu_b(vec0, filt0);
646                   src4, mask2, src5, src4, mask3, vec0, vec1, vec2, vec3);
647         dst2 = __lsx_vdp2_h_bu_b(vec0, filt0);
652                   src6, mask2, src7, src6, mask3, vec0, vec1, vec2, vec3);
653         dst3 = __lsx_vdp2_h_bu_b(vec0, filt0);
673                   src0, mask2, src1, src0, mask3, vec0, vec1, vec2, vec3);
674         dst0 = __lsx_vdp2_h_bu_b(vec0, filt0);
699     __m128i vec0, vec1, vec2, vec3;
717                   src0, mask2, src0, src0, mask3, vec0, vec1, vec2, vec3);
718         dst0 = __lsx_vdp2_h_bu_b(vec0, filt0);
723                   src1, mask2, src1, src1, mask3, vec0, vec1, vec2, vec3);
724         dst1 = __lsx_vdp2_h_bu_b(vec0, filt0);
729                   src2, mask2, src2, src2, mask3, vec0, vec1, vec2, vec3);
730         dst2 = __lsx_vdp2_h_bu_b(vec0, filt0);
735                   src3, mask2, src3, src3, mask3, vec0, vec1, vec2, vec3);
736         dst3 = __lsx_vdp2_h_bu_b(vec0, filt0);
760     __m128i vec0, vec1, vec2, vec3, vec4, vec5;
787                   vec0, vec1);
792         DUP4_ARG2(__lsx_vdp2_h_bu_b, vec0, filt0, vec1, filt0, vec2, filt0,
796                   vec0, vec1);
801         DUP4_ARG3(__lsx_vdp2add_h_bu_b, dst0, vec0, filt1, dst1, vec1, filt1,
806                   vec0, vec1);
811         DUP4_ARG3(__lsx_vdp2add_h_bu_b, dst0, vec0, filt2, dst1, vec1, filt2,
816                   vec0, vec1);
821         DUP4_ARG3(__lsx_vdp2add_h_bu_b, dst0, vec0, filt3, dst1, vec1, filt3,
849     __m128i vec0, vec1, vec2, vec3;
868                   vec0, vec1);
871         DUP4_ARG2(__lsx_vdp2_h_bu_b, vec0, filt0, vec1, filt0, vec2, filt0,
874                   vec0, vec1);
877         DUP4_ARG3(__lsx_vdp2add_h_bu_b, dst0, vec0, filt1, dst1, vec1, filt1,
880                   vec0, vec1);
883         DUP4_ARG3(__lsx_vdp2add_h_bu_b, dst0, vec0, filt2, dst1, vec1, filt2,
886                   vec0, vec1);
889         DUP4_ARG3(__lsx_vdp2add_h_bu_b, dst0, vec0, filt3, dst1, vec1, filt3,
909     __m128i vec0, vec1, vec2, vec3, vec4, vec5;
929                   src1, mask0, src2, src2, mask0, vec0, vec1, vec2, vec3);
932         DUP4_ARG2(__lsx_vdp2_h_bu_b, vec0, filt0, vec1, filt0, vec2, filt0,
936                   src1, mask1, src2, src2, mask1, vec0, vec1, vec2, vec3);
939         DUP4_ARG3(__lsx_vdp2add_h_bu_b, dst0, vec0, filt1, dst1, vec1, filt1,
944                   src1, mask2, src2, src2, mask2, vec0, vec1, vec2, vec3);
947         DUP4_ARG3(__lsx_vdp2add_h_bu_b, dst0, vec0, filt2, dst1, vec1, filt2,
952                   src1, mask3, src2, src2, mask3, vec0, vec1, vec2, vec3);
955         DUP4_ARG3(__lsx_vdp2add_h_bu_b, dst0, vec0, filt3, dst1, vec1, filt3,
979     __m128i vec0, vec1, vec2, vec3;
998                   src0, mask2, src0, src0, mask3, vec0, vec1, vec2, vec3);
999         dst0 = __lsx_vdp2_h_bu_b(vec0, filt0);
1005                   src0, mask6, src1, src0, mask7, vec0, vec1, vec2, vec3);
1006         dst1 = __lsx_vdp2_h_bu_b(vec0, filt0);
1011                   mask2, src1, src1, mask3, vec0, vec1, vec2, vec3);
1012         dst2 = __lsx_vdp2_h_bu_b(vec0, filt0);
1017                   mask2, src2, src2, mask3, vec0, vec1, vec2, vec3);
1018         dst3 = __lsx_vdp2_h_bu_b(vec0, filt0);
1039     __m128i vec0, vec1, vec2, vec3, vec4, vec5;
1059                   mask0, src2, src1, mask4, vec0, vec1, vec2, vec3);
1060         DUP4_ARG2(__lsx_vdp2_h_bu_b, vec0, filt0, vec1, filt0, vec2, filt0,
1063                   src1, mask1, src2, src1, mask5, vec0, vec1, vec2, vec3);
1064         DUP4_ARG3(__lsx_vdp2add_h_bu_b, dst0, vec0, filt1, dst1, vec1, filt1,
1067                   src1, mask2, src2, src1, mask6, vec0, vec1, vec2, vec3);
1068         DUP4_ARG3(__lsx_vdp2add_h_bu_b, dst0, vec0, filt2, dst1, vec1, filt2,
1071                   src1, mask3, src2, src1, mask7, vec0, vec1, vec2, vec3);
1072         DUP4_ARG3(__lsx_vdp2add_h_bu_b, dst0, vec0, filt3, dst1, vec1, filt3,
1108     __m128i vec0, vec1, vec2, vec3;
1128                   src0, mask2, src0, src0, mask3, vec0, vec1, vec2, vec3);
1129         dst0 = __lsx_vdp2_h_bu_b(vec0, filt0);
1136                   src0, mask6, src1, src0, mask7, vec0, vec1, vec2, vec3);
1137         dst1 = __lsx_vdp2_h_bu_b(vec0, filt0);
1144                   src1, mask2, src1, src1, mask3, vec0, vec1, vec2, vec3);
1145         dst2 = __lsx_vdp2_h_bu_b(vec0, filt0);
1152                   src1, mask6, src2, src1, mask7, vec0, vec1, vec2, vec3);
1153         dst3 = __lsx_vdp2_h_bu_b(vec0, filt0);
1160                   src2, mask2, src2, src2, mask3, vec0, vec1, vec2, vec3);
1161         dst4 = __lsx_vdp2_h_bu_b(vec0, filt0);
1168                   src2, mask6, src3, src2, mask7, vec0, vec1, vec2, vec3);
1169         dst5 = __lsx_vdp2_h_bu_b(vec0, filt0);
1176                   src3, mask2, src3, src3, mask3, vec0, vec1, vec2, vec3);
1177         dst6 = __lsx_vdp2_h_bu_b(vec0, filt0);
1184                   src4, mask2, src4, src4, mask3, vec0, vec1, vec2, vec3);
1185         dst7 = __lsx_vdp2_h_bu_b(vec0, filt0);
1670     __m128i vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
1700               mask2, src3, src0, mask3, vec0, vec1, vec2, vec3);
1707     dst30 = __lsx_vdp2_h_bu_b(vec0, filt0);
1737                   mask2, src9, src7, mask3, vec0, vec1, vec2, vec3);
1741         dst97 = __lsx_vdp2_h_bu_b(vec0, filt0);
1808     __m128i vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
1844                   src0, mask2, src0, src0, mask3, vec0, vec1, vec2, vec3);
1851         dst0 = __lsx_vdp2_h_bu_b(vec0, filt0);
1870                   src4, mask2, src4, src4, mask3, vec0, vec1, vec2, vec3);
1875         dst4 = __lsx_vdp2_h_bu_b(vec0, filt0);
1893                       src7, mask2, src7, src7, mask3, vec0, vec1, vec2, vec3);
1894             dst7 = __lsx_vdp2_h_bu_b(vec0, filt0);
1953     __m128i vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
1992               mask2, src0, src0, mask3, vec0, vec1, vec2, vec3);
1999     dst0 = __lsx_vdp2_h_bu_b(vec0, filt0);
2018               mask2, src4, src4, mask3, vec0, vec1, vec2, vec3);
2023     dst4 = __lsx_vdp2_h_bu_b(vec0, filt0);
2041                   src7, mask2, src7, src7, mask3, vec0, vec1, vec2, vec3);
2042         dst7 = __lsx_vdp2_h_bu_b(vec0, filt0);
2088               mask6, src3, src0, mask7, vec0, vec1, vec2, vec3);
2095     dst30 = __lsx_vdp2_h_bu_b(vec0, filt0);
2126                   src7, mask6, src9, src7, mask7, vec0, vec1, vec2, vec3);
2129         dst97 = __lsx_vdp2_h_bu_b(vec0, filt0);
2235     __m128i vec0, vec1, vec2, vec3;
2249                   vec0, vec1);
2252         DUP4_ARG2(__lsx_vdp2_h_bu_b, vec0, filt0, vec1, filt0, vec2, filt0,
2255                   vec0, vec1);
2258         DUP4_ARG3(__lsx_vdp2add_h_bu_b, dst0, vec0, filt1, dst1, vec1, filt1,
2547     __m128i vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, vec8, vec9;
2567     DUP2_ARG3(__lsx_vshuf_b, src0, src0, mask0, src0, src0, mask1, vec0, vec1);
2573     DUP4_ARG2(__lsx_vdp2_h_bu_b,  vec0, filt0, vec2, filt0, vec4, filt0,
2611     __m128i vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
2637                   vec0, vec1);
2643         DUP2_ARG2(__lsx_vdp2_h_bu_b, vec0, filt0, vec2, filt0, dst0, dst1);
2653                   vec0, vec1);
2660         DUP4_ARG2(__lsx_vdp2_h_bu_b, vec0, filt0, vec2, filt0, vec4, filt0,
2713     __m128i vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, vec8, vec9;
2740               mask0, src1, src1, mask1, vec0, vec1, vec2, vec3);
2750     DUP4_ARG2(__lsx_vdp2_h_bu_b, vec0, filt0, vec2, filt0, vec4, filt0, vec6,
2830     __m128i vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
2855                   vec0, vec1);
2861         DUP2_ARG2(__lsx_vdp2_h_bu_b, vec0, filt0, vec2, filt0, dst0, dst1);
2878                       vec0, vec1);
2886             DUP4_ARG2(__lsx_vdp2_h_bu_b, vec0, filt0, vec2, filt0, vec4, filt0,
2976     __m128i vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
3004     DUP2_ARG3(__lsx_vshuf_b, src0, src0, mask0, src0, src0, mask1, vec0, vec1);
3008     DUP2_ARG2(__lsx_vdp2_h_bu_b, vec0, filt0, vec2, filt0, dst0, dst1);
3025                   vec0, vec1);
3033         DUP4_ARG2(__lsx_vdp2_h_bu_b, vec0, filt0, vec2, filt0, vec4, filt0,
3082     DUP2_ARG3(__lsx_vshuf_b, src1, src0, mask2, src1, src0, mask3, vec0, vec1);
3084     DUP2_ARG2(__lsx_vdp2_h_bu_b, vec0, filt0, vec2, filt0, dst10, dst21);
3101                   vec0, vec1);
3109         DUP4_ARG2(__lsx_vdp2_h_bu_b, vec0, filt0, vec2, filt0, vec4, filt0,

Indexes created Thu Nov 07 10:32:03 CST 2024