Lines Matching defs:vec0

52 #define FILT_8TAP_DPADD_S_H(vec0, vec1, vec2, vec3,             \
57 tmp0 = __msa_dotp_s_h((v16i8) vec0, (v16i8) filt0); \
846 v16u8 mask0, mask1, mask2, mask3, vec0, vec1;
920 vec0 = PCKEV_XORI128_UB(tmp0, tmp1);
922 ST_D4(vec0, vec1, 0, 1, 0, 1, dst, dst_stride);
1030 v8i16 filt, vec0, vec1, vec2, vec3;
1051 mask3, filt0, filt1, filt2, filt3, vec0, vec1);
1056 SRARI_H4_SH(vec0, vec1, vec2, vec3, 7);
1057 SAT_SH4_SH(vec0, vec1, vec2, vec3, 7);
1058 PCKEV_B4_UB(vec0, vec0, vec1, vec1, vec2, vec2, vec3, vec3,
1132 v8i16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
1152 VSHF_B4_SH(src0, src0, mask0, mask1, mask2, mask3, vec0, vec4, vec8,
1160 DOTP_SB4_SH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec0,
1164 DPADD_SB4_SH(vec4, vec5, vec6, vec7, filt1, filt1, filt1, filt1, vec0,
1168 ADDS_SH4_SH(vec0, vec8, vec1, vec9, vec2, vec10, vec3, vec11, out0,
1190 v8i16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
1212 VSHF_B4_SH(src0, src0, mask0, mask1, mask2, mask3, vec0, vec4, vec8,
1220 DOTP_SB4_SH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec0,
1224 DPADD_SB4_SH(vec4, vec5, vec6, vec7, filt1, filt1, filt1, filt1, vec0,
1228 ADDS_SH4_SH(vec0, vec8, vec1, vec9, vec2, vec10, vec3, vec11, out0,
1249 v8i16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
1271 VSHF_B4_SH(src0, src0, mask0, mask1, mask2, mask3, vec0, vec4, vec8,
1279 DOTP_SB4_SH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0,
1280 vec0, vec1, vec2, vec3);
1284 vec0, vec1, vec2, vec3);
1287 ADDS_SH4_SH(vec0, vec8, vec1, vec9, vec2, vec10, vec3, vec11, out0,
1562 v8i16 hz_out7, hz_out8, hz_out9, res0, res1, vec0, vec1, vec2, vec3, vec4;
1593 ILVEV_B2_SH(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
1607 res0 = FILT_8TAP_DPADD_S_H(vec0, vec1, vec2, vec3, filt_vt0, filt_vt1,
1625 vec0 = vec2;
1800 v16u8 filt0, vec0, vec1, res0, res1;
1810 VSHF_B2_UB(src0, src1, src2, src3, mask, mask, vec0, vec1);
1811 DOTP_UB2_UH(vec0, vec1, filt0, filt0, vec2, vec3);
1822 v16u8 vec0, vec1, vec2, vec3, filt0;
1834 VSHF_B2_UB(src0, src1, src2, src3, mask, mask, vec0, vec1);
1836 DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0,
1866 v8u16 vec0, vec1, vec2, vec3, filt;
1875 VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
1877 DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0,
1878 vec0, vec1, vec2, vec3);
1879 SRARI_H4_UH(vec0, vec1, vec2, vec3, 7);
1880 PCKEV_B2_SB(vec1, vec0, vec3, vec2, src0, src1);
1890 v8u16 vec0, vec1, vec2, vec3, filt;
1901 VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
1903 DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0,
1904 vec0, vec1, vec2, vec3);
1905 SRARI_H4_UH(vec0, vec1, vec2, vec3, 7);
1909 PCKEV_B2_SB(vec1, vec0, vec3, vec2, out0, out1);
1912 VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
1914 DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0,
1915 vec0, vec1, vec2, vec3);
1916 SRARI_H4_UH(vec0, vec1, vec2, vec3, 7);
1917 PCKEV_B2_SB(vec1, vec0, vec3, vec2, out0, out1);
1925 VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
1927 DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0,
1928 vec0, vec1, vec2, vec3);
1929 SRARI_H4_UH(vec0, vec1, vec2, vec3, 7);
1933 PCKEV_B2_SB(vec1, vec0, vec3, vec2, out0, out1);
1936 VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
1938 DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0,
1939 vec0, vec1, vec2, vec3);
1940 SRARI_H4_UH(vec0, vec1, vec2, vec3, 7);
1941 PCKEV_B2_SB(vec1, vec0, vec3, vec2, out0, out1);
1967 v16u8 filt0, vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
1982 VSHF_B2_UB(src0, src0, src1, src1, mask, mask, vec0, vec1);
1986 DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0,
2006 VSHF_B2_UB(src0, src0, src1, src1, mask, mask, vec0, vec1);
2010 DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0,
2034 v16u8 filt0, vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
2055 VSHF_B2_UB(src0, src0, src1, src1, mask, mask, vec0, vec1);
2059 DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0,
2081 v16u8 filt0, vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
2099 VSHF_B2_UB(src0, src0, src1, src1, mask, mask, vec0, vec1);
2103 DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0,
2194 v16u8 src0, src1, src2, src3, src4, vec0, vec1, vec2, vec3, filt0;
2204 ILVR_B2_UB(src1, src0, src2, src1, vec0, vec1);
2206 DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0,
2220 v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, filt0;
2237 vec0, vec1, vec2, vec3);
2240 DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0,
2280 v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, filt0;
2295 ILVR_B2_UB(src1, src0, src2, src1, vec0, vec2);
2297 DOTP_UB2_UH(vec0, vec1, filt0, filt0, tmp0, tmp1);
2334 v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, filt0;
2348 ILVR_B2_UB(src1, src0, src2, src1, vec0, vec2);
2354 DOTP_UB2_UH(vec0, vec1, filt0, filt0, tmp0, tmp1);
2375 ILVR_B2_UB(src6, src5, src7, src6, vec0, vec2);
2377 DOTP_UB2_UH(vec0, vec1, filt0, filt0, tmp0, tmp1);
2412 v16u8 src11, vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, filt0;
2430 ILVR_B2_UB(src1, src0, src2, src1, vec0, vec2);
2432 DOTP_UB2_UH(vec0, vec1, filt0, filt0, tmp0, tmp1);
2454 ILVR_B2_UB(src7, src6, src8, src7, vec0, vec2);
2456 DOTP_UB2_UH(vec0, vec1, filt0, filt0, tmp0, tmp1);
2491 v16u8 filt_vt, filt_hz, vec0, vec1, res0, res1;
2510 ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
2511 DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1);
2525 v16u8 filt_hz, filt_vt, vec0, vec1, vec2, vec3;
2551 ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
2553 DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt_vt, filt_vt, filt_vt, filt_vt,
2586 v16u8 filt_hz, filt_vt, vec0, vec1, vec2, vec3;
2603 vec0 = (v16u8) __msa_ilvev_b((v16i8) hz_out1, (v16i8) hz_out0);
2604 tmp0 = __msa_dotp_u_h(vec0, filt_vt);
2631 v16u8 filt_hz, filt_vt, vec0;
2654 vec0 = (v16u8) __msa_ilvev_b((v16i8) hz_out1, (v16i8) hz_out0);
2655 tmp1 = __msa_dotp_u_h(vec0, filt_vt);
2658 vec0 = (v16u8) __msa_ilvev_b((v16i8) hz_out0, (v16i8) hz_out1);
2659 tmp2 = __msa_dotp_u_h(vec0, filt_vt);
2665 vec0 = (v16u8) __msa_ilvev_b((v16i8) hz_out1, (v16i8) hz_out0);
2666 tmp3 = __msa_dotp_u_h(vec0, filt_vt);
2671 vec0 = (v16u8) __msa_ilvev_b((v16i8) hz_out0, (v16i8) hz_out1);
2672 tmp4 = __msa_dotp_u_h(vec0, filt_vt);
2680 vec0 = (v16u8) __msa_ilvev_b((v16i8) hz_out1, (v16i8) hz_out0);
2681 tmp5 = __msa_dotp_u_h(vec0, filt_vt);
2684 vec0 = (v16u8) __msa_ilvev_b((v16i8) hz_out0, (v16i8) hz_out1);
2685 tmp6 = __msa_dotp_u_h(vec0, filt_vt);
2688 vec0 = (v16u8) __msa_ilvev_b((v16i8) hz_out1, (v16i8) hz_out0);
2689 tmp7 = __msa_dotp_u_h(vec0, filt_vt);
2692 vec0 = (v16u8) __msa_ilvev_b((v16i8) hz_out0, (v16i8) hz_out1);
2693 tmp8 = __msa_dotp_u_h(vec0, filt_vt);
2727 v16u8 filt_hz, filt_vt, vec0, vec1;
2754 ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
2755 DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp1, tmp2);
2763 ILVEV_B2_UB(hz_out1, hz_out0, hz_out3, hz_out2, vec0, vec1);
2764 DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp1, tmp2);
2772 ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
2773 DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp1, tmp2);
2781 ILVEV_B2_UB(hz_out1, hz_out0, hz_out3, hz_out2, vec0, vec1);
2782 DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp1, tmp2);
2825 v16u8 filt0, dst0, vec0, vec1, res;
2837 VSHF_B2_UB(src0, src1, src2, src3, mask, mask, vec0, vec1);
2838 DOTP_UB2_UH(vec0, vec1, filt0, filt0, vec2, vec3);
2854 v16u8 filt0, vec0, vec1, vec2, vec3, res0, res1, res2, res3;
2869 VSHF_B2_UB(src0, src1, src2, src3, mask, mask, vec0, vec1);
2871 DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec4, vec5,
2904 v8u16 vec0, vec1, vec2, vec3, filt;
2913 VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
2915 DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0,
2916 vec0, vec1, vec2, vec3);
2917 SRARI_H4_UH(vec0, vec1, vec2, vec3, 7);
2921 PCKEV_AVG_ST8x4_UB(vec0, vec1, vec2, vec3, dst0, dst1, dst, dst_stride);
2934 v8u16 vec0, vec1, vec2, vec3, filt;
2944 VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
2946 DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec0, vec1,
2948 SRARI_H4_UH(vec0, vec1, vec2, vec3, 7);
2954 PCKEV_AVG_ST8x4_UB(vec0, vec1, vec2, vec3, dst0, dst1, dst, dst_stride);
2957 VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
2959 DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec0, vec1,
2961 SRARI_H4_UH(vec0, vec1, vec2, vec3, 7);
2965 PCKEV_AVG_ST8x4_UB(vec0, vec1, vec2, vec3, dst0, dst1, dst, dst_stride);
2972 VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
2974 DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec0,
2976 SRARI_H4_UH(vec0, vec1, vec2, vec3, 7);
2981 PCKEV_AVG_ST8x4_UB(vec0, vec1, vec2, vec3, dst0, dst1, dst, dst_stride);
2984 VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
2986 DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec0,
2988 SRARI_H4_UH(vec0, vec1, vec2, vec3, 7);
2992 PCKEV_AVG_ST8x4_UB(vec0, vec1, vec2, vec3, dst0, dst1, dst, dst_stride);
3019 v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
3032 VSHF_B2_UB(src0, src0, src1, src1, mask, mask, vec0, vec1);
3036 DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, res0, res1,
3057 VSHF_B2_UB(src0, src0, src1, src1, mask, mask, vec0, vec1);
3061 DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, res0,
3087 v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
3108 VSHF_B2_UB(src0, src0, src1, src1, mask, mask, vec0, vec1);
3112 DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0,
3137 v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
3152 VSHF_B2_UB(src0, src0, src1, src1, mask, mask, vec0, vec1);
3156 DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0,
3269 v16u8 dst0, dst1, vec0, vec1, vec2, vec3, filt0;
3281 ILVR_B2_UB(src1, src0, src2, src1, vec0, vec1);
3283 DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0,
3301 v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, filt0;
3324 vec0, vec1, vec2, vec3);
3327 DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0,
3367 v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
3382 ILVR_B2_UB(src1, src0, src2, src1, vec0, vec2);
3384 DOTP_UB2_UH(vec0, vec1, filt0, filt0, tmp0, tmp1);
3422 v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, filt0;
3435 ILVR_B2_UB(src1, src0, src2, src1, vec0, vec2);
3442 DOTP_UB2_UH(vec0, vec1, filt0, filt0, tmp0, tmp1);
3464 ILVR_B2_UB(src6, src5, src7, src6, vec0, vec2);
3466 DOTP_UB2_UH(vec0, vec1, filt0, filt0, tmp0, tmp1);
3503 v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
3525 ILVR_B2_UB(src1, src0, src2, src1, vec0, vec2);
3527 DOTP_UB2_UH(vec0, vec1, filt0, filt0, tmp0, tmp1);
3549 ILVR_B2_UB(src7, src6, src8, src7, vec0, vec2);
3551 DOTP_UB2_UH(vec0, vec1, filt0, filt0, tmp0, tmp1);
3590 v16u8 filt_hz, filt_vt, vec0, vec1;
3610 ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
3615 DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1);
3634 v16u8 filt_hz, filt_vt, vec0, vec1, vec2, vec3, res0, res1;
3666 ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
3668 DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt_vt, filt_vt, filt_vt, filt_vt,
3702 v16u8 filt_hz, filt_vt, dst0, dst1, vec0, vec1, vec2, vec3;
3723 vec0 = (v16u8) __msa_ilvev_b((v16i8) hz_out1, (v16i8) hz_out0);
3724 tmp0 = __msa_dotp_u_h(vec0, filt_vt);
3754 v16u8 filt_hz, filt_vt, vec0, dst0, dst1;
3777 vec0 = (v16u8) __msa_ilvev_b((v16i8) hz_out1, (v16i8) hz_out0);
3778 tmp0 = __msa_dotp_u_h(vec0, filt_vt);
3781 vec0 = (v16u8) __msa_ilvev_b((v16i8) hz_out0, (v16i8) hz_out1);
3782 tmp1 = __msa_dotp_u_h(vec0, filt_vt);
3788 vec0 = (v16u8) __msa_ilvev_b((v16i8) hz_out1, (v16i8) hz_out0);
3789 tmp2 = __msa_dotp_u_h(vec0, filt_vt);
3792 vec0 = (v16u8) __msa_ilvev_b((v16i8) hz_out0, (v16i8) hz_out1);
3793 tmp3 = __msa_dotp_u_h(vec0, filt_vt);
3831 v16u8 filt_hz, filt_vt, vec0, vec1, dst0, dst1, dst2, dst3;
3858 ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
3859 DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1);
3867 ILVEV_B2_UB(hz_out1, hz_out0, hz_out3, hz_out2, vec0, vec1);
3868 DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1);
3876 ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
3877 DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1);
3885 ILVEV_B2_UB(hz_out1, hz_out0, hz_out3, hz_out2, vec0, vec1);
3886 DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1);