Lines Matching refs:src3
83 #define HORIZ_8TAP_4WID_4VECS_FILT(src0, src1, src2, src3, \
91 VSHF_B2_SB(src0, src1, src2, src3, mask0, mask0, vec0_m, vec1_m); \
93 VSHF_B2_SB(src0, src1, src2, src3, mask1, mask1, vec2_m, vec3_m); \
95 VSHF_B2_SB(src0, src1, src2, src3, mask2, mask2, vec4_m, vec5_m); \
97 VSHF_B2_SB(src0, src1, src2, src3, mask3, mask3, vec6_m, vec7_m); \
102 #define HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, \
111 VSHF_B2_SB(src2, src2, src3, src3, mask0, mask0, vec2_m, vec3_m); \
115 VSHF_B2_SB(src2, src2, src3, src3, mask2, mask2, vec2_m, vec3_m); \
119 VSHF_B2_SB(src2, src2, src3, src3, mask1, mask1, vec6_m, vec7_m); \
123 VSHF_B2_SB(src2, src2, src3, src3, mask3, mask3, vec6_m, vec7_m); \
164 v16i8 src0, src1, src2, src3, filt0, filt1, filt2, filt3;
178 LD_SB4(src, src_stride, src0, src1, src2, src3);
179 XORI_B4_128_SB(src0, src1, src2, src3);
180 HORIZ_8TAP_4WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2,
193 v16i8 src0, src1, src2, src3;
208 LD_SB4(src, src_stride, src0, src1, src2, src3);
209 XORI_B4_128_SB(src0, src1, src2, src3);
211 HORIZ_8TAP_4WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2,
213 LD_SB4(src, src_stride, src0, src1, src2, src3);
214 XORI_B4_128_SB(src0, src1, src2, src3);
215 HORIZ_8TAP_4WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2,
240 v16i8 src0, src1, src2, src3, filt0, filt1, filt2, filt3;
255 LD_SB4(src, src_stride, src0, src1, src2, src3);
256 XORI_B4_128_SB(src0, src1, src2, src3);
257 HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2,
272 v16i8 src0, src1, src2, src3, filt0, filt1, filt2, filt3;
288 LD_SB4(src, src_stride, src0, src1, src2, src3);
289 XORI_B4_128_SB(src0, src1, src2, src3);
291 HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2,
320 v16i8 src0, src1, src2, src3, filt0, filt1, filt2, filt3;
337 LD_SB2(src + 8, src_stride, src1, src3);
338 XORI_B4_128_SB(src0, src1, src2, src3);
340 HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2,
359 v16i8 src0, src1, src2, src3, filt0, filt1, filt2, filt3;
377 src3 = LD_SB(src + 24);
380 XORI_B4_128_SB(src0, src1, src2, src3);
381 HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2,
389 src3 = LD_SB(src + 24);
399 XORI_B4_128_SB(src0, src1, src2, src3);
400 HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2,
418 v16i8 src0, src1, src2, src3, filt0, filt1, filt2, filt3;
436 src3 = LD_SB(src + 24);
439 XORI_B4_128_SB(src0, src1, src2, src3);
440 HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1,
452 src3 = LD_SB(src + 56);
456 XORI_B4_128_SB(src0, src1, src2, src3);
457 HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1,
475 v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
487 LD_SB7(src, src_stride, src0, src1, src2, src3, src4, src5, src6);
490 ILVR_B4_SB(src1, src0, src3, src2, src5, src4, src2, src1, src10_r, src32_r,
492 ILVR_B2_SB(src4, src3, src6, src5, src43_r, src65_r);
527 v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
538 LD_SB7(src, src_stride, src0, src1, src2, src3, src4, src5, src6);
539 XORI_B7_128_SB(src0, src1, src2, src3, src4, src5, src6);
541 ILVR_B4_SB(src1, src0, src3, src2, src5, src4, src2, src1, src10_r, src32_r,
543 ILVR_B2_SB(src4, src3, src6, src5, src43_r, src65_r);
582 v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
595 LD_SB7(src, src_stride, src0, src1, src2, src3, src4, src5, src6);
596 XORI_B7_128_SB(src0, src1, src2, src3, src4, src5, src6);
598 ILVR_B4_SB(src1, src0, src3, src2, src5, src4, src2, src1, src10_r, src32_r,
600 ILVR_B2_SB(src4, src3, src6, src5, src43_r, src65_r);
601 ILVL_B4_SB(src1, src0, src3, src2, src5, src4, src2, src1, src10_l, src32_l,
603 ILVL_B2_SB(src4, src3, src6, src5, src43_l, src65_l);
664 v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
681 LD_SB7(src_tmp, src_stride, src0, src1, src2, src3, src4, src5, src6);
682 XORI_B7_128_SB(src0, src1, src2, src3, src4, src5, src6);
684 ILVR_B4_SB(src1, src0, src3, src2, src5, src4, src2, src1, src10_r,
686 ILVR_B2_SB(src4, src3, src6, src5, src43_r, src65_r);
687 ILVL_B4_SB(src1, src0, src3, src2, src5, src4, src2, src1, src10_l,
689 ILVL_B2_SB(src4, src3, src6, src5, src43_l, src65_l);
768 v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
786 LD_SB7(src, src_stride, src0, src1, src2, src3, src4, src5, src6);
787 XORI_B7_128_SB(src0, src1, src2, src3, src4, src5, src6);
792 hz_out2 = HORIZ_8TAP_FILT(src2, src3, mask0, mask1, mask2, mask3, filt_hz0,
844 v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
863 LD_SB7(src, src_stride, src0, src1, src2, src3, src4, src5, src6);
866 XORI_B7_128_SB(src0, src1, src2, src3, src4, src5, src6);
873 hz_out3 = HORIZ_8TAP_FILT(src3, src3, mask0, mask1, mask2, mask3, filt_hz0,
992 v16i8 src0, src1, src2, src3, filt0, filt1, filt2, filt3;
1008 LD_SB4(src, src_stride, src0, src1, src2, src3);
1009 XORI_B4_128_SB(src0, src1, src2, src3);
1010 HORIZ_8TAP_4WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2,
1027 v16i8 src0, src1, src2, src3, filt0, filt1, filt2, filt3;
1043 LD_SB4(src, src_stride, src0, src1, src2, src3);
1044 XORI_B4_128_SB(src0, src1, src2, src3);
1050 HORIZ_8TAP_4WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2,
1052 LD_SB4(src, src_stride, src0, src1, src2, src3);
1053 XORI_B4_128_SB(src0, src1, src2, src3);
1054 HORIZ_8TAP_4WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2,
1089 v16i8 src0, src1, src2, src3, filt0, filt1, filt2, filt3;
1105 LD_SB4(src, src_stride, src0, src1, src2, src3);
1106 XORI_B4_128_SB(src0, src1, src2, src3);
1108 HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2,
1129 v16i8 src0, src1, src2, src3, filt0, filt1, filt2, filt3;
1148 LD_SB2(src + 8, src_stride, src1, src3);
1151 XORI_B4_128_SB(src0, src1, src2, src3);
1158 VSHF_B4_SH(src3, src3, mask0, mask1, mask2, mask3, vec3, vec7, vec11,
1187 v16i8 src0, src1, src2, src3, filt0, filt1, filt2, filt3;
1207 src3 = LD_SB(src + 24);
1211 XORI_B4_128_SB(src0, src1, src2, src3);
1218 VSHF_B4_SH(src3, src3, mask0, mask1, mask2, mask3, vec3, vec7, vec11,
1246 v16i8 src0, src1, src2, src3, filt0, filt1, filt2, filt3;
1267 src3 = LD_SB(&src[24 + (cnt << 5)]);
1270 XORI_B4_128_SB(src0, src1, src2, src3);
1277 VSHF_B4_SH(src3, src3, mask0, mask1, mask2, mask3, vec3, vec7,
1309 v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
1321 LD_SB7(src, src_stride, src0, src1, src2, src3, src4, src5, src6);
1324 ILVR_B4_SB(src1, src0, src3, src2, src5, src4, src2, src1, src10_r, src32_r,
1326 ILVR_B2_SB(src4, src3, src6, src5, src43_r, src65_r);
1368 v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
1379 LD_SB7(src, src_stride, src0, src1, src2, src3, src4, src5, src6);
1382 XORI_B7_128_SB(src0, src1, src2, src3, src4, src5, src6);
1383 ILVR_B4_SB(src1, src0, src3, src2, src5, src4, src2, src1, src10_r, src32_r,
1385 ILVR_B2_SB(src4, src3, src6, src5, src43_r, src65_r);
1432 v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
1449 LD_SB7(src_tmp, src_stride, src0, src1, src2, src3, src4, src5, src6);
1450 XORI_B7_128_SB(src0, src1, src2, src3, src4, src5, src6);
1453 ILVR_B4_SB(src1, src0, src3, src2, src5, src4, src2, src1, src10_r,
1455 ILVR_B2_SB(src4, src3, src6, src5, src43_r, src65_r);
1456 ILVL_B4_SB(src1, src0, src3, src2, src5, src4, src2, src1, src10_l,
1458 ILVL_B2_SB(src4, src3, src6, src5, src43_l, src65_l);
1558 v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
1576 LD_SB7(src, src_stride, src0, src1, src2, src3, src4, src5, src6);
1577 XORI_B7_128_SB(src0, src1, src2, src3, src4, src5, src6);
1582 hz_out2 = HORIZ_8TAP_FILT(src2, src3, mask0, mask1, mask2, mask3, filt_hz0,
1641 v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
1660 LD_SB7(src, src_stride, src0, src1, src2, src3, src4, src5, src6);
1663 XORI_B7_128_SB(src0, src1, src2, src3, src4, src5, src6);
1670 hz_out3 = HORIZ_8TAP_FILT(src3, src3, mask0, mask1, mask2, mask3, filt_hz0,
1799 v16i8 src0, src1, src2, src3, mask;
1809 LD_SB4(src, src_stride, src0, src1, src2, src3);
1810 VSHF_B2_UB(src0, src1, src2, src3, mask, mask, vec0, vec1);
1823 v16i8 src0, src1, src2, src3, src4, src5, src6, src7, mask;
1833 LD_SB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
1834 VSHF_B2_UB(src0, src1, src2, src3, mask, mask, vec0, vec1);
1865 v16i8 src0, src1, src2, src3, mask;
1874 LD_SB4(src, src_stride, src0, src1, src2, src3);
1876 VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
1889 v16i8 src0, src1, src2, src3, mask, out0, out1;
1898 LD_SB4(src, src_stride, src0, src1, src2, src3);
1902 VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
1906 LD_SB4(src, src_stride, src0, src1, src2, src3);
1913 VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
1922 LD_SB4(src, src_stride, src0, src1, src2, src3);
1926 VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
1930 LD_SB4(src, src_stride, src0, src1, src2, src3);
1937 VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
1966 v16i8 src0, src1, src2, src3, src4, src5, src6, src7, mask;
1979 LD_SB4(src + 8, src_stride, src1, src3, src5, src7);
1983 VSHF_B2_UB(src2, src2, src3, src3, mask, mask, vec2, vec3);
2003 LD_SB4(src + 8, src_stride, src1, src3, src5, src7);
2007 VSHF_B2_UB(src2, src2, src3, src3, mask, mask, vec2, vec3);
2033 v16i8 src0, src1, src2, src3, src4, src5, src6, src7, mask;
2046 src3 = LD_SB(src + 24);
2056 VSHF_B2_UB(src2, src2, src3, src3, mask, mask, vec2, vec3);
2080 v16i8 src0, src1, src2, src3, src4, src5, src6, src7, mask;
2096 SLDI_B3_SB(src2, src0, src4, src2, src6, src4, 8, src1, src3, src5);
2100 VSHF_B2_UB(src2, src2, src3, src3, mask, mask, vec2, vec3);
2121 v16i8 src0, src1, src2, src3, src4;
2130 LD_SB5(src, src_stride, src0, src1, src2, src3, src4);
2133 ILVR_B4_SB(src1, src0, src2, src1, src3, src2, src4, src3,
2147 v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8;
2157 LD_SB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
2163 ILVR_B4_SB(src1, src0, src2, src1, src3, src2, src4, src3, src10_r, src21_r,
2194 v16u8 src0, src1, src2, src3, src4, vec0, vec1, vec2, vec3, filt0;
2203 LD_UB5(src, src_stride, src0, src1, src2, src3, src4);
2205 ILVR_B2_UB(src3, src2, src4, src3, vec2, vec3);
2219 v16u8 src0, src1, src2, src3, src4, src5, src6, src7, src8;
2233 LD_UB8(src, src_stride, src1, src2, src3, src4, src5, src6, src7, src8);
2236 ILVR_B4_UB(src1, src0, src2, src1, src3, src2, src4, src3,
2279 v16u8 src0, src1, src2, src3, src4;
2292 LD_UB4(src, src_stride, src1, src2, src3, src4);
2303 ILVR_B2_UB(src3, src2, src4, src3, vec4, vec6);
2304 ILVL_B2_UB(src3, src2, src4, src3, vec5, vec7);
2333 v16u8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9;
2347 LD_UB4(src, src_stride, src1, src2, src3, src4);
2363 ILVR_B2_UB(src3, src2, src4, src3, vec4, vec6);
2364 ILVL_B2_UB(src3, src2, src4, src3, vec5, vec7);
2411 v16u8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
2420 LD_UB4(src, 16, src0, src3, src6, src9);
2442 ILVR_B2_UB(src4, src3, src5, src4, vec4, vec6);
2443 ILVL_B2_UB(src4, src3, src5, src4, vec5, vec7);
2480 src3 = src5;
2490 v16i8 src0, src1, src2, src3, src4, mask;
2503 LD_SB5(src, src_stride, src0, src1, src2, src3, src4);
2505 hz_out2 = HORIZ_2TAP_FILT_UH(src2, src3, mask, filt_hz, 7);
2523 v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, mask;
2538 LD_SB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
2543 hz_out2 = HORIZ_2TAP_FILT_UH(src2, src3, mask, filt_hz, 7);
2585 v16i8 src0, src1, src2, src3, src4, mask, out0, out1;
2599 LD_SB5(src, src_stride, src0, src1, src2, src3, src4);
2610 hz_out1 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz, 7);
2630 v16i8 src0, src1, src2, src3, src4, mask, out0, out1;
2650 LD_SB4(src, src_stride, src1, src2, src3, src4);
2664 hz_out1 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz, 7);
2669 LD_SB4(src, src_stride, src1, src2, src3, src4);
2687 hz_out1 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz, 7);
2726 v16i8 src0, src1, src2, src3, src4, src5, src6, src7, mask;
2749 LD_SB4(src + 8, src_stride, src1, src3, src5, src7);
2762 hz_out2 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz, 7);
2824 v16i8 src0, src1, src2, src3, mask;
2834 LD_SB4(src, src_stride, src0, src1, src2, src3);
2837 VSHF_B2_UB(src0, src1, src2, src3, mask, mask, vec0, vec1);
2853 v16i8 src0, src1, src2, src3, src4, src5, src6, src7, mask;
2864 LD_SB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
2869 VSHF_B2_UB(src0, src1, src2, src3, mask, mask, vec0, vec1);
2902 v16i8 src0, src1, src2, src3, mask;
2912 LD_SB4(src, src_stride, src0, src1, src2, src3);
2914 VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
2932 v16i8 src0, src1, src2, src3, mask;
2942 LD_SB4(src, src_stride, src0, src1, src2, src3);
2945 VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
2952 LD_SB4(src, src_stride, src0, src1, src2, src3);
2958 VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
2969 LD_SB4(src, src_stride, src0, src1, src2, src3);
2973 VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
2980 LD_SB4(src, src_stride, src0, src1, src2, src3);
2985 VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
3017 v16i8 src0, src1, src2, src3, src4, src5, src6, src7, mask;
3029 LD_SB4(src + 8, src_stride, src1, src3, src5, src7);
3033 VSHF_B2_UB(src2, src2, src3, src3, mask, mask, vec2, vec3);
3054 LD_SB4(src + 8, src_stride, src1, src3, src5, src7);
3058 VSHF_B2_UB(src2, src2, src3, src3, mask, mask, vec2, vec3);
3085 v16i8 src0, src1, src2, src3, src4, src5, src6, src7, mask;
3099 src3 = LD_SB(src + 24);
3109 VSHF_B2_UB(src2, src2, src3, src3, mask, mask, vec2, vec3);
3135 v16i8 src0, src1, src2, src3, src4, src5, src6, src7, mask;
3149 SLDI_B3_SB(src2, src0, src4, src2, src6, src4, 8, src1, src3, src5);
3153 VSHF_B2_UB(src2, src2, src3, src3, mask, mask, vec2, vec3);
3177 v16i8 src0, src1, src2, src3, src4;
3186 LD_SB4(src, src_stride, src0, src1, src2, src3);
3194 ILVR_B4_SB(src1, src0, src2, src1, src3, src2, src4, src3,
3214 v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src87_r;
3223 LD_SB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
3231 ILVR_B4_SB(src1, src0, src2, src1, src3, src2, src4, src3, src10_r, src21_r,
3268 v16u8 src0, src1, src2, src3, src4;
3277 LD_UB5(src, src_stride, src0, src1, src2, src3, src4);
3282 ILVR_B2_UB(src3, src2, src4, src3, vec2, vec3);
3299 v16u8 src0, src1, src2, src3, src4, src5, src6, src7, src8;
3313 LD_UB8(src, src_stride, src1, src2, src3, src4, src5, src6, src7, src8);
3323 ILVR_B4_UB(src1, src0, src2, src1, src3, src2, src4, src3,
3366 v16u8 src0, src1, src2, src3, src4, dst0, dst1, dst2, dst3, filt0;
3378 LD_UB4(src, src_stride, src1, src2, src3, src4);
3390 ILVR_B2_UB(src3, src2, src4, src3, vec4, vec6);
3391 ILVL_B2_UB(src3, src2, src4, src3, vec5, vec7);
3420 v16u8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9;
3433 LD_UB4(src, src_stride, src1, src2, src3, src4);
3452 ILVR_B2_UB(src3, src2, src4, src3, vec4, vec6);
3453 ILVL_B2_UB(src3, src2, src4, src3, vec5, vec7);
3500 v16u8 src0, src1, src2, src3, src4, src5;
3511 LD_UB4(src, 16, src0, src3, src6, src9);
3537 ILVR_B2_UB(src4, src3, src5, src4, vec4, vec6);
3538 ILVL_B2_UB(src4, src3, src5, src4, vec5, vec7);
3575 src3 = src5;
3589 v16i8 src0, src1, src2, src3, src4, mask;
3603 LD_SB5(src, src_stride, src0, src1, src2, src3, src4);
3606 hz_out2 = HORIZ_2TAP_FILT_UH(src2, src3, mask, filt_hz, 7);
3633 v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, mask;
3649 LD_SB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
3654 hz_out2 = HORIZ_2TAP_FILT_UH(src2, src3, mask, filt_hz, 7);
3701 v16i8 src0, src1, src2, src3, src4, mask;
3715 LD_SB5(src, src_stride, src0, src1, src2, src3, src4);
3730 hz_out1 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz, 7);
3753 v16i8 src0, src1, src2, src3, src4, mask;
3773 LD_SB4(src, src_stride, src1, src2, src3, src4);
3787 hz_out1 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz, 7);
3830 v16i8 src0, src1, src2, src3, src4, src5, src6, src7, mask;
3852 LD_SB4(src + 8, src_stride, src1, src3, src5, src7);
3866 hz_out2 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz, 7);
3957 v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
3960 LD_UB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
3961 ST_UB8(src0, src1, src2, src3, src4, src5, src6, src7, dst, dst_stride);
3963 LD_UB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
3965 ST_UB8(src0, src1, src2, src3, src4, src5, src6, src7, dst, dst_stride);
3967 LD_UB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
3969 ST_UB8(src0, src1, src2, src3, src4, src5, src6, src7, dst, dst_stride);
3972 LD_UB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
3974 ST_UB8(src0, src1, src2, src3, src4, src5, src6, src7, dst, dst_stride);
3976 LD_UB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
3978 ST_UB8(src0, src1, src2, src3, src4, src5, src6, src7, dst, dst_stride);
3980 LD_UB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
3982 ST_UB8(src0, src1, src2, src3, src4, src5, src6, src7, dst, dst_stride);
3984 LD_UB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
3985 ST_UB8(src0, src1, src2, src3, src4, src5, src6, src7, dst, dst_stride);
3988 LD_UB4(src, src_stride, src0, src1, src2, src3);
3990 ST_UB4(src0, src1, src2, src3, dst, dst_stride);
4001 v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
4005 LD_UB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
4006 ST_UB8(src0, src1, src2, src3, src4, src5, src6, src7, dst, dst_stride);
4007 LD_UB8(src + 16, src_stride, src0, src1, src2, src3, src4, src5, src6,
4010 ST_UB8(src0, src1, src2, src3, src4, src5, src6, src7, dst + 16,
4016 LD_UB4(src, src_stride, src0, src1, src2, src3);
4019 ST_UB4(src0, src1, src2, src3, dst, dst_stride);
4031 v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
4035 LD_UB4(src, 16, src0, src1, src2, src3);
4044 ST_UB4(src0, src1, src2, src3, dst, 16);
4090 v16u8 src0, src1, src2, src3;
4102 INSERT_D2_UB(tp6, tp7, src3);
4109 AVER_UB4_UB(src0, dst0, src1, dst1, src2, dst2, src3, dst3, dst0,
4131 v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
4136 LD_UB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
4140 AVER_UB4_UB(src0, dst0, src1, dst1, src2, dst2, src3, dst3,
4149 LD_UB4(src, src_stride, src0, src1, src2, src3);
4153 AVER_UB4_UB(src0, dst0, src1, dst1, src2, dst2, src3, dst3,
4167 v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
4175 LD_UB4(src + 16, src_stride, src1, src3, src5, src7);
4187 AVER_UB4_UB(src0, dst0, src1, dst1, src2, dst2, src3, dst3,
4206 LD_UB4(src + 16, src_stride, src1, src3, src5, src7);
4212 AVER_UB4_UB(src0, dst0, src1, dst1, src2, dst2, src3, dst3,
4230 v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
4236 LD_UB4(src, 16, src0, src1, src2, src3);
4254 AVER_UB4_UB(src0, dst0, src1, dst1, src2, dst2, src3, dst3,