Lines Matching defs:stride
94 uint8_t *dst, int32_t stride)
112 LD_SB5(src_y, stride, src_vt0, src_vt1, src_vt2, src_vt3, src_vt4);
113 src_y += (5 * stride);
122 LD_SB4(src_x, stride, src_hz0, src_hz1, src_hz2, src_hz3);
130 LD_SB4(src_y, stride, src_vt5, src_vt6, src_vt7, src_vt8);
152 ST_W4(out, 0, 1, 2, 3, dst, stride);
156 uint8_t *dst, int32_t stride)
176 LD_SB5(src_y, stride, src_vt0, src_vt1, src_vt2, src_vt3, src_vt4);
177 src_y += (5 * stride);
181 LD_SB4(src_x, stride, src_hz0, src_hz1, src_hz2, src_hz3);
183 src_x += (4 * stride);
193 LD_SB4(src_y, stride, src_vt5, src_vt6, src_vt7, src_vt8);
194 src_y += (4 * stride);
217 LD_SB4(src_x, stride, src_hz0, src_hz1, src_hz2, src_hz3);
223 ST_D4(out0, out1, 0, 1, 0, 1, dst, stride);
224 dst += (4 * stride);
226 LD_SB4(src_y, stride, src_vt9, src_vt10, src_vt11, src_vt12);
259 ST_D4(out0, out1, 0, 1, 0, 1, dst, stride);
264 int32_t stride)
293 LD_SB5(src_y, stride, src_vt0, src_vt1, src_vt2, src_vt3, src_vt4);
294 src_y += (5 * stride);
299 LD_SB4(src_x, stride, src_hz0, src_hz1, src_hz2, src_hz3);
301 src_x += (4 * stride);
310 LD_SB4(src_y, stride, src_vt5, src_vt6, src_vt7, src_vt8);
311 src_y += (4 * stride);
339 ST_D4(tmp0, tmp1, 0, 1, 0, 1, dst, stride);
340 dst += (4 * stride);
358 int32_t stride)
377 LD_SB5(src_y, stride, src_vt0, src_vt1, src_vt2, src_vt3, src_vt4);
378 src_y += (5 * stride);
387 LD_SB4(src_x, stride, src_hz0, src_hz1, src_hz2, src_hz3);
395 LD_SB4(src_y, stride, src_vt5, src_vt6, src_vt7, src_vt8);
411 LW4(dst, stride, tp0, tp1, tp2, tp3);
421 ST_W4(dst0, 0, 1, 2, 3, dst, stride);
427 int32_t stride)
448 LD_SB5(src_y, stride, src_vt0, src_vt1, src_vt2, src_vt3, src_vt4);
449 src_y += (5 * stride);
453 LD_SB4(src_x, stride, src_hz0, src_hz1, src_hz2, src_hz3);
455 src_x += (4 * stride);
465 LD_SB4(src_y, stride, src_vt5, src_vt6, src_vt7, src_vt8);
466 src_y += (4 * stride);
489 LD_SB4(src_x, stride, src_hz0, src_hz1, src_hz2, src_hz3);
492 LD4(dst, stride, tp0, tp1, tp2, tp3);
500 ST_D4(dst0, dst1, 0, 1, 0, 1, dst, stride);
501 dst += (4 * stride);
503 LD_SB4(src_y, stride, src_vt9, src_vt10, src_vt11, src_vt12);
533 LD4(dst, stride, tp0, tp1, tp2, tp3);
541 ST_D4(dst0, dst1, 0, 1, 0, 1, dst, stride);
547 int32_t stride)
577 LD_SB5(src_y, stride, src_vt0, src_vt1, src_vt2, src_vt3, src_vt4);
578 src_y += (5 * stride);
583 LD_SB4(src_x, stride, src_hz0, src_hz1, src_hz2, src_hz3);
585 src_x += (4 * stride);
594 LD_SB4(src_y, stride, src_vt5, src_vt6, src_vt7, src_vt8);
595 src_y += (4 * stride);
620 LD4(dst, stride, tp0, tp1, tp2, tp3);
628 ST_D4(dst0, dst1, 0, 1, 0, 1, dst, stride);
629 dst += (4 * stride);
645 ptrdiff_t stride)
650 LD_UB8(src, stride, src0, src1, src2, src3, src4, src5, src6, src7);
651 src += (8 * stride);
652 LD_UB8(src, stride, src8, src9, src10, src11, src12, src13, src14, src15);
654 ST_UB8(src0, src1, src2, src3, src4, src5, src6, src7, dst, stride);
655 dst += (8 * stride);
656 ST_UB8(src8, src9, src10, src11, src12, src13, src14, src15, dst, stride);
660 ptrdiff_t stride)
664 LD4(src, stride, src0, src1, src2, src3);
665 src += 4 * stride;
666 LD4(src, stride, src4, src5, src6, src7);
667 SD4(src0, src1, src2, src3, dst, stride);
668 dst += 4 * stride;
669 SD4(src4, src5, src6, src7, dst, stride);
673 ptrdiff_t stride)
678 LD_UB8(src, stride, src0, src1, src2, src3, src4, src5, src6, src7);
679 src += (8 * stride);
680 LD_UB8(dst, stride, dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7);
686 ST_UB8(dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7, dst, stride);
687 dst += (8 * stride);
689 LD_UB8(src, stride, src0, src1, src2, src3, src4, src5, src6, src7);
690 LD_UB8(dst, stride, dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7);
696 ST_UB8(dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7, dst, stride);
700 ptrdiff_t stride)
706 LD4(src, stride, tp0, tp1, tp2, tp3);
707 src += 4 * stride;
708 LD4(src, stride, tp4, tp5, tp6, tp7);
714 LD4(dst, stride, tp0, tp1, tp2, tp3);
715 LD4(dst + 4 * stride, stride, tp4, tp5, tp6, tp7);
724 ST_D8(dst0, dst1, dst2, dst3, 0, 1, 0, 1, 0, 1, 0, 1, dst, stride);
728 ptrdiff_t stride)
733 LW4(src, stride, tp0, tp1, tp2, tp3);
735 LW4(dst, stride, tp0, tp1, tp2, tp3);
740 ST_W4(dst0, 0, 1, 2, 3, dst, stride);
744 ptrdiff_t stride)
762 src += stride;
764 src += stride;
766 src += stride;
768 src += stride;
806 ST_SB4(dst0, dst1, dst2, dst3, dst, stride);
807 dst += (4 * stride);
812 ptrdiff_t stride)
830 src += stride;
832 src += stride;
834 src += stride;
836 src += stride;
874 ST_SB4(dst0, dst1, dst2, dst3, dst, stride);
875 dst += (4 * stride);
880 ptrdiff_t stride)
890 LD_SB8(src - 2, stride, src0, src1, src2, src3, src4, src5, src6, src7);
931 ST_D8(tmp0, tmp1, tmp2, tmp3, 0, 1, 0, 1, 0, 1, 0, 1, dst, stride);
935 ptrdiff_t stride)
945 LD_SB8(src - 2, stride, src0, src1, src2, src3, src4, src5, src6, src7);
986 ST_D8(tmp0, tmp1, tmp2, tmp3, 0, 1, 0, 1, 0, 1, 0, 1, dst, stride);
990 ptrdiff_t stride)
999 LD_SB4(src - 2, stride, src0, src1, src2, src3);
1017 ST_W4(res, 0, 1, 2, 3, dst, stride);
1021 ptrdiff_t stride)
1030 LD_SB4(src - 2, stride, src0, src1, src2, src3);
1048 ST_W4(res, 0, 1, 2, 3, dst, stride);
1052 ptrdiff_t stride)
1067 src += stride;
1069 src += stride;
1071 src += stride;
1073 src += stride;
1105 ST_SB4(vec0, vec1, vec2, vec3, dst, stride);
1106 dst += (4 * stride);
1111 ptrdiff_t stride)
1122 LD_SB8(src - 2, stride, src0, src1, src2, src3, src4, src5, src6, src7);
1154 ST_D8(out0, out1, out2, out3, 0, 1, 0, 1, 0, 1, 0, 1, dst, stride);
1158 ptrdiff_t stride)
1168 LD_SB4(src - 2, stride, src0, src1, src2, src3);
1179 ST_W4(out, 0, 1, 2, 3, dst, stride);
1183 ptrdiff_t stride)
1200 src -= (stride * 2);
1202 LD_SB5(src, stride, src0, src1, src2, src3, src4);
1203 src += (5 * stride);
1212 LD_SB4(src, stride, src5, src6, src7, src8);
1213 src += (4 * stride);
1239 ST_UB4(res0, res1, res2, res3, dst, stride);
1240 dst += (4 * stride);
1257 ptrdiff_t stride)
1274 src -= (stride * 2);
1276 LD_SB5(src, stride, src0, src1, src2, src3, src4);
1277 src += (5 * stride);
1286 LD_SB4(src, stride, src5, src6, src7, src8);
1287 src += (4 * stride);
1313 ST_UB4(res0, res1, res2, res3, dst, stride);
1314 dst += (4 * stride);
1330 ptrdiff_t stride)
1345 src -= (stride * 2);
1347 LD_SB5(src, stride, src0, src1, src2, src3, src4);
1348 src += (5 * stride);
1349 LD_SB8(src, stride, src5, src6, src7, src8, src9, src10, src11, src12);
1379 ST_D8(out0, out1, out2, out3, 0, 1, 0, 1, 0, 1, 0, 1, dst, stride);
1383 ptrdiff_t stride)
1398 src -= (stride * 2);
1400 LD_SB5(src, stride, src0, src1, src2, src3, src4);
1401 src += (5 * stride);
1402 LD_SB8(src, stride, src5, src6, src7, src8, src9, src10, src11, src12);
1432 ST_D8(out0, out1, out2, out3, 0, 1, 0, 1, 0, 1, 0, 1, dst, stride);
1436 ptrdiff_t stride)
1451 src -= (stride * 2);
1453 LD_SB5(src, stride, src0, src1, src2, src3, src4);
1454 src += (5 * stride);
1459 LD_SB4(src, stride, src5, src6, src7, src8);
1473 ST_W4(out, 0, 1, 2, 3, dst, stride);
1477 ptrdiff_t stride)
1492 src -= (stride * 2);
1494 LD_SB5(src, stride, src0, src1, src2, src3, src4);
1495 src += (5 * stride);
1500 LD_SB4(src, stride, src5, src6, src7, src8);
1514 ST_W4(out, 0, 1, 2, 3, dst, stride);
1518 ptrdiff_t stride)
1520 avc_luma_hv_qrt_16x16_msa(src - 2, src - (stride * 2), dst, stride);
1524 ptrdiff_t stride)
1526 avc_luma_hv_qrt_16x16_msa(src - 2, src - (stride * 2) + 1, dst, stride);
1530 ptrdiff_t stride)
1532 avc_luma_hv_qrt_16x16_msa(src + stride - 2, src - (stride * 2), dst,
1533 stride);
1537 ptrdiff_t stride)
1539 avc_luma_hv_qrt_16x16_msa(src + stride - 2, src - (stride * 2) + 1, dst,
1540 stride);
1544 ptrdiff_t stride)
1546 avc_luma_hv_qrt_8x8_msa(src - 2, src - (stride * 2), dst, stride);
1550 ptrdiff_t stride)
1552 avc_luma_hv_qrt_8x8_msa(src - 2, src - (stride * 2) + 1, dst, stride);
1556 ptrdiff_t stride)
1558 avc_luma_hv_qrt_8x8_msa(src + stride - 2, src - (stride * 2), dst, stride);
1562 ptrdiff_t stride)
1564 avc_luma_hv_qrt_8x8_msa(src + stride - 2, src - (stride * 2) + 1, dst,
1565 stride);
1570 ptrdiff_t stride)
1572 avc_luma_hv_qrt_4x4_msa(src - 2, src - (stride * 2), dst, stride);
1576 ptrdiff_t stride)
1578 avc_luma_hv_qrt_4x4_msa(src - 2, src - (stride * 2) + 1, dst, stride);
1582 ptrdiff_t stride)
1584 avc_luma_hv_qrt_4x4_msa(src + stride - 2, src - (stride * 2), dst, stride);
1588 ptrdiff_t stride)
1590 avc_luma_hv_qrt_4x4_msa(src + stride - 2, src - (stride * 2) + 1, dst,
1591 stride);
1595 ptrdiff_t stride)
1598 const uint8_t *src_tmp = src - (2 * stride) - 2;
1624 LD_SB5(src, stride, src0, src1, src2, src3, src4);
1626 src += (5 * stride);
1635 LD_SB4(src, stride, src5, src6, src7, src8);
1636 src += (4 * stride);
1692 ST_D4(out0, out1, 0, 1, 0, 1, dst, stride);
1693 dst += (4 * stride);
1708 ptrdiff_t stride)
1711 const uint8_t *src_tmp = src - (2 * stride) - 2;
1737 LD_SB5(src, stride, src0, src1, src2, src3, src4);
1739 src += (5 * stride);
1748 LD_SB4(src, stride, src5, src6, src7, src8);
1749 src += (4 * stride);
1805 ST_D4(out0, out1, 0, 1, 0, 1, dst, stride);
1806 dst += (4 * stride);
1821 ptrdiff_t stride)
1845 src -= ((2 * stride) + 2);
1847 LD_SB5(src, stride, src0, src1, src2, src3, src4);
1849 src += (5 * stride);
1857 LD_SB4(src, stride, src5, src6, src7, src8);
1858 src += (4 * stride);
1906 ST_D4(out0, out1, 0, 1, 0, 1, dst, stride);
1907 dst += (4 * stride);
1909 LD_SB4(src, stride, src9, src10, src11, src12);
1952 ST_D4(out0, out1, 0, 1, 0, 1, dst, stride);
1956 ptrdiff_t stride)
1980 src -= ((2 * stride) + 2);
1982 LD_SB5(src, stride, src0, src1, src2, src3, src4);
1984 src += (5 * stride);
1992 LD_SB4(src, stride, src5, src6, src7, src8);
1993 src += (4 * stride);
2041 ST_D4(out0, out1, 0, 1, 0, 1, dst, stride);
2042 dst += (4 * stride);
2044 LD_SB4(src, stride, src9, src10, src11, src12);
2087 ST_D4(out0, out1, 0, 1, 0, 1, dst, stride);
2091 ptrdiff_t stride)
2111 src -= ((2 * stride) + 2);
2113 LD_SB5(src, stride, src0, src1, src2, src3, src4);
2114 src += (5 * stride);
2115 LD_SB4(src, stride, src5, src6, src7, src8);
2151 ST_W4(res, 0, 1, 2, 3, dst, stride);
2155 ptrdiff_t stride)
2175 src -= ((2 * stride) + 2);
2177 LD_SB5(src, stride, src0, src1, src2, src3, src4);
2178 src += (5 * stride);
2179 LD_SB4(src, stride, src5, src6, src7, src8);
2216 ST_W4(res, 0, 1, 2, 3, dst, stride);
2220 ptrdiff_t stride)
2236 src -= (stride * 2);
2238 LD_SB5(src, stride, src0, src1, src2, src3, src4);
2239 src += (5 * stride);
2248 LD_SB4(src, stride, src5, src6, src7, src8);
2249 src += (4 * stride);
2271 ST_UB4(res0, res1, res2, res3, dst, stride);
2272 dst += (4 * stride);
2287 ptrdiff_t stride)
2303 src -= (stride * 2);
2305 LD_SB8(src, stride, src0, src1, src2, src3, src4, src5, src6, src7);
2306 src += (8 * stride);
2307 LD_SB5(src, stride, src8, src9, src10, src11, src12);
2333 ST_D8(out0, out1, out2, out3, 0, 1, 0, 1, 0, 1, 0, 1, dst, stride);
2337 ptrdiff_t stride)
2352 src -= (stride * 2);
2354 LD_SB5(src, stride, src0, src1, src2, src3, src4);
2355 src += (5 * stride);
2356 LD_SB4(src, stride, src5, src6, src7, src8);
2370 ST_W4(out, 0, 1, 2, 3, dst, stride);
2374 ptrdiff_t stride)
2394 src -= ((2 * stride) + 2);
2396 LD_SB5(src, stride, src0, src1, src2, src3, src4);
2397 LD_SB5(src + 8, stride, src7, src8, src9, src10, src11);
2398 src += (5 * stride);
2404 src += stride;
2440 dst += stride;
2456 ptrdiff_t stride)
2476 src -= ((2 * stride) + 2);
2478 LD_SB5(src, stride, src0, src1, src2, src3, src4);
2479 LD_SB5(src + 8, stride, src7, src8, src9, src10, src11);
2480 src += (5 * stride);
2486 src += stride;
2523 dst += stride;
2539 ptrdiff_t stride)
2559 src -= ((2 * stride) + 2);
2561 LD_SB5(src, stride, src0, src1, src2, src3, src4);
2562 src += (5 * stride);
2566 LD_SB2(src, stride, src5, src6);
2567 src += (2 * stride);
2602 ST_D2(out, 0, 1, dst, stride);
2603 dst += (2 * stride);
2614 ptrdiff_t stride)
2634 src -= ((2 * stride) + 2);
2636 LD_SB5(src, stride, src0, src1, src2, src3, src4);
2637 src += (5 * stride);
2641 LD_SB2(src, stride, src5, src6);
2642 src += (2 * stride);
2678 ST_D2(out, 0, 1, dst, stride);
2679 dst += (2 * stride);
2690 ptrdiff_t stride)
2714 src -= ((2 * stride) + 2);
2716 LD_SB5(src, stride, src0, src1, src2, src3, src4);
2717 src += (5 * stride);
2719 LD_SB4(src, stride, src5, src6, src7, src8);
2778 ST_W4(out, 0, 1, 2, 3, dst, stride);
2782 ptrdiff_t stride)
2806 src -= ((2 * stride) + 2);
2808 LD_SB5(src, stride, src0, src1, src2, src3, src4);
2809 src += (5 * stride);
2811 LD_SB4(src, stride, src5, src6, src7, src8);
2874 ST_W4(out, 0, 1, 2, 3, dst, stride);
2878 ptrdiff_t stride)
2883 const uint8_t *src_tmp = src - (2 * stride) - 2;
2906 LD_SB5(src, stride, src0, src1, src2, src3, src4);
2908 src += (5 * stride);
2917 LD_SB4(src, stride, src0, src1, src2, src3);
2919 src += (4 * stride);
2962 ST_D4(out0, out1, 0, 1, 0, 1, dst, stride);
2963 dst += (4 * stride);
2978 ptrdiff_t stride)
3001 src -= ((2 * stride) + 2);
3002 LD_SB5(src, stride, src0, src1, src2, src3, src4);
3004 src += (5 * stride);
3012 LD_SB4(src, stride, src0, src1, src2, src3);
3014 src += (4 * stride);
3050 ST_D4(out0, out1, 0, 1, 0, 1, dst, stride);
3051 dst += (4 * stride);
3053 LD_SB4(src, stride, src0, src1, src2, src3);
3087 ST_D4(out0, out1, 0, 1, 0, 1, dst, stride);
3091 ptrdiff_t stride)
3111 src -= ((2 * stride) + 2);
3113 LD_SB5(src, stride, src0, src1, src2, src3, src4);
3114 src += (5 * stride);
3115 LD_SB4(src, stride, src5, src6, src7, src8);
3142 ST_W4(res, 0, 1, 2, 3, dst, stride);
3146 ptrdiff_t stride)
3165 src += stride;
3167 src += stride;
3169 src += stride;
3171 src += stride;
3173 LD_UB4(dst, stride, dst0, dst1, dst2, dst3);
3212 ST_UB4(dst0, dst1, dst2, dst3, dst, stride);
3213 dst += (4 * stride);
3218 ptrdiff_t stride)
3237 src += stride;
3239 src += stride;
3241 src += stride;
3243 src += stride;
3245 LD_UB4(dst, stride, dst0, dst1, dst2, dst3);
3284 ST_UB4(dst0, dst1, dst2, dst3, dst, stride);
3285 dst += (4 * stride);
3290 ptrdiff_t stride)
3302 LD_SB8(src - 2, stride, src0, src1, src2, src3, src4, src5, src6, src7);
3343 LD4(dst, stride, tp0, tp1, tp2, tp3);
3346 LD4(dst + 4 * stride, stride, tp0, tp1, tp2, tp3);
3351 ST_D8(dst0, dst1, dst2, dst3, 0, 1, 0, 1, 0, 1, 0, 1, dst, stride);
3355 ptrdiff_t stride)
3367 LD_SB8(src - 2, stride, src0, src1, src2, src3, src4, src5, src6, src7);
3408 LD4(dst, stride, tp0, tp1, tp2, tp3);
3411 LD4(dst + 4 * stride, stride, tp0, tp1, tp2, tp3);
3416 ST_D8(dst0, dst1, dst2, dst3, 0, 1, 0, 1, 0, 1, 0, 1, dst, stride);
3420 ptrdiff_t stride)
3431 LD_SB4(src - 2, stride, src0, src1, src2, src3);
3449 LW4(dst, stride, tp0, tp1, tp2, tp3);
3452 ST_W4(dst0, 0, 1, 2, 3, dst, stride);
3456 ptrdiff_t stride)
3467 LD_SB4(src - 2, stride, src0, src1, src2, src3);
3485 LW4(dst, stride, tp0, tp1, tp2, tp3);
3488 ST_W4(dst0, 0, 1, 2, 3, dst, stride);
3492 ptrdiff_t stride)
3508 src += stride;
3510 src += stride;
3512 src += stride;
3514 src += stride;
3516 LD_UB4(dst, stride, dst0, dst1, dst2, dst3);
3549 ST_UB4(dst0, dst1, dst2, dst3, dst, stride);
3550 dst += (4 * stride);
3555 ptrdiff_t stride)
3569 LD_SB8(src - 2, stride, src0, src1, src2, src3, src4, src5, src6, src7);
3601 LD4(dst, stride, tp0, tp1, tp2, tp3);
3604 LD4(dst + 4 * stride, stride, tp0, tp1, tp2, tp3);
3609 ST_D8(out0, out1, out4, out5, 0, 1, 0, 1, 0, 1, 0, 1, dst, stride);
3613 ptrdiff_t stride)
3624 LD_SB4(src - 2, stride, src0, src1, src2, src3);
3635 LW4(dst, stride, tp0, tp1, tp2, tp3);
3638 ST_W4(res, 0, 1, 2, 3, dst, stride);
3642 ptrdiff_t stride)
3659 src -= (stride * 2);
3661 LD_SB5(src, stride, src0, src1, src2, src3, src4);
3662 src += (5 * stride);
3671 LD_SB4(src, stride, src5, src6, src7, src8);
3672 src += (4 * stride);
3697 LD_UB4(dst, stride, dst0, dst1, dst2, dst3);
3701 ST_UB4(dst0, dst1, dst2, dst3, dst, stride);
3702 dst += (4 * stride);
3719 ptrdiff_t stride)
3736 src -= (stride * 2);
3738 LD_SB5(src, stride, src0, src1, src2, src3, src4);
3739 src += (5 * stride);
3748 LD_SB4(src, stride, src5, src6, src7, src8);
3749 src += (4 * stride);
3774 LD_UB4(dst, stride, dst0, dst1, dst2, dst3);
3778 ST_UB4(dst0, dst1, dst2, dst3, dst, stride);
3779 dst += (4 * stride);
3795 ptrdiff_t stride)
3812 src -= (stride * 2);
3814 LD_SB5(src, stride, src0, src1, src2, src3, src4);
3815 src += (5 * stride);
3820 LD_SB8(src, stride, src7, src8, src9, src10, src11, src12, src13, src14);
3841 LD4(dst, stride, tp0, tp1, tp2, tp3);
3844 LD4(dst + 4 * stride, stride, tp0, tp1, tp2, tp3);
3857 ST_D8(dst0, dst1, dst2, dst3, 0, 1, 0, 1, 0, 1, 0, 1, dst, stride);
3861 ptrdiff_t stride)
3878 src -= (stride * 2);
3880 LD_SB5(src, stride, src0, src1, src2, src3, src4);
3881 src += (5 * stride);
3886 LD_SB8(src, stride, src7, src8, src9, src10, src11, src12, src13, src14);
3907 LD4(dst, stride, tp0, tp1, tp2, tp3);
3910 LD4(dst + 4 * stride, stride, tp0, tp1, tp2, tp3);
3923 ST_D8(dst0, dst1, dst2, dst3, 0, 1, 0, 1, 0, 1, 0, 1, dst, stride);
3927 ptrdiff_t stride)
3943 src -= (stride * 2);
3944 LD_SB5(src, stride, src0, src1, src2, src3, src4);
3945 src += (5 * stride);
3951 LD_SB4(src, stride, src5, src6, src7, src8);
3963 LW4(dst, stride, tp0, tp1, tp2, tp3);
3968 ST_W4(dst0, 0, 1, 2, 3, dst, stride);
3972 ptrdiff_t stride)
3988 src -= (stride * 2);
3990 LD_SB5(src, stride, src0, src1, src2, src3, src4);
3991 src += (5 * stride);
3997 LD_SB4(src, stride, src5, src6, src7, src8);
4006 LW4(dst, stride, tp0, tp1, tp2, tp3);
4014 ST_W4(dst0, 0, 1, 2, 3, dst, stride);
4018 ptrdiff_t stride)
4021 src - (stride * 2),
4022 dst, stride);
4026 ptrdiff_t stride)
4029 src - (stride * 2) +
4031 dst, stride);
4035 ptrdiff_t stride)
4037 avc_luma_hv_qrt_and_aver_dst_16x16_msa(src + stride - 2,
4038 src - (stride * 2),
4039 dst, stride);
4043 ptrdiff_t stride)
4045 avc_luma_hv_qrt_and_aver_dst_16x16_msa(src + stride - 2,
4046 src - (stride * 2) +
4048 dst, stride);
4052 ptrdiff_t stride)
4055 src - (stride * 2),
4056 dst, stride);
4060 ptrdiff_t stride)
4063 src - (stride * 2) +
4064 sizeof(uint8_t), dst, stride);
4068 ptrdiff_t stride)
4070 avc_luma_hv_qrt_and_aver_dst_8x8_msa(src + stride - 2,
4071 src - (stride * 2),
4072 dst, stride);
4076 ptrdiff_t stride)
4078 avc_luma_hv_qrt_and_aver_dst_8x8_msa(src + stride - 2,
4079 src - (stride * 2) +
4080 sizeof(uint8_t), dst, stride);
4085 ptrdiff_t stride)
4088 src - (stride * 2),
4089 dst, stride);
4093 ptrdiff_t stride)
4096 src - (stride * 2) +
4097 sizeof(uint8_t), dst, stride);
4101 ptrdiff_t stride)
4103 avc_luma_hv_qrt_and_aver_dst_4x4_msa(src + stride - 2,
4104 src - (stride * 2),
4105 dst, stride);
4109 ptrdiff_t stride)
4111 avc_luma_hv_qrt_and_aver_dst_4x4_msa(src + stride - 2,
4112 src - (stride * 2) +
4113 sizeof(uint8_t), dst, stride);
4117 ptrdiff_t stride)
4121 const uint8_t *src_tmp = src - (2 * stride) - 2;
4147 LD_SB5(src, stride, src0, src1, src2, src3, src4);
4149 src += (5 * stride);
4158 LD_SB2(src, stride, src5, src6);
4159 src += (2 * stride);
4192 LD2(dst, stride, tp0, tp1);
4197 ST_D2(dst0, 0, 1, dst, stride);
4198 dst += (2 * stride);
4200 LD_SB2(src, stride, src7, src8);
4201 src += (2 * stride);
4228 LD2(dst, stride, tp2, tp3);
4233 ST_D2(dst1, 0, 1, dst, stride);
4234 dst += (2 * stride);
4249 ptrdiff_t stride)
4253 const uint8_t *src_tmp = src - (2 * stride) - 2;
4279 LD_SB5(src, stride, src0, src1, src2, src3, src4);
4281 src += (5 * stride);
4290 LD_SB2(src, stride, src5, src6);
4291 src += (2 * stride);
4323 LD2(dst, stride, tp0, tp1);
4327 ST_D2(dst0, 0, 1, dst, stride);
4328 dst += (2 * stride);
4330 LD_SB2(src, stride, src7, src8);
4331 src += (2 * stride);
4358 LD2(dst, stride, tp2, tp3);
4362 ST_D2(dst1, 0, 1, dst, stride);
4363 dst += (2 * stride);
4378 ptrdiff_t stride)
4403 src -= ((2 * stride) + 2);
4405 LD_SB5(src, stride, src0, src1, src2, src3, src4);
4407 src += (5 * stride);
4415 LD_SB4(src, stride, src5, src6, src7, src8);
4416 src += (4 * stride);
4457 LD4(dst, stride, tp0, tp1, tp2, tp3);
4469 ST_D4(dst0, dst1, 0, 1, 0, 1, dst, stride);
4470 dst += (4 * stride);
4472 LD_SB4(src, stride, src9, src10, src11, src12);
4508 LD4(dst, stride, tp0, tp1, tp2, tp3);
4520 ST_D4(dst0, dst1, 0, 1, 0, 1, dst, stride);
4524 ptrdiff_t stride)
4549 src -= ((2 * stride) + 2);
4551 LD_SB5(src, stride, src0, src1, src2, src3, src4);
4553 src += (5 * stride);
4561 LD_SB4(src, stride, src5, src6, src7, src8);
4562 src += (4 * stride);
4603 LD4(dst, stride, tp0, tp1, tp2, tp3);
4615 ST_D4(dst0, dst1, 0, 1, 0, 1, dst, stride);
4616 dst += (4 * stride);
4618 LD_SB4(src, stride, src9, src10, src11, src12);
4654 LD4(dst, stride, tp0, tp1, tp2, tp3);
4666 ST_D4(dst0, dst1, 0, 1, 0, 1, dst, stride);
4670 ptrdiff_t stride)
4691 src -= ((2 * stride) + 2);
4693 LD_SB5(src, stride, src0, src1, src2, src3, src4);
4694 src += (5 * stride);
4695 LD_SB4(src, stride, src5, src6, src7, src8);
4729 LW4(dst, stride, tp0, tp1, tp2, tp3);
4733 ST_W4(res, 0, 1, 2, 3, dst, stride);
4737 ptrdiff_t stride)
4758 src -= ((2 * stride) + 2);
4760 LD_SB5(src, stride, src0, src1, src2, src3, src4);
4761 src += (5 * stride);
4762 LD_SB4(src, stride, src5, src6, src7, src8);
4797 LW4(dst, stride, tp0, tp1, tp2, tp3);
4801 ST_W4(res, 0, 1, 2, 3, dst, stride);
4805 ptrdiff_t stride)
4821 src -= (stride * 2);
4823 LD_SB5(src, stride, src0, src1, src2, src3, src4);
4824 src += (5 * stride);
4833 LD_SB4(src, stride, src5, src6, src7, src8);
4834 src += (4 * stride);
4853 LD_UB4(dst, stride, dst0, dst1, dst2, dst3);
4859 ST_UB4(res0, res1, res2, res3, dst, stride);
4860 dst += (4 * stride);
4875 ptrdiff_t stride)
4892 src -= (stride * 2);
4894 LD_SB5(src, stride, src0, src1, src2, src3, src4);
4895 src += (5 * stride);
4901 LD_SB4(src, stride, src7, src8, src9, src10);
4902 src += (4 * stride);
4911 LD_SB4(src, stride, src0, src1, src2, src3);
4920 LD4(dst, stride, tp0, tp1, tp2, tp3);
4923 LD4(dst + 4 * stride, stride, tp0, tp1, tp2, tp3);
4937 ST_D8(dst0, dst1, dst2, dst3, 0, 1, 0, 1, 0, 1, 0, 1, dst, stride);
4941 ptrdiff_t stride)
4957 src -= (stride * 2);
4958 LD_SB5(src, stride, src0, src1, src2, src3, src4);
4959 src += (5 * stride);
4965 LD_SB4(src, stride, src5, src6, src7, src8);
4974 LW4(dst, stride, tp0, tp1, tp2, tp3);
4978 ST_W4(dst0, 0, 1, 2, 3, dst, stride);
4982 ptrdiff_t stride)
5002 src -= ((2 * stride) + 2);
5004 LD_SB5(src, stride, src0, src1, src2, src3, src4);
5005 LD_SB5(src + 8, stride, src7, src8, src9, src10, src11);
5006 src += (5 * stride);
5012 src += stride;
5050 dst += stride;
5066 ptrdiff_t stride)
5086 src -= ((2 * stride) + 2);
5088 LD_SB5(src, stride, src0, src1, src2, src3, src4);
5089 LD_SB5(src + 8, stride, src7, src8, src9, src10, src11);
5090 src += (5 * stride);
5096 src += stride;
5135 dst += stride;
5151 ptrdiff_t stride)
5172 src -= ((2 * stride) + 2);
5174 LD_SB5(src, stride, src0, src1, src2, src3, src4);
5175 src += (5 * stride);
5179 LD_SB2(src, stride, src5, src6);
5180 src += (2 * stride);
5209 LD2(dst, stride, tp0, tp1);
5218 ST_D2(out, 0, 1, dst, stride);
5219 dst += (2 * stride);
5230 ptrdiff_t stride)
5251 src -= ((2 * stride) + 2);
5253 LD_SB5(src, stride, src0, src1, src2, src3, src4);
5254 src += (5 * stride);
5258 LD_SB2(src, stride, src5, src6);
5259 src += (2 * stride);
5288 LD2(dst, stride, tp0, tp1);
5298 ST_D2(out, 0, 1, dst, stride);
5299 dst += (2 * stride);
5310 ptrdiff_t stride)
5335 src -= ((2 * stride) + 2);
5337 LD_SB5(src, stride, src0, src1, src2, src3, src4);
5338 src += (5 * stride);
5340 LD_SB4(src, stride, src5, src6, src7, src8);
5397 LW4(dst, stride, tp0, tp1, tp2, tp3);
5402 ST_W4(out, 0, 1, 2, 3, dst, stride);
5406 ptrdiff_t stride)
5431 src -= ((2 * stride) + 2);
5433 LD_SB5(src, stride, src0, src1, src2, src3, src4);
5434 src += (5 * stride);
5436 LD_SB4(src, stride, src5, src6, src7, src8);
5496 LW4(dst, stride, tp0, tp1, tp2, tp3);
5501 ST_W4(out, 0, 1, 2, 3, dst, stride);
5505 ptrdiff_t stride)
5510 const uint8_t *src_tmp = src - (2 * stride) - 2;
5534 LD_SB5(src, stride, src0, src1, src2, src3, src4);
5536 src += (5 * stride);
5545 LD_SB4(src, stride, src0, src1, src2, src3);
5547 src += (4 * stride);
5587 LD4(dst, stride, tp0, tp1, tp2, tp3);
5593 ST_D4(out0, out1, 0, 1, 0, 1, dst, stride);
5594 dst += (4 * stride);
5609 ptrdiff_t stride)
5633 src -= ((2 * stride) + 2);
5634 LD_SB5(src, stride, src0, src1, src2, src3, src4);
5636 src += (5 * stride);
5644 LD_SB4(src, stride, src0, src1, src2, src3);
5646 src += (4 * stride);
5680 LD4(dst, stride, tp0, tp1, tp2, tp3);
5686 ST_D4(dst0, dst1, 0, 1, 0, 1, dst, stride);
5687 dst += (4 * stride);
5689 LD_SB4(src, stride, src0, src1, src2, src3);
5721 LD4(dst, stride, tp0, tp1, tp2, tp3);
5727 ST_D4(dst0, dst1, 0, 1, 0, 1, dst, stride);
5731 ptrdiff_t stride)
5752 src -= ((2 * stride) + 2);
5754 LD_SB5(src, stride, src0, src1, src2, src3, src4);
5755 src += (5 * stride);
5756 LD_SB4(src, stride, src5, src6, src7, src8);
5782 LW4(dst, stride, tp0, tp1, tp2, tp3);
5786 ST_W4(res, 0, 1, 2, 3, dst, stride);