diff options
Diffstat (limited to 'libavcodec/loongarch/hevc_mc.S')
| -rw-r--r-- | libavcodec/loongarch/hevc_mc.S | 224 |
1 files changed, 48 insertions, 176 deletions
diff --git a/libavcodec/loongarch/hevc_mc.S b/libavcodec/loongarch/hevc_mc.S index a0e5938fbd..12d92e32e9 100644 --- a/libavcodec/loongarch/hevc_mc.S +++ b/libavcodec/loongarch/hevc_mc.S @@ -498,7 +498,6 @@ endfunc function ff_hevc_put_hevc_qpel_uni_w_v4_8_lsx LOAD_VAR 128 ld.d t0, sp, 8 //my - addi.d t0, t0, -1 slli.w t0, t0, 4 la.local t1, ff_hevc_qpel_filters vldx vr5, t1, t0 //filter @@ -570,7 +569,6 @@ endfunc function ff_hevc_put_hevc_qpel_uni_w_v6_8_lsx LOAD_VAR 128 ld.d t0, sp, 8 //my - addi.d t0, t0, -1 slli.w t0, t0, 4 la.local t1, ff_hevc_qpel_filters vldx vr5, t1, t0 //filter @@ -699,7 +697,6 @@ endfunc function ff_hevc_put_hevc_qpel_uni_w_v8_8_lsx LOAD_VAR 128 ld.d t0, sp, 8 //my - addi.d t0, t0, -1 slli.w t0, t0, 4 la.local t1, ff_hevc_qpel_filters vldx vr5, t1, t0 //filter @@ -778,7 +775,6 @@ endfunc function ff_hevc_put_hevc_qpel_uni_w_v8_8_lasx LOAD_VAR 256 ld.d t0, sp, 8 //my - addi.d t0, t0, -1 slli.w t0, t0, 4 la.local t1, ff_hevc_qpel_filters vldx vr5, t1, t0 //filter @@ -845,7 +841,6 @@ endfunc function ff_hevc_put_hevc_qpel_uni_w_v16_8_lsx LOAD_VAR 128 ld.d t0, sp, 8 //my - addi.d t0, t0, -1 slli.w t0, t0, 4 la.local t1, ff_hevc_qpel_filters vldx vr5, t1, t0 //filter @@ -940,7 +935,6 @@ endfunc function ff_hevc_put_hevc_qpel_uni_w_v16_8_lasx LOAD_VAR 256 ld.d t0, sp, 8 //my - addi.d t0, t0, -1 slli.w t0, t0, 4 la.local t1, ff_hevc_qpel_filters vldx vr5, t1, t0 //filter @@ -955,7 +949,6 @@ endfunc function ff_hevc_put_hevc_qpel_uni_w_v12_8_lsx LOAD_VAR 128 ld.d t0, sp, 8 //my - addi.d t0, t0, -1 slli.w t0, t0, 4 la.local t1, ff_hevc_qpel_filters vldx vr5, t1, t0 //filter @@ -969,7 +962,6 @@ endfunc function ff_hevc_put_hevc_qpel_uni_w_v12_8_lasx LOAD_VAR 256 ld.d t0, sp, 8 //my - addi.d t0, t0, -1 slli.w t0, t0, 4 la.local t1, ff_hevc_qpel_filters vldx vr5, t1, t0 //filter @@ -984,7 +976,6 @@ endfunc function ff_hevc_put_hevc_qpel_uni_w_v24_8_lsx LOAD_VAR 128 ld.d t0, sp, 8 //my - addi.d t0, t0, -1 slli.w t0, t0, 4 la.local t1, ff_hevc_qpel_filters vldx vr5, t1, t0 //filter @@ -1005,7 +996,6 @@ endfunc function ff_hevc_put_hevc_qpel_uni_w_v24_8_lasx LOAD_VAR 256 ld.d t0, sp, 8 //my - addi.d t0, t0, -1 slli.w t0, t0, 4 la.local t1, ff_hevc_qpel_filters vldx vr5, t1, t0 //filter @@ -1027,7 +1017,6 @@ endfunc function ff_hevc_put_hevc_qpel_uni_w_v32_8_lsx LOAD_VAR 128 ld.d t0, sp, 8 //my - addi.d t0, t0, -1 slli.w t0, t0, 4 la.local t1, ff_hevc_qpel_filters vldx vr5, t1, t0 //filter @@ -1051,7 +1040,6 @@ endfunc function ff_hevc_put_hevc_qpel_uni_w_v32_8_lasx LOAD_VAR 256 ld.d t0, sp, 8 //my - addi.d t0, t0, -1 slli.w t0, t0, 4 la.local t1, ff_hevc_qpel_filters vldx vr5, t1, t0 //filter @@ -1076,7 +1064,6 @@ endfunc function ff_hevc_put_hevc_qpel_uni_w_v48_8_lsx LOAD_VAR 128 ld.d t0, sp, 8 //my - addi.d t0, t0, -1 slli.w t0, t0, 4 la.local t1, ff_hevc_qpel_filters vldx vr5, t1, t0 //filter @@ -1102,7 +1089,6 @@ endfunc function ff_hevc_put_hevc_qpel_uni_w_v48_8_lasx LOAD_VAR 256 ld.d t0, sp, 8 //my - addi.d t0, t0, -1 slli.w t0, t0, 4 la.local t1, ff_hevc_qpel_filters vldx vr5, t1, t0 //filter @@ -1129,7 +1115,6 @@ endfunc function ff_hevc_put_hevc_qpel_uni_w_v64_8_lsx LOAD_VAR 128 ld.d t0, sp, 8 //my - addi.d t0, t0, -1 slli.w t0, t0, 4 la.local t1, ff_hevc_qpel_filters vldx vr5, t1, t0 //filter @@ -1155,7 +1140,6 @@ endfunc function ff_hevc_put_hevc_qpel_uni_w_v64_8_lasx LOAD_VAR 256 ld.d t0, sp, 8 //my - addi.d t0, t0, -1 slli.w t0, t0, 4 la.local t1, ff_hevc_qpel_filters vldx vr5, t1, t0 //filter @@ -1275,7 +1259,6 @@ endfunc function ff_hevc_put_hevc_qpel_uni_w_h4_8_lsx LOAD_VAR 128 ld.d t0, sp, 0 //mx - addi.d t0, t0, -1 slli.w t0, t0, 4 la.local t1, ff_hevc_qpel_filters vldx vr5, t1, t0 //filter @@ -1325,7 +1308,6 @@ endfunc function ff_hevc_put_hevc_qpel_uni_w_h4_8_lasx LOAD_VAR 256 ld.d t0, sp, 0 //mx - addi.d t0, t0, -1 slli.w t0, t0, 4 la.local t1, ff_hevc_qpel_filters vldx vr5, t1, t0 //filter @@ -1364,7 +1346,6 @@ endfunc function ff_hevc_put_hevc_qpel_uni_w_h6_8_lsx LOAD_VAR 128 ld.d t0, sp, 0 //mx - addi.d t0, t0, -1 slli.w t0, t0, 4 la.local t1, ff_hevc_qpel_filters vldx vr5, t1, t0 //filter @@ -1385,7 +1366,6 @@ endfunc function ff_hevc_put_hevc_qpel_uni_w_h6_8_lasx LOAD_VAR 256 ld.d t0, sp, 0 //mx - addi.d t0, t0, -1 slli.w t0, t0, 4 la.local t1, ff_hevc_qpel_filters vldx vr5, t1, t0 //filter @@ -1408,7 +1388,6 @@ endfunc function ff_hevc_put_hevc_qpel_uni_w_h8_8_lsx LOAD_VAR 128 ld.d t0, sp, 0 //mx - addi.d t0, t0, -1 slli.w t0, t0, 4 la.local t1, ff_hevc_qpel_filters vldx vr5, t1, t0 //filter @@ -1428,7 +1407,6 @@ endfunc function ff_hevc_put_hevc_qpel_uni_w_h8_8_lasx LOAD_VAR 256 ld.d t0, sp, 0 //mx - addi.d t0, t0, -1 slli.w t0, t0, 4 la.local t1, ff_hevc_qpel_filters vldx vr5, t1, t0 //filter @@ -1450,7 +1428,6 @@ endfunc function ff_hevc_put_hevc_qpel_uni_w_h12_8_lsx LOAD_VAR 128 ld.d t0, sp, 0 //mx - addi.d t0, t0, -1 slli.w t0, t0, 4 la.local t1, ff_hevc_qpel_filters vldx vr5, t1, t0 //filter @@ -1475,7 +1452,6 @@ endfunc function ff_hevc_put_hevc_qpel_uni_w_h12_8_lasx LOAD_VAR 256 ld.d t0, sp, 0 //mx - addi.d t0, t0, -1 slli.w t0, t0, 4 la.local t1, ff_hevc_qpel_filters vldx vr5, t1, t0 //filter @@ -1495,7 +1471,6 @@ endfunc function ff_hevc_put_hevc_qpel_uni_w_h16_8_lsx LOAD_VAR 128 ld.d t0, sp, 0 //mx - addi.d t0, t0, -1 slli.w t0, t0, 4 la.local t1, ff_hevc_qpel_filters vldx vr5, t1, t0 //filter @@ -1518,7 +1493,6 @@ endfunc function ff_hevc_put_hevc_qpel_uni_w_h16_8_lasx LOAD_VAR 256 ld.d t0, sp, 0 //mx - addi.d t0, t0, -1 slli.w t0, t0, 4 la.local t1, ff_hevc_qpel_filters vldx vr5, t1, t0 //filter @@ -1537,7 +1511,6 @@ endfunc function ff_hevc_put_hevc_qpel_uni_w_h24_8_lsx LOAD_VAR 128 ld.d t0, sp, 0 //mx - addi.d t0, t0, -1 slli.w t0, t0, 4 la.local t1, ff_hevc_qpel_filters vldx vr5, t1, t0 //filter @@ -1565,7 +1538,6 @@ endfunc function ff_hevc_put_hevc_qpel_uni_w_h24_8_lasx LOAD_VAR 256 ld.d t0, sp, 0 //mx - addi.d t0, t0, -1 slli.w t0, t0, 4 la.local t1, ff_hevc_qpel_filters vldx vr5, t1, t0 //filter @@ -1590,7 +1562,6 @@ endfunc function ff_hevc_put_hevc_qpel_uni_w_h32_8_lsx LOAD_VAR 128 ld.d t0, sp, 0 //mx - addi.d t0, t0, -1 slli.w t0, t0, 4 la.local t1, ff_hevc_qpel_filters vldx vr5, t1, t0 //filter @@ -1622,7 +1593,6 @@ endfunc function ff_hevc_put_hevc_qpel_uni_w_h32_8_lasx LOAD_VAR 256 ld.d t0, sp, 0 //mx - addi.d t0, t0, -1 slli.w t0, t0, 4 la.local t1, ff_hevc_qpel_filters vldx vr5, t1, t0 //filter @@ -1644,7 +1614,6 @@ endfunc function ff_hevc_put_hevc_qpel_uni_w_h48_8_lsx LOAD_VAR 128 ld.d t0, sp, 0 //mx - addi.d t0, t0, -1 slli.w t0, t0, 4 la.local t1, ff_hevc_qpel_filters vldx vr5, t1, t0 //filter @@ -1684,7 +1653,6 @@ endfunc function ff_hevc_put_hevc_qpel_uni_w_h48_8_lasx LOAD_VAR 256 ld.d t0, sp, 0 //mx - addi.d t0, t0, -1 slli.w t0, t0, 4 la.local t1, ff_hevc_qpel_filters vldx vr5, t1, t0 //filter @@ -1709,7 +1677,6 @@ endfunc function ff_hevc_put_hevc_qpel_uni_w_h64_8_lsx LOAD_VAR 128 ld.d t0, sp, 0 //mx - addi.d t0, t0, -1 slli.w t0, t0, 4 la.local t1, ff_hevc_qpel_filters vldx vr5, t1, t0 //filter @@ -1757,7 +1724,6 @@ endfunc function ff_hevc_put_hevc_qpel_uni_w_h64_8_lasx LOAD_VAR 256 ld.d t0, sp, 0 //mx - addi.d t0, t0, -1 slli.w t0, t0, 4 la.local t1, ff_hevc_qpel_filters vldx vr5, t1, t0 //filter @@ -1841,15 +1807,13 @@ endconst function ff_hevc_put_hevc_epel_uni_w_hv4_8_lsx LOAD_VAR 128 ld.d t0, sp, 0 // mx - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters - vldx vr5, t1, t0 // ff_hevc_epel_filters[mx - 1]; + vldx vr5, t1, t0 // ff_hevc_epel_filters[mx]; vreplvei.w vr5, vr5, 0 ld.d t0, sp, 8 // my - addi.d t0, t0, -1 slli.w t0, t0, 2 - vldx vr6, t1, t0 // ff_hevc_epel_filters[my - 1]; + vldx vr6, t1, t0 // ff_hevc_epel_filters[my]; vsllwil.h.b vr6, vr6, 0 vsllwil.w.h vr6, vr6, 0 vreplvei.w vr16, vr6, 0 @@ -2068,15 +2032,13 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_hv6_8_lsx LOAD_VAR 128 ld.d t0, sp, 0 // mx - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters - vldx vr5, t1, t0 // ff_hevc_epel_filters[mx - 1]; + vldx vr5, t1, t0 // ff_hevc_epel_filters[mx]; vreplvei.w vr5, vr5, 0 ld.d t0, sp, 8 // my - addi.d t0, t0, -1 slli.w t0, t0, 2 - vldx vr6, t1, t0 // ff_hevc_epel_filters[my - 1]; + vldx vr6, t1, t0 // ff_hevc_epel_filters[my]; vsllwil.h.b vr6, vr6, 0 vsllwil.w.h vr6, vr6, 0 vreplvei.w vr16, vr6, 0 @@ -2094,15 +2056,13 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_hv6_8_lasx LOAD_VAR 256 ld.d t0, sp, 0 // mx - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters - vldx vr5, t1, t0 // ff_hevc_epel_filters[mx - 1]; + vldx vr5, t1, t0 // ff_hevc_epel_filters[mx]; xvreplve0.w xr5, xr5 ld.d t0, sp, 8 // my - addi.d t0, t0, -1 slli.w t0, t0, 2 - vldx vr6, t1, t0 // ff_hevc_epel_filters[my - 1]; + vldx vr6, t1, t0 // ff_hevc_epel_filters[my]; vsllwil.h.b vr6, vr6, 0 vsllwil.w.h vr6, vr6, 0 xvreplve0.q xr6, xr6 @@ -2120,15 +2080,13 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_hv8_8_lsx LOAD_VAR 128 ld.d t0, sp, 0 // mx - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters - vldx vr5, t1, t0 // ff_hevc_epel_filters[mx - 1]; + vldx vr5, t1, t0 // ff_hevc_epel_filters[mx]; vreplvei.w vr5, vr5, 0 ld.d t0, sp, 8 // my - addi.d t0, t0, -1 slli.w t0, t0, 2 - vldx vr6, t1, t0 // ff_hevc_epel_filters[my - 1]; + vldx vr6, t1, t0 // ff_hevc_epel_filters[my]; vsllwil.h.b vr6, vr6, 0 vsllwil.w.h vr6, vr6, 0 vreplvei.w vr16, vr6, 0 @@ -2146,15 +2104,13 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_hv8_8_lasx LOAD_VAR 256 ld.d t0, sp, 0 // mx - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters - vldx vr5, t1, t0 // ff_hevc_epel_filters[mx - 1]; + vldx vr5, t1, t0 // ff_hevc_epel_filters[mx]; xvreplve0.w xr5, xr5 ld.d t0, sp, 8 // my - addi.d t0, t0, -1 slli.w t0, t0, 2 - vldx vr6, t1, t0 // ff_hevc_epel_filters[my - 1]; + vldx vr6, t1, t0 // ff_hevc_epel_filters[my]; vsllwil.h.b vr6, vr6, 0 vsllwil.w.h vr6, vr6, 0 xvreplve0.q xr6, xr6 @@ -2172,15 +2128,13 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_hv12_8_lsx LOAD_VAR 128 ld.d t0, sp, 0 // mx - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters - vldx vr5, t1, t0 // ff_hevc_epel_filters[mx - 1]; + vldx vr5, t1, t0 // ff_hevc_epel_filters[mx]; vreplvei.w vr5, vr5, 0 ld.d t0, sp, 8 // my - addi.d t0, t0, -1 slli.w t0, t0, 2 - vldx vr6, t1, t0 // ff_hevc_epel_filters[my - 1]; + vldx vr6, t1, t0 // ff_hevc_epel_filters[my]; vsllwil.h.b vr6, vr6, 0 vsllwil.w.h vr6, vr6, 0 vreplvei.w vr16, vr6, 0 @@ -2205,15 +2159,13 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_hv12_8_lasx LOAD_VAR 256 ld.d t0, sp, 0 // mx - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters - vldx vr5, t1, t0 // ff_hevc_epel_filters[mx - 1]; + vldx vr5, t1, t0 // ff_hevc_epel_filters[mx]; xvreplve0.w xr5, xr5 ld.d t0, sp, 8 // my - addi.d t0, t0, -1 slli.w t0, t0, 2 - vldx vr6, t1, t0 // ff_hevc_epel_filters[my - 1]; + vldx vr6, t1, t0 // ff_hevc_epel_filters[my]; vsllwil.h.b vr6, vr6, 0 vsllwil.w.h vr6, vr6, 0 xvreplve0.q xr6, xr6 @@ -2231,15 +2183,13 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_hv16_8_lsx LOAD_VAR 128 ld.d t0, sp, 0 // mx - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters - vldx vr5, t1, t0 // ff_hevc_epel_filters[mx - 1]; + vldx vr5, t1, t0 // ff_hevc_epel_filters[mx]; vreplvei.w vr5, vr5, 0 ld.d t0, sp, 8 // my - addi.d t0, t0, -1 slli.w t0, t0, 2 - vldx vr6, t1, t0 // ff_hevc_epel_filters[my - 1]; + vldx vr6, t1, t0 // ff_hevc_epel_filters[my]; vsllwil.h.b vr6, vr6, 0 vsllwil.w.h vr6, vr6, 0 vreplvei.w vr16, vr6, 0 @@ -2267,15 +2217,13 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_hv16_8_lasx LOAD_VAR 256 ld.d t0, sp, 0 // mx - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters - vldx vr5, t1, t0 // ff_hevc_epel_filters[mx - 1]; + vldx vr5, t1, t0 // ff_hevc_epel_filters[mx]; xvreplve0.w xr5, xr5 ld.d t0, sp, 8 // my - addi.d t0, t0, -1 slli.w t0, t0, 2 - vldx vr6, t1, t0 // ff_hevc_epel_filters[my - 1]; + vldx vr6, t1, t0 // ff_hevc_epel_filters[my]; vsllwil.h.b vr6, vr6, 0 vsllwil.w.h vr6, vr6, 0 xvreplve0.q xr6, xr6 @@ -2293,15 +2241,13 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_hv24_8_lsx LOAD_VAR 128 ld.d t0, sp, 0 // mx - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters - vldx vr5, t1, t0 // ff_hevc_epel_filters[mx - 1]; + vldx vr5, t1, t0 // ff_hevc_epel_filters[mx]; vreplvei.w vr5, vr5, 0 ld.d t0, sp, 8 // my - addi.d t0, t0, -1 slli.w t0, t0, 2 - vldx vr6, t1, t0 // ff_hevc_epel_filters[my - 1]; + vldx vr6, t1, t0 // ff_hevc_epel_filters[my]; vsllwil.h.b vr6, vr6, 0 vsllwil.w.h vr6, vr6, 0 vreplvei.w vr16, vr6, 0 @@ -2331,15 +2277,13 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_hv24_8_lasx LOAD_VAR 256 ld.d t0, sp, 0 // mx - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters - vldx vr5, t1, t0 // ff_hevc_epel_filters[mx - 1]; + vldx vr5, t1, t0 // ff_hevc_epel_filters[mx]; xvreplve0.w xr5, xr5 ld.d t0, sp, 8 // my - addi.d t0, t0, -1 slli.w t0, t0, 2 - vldx vr6, t1, t0 // ff_hevc_epel_filters[my - 1]; + vldx vr6, t1, t0 // ff_hevc_epel_filters[my]; vsllwil.h.b vr6, vr6, 0 vsllwil.w.h vr6, vr6, 0 xvreplve0.q xr6, xr6 @@ -2364,15 +2308,13 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_hv32_8_lsx LOAD_VAR 128 ld.d t0, sp, 0 // mx - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters - vldx vr5, t1, t0 // ff_hevc_epel_filters[mx - 1]; + vldx vr5, t1, t0 // ff_hevc_epel_filters[mx]; vreplvei.w vr5, vr5, 0 ld.d t0, sp, 8 // my - addi.d t0, t0, -1 slli.w t0, t0, 2 - vldx vr6, t1, t0 // ff_hevc_epel_filters[my - 1]; + vldx vr6, t1, t0 // ff_hevc_epel_filters[my]; vsllwil.h.b vr6, vr6, 0 vsllwil.w.h vr6, vr6, 0 vreplvei.w vr16, vr6, 0 @@ -2402,15 +2344,13 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_hv32_8_lasx LOAD_VAR 256 ld.d t0, sp, 0 // mx - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters - vldx vr5, t1, t0 // ff_hevc_epel_filters[mx - 1]; + vldx vr5, t1, t0 // ff_hevc_epel_filters[mx]; xvreplve0.w xr5, xr5 ld.d t0, sp, 8 // my - addi.d t0, t0, -1 slli.w t0, t0, 2 - vldx vr6, t1, t0 // ff_hevc_epel_filters[my - 1]; + vldx vr6, t1, t0 // ff_hevc_epel_filters[my]; vsllwil.h.b vr6, vr6, 0 vsllwil.w.h vr6, vr6, 0 xvreplve0.q xr6, xr6 @@ -2440,15 +2380,13 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_hv48_8_lsx LOAD_VAR 128 ld.d t0, sp, 0 // mx - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters - vldx vr5, t1, t0 // ff_hevc_epel_filters[mx - 1]; + vldx vr5, t1, t0 // ff_hevc_epel_filters[mx]; vreplvei.w vr5, vr5, 0 ld.d t0, sp, 8 // my - addi.d t0, t0, -1 slli.w t0, t0, 2 - vldx vr6, t1, t0 // ff_hevc_epel_filters[my - 1]; + vldx vr6, t1, t0 // ff_hevc_epel_filters[my]; vsllwil.h.b vr6, vr6, 0 vsllwil.w.h vr6, vr6, 0 vreplvei.w vr16, vr6, 0 @@ -2478,15 +2416,13 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_hv48_8_lasx LOAD_VAR 256 ld.d t0, sp, 0 // mx - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters - vldx vr5, t1, t0 // ff_hevc_epel_filters[mx - 1]; + vldx vr5, t1, t0 // ff_hevc_epel_filters[mx]; xvreplve0.w xr5, xr5 ld.d t0, sp, 8 // my - addi.d t0, t0, -1 slli.w t0, t0, 2 - vldx vr6, t1, t0 // ff_hevc_epel_filters[my - 1]; + vldx vr6, t1, t0 // ff_hevc_epel_filters[my]; vsllwil.h.b vr6, vr6, 0 vsllwil.w.h vr6, vr6, 0 xvreplve0.q xr6, xr6 @@ -2516,15 +2452,13 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_hv64_8_lsx LOAD_VAR 128 ld.d t0, sp, 0 // mx - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters - vldx vr5, t1, t0 // ff_hevc_epel_filters[mx - 1]; + vldx vr5, t1, t0 // ff_hevc_epel_filters[mx]; vreplvei.w vr5, vr5, 0 ld.d t0, sp, 8 // my - addi.d t0, t0, -1 slli.w t0, t0, 2 - vldx vr6, t1, t0 // ff_hevc_epel_filters[my - 1]; + vldx vr6, t1, t0 // ff_hevc_epel_filters[my]; vsllwil.h.b vr6, vr6, 0 vsllwil.w.h vr6, vr6, 0 vreplvei.w vr16, vr6, 0 @@ -2554,15 +2488,13 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_hv64_8_lasx LOAD_VAR 256 ld.d t0, sp, 0 // mx - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters - vldx vr5, t1, t0 // ff_hevc_epel_filters[mx - 1]; + vldx vr5, t1, t0 // ff_hevc_epel_filters[mx]; xvreplve0.w xr5, xr5 ld.d t0, sp, 8 // my - addi.d t0, t0, -1 slli.w t0, t0, 2 - vldx vr6, t1, t0 // ff_hevc_epel_filters[my - 1]; + vldx vr6, t1, t0 // ff_hevc_epel_filters[my]; vsllwil.h.b vr6, vr6, 0 vsllwil.w.h vr6, vr6, 0 xvreplve0.q xr6, xr6 @@ -2596,8 +2528,7 @@ endfunc * int width) */ function ff_hevc_put_hevc_uni_qpel_h4_8_lsx - addi.d t0, a5, -1 - slli.w t0, t0, 4 + slli.w t0, a5, 4 la.local t1, ff_hevc_qpel_filters vldx vr5, t1, t0 //filter addi.d a2, a2, -3 //src -= 3 @@ -2663,8 +2594,7 @@ endfunc .endm function ff_hevc_put_hevc_uni_qpel_h6_8_lsx - addi.d t0, a5, -1 - slli.w t0, t0, 4 + slli.w t0, a5, 4 la.local t1, ff_hevc_qpel_filters vldx vr0, t1, t0 //filter abcdefgh vreplvei.h vr1, vr0, 1 //cd... @@ -2692,8 +2622,7 @@ function ff_hevc_put_hevc_uni_qpel_h6_8_lsx endfunc function ff_hevc_put_hevc_uni_qpel_h8_8_lsx - addi.d t0, a5, -1 - slli.w t0, t0, 4 + slli.w t0, a5, 4 la.local t1, ff_hevc_qpel_filters vldx vr0, t1, t0 //filter abcdefgh vreplvei.h vr1, vr0, 1 //cd... @@ -2720,8 +2649,7 @@ function ff_hevc_put_hevc_uni_qpel_h8_8_lsx endfunc function ff_hevc_put_hevc_uni_qpel_h12_8_lsx - addi.d t0, a5, -1 - slli.w t0, t0, 4 + slli.w t0, a5, 4 la.local t1, ff_hevc_qpel_filters vldx vr0, t1, t0 //filter abcdefgh vreplvei.h vr1, vr0, 1 //cd... @@ -2751,8 +2679,7 @@ function ff_hevc_put_hevc_uni_qpel_h12_8_lsx endfunc function ff_hevc_put_hevc_uni_qpel_h12_8_lasx - addi.d t0, a5, -1 - slli.w t0, t0, 4 + slli.w t0, a5, 4 la.local t1, ff_hevc_qpel_filters vldx vr0, t1, t0 //filter abcdefgh xvreplve0.q xr0, xr0 @@ -2784,8 +2711,7 @@ function ff_hevc_put_hevc_uni_qpel_h12_8_lasx endfunc function ff_hevc_put_hevc_uni_qpel_h16_8_lsx - addi.d t0, a5, -1 - slli.w t0, t0, 4 + slli.w t0, a5, 4 la.local t1, ff_hevc_qpel_filters vldx vr0, t1, t0 //filter abcdefgh vreplvei.h vr1, vr0, 1 //cd... @@ -2814,8 +2740,7 @@ function ff_hevc_put_hevc_uni_qpel_h16_8_lsx endfunc function ff_hevc_put_hevc_uni_qpel_h16_8_lasx - addi.d t0, a5, -1 - slli.w t0, t0, 4 + slli.w t0, a5, 4 la.local t1, ff_hevc_qpel_filters vldx vr0, t1, t0 //filter abcdefgh xvreplve0.q xr0, xr0 @@ -2846,8 +2771,7 @@ function ff_hevc_put_hevc_uni_qpel_h16_8_lasx endfunc function ff_hevc_put_hevc_uni_qpel_h24_8_lsx - addi.d t0, a5, -1 - slli.w t0, t0, 4 + slli.w t0, a5, 4 la.local t1, ff_hevc_qpel_filters vldx vr0, t1, t0 //filter abcdefgh vreplvei.h vr1, vr0, 1 //cd... @@ -2880,8 +2804,7 @@ function ff_hevc_put_hevc_uni_qpel_h24_8_lsx endfunc function ff_hevc_put_hevc_uni_qpel_h24_8_lasx - addi.d t0, a5, -1 - slli.w t0, t0, 4 + slli.w t0, a5, 4 la.local t1, ff_hevc_qpel_filters vldx vr0, t1, t0 //filter abcdefgh xvreplve0.q xr0, xr0 @@ -2916,8 +2839,7 @@ function ff_hevc_put_hevc_uni_qpel_h24_8_lasx endfunc function ff_hevc_put_hevc_uni_qpel_h32_8_lsx - addi.d t0, a5, -1 - slli.w t0, t0, 4 + slli.w t0, a5, 4 la.local t1, ff_hevc_qpel_filters vldx vr0, t1, t0 //filter abcdefgh vreplvei.h vr1, vr0, 1 //cd... @@ -2952,8 +2874,7 @@ function ff_hevc_put_hevc_uni_qpel_h32_8_lsx endfunc function ff_hevc_put_hevc_uni_qpel_h32_8_lasx - addi.d t0, a5, -1 - slli.w t0, t0, 4 + slli.w t0, a5, 4 la.local t1, ff_hevc_qpel_filters vldx vr0, t1, t0 //filter abcdefgh xvreplve0.q xr0, xr0 @@ -2987,8 +2908,7 @@ function ff_hevc_put_hevc_uni_qpel_h32_8_lasx endfunc function ff_hevc_put_hevc_uni_qpel_h48_8_lsx - addi.d t0, a5, -1 - slli.w t0, t0, 4 + slli.w t0, a5, 4 la.local t1, ff_hevc_qpel_filters vldx vr0, t1, t0 //filter abcdefgh vreplvei.h vr1, vr0, 1 //cd... @@ -3029,8 +2949,7 @@ function ff_hevc_put_hevc_uni_qpel_h48_8_lsx endfunc function ff_hevc_put_hevc_uni_qpel_h48_8_lasx - addi.d t0, a5, -1 - slli.w t0, t0, 4 + slli.w t0, a5, 4 la.local t1, ff_hevc_qpel_filters vldx vr0, t1, t0 //filter abcdefgh xvreplve0.q xr0, xr0 @@ -3070,8 +2989,7 @@ function ff_hevc_put_hevc_uni_qpel_h48_8_lasx endfunc function ff_hevc_put_hevc_uni_qpel_h64_8_lasx - addi.d t0, a5, -1 - slli.w t0, t0, 4 + slli.w t0, a5, 4 la.local t1, ff_hevc_qpel_filters vldx vr0, t1, t0 //filter abcdefgh xvreplve0.q xr0, xr0 @@ -3122,7 +3040,6 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_v4_8_lsx LOAD_VAR 128 ld.d t0, sp, 8 //my - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters vldx vr0, t1, t0 //filter @@ -3260,7 +3177,6 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_v6_8_lsx LOAD_VAR 128 ld.d t0, sp, 8 //my - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters vldx vr0, t1, t0 //filter @@ -3275,7 +3191,6 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_v6_8_lasx LOAD_VAR 256 ld.d t0, sp, 8 //my - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters vldx vr0, t1, t0 //filter @@ -3289,7 +3204,6 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_v8_8_lsx LOAD_VAR 128 ld.d t0, sp, 8 //my - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters vldx vr0, t1, t0 //filter @@ -3304,7 +3218,6 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_v8_8_lasx LOAD_VAR 256 ld.d t0, sp, 8 //my - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters vldx vr0, t1, t0 //filter @@ -3394,7 +3307,6 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_v12_8_lsx LOAD_VAR 128 ld.d t0, sp, 8 //my - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters vldx vr0, t1, t0 //filter @@ -3409,7 +3321,6 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_v12_8_lasx LOAD_VAR 256 ld.d t0, sp, 8 //my - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters vldx vr0, t1, t0 //filter @@ -3425,7 +3336,6 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_v16_8_lsx LOAD_VAR 128 ld.d t0, sp, 8 //my - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters vldx vr0, t1, t0 //filter @@ -3440,7 +3350,6 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_v16_8_lasx LOAD_VAR 256 ld.d t0, sp, 8 //my - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters vldx vr0, t1, t0 //filter @@ -3456,7 +3365,6 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_v24_8_lsx LOAD_VAR 128 ld.d t0, sp, 8 //my - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters vldx vr0, t1, t0 //filter @@ -3478,7 +3386,6 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_v24_8_lasx LOAD_VAR 256 ld.d t0, sp, 8 //my - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters vldx vr0, t1, t0 //filter @@ -3503,7 +3410,6 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_v32_8_lsx LOAD_VAR 128 ld.d t0, sp, 8 //my - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters vldx vr0, t1, t0 //filter @@ -3525,7 +3431,6 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_v32_8_lasx LOAD_VAR 256 ld.d t0, sp, 8 //my - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters vldx vr0, t1, t0 //filter @@ -3548,7 +3453,6 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_v48_8_lsx LOAD_VAR 128 ld.d t0, sp, 8 //my - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters vldx vr0, t1, t0 //filter @@ -3574,7 +3478,6 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_v48_8_lasx LOAD_VAR 256 ld.d t0, sp, 8 //my - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters vldx vr0, t1, t0 //filter @@ -3601,7 +3504,6 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_v64_8_lsx LOAD_VAR 128 ld.d t0, sp, 8 //my - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters vldx vr0, t1, t0 //filter @@ -3631,7 +3533,6 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_v64_8_lasx LOAD_VAR 256 ld.d t0, sp, 8 //my - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters vldx vr0, t1, t0 //filter @@ -3668,7 +3569,6 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_h4_8_lsx LOAD_VAR 128 ld.d t0, sp, 0 //mx - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters vldx vr0, t1, t0 //filter @@ -3699,7 +3599,6 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_h6_8_lsx LOAD_VAR 128 ld.d t0, sp, 0 //mx - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters vldx vr0, t1, t0 //filter @@ -3730,7 +3629,6 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_h6_8_lasx LOAD_VAR 256 ld.d t0, sp, 0 //mx - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters vldx vr0, t1, t0 //filter @@ -3759,7 +3657,6 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_h8_8_lsx LOAD_VAR 128 ld.d t0, sp, 0 //mx - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters vldx vr0, t1, t0 //filter @@ -3789,7 +3686,6 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_h8_8_lasx LOAD_VAR 256 ld.d t0, sp, 0 //mx - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters vldx vr0, t1, t0 //filter @@ -3853,7 +3749,6 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_h12_8_lsx LOAD_VAR 128 ld.d t0, sp, 0 //mx - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters vldx vr0, t1, t0 //filter @@ -3889,7 +3784,6 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_h12_8_lasx LOAD_VAR 256 ld.d t0, sp, 0 //mx - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters vldx vr0, t1, t0 //filter @@ -3910,7 +3804,6 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_h16_8_lsx LOAD_VAR 128 ld.d t0, sp, 0 //mx - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters vldx vr0, t1, t0 //filter @@ -3934,7 +3827,6 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_h16_8_lasx LOAD_VAR 256 ld.d t0, sp, 0 //mx - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters vldx vr0, t1, t0 //filter @@ -3955,7 +3847,6 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_h24_8_lsx LOAD_VAR 128 ld.d t0, sp, 0 //mx - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters vldx vr0, t1, t0 //filter @@ -3986,7 +3877,6 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_h24_8_lasx LOAD_VAR 256 ld.d t0, sp, 0 //mx - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters vldx vr0, t1, t0 //filter @@ -4015,7 +3905,6 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_h32_8_lsx LOAD_VAR 128 ld.d t0, sp, 0 //mx - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters vldx vr0, t1, t0 //filter @@ -4040,7 +3929,6 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_h32_8_lasx LOAD_VAR 256 ld.d t0, sp, 0 //mx - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters vldx vr0, t1, t0 //filter @@ -4062,7 +3950,6 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_h48_8_lsx LOAD_VAR 128 ld.d t0, sp, 0 //mx - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters vldx vr0, t1, t0 //filter @@ -4088,7 +3975,6 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_h48_8_lasx LOAD_VAR 256 ld.d t0, sp, 0 //mx - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters vldx vr0, t1, t0 //filter @@ -4111,7 +3997,6 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_h64_8_lsx LOAD_VAR 128 ld.d t0, sp, 0 //mx - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters vldx vr0, t1, t0 //filter @@ -4138,7 +4023,6 @@ endfunc function ff_hevc_put_hevc_epel_uni_w_h64_8_lasx LOAD_VAR 256 ld.d t0, sp, 0 //mx - addi.d t0, t0, -1 slli.w t0, t0, 2 la.local t1, ff_hevc_epel_filters vldx vr0, t1, t0 //filter @@ -4166,7 +4050,6 @@ endfunc * intptr_t my, int width) */ function ff_hevc_put_hevc_bi_epel_h4_8_lsx - addi.d a6, a6, -1 slli.w a6, a6, 2 la.local t0, ff_hevc_epel_filters vldx vr0, t0, a6 // filter @@ -4209,7 +4092,6 @@ endfunc .endm function ff_hevc_put_hevc_bi_epel_h6_8_lsx - addi.d a6, a6, -1 slli.w a6, a6, 2 la.local t0, ff_hevc_epel_filters vldx vr0, t0, a6 // filter @@ -4234,7 +4116,6 @@ function ff_hevc_put_hevc_bi_epel_h6_8_lsx endfunc function ff_hevc_put_hevc_bi_epel_h8_8_lsx - addi.d a6, a6, -1 slli.w a6, a6, 2 la.local t0, ff_hevc_epel_filters vldx vr0, t0, a6 // filter @@ -4258,7 +4139,6 @@ function ff_hevc_put_hevc_bi_epel_h8_8_lsx endfunc function ff_hevc_put_hevc_bi_epel_h12_8_lsx - addi.d a6, a6, -1 slli.w a6, a6, 2 la.local t0, ff_hevc_epel_filters vldx vr0, t0, a6 // filter @@ -4286,7 +4166,6 @@ function ff_hevc_put_hevc_bi_epel_h12_8_lsx endfunc function ff_hevc_put_hevc_bi_epel_h12_8_lasx - addi.d a6, a6, -1 slli.w a6, a6, 2 la.local t0, ff_hevc_epel_filters vldx vr0, t0, a6 // filter @@ -4314,7 +4193,6 @@ function ff_hevc_put_hevc_bi_epel_h12_8_lasx endfunc function ff_hevc_put_hevc_bi_epel_h16_8_lsx - addi.d a6, a6, -1 slli.w a6, a6, 2 la.local t0, ff_hevc_epel_filters vldx vr0, t0, a6 // filter @@ -4341,7 +4219,6 @@ function ff_hevc_put_hevc_bi_epel_h16_8_lsx endfunc function ff_hevc_put_hevc_bi_epel_h16_8_lasx - addi.d a6, a6, -1 slli.w a6, a6, 2 la.local t0, ff_hevc_epel_filters vldx vr0, t0, a6 // filter @@ -4368,7 +4245,6 @@ function ff_hevc_put_hevc_bi_epel_h16_8_lasx endfunc function ff_hevc_put_hevc_bi_epel_h32_8_lasx - addi.d a6, a6, -1 slli.w a6, a6, 2 la.local t0, ff_hevc_epel_filters vldx vr0, t0, a6 // filter @@ -4400,7 +4276,6 @@ function ff_hevc_put_hevc_bi_epel_h32_8_lasx endfunc function ff_hevc_put_hevc_bi_epel_h48_8_lsx - addi.d a6, a6, -1 slli.w a6, a6, 2 la.local t0, ff_hevc_epel_filters vldx vr0, t0, a6// filter @@ -4443,7 +4318,6 @@ function ff_hevc_put_hevc_bi_epel_h48_8_lsx endfunc function ff_hevc_put_hevc_bi_epel_h48_8_lasx - addi.d a6, a6, -1 slli.w a6, a6, 2 la.local t0, ff_hevc_epel_filters vldx vr0, t0, a6 // filter @@ -4481,7 +4355,6 @@ function ff_hevc_put_hevc_bi_epel_h48_8_lasx endfunc function ff_hevc_put_hevc_bi_epel_h64_8_lsx - addi.d a6, a6, -1 slli.w a6, a6, 2 la.local t0, ff_hevc_epel_filters vldx vr0, t0, a6// filter @@ -4531,7 +4404,6 @@ function ff_hevc_put_hevc_bi_epel_h64_8_lsx endfunc function ff_hevc_put_hevc_bi_epel_h64_8_lasx - addi.d a6, a6, -1 slli.w a6, a6, 2 la.local t0, ff_hevc_epel_filters vldx vr0, t0, a6 // filter |