766     __m128i t0, t1, t2, t3;  in rsdIntrinsicBlendSrcOver_K()  local
788         t1 = _mm_unpackhi_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendSrcOver_K()
789         t1 = _mm_mullo_epi16(t1, _mm_sub_epi16(all1s, ina));  in rsdIntrinsicBlendSrcOver_K()
790         t1 = _mm_srli_epi16(t1, 8);  in rsdIntrinsicBlendSrcOver_K()
791         t1 = _mm_add_epi16(t1, ins);  in rsdIntrinsicBlendSrcOver_K()
809         t0 = _mm_packus_epi16(t0, t1);  in rsdIntrinsicBlendSrcOver_K()
822     __m128i t0, t1, t2, t3;  in rsdIntrinsicBlendDstOver_K()  local
845         t1 = _mm_unpackhi_epi8(in0, _mm_setzero_si128());  in rsdIntrinsicBlendDstOver_K()
846         t1 = _mm_mullo_epi16(t1, _mm_sub_epi16(all1s, outa));  in rsdIntrinsicBlendDstOver_K()
847         t1 = _mm_srli_epi16(t1, 8);  in rsdIntrinsicBlendDstOver_K()
848         t1 = _mm_add_epi16(t1, outs);  in rsdIntrinsicBlendDstOver_K()
866         t0 = _mm_packus_epi16(t0, t1);  in rsdIntrinsicBlendDstOver_K()
879     __m128i t0, t1, t2, t3;  in rsdIntrinsicBlendSrcIn_K()  local
898         t1 = _mm_unpackhi_epi8(in0, _mm_setzero_si128());  in rsdIntrinsicBlendSrcIn_K()
899         t1 = _mm_mullo_epi16(t1, outa);  in rsdIntrinsicBlendSrcIn_K()
900         t1 = _mm_srli_epi16(t1, 8);  in rsdIntrinsicBlendSrcIn_K()
916         t0 = _mm_packus_epi16(t0, t1);  in rsdIntrinsicBlendSrcIn_K()
929     __m128i t0, t1, t2, t3;  in rsdIntrinsicBlendDstIn_K()  local
948         t1 = _mm_unpackhi_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendDstIn_K()
949         t1 = _mm_mullo_epi16(t1, ina);  in rsdIntrinsicBlendDstIn_K()
950         t1 = _mm_srli_epi16(t1, 8);  in rsdIntrinsicBlendDstIn_K()
966         t0 = _mm_packus_epi16(t0, t1);  in rsdIntrinsicBlendDstIn_K()
979     __m128i t0, t1, t2, t3;  in rsdIntrinsicBlendSrcOut_K()  local
1000         t1 = _mm_unpackhi_epi8(in0, _mm_setzero_si128());  in rsdIntrinsicBlendSrcOut_K()
1001         t1 = _mm_mullo_epi16(t1, _mm_sub_epi16(all1s, outa));  in rsdIntrinsicBlendSrcOut_K()
1002         t1 = _mm_srli_epi16(t1, 8);  in rsdIntrinsicBlendSrcOut_K()
1018         t0 = _mm_packus_epi16(t0, t1);  in rsdIntrinsicBlendSrcOut_K()
1031     __m128i t0, t1, t2, t3;  in rsdIntrinsicBlendDstOut_K()  local
1052         t1 = _mm_unpackhi_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendDstOut_K()
1053         t1 = _mm_mullo_epi16(t1, _mm_sub_epi16(all1s, ina));  in rsdIntrinsicBlendDstOut_K()
1054         t1 = _mm_srli_epi16(t1, 8);  in rsdIntrinsicBlendDstOut_K()
1070         t0 = _mm_packus_epi16(t0, t1);  in rsdIntrinsicBlendDstOut_K()
1084     __m128i t0, t1, t2, t3;  in rsdIntrinsicBlendSrcAtop_K()  local
1112         t1 = _mm_sub_epi16(all1s, ina);  in rsdIntrinsicBlendSrcAtop_K()
1113         t1 = _mm_mullo_epi16(t1, outs);  in rsdIntrinsicBlendSrcAtop_K()
1114         t1 = _mm_adds_epu16(t1, _mm_mullo_epi16(outa, ins));  in rsdIntrinsicBlendSrcAtop_K()
1115         t1 = _mm_srli_epi16(t1, 8);  in rsdIntrinsicBlendSrcAtop_K()
1139         t0 = _mm_packus_epi16(t0, t1);  in rsdIntrinsicBlendSrcAtop_K()
1155     __m128i t0, t1, t2, t3;  in rsdIntrinsicBlendDstAtop_K()  local
1183         t1 = _mm_sub_epi16(all1s, outa);  in rsdIntrinsicBlendDstAtop_K()
1184         t1 = _mm_mullo_epi16(t1, ins);  in rsdIntrinsicBlendDstAtop_K()
1185         t1 = _mm_adds_epu16(t1, _mm_mullo_epi16(ina, outs));  in rsdIntrinsicBlendDstAtop_K()
1186         t1 = _mm_srli_epi16(t1, 8);  in rsdIntrinsicBlendDstAtop_K()
1210         t0 = _mm_packus_epi16(t0, t1);  in rsdIntrinsicBlendDstAtop_K()
1245     __m128i t0, t1, t2, t3;  in rsdIntrinsicBlendMultiply_K()  local
1258         t1 = _mm_unpackhi_epi8(in0, _mm_setzero_si128());  in rsdIntrinsicBlendMultiply_K()
1259         t1 = _mm_mullo_epi16(t1, _mm_unpackhi_epi8(out0, _mm_setzero_si128()));  in rsdIntrinsicBlendMultiply_K()
1260         t1 = _mm_srli_epi16(t1, 8);  in rsdIntrinsicBlendMultiply_K()
1270         t0 = _mm_packus_epi16(t0, t1);  in rsdIntrinsicBlendMultiply_K()