763     __m128i t0, t1, t2, t3;  in rsdIntrinsicBlendSrcOver_K()  local
785         t1 = _mm_unpackhi_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendSrcOver_K()
786         t1 = _mm_mullo_epi16(t1, _mm_sub_epi16(all1s, ina));  in rsdIntrinsicBlendSrcOver_K()
787         t1 = _mm_srli_epi16(t1, 8);  in rsdIntrinsicBlendSrcOver_K()
788         t1 = _mm_add_epi16(t1, ins);  in rsdIntrinsicBlendSrcOver_K()
806         t0 = _mm_packus_epi16(t0, t1);  in rsdIntrinsicBlendSrcOver_K()
819     __m128i t0, t1, t2, t3;  in rsdIntrinsicBlendDstOver_K()  local
842         t1 = _mm_unpackhi_epi8(in0, _mm_setzero_si128());  in rsdIntrinsicBlendDstOver_K()
843         t1 = _mm_mullo_epi16(t1, _mm_sub_epi16(all1s, outa));  in rsdIntrinsicBlendDstOver_K()
844         t1 = _mm_srli_epi16(t1, 8);  in rsdIntrinsicBlendDstOver_K()
845         t1 = _mm_add_epi16(t1, outs);  in rsdIntrinsicBlendDstOver_K()
863         t0 = _mm_packus_epi16(t0, t1);  in rsdIntrinsicBlendDstOver_K()
876     __m128i t0, t1, t2, t3;  in rsdIntrinsicBlendSrcIn_K()  local
895         t1 = _mm_unpackhi_epi8(in0, _mm_setzero_si128());  in rsdIntrinsicBlendSrcIn_K()
896         t1 = _mm_mullo_epi16(t1, outa);  in rsdIntrinsicBlendSrcIn_K()
897         t1 = _mm_srli_epi16(t1, 8);  in rsdIntrinsicBlendSrcIn_K()
913         t0 = _mm_packus_epi16(t0, t1);  in rsdIntrinsicBlendSrcIn_K()
926     __m128i t0, t1, t2, t3;  in rsdIntrinsicBlendDstIn_K()  local
945         t1 = _mm_unpackhi_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendDstIn_K()
946         t1 = _mm_mullo_epi16(t1, ina);  in rsdIntrinsicBlendDstIn_K()
947         t1 = _mm_srli_epi16(t1, 8);  in rsdIntrinsicBlendDstIn_K()
963         t0 = _mm_packus_epi16(t0, t1);  in rsdIntrinsicBlendDstIn_K()
976     __m128i t0, t1, t2, t3;  in rsdIntrinsicBlendSrcOut_K()  local
997         t1 = _mm_unpackhi_epi8(in0, _mm_setzero_si128());  in rsdIntrinsicBlendSrcOut_K()
998         t1 = _mm_mullo_epi16(t1, _mm_sub_epi16(all1s, outa));  in rsdIntrinsicBlendSrcOut_K()
999         t1 = _mm_srli_epi16(t1, 8);  in rsdIntrinsicBlendSrcOut_K()
1015         t0 = _mm_packus_epi16(t0, t1);  in rsdIntrinsicBlendSrcOut_K()
1028     __m128i t0, t1, t2, t3;  in rsdIntrinsicBlendDstOut_K()  local
1049         t1 = _mm_unpackhi_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendDstOut_K()
1050         t1 = _mm_mullo_epi16(t1, _mm_sub_epi16(all1s, ina));  in rsdIntrinsicBlendDstOut_K()
1051         t1 = _mm_srli_epi16(t1, 8);  in rsdIntrinsicBlendDstOut_K()
1067         t0 = _mm_packus_epi16(t0, t1);  in rsdIntrinsicBlendDstOut_K()
1081     __m128i t0, t1, t2, t3;  in rsdIntrinsicBlendSrcAtop_K()  local
1109         t1 = _mm_sub_epi16(all1s, ina);  in rsdIntrinsicBlendSrcAtop_K()
1110         t1 = _mm_mullo_epi16(t1, outs);  in rsdIntrinsicBlendSrcAtop_K()
1111         t1 = _mm_adds_epu16(t1, _mm_mullo_epi16(outa, ins));  in rsdIntrinsicBlendSrcAtop_K()
1112         t1 = _mm_srli_epi16(t1, 8);  in rsdIntrinsicBlendSrcAtop_K()
1136         t0 = _mm_packus_epi16(t0, t1);  in rsdIntrinsicBlendSrcAtop_K()
1152     __m128i t0, t1, t2, t3;  in rsdIntrinsicBlendDstAtop_K()  local
1180         t1 = _mm_sub_epi16(all1s, outa);  in rsdIntrinsicBlendDstAtop_K()
1181         t1 = _mm_mullo_epi16(t1, ins);  in rsdIntrinsicBlendDstAtop_K()
1182         t1 = _mm_adds_epu16(t1, _mm_mullo_epi16(ina, outs));  in rsdIntrinsicBlendDstAtop_K()
1183         t1 = _mm_srli_epi16(t1, 8);  in rsdIntrinsicBlendDstAtop_K()
1207         t0 = _mm_packus_epi16(t0, t1);  in rsdIntrinsicBlendDstAtop_K()
1242     __m128i t0, t1, t2, t3;  in rsdIntrinsicBlendMultiply_K()  local
1255         t1 = _mm_unpackhi_epi8(in0, _mm_setzero_si128());  in rsdIntrinsicBlendMultiply_K()
1256         t1 = _mm_mullo_epi16(t1, _mm_unpackhi_epi8(out0, _mm_setzero_si128()));  in rsdIntrinsicBlendMultiply_K()
1257         t1 = _mm_srli_epi16(t1, 8);  in rsdIntrinsicBlendMultiply_K()
1267         t0 = _mm_packus_epi16(t0, t1);  in rsdIntrinsicBlendMultiply_K()