763     __m128i t0, t1, t2, t3;  in rsdIntrinsicBlendSrcOver_K()  local
801         t3 = _mm_unpackhi_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcOver_K()
802         t3 = _mm_mullo_epi16(t3, _mm_sub_epi16(all1s, ina));  in rsdIntrinsicBlendSrcOver_K()
803         t3 = _mm_srli_epi16(t3, 8);  in rsdIntrinsicBlendSrcOver_K()
804         t3 = _mm_add_epi16(t3, ins);  in rsdIntrinsicBlendSrcOver_K()
807         t2 = _mm_packus_epi16(t2, t3);  in rsdIntrinsicBlendSrcOver_K()
819     __m128i t0, t1, t2, t3;  in rsdIntrinsicBlendDstOver_K()  local
858         t3 = _mm_unpackhi_epi8(in1, _mm_setzero_si128());  in rsdIntrinsicBlendDstOver_K()
859         t3 = _mm_mullo_epi16(t3, _mm_sub_epi16(all1s, outa));  in rsdIntrinsicBlendDstOver_K()
860         t3 = _mm_srli_epi16(t3, 8);  in rsdIntrinsicBlendDstOver_K()
861         t3 = _mm_add_epi16(t3, outs);  in rsdIntrinsicBlendDstOver_K()
864         t2 = _mm_packus_epi16(t2, t3);  in rsdIntrinsicBlendDstOver_K()
876     __m128i t0, t1, t2, t3;  in rsdIntrinsicBlendSrcIn_K()  local
909         t3 = _mm_unpackhi_epi8(in1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcIn_K()
910         t3 = _mm_mullo_epi16(t3, outa);  in rsdIntrinsicBlendSrcIn_K()
911         t3 = _mm_srli_epi16(t3, 8);  in rsdIntrinsicBlendSrcIn_K()
914         t2 = _mm_packus_epi16(t2, t3);  in rsdIntrinsicBlendSrcIn_K()
926     __m128i t0, t1, t2, t3;  in rsdIntrinsicBlendDstIn_K()  local
959         t3 = _mm_unpackhi_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendDstIn_K()
960         t3 = _mm_mullo_epi16(t3, ina);  in rsdIntrinsicBlendDstIn_K()
961         t3 = _mm_srli_epi16(t3, 8);  in rsdIntrinsicBlendDstIn_K()
964         t2 = _mm_packus_epi16(t2, t3);  in rsdIntrinsicBlendDstIn_K()
976     __m128i t0, t1, t2, t3;  in rsdIntrinsicBlendSrcOut_K()  local
1011         t3 = _mm_unpackhi_epi8(in1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcOut_K()
1012         t3 = _mm_mullo_epi16(t3, _mm_sub_epi16(all1s, outa));  in rsdIntrinsicBlendSrcOut_K()
1013         t3 = _mm_srli_epi16(t3, 8);  in rsdIntrinsicBlendSrcOut_K()
1016         t2 = _mm_packus_epi16(t2, t3);  in rsdIntrinsicBlendSrcOut_K()
1028     __m128i t0, t1, t2, t3;  in rsdIntrinsicBlendDstOut_K()  local
1063         t3 = _mm_unpackhi_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendDstOut_K()
1064         t3 = _mm_mullo_epi16(t3, _mm_sub_epi16(all1s, ina));  in rsdIntrinsicBlendDstOut_K()
1065         t3 = _mm_srli_epi16(t3, 8);  in rsdIntrinsicBlendDstOut_K()
1068         t2 = _mm_packus_epi16(t2, t3);  in rsdIntrinsicBlendDstOut_K()
1081     __m128i t0, t1, t2, t3;  in rsdIntrinsicBlendSrcAtop_K()  local
1131         t3 = _mm_sub_epi16(all1s, ina);  in rsdIntrinsicBlendSrcAtop_K()
1132         t3 = _mm_mullo_epi16(t3, outs);  in rsdIntrinsicBlendSrcAtop_K()
1133         t3 = _mm_adds_epu16(t3, _mm_mullo_epi16(outa, ins));  in rsdIntrinsicBlendSrcAtop_K()
1134         t3 = _mm_srli_epi16(t3, 8);  in rsdIntrinsicBlendSrcAtop_K()
1138         t2 = _mm_packus_epi16(t2, t3);  in rsdIntrinsicBlendSrcAtop_K()
1152     __m128i t0, t1, t2, t3;  in rsdIntrinsicBlendDstAtop_K()  local
1202         t3 = _mm_sub_epi16(all1s, outa);  in rsdIntrinsicBlendDstAtop_K()
1203         t3 = _mm_mullo_epi16(t3, ins);  in rsdIntrinsicBlendDstAtop_K()
1204         t3 = _mm_adds_epu16(t3, _mm_mullo_epi16(ina, outs));  in rsdIntrinsicBlendDstAtop_K()
1205         t3 = _mm_srli_epi16(t3, 8);  in rsdIntrinsicBlendDstAtop_K()
1209         t2 = _mm_packus_epi16(t2, t3);  in rsdIntrinsicBlendDstAtop_K()
1242     __m128i t0, t1, t2, t3;  in rsdIntrinsicBlendMultiply_K()  local
1263         t3 = _mm_unpackhi_epi8(in1, _mm_setzero_si128());  in rsdIntrinsicBlendMultiply_K()
1264         t3 = _mm_mullo_epi16(t3, _mm_unpackhi_epi8(out1, _mm_setzero_si128()));  in rsdIntrinsicBlendMultiply_K()
1265         t3 = _mm_srli_epi16(t3, 8);  in rsdIntrinsicBlendMultiply_K()
1268         t2 = _mm_packus_epi16(t2, t3);  in rsdIntrinsicBlendMultiply_K()