820     __m128i all1s, outa, outs;  in rsdIntrinsicBlendDstOver_K()  local
834         outs = _mm_unpacklo_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendDstOver_K()
835         outa = _mm_shufflelo_epi16(outs, 0xFF);  in rsdIntrinsicBlendDstOver_K()
840         t0 = _mm_add_epi16(t0, outs);  in rsdIntrinsicBlendDstOver_K()
842         outs = _mm_unpackhi_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendDstOver_K()
843         outa = _mm_shufflelo_epi16(outs, 0xFF);  in rsdIntrinsicBlendDstOver_K()
848         t1 = _mm_add_epi16(t1, outs);  in rsdIntrinsicBlendDstOver_K()
850         outs = _mm_unpacklo_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendDstOver_K()
851         outa = _mm_shufflelo_epi16(outs, 0xFF);  in rsdIntrinsicBlendDstOver_K()
856         t2 = _mm_add_epi16(t2, outs);  in rsdIntrinsicBlendDstOver_K()
858         outs = _mm_unpackhi_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendDstOver_K()
859         outa = _mm_shufflelo_epi16(outs, 0xFF);  in rsdIntrinsicBlendDstOver_K()
864         t3 = _mm_add_epi16(t3, outs);  in rsdIntrinsicBlendDstOver_K()
1082     __m128i all1s, ina, outa, ins, outs;  in rsdIntrinsicBlendSrcAtop_K()  local
1098         outs = _mm_unpacklo_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendSrcAtop_K()
1099         outa = _mm_shufflelo_epi16(outs, 0xFF);  in rsdIntrinsicBlendSrcAtop_K()
1102         t0 = _mm_mullo_epi16(t0, outs);  in rsdIntrinsicBlendSrcAtop_K()
1109         outs = _mm_unpackhi_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendSrcAtop_K()
1110         outa = _mm_shufflelo_epi16(outs, 0xFF);  in rsdIntrinsicBlendSrcAtop_K()
1113         t1 = _mm_mullo_epi16(t1, outs);  in rsdIntrinsicBlendSrcAtop_K()
1120         outs = _mm_unpacklo_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcAtop_K()
1121         outa = _mm_shufflelo_epi16(outs, 0xFF);  in rsdIntrinsicBlendSrcAtop_K()
1124         t2 = _mm_mullo_epi16(t2, outs);  in rsdIntrinsicBlendSrcAtop_K()
1131         outs = _mm_unpackhi_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcAtop_K()
1132         outa = _mm_shufflelo_epi16(outs, 0xFF);  in rsdIntrinsicBlendSrcAtop_K()
1135         t3 = _mm_mullo_epi16(t3, outs);  in rsdIntrinsicBlendSrcAtop_K()
1153     __m128i all1s, ina, ins, outa, outs;  in rsdIntrinsicBlendDstAtop_K()  local
1169         outs = _mm_unpacklo_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendDstAtop_K()
1170         outa = _mm_shufflelo_epi16(outs, 0xFF);  in rsdIntrinsicBlendDstAtop_K()
1174         t0 = _mm_adds_epu16(t0, _mm_mullo_epi16(ina, outs));  in rsdIntrinsicBlendDstAtop_K()
1180         outs = _mm_unpackhi_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendDstAtop_K()
1181         outa = _mm_shufflelo_epi16(outs, 0xFF);  in rsdIntrinsicBlendDstAtop_K()
1185         t1 = _mm_adds_epu16(t1, _mm_mullo_epi16(ina, outs));  in rsdIntrinsicBlendDstAtop_K()
1191         outs = _mm_unpacklo_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendDstAtop_K()
1192         outa = _mm_shufflelo_epi16(outs, 0xFF);  in rsdIntrinsicBlendDstAtop_K()
1196         t2 = _mm_adds_epu16(t2, _mm_mullo_epi16(ina, outs));  in rsdIntrinsicBlendDstAtop_K()
1202         outs = _mm_unpackhi_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendDstAtop_K()
1203         outa = _mm_shufflelo_epi16(outs, 0xFF);  in rsdIntrinsicBlendDstAtop_K()
1207         t3 = _mm_adds_epu16(t3, _mm_mullo_epi16(ina, outs));  in rsdIntrinsicBlendDstAtop_K()