816     __m128i all1s, outa, outs;  in rsdIntrinsicBlendDstOver_K()  local
830         outs = _mm_unpacklo_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendDstOver_K()
831         outa = _mm_shufflelo_epi16(outs, 0xFF);  in rsdIntrinsicBlendDstOver_K()
836         t0 = _mm_add_epi16(t0, outs);  in rsdIntrinsicBlendDstOver_K()
838         outs = _mm_unpackhi_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendDstOver_K()
839         outa = _mm_shufflelo_epi16(outs, 0xFF);  in rsdIntrinsicBlendDstOver_K()
844         t1 = _mm_add_epi16(t1, outs);  in rsdIntrinsicBlendDstOver_K()
846         outs = _mm_unpacklo_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendDstOver_K()
847         outa = _mm_shufflelo_epi16(outs, 0xFF);  in rsdIntrinsicBlendDstOver_K()
852         t2 = _mm_add_epi16(t2, outs);  in rsdIntrinsicBlendDstOver_K()
854         outs = _mm_unpackhi_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendDstOver_K()
855         outa = _mm_shufflelo_epi16(outs, 0xFF);  in rsdIntrinsicBlendDstOver_K()
860         t3 = _mm_add_epi16(t3, outs);  in rsdIntrinsicBlendDstOver_K()
1078     __m128i all1s, ina, outa, ins, outs;  in rsdIntrinsicBlendSrcAtop_K()  local
1094         outs = _mm_unpacklo_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendSrcAtop_K()
1095         outa = _mm_shufflelo_epi16(outs, 0xFF);  in rsdIntrinsicBlendSrcAtop_K()
1098         t0 = _mm_mullo_epi16(t0, outs);  in rsdIntrinsicBlendSrcAtop_K()
1105         outs = _mm_unpackhi_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendSrcAtop_K()
1106         outa = _mm_shufflelo_epi16(outs, 0xFF);  in rsdIntrinsicBlendSrcAtop_K()
1109         t1 = _mm_mullo_epi16(t1, outs);  in rsdIntrinsicBlendSrcAtop_K()
1116         outs = _mm_unpacklo_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcAtop_K()
1117         outa = _mm_shufflelo_epi16(outs, 0xFF);  in rsdIntrinsicBlendSrcAtop_K()
1120         t2 = _mm_mullo_epi16(t2, outs);  in rsdIntrinsicBlendSrcAtop_K()
1127         outs = _mm_unpackhi_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcAtop_K()
1128         outa = _mm_shufflelo_epi16(outs, 0xFF);  in rsdIntrinsicBlendSrcAtop_K()
1131         t3 = _mm_mullo_epi16(t3, outs);  in rsdIntrinsicBlendSrcAtop_K()
1149     __m128i all1s, ina, ins, outa, outs;  in rsdIntrinsicBlendDstAtop_K()  local
1165         outs = _mm_unpacklo_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendDstAtop_K()
1166         outa = _mm_shufflelo_epi16(outs, 0xFF);  in rsdIntrinsicBlendDstAtop_K()
1170         t0 = _mm_adds_epu16(t0, _mm_mullo_epi16(ina, outs));  in rsdIntrinsicBlendDstAtop_K()
1176         outs = _mm_unpackhi_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendDstAtop_K()
1177         outa = _mm_shufflelo_epi16(outs, 0xFF);  in rsdIntrinsicBlendDstAtop_K()
1181         t1 = _mm_adds_epu16(t1, _mm_mullo_epi16(ina, outs));  in rsdIntrinsicBlendDstAtop_K()
1187         outs = _mm_unpacklo_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendDstAtop_K()
1188         outa = _mm_shufflelo_epi16(outs, 0xFF);  in rsdIntrinsicBlendDstAtop_K()
1192         t2 = _mm_adds_epu16(t2, _mm_mullo_epi16(ina, outs));  in rsdIntrinsicBlendDstAtop_K()
1198         outs = _mm_unpackhi_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendDstAtop_K()
1199         outa = _mm_shufflelo_epi16(outs, 0xFF);  in rsdIntrinsicBlendDstAtop_K()
1203         t3 = _mm_adds_epu16(t3, _mm_mullo_epi16(ina, outs));  in rsdIntrinsicBlendDstAtop_K()