817     __m128i all1s, outa, outs;  in rsdIntrinsicBlendDstOver_K()  local
831         outs = _mm_unpacklo_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendDstOver_K()
832         outa = _mm_shufflelo_epi16(outs, 0xFF);  in rsdIntrinsicBlendDstOver_K()
837         t0 = _mm_add_epi16(t0, outs);  in rsdIntrinsicBlendDstOver_K()
839         outs = _mm_unpackhi_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendDstOver_K()
840         outa = _mm_shufflelo_epi16(outs, 0xFF);  in rsdIntrinsicBlendDstOver_K()
845         t1 = _mm_add_epi16(t1, outs);  in rsdIntrinsicBlendDstOver_K()
847         outs = _mm_unpacklo_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendDstOver_K()
848         outa = _mm_shufflelo_epi16(outs, 0xFF);  in rsdIntrinsicBlendDstOver_K()
853         t2 = _mm_add_epi16(t2, outs);  in rsdIntrinsicBlendDstOver_K()
855         outs = _mm_unpackhi_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendDstOver_K()
856         outa = _mm_shufflelo_epi16(outs, 0xFF);  in rsdIntrinsicBlendDstOver_K()
861         t3 = _mm_add_epi16(t3, outs);  in rsdIntrinsicBlendDstOver_K()
1079     __m128i all1s, ina, outa, ins, outs;  in rsdIntrinsicBlendSrcAtop_K()  local
1095         outs = _mm_unpacklo_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendSrcAtop_K()
1096         outa = _mm_shufflelo_epi16(outs, 0xFF);  in rsdIntrinsicBlendSrcAtop_K()
1099         t0 = _mm_mullo_epi16(t0, outs);  in rsdIntrinsicBlendSrcAtop_K()
1106         outs = _mm_unpackhi_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendSrcAtop_K()
1107         outa = _mm_shufflelo_epi16(outs, 0xFF);  in rsdIntrinsicBlendSrcAtop_K()
1110         t1 = _mm_mullo_epi16(t1, outs);  in rsdIntrinsicBlendSrcAtop_K()
1117         outs = _mm_unpacklo_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcAtop_K()
1118         outa = _mm_shufflelo_epi16(outs, 0xFF);  in rsdIntrinsicBlendSrcAtop_K()
1121         t2 = _mm_mullo_epi16(t2, outs);  in rsdIntrinsicBlendSrcAtop_K()
1128         outs = _mm_unpackhi_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcAtop_K()
1129         outa = _mm_shufflelo_epi16(outs, 0xFF);  in rsdIntrinsicBlendSrcAtop_K()
1132         t3 = _mm_mullo_epi16(t3, outs);  in rsdIntrinsicBlendSrcAtop_K()
1150     __m128i all1s, ina, ins, outa, outs;  in rsdIntrinsicBlendDstAtop_K()  local
1166         outs = _mm_unpacklo_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendDstAtop_K()
1167         outa = _mm_shufflelo_epi16(outs, 0xFF);  in rsdIntrinsicBlendDstAtop_K()
1171         t0 = _mm_adds_epu16(t0, _mm_mullo_epi16(ina, outs));  in rsdIntrinsicBlendDstAtop_K()
1177         outs = _mm_unpackhi_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendDstAtop_K()
1178         outa = _mm_shufflelo_epi16(outs, 0xFF);  in rsdIntrinsicBlendDstAtop_K()
1182         t1 = _mm_adds_epu16(t1, _mm_mullo_epi16(ina, outs));  in rsdIntrinsicBlendDstAtop_K()
1188         outs = _mm_unpacklo_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendDstAtop_K()
1189         outa = _mm_shufflelo_epi16(outs, 0xFF);  in rsdIntrinsicBlendDstAtop_K()
1193         t2 = _mm_adds_epu16(t2, _mm_mullo_epi16(ina, outs));  in rsdIntrinsicBlendDstAtop_K()
1199         outs = _mm_unpackhi_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendDstAtop_K()
1200         outa = _mm_shufflelo_epi16(outs, 0xFF);  in rsdIntrinsicBlendDstAtop_K()
1204         t3 = _mm_adds_epu16(t3, _mm_mullo_epi16(ina, outs));  in rsdIntrinsicBlendDstAtop_K()