761     __m128i all1s, ina, ins;  in rsdIntrinsicBlendSrcOver_K()  local
774         ins = _mm_unpacklo_epi8(in0, _mm_setzero_si128());  in rsdIntrinsicBlendSrcOver_K()
775         ina = _mm_shufflelo_epi16(ins, 0xFF);  in rsdIntrinsicBlendSrcOver_K()
780         t0 = _mm_add_epi16(t0, ins);  in rsdIntrinsicBlendSrcOver_K()
782         ins = _mm_unpackhi_epi8(in0, _mm_setzero_si128());  in rsdIntrinsicBlendSrcOver_K()
783         ina = _mm_shufflelo_epi16(ins, 0xFF);  in rsdIntrinsicBlendSrcOver_K()
788         t1 = _mm_add_epi16(t1, ins);  in rsdIntrinsicBlendSrcOver_K()
790         ins = _mm_unpacklo_epi8(in1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcOver_K()
791         ina = _mm_shufflelo_epi16(ins, 0xFF);  in rsdIntrinsicBlendSrcOver_K()
796         t2 = _mm_add_epi16(t2, ins);  in rsdIntrinsicBlendSrcOver_K()
798         ins = _mm_unpackhi_epi8(in1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcOver_K()
799         ina = _mm_shufflelo_epi16(ins, 0xFF);  in rsdIntrinsicBlendSrcOver_K()
804         t3 = _mm_add_epi16(t3, ins);  in rsdIntrinsicBlendSrcOver_K()
1079     __m128i all1s, ina, outa, ins, outs;  in rsdIntrinsicBlendSrcAtop_K()  local
1092         ins = _mm_unpacklo_epi8(in0, _mm_setzero_si128());  in rsdIntrinsicBlendSrcAtop_K()
1093         ina = _mm_shufflelo_epi16(ins, 0xFF);  in rsdIntrinsicBlendSrcAtop_K()
1100         t0 = _mm_adds_epu16(t0, _mm_mullo_epi16(outa, ins));  in rsdIntrinsicBlendSrcAtop_K()
1103         ins = _mm_unpackhi_epi8(in0, _mm_setzero_si128());  in rsdIntrinsicBlendSrcAtop_K()
1104         ina = _mm_shufflelo_epi16(ins, 0xFF);  in rsdIntrinsicBlendSrcAtop_K()
1111         t1 = _mm_adds_epu16(t1, _mm_mullo_epi16(outa, ins));  in rsdIntrinsicBlendSrcAtop_K()
1114         ins = _mm_unpacklo_epi8(in1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcAtop_K()
1115         ina = _mm_shufflelo_epi16(ins, 0xFF);  in rsdIntrinsicBlendSrcAtop_K()
1122         t2 = _mm_adds_epu16(t2, _mm_mullo_epi16(outa, ins));  in rsdIntrinsicBlendSrcAtop_K()
1125         ins = _mm_unpackhi_epi8(in1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcAtop_K()
1126         ina = _mm_shufflelo_epi16(ins, 0xFF);  in rsdIntrinsicBlendSrcAtop_K()
1133         t3 = _mm_adds_epu16(t3, _mm_mullo_epi16(outa, ins));  in rsdIntrinsicBlendSrcAtop_K()
1150     __m128i all1s, ina, ins, outa, outs;  in rsdIntrinsicBlendDstAtop_K()  local
1163         ins = _mm_unpacklo_epi8(in0, _mm_setzero_si128());  in rsdIntrinsicBlendDstAtop_K()
1164         ina = _mm_shufflelo_epi16(ins, 0xFF);  in rsdIntrinsicBlendDstAtop_K()
1170         t0 = _mm_mullo_epi16(t0, ins);  in rsdIntrinsicBlendDstAtop_K()
1174         ins = _mm_unpackhi_epi8(in0, _mm_setzero_si128());  in rsdIntrinsicBlendDstAtop_K()
1175         ina = _mm_shufflelo_epi16(ins, 0xFF);  in rsdIntrinsicBlendDstAtop_K()
1181         t1 = _mm_mullo_epi16(t1, ins);  in rsdIntrinsicBlendDstAtop_K()
1185         ins = _mm_unpacklo_epi8(in1, _mm_setzero_si128());  in rsdIntrinsicBlendDstAtop_K()
1186         ina = _mm_shufflelo_epi16(ins, 0xFF);  in rsdIntrinsicBlendDstAtop_K()
1192         t2 = _mm_mullo_epi16(t2, ins);  in rsdIntrinsicBlendDstAtop_K()
1196         ins = _mm_unpackhi_epi8(in1, _mm_setzero_si128());  in rsdIntrinsicBlendDstAtop_K()
1197         ina = _mm_shufflelo_epi16(ins, 0xFF);  in rsdIntrinsicBlendDstAtop_K()
1203         t3 = _mm_mullo_epi16(t3, ins);  in rsdIntrinsicBlendDstAtop_K()