761     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendSrcOver_K()  local
771         out1 = _mm_loadu_si128((const __m128i *)dst + 1);  in rsdIntrinsicBlendSrcOver_K()
792         t2 = _mm_unpacklo_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcOver_K()
800         t3 = _mm_unpackhi_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcOver_K()
817     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendDstOver_K()  local
827         out1 = _mm_loadu_si128((const __m128i *)dst + 1);  in rsdIntrinsicBlendDstOver_K()
846         outs = _mm_unpacklo_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendDstOver_K()
854         outs = _mm_unpackhi_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendDstOver_K()
874     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendSrcIn_K()  local
882         out1 = _mm_loadu_si128((const __m128i *)dst + 1);  in rsdIntrinsicBlendSrcIn_K()
898         outa = _mm_unpacklo_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcIn_K()
905         outa = _mm_unpackhi_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcIn_K()
924     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendDstIn_K()  local
932         out1 = _mm_loadu_si128((const __m128i *)dst + 1);  in rsdIntrinsicBlendDstIn_K()
951         t2 = _mm_unpacklo_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendDstIn_K()
958         t3 = _mm_unpackhi_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendDstIn_K()
974     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendSrcOut_K()  local
984         out1 = _mm_loadu_si128((const __m128i *)dst + 1);  in rsdIntrinsicBlendSrcOut_K()
1000         outa = _mm_unpacklo_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcOut_K()
1007         outa = _mm_unpackhi_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcOut_K()
1026     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendDstOut_K()  local
1036         out1 = _mm_loadu_si128((const __m128i *)dst + 1);  in rsdIntrinsicBlendDstOut_K()
1055         t2 = _mm_unpacklo_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendDstOut_K()
1062         t3 = _mm_unpackhi_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendDstOut_K()
1079     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendSrcAtop_K()  local
1089         out1 = _mm_loadu_si128((const __m128i *)dst + 1);  in rsdIntrinsicBlendSrcAtop_K()
1116         outs = _mm_unpacklo_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcAtop_K()
1127         outs = _mm_unpackhi_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcAtop_K()
1138         t2 = blendv_epi8(t2, out1, M0001);  in rsdIntrinsicBlendSrcAtop_K()
1150     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendDstAtop_K()  local
1160         out1 = _mm_loadu_si128((const __m128i *)dst + 1);  in rsdIntrinsicBlendDstAtop_K()
1187         outs = _mm_unpacklo_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendDstAtop_K()
1198         outs = _mm_unpackhi_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendDstAtop_K()
1209         t2 = blendv_epi8(t2, out1, M0001);  in rsdIntrinsicBlendDstAtop_K()
1219     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendXor_K()  local
1226         out1 = _mm_loadu_si128((const __m128i *)dst + 1);  in rsdIntrinsicBlendXor_K()
1229         out1 = _mm_xor_si128(out1, in1);  in rsdIntrinsicBlendXor_K()
1232         _mm_storeu_si128((__m128i *)dst + 1, out1);  in rsdIntrinsicBlendXor_K()
1240     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendMultiply_K()  local
1248         out1 = _mm_loadu_si128((const __m128i *)dst + 1);  in rsdIntrinsicBlendMultiply_K()
1259         t2 = _mm_mullo_epi16(t2, _mm_unpacklo_epi8(out1, _mm_setzero_si128()));  in rsdIntrinsicBlendMultiply_K()
1263         t3 = _mm_mullo_epi16(t3, _mm_unpackhi_epi8(out1, _mm_setzero_si128()));  in rsdIntrinsicBlendMultiply_K()
1277     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendAdd_K()  local
1284         out1 = _mm_loadu_si128((const __m128i *)dst + 1);  in rsdIntrinsicBlendAdd_K()
1287         out1 = _mm_adds_epu8(out1, in1);  in rsdIntrinsicBlendAdd_K()
1290         _mm_storeu_si128((__m128i *)dst + 1, out1);  in rsdIntrinsicBlendAdd_K()
1298     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendSub_K()  local
1305         out1 = _mm_loadu_si128((const __m128i *)dst + 1);  in rsdIntrinsicBlendSub_K()
1308         out1 = _mm_subs_epu8(out1, in1);  in rsdIntrinsicBlendSub_K()
1311         _mm_storeu_si128((__m128i *)dst + 1, out1);  in rsdIntrinsicBlendSub_K()