762     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendSrcOver_K()  local
771         out0 = _mm_loadu_si128((const __m128i *)dst);  in rsdIntrinsicBlendSrcOver_K()
777         t0 = _mm_unpacklo_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendSrcOver_K()
785         t1 = _mm_unpackhi_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendSrcOver_K()
818     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendDstOver_K()  local
827         out0 = _mm_loadu_si128((const __m128i *)dst);  in rsdIntrinsicBlendDstOver_K()
831         outs = _mm_unpacklo_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendDstOver_K()
839         outs = _mm_unpackhi_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendDstOver_K()
875     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendSrcIn_K()  local
882         out0 = _mm_loadu_si128((const __m128i *)dst);  in rsdIntrinsicBlendSrcIn_K()
885         outa = _mm_unpacklo_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendSrcIn_K()
892         outa = _mm_unpackhi_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendSrcIn_K()
925     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendDstIn_K()  local
932         out0 = _mm_loadu_si128((const __m128i *)dst);  in rsdIntrinsicBlendDstIn_K()
938         t0 = _mm_unpacklo_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendDstIn_K()
945         t1 = _mm_unpackhi_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendDstIn_K()
975     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendSrcOut_K()  local
984         out0 = _mm_loadu_si128((const __m128i *)dst);  in rsdIntrinsicBlendSrcOut_K()
987         outa = _mm_unpacklo_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendSrcOut_K()
994         outa = _mm_unpackhi_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendSrcOut_K()
1027     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendDstOut_K()  local
1036         out0 = _mm_loadu_si128((const __m128i *)dst);  in rsdIntrinsicBlendDstOut_K()
1042         t0 = _mm_unpacklo_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendDstOut_K()
1049         t1 = _mm_unpackhi_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendDstOut_K()
1080     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendSrcAtop_K()  local
1089         out0 = _mm_loadu_si128((const __m128i *)dst);  in rsdIntrinsicBlendSrcAtop_K()
1095         outs = _mm_unpacklo_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendSrcAtop_K()
1106         outs = _mm_unpackhi_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendSrcAtop_K()
1137         t0 = blendv_epi8(t0, out0, M0001);  in rsdIntrinsicBlendSrcAtop_K()
1151     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendDstAtop_K()  local
1160         out0 = _mm_loadu_si128((const __m128i *)dst);  in rsdIntrinsicBlendDstAtop_K()
1166         outs = _mm_unpacklo_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendDstAtop_K()
1177         outs = _mm_unpackhi_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendDstAtop_K()
1220     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendXor_K()  local
1226         out0 = _mm_loadu_si128((const __m128i *)dst);  in rsdIntrinsicBlendXor_K()
1229         out0 = _mm_xor_si128(out0, in0);  in rsdIntrinsicBlendXor_K()
1232         _mm_storeu_si128((__m128i *)dst, out0);  in rsdIntrinsicBlendXor_K()
1241     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendMultiply_K()  local
1248         out0 = _mm_loadu_si128((const __m128i *)dst);  in rsdIntrinsicBlendMultiply_K()
1252         t0 = _mm_mullo_epi16(t0, _mm_unpacklo_epi8(out0, _mm_setzero_si128()));  in rsdIntrinsicBlendMultiply_K()
1256         t1 = _mm_mullo_epi16(t1, _mm_unpackhi_epi8(out0, _mm_setzero_si128()));  in rsdIntrinsicBlendMultiply_K()
1278     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendAdd_K()  local
1284         out0 = _mm_loadu_si128((const __m128i *)dst);  in rsdIntrinsicBlendAdd_K()
1287         out0 = _mm_adds_epu8(out0, in0);  in rsdIntrinsicBlendAdd_K()
1290         _mm_storeu_si128((__m128i *)dst, out0);  in rsdIntrinsicBlendAdd_K()
1299     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendSub_K()  local
1305         out0 = _mm_loadu_si128((const __m128i *)dst);  in rsdIntrinsicBlendSub_K()
1308         out0 = _mm_subs_epu8(out0, in0);  in rsdIntrinsicBlendSub_K()
1311         _mm_storeu_si128((__m128i *)dst, out0);  in rsdIntrinsicBlendSub_K()