765     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendSrcOver_K()  local
772         in0 = _mm_loadu_si128((const __m128i *)src);  in rsdIntrinsicBlendSrcOver_K()
777         ins = _mm_unpacklo_epi8(in0, _mm_setzero_si128());  in rsdIntrinsicBlendSrcOver_K()
785         ins = _mm_unpackhi_epi8(in0, _mm_setzero_si128());  in rsdIntrinsicBlendSrcOver_K()
821     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendDstOver_K()  local
828         in0 = _mm_loadu_si128((const __m128i *)src);  in rsdIntrinsicBlendDstOver_K()
837         t0 = _mm_unpacklo_epi8(in0, _mm_setzero_si128());  in rsdIntrinsicBlendDstOver_K()
845         t1 = _mm_unpackhi_epi8(in0, _mm_setzero_si128());  in rsdIntrinsicBlendDstOver_K()
878     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendSrcIn_K()  local
883         in0 = _mm_loadu_si128((const __m128i *)src);  in rsdIntrinsicBlendSrcIn_K()
891         t0 = _mm_unpacklo_epi8(in0, _mm_setzero_si128());  in rsdIntrinsicBlendSrcIn_K()
898         t1 = _mm_unpackhi_epi8(in0, _mm_setzero_si128());  in rsdIntrinsicBlendSrcIn_K()
928     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendDstIn_K()  local
933         in0 = _mm_loadu_si128((const __m128i *)src);  in rsdIntrinsicBlendDstIn_K()
938         ina = _mm_unpacklo_epi8(in0, _mm_setzero_si128());  in rsdIntrinsicBlendDstIn_K()
945         ina = _mm_unpackhi_epi8(in0, _mm_setzero_si128());  in rsdIntrinsicBlendDstIn_K()
978     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendSrcOut_K()  local
985         in0 = _mm_loadu_si128((const __m128i *)src);  in rsdIntrinsicBlendSrcOut_K()
993         t0 = _mm_unpacklo_epi8(in0, _mm_setzero_si128());  in rsdIntrinsicBlendSrcOut_K()
1000         t1 = _mm_unpackhi_epi8(in0, _mm_setzero_si128());  in rsdIntrinsicBlendSrcOut_K()
1030     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendDstOut_K()  local
1037         in0 = _mm_loadu_si128((const __m128i *)src);  in rsdIntrinsicBlendDstOut_K()
1042         ina = _mm_unpacklo_epi8(in0, _mm_setzero_si128());  in rsdIntrinsicBlendDstOut_K()
1049         ina = _mm_unpackhi_epi8(in0, _mm_setzero_si128());  in rsdIntrinsicBlendDstOut_K()
1083     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendSrcAtop_K()  local
1090         in0 = _mm_loadu_si128((const __m128i *)src);  in rsdIntrinsicBlendSrcAtop_K()
1095         ins = _mm_unpacklo_epi8(in0, _mm_setzero_si128());  in rsdIntrinsicBlendSrcAtop_K()
1106         ins = _mm_unpackhi_epi8(in0, _mm_setzero_si128());  in rsdIntrinsicBlendSrcAtop_K()
1154     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendDstAtop_K()  local
1161         in0 = _mm_loadu_si128((const __m128i *)src);  in rsdIntrinsicBlendDstAtop_K()
1166         ins = _mm_unpacklo_epi8(in0, _mm_setzero_si128());  in rsdIntrinsicBlendDstAtop_K()
1177         ins = _mm_unpackhi_epi8(in0, _mm_setzero_si128());  in rsdIntrinsicBlendDstAtop_K()
1211         t0 = blendv_epi8(t0, in0, M0001);  in rsdIntrinsicBlendDstAtop_K()
1223     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendXor_K()  local
1227         in0 = _mm_loadu_si128((const __m128i *)src);  in rsdIntrinsicBlendXor_K()
1232         out0 = _mm_xor_si128(out0, in0);  in rsdIntrinsicBlendXor_K()
1244     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendMultiply_K()  local
1249         in0 = _mm_loadu_si128((const __m128i *)src);  in rsdIntrinsicBlendMultiply_K()
1254         t0 = _mm_unpacklo_epi8(in0, _mm_setzero_si128());  in rsdIntrinsicBlendMultiply_K()
1258         t1 = _mm_unpackhi_epi8(in0, _mm_setzero_si128());  in rsdIntrinsicBlendMultiply_K()
1281     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendAdd_K()  local
1285         in0 = _mm_loadu_si128((const __m128i *)src);  in rsdIntrinsicBlendAdd_K()
1290         out0 = _mm_adds_epu8(out0, in0);  in rsdIntrinsicBlendAdd_K()
1302     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendSub_K()  local
1306         in0 = _mm_loadu_si128((const __m128i *)src);  in rsdIntrinsicBlendSub_K()
1311         out0 = _mm_subs_epu8(out0, in0);  in rsdIntrinsicBlendSub_K()