762     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendSrcOver_K()  local
772         out1 = _mm_loadu_si128((const __m128i *)dst + 1);  in rsdIntrinsicBlendSrcOver_K()
793         t2 = _mm_unpacklo_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcOver_K()
801         t3 = _mm_unpackhi_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcOver_K()
818     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendDstOver_K()  local
828         out1 = _mm_loadu_si128((const __m128i *)dst + 1);  in rsdIntrinsicBlendDstOver_K()
847         outs = _mm_unpacklo_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendDstOver_K()
855         outs = _mm_unpackhi_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendDstOver_K()
875     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendSrcIn_K()  local
883         out1 = _mm_loadu_si128((const __m128i *)dst + 1);  in rsdIntrinsicBlendSrcIn_K()
899         outa = _mm_unpacklo_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcIn_K()
906         outa = _mm_unpackhi_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcIn_K()
925     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendDstIn_K()  local
933         out1 = _mm_loadu_si128((const __m128i *)dst + 1);  in rsdIntrinsicBlendDstIn_K()
952         t2 = _mm_unpacklo_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendDstIn_K()
959         t3 = _mm_unpackhi_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendDstIn_K()
975     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendSrcOut_K()  local
985         out1 = _mm_loadu_si128((const __m128i *)dst + 1);  in rsdIntrinsicBlendSrcOut_K()
1001         outa = _mm_unpacklo_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcOut_K()
1008         outa = _mm_unpackhi_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcOut_K()
1027     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendDstOut_K()  local
1037         out1 = _mm_loadu_si128((const __m128i *)dst + 1);  in rsdIntrinsicBlendDstOut_K()
1056         t2 = _mm_unpacklo_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendDstOut_K()
1063         t3 = _mm_unpackhi_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendDstOut_K()
1080     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendSrcAtop_K()  local
1090         out1 = _mm_loadu_si128((const __m128i *)dst + 1);  in rsdIntrinsicBlendSrcAtop_K()
1117         outs = _mm_unpacklo_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcAtop_K()
1128         outs = _mm_unpackhi_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcAtop_K()
1139         t2 = blendv_epi8(t2, out1, M0001);  in rsdIntrinsicBlendSrcAtop_K()
1151     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendDstAtop_K()  local
1161         out1 = _mm_loadu_si128((const __m128i *)dst + 1);  in rsdIntrinsicBlendDstAtop_K()
1188         outs = _mm_unpacklo_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendDstAtop_K()
1199         outs = _mm_unpackhi_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendDstAtop_K()
1220     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendXor_K()  local
1227         out1 = _mm_loadu_si128((const __m128i *)dst + 1);  in rsdIntrinsicBlendXor_K()
1230         out1 = _mm_xor_si128(out1, in1);  in rsdIntrinsicBlendXor_K()
1233         _mm_storeu_si128((__m128i *)dst + 1, out1);  in rsdIntrinsicBlendXor_K()
1241     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendMultiply_K()  local
1249         out1 = _mm_loadu_si128((const __m128i *)dst + 1);  in rsdIntrinsicBlendMultiply_K()
1260         t2 = _mm_mullo_epi16(t2, _mm_unpacklo_epi8(out1, _mm_setzero_si128()));  in rsdIntrinsicBlendMultiply_K()
1264         t3 = _mm_mullo_epi16(t3, _mm_unpackhi_epi8(out1, _mm_setzero_si128()));  in rsdIntrinsicBlendMultiply_K()
1278     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendAdd_K()  local
1285         out1 = _mm_loadu_si128((const __m128i *)dst + 1);  in rsdIntrinsicBlendAdd_K()
1288         out1 = _mm_adds_epu8(out1, in1);  in rsdIntrinsicBlendAdd_K()
1291         _mm_storeu_si128((__m128i *)dst + 1, out1);  in rsdIntrinsicBlendAdd_K()
1299     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendSub_K()  local
1306         out1 = _mm_loadu_si128((const __m128i *)dst + 1);  in rsdIntrinsicBlendSub_K()
1309         out1 = _mm_subs_epu8(out1, in1);  in rsdIntrinsicBlendSub_K()
1312         _mm_storeu_si128((__m128i *)dst + 1, out1);  in rsdIntrinsicBlendSub_K()