762     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendSrcOver_K()  local
770         in1 = _mm_loadu_si128((const __m128i *)src + 1);  in rsdIntrinsicBlendSrcOver_K()
790         ins = _mm_unpacklo_epi8(in1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcOver_K()
798         ins = _mm_unpackhi_epi8(in1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcOver_K()
818     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendDstOver_K()  local
826         in1 = _mm_loadu_si128((const __m128i *)src + 1);  in rsdIntrinsicBlendDstOver_K()
850         t2 = _mm_unpacklo_epi8(in1, _mm_setzero_si128());  in rsdIntrinsicBlendDstOver_K()
858         t3 = _mm_unpackhi_epi8(in1, _mm_setzero_si128());  in rsdIntrinsicBlendDstOver_K()
875     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendSrcIn_K()  local
881         in1 = _mm_loadu_si128((const __m128i *)src + 1);  in rsdIntrinsicBlendSrcIn_K()
902         t2 = _mm_unpacklo_epi8(in1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcIn_K()
909         t3 = _mm_unpackhi_epi8(in1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcIn_K()
925     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendDstIn_K()  local
931         in1 = _mm_loadu_si128((const __m128i *)src + 1);  in rsdIntrinsicBlendDstIn_K()
949         ina = _mm_unpacklo_epi8(in1, _mm_setzero_si128());  in rsdIntrinsicBlendDstIn_K()
956         ina = _mm_unpackhi_epi8(in1, _mm_setzero_si128());  in rsdIntrinsicBlendDstIn_K()
975     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendSrcOut_K()  local
983         in1 = _mm_loadu_si128((const __m128i *)src + 1);  in rsdIntrinsicBlendSrcOut_K()
1004         t2 = _mm_unpacklo_epi8(in1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcOut_K()
1011         t3 = _mm_unpackhi_epi8(in1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcOut_K()
1027     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendDstOut_K()  local
1035         in1 = _mm_loadu_si128((const __m128i *)src + 1);  in rsdIntrinsicBlendDstOut_K()
1053         ina = _mm_unpacklo_epi8(in1, _mm_setzero_si128());  in rsdIntrinsicBlendDstOut_K()
1060         ina = _mm_unpackhi_epi8(in1, _mm_setzero_si128());  in rsdIntrinsicBlendDstOut_K()
1080     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendSrcAtop_K()  local
1088         in1 = _mm_loadu_si128((const __m128i *)src + 1);  in rsdIntrinsicBlendSrcAtop_K()
1114         ins = _mm_unpacklo_epi8(in1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcAtop_K()
1125         ins = _mm_unpackhi_epi8(in1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcAtop_K()
1151     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendDstAtop_K()  local
1159         in1 = _mm_loadu_si128((const __m128i *)src + 1);  in rsdIntrinsicBlendDstAtop_K()
1185         ins = _mm_unpacklo_epi8(in1, _mm_setzero_si128());  in rsdIntrinsicBlendDstAtop_K()
1196         ins = _mm_unpackhi_epi8(in1, _mm_setzero_si128());  in rsdIntrinsicBlendDstAtop_K()
1210         t2 = blendv_epi8(t2, in1, M0001);  in rsdIntrinsicBlendDstAtop_K()
1220     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendXor_K()  local
1225         in1 = _mm_loadu_si128((const __m128i *)src + 1);  in rsdIntrinsicBlendXor_K()
1230         out1 = _mm_xor_si128(out1, in1);  in rsdIntrinsicBlendXor_K()
1241     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendMultiply_K()  local
1247         in1 = _mm_loadu_si128((const __m128i *)src + 1);  in rsdIntrinsicBlendMultiply_K()
1259         t2 = _mm_unpacklo_epi8(in1, _mm_setzero_si128());  in rsdIntrinsicBlendMultiply_K()
1263         t3 = _mm_unpackhi_epi8(in1, _mm_setzero_si128());  in rsdIntrinsicBlendMultiply_K()
1278     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendAdd_K()  local
1283         in1 = _mm_loadu_si128((const __m128i *)src + 1);  in rsdIntrinsicBlendAdd_K()
1288         out1 = _mm_adds_epu8(out1, in1);  in rsdIntrinsicBlendAdd_K()
1299     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendSub_K()  local
1304         in1 = _mm_loadu_si128((const __m128i *)src + 1);  in rsdIntrinsicBlendSub_K()
1309         out1 = _mm_subs_epu8(out1, in1);  in rsdIntrinsicBlendSub_K()