765     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendSrcOver_K()  local
775         out1 = _mm_loadu_si128((const __m128i *)dst + 1);  in rsdIntrinsicBlendSrcOver_K()
796         t2 = _mm_unpacklo_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcOver_K()
804         t3 = _mm_unpackhi_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcOver_K()
821     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendDstOver_K()  local
831         out1 = _mm_loadu_si128((const __m128i *)dst + 1);  in rsdIntrinsicBlendDstOver_K()
850         outs = _mm_unpacklo_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendDstOver_K()
858         outs = _mm_unpackhi_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendDstOver_K()
878     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendSrcIn_K()  local
886         out1 = _mm_loadu_si128((const __m128i *)dst + 1);  in rsdIntrinsicBlendSrcIn_K()
902         outa = _mm_unpacklo_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcIn_K()
909         outa = _mm_unpackhi_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcIn_K()
928     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendDstIn_K()  local
936         out1 = _mm_loadu_si128((const __m128i *)dst + 1);  in rsdIntrinsicBlendDstIn_K()
955         t2 = _mm_unpacklo_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendDstIn_K()
962         t3 = _mm_unpackhi_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendDstIn_K()
978     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendSrcOut_K()  local
988         out1 = _mm_loadu_si128((const __m128i *)dst + 1);  in rsdIntrinsicBlendSrcOut_K()
1004         outa = _mm_unpacklo_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcOut_K()
1011         outa = _mm_unpackhi_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcOut_K()
1030     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendDstOut_K()  local
1040         out1 = _mm_loadu_si128((const __m128i *)dst + 1);  in rsdIntrinsicBlendDstOut_K()
1059         t2 = _mm_unpacklo_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendDstOut_K()
1066         t3 = _mm_unpackhi_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendDstOut_K()
1083     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendSrcAtop_K()  local
1093         out1 = _mm_loadu_si128((const __m128i *)dst + 1);  in rsdIntrinsicBlendSrcAtop_K()
1120         outs = _mm_unpacklo_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcAtop_K()
1131         outs = _mm_unpackhi_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcAtop_K()
1142         t2 = blendv_epi8(t2, out1, M0001);  in rsdIntrinsicBlendSrcAtop_K()
1154     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendDstAtop_K()  local
1164         out1 = _mm_loadu_si128((const __m128i *)dst + 1);  in rsdIntrinsicBlendDstAtop_K()
1191         outs = _mm_unpacklo_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendDstAtop_K()
1202         outs = _mm_unpackhi_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendDstAtop_K()
1223     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendXor_K()  local
1230         out1 = _mm_loadu_si128((const __m128i *)dst + 1);  in rsdIntrinsicBlendXor_K()
1233         out1 = _mm_xor_si128(out1, in1);  in rsdIntrinsicBlendXor_K()
1236         _mm_storeu_si128((__m128i *)dst + 1, out1);  in rsdIntrinsicBlendXor_K()
1244     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendMultiply_K()  local
1252         out1 = _mm_loadu_si128((const __m128i *)dst + 1);  in rsdIntrinsicBlendMultiply_K()
1263         t2 = _mm_mullo_epi16(t2, _mm_unpacklo_epi8(out1, _mm_setzero_si128()));  in rsdIntrinsicBlendMultiply_K()
1267         t3 = _mm_mullo_epi16(t3, _mm_unpackhi_epi8(out1, _mm_setzero_si128()));  in rsdIntrinsicBlendMultiply_K()
1281     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendAdd_K()  local
1288         out1 = _mm_loadu_si128((const __m128i *)dst + 1);  in rsdIntrinsicBlendAdd_K()
1291         out1 = _mm_adds_epu8(out1, in1);  in rsdIntrinsicBlendAdd_K()
1294         _mm_storeu_si128((__m128i *)dst + 1, out1);  in rsdIntrinsicBlendAdd_K()
1302     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendSub_K()  local
1309         out1 = _mm_loadu_si128((const __m128i *)dst + 1);  in rsdIntrinsicBlendSub_K()
1312         out1 = _mm_subs_epu8(out1, in1);  in rsdIntrinsicBlendSub_K()
1315         _mm_storeu_si128((__m128i *)dst + 1, out1);  in rsdIntrinsicBlendSub_K()