765     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendSrcOver_K()  local
774         out0 = _mm_loadu_si128((const __m128i *)dst);  in rsdIntrinsicBlendSrcOver_K()
780         t0 = _mm_unpacklo_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendSrcOver_K()
788         t1 = _mm_unpackhi_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendSrcOver_K()
821     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendDstOver_K()  local
830         out0 = _mm_loadu_si128((const __m128i *)dst);  in rsdIntrinsicBlendDstOver_K()
834         outs = _mm_unpacklo_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendDstOver_K()
842         outs = _mm_unpackhi_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendDstOver_K()
878     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendSrcIn_K()  local
885         out0 = _mm_loadu_si128((const __m128i *)dst);  in rsdIntrinsicBlendSrcIn_K()
888         outa = _mm_unpacklo_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendSrcIn_K()
895         outa = _mm_unpackhi_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendSrcIn_K()
928     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendDstIn_K()  local
935         out0 = _mm_loadu_si128((const __m128i *)dst);  in rsdIntrinsicBlendDstIn_K()
941         t0 = _mm_unpacklo_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendDstIn_K()
948         t1 = _mm_unpackhi_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendDstIn_K()
978     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendSrcOut_K()  local
987         out0 = _mm_loadu_si128((const __m128i *)dst);  in rsdIntrinsicBlendSrcOut_K()
990         outa = _mm_unpacklo_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendSrcOut_K()
997         outa = _mm_unpackhi_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendSrcOut_K()
1030     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendDstOut_K()  local
1039         out0 = _mm_loadu_si128((const __m128i *)dst);  in rsdIntrinsicBlendDstOut_K()
1045         t0 = _mm_unpacklo_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendDstOut_K()
1052         t1 = _mm_unpackhi_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendDstOut_K()
1083     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendSrcAtop_K()  local
1092         out0 = _mm_loadu_si128((const __m128i *)dst);  in rsdIntrinsicBlendSrcAtop_K()
1098         outs = _mm_unpacklo_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendSrcAtop_K()
1109         outs = _mm_unpackhi_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendSrcAtop_K()
1140         t0 = blendv_epi8(t0, out0, M0001);  in rsdIntrinsicBlendSrcAtop_K()
1154     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendDstAtop_K()  local
1163         out0 = _mm_loadu_si128((const __m128i *)dst);  in rsdIntrinsicBlendDstAtop_K()
1169         outs = _mm_unpacklo_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendDstAtop_K()
1180         outs = _mm_unpackhi_epi8(out0, _mm_setzero_si128());  in rsdIntrinsicBlendDstAtop_K()
1223     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendXor_K()  local
1229         out0 = _mm_loadu_si128((const __m128i *)dst);  in rsdIntrinsicBlendXor_K()
1232         out0 = _mm_xor_si128(out0, in0);  in rsdIntrinsicBlendXor_K()
1235         _mm_storeu_si128((__m128i *)dst, out0);  in rsdIntrinsicBlendXor_K()
1244     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendMultiply_K()  local
1251         out0 = _mm_loadu_si128((const __m128i *)dst);  in rsdIntrinsicBlendMultiply_K()
1255         t0 = _mm_mullo_epi16(t0, _mm_unpacklo_epi8(out0, _mm_setzero_si128()));  in rsdIntrinsicBlendMultiply_K()
1259         t1 = _mm_mullo_epi16(t1, _mm_unpackhi_epi8(out0, _mm_setzero_si128()));  in rsdIntrinsicBlendMultiply_K()
1281     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendAdd_K()  local
1287         out0 = _mm_loadu_si128((const __m128i *)dst);  in rsdIntrinsicBlendAdd_K()
1290         out0 = _mm_adds_epu8(out0, in0);  in rsdIntrinsicBlendAdd_K()
1293         _mm_storeu_si128((__m128i *)dst, out0);  in rsdIntrinsicBlendAdd_K()
1302     __m128i in0, in1, out0, out1;  in rsdIntrinsicBlendSub_K()  local
1308         out0 = _mm_loadu_si128((const __m128i *)dst);  in rsdIntrinsicBlendSub_K()
1311         out0 = _mm_subs_epu8(out0, in0);  in rsdIntrinsicBlendSub_K()
1314         _mm_storeu_si128((__m128i *)dst, out0);  in rsdIntrinsicBlendSub_K()