763     __m128i t0, t1, t2, t3;  in rsdIntrinsicBlendSrcOver_K()  local
793         t2 = _mm_unpacklo_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcOver_K()
794         t2 = _mm_mullo_epi16(t2, _mm_sub_epi16(all1s, ina));  in rsdIntrinsicBlendSrcOver_K()
795         t2 = _mm_srli_epi16(t2, 8);  in rsdIntrinsicBlendSrcOver_K()
796         t2 = _mm_add_epi16(t2, ins);  in rsdIntrinsicBlendSrcOver_K()
807         t2 = _mm_packus_epi16(t2, t3);  in rsdIntrinsicBlendSrcOver_K()
809         _mm_storeu_si128((__m128i *)dst + 1, t2);  in rsdIntrinsicBlendSrcOver_K()
819     __m128i t0, t1, t2, t3;  in rsdIntrinsicBlendDstOver_K()  local
850         t2 = _mm_unpacklo_epi8(in1, _mm_setzero_si128());  in rsdIntrinsicBlendDstOver_K()
851         t2 = _mm_mullo_epi16(t2, _mm_sub_epi16(all1s, outa));  in rsdIntrinsicBlendDstOver_K()
852         t2 = _mm_srli_epi16(t2, 8);  in rsdIntrinsicBlendDstOver_K()
853         t2 = _mm_add_epi16(t2, outs);  in rsdIntrinsicBlendDstOver_K()
864         t2 = _mm_packus_epi16(t2, t3);  in rsdIntrinsicBlendDstOver_K()
866         _mm_storeu_si128((__m128i *)dst + 1, t2);  in rsdIntrinsicBlendDstOver_K()
876     __m128i t0, t1, t2, t3;  in rsdIntrinsicBlendSrcIn_K()  local
902         t2 = _mm_unpacklo_epi8(in1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcIn_K()
903         t2 = _mm_mullo_epi16(t2, outa);  in rsdIntrinsicBlendSrcIn_K()
904         t2 = _mm_srli_epi16(t2, 8);  in rsdIntrinsicBlendSrcIn_K()
914         t2 = _mm_packus_epi16(t2, t3);  in rsdIntrinsicBlendSrcIn_K()
916         _mm_storeu_si128((__m128i *)dst + 1, t2);  in rsdIntrinsicBlendSrcIn_K()
926     __m128i t0, t1, t2, t3;  in rsdIntrinsicBlendDstIn_K()  local
952         t2 = _mm_unpacklo_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendDstIn_K()
953         t2 = _mm_mullo_epi16(t2, ina);  in rsdIntrinsicBlendDstIn_K()
954         t2 = _mm_srli_epi16(t2, 8);  in rsdIntrinsicBlendDstIn_K()
964         t2 = _mm_packus_epi16(t2, t3);  in rsdIntrinsicBlendDstIn_K()
966         _mm_storeu_si128((__m128i *)dst + 1, t2);  in rsdIntrinsicBlendDstIn_K()
976     __m128i t0, t1, t2, t3;  in rsdIntrinsicBlendSrcOut_K()  local
1004         t2 = _mm_unpacklo_epi8(in1, _mm_setzero_si128());  in rsdIntrinsicBlendSrcOut_K()
1005         t2 = _mm_mullo_epi16(t2, _mm_sub_epi16(all1s, outa));  in rsdIntrinsicBlendSrcOut_K()
1006         t2 = _mm_srli_epi16(t2, 8);  in rsdIntrinsicBlendSrcOut_K()
1016         t2 = _mm_packus_epi16(t2, t3);  in rsdIntrinsicBlendSrcOut_K()
1018         _mm_storeu_si128((__m128i *)dst + 1, t2);  in rsdIntrinsicBlendSrcOut_K()
1028     __m128i t0, t1, t2, t3;  in rsdIntrinsicBlendDstOut_K()  local
1056         t2 = _mm_unpacklo_epi8(out1, _mm_setzero_si128());  in rsdIntrinsicBlendDstOut_K()
1057         t2 = _mm_mullo_epi16(t2, _mm_sub_epi16(all1s, ina));  in rsdIntrinsicBlendDstOut_K()
1058         t2 = _mm_srli_epi16(t2, 8);  in rsdIntrinsicBlendDstOut_K()
1068         t2 = _mm_packus_epi16(t2, t3);  in rsdIntrinsicBlendDstOut_K()
1070         _mm_storeu_si128((__m128i *)dst + 1, t2);  in rsdIntrinsicBlendDstOut_K()
1081     __m128i t0, t1, t2, t3;  in rsdIntrinsicBlendSrcAtop_K()  local
1120         t2 = _mm_sub_epi16(all1s, ina);  in rsdIntrinsicBlendSrcAtop_K()
1121         t2 = _mm_mullo_epi16(t2, outs);  in rsdIntrinsicBlendSrcAtop_K()
1122         t2 = _mm_adds_epu16(t2, _mm_mullo_epi16(outa, ins));  in rsdIntrinsicBlendSrcAtop_K()
1123         t2 = _mm_srli_epi16(t2, 8);  in rsdIntrinsicBlendSrcAtop_K()
1138         t2 = _mm_packus_epi16(t2, t3);  in rsdIntrinsicBlendSrcAtop_K()
1139         t2 = blendv_epi8(t2, out1, M0001);  in rsdIntrinsicBlendSrcAtop_K()
1141         _mm_storeu_si128((__m128i *)dst + 1, t2);  in rsdIntrinsicBlendSrcAtop_K()
1152     __m128i t0, t1, t2, t3;  in rsdIntrinsicBlendDstAtop_K()  local
1191         t2 = _mm_sub_epi16(all1s, outa);  in rsdIntrinsicBlendDstAtop_K()
1192         t2 = _mm_mullo_epi16(t2, ins);  in rsdIntrinsicBlendDstAtop_K()
1193         t2 = _mm_adds_epu16(t2, _mm_mullo_epi16(ina, outs));  in rsdIntrinsicBlendDstAtop_K()
1194         t2 = _mm_srli_epi16(t2, 8);  in rsdIntrinsicBlendDstAtop_K()
1209         t2 = _mm_packus_epi16(t2, t3);  in rsdIntrinsicBlendDstAtop_K()
1210         t2 = blendv_epi8(t2, in1, M0001);  in rsdIntrinsicBlendDstAtop_K()
1212         _mm_storeu_si128((__m128i *)dst + 1, t2);  in rsdIntrinsicBlendDstAtop_K()
1242     __m128i t0, t1, t2, t3;  in rsdIntrinsicBlendMultiply_K()  local
1259         t2 = _mm_unpacklo_epi8(in1, _mm_setzero_si128());  in rsdIntrinsicBlendMultiply_K()
1260         t2 = _mm_mullo_epi16(t2, _mm_unpacklo_epi8(out1, _mm_setzero_si128()));  in rsdIntrinsicBlendMultiply_K()
1261         t2 = _mm_srli_epi16(t2, 8);  in rsdIntrinsicBlendMultiply_K()
1268         t2 = _mm_packus_epi16(t2, t3);  in rsdIntrinsicBlendMultiply_K()
1270         _mm_storeu_si128((__m128i *)dst + 1, t2);  in rsdIntrinsicBlendMultiply_K()