Lines Matching refs:t0
763 __m128i t0, t1, t2, t3; in rsdIntrinsicBlendSrcOver_K() local
777 t0 = _mm_unpacklo_epi8(out0, _mm_setzero_si128()); in rsdIntrinsicBlendSrcOver_K()
778 t0 = _mm_mullo_epi16(t0, _mm_sub_epi16(all1s, ina)); in rsdIntrinsicBlendSrcOver_K()
779 t0 = _mm_srli_epi16(t0, 8); in rsdIntrinsicBlendSrcOver_K()
780 t0 = _mm_add_epi16(t0, ins); in rsdIntrinsicBlendSrcOver_K()
806 t0 = _mm_packus_epi16(t0, t1); in rsdIntrinsicBlendSrcOver_K()
808 _mm_storeu_si128((__m128i *)dst, t0); in rsdIntrinsicBlendSrcOver_K()
819 __m128i t0, t1, t2, t3; in rsdIntrinsicBlendDstOver_K() local
834 t0 = _mm_unpacklo_epi8(in0, _mm_setzero_si128()); in rsdIntrinsicBlendDstOver_K()
835 t0 = _mm_mullo_epi16(t0, _mm_sub_epi16(all1s, outa)); in rsdIntrinsicBlendDstOver_K()
836 t0 = _mm_srli_epi16(t0, 8); in rsdIntrinsicBlendDstOver_K()
837 t0 = _mm_add_epi16(t0, outs); in rsdIntrinsicBlendDstOver_K()
863 t0 = _mm_packus_epi16(t0, t1); in rsdIntrinsicBlendDstOver_K()
865 _mm_storeu_si128((__m128i *)dst, t0); in rsdIntrinsicBlendDstOver_K()
876 __m128i t0, t1, t2, t3; in rsdIntrinsicBlendSrcIn_K() local
888 t0 = _mm_unpacklo_epi8(in0, _mm_setzero_si128()); in rsdIntrinsicBlendSrcIn_K()
889 t0 = _mm_mullo_epi16(t0, outa); in rsdIntrinsicBlendSrcIn_K()
890 t0 = _mm_srli_epi16(t0, 8); in rsdIntrinsicBlendSrcIn_K()
913 t0 = _mm_packus_epi16(t0, t1); in rsdIntrinsicBlendSrcIn_K()
915 _mm_storeu_si128((__m128i *)dst, t0); in rsdIntrinsicBlendSrcIn_K()
926 __m128i t0, t1, t2, t3; in rsdIntrinsicBlendDstIn_K() local
938 t0 = _mm_unpacklo_epi8(out0, _mm_setzero_si128()); in rsdIntrinsicBlendDstIn_K()
939 t0 = _mm_mullo_epi16(t0, ina); in rsdIntrinsicBlendDstIn_K()
940 t0 = _mm_srli_epi16(t0, 8); in rsdIntrinsicBlendDstIn_K()
963 t0 = _mm_packus_epi16(t0, t1); in rsdIntrinsicBlendDstIn_K()
965 _mm_storeu_si128((__m128i *)dst, t0); in rsdIntrinsicBlendDstIn_K()
976 __m128i t0, t1, t2, t3; in rsdIntrinsicBlendSrcOut_K() local
990 t0 = _mm_unpacklo_epi8(in0, _mm_setzero_si128()); in rsdIntrinsicBlendSrcOut_K()
991 t0 = _mm_mullo_epi16(t0, _mm_sub_epi16(all1s, outa)); in rsdIntrinsicBlendSrcOut_K()
992 t0 = _mm_srli_epi16(t0, 8); in rsdIntrinsicBlendSrcOut_K()
1015 t0 = _mm_packus_epi16(t0, t1); in rsdIntrinsicBlendSrcOut_K()
1017 _mm_storeu_si128((__m128i *)dst, t0); in rsdIntrinsicBlendSrcOut_K()
1028 __m128i t0, t1, t2, t3; in rsdIntrinsicBlendDstOut_K() local
1042 t0 = _mm_unpacklo_epi8(out0, _mm_setzero_si128()); in rsdIntrinsicBlendDstOut_K()
1043 t0 = _mm_mullo_epi16(t0, _mm_sub_epi16(all1s, ina)); in rsdIntrinsicBlendDstOut_K()
1044 t0 = _mm_srli_epi16(t0, 8); in rsdIntrinsicBlendDstOut_K()
1067 t0 = _mm_packus_epi16(t0, t1); in rsdIntrinsicBlendDstOut_K()
1069 _mm_storeu_si128((__m128i *)dst, t0); in rsdIntrinsicBlendDstOut_K()
1081 __m128i t0, t1, t2, t3; in rsdIntrinsicBlendSrcAtop_K() local
1098 t0 = _mm_sub_epi16(all1s, ina); in rsdIntrinsicBlendSrcAtop_K()
1099 t0 = _mm_mullo_epi16(t0, outs); in rsdIntrinsicBlendSrcAtop_K()
1100 t0 = _mm_adds_epu16(t0, _mm_mullo_epi16(outa, ins)); in rsdIntrinsicBlendSrcAtop_K()
1101 t0 = _mm_srli_epi16(t0, 8); in rsdIntrinsicBlendSrcAtop_K()
1136 t0 = _mm_packus_epi16(t0, t1); in rsdIntrinsicBlendSrcAtop_K()
1137 t0 = blendv_epi8(t0, out0, M0001); in rsdIntrinsicBlendSrcAtop_K()
1140 _mm_storeu_si128((__m128i *)dst, t0); in rsdIntrinsicBlendSrcAtop_K()
1152 __m128i t0, t1, t2, t3; in rsdIntrinsicBlendDstAtop_K() local
1169 t0 = _mm_sub_epi16(all1s, outa); in rsdIntrinsicBlendDstAtop_K()
1170 t0 = _mm_mullo_epi16(t0, ins); in rsdIntrinsicBlendDstAtop_K()
1171 t0 = _mm_adds_epu16(t0, _mm_mullo_epi16(ina, outs)); in rsdIntrinsicBlendDstAtop_K()
1172 t0 = _mm_srli_epi16(t0, 8); in rsdIntrinsicBlendDstAtop_K()
1207 t0 = _mm_packus_epi16(t0, t1); in rsdIntrinsicBlendDstAtop_K()
1208 t0 = blendv_epi8(t0, in0, M0001); in rsdIntrinsicBlendDstAtop_K()
1211 _mm_storeu_si128((__m128i *)dst, t0); in rsdIntrinsicBlendDstAtop_K()
1242 __m128i t0, t1, t2, t3; in rsdIntrinsicBlendMultiply_K() local
1251 t0 = _mm_unpacklo_epi8(in0, _mm_setzero_si128()); in rsdIntrinsicBlendMultiply_K()
1252 t0 = _mm_mullo_epi16(t0, _mm_unpacklo_epi8(out0, _mm_setzero_si128())); in rsdIntrinsicBlendMultiply_K()
1253 t0 = _mm_srli_epi16(t0, 8); in rsdIntrinsicBlendMultiply_K()
1267 t0 = _mm_packus_epi16(t0, t1); in rsdIntrinsicBlendMultiply_K()
1269 _mm_storeu_si128((__m128i *)dst, t0); in rsdIntrinsicBlendMultiply_K()