Lines Matching refs:outa
817 __m128i all1s, outa, outs;
832 outa = _mm_shufflelo_epi16(outs, 0xFF);
833 outa = _mm_shufflehi_epi16(outa, 0xFF);
835 t0 = _mm_mullo_epi16(t0, _mm_sub_epi16(all1s, outa));
840 outa = _mm_shufflelo_epi16(outs, 0xFF);
841 outa = _mm_shufflehi_epi16(outa, 0xFF);
843 t1 = _mm_mullo_epi16(t1, _mm_sub_epi16(all1s, outa));
848 outa = _mm_shufflelo_epi16(outs, 0xFF);
849 outa = _mm_shufflehi_epi16(outa, 0xFF);
851 t2 = _mm_mullo_epi16(t2, _mm_sub_epi16(all1s, outa));
856 outa = _mm_shufflelo_epi16(outs, 0xFF);
857 outa = _mm_shufflehi_epi16(outa, 0xFF);
859 t3 = _mm_mullo_epi16(t3, _mm_sub_epi16(all1s, outa));
874 __m128i outa;
885 outa = _mm_unpacklo_epi8(out0, _mm_setzero_si128());
886 outa = _mm_shufflelo_epi16(outa, 0xFF);
887 outa = _mm_shufflehi_epi16(outa, 0xFF);
889 t0 = _mm_mullo_epi16(t0, outa);
892 outa = _mm_unpackhi_epi8(out0, _mm_setzero_si128());
893 outa = _mm_shufflelo_epi16(outa, 0xFF);
894 outa = _mm_shufflehi_epi16(outa, 0xFF);
896 t1 = _mm_mullo_epi16(t1, outa);
899 outa = _mm_unpacklo_epi8(out1, _mm_setzero_si128());
900 outa = _mm_shufflelo_epi16(outa, 0xFF);
901 outa = _mm_shufflehi_epi16(outa, 0xFF);
903 t2 = _mm_mullo_epi16(t2, outa);
906 outa = _mm_unpackhi_epi8(out1, _mm_setzero_si128());
907 outa = _mm_shufflelo_epi16(outa, 0xFF);
908 outa = _mm_shufflehi_epi16(outa, 0xFF);
910 t3 = _mm_mullo_epi16(t3, outa);
974 __m128i all1s, outa;
987 outa = _mm_unpacklo_epi8(out0, _mm_setzero_si128());
988 outa = _mm_shufflelo_epi16(outa, 0xFF);
989 outa = _mm_shufflehi_epi16(outa, 0xFF);
991 t0 = _mm_mullo_epi16(t0, _mm_sub_epi16(all1s, outa));
994 outa = _mm_unpackhi_epi8(out0, _mm_setzero_si128());
995 outa = _mm_shufflelo_epi16(outa, 0xFF);
996 outa = _mm_shufflehi_epi16(outa, 0xFF);
998 t1 = _mm_mullo_epi16(t1, _mm_sub_epi16(all1s, outa));
1001 outa = _mm_unpacklo_epi8(out1, _mm_setzero_si128());
1002 outa = _mm_shufflelo_epi16(outa, 0xFF);
1003 outa = _mm_shufflehi_epi16(outa, 0xFF);
1005 t2 = _mm_mullo_epi16(t2, _mm_sub_epi16(all1s, outa));
1008 outa = _mm_unpackhi_epi8(out1, _mm_setzero_si128());
1009 outa = _mm_shufflelo_epi16(outa, 0xFF);
1010 outa = _mm_shufflehi_epi16(outa, 0xFF);
1012 t3 = _mm_mullo_epi16(t3, _mm_sub_epi16(all1s, outa));
1079 __m128i all1s, ina, outa, ins, outs;
1096 outa = _mm_shufflelo_epi16(outs, 0xFF);
1097 outa = _mm_shufflehi_epi16(outa, 0xFF);
1100 t0 = _mm_adds_epu16(t0, _mm_mullo_epi16(outa, ins));
1107 outa = _mm_shufflelo_epi16(outs, 0xFF);
1108 outa = _mm_shufflehi_epi16(outa, 0xFF);
1111 t1 = _mm_adds_epu16(t1, _mm_mullo_epi16(outa, ins));
1118 outa = _mm_shufflelo_epi16(outs, 0xFF);
1119 outa = _mm_shufflehi_epi16(outa, 0xFF);
1122 t2 = _mm_adds_epu16(t2, _mm_mullo_epi16(outa, ins));
1129 outa = _mm_shufflelo_epi16(outs, 0xFF);
1130 outa = _mm_shufflehi_epi16(outa, 0xFF);
1133 t3 = _mm_adds_epu16(t3, _mm_mullo_epi16(outa, ins));
1150 __m128i all1s, ina, ins, outa, outs;
1167 outa = _mm_shufflelo_epi16(outs, 0xFF);
1168 outa = _mm_shufflehi_epi16(outa, 0xFF);
1169 t0 = _mm_sub_epi16(all1s, outa);
1178 outa = _mm_shufflelo_epi16(outs, 0xFF);
1179 outa = _mm_shufflehi_epi16(outa, 0xFF);
1180 t1 = _mm_sub_epi16(all1s, outa);
1189 outa = _mm_shufflelo_epi16(outs, 0xFF);
1190 outa = _mm_shufflehi_epi16(outa, 0xFF);
1191 t2 = _mm_sub_epi16(all1s, outa);
1200 outa = _mm_shufflelo_epi16(outs, 0xFF);
1201 outa = _mm_shufflehi_epi16(outa, 0xFF);
1202 t3 = _mm_sub_epi16(all1s, outa);
