Lines Matching refs:res_temp1_8x16b
730 __m128i res_temp_8x16b, res_temp1_8x16b, res_temp2_8x16b, res_temp3_8x16b;
736 res_temp1_8x16b = _mm_set1_epi8(pu1_ref[two_nt - 1 - row]);
748 res_temp4_8x16b = _mm_unpacklo_epi8(res_temp1_8x16b, src_temp_8x16b); /* row=0*/
749 res_temp5_8x16b = _mm_unpackhi_epi8(res_temp1_8x16b, src_temp_8x16b); /* row=1*/
750 res_temp6_8x16b = _mm_unpacklo_epi8(res_temp1_8x16b, src_temp1_8x16b); /* row=2*/
751 res_temp7_8x16b = _mm_unpackhi_epi8(res_temp1_8x16b, src_temp1_8x16b); /* row=3*/
755 res_temp1_8x16b = _mm_maddubs_epi16(pu1_ref_16x8b, row1_16x8b);
765 res_temp1_8x16b = _mm_add_epi16(res_temp1_8x16b, res_temp5_8x16b);
770 res_temp1_8x16b = _mm_add_epi16(res_temp1_8x16b, const_temp6_4x32b);
775 res_temp1_8x16b = _mm_srli_epi16(res_temp1_8x16b, 6);
779 res_temp_8x16b = _mm_packus_epi16(res_temp_8x16b, res_temp1_8x16b);
780 res_temp1_8x16b = _mm_packus_epi16(res_temp2_8x16b, res_temp3_8x16b);
784 _mm_storeu_si128((__m128i *)(pu1_dst + (row * dst_strd) + 16), res_temp1_8x16b);
806 __m128i res_temp_8x16b, res_temp1_8x16b, res_temp2_8x16b, res_temp3_8x16b;
812 res_temp1_8x16b = _mm_set1_epi8(pu1_ref[two_nt - 1 - row]);
826 res_temp4_8x16b = _mm_unpacklo_epi8(res_temp1_8x16b, src_temp_8x16b); /* row=0*/
828 res_temp6_8x16b = _mm_unpackhi_epi8(res_temp1_8x16b, src_temp_8x16b); /* row=2*/
833 res_temp1_8x16b = _mm_maddubs_epi16(pu1_ref_16x8b, row1_16x8b);
843 res_temp1_8x16b = _mm_add_epi16(res_temp1_8x16b, res_temp5_8x16b);
848 res_temp1_8x16b = _mm_add_epi16(res_temp1_8x16b, const_temp6_4x32b);
853 res_temp1_8x16b = _mm_srli_epi16(res_temp1_8x16b, 5);
858 res_temp1_8x16b = _mm_packus_epi16(res_temp1_8x16b, res_temp3_8x16b);
861 _mm_storeu_si128((__m128i *)(pu1_dst + ((row + 1) * dst_strd)), res_temp1_8x16b);
884 __m128i res_temp_8x16b, res_temp1_8x16b, res_temp2_8x16b, res_temp3_8x16b;
911 res_temp1_8x16b = _mm_maddubs_epi16(pu1_ref_16x8b, row1_16x8b);
921 res_temp1_8x16b = _mm_add_epi16(res_temp1_8x16b, res_temp5_8x16b);
926 res_temp1_8x16b = _mm_add_epi16(res_temp1_8x16b, const_temp6_4x32b);
931 res_temp1_8x16b = _mm_srli_epi16(res_temp1_8x16b, 4);
936 res_temp1_8x16b = _mm_packus_epi16(res_temp1_8x16b, zero_8x16b);
941 _mm_storel_epi64((__m128i *)(pu1_dst + ((row + 1) * dst_strd)), res_temp1_8x16b);
961 __m128i res_temp_8x16b, row_8x16b, res_temp1_8x16b, res_temp2_8x16b;
992 res_temp1_8x16b = _mm_mullo_epi16(src_temp_8x16b, row_8x16b);
1000 res_temp1_8x16b = _mm_add_epi16(res_temp_8x16b, res_temp1_8x16b);
1001 res_temp1_8x16b = _mm_add_epi16(res_temp1_8x16b, res_temp2_8x16b);
1002 res_temp1_8x16b = _mm_add_epi16(res_temp1_8x16b, res_temp3_8x16b);
1004 res_temp1_8x16b = _mm_srli_epi16(res_temp1_8x16b, 3); //log2(16)+1
1005 res_temp1_8x16b = _mm_packus_epi16(res_temp1_8x16b, zero_8x16b);
1007 temp1 = _mm_cvtsi128_si32(res_temp1_8x16b);
