Lines Matching refs:_mm_shuffle_epi8
373 src_temp4 = _mm_shuffle_epi8(src_temp4, m_mask);
404 src_temp4 = _mm_shuffle_epi8(src_temp4, m_mask);
428 src_temp4 = _mm_shuffle_epi8(src_temp4, m_mask);
842 _mm_storel_epi64((__m128i *)(pu1_dst + (0 * dst_strd)), _mm_shuffle_epi8(src_temp1, sm2));
843 _mm_storel_epi64((__m128i *)(pu1_dst + (1 * dst_strd)), _mm_shuffle_epi8(src_temp2, sm2));
844 _mm_storel_epi64((__m128i *)(pu1_dst + (2 * dst_strd)), _mm_shuffle_epi8(src_temp3, sm2));
845 _mm_storel_epi64((__m128i *)(pu1_dst + (3 * dst_strd)), _mm_shuffle_epi8(src_temp4, sm2));
860 _mm_storeu_si128((__m128i *)(pu1_dst + (0 * dst_strd)), _mm_shuffle_epi8(src_temp1, sm3));
861 _mm_storeu_si128((__m128i *)(pu1_dst + (1 * dst_strd)), _mm_shuffle_epi8(src_temp2, sm3));
862 _mm_storeu_si128((__m128i *)(pu1_dst + (2 * dst_strd)), _mm_shuffle_epi8(src_temp3, sm3));
863 _mm_storeu_si128((__m128i *)(pu1_dst + (3 * dst_strd)), _mm_shuffle_epi8(src_temp4, sm3));
864 _mm_storeu_si128((__m128i *)(pu1_dst + (4 * dst_strd)), _mm_shuffle_epi8(src_temp5, sm3));
865 _mm_storeu_si128((__m128i *)(pu1_dst + (5 * dst_strd)), _mm_shuffle_epi8(src_temp6, sm3));
866 _mm_storeu_si128((__m128i *)(pu1_dst + (6 * dst_strd)), _mm_shuffle_epi8(src_temp7, sm3));
867 _mm_storeu_si128((__m128i *)(pu1_dst + (7 * dst_strd)), _mm_shuffle_epi8(src_temp8, sm3));
886 _mm_shuffle_epi8(src_temp1, sm3));
887 _mm_storeu_si128((__m128i *)(pu1_dst + col + ((row + 1) * dst_strd)), _mm_shuffle_epi8(src_temp2, sm3));
888 _mm_storeu_si128((__m128i *)(pu1_dst + col + ((row + 2) * dst_strd)), _mm_shuffle_epi8(src_temp3, sm3));
889 _mm_storeu_si128((__m128i *)(pu1_dst + col + ((row + 3) * dst_strd)), _mm_shuffle_epi8(src_temp4, sm3));
890 _mm_storeu_si128((__m128i *)(pu1_dst + col + ((row + 4) * dst_strd)), _mm_shuffle_epi8(src_temp5, sm3));
891 _mm_storeu_si128((__m128i *)(pu1_dst + col + ((row + 5) * dst_strd)), _mm_shuffle_epi8(src_temp6, sm3));
892 _mm_storeu_si128((__m128i *)(pu1_dst + col + ((row + 6) * dst_strd)), _mm_shuffle_epi8(src_temp7, sm3));
893 _mm_storeu_si128((__m128i *)(pu1_dst + col + ((row + 7) * dst_strd)), _mm_shuffle_epi8(src_temp8, sm3));
1301 src_temp1_8x16b = _mm_shuffle_epi8(src_temp1_8x16b, sm1);
1302 src_temp2_8x16b = _mm_shuffle_epi8(src_temp2_8x16b, sm1);
1303 src_temp3_8x16b = _mm_shuffle_epi8(src_temp3_8x16b, sm1);
1304 src_temp4_8x16b = _mm_shuffle_epi8(src_temp4_8x16b, sm1);
1487 src_temp1_8x16b = _mm_shuffle_epi8(src_temp1_8x16b, sm1);
1488 src_temp2_8x16b = _mm_shuffle_epi8(src_temp2_8x16b, sm1);
1489 src_temp3_8x16b = _mm_shuffle_epi8(src_temp3_8x16b, sm1);
1490 src_temp4_8x16b = _mm_shuffle_epi8(src_temp4_8x16b, sm1);
1498 src_temp11_8x16b = _mm_shuffle_epi8(src_temp11_8x16b, sm1);
1499 src_temp12_8x16b = _mm_shuffle_epi8(src_temp12_8x16b, sm1);
1500 src_temp13_8x16b = _mm_shuffle_epi8(src_temp13_8x16b, sm1);
1501 src_temp14_8x16b = _mm_shuffle_epi8(src_temp14_8x16b, sm1);
