| /external/libhevc/common/x86/ |
| H A D | ihevc_weighted_pred_ssse3_intr.c | 754 __m128i src_temp1_8x16b, src_temp2_8x16b, src_temp3_8x16b, src_temp4_8x16b; local
796 src_temp4_8x16b = _mm_loadu_si128((__m128i *)(pi2_src2 + 1 * src_strd2)); /* row = 1 */
802 res_temp4_4x32b = _mm_mullo_epi16(src_temp4_8x16b, wgt1_8x16b);
807 src_temp4_8x16b = _mm_mulhi_epi16(src_temp4_8x16b, wgt1_8x16b);
813 res_temp8_4x32b = _mm_unpackhi_epi16(res_temp4_4x32b, src_temp4_8x16b);
818 res_temp4_4x32b = _mm_unpacklo_epi16(res_temp4_4x32b, src_temp4_8x16b);
883 src_temp4_8x16b = _mm_loadl_epi64((__m128i *)(pi2_src2 + 1 * src_strd2)); /* row = 1 */
887 src_temp2_8x16b = _mm_unpacklo_epi64(src_temp2_8x16b, src_temp4_8x16b);
1042 __m128i src_temp1_8x16b, src_temp2_8x16b, src_temp3_8x16b, src_temp4_8x16b; local
1309 __m128i src_temp1_8x16b, src_temp2_8x16b, src_temp3_8x16b, src_temp4_8x16b; local
1699 __m128i src_temp1_8x16b, src_temp2_8x16b, src_temp3_8x16b, src_temp4_8x16b; local
[all...] |
| H A D | ihevc_intra_pred_filters_sse42_intr.c | 1465 __m128i src_temp1_8x16b, src_temp2_8x16b, src_temp3_8x16b, src_temp4_8x16b; local
1507 src_temp4_8x16b = _mm_loadu_si128((__m128i *)(pu1_ref + ref_main_idx4 - 1)); /* col=3*/
1512 src_temp4_8x16b = _mm_shuffle_epi8(src_temp4_8x16b, sm3); /* col=3*/
1518 src_temp4_8x16b = _mm_maddubs_epi16(src_temp4_8x16b, temp4_8x16b);
1524 src_temp4_8x16b = _mm_add_epi16(src_temp4_8x16b, const_temp_4x32b);
1530 src_temp4_8x16b = _mm_srai_epi16(src_temp4_8x16b,
1624 __m128i src_temp1_8x16b, src_temp2_8x16b, src_temp3_8x16b, src_temp4_8x16b; local
1800 __m128i src_temp1_8x16b, src_temp2_8x16b, src_temp3_8x16b, src_temp4_8x16b; local
2010 __m128i src_temp1_8x16b, src_temp2_8x16b, src_temp3_8x16b, src_temp4_8x16b; local
2222 __m128i src_temp1_8x16b, src_temp2_8x16b, src_temp3_8x16b, src_temp4_8x16b; local
2426 __m128i src_temp1_8x16b, src_temp2_8x16b, src_temp3_8x16b, src_temp4_8x16b; local
2631 __m128i src_temp1_8x16b, src_temp2_8x16b, src_temp3_8x16b, src_temp4_8x16b; local
[all...] |
| H A D | ihevc_chroma_intra_pred_filters_ssse3_intr.c | 1258 __m128i src_temp1_8x16b, src_temp2_8x16b, src_temp3_8x16b, src_temp4_8x16b; local
1270 src_temp4_8x16b = _mm_srli_si128(src_temp8_8x16b, 2); /* col=3*/
1275 src_temp4_8x16b = _mm_unpacklo_epi8(src_temp4_8x16b, src_temp8_8x16b); /* col=3*/
1281 src_temp4_8x16b = _mm_maddubs_epi16(src_temp4_8x16b, temp4_8x16b);
1287 src_temp4_8x16b = _mm_add_epi16(src_temp4_8x16b, const_temp_4x32b);
1293 src_temp4_8x16b = _mm_srai_epi16(src_temp4_8x16b,
1407 __m128i src_temp1_8x16b, src_temp2_8x16b, src_temp3_8x16b, src_temp4_8x16b; local
1697 __m128i src_temp1_8x16b, src_temp2_8x16b, src_temp3_8x16b, src_temp4_8x16b; local
1838 __m128i src_temp1_8x16b, src_temp2_8x16b, src_temp3_8x16b, src_temp4_8x16b; local
[all...] |
| H A D | ihevc_intra_pred_filters_ssse3_intr.c | 2378 __m128i src_temp1_8x16b, src_temp2_8x16b, src_temp3_8x16b, src_temp4_8x16b; local
2426 src_temp4_8x16b = _mm_loadu_si128((__m128i *)(pu1_ref + ref_main_idx4 - 1)); /* col=3*/
2431 src_temp4_8x16b = _mm_shuffle_epi8(src_temp4_8x16b, sm3); /* col=3*/
2437 src_temp4_8x16b = _mm_maddubs_epi16(src_temp4_8x16b, temp4_8x16b);
2443 src_temp4_8x16b = _mm_add_epi16(src_temp4_8x16b, const_temp_4x32b);
2449 src_temp4_8x16b = _mm_srai_epi16(src_temp4_8x16b,
2543 __m128i src_temp1_8x16b, src_temp2_8x16b, src_temp3_8x16b, src_temp4_8x16b; local
2719 __m128i src_temp1_8x16b, src_temp2_8x16b, src_temp3_8x16b, src_temp4_8x16b; local
2930 __m128i src_temp1_8x16b, src_temp2_8x16b, src_temp3_8x16b, src_temp4_8x16b; local
3148 __m128i src_temp1_8x16b, src_temp2_8x16b, src_temp3_8x16b, src_temp4_8x16b; local
3352 __m128i src_temp1_8x16b, src_temp2_8x16b, src_temp3_8x16b, src_temp4_8x16b; local
3557 __m128i src_temp1_8x16b, src_temp2_8x16b, src_temp3_8x16b, src_temp4_8x16b; local
[all...] |
| H A D | ihevc_weighted_pred_sse42_intr.c | 1795 __m128i src_temp1_8x16b, src_temp2_8x16b, src_temp3_8x16b, src_temp4_8x16b; local
1829 src_temp4_8x16b = _mm_loadu_si128((__m128i *)(pi2_src2 + src_strd2));
1849 src_temp3_8x16b = _mm_adds_epi16(src_temp3_8x16b, src_temp4_8x16b);
1925 src_temp4_8x16b = _mm_loadu_si128((__m128i *)(pi2_src2 + src_strd2));
1935 src_temp3_8x16b = _mm_adds_epi16(src_temp3_8x16b, src_temp4_8x16b);
1992 src_temp4_8x16b = _mm_loadl_epi64((__m128i *)(pi2_src2 + src_strd2));
2002 src_temp2_8x16b = _mm_unpacklo_epi64(src_temp2_8x16b, src_temp4_8x16b);
2025 src_temp4_8x16b = _mm_shuffle_epi32(src_temp5_8x16b, 1);
2032 dst3 = _mm_cvtsi128_si32(src_temp4_8x16b);
2070 src_temp4_8x16b
[all...] |