/aosp_15_r20/external/libhevc/common/x86/ |
H A D | ihevc_weighted_pred_ssse3_intr.c | 128 __m128i src_temp0_8x16b, src_temp1_8x16b, src_temp2_8x16b, src_temp3_8x16b; in ihevc_weighted_pred_uni_ssse3() local 167 src_temp1_8x16b = _mm_loadu_si128((__m128i *)(pi2_src + src_strd)); in ihevc_weighted_pred_uni_ssse3() 175 res_temp1_4x32b = _mm_mullo_epi16(src_temp1_8x16b, wgt0_8x16b); in ihevc_weighted_pred_uni_ssse3() 181 src_temp1_8x16b = _mm_mulhi_epi16(src_temp1_8x16b, wgt0_8x16b); in ihevc_weighted_pred_uni_ssse3() 187 res_temp5_4x32b = _mm_unpackhi_epi16(res_temp1_4x32b, src_temp1_8x16b); in ihevc_weighted_pred_uni_ssse3() 192 res_temp1_4x32b = _mm_unpacklo_epi16(res_temp1_4x32b, src_temp1_8x16b); in ihevc_weighted_pred_uni_ssse3() 269 src_temp1_8x16b = _mm_loadl_epi64((__m128i *)(pi2_src + src_strd)); in ihevc_weighted_pred_uni_ssse3() 277 src_temp1_8x16b = _mm_unpacklo_epi64(src_temp1_8x16b, src_temp3_8x16b); in ihevc_weighted_pred_uni_ssse3() 281 res_temp1_4x32b = _mm_mullo_epi16(src_temp1_8x16b, wgt0_8x16b); in ihevc_weighted_pred_uni_ssse3() 284 src_temp1_8x16b = _mm_mulhi_epi16(src_temp1_8x16b, wgt0_8x16b); in ihevc_weighted_pred_uni_ssse3() [all …]
|
H A D | ihevc_weighted_pred_sse42_intr.c | 1795 __m128i src_temp1_8x16b, src_temp2_8x16b, src_temp3_8x16b, src_temp4_8x16b; in ihevc_weighted_pred_bi_default_sse42() local 1825 src_temp1_8x16b = _mm_loadu_si128((__m128i *)(pi2_src1)); in ihevc_weighted_pred_bi_default_sse42() 1848 src_temp1_8x16b = _mm_adds_epi16(src_temp1_8x16b, src_temp2_8x16b); in ihevc_weighted_pred_bi_default_sse42() 1859 src_temp1_8x16b = _mm_adds_epi16(src_temp1_8x16b, lvl_shift1_8x16b); in ihevc_weighted_pred_bi_default_sse42() 1871 src_temp1_8x16b = _mm_srai_epi16(src_temp1_8x16b, shift); in ihevc_weighted_pred_bi_default_sse42() 1889 src_temp1_8x16b = _mm_packus_epi16(src_temp1_8x16b, src_temp9_8x16b); in ihevc_weighted_pred_bi_default_sse42() 1895 … _mm_storeu_si128((__m128i *)(pu1_dst + 0 * dst_strd), src_temp1_8x16b); /* row = 0*/ in ihevc_weighted_pred_bi_default_sse42() 1921 src_temp1_8x16b = _mm_loadu_si128((__m128i *)(pi2_src1)); in ihevc_weighted_pred_bi_default_sse42() 1934 src_temp1_8x16b = _mm_adds_epi16(src_temp1_8x16b, src_temp2_8x16b); in ihevc_weighted_pred_bi_default_sse42() 1940 src_temp1_8x16b = _mm_adds_epi16(src_temp1_8x16b, lvl_shift1_8x16b); in ihevc_weighted_pred_bi_default_sse42() [all …]
|
H A D | ihevc_intra_pred_filters_sse42_intr.c | 1447 __m128i src_temp1_8x16b, src_temp2_8x16b, src_temp3_8x16b, src_temp4_8x16b; in ihevc_intra_pred_luma_mode_3_to_9_sse42() local 1486 src_temp1_8x16b = _mm_loadu_si128((__m128i *)(pu1_ref + ref_main_idx1 - 1)); /* col=0*/ in ihevc_intra_pred_luma_mode_3_to_9_sse42() 1491 src_temp1_8x16b = _mm_shuffle_epi8(src_temp1_8x16b, sm3); /* col=0*/ in ihevc_intra_pred_luma_mode_3_to_9_sse42() 1497 src_temp1_8x16b = _mm_maddubs_epi16(src_temp1_8x16b, temp1_8x16b); in ihevc_intra_pred_luma_mode_3_to_9_sse42() 1503 src_temp1_8x16b = _mm_add_epi16(src_temp1_8x16b, const_temp_4x32b); in ihevc_intra_pred_luma_mode_3_to_9_sse42() 1509 src_temp1_8x16b = _mm_srai_epi16(src_temp1_8x16b, 5); /* col=0*/ in ihevc_intra_pred_luma_mode_3_to_9_sse42() 1515 src_temp1_8x16b = _mm_packus_epi16(src_temp1_8x16b, src_temp3_8x16b); /* col=0*/ in ihevc_intra_pred_luma_mode_3_to_9_sse42() 1519 src_temp5_8x16b = _mm_unpacklo_epi8(src_temp1_8x16b, src_temp2_8x16b); in ihevc_intra_pred_luma_mode_3_to_9_sse42() 1520 src_temp6_8x16b = _mm_unpackhi_epi8(src_temp1_8x16b, src_temp2_8x16b); in ihevc_intra_pred_luma_mode_3_to_9_sse42() 1524 src_temp1_8x16b = _mm_srli_si128(src_temp3_8x16b, 8); in ihevc_intra_pred_luma_mode_3_to_9_sse42() [all …]
|
H A D | ihevc_intra_pred_filters_ssse3_intr.c | 714 __m128i src_temp_8x16b, src_temp1_8x16b; in ihevc_intra_pred_luma_planar_ssse3() local 727 src_temp1_8x16b = _mm_loadu_si128((__m128i *)(pu1_ref + two_nt + 17)); in ihevc_intra_pred_luma_planar_ssse3() 731 res_temp6_8x16b = _mm_unpacklo_epi8(res_temp1_8x16b, src_temp1_8x16b); /* row=2*/ in ihevc_intra_pred_luma_planar_ssse3() 732 res_temp7_8x16b = _mm_unpackhi_epi8(res_temp1_8x16b, src_temp1_8x16b); /* row=3*/ in ihevc_intra_pred_luma_planar_ssse3() 2359 __m128i src_temp1_8x16b, src_temp2_8x16b, src_temp3_8x16b, src_temp4_8x16b; in ihevc_intra_pred_luma_mode_3_to_9_ssse3() local 2404 src_temp1_8x16b = _mm_loadu_si128((__m128i *)(pu1_ref + ref_main_idx1 - 1)); /* col=0*/ in ihevc_intra_pred_luma_mode_3_to_9_ssse3() 2409 src_temp1_8x16b = _mm_shuffle_epi8(src_temp1_8x16b, sm3); /* col=0*/ in ihevc_intra_pred_luma_mode_3_to_9_ssse3() 2415 src_temp1_8x16b = _mm_maddubs_epi16(src_temp1_8x16b, temp1_8x16b); in ihevc_intra_pred_luma_mode_3_to_9_ssse3() 2421 src_temp1_8x16b = _mm_add_epi16(src_temp1_8x16b, const_temp_4x32b); in ihevc_intra_pred_luma_mode_3_to_9_ssse3() 2427 src_temp1_8x16b = _mm_srai_epi16(src_temp1_8x16b, 5); /* col=0*/ in ihevc_intra_pred_luma_mode_3_to_9_ssse3() [all …]
|
H A D | ihevc_chroma_intra_pred_filters_ssse3_intr.c | 1245 __m128i src_temp1_8x16b, src_temp2_8x16b, src_temp3_8x16b, src_temp4_8x16b; in ihevc_intra_pred_chroma_mode_3_to_9_ssse3() local 1254 src_temp1_8x16b = _mm_srli_si128(src_temp5_8x16b, 2); /* col=0*/ in ihevc_intra_pred_chroma_mode_3_to_9_ssse3() 1259 src_temp1_8x16b = _mm_unpacklo_epi8(src_temp1_8x16b, src_temp5_8x16b); /* col=0*/ in ihevc_intra_pred_chroma_mode_3_to_9_ssse3() 1265 src_temp1_8x16b = _mm_maddubs_epi16(src_temp1_8x16b, temp1_8x16b); in ihevc_intra_pred_chroma_mode_3_to_9_ssse3() 1271 src_temp1_8x16b = _mm_add_epi16(src_temp1_8x16b, const_temp_4x32b); in ihevc_intra_pred_chroma_mode_3_to_9_ssse3() 1277 src_temp1_8x16b = _mm_srai_epi16(src_temp1_8x16b, 5); /* col=0*/ in ihevc_intra_pred_chroma_mode_3_to_9_ssse3() 1283 src_temp1_8x16b = _mm_packus_epi16(src_temp1_8x16b, zero_8x16b); /* col=0*/ in ihevc_intra_pred_chroma_mode_3_to_9_ssse3() 1288 src_temp1_8x16b = _mm_shuffle_epi8(src_temp1_8x16b, sm1); in ihevc_intra_pred_chroma_mode_3_to_9_ssse3() 1293 src_temp5_8x16b = _mm_unpacklo_epi16(src_temp1_8x16b, src_temp2_8x16b); in ihevc_intra_pred_chroma_mode_3_to_9_ssse3() 1394 __m128i src_temp1_8x16b, src_temp2_8x16b, src_temp3_8x16b, src_temp4_8x16b; in ihevc_intra_pred_chroma_mode_3_to_9_ssse3() local [all …]
|
H A D | ihevc_sao_ssse3_intr.c | 136 __m128i src_temp0_8x16b, src_temp1_8x16b, src_temp2_8x16b, src_temp3_8x16b; in ihevc_sao_band_offset_luma_ssse3() local 346 src_temp1_8x16b = _mm_loadl_epi64((__m128i *)(pu1_src_cpy + src_strd)); in ihevc_sao_band_offset_luma_ssse3() 353 src_temp0_8x16b = _mm_unpacklo_epi64(src_temp0_8x16b, src_temp1_8x16b); in ihevc_sao_band_offset_luma_ssse3() 416 src_temp1_8x16b = _mm_srli_si128(src_temp0_8x16b, 8); in ihevc_sao_band_offset_luma_ssse3() 423 _mm_storel_epi64((__m128i *)(pu1_src_cpy + src_strd), src_temp1_8x16b); in ihevc_sao_band_offset_luma_ssse3() 454 __m128i src_temp0_8x16b, src_temp1_8x16b, src_temp2_8x16b, src_temp3_8x16b; in ihevc_sao_band_offset_chroma_ssse3() local 676 src_temp1_8x16b = _mm_srli_epi16(src_temp0_8x16b, 8); in ihevc_sao_band_offset_chroma_ssse3() 684 src_temp2_8x16b = _mm_packus_epi16(src_temp1_8x16b, src_temp2_8x16b); in ihevc_sao_band_offset_chroma_ssse3() 703 src_temp1_8x16b = _mm_cmpgt_epi8(tmp_set_128i_1, cmp_mask); in ihevc_sao_band_offset_chroma_ssse3() 708 tmp_set_128i_2 = _mm_and_si128(tmp_set_128i_1, src_temp1_8x16b); in ihevc_sao_band_offset_chroma_ssse3() [all …]
|