/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 171 src_temp3_8x16b = _mm_loadu_si128((__m128i *)(pi2_src + 3 * src_strd)); in ihevc_weighted_pred_uni_ssse3() 177 res_temp3_4x32b = _mm_mullo_epi16(src_temp3_8x16b, wgt0_8x16b); in ihevc_weighted_pred_uni_ssse3() 183 src_temp3_8x16b = _mm_mulhi_epi16(src_temp3_8x16b, wgt0_8x16b); in ihevc_weighted_pred_uni_ssse3() 189 res_temp7_4x32b = _mm_unpackhi_epi16(res_temp3_4x32b, src_temp3_8x16b); in ihevc_weighted_pred_uni_ssse3() 194 res_temp3_4x32b = _mm_unpacklo_epi16(res_temp3_4x32b, src_temp3_8x16b); in ihevc_weighted_pred_uni_ssse3() 273 src_temp3_8x16b = _mm_loadl_epi64((__m128i *)(pi2_src + 3 * 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() 447 __m128i src_temp2_8x16b, src_temp3_8x16b; in ihevc_weighted_pred_chroma_uni_ssse3() local 463 src_temp3_8x16b = _mm_loadu_si128((__m128i *)(pi2_src + src_strd + 8)); in ihevc_weighted_pred_chroma_uni_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 1256 src_temp3_8x16b = _mm_srli_si128(src_temp7_8x16b, 2); /* col=2*/ in ihevc_intra_pred_chroma_mode_3_to_9_ssse3() 1261 src_temp3_8x16b = _mm_unpacklo_epi8(src_temp3_8x16b, src_temp7_8x16b); /* col=2*/ in ihevc_intra_pred_chroma_mode_3_to_9_ssse3() 1267 src_temp3_8x16b = _mm_maddubs_epi16(src_temp3_8x16b, temp3_8x16b); in ihevc_intra_pred_chroma_mode_3_to_9_ssse3() 1273 src_temp3_8x16b = _mm_add_epi16(src_temp3_8x16b, const_temp_4x32b); in ihevc_intra_pred_chroma_mode_3_to_9_ssse3() 1279 src_temp3_8x16b = _mm_srai_epi16(src_temp3_8x16b, 5); /* col=2*/ in ihevc_intra_pred_chroma_mode_3_to_9_ssse3() 1285 src_temp3_8x16b = _mm_packus_epi16(src_temp3_8x16b, zero_8x16b); /* col=2*/ in ihevc_intra_pred_chroma_mode_3_to_9_ssse3() 1290 src_temp3_8x16b = _mm_shuffle_epi8(src_temp3_8x16b, sm1); in ihevc_intra_pred_chroma_mode_3_to_9_ssse3() 1294 src_temp6_8x16b = _mm_unpacklo_epi16(src_temp3_8x16b, src_temp4_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_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 1488 src_temp3_8x16b = _mm_loadu_si128((__m128i *)(pu1_ref + ref_main_idx3 - 1)); /* col=2*/ in ihevc_intra_pred_luma_mode_3_to_9_sse42() 1493 src_temp3_8x16b = _mm_shuffle_epi8(src_temp3_8x16b, sm3); /* col=2*/ in ihevc_intra_pred_luma_mode_3_to_9_sse42() 1499 src_temp3_8x16b = _mm_maddubs_epi16(src_temp3_8x16b, temp3_8x16b); in ihevc_intra_pred_luma_mode_3_to_9_sse42() 1505 src_temp3_8x16b = _mm_add_epi16(src_temp3_8x16b, const_temp_4x32b); in ihevc_intra_pred_luma_mode_3_to_9_sse42() 1511 src_temp3_8x16b = _mm_srai_epi16(src_temp3_8x16b, 5); /* col=2*/ 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() 1522 src_temp3_8x16b = _mm_unpacklo_epi16(src_temp5_8x16b, src_temp6_8x16b); in ihevc_intra_pred_luma_mode_3_to_9_sse42() 1523 src_temp2_8x16b = _mm_srli_si128(src_temp3_8x16b, 4); 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 | 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 2406 src_temp3_8x16b = _mm_loadu_si128((__m128i *)(pu1_ref + ref_main_idx3 - 1)); /* col=2*/ in ihevc_intra_pred_luma_mode_3_to_9_ssse3() 2411 src_temp3_8x16b = _mm_shuffle_epi8(src_temp3_8x16b, sm3); /* col=2*/ in ihevc_intra_pred_luma_mode_3_to_9_ssse3() 2417 src_temp3_8x16b = _mm_maddubs_epi16(src_temp3_8x16b, temp3_8x16b); in ihevc_intra_pred_luma_mode_3_to_9_ssse3() 2423 src_temp3_8x16b = _mm_add_epi16(src_temp3_8x16b, const_temp_4x32b); in ihevc_intra_pred_luma_mode_3_to_9_ssse3() 2429 src_temp3_8x16b = _mm_srai_epi16(src_temp3_8x16b, 5); /* col=2*/ in ihevc_intra_pred_luma_mode_3_to_9_ssse3() 2433 src_temp1_8x16b = _mm_packus_epi16(src_temp1_8x16b, src_temp3_8x16b); /* col=0*/ in ihevc_intra_pred_luma_mode_3_to_9_ssse3() 2440 src_temp3_8x16b = _mm_unpacklo_epi16(src_temp5_8x16b, src_temp6_8x16b); in ihevc_intra_pred_luma_mode_3_to_9_ssse3() 2441 src_temp2_8x16b = _mm_srli_si128(src_temp3_8x16b, 4); in ihevc_intra_pred_luma_mode_3_to_9_ssse3() 2442 src_temp1_8x16b = _mm_srli_si128(src_temp3_8x16b, 8); in ihevc_intra_pred_luma_mode_3_to_9_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 1828 src_temp3_8x16b = _mm_loadu_si128((__m128i *)(pi2_src1 + src_strd1)); in ihevc_weighted_pred_bi_default_sse42() 1849 src_temp3_8x16b = _mm_adds_epi16(src_temp3_8x16b, src_temp4_8x16b); in ihevc_weighted_pred_bi_default_sse42() 1860 src_temp3_8x16b = _mm_adds_epi16(src_temp3_8x16b, lvl_shift1_8x16b); in ihevc_weighted_pred_bi_default_sse42() 1872 src_temp3_8x16b = _mm_srai_epi16(src_temp3_8x16b, shift); in ihevc_weighted_pred_bi_default_sse42() 1890 src_temp3_8x16b = _mm_packus_epi16(src_temp3_8x16b, src_temp11_8x16b); in ihevc_weighted_pred_bi_default_sse42() 1896 … _mm_storeu_si128((__m128i *)(pu1_dst + 1 * dst_strd), src_temp3_8x16b); /* row = 2*/ in ihevc_weighted_pred_bi_default_sse42() 1924 src_temp3_8x16b = _mm_loadu_si128((__m128i *)(pi2_src1 + src_strd1)); in ihevc_weighted_pred_bi_default_sse42() 1935 src_temp3_8x16b = _mm_adds_epi16(src_temp3_8x16b, src_temp4_8x16b); in ihevc_weighted_pred_bi_default_sse42() 1941 src_temp3_8x16b = _mm_adds_epi16(src_temp3_8x16b, lvl_shift1_8x16b); in ihevc_weighted_pred_bi_default_sse42() [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 350 src_temp3_8x16b = _mm_loadl_epi64((__m128i *)(pu1_src_cpy + 3 * src_strd)); in ihevc_sao_band_offset_luma_ssse3() 354 src_temp2_8x16b = _mm_unpacklo_epi64(src_temp2_8x16b, src_temp3_8x16b); in ihevc_sao_band_offset_luma_ssse3() 418 src_temp3_8x16b = _mm_srli_si128(src_temp2_8x16b, 8); in ihevc_sao_band_offset_luma_ssse3() 427 _mm_storel_epi64((__m128i *)(pu1_src_cpy + 3 * src_strd), src_temp3_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 672 src_temp3_8x16b = _mm_loadu_si128((__m128i *)(pu1_src_cpy + src_strd)); in ihevc_sao_band_offset_chroma_ssse3() 677 src_temp2_8x16b = _mm_srli_epi16(src_temp3_8x16b, 8); in ihevc_sao_band_offset_chroma_ssse3() 680 src_temp3_8x16b = _mm_slli_epi16(src_temp3_8x16b, 8); in ihevc_sao_band_offset_chroma_ssse3() 682 src_temp3_8x16b = _mm_srli_epi16(src_temp3_8x16b, 8); in ihevc_sao_band_offset_chroma_ssse3() [all …]
|