/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 169 src_temp2_8x16b = _mm_loadu_si128((__m128i *)(pi2_src + 2 * src_strd)); in ihevc_weighted_pred_uni_ssse3() 176 res_temp2_4x32b = _mm_mullo_epi16(src_temp2_8x16b, wgt0_8x16b); in ihevc_weighted_pred_uni_ssse3() 182 src_temp2_8x16b = _mm_mulhi_epi16(src_temp2_8x16b, wgt0_8x16b); in ihevc_weighted_pred_uni_ssse3() 188 res_temp6_4x32b = _mm_unpackhi_epi16(res_temp2_4x32b, src_temp2_8x16b); in ihevc_weighted_pred_uni_ssse3() 193 res_temp2_4x32b = _mm_unpacklo_epi16(res_temp2_4x32b, src_temp2_8x16b); in ihevc_weighted_pred_uni_ssse3() 271 src_temp2_8x16b = _mm_loadl_epi64((__m128i *)(pi2_src + 2 * src_strd)); in ihevc_weighted_pred_uni_ssse3() 276 src_temp0_8x16b = _mm_unpacklo_epi64(src_temp0_8x16b, src_temp2_8x16b); in ihevc_weighted_pred_uni_ssse3() 447 __m128i src_temp2_8x16b, src_temp3_8x16b; in ihevc_weighted_pred_chroma_uni_ssse3() local 461 src_temp2_8x16b = _mm_loadu_si128((__m128i *)(pi2_src + 8)); in ihevc_weighted_pred_chroma_uni_ssse3() [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 1487 src_temp2_8x16b = _mm_loadu_si128((__m128i *)(pu1_ref + ref_main_idx2 - 1)); /* col=1*/ in ihevc_intra_pred_luma_mode_3_to_9_sse42() 1492 src_temp2_8x16b = _mm_shuffle_epi8(src_temp2_8x16b, sm3); /* col=1*/ in ihevc_intra_pred_luma_mode_3_to_9_sse42() 1498 src_temp2_8x16b = _mm_maddubs_epi16(src_temp2_8x16b, temp2_8x16b); in ihevc_intra_pred_luma_mode_3_to_9_sse42() 1504 src_temp2_8x16b = _mm_add_epi16(src_temp2_8x16b, const_temp_4x32b); in ihevc_intra_pred_luma_mode_3_to_9_sse42() 1510 src_temp2_8x16b = _mm_srai_epi16(src_temp2_8x16b, 5); /* col=1*/ in ihevc_intra_pred_luma_mode_3_to_9_sse42() 1516 src_temp2_8x16b = _mm_packus_epi16(src_temp2_8x16b, src_temp4_8x16b); /* col=1*/ 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() 1523 src_temp2_8x16b = _mm_srli_si128(src_temp3_8x16b, 4); in ihevc_intra_pred_luma_mode_3_to_9_sse42() [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 1255 src_temp2_8x16b = _mm_srli_si128(src_temp6_8x16b, 2); /* col=1*/ in ihevc_intra_pred_chroma_mode_3_to_9_ssse3() 1260 src_temp2_8x16b = _mm_unpacklo_epi8(src_temp2_8x16b, src_temp6_8x16b); /* col=1*/ in ihevc_intra_pred_chroma_mode_3_to_9_ssse3() 1266 src_temp2_8x16b = _mm_maddubs_epi16(src_temp2_8x16b, temp2_8x16b); in ihevc_intra_pred_chroma_mode_3_to_9_ssse3() 1272 src_temp2_8x16b = _mm_add_epi16(src_temp2_8x16b, const_temp_4x32b); in ihevc_intra_pred_chroma_mode_3_to_9_ssse3() 1278 src_temp2_8x16b = _mm_srai_epi16(src_temp2_8x16b, 5); /* col=1*/ in ihevc_intra_pred_chroma_mode_3_to_9_ssse3() 1284 src_temp2_8x16b = _mm_packus_epi16(src_temp2_8x16b, zero_8x16b); /* col=1*/ in ihevc_intra_pred_chroma_mode_3_to_9_ssse3() 1289 src_temp2_8x16b = _mm_shuffle_epi8(src_temp2_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() 1301 src_temp2_8x16b = _mm_shuffle_epi32(src_temp8_8x16b, _MM_SHUFFLE(3, 2, 3, 2)); in ihevc_intra_pred_chroma_mode_3_to_9_ssse3() [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 2405 src_temp2_8x16b = _mm_loadu_si128((__m128i *)(pu1_ref + ref_main_idx2 - 1)); /* col=1*/ in ihevc_intra_pred_luma_mode_3_to_9_ssse3() 2410 src_temp2_8x16b = _mm_shuffle_epi8(src_temp2_8x16b, sm3); /* col=1*/ in ihevc_intra_pred_luma_mode_3_to_9_ssse3() 2416 src_temp2_8x16b = _mm_maddubs_epi16(src_temp2_8x16b, temp2_8x16b); in ihevc_intra_pred_luma_mode_3_to_9_ssse3() 2422 src_temp2_8x16b = _mm_add_epi16(src_temp2_8x16b, const_temp_4x32b); in ihevc_intra_pred_luma_mode_3_to_9_ssse3() 2428 src_temp2_8x16b = _mm_srai_epi16(src_temp2_8x16b, 5); /* col=1*/ in ihevc_intra_pred_luma_mode_3_to_9_ssse3() 2434 src_temp2_8x16b = _mm_packus_epi16(src_temp2_8x16b, src_temp4_8x16b); /* col=1*/ in ihevc_intra_pred_luma_mode_3_to_9_ssse3() 2437 src_temp5_8x16b = _mm_unpacklo_epi8(src_temp1_8x16b, src_temp2_8x16b); in ihevc_intra_pred_luma_mode_3_to_9_ssse3() 2438 src_temp6_8x16b = _mm_unpackhi_epi8(src_temp1_8x16b, src_temp2_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() [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 1826 src_temp2_8x16b = _mm_loadu_si128((__m128i *)(pi2_src2)); 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() 1922 src_temp2_8x16b = _mm_loadu_si128((__m128i *)(pi2_src2)); 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() 1988 src_temp2_8x16b = _mm_loadl_epi64((__m128i *)(pi2_src2)); in ihevc_weighted_pred_bi_default_sse42() 2002 src_temp2_8x16b = _mm_unpacklo_epi64(src_temp2_8x16b, src_temp4_8x16b); in ihevc_weighted_pred_bi_default_sse42() 2007 src_temp1_8x16b = _mm_adds_epi16(src_temp1_8x16b, src_temp2_8x16b); in ihevc_weighted_pred_bi_default_sse42() 2024 src_temp2_8x16b = _mm_shuffle_epi32(src_temp1_8x16b, 1); in ihevc_weighted_pred_bi_default_sse42() 2030 dst1 = _mm_cvtsi128_si32(src_temp2_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 264 src_temp2_8x16b = _mm_loadu_si128((__m128i *)(pu1_src_cpy + src_strd)); in ihevc_sao_band_offset_luma_ssse3() 270 tmp_set_128i_3 = _mm_sub_epi8(src_temp2_8x16b, band_pos_16x8b); in ihevc_sao_band_offset_luma_ssse3() 312 src_temp2_8x16b = _mm_and_si128(src_temp2_8x16b, cmp_store); in ihevc_sao_band_offset_luma_ssse3() 324 src_temp2_8x16b = _mm_or_si128(src_temp2_8x16b, tmp_set_128i_3); in ihevc_sao_band_offset_luma_ssse3() 330 _mm_storeu_si128((__m128i *)(pu1_src_cpy + src_strd), src_temp2_8x16b); in ihevc_sao_band_offset_luma_ssse3() 348 src_temp2_8x16b = _mm_loadl_epi64((__m128i *)(pu1_src_cpy + 2 * 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() 358 tmp_set_128i_3 = _mm_sub_epi8(src_temp2_8x16b, band_pos_16x8b); in ihevc_sao_band_offset_luma_ssse3() 401 src_temp2_8x16b = _mm_and_si128(src_temp2_8x16b, cmp_store); in ihevc_sao_band_offset_luma_ssse3() [all …]
|