/aosp_15_r20/external/libavc/common/x86/svc/ |
H A D | isvc_resi_trans_quant_sse42.c | 148 __m128i src_r0, src_r1, src_r2, src_r3; in isvc_resi_trans_quant_4x4_sse42() local 201 src_r3 = _mm_loadl_epi64((__m128i *) (&pu1_src[3 * i4_src_stride])); in isvc_resi_trans_quant_4x4_sse42() 206 src_r3 = _mm_cvtepu8_epi16(src_r3); in isvc_resi_trans_quant_4x4_sse42() 232 src_r3 = _mm_sub_epi16(src_r3, pred_r3); in isvc_resi_trans_quant_4x4_sse42() 248 temp2 = _mm_unpacklo_epi16(src_r2, src_r3); in isvc_resi_trans_quant_4x4_sse42() 261 src_r3 = _mm_unpackhi_epi64(temp3, zero_8x16b); in isvc_resi_trans_quant_4x4_sse42() 265 temp0 = _mm_add_epi16(src_r0, src_r3); in isvc_resi_trans_quant_4x4_sse42() 271 temp3 = _mm_sub_epi16(src_r0, src_r3); in isvc_resi_trans_quant_4x4_sse42() 281 src_r3 = _mm_slli_epi16(temp2, 1); in isvc_resi_trans_quant_4x4_sse42() 282 src_r3 = _mm_sub_epi16(temp3, src_r3); in isvc_resi_trans_quant_4x4_sse42() [all …]
|
H A D | isvc_iquant_itrans_recon_sse42.c | 121 __m128i src_r0, src_r1, src_r2, src_r3; in isvc_iquant_itrans_recon_4x4_sse42() local 198 src_r3 = _mm_unpackhi_epi16(src_r2_r3, zero_8x16b); in isvc_iquant_itrans_recon_4x4_sse42() 203 temp7 = _mm_madd_epi16(src_r3, temp7); in isvc_iquant_itrans_recon_4x4_sse42() 416 __m128i src_r0, src_r1, src_r2, src_r3; in isvc_iquant_itrans_recon_res_4x4_sse42() local 492 src_r3 = _mm_unpackhi_epi16(src_r2_r3, zero_8x16b); in isvc_iquant_itrans_recon_res_4x4_sse42() 497 temp7 = _mm_madd_epi16(src_r3, temp7); in isvc_iquant_itrans_recon_res_4x4_sse42() 727 __m128i src_r0, src_r1, src_r2, src_r3; in isvc_iquant_itrans_recon_res_4x4_with_res_acc_sse42() local 805 src_r3 = _mm_unpackhi_epi16(src_r2_r3, zero_8x16b); in isvc_iquant_itrans_recon_res_4x4_with_res_acc_sse42() 810 temp7 = _mm_madd_epi16(src_r3, temp7); in isvc_iquant_itrans_recon_res_4x4_with_res_acc_sse42() 1056 __m128i src_r0, src_r1, src_r2, src_r3; in isvc_iquant_itrans_recon_res_chroma_4x4_sse42() local [all …]
|
H A D | isvc_iquant_itrans_recon_ssse3.c | 123 __m128i src_r0, src_r1, src_r2, src_r3; in isvc_iquant_itrans_recon_4x4_ssse3() local 188 src_r3 = _mm_unpackhi_epi16(src_r2_r3, zero_8x16b); // a30 0 a31 0 a32 0 a33 0 -- 16 bit long in isvc_iquant_itrans_recon_4x4_ssse3() 194 temp7 = _mm_madd_epi16(src_r3, temp7); in isvc_iquant_itrans_recon_4x4_ssse3()
|
/aosp_15_r20/external/libavc/common/x86/ |
H A D | ih264_resi_trans_quant_sse42.c | 119 __m128i src_r0, src_r1, src_r2, src_r3; in ih264_resi_trans_quant_4x4_sse42() local 133 …src_r3 = _mm_loadl_epi64((__m128i *) (&pu1_src[3 * src_strd])); //a30 a31 a32 a33 0 0 0 0 0 0 0 0 … in ih264_resi_trans_quant_4x4_sse42() 138 src_r3 = _mm_cvtepu8_epi16(src_r3); in ih264_resi_trans_quant_4x4_sse42() 153 src_r3 = _mm_sub_epi16(src_r3, pred_r3); in ih264_resi_trans_quant_4x4_sse42() 167 temp2 = _mm_unpacklo_epi16(src_r2, src_r3); //c0 d0 c1 d1 c2 d2 c3 d3 in ih264_resi_trans_quant_4x4_sse42() 174 src_r3 = _mm_unpackhi_epi64(temp3, zero_8x16b); //a3 b3 c3 d3 in ih264_resi_trans_quant_4x4_sse42() 178 temp0 = _mm_add_epi16(src_r0, src_r3); in ih264_resi_trans_quant_4x4_sse42() 184 temp3 = _mm_sub_epi16(src_r0, src_r3); in ih264_resi_trans_quant_4x4_sse42() 194 src_r3 = _mm_slli_epi16(temp2, 1); //(x2<<1) in ih264_resi_trans_quant_4x4_sse42() 195 src_r3 = _mm_sub_epi16(temp3, src_r3); in ih264_resi_trans_quant_4x4_sse42() [all …]
|
H A D | ih264_ihadamard_scaling_ssse3.c | 97 __m128i src_r0, src_r1, src_r2, src_r3; in ih264_ihadamard_scaling_4x4_ssse3() local 114 src_r3 = _mm_unpackhi_epi16(src_r2_r3, sign_reg); in ih264_ihadamard_scaling_4x4_ssse3() 128 temp2 = _mm_unpacklo_epi32(src_r2, src_r3); //c0 d0 c1 d1 in ih264_ihadamard_scaling_4x4_ssse3() 130 temp3 = _mm_unpackhi_epi32(src_r2, src_r3); //c2 d2 c3 d3 in ih264_ihadamard_scaling_4x4_ssse3() 134 src_r3 = _mm_unpackhi_epi64(temp1, temp3); //a3 b3 c3 d3 in ih264_ihadamard_scaling_4x4_ssse3() 136 temp0 = _mm_add_epi32(src_r0, src_r3); in ih264_ihadamard_scaling_4x4_ssse3() 139 temp3 = _mm_sub_epi32(src_r0, src_r3); in ih264_ihadamard_scaling_4x4_ssse3() 144 src_r3 = _mm_sub_epi32(temp3, temp2); in ih264_ihadamard_scaling_4x4_ssse3() 157 temp2 = _mm_unpacklo_epi32(src_r2, src_r3); //a2 a3 b2 b3 in ih264_ihadamard_scaling_4x4_ssse3() 159 temp3 = _mm_unpackhi_epi32(src_r2, src_r3); //c2 c3 d2 d3 in ih264_ihadamard_scaling_4x4_ssse3() [all …]
|
H A D | ih264_ihadamard_scaling_sse42.c | 97 __m128i src_r0, src_r1, src_r2, src_r3; in ih264_ihadamard_scaling_4x4_sse42() local 112 src_r3 = _mm_cvtepi16_epi32(src_r2_r3); in ih264_ihadamard_scaling_4x4_sse42() 126 temp2 = _mm_unpacklo_epi32(src_r2, src_r3); //c0 d0 c1 d1 in ih264_ihadamard_scaling_4x4_sse42() 128 temp3 = _mm_unpackhi_epi32(src_r2, src_r3); //c2 d2 c3 d3 in ih264_ihadamard_scaling_4x4_sse42() 132 src_r3 = _mm_unpackhi_epi64(temp1, temp3); //a3 b3 c3 d3 in ih264_ihadamard_scaling_4x4_sse42() 134 temp0 = _mm_add_epi32(src_r0, src_r3); in ih264_ihadamard_scaling_4x4_sse42() 137 temp3 = _mm_sub_epi32(src_r0, src_r3); in ih264_ihadamard_scaling_4x4_sse42() 142 src_r3 = _mm_sub_epi32(temp3, temp2); in ih264_ihadamard_scaling_4x4_sse42() 155 temp2 = _mm_unpacklo_epi32(src_r2, src_r3); //a2 a3 b2 b3 in ih264_ihadamard_scaling_4x4_sse42() 157 temp3 = _mm_unpackhi_epi32(src_r2, src_r3); //c2 c3 d2 d3 in ih264_ihadamard_scaling_4x4_sse42() [all …]
|
H A D | ih264_iquant_itrans_recon_sse42.c | 114 __m128i src_r0, src_r1, src_r2, src_r3; in ih264_iquant_itrans_recon_4x4_sse42() local 147 src_r3 = _mm_unpackhi_epi16(src_r2_r3, zero_8x16b); // a30 0 a31 0 a32 0 a33 0 -- 16 bit long in ih264_iquant_itrans_recon_4x4_sse42() 152 temp7 = _mm_madd_epi16(src_r3, temp7); in ih264_iquant_itrans_recon_4x4_sse42() 363 __m128i src_r0, src_r1, src_r2, src_r3; in ih264_iquant_itrans_recon_chroma_4x4_sse42() local 398 src_r3 = _mm_unpackhi_epi16(src_r2_r3, zero_8x16b); // a30 0 a31 0 a32 0 a33 0 -- 16 bit long in ih264_iquant_itrans_recon_chroma_4x4_sse42() 403 temp7 = _mm_madd_epi16(src_r3, temp7); in ih264_iquant_itrans_recon_chroma_4x4_sse42()
|
H A D | ih264_iquant_itrans_recon_ssse3.c | 114 __m128i src_r0, src_r1, src_r2, src_r3; in ih264_iquant_itrans_recon_4x4_ssse3() local 148 src_r3 = _mm_unpackhi_epi16(src_r2_r3, zero_8x16b); // a30 0 a31 0 a32 0 a33 0 -- 16 bit long in ih264_iquant_itrans_recon_4x4_ssse3() 153 temp7 = _mm_madd_epi16(src_r3, temp7); in ih264_iquant_itrans_recon_4x4_ssse3()
|
/aosp_15_r20/external/libavc/encoder/x86/ |
H A D | ime_distortion_metrics_sse42.c | 108 __m128i src_r0, src_r1, src_r2, src_r3; in ime_compute_sad_16x16_sse42() local 119 src_r3 = _mm_loadu_si128((__m128i *) (pu1_src + 3*src_strd)); in ime_compute_sad_16x16_sse42() 129 res_r3 = _mm_sad_epu8(src_r3, est_r3); in ime_compute_sad_16x16_sse42() 142 src_r3 = _mm_loadu_si128((__m128i *) (pu1_src + 3*src_strd)); in ime_compute_sad_16x16_sse42() 152 res_r3 = _mm_sad_epu8(src_r3, est_r3); in ime_compute_sad_16x16_sse42() 165 src_r3 = _mm_loadu_si128((__m128i *) (pu1_src + 3*src_strd)); in ime_compute_sad_16x16_sse42() 175 res_r3 = _mm_sad_epu8(src_r3, est_r3); in ime_compute_sad_16x16_sse42() 188 src_r3 = _mm_loadu_si128((__m128i *) (pu1_src + 3*src_strd)); in ime_compute_sad_16x16_sse42() 198 res_r3 = _mm_sad_epu8(src_r3, est_r3); in ime_compute_sad_16x16_sse42() 252 __m128i src_r0, src_r1, src_r2, src_r3; in ime_compute_sad_16x8_sse42() local [all …]
|
/aosp_15_r20/external/libmpeg2/common/x86/ |
H A D | impeg2_inter_pred_sse42_intr.c | 76 __m128i src_r0, src_r1, src_r2, src_r3; in impeg2_copy_mb_sse42() local 87 src_r3 = _mm_loadu_si128((__m128i *) (src + 3 * src_wd)); in impeg2_copy_mb_sse42() 92 _mm_storeu_si128((__m128i *) (dst + 3 * dst_wd), src_r3); in impeg2_copy_mb_sse42() 100 src_r3 = _mm_loadu_si128((__m128i *) (src + 3 * src_wd)); in impeg2_copy_mb_sse42() 105 _mm_storeu_si128((__m128i *) (dst + 3 * dst_wd), src_r3); in impeg2_copy_mb_sse42() 113 src_r3 = _mm_loadu_si128((__m128i *) (src + 3 * src_wd)); in impeg2_copy_mb_sse42() 118 _mm_storeu_si128((__m128i *) (dst + 3 * dst_wd), src_r3); in impeg2_copy_mb_sse42() 126 src_r3 = _mm_loadu_si128((__m128i *) (src + 3 * src_wd)); in impeg2_copy_mb_sse42() 131 _mm_storeu_si128((__m128i *) (dst + 3 * dst_wd), src_r3); in impeg2_copy_mb_sse42() 146 src_r3 = _mm_loadl_epi64((__m128i *)(src + 3 * src_wd)); in impeg2_copy_mb_sse42() [all …]
|
/aosp_15_r20/external/libavc/encoder/x86/svc/ |
H A D | isvce_residual_pred_sse42.c | 579 __m128i src_r0, src_r1, src_r2, src_r3; in isvce_get_sad_with_residual_pred_sse42() local 599 src_r3 = _mm_loadl_epi64((__m128i *) (pu1_src + 8)); in isvce_get_sad_with_residual_pred_sse42() 636 src_r3 = _mm_cvtepu8_epi16(src_r3); in isvce_get_sad_with_residual_pred_sse42() 674 src_r3 = _mm_sub_epi16(src_r3, pred_r3); in isvce_get_sad_with_residual_pred_sse42() 683 src_r3 = _mm_sub_epi16(src_r3, res_r3); in isvce_get_sad_with_residual_pred_sse42() 692 src_r3 = _mm_abs_epi16(src_r3); in isvce_get_sad_with_residual_pred_sse42() 699 src_r1 = _mm_adds_epu16(src_r2, src_r3); in isvce_get_sad_with_residual_pred_sse42() 701 src_r3 = _mm_adds_epu16(src_r6, src_r7); in isvce_get_sad_with_residual_pred_sse42() 704 src_r1 = _mm_adds_epu16(src_r2, src_r3); in isvce_get_sad_with_residual_pred_sse42()
|
/aosp_15_r20/external/libavc/decoder/x86/svc/ |
H A D | isvcd_iquant_itrans_sse42.c | 205 __m128i src_r0, src_r1, src_r2, src_r3; in isvcd_iquant_itrans_chroma_4x4_dc_sse42() local 225 src_r3 = _mm_loadu_si128((__m128i *) (pi2_out + (3 * out_strd))); in isvcd_iquant_itrans_chroma_4x4_dc_sse42() 234 src_r3 = _mm_and_si128(src_r3, chroma_mask_odd); in isvcd_iquant_itrans_chroma_4x4_dc_sse42() 241 src_r3 = _mm_add_epi16(src_r3, i_macro); in isvcd_iquant_itrans_chroma_4x4_dc_sse42() 246 _mm_storeu_si128((__m128i *) (&pi2_out[3 * out_strd]), src_r3); in isvcd_iquant_itrans_chroma_4x4_dc_sse42() 277 __m128i src_r0, src_r1, src_r2, src_r3; in isvcd_iquant_itrans_4x4_sse42() local 323 src_r3 = _mm_unpackhi_epi16(src_r2_r3, zero_8x16b); // a30 0 a31 0 a32 0 a33 0 -- 16 bit long in isvcd_iquant_itrans_4x4_sse42() 329 temp7 = _mm_madd_epi16(src_r3, temp7); in isvcd_iquant_itrans_4x4_sse42() 1251 __m128i src_r0, src_r1, src_r2, src_r3; in isvcd_iquant_itrans_chroma_4x4_sse42() local 1301 src_r3 = _mm_unpackhi_epi16(src_r2_r3, zero_8x16b); // a30 0 a31 0 a32 0 a33 0 -- 16 bit long in isvcd_iquant_itrans_chroma_4x4_sse42() [all …]
|
H A D | isvcd_pred_residual_recon_sse42.c | 682 __m128i src_r0, src_r1, src_r2, src_r3; in isvcd_pred_residual_recon_chroma_4x4_sse42() local 705 src_r3 = _mm_loadu_si128((__m128i *) (pu1_out + (3 * out_strd))); in isvcd_pred_residual_recon_chroma_4x4_sse42() 730 src_r3 = _mm_and_si128(src_r3, chroma_mask_odd); in isvcd_pred_residual_recon_chroma_4x4_sse42() 740 src_r3 = _mm_add_epi8(src_r3, pred3); in isvcd_pred_residual_recon_chroma_4x4_sse42() 745 _mm_storel_epi64((__m128i *) (&pu1_out[3 * out_strd]), src_r3); in isvcd_pred_residual_recon_chroma_4x4_sse42() 773 __m128i src_r0, src_r1, src_r2, src_r3, src_r4, src_r5, src_r6, src_r7; in isvcd_pred_residual_recon_chroma_8x8_sse42() local 803 src_r3 = _mm_loadu_si128((__m128i *) (pu1_out + (3 * out_strd))); in isvcd_pred_residual_recon_chroma_8x8_sse42() 844 src_r3 = _mm_and_si128(src_r3, chroma_mask_odd); in isvcd_pred_residual_recon_chroma_8x8_sse42() 862 src_r3 = _mm_add_epi8(src_r3, pred3); in isvcd_pred_residual_recon_chroma_8x8_sse42() 871 _mm_storel_epi64((__m128i *) (&pu1_out[3 * out_strd]), src_r3); in isvcd_pred_residual_recon_chroma_8x8_sse42() [all …]
|
H A D | isvcd_iquant_itrans_residual_sse42.c | 88 __m128i src_r0, src_r1, src_r2, src_r3; in isvcd_iquant_itrans_residual_4x4_sse42() local 136 src_r3 = _mm_unpackhi_epi16(src_r2_r3, zero_8x16b); // a30 0 a31 0 a32 0 a33 0 -- 16 bit long in isvcd_iquant_itrans_residual_4x4_sse42() 142 temp7 = _mm_madd_epi16(src_r3, temp7); in isvcd_iquant_itrans_residual_4x4_sse42() 1342 __m128i src_r0, src_r1, src_r2, src_r3; in isvcd_iquant_itrans_residual_chroma_4x4_sse42() local 1394 src_r3 = _mm_unpackhi_epi16(src_r2_r3, zero_8x16b); // a30 0 a31 0 a32 0 a33 0 -- 16 bit long in isvcd_iquant_itrans_residual_chroma_4x4_sse42() 1400 temp7 = _mm_madd_epi16(src_r3, temp7); in isvcd_iquant_itrans_residual_chroma_4x4_sse42()
|
H A D | isvcd_iquant_itrans_residual_recon_sse42.c | 87 __m128i src_r0, src_r1, src_r2, src_r3; in isvcd_iquant_itrans_residual_recon_4x4_sse42() local 133 src_r3 = _mm_unpackhi_epi16(src_r2_r3, zero_8x16b); // a30 0 a31 0 a32 0 a33 0 -- 16 bit long in isvcd_iquant_itrans_residual_recon_4x4_sse42() 138 temp7 = _mm_madd_epi16(src_r3, temp7); in isvcd_iquant_itrans_residual_recon_4x4_sse42() 1643 __m128i src_r0, src_r1, src_r2, src_r3; in isvcd_iquant_itrans_residual_recon_chroma_4x4_sse42() local 1695 src_r3 = _mm_unpackhi_epi16(src_r2_r3, zero_8x16b); // a30 0 a31 0 a32 0 a33 0 -- 16 bit long in isvcd_iquant_itrans_residual_recon_chroma_4x4_sse42() 1701 temp7 = _mm_madd_epi16(src_r3, temp7); in isvcd_iquant_itrans_residual_recon_chroma_4x4_sse42()
|