/external/libavc/common/x86/ |
D | ih264_ihadamard_scaling_ssse3.c | 97 __m128i src_r0, src_r1, src_r2, src_r3; in ih264_ihadamard_scaling_4x4_ssse3() local 113 src_r2 = _mm_unpacklo_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() 133 src_r2 = _mm_unpacklo_epi64(temp1, temp3); //a2 b2 c2 d2 in ih264_ihadamard_scaling_4x4_ssse3() 137 temp1 = _mm_add_epi32(src_r1, src_r2); in ih264_ihadamard_scaling_4x4_ssse3() 138 temp2 = _mm_sub_epi32(src_r1, src_r2); in ih264_ihadamard_scaling_4x4_ssse3() 143 src_r2 = _mm_sub_epi32(temp0, temp1); 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 …]
|
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 132 …src_r2 = _mm_loadl_epi64((__m128i *) (&pu1_src[2 * src_strd])); //a20 a21 a22 a23 0 0 0 0 0 0 0 0 … in ih264_resi_trans_quant_4x4_sse42() 137 src_r2 = _mm_cvtepu8_epi16(src_r2); in ih264_resi_trans_quant_4x4_sse42() 152 src_r2 = _mm_sub_epi16(src_r2, pred_r2); 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() 173 src_r2 = _mm_unpacklo_epi64(temp3, zero_8x16b); //a2 b2 c2 d2 in ih264_resi_trans_quant_4x4_sse42() 180 temp1 = _mm_add_epi16(src_r1, src_r2); in ih264_resi_trans_quant_4x4_sse42() 182 temp2 = _mm_sub_epi16(src_r1, src_r2); in ih264_resi_trans_quant_4x4_sse42() 192 src_r2 = _mm_sub_epi16(temp0, temp1); in ih264_resi_trans_quant_4x4_sse42() 205 temp2 = _mm_unpacklo_epi16(src_r2, src_r3); //a2 a3 b2 b3 c2 c3 d2 d3 in ih264_resi_trans_quant_4x4_sse42() [all …]
|
D | ih264_ihadamard_scaling_sse42.c | 97 __m128i src_r0, src_r1, src_r2, src_r3; in ih264_ihadamard_scaling_4x4_sse42() local 110 src_r2 = _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() 131 src_r2 = _mm_unpacklo_epi64(temp1, temp3); //a2 b2 c2 d2 in ih264_ihadamard_scaling_4x4_sse42() 135 temp1 = _mm_add_epi32(src_r1, src_r2); in ih264_ihadamard_scaling_4x4_sse42() 136 temp2 = _mm_sub_epi32(src_r1, src_r2); in ih264_ihadamard_scaling_4x4_sse42() 141 src_r2 = _mm_sub_epi32(temp0, temp1); 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 …]
|
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 146 src_r2 = _mm_unpacklo_epi16(src_r2_r3, zero_8x16b); // a20 0 a21 0 a22 0 a23 0 -- 16 bit long in ih264_iquant_itrans_recon_4x4_sse42() 151 temp6 = _mm_madd_epi16(src_r2, temp6); 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 397 src_r2 = _mm_unpacklo_epi16(src_r2_r3, zero_8x16b); // a20 0 a21 0 a22 0 a23 0 -- 16 bit long in ih264_iquant_itrans_recon_chroma_4x4_sse42() 402 temp6 = _mm_madd_epi16(src_r2, temp6); in ih264_iquant_itrans_recon_chroma_4x4_sse42()
|
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 147 src_r2 = _mm_unpacklo_epi16(src_r2_r3, zero_8x16b); // a20 0 a21 0 a22 0 a23 0 -- 16 bit long in ih264_iquant_itrans_recon_4x4_ssse3() 152 temp6 = _mm_madd_epi16(src_r2, temp6); in ih264_iquant_itrans_recon_4x4_ssse3()
|
/external/libavc/common/x86/svc/ |
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 197 src_r2 = _mm_loadl_epi64((__m128i *) (&pu1_src[2 * i4_src_stride])); in isvc_resi_trans_quant_4x4_sse42() 205 src_r2 = _mm_cvtepu8_epi16(src_r2); in isvc_resi_trans_quant_4x4_sse42() 231 src_r2 = _mm_sub_epi16(src_r2, pred_r2); in isvc_resi_trans_quant_4x4_sse42() 248 temp2 = _mm_unpacklo_epi16(src_r2, src_r3); in isvc_resi_trans_quant_4x4_sse42() 259 src_r2 = _mm_unpacklo_epi64(temp3, zero_8x16b); in isvc_resi_trans_quant_4x4_sse42() 267 temp1 = _mm_add_epi16(src_r1, src_r2); in isvc_resi_trans_quant_4x4_sse42() 269 temp2 = _mm_sub_epi16(src_r1, src_r2); in isvc_resi_trans_quant_4x4_sse42() 279 src_r2 = _mm_sub_epi16(temp0, temp1); in isvc_resi_trans_quant_4x4_sse42() 294 temp2 = _mm_unpacklo_epi16(src_r2, src_r3); in isvc_resi_trans_quant_4x4_sse42() [all …]
|
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 196 src_r2 = _mm_unpacklo_epi16(src_r2_r3, zero_8x16b); in isvc_iquant_itrans_recon_4x4_sse42() 202 temp6 = _mm_madd_epi16(src_r2, temp6); 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 490 src_r2 = _mm_unpacklo_epi16(src_r2_r3, zero_8x16b); in isvc_iquant_itrans_recon_res_4x4_sse42() 496 temp6 = _mm_madd_epi16(src_r2, temp6); 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 803 src_r2 = _mm_unpacklo_epi16(src_r2_r3, zero_8x16b); in isvc_iquant_itrans_recon_res_4x4_with_res_acc_sse42() 809 temp6 = _mm_madd_epi16(src_r2, temp6); 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 …]
|
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 187 src_r2 = _mm_unpacklo_epi16(src_r2_r3, zero_8x16b); // a20 0 a21 0 a22 0 a23 0 -- 16 bit long in isvc_iquant_itrans_recon_4x4_ssse3() 193 temp6 = _mm_madd_epi16(src_r2, temp6); in isvc_iquant_itrans_recon_4x4_ssse3()
|
/external/libmpeg2/common/x86/ |
D | impeg2_inter_pred_sse42_intr.c | 76 __m128i src_r0, src_r1, src_r2, src_r3; in impeg2_copy_mb_sse42() local 86 src_r2 = _mm_loadu_si128((__m128i *) (src + 2 * src_wd)); in impeg2_copy_mb_sse42() 91 _mm_storeu_si128((__m128i *) (dst + 2 * dst_wd), src_r2); in impeg2_copy_mb_sse42() 99 src_r2 = _mm_loadu_si128((__m128i *) (src + 2 * src_wd)); in impeg2_copy_mb_sse42() 104 _mm_storeu_si128((__m128i *) (dst + 2 * dst_wd), src_r2); in impeg2_copy_mb_sse42() 112 src_r2 = _mm_loadu_si128((__m128i *) (src + 2 * src_wd)); in impeg2_copy_mb_sse42() 117 _mm_storeu_si128((__m128i *) (dst + 2 * dst_wd), src_r2); in impeg2_copy_mb_sse42() 125 src_r2 = _mm_loadu_si128((__m128i *) (src + 2 * src_wd)); in impeg2_copy_mb_sse42() 130 _mm_storeu_si128((__m128i *) (dst + 2 * dst_wd), src_r2); in impeg2_copy_mb_sse42() 145 src_r2 = _mm_loadl_epi64((__m128i *)(src + 2 * src_wd)); in impeg2_copy_mb_sse42() [all …]
|
/external/libavc/encoder/x86/ |
D | ime_distortion_metrics_sse42.c | 108 __m128i src_r0, src_r1, src_r2, src_r3; in ime_compute_sad_16x16_sse42() local 118 src_r2 = _mm_loadu_si128((__m128i *) (pu1_src + 2*src_strd)); in ime_compute_sad_16x16_sse42() 128 res_r2 = _mm_sad_epu8(src_r2, est_r2); in ime_compute_sad_16x16_sse42() 141 src_r2 = _mm_loadu_si128((__m128i *) (pu1_src + 2*src_strd)); in ime_compute_sad_16x16_sse42() 151 res_r2 = _mm_sad_epu8(src_r2, est_r2); in ime_compute_sad_16x16_sse42() 164 src_r2 = _mm_loadu_si128((__m128i *) (pu1_src + 2*src_strd)); in ime_compute_sad_16x16_sse42() 174 res_r2 = _mm_sad_epu8(src_r2, est_r2); in ime_compute_sad_16x16_sse42() 187 src_r2 = _mm_loadu_si128((__m128i *) (pu1_src + 2*src_strd)); in ime_compute_sad_16x16_sse42() 197 res_r2 = _mm_sad_epu8(src_r2, est_r2); in ime_compute_sad_16x16_sse42() 252 __m128i src_r0, src_r1, src_r2, src_r3; in ime_compute_sad_16x8_sse42() local [all …]
|
/external/libavc/encoder/x86/svc/ |
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 598 src_r2 = _mm_loadl_epi64((__m128i *) (pu1_src)); in isvce_get_sad_with_residual_pred_sse42() 635 src_r2 = _mm_cvtepu8_epi16(src_r2); in isvce_get_sad_with_residual_pred_sse42() 673 src_r2 = _mm_sub_epi16(src_r2, pred_r2); in isvce_get_sad_with_residual_pred_sse42() 682 src_r2 = _mm_sub_epi16(src_r2, res_r2); in isvce_get_sad_with_residual_pred_sse42() 691 src_r2 = _mm_abs_epi16(src_r2); 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() 700 src_r2 = _mm_adds_epu16(src_r4, src_r5); 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() 709 src_r2 = _mm_srli_si128(src_r0, 8); in isvce_get_sad_with_residual_pred_sse42() [all …]
|
/external/libavc/decoder/x86/svc/ |
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 224 src_r2 = _mm_loadu_si128((__m128i *) (pi2_out + (2 * out_strd))); in isvcd_iquant_itrans_chroma_4x4_dc_sse42() 233 src_r2 = _mm_and_si128(src_r2, chroma_mask_odd); in isvcd_iquant_itrans_chroma_4x4_dc_sse42() 240 src_r2 = _mm_add_epi16(src_r2, i_macro); in isvcd_iquant_itrans_chroma_4x4_dc_sse42() 245 _mm_storeu_si128((__m128i *) (&pi2_out[2 * out_strd]), src_r2); 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 322 src_r2 = _mm_unpacklo_epi16(src_r2_r3, zero_8x16b); // a20 0 a21 0 a22 0 a23 0 -- 16 bit long in isvcd_iquant_itrans_4x4_sse42() 328 temp6 = _mm_madd_epi16(src_r2, temp6); in isvcd_iquant_itrans_4x4_sse42() 1251 __m128i src_r0, src_r1, src_r2, src_r3; in isvcd_iquant_itrans_chroma_4x4_sse42() local 1300 src_r2 = _mm_unpacklo_epi16(src_r2_r3, zero_8x16b); // a20 0 a21 0 a22 0 a23 0 -- 16 bit long in isvcd_iquant_itrans_chroma_4x4_sse42() [all …]
|
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 704 src_r2 = _mm_loadu_si128((__m128i *) (pu1_out + (2 * out_strd))); in isvcd_pred_residual_recon_chroma_4x4_sse42() 729 src_r2 = _mm_and_si128(src_r2, chroma_mask_odd); in isvcd_pred_residual_recon_chroma_4x4_sse42() 739 src_r2 = _mm_add_epi8(src_r2, pred2); in isvcd_pred_residual_recon_chroma_4x4_sse42() 744 _mm_storel_epi64((__m128i *) (&pu1_out[2 * out_strd]), src_r2); 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 802 src_r2 = _mm_loadu_si128((__m128i *) (pu1_out + (2 * out_strd))); in isvcd_pred_residual_recon_chroma_8x8_sse42() 843 src_r2 = _mm_and_si128(src_r2, chroma_mask_odd); in isvcd_pred_residual_recon_chroma_8x8_sse42() 861 src_r2 = _mm_add_epi8(src_r2, pred2); in isvcd_pred_residual_recon_chroma_8x8_sse42() 870 _mm_storel_epi64((__m128i *) (&pu1_out[2 * out_strd]), src_r2); in isvcd_pred_residual_recon_chroma_8x8_sse42() [all …]
|
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 135 src_r2 = _mm_unpacklo_epi16(src_r2_r3, zero_8x16b); // a20 0 a21 0 a22 0 a23 0 -- 16 bit long in isvcd_iquant_itrans_residual_4x4_sse42() 141 temp6 = _mm_madd_epi16(src_r2, temp6); 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 1393 src_r2 = _mm_unpacklo_epi16(src_r2_r3, zero_8x16b); // a20 0 a21 0 a22 0 a23 0 -- 16 bit long in isvcd_iquant_itrans_residual_chroma_4x4_sse42() 1399 temp6 = _mm_madd_epi16(src_r2, temp6); in isvcd_iquant_itrans_residual_chroma_4x4_sse42()
|
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 132 src_r2 = _mm_unpacklo_epi16(src_r2_r3, zero_8x16b); // a20 0 a21 0 a22 0 a23 0 -- 16 bit long in isvcd_iquant_itrans_residual_recon_4x4_sse42() 137 temp6 = _mm_madd_epi16(src_r2, temp6); 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 1694 src_r2 = _mm_unpacklo_epi16(src_r2_r3, zero_8x16b); // a20 0 a21 0 a22 0 a23 0 -- 16 bit long in isvcd_iquant_itrans_residual_recon_chroma_4x4_sse42() 1700 temp6 = _mm_madd_epi16(src_r2, temp6); in isvcd_iquant_itrans_residual_recon_chroma_4x4_sse42()
|