/external/libmpeg2/common/x86/ |
D | impeg2_inter_pred_sse42_intr.c | 143 src_r0 = _mm_loadl_epi64((__m128i *)src); in impeg2_copy_mb_sse42() 144 src_r1 = _mm_loadl_epi64((__m128i *)(src + src_wd)); in impeg2_copy_mb_sse42() 145 src_r2 = _mm_loadl_epi64((__m128i *)(src + 2 * src_wd)); in impeg2_copy_mb_sse42() 146 src_r3 = _mm_loadl_epi64((__m128i *)(src + 3 * src_wd)); in impeg2_copy_mb_sse42() 157 src_r0 = _mm_loadl_epi64((__m128i *)src); in impeg2_copy_mb_sse42() 158 src_r1 = _mm_loadl_epi64((__m128i *)(src + src_wd)); in impeg2_copy_mb_sse42() 159 src_r2 = _mm_loadl_epi64((__m128i *)(src + 2 * src_wd)); in impeg2_copy_mb_sse42() 160 src_r3 = _mm_loadl_epi64((__m128i *)(src + 3 * src_wd)); in impeg2_copy_mb_sse42() 173 src_r0 = _mm_loadl_epi64((__m128i *)src); in impeg2_copy_mb_sse42() 174 src_r1 = _mm_loadl_epi64((__m128i *)(src + src_wd)); in impeg2_copy_mb_sse42() [all …]
|
D | ideint_spatial_filter_ssse3.c | 119 row1_m1 = _mm_loadl_epi64((__m128i *) (pu1_src - 1)); in ideint_spatial_filter_ssse3() 120 row1_0 = _mm_loadl_epi64((__m128i *) (pu1_src)); in ideint_spatial_filter_ssse3() 121 row1_p1 = _mm_loadl_epi64((__m128i *) (pu1_src + 1)); in ideint_spatial_filter_ssse3() 134 row2_m1 = _mm_loadl_epi64((__m128i *) (pu1_src - 1)); in ideint_spatial_filter_ssse3() 135 row2_0 = _mm_loadl_epi64((__m128i *) (pu1_src)); in ideint_spatial_filter_ssse3() 136 row2_p1 = _mm_loadl_epi64((__m128i *) (pu1_src + 1)); in ideint_spatial_filter_ssse3()
|
D | icv_sad_ssse3.c | 108 src_r0 = (__m128)_mm_loadl_epi64((__m128i *) (pu1_src)); in icv_sad_8x4_ssse3() 111 src_r1 = (__m128)_mm_loadl_epi64((__m128i *) (pu1_src)); in icv_sad_8x4_ssse3() 122 ref_r0 = (__m128)_mm_loadl_epi64((__m128i *) (pu1_ref)); in icv_sad_8x4_ssse3() 125 ref_r1 = (__m128)_mm_loadl_epi64((__m128i *) (pu1_ref)); in icv_sad_8x4_ssse3()
|
/external/libhevc/common/x86/ |
D | ihevc_inter_pred_filters_sse42_intr.c | 130 src0_16x8b = _mm_loadl_epi64((__m128i *)(pu1_src)); /* row =0 */ in ihevc_inter_pred_luma_copy_w16out_sse42() 131 src1_16x8b = _mm_loadl_epi64((__m128i *)(pu1_src + 1 * src_strd)); /* row =1 */ in ihevc_inter_pred_luma_copy_w16out_sse42() 132 src2_16x8b = _mm_loadl_epi64((__m128i *)(pu1_src + 2 * src_strd)); /* row =2 */ in ihevc_inter_pred_luma_copy_w16out_sse42() 133 src3_16x8b = _mm_loadl_epi64((__m128i *)(pu1_src + 3 * src_strd)); /* row =3 */ in ihevc_inter_pred_luma_copy_w16out_sse42() 166 src0_16x8b = _mm_loadl_epi64((__m128i *)(pu1_src)); /* row =0 */ in ihevc_inter_pred_luma_copy_w16out_sse42() 167 src1_16x8b = _mm_loadl_epi64((__m128i *)(pu1_src + 1 * src_strd)); /* row =1 */ in ihevc_inter_pred_luma_copy_w16out_sse42() 168 src2_16x8b = _mm_loadl_epi64((__m128i *)(pu1_src + 2 * src_strd)); /* row =2 */ in ihevc_inter_pred_luma_copy_w16out_sse42() 169 src3_16x8b = _mm_loadl_epi64((__m128i *)(pu1_src + 3 * src_strd)); /* row =3 */ in ihevc_inter_pred_luma_copy_w16out_sse42() 288 src0_16x8b = _mm_loadl_epi64((__m128i *)(pu1_src)); /* row =0 */ in ihevc_inter_pred_chroma_copy_sse42() 289 src1_16x8b = _mm_loadl_epi64((__m128i *)(pu1_src + 1 * src_strd)); /* row =1 */ in ihevc_inter_pred_chroma_copy_sse42() [all …]
|
D | ihevc_inter_pred_filters_ssse3_intr.c | 168 src0_16x8b = _mm_loadl_epi64((__m128i *)(pu1_src)); /* row =0 */ in ihevc_inter_pred_luma_copy_ssse3() 169 src1_16x8b = _mm_loadl_epi64((__m128i *)(pu1_src + 1 * src_strd)); /* row =1 */ in ihevc_inter_pred_luma_copy_ssse3() 170 src2_16x8b = _mm_loadl_epi64((__m128i *)(pu1_src + 2 * src_strd)); /* row =2 */ in ihevc_inter_pred_luma_copy_ssse3() 171 src3_16x8b = _mm_loadl_epi64((__m128i *)(pu1_src + 3 * src_strd)); /* row =3 */ in ihevc_inter_pred_luma_copy_ssse3() 195 src0_16x8b = _mm_loadl_epi64((__m128i *)(pu1_src)); /* row =0 */ in ihevc_inter_pred_luma_copy_ssse3() 196 src1_16x8b = _mm_loadl_epi64((__m128i *)(pu1_src + 1 * src_strd)); /* row =1 */ in ihevc_inter_pred_luma_copy_ssse3() 197 src2_16x8b = _mm_loadl_epi64((__m128i *)(pu1_src + 2 * src_strd)); /* row =2 */ in ihevc_inter_pred_luma_copy_ssse3() 198 src3_16x8b = _mm_loadl_epi64((__m128i *)(pu1_src + 3 * src_strd)); /* row =3 */ in ihevc_inter_pred_luma_copy_ssse3() 292 src_temp1_16x8b = _mm_loadl_epi64((__m128i *)pi1_coeff); in ihevc_inter_pred_luma_horz_ssse3() 453 res_temp7_8x16b = _mm_loadl_epi64((__m128i *)(pu1_dst + offset)); in ihevc_inter_pred_luma_horz_ssse3() [all …]
|
D | ihevc_deblk_ssse3_intr.c | 144 src_row0_8x16b = _mm_loadl_epi64((__m128i *)(pu1_src - 4)); in ihevc_deblk_luma_vert_ssse3() 145 src_row3_8x16b = _mm_loadl_epi64((__m128i *)((pu1_src - 4) + 3 * src_strd)); in ihevc_deblk_luma_vert_ssse3() 241 src_row1_8x16b = _mm_loadl_epi64((__m128i *)((pu1_src - 4) + src_strd)); in ihevc_deblk_luma_vert_ssse3() 242 src_row2_8x16b = _mm_loadl_epi64((__m128i *)((pu1_src - 4) + 2 * src_strd)); in ihevc_deblk_luma_vert_ssse3() 295 temp_str0_16x8b = _mm_loadl_epi64((__m128i *)(&mask)); in ihevc_deblk_luma_vert_ssse3() 296 const2tc_8x16b = _mm_loadl_epi64((__m128i *)(&tc2)); in ihevc_deblk_luma_vert_ssse3() 429 mask_pq_8x16b = _mm_loadl_epi64((__m128i *)(&mask1)); in ihevc_deblk_luma_vert_ssse3() 459 tmp2_const_8x16b = _mm_loadl_epi64((__m128i *)(shuffle0)); in ihevc_deblk_luma_vert_ssse3() 584 src_q0_8x16b = _mm_loadl_epi64((__m128i *)(pu1_src)); in ihevc_deblk_luma_horz_ssse3() 585 src_q1_8x16b = _mm_loadl_epi64((__m128i *)(pu1_src + src_strd)); in ihevc_deblk_luma_horz_ssse3() [all …]
|
D | ihevc_weighted_pred_ssse3_intr.c | 267 src_temp0_8x16b = _mm_loadl_epi64((__m128i *)(pi2_src)); in ihevc_weighted_pred_uni_ssse3() 269 src_temp1_8x16b = _mm_loadl_epi64((__m128i *)(pi2_src + src_strd)); 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() 273 src_temp3_8x16b = _mm_loadl_epi64((__m128i *)(pi2_src + 3 * src_strd)); in ihevc_weighted_pred_uni_ssse3() 616 src_temp0_8x16b = _mm_loadl_epi64((__m128i *)(pi2_src)); in ihevc_weighted_pred_chroma_uni_ssse3() 618 src_temp1_8x16b = _mm_loadl_epi64((__m128i *)(pi2_src + src_strd)); in ihevc_weighted_pred_chroma_uni_ssse3() 880 src_temp1_8x16b = _mm_loadl_epi64((__m128i *)(pi2_src1)); /* row = 0 */ in ihevc_weighted_pred_bi_ssse3() 881 src_temp2_8x16b = _mm_loadl_epi64((__m128i *)(pi2_src2)); /* row = 0 */ in ihevc_weighted_pred_bi_ssse3() 882 … src_temp3_8x16b = _mm_loadl_epi64((__m128i *)(pi2_src1 + 1 * src_strd1)); /* row = 1 */ in ihevc_weighted_pred_bi_ssse3() 883 … src_temp4_8x16b = _mm_loadl_epi64((__m128i *)(pi2_src2 + 1 * src_strd2)); /* row = 1 */ in ihevc_weighted_pred_bi_ssse3() [all …]
|
D | ihevc_sao_ssse3_intr.c | 151 tmp_set_128i_1 = _mm_loadl_epi64((__m128i *)(pu1_src + (ht - 1) * src_strd + offset)); in ihevc_sao_band_offset_luma_ssse3() 167 sao_offset = _mm_loadl_epi64((__m128i *)pi1_sao_offset); in ihevc_sao_band_offset_luma_ssse3() 344 src_temp0_8x16b = _mm_loadl_epi64((__m128i *)(pu1_src_cpy)); in ihevc_sao_band_offset_luma_ssse3() 346 src_temp1_8x16b = _mm_loadl_epi64((__m128i *)(pu1_src_cpy + src_strd)); 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() 350 src_temp3_8x16b = _mm_loadl_epi64((__m128i *)(pu1_src_cpy + 3 * src_strd)); in ihevc_sao_band_offset_luma_ssse3() 473 tmp_set_128i_1 = _mm_loadl_epi64((__m128i *)(pu1_src + (ht - 1) * src_strd + offset)); in ihevc_sao_band_offset_chroma_ssse3() 490 sao_offset = _mm_loadl_epi64((__m128i *)pi1_sao_offset_u); in ihevc_sao_band_offset_chroma_ssse3() 572 sao_offset = _mm_loadl_epi64((__m128i *)pi1_sao_offset_v); in ihevc_sao_band_offset_chroma_ssse3() 768 src_temp0_8x16b = _mm_loadl_epi64((__m128i *)(pu1_src_cpy)); in ihevc_sao_band_offset_chroma_ssse3() [all …]
|
/external/libavc/common/x86/ |
D | ih264_iquant_itrans_recon_dc_ssse3.c | 136 …predload_r = _mm_loadl_epi64((__m128i *) (&pu1_pred[0])); //p00 p01 p02 p03 0 0 0 0 0 0 0 0 -- all… in ih264_iquant_itrans_recon_4x4_dc_ssse3() 138 …predload_r = _mm_loadl_epi64((__m128i *) (&pu1_pred[pred_strd])); //p10 p11 p12 p13 0 0 0 0 0 0 0 … in ih264_iquant_itrans_recon_4x4_dc_ssse3() 140 …predload_r = _mm_loadl_epi64((__m128i *) (&pu1_pred[2*pred_strd])); //p20 p21 p22 p23 0 0 0 0 0 0 … in ih264_iquant_itrans_recon_4x4_dc_ssse3() 142 …predload_r = _mm_loadl_epi64((__m128i *) (&pu1_pred[3*pred_strd])); //p30 p31 p32 p33 0 0 0 0 0 0 … in ih264_iquant_itrans_recon_4x4_dc_ssse3() 258 …predload_r = _mm_loadl_epi64((__m128i *)(&pu1_pred[0])); //p0 p1 p2 p3 p4 p5 p6 p7 0 0 0 0 0 0 0 0… in ih264_iquant_itrans_recon_8x8_dc_ssse3() 261 …predload_r = _mm_loadl_epi64((__m128i *)(&pu1_pred[pred_strd])); //p0 p1 p2 p3 p4 p5 p6 p7 0 0 0 0… in ih264_iquant_itrans_recon_8x8_dc_ssse3() 264 predload_r = _mm_loadl_epi64( in ih264_iquant_itrans_recon_8x8_dc_ssse3() 268 predload_r = _mm_loadl_epi64( in ih264_iquant_itrans_recon_8x8_dc_ssse3() 272 predload_r = _mm_loadl_epi64( in ih264_iquant_itrans_recon_8x8_dc_ssse3() 276 predload_r = _mm_loadl_epi64( in ih264_iquant_itrans_recon_8x8_dc_ssse3() [all …]
|
D | ih264_weighted_pred_sse42.c | 101 y0_0_16x8b = _mm_loadl_epi64((__m128i *)pu1_src1); in ih264_default_weighted_pred_luma_sse42() 102 y0_1_16x8b = _mm_loadl_epi64((__m128i *)(pu1_src1 + src_strd1)); in ih264_default_weighted_pred_luma_sse42() 103 y0_2_16x8b = _mm_loadl_epi64( in ih264_default_weighted_pred_luma_sse42() 105 y0_3_16x8b = _mm_loadl_epi64((__m128i *)(pu1_src1 + src_strd1 * 3)); in ih264_default_weighted_pred_luma_sse42() 107 y1_0_16x8b = _mm_loadl_epi64((__m128i *)pu1_src2); in ih264_default_weighted_pred_luma_sse42() 108 y1_1_16x8b = _mm_loadl_epi64((__m128i *)(pu1_src2 + src_strd2)); in ih264_default_weighted_pred_luma_sse42() 109 y1_2_16x8b = _mm_loadl_epi64( in ih264_default_weighted_pred_luma_sse42() 111 y1_3_16x8b = _mm_loadl_epi64((__m128i *)(pu1_src2 + src_strd2 * 3)); in ih264_default_weighted_pred_luma_sse42() 134 y0_0_16x8b = _mm_loadl_epi64((__m128i *)pu1_src1); in ih264_default_weighted_pred_luma_sse42() 135 y0_1_16x8b = _mm_loadl_epi64((__m128i *)(pu1_src1 + src_strd1)); in ih264_default_weighted_pred_luma_sse42() [all …]
|
D | ih264_inter_pred_filters_ssse3.c | 131 y_0_16x8b = _mm_loadl_epi64((__m128i *)pu1_src); in ih264_inter_pred_luma_copy_ssse3() 132 y_1_16x8b = _mm_loadl_epi64((__m128i *)(pu1_src + src_strd)); in ih264_inter_pred_luma_copy_ssse3() 133 y_2_16x8b = _mm_loadl_epi64((__m128i *)(pu1_src + src_strd2)); in ih264_inter_pred_luma_copy_ssse3() 134 y_3_16x8b = _mm_loadl_epi64((__m128i *)(pu1_src + src_strd3)); in ih264_inter_pred_luma_copy_ssse3() 514 src_r0_16x8b = _mm_loadl_epi64((__m128i *)pu1_src); in ih264_inter_pred_luma_vert_ssse3() 516 src_r1_16x8b = _mm_loadl_epi64((__m128i *)pu1_src); in ih264_inter_pred_luma_vert_ssse3() 518 src_r2_16x8b = _mm_loadl_epi64((__m128i *)pu1_src); in ih264_inter_pred_luma_vert_ssse3() 520 src_r3_16x8b = _mm_loadl_epi64((__m128i *)pu1_src); in ih264_inter_pred_luma_vert_ssse3() 522 src_r4_16x8b = _mm_loadl_epi64((__m128i *)pu1_src); in ih264_inter_pred_luma_vert_ssse3() 532 src_r5_16x8b = _mm_loadl_epi64((__m128i *)pu1_src); in ih264_inter_pred_luma_vert_ssse3() [all …]
|
D | ih264_deblk_chroma_ssse3.c | 115 linea = _mm_loadl_epi64((__m128i *)(pu1_src_uv - 4)); in ih264_deblk_chroma_vert_bs4_ssse3() 116 lineb = _mm_loadl_epi64((__m128i *)(pu1_src_uv - 4 + src_strd)); in ih264_deblk_chroma_vert_bs4_ssse3() 117 linec = _mm_loadl_epi64((__m128i *)(pu1_src_uv - 4 + 2 * src_strd)); in ih264_deblk_chroma_vert_bs4_ssse3() 118 lined = _mm_loadl_epi64((__m128i *)(pu1_src_uv - 4 + 3 * src_strd)); in ih264_deblk_chroma_vert_bs4_ssse3() 119 linee = _mm_loadl_epi64((__m128i *)(pu1_src_uv - 4 + 4 * src_strd)); in ih264_deblk_chroma_vert_bs4_ssse3() 120 linef = _mm_loadl_epi64((__m128i *)(pu1_src_uv - 4 + 5 * src_strd)); in ih264_deblk_chroma_vert_bs4_ssse3() 121 lineg = _mm_loadl_epi64((__m128i *)(pu1_src_uv - 4 + 6 * src_strd)); in ih264_deblk_chroma_vert_bs4_ssse3() 122 lineh = _mm_loadl_epi64((__m128i *)(pu1_src_uv - 4 + 7 * src_strd)); in ih264_deblk_chroma_vert_bs4_ssse3() 464 linea = _mm_loadl_epi64((__m128i *)(pu1_src_uv - 4)); in ih264_deblk_chroma_vert_bslt4_ssse3() 465 lineb = _mm_loadl_epi64((__m128i *)(pu1_src_uv - 4 + src_strd)); in ih264_deblk_chroma_vert_bslt4_ssse3() [all …]
|
D | ih264_iquant_itrans_recon_sse42.c | 230 …pred_r0 = _mm_loadl_epi64((__m128i *) (&pu1_pred[0])); //p00 p01 p02 p03 0 0 0 0 0 0 0 0 -- all 8 … in ih264_iquant_itrans_recon_4x4_sse42() 231 …pred_r1 = _mm_loadl_epi64((__m128i *) (&pu1_pred[pred_strd])); //p10 p11 p12 p13 0 0 0 0 0 0 0 0 -… in ih264_iquant_itrans_recon_4x4_sse42() 232 …pred_r2 = _mm_loadl_epi64((__m128i *) (&pu1_pred[2 * pred_strd])); //p20 p21 p22 p23 0 0 0 0 0 0 0… in ih264_iquant_itrans_recon_4x4_sse42() 233 …pred_r3 = _mm_loadl_epi64((__m128i *) (&pu1_pred[3 * pred_strd])); //p30 p31 p32 p33 0 0 0 0 0 0 0… in ih264_iquant_itrans_recon_4x4_sse42() 480 …pred_r0 = _mm_loadl_epi64((__m128i *) (&pu1_pred[0])); //p00 p01 p02 p03 0 0 0 0 0 0 0 0 -- all 8 … in ih264_iquant_itrans_recon_chroma_4x4_sse42() 481 …pred_r1 = _mm_loadl_epi64((__m128i *) (&pu1_pred[pred_strd])); //p10 p11 p12 p13 0 0 0 0 0 0 0 0 -… in ih264_iquant_itrans_recon_chroma_4x4_sse42() 482 …pred_r2 = _mm_loadl_epi64((__m128i *) (&pu1_pred[2 * pred_strd])); //p20 p21 p22 p23 0 0 0 0 0 0 0… in ih264_iquant_itrans_recon_chroma_4x4_sse42() 483 …pred_r3 = _mm_loadl_epi64((__m128i *) (&pu1_pred[3 * pred_strd])); //p30 p31 p32 p33 0 0 0 0 0 0 0… in ih264_iquant_itrans_recon_chroma_4x4_sse42() 553 out_r0 = _mm_loadl_epi64((__m128i *) (&pu1_out[0])); in ih264_iquant_itrans_recon_chroma_4x4_sse42() 554 out_r1 = _mm_loadl_epi64((__m128i *) (&pu1_out[out_strd])); in ih264_iquant_itrans_recon_chroma_4x4_sse42() [all …]
|
D | ih264_resi_trans_quant_sse42.c | 129 …src_r0 = _mm_loadl_epi64((__m128i *) (&pu1_src[0])); //a00 a01 a02 a03 0 0 0 0 0 0 0 0 -- all 8 bi… in ih264_resi_trans_quant_4x4_sse42() 130 …src_r1 = _mm_loadl_epi64((__m128i *) (&pu1_src[src_strd])); //a10 a11 a12 a13 0 0 0 0 0 0 0 0 -- a… in ih264_resi_trans_quant_4x4_sse42() 131 …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() 132 …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() 139 …pred_r0 = _mm_loadl_epi64((__m128i *) (&pu1_pred[0])); //p00 p01 p02 p03 0 0 0 0 0 0 0 0 -- all 8 … in ih264_resi_trans_quant_4x4_sse42() 140 …pred_r1 = _mm_loadl_epi64((__m128i *) (&pu1_pred[pred_strd])); //p10 p11 p12 p13 0 0 0 0 0 0 0 0 -… in ih264_resi_trans_quant_4x4_sse42() 141 …pred_r2 = _mm_loadl_epi64((__m128i *) (&pu1_pred[2 * pred_strd])); //p20 p21 p22 p23 0 0 0 0 0 0 0… in ih264_resi_trans_quant_4x4_sse42() 142 …pred_r3 = _mm_loadl_epi64((__m128i *) (&pu1_pred[3 * pred_strd])); //p30 p31 p32 p33 0 0 0 0 0 0 0… in ih264_resi_trans_quant_4x4_sse42() 404 …src_r0 = _mm_loadl_epi64((__m128i *) (&pu1_src[0])); //a00 a01 a02 a03 0 0 0 0 0 0 0 0 -- all 8 bi… in ih264_resi_trans_quant_chroma_4x4_sse42() 405 …src_r1 = _mm_loadl_epi64((__m128i *) (&pu1_src[src_strd])); //a10 a11 a12 a13 0 0 0 0 0 0 0 0 -- a… in ih264_resi_trans_quant_chroma_4x4_sse42() [all …]
|
D | ih264_deblk_luma_ssse3.c | 113 line1 = _mm_loadl_epi64((__m128i *)(pu1_src - 4 + 0 * src_strd)); in ih264_deblk_luma_vert_bs4_ssse3() 114 line2 = _mm_loadl_epi64((__m128i *)(pu1_src - 4 + 1 * src_strd)); in ih264_deblk_luma_vert_bs4_ssse3() 115 line3 = _mm_loadl_epi64((__m128i *)(pu1_src - 4 + 2 * src_strd)); in ih264_deblk_luma_vert_bs4_ssse3() 116 line4 = _mm_loadl_epi64((__m128i *)(pu1_src - 4 + 3 * src_strd)); in ih264_deblk_luma_vert_bs4_ssse3() 117 line5 = _mm_loadl_epi64((__m128i *)(pu1_src - 4 + 4 * src_strd)); in ih264_deblk_luma_vert_bs4_ssse3() 118 line6 = _mm_loadl_epi64((__m128i *)(pu1_src - 4 + 5 * src_strd)); in ih264_deblk_luma_vert_bs4_ssse3() 119 line7 = _mm_loadl_epi64((__m128i *)(pu1_src - 4 + 6 * src_strd)); in ih264_deblk_luma_vert_bs4_ssse3() 120 line8 = _mm_loadl_epi64((__m128i *)(pu1_src - 4 + 7 * src_strd)); in ih264_deblk_luma_vert_bs4_ssse3() 137 line1 = _mm_loadl_epi64((__m128i *)(pu1_src - 4 + 8 * src_strd)); in ih264_deblk_luma_vert_bs4_ssse3() 138 line2 = _mm_loadl_epi64((__m128i *)(pu1_src - 4 + 9 * src_strd)); in ih264_deblk_luma_vert_bs4_ssse3() [all …]
|
/external/libhevc/decoder/x86/ |
D | ihevcd_it_rec_dc_sse42_intr.c | 97 m_temp_reg_0 = _mm_loadl_epi64((__m128i *)(pu1_pred)); in ihevcd_itrans_recon_dc_luma_sse42() 98 m_temp_reg_1 = _mm_loadl_epi64((__m128i *)(pu1_pred + pred_strd)); in ihevcd_itrans_recon_dc_luma_sse42() 99 m_temp_reg_2 = _mm_loadl_epi64((__m128i *)(pu1_pred + 2 * pred_strd)); in ihevcd_itrans_recon_dc_luma_sse42() 100 m_temp_reg_3 = _mm_loadl_epi64((__m128i *)(pu1_pred + 3 * pred_strd)); in ihevcd_itrans_recon_dc_luma_sse42() 140 m_temp_reg_0 = _mm_loadl_epi64((__m128i *)pu1_pred); in ihevcd_itrans_recon_dc_luma_sse42() 141 m_temp_reg_1 = _mm_loadl_epi64((__m128i *)(pu1_pred + pred_strd)); in ihevcd_itrans_recon_dc_luma_sse42() 142 m_temp_reg_2 = _mm_loadl_epi64((__m128i *)(pu1_pred + 2 * pred_strd)); in ihevcd_itrans_recon_dc_luma_sse42() 143 m_temp_reg_3 = _mm_loadl_epi64((__m128i *)(pu1_pred + 3 * pred_strd)); in ihevcd_itrans_recon_dc_luma_sse42() 244 m_temp_reg_0 = _mm_loadl_epi64((__m128i *)(pu1_pred)); in ihevcd_itrans_recon_dc_chroma_sse42() 245 m_temp_reg_1 = _mm_loadl_epi64((__m128i *)(pu1_pred + pred_strd)); in ihevcd_itrans_recon_dc_chroma_sse42() [all …]
|
D | ihevcd_it_rec_dc_ssse3_intr.c | 98 m_temp_reg_0 = _mm_loadl_epi64((__m128i *)(pu1_pred)); in ihevcd_itrans_recon_dc_luma_ssse3() 99 m_temp_reg_1 = _mm_loadl_epi64((__m128i *)(pu1_pred + pred_strd)); in ihevcd_itrans_recon_dc_luma_ssse3() 100 m_temp_reg_2 = _mm_loadl_epi64((__m128i *)(pu1_pred + 2 * pred_strd)); in ihevcd_itrans_recon_dc_luma_ssse3() 101 m_temp_reg_3 = _mm_loadl_epi64((__m128i *)(pu1_pred + 3 * pred_strd)); in ihevcd_itrans_recon_dc_luma_ssse3() 141 m_temp_reg_0 = _mm_loadl_epi64((__m128i *)pu1_pred); in ihevcd_itrans_recon_dc_luma_ssse3() 142 m_temp_reg_1 = _mm_loadl_epi64((__m128i *)(pu1_pred + pred_strd)); in ihevcd_itrans_recon_dc_luma_ssse3() 143 m_temp_reg_2 = _mm_loadl_epi64((__m128i *)(pu1_pred + 2 * pred_strd)); in ihevcd_itrans_recon_dc_luma_ssse3() 144 m_temp_reg_3 = _mm_loadl_epi64((__m128i *)(pu1_pred + 3 * pred_strd)); in ihevcd_itrans_recon_dc_luma_ssse3() 244 m_temp_reg_0 = _mm_loadl_epi64((__m128i *)(pu1_pred)); in ihevcd_itrans_recon_dc_chroma_ssse3() 245 m_temp_reg_1 = _mm_loadl_epi64((__m128i *)(pu1_pred + pred_strd)); in ihevcd_itrans_recon_dc_chroma_ssse3() [all …]
|
/external/libvpx/libvpx/vp9/encoder/x86/ |
D | vp9_avg_intrin_sse2.c | 21 s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s)), u0); in vp9_minmax_8x8_sse2() 22 d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d)), u0); in vp9_minmax_8x8_sse2() 27 s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + p)), u0); in vp9_minmax_8x8_sse2() 28 d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + dp)), u0); in vp9_minmax_8x8_sse2() 35 s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 2 * p)), u0); in vp9_minmax_8x8_sse2() 36 d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 2 * dp)), u0); in vp9_minmax_8x8_sse2() 43 s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 3 * p)), u0); in vp9_minmax_8x8_sse2() 44 d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 3 * dp)), u0); in vp9_minmax_8x8_sse2() 51 s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 4 * p)), u0); in vp9_minmax_8x8_sse2() 52 d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 4 * dp)), u0); in vp9_minmax_8x8_sse2() [all …]
|
/external/opencv3/3rdparty/libwebp/dsp/ |
D | enc_sse2.c | 133 in0 = _mm_loadl_epi64((__m128i*)&in[0]); in ITransformSSE2() 134 in1 = _mm_loadl_epi64((__m128i*)&in[4]); in ITransformSSE2() 135 in2 = _mm_loadl_epi64((__m128i*)&in[8]); in ITransformSSE2() 136 in3 = _mm_loadl_epi64((__m128i*)&in[12]); in ITransformSSE2() 142 const __m128i inB0 = _mm_loadl_epi64((__m128i*)&in[16]); in ITransformSSE2() 143 const __m128i inB1 = _mm_loadl_epi64((__m128i*)&in[20]); in ITransformSSE2() 144 const __m128i inB2 = _mm_loadl_epi64((__m128i*)&in[24]); in ITransformSSE2() 145 const __m128i inB3 = _mm_loadl_epi64((__m128i*)&in[28]); in ITransformSSE2() 282 ref0 = _mm_loadl_epi64((__m128i*)&ref[0 * BPS]); in ITransformSSE2() 283 ref1 = _mm_loadl_epi64((__m128i*)&ref[1 * BPS]); in ITransformSSE2() [all …]
|
/external/libvpx/libvpx/vpx_dsp/x86/ |
D | vpx_subpixel_8t_intrin_ssse3.c | 236 srcReg1 = _mm_loadl_epi64((const __m128i *)src_ptr); in vpx_filter_block1d8_v8_intrin_ssse3() 237 srcReg2 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch)); in vpx_filter_block1d8_v8_intrin_ssse3() 238 srcReg3 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 2)); in vpx_filter_block1d8_v8_intrin_ssse3() 239 srcReg4 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 3)); in vpx_filter_block1d8_v8_intrin_ssse3() 240 srcReg5 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 4)); in vpx_filter_block1d8_v8_intrin_ssse3() 241 srcReg6 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 5)); in vpx_filter_block1d8_v8_intrin_ssse3() 242 srcReg7 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 6)); in vpx_filter_block1d8_v8_intrin_ssse3() 246 srcReg8 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 7)); in vpx_filter_block1d8_v8_intrin_ssse3() 382 const __m128i A = _mm_loadl_epi64((const __m128i *)src_x); in filter_horiz_w8_ssse3() 383 const __m128i B = _mm_loadl_epi64((const __m128i *)(src_x + src_pitch)); in filter_horiz_w8_ssse3() [all …]
|
D | loopfilter_sse2.c | 35 q4p4 = _mm_loadl_epi64((__m128i *)(s - 5 * p)); in mb_lpf_horizontal_edge_w_sse2_8() 38 q3p3 = _mm_loadl_epi64((__m128i *)(s - 4 * p)); in mb_lpf_horizontal_edge_w_sse2_8() 41 q2p2 = _mm_loadl_epi64((__m128i *)(s - 3 * p)); in mb_lpf_horizontal_edge_w_sse2_8() 44 q1p1 = _mm_loadl_epi64((__m128i *)(s - 2 * p)); in mb_lpf_horizontal_edge_w_sse2_8() 48 q0p0 = _mm_loadl_epi64((__m128i *)(s - 1 * p)); in mb_lpf_horizontal_edge_w_sse2_8() 136 q5p5 = _mm_loadl_epi64((__m128i *)(s - 6 * p)); in mb_lpf_horizontal_edge_w_sse2_8() 140 q6p6 = _mm_loadl_epi64((__m128i *)(s - 7 * p)); in mb_lpf_horizontal_edge_w_sse2_8() 145 q7p7 = _mm_loadl_epi64((__m128i *)(s - 8 * p)); in mb_lpf_horizontal_edge_w_sse2_8() 750 q3p3 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 4 * p)), in vpx_lpf_horizontal_8_sse2() 751 _mm_loadl_epi64((__m128i *)(s + 3 * p))); in vpx_lpf_horizontal_8_sse2() [all …]
|
/external/webp/src/dsp/ |
D | enc_sse2.c | 84 in0 = _mm_loadl_epi64((const __m128i*)&in[0]); in ITransform() 85 in1 = _mm_loadl_epi64((const __m128i*)&in[4]); in ITransform() 86 in2 = _mm_loadl_epi64((const __m128i*)&in[8]); in ITransform() 87 in3 = _mm_loadl_epi64((const __m128i*)&in[12]); in ITransform() 93 const __m128i inB0 = _mm_loadl_epi64((const __m128i*)&in[16]); in ITransform() 94 const __m128i inB1 = _mm_loadl_epi64((const __m128i*)&in[20]); in ITransform() 95 const __m128i inB2 = _mm_loadl_epi64((const __m128i*)&in[24]); in ITransform() 96 const __m128i inB3 = _mm_loadl_epi64((const __m128i*)&in[28]); in ITransform() 233 ref0 = _mm_loadl_epi64((const __m128i*)&ref[0 * BPS]); in ITransform() 234 ref1 = _mm_loadl_epi64((const __m128i*)&ref[1 * BPS]); in ITransform() [all …]
|
D | enc_sse41.c | 78 const __m128i inA_0 = _mm_loadl_epi64((const __m128i*)&inA[BPS * 0]); in TTransform() 79 const __m128i inA_1 = _mm_loadl_epi64((const __m128i*)&inA[BPS * 1]); in TTransform() 80 const __m128i inA_2 = _mm_loadl_epi64((const __m128i*)&inA[BPS * 2]); in TTransform() 81 const __m128i inA_3 = _mm_loadl_epi64((const __m128i*)&inA[BPS * 3]); in TTransform() 82 const __m128i inB_0 = _mm_loadl_epi64((const __m128i*)&inB[BPS * 0]); in TTransform() 83 const __m128i inB_1 = _mm_loadl_epi64((const __m128i*)&inB[BPS * 1]); in TTransform() 84 const __m128i inB_2 = _mm_loadl_epi64((const __m128i*)&inB[BPS * 2]); in TTransform() 85 const __m128i inB_3 = _mm_loadl_epi64((const __m128i*)&inB[BPS * 3]); in TTransform()
|
/external/neven/Embedded/common/src/b_BasicEm/ |
D | MathSSE2.c | 55 m_XMM0 = _mm_loadl_epi64( (__m128i *)&0[vec1L] ); in bbs_dotProduct_64SSE2() 58 m_XMM1 = _mm_loadl_epi64( (__m128i *)&0[vec2L] ); in bbs_dotProduct_64SSE2() 61 m_XMM2 = _mm_loadl_epi64( (__m128i *)&4[vec1L] ); in bbs_dotProduct_64SSE2() 65 m_XMM3 = _mm_loadl_epi64( (__m128i *)&4[vec2L] ); in bbs_dotProduct_64SSE2() 66 m_XMM4 = _mm_loadl_epi64( (__m128i *)&8[vec1L] ); in bbs_dotProduct_64SSE2() 70 m_XMM5 = _mm_loadl_epi64( (__m128i *)&8[vec2L] ); in bbs_dotProduct_64SSE2() 74 m_XMM6 = _mm_loadl_epi64( (__m128i *)&12[vec1L] ); in bbs_dotProduct_64SSE2() 78 m_XMM8 = _mm_loadl_epi64( (__m128i *)&12[vec2L] ); in bbs_dotProduct_64SSE2() 93 m_XMM0 = _mm_loadl_epi64( (__m128i *)&m_XMM7 ); in bbs_dotProduct_64SSE2()
|
/external/webrtc/webrtc/modules/audio_processing/aec/ |
D | aec_rdft_sse2.c | 88 const __m128i a_00 = _mm_loadl_epi64((__m128i*)&a[j0 + 0]); in cftmdl_128_SSE2() 89 const __m128i a_08 = _mm_loadl_epi64((__m128i*)&a[j0 + 8]); in cftmdl_128_SSE2() 90 const __m128i a_32 = _mm_loadl_epi64((__m128i*)&a[j0 + 32]); in cftmdl_128_SSE2() 91 const __m128i a_40 = _mm_loadl_epi64((__m128i*)&a[j0 + 40]); in cftmdl_128_SSE2() 101 const __m128i a_16 = _mm_loadl_epi64((__m128i*)&a[j0 + 16]); in cftmdl_128_SSE2() 102 const __m128i a_24 = _mm_loadl_epi64((__m128i*)&a[j0 + 24]); in cftmdl_128_SSE2() 103 const __m128i a_48 = _mm_loadl_epi64((__m128i*)&a[j0 + 48]); in cftmdl_128_SSE2() 104 const __m128i a_56 = _mm_loadl_epi64((__m128i*)&a[j0 + 56]); in cftmdl_128_SSE2() 162 const __m128i a_00 = _mm_loadl_epi64((__m128i*)&a[j0 + 0]); in cftmdl_128_SSE2() 163 const __m128i a_08 = _mm_loadl_epi64((__m128i*)&a[j0 + 8]); in cftmdl_128_SSE2() [all …]
|