Searched refs:_mm_loadu_si128 (Results 1 - 25 of 64) sorted by relevance

123

/external/libavc/encoder/x86/
H A Dime_distortion_metrics_sse42.c116 src_r0 = _mm_loadu_si128((__m128i *) (pu1_src));
117 src_r1 = _mm_loadu_si128((__m128i *) (pu1_src + src_strd));
118 src_r2 = _mm_loadu_si128((__m128i *) (pu1_src + 2*src_strd));
119 src_r3 = _mm_loadu_si128((__m128i *) (pu1_src + 3*src_strd));
121 est_r0 = _mm_loadu_si128((__m128i *) (pu1_est));
122 est_r1 = _mm_loadu_si128((__m128i *) (pu1_est + est_strd));
123 est_r2 = _mm_loadu_si128((__m128i *) (pu1_est + 2*est_strd));
124 est_r3 = _mm_loadu_si128((__m128i *) (pu1_est + 3*est_strd));
139 src_r0 = _mm_loadu_si128((__m128i *) (pu1_src));
140 src_r1 = _mm_loadu_si128((__m128
[all...]
H A Dih264e_half_pel_ssse3.c132 src_r0_16x8b = _mm_loadu_si128((__m128i *)pu1_src); //a0 a1 a2 a3 a4 a5 a6 a7 a8 a9....a15
133 src_r1_16x8b = _mm_loadu_si128((__m128i *)(pu1_src + 8)); //b0 b1 b2 b3 b4 b5 b6 b7 b8 b9....b15
283 src1_r0_16x8b = _mm_loadu_si128((__m128i *)pu1_src);
287 src1_r1_16x8b = _mm_loadu_si128((__m128i *)pu1_src);
291 src1_r2_16x8b = _mm_loadu_si128((__m128i *)pu1_src);
295 src1_r3_16x8b = _mm_loadu_si128((__m128i *)pu1_src);
299 src1_r4_16x8b = _mm_loadu_si128((__m128i *)pu1_src);
305 src1_r5_16x8b = _mm_loadu_si128((__m128i *)pu1_src);
393 src_r0_8x16b = _mm_loadu_si128((__m128i *)(pi2_pred1));
394 src_r1_8x16b = _mm_loadu_si128((__m128
[all...]
H A Dih264e_intra_modes_eval_ssse3.c190 src1_16x8b = _mm_loadu_si128((__m128i *)pu1_src_temp);
191 src2_16x8b = _mm_loadu_si128((__m128i *)(pu1_src_temp + src_strd));
192 src3_16x8b = _mm_loadu_si128((__m128i *)(pu1_src_temp + src_strd2));
193 src4_16x8b = _mm_loadu_si128((__m128i *)(pu1_src_temp + src_strd3));
214 src1_16x8b = _mm_loadu_si128((__m128i *)pu1_src_temp);
215 src2_16x8b = _mm_loadu_si128((__m128i *)(pu1_src_temp + src_strd));
216 src3_16x8b = _mm_loadu_si128((__m128i *)(pu1_src_temp + src_strd2));
217 src4_16x8b = _mm_loadu_si128((__m128i *)(pu1_src_temp + src_strd3));
244 val1_16x8b = _mm_loadu_si128((__m128i *)(pu1_ngbr_pels_i16 + 17));
246 src1_16x8b = _mm_loadu_si128((__m128
[all...]
/external/skia/src/opts/
H A DSkBlitRow_opts_SSE4.cpp28 __m128i s0 = _mm_loadu_si128(src4+i+0),
29 s1 = _mm_loadu_si128(src4+i+1),
30 s2 = _mm_loadu_si128(src4+i+2),
31 s3 = _mm_loadu_si128(src4+i+3);
49 _mm_storeu_si128(dst4+i+0, SkPMSrcOver_SSE2(s0, _mm_loadu_si128(dst4+i+0)));
50 _mm_storeu_si128(dst4+i+1, SkPMSrcOver_SSE2(s1, _mm_loadu_si128(dst4+i+1)));
51 _mm_storeu_si128(dst4+i+2, SkPMSrcOver_SSE2(s2, _mm_loadu_si128(dst4+i+2)));
52 _mm_storeu_si128(dst4+i+3, SkPMSrcOver_SSE2(s3, _mm_loadu_si128(dst4+i+3)));
H A DSkUtils_opts_SSE2.cpp81 __m128i a = _mm_loadu_si128(src128++);
82 __m128i b = _mm_loadu_si128(src128++);
83 __m128i c = _mm_loadu_si128(src128++);
84 __m128i d = _mm_loadu_si128(src128++);
/external/libhevc/common/x86/
H A Dihevc_32x32_itrans_recon_sse42_intr.c251 m_temp_reg_70 = _mm_loadu_si128((__m128i *)pi2_tmp_src);
253 m_temp_reg_71 = _mm_loadu_si128((__m128i *)pi2_tmp_src);
255 m_temp_reg_72 = _mm_loadu_si128((__m128i *)pi2_tmp_src);
257 m_temp_reg_73 = _mm_loadu_si128((__m128i *)pi2_tmp_src);
259 m_temp_reg_74 = _mm_loadu_si128((__m128i *)pi2_tmp_src);
261 m_temp_reg_75 = _mm_loadu_si128((__m128i *)pi2_tmp_src);
263 m_temp_reg_76 = _mm_loadu_si128((__m128i *)pi2_tmp_src);
265 m_temp_reg_77 = _mm_loadu_si128((__m128i *)pi2_tmp_src);
268 m_temp_reg_80 = _mm_loadu_si128((__m128i *)pi2_tmp_src);
270 m_temp_reg_81 = _mm_loadu_si128((__m128
[all...]
H A Dihevc_16x16_itrans_recon_sse42_intr.c204 m_temp_reg_70 = _mm_loadu_si128((__m128i *)pi2_tmp_src);
206 m_temp_reg_71 = _mm_loadu_si128((__m128i *)pi2_tmp_src);
208 m_temp_reg_72 = _mm_loadu_si128((__m128i *)pi2_tmp_src);
210 m_temp_reg_73 = _mm_loadu_si128((__m128i *)pi2_tmp_src);
212 m_temp_reg_74 = _mm_loadu_si128((__m128i *)pi2_tmp_src);
214 m_temp_reg_75 = _mm_loadu_si128((__m128i *)pi2_tmp_src);
216 m_temp_reg_76 = _mm_loadu_si128((__m128i *)pi2_tmp_src);
218 m_temp_reg_77 = _mm_loadu_si128((__m128i *)pi2_tmp_src);
243 m_coeff1 = _mm_loadu_si128((__m128i *)&g_ai2_ihevc_trans_16_even[2][0]); //89 75
289 m_coeff3 = _mm_loadu_si128((__m128
[all...]
H A Dihevc_intra_pred_filters_sse42_intr.c153 src_temp1 = _mm_loadu_si128((__m128i *)(pu1_src));
162 src_temp1 = _mm_loadu_si128((__m128i *)(pu1_src));
163 src_temp2 = _mm_loadu_si128((__m128i *)(pu1_src + 16));
174 src_temp1 = _mm_loadu_si128((__m128i *)(pu1_src));
175 src_temp2 = _mm_loadu_si128((__m128i *)(pu1_src + 16));
176 src_temp3 = _mm_loadu_si128((__m128i *)(pu1_src + 32));
177 src_temp4 = _mm_loadu_si128((__m128i *)(pu1_src + 48));
189 src_temp1 = _mm_loadu_si128((__m128i *)(pu1_src));
190 src_temp2 = _mm_loadu_si128((__m128i *)(pu1_src + 16));
191 src_temp3 = _mm_loadu_si128((__m128
[all...]
H A Dihevc_intra_pred_filters_ssse3_intr.c437 src_temp1 = _mm_loadu_si128((__m128i *)(pu1_src));
446 src_temp1 = _mm_loadu_si128((__m128i *)(pu1_src));
447 src_temp2 = _mm_loadu_si128((__m128i *)(pu1_src + 16));
458 src_temp1 = _mm_loadu_si128((__m128i *)(pu1_src));
459 src_temp2 = _mm_loadu_si128((__m128i *)(pu1_src + 16));
460 src_temp3 = _mm_loadu_si128((__m128i *)(pu1_src + 32));
461 src_temp4 = _mm_loadu_si128((__m128i *)(pu1_src + 48));
473 src_temp1 = _mm_loadu_si128((__m128i *)(pu1_src));
474 src_temp2 = _mm_loadu_si128((__m128i *)(pu1_src + 16));
475 src_temp3 = _mm_loadu_si128((__m128
[all...]
H A Dihevc_weighted_pred_sse42_intr.c152 src_temp0_4x32b = _mm_loadu_si128((__m128i *)(pi2_src));
154 src_temp1_4x32b = _mm_loadu_si128((__m128i *)(pi2_src + src_strd));
156 src_temp2_4x32b = _mm_loadu_si128((__m128i *)(pi2_src + 2 * src_strd));
158 src_temp3_4x32b = _mm_loadu_si128((__m128i *)(pi2_src + 3 * src_strd));
161 src_temp4_4x32b = _mm_loadu_si128((__m128i *)(pi2_src + 4));
163 src_temp5_4x32b = _mm_loadu_si128((__m128i *)(pi2_src + src_strd + 4));
165 src_temp6_4x32b = _mm_loadu_si128((__m128i *)(pi2_src + 2 * src_strd + 4));
167 src_temp7_4x32b = _mm_loadu_si128((__m128i *)(pi2_src + 3 * src_strd + 4));
277 src_temp0_4x32b = _mm_loadu_si128((__m128i *)(pi2_src));
279 src_temp1_4x32b = _mm_loadu_si128((__m128
[all...]
H A Dihevc_padding_ssse3_intr.c109 src_temp0_16x8b = _mm_loadu_si128((__m128i *)pu1_src);
180 src_temp0_16x8b = _mm_loadu_si128((__m128i *)pu1_src);
250 src_temp0_16x8b = _mm_loadu_si128((__m128i *)(pu1_src - 1));
323 src_temp0_16x8b = _mm_loadu_si128((__m128i *)(pu1_src - 2));
H A Dihevc_chroma_intra_pred_filters_ssse3_intr.c229 src_temp_8x16b = _mm_loadu_si128((__m128i *)(pu1_ref + 2 * (two_nt + 1) + col));
344 src_temp3 = _mm_loadu_si128((__m128i *)(pu1_ref + (2 * nt)));
345 src_temp4 = _mm_loadu_si128((__m128i *)(pu1_ref + (2 * nt) + 16));
346 src_temp7 = _mm_loadu_si128((__m128i *)(pu1_ref + (2 * nt) + 32));
347 src_temp8 = _mm_loadu_si128((__m128i *)(pu1_ref + (2 * nt) + 48));
388 src_temp3 = _mm_loadu_si128((__m128i *)(pu1_ref + (2 * nt)));
389 src_temp4 = _mm_loadu_si128((__m128i *)(pu1_ref + (2 * nt) + 16));
419 src_temp3 = _mm_loadu_si128((__m128i *)(pu1_ref + (2 * nt)));
705 src_temp1 = _mm_loadu_si128((__m128i *)(pu1_ref + (4 * nt) + 2 + 0));
722 temp1 = _mm_loadu_si128((__m128
[all...]
H A Dihevc_weighted_pred_ssse3_intr.c165 src_temp0_8x16b = _mm_loadu_si128((__m128i *)(pi2_src));
167 src_temp1_8x16b = _mm_loadu_si128((__m128i *)(pi2_src + src_strd));
169 src_temp2_8x16b = _mm_loadu_si128((__m128i *)(pi2_src + 2 * src_strd));
171 src_temp3_8x16b = _mm_loadu_si128((__m128i *)(pi2_src + 3 * src_strd));
457 src_temp0_8x16b = _mm_loadu_si128((__m128i *)(pi2_src));
459 src_temp1_8x16b = _mm_loadu_si128((__m128i *)(pi2_src + src_strd));
461 src_temp2_8x16b = _mm_loadu_si128((__m128i *)(pi2_src + 8));
463 src_temp3_8x16b = _mm_loadu_si128((__m128i *)(pi2_src + src_strd + 8));
551 src_temp0_8x16b = _mm_loadu_si128((__m128i *)(pi2_src));
553 src_temp1_8x16b = _mm_loadu_si128((__m128
[all...]
H A Dihevc_itrans_recon_sse42_intr.c153 m_coeff3 = _mm_loadu_si128((__m128i *)&g_ai4_ihevc_trans_4_ttype1[2][0]); //74
201 m_coeff1 = _mm_loadu_si128((__m128i *)&g_ai4_ihevc_trans_4_ttype1[1][0]); //29
202 m_coeff2 = _mm_loadu_si128((__m128i *)&g_ai4_ihevc_trans_4_ttype1[0][0]); //55
499 m_coeff1 = _mm_loadu_si128((__m128i *)&g_ai4_ihevc_trans_4_ttype0[0][0]); //36
500 m_coeff3 = _mm_loadu_si128((__m128i *)&g_ai4_ihevc_trans_4_ttype0[2][0]); //83
856 m_temp_reg_70 = _mm_loadu_si128((__m128i *)pi2_src);
858 m_temp_reg_71 = _mm_loadu_si128((__m128i *)pi2_src);
860 m_temp_reg_72 = _mm_loadu_si128((__m128i *)pi2_src);
862 m_temp_reg_73 = _mm_loadu_si128((__m128i *)pi2_src);
865 m_temp_reg_74 = _mm_loadu_si128((__m128
[all...]
/external/libmpeg2/common/x86/
H A Dimpeg2_inter_pred_sse42_intr.c84 src_r0 = _mm_loadu_si128((__m128i *) (src));
85 src_r1 = _mm_loadu_si128((__m128i *) (src + src_wd));
86 src_r2 = _mm_loadu_si128((__m128i *) (src + 2 * src_wd));
87 src_r3 = _mm_loadu_si128((__m128i *) (src + 3 * src_wd));
97 src_r0 = _mm_loadu_si128((__m128i *) (src));
98 src_r1 = _mm_loadu_si128((__m128i *) (src + src_wd));
99 src_r2 = _mm_loadu_si128((__m128i *) (src + 2 * src_wd));
100 src_r3 = _mm_loadu_si128((__m128i *) (src + 3 * src_wd));
110 src_r0 = _mm_loadu_si128((__m128i *) (src));
111 src_r1 = _mm_loadu_si128((__m128
[all...]
H A Dimpeg2_idct_recon_sse42_intr.c194 m_temp_reg_70 = _mm_loadu_si128((__m128i *)pi2_src);
196 m_temp_reg_71 = _mm_loadu_si128((__m128i *)pi2_src);
198 m_temp_reg_72 = _mm_loadu_si128((__m128i *)pi2_src);
200 m_temp_reg_73 = _mm_loadu_si128((__m128i *)pi2_src);
203 m_temp_reg_74 = _mm_loadu_si128((__m128i *)pi2_src);
205 m_temp_reg_75 = _mm_loadu_si128((__m128i *)pi2_src);
207 m_temp_reg_76 = _mm_loadu_si128((__m128i *)pi2_src);
209 m_temp_reg_77 = _mm_loadu_si128((__m128i *)pi2_src);
220 m_coeff2 = _mm_loadu_si128((__m128i *)&gai2_impeg2_idct_even_8_q15[3][0]);
221 m_coeff1 = _mm_loadu_si128((__m128
[all...]
/external/libhevc/decoder/x86/
H A Dihevcd_fmt_conv_ssse3_intr.c124 alt_first_mask = _mm_loadu_si128((__m128i *)&FIRST_ALT_SHUFFLE[0]);
147 src_uv0_8x16b = _mm_loadu_si128((__m128i *)pu1_uv_src_temp);
148 src_uv1_8x16b = _mm_loadu_si128((__m128i *)(pu1_uv_src_temp + 16));
160 src_uv0_8x16b = _mm_loadu_si128((__m128i *)(pu1_uv_src_temp + (1 * src_strd)));
161 src_uv1_8x16b = _mm_loadu_si128((__m128i *)(pu1_uv_src_temp + (1 * src_strd) + 16));
173 src_uv0_8x16b = _mm_loadu_si128((__m128i *)(pu1_uv_src_temp + (2 * src_strd)));
174 src_uv1_8x16b = _mm_loadu_si128((__m128i *)(pu1_uv_src_temp + (2 * src_strd) + 16));
186 src_uv0_8x16b = _mm_loadu_si128((__m128i *)(pu1_uv_src_temp + (3 * src_strd)));
187 src_uv1_8x16b = _mm_loadu_si128((__m128i *)(pu1_uv_src_temp + (3 * src_strd) + 16));
222 src_uv0_8x16b = _mm_loadu_si128((__m128
[all...]
/external/libvpx/libvpx/vp9/common/x86/
H A Dvp9_subpixel_8t_intrin_avx2.c72 filtersReg = _mm_loadu_si128((__m128i *)filter);
107 _mm_loadu_si128((__m128i *)(src_ptr-3)));
109 _mm_loadu_si128((__m128i *)
138 _mm_loadu_si128((__m128i *)(src_ptr+5)));
140 _mm_loadu_si128((__m128i *)
205 srcReg1 = _mm_loadu_si128((__m128i *)(src_ptr-3));
240 srcReg2 = _mm_loadu_si128((__m128i *)(src_ptr+5));
317 filtersReg = _mm_loadu_si128((__m128i *)filter);
347 _mm_loadu_si128((__m128i *)(src_ptr)));
349 _mm_loadu_si128((__m128
[all...]
H A Dvp9_subpixel_8t_intrin_ssse3.c54 filtersReg = _mm_loadu_si128((__m128i *)filter);
75 srcReg = _mm_loadu_si128((__m128i *)(src_ptr-3));
126 filtersReg = _mm_loadu_si128((__m128i *)filter);
150 srcReg = _mm_loadu_si128((__m128i *)(src_ptr-3));
206 filtersReg = _mm_loadu_si128((__m128i *)filter);
230 srcReg1 = _mm_loadu_si128((__m128i *)(src_ptr-3));
257 srcReg2 = _mm_loadu_si128((__m128i *)(src_ptr+5));
322 filtersReg = _mm_loadu_si128((__m128i *)filter);
400 filtersReg = _mm_loadu_si128((__m128i *)filter);
416 srcRegFilt1 = _mm_loadu_si128((__m128
[all...]
/external/libavc/common/x86/
H A Dih264_inter_pred_filters_ssse3.c165 y_0_16x8b = _mm_loadu_si128((__m128i *)pu1_src);
166 y_1_16x8b = _mm_loadu_si128((__m128i *)(pu1_src + src_strd));
167 y_2_16x8b = _mm_loadu_si128((__m128i *)(pu1_src + src_strd2));
168 y_3_16x8b = _mm_loadu_si128((__m128i *)(pu1_src + src_strd3));
169 y_4_16x8b = _mm_loadu_si128((__m128i *)(pu1_src + src_strd4));
170 y_5_16x8b = _mm_loadu_si128((__m128i *)(pu1_src + src_strd5));
171 y_6_16x8b = _mm_loadu_si128((__m128i *)(pu1_src + src_strd6));
172 y_7_16x8b = _mm_loadu_si128((__m128i *)(pu1_src + src_strd7));
252 src_r0_16x8b = _mm_loadu_si128((__m128i *)pu1_src); //a0 a1 a2 a3 a4 a5 a6 a7 a8 a9....a15
253 src_r1_16x8b = _mm_loadu_si128((__m128
[all...]
H A Dih264_weighted_pred_sse42.c170 y0_0_16x8b = _mm_loadu_si128((__m128i *)pu1_src1);
171 y0_1_16x8b = _mm_loadu_si128((__m128i *)(pu1_src1 + src_strd1));
172 y0_2_16x8b = _mm_loadu_si128(
174 y0_3_16x8b = _mm_loadu_si128((__m128i *)(pu1_src1 + src_strd1 * 3));
175 y0_4_16x8b = _mm_loadu_si128(
177 y0_5_16x8b = _mm_loadu_si128((__m128i *)(pu1_src1 + src_strd1 * 5));
178 y0_6_16x8b = _mm_loadu_si128((__m128i *)(pu1_src1 + src_strd1 * 6));
179 y0_7_16x8b = _mm_loadu_si128((__m128i *)(pu1_src1 + src_strd1 * 7));
181 y1_0_16x8b = _mm_loadu_si128((__m128i *)pu1_src2);
182 y1_1_16x8b = _mm_loadu_si128((__m128
[all...]
H A Dih264_iquant_itrans_recon_ssse3.c130 src_r0_r1 = _mm_loadu_si128((__m128i *) (pi2_src)); //a00 a01 a02 a03 a10 a11 a12 a13 -- the source matrix 0th,1st row
131 src_r2_r3 = _mm_loadu_si128((__m128i *) (pi2_src + 8)); //a20 a21 a22 a23 a30 a31 a32 a33 -- the source matrix 2nd,3rd row
132 scalemat_r0_r1 = _mm_loadu_si128((__m128i *) (pu2_iscal_mat)); //b00 b01 b02 b03 b10 b11 b12 b13 -- the scaling matrix 0th,1st row
133 scalemat_r2_r3 = _mm_loadu_si128((__m128i *) (pu2_iscal_mat + 8)); //b20 b21 b22 b23 b30 b31 b32 b33 -- the scaling matrix 2nd,3rd row
134 dequant_r0_r1 = _mm_loadu_si128((__m128i *) (pu2_weigh_mat)); //q00 q01 q02 q03 q10 q11 q12 q13 -- all 16 bits
135 dequant_r2_r3 = _mm_loadu_si128((__m128i *) (pu2_weigh_mat + 8)); //q20 q21 q22 q23 q30 q31 q32 q33 -- all 16 bits
412 src_r0 = _mm_loadu_si128((__m128i *) (pi2_src)); //a00 a01 a02 a03 a04 a05 a06 a07 -- the source matrix 0th row
413 scalemat_r0 = _mm_loadu_si128((__m128i *) (pu2_iscale_mat)); //b00 b01 b02 b03 b04 b05 b06 b07 -- the scaling matrix 0th row
414 dequant_r0 = _mm_loadu_si128((__m128i *) (&pu2_weigh_mat[0])); //q0 q1 q2 q3 q4 q5 q6 q7 -- all 16 bits
435 src_r0 = _mm_loadu_si128((__m128
[all...]
/external/webp/src/dsp/
H A Dalpha_processing_sse2.c41 const __m128i a0 = _mm_loadu_si128(src + 0);
42 const __m128i a1 = _mm_loadu_si128(src + 1);
H A Dlossless_sse2.c163 const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]);
179 const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]);
218 const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]);
257 const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]);
294 const __m128i bgra0 = _mm_loadu_si128(in++); // bgra0|bgra1|bgra2|bgra3
295 const __m128i bgra4 = _mm_loadu_si128(in++); // bgra4|bgra5|bgra6|bgra7
323 const __m128i bgra0 = _mm_loadu_si128(in++); // bgra0|bgra1|bgra2|bgra3
324 const __m128i bgra4 = _mm_loadu_si128(in++); // bgra4|bgra5|bgra6|bgra7
358 const __m128i bgra0 = _mm_loadu_si128(in++); // bgra0|bgra1|bgra2|bgra3
359 const __m128i bgra4 = _mm_loadu_si128(i
[all...]
H A Denc_sse2.c71 const __m128i out0 = _mm_loadu_si128((__m128i*)&out[0]);
72 const __m128i out1 = _mm_loadu_si128((__m128i*)&out[8]);
469 const __m128i src0 = _mm_loadu_si128((__m128i*)&tmp[0]);
470 const __m128i src1 = _mm_loadu_si128((__m128i*)&tmp[4]);
471 const __m128i src2 = _mm_loadu_si128((__m128i*)&tmp[8]);
472 const __m128i src3 = _mm_loadu_si128((__m128i*)&tmp[12]);
500 const __m128i a0 = _mm_loadu_si128((__m128i*)&a[BPS * 0]);
501 const __m128i a1 = _mm_loadu_si128((__m128i*)&a[BPS * 1]);
502 const __m128i a2 = _mm_loadu_si128((__m128i*)&a[BPS * 2]);
503 const __m128i a3 = _mm_loadu_si128((__m128
[all...]

Completed in 457 milliseconds

123