Home
last modified time | relevance | path

Searched refs:_mm_loadu_si128 (Results 1 – 25 of 125) sorted by relevance

12345

/external/libavc/encoder/x86/
Dime_distortion_metrics_sse42.c116 src_r0 = _mm_loadu_si128((__m128i *) (pu1_src)); in ime_compute_sad_16x16_sse42()
117 src_r1 = _mm_loadu_si128((__m128i *) (pu1_src + src_strd)); in ime_compute_sad_16x16_sse42()
118 src_r2 = _mm_loadu_si128((__m128i *) (pu1_src + 2*src_strd)); in ime_compute_sad_16x16_sse42()
119 src_r3 = _mm_loadu_si128((__m128i *) (pu1_src + 3*src_strd)); in ime_compute_sad_16x16_sse42()
121 est_r0 = _mm_loadu_si128((__m128i *) (pu1_est)); in ime_compute_sad_16x16_sse42()
122 est_r1 = _mm_loadu_si128((__m128i *) (pu1_est + est_strd)); in ime_compute_sad_16x16_sse42()
123 est_r2 = _mm_loadu_si128((__m128i *) (pu1_est + 2*est_strd)); in ime_compute_sad_16x16_sse42()
124 est_r3 = _mm_loadu_si128((__m128i *) (pu1_est + 3*est_strd)); in ime_compute_sad_16x16_sse42()
139 src_r0 = _mm_loadu_si128((__m128i *) (pu1_src)); in ime_compute_sad_16x16_sse42()
140 src_r1 = _mm_loadu_si128((__m128i *) (pu1_src + src_strd)); in ime_compute_sad_16x16_sse42()
[all …]
Dih264e_half_pel_ssse3.c132 …src_r0_16x8b = _mm_loadu_si128((__m128i *)pu1_src); //a0 a1 a2 a3 a4 a5 a6 a7 … in ih264e_sixtapfilter_horz_ssse3()
133 …src_r1_16x8b = _mm_loadu_si128((__m128i *)(pu1_src + 8)); //b0 b1 b2 b3 b4 b5 b6 b7 … in ih264e_sixtapfilter_horz_ssse3()
283 src1_r0_16x8b = _mm_loadu_si128((__m128i *)pu1_src); in ih264e_sixtap_filter_2dvh_vert_ssse3()
287 src1_r1_16x8b = _mm_loadu_si128((__m128i *)pu1_src); in ih264e_sixtap_filter_2dvh_vert_ssse3()
291 src1_r2_16x8b = _mm_loadu_si128((__m128i *)pu1_src); in ih264e_sixtap_filter_2dvh_vert_ssse3()
295 src1_r3_16x8b = _mm_loadu_si128((__m128i *)pu1_src); in ih264e_sixtap_filter_2dvh_vert_ssse3()
299 src1_r4_16x8b = _mm_loadu_si128((__m128i *)pu1_src); in ih264e_sixtap_filter_2dvh_vert_ssse3()
305 src1_r5_16x8b = _mm_loadu_si128((__m128i *)pu1_src); in ih264e_sixtap_filter_2dvh_vert_ssse3()
393 src_r0_8x16b = _mm_loadu_si128((__m128i *)(pi2_pred1)); in ih264e_sixtap_filter_2dvh_vert_ssse3()
394 src_r1_8x16b = _mm_loadu_si128((__m128i *)(pi2_pred1 + 1)); in ih264e_sixtap_filter_2dvh_vert_ssse3()
[all …]
/external/libhevc/common/x86/
Dihevc_32x32_itrans_recon_sse42_intr.c251 m_temp_reg_70 = _mm_loadu_si128((__m128i *)pi2_tmp_src); in ihevc_itrans_recon_32x32_sse42()
253 m_temp_reg_71 = _mm_loadu_si128((__m128i *)pi2_tmp_src); in ihevc_itrans_recon_32x32_sse42()
255 m_temp_reg_72 = _mm_loadu_si128((__m128i *)pi2_tmp_src); in ihevc_itrans_recon_32x32_sse42()
257 m_temp_reg_73 = _mm_loadu_si128((__m128i *)pi2_tmp_src); in ihevc_itrans_recon_32x32_sse42()
259 m_temp_reg_74 = _mm_loadu_si128((__m128i *)pi2_tmp_src); in ihevc_itrans_recon_32x32_sse42()
261 m_temp_reg_75 = _mm_loadu_si128((__m128i *)pi2_tmp_src); in ihevc_itrans_recon_32x32_sse42()
263 m_temp_reg_76 = _mm_loadu_si128((__m128i *)pi2_tmp_src); in ihevc_itrans_recon_32x32_sse42()
265 m_temp_reg_77 = _mm_loadu_si128((__m128i *)pi2_tmp_src); in ihevc_itrans_recon_32x32_sse42()
268 m_temp_reg_80 = _mm_loadu_si128((__m128i *)pi2_tmp_src); in ihevc_itrans_recon_32x32_sse42()
270 m_temp_reg_81 = _mm_loadu_si128((__m128i *)pi2_tmp_src); in ihevc_itrans_recon_32x32_sse42()
[all …]
Dihevc_16x16_itrans_recon_sse42_intr.c204 m_temp_reg_70 = _mm_loadu_si128((__m128i *)pi2_tmp_src); in ihevc_itrans_recon_16x16_sse42()
206 m_temp_reg_71 = _mm_loadu_si128((__m128i *)pi2_tmp_src); in ihevc_itrans_recon_16x16_sse42()
208 m_temp_reg_72 = _mm_loadu_si128((__m128i *)pi2_tmp_src); in ihevc_itrans_recon_16x16_sse42()
210 m_temp_reg_73 = _mm_loadu_si128((__m128i *)pi2_tmp_src); in ihevc_itrans_recon_16x16_sse42()
212 m_temp_reg_74 = _mm_loadu_si128((__m128i *)pi2_tmp_src); in ihevc_itrans_recon_16x16_sse42()
214 m_temp_reg_75 = _mm_loadu_si128((__m128i *)pi2_tmp_src); in ihevc_itrans_recon_16x16_sse42()
216 m_temp_reg_76 = _mm_loadu_si128((__m128i *)pi2_tmp_src); in ihevc_itrans_recon_16x16_sse42()
218 m_temp_reg_77 = _mm_loadu_si128((__m128i *)pi2_tmp_src); in ihevc_itrans_recon_16x16_sse42()
243 m_coeff1 = _mm_loadu_si128((__m128i *)&g_ai2_ihevc_trans_16_even[2][0]); //89 75 in ihevc_itrans_recon_16x16_sse42()
289 … m_coeff3 = _mm_loadu_si128((__m128i *)&g_ai2_ihevc_trans_16_even[3][0]); //75 -18 in ihevc_itrans_recon_16x16_sse42()
[all …]
Dihevc_intra_pred_filters_ssse3_intr.c437 src_temp1 = _mm_loadu_si128((__m128i *)(pu1_src)); in ihevc_intra_pred_ref_filtering_ssse3()
446 src_temp1 = _mm_loadu_si128((__m128i *)(pu1_src)); in ihevc_intra_pred_ref_filtering_ssse3()
447 src_temp2 = _mm_loadu_si128((__m128i *)(pu1_src + 16)); in ihevc_intra_pred_ref_filtering_ssse3()
458 src_temp1 = _mm_loadu_si128((__m128i *)(pu1_src)); in ihevc_intra_pred_ref_filtering_ssse3()
459 src_temp2 = _mm_loadu_si128((__m128i *)(pu1_src + 16)); in ihevc_intra_pred_ref_filtering_ssse3()
460 src_temp3 = _mm_loadu_si128((__m128i *)(pu1_src + 32)); in ihevc_intra_pred_ref_filtering_ssse3()
461 src_temp4 = _mm_loadu_si128((__m128i *)(pu1_src + 48)); in ihevc_intra_pred_ref_filtering_ssse3()
473 src_temp1 = _mm_loadu_si128((__m128i *)(pu1_src)); in ihevc_intra_pred_ref_filtering_ssse3()
474 src_temp2 = _mm_loadu_si128((__m128i *)(pu1_src + 16)); in ihevc_intra_pred_ref_filtering_ssse3()
475 src_temp3 = _mm_loadu_si128((__m128i *)(pu1_src + 32)); in ihevc_intra_pred_ref_filtering_ssse3()
[all …]
Dihevc_weighted_pred_sse42_intr.c152 src_temp0_4x32b = _mm_loadu_si128((__m128i *)(pi2_src)); in ihevc_weighted_pred_uni_sse42()
154 src_temp1_4x32b = _mm_loadu_si128((__m128i *)(pi2_src + src_strd)); in ihevc_weighted_pred_uni_sse42()
156 src_temp2_4x32b = _mm_loadu_si128((__m128i *)(pi2_src + 2 * src_strd)); in ihevc_weighted_pred_uni_sse42()
158 src_temp3_4x32b = _mm_loadu_si128((__m128i *)(pi2_src + 3 * src_strd)); in ihevc_weighted_pred_uni_sse42()
161 src_temp4_4x32b = _mm_loadu_si128((__m128i *)(pi2_src + 4)); in ihevc_weighted_pred_uni_sse42()
163 src_temp5_4x32b = _mm_loadu_si128((__m128i *)(pi2_src + src_strd + 4)); in ihevc_weighted_pred_uni_sse42()
165 src_temp6_4x32b = _mm_loadu_si128((__m128i *)(pi2_src + 2 * src_strd + 4)); in ihevc_weighted_pred_uni_sse42()
167 src_temp7_4x32b = _mm_loadu_si128((__m128i *)(pi2_src + 3 * src_strd + 4)); in ihevc_weighted_pred_uni_sse42()
277 src_temp0_4x32b = _mm_loadu_si128((__m128i *)(pi2_src)); in ihevc_weighted_pred_uni_sse42()
279 src_temp1_4x32b = _mm_loadu_si128((__m128i *)(pi2_src + src_strd)); in ihevc_weighted_pred_uni_sse42()
[all …]
Dihevc_chroma_intra_pred_filters_ssse3_intr.c229 src_temp_8x16b = _mm_loadu_si128((__m128i *)(pu1_ref + 2 * (two_nt + 1) + col)); in ihevc_intra_pred_chroma_planar_ssse3()
344 src_temp3 = _mm_loadu_si128((__m128i *)(pu1_ref + (2 * nt))); in ihevc_intra_pred_chroma_dc_ssse3()
345 src_temp4 = _mm_loadu_si128((__m128i *)(pu1_ref + (2 * nt) + 16)); in ihevc_intra_pred_chroma_dc_ssse3()
346 src_temp7 = _mm_loadu_si128((__m128i *)(pu1_ref + (2 * nt) + 32)); in ihevc_intra_pred_chroma_dc_ssse3()
347 src_temp8 = _mm_loadu_si128((__m128i *)(pu1_ref + (2 * nt) + 48)); in ihevc_intra_pred_chroma_dc_ssse3()
388 src_temp3 = _mm_loadu_si128((__m128i *)(pu1_ref + (2 * nt))); in ihevc_intra_pred_chroma_dc_ssse3()
389 src_temp4 = _mm_loadu_si128((__m128i *)(pu1_ref + (2 * nt) + 16)); in ihevc_intra_pred_chroma_dc_ssse3()
419 src_temp3 = _mm_loadu_si128((__m128i *)(pu1_ref + (2 * nt))); in ihevc_intra_pred_chroma_dc_ssse3()
705 src_temp1 = _mm_loadu_si128((__m128i *)(pu1_ref + (4 * nt) + 2 + 0)); in ihevc_intra_pred_chroma_ver_ssse3()
722 temp1 = _mm_loadu_si128((__m128i *)(pu1_ref + (4 * nt) + 2 + 0)); in ihevc_intra_pred_chroma_ver_ssse3()
[all …]
Dihevc_intra_pred_filters_sse42_intr.c153 src_temp1 = _mm_loadu_si128((__m128i *)(pu1_src)); in ihevc_intra_pred_ref_filtering_sse42()
162 src_temp1 = _mm_loadu_si128((__m128i *)(pu1_src)); in ihevc_intra_pred_ref_filtering_sse42()
163 src_temp2 = _mm_loadu_si128((__m128i *)(pu1_src + 16)); in ihevc_intra_pred_ref_filtering_sse42()
174 src_temp1 = _mm_loadu_si128((__m128i *)(pu1_src)); in ihevc_intra_pred_ref_filtering_sse42()
175 src_temp2 = _mm_loadu_si128((__m128i *)(pu1_src + 16)); in ihevc_intra_pred_ref_filtering_sse42()
176 src_temp3 = _mm_loadu_si128((__m128i *)(pu1_src + 32)); in ihevc_intra_pred_ref_filtering_sse42()
177 src_temp4 = _mm_loadu_si128((__m128i *)(pu1_src + 48)); in ihevc_intra_pred_ref_filtering_sse42()
189 src_temp1 = _mm_loadu_si128((__m128i *)(pu1_src)); in ihevc_intra_pred_ref_filtering_sse42()
190 src_temp2 = _mm_loadu_si128((__m128i *)(pu1_src + 16)); in ihevc_intra_pred_ref_filtering_sse42()
191 src_temp3 = _mm_loadu_si128((__m128i *)(pu1_src + 32)); in ihevc_intra_pred_ref_filtering_sse42()
[all …]
Dihevc_weighted_pred_ssse3_intr.c165 src_temp0_8x16b = _mm_loadu_si128((__m128i *)(pi2_src)); in ihevc_weighted_pred_uni_ssse3()
167 src_temp1_8x16b = _mm_loadu_si128((__m128i *)(pi2_src + src_strd)); in ihevc_weighted_pred_uni_ssse3()
169 src_temp2_8x16b = _mm_loadu_si128((__m128i *)(pi2_src + 2 * src_strd)); in ihevc_weighted_pred_uni_ssse3()
171 src_temp3_8x16b = _mm_loadu_si128((__m128i *)(pi2_src + 3 * src_strd)); in ihevc_weighted_pred_uni_ssse3()
457 src_temp0_8x16b = _mm_loadu_si128((__m128i *)(pi2_src)); in ihevc_weighted_pred_chroma_uni_ssse3()
459 src_temp1_8x16b = _mm_loadu_si128((__m128i *)(pi2_src + src_strd)); in ihevc_weighted_pred_chroma_uni_ssse3()
461 src_temp2_8x16b = _mm_loadu_si128((__m128i *)(pi2_src + 8)); in ihevc_weighted_pred_chroma_uni_ssse3()
463 src_temp3_8x16b = _mm_loadu_si128((__m128i *)(pi2_src + src_strd + 8)); in ihevc_weighted_pred_chroma_uni_ssse3()
551 src_temp0_8x16b = _mm_loadu_si128((__m128i *)(pi2_src)); in ihevc_weighted_pred_chroma_uni_ssse3()
553 src_temp1_8x16b = _mm_loadu_si128((__m128i *)(pi2_src + src_strd)); in ihevc_weighted_pred_chroma_uni_ssse3()
[all …]
Dihevc_itrans_recon_sse42_intr.c153 m_coeff3 = _mm_loadu_si128((__m128i *)&g_ai4_ihevc_trans_4_ttype1[2][0]); //74 in ihevc_itrans_recon_4x4_ttype1_sse42()
201 m_coeff1 = _mm_loadu_si128((__m128i *)&g_ai4_ihevc_trans_4_ttype1[1][0]); //29 in ihevc_itrans_recon_4x4_ttype1_sse42()
202 m_coeff2 = _mm_loadu_si128((__m128i *)&g_ai4_ihevc_trans_4_ttype1[0][0]); //55 in ihevc_itrans_recon_4x4_ttype1_sse42()
499 m_coeff1 = _mm_loadu_si128((__m128i *)&g_ai4_ihevc_trans_4_ttype0[0][0]); //36 in ihevc_itrans_recon_4x4_sse42()
500 m_coeff3 = _mm_loadu_si128((__m128i *)&g_ai4_ihevc_trans_4_ttype0[2][0]); //83 in ihevc_itrans_recon_4x4_sse42()
856 m_temp_reg_70 = _mm_loadu_si128((__m128i *)pi2_src); in ihevc_itrans_recon_8x8_sse42()
858 m_temp_reg_71 = _mm_loadu_si128((__m128i *)pi2_src); in ihevc_itrans_recon_8x8_sse42()
860 m_temp_reg_72 = _mm_loadu_si128((__m128i *)pi2_src); in ihevc_itrans_recon_8x8_sse42()
862 m_temp_reg_73 = _mm_loadu_si128((__m128i *)pi2_src); in ihevc_itrans_recon_8x8_sse42()
865 m_temp_reg_74 = _mm_loadu_si128((__m128i *)pi2_src); in ihevc_itrans_recon_8x8_sse42()
[all …]
/external/webrtc/webrtc/modules/desktop_capture/
Ddiffer_block_sse2.cc34 v0 = _mm_loadu_si128(i1); in BlockDifference_SSE2_W16()
35 v1 = _mm_loadu_si128(i2); in BlockDifference_SSE2_W16()
38 v0 = _mm_loadu_si128(i1 + 1); in BlockDifference_SSE2_W16()
39 v1 = _mm_loadu_si128(i2 + 1); in BlockDifference_SSE2_W16()
42 v0 = _mm_loadu_si128(i1 + 2); in BlockDifference_SSE2_W16()
43 v1 = _mm_loadu_si128(i2 + 2); in BlockDifference_SSE2_W16()
46 v0 = _mm_loadu_si128(i1 + 3); in BlockDifference_SSE2_W16()
47 v1 = _mm_loadu_si128(i2 + 3); in BlockDifference_SSE2_W16()
74 v0 = _mm_loadu_si128(i1); in BlockDifference_SSE2_W32()
75 v1 = _mm_loadu_si128(i2); in BlockDifference_SSE2_W32()
[all …]
/external/webp/src/dsp/
Dlossless_enc_sse2.c33 const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]); // argb in SubtractGreenFromBlueAndRed_SSE2()
63 const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]); // argb in TransformColor_SSE2()
102 const __m128i in0 = _mm_loadu_si128((__m128i*)&src[x + 0]); in CollectColorBlueTransforms_SSE2()
103 const __m128i in1 = _mm_loadu_si128((__m128i*)&src[x + SPAN / 2]); in CollectColorBlueTransforms_SSE2()
150 const __m128i in0 = _mm_loadu_si128((__m128i*)&src[x + 0]); in CollectColorRedTransforms_SSE2()
151 const __m128i in1 = _mm_loadu_si128((__m128i*)&src[x + SPAN / 2]); in CollectColorRedTransforms_SSE2()
186 const __m128i a0 = _mm_loadu_si128((const __m128i*)&a[i + 0]); in AddVector_SSE2()
187 const __m128i a1 = _mm_loadu_si128((const __m128i*)&a[i + 4]); in AddVector_SSE2()
189 const __m128i a2 = _mm_loadu_si128((const __m128i*)&a[i + 8]); in AddVector_SSE2()
190 const __m128i a3 = _mm_loadu_si128((const __m128i*)&a[i + 12]); in AddVector_SSE2()
[all …]
Denc_sse41.c41 const __m128i out0 = _mm_loadu_si128((__m128i*)&out[0]); in CollectHistogram_SSE41()
42 const __m128i out1 = _mm_loadu_si128((__m128i*)&out[8]); in CollectHistogram_SSE41()
80 const __m128i inA_0 = _mm_loadu_si128((const __m128i*)&inA[BPS * 0]); in TTransform_SSE41()
81 const __m128i inA_1 = _mm_loadu_si128((const __m128i*)&inA[BPS * 1]); in TTransform_SSE41()
82 const __m128i inA_2 = _mm_loadu_si128((const __m128i*)&inA[BPS * 2]); in TTransform_SSE41()
88 const __m128i inB_0 = _mm_loadu_si128((const __m128i*)&inB[BPS * 0]); in TTransform_SSE41()
89 const __m128i inB_1 = _mm_loadu_si128((const __m128i*)&inB[BPS * 1]); in TTransform_SSE41()
90 const __m128i inB_2 = _mm_loadu_si128((const __m128i*)&inB[BPS * 2]); in TTransform_SSE41()
132 const __m128i w_0 = _mm_loadu_si128((const __m128i*)&w[0]); in TTransform_SSE41()
133 const __m128i w_8 = _mm_loadu_si128((const __m128i*)&w[8]); in TTransform_SSE41()
[all …]
Dlossless_sse2.c189 const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]); in PredictorAdd0_SSE2()
205 const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]); in PredictorAdd1_SSE2()
231 const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]); \
232 const __m128i other = _mm_loadu_si128((const __m128i*)&(IN)); \
260 const __m128i Tother = _mm_loadu_si128((const __m128i*)&(IN)); \ in GENERATE_PREDICTOR_ADD()
261 const __m128i T = _mm_loadu_si128((const __m128i*)&upper[i]); \ in GENERATE_PREDICTOR_ADD()
262 const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]); \ in GENERATE_PREDICTOR_ADD()
299 __m128i src = _mm_loadu_si128((const __m128i*)&in[i]);
300 __m128i TL = _mm_loadu_si128((const __m128i*)&upper[i - 1]);
301 const __m128i T = _mm_loadu_si128((const __m128i*)&upper[i]);
[all …]
/external/flac/libFLAC/
Dlpc_intrin_sse2.c86 summ = _mm_madd_epi16(q11, _mm_loadu_si128((const __m128i*)(data+i-12))); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
87 …mull = _mm_madd_epi16(q10, _mm_loadu_si128((const __m128i*)(data+i-11))); summ = _mm_add_epi32(sum… in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
88 …mull = _mm_madd_epi16(q9, _mm_loadu_si128((const __m128i*)(data+i-10))); summ = _mm_add_epi32(summ… in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
89 …mull = _mm_madd_epi16(q8, _mm_loadu_si128((const __m128i*)(data+i-9))); summ = _mm_add_epi32(summ,… in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
90 …mull = _mm_madd_epi16(q7, _mm_loadu_si128((const __m128i*)(data+i-8))); summ = _mm_add_epi32(summ,… in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
91 …mull = _mm_madd_epi16(q6, _mm_loadu_si128((const __m128i*)(data+i-7))); summ = _mm_add_epi32(summ,… in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
92 …mull = _mm_madd_epi16(q5, _mm_loadu_si128((const __m128i*)(data+i-6))); summ = _mm_add_epi32(summ,… in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
93 …mull = _mm_madd_epi16(q4, _mm_loadu_si128((const __m128i*)(data+i-5))); summ = _mm_add_epi32(summ,… in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
94 …mull = _mm_madd_epi16(q3, _mm_loadu_si128((const __m128i*)(data+i-4))); summ = _mm_add_epi32(summ,… in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
95 …mull = _mm_madd_epi16(q2, _mm_loadu_si128((const __m128i*)(data+i-3))); summ = _mm_add_epi32(summ,… in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2()
[all …]
Dlpc_intrin_sse41.c987 summ = _mm_mullo_epi32(q11, _mm_loadu_si128((const __m128i*)(data+i-12))); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
988 …mull = _mm_mullo_epi32(q10, _mm_loadu_si128((const __m128i*)(data+i-11))); summ = _mm_add_epi32(su… in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
989 …mull = _mm_mullo_epi32(q9, _mm_loadu_si128((const __m128i*)(data+i-10))); summ = _mm_add_epi32(sum… in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
990 …mull = _mm_mullo_epi32(q8, _mm_loadu_si128((const __m128i*)(data+i-9))); summ = _mm_add_epi32(summ… in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
991 …mull = _mm_mullo_epi32(q7, _mm_loadu_si128((const __m128i*)(data+i-8))); summ = _mm_add_epi32(summ… in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
992 …mull = _mm_mullo_epi32(q6, _mm_loadu_si128((const __m128i*)(data+i-7))); summ = _mm_add_epi32(summ… in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
993 …mull = _mm_mullo_epi32(q5, _mm_loadu_si128((const __m128i*)(data+i-6))); summ = _mm_add_epi32(summ… in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
994 …mull = _mm_mullo_epi32(q4, _mm_loadu_si128((const __m128i*)(data+i-5))); summ = _mm_add_epi32(summ… in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
995 …mull = _mm_mullo_epi32(q3, _mm_loadu_si128((const __m128i*)(data+i-4))); summ = _mm_add_epi32(summ… in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
996 …mull = _mm_mullo_epi32(q2, _mm_loadu_si128((const __m128i*)(data+i-3))); summ = _mm_add_epi32(summ… in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
[all …]
Dlpc_intrin_avx2.c794 … summ = _mm256_mul_epi32(q11, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-12)))); in FLAC__lpc_compute_residual_from_qlp_coefficients_wide_intrin_avx2()
795 …mull = _mm256_mul_epi32(q10, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-11))));… in FLAC__lpc_compute_residual_from_qlp_coefficients_wide_intrin_avx2()
796 …mull = _mm256_mul_epi32(q9, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-10))));… in FLAC__lpc_compute_residual_from_qlp_coefficients_wide_intrin_avx2()
797 …mull = _mm256_mul_epi32(q8, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-9 ))));… in FLAC__lpc_compute_residual_from_qlp_coefficients_wide_intrin_avx2()
798 …mull = _mm256_mul_epi32(q7, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-8 ))));… in FLAC__lpc_compute_residual_from_qlp_coefficients_wide_intrin_avx2()
799 …mull = _mm256_mul_epi32(q6, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-7 ))));… in FLAC__lpc_compute_residual_from_qlp_coefficients_wide_intrin_avx2()
800 …mull = _mm256_mul_epi32(q5, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-6 ))));… in FLAC__lpc_compute_residual_from_qlp_coefficients_wide_intrin_avx2()
801 …mull = _mm256_mul_epi32(q4, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-5 ))));… in FLAC__lpc_compute_residual_from_qlp_coefficients_wide_intrin_avx2()
802 …mull = _mm256_mul_epi32(q3, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-4 ))));… in FLAC__lpc_compute_residual_from_qlp_coefficients_wide_intrin_avx2()
803 …mull = _mm256_mul_epi32(q2, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-3 ))));… in FLAC__lpc_compute_residual_from_qlp_coefficients_wide_intrin_avx2()
[all …]
/external/libmpeg2/common/x86/
Dimpeg2_inter_pred_sse42_intr.c84 src_r0 = _mm_loadu_si128((__m128i *) (src)); in impeg2_copy_mb_sse42()
85 src_r1 = _mm_loadu_si128((__m128i *) (src + src_wd)); in impeg2_copy_mb_sse42()
86 src_r2 = _mm_loadu_si128((__m128i *) (src + 2 * src_wd)); in impeg2_copy_mb_sse42()
87 src_r3 = _mm_loadu_si128((__m128i *) (src + 3 * src_wd)); in impeg2_copy_mb_sse42()
97 src_r0 = _mm_loadu_si128((__m128i *) (src)); in impeg2_copy_mb_sse42()
98 src_r1 = _mm_loadu_si128((__m128i *) (src + src_wd)); in impeg2_copy_mb_sse42()
99 src_r2 = _mm_loadu_si128((__m128i *) (src + 2 * src_wd)); in impeg2_copy_mb_sse42()
100 src_r3 = _mm_loadu_si128((__m128i *) (src + 3 * src_wd)); in impeg2_copy_mb_sse42()
110 src_r0 = _mm_loadu_si128((__m128i *) (src)); in impeg2_copy_mb_sse42()
111 src_r1 = _mm_loadu_si128((__m128i *) (src + src_wd)); in impeg2_copy_mb_sse42()
[all …]
Dimpeg2_idct_recon_sse42_intr.c194 m_temp_reg_70 = _mm_loadu_si128((__m128i *)pi2_src); in impeg2_idct_recon_sse42()
196 m_temp_reg_71 = _mm_loadu_si128((__m128i *)pi2_src); in impeg2_idct_recon_sse42()
198 m_temp_reg_72 = _mm_loadu_si128((__m128i *)pi2_src); in impeg2_idct_recon_sse42()
200 m_temp_reg_73 = _mm_loadu_si128((__m128i *)pi2_src); in impeg2_idct_recon_sse42()
203 m_temp_reg_74 = _mm_loadu_si128((__m128i *)pi2_src); in impeg2_idct_recon_sse42()
205 m_temp_reg_75 = _mm_loadu_si128((__m128i *)pi2_src); in impeg2_idct_recon_sse42()
207 m_temp_reg_76 = _mm_loadu_si128((__m128i *)pi2_src); in impeg2_idct_recon_sse42()
209 m_temp_reg_77 = _mm_loadu_si128((__m128i *)pi2_src); in impeg2_idct_recon_sse42()
220 m_coeff2 = _mm_loadu_si128((__m128i *)&gai2_impeg2_idct_even_8_q15[3][0]); in impeg2_idct_recon_sse42()
221 m_coeff1 = _mm_loadu_si128((__m128i *)&gai2_impeg2_idct_even_8_q15[0][0]); in impeg2_idct_recon_sse42()
[all …]
/external/skqp/src/opts/
DSkBlitRow_opts.h88 auto s0 = _mm_loadu_si128((const __m128i*)(src) + 0), in blit_row_s32a_opaque()
89 s1 = _mm_loadu_si128((const __m128i*)(src) + 1), in blit_row_s32a_opaque()
90 s2 = _mm_loadu_si128((const __m128i*)(src) + 2), in blit_row_s32a_opaque()
91 s3 = _mm_loadu_si128((const __m128i*)(src) + 3); in blit_row_s32a_opaque()
124 _mm_storeu_si128(d0, SkPMSrcOver_SSE2(s0, _mm_loadu_si128(d0))); in blit_row_s32a_opaque()
125 _mm_storeu_si128(d1, SkPMSrcOver_SSE2(s1, _mm_loadu_si128(d1))); in blit_row_s32a_opaque()
126 _mm_storeu_si128(d2, SkPMSrcOver_SSE2(s2, _mm_loadu_si128(d2))); in blit_row_s32a_opaque()
127 _mm_storeu_si128(d3, SkPMSrcOver_SSE2(s3, _mm_loadu_si128(d3))); in blit_row_s32a_opaque()
136 auto s0 = _mm_loadu_si128((const __m128i*)(src) + 0), in blit_row_s32a_opaque()
137 s1 = _mm_loadu_si128((const __m128i*)(src) + 1), in blit_row_s32a_opaque()
[all …]
/external/skia/src/opts/
DSkBlitRow_opts.h88 auto s0 = _mm_loadu_si128((const __m128i*)(src) + 0), in blit_row_s32a_opaque()
89 s1 = _mm_loadu_si128((const __m128i*)(src) + 1), in blit_row_s32a_opaque()
90 s2 = _mm_loadu_si128((const __m128i*)(src) + 2), in blit_row_s32a_opaque()
91 s3 = _mm_loadu_si128((const __m128i*)(src) + 3); in blit_row_s32a_opaque()
124 _mm_storeu_si128(d0, SkPMSrcOver_SSE2(s0, _mm_loadu_si128(d0))); in blit_row_s32a_opaque()
125 _mm_storeu_si128(d1, SkPMSrcOver_SSE2(s1, _mm_loadu_si128(d1))); in blit_row_s32a_opaque()
126 _mm_storeu_si128(d2, SkPMSrcOver_SSE2(s2, _mm_loadu_si128(d2))); in blit_row_s32a_opaque()
127 _mm_storeu_si128(d3, SkPMSrcOver_SSE2(s3, _mm_loadu_si128(d3))); in blit_row_s32a_opaque()
136 auto s0 = _mm_loadu_si128((const __m128i*)(src) + 0), in blit_row_s32a_opaque()
137 s1 = _mm_loadu_si128((const __m128i*)(src) + 1), in blit_row_s32a_opaque()
[all …]
/external/libavc/common/x86/
Dih264_inter_pred_filters_ssse3.c165 y_0_16x8b = _mm_loadu_si128((__m128i *)pu1_src); in ih264_inter_pred_luma_copy_ssse3()
166 y_1_16x8b = _mm_loadu_si128((__m128i *)(pu1_src + src_strd)); in ih264_inter_pred_luma_copy_ssse3()
167 y_2_16x8b = _mm_loadu_si128((__m128i *)(pu1_src + src_strd2)); in ih264_inter_pred_luma_copy_ssse3()
168 y_3_16x8b = _mm_loadu_si128((__m128i *)(pu1_src + src_strd3)); in ih264_inter_pred_luma_copy_ssse3()
169 y_4_16x8b = _mm_loadu_si128((__m128i *)(pu1_src + src_strd4)); in ih264_inter_pred_luma_copy_ssse3()
170 y_5_16x8b = _mm_loadu_si128((__m128i *)(pu1_src + src_strd5)); in ih264_inter_pred_luma_copy_ssse3()
171 y_6_16x8b = _mm_loadu_si128((__m128i *)(pu1_src + src_strd6)); in ih264_inter_pred_luma_copy_ssse3()
172 y_7_16x8b = _mm_loadu_si128((__m128i *)(pu1_src + src_strd7)); in ih264_inter_pred_luma_copy_ssse3()
252 …src_r0_16x8b = _mm_loadu_si128((__m128i *)pu1_src); //a0 a1 a2 a3 a4 a5 a6 a7 … in ih264_inter_pred_luma_horz_ssse3()
253 …src_r1_16x8b = _mm_loadu_si128((__m128i *)(pu1_src + src_strd)); //b0 b1 b2 b3 b4 b5 b6 b7 … in ih264_inter_pred_luma_horz_ssse3()
[all …]
/external/libopus/celt/x86/
Dpitch_sse4_1.c61 inVec1_76543210 = _mm_loadu_si128((__m128i *)(&x[i + 0])); in celt_inner_prod_sse4_1()
62 inVec2_76543210 = _mm_loadu_si128((__m128i *)(&y[i + 0])); in celt_inner_prod_sse4_1()
64 inVec1_FEDCBA98 = _mm_loadu_si128((__m128i *)(&x[i + 8])); in celt_inner_prod_sse4_1()
65 inVec2_FEDCBA98 = _mm_loadu_si128((__m128i *)(&y[i + 8])); in celt_inner_prod_sse4_1()
78 inVec1_76543210 = _mm_loadu_si128((__m128i *)(&x[i + 0])); in celt_inner_prod_sse4_1()
79 inVec2_76543210 = _mm_loadu_si128((__m128i *)(&y[i + 0])); in celt_inner_prod_sse4_1()
129 vecX = _mm_loadu_si128((__m128i *)(&x[j + 0])); in xcorr_kernel_sse4_1()
130 vecY0 = _mm_loadu_si128((__m128i *)(&y[j + 0])); in xcorr_kernel_sse4_1()
131 vecY1 = _mm_loadu_si128((__m128i *)(&y[j + 1])); in xcorr_kernel_sse4_1()
132 vecY2 = _mm_loadu_si128((__m128i *)(&y[j + 2])); in xcorr_kernel_sse4_1()
[all …]
/external/libhevc/decoder/x86/
Dihevcd_fmt_conv_ssse3_intr.c124 alt_first_mask = _mm_loadu_si128((__m128i *)&FIRST_ALT_SHUFFLE[0]); in ihevcd_fmt_conv_420sp_to_420p_ssse3()
147 src_uv0_8x16b = _mm_loadu_si128((__m128i *)pu1_uv_src_temp); in ihevcd_fmt_conv_420sp_to_420p_ssse3()
148 src_uv1_8x16b = _mm_loadu_si128((__m128i *)(pu1_uv_src_temp + 16)); in ihevcd_fmt_conv_420sp_to_420p_ssse3()
160 src_uv0_8x16b = _mm_loadu_si128((__m128i *)(pu1_uv_src_temp + (1 * src_strd))); in ihevcd_fmt_conv_420sp_to_420p_ssse3()
161 … src_uv1_8x16b = _mm_loadu_si128((__m128i *)(pu1_uv_src_temp + (1 * src_strd) + 16)); in ihevcd_fmt_conv_420sp_to_420p_ssse3()
173 src_uv0_8x16b = _mm_loadu_si128((__m128i *)(pu1_uv_src_temp + (2 * src_strd))); in ihevcd_fmt_conv_420sp_to_420p_ssse3()
174 … src_uv1_8x16b = _mm_loadu_si128((__m128i *)(pu1_uv_src_temp + (2 * src_strd) + 16)); in ihevcd_fmt_conv_420sp_to_420p_ssse3()
186 src_uv0_8x16b = _mm_loadu_si128((__m128i *)(pu1_uv_src_temp + (3 * src_strd))); in ihevcd_fmt_conv_420sp_to_420p_ssse3()
187 … src_uv1_8x16b = _mm_loadu_si128((__m128i *)(pu1_uv_src_temp + (3 * src_strd) + 16)); in ihevcd_fmt_conv_420sp_to_420p_ssse3()
222 src_uv0_8x16b = _mm_loadu_si128((__m128i *)pu1_uv_src_temp); in ihevcd_fmt_conv_420sp_to_420p_ssse3()
[all …]
/external/libvpx/libvpx/vpx_dsp/x86/
Dvpx_subpixel_8t_intrin_avx2.c63 _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(src_ptr - 3))); in vpx_filter_block1d16_h8_x_avx2()
66 _mm_loadu_si128((const __m128i *)(src_ptr + src_pixels_per_line - 3)), in vpx_filter_block1d16_h8_x_avx2()
79 _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(src_ptr + 5))); in vpx_filter_block1d16_h8_x_avx2()
82 _mm_loadu_si128((const __m128i *)(src_ptr + src_pixels_per_line + 5)), in vpx_filter_block1d16_h8_x_avx2()
122 srcReg = _mm_loadu_si128((const __m128i *)(src_ptr - 3)); in vpx_filter_block1d16_h8_x_avx2()
137 srcReg = _mm_loadu_si128((const __m128i *)(src_ptr + 5)); in vpx_filter_block1d16_h8_x_avx2()
199 s[0] = _mm_loadu_si128((const __m128i *)(src_ptr + 0 * src_pitch)); in vpx_filter_block1d16_v8_x_avx2()
200 s[1] = _mm_loadu_si128((const __m128i *)(src_ptr + 1 * src_pitch)); in vpx_filter_block1d16_v8_x_avx2()
201 s[2] = _mm_loadu_si128((const __m128i *)(src_ptr + 2 * src_pitch)); in vpx_filter_block1d16_v8_x_avx2()
202 s[3] = _mm_loadu_si128((const __m128i *)(src_ptr + 3 * src_pitch)); in vpx_filter_block1d16_v8_x_avx2()
[all …]

12345