Searched refs:_mm_load_si128 (Results 1 - 25 of 43) sorted by relevance

12

/external/libvpx/libvpx/vp8/encoder/x86/
H A Dvp8_quantize_sse2.c44 __m128i quant_shift0 = _mm_load_si128((__m128i *)(b->quant_shift));
45 __m128i quant_shift1 = _mm_load_si128((__m128i *)(b->quant_shift + 8));
46 __m128i z0 = _mm_load_si128((__m128i *)(b->coeff));
47 __m128i z1 = _mm_load_si128((__m128i *)(b->coeff+8));
49 __m128i zbin0 = _mm_load_si128((__m128i *)(b->zbin));
50 __m128i zbin1 = _mm_load_si128((__m128i *)(b->zbin + 8));
51 __m128i round0 = _mm_load_si128((__m128i *)(b->round));
52 __m128i round1 = _mm_load_si128((__m128i *)(b->round + 8));
53 __m128i quant0 = _mm_load_si128((__m128i *)(b->quant));
54 __m128i quant1 = _mm_load_si128((__m128
[all...]
H A Dquantize_ssse3.c45 __m128i z0 = _mm_load_si128((__m128i *)(b->coeff));
46 __m128i z1 = _mm_load_si128((__m128i *)(b->coeff + 8));
47 __m128i round0 = _mm_load_si128((__m128i *)(b->round));
48 __m128i round1 = _mm_load_si128((__m128i *)(b->round + 8));
49 __m128i quant_fast0 = _mm_load_si128((__m128i *)(b->quant_fast));
50 __m128i quant_fast1 = _mm_load_si128((__m128i *)(b->quant_fast + 8));
51 __m128i dequant0 = _mm_load_si128((__m128i *)(d->dequant));
52 __m128i dequant1 = _mm_load_si128((__m128i *)(d->dequant + 8));
58 __m128i zig_zag = _mm_load_si128((const __m128i *)pshufb_zig_zag_mask);
H A Dquantize_sse4.c38 __m128i quant_shift0 = _mm_load_si128((__m128i *)(b->quant_shift));
39 __m128i quant_shift1 = _mm_load_si128((__m128i *)(b->quant_shift + 8));
40 __m128i z0 = _mm_load_si128((__m128i *)(b->coeff));
41 __m128i z1 = _mm_load_si128((__m128i *)(b->coeff+8));
43 __m128i zbin0 = _mm_load_si128((__m128i *)(b->zbin));
44 __m128i zbin1 = _mm_load_si128((__m128i *)(b->zbin + 8));
45 __m128i round0 = _mm_load_si128((__m128i *)(b->round));
46 __m128i round1 = _mm_load_si128((__m128i *)(b->round + 8));
47 __m128i quant0 = _mm_load_si128((__m128i *)(b->quant));
48 __m128i quant1 = _mm_load_si128((__m128
[all...]
/external/libvpx/libvpx/vpx_dsp/x86/
H A Dfwd_txfm_sse2.c47 __m128i in0 = _mm_load_si128((const __m128i *)(input + 0 * stride));
48 __m128i in1 = _mm_load_si128((const __m128i *)(input + 1 * stride));
49 __m128i in2 = _mm_load_si128((const __m128i *)(input + 2 * stride));
50 __m128i in3 = _mm_load_si128((const __m128i *)(input + 3 * stride));
56 in0 = _mm_load_si128((const __m128i *)(input + 4 * stride));
57 in1 = _mm_load_si128((const __m128i *)(input + 5 * stride));
58 in2 = _mm_load_si128((const __m128i *)(input + 6 * stride));
59 in3 = _mm_load_si128((const __m128i *)(input + 7 * stride));
95 in0 = _mm_load_si128((const __m128i *)(input + 0 * stride));
96 in1 = _mm_load_si128((cons
[all...]
H A Dinv_txfm_sse2.h93 in[0] = _mm_load_si128((const __m128i *)(input + 0 * 16));
94 in[1] = _mm_load_si128((const __m128i *)(input + 1 * 16));
95 in[2] = _mm_load_si128((const __m128i *)(input + 2 * 16));
96 in[3] = _mm_load_si128((const __m128i *)(input + 3 * 16));
97 in[4] = _mm_load_si128((const __m128i *)(input + 4 * 16));
98 in[5] = _mm_load_si128((const __m128i *)(input + 5 * 16));
99 in[6] = _mm_load_si128((const __m128i *)(input + 6 * 16));
100 in[7] = _mm_load_si128((const __m128i *)(input + 7 * 16));
102 in[8] = _mm_load_si128((const __m128i *)(input + 8 * 16));
103 in[9] = _mm_load_si128((cons
[all...]
H A Dquantize_sse2.c44 zbin = _mm_load_si128((const __m128i*)zbin_ptr);
45 round = _mm_load_si128((const __m128i*)round_ptr);
46 quant = _mm_load_si128((const __m128i*)quant_ptr);
49 dequant = _mm_load_si128((const __m128i*)dequant_ptr);
50 shift = _mm_load_si128((const __m128i*)quant_shift_ptr);
59 coeff0 = _mm_load_si128((const __m128i*)(coeff_ptr + n_coeffs));
60 coeff1 = _mm_load_si128((const __m128i*)(coeff_ptr + n_coeffs) + 1);
116 iscan0 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs));
117 iscan1 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs) + 1);
137 coeff0 = _mm_load_si128((cons
[all...]
H A Dhighbd_loopfilter_sse2.c79 blimit = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero);
80 limit = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero);
81 thresh = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero);
84 _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero), 2);
86 _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero), 2);
88 _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero), 2);
91 _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero), 4);
93 _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero), 4);
95 _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero), 4);
98 q4 = _mm_load_si128((__m128
[all...]
H A Dfwd_txfm_impl_sse2.h286 __m128i in0 = _mm_load_si128((const __m128i *)(input + 0 * stride));
287 __m128i in1 = _mm_load_si128((const __m128i *)(input + 1 * stride));
288 __m128i in2 = _mm_load_si128((const __m128i *)(input + 2 * stride));
289 __m128i in3 = _mm_load_si128((const __m128i *)(input + 3 * stride));
290 __m128i in4 = _mm_load_si128((const __m128i *)(input + 4 * stride));
291 __m128i in5 = _mm_load_si128((const __m128i *)(input + 5 * stride));
292 __m128i in6 = _mm_load_si128((const __m128i *)(input + 6 * stride));
293 __m128i in7 = _mm_load_si128((const __m128i *)(input + 7 * stride));
633 in00 = _mm_load_si128((const __m128i *)(in + 0 * stride));
634 in01 = _mm_load_si128((cons
[all...]
H A Dhighbd_quantize_intrin_sse2.c56 coeffs = _mm_load_si128((const __m128i *)(coeff_ptr + i * 4));
74 coeffs = _mm_load_si128((const __m128i *)(coeff_ptr + i * 4));
144 coeffs = _mm_load_si128((const __m128i *)(coeff_ptr + i * 4));
H A Dloopfilter_sse2.c28 const __m128i blimit = _mm_load_si128((const __m128i *)_blimit);
29 const __m128i limit = _mm_load_si128((const __m128i *)_limit);
30 const __m128i thresh = _mm_load_si128((const __m128i *)_thresh);
393 const __m128i blimit = _mm_load_si128((const __m128i *)_blimit);
394 const __m128i limit = _mm_load_si128((const __m128i *)_limit);
395 const __m128i thresh = _mm_load_si128((const __m128i *)_thresh);
741 const __m128i blimit = _mm_load_si128((const __m128i *)_blimit);
742 const __m128i limit = _mm_load_si128((const __m128i *)_limit);
743 const __m128i thresh = _mm_load_si128((const __m128i *)_thresh);
961 _mm_unpacklo_epi64(_mm_load_si128((cons
[all...]
H A Dvpx_subpixel_8t_intrin_ssse3.c85 shuffle1 =_mm_load_si128((__m128i const *)filt1_4_h8);
86 shuffle2 = _mm_load_si128((__m128i const *)filt2_4_h8);
158 filt1Reg = _mm_load_si128((__m128i const *)filt1_global);
159 filt2Reg = _mm_load_si128((__m128i const *)filt2_global);
160 filt3Reg = _mm_load_si128((__m128i const *)filt3_global);
161 filt4Reg = _mm_load_si128((__m128i const *)filt4_global);
238 filt1Reg = _mm_load_si128((__m128i const *)filt1_global);
239 filt2Reg = _mm_load_si128((__m128i const *)filt2_global);
240 filt3Reg = _mm_load_si128((__m128i const *)filt3_global);
241 filt4Reg = _mm_load_si128((__m128
[all...]
/external/libhevc/common/x86/
H A Dihevc_itrans_recon_32x32_ssse3_intr.c260 m_temp_reg_70 = _mm_load_si128((__m128i *)pi2_tmp_src);
262 m_temp_reg_71 = _mm_load_si128((__m128i *)pi2_tmp_src);
264 m_temp_reg_72 = _mm_load_si128((__m128i *)pi2_tmp_src);
266 m_temp_reg_73 = _mm_load_si128((__m128i *)pi2_tmp_src);
268 m_temp_reg_74 = _mm_load_si128((__m128i *)pi2_tmp_src);
270 m_temp_reg_75 = _mm_load_si128((__m128i *)pi2_tmp_src);
272 m_temp_reg_76 = _mm_load_si128((__m128i *)pi2_tmp_src);
274 m_temp_reg_77 = _mm_load_si128((__m128i *)pi2_tmp_src);
277 m_temp_reg_80 = _mm_load_si128((__m128i *)pi2_tmp_src);
279 m_temp_reg_81 = _mm_load_si128((__m128
[all...]
H A Dihevc_itrans_recon_16x16_ssse3_intr.c210 m_temp_reg_70 = _mm_load_si128((__m128i *)pi2_tmp_src);
212 m_temp_reg_71 = _mm_load_si128((__m128i *)pi2_tmp_src);
214 m_temp_reg_72 = _mm_load_si128((__m128i *)pi2_tmp_src);
216 m_temp_reg_73 = _mm_load_si128((__m128i *)pi2_tmp_src);
218 m_temp_reg_74 = _mm_load_si128((__m128i *)pi2_tmp_src);
220 m_temp_reg_75 = _mm_load_si128((__m128i *)pi2_tmp_src);
222 m_temp_reg_76 = _mm_load_si128((__m128i *)pi2_tmp_src);
224 m_temp_reg_77 = _mm_load_si128((__m128i *)pi2_tmp_src);
249 m_coeff1 = _mm_load_si128((__m128i *)&g_ai2_ihevc_trans_16_even[2][0]); //89 75
295 m_coeff3 = _mm_load_si128((__m128
[all...]
H A Dihevc_itrans_recon_ssse3_intr.c1035 m_temp_reg_70 = _mm_load_si128((__m128i *)pi2_src);
1037 m_temp_reg_71 = _mm_load_si128((__m128i *)pi2_src);
1039 m_temp_reg_72 = _mm_load_si128((__m128i *)pi2_src);
1041 m_temp_reg_73 = _mm_load_si128((__m128i *)pi2_src);
1044 m_temp_reg_74 = _mm_load_si128((__m128i *)pi2_src);
1046 m_temp_reg_75 = _mm_load_si128((__m128i *)pi2_src);
1048 m_temp_reg_76 = _mm_load_si128((__m128i *)pi2_src);
1050 m_temp_reg_77 = _mm_load_si128((__m128i *)pi2_src);
1061 m_coeff2 = _mm_load_si128((__m128i *)&g_ai2_ihevc_trans_intr_even_8[3][0]);
1062 m_coeff1 = _mm_load_si128((__m128
[all...]
H A Dihevc_deblk_ssse3_intr.c147 coef_8x16b = _mm_load_si128((__m128i *)(coef_d));
148 mask_16x8b = _mm_load_si128((__m128i *)(shuffle_d));
409 coefdelta_0_8x16b = _mm_load_si128((__m128i *)coef_de1);
431 coefdelta_0_8x16b = _mm_load_si128((__m128i *)coef_dep1);
474 tmp3_const_8x16b = _mm_load_si128((__m128i *)(shuffle1));
508 tmp0_const_8x16b = _mm_load_si128((__m128i *)shuffle2);
509 tmp1_const_8x16b = _mm_load_si128((__m128i *)shuffle3);
606 coef_8x16b = _mm_load_si128((__m128i *)(coef_d));
607 mask_16x8b = _mm_load_si128((__m128i *)(shuffle_d));
843 coefdelta_0_8x16b = _mm_load_si128((__m128
[all...]
/external/libvpx/libvpx/vp9/common/x86/
H A Dvp9_idct_intrin_sse2.c87 in[0] = _mm_load_si128((const __m128i *)input);
88 in[1] = _mm_load_si128((const __m128i *)(input + 8 * 1));
89 in[2] = _mm_load_si128((const __m128i *)(input + 8 * 2));
90 in[3] = _mm_load_si128((const __m128i *)(input + 8 * 3));
91 in[4] = _mm_load_si128((const __m128i *)(input + 8 * 4));
92 in[5] = _mm_load_si128((const __m128i *)(input + 8 * 5));
93 in[6] = _mm_load_si128((const __m128i *)(input + 8 * 6));
94 in[7] = _mm_load_si128((const __m128i *)(input + 8 * 7));
/external/libvpx/libvpx/vp9/encoder/x86/
H A Dvp9_quantize_sse2.c47 round = _mm_load_si128((const __m128i*)round_ptr);
48 quant = _mm_load_si128((const __m128i*)quant_ptr);
49 dequant = _mm_load_si128((const __m128i*)dequant_ptr);
57 coeff0 = _mm_load_si128((const __m128i*)(coeff_ptr + n_coeffs));
58 coeff1 = _mm_load_si128((const __m128i*)(coeff_ptr + n_coeffs) + 1);
102 iscan0 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs));
103 iscan1 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs) + 1);
124 coeff0 = _mm_load_si128((const __m128i*)(coeff_ptr + n_coeffs));
125 coeff1 = _mm_load_si128((const __m128i*)(coeff_ptr + n_coeffs) + 1);
177 iscan0 = _mm_load_si128((cons
[all...]
H A Dvp9_highbd_block_error_intrin_sse2.c28 __m128i mm_coeff = _mm_load_si128((__m128i*) (coeff + i));
29 __m128i mm_coeff2 = _mm_load_si128((__m128i*) (coeff + i + 4));
30 __m128i mm_dqcoeff = _mm_load_si128((__m128i*) (dqcoeff + i));
31 __m128i mm_dqcoeff2 = _mm_load_si128((__m128i*) (dqcoeff + i + 4));
H A Dvp9_dct_ssse3.c50 __m128i in0 = _mm_load_si128((const __m128i *)(input + 0 * stride));
51 __m128i in1 = _mm_load_si128((const __m128i *)(input + 1 * stride));
52 __m128i in2 = _mm_load_si128((const __m128i *)(input + 2 * stride));
53 __m128i in3 = _mm_load_si128((const __m128i *)(input + 3 * stride));
54 __m128i in4 = _mm_load_si128((const __m128i *)(input + 4 * stride));
55 __m128i in5 = _mm_load_si128((const __m128i *)(input + 5 * stride));
56 __m128i in6 = _mm_load_si128((const __m128i *)(input + 6 * stride));
57 __m128i in7 = _mm_load_si128((const __m128i *)(input + 7 * stride));
306 round = _mm_load_si128((const __m128i*)round_ptr);
307 quant = _mm_load_si128((cons
[all...]
H A Dvp9_avg_intrin_sse2.c218 src[0] = _mm_load_si128((const __m128i *)src_diff);
219 src[1] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
220 src[2] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
221 src[3] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
222 src[4] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
223 src[5] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
224 src[6] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
225 src[7] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
257 __m128i coeff0 = _mm_load_si128((const __m128i *)coeff);
258 __m128i coeff1 = _mm_load_si128((cons
[all...]
/external/jemalloc/test/include/test/
H A DSFMT-sse2.h67 x = _mm_load_si128(a);
89 r1 = _mm_load_si128(&ctx->sfmt[N - 2].si);
90 r2 = _mm_load_si128(&ctx->sfmt[N - 1].si);
119 r1 = _mm_load_si128(&ctx->sfmt[N - 2].si);
120 r2 = _mm_load_si128(&ctx->sfmt[N - 1].si);
144 r = _mm_load_si128(&array[j + size - N].si);
/external/jpeg/
H A Djidctintelsse.c171 x3 = _mm_load_si128(( __m128i*)(wsptr+24));\
172 x1 = _mm_load_si128(( __m128i*)(wsptr+8));\
200 x0 = _mm_load_si128(( __m128i*)(wsptr));\
201 x4 = _mm_load_si128(( __m128i*)(wsptr+32));\
202 x2 = _mm_load_si128(( __m128i*)(wsptr+16));\
203 x6 = _mm_load_si128(( __m128i*)(wsptr+48));\
281 row0 = _mm_load_si128((__m128i const*)(coef_blockSSE));
282 row2 = _mm_load_si128((__m128i const*)(coef_blockSSE+8*2));
292 row0 = _mm_load_si128((__m128i const*)(coef_blockSSE+8*4));
293 row2 = _mm_load_si128((__m128
[all...]
/external/neven/Embedded/common/src/b_BasicEm/
H A DMathSSE2.c165 m_XMM0 = _mm_load_si128( (__m128i *)&0[vec1L] );
168 m_XMM2 = _mm_load_si128( (__m128i *)&0[vec2L] );
170 m_XMM6 = _mm_load_si128( (__m128i *)&8[vec1L] );
176 m_XMM3 = _mm_load_si128( (__m128i *)&8[vec2L] );
188 m_XMM0 = _mm_load_si128( (__m128i *)&m_XMM5 );
/external/pdfium/core/src/fxcodec/fx_libopenjpeg/libopenjpeg20/
H A Dmct.c87 __m128i r = _mm_load_si128((const __m128i *)&(c0[i]));
88 __m128i g = _mm_load_si128((const __m128i *)&(c1[i]));
89 __m128i b = _mm_load_si128((const __m128i *)&(c2[i]));
152 __m128i y = _mm_load_si128((const __m128i *)&(c0[i]));
153 __m128i u = _mm_load_si128((const __m128i *)&(c1[i]));
154 __m128i v = _mm_load_si128((const __m128i *)&(c2[i]));
231 __m128i r = _mm_load_si128((const __m128i *)&(c0[i]));
232 __m128i g = _mm_load_si128((const __m128i *)&(c1[i]));
233 __m128i b = _mm_load_si128((const __m128i *)&(c2[i]));
/external/mesa3d/src/gallium/drivers/llvmpipe/
H A Dlp_rast_tri.c264 __m128i p0 = _mm_load_si128((__m128i *)&plane[0]); /* c, dcdx, dcdy, eo */
265 __m128i p1 = _mm_load_si128((__m128i *)&plane[1]); /* c, dcdx, dcdy, eo */
266 __m128i p2 = _mm_load_si128((__m128i *)&plane[2]); /* c, dcdx, dcdy, eo */
374 __m128i p0 = _mm_load_si128((__m128i *)&plane[0]); /* c, dcdx, dcdy, eo */
375 __m128i p1 = _mm_load_si128((__m128i *)&plane[1]); /* c, dcdx, dcdy, eo */
376 __m128i p2 = _mm_load_si128((__m128i *)&plane[2]); /* c, dcdx, dcdy, eo */

Completed in 527 milliseconds

12