| /external/python/cpython3/Modules/_blake2/impl/ |
| blake2b-load-sse41.h | 22 b0 = _mm_unpacklo_epi64(m0, m1); \
23 b1 = _mm_unpacklo_epi64(m2, m3); \
38 b0 = _mm_unpacklo_epi64(m4, m5); \
39 b1 = _mm_unpacklo_epi64(m6, m7); \
54 b0 = _mm_unpacklo_epi64(m7, m2); \
62 b0 = _mm_unpacklo_epi64(m5, m4); \
78 b0 = _mm_unpacklo_epi64(m6, m1); \
94 b0 = _mm_unpacklo_epi64(m4, m0); \
110 b0 = _mm_unpacklo_epi64(m7, m3); \
127 b1 = _mm_unpacklo_epi64(m6, m7);
[all...] |
| blake2b-round.h | 122 row4l = _mm_unpackhi_epi64(row4h, _mm_unpacklo_epi64(t0, t0)); \
123 row4h = _mm_unpackhi_epi64(t0, _mm_unpacklo_epi64(row4h, row4h)); \
124 row2l = _mm_unpackhi_epi64(row2l, _mm_unpacklo_epi64(row2h, row2h)); \
125 row2h = _mm_unpackhi_epi64(row2h, _mm_unpacklo_epi64(t1, t1))
133 row2l = _mm_unpackhi_epi64(row2h, _mm_unpacklo_epi64(row2l, row2l)); \
134 row2h = _mm_unpackhi_epi64(t0, _mm_unpacklo_epi64(row2h, row2h)); \
135 row4l = _mm_unpackhi_epi64(row4l, _mm_unpacklo_epi64(row4h, row4h)); \
136 row4h = _mm_unpackhi_epi64(row4h, _mm_unpacklo_epi64(t1, t1))
|
| blake2s-load-sse41.h | 99 buf = _mm_unpacklo_epi64(t1,t0);
102 t0 = _mm_unpacklo_epi64(m1,m2); \
109 t1 = _mm_unpacklo_epi64(m0,m1); \
126 buf = _mm_unpacklo_epi64(t0,t1);
140 t0 = _mm_unpacklo_epi64(m1,m2); \
156 t0 = _mm_unpacklo_epi64(m0,m2); \
178 t1 = _mm_unpacklo_epi64(m1,m2); \
185 buf = _mm_unpacklo_epi64(t0,t1);
189 t1 = _mm_unpacklo_epi64(t0,m0); \
|
| /external/libvpx/libvpx/vpx_dsp/x86/ |
| transpose_sse2.h | 85 out[0] = _mm_unpacklo_epi64(c0, c0);
87 out[2] = _mm_unpacklo_epi64(c1, c1);
89 out[4] = _mm_unpacklo_epi64(c2, c2);
91 out[6] = _mm_unpacklo_epi64(c3, c3);
151 out[0] = _mm_unpacklo_epi64(b0, b1);
153 out[2] = _mm_unpacklo_epi64(b2, b3);
213 out[0] = _mm_unpacklo_epi64(b0, b1);
215 out[2] = _mm_unpacklo_epi64(b4, b5);
217 out[4] = _mm_unpacklo_epi64(b2, b3);
219 out[6] = _mm_unpacklo_epi64(b6, b7)
[all...] |
| inv_txfm_ssse3.h | 43 tmp[0] = _mm_unpacklo_epi64(io[0], io[0]);
45 tmp[2] = _mm_unpacklo_epi64(io[1], io[1]);
63 step1[3] = _mm_unpacklo_epi64(tmp[1], tmp[0]); // step1 3&0
|
| fwd_txfm_sse2.h | 33 return _mm_unpacklo_epi64(buf0, buf1);
159 __m128i top_dwords_01 = _mm_unpacklo_epi64(reg0_top_dwords, reg1_top_dwords);
160 __m128i top_dwords_23 = _mm_unpacklo_epi64(reg2_top_dwords, reg3_top_dwords);
330 const __m128i tr2_0 = _mm_unpacklo_epi64(tr1_0, tr1_4);
332 const __m128i tr2_2 = _mm_unpacklo_epi64(tr1_2, tr1_6);
334 const __m128i tr2_4 = _mm_unpacklo_epi64(tr1_1, tr1_5);
336 const __m128i tr2_6 = _mm_unpacklo_epi64(tr1_3, tr1_7);
|
| highbd_intrapred_intrin_sse2.c | 52 _mm_store_si128((__m128i *)dst, _mm_unpacklo_epi64(row0, row0));
54 _mm_store_si128((__m128i *)dst, _mm_unpacklo_epi64(row1, row1));
56 _mm_store_si128((__m128i *)dst, _mm_unpacklo_epi64(row2, row2));
58 _mm_store_si128((__m128i *)dst, _mm_unpacklo_epi64(row3, row3));
71 const __m128i val = _mm_unpacklo_epi64(*row, *row);
115 const __m128i val = _mm_unpacklo_epi64(*row, *row);
227 const __m128i dc_dup = _mm_unpacklo_epi64(dc_dup_lo, dc_dup_lo);
278 const __m128i dc_dup = _mm_unpacklo_epi64(dc_dup_lo, dc_dup_lo);
333 const __m128i dc_dup = _mm_unpacklo_epi64(dc_dup_lo, dc_dup_lo);
492 const __m128i IJKLLLLL = _mm_unpacklo_epi64(IJKL0000, LLLL0000)
[all...] |
| loopfilter_sse2.c | 53 mask = _mm_unpacklo_epi64(mask, flat); \
77 filter = _mm_unpacklo_epi64(filter, filter); \
96 filter2filter1 = _mm_unpacklo_epi64(filter2filter1, filter); \
111 _mm_unpacklo_epi64(_mm_loadl_epi64((const __m128i *)_blimit),
119 p3p2 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 3 * p)),
121 q1p1 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 2 * p)),
123 q0p0 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 1 * p)),
125 q3q2 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s + 2 * p)),
127 p1p0 = _mm_unpacklo_epi64(q0p0, q1p1);
128 p2p1 = _mm_unpacklo_epi64(q1p1, p3p2)
[all...] |
| /external/libavc/common/x86/ |
| ih264_ihadamard_scaling_ssse3.c | 131 src_r0 = _mm_unpacklo_epi64(temp0, temp2); //a0 b0 c0 d0
133 src_r2 = _mm_unpacklo_epi64(temp1, temp3); //a2 b2 c2 d2
160 src_r0 = _mm_unpacklo_epi64(temp0, temp2); //a0 a1 a2 a3
162 src_r2 = _mm_unpacklo_epi64(temp1, temp3); //c0 c1 c2 c3
|
| ih264_ihadamard_scaling_sse42.c | 129 src_r0 = _mm_unpacklo_epi64(temp0, temp2); //a0 b0 c0 d0
131 src_r2 = _mm_unpacklo_epi64(temp1, temp3); //a2 b2 c2 d2
158 src_r0 = _mm_unpacklo_epi64(temp0, temp2); //a0 a1 a2 a3
160 src_r2 = _mm_unpacklo_epi64(temp1, temp3); //c0 c1 c2 c3
228 plane_0 = _mm_unpacklo_epi64(temp0, temp1); //a0+a1+a2+a3 a0+a1-a2-a3 a0-a1+a2-a3 a0-a1-a2+a3
|
| ih264_resi_trans_quant_sse42.c | 170 src_r0 = _mm_unpacklo_epi64(temp1, zero_8x16b); //a0 b0 c0 d0
172 src_r2 = _mm_unpacklo_epi64(temp3, zero_8x16b); //a2 b2 c2 d2
208 src_r0 = _mm_unpacklo_epi64(temp1, zero_8x16b); //a0 a1 a2 a3
210 src_r2 = _mm_unpacklo_epi64(temp3, zero_8x16b); //c0 c1 c2 c3
237 src_r0 = _mm_unpacklo_epi64(src_r0, src_r1); //a0 a1 a2 a3 b0 b1 b2 b3
238 src_r2 = _mm_unpacklo_epi64(src_r2, src_r3); //c0 c1 c2 c3 d0 d1 d2 d3
453 src_r0 = _mm_unpacklo_epi64(temp1, zero_8x16b); //a0 b0 c0 d0
455 src_r2 = _mm_unpacklo_epi64(temp3, zero_8x16b); //a2 b2 c2 d2
491 src_r0 = _mm_unpacklo_epi64(temp1, zero_8x16b); //a0 a1 a2 a3
493 src_r2 = _mm_unpacklo_epi64(temp3, zero_8x16b); //c0 c1 c2 c
[all...] |
| ih264_iquant_itrans_recon_sse42.c | 187 resq_r0 = _mm_unpacklo_epi64(temp1, temp3); //a0 b0 c0 d0
189 resq_r2 = _mm_unpacklo_epi64(temp2, temp4); //a2 b2 c2 d2
223 resq_r0 = _mm_unpacklo_epi64(temp1, temp3); //a0 a1 a2 a3
225 resq_r2 = _mm_unpacklo_epi64(temp2, temp4); //c0 c1 c2 c3
437 resq_r0 = _mm_unpacklo_epi64(temp1, temp3); //a0 b0 c0 d0
439 resq_r2 = _mm_unpacklo_epi64(temp2, temp4); //a2 b2 c2 d2
473 resq_r0 = _mm_unpacklo_epi64(temp1, temp3); //a0 a1 a2 a3
475 resq_r2 = _mm_unpacklo_epi64(temp2, temp4); //c0 c1 c2 c3
|
| ih264_iquant_itrans_recon_ssse3.c | 194 resq_r0 = _mm_unpacklo_epi64(temp1, temp3); //a0 b0 c0 d0
196 resq_r2 = _mm_unpacklo_epi64(temp2, temp4); //a2 b2 c2 d2
230 resq_r0 = _mm_unpacklo_epi64(temp1, temp3); //a0 a1 a2 a3
232 resq_r2 = _mm_unpacklo_epi64(temp2, temp4); //c0 c1 c2 c3
[all...] |
| ih264_iquant_itrans_recon_dc_ssse3.c | 145 pred_r0 = _mm_unpacklo_epi64(pred_r0, pred_r1); //p00 p01 p02 p03 p10 p11 p12 p13
146 pred_r2 = _mm_unpacklo_epi64(pred_r2, pred_r3); //p20 p21 p22p p23 p30 p31 p32 p33
425 pred_r0 = _mm_unpacklo_epi64(pred_r0, pred_r1); //p00 p01 p02 p03 p10 p11 p12 p13
426 pred_r2 = _mm_unpacklo_epi64(pred_r2, pred_r3); //p20 p21 p22p p23 p30 p31 p32 p33
|
| ih264_inter_pred_filters_ssse3.c | 261 src_r0r1_16x8b = _mm_unpacklo_epi64(src_r0_16x8b, src_r1_16x8b); //a0 a1 a1 a2 a2 a3 a3 a4 b0 b1 b1 b2 b2 b3 b3 b4
268 src_r0r1_16x8b = _mm_unpacklo_epi64(src_r0_16x8b, src_r1_16x8b); //a2 a3 a3 a4 a4 a5 a5 a6 b2 b3 b3 b4 b4 b5 b5 b6
275 src_r0r1_16x8b = _mm_unpacklo_epi64(src_r0_16x8b, src_r1_16x8b); //a4 a5 a5 a6 a6 a7 a7 a8 b4 b5 b5 b6 b6 b7 b7 b8
[all...] |
| /external/libhevc/decoder/x86/ |
| ihevcd_fmt_conv_ssse3_intr.c | 153 src_u_8x16b = _mm_unpacklo_epi64(temp0_8x16b, temp1_8x16b);
166 src_u_8x16b = _mm_unpacklo_epi64(temp0_8x16b, temp1_8x16b);
179 src_u_8x16b = _mm_unpacklo_epi64(temp0_8x16b, temp1_8x16b);
192 src_u_8x16b = _mm_unpacklo_epi64(temp0_8x16b, temp1_8x16b);
228 src_u_8x16b = _mm_unpacklo_epi64(temp0_8x16b, temp1_8x16b);
|
| /external/mesa3d/src/gallium/auxiliary/util/ |
| u_sse.h | 263 *o = _mm_unpacklo_epi64(t0, t1);
265 *q = _mm_unpacklo_epi64(t2, t3);
289 *o = _mm_unpacklo_epi64(t0, t1);
291 *q = _mm_unpacklo_epi64(t2, t3);
|
| /external/webp/src/dsp/ |
| common_sse2.h | 93 *out0 = _mm_unpacklo_epi64(transpose1_0, transpose1_1);
95 *out2 = _mm_unpacklo_epi64(transpose1_2, transpose1_3);
183 *in1 = _mm_unpacklo_epi64(C1, C3);
185 *in3 = _mm_unpacklo_epi64(C0, C2);
|
| /external/libhevc/common/x86/ |
| ihevc_32x32_itrans_recon_sse42_intr.c | [all...] |
| ihevc_itrans_recon_32x32_ssse3_intr.c | [all...] |
| ihevc_deblk_ssse3_intr.c | 150 src_tmp_8x16b = _mm_unpacklo_epi64(src_row0_8x16b, src_row3_8x16b);
260 src_row0_8x16b = _mm_unpacklo_epi64(src_row0_8x16b, src_row2_8x16b);
261 src_row1_8x16b = _mm_unpacklo_epi64(src_row1_8x16b, src_row3_8x16b);
365 src_row0_8x16b = _mm_unpacklo_epi64(temp_pq_str1_16x8b, temp_str0_16x8b);
367 src_row2_8x16b = _mm_unpacklo_epi64(temp_str2_16x8b, temp_pq_str0_16x8b);
398 src_row0_8x16b = _mm_unpacklo_epi64(src_row0_8x16b, src_row1_8x16b);
399 src_row2_8x16b = _mm_unpacklo_epi64(src_row2_8x16b, src_row3_8x16b);
407 tmp_delta2_8x16b = _mm_unpacklo_epi64(tmp_delta2_8x16b, tmp_delta3_8x16b);
437 mask_pq_8x16b = _mm_unpacklo_epi64(mask_pq_8x16b, mask_pq_8x16b);
720 temp_pq0_str0_16x8b = _mm_unpacklo_epi64(temp_pq0_str0_16x8b, temp_str0_16x8b)
[all...] |
| ihevc_itrans_recon_sse42_intr.c | 220 m_rdng_factor = _mm_unpacklo_epi64(m_rdng_factor, m_rdng_factor);
311 m_rdng_factor = _mm_unpacklo_epi64(m_rdng_factor, m_rdng_factor);
360 m_temp_reg_0 = _mm_unpacklo_epi64(m_temp_reg_0, m_temp_reg_1);
361 m_temp_reg_1 = _mm_unpacklo_epi64(m_temp_reg_2, m_temp_reg_3);
530 m_rdng_factor = _mm_unpacklo_epi64(m_rdng_factor, m_rdng_factor);
624 m_rdng_factor = _mm_unpacklo_epi64(m_rdng_factor, m_rdng_factor);
685 m_temp_reg_0 = _mm_unpacklo_epi64(m_temp_reg_0, m_temp_reg_1);
688 m_temp_reg_1 = _mm_unpacklo_epi64(m_temp_reg_2, m_temp_reg_3);
[all...] |
| ihevc_itrans_recon_ssse3_intr.c | 269 m_rdng_factor = _mm_unpacklo_epi64(m_rdng_factor, m_rdng_factor);
407 m_rdng_factor = _mm_unpacklo_epi64(m_rdng_factor, m_rdng_factor);
462 m_temp_reg_0 = _mm_unpacklo_epi64(m_temp_reg_0, m_temp_reg_1);
463 m_temp_reg_1 = _mm_unpacklo_epi64(m_temp_reg_2, m_temp_reg_3);
663 m_rdng_factor = _mm_unpacklo_epi64(m_rdng_factor, m_rdng_factor);
792 m_rdng_factor = _mm_unpacklo_epi64(m_rdng_factor, m_rdng_factor);
862 m_temp_reg_0 = _mm_unpacklo_epi64(m_temp_reg_0, m_temp_reg_1);
863 m_temp_reg_1 = _mm_unpacklo_epi64(m_temp_reg_2, m_temp_reg_3);
[all...] |
| ihevc_weighted_pred_ssse3_intr.c | 276 src_temp0_8x16b = _mm_unpacklo_epi64(src_temp0_8x16b, src_temp2_8x16b);
277 src_temp1_8x16b = _mm_unpacklo_epi64(src_temp1_8x16b, src_temp3_8x16b);
621 src_temp0_8x16b = _mm_unpacklo_epi64(src_temp0_8x16b, src_temp1_8x16b);
[all...] |
| /external/libmpeg2/common/x86/ |
| impeg2_idct_recon_sse42_intr.c | 410 m_temp_reg_50 = _mm_unpacklo_epi64(m_temp_reg_0, m_temp_reg_4);
412 m_temp_reg_52 = _mm_unpacklo_epi64(m_temp_reg_1, m_temp_reg_5);
627 m_temp_reg_50 = _mm_unpacklo_epi64(m_temp_reg_0, m_temp_reg_4);
629 m_temp_reg_52 = _mm_unpacklo_epi64(m_temp_reg_1, m_temp_reg_5);
[all...] |
