Home
last modified time | relevance | path

Searched refs:_mm_unpacklo_epi64 (Results 1 – 25 of 189) sorted by relevance

12345678

/external/python/cpython3/Modules/_blake2/impl/
Dblake2b-load-sse41.h20 b0 = _mm_unpacklo_epi64(m0, m1); \
21 b1 = _mm_unpacklo_epi64(m2, m3); \
36 b0 = _mm_unpacklo_epi64(m4, m5); \
37 b1 = _mm_unpacklo_epi64(m6, m7); \
52 b0 = _mm_unpacklo_epi64(m7, m2); \
60 b0 = _mm_unpacklo_epi64(m5, m4); \
76 b0 = _mm_unpacklo_epi64(m6, m1); \
92 b0 = _mm_unpacklo_epi64(m4, m0); \
108 b0 = _mm_unpacklo_epi64(m7, m3); \
125 b1 = _mm_unpacklo_epi64(m6, m7); \
[all …]
Dblake2b-round.h123 row4l = _mm_unpackhi_epi64(row4h, _mm_unpacklo_epi64(t0, t0)); \
124 row4h = _mm_unpackhi_epi64(t0, _mm_unpacklo_epi64(row4h, row4h)); \
125 row2l = _mm_unpackhi_epi64(row2l, _mm_unpacklo_epi64(row2h, row2h)); \
126 row2h = _mm_unpackhi_epi64(row2h, _mm_unpacklo_epi64(t1, t1))
134 row2l = _mm_unpackhi_epi64(row2h, _mm_unpacklo_epi64(row2l, row2l)); \
135 row2h = _mm_unpackhi_epi64(t0, _mm_unpacklo_epi64(row2h, row2h)); \
136 row4l = _mm_unpackhi_epi64(row4l, _mm_unpacklo_epi64(row4h, row4h)); \
137 row4h = _mm_unpackhi_epi64(row4h, _mm_unpacklo_epi64(t1, t1))
Dblake2s-load-sse41.h97 buf = _mm_unpacklo_epi64(t1,t0);
100 t0 = _mm_unpacklo_epi64(m1,m2); \
107 t1 = _mm_unpacklo_epi64(m0,m1); \
124 buf = _mm_unpacklo_epi64(t0,t1);
138 t0 = _mm_unpacklo_epi64(m1,m2); \
154 t0 = _mm_unpacklo_epi64(m0,m2); \
176 t1 = _mm_unpacklo_epi64(m1,m2); \
183 buf = _mm_unpacklo_epi64(t0,t1);
187 t1 = _mm_unpacklo_epi64(t0,m0); \
/external/libaom/libaom/aom_dsp/x86/
Dtranspose_sse2.h86 out[0] = _mm_unpacklo_epi64(c0, c0); in transpose_8bit_8x8()
88 out[2] = _mm_unpacklo_epi64(c1, c1); in transpose_8bit_8x8()
90 out[4] = _mm_unpacklo_epi64(c2, c2); in transpose_8bit_8x8()
92 out[6] = _mm_unpacklo_epi64(c3, c3); in transpose_8bit_8x8()
156 out[0] = _mm_unpacklo_epi64(b0, b1); in transpose_16bit_4x8()
158 out[2] = _mm_unpacklo_epi64(b2, b3); in transpose_16bit_4x8()
200 out[0] = _mm_unpacklo_epi64(b0, zeros); in transpose_16bit_8x4()
202 out[2] = _mm_unpacklo_epi64(b4, zeros); in transpose_16bit_8x4()
204 out[4] = _mm_unpacklo_epi64(b2, zeros); in transpose_16bit_8x4()
206 out[6] = _mm_unpacklo_epi64(b6, zeros); in transpose_16bit_8x4()
[all …]
Dlpf_common_sse2.h39 *d0 = _mm_unpacklo_epi64(ww0, w2); // 00 10 20 30 40 50 41 51 in highbd_transpose6x6_sse2()
44 *d2 = _mm_unpacklo_epi64(ww0, in highbd_transpose6x6_sse2()
54 *d4 = _mm_unpacklo_epi64(ww0, w5); // 04 14 24 34 44 54 45 55 in highbd_transpose6x6_sse2()
72 *d0 = _mm_unpacklo_epi64(ww0, zero); // 00 10 20 30 xx xx xx xx in highbd_transpose4x8_8x4_low_sse2()
74 *d2 = _mm_unpacklo_epi64(ww1, zero); // 02 12 22 32 xx xx xx xx in highbd_transpose4x8_8x4_low_sse2()
91 *d4 = _mm_unpacklo_epi64(ww2, zero); // 04 14 24 34 xx xx xx xx in highbd_transpose4x8_8x4_high_sse2()
93 *d6 = _mm_unpacklo_epi64(ww3, zero); // 06 16 26 36 xx xx xx xx in highbd_transpose4x8_8x4_high_sse2()
147 *d0 = _mm_unpacklo_epi64(ww0, ww1); // 00 10 20 30 40 50 60 70 in highbd_transpose8x8_low_sse2()
153 *d2 = _mm_unpacklo_epi64(ww0, ww1); // 02 12 22 32 42 52 62 72 in highbd_transpose8x8_low_sse2()
180 *d4 = _mm_unpacklo_epi64(ww0, ww1); // 04 14 24 34 44 54 64 74 in highbd_transpose8x8_high_sse2()
[all …]
Dloopfilter_sse2.c135 *pq0 = _mm_unpacklo_epi64(d0, d1); // pq in transpose_pq_14_inv_sse2()
137 *pq2 = _mm_unpacklo_epi64(d2, d3); // pq in transpose_pq_14_inv_sse2()
215 filter = _mm_unpacklo_epi64(filter, filter); in filter4_dual_sse2()
234 filter2filter1 = _mm_unpacklo_epi64(filter2filter1, filter); in filter4_dual_sse2()
295 q1p1 = _mm_unpacklo_epi64(*p1, *q1); in lpf_internal_4_dual_sse2()
296 q0p0 = _mm_unpacklo_epi64(*p0, *q0); in lpf_internal_4_dual_sse2()
298 p1p0 = _mm_unpacklo_epi64(q0p0, q1p1); in lpf_internal_4_dual_sse2()
321 mask = _mm_unpacklo_epi64(mask, flat); in lpf_internal_4_dual_sse2()
405 p1p0 = _mm_unpacklo_epi64(*q0p0, *q1p1); in lpf_internal_14_dual_sse2()
417 abs_p0q0 = _mm_unpacklo_epi64(abs_p0q0, zero); in lpf_internal_14_dual_sse2()
[all …]
/external/libvpx/libvpx/vpx_dsp/x86/
Dtranspose_sse2.h85 out[0] = _mm_unpacklo_epi64(c0, c0); in transpose_8bit_8x8()
87 out[2] = _mm_unpacklo_epi64(c1, c1); in transpose_8bit_8x8()
89 out[4] = _mm_unpacklo_epi64(c2, c2); in transpose_8bit_8x8()
91 out[6] = _mm_unpacklo_epi64(c3, c3); in transpose_8bit_8x8()
151 out[0] = _mm_unpacklo_epi64(b0, b1); in transpose_16bit_4x8()
153 out[2] = _mm_unpacklo_epi64(b2, b3); in transpose_16bit_4x8()
213 out[0] = _mm_unpacklo_epi64(b0, b1); in transpose_16bit_8x8()
215 out[2] = _mm_unpacklo_epi64(b4, b5); in transpose_16bit_8x8()
217 out[4] = _mm_unpacklo_epi64(b2, b3); in transpose_16bit_8x8()
219 out[6] = _mm_unpacklo_epi64(b6, b7); in transpose_16bit_8x8()
[all …]
/external/libgav1/libgav1/src/dsp/x86/
Dtranspose_sse4.h183 out[0] = _mm_unpacklo_epi64(b0, b1); in Transpose4x8To8x4_U16()
185 out[2] = _mm_unpacklo_epi64(b2, b3); in Transpose4x8To8x4_U16()
227 out[0] = _mm_unpacklo_epi64(b0, zeros); in Transpose8x4To4x8_U16()
229 out[2] = _mm_unpacklo_epi64(b4, zeros); in Transpose8x4To4x8_U16()
231 out[4] = _mm_unpacklo_epi64(b2, zeros); in Transpose8x4To4x8_U16()
233 out[6] = _mm_unpacklo_epi64(b6, zeros); in Transpose8x4To4x8_U16()
293 out[0] = _mm_unpacklo_epi64(b0, b1); in Transpose8x8_U16()
295 out[2] = _mm_unpacklo_epi64(b4, b5); in Transpose8x8_U16()
297 out[4] = _mm_unpacklo_epi64(b2, b3); in Transpose8x8_U16()
299 out[6] = _mm_unpacklo_epi64(b6, b7); in Transpose8x8_U16()
Dconvolve_sse4.inc121 v_tap[0] = _mm_unpacklo_epi64(v_tap[0], v_tap[0]);
122 v_tap[1] = _mm_unpacklo_epi64(v_tap[1], v_tap[1]);
123 v_tap[2] = _mm_unpacklo_epi64(v_tap[2], v_tap[2]);
124 v_tap[3] = _mm_unpacklo_epi64(v_tap[3], v_tap[3]);
136 v_tap[0] = _mm_unpacklo_epi64(v_tap[0], v_tap[0]);
137 v_tap[1] = _mm_unpacklo_epi64(v_tap[1], v_tap[1]);
138 v_tap[2] = _mm_unpacklo_epi64(v_tap[2], v_tap[2]);
147 v_tap[0] = _mm_unpacklo_epi64(v_tap[0], v_tap[0]);
148 v_tap[1] = _mm_unpacklo_epi64(v_tap[1], v_tap[1]);
156 v_tap[0] = _mm_unpacklo_epi64(v_tap[0], v_tap[0]);
[all …]
Dloop_filter_sse4.cc110 const __m128i qp1qp0 = _mm_unpacklo_epi64(qp0, qp1); in Filter4()
132 const __m128i a3a3a1a2 = _mm_unpacklo_epi64(a1a2, a3a3); in Filter4()
1003 *x0 = _mm_unpacklo_epi64(d0, d1); in DualTranspose4x8To8x4()
1007 *x2 = _mm_unpacklo_epi64(d2, d3); in DualTranspose4x8To8x4()
1035 const __m128i qp1qp0 = _mm_unpacklo_epi64(qp0, qp1); in Vertical14()
1256 __m128i oqps1 = _mm_unpacklo_epi64(ops1, oqs1); in Filter4()
1257 __m128i oqps0 = _mm_unpacklo_epi64(ops0, oqs0); in Filter4()
1285 const __m128i p1p0 = _mm_unpacklo_epi64(qp0, qp1); in Horizontal4()
1327 const __m128i qp0 = _mm_unpacklo_epi64(p1p0, q1q0); in Vertical4()
1358 const __m128i p1p0 = _mm_unpacklo_epi64(qp0, qp1); in CheckOuterThreshF6()
[all …]
/external/libavc/common/x86/
Dih264_resi_trans_quant_sse42.c170 src_r0 = _mm_unpacklo_epi64(temp1, zero_8x16b); //a0 b0 c0 d0 in ih264_resi_trans_quant_4x4_sse42()
172 src_r2 = _mm_unpacklo_epi64(temp3, zero_8x16b); //a2 b2 c2 d2 in ih264_resi_trans_quant_4x4_sse42()
208 src_r0 = _mm_unpacklo_epi64(temp1, zero_8x16b); //a0 a1 a2 a3 in ih264_resi_trans_quant_4x4_sse42()
210 src_r2 = _mm_unpacklo_epi64(temp3, zero_8x16b); //c0 c1 c2 c3 in ih264_resi_trans_quant_4x4_sse42()
237 src_r0 = _mm_unpacklo_epi64(src_r0, src_r1); //a0 a1 a2 a3 b0 b1 b2 b3 in ih264_resi_trans_quant_4x4_sse42()
238 src_r2 = _mm_unpacklo_epi64(src_r2, src_r3); //c0 c1 c2 c3 d0 d1 d2 d3 in ih264_resi_trans_quant_4x4_sse42()
453 src_r0 = _mm_unpacklo_epi64(temp1, zero_8x16b); //a0 b0 c0 d0 in ih264_resi_trans_quant_chroma_4x4_sse42()
455 src_r2 = _mm_unpacklo_epi64(temp3, zero_8x16b); //a2 b2 c2 d2 in ih264_resi_trans_quant_chroma_4x4_sse42()
491 src_r0 = _mm_unpacklo_epi64(temp1, zero_8x16b); //a0 a1 a2 a3 in ih264_resi_trans_quant_chroma_4x4_sse42()
493 src_r2 = _mm_unpacklo_epi64(temp3, zero_8x16b); //c0 c1 c2 c3 in ih264_resi_trans_quant_chroma_4x4_sse42()
[all …]
Dih264_ihadamard_scaling_sse42.c129 src_r0 = _mm_unpacklo_epi64(temp0, temp2); //a0 b0 c0 d0 in ih264_ihadamard_scaling_4x4_sse42()
131 src_r2 = _mm_unpacklo_epi64(temp1, temp3); //a2 b2 c2 d2 in ih264_ihadamard_scaling_4x4_sse42()
158 src_r0 = _mm_unpacklo_epi64(temp0, temp2); //a0 a1 a2 a3 in ih264_ihadamard_scaling_4x4_sse42()
160 src_r2 = _mm_unpacklo_epi64(temp1, temp3); //c0 c1 c2 c3 in ih264_ihadamard_scaling_4x4_sse42()
228 …plane_0 = _mm_unpacklo_epi64(temp0, temp1); //a0+a1+a2+a3 a0+a1-a2-a3 a0-a1+a2-a3 a0-a1-a2… in ih264_ihadamard_scaling_2x2_uv_sse42()
Dih264_ihadamard_scaling_ssse3.c131 src_r0 = _mm_unpacklo_epi64(temp0, temp2); //a0 b0 c0 d0 in ih264_ihadamard_scaling_4x4_ssse3()
133 src_r2 = _mm_unpacklo_epi64(temp1, temp3); //a2 b2 c2 d2 in ih264_ihadamard_scaling_4x4_ssse3()
160 src_r0 = _mm_unpacklo_epi64(temp0, temp2); //a0 a1 a2 a3 in ih264_ihadamard_scaling_4x4_ssse3()
162 src_r2 = _mm_unpacklo_epi64(temp1, temp3); //c0 c1 c2 c3 in ih264_ihadamard_scaling_4x4_ssse3()
/external/XNNPACK/src/f32-sigmoid/gen/
Dsse41-lut64-p2-div-x24.c82 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x24()
89 const __m128i vl4567 = _mm_unpacklo_epi64(vl45, vl67); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x24()
96 const __m128i vl89AB = _mm_unpacklo_epi64(vl89, vlAB); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x24()
103 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x24()
110 const __m128i vlGHIJ = _mm_unpacklo_epi64(vlGH, vlIJ); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x24()
117 const __m128i vlKLMN = _mm_unpacklo_epi64(vlKL, vlMN); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x24()
127 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x24()
136 const __m128i vl4567 = _mm_unpacklo_epi64(vl45, vl67); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x24()
145 const __m128i vl89AB = _mm_unpacklo_epi64(vl89, vlAB); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x24()
154 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x24()
[all …]
Dsse41-lut64-p2-div-x16.c72 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x16()
79 const __m128i vl4567 = _mm_unpacklo_epi64(vl45, vl67); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x16()
86 const __m128i vl89AB = _mm_unpacklo_epi64(vl89, vlAB); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x16()
93 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x16()
103 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x16()
112 const __m128i vl4567 = _mm_unpacklo_epi64(vl45, vl67); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x16()
121 const __m128i vl89AB = _mm_unpacklo_epi64(vl89, vlAB); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x16()
130 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x16()
212 const __m128i vl = _mm_unpacklo_epi64(vl_lo, vl_hi); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x16()
254 const __m128i vl = _mm_unpacklo_epi64(vl_lo, vl_hi); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x16()
Dsse41-lut64-p2-div-x20.c77 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x20()
84 const __m128i vl4567 = _mm_unpacklo_epi64(vl45, vl67); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x20()
91 const __m128i vl89AB = _mm_unpacklo_epi64(vl89, vlAB); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x20()
98 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x20()
105 const __m128i vlGHIJ = _mm_unpacklo_epi64(vlGH, vlIJ); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x20()
115 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x20()
124 const __m128i vl4567 = _mm_unpacklo_epi64(vl45, vl67); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x20()
133 const __m128i vl89AB = _mm_unpacklo_epi64(vl89, vlAB); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x20()
142 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x20()
151 const __m128i vlGHIJ = _mm_unpacklo_epi64(vlGH, vlIJ); in xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x20()
[all …]
Dsse2-lut64-p2-div-x24.c84 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x24()
93 const __m128i vl4567 = _mm_unpacklo_epi64(vl45, vl67); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x24()
102 const __m128i vl89AB = _mm_unpacklo_epi64(vl89, vlAB); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x24()
111 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x24()
120 const __m128i vlGHIJ = _mm_unpacklo_epi64(vlGH, vlIJ); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x24()
129 const __m128i vlKLMN = _mm_unpacklo_epi64(vlKL, vlMN); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x24()
141 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x24()
152 const __m128i vl4567 = _mm_unpacklo_epi64(vl45, vl67); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x24()
163 const __m128i vl89AB = _mm_unpacklo_epi64(vl89, vlAB); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x24()
174 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF); in xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x24()
[all …]
/external/libhevc/common/x86/
Dihevc_32x32_itrans_recon_sse42_intr.c1833 m_rdng_factor = _mm_unpacklo_epi64(m_rdng_factor, m_rdng_factor); in ihevc_itrans_recon_32x32_sse42()
1863 m_rdng_factor = _mm_unpacklo_epi64(m_rdng_factor, m_rdng_factor); in ihevc_itrans_recon_32x32_sse42()
1893 m_rdng_factor = _mm_unpacklo_epi64(m_rdng_factor, m_rdng_factor); in ihevc_itrans_recon_32x32_sse42()
1922 m_rdng_factor = _mm_unpacklo_epi64(m_rdng_factor, m_rdng_factor); in ihevc_itrans_recon_32x32_sse42()
1951 m_rdng_factor = _mm_unpacklo_epi64(m_rdng_factor, m_rdng_factor); in ihevc_itrans_recon_32x32_sse42()
1981 m_rdng_factor = _mm_unpacklo_epi64(m_rdng_factor, m_rdng_factor); in ihevc_itrans_recon_32x32_sse42()
2010 m_rdng_factor = _mm_unpacklo_epi64(m_rdng_factor, m_rdng_factor); in ihevc_itrans_recon_32x32_sse42()
2040 m_rdng_factor = _mm_unpacklo_epi64(m_rdng_factor, m_rdng_factor); in ihevc_itrans_recon_32x32_sse42()
2069 m_rdng_factor = _mm_unpacklo_epi64(m_rdng_factor, m_rdng_factor); in ihevc_itrans_recon_32x32_sse42()
2097 m_rdng_factor = _mm_unpacklo_epi64(m_rdng_factor, m_rdng_factor); in ihevc_itrans_recon_32x32_sse42()
[all …]
Dihevc_itrans_recon_32x32_ssse3_intr.c1844 m_rdng_factor = _mm_unpacklo_epi64(m_rdng_factor, m_rdng_factor); in ihevc_itrans_recon_32x32_ssse3()
1874 m_rdng_factor = _mm_unpacklo_epi64(m_rdng_factor, m_rdng_factor); in ihevc_itrans_recon_32x32_ssse3()
1904 m_rdng_factor = _mm_unpacklo_epi64(m_rdng_factor, m_rdng_factor); in ihevc_itrans_recon_32x32_ssse3()
1933 m_rdng_factor = _mm_unpacklo_epi64(m_rdng_factor, m_rdng_factor); in ihevc_itrans_recon_32x32_ssse3()
1962 m_rdng_factor = _mm_unpacklo_epi64(m_rdng_factor, m_rdng_factor); in ihevc_itrans_recon_32x32_ssse3()
1992 m_rdng_factor = _mm_unpacklo_epi64(m_rdng_factor, m_rdng_factor); in ihevc_itrans_recon_32x32_ssse3()
2021 m_rdng_factor = _mm_unpacklo_epi64(m_rdng_factor, m_rdng_factor); in ihevc_itrans_recon_32x32_ssse3()
2051 m_rdng_factor = _mm_unpacklo_epi64(m_rdng_factor, m_rdng_factor); in ihevc_itrans_recon_32x32_ssse3()
2080 m_rdng_factor = _mm_unpacklo_epi64(m_rdng_factor, m_rdng_factor); in ihevc_itrans_recon_32x32_ssse3()
2108 m_rdng_factor = _mm_unpacklo_epi64(m_rdng_factor, m_rdng_factor); in ihevc_itrans_recon_32x32_ssse3()
[all …]
Dihevc_deblk_ssse3_intr.c150 src_tmp_8x16b = _mm_unpacklo_epi64(src_row0_8x16b, src_row3_8x16b); in ihevc_deblk_luma_vert_ssse3()
260 src_row0_8x16b = _mm_unpacklo_epi64(src_row0_8x16b, src_row2_8x16b); in ihevc_deblk_luma_vert_ssse3()
261 src_row1_8x16b = _mm_unpacklo_epi64(src_row1_8x16b, src_row3_8x16b); in ihevc_deblk_luma_vert_ssse3()
365 src_row0_8x16b = _mm_unpacklo_epi64(temp_pq_str1_16x8b, temp_str0_16x8b); in ihevc_deblk_luma_vert_ssse3()
367 src_row2_8x16b = _mm_unpacklo_epi64(temp_str2_16x8b, temp_pq_str0_16x8b); in ihevc_deblk_luma_vert_ssse3()
398 src_row0_8x16b = _mm_unpacklo_epi64(src_row0_8x16b, src_row1_8x16b); in ihevc_deblk_luma_vert_ssse3()
399 src_row2_8x16b = _mm_unpacklo_epi64(src_row2_8x16b, src_row3_8x16b); in ihevc_deblk_luma_vert_ssse3()
407 tmp_delta2_8x16b = _mm_unpacklo_epi64(tmp_delta2_8x16b, tmp_delta3_8x16b); in ihevc_deblk_luma_vert_ssse3()
437 mask_pq_8x16b = _mm_unpacklo_epi64(mask_pq_8x16b, mask_pq_8x16b); in ihevc_deblk_luma_vert_ssse3()
720 temp_pq0_str0_16x8b = _mm_unpacklo_epi64(temp_pq0_str0_16x8b, temp_str0_16x8b); in ihevc_deblk_luma_horz_ssse3()
[all …]
/external/libhevc/decoder/x86/
Dihevcd_fmt_conv_ssse3_intr.c153 src_u_8x16b = _mm_unpacklo_epi64(temp0_8x16b, temp1_8x16b); in ihevcd_fmt_conv_420sp_to_420p_ssse3()
166 src_u_8x16b = _mm_unpacklo_epi64(temp0_8x16b, temp1_8x16b); in ihevcd_fmt_conv_420sp_to_420p_ssse3()
179 src_u_8x16b = _mm_unpacklo_epi64(temp0_8x16b, temp1_8x16b); in ihevcd_fmt_conv_420sp_to_420p_ssse3()
192 src_u_8x16b = _mm_unpacklo_epi64(temp0_8x16b, temp1_8x16b); in ihevcd_fmt_conv_420sp_to_420p_ssse3()
228 src_u_8x16b = _mm_unpacklo_epi64(temp0_8x16b, temp1_8x16b); in ihevcd_fmt_conv_420sp_to_420p_ssse3()
/external/libaom/libaom/aom_dsp/simd/
Dv64_intrinsics_x86.h95 ((c) ? _mm_srli_si128(_mm_unpacklo_epi64(b, a), (c)) : b)
172 __m128i t = _mm_unpacklo_epi64(b, a); in v64_pack_s32_s16()
178 __m128i t = _mm_unpacklo_epi64(b, a); in v64_pack_s32_u16()
193 __m128i t = _mm_unpacklo_epi64(b, a); in v64_pack_s16_u8()
198 __m128i t = _mm_unpacklo_epi64(b, a); in v64_pack_s16_s8()
204 return _mm_shuffle_epi8(_mm_unpacklo_epi64(b, a), in v64_unziphi_8()
208 _mm_unpacklo_epi64(_mm_srli_epi16(b, 8), _mm_srli_epi16(a, 8)), in v64_unziphi_8()
215 return _mm_shuffle_epi8(_mm_unpacklo_epi64(b, a), in v64_unziplo_8()
224 return _mm_shuffle_epi8(_mm_unpacklo_epi64(b, a), in v64_unziphi_16()
228 _mm_unpacklo_epi64(_mm_srai_epi32(b, 16), _mm_srai_epi32(a, 16)), in v64_unziphi_16()
[all …]
/external/XNNPACK/src/f32-velu/gen/
Dvelu-sse41-rr2-lut16-p3-x20.c83 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x20()
90 const __m128i vl4567 = _mm_unpacklo_epi64(vl45, vl67); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x20()
97 const __m128i vl89AB = _mm_unpacklo_epi64(vl89, vlAB); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x20()
104 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x20()
111 const __m128i vlGHIJ = _mm_unpacklo_epi64(vlGH, vlIJ); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x20()
121 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x20()
130 const __m128i vl4567 = _mm_unpacklo_epi64(vl45, vl67); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x20()
139 const __m128i vl89AB = _mm_unpacklo_epi64(vl89, vlAB); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x20()
148 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x20()
157 const __m128i vlGHIJ = _mm_unpacklo_epi64(vlGH, vlIJ); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x20()
[all …]
Dvelu-sse41-rr2-lut16-p3-x24.c88 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x24()
95 const __m128i vl4567 = _mm_unpacklo_epi64(vl45, vl67); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x24()
102 const __m128i vl89AB = _mm_unpacklo_epi64(vl89, vlAB); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x24()
109 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x24()
116 const __m128i vlGHIJ = _mm_unpacklo_epi64(vlGH, vlIJ); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x24()
123 const __m128i vlKLMN = _mm_unpacklo_epi64(vlKL, vlMN); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x24()
133 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x24()
142 const __m128i vl4567 = _mm_unpacklo_epi64(vl45, vl67); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x24()
151 const __m128i vl89AB = _mm_unpacklo_epi64(vl89, vlAB); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x24()
160 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF); in xnn_f32_velu_ukernel__sse41_rr2_lut16_p3_x24()
[all …]
/external/webp/src/dsp/
Dcommon_sse2.h93 *out0 = _mm_unpacklo_epi64(transpose1_0, transpose1_1); in VP8Transpose_2_4x4_16b()
95 *out2 = _mm_unpacklo_epi64(transpose1_2, transpose1_3); in VP8Transpose_2_4x4_16b()
183 *in1 = _mm_unpacklo_epi64(C1, C3); in VP8L32bToPlanar_SSE2()
185 *in3 = _mm_unpacklo_epi64(C0, C2); in VP8L32bToPlanar_SSE2()

12345678