Home
last modified time | relevance | path

Searched refs:_mm_shuffle_ps (Results 1 – 25 of 30) sorted by relevance

12

/external/vulkan-validation-layers/libs/glm/detail/
Dintrinsic_matrix.inl71 __m128 v0 = _mm_shuffle_ps(v, v, _MM_SHUFFLE(0, 0, 0, 0));
72 __m128 v1 = _mm_shuffle_ps(v, v, _MM_SHUFFLE(1, 1, 1, 1));
73 __m128 v2 = _mm_shuffle_ps(v, v, _MM_SHUFFLE(2, 2, 2, 2));
74 __m128 v3 = _mm_shuffle_ps(v, v, _MM_SHUFFLE(3, 3, 3, 3));
118 __m128 e0 = _mm_shuffle_ps(in2[0], in2[0], _MM_SHUFFLE(0, 0, 0, 0));
119 __m128 e1 = _mm_shuffle_ps(in2[0], in2[0], _MM_SHUFFLE(1, 1, 1, 1));
120 __m128 e2 = _mm_shuffle_ps(in2[0], in2[0], _MM_SHUFFLE(2, 2, 2, 2));
121 __m128 e3 = _mm_shuffle_ps(in2[0], in2[0], _MM_SHUFFLE(3, 3, 3, 3));
136 __m128 e0 = _mm_shuffle_ps(in2[1], in2[1], _MM_SHUFFLE(0, 0, 0, 0));
137 __m128 e1 = _mm_shuffle_ps(in2[1], in2[1], _MM_SHUFFLE(1, 1, 1, 1));
[all …]
Dintrinsic_geometric.inl55 __m128 swp0 = _mm_shuffle_ps(mul0, mul0, _MM_SHUFFLE(2, 3, 0, 1));
57 __m128 swp1 = _mm_shuffle_ps(add0, add0, _MM_SHUFFLE(0, 1, 2, 3));
69 __m128 swp1 = _mm_shuffle_ps(add0, add0, 1);
77 __m128 swp0 = _mm_shuffle_ps(v1, v1, _MM_SHUFFLE(3, 0, 2, 1));
78 __m128 swp1 = _mm_shuffle_ps(v1, v1, _MM_SHUFFLE(3, 1, 0, 2));
79 __m128 swp2 = _mm_shuffle_ps(v2, v2, _MM_SHUFFLE(3, 0, 2, 1));
80 __m128 swp3 = _mm_shuffle_ps(v2, v2, _MM_SHUFFLE(3, 1, 0, 2));
Dtype_mat4x4.inl658 __m128 v0 = _mm_shuffle_ps(v.data, v.data, _MM_SHUFFLE(0, 0, 0, 0));
659 __m128 v1 = _mm_shuffle_ps(v.data, v.data, _MM_SHUFFLE(1, 1, 1, 1));
660 __m128 v2 = _mm_shuffle_ps(v.data, v.data, _MM_SHUFFLE(2, 2, 2, 2));
661 __m128 v3 = _mm_shuffle_ps(v.data, v.data, _MM_SHUFFLE(3, 3, 3, 3));
/external/speex/libspeex/
Dfilters_sse.h62 yy = _mm_shuffle_ps(yy, yy, 0); in filter_mem16_10()
66 mem[0] = _mm_shuffle_ps(mem[0], mem[0], 0x39); in filter_mem16_10()
72 mem[1] = _mm_shuffle_ps(mem[1], mem[1], 0x39); in filter_mem16_10()
77 mem[2] = _mm_shuffle_ps(mem[2], mem[2], 0xfd); in filter_mem16_10()
86 mem[2] = _mm_shuffle_ps(mem[2], mem[2], 0x55); in filter_mem16_10()
112 yy = _mm_shuffle_ps(yy, yy, 0); in filter_mem16_8()
116 mem[0] = _mm_shuffle_ps(mem[0], mem[0], 0x39); in filter_mem16_8()
122 mem[1] = _mm_shuffle_ps(mem[1], mem[1], 0x39); in filter_mem16_8()
167 yy = _mm_shuffle_ps(yy, yy, 0); in iir_mem16_10()
171 mem[0] = _mm_shuffle_ps(mem[0], mem[0], 0x39); in iir_mem16_10()
[all …]
Dltp_sse.h51 sum = _mm_add_ss(sum, _mm_shuffle_ps(sum, sum, 0x55)); in inner_prod()
88 sum = _mm_add_ss(sum, _mm_shuffle_ps(sum, sum, 0x55)); in pitch_xcorr()
Dresample_sse.h51 sum = _mm_add_ss(sum, _mm_shuffle_ps(sum, sum, 0x55)); in inner_product_single()
69 sum = _mm_add_ss(sum, _mm_shuffle_ps(sum, sum, 0x55)); in interpolate_product_single()
/external/flac/libFLAC/
Dlpc_intrin_sse.c72 d = d0; d = _mm_shuffle_ps(d, d, 0); in FLAC__lpc_compute_autocorrelation_intrin_sse_lag_4_new()
82 d = _mm_load_ss(data+i); d = _mm_shuffle_ps(d, d, 0); in FLAC__lpc_compute_autocorrelation_intrin_sse_lag_4_new()
83 d0 = _mm_shuffle_ps(d0, d0, _MM_SHUFFLE(2,1,0,3)); in FLAC__lpc_compute_autocorrelation_intrin_sse_lag_4_new()
110 d = d0; d = _mm_shuffle_ps(d, d, 0); in FLAC__lpc_compute_autocorrelation_intrin_sse_lag_8_new()
122 d = _mm_load_ss(data+i); d = _mm_shuffle_ps(d, d, 0); in FLAC__lpc_compute_autocorrelation_intrin_sse_lag_8_new()
123 d1 = _mm_shuffle_ps(d1, d1, _MM_SHUFFLE(2,1,0,3)); in FLAC__lpc_compute_autocorrelation_intrin_sse_lag_8_new()
124 d0 = _mm_shuffle_ps(d0, d0, _MM_SHUFFLE(2,1,0,3)); in FLAC__lpc_compute_autocorrelation_intrin_sse_lag_8_new()
156 d = d0; d = _mm_shuffle_ps(d, d, 0); in FLAC__lpc_compute_autocorrelation_intrin_sse_lag_12_new()
170 d = _mm_load_ss(data+i); d = _mm_shuffle_ps(d, d, 0); in FLAC__lpc_compute_autocorrelation_intrin_sse_lag_12_new()
171 d2 = _mm_shuffle_ps(d2, d2, _MM_SHUFFLE(2,1,0,3)); in FLAC__lpc_compute_autocorrelation_intrin_sse_lag_12_new()
[all …]
/external/webrtc/webrtc/modules/audio_processing/aec/
Daec_rdft_sse2.c27 __m128 a01v = _mm_shuffle_ps(a00v, a08v, _MM_SHUFFLE(1, 0, 1, 0)); in cft1st_128_SSE2()
28 __m128 a23v = _mm_shuffle_ps(a00v, a08v, _MM_SHUFFLE(3, 2, 3, 2)); in cft1st_128_SSE2()
29 __m128 a45v = _mm_shuffle_ps(a04v, a12v, _MM_SHUFFLE(1, 0, 1, 0)); in cft1st_128_SSE2()
30 __m128 a67v = _mm_shuffle_ps(a04v, a12v, _MM_SHUFFLE(3, 2, 3, 2)); in cft1st_128_SSE2()
45 x0w = _mm_shuffle_ps(x0v, x0v, _MM_SHUFFLE(2, 3, 0, 1)); in cft1st_128_SSE2()
53 const __m128 x3w = _mm_shuffle_ps(x3v, x3v, _MM_SHUFFLE(2, 3, 0, 1)); in cft1st_128_SSE2()
56 x0w = _mm_shuffle_ps(x0v, x0v, _MM_SHUFFLE(2, 3, 0, 1)); in cft1st_128_SSE2()
62 x0w = _mm_shuffle_ps(x0v, x0v, _MM_SHUFFLE(2, 3, 0, 1)); in cft1st_128_SSE2()
70 a00v = _mm_shuffle_ps(a01v, a23v, _MM_SHUFFLE(1, 0, 1, 0)); in cft1st_128_SSE2()
71 a04v = _mm_shuffle_ps(a45v, a67v, _MM_SHUFFLE(1, 0, 1, 0)); in cft1st_128_SSE2()
[all …]
Daec_core_sse2.c220 _mm_shuffle_ps(fft0, fft4, _MM_SHUFFLE(2, 0, 2, 0)); in FilterAdaptationSSE2()
222 _mm_shuffle_ps(fft0, fft4, _MM_SHUFFLE(3, 1, 3, 1)); in FilterAdaptationSSE2()
437 sum = _mm_add_ps(sum, _mm_shuffle_ps(sum, sum, _MM_SHUFFLE(0, 0, 3, 2))); in _mm_add_ps_4x1()
439 sum = _mm_add_ps(sum, _mm_shuffle_ps(sum, sum, _MM_SHUFFLE(1, 1, 1, 1))); in _mm_add_ps_4x1()
533 __m128 vec_a = _mm_shuffle_ps(vec_3210, vec_7654, in SmoothedPSD()
535 __m128 vec_b = _mm_shuffle_ps(vec_3210, vec_7654, in SmoothedPSD()
554 __m128 vec_a = _mm_shuffle_ps(vec_3210, vec_7654, in SmoothedPSD()
556 __m128 vec_b = _mm_shuffle_ps(vec_3210, vec_7654, in SmoothedPSD()
632 _mm_shuffle_ps(vec_sqrtHanning_rev, vec_sqrtHanning_rev, in WindowDataSSE2()
647 const __m128 vec_a = _mm_shuffle_ps(vec_fft0, vec_fft4, in StoreAsComplexSSE2()
[all …]
/external/eigen/Eigen/src/LU/arch/
DInverse_SSE.h84 AB = _mm_mul_ps(_mm_shuffle_ps(A,A,0x0F), B);
85 AB = _mm_sub_ps(AB,_mm_mul_ps(_mm_shuffle_ps(A,A,0xA5), _mm_shuffle_ps(B,B,0x4E)));
87 DC = _mm_mul_ps(_mm_shuffle_ps(D,D,0x0F), C);
88 DC = _mm_sub_ps(DC,_mm_mul_ps(_mm_shuffle_ps(D,D,0xA5), _mm_shuffle_ps(C,C,0x4E)));
91 dA = _mm_mul_ps(_mm_shuffle_ps(A, A, 0x5F),A);
94 dB = _mm_mul_ps(_mm_shuffle_ps(B, B, 0x5F),B);
98 dC = _mm_mul_ps(_mm_shuffle_ps(C, C, 0x5F),C);
101 dD = _mm_mul_ps(_mm_shuffle_ps(D, D, 0x5F),D);
105 d = _mm_mul_ps(_mm_shuffle_ps(DC,DC,0xD8),AB);
108 iD = _mm_mul_ps(_mm_shuffle_ps(C,C,0xA0), _mm_movelh_ps(AB,AB));
[all …]
/external/libopus/celt/x86/
Dpitch_sse.c56 xsum1 = _mm_add_ps(xsum1,_mm_mul_ps(_mm_shuffle_ps(x0,x0,0x00),yj)); in xcorr_kernel_sse()
57 xsum2 = _mm_add_ps(xsum2,_mm_mul_ps(_mm_shuffle_ps(x0,x0,0x55), in xcorr_kernel_sse()
58 _mm_shuffle_ps(yj,y3,0x49))); in xcorr_kernel_sse()
59 xsum1 = _mm_add_ps(xsum1,_mm_mul_ps(_mm_shuffle_ps(x0,x0,0xaa), in xcorr_kernel_sse()
60 _mm_shuffle_ps(yj,y3,0x9e))); in xcorr_kernel_sse()
61 xsum2 = _mm_add_ps(xsum2,_mm_mul_ps(_mm_shuffle_ps(x0,x0,0xff),y3)); in xcorr_kernel_sse()
96 xsum1 = _mm_add_ss(xsum1, _mm_shuffle_ps(xsum1, xsum1, 0x55)); in dual_inner_prod_sse()
99 xsum2 = _mm_add_ss(xsum2, _mm_shuffle_ps(xsum2, xsum2, 0x55)); in dual_inner_prod_sse()
124 sum = _mm_add_ss(sum, _mm_shuffle_ps(sum, sum, 0x55)); in celt_inner_prod_sse()
155 x2v = _mm_shuffle_ps(x0v, x4v, 0x4e); in comb_filter_const_sse()
[all …]
Dvq_sse2.c82 sums = _mm_add_ps(sums, _mm_shuffle_ps(sums, sums, _MM_SHUFFLE(1, 0, 3, 2))); in op_pvq_search_sse2()
83 sums = _mm_add_ps(sums, _mm_shuffle_ps(sums, sums, _MM_SHUFFLE(2, 3, 0, 1))); in op_pvq_search_sse2()
129 xy4 = _mm_add_ps(xy4, _mm_shuffle_ps(xy4, xy4, _MM_SHUFFLE(1, 0, 3, 2))); in op_pvq_search_sse2()
130 xy4 = _mm_add_ps(xy4, _mm_shuffle_ps(xy4, xy4, _MM_SHUFFLE(2, 3, 0, 1))); in op_pvq_search_sse2()
132 yy4 = _mm_add_ps(yy4, _mm_shuffle_ps(yy4, yy4, _MM_SHUFFLE(1, 0, 3, 2))); in op_pvq_search_sse2()
133 yy4 = _mm_add_ps(yy4, _mm_shuffle_ps(yy4, yy4, _MM_SHUFFLE(2, 3, 0, 1))); in op_pvq_search_sse2()
183 max2 = _mm_max_ps(max, _mm_shuffle_ps(max, max, _MM_SHUFFLE(1, 0, 3, 2))); in op_pvq_search_sse2()
184 max2 = _mm_max_ps(max2, _mm_shuffle_ps(max2, max2, _MM_SHUFFLE(2, 3, 0, 1))); in op_pvq_search_sse2()
/external/vulkan-validation-layers/libs/glm/gtx/
Dsimd_quat.inl123 __m128 mul0 = _mm_mul_ps(q1.Data, _mm_shuffle_ps(q2.Data, q2.Data, _MM_SHUFFLE(0, 1, 2, 3)));
124 __m128 mul1 = _mm_mul_ps(q1.Data, _mm_shuffle_ps(q2.Data, q2.Data, _MM_SHUFFLE(1, 0, 3, 2)));
125 __m128 mul2 = _mm_mul_ps(q1.Data, _mm_shuffle_ps(q2.Data, q2.Data, _MM_SHUFFLE(2, 3, 0, 1)));
136 add0 = _mm_add_ss(add0, _mm_shuffle_ps(add0, add0, 1));
140 add1 = _mm_add_ss(add1, _mm_shuffle_ps(add1, add1, 1));
144 add2 = _mm_add_ss(add2, _mm_shuffle_ps(add2, add2, 1));
148 add3 = _mm_add_ss(add3, _mm_shuffle_ps(add3, add3, 1));
155 //__m128 xxyy = _mm_shuffle_ps(add0, add1, _MM_SHUFFLE(0, 0, 0, 0));
156 //__m128 zzww = _mm_shuffle_ps(add2, add3, _MM_SHUFFLE(0, 0, 0, 0));
158 //return _mm_shuffle_ps(xxyy, zzww, _MM_SHUFFLE(2, 0, 2, 0));
[all …]
Dsimd_mat4.inl552 __m128 Shu0 = _mm_shuffle_ps(r.Data, r.Data, _MM_SHUFFLE(0, 0, 0, 0));
553 __m128 Shu1 = _mm_shuffle_ps(r.Data, r.Data, _MM_SHUFFLE(1, 1, 1, 1));
554 __m128 Shu2 = _mm_shuffle_ps(r.Data, r.Data, _MM_SHUFFLE(2, 2, 2, 2));
555 __m128 Shu3 = _mm_shuffle_ps(r.Data, r.Data, _MM_SHUFFLE(3, 3, 3, 3));
/external/tensorflow/tensorflow/core/kernels/
Dsparse_matmul_op.h234 return _mm_set1_ps(_mm_cvtss_f32(_mm_shuffle_ps(a, a, 1)));
240 return _mm_set1_ps(_mm_cvtss_f32(_mm_shuffle_ps(a, a, 2)));
246 return _mm_set1_ps(_mm_cvtss_f32(_mm_shuffle_ps(a, a, 3)));
262 return _mm512_broadcastss_ps(_mm_shuffle_ps(a, a, _MM_SHUFFLE(1, 1, 1, 1)));
268 return _mm512_broadcastss_ps(_mm_shuffle_ps(a, a, _MM_SHUFFLE(2, 2, 2, 2)));
274 return _mm512_broadcastss_ps(_mm_shuffle_ps(a, a, _MM_SHUFFLE(3, 3, 3, 3)));
/external/eigen/Eigen/src/Core/arch/SSE/
DPacketMath.h74 (_mm_shuffle_ps( (a), (b), ((s)<<6|(r)<<4|(q)<<2|(p))))
77 …(_mm_castps_si128( (_mm_shuffle_ps( _mm_castsi128_ps(a), _mm_castsi128_ps(b), ((s)<<6|(r)<<4|(q)<<…
382 to[stride*1] = _mm_cvtss_f32(_mm_shuffle_ps(from, from, 1));
383 to[stride*2] = _mm_cvtss_f32(_mm_shuffle_ps(from, from, 2));
384 to[stride*3] = _mm_cvtss_f32(_mm_shuffle_ps(from, from, 3));
436 { return _mm_shuffle_ps(a,a,0x1B); }
543 return pfirst<Packet4f>(_mm_add_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1)));
595 return pfirst<Packet4f>(_mm_mul_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1)));
615 return pfirst<Packet4f>(_mm_min_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1)));
641 return pfirst<Packet4f>(_mm_max_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1)));
[all …]
DTypeCasting.h55 return _mm_shuffle_ps(_mm_cvtpd_ps(a), _mm_cvtpd_ps(b), (1 << 2) | (1 << 6));
DComplex.h125 to[stride*0] = std::complex<float>(_mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 0)),
126 _mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 1)));
127 to[stride*1] = std::complex<float>(_mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 2)),
128 _mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 3)));
/external/python/cpython3/Modules/_blake2/impl/
Dblake2s-load-sse41.h20 buf = TOI(_mm_shuffle_ps(TOF(m0), TOF(m1), _MM_SHUFFLE(2,0,2,0)));
23 buf = TOI(_mm_shuffle_ps(TOF(m0), TOF(m1), _MM_SHUFFLE(3,1,3,1)));
26 buf = TOI(_mm_shuffle_ps(TOF(m2), TOF(m3), _MM_SHUFFLE(2,0,2,0)));
29 buf = TOI(_mm_shuffle_ps(TOF(m2), TOF(m3), _MM_SHUFFLE(3,1,3,1)));
/external/eigen/Eigen/src/Core/arch/AVX/
DComplex.h106 to[stride*0] = std::complex<float>(_mm_cvtss_f32(_mm_shuffle_ps(low, low, 0)),
107 _mm_cvtss_f32(_mm_shuffle_ps(low, low, 1)));
108 to[stride*1] = std::complex<float>(_mm_cvtss_f32(_mm_shuffle_ps(low, low, 2)),
109 _mm_cvtss_f32(_mm_shuffle_ps(low, low, 3)));
112 to[stride*2] = std::complex<float>(_mm_cvtss_f32(_mm_shuffle_ps(high, high, 0)),
113 _mm_cvtss_f32(_mm_shuffle_ps(high, high, 1)));
114 to[stride*3] = std::complex<float>(_mm_cvtss_f32(_mm_shuffle_ps(high, high, 2)),
115 _mm_cvtss_f32(_mm_shuffle_ps(high, high, 3)));
DPacketMath.h274 to[stride*1] = _mm_cvtss_f32(_mm_shuffle_ps(low, low, 1));
275 to[stride*2] = _mm_cvtss_f32(_mm_shuffle_ps(low, low, 2));
276 to[stride*3] = _mm_cvtss_f32(_mm_shuffle_ps(low, low, 3));
280 to[stride*5] = _mm_cvtss_f32(_mm_shuffle_ps(high, high, 1));
281 to[stride*6] = _mm_cvtss_f32(_mm_shuffle_ps(high, high, 2));
282 to[stride*7] = _mm_cvtss_f32(_mm_shuffle_ps(high, high, 3));
/external/webrtc/webrtc/common_audio/resampler/
Dsinc_resampler_sse.cc53 _mm_store_ss(&result, _mm_add_ss(m_sums2, _mm_shuffle_ps( in Convolve_SSE()
/external/webrtc/webrtc/common_audio/
Dfir_filter_sse.cc73 _mm_store_ss(out + i, _mm_add_ss(m_sum, _mm_shuffle_ps(m_sum, m_sum, 1))); in Filter()
/external/mesa3d/src/gallium/drivers/swr/rasterizer/core/
Drasterizer.cpp1615 __m128 vXa = _mm_shuffle_ps(vX, vX, _MM_SHUFFLE(1, 1, 0, 0)); in RasterizeLine()
1616 __m128 vYa = _mm_shuffle_ps(vY, vY, _MM_SHUFFLE(1, 1, 0, 0)); in RasterizeLine()
1617 __m128 vZa = _mm_shuffle_ps(vZ, vZ, _MM_SHUFFLE(1, 1, 0, 0)); in RasterizeLine()
1618 __m128 vRecipWa = _mm_shuffle_ps(vRecipW, vRecipW, _MM_SHUFFLE(1, 1, 0, 0)); in RasterizeLine()
1701 vXa = _mm_shuffle_ps(vX, vX, _MM_SHUFFLE(1, 0, 1, 1)); in RasterizeLine()
1702 vYa = _mm_shuffle_ps(vY, vY, _MM_SHUFFLE(1, 0, 1, 1)); in RasterizeLine()
1703 vZa = _mm_shuffle_ps(vZ, vZ, _MM_SHUFFLE(1, 0, 1, 1)); in RasterizeLine()
1704 vRecipWa = _mm_shuffle_ps(vRecipW, vRecipW, _MM_SHUFFLE(1, 0, 1, 1)); in RasterizeLine()
Dfrontend.h45 __m128 vYsub = _mm_shuffle_ps(vY, vY, _MM_SHUFFLE(3, 0, 2, 1)); in triangleSetupAB()
55 __m128 vXsub = _mm_shuffle_ps(vX, vX, _MM_SHUFFLE(3, 0, 2, 1)); in triangleSetupAB()

12