Home
last modified time | relevance | path

Searched refs:_mm_mul_ps (Results 1 – 25 of 32) sorted by relevance

12

/external/webrtc/webrtc/modules/audio_processing/aec/
Daec_core_sse2.c57 const __m128 a = _mm_mul_ps(x_fft_buf_re, h_fft_buf_re); in FilterFarSSE2()
58 const __m128 b = _mm_mul_ps(x_fft_buf_im, h_fft_buf_im); in FilterFarSSE2()
59 const __m128 c = _mm_mul_ps(x_fft_buf_re, h_fft_buf_im); in FilterFarSSE2()
60 const __m128 d = _mm_mul_ps(x_fft_buf_im, h_fft_buf_re); in FilterFarSSE2()
104 const __m128 ef_re2 = _mm_mul_ps(ef_re, ef_re); in ScaleErrorSignalSSE2()
105 const __m128 ef_im2 = _mm_mul_ps(ef_im, ef_im); in ScaleErrorSignalSSE2()
111 __m128 ef_re_if = _mm_mul_ps(ef_re, absEfInv); in ScaleErrorSignalSSE2()
112 __m128 ef_im_if = _mm_mul_ps(ef_im, absEfInv); in ScaleErrorSignalSSE2()
119 ef_re = _mm_mul_ps(ef_re, kMu); in ScaleErrorSignalSSE2()
120 ef_im = _mm_mul_ps(ef_im, kMu); in ScaleErrorSignalSSE2()
[all …]
Daec_rdft_sse2.c47 const __m128 a45_0v = _mm_mul_ps(wk2rv, x0v); in cft1st_128_SSE2()
48 const __m128 a45_1v = _mm_mul_ps(wk2iv, x0w); in cft1st_128_SSE2()
54 const __m128 x3s = _mm_mul_ps(mm_swap_sign, x3w); in cft1st_128_SSE2()
57 a23_0v = _mm_mul_ps(wk1rv, x0v); in cft1st_128_SSE2()
58 a23_1v = _mm_mul_ps(wk1iv, x0w); in cft1st_128_SSE2()
65 const __m128 a67_0v = _mm_mul_ps(wk3rv, x0v); in cft1st_128_SSE2()
66 const __m128 a67_1v = _mm_mul_ps(wk3iv, x0w); in cft1st_128_SSE2()
119 const __m128 x3_swapped = _mm_mul_ps(mm_swap_sign, x3i0_3r0_3i1_x3r1); in cftmdl_128_SSE2()
127 const __m128 yy2 = _mm_mul_ps(mm_swap_sign, yy1); in cftmdl_128_SSE2()
129 const __m128 yy4 = _mm_mul_ps(wk1rv, yy3); in cftmdl_128_SSE2()
[all …]
/external/flac/libFLAC/
Dlpc_intrin_sse.c73 sum0 = _mm_add_ps(sum0, _mm_mul_ps(d0, d)); in FLAC__lpc_compute_autocorrelation_intrin_sse_lag_4_new()
85 sum0 = _mm_add_ps(sum0, _mm_mul_ps(d, d0)); in FLAC__lpc_compute_autocorrelation_intrin_sse_lag_4_new()
111 sum0 = _mm_add_ps(sum0, _mm_mul_ps(d0, d)); in FLAC__lpc_compute_autocorrelation_intrin_sse_lag_8_new()
112 sum1 = _mm_add_ps(sum1, _mm_mul_ps(d1, d)); in FLAC__lpc_compute_autocorrelation_intrin_sse_lag_8_new()
127 sum1 = _mm_add_ps(sum1, _mm_mul_ps(d, d1)); in FLAC__lpc_compute_autocorrelation_intrin_sse_lag_8_new()
128 sum0 = _mm_add_ps(sum0, _mm_mul_ps(d, d0)); in FLAC__lpc_compute_autocorrelation_intrin_sse_lag_8_new()
157 sum0 = _mm_add_ps(sum0, _mm_mul_ps(d0, d)); in FLAC__lpc_compute_autocorrelation_intrin_sse_lag_12_new()
158 sum1 = _mm_add_ps(sum1, _mm_mul_ps(d1, d)); in FLAC__lpc_compute_autocorrelation_intrin_sse_lag_12_new()
159 sum2 = _mm_add_ps(sum2, _mm_mul_ps(d2, d)); in FLAC__lpc_compute_autocorrelation_intrin_sse_lag_12_new()
177 sum2 = _mm_add_ps(sum2, _mm_mul_ps(d, d2)); in FLAC__lpc_compute_autocorrelation_intrin_sse_lag_12_new()
[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()
59 xsum1 = _mm_add_ps(xsum1,_mm_mul_ps(_mm_shuffle_ps(x0,x0,0xaa), in xcorr_kernel_sse()
61 xsum2 = _mm_add_ps(xsum2,_mm_mul_ps(_mm_shuffle_ps(x0,x0,0xff),y3)); in xcorr_kernel_sse()
65 xsum1 = _mm_add_ps(xsum1,_mm_mul_ps(_mm_load1_ps(x+j),_mm_loadu_ps(y+j))); in xcorr_kernel_sse()
68 xsum2 = _mm_add_ps(xsum2,_mm_mul_ps(_mm_load1_ps(x+j),_mm_loadu_ps(y+j))); in xcorr_kernel_sse()
71 xsum1 = _mm_add_ps(xsum1,_mm_mul_ps(_mm_load1_ps(x+j),_mm_loadu_ps(y+j))); in xcorr_kernel_sse()
91 xsum1 = _mm_add_ps(xsum1,_mm_mul_ps(xi, y1i)); in dual_inner_prod_sse()
92 xsum2 = _mm_add_ps(xsum2,_mm_mul_ps(xi, y2i)); in dual_inner_prod_sse()
120 sum = _mm_add_ps(sum,_mm_mul_ps(xi, yi)); in celt_inner_prod_sse()
[all …]
Dvq_sse2.c108 rcp4 = _mm_mul_ps(_mm_set_ps1((float)(K+.8)), _mm_rcp_ps(sums)); in op_pvq_search_sse2()
116 rx4 = _mm_mul_ps(x4, rcp4); in op_pvq_search_sse2()
121 xy4 = _mm_add_ps(xy4, _mm_mul_ps(x4, y4)); in op_pvq_search_sse2()
122 yy4 = _mm_add_ps(yy4, _mm_mul_ps(y4, y4)); in op_pvq_search_sse2()
174 r4 = _mm_mul_ps(x4, y4); in op_pvq_search_sse2()
/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/speex/libspeexdsp/
Dresample_sse.h47 sum = _mm_add_ps(sum, _mm_mul_ps(_mm_loadu_ps(a+i), _mm_loadu_ps(b+i))); in inner_product_single()
48 sum = _mm_add_ps(sum, _mm_mul_ps(_mm_loadu_ps(a+i+4), _mm_loadu_ps(b+i+4))); in inner_product_single()
64 sum = _mm_add_ps(sum, _mm_mul_ps(_mm_load1_ps(a+i), _mm_loadu_ps(b+i*oversample))); in interpolate_product_single()
65 sum = _mm_add_ps(sum, _mm_mul_ps(_mm_load1_ps(a+i+1), _mm_loadu_ps(b+(i+1)*oversample))); in interpolate_product_single()
67 sum = _mm_mul_ps(f, sum); in interpolate_product_single()
86 t = _mm_mul_ps(_mm_loadu_ps(a+i), _mm_loadu_ps(b+i)); in inner_product_double()
90 t = _mm_mul_ps(_mm_loadu_ps(a+i+4), _mm_loadu_ps(b+i+4)); in inner_product_double()
112 t = _mm_mul_ps(_mm_load1_ps(a+i), _mm_loadu_ps(b+i*oversample)); in interpolate_product_double()
116 t = _mm_mul_ps(_mm_load1_ps(a+i+1), _mm_loadu_ps(b+(i+1)*oversample)); in interpolate_product_double()
/external/webrtc/webrtc/common_audio/resampler/
Dsinc_resampler_sse.cc32 m_sums1 = _mm_add_ps(m_sums1, _mm_mul_ps(m_input, _mm_load_ps(k1 + i))); in Convolve_SSE()
33 m_sums2 = _mm_add_ps(m_sums2, _mm_mul_ps(m_input, _mm_load_ps(k2 + i))); in Convolve_SSE()
38 m_sums1 = _mm_add_ps(m_sums1, _mm_mul_ps(m_input, _mm_load_ps(k1 + i))); in Convolve_SSE()
39 m_sums2 = _mm_add_ps(m_sums2, _mm_mul_ps(m_input, _mm_load_ps(k2 + i))); in Convolve_SSE()
44 m_sums1 = _mm_mul_ps(m_sums1, _mm_set_ps1( in Convolve_SSE()
46 m_sums2 = _mm_mul_ps(m_sums2, _mm_set_ps1( in Convolve_SSE()
/external/libaom/libaom/av1/encoder/x86/
Dml_sse3.c31 const __m128 mul_h = _mm_mul_ps(inputs_h, weights_h); in nn_propagate_8to1()
32 const __m128 mul_l = _mm_mul_ps(inputs_l, weights_l); in nn_propagate_8to1()
51 const __m128 mul = _mm_mul_ps(inputs128, weights128); in nn_propagate_4to1()
69 const __m128 mul0 = _mm_mul_ps(weight0, inputs128); in nn_propagate_4to4()
71 const __m128 mul1 = _mm_mul_ps(weight1, inputs128); in nn_propagate_4to4()
92 const __m128 mul0 = _mm_mul_ps(inputs128, weight0); in nn_propagate_4to8()
93 const __m128 mul1 = _mm_mul_ps(inputs128, weight1); in nn_propagate_4to8()
121 const __m128 mul_h = _mm_mul_ps(inputs_h, weight_h); in nn_propagate_8to4()
122 const __m128 mul_l = _mm_mul_ps(inputs_l, weight_l); in nn_propagate_8to4()
233 total = _mm_add_ps(total, _mm_mul_ps(input, weight)); in av1_nn_predict_sse3()
/external/eigen/Eigen/src/Geometry/arch/
DGeometry_SSE.h32 __m128 s1 = _mm_mul_ps(vec4f_swizzle1(a,1,2,0,2),vec4f_swizzle1(b,2,0,1,2));
33 __m128 s2 = _mm_mul_ps(vec4f_swizzle1(a,3,3,3,1),vec4f_swizzle1(b,0,1,2,1));
36 _mm_add_ps(_mm_sub_ps(_mm_mul_ps(a,vec4f_swizzle1(b,3,3,3,3)),
37 _mm_mul_ps(vec4f_swizzle1(a,2,0,1,0),
72 __m128 mul1=_mm_mul_ps(vec4f_swizzle1(a,1,2,0,3),vec4f_swizzle1(b,2,0,1,3));
73 __m128 mul2=_mm_mul_ps(vec4f_swizzle1(a,2,0,1,3),vec4f_swizzle1(b,1,2,0,3));
/external/libaom/libaom/aom_dsp/x86/
Dfft_sse2.c111 _mm_set1_ps, _mm_add_ps, _mm_sub_ps, _mm_mul_ps);
113 _mm_set1_ps, _mm_add_ps, _mm_sub_ps, _mm_mul_ps);
115 _mm_set1_ps, _mm_add_ps, _mm_sub_ps, _mm_mul_ps);
141 _mm_set1_ps, _mm_add_ps, _mm_sub_ps, _mm_mul_ps);
143 _mm_set1_ps, _mm_add_ps, _mm_sub_ps, _mm_mul_ps);
145 _mm_set1_ps, _mm_add_ps, _mm_sub_ps, _mm_mul_ps);
/external/eigen/Eigen/src/Core/arch/SSE/
DComplex.h71 return Packet2cf(_mm_addsub_ps(_mm_mul_ps(_mm_moveldup_ps(a.v), b.v),
72 _mm_mul_ps(_mm_movehdup_ps(a.v),
79 return Packet2cf(_mm_add_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v),
80 _mm_xor_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3),
189 return Packet2cf(_mm_add_ps(_mm_xor_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v), mask),
190 _mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3),
207 return Packet2cf(_mm_add_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v),
208 _mm_xor_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3),
225 return Packet2cf(_mm_sub_ps(_mm_xor_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v), mask),
226 _mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3),
[all …]
DMathFunctions.h330 Packet4f z = _mm_mul_ps(x,x);
425 Packet4f tmp = _mm_mul_ps(z, p4f_half);
/external/pdfium/third_party/libopenjpeg20/
Dmct.c408 vr = _mm_add_ps(vy, _mm_mul_ps(vv, vrv)); in opj_mct_decode_real()
409 vg = _mm_sub_ps(_mm_sub_ps(vy, _mm_mul_ps(vu, vgu)), _mm_mul_ps(vv, vgv)); in opj_mct_decode_real()
410 vb = _mm_add_ps(vy, _mm_mul_ps(vu, vbu)); in opj_mct_decode_real()
421 vr = _mm_add_ps(vy, _mm_mul_ps(vv, vrv)); in opj_mct_decode_real()
422 vg = _mm_sub_ps(_mm_sub_ps(vy, _mm_mul_ps(vu, vgu)), _mm_mul_ps(vv, vgv)); in opj_mct_decode_real()
423 vb = _mm_add_ps(vy, _mm_mul_ps(vu, vbu)); in opj_mct_decode_real()
Ddwt.c2317 __m128 xmm0 = _mm_mul_ps(vw[0], c); in opj_v4dwt_decode_step1_sse()
2318 __m128 xmm2 = _mm_mul_ps(vw[2], c); in opj_v4dwt_decode_step1_sse()
2319 __m128 xmm4 = _mm_mul_ps(vw[4], c); in opj_v4dwt_decode_step1_sse()
2320 __m128 xmm6 = _mm_mul_ps(vw[6], c); in opj_v4dwt_decode_step1_sse()
2327 vw[0] = _mm_mul_ps(vw[0], c); in opj_v4dwt_decode_step1_sse()
2362 vw[-1] = _mm_add_ps(tmp2, _mm_mul_ps(_mm_add_ps(tmp1, tmp3), c)); in opj_v4dwt_decode_step2_sse()
2363 vw[ 1] = _mm_add_ps(tmp4, _mm_mul_ps(_mm_add_ps(tmp3, tmp5), c)); in opj_v4dwt_decode_step2_sse()
2364 vw[ 3] = _mm_add_ps(tmp6, _mm_mul_ps(_mm_add_ps(tmp5, tmp7), c)); in opj_v4dwt_decode_step2_sse()
2365 vw[ 5] = _mm_add_ps(tmp8, _mm_mul_ps(_mm_add_ps(tmp7, tmp9), c)); in opj_v4dwt_decode_step2_sse()
2373 vw[-1] = _mm_add_ps(tmp2, _mm_mul_ps(_mm_add_ps(tmp1, tmp3), c)); in opj_v4dwt_decode_step2_sse()
[all …]
Dt1.c1741 _mm_store_ps((float*)(datap + 0), _mm_mul_ps(xmm0_data, xmm_stepsize)); in opj_t1_clbl_decode_processor()
1742 _mm_store_ps((float*)(datap + 4), _mm_mul_ps(xmm1_data, xmm_stepsize)); in opj_t1_clbl_decode_processor()
1743 _mm_store_ps((float*)(datap + 8), _mm_mul_ps(xmm2_data, xmm_stepsize)); in opj_t1_clbl_decode_processor()
1744 _mm_store_ps((float*)(datap + 12), _mm_mul_ps(xmm3_data, xmm_stepsize)); in opj_t1_clbl_decode_processor()
/external/webrtc/webrtc/common_audio/
Dfir_filter_sse.cc64 m_sum = _mm_add_ps(m_sum, _mm_mul_ps(m_in, _mm_load_ps(coef_ptr + j))); in Filter()
69 m_sum = _mm_add_ps(m_sum, _mm_mul_ps(m_in, _mm_load_ps(coef_ptr + j))); in Filter()
/external/mesa3d/src/gallium/drivers/swr/rasterizer/core/
Dformat_types.h708 ret = _mm_mul_ps(ret, factor);
715 ret = _mm_mul_ps(ret, _mm_set1_ps(1.0f * expnum / expden));
729 __m128 xover = _mm_mul_ps(arg, xf);
732 __m128 x2 = _mm_mul_ps(arg, arg);
733 __m128 xunder = _mm_mul_ps(x2, xfm1);
736 __m128 xavg = _mm_mul_ps(_mm_set1_ps(1.0f / (3.0f * 0.629960524947437f) * 0.999852f),
739 xavg = _mm_mul_ps(xavg, _mm_rsqrt_ps(xavg));
740 xavg = _mm_mul_ps(xavg, _mm_rsqrt_ps(xavg));
768 Result = _mm_mul_ps(Src, _mm_set1_ps(12.92f));
782 f = _mm_mul_ps(f, _mm_set1_ps(1.055f));
[all …]
Dfrontend.h230 vC = _mm_mul_ps(vA, vX); in triangleSetupC()
231 __m128 vCy = _mm_mul_ps(vB, vY); in triangleSetupC()
232 vC = _mm_mul_ps(vC, _mm_set1_ps(-1.0f)); in triangleSetupC()
Drasterizer_impl.h876 vX = _mm_mul_ps(_mm_cvtepi32_ps(vXi), _mm_set1_ps(1.0f / FIXED_POINT_SCALE));
877 vY = _mm_mul_ps(_mm_cvtepi32_ps(vYi), _mm_set1_ps(1.0f / FIXED_POINT_SCALE));
894 vA = _mm_mul_ps(vA, _mm_set1_ps(-1));
895 vB = _mm_mul_ps(vB, _mm_set1_ps(-1));
958 attribA = _mm_mul_ps(attribA, vOneOverWV0);
959 attribB = _mm_mul_ps(attribB, vOneOverWV1);
960 attribC = _mm_mul_ps(attribC, vOneOverWV2);
Drasterizer.cpp98 __m128 vAdjust = _mm_mul_ps(vLineWidth, vBloat0); in RasterizeLine()
183 vAdjust = _mm_mul_ps(vLineWidth, vBloat1); in RasterizeLine()
Drasterizer.h43 __m128 vFixed = _mm_mul_ps(vIn, _mm_set1_ps(FIXED_POINT_SCALE)); in fpToFixedPoint()
/external/swiftshader/src/Device/
DSurface.cpp4259 c0 = _mm_mul_ps(c0, _mm_set1_ps(1.0f / 2.0f)); in resolve()
4282 c0 = _mm_mul_ps(c0, _mm_set1_ps(1.0f / 4.0f)); in resolve()
4315 c0 = _mm_mul_ps(c0, _mm_set1_ps(1.0f / 8.0f)); in resolve()
4368 c0 = _mm_mul_ps(c0, _mm_set1_ps(1.0f / 16.0f)); in resolve()
4556 c0 = _mm_mul_ps(c0, _mm_set1_ps(1.0f / 2.0f)); in resolve()
4579 c0 = _mm_mul_ps(c0, _mm_set1_ps(1.0f / 4.0f)); in resolve()
4612 c0 = _mm_mul_ps(c0, _mm_set1_ps(1.0f / 8.0f)); in resolve()
4665 c0 = _mm_mul_ps(c0, _mm_set1_ps(1.0f / 16.0f)); in resolve()
4853 c0 = _mm_mul_ps(c0, _mm_set1_ps(1.0f / 2.0f)); in resolve()
4876 c0 = _mm_mul_ps(c0, _mm_set1_ps(1.0f / 4.0f)); in resolve()
[all …]
/external/swiftshader/src/Renderer/
DSurface.cpp4931 c0 = _mm_mul_ps(c0, _mm_set1_ps(1.0f / 2.0f)); in resolve()
4954 c0 = _mm_mul_ps(c0, _mm_set1_ps(1.0f / 4.0f)); in resolve()
4987 c0 = _mm_mul_ps(c0, _mm_set1_ps(1.0f / 8.0f)); in resolve()
5040 c0 = _mm_mul_ps(c0, _mm_set1_ps(1.0f / 16.0f)); in resolve()
5228 c0 = _mm_mul_ps(c0, _mm_set1_ps(1.0f / 2.0f)); in resolve()
5251 c0 = _mm_mul_ps(c0, _mm_set1_ps(1.0f / 4.0f)); in resolve()
5284 c0 = _mm_mul_ps(c0, _mm_set1_ps(1.0f / 8.0f)); in resolve()
5337 c0 = _mm_mul_ps(c0, _mm_set1_ps(1.0f / 16.0f)); in resolve()
5527 c0 = _mm_mul_ps(c0, _mm_set1_ps(1.0f / 2.0f)); in resolve()
5550 c0 = _mm_mul_ps(c0, _mm_set1_ps(1.0f / 4.0f)); in resolve()
[all …]
/external/clang/test/CodeGen/
Dsse-builtins.c483 return _mm_mul_ps(A, B); in test_mm_mul_ps()

12