Home
last modified time | relevance | path

Searched refs:_mm_add_ps (Results 1 – 25 of 28) sorted by relevance

12

/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()
75 _mm_storeu_ps(sum,_mm_add_ps(xsum1,xsum2)); 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()
[all …]
Dvq_sse2.c75 sums = _mm_add_ps(sums, x4); in op_pvq_search_sse2()
82 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()
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()
124 _mm_storeu_ps(&y[j], _mm_add_ps(y4, y4)); 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()
[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/webrtc/webrtc/modules/audio_processing/aec/
Daec_core_sse2.c62 const __m128 f = _mm_add_ps(c, d); in FilterFarSSE2()
63 const __m128 g = _mm_add_ps(y_fft_re, e); in FilterFarSSE2()
64 const __m128 h = _mm_add_ps(y_fft_im, f); in FilterFarSSE2()
101 const __m128 xPowPlus = _mm_add_ps(x_pow_local, k1e_10f); in ScaleErrorSignalSSE2()
106 const __m128 ef_sum2 = _mm_add_ps(ef_re2, ef_im2); in ScaleErrorSignalSSE2()
109 __m128 absEfPlus = _mm_add_ps(absEf, k1e_10f); in ScaleErrorSignalSSE2()
181 const __m128 e = _mm_add_ps(a, b); in FilterAdaptationSSE2()
223 wtBuf_re = _mm_add_ps(wtBuf_re, fft_re); in FilterAdaptationSSE2()
224 wtBuf_im = _mm_add_ps(wtBuf_im, fft_im); in FilterAdaptationSSE2()
295 const __m128 pol5_y_1 = _mm_add_ps(pol5_y_0, *((__m128*)C4)); in mm_pow_ps()
[all …]
Daec_rdft_sse2.c38 __m128 x0v = _mm_add_ps(a01v, a23v); in cft1st_128_SSE2()
40 const __m128 x2v = _mm_add_ps(a45v, a67v); in cft1st_128_SSE2()
43 a01v = _mm_add_ps(x0v, x2v); in cft1st_128_SSE2()
49 a45v = _mm_add_ps(a45_0v, a45_1v); in cft1st_128_SSE2()
55 x0v = _mm_add_ps(x1v, x3s); in cft1st_128_SSE2()
59 a23v = _mm_add_ps(a23_0v, a23_1v); in cft1st_128_SSE2()
67 a67v = _mm_add_ps(a67_0v, a67_1v); in cft1st_128_SSE2()
98 __m128 x0r0_0i0_0r1_x0i1 = _mm_add_ps(a_00_32, a_08_40); in cftmdl_128_SSE2()
111 const __m128 x2r0_2i0_2r1_x2i1 = _mm_add_ps(a_16_48, a_24_56); in cftmdl_128_SSE2()
114 const __m128 xx0 = _mm_add_ps(x0r0_0i0_0r1_x0i1, x2r0_2i0_2r1_x2i1); in cftmdl_128_SSE2()
[all …]
/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()
48 m_sums1 = _mm_add_ps(m_sums1, m_sums2); in Convolve_SSE()
52 m_sums2 = _mm_add_ps(_mm_movehl_ps(m_sums1, m_sums1), m_sums1); in Convolve_SSE()
/external/libaom/libaom/aom_dsp/x86/
Dfft_sse2.c76 imag1 = _mm_add_ps(imag1, imag2); in aom_fft_unpack_2d_output_sse2()
98 real1 = _mm_add_ps(real1, real2); in aom_fft_unpack_2d_output_sse2()
109 _mm_set1_ps, _mm_add_ps, _mm_sub_ps);
111 _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);
139 _mm_set1_ps, _mm_add_ps, _mm_sub_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/libaom/libaom/av1/encoder/x86/
Dml_sse3.c35 const __m128 vadd = _mm_add_ps(mul_l, mul_h); in nn_propagate_8to1()
41 *output = _mm_add_ps(*output, hadd2); in nn_propagate_8to1()
58 *output = _mm_add_ps(*output, hadd2); in nn_propagate_4to1()
80 *outputs = _mm_add_ps(*outputs, hh); in nn_propagate_4to4()
106 *out_h = _mm_add_ps(*out_h, hh1); in nn_propagate_4to8()
107 *out_l = _mm_add_ps(*out_l, hh0); in nn_propagate_4to8()
123 add[i] = _mm_add_ps(mul_l, mul_h); in nn_propagate_8to4()
139 *outputs = _mm_add_ps(*outputs, haddhadd); in nn_propagate_8to4()
233 total = _mm_add_ps(total, _mm_mul_ps(input, weight)); in av1_nn_predict_sse3()
/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()
50 sum = _mm_add_ps(sum, _mm_movehl_ps(sum, sum)); 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()
68 sum = _mm_add_ps(sum, _mm_movehl_ps(sum, sum)); in interpolate_product_single()
/external/swiftshader/src/Device/
DSurface.cpp4258 c0 = _mm_add_ps(c0, c1); in resolve()
4279 c0 = _mm_add_ps(c0, c1); in resolve()
4280 c2 = _mm_add_ps(c2, c3); in resolve()
4281 c0 = _mm_add_ps(c0, c2); in resolve()
4308 c0 = _mm_add_ps(c0, c1); in resolve()
4309 c2 = _mm_add_ps(c2, c3); in resolve()
4310 c4 = _mm_add_ps(c4, c5); in resolve()
4311 c6 = _mm_add_ps(c6, c7); in resolve()
4312 c0 = _mm_add_ps(c0, c2); in resolve()
4313 c4 = _mm_add_ps(c4, c6); in resolve()
[all …]
/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()
72 m_sum = _mm_add_ps(_mm_movehl_ps(m_sum, m_sum), m_sum); in Filter()
/external/swiftshader/src/Renderer/
DSurface.cpp4930 c0 = _mm_add_ps(c0, c1); in resolve()
4951 c0 = _mm_add_ps(c0, c1); in resolve()
4952 c2 = _mm_add_ps(c2, c3); in resolve()
4953 c0 = _mm_add_ps(c0, c2); in resolve()
4980 c0 = _mm_add_ps(c0, c1); in resolve()
4981 c2 = _mm_add_ps(c2, c3); in resolve()
4982 c4 = _mm_add_ps(c4, c5); in resolve()
4983 c6 = _mm_add_ps(c6, c7); in resolve()
4984 c0 = _mm_add_ps(c0, c2); in resolve()
4985 c4 = _mm_add_ps(c4, c6); in resolve()
[all …]
/external/eigen/Eigen/src/Core/arch/SSE/
DComplex.h55 … padd<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_add_ps(a.v,b.v)); }
79 return Packet2cf(_mm_add_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v),
152 return pfirst(Packet2cf(_mm_add_ps(a.v, _mm_movehl_ps(a.v,a.v))));
157 …return Packet2cf(_mm_add_ps(_mm_movelh_ps(vecs[0].v,vecs[1].v), _mm_movehl_ps(vecs[1].v,vecs[0].v)…
189 return Packet2cf(_mm_add_ps(_mm_xor_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v), mask),
207 return Packet2cf(_mm_add_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v),
255 …return Packet2cf(_mm_div_ps(res.v,_mm_add_ps(s,_mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128…
DPacketMath.h194 template<> EIGEN_STRONG_INLINE Packet4f plset<Packet4f>(const float& a) { return _mm_add_ps(pset1<P…
198 …NG_INLINE Packet4f padd<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_add_ps(a,b); }
520 tmp0 = _mm_add_ps(tmp0, tmp1);
522 tmp1 = _mm_add_ps(tmp1, tmp2);
525 return _mm_add_ps(tmp0, tmp2);
542 Packet4f tmp = _mm_add_ps(a, _mm_movehl_ps(a,a));
/external/mesa3d/src/gallium/drivers/swr/rasterizer/core/
Drasterizer.cpp101 vXa = _mm_add_ps(vAdjust, vXa); in RasterizeLine()
105 vYa = _mm_add_ps(vAdjust, vYa); in RasterizeLine()
186 vXa = _mm_add_ps(vAdjust, vXa); in RasterizeLine()
190 vYa = _mm_add_ps(vAdjust, vYa); in RasterizeLine()
/external/eigen/Eigen/src/Geometry/arch/
DGeometry_SSE.h36 _mm_add_ps(_mm_sub_ps(_mm_mul_ps(a,vec4f_swizzle1(b,3,3,3,3)),
39 _mm_xor_ps(mask,_mm_add_ps(s1,s2))));
/external/pdfium/third_party/libopenjpeg20/
Dmct.c408 vr = _mm_add_ps(vy, _mm_mul_ps(vv, vrv)); 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()
423 vb = _mm_add_ps(vy, _mm_mul_ps(vu, vbu)); in opj_mct_decode_real()
Ddwt.c2362 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()
2379 c = _mm_add_ps(c, c); in opj_v4dwt_decode_step2_sse()
2381 vw[-1] = _mm_add_ps(vw[-1], c); in opj_v4dwt_decode_step2_sse()
/external/eigen/Eigen/src/LU/arch/
DInverse_SSE.h109 iD = _mm_add_ps(iD,_mm_mul_ps(_mm_shuffle_ps(C,C,0xF5), _mm_movehl_ps(AB,AB)));
112 iA = _mm_add_ps(iA,_mm_mul_ps(_mm_shuffle_ps(B,B,0xF5), _mm_movehl_ps(DC,DC)));
115 d = _mm_add_ps(d, _mm_movehl_ps(d, d));
/external/clang/test/CodeGen/
Dsse-builtins.c13 return _mm_add_ps(A, B); in test_mm_add_ps()
/external/eigen/Eigen/src/Core/arch/AVX/
DPacketMath.h398 return predux(Packet4f(_mm_add_ps(_mm256_castps256_ps128(a),_mm256_extractf128_ps(a,1))));
407 return _mm_add_ps(_mm256_castps256_ps128(a),_mm256_extractf128_ps(a,1));
/external/speex/
Dconfigure.ac72 return _mm_add_ps(_mm_loadu_ps(a), _mm_loadu_ps(b));
/external/skia/include/private/
DSkNx_sse.h78 AI SkNx operator + (const SkNx& o) const { return _mm_add_ps(fVec, o.fVec); }
171 AI SkNx operator + (const SkNx& o) const { return _mm_add_ps(fVec, o.fVec); }
/external/skqp/include/private/
DSkNx_sse.h78 AI SkNx operator + (const SkNx& o) const { return _mm_add_ps(fVec, o.fVec); }
171 AI SkNx operator + (const SkNx& o) const { return _mm_add_ps(fVec, o.fVec); }
/external/clang/lib/Headers/
Dxmmintrin.h81 _mm_add_ps(__m128 __a, __m128 __b) in _mm_add_ps() function

12