GitHub-Proxy/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-08-11 19:29:02 +08:00
Code Issues Packages Projects Releases Wiki Activity

Vectorized the multiplication and division of complex numbers using AVX instructions.

Browse Source
This commit is contained in:
Benoit Steiner 2014-03-26 15:11:18 -07:00
parent cf1a7bfbe1
commit a078f442a3
1 changed files with 15 additions and 21 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

36
Eigen/src/Core/arch/AVX/Complex.h
View File

@ -221,13 +221,11 @@ template<> struct conj_helper<Packet4cf, Packet8f, false,false>
template<> EIGEN_STRONG_INLINE Packet4cf pdiv<Packet4cf>(const Packet4cf& a, const Packet4cf& b) template<> EIGEN_STRONG_INLINE Packet4cf pdiv<Packet4cf>(const Packet4cf& a, const Packet4cf& b)
{ {
Packet4cf res; Packet4cf num = pmul(a, pconj(b));
for (int i = 0; i < 8; i+=2) { __m256 tmp = _mm256_mul_ps(b.v, b.v);
std::complex<float> result = std::complex<float>(a.v[i], a.v[i+1]) / std::complex<float>(b.v[i], b.v[i+1]); __m256 tmp2 = _mm256_shuffle_ps(tmp,tmp,0xB1);
res.v[i] = std::real(result); __m256 denom = _mm256_add_ps(tmp, tmp2);
res.v[i+1] = std::imag(result); return Packet4cf(_mm256_div_ps(num.v, denom));
}
return res;
} }
template<> EIGEN_STRONG_INLINE Packet4cf pcplxflip<Packet4cf>(const Packet4cf& x) template<> EIGEN_STRONG_INLINE Packet4cf pcplxflip<Packet4cf>(const Packet4cf& x)
@ -282,13 +280,12 @@ template<> EIGEN_STRONG_INLINE Packet2cd pconj(const Packet2cd& a)
template<> EIGEN_STRONG_INLINE Packet2cd pmul<Packet2cd>(const Packet2cd& a, const Packet2cd& b) template<> EIGEN_STRONG_INLINE Packet2cd pmul<Packet2cd>(const Packet2cd& a, const Packet2cd& b)
{ {
__m256d tmp1 = _mm256_mul_pd(_mm256_permute_pd(a.v, 0), b.v); __m256d tmp1 = _mm256_shuffle_pd(a.v,a.v,0x0);
// FIXME: _mm256_permute_pd(b.v, _MM_SHUFFLE2(1,0) won't work as expected, figure out an alternative. __m256d even = _mm256_mul_pd(tmp1, b.v);
__m256d op = {b.v[1], b.v[0], b.v[3], b.v[2]}; __m256d tmp2 = _mm256_shuffle_pd(a.v,a.v,0xF);
__m256d tmp2 = _mm256_mul_pd(_mm256_permute_pd(a.v, 15), op); __m256d tmp3 = _mm256_shuffle_pd(b.v,b.v,0x5);
__m256d result = _mm256_addsub_pd(tmp1, tmp2); __m256d odd = _mm256_mul_pd(tmp2, tmp3);
return Packet2cd(_mm256_addsub_pd(even, odd));
return Packet2cd(result);
} }
template<> EIGEN_STRONG_INLINE Packet2cd pand <Packet2cd>(const Packet2cd& a, const Packet2cd& b) { return Packet2cd(_mm256_and_pd(a.v,b.v)); } template<> EIGEN_STRONG_INLINE Packet2cd pand <Packet2cd>(const Packet2cd& a, const Packet2cd& b) { return Packet2cd(_mm256_and_pd(a.v,b.v)); }
@ -418,13 +415,10 @@ template<> struct conj_helper<Packet2cd, Packet4d, false,false>
template<> EIGEN_STRONG_INLINE Packet2cd pdiv<Packet2cd>(const Packet2cd& a, const Packet2cd& b) template<> EIGEN_STRONG_INLINE Packet2cd pdiv<Packet2cd>(const Packet2cd& a, const Packet2cd& b)
{ {
Packet2cd res; Packet2cd num = pmul(a, pconj(b));
for (int i = 0; i < 4; i+=2) { __m256d tmp = _mm256_mul_pd(b.v, b.v);
std::complex<double> result = std::complex<double>(a.v[i], a.v[i+1]) / std::complex<double>(b.v[i], b.v[i+1]); __m256d denom = _mm256_hadd_pd(tmp, tmp);
res.v[i] = std::real(result); return Packet2cd(_mm256_div_pd(num.v, denom));
res.v[i+1] = std::imag(result);
}
return res;
} }
template<> EIGEN_STRONG_INLINE Packet2cd pcplxflip<Packet2cd>(const Packet2cd& x) template<> EIGEN_STRONG_INLINE Packet2cd pcplxflip<Packet2cd>(const Packet2cd& x)