devtools/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2026-04-10 11:34:33 +08:00
Code Issues Packages Projects Releases Wiki Activity

Fix x86 complex vectorized fma

Browse Source
This commit is contained in:
Charles Schlosser
2025-03-12 17:06:32 +00:00
parent 464c1d0978
commit 10e62ccd22
3 changed files with 34 additions and 48 deletions
Download Patch File Download Diff File Expand all files Collapse all files

24
Eigen/src/Core/arch/AVX/Complex.h
View File

@@ -475,19 +475,11 @@ EIGEN_STRONG_INLINE Packet4cf pmsub(const Packet4cf& a, const Packet4cf& b, cons
}
template <>
EIGEN_STRONG_INLINE Packet4cf pnmadd(const Packet4cf& a, const Packet4cf& b, const Packet4cf& c) {
__m256 a_odd = _mm256_movehdup_ps(a.v);
__m256 a_even = _mm256_moveldup_ps(a.v);
__m256 b_swap = _mm256_permute_ps(b.v, _MM_SHUFFLE(2, 3, 0, 1));
__m256 result = _mm256_fmaddsub_ps(a_odd, b_swap, _mm256_fmaddsub_ps(a_even, b.v, c.v));
return Packet4cf(result);
return pnegate(pmsub(a, b, c));
}
template <>
EIGEN_STRONG_INLINE Packet4cf pnmsub(const Packet4cf& a, const Packet4cf& b, const Packet4cf& c) {
__m256 a_odd = _mm256_movehdup_ps(a.v);
__m256 a_even = _mm256_moveldup_ps(a.v);
__m256 b_swap = _mm256_permute_ps(b.v, _MM_SHUFFLE(2, 3, 0, 1));
__m256 result = _mm256_fmaddsub_ps(a_odd, b_swap, _mm256_fmsubadd_ps(a_even, b.v, c.v));
return Packet4cf(result);
return pnegate(pmadd(a, b, c));
}
// std::complex<double>
template <>
@@ -508,19 +500,11 @@ EIGEN_STRONG_INLINE Packet2cd pmsub(const Packet2cd& a, const Packet2cd& b, cons
}
template <>
EIGEN_STRONG_INLINE Packet2cd pnmadd(const Packet2cd& a, const Packet2cd& b, const Packet2cd& c) {
__m256d a_odd = _mm256_permute_pd(a.v, 0xF);
__m256d a_even = _mm256_movedup_pd(a.v);
__m256d b_swap = _mm256_permute_pd(b.v, 0x5);
__m256d result = _mm256_fmaddsub_pd(a_odd, b_swap, _mm256_fmaddsub_pd(a_even, b.v, c.v));
return Packet2cd(result);
return pnegate(pmsub(a, b, c));
}
template <>
EIGEN_STRONG_INLINE Packet2cd pnmsub(const Packet2cd& a, const Packet2cd& b, const Packet2cd& c) {
__m256d a_odd = _mm256_permute_pd(a.v, 0xF);
__m256d a_even = _mm256_movedup_pd(a.v);
__m256d b_swap = _mm256_permute_pd(b.v, 0x5);
__m256d result = _mm256_fmaddsub_pd(a_odd, b_swap, _mm256_fmsubadd_pd(a_even, b.v, c.v));
return Packet2cd(result);
return pnegate(pmadd(a, b, c));
}
#endif
} // end namespace internal