From 1e74f93d55bf517d8db52c3d7a9680144c48082f Mon Sep 17 00:00:00 2001 From: Andreas Krebbel Date: Wed, 25 Nov 2020 14:11:23 +0000 Subject: Fix some packet-functions in the IBM ZVector packet-math. --- Eigen/src/Core/arch/ZVector/Complex.h | 21 +++++++++++++++++++++ 1 file changed, 21 insertions(+) (limited to 'Eigen/src/Core/arch/ZVector/Complex.h') diff --git a/Eigen/src/Core/arch/ZVector/Complex.h b/Eigen/src/Core/arch/ZVector/Complex.h index d3e41b43e..ddf5a97d8 100644 --- a/Eigen/src/Core/arch/ZVector/Complex.h +++ b/Eigen/src/Core/arch/ZVector/Complex.h @@ -140,6 +140,11 @@ template<> EIGEN_STRONG_INLINE Packet1cd por (const Packet1cd& a, template<> EIGEN_STRONG_INLINE Packet1cd pxor (const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_xor(a.v,b.v)); } template<> EIGEN_STRONG_INLINE Packet1cd pandnot (const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_and(a.v, vec_nor(b.v,b.v))); } template<> EIGEN_STRONG_INLINE Packet1cd ploaddup(const std::complex* from) { return pset1(*from); } +template<> EIGEN_STRONG_INLINE Packet1cd pcmp_eq(const Packet1cd& a, const Packet1cd& b) { + Packet2d eq = vec_cmpeq (a.v, b.v); + Packet2d tmp = { eq[1], eq[0] }; + return (Packet1cd)pand(eq, tmp); +} template<> EIGEN_STRONG_INLINE void prefetch >(const std::complex * addr) { EIGEN_ZVECTOR_PREFETCH(addr); } @@ -281,6 +286,17 @@ template<> EIGEN_STRONG_INLINE void prefetch >(const std::co #if !defined(__ARCH__) || (defined(__ARCH__) && __ARCH__ < 12) + +template<> EIGEN_STRONG_INLINE Packet2cf pcmp_eq(const Packet2cf& a, const Packet2cf& b) { + Packet4f eq = pcmp_eq (a.v, b.v); + Packet2cf res; + Packet2d tmp1 = { eq.v4f[0][1], eq.v4f[0][0] }; + Packet2d tmp2 = { eq.v4f[1][1], eq.v4f[1][0] }; + res.v.v4f[0] = pand(eq.v4f[0], tmp1); + res.v.v4f[1] = pand(eq.v4f[1], tmp2); + return res; +} + template<> EIGEN_STRONG_INLINE Packet2cf pconj(const Packet2cf& a) { Packet2cf res; @@ -387,6 +403,11 @@ template<> EIGEN_STRONG_INLINE Packet2cf pblend(const Selector<2>& ifPacket, con return result; } #else +template<> EIGEN_STRONG_INLINE Packet2cf pcmp_eq(const Packet2cf& a, const Packet2cf& b) { + Packet4f eq = vec_cmpeq (a.v, b.v); + Packet4f tmp = { eq[1], eq[0], eq[3], eq[2] }; + return (Packet2cf)pand(eq, tmp); +} template<> EIGEN_STRONG_INLINE Packet2cf pconj(const Packet2cf& a) { return Packet2cf(pxor(a.v, reinterpret_cast(p4ui_CONJ_XOR))); } template<> EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) { -- cgit v1.2.3