From 7ff266e3ce592ec1a6284cf16811965eec775f25 Mon Sep 17 00:00:00 2001 From: Konstantinos Margaritis Date: Fri, 29 Aug 2014 20:03:49 +0000 Subject: Initial VSX commit --- Eigen/src/Core/arch/AltiVec/Complex.h | 186 ++++++++++++++++++++++++++++++++++ 1 file changed, 186 insertions(+) (limited to 'Eigen/src/Core/arch/AltiVec/Complex.h') diff --git a/Eigen/src/Core/arch/AltiVec/Complex.h b/Eigen/src/Core/arch/AltiVec/Complex.h index 13b874d0c..064341a3b 100644 --- a/Eigen/src/Core/arch/AltiVec/Complex.h +++ b/Eigen/src/Core/arch/AltiVec/Complex.h @@ -14,7 +14,13 @@ namespace Eigen { namespace internal { +#ifdef _BIG_ENDIAN static Packet4ui p4ui_CONJ_XOR = vec_mergeh((Packet4ui)p4i_ZERO, (Packet4ui)p4f_ZERO_);//{ 0x00000000, 0x80000000, 0x00000000, 0x80000000 }; +#else +static Packet4ui p4ui_CONJ_XOR = vec_mergeh((Packet4ui)p4f_ZERO_, (Packet4ui)p4i_ZERO);//{ 0x80000000, 0x00000000, 0x80000000, 0x00000000 }; +static Packet2ul p2ul_CONJ_XOR = (Packet2ul) vec_sld((Packet4ui)p2d_ZERO_, (Packet4ui)p2l_ZERO, 8);//{ 0x8000000000000000, 0x0000000000000000 }; +#endif + static Packet16uc p16uc_COMPLEX_RE = vec_sld((Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 0), (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 2), 8);//{ 0,1,2,3, 0,1,2,3, 8,9,10,11, 8,9,10,11 }; static Packet16uc p16uc_COMPLEX_IM = vec_sld(p16uc_DUPLICATE, (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 3), 8);//{ 4,5,6,7, 4,5,6,7, 12,13,14,15, 12,13,14,15 }; static Packet16uc p16uc_COMPLEX_REV = vec_sld(p16uc_REVERSE, p16uc_REVERSE, 8);//{ 4,5,6,7, 0,1,2,3, 12,13,14,15, 8,9,10,11 }; @@ -237,6 +243,186 @@ EIGEN_STRONG_INLINE void ptranspose(PacketBlock& kernel) kernel.packet[0].v = tmp; } +//---------- double ---------- +struct Packet1cd +{ + EIGEN_STRONG_INLINE Packet1cd() {} + EIGEN_STRONG_INLINE explicit Packet1cd(const Packet2d& a) : v(a) {} + Packet2d v; +}; + +template<> struct packet_traits > : default_packet_traits +{ + typedef Packet1cd type; + typedef Packet1cd half; + enum { + Vectorizable = 1, + AlignedOnScalar = 0, + size = 1, + HasHalfPacket = 0, + + HasAdd = 1, + HasSub = 1, + HasMul = 1, + HasDiv = 1, + HasNegate = 1, + HasAbs = 0, + HasAbs2 = 0, + HasMin = 0, + HasMax = 0, + HasSetLinear = 0 + }; +}; + +template<> struct unpacket_traits { typedef std::complex type; enum {size=1}; typedef Packet1cd half; }; + +template<> EIGEN_STRONG_INLINE Packet1cd padd(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_add(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet1cd psub(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_sub(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet1cd pnegate(const Packet1cd& a) { return Packet1cd(pnegate(Packet2d(a.v))); } +template<> EIGEN_STRONG_INLINE Packet1cd pconj(const Packet1cd& a) { return Packet1cd((Packet2d)vec_xor((Packet2ul)a.v, p2ul_CONJ_XOR)); } + +template<> EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) +{ + Packet2d v1, v2; + + // Permute and multiply the real parts of a and b + v1 = vec_perm(a.v, a.v, p16uc_COMPLEX_RE); + // Get the imaginary parts of a + v2 = vec_perm(a.v, a.v, p16uc_COMPLEX_IM); + // multiply a_re * b + v1 = vec_madd(v1, b.v, p4f_ZERO); + // multiply a_im * b and get the conjugate result + v2 = vec_madd(v2, b.v, p4f_ZERO); + v2 = (Packet4f) vec_xor((Packet4ui)v2, p4ui_CONJ_XOR); + // permute back to a proper order + v2 = vec_perm(v2, v2, p16uc_COMPLEX_REV); + + return Packet2cf(vec_add(v1, v2)); +} + +template<> EIGEN_STRONG_INLINE Packet1cd pand (const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_and(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet1cd por (const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_or(a.v,b.v)); } +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 pload (const std::complex* from) { EIGEN_DEBUG_ALIGNED_LOAD return Packet1cd(pload((const double*)from)); } +template<> EIGEN_STRONG_INLINE Packet1cd ploadu(const std::complex* from) { EIGEN_DEBUG_UNALIGNED_LOAD return Packet1cd(ploadu((const double*)from)); } + +template<> EIGEN_STRONG_INLINE Packet1cd ploaddup(const std::complex* from) +{ + return pset1(*from); +} + +template<> EIGEN_STRONG_INLINE void pstore >(std::complex * to, const Packet1cd& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((double*)to, from.v); } +template<> EIGEN_STRONG_INLINE void pstoreu >(std::complex * to, const Packet1cd& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((double*)to, from.v); } + +template<> EIGEN_STRONG_INLINE void prefetch >(const std::complex * addr) { vec_dstt((double *)addr, DST_CTRL(2,2,32), DST_CHAN); } + +template<> EIGEN_STRONG_INLINE std::complex pfirst(const Packet1cd& a) +{ + std::complex EIGEN_ALIGN16 res[2]; + pstore((float *)&res, a.v); + + return res[0]; +} + +template<> EIGEN_STRONG_INLINE Packet1cd preverse(const Packet1cd& a) +{ + Packet2d rev_a; + rev_a = vec_perm(a.v, a.v, p16uc_COMPLEX_REV2); + return Packet1cd(rev_a); +} + +template<> EIGEN_STRONG_INLINE std::complex predux(const Packet1cd& a) +{ + Packet2d b; + b = (Packet2d) vec_sld(a.v, a.v, 8); + b = padd(a.v, b); + return pfirst(Packet1cd(b)); +} + +template<> EIGEN_STRONG_INLINE Packet1cd preduxp(const Packet1cd* vecs) +{ + Packet2d b1, b2; + + b1 = (Packet2d) vec_sld(vecs[0].v, vecs[1].v, 8); + b2 = (Packet2d) vec_sld(vecs[1].v, vecs[0].v, 8); + b2 = (Packet2d) vec_sld(b2, b2, 8); + b2 = padd(b1, b2); + + return Packet1cd(b2); +} + +template<> EIGEN_STRONG_INLINE std::complex predux_mul(const Packet1cd& a) +{ + Packet2d b; + Packet1cd prod; + b = (Packet2d) vec_sld(a.v, a.v, 8); + prod = pmul(a, Packet1cd(b)); + + return pfirst(prod); +} + +template +struct palign_impl +{ + static EIGEN_STRONG_INLINE void run(Packet1cd& first, const Packet1cd& second) + { } +}; + +template<> struct conj_helper +{ + EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const + { + return internal::pmul(a, pconj(b)); + } +}; + +template<> struct conj_helper +{ + EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const + { + return internal::pmul(pconj(a), b); + } +}; + +template<> struct conj_helper +{ + EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const + { + return pconj(internal::pmul(a, b)); + } +}; + +template<> EIGEN_STRONG_INLINE Packet1cd pdiv(const Packet1cd& a, const Packet1cd& b) +{ + // TODO optimize it for AltiVec + Packet1cd res = conj_helper().pmul(a,b); + Packet2d s = vec_madd(b.v, b.v, p4f_ZERO); + return Packet1cd(pdiv(res.v, vec_add(s,vec_perm(s, s, p16uc_COMPLEX_REV)))); +} + +template<> EIGEN_STRONG_INLINE Packet1cd pcplxflip(const Packet1cd& x) +{ + return Packet1cd(vec_perm(x.v, x.v, p16uc_COMPLEX_REV)); +} + +EIGEN_STRONG_INLINE void ptranspose(PacketBlock& kernel) +{ + Packet2d tmp = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_COMPLEX_TRANSPOSE_0); + kernel.packet[1].v = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_COMPLEX_TRANSPOSE_1); + kernel.packet[0].v = tmp; +} + } // end namespace internal } // end namespace Eigen -- cgit v1.2.3