mostly cleanups and modernizing code

This commit is contained in:
Konstantinos Margaritis 2016-06-19 16:13:17 -03:00
parent b410d46482
commit 8c34b5a0e3

View File

@ -42,6 +42,7 @@ template<> struct packet_traits<std::complex<float> > : default_packet_traits
Vectorizable = 1, Vectorizable = 1,
AlignedOnScalar = 1, AlignedOnScalar = 1,
size = 2, size = 2,
HasHalfPacket = 0,
HasAdd = 1, HasAdd = 1,
HasSub = 1, HasSub = 1,
@ -52,7 +53,8 @@ template<> struct packet_traits<std::complex<float> > : default_packet_traits
HasAbs2 = 0, HasAbs2 = 0,
HasMin = 0, HasMin = 0,
HasMax = 0, HasMax = 0,
HasSetLinear = 0 HasSetLinear = 0,
HasBlend = 1
}; };
}; };
@ -61,7 +63,6 @@ template<> struct unpacket_traits<Packet2cf> { typedef std::complex<float> type;
template<> EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<float>& from) template<> EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<float>& from)
{ {
Packet2cf res; Packet2cf res;
/* On AltiVec we cannot load 64-bit registers, so wa have to take care of alignment */
if((ptrdiff_t(&from) % 16) == 0) if((ptrdiff_t(&from) % 16) == 0)
res.v = pload<Packet4f>((const float *)&from); res.v = pload<Packet4f>((const float *)&from);
else else
@ -70,26 +71,32 @@ template<> EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<flo
return res; return res;
} }
template<> EIGEN_STRONG_INLINE Packet2cf pload <Packet2cf>(const std::complex<float>* from) { EIGEN_DEBUG_ALIGNED_LOAD return Packet2cf(pload<Packet4f>((const float*)from)); }
template<> EIGEN_STRONG_INLINE Packet2cf ploadu<Packet2cf>(const std::complex<float>* from) { EIGEN_DEBUG_UNALIGNED_LOAD return Packet2cf(ploadu<Packet4f>((const float*)from)); }
template<> EIGEN_STRONG_INLINE Packet2cf ploaddup<Packet2cf>(const std::complex<float>* from) { return pset1<Packet2cf>(*from); }
template<> EIGEN_STRONG_INLINE void pstore <std::complex<float> >(std::complex<float> * to, const Packet2cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((float*)to, from.v); }
template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<float> >(std::complex<float> * to, const Packet2cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((float*)to, from.v); }
template<> EIGEN_DEVICE_FUNC inline Packet2cf pgather<std::complex<float>, Packet2cf>(const std::complex<float>* from, Index stride) template<> EIGEN_DEVICE_FUNC inline Packet2cf pgather<std::complex<float>, Packet2cf>(const std::complex<float>* from, Index stride)
{ {
std::complex<float> EIGEN_ALIGN16 af[2]; std::complex<float> EIGEN_ALIGN16 af[2];
af[0] = from[0*stride]; af[0] = from[0*stride];
af[1] = from[1*stride]; af[1] = from[1*stride];
return Packet2cf(vec_ld(0, (const float*)af)); return pload<Packet2cf>(af);
} }
template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<float>, Packet2cf>(std::complex<float>* to, const Packet2cf& from, Index stride) template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<float>, Packet2cf>(std::complex<float>* to, const Packet2cf& from, Index stride)
{ {
std::complex<float> EIGEN_ALIGN16 af[2]; std::complex<float> EIGEN_ALIGN16 af[2];
vec_st(from.v, 0, (float*)af); pstore<std::complex<float> >((std::complex<float> *) af, from);
to[0*stride] = af[0]; to[0*stride] = af[0];
to[1*stride] = af[1]; to[1*stride] = af[1];
} }
template<> EIGEN_STRONG_INLINE Packet2cf padd<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(padd<Packet4f>(a.v, b.v)); }
template<> EIGEN_STRONG_INLINE Packet2cf padd<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(vec_add(a.v,b.v)); } template<> EIGEN_STRONG_INLINE Packet2cf psub<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(psub<Packet4f>(a.v, b.v)); }
template<> EIGEN_STRONG_INLINE Packet2cf psub<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(vec_sub(a.v,b.v)); }
template<> EIGEN_STRONG_INLINE Packet2cf pnegate(const Packet2cf& a) { return Packet2cf(pnegate(a.v)); } template<> EIGEN_STRONG_INLINE Packet2cf pnegate(const Packet2cf& a) { return Packet2cf(pnegate(a.v)); }
template<> EIGEN_STRONG_INLINE Packet2cf pconj(const Packet2cf& a) { return Packet2cf((Packet4f)vec_xor((Packet4ui)a.v, p4ui_CONJ_XOR)); } template<> EIGEN_STRONG_INLINE Packet2cf pconj(const Packet2cf& a) { return Packet2cf(pxor<Packet4f>(a.v, reinterpret_cast<Packet4f>(p4ui_CONJ_XOR))); }
template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, const Packet2cf& b) template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
{ {
@ -103,30 +110,19 @@ template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, con
v1 = vec_madd(v1, b.v, p4f_ZERO); v1 = vec_madd(v1, b.v, p4f_ZERO);
// multiply a_im * b and get the conjugate result // multiply a_im * b and get the conjugate result
v2 = vec_madd(v2, b.v, p4f_ZERO); v2 = vec_madd(v2, b.v, p4f_ZERO);
v2 = (Packet4f) vec_xor((Packet4ui)v2, p4ui_CONJ_XOR); v2 = reinterpret_cast<Packet4f>(pxor(v2, reinterpret_cast<Packet4f>(p4ui_CONJ_XOR)));
// permute back to a proper order // permute back to a proper order
v2 = vec_perm(v2, v2, p16uc_COMPLEX32_REV); v2 = vec_perm(v2, v2, p16uc_COMPLEX32_REV);
return Packet2cf(vec_add(v1, v2)); return Packet2cf(vec_add(v1, v2));
} }
template<> EIGEN_STRONG_INLINE Packet2cf pand <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(vec_and(a.v,b.v)); } template<> EIGEN_STRONG_INLINE Packet2cf pand <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(pand<Packet4f>(a.v, b.v)); }
template<> EIGEN_STRONG_INLINE Packet2cf por <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(vec_or(a.v,b.v)); } template<> EIGEN_STRONG_INLINE Packet2cf por <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(por<Packet4f>(a.v, b.v)); }
template<> EIGEN_STRONG_INLINE Packet2cf pxor <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(vec_xor(a.v,b.v)); } template<> EIGEN_STRONG_INLINE Packet2cf pxor <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(pxor<Packet4f>(a.v, b.v)); }
template<> EIGEN_STRONG_INLINE Packet2cf pandnot<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(vec_and(a.v, vec_nor(b.v,b.v))); } template<> EIGEN_STRONG_INLINE Packet2cf pandnot<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(pandnot<Packet4f>(a.v, b.v)); }
template<> EIGEN_STRONG_INLINE Packet2cf pload <Packet2cf>(const std::complex<float>* from) { EIGEN_DEBUG_ALIGNED_LOAD return Packet2cf(pload<Packet4f>((const float*)from)); } template<> EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::complex<float> * addr) { EIGEN_PPC_PREFETCH(addr); }
template<> EIGEN_STRONG_INLINE Packet2cf ploadu<Packet2cf>(const std::complex<float>* from) { EIGEN_DEBUG_UNALIGNED_LOAD return Packet2cf(ploadu<Packet4f>((const float*)from)); }
template<> EIGEN_STRONG_INLINE Packet2cf ploaddup<Packet2cf>(const std::complex<float>* from)
{
return pset1<Packet2cf>(*from);
}
template<> EIGEN_STRONG_INLINE void pstore <std::complex<float> >(std::complex<float> * to, const Packet2cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((float*)to, from.v); }
template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<float> >(std::complex<float> * to, const Packet2cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((float*)to, from.v); }
template<> EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::complex<float> * addr) { vec_dstt((float *)addr, DST_CTRL(2,2,32), DST_CHAN); }
template<> EIGEN_STRONG_INLINE std::complex<float> pfirst<Packet2cf>(const Packet2cf& a) template<> EIGEN_STRONG_INLINE std::complex<float> pfirst<Packet2cf>(const Packet2cf& a)
{ {
@ -146,23 +142,23 @@ template<> EIGEN_STRONG_INLINE Packet2cf preverse(const Packet2cf& a)
template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(const Packet2cf& a) template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(const Packet2cf& a)
{ {
Packet4f b; Packet4f b;
b = (Packet4f) vec_sld(a.v, a.v, 8); b = vec_sld(a.v, a.v, 8);
b = padd(a.v, b); b = padd<Packet4f>(a.v, b);
return pfirst(Packet2cf(b)); return pfirst<Packet2cf>(Packet2cf(b));
} }
template<> EIGEN_STRONG_INLINE Packet2cf preduxp<Packet2cf>(const Packet2cf* vecs) template<> EIGEN_STRONG_INLINE Packet2cf preduxp<Packet2cf>(const Packet2cf* vecs)
{ {
Packet4f b1, b2; Packet4f b1, b2;
#ifdef _BIG_ENDIAN #ifdef _BIG_ENDIAN
b1 = (Packet4f) vec_sld(vecs[0].v, vecs[1].v, 8); b1 = vec_sld(vecs[0].v, vecs[1].v, 8);
b2 = (Packet4f) vec_sld(vecs[1].v, vecs[0].v, 8); b2 = vec_sld(vecs[1].v, vecs[0].v, 8);
#else #else
b1 = (Packet4f) vec_sld(vecs[1].v, vecs[0].v, 8); b1 = vec_sld(vecs[1].v, vecs[0].v, 8);
b2 = (Packet4f) vec_sld(vecs[0].v, vecs[1].v, 8); b2 = vec_sld(vecs[0].v, vecs[1].v, 8);
#endif #endif
b2 = (Packet4f) vec_sld(b2, b2, 8); b2 = vec_sld(b2, b2, 8);
b2 = padd(b1, b2); b2 = padd<Packet4f>(b1, b2);
return Packet2cf(b2); return Packet2cf(b2);
} }
@ -171,10 +167,10 @@ template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(const P
{ {
Packet4f b; Packet4f b;
Packet2cf prod; Packet2cf prod;
b = (Packet4f) vec_sld(a.v, a.v, 8); b = vec_sld(a.v, a.v, 8);
prod = pmul(a, Packet2cf(b)); prod = pmul<Packet2cf>(a, Packet2cf(b));
return pfirst(prod); return pfirst<Packet2cf>(prod);
} }
template<int Offset> template<int Offset>
@ -226,12 +222,30 @@ template<> struct conj_helper<Packet2cf, Packet2cf, true,true>
} }
}; };
template<> struct conj_helper<Packet4f, Packet2cf, false,false>
{
EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet4f& x, const Packet2cf& y, const Packet2cf& c) const
{ return padd(c, pmul(x,y)); }
EIGEN_STRONG_INLINE Packet2cf pmul(const Packet4f& x, const Packet2cf& y) const
{ return Packet2cf(internal::pmul<Packet4f>(x, y.v)); }
};
template<> struct conj_helper<Packet2cf, Packet4f, false,false>
{
EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet4f& y, const Packet2cf& c) const
{ return padd(c, pmul(x,y)); }
EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& x, const Packet4f& y) const
{ return Packet2cf(internal::pmul<Packet4f>(x.v, y)); }
};
template<> EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, const Packet2cf& b) template<> EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
{ {
// TODO optimize it for AltiVec // TODO optimize it for AltiVec
Packet2cf res = conj_helper<Packet2cf,Packet2cf,false,true>().pmul(a, b); Packet2cf res = conj_helper<Packet2cf,Packet2cf,false,true>().pmul(a, b);
Packet4f s = vec_madd(b.v, b.v, p4f_ZERO); Packet4f s = pmul<Packet4f>(b.v, b.v);
return Packet2cf(pdiv(res.v, vec_add(s,vec_perm(s, s, p16uc_COMPLEX32_REV)))); return Packet2cf(pdiv(res.v, padd<Packet4f>(s, vec_perm(s, s, p16uc_COMPLEX32_REV))));
} }
template<> EIGEN_STRONG_INLINE Packet2cf pcplxflip<Packet2cf>(const Packet2cf& x) template<> EIGEN_STRONG_INLINE Packet2cf pcplxflip<Packet2cf>(const Packet2cf& x)
@ -246,6 +260,12 @@ EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet2cf,2>& kernel)
kernel.packet[0].v = tmp; kernel.packet[0].v = tmp;
} }
template<> EIGEN_STRONG_INLINE Packet2cf pblend(const Selector<2>& ifPacket, const Packet2cf& thenPacket, const Packet2cf& elsePacket) {
Packet2cf result;
result.v = reinterpret_cast<Packet4f>(pblend<Packet2d>(ifPacket, reinterpret_cast<Packet2d>(thenPacket.v), reinterpret_cast<Packet2d>(elsePacket.v)));
return result;
}
//---------- double ---------- //---------- double ----------
#ifdef __VSX__ #ifdef __VSX__
struct Packet1cd struct Packet1cd
@ -303,10 +323,10 @@ template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<double>, Packet1c
to[1*stride] = af[1]; to[1*stride] = af[1];
} }
template<> EIGEN_STRONG_INLINE Packet1cd padd<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_add(a.v,b.v)); } template<> EIGEN_STRONG_INLINE Packet1cd padd<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(padd<Packet2d>(a.v,b.v)); }
template<> EIGEN_STRONG_INLINE Packet1cd psub<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_sub(a.v,b.v)); } template<> EIGEN_STRONG_INLINE Packet1cd psub<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(psub<Packet2d>(a.v,b.v)); }
template<> EIGEN_STRONG_INLINE Packet1cd pnegate(const Packet1cd& a) { return Packet1cd(pnegate(Packet2d(a.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((Packet2d)a.v, (Packet2d)p2ul_CONJ_XOR2)); } template<> EIGEN_STRONG_INLINE Packet1cd pconj(const Packet1cd& a) { return Packet1cd(pxor(a.v, reinterpret_cast<Packet2d>(p2ul_CONJ_XOR2))); }
template<> EIGEN_STRONG_INLINE Packet1cd pmul<Packet1cd>(const Packet1cd& a, const Packet1cd& b) template<> EIGEN_STRONG_INLINE Packet1cd pmul<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
{ {
@ -320,23 +340,20 @@ template<> EIGEN_STRONG_INLINE Packet1cd pmul<Packet1cd>(const Packet1cd& a, con
v1 = vec_madd(a_re, b.v, p2d_ZERO); v1 = vec_madd(a_re, b.v, p2d_ZERO);
// multiply a_im * b and get the conjugate result // multiply a_im * b and get the conjugate result
v2 = vec_madd(a_im, b.v, p2d_ZERO); v2 = vec_madd(a_im, b.v, p2d_ZERO);
v2 = (Packet2d) vec_sld((Packet4ui)v2, (Packet4ui)v2, 8); v2 = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4ui>(v2), reinterpret_cast<Packet4ui>(v2), 8));
v2 = (Packet2d) vec_xor((Packet2d)v2, (Packet2d) p2ul_CONJ_XOR1); v2 = pxor(v2, reinterpret_cast<Packet2d>(p2ul_CONJ_XOR1));
return Packet1cd(vec_add(v1, v2)); return Packet1cd(vec_add(v1, v2));
} }
template<> EIGEN_STRONG_INLINE Packet1cd pand <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_and(a.v,b.v)); } template<> EIGEN_STRONG_INLINE Packet1cd pand <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(pand(a.v,b.v)); }
template<> EIGEN_STRONG_INLINE Packet1cd por <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_or(a.v,b.v)); } template<> EIGEN_STRONG_INLINE Packet1cd por <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(por(a.v,b.v)); }
template<> EIGEN_STRONG_INLINE Packet1cd pxor <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_xor(a.v,b.v)); } template<> EIGEN_STRONG_INLINE Packet1cd pxor <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(pxor(a.v,b.v)); }
template<> EIGEN_STRONG_INLINE Packet1cd pandnot<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_and(a.v, vec_nor(b.v,b.v))); } template<> EIGEN_STRONG_INLINE Packet1cd pandnot<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(pandnot(a.v, b.v)); }
template<> EIGEN_STRONG_INLINE Packet1cd ploaddup<Packet1cd>(const std::complex<double>* from) template<> EIGEN_STRONG_INLINE Packet1cd ploaddup<Packet1cd>(const std::complex<double>* from) { return pset1<Packet1cd>(*from); }
{
return pset1<Packet1cd>(*from);
}
template<> EIGEN_STRONG_INLINE void prefetch<std::complex<double> >(const std::complex<double> * addr) { vec_dstt((long *)addr, DST_CTRL(2,2,32), DST_CHAN); } template<> EIGEN_STRONG_INLINE void prefetch<std::complex<double> >(const std::complex<double> * addr) { EIGEN_PPC_PREFETCH(addr); }
template<> EIGEN_STRONG_INLINE std::complex<double> pfirst<Packet1cd>(const Packet1cd& a) template<> EIGEN_STRONG_INLINE std::complex<double> pfirst<Packet1cd>(const Packet1cd& a)
{ {
@ -348,20 +365,10 @@ template<> EIGEN_STRONG_INLINE std::complex<double> pfirst<Packet1cd>(const Pac
template<> EIGEN_STRONG_INLINE Packet1cd preverse(const Packet1cd& a) { return a; } template<> EIGEN_STRONG_INLINE Packet1cd preverse(const Packet1cd& a) { return a; }
template<> EIGEN_STRONG_INLINE std::complex<double> predux<Packet1cd>(const Packet1cd& a) template<> EIGEN_STRONG_INLINE std::complex<double> predux<Packet1cd>(const Packet1cd& a) { return pfirst(a); }
{ template<> EIGEN_STRONG_INLINE Packet1cd preduxp<Packet1cd>(const Packet1cd* vecs) { return vecs[0]; }
return pfirst(a);
}
template<> EIGEN_STRONG_INLINE Packet1cd preduxp<Packet1cd>(const Packet1cd* vecs) template<> EIGEN_STRONG_INLINE std::complex<double> predux_mul<Packet1cd>(const Packet1cd& a) { return pfirst(a); }
{
return vecs[0];
}
template<> EIGEN_STRONG_INLINE std::complex<double> predux_mul<Packet1cd>(const Packet1cd& a)
{
return pfirst(a);
}
template<int Offset> template<int Offset>
struct palign_impl<Offset,Packet1cd> struct palign_impl<Offset,Packet1cd>
@ -405,6 +412,23 @@ template<> struct conj_helper<Packet1cd, Packet1cd, true,true>
return pconj(internal::pmul(a, b)); return pconj(internal::pmul(a, b));
} }
}; };
template<> struct conj_helper<Packet2d, Packet1cd, false,false>
{
EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet2d& x, const Packet1cd& y, const Packet1cd& c) const
{ return padd(c, pmul(x,y)); }
EIGEN_STRONG_INLINE Packet1cd pmul(const Packet2d& x, const Packet1cd& y) const
{ return Packet1cd(internal::pmul<Packet2d>(x, y.v)); }
};
template<> struct conj_helper<Packet1cd, Packet2d, false,false>
{
EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet2d& y, const Packet1cd& c) const
{ return padd(c, pmul(x,y)); }
EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& x, const Packet2d& y) const
{ return Packet1cd(internal::pmul<Packet2d>(x.v, y)); }
};
template<> EIGEN_STRONG_INLINE Packet1cd pdiv<Packet1cd>(const Packet1cd& a, const Packet1cd& b) template<> EIGEN_STRONG_INLINE Packet1cd pdiv<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
{ {