diff options
Diffstat (limited to 'Eigen/src/Core/arch/ZVector')
| -rw-r--r-- | Eigen/src/Core/arch/ZVector/Complex.h | 392 | ||||
| -rw-r--r-- | Eigen/src/Core/arch/ZVector/MathFunctions.h | 106 | ||||
| -rwxr-xr-x | Eigen/src/Core/arch/ZVector/PacketMath.h | 883 |
3 files changed, 812 insertions, 569 deletions
diff --git a/Eigen/src/Core/arch/ZVector/Complex.h b/Eigen/src/Core/arch/ZVector/Complex.h index d39d2d105..0b9b33d99 100644 --- a/Eigen/src/Core/arch/ZVector/Complex.h +++ b/Eigen/src/Core/arch/ZVector/Complex.h @@ -15,6 +15,10 @@ namespace Eigen { namespace internal { +#if !defined(__ARCH__) || (defined(__ARCH__) && __ARCH__ >= 12) +static Packet4ui p4ui_CONJ_XOR = { 0x00000000, 0x80000000, 0x00000000, 0x80000000 }; //vec_mergeh((Packet4ui)p4i_ZERO, (Packet4ui)p4f_MZERO); +#endif + static Packet2ul p2ul_CONJ_XOR1 = (Packet2ul) vec_sld((Packet4ui) p2d_ZERO_, (Packet4ui) p2l_ZERO, 8);//{ 0x8000000000000000, 0x0000000000000000 }; static Packet2ul p2ul_CONJ_XOR2 = (Packet2ul) vec_sld((Packet4ui) p2l_ZERO, (Packet4ui) p2d_ZERO_, 8);//{ 0x8000000000000000, 0x0000000000000000 }; @@ -29,10 +33,14 @@ struct Packet2cf { EIGEN_STRONG_INLINE Packet2cf() {} EIGEN_STRONG_INLINE explicit Packet2cf(const Packet4f& a) : v(a) {} +#if !defined(__ARCH__) || (defined(__ARCH__) && __ARCH__ < 12) union { Packet4f v; Packet1cd cd[2]; }; +#else + Packet4f v; +#endif }; template<> struct packet_traits<std::complex<float> > : default_packet_traits @@ -83,69 +91,33 @@ template<> struct packet_traits<std::complex<double> > : default_packet_traits }; }; -template<> struct unpacket_traits<Packet2cf> { typedef std::complex<float> type; enum {size=2, alignment=Aligned16}; typedef Packet2cf half; }; -template<> struct unpacket_traits<Packet1cd> { typedef std::complex<double> type; enum {size=1, alignment=Aligned16}; typedef Packet1cd half; }; +template<> struct unpacket_traits<Packet2cf> { typedef std::complex<float> type; enum {size=2, alignment=Aligned16, vectorizable=true, masked_load_available=false, masked_store_available=false}; typedef Packet2cf half; }; +template<> struct unpacket_traits<Packet1cd> { typedef std::complex<double> type; enum {size=1, alignment=Aligned16, vectorizable=true, masked_load_available=false, masked_store_available=false}; typedef Packet1cd half; }; /* Forward declaration */ EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet2cf,2>& kernel); -template<> EIGEN_STRONG_INLINE Packet2cf pload <Packet2cf>(const std::complex<float>* from) { EIGEN_DEBUG_ALIGNED_LOAD return Packet2cf(pload<Packet4f>((const float*)from)); } +/* complex<double> first */ template<> EIGEN_STRONG_INLINE Packet1cd pload <Packet1cd>(const std::complex<double>* from) { EIGEN_DEBUG_ALIGNED_LOAD return Packet1cd(pload<Packet2d>((const double*)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 Packet1cd ploadu<Packet1cd>(const std::complex<double>* from) { EIGEN_DEBUG_UNALIGNED_LOAD return Packet1cd(ploadu<Packet2d>((const double*)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 pstore <std::complex<double> >(std::complex<double> * to, const Packet1cd& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((double*)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 pstoreu<std::complex<double> >(std::complex<double> * to, const Packet1cd& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((double*)to, from.v); } template<> EIGEN_STRONG_INLINE Packet1cd pset1<Packet1cd>(const std::complex<double>& from) { /* here we really have to use unaligned loads :( */ return ploadu<Packet1cd>(&from); } -template<> EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<float>& from) -{ - Packet2cf res; - res.cd[0] = Packet1cd(vec_ld2f((const float *)&from)); - res.cd[1] = res.cd[0]; - return res; -} -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]; - af[0] = from[0*stride]; - af[1] = from[1*stride]; - return pload<Packet2cf>(af); -} template<> EIGEN_DEVICE_FUNC inline Packet1cd pgather<std::complex<double>, Packet1cd>(const std::complex<double>* from, Index stride EIGEN_UNUSED) { return pload<Packet1cd>(from); } -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]; - pstore<std::complex<float> >((std::complex<float> *) af, from); - to[0*stride] = af[0]; - to[1*stride] = af[1]; -} template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<double>, Packet1cd>(std::complex<double>* to, const Packet1cd& from, Index stride EIGEN_UNUSED) { pstore<std::complex<double> >(to, from); } - -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 Packet1cd padd<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(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 Packet1cd psub<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(a.v - b.v); } template<> EIGEN_STRONG_INLINE Packet1cd pnegate(const Packet1cd& a) { return Packet1cd(pnegate(Packet2d(a.v))); } -template<> EIGEN_STRONG_INLINE Packet2cf pnegate(const Packet2cf& a) { return Packet2cf(pnegate(Packet4f(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 Packet2cf pconj(const Packet2cf& a) -{ - Packet2cf res; - res.v.v4f[0] = pconj(Packet1cd(reinterpret_cast<Packet2d>(a.v.v4f[0]))).v; - res.v.v4f[1] = pconj(Packet1cd(reinterpret_cast<Packet2d>(a.v.v4f[1]))).v; - return res; -} - template<> EIGEN_STRONG_INLINE Packet1cd pmul<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { Packet2d a_re, a_im, v1, v2; @@ -163,27 +135,17 @@ template<> EIGEN_STRONG_INLINE Packet1cd pmul<Packet1cd>(const Packet1cd& a, con return Packet1cd(v1 + v2); } -template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, const Packet2cf& b) -{ - Packet2cf res; - res.v.v4f[0] = pmul(Packet1cd(reinterpret_cast<Packet2d>(a.v.v4f[0])), Packet1cd(reinterpret_cast<Packet2d>(b.v.v4f[0]))).v; - res.v.v4f[1] = pmul(Packet1cd(reinterpret_cast<Packet2d>(a.v.v4f[1])), Packet1cd(reinterpret_cast<Packet2d>(b.v.v4f[1]))).v; - return res; -} - -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 Packet2cf pand <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(pand<Packet4f>(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 Packet2cf por <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(por<Packet4f>(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 Packet2cf pxor <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(pxor<Packet4f>(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 Packet2cf pandnot<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(pandnot<Packet4f>(a.v,b.v)); } - +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 por <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_or(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 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 ploaddup<Packet1cd>(const std::complex<double>* from) { return pset1<Packet1cd>(*from); } -template<> EIGEN_STRONG_INLINE Packet2cf ploaddup<Packet2cf>(const std::complex<float>* from) { return pset1<Packet2cf>(*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<Packet2d>(eq, tmp); +} -template<> EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::complex<float> * addr) { EIGEN_ZVECTOR_PREFETCH(addr); } template<> EIGEN_STRONG_INLINE void prefetch<std::complex<double> >(const std::complex<double> * addr) { EIGEN_ZVECTOR_PREFETCH(addr); } template<> EIGEN_STRONG_INLINE std::complex<double> pfirst<Packet1cd>(const Packet1cd& a) @@ -193,157 +155,157 @@ template<> EIGEN_STRONG_INLINE std::complex<double> pfirst<Packet1cd>(const Pac return res; } -template<> EIGEN_STRONG_INLINE std::complex<float> pfirst<Packet2cf>(const Packet2cf& a) -{ - std::complex<float> EIGEN_ALIGN16 res[2]; - pstore<std::complex<float> >(res, a); - - return res[0]; -} template<> EIGEN_STRONG_INLINE Packet1cd preverse(const Packet1cd& a) { return a; } -template<> EIGEN_STRONG_INLINE Packet2cf preverse(const Packet2cf& a) +template<> EIGEN_STRONG_INLINE std::complex<double> predux<Packet1cd>(const Packet1cd& a) { - Packet2cf res; - res.cd[0] = a.cd[1]; - res.cd[1] = a.cd[0]; - return res; + return pfirst(a); } - -template<> EIGEN_STRONG_INLINE std::complex<double> predux<Packet1cd>(const Packet1cd& a) +template<> EIGEN_STRONG_INLINE std::complex<double> predux_mul<Packet1cd>(const Packet1cd& a) { return pfirst(a); } -template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(const Packet2cf& a) +EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet1cd,Packet2d) + +template<> EIGEN_STRONG_INLINE Packet1cd pdiv<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { - std::complex<float> res; - Packet1cd b = padd<Packet1cd>(a.cd[0], a.cd[1]); - vec_st2f(b.v, (float*)&res); - return res; + // TODO optimize it for AltiVec + Packet1cd res = pmul(a,pconj(b)); + Packet2d s = vec_madd(b.v, b.v, p2d_ZERO_); + return Packet1cd(pdiv(res.v, s + vec_perm(s, s, p16uc_REVERSE64))); } -template<> EIGEN_STRONG_INLINE Packet1cd preduxp<Packet1cd>(const Packet1cd* vecs) +EIGEN_STRONG_INLINE Packet1cd pcplxflip/*<Packet1cd>*/(const Packet1cd& x) { - return vecs[0]; + return Packet1cd(preverse(Packet2d(x.v))); } -template<> EIGEN_STRONG_INLINE Packet2cf preduxp<Packet2cf>(const Packet2cf* vecs) + +EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet1cd,2>& kernel) { - PacketBlock<Packet2cf,2> transpose; - transpose.packet[0] = vecs[0]; - transpose.packet[1] = vecs[1]; - ptranspose(transpose); + Packet2d tmp = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_HI); + kernel.packet[1].v = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_LO); + kernel.packet[0].v = tmp; +} - return padd<Packet2cf>(transpose.packet[0], transpose.packet[1]); -} +/* complex<float> follows */ +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 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 std::complex<double> predux_mul<Packet1cd>(const Packet1cd& a) +template<> EIGEN_STRONG_INLINE std::complex<float> pfirst<Packet2cf>(const Packet2cf& a) { - return pfirst(a); + std::complex<float> EIGEN_ALIGN16 res[2]; + pstore<std::complex<float> >(res, a); + + return res[0]; } -template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(const Packet2cf& a) + + +#if !defined(__ARCH__) || (defined(__ARCH__) && __ARCH__ < 12) +template<> EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<float>& from) { - std::complex<float> res; - Packet1cd b = pmul<Packet1cd>(a.cd[0], a.cd[1]); - vec_st2f(b.v, (float*)&res); + Packet2cf res; + res.cd[0] = Packet1cd(vec_ld2f((const float *)&from)); + res.cd[1] = res.cd[0]; return res; } - -template<int Offset> -struct palign_impl<Offset,Packet1cd> +#else +template<> EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<float>& from) { - static EIGEN_STRONG_INLINE void run(Packet1cd& /*first*/, const Packet1cd& /*second*/) - { - // FIXME is it sure we never have to align a Packet1cd? - // Even though a std::complex<double> has 16 bytes, it is not necessarily aligned on a 16 bytes boundary... - } -}; + Packet2cf res; + if((std::ptrdiff_t(&from) % 16) == 0) + res.v = pload<Packet4f>((const float *)&from); + else + res.v = ploadu<Packet4f>((const float *)&from); + res.v = vec_perm(res.v, res.v, p16uc_PSET64_HI); + return res; +} +#endif -template<int Offset> -struct palign_impl<Offset,Packet2cf> +template<> EIGEN_DEVICE_FUNC inline Packet2cf pgather<std::complex<float>, Packet2cf>(const std::complex<float>* from, Index stride) { - static EIGEN_STRONG_INLINE void run(Packet2cf& first, const Packet2cf& second) - { - if (Offset == 1) { - first.cd[0] = first.cd[1]; - first.cd[1] = second.cd[0]; - } - } -}; - -template<> struct conj_helper<Packet1cd, Packet1cd, false,true> + std::complex<float> EIGEN_ALIGN16 af[2]; + af[0] = from[0*stride]; + af[1] = from[1*stride]; + return pload<Packet2cf>(af); +} +template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<float>, Packet2cf>(std::complex<float>* to, const Packet2cf& from, Index stride) { - EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const - { return padd(pmul(x,y),c); } + std::complex<float> EIGEN_ALIGN16 af[2]; + pstore<std::complex<float> >((std::complex<float> *) af, from); + to[0*stride] = af[0]; + to[1*stride] = af[1]; +} - EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const - { - return internal::pmul(a, pconj(b)); - } -}; +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 psub<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(psub<Packet4f>(a.v, b.v)); } +template<> EIGEN_STRONG_INLINE Packet2cf pnegate(const Packet2cf& a) { return Packet2cf(pnegate(Packet4f(a.v))); } -template<> struct conj_helper<Packet1cd, Packet1cd, true,false> -{ - EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const - { return padd(pmul(x,y),c); } +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(por<Packet4f>(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(pandnot<Packet4f>(a.v,b.v)); } - EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const - { - return internal::pmul(pconj(a), b); - } -}; +template<> EIGEN_STRONG_INLINE Packet2cf ploaddup<Packet2cf>(const std::complex<float>* from) { return pset1<Packet2cf>(*from); } -template<> struct conj_helper<Packet1cd, Packet1cd, true,true> -{ - EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const - { return padd(pmul(x,y),c); } +template<> EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::complex<float> * addr) { EIGEN_ZVECTOR_PREFETCH(addr); } - EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const - { - return pconj(internal::pmul(a, b)); - } -}; -template<> struct conj_helper<Packet2cf, Packet2cf, false,true> -{ - EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const - { return padd(pmul(x,y),c); } +#if !defined(__ARCH__) || (defined(__ARCH__) && __ARCH__ < 12) - EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const - { - return internal::pmul(a, pconj(b)); - } -}; +template<> EIGEN_STRONG_INLINE Packet2cf pcmp_eq(const Packet2cf& a, const Packet2cf& b) { + Packet4f eq = pcmp_eq<Packet4f> (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<Packet2d>(eq.v4f[0], tmp1); + res.v.v4f[1] = pand<Packet2d>(eq.v4f[1], tmp2); + return res; +} -template<> struct conj_helper<Packet2cf, Packet2cf, true,false> +template<> EIGEN_STRONG_INLINE Packet2cf pconj(const Packet2cf& a) { - EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const - { return padd(pmul(x,y),c); } + Packet2cf res; + res.v.v4f[0] = pconj(Packet1cd(reinterpret_cast<Packet2d>(a.v.v4f[0]))).v; + res.v.v4f[1] = pconj(Packet1cd(reinterpret_cast<Packet2d>(a.v.v4f[1]))).v; + return res; +} - EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const - { - return internal::pmul(pconj(a), b); - } -}; +template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, const Packet2cf& b) +{ + Packet2cf res; + res.v.v4f[0] = pmul(Packet1cd(reinterpret_cast<Packet2d>(a.v.v4f[0])), Packet1cd(reinterpret_cast<Packet2d>(b.v.v4f[0]))).v; + res.v.v4f[1] = pmul(Packet1cd(reinterpret_cast<Packet2d>(a.v.v4f[1])), Packet1cd(reinterpret_cast<Packet2d>(b.v.v4f[1]))).v; + return res; +} -template<> struct conj_helper<Packet2cf, Packet2cf, true,true> +template<> EIGEN_STRONG_INLINE Packet2cf preverse(const Packet2cf& a) { - EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const - { return padd(pmul(x,y),c); } + Packet2cf res; + res.cd[0] = a.cd[1]; + res.cd[1] = a.cd[0]; + return res; +} - EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const - { - return pconj(internal::pmul(a, b)); - } -}; +template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(const Packet2cf& a) +{ + std::complex<float> res; + Packet1cd b = padd<Packet1cd>(a.cd[0], a.cd[1]); + vec_st2f(b.v, (float*)&res); + return res; +} -template<> EIGEN_STRONG_INLINE Packet1cd pdiv<Packet1cd>(const Packet1cd& a, const Packet1cd& b) +template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(const Packet2cf& a) { - // TODO optimize it for AltiVec - Packet1cd res = conj_helper<Packet1cd,Packet1cd,false,true>().pmul(a,b); - Packet2d s = vec_madd(b.v, b.v, p2d_ZERO_); - return Packet1cd(pdiv(res.v, s + vec_perm(s, s, p16uc_REVERSE64))); + std::complex<float> res; + Packet1cd b = pmul<Packet1cd>(a.cd[0], a.cd[1]); + vec_st2f(b.v, (float*)&res); + return res; } +EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet2cf,Packet4f) + template<> EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { // TODO optimize it for AltiVec @@ -353,11 +315,6 @@ template<> EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, con return res; } -EIGEN_STRONG_INLINE Packet1cd pcplxflip/*<Packet1cd>*/(const Packet1cd& x) -{ - return Packet1cd(preverse(Packet2d(x.v))); -} - EIGEN_STRONG_INLINE Packet2cf pcplxflip/*<Packet2cf>*/(const Packet2cf& x) { Packet2cf res; @@ -366,13 +323,6 @@ EIGEN_STRONG_INLINE Packet2cf pcplxflip/*<Packet2cf>*/(const Packet2cf& x) return res; } -EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet1cd,2>& kernel) -{ - Packet2d tmp = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_HI); - kernel.packet[1].v = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_LO); - kernel.packet[0].v = tmp; -} - EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet2cf,2>& kernel) { Packet1cd tmp = kernel.packet[0].cd[1]; @@ -386,6 +336,88 @@ template<> EIGEN_STRONG_INLINE Packet2cf pblend(const Selector<2>& ifPacket, con result.v = pblend<Packet4f>(ifPacket4, thenPacket.v, elsePacket.v); 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<Packet4f>(eq, tmp); +} +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) +{ + Packet4f a_re, a_im, prod, prod_im; + + // Permute and multiply the real parts of a and b + a_re = vec_perm(a.v, a.v, p16uc_PSET32_WODD); + + // Get the imaginary parts of a + a_im = vec_perm(a.v, a.v, p16uc_PSET32_WEVEN); + + // multiply a_im * b and get the conjugate result + prod_im = a_im * b.v; + prod_im = pxor<Packet4f>(prod_im, reinterpret_cast<Packet4f>(p4ui_CONJ_XOR)); + // permute back to a proper order + prod_im = vec_perm(prod_im, prod_im, p16uc_COMPLEX32_REV); + + // multiply a_re * b, add prod_im + prod = pmadd<Packet4f>(a_re, b.v, prod_im); + + return Packet2cf(prod); +} + +template<> EIGEN_STRONG_INLINE Packet2cf preverse(const Packet2cf& a) +{ + Packet4f rev_a; + rev_a = vec_perm(a.v, a.v, p16uc_COMPLEX32_REV2); + return Packet2cf(rev_a); +} + +template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(const Packet2cf& a) +{ + Packet4f b; + b = vec_sld(a.v, a.v, 8); + b = padd<Packet4f>(a.v, b); + return pfirst<Packet2cf>(Packet2cf(b)); +} + +template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(const Packet2cf& a) +{ + Packet4f b; + Packet2cf prod; + b = vec_sld(a.v, a.v, 8); + prod = pmul<Packet2cf>(a, Packet2cf(b)); + + return pfirst<Packet2cf>(prod); +} + +EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet2cf,Packet4f) + +template<> EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, const Packet2cf& b) +{ + // TODO optimize it for AltiVec + Packet2cf res = pmul(a, pconj(b)); + Packet4f s = pmul<Packet4f>(b.v, b.v); + 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) +{ + return Packet2cf(vec_perm(x.v, x.v, p16uc_COMPLEX32_REV)); +} + +EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet2cf,2>& kernel) +{ + Packet4f tmp = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_HI); + kernel.packet[1].v = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_LO); + 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; +} +#endif } // end namespace internal diff --git a/Eigen/src/Core/arch/ZVector/MathFunctions.h b/Eigen/src/Core/arch/ZVector/MathFunctions.h index 5c7aa7256..1635e128c 100644 --- a/Eigen/src/Core/arch/ZVector/MathFunctions.h +++ b/Eigen/src/Core/arch/ZVector/MathFunctions.h @@ -20,6 +20,50 @@ namespace Eigen { namespace internal { +#if !defined(__ARCH__) || (defined(__ARCH__) && __ARCH__ >= 12) +static _EIGEN_DECLARE_CONST_Packet4f(1 , 1.0f); +static _EIGEN_DECLARE_CONST_Packet4f(half, 0.5f); +static _EIGEN_DECLARE_CONST_Packet4i(0x7f, 0x7f); +static _EIGEN_DECLARE_CONST_Packet4i(23, 23); + +static _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(inv_mant_mask, ~0x7f800000); + +/* the smallest non denormalized float number */ +static _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(min_norm_pos, 0x00800000); +static _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(minus_inf, 0xff800000); // -1.f/0.f +static _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(minus_nan, 0xffffffff); + +/* natural logarithm computed for 4 simultaneous float + return NaN for x <= 0 +*/ +static _EIGEN_DECLARE_CONST_Packet4f(cephes_SQRTHF, 0.707106781186547524f); +static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p0, 7.0376836292E-2f); +static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p1, - 1.1514610310E-1f); +static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p2, 1.1676998740E-1f); +static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p3, - 1.2420140846E-1f); +static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p4, + 1.4249322787E-1f); +static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p5, - 1.6668057665E-1f); +static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p6, + 2.0000714765E-1f); +static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p7, - 2.4999993993E-1f); +static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p8, + 3.3333331174E-1f); +static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_q1, -2.12194440e-4f); +static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_q2, 0.693359375f); + +static _EIGEN_DECLARE_CONST_Packet4f(exp_hi, 88.3762626647950f); +static _EIGEN_DECLARE_CONST_Packet4f(exp_lo, -88.3762626647949f); + +static _EIGEN_DECLARE_CONST_Packet4f(cephes_LOG2EF, 1.44269504088896341f); +static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_C1, 0.693359375f); +static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_C2, -2.12194440e-4f); + +static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p0, 1.9875691500E-4f); +static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p1, 1.3981999507E-3f); +static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p2, 8.3334519073E-3f); +static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p3, 4.1665795894E-2f); +static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p4, 1.6666665459E-1f); +static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p5, 5.0000001201E-1f); +#endif + static _EIGEN_DECLARE_CONST_Packet2d(1 , 1.0); static _EIGEN_DECLARE_CONST_Packet2d(2 , 2.0); static _EIGEN_DECLARE_CONST_Packet2d(half, 0.5); @@ -93,43 +137,95 @@ Packet2d pexp<Packet2d>(const Packet2d& _x) } template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED -Packet4f pexp<Packet4f>(const Packet4f& x) +Packet4f pexp<Packet4f>(const Packet4f& _x) { +#if !defined(__ARCH__) || (defined(__ARCH__) && __ARCH__ >= 12) + Packet4f x = _x; + + Packet4f tmp, fx; + Packet4i emm0; + + // clamp x + x = pmax(pmin(x, p4f_exp_hi), p4f_exp_lo); + + // express exp(x) as exp(g + n*log(2)) + fx = pmadd(x, p4f_cephes_LOG2EF, p4f_half); + + fx = pfloor(fx); + + tmp = pmul(fx, p4f_cephes_exp_C1); + Packet4f z = pmul(fx, p4f_cephes_exp_C2); + x = psub(x, tmp); + x = psub(x, z); + + z = pmul(x,x); + + Packet4f y = p4f_cephes_exp_p0; + y = pmadd(y, x, p4f_cephes_exp_p1); + y = pmadd(y, x, p4f_cephes_exp_p2); + y = pmadd(y, x, p4f_cephes_exp_p3); + y = pmadd(y, x, p4f_cephes_exp_p4); + y = pmadd(y, x, p4f_cephes_exp_p5); + y = pmadd(y, z, x); + y = padd(y, p4f_1); + + // build 2^n + emm0 = (Packet4i){ (int)fx[0], (int)fx[1], (int)fx[2], (int)fx[3] }; + emm0 = emm0 + p4i_0x7f; + emm0 = emm0 << reinterpret_cast<Packet4i>(p4i_23); + + return pmax(pmul(y, reinterpret_cast<Packet4f>(emm0)), _x); +#else Packet4f res; - res.v4f[0] = pexp<Packet2d>(x.v4f[0]); - res.v4f[1] = pexp<Packet2d>(x.v4f[1]); + res.v4f[0] = pexp<Packet2d>(_x.v4f[0]); + res.v4f[1] = pexp<Packet2d>(_x.v4f[1]); return res; +#endif } template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet2d psqrt<Packet2d>(const Packet2d& x) { - return __builtin_s390_vfsqdb(x); + return vec_sqrt(x); } template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet4f psqrt<Packet4f>(const Packet4f& x) { Packet4f res; +#if !defined(__ARCH__) || (defined(__ARCH__) && __ARCH__ >= 12) + res = vec_sqrt(x); +#else res.v4f[0] = psqrt<Packet2d>(x.v4f[0]); res.v4f[1] = psqrt<Packet2d>(x.v4f[1]); +#endif return res; } template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet2d prsqrt<Packet2d>(const Packet2d& x) { - // Unfortunately we can't use the much faster mm_rqsrt_pd since it only provides an approximation. return pset1<Packet2d>(1.0) / psqrt<Packet2d>(x); } template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet4f prsqrt<Packet4f>(const Packet4f& x) { Packet4f res; +#if !defined(__ARCH__) || (defined(__ARCH__) && __ARCH__ >= 12) + res = pset1<Packet4f>(1.0) / psqrt<Packet4f>(x); +#else res.v4f[0] = prsqrt<Packet2d>(x.v4f[0]); res.v4f[1] = prsqrt<Packet2d>(x.v4f[1]); +#endif return res; } +// Hyperbolic Tangent function. +template <> +EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet4f +ptanh<Packet4f>(const Packet4f& x) { + return internal::generic_fast_tanh_float(x); +} + } // end namespace internal } // end namespace Eigen diff --git a/Eigen/src/Core/arch/ZVector/PacketMath.h b/Eigen/src/Core/arch/ZVector/PacketMath.h index 57b01fc63..1f55a90a5 100755 --- a/Eigen/src/Core/arch/ZVector/PacketMath.h +++ b/Eigen/src/Core/arch/ZVector/PacketMath.h @@ -10,26 +10,20 @@ #ifndef EIGEN_PACKET_MATH_ZVECTOR_H #define EIGEN_PACKET_MATH_ZVECTOR_H -#include <stdint.h> - namespace Eigen { namespace internal { #ifndef EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD -#define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 4 +#define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 16 #endif #ifndef EIGEN_HAS_SINGLE_INSTRUCTION_MADD #define EIGEN_HAS_SINGLE_INSTRUCTION_MADD #endif -#ifndef EIGEN_HAS_SINGLE_INSTRUCTION_CJMADD -#define EIGEN_HAS_SINGLE_INSTRUCTION_CJMADD -#endif - #ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS -#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 16 +#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 32 #endif typedef __vector int Packet4i; @@ -41,21 +35,30 @@ typedef __vector double Packet2d; typedef __vector unsigned long long Packet2ul; typedef __vector long long Packet2l; +// Z14 has builtin support for float vectors +#if !defined(__ARCH__) || (defined(__ARCH__) && __ARCH__ >= 12) +typedef __vector float Packet4f; +#else typedef struct { Packet2d v4f[2]; } Packet4f; +#endif typedef union { - int32_t i[4]; - uint32_t ui[4]; - int64_t l[2]; - uint64_t ul[2]; + numext::int32_t i[4]; + numext::uint32_t ui[4]; + numext::int64_t l[2]; + numext::uint64_t ul[2]; double d[2]; + float f[4]; Packet4i v4i; Packet4ui v4ui; Packet2l v2l; Packet2ul v2ul; Packet2d v2d; +#if !defined(__ARCH__) || (defined(__ARCH__) && __ARCH__ >= 12) + Packet4f v4f; +#endif } Packet; // We don't want to write the same code all the time, but we need to reuse the constants @@ -80,15 +83,31 @@ typedef union { Packet2l p2l_##NAME = pset1<Packet2l>(X) // These constants are endian-agnostic -//static _EIGEN_DECLARE_CONST_FAST_Packet4i(ZERO, 0); //{ 0, 0, 0, 0,} +static _EIGEN_DECLARE_CONST_FAST_Packet4i(ZERO, 0); //{ 0, 0, 0, 0,} static _EIGEN_DECLARE_CONST_FAST_Packet4i(ONE, 1); //{ 1, 1, 1, 1} static _EIGEN_DECLARE_CONST_FAST_Packet2d(ZERO, 0); static _EIGEN_DECLARE_CONST_FAST_Packet2l(ZERO, 0); static _EIGEN_DECLARE_CONST_FAST_Packet2l(ONE, 1); -static Packet2d p2d_ONE = { 1.0, 1.0 }; -static Packet2d p2d_ZERO_ = { -0.0, -0.0 }; +static Packet2d p2d_ONE = { 1.0, 1.0 }; +static Packet2d p2d_ZERO_ = { numext::bit_cast<double>0x8000000000000000ull), + numext::bit_cast<double>0x8000000000000000ull) }; + +#if !defined(__ARCH__) || (defined(__ARCH__) && __ARCH__ >= 12) +#define _EIGEN_DECLARE_CONST_FAST_Packet4f(NAME,X) \ + Packet4f p4f_##NAME = reinterpret_cast<Packet4f>(vec_splat_s32(X)) + +#define _EIGEN_DECLARE_CONST_Packet4f(NAME,X) \ + Packet4f p4f_##NAME = pset1<Packet4f>(X) + +#define _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(NAME,X) \ + const Packet4f p4f_##NAME = reinterpret_cast<Packet4f>(pset1<Packet4i>(X)) + +static _EIGEN_DECLARE_CONST_FAST_Packet4f(ZERO, 0); //{ 0.0, 0.0, 0.0, 0.0} +static _EIGEN_DECLARE_CONST_FAST_Packet4i(MINUS1,-1); //{ -1, -1, -1, -1} +static Packet4f p4f_MZERO = { 0x80000000, 0x80000000, 0x80000000, 0x80000000}; +#endif static Packet4i p4i_COUNTDOWN = { 0, 1, 2, 3 }; static Packet4f p4f_COUNTDOWN = { 0.0, 1.0, 2.0, 3.0 }; @@ -120,9 +139,9 @@ static Packet16uc p16uc_TRANSPOSE64_LO = vec_add(p16uc_PSET64_LO, p16uc_HALF64_0 static Packet16uc p16uc_TRANSPOSE64_HI = { 0,1,2,3, 4,5,6,7, 16,17,18,19, 20,21,22,23}; static Packet16uc p16uc_TRANSPOSE64_LO = { 8,9,10,11, 12,13,14,15, 24,25,26,27, 28,29,30,31}; -//static Packet16uc p16uc_COMPLEX32_REV = vec_sld(p16uc_REVERSE32, p16uc_REVERSE32, 8); //{ 4,5,6,7, 0,1,2,3, 12,13,14,15, 8,9,10,11 }; +static Packet16uc p16uc_COMPLEX32_REV = vec_sld(p16uc_REVERSE32, p16uc_REVERSE32, 8); //{ 4,5,6,7, 0,1,2,3, 12,13,14,15, 8,9,10,11 }; -//static Packet16uc p16uc_COMPLEX32_REV2 = vec_sld(p16uc_FORWARD, p16uc_FORWARD, 8); //{ 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 }; +static Packet16uc p16uc_COMPLEX32_REV2 = vec_sld(p16uc_FORWARD, p16uc_FORWARD, 8); //{ 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 }; #if EIGEN_HAS_BUILTIN(__builtin_prefetch) || EIGEN_COMP_GNUC @@ -149,29 +168,31 @@ template<> struct packet_traits<int> : default_packet_traits }; }; -template<> struct packet_traits<float> : default_packet_traits -{ +template <> +struct packet_traits<float> : default_packet_traits { typedef Packet4f type; typedef Packet4f half; enum { Vectorizable = 1, AlignedOnScalar = 1, - size=4, + size = 4, HasHalfPacket = 0, - HasAdd = 1, - HasSub = 1, - HasMul = 1, - HasDiv = 1, - HasMin = 1, - HasMax = 1, - HasAbs = 1, - HasSin = 0, - HasCos = 0, - HasLog = 0, - HasExp = 1, + HasAdd = 1, + HasSub = 1, + HasMul = 1, + HasDiv = 1, + HasMin = 1, + HasMax = 1, + HasAbs = 1, + HasSin = 0, + HasCos = 0, + HasLog = 0, + HasExp = 1, HasSqrt = 1, HasRsqrt = 1, + HasTanh = 1, + HasErf = 1, HasRound = 1, HasFloor = 1, HasCeil = 1, @@ -211,9 +232,9 @@ template<> struct packet_traits<double> : default_packet_traits }; }; -template<> struct unpacket_traits<Packet4i> { typedef int type; enum {size=4, alignment=Aligned16}; typedef Packet4i half; }; -template<> struct unpacket_traits<Packet4f> { typedef float type; enum {size=4, alignment=Aligned16}; typedef Packet4f half; }; -template<> struct unpacket_traits<Packet2d> { typedef double type; enum {size=2, alignment=Aligned16}; typedef Packet2d half; }; +template<> struct unpacket_traits<Packet4i> { typedef int type; enum {size=4, alignment=Aligned16, vectorizable=true, masked_load_available=false, masked_store_available=false}; typedef Packet4i half; }; +template<> struct unpacket_traits<Packet4f> { typedef float type; enum {size=4, alignment=Aligned16, vectorizable=true, masked_load_available=false, masked_store_available=false}; typedef Packet4f half; }; +template<> struct unpacket_traits<Packet2d> { typedef double type; enum {size=2, alignment=Aligned16, vectorizable=true, masked_load_available=false, masked_store_available=false}; typedef Packet2d half; }; /* Forward declaration */ EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet4f,4>& kernel); @@ -258,82 +279,15 @@ inline std::ostream & operator <<(std::ostream & s, const Packet2d & v) return s; } -/* Helper function to simulate a vec_splat_packet4f - */ -template<int element> EIGEN_STRONG_INLINE Packet4f vec_splat_packet4f(const Packet4f& from) +#if !defined(__ARCH__) || (defined(__ARCH__) && __ARCH__ >= 12) +inline std::ostream & operator <<(std::ostream & s, const Packet4f & v) { - Packet4f splat; - switch (element) { - case 0: - splat.v4f[0] = vec_splat(from.v4f[0], 0); - splat.v4f[1] = splat.v4f[0]; - break; - case 1: - splat.v4f[0] = vec_splat(from.v4f[0], 1); - splat.v4f[1] = splat.v4f[0]; - break; - case 2: - splat.v4f[0] = vec_splat(from.v4f[1], 0); - splat.v4f[1] = splat.v4f[0]; - break; - case 3: - splat.v4f[0] = vec_splat(from.v4f[1], 1); - splat.v4f[1] = splat.v4f[0]; - break; - } - return splat; + Packet vt; + vt.v4f = v; + s << vt.f[0] << ", " << vt.f[1] << ", " << vt.f[2] << ", " << vt.f[3]; + return s; } - -template<int Offset> -struct palign_impl<Offset,Packet4i> -{ - static EIGEN_STRONG_INLINE void run(Packet4i& first, const Packet4i& second) - { - switch (Offset % 4) { - case 1: - first = vec_sld(first, second, 4); break; - case 2: - first = vec_sld(first, second, 8); break; - case 3: - first = vec_sld(first, second, 12); break; - } - } -}; - -/* This is a tricky one, we have to translate float alignment to vector elements of sizeof double - */ -template<int Offset> -struct palign_impl<Offset,Packet4f> -{ - static EIGEN_STRONG_INLINE void run(Packet4f& first, const Packet4f& second) - { - switch (Offset % 4) { - case 1: - first.v4f[0] = vec_sld(first.v4f[0], first.v4f[1], 8); - first.v4f[1] = vec_sld(first.v4f[1], second.v4f[0], 8); - break; - case 2: - first.v4f[0] = first.v4f[1]; - first.v4f[1] = second.v4f[0]; - break; - case 3: - first.v4f[0] = vec_sld(first.v4f[1], second.v4f[0], 8); - first.v4f[1] = vec_sld(second.v4f[0], second.v4f[1], 8); - break; - } - } -}; - - -template<int Offset> -struct palign_impl<Offset,Packet2d> -{ - static EIGEN_STRONG_INLINE void run(Packet2d& first, const Packet2d& second) - { - if (Offset == 1) - first = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(first), reinterpret_cast<Packet4i>(second), 8)); - } -}; +#endif template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int* from) { @@ -344,16 +298,6 @@ template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int* from) return vfrom->v4i; } -template<> EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(const float* from) -{ - // FIXME: No intrinsic yet - EIGEN_DEBUG_ALIGNED_LOAD - Packet4f vfrom; - vfrom.v4f[0] = vec_ld2f(&from[0]); - vfrom.v4f[1] = vec_ld2f(&from[2]); - return vfrom; -} - template<> EIGEN_STRONG_INLINE Packet2d pload<Packet2d>(const double* from) { // FIXME: No intrinsic yet @@ -372,15 +316,6 @@ template<> EIGEN_STRONG_INLINE void pstore<int>(int* to, const Packet4i& f vto->v4i = from; } -template<> EIGEN_STRONG_INLINE void pstore<float>(float* to, const Packet4f& from) -{ - // FIXME: No intrinsic yet - EIGEN_DEBUG_ALIGNED_STORE - vec_st2f(from.v4f[0], &to[0]); - vec_st2f(from.v4f[1], &to[2]); -} - - template<> EIGEN_STRONG_INLINE void pstore<double>(double* to, const Packet2d& from) { // FIXME: No intrinsic yet @@ -397,13 +332,6 @@ template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int& from) template<> EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(const double& from) { return vec_splats(from); } -template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float& from) -{ - Packet4f to; - to.v4f[0] = pset1<Packet2d>(static_cast<const double&>(from)); - to.v4f[1] = to.v4f[0]; - return to; -} template<> EIGEN_STRONG_INLINE void pbroadcast4<Packet4i>(const int *a, @@ -417,17 +345,6 @@ pbroadcast4<Packet4i>(const int *a, } template<> EIGEN_STRONG_INLINE void -pbroadcast4<Packet4f>(const float *a, - Packet4f& a0, Packet4f& a1, Packet4f& a2, Packet4f& a3) -{ - a3 = pload<Packet4f>(a); - a0 = vec_splat_packet4f<0>(a3); - a1 = vec_splat_packet4f<1>(a3); - a2 = vec_splat_packet4f<2>(a3); - a3 = vec_splat_packet4f<3>(a3); -} - -template<> EIGEN_STRONG_INLINE void pbroadcast4<Packet2d>(const double *a, Packet2d& a0, Packet2d& a1, Packet2d& a2, Packet2d& a3) { @@ -449,16 +366,6 @@ template<> EIGEN_DEVICE_FUNC inline Packet4i pgather<int, Packet4i>(const int* f return pload<Packet4i>(ai); } -template<> EIGEN_DEVICE_FUNC inline Packet4f pgather<float, Packet4f>(const float* from, Index stride) -{ - float EIGEN_ALIGN16 ai[4]; - ai[0] = from[0*stride]; - ai[1] = from[1*stride]; - ai[2] = from[2*stride]; - ai[3] = from[3*stride]; - return pload<Packet4f>(ai); -} - template<> EIGEN_DEVICE_FUNC inline Packet2d pgather<double, Packet2d>(const double* from, Index stride) { double EIGEN_ALIGN16 af[2]; @@ -477,6 +384,269 @@ template<> EIGEN_DEVICE_FUNC inline void pscatter<int, Packet4i>(int* to, const to[3*stride] = ai[3]; } +template<> EIGEN_DEVICE_FUNC inline void pscatter<double, Packet2d>(double* to, const Packet2d& from, Index stride) +{ + double EIGEN_ALIGN16 af[2]; + pstore<double>(af, from); + to[0*stride] = af[0]; + to[1*stride] = af[1]; +} + +template<> EIGEN_STRONG_INLINE Packet4i padd<Packet4i>(const Packet4i& a, const Packet4i& b) { return (a + b); } +template<> EIGEN_STRONG_INLINE Packet2d padd<Packet2d>(const Packet2d& a, const Packet2d& b) { return (a + b); } + +template<> EIGEN_STRONG_INLINE Packet4i psub<Packet4i>(const Packet4i& a, const Packet4i& b) { return (a - b); } +template<> EIGEN_STRONG_INLINE Packet2d psub<Packet2d>(const Packet2d& a, const Packet2d& b) { return (a - b); } + +template<> EIGEN_STRONG_INLINE Packet4i pmul<Packet4i>(const Packet4i& a, const Packet4i& b) { return (a * b); } +template<> EIGEN_STRONG_INLINE Packet2d pmul<Packet2d>(const Packet2d& a, const Packet2d& b) { return (a * b); } + +template<> EIGEN_STRONG_INLINE Packet4i pdiv<Packet4i>(const Packet4i& a, const Packet4i& b) { return (a / b); } +template<> EIGEN_STRONG_INLINE Packet2d pdiv<Packet2d>(const Packet2d& a, const Packet2d& b) { return (a / b); } + +template<> EIGEN_STRONG_INLINE Packet4i pnegate(const Packet4i& a) { return (-a); } +template<> EIGEN_STRONG_INLINE Packet2d pnegate(const Packet2d& a) { return (-a); } + +template<> EIGEN_STRONG_INLINE Packet4i pconj(const Packet4i& a) { return a; } +template<> EIGEN_STRONG_INLINE Packet2d pconj(const Packet2d& a) { return a; } + +template<> EIGEN_STRONG_INLINE Packet4i pmadd(const Packet4i& a, const Packet4i& b, const Packet4i& c) { return padd<Packet4i>(pmul<Packet4i>(a, b), c); } +template<> EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& b, const Packet2d& c) { return vec_madd(a, b, c); } + +template<> EIGEN_STRONG_INLINE Packet4i plset<Packet4i>(const int& a) { return padd<Packet4i>(pset1<Packet4i>(a), p4i_COUNTDOWN); } +template<> EIGEN_STRONG_INLINE Packet2d plset<Packet2d>(const double& a) { return padd<Packet2d>(pset1<Packet2d>(a), p2d_COUNTDOWN); } + +template<> EIGEN_STRONG_INLINE Packet4i pmin<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_min(a, b); } +template<> EIGEN_STRONG_INLINE Packet2d pmin<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_min(a, b); } + +template<> EIGEN_STRONG_INLINE Packet4i pmax<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_max(a, b); } +template<> EIGEN_STRONG_INLINE Packet2d pmax<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_max(a, b); } + +template<> EIGEN_STRONG_INLINE Packet4i pand<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_and(a, b); } +template<> EIGEN_STRONG_INLINE Packet2d pand<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_and(a, b); } + +template<> EIGEN_STRONG_INLINE Packet4i por<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_or(a, b); } +template<> EIGEN_STRONG_INLINE Packet2d por<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_or(a, b); } + +template<> EIGEN_STRONG_INLINE Packet4i pxor<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_xor(a, b); } +template<> EIGEN_STRONG_INLINE Packet2d pxor<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_xor(a, b); } + +template<> EIGEN_STRONG_INLINE Packet4i pandnot<Packet4i>(const Packet4i& a, const Packet4i& b) { return pand<Packet4i>(a, vec_nor(b, b)); } +template<> EIGEN_STRONG_INLINE Packet2d pandnot<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_and(a, vec_nor(b, b)); } + +template<> EIGEN_STRONG_INLINE Packet2d pround<Packet2d>(const Packet2d& a) { return vec_round(a); } +template<> EIGEN_STRONG_INLINE Packet2d pceil<Packet2d>(const Packet2d& a) { return vec_ceil(a); } +template<> EIGEN_STRONG_INLINE Packet2d pfloor<Packet2d>(const Packet2d& a) { return vec_floor(a); } + +template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from) { return pload<Packet4i>(from); } +template<> EIGEN_STRONG_INLINE Packet2d ploadu<Packet2d>(const double* from) { return pload<Packet2d>(from); } + + +template<> EIGEN_STRONG_INLINE Packet4i ploaddup<Packet4i>(const int* from) +{ + Packet4i p = pload<Packet4i>(from); + return vec_perm(p, p, p16uc_DUPLICATE32_HI); +} + +template<> EIGEN_STRONG_INLINE Packet2d ploaddup<Packet2d>(const double* from) +{ + Packet2d p = pload<Packet2d>(from); + return vec_perm(p, p, p16uc_PSET64_HI); +} + +template<> EIGEN_STRONG_INLINE void pstoreu<int>(int* to, const Packet4i& from) { pstore<int>(to, from); } +template<> EIGEN_STRONG_INLINE void pstoreu<double>(double* to, const Packet2d& from) { pstore<double>(to, from); } + +template<> EIGEN_STRONG_INLINE void prefetch<int>(const int* addr) { EIGEN_ZVECTOR_PREFETCH(addr); } +template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { EIGEN_ZVECTOR_PREFETCH(addr); } + +template<> EIGEN_STRONG_INLINE int pfirst<Packet4i>(const Packet4i& a) { int EIGEN_ALIGN16 x[4]; pstore(x, a); return x[0]; } +template<> EIGEN_STRONG_INLINE double pfirst<Packet2d>(const Packet2d& a) { double EIGEN_ALIGN16 x[2]; pstore(x, a); return x[0]; } + +template<> EIGEN_STRONG_INLINE Packet4i preverse(const Packet4i& a) +{ + return reinterpret_cast<Packet4i>(vec_perm(reinterpret_cast<Packet16uc>(a), reinterpret_cast<Packet16uc>(a), p16uc_REVERSE32)); +} + +template<> EIGEN_STRONG_INLINE Packet2d preverse(const Packet2d& a) +{ + return reinterpret_cast<Packet2d>(vec_perm(reinterpret_cast<Packet16uc>(a), reinterpret_cast<Packet16uc>(a), p16uc_REVERSE64)); +} + +template<> EIGEN_STRONG_INLINE Packet4i pabs<Packet4i>(const Packet4i& a) { return vec_abs(a); } +template<> EIGEN_STRONG_INLINE Packet2d pabs<Packet2d>(const Packet2d& a) { return vec_abs(a); } + +template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a) +{ + Packet4i b, sum; + b = vec_sld(a, a, 8); + sum = padd<Packet4i>(a, b); + b = vec_sld(sum, sum, 4); + sum = padd<Packet4i>(sum, b); + return pfirst(sum); +} + +template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a) +{ + Packet2d b, sum; + b = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(a), reinterpret_cast<Packet4i>(a), 8)); + sum = padd<Packet2d>(a, b); + return pfirst(sum); +} + +// Other reduction functions: +// mul +template<> EIGEN_STRONG_INLINE int predux_mul<Packet4i>(const Packet4i& a) +{ + EIGEN_ALIGN16 int aux[4]; + pstore(aux, a); + return aux[0] * aux[1] * aux[2] * aux[3]; +} + +template<> EIGEN_STRONG_INLINE double predux_mul<Packet2d>(const Packet2d& a) +{ + return pfirst(pmul(a, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(a), reinterpret_cast<Packet4i>(a), 8)))); +} + +// min +template<> EIGEN_STRONG_INLINE int predux_min<Packet4i>(const Packet4i& a) +{ + Packet4i b, res; + b = pmin<Packet4i>(a, vec_sld(a, a, 8)); + res = pmin<Packet4i>(b, vec_sld(b, b, 4)); + return pfirst(res); +} + +template<> EIGEN_STRONG_INLINE double predux_min<Packet2d>(const Packet2d& a) +{ + return pfirst(pmin<Packet2d>(a, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(a), reinterpret_cast<Packet4i>(a), 8)))); +} + +// max +template<> EIGEN_STRONG_INLINE int predux_max<Packet4i>(const Packet4i& a) +{ + Packet4i b, res; + b = pmax<Packet4i>(a, vec_sld(a, a, 8)); + res = pmax<Packet4i>(b, vec_sld(b, b, 4)); + return pfirst(res); +} + +// max +template<> EIGEN_STRONG_INLINE double predux_max<Packet2d>(const Packet2d& a) +{ + return pfirst(pmax<Packet2d>(a, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(a), reinterpret_cast<Packet4i>(a), 8)))); +} + +EIGEN_DEVICE_FUNC inline void +ptranspose(PacketBlock<Packet4i,4>& kernel) { + Packet4i t0 = vec_mergeh(kernel.packet[0], kernel.packet[2]); + Packet4i t1 = vec_mergel(kernel.packet[0], kernel.packet[2]); + Packet4i t2 = vec_mergeh(kernel.packet[1], kernel.packet[3]); + Packet4i t3 = vec_mergel(kernel.packet[1], kernel.packet[3]); + kernel.packet[0] = vec_mergeh(t0, t2); + kernel.packet[1] = vec_mergel(t0, t2); + kernel.packet[2] = vec_mergeh(t1, t3); + kernel.packet[3] = vec_mergel(t1, t3); +} + +EIGEN_DEVICE_FUNC inline void +ptranspose(PacketBlock<Packet2d,2>& kernel) { + Packet2d t0 = vec_perm(kernel.packet[0], kernel.packet[1], p16uc_TRANSPOSE64_HI); + Packet2d t1 = vec_perm(kernel.packet[0], kernel.packet[1], p16uc_TRANSPOSE64_LO); + kernel.packet[0] = t0; + kernel.packet[1] = t1; +} + +template<> EIGEN_STRONG_INLINE Packet4i pblend(const Selector<4>& ifPacket, const Packet4i& thenPacket, const Packet4i& elsePacket) { + Packet4ui select = { ifPacket.select[0], ifPacket.select[1], ifPacket.select[2], ifPacket.select[3] }; + Packet4ui mask = vec_cmpeq(select, reinterpret_cast<Packet4ui>(p4i_ONE)); + return vec_sel(elsePacket, thenPacket, mask); +} + + +template<> EIGEN_STRONG_INLINE Packet2d pblend(const Selector<2>& ifPacket, const Packet2d& thenPacket, const Packet2d& elsePacket) { + Packet2ul select = { ifPacket.select[0], ifPacket.select[1] }; + Packet2ul mask = vec_cmpeq(select, reinterpret_cast<Packet2ul>(p2l_ONE)); + return vec_sel(elsePacket, thenPacket, mask); +} + +/* z13 has no vector float support so we emulate that with double + z14 has proper vector float support. +*/ +#if !defined(__ARCH__) || (defined(__ARCH__) && __ARCH__ < 12) +/* Helper function to simulate a vec_splat_packet4f + */ +template<int element> EIGEN_STRONG_INLINE Packet4f vec_splat_packet4f(const Packet4f& from) +{ + Packet4f splat; + switch (element) { + case 0: + splat.v4f[0] = vec_splat(from.v4f[0], 0); + splat.v4f[1] = splat.v4f[0]; + break; + case 1: + splat.v4f[0] = vec_splat(from.v4f[0], 1); + splat.v4f[1] = splat.v4f[0]; + break; + case 2: + splat.v4f[0] = vec_splat(from.v4f[1], 0); + splat.v4f[1] = splat.v4f[0]; + break; + case 3: + splat.v4f[0] = vec_splat(from.v4f[1], 1); + splat.v4f[1] = splat.v4f[0]; + break; + } + return splat; +} + +template<> EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(const float* from) +{ + // FIXME: No intrinsic yet + EIGEN_DEBUG_ALIGNED_LOAD + Packet4f vfrom; + vfrom.v4f[0] = vec_ld2f(&from[0]); + vfrom.v4f[1] = vec_ld2f(&from[2]); + return vfrom; +} + +template<> EIGEN_STRONG_INLINE void pstore<float>(float* to, const Packet4f& from) +{ + // FIXME: No intrinsic yet + EIGEN_DEBUG_ALIGNED_STORE + vec_st2f(from.v4f[0], &to[0]); + vec_st2f(from.v4f[1], &to[2]); +} + +template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float& from) +{ + Packet4f to; + to.v4f[0] = pset1<Packet2d>(static_cast<const double&>(from)); + to.v4f[1] = to.v4f[0]; + return to; +} + +template<> EIGEN_STRONG_INLINE void +pbroadcast4<Packet4f>(const float *a, + Packet4f& a0, Packet4f& a1, Packet4f& a2, Packet4f& a3) +{ + a3 = pload<Packet4f>(a); + a0 = vec_splat_packet4f<0>(a3); + a1 = vec_splat_packet4f<1>(a3); + a2 = vec_splat_packet4f<2>(a3); + a3 = vec_splat_packet4f<3>(a3); +} + +template<> EIGEN_DEVICE_FUNC inline Packet4f pgather<float, Packet4f>(const float* from, Index stride) +{ + float EIGEN_ALIGN16 ai[4]; + ai[0] = from[0*stride]; + ai[1] = from[1*stride]; + ai[2] = from[2*stride]; + ai[3] = from[3*stride]; + return pload<Packet4f>(ai); +} + template<> EIGEN_DEVICE_FUNC inline void pscatter<float, Packet4f>(float* to, const Packet4f& from, Index stride) { float EIGEN_ALIGN16 ai[4]; @@ -487,15 +657,6 @@ template<> EIGEN_DEVICE_FUNC inline void pscatter<float, Packet4f>(float* to, co to[3*stride] = ai[3]; } -template<> EIGEN_DEVICE_FUNC inline void pscatter<double, Packet2d>(double* to, const Packet2d& from, Index stride) -{ - double EIGEN_ALIGN16 af[2]; - pstore<double>(af, from); - to[0*stride] = af[0]; - to[1*stride] = af[1]; -} - -template<> EIGEN_STRONG_INLINE Packet4i padd<Packet4i>(const Packet4i& a, const Packet4i& b) { return (a + b); } template<> EIGEN_STRONG_INLINE Packet4f padd<Packet4f>(const Packet4f& a, const Packet4f& b) { Packet4f c; @@ -503,9 +664,7 @@ template<> EIGEN_STRONG_INLINE Packet4f padd<Packet4f>(const Packet4f& a, const c.v4f[1] = a.v4f[1] + b.v4f[1]; return c; } -template<> EIGEN_STRONG_INLINE Packet2d padd<Packet2d>(const Packet2d& a, const Packet2d& b) { return (a + b); } -template<> EIGEN_STRONG_INLINE Packet4i psub<Packet4i>(const Packet4i& a, const Packet4i& b) { return (a - b); } template<> EIGEN_STRONG_INLINE Packet4f psub<Packet4f>(const Packet4f& a, const Packet4f& b) { Packet4f c; @@ -513,9 +672,7 @@ template<> EIGEN_STRONG_INLINE Packet4f psub<Packet4f>(const Packet4f& a, const c.v4f[1] = a.v4f[1] - b.v4f[1]; return c; } -template<> EIGEN_STRONG_INLINE Packet2d psub<Packet2d>(const Packet2d& a, const Packet2d& b) { return (a - b); } -template<> EIGEN_STRONG_INLINE Packet4i pmul<Packet4i>(const Packet4i& a, const Packet4i& b) { return (a * b); } template<> EIGEN_STRONG_INLINE Packet4f pmul<Packet4f>(const Packet4f& a, const Packet4f& b) { Packet4f c; @@ -523,9 +680,7 @@ template<> EIGEN_STRONG_INLINE Packet4f pmul<Packet4f>(const Packet4f& a, const c.v4f[1] = a.v4f[1] * b.v4f[1]; return c; } -template<> EIGEN_STRONG_INLINE Packet2d pmul<Packet2d>(const Packet2d& a, const Packet2d& b) { return (a * b); } -template<> EIGEN_STRONG_INLINE Packet4i pdiv<Packet4i>(const Packet4i& a, const Packet4i& b) { return (a / b); } template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const Packet4f& b) { Packet4f c; @@ -533,9 +688,7 @@ template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const c.v4f[1] = a.v4f[1] / b.v4f[1]; return c; } -template<> EIGEN_STRONG_INLINE Packet2d pdiv<Packet2d>(const Packet2d& a, const Packet2d& b) { return (a / b); } -template<> EIGEN_STRONG_INLINE Packet4i pnegate(const Packet4i& a) { return (-a); } template<> EIGEN_STRONG_INLINE Packet4f pnegate(const Packet4f& a) { Packet4f c; @@ -543,13 +696,7 @@ template<> EIGEN_STRONG_INLINE Packet4f pnegate(const Packet4f& a) c.v4f[1] = -a.v4f[1]; return c; } -template<> EIGEN_STRONG_INLINE Packet2d pnegate(const Packet2d& a) { return (-a); } - -template<> EIGEN_STRONG_INLINE Packet4i pconj(const Packet4i& a) { return a; } -template<> EIGEN_STRONG_INLINE Packet4f pconj(const Packet4f& a) { return a; } -template<> EIGEN_STRONG_INLINE Packet2d pconj(const Packet2d& a) { return a; } -template<> EIGEN_STRONG_INLINE Packet4i pmadd(const Packet4i& a, const Packet4i& b, const Packet4i& c) { return padd<Packet4i>(pmul<Packet4i>(a, b), c); } template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) { Packet4f res; @@ -557,14 +704,7 @@ template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& res.v4f[1] = vec_madd(a.v4f[1], b.v4f[1], c.v4f[1]); return res; } -template<> EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& b, const Packet2d& c) { return vec_madd(a, b, c); } -template<> EIGEN_STRONG_INLINE Packet4i plset<Packet4i>(const int& a) { return padd<Packet4i>(pset1<Packet4i>(a), p4i_COUNTDOWN); } -template<> EIGEN_STRONG_INLINE Packet4f plset<Packet4f>(const float& a) { return padd<Packet4f>(pset1<Packet4f>(a), p4f_COUNTDOWN); } -template<> EIGEN_STRONG_INLINE Packet2d plset<Packet2d>(const double& a) { return padd<Packet2d>(pset1<Packet2d>(a), p2d_COUNTDOWN); } - -template<> EIGEN_STRONG_INLINE Packet4i pmin<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_min(a, b); } -template<> EIGEN_STRONG_INLINE Packet2d pmin<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_min(a, b); } template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const Packet4f& b) { Packet4f res; @@ -573,8 +713,6 @@ template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const return res; } -template<> EIGEN_STRONG_INLINE Packet4i pmax<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_max(a, b); } -template<> EIGEN_STRONG_INLINE Packet2d pmax<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_max(a, b); } template<> EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>(const Packet4f& a, const Packet4f& b) { Packet4f res; @@ -583,8 +721,6 @@ template<> EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>(const Packet4f& a, const return res; } -template<> EIGEN_STRONG_INLINE Packet4i pand<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_and(a, b); } -template<> EIGEN_STRONG_INLINE Packet2d pand<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_and(a, b); } template<> EIGEN_STRONG_INLINE Packet4f pand<Packet4f>(const Packet4f& a, const Packet4f& b) { Packet4f res; @@ -593,28 +729,22 @@ template<> EIGEN_STRONG_INLINE Packet4f pand<Packet4f>(const Packet4f& a, const return res; } -template<> EIGEN_STRONG_INLINE Packet4i por<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_or(a, b); } -template<> EIGEN_STRONG_INLINE Packet2d por<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_or(a, b); } template<> EIGEN_STRONG_INLINE Packet4f por<Packet4f>(const Packet4f& a, const Packet4f& b) { Packet4f res; - res.v4f[0] = pand(a.v4f[0], b.v4f[0]); - res.v4f[1] = pand(a.v4f[1], b.v4f[1]); + res.v4f[0] = por(a.v4f[0], b.v4f[0]); + res.v4f[1] = por(a.v4f[1], b.v4f[1]); return res; } -template<> EIGEN_STRONG_INLINE Packet4i pxor<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_xor(a, b); } -template<> EIGEN_STRONG_INLINE Packet2d pxor<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_xor(a, b); } template<> EIGEN_STRONG_INLINE Packet4f pxor<Packet4f>(const Packet4f& a, const Packet4f& b) { Packet4f res; - res.v4f[0] = pand(a.v4f[0], b.v4f[0]); - res.v4f[1] = pand(a.v4f[1], b.v4f[1]); + res.v4f[0] = pxor(a.v4f[0], b.v4f[0]); + res.v4f[1] = pxor(a.v4f[1], b.v4f[1]); return res; } -template<> EIGEN_STRONG_INLINE Packet4i pandnot<Packet4i>(const Packet4i& a, const Packet4i& b) { return pand<Packet4i>(a, vec_nor(b, b)); } -template<> EIGEN_STRONG_INLINE Packet2d pandnot<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_and(a, vec_nor(b, b)); } template<> EIGEN_STRONG_INLINE Packet4f pandnot<Packet4f>(const Packet4f& a, const Packet4f& b) { Packet4f res; @@ -630,7 +760,7 @@ template<> EIGEN_STRONG_INLINE Packet4f pround<Packet4f>(const Packet4f& a) res.v4f[1] = vec_round(a.v4f[1]); return res; } -template<> EIGEN_STRONG_INLINE Packet2d pround<Packet2d>(const Packet2d& a) { return vec_round(a); } + template<> EIGEN_STRONG_INLINE Packet4f pceil<Packet4f>(const Packet4f& a) { Packet4f res; @@ -638,7 +768,7 @@ template<> EIGEN_STRONG_INLINE Packet4f pceil<Packet4f>(const Packet4f& a) res.v4f[1] = vec_ceil(a.v4f[1]); return res; } -template<> EIGEN_STRONG_INLINE Packet2d pceil<Packet2d>(const Packet2d& a) { return vec_ceil(a); } + template<> EIGEN_STRONG_INLINE Packet4f pfloor<Packet4f>(const Packet4f& a) { Packet4f res; @@ -646,18 +776,6 @@ template<> EIGEN_STRONG_INLINE Packet4f pfloor<Packet4f>(const Packet4f& a) res.v4f[1] = vec_floor(a.v4f[1]); return res; } -template<> EIGEN_STRONG_INLINE Packet2d pfloor<Packet2d>(const Packet2d& a) { return vec_floor(a); } - -template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from) { return pload<Packet4i>(from); } -template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from) { return pload<Packet4f>(from); } -template<> EIGEN_STRONG_INLINE Packet2d ploadu<Packet2d>(const double* from) { return pload<Packet2d>(from); } - - -template<> EIGEN_STRONG_INLINE Packet4i ploaddup<Packet4i>(const int* from) -{ - Packet4i p = pload<Packet4i>(from); - return vec_perm(p, p, p16uc_DUPLICATE32_HI); -} template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float* from) { @@ -667,33 +785,7 @@ template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float* from) return p; } -template<> EIGEN_STRONG_INLINE Packet2d ploaddup<Packet2d>(const double* from) -{ - Packet2d p = pload<Packet2d>(from); - return vec_perm(p, p, p16uc_PSET64_HI); -} - -template<> EIGEN_STRONG_INLINE void pstoreu<int>(int* to, const Packet4i& from) { pstore<int>(to, from); } -template<> EIGEN_STRONG_INLINE void pstoreu<float>(float* to, const Packet4f& from) { pstore<float>(to, from); } -template<> EIGEN_STRONG_INLINE void pstoreu<double>(double* to, const Packet2d& from) { pstore<double>(to, from); } - -template<> EIGEN_STRONG_INLINE void prefetch<int>(const int* addr) { EIGEN_ZVECTOR_PREFETCH(addr); } -template<> EIGEN_STRONG_INLINE void prefetch<float>(const float* addr) { EIGEN_ZVECTOR_PREFETCH(addr); } -template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { EIGEN_ZVECTOR_PREFETCH(addr); } - -template<> EIGEN_STRONG_INLINE int pfirst<Packet4i>(const Packet4i& a) { int EIGEN_ALIGN16 x[4]; pstore(x, a); return x[0]; } template<> EIGEN_STRONG_INLINE float pfirst<Packet4f>(const Packet4f& a) { float EIGEN_ALIGN16 x[2]; vec_st2f(a.v4f[0], &x[0]); return x[0]; } -template<> EIGEN_STRONG_INLINE double pfirst<Packet2d>(const Packet2d& a) { double EIGEN_ALIGN16 x[2]; pstore(x, a); return x[0]; } - -template<> EIGEN_STRONG_INLINE Packet4i preverse(const Packet4i& a) -{ - return reinterpret_cast<Packet4i>(vec_perm(reinterpret_cast<Packet16uc>(a), reinterpret_cast<Packet16uc>(a), p16uc_REVERSE32)); -} - -template<> EIGEN_STRONG_INLINE Packet2d preverse(const Packet2d& a) -{ - return reinterpret_cast<Packet2d>(vec_perm(reinterpret_cast<Packet16uc>(a), reinterpret_cast<Packet16uc>(a), p16uc_REVERSE64)); -} template<> EIGEN_STRONG_INLINE Packet4f preverse(const Packet4f& a) { @@ -703,8 +795,6 @@ template<> EIGEN_STRONG_INLINE Packet4f preverse(const Packet4f& a) return rev; } -template<> EIGEN_STRONG_INLINE Packet4i pabs<Packet4i>(const Packet4i& a) { return vec_abs(a); } -template<> EIGEN_STRONG_INLINE Packet2d pabs<Packet2d>(const Packet2d& a) { return vec_abs(a); } template<> EIGEN_STRONG_INLINE Packet4f pabs<Packet4f>(const Packet4f& a) { Packet4f res; @@ -713,23 +803,6 @@ template<> EIGEN_STRONG_INLINE Packet4f pabs<Packet4f>(const Packet4f& a) return res; } -template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a) -{ - Packet4i b, sum; - b = vec_sld(a, a, 8); - sum = padd<Packet4i>(a, b); - b = vec_sld(sum, sum, 4); - sum = padd<Packet4i>(sum, b); - return pfirst(sum); -} - -template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a) -{ - Packet2d b, sum; - b = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(a), reinterpret_cast<Packet4i>(a), 8)); - sum = padd<Packet2d>(a, b); - return pfirst(sum); -} template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a) { Packet2d sum; @@ -738,94 +811,12 @@ template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a) return static_cast<float>(first); } -template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs) -{ - Packet4i v[4], sum[4]; - - // It's easier and faster to transpose then add as columns - // Check: http://www.freevec.org/function/matrix_4x4_transpose_floats for explanation - // Do the transpose, first set of moves - v[0] = vec_mergeh(vecs[0], vecs[2]); - v[1] = vec_mergel(vecs[0], vecs[2]); - v[2] = vec_mergeh(vecs[1], vecs[3]); - v[3] = vec_mergel(vecs[1], vecs[3]); - // Get the resulting vectors - sum[0] = vec_mergeh(v[0], v[2]); - sum[1] = vec_mergel(v[0], v[2]); - sum[2] = vec_mergeh(v[1], v[3]); - sum[3] = vec_mergel(v[1], v[3]); - - // Now do the summation: - // Lines 0+1 - sum[0] = padd<Packet4i>(sum[0], sum[1]); - // Lines 2+3 - sum[1] = padd<Packet4i>(sum[2], sum[3]); - // Add the results - sum[0] = padd<Packet4i>(sum[0], sum[1]); - - return sum[0]; -} - -template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs) -{ - Packet2d v[2], sum; - v[0] = padd<Packet2d>(vecs[0], reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4ui>(vecs[0]), reinterpret_cast<Packet4ui>(vecs[0]), 8))); - v[1] = padd<Packet2d>(vecs[1], reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4ui>(vecs[1]), reinterpret_cast<Packet4ui>(vecs[1]), 8))); - - sum = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4ui>(v[0]), reinterpret_cast<Packet4ui>(v[1]), 8)); - - return sum; -} - -template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs) -{ - PacketBlock<Packet4f,4> transpose; - transpose.packet[0] = vecs[0]; - transpose.packet[1] = vecs[1]; - transpose.packet[2] = vecs[2]; - transpose.packet[3] = vecs[3]; - ptranspose(transpose); - - Packet4f sum = padd(transpose.packet[0], transpose.packet[1]); - sum = padd(sum, transpose.packet[2]); - sum = padd(sum, transpose.packet[3]); - return sum; -} - -// Other reduction functions: -// mul -template<> EIGEN_STRONG_INLINE int predux_mul<Packet4i>(const Packet4i& a) -{ - EIGEN_ALIGN16 int aux[4]; - pstore(aux, a); - return aux[0] * aux[1] * aux[2] * aux[3]; -} - -template<> EIGEN_STRONG_INLINE double predux_mul<Packet2d>(const Packet2d& a) -{ - return pfirst(pmul(a, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(a), reinterpret_cast<Packet4i>(a), 8)))); -} - template<> EIGEN_STRONG_INLINE float predux_mul<Packet4f>(const Packet4f& a) { // Return predux_mul<Packet2d> of the subvectors product return static_cast<float>(pfirst(predux_mul(pmul(a.v4f[0], a.v4f[1])))); } -// min -template<> EIGEN_STRONG_INLINE int predux_min<Packet4i>(const Packet4i& a) -{ - Packet4i b, res; - b = pmin<Packet4i>(a, vec_sld(a, a, 8)); - res = pmin<Packet4i>(b, vec_sld(b, b, 4)); - return pfirst(res); -} - -template<> EIGEN_STRONG_INLINE double predux_min<Packet2d>(const Packet2d& a) -{ - return pfirst(pmin<Packet2d>(a, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(a), reinterpret_cast<Packet4i>(a), 8)))); -} - template<> EIGEN_STRONG_INLINE float predux_min<Packet4f>(const Packet4f& a) { Packet2d b, res; @@ -834,21 +825,6 @@ template<> EIGEN_STRONG_INLINE float predux_min<Packet4f>(const Packet4f& a) return static_cast<float>(pfirst(res)); } -// max -template<> EIGEN_STRONG_INLINE int predux_max<Packet4i>(const Packet4i& a) -{ - Packet4i b, res; - b = pmax<Packet4i>(a, vec_sld(a, a, 8)); - res = pmax<Packet4i>(b, vec_sld(b, b, 4)); - return pfirst(res); -} - -// max -template<> EIGEN_STRONG_INLINE double predux_max<Packet2d>(const Packet2d& a) -{ - return pfirst(pmax<Packet2d>(a, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(a), reinterpret_cast<Packet4i>(a), 8)))); -} - template<> EIGEN_STRONG_INLINE float predux_max<Packet4f>(const Packet4f& a) { Packet2d b, res; @@ -857,26 +833,6 @@ template<> EIGEN_STRONG_INLINE float predux_max<Packet4f>(const Packet4f& a) return static_cast<float>(pfirst(res)); } -EIGEN_DEVICE_FUNC inline void -ptranspose(PacketBlock<Packet4i,4>& kernel) { - Packet4i t0 = vec_mergeh(kernel.packet[0], kernel.packet[2]); - Packet4i t1 = vec_mergel(kernel.packet[0], kernel.packet[2]); - Packet4i t2 = vec_mergeh(kernel.packet[1], kernel.packet[3]); - Packet4i t3 = vec_mergel(kernel.packet[1], kernel.packet[3]); - kernel.packet[0] = vec_mergeh(t0, t2); - kernel.packet[1] = vec_mergel(t0, t2); - kernel.packet[2] = vec_mergeh(t1, t3); - kernel.packet[3] = vec_mergel(t1, t3); -} - -EIGEN_DEVICE_FUNC inline void -ptranspose(PacketBlock<Packet2d,2>& kernel) { - Packet2d t0 = vec_perm(kernel.packet[0], kernel.packet[1], p16uc_TRANSPOSE64_HI); - Packet2d t1 = vec_perm(kernel.packet[0], kernel.packet[1], p16uc_TRANSPOSE64_LO); - kernel.packet[0] = t0; - kernel.packet[1] = t1; -} - /* Split the Packet4f PacketBlock into 4 Packet2d PacketBlocks and transpose each one */ EIGEN_DEVICE_FUNC inline void @@ -915,12 +871,6 @@ ptranspose(PacketBlock<Packet4f,4>& kernel) { kernel.packet[3].v4f[1] = t3.packet[1]; } -template<> EIGEN_STRONG_INLINE Packet4i pblend(const Selector<4>& ifPacket, const Packet4i& thenPacket, const Packet4i& elsePacket) { - Packet4ui select = { ifPacket.select[0], ifPacket.select[1], ifPacket.select[2], ifPacket.select[3] }; - Packet4ui mask = vec_cmpeq(select, reinterpret_cast<Packet4ui>(p4i_ONE)); - return vec_sel(elsePacket, thenPacket, mask); -} - template<> EIGEN_STRONG_INLINE Packet4f pblend(const Selector<4>& ifPacket, const Packet4f& thenPacket, const Packet4f& elsePacket) { Packet2ul select_hi = { ifPacket.select[0], ifPacket.select[1] }; Packet2ul select_lo = { ifPacket.select[2], ifPacket.select[3] }; @@ -932,12 +882,177 @@ template<> EIGEN_STRONG_INLINE Packet4f pblend(const Selector<4>& ifPacket, cons return result; } -template<> EIGEN_STRONG_INLINE Packet2d pblend(const Selector<2>& ifPacket, const Packet2d& thenPacket, const Packet2d& elsePacket) { - Packet2ul select = { ifPacket.select[0], ifPacket.select[1] }; - Packet2ul mask = vec_cmpeq(select, reinterpret_cast<Packet2ul>(p2l_ONE)); +template<> Packet4f EIGEN_STRONG_INLINE pcmp_le<Packet4f>(const Packet4f& a, const Packet4f& b) +{ + Packet4f res; + res.v4f[0] = pcmp_le(a.v4f[0], b.v4f[0]); + res.v4f[1] = pcmp_le(a.v4f[1], b.v4f[1]); + return res; +} + +template<> Packet4f EIGEN_STRONG_INLINE pcmp_lt<Packet4f>(const Packet4f& a, const Packet4f& b) +{ + Packet4f res; + res.v4f[0] = pcmp_lt(a.v4f[0], b.v4f[0]); + res.v4f[1] = pcmp_lt(a.v4f[1], b.v4f[1]); + return res; +} + +template<> Packet4f EIGEN_STRONG_INLINE pcmp_eq<Packet4f>(const Packet4f& a, const Packet4f& b) +{ + Packet4f res; + res.v4f[0] = pcmp_eq(a.v4f[0], b.v4f[0]); + res.v4f[1] = pcmp_eq(a.v4f[1], b.v4f[1]); + return res; +} + +#else +template<> EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(const float* from) +{ + // FIXME: No intrinsic yet + EIGEN_DEBUG_ALIGNED_LOAD + Packet *vfrom; + vfrom = (Packet *) from; + return vfrom->v4f; +} + +template<> EIGEN_STRONG_INLINE void pstore<float>(float* to, const Packet4f& from) +{ + // FIXME: No intrinsic yet + EIGEN_DEBUG_ALIGNED_STORE + Packet *vto; + vto = (Packet *) to; + vto->v4f = from; +} + +template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float& from) +{ + return vec_splats(from); +} + +template<> EIGEN_STRONG_INLINE void +pbroadcast4<Packet4f>(const float *a, + Packet4f& a0, Packet4f& a1, Packet4f& a2, Packet4f& a3) +{ + a3 = pload<Packet4f>(a); + a0 = vec_splat(a3, 0); + a1 = vec_splat(a3, 1); + a2 = vec_splat(a3, 2); + a3 = vec_splat(a3, 3); +} + +template<> EIGEN_DEVICE_FUNC inline Packet4f pgather<float, Packet4f>(const float* from, Index stride) +{ + float EIGEN_ALIGN16 af[4]; + af[0] = from[0*stride]; + af[1] = from[1*stride]; + af[2] = from[2*stride]; + af[3] = from[3*stride]; + return pload<Packet4f>(af); +} + +template<> EIGEN_DEVICE_FUNC inline void pscatter<float, Packet4f>(float* to, const Packet4f& from, Index stride) +{ + float EIGEN_ALIGN16 af[4]; + pstore<float>((float*)af, from); + to[0*stride] = af[0]; + to[1*stride] = af[1]; + to[2*stride] = af[2]; + to[3*stride] = af[3]; +} + +template<> EIGEN_STRONG_INLINE Packet4f padd<Packet4f>(const Packet4f& a, const Packet4f& b) { return (a + b); } +template<> EIGEN_STRONG_INLINE Packet4f psub<Packet4f>(const Packet4f& a, const Packet4f& b) { return (a - b); } +template<> EIGEN_STRONG_INLINE Packet4f pmul<Packet4f>(const Packet4f& a, const Packet4f& b) { return (a * b); } +template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const Packet4f& b) { return (a / b); } +template<> EIGEN_STRONG_INLINE Packet4f pnegate<Packet4f>(const Packet4f& a) { return (-a); } +template<> EIGEN_STRONG_INLINE Packet4f pconj<Packet4f> (const Packet4f& a) { return a; } +template<> EIGEN_STRONG_INLINE Packet4f pmadd<Packet4f> (const Packet4f& a, const Packet4f& b, const Packet4f& c) { return vec_madd(a, b, c); } +template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f> (const Packet4f& a, const Packet4f& b) { return vec_min(a, b); } +template<> EIGEN_STRONG_INLINE Packet4f pmax<Packet4f> (const Packet4f& a, const Packet4f& b) { return vec_max(a, b); } +template<> EIGEN_STRONG_INLINE Packet4f pand<Packet4f> (const Packet4f& a, const Packet4f& b) { return vec_and(a, b); } +template<> EIGEN_STRONG_INLINE Packet4f por<Packet4f> (const Packet4f& a, const Packet4f& b) { return vec_or(a, b); } +template<> EIGEN_STRONG_INLINE Packet4f pxor<Packet4f> (const Packet4f& a, const Packet4f& b) { return vec_xor(a, b); } +template<> EIGEN_STRONG_INLINE Packet4f pandnot<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_and(a, vec_nor(b, b)); } +template<> EIGEN_STRONG_INLINE Packet4f pround<Packet4f> (const Packet4f& a) { return vec_round(a); } +template<> EIGEN_STRONG_INLINE Packet4f pceil<Packet4f> (const Packet4f& a) { return vec_ceil(a); } +template<> EIGEN_STRONG_INLINE Packet4f pfloor<Packet4f> (const Packet4f& a) { return vec_floor(a); } +template<> EIGEN_STRONG_INLINE Packet4f pabs<Packet4f> (const Packet4f& a) { return vec_abs(a); } +template<> EIGEN_STRONG_INLINE float pfirst<Packet4f>(const Packet4f& a) { float EIGEN_ALIGN16 x[4]; pstore(x, a); return x[0]; } + +template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float* from) +{ + Packet4f p = pload<Packet4f>(from); + return vec_perm(p, p, p16uc_DUPLICATE32_HI); +} + +template<> EIGEN_STRONG_INLINE Packet4f preverse(const Packet4f& a) +{ + return reinterpret_cast<Packet4f>(vec_perm(reinterpret_cast<Packet16uc>(a), reinterpret_cast<Packet16uc>(a), p16uc_REVERSE32)); +} + +template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a) +{ + Packet4f b, sum; + b = vec_sld(a, a, 8); + sum = padd<Packet4f>(a, b); + b = vec_sld(sum, sum, 4); + sum = padd<Packet4f>(sum, b); + return pfirst(sum); +} + +// Other reduction functions: +// mul +template<> EIGEN_STRONG_INLINE float predux_mul<Packet4f>(const Packet4f& a) +{ + Packet4f prod; + prod = pmul(a, vec_sld(a, a, 8)); + return pfirst(pmul(prod, vec_sld(prod, prod, 4))); +} + +// min +template<> EIGEN_STRONG_INLINE float predux_min<Packet4f>(const Packet4f& a) +{ + Packet4f b, res; + b = pmin<Packet4f>(a, vec_sld(a, a, 8)); + res = pmin<Packet4f>(b, vec_sld(b, b, 4)); + return pfirst(res); +} + +// max +template<> EIGEN_STRONG_INLINE float predux_max<Packet4f>(const Packet4f& a) +{ + Packet4f b, res; + b = pmax<Packet4f>(a, vec_sld(a, a, 8)); + res = pmax<Packet4f>(b, vec_sld(b, b, 4)); + return pfirst(res); +} + +EIGEN_DEVICE_FUNC inline void +ptranspose(PacketBlock<Packet4f,4>& kernel) { + Packet4f t0 = vec_mergeh(kernel.packet[0], kernel.packet[2]); + Packet4f t1 = vec_mergel(kernel.packet[0], kernel.packet[2]); + Packet4f t2 = vec_mergeh(kernel.packet[1], kernel.packet[3]); + Packet4f t3 = vec_mergel(kernel.packet[1], kernel.packet[3]); + kernel.packet[0] = vec_mergeh(t0, t2); + kernel.packet[1] = vec_mergel(t0, t2); + kernel.packet[2] = vec_mergeh(t1, t3); + kernel.packet[3] = vec_mergel(t1, t3); +} + +template<> EIGEN_STRONG_INLINE Packet4f pblend(const Selector<4>& ifPacket, const Packet4f& thenPacket, const Packet4f& elsePacket) { + Packet4ui select = { ifPacket.select[0], ifPacket.select[1], ifPacket.select[2], ifPacket.select[3] }; + Packet4ui mask = vec_cmpeq(select, reinterpret_cast<Packet4ui>(p4i_ONE)); return vec_sel(elsePacket, thenPacket, mask); } +#endif + +template<> EIGEN_STRONG_INLINE void prefetch<float>(const float* addr) { EIGEN_ZVECTOR_PREFETCH(addr); } +template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f> (const float* from) { return pload<Packet4f>(from); } +template<> EIGEN_STRONG_INLINE void pstoreu<float>(float* to, const Packet4f& from) { pstore<float>(to, from); } +template<> EIGEN_STRONG_INLINE Packet4f plset<Packet4f> (const float& a) { return padd<Packet4f>(pset1<Packet4f>(a), p4f_COUNTDOWN); } + } // end namespace internal } // end namespace Eigen |