diff options
Diffstat (limited to 'Eigen/src')
-rw-r--r-- | Eigen/src/Core/arch/CUDA/Half.h | 22 | ||||
-rw-r--r-- | Eigen/src/Core/arch/ZVector/Complex.h | 222 | ||||
-rw-r--r-- | Eigen/src/Core/arch/ZVector/MathFunctions.h | 65 | ||||
-rwxr-xr-x | Eigen/src/Core/arch/ZVector/PacketMath.h | 386 | ||||
-rw-r--r-- | Eigen/src/Core/util/Macros.h | 2 |
5 files changed, 659 insertions, 38 deletions
diff --git a/Eigen/src/Core/arch/CUDA/Half.h b/Eigen/src/Core/arch/CUDA/Half.h index 52892db38..63371bdf4 100644 --- a/Eigen/src/Core/arch/CUDA/Half.h +++ b/Eigen/src/Core/arch/CUDA/Half.h @@ -386,11 +386,15 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half abs(const half& a) { return result; } EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half exp(const half& a) { - return half(::expf(float(a))); +#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 530 + return half(hexp(a)); +#else + return half(::expf(float(a))); +#endif } EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half log(const half& a) { #if defined(EIGEN_HAS_CUDA_FP16) && defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530 - return Eigen::half(::hlog(a)); + return half(::hlog(a)); #else return half(::logf(float(a))); #endif @@ -402,7 +406,11 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half log10(const half& a) { return half(::log10f(float(a))); } EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half sqrt(const half& a) { - return half(::sqrtf(float(a))); +#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 530 + return half(hsqrt(a)); +#else + return half(::sqrtf(float(a))); +#endif } EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half pow(const half& a, const half& b) { return half(::powf(float(a), float(b))); @@ -420,10 +428,18 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half tanh(const half& a) { return half(::tanhf(float(a))); } EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half floor(const half& a) { +#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 300 + return half(hfloor(a)); +#else return half(::floorf(float(a))); +#endif } EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half ceil(const half& a) { +#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 300 + return half(hceil(a)); +#else return half(::ceilf(float(a))); +#endif } EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half (min)(const half& a, const half& b) { diff --git a/Eigen/src/Core/arch/ZVector/Complex.h b/Eigen/src/Core/arch/ZVector/Complex.h index e9d83eca6..d39d2d105 100644 --- a/Eigen/src/Core/arch/ZVector/Complex.h +++ b/Eigen/src/Core/arch/ZVector/Complex.h @@ -2,6 +2,7 @@ // for linear algebra. // // Copyright (C) 2010 Gael Guennebaud <gael.guennebaud@inria.fr> +// Copyright (C) 2016 Konstantinos Margaritis <markos@freevec.org> // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed @@ -24,13 +25,48 @@ struct Packet1cd Packet2d v; }; +struct Packet2cf +{ + EIGEN_STRONG_INLINE Packet2cf() {} + EIGEN_STRONG_INLINE explicit Packet2cf(const Packet4f& a) : v(a) {} + union { + Packet4f v; + Packet1cd cd[2]; + }; +}; + +template<> struct packet_traits<std::complex<float> > : default_packet_traits +{ + typedef Packet2cf type; + typedef Packet2cf half; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size = 2, + HasHalfPacket = 0, + + HasAdd = 1, + HasSub = 1, + HasMul = 1, + HasDiv = 1, + HasNegate = 1, + HasAbs = 0, + HasAbs2 = 0, + HasMin = 0, + HasMax = 0, + HasBlend = 1, + HasSetLinear = 0 + }; +}; + + template<> struct packet_traits<std::complex<double> > : default_packet_traits { typedef Packet1cd type; typedef Packet1cd half; enum { Vectorizable = 1, - AlignedOnScalar = 0, + AlignedOnScalar = 1, size = 1, HasHalfPacket = 0, @@ -47,20 +83,68 @@ 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; }; +/* 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)); } 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) { @@ -79,43 +163,90 @@ 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 ploaddup<Packet1cd>(const std::complex<double>* from) -{ - return pset1<Packet1cd>(*from); -} +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 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) { - std::complex<double> EIGEN_ALIGN16 res[2]; - pstore<std::complex<double> >(res, a); + std::complex<double> EIGEN_ALIGN16 res; + pstore<std::complex<double> >(&res, a); + + 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) +{ + Packet2cf res; + res.cd[0] = a.cd[1]; + res.cd[1] = a.cd[0]; + return res; +} template<> EIGEN_STRONG_INLINE std::complex<double> predux<Packet1cd>(const Packet1cd& a) { return pfirst(a); } +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 preduxp<Packet1cd>(const Packet1cd* vecs) { return vecs[0]; } +template<> EIGEN_STRONG_INLINE Packet2cf preduxp<Packet2cf>(const Packet2cf* vecs) +{ + PacketBlock<Packet2cf,2> transpose; + transpose.packet[0] = vecs[0]; + transpose.packet[1] = vecs[1]; + ptranspose(transpose); + + return padd<Packet2cf>(transpose.packet[0], transpose.packet[1]); +} template<> EIGEN_STRONG_INLINE std::complex<double> predux_mul<Packet1cd>(const Packet1cd& a) { return pfirst(a); } +template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(const Packet2cf& a) +{ + std::complex<float> res; + Packet1cd b = pmul<Packet1cd>(a.cd[0], a.cd[1]); + vec_st2f(b.v, (float*)&res); + return res; +} template<int Offset> struct palign_impl<Offset,Packet1cd> @@ -127,6 +258,18 @@ struct palign_impl<Offset,Packet1cd> } }; +template<int Offset> +struct palign_impl<Offset,Packet2cf> +{ + 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> { EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const @@ -160,6 +303,39 @@ template<> struct conj_helper<Packet1cd, Packet1cd, true,true> } }; +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); } + + EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const + { + return internal::pmul(a, pconj(b)); + } +}; + +template<> struct conj_helper<Packet2cf, Packet2cf, true,false> +{ + EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const + { + return internal::pmul(pconj(a), b); + } +}; + +template<> struct conj_helper<Packet2cf, Packet2cf, true,true> +{ + EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const + { + return pconj(internal::pmul(a, b)); + } +}; + template<> EIGEN_STRONG_INLINE Packet1cd pdiv<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { // TODO optimize it for AltiVec @@ -168,17 +344,49 @@ template<> EIGEN_STRONG_INLINE Packet1cd pdiv<Packet1cd>(const Packet1cd& a, con return Packet1cd(pdiv(res.v, s + vec_perm(s, s, p16uc_REVERSE64))); } +template<> EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, const Packet2cf& b) +{ + // TODO optimize it for AltiVec + Packet2cf res; + res.cd[0] = pdiv<Packet1cd>(a.cd[0], b.cd[0]); + res.cd[1] = pdiv<Packet1cd>(a.cd[1], b.cd[1]); + 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; + res.cd[0] = pcplxflip(x.cd[0]); + res.cd[1] = pcplxflip(x.cd[1]); + 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]; + kernel.packet[0].cd[1] = kernel.packet[1].cd[0]; + kernel.packet[1].cd[0] = tmp; +} + +template<> EIGEN_STRONG_INLINE Packet2cf pblend(const Selector<2>& ifPacket, const Packet2cf& thenPacket, const Packet2cf& elsePacket) { + Packet2cf result; + const Selector<4> ifPacket4 = { ifPacket.select[0], ifPacket.select[0], ifPacket.select[1], ifPacket.select[1] }; + result.v = pblend<Packet4f>(ifPacket4, thenPacket.v, elsePacket.v); + return result; +} + } // end namespace internal } // end namespace Eigen diff --git a/Eigen/src/Core/arch/ZVector/MathFunctions.h b/Eigen/src/Core/arch/ZVector/MathFunctions.h index 6fff8524e..5c7aa7256 100644 --- a/Eigen/src/Core/arch/ZVector/MathFunctions.h +++ b/Eigen/src/Core/arch/ZVector/MathFunctions.h @@ -3,6 +3,7 @@ // // Copyright (C) 2007 Julien Pommier // Copyright (C) 2009 Gael Guennebaud <gael.guennebaud@inria.fr> +// Copyright (C) 2016 Konstantinos Margaritis <markos@freevec.org> // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed @@ -19,31 +20,31 @@ namespace Eigen { namespace internal { -template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED -Packet2d pexp<Packet2d>(const Packet2d& _x) -{ - Packet2d x = _x; +static _EIGEN_DECLARE_CONST_Packet2d(1 , 1.0); +static _EIGEN_DECLARE_CONST_Packet2d(2 , 2.0); +static _EIGEN_DECLARE_CONST_Packet2d(half, 0.5); - _EIGEN_DECLARE_CONST_Packet2d(1 , 1.0); - _EIGEN_DECLARE_CONST_Packet2d(2 , 2.0); - _EIGEN_DECLARE_CONST_Packet2d(half, 0.5); +static _EIGEN_DECLARE_CONST_Packet2d(exp_hi, 709.437); +static _EIGEN_DECLARE_CONST_Packet2d(exp_lo, -709.436139303); - _EIGEN_DECLARE_CONST_Packet2d(exp_hi, 709.437); - _EIGEN_DECLARE_CONST_Packet2d(exp_lo, -709.436139303); +static _EIGEN_DECLARE_CONST_Packet2d(cephes_LOG2EF, 1.4426950408889634073599); - _EIGEN_DECLARE_CONST_Packet2d(cephes_LOG2EF, 1.4426950408889634073599); +static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p0, 1.26177193074810590878e-4); +static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p1, 3.02994407707441961300e-2); +static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p2, 9.99999999999999999910e-1); - _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p0, 1.26177193074810590878e-4); - _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p1, 3.02994407707441961300e-2); - _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p2, 9.99999999999999999910e-1); +static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q0, 3.00198505138664455042e-6); +static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q1, 2.52448340349684104192e-3); +static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q2, 2.27265548208155028766e-1); +static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q3, 2.00000000000000000009e0); - _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q0, 3.00198505138664455042e-6); - _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q1, 2.52448340349684104192e-3); - _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q2, 2.27265548208155028766e-1); - _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q3, 2.00000000000000000009e0); +static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C1, 0.693145751953125); +static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C2, 1.42860682030941723212e-6); - _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C1, 0.693145751953125); - _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C2, 1.42860682030941723212e-6); +template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED +Packet2d pexp<Packet2d>(const Packet2d& _x) +{ + Packet2d x = _x; Packet2d tmp, fx; Packet2l emm0; @@ -92,17 +93,43 @@ Packet2d pexp<Packet2d>(const Packet2d& _x) } template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED +Packet4f pexp<Packet4f>(const Packet4f& x) +{ + Packet4f res; + res.v4f[0] = pexp<Packet2d>(x.v4f[0]); + res.v4f[1] = pexp<Packet2d>(x.v4f[1]); + return res; +} + +template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet2d psqrt<Packet2d>(const Packet2d& x) { return __builtin_s390_vfsqdb(x); } template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED +Packet4f psqrt<Packet4f>(const Packet4f& x) +{ + Packet4f res; + res.v4f[0] = psqrt<Packet2d>(x.v4f[0]); + res.v4f[1] = psqrt<Packet2d>(x.v4f[1]); + 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; + res.v4f[0] = prsqrt<Packet2d>(x.v4f[0]); + res.v4f[1] = prsqrt<Packet2d>(x.v4f[1]); + return res; +} + } // 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 5a7226be6..e2deb25c8 100755 --- a/Eigen/src/Core/arch/ZVector/PacketMath.h +++ b/Eigen/src/Core/arch/ZVector/PacketMath.h @@ -28,9 +28,8 @@ namespace internal { #define EIGEN_HAS_SINGLE_INSTRUCTION_CJMADD #endif -// NOTE Altivec has 32 registers, but Eigen only accepts a value of 8 or 16 #ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS -#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 32 +#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 16 #endif typedef __vector int Packet4i; @@ -42,6 +41,10 @@ typedef __vector double Packet2d; typedef __vector unsigned long long Packet2ul; typedef __vector long long Packet2l; +typedef struct { + Packet2d v4f[2]; +} Packet4f; + typedef union { int32_t i[4]; uint32_t ui[4]; @@ -88,6 +91,7 @@ static Packet2d p2d_ONE = { 1.0, 1.0 }; static Packet2d p2d_ZERO_ = { -0.0, -0.0 }; static Packet4i p4i_COUNTDOWN = { 0, 1, 2, 3 }; +static Packet4f p4f_COUNTDOWN = { 0.0, 1.0, 2.0, 3.0 }; static Packet2d p2d_COUNTDOWN = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet16uc>(p2d_ZERO), reinterpret_cast<Packet16uc>(p2d_ONE), 8)); static Packet16uc p16uc_PSET64_HI = { 0,1,2,3, 4,5,6,7, 0,1,2,3, 4,5,6,7 }; @@ -132,13 +136,11 @@ template<> struct packet_traits<int> : default_packet_traits typedef Packet4i type; typedef Packet4i half; enum { - // FIXME check the Has* Vectorizable = 1, AlignedOnScalar = 1, size = 4, HasHalfPacket = 0, - // FIXME check the Has* HasAdd = 1, HasSub = 1, HasMul = 1, @@ -147,6 +149,37 @@ template<> struct packet_traits<int> : default_packet_traits }; }; +template<> struct packet_traits<float> : default_packet_traits +{ + typedef Packet4f type; + typedef Packet4f half; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + 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, + HasSqrt = 1, + HasRsqrt = 1, + HasRound = 1, + HasFloor = 1, + HasCeil = 1, + HasNegate = 1, + HasBlend = 1 + }; +}; + template<> struct packet_traits<double> : default_packet_traits { typedef Packet2d type; @@ -157,7 +190,6 @@ template<> struct packet_traits<double> : default_packet_traits size=2, HasHalfPacket = 1, - // FIXME check the Has* HasAdd = 1, HasSub = 1, HasMul = 1, @@ -180,8 +212,12 @@ 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; }; +/* Forward declaration */ +EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet4f,4>& kernel); + inline std::ostream & operator <<(std::ostream & s, const Packet4i & v) { Packet vt; @@ -222,6 +258,32 @@ 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) +{ + 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<int Offset> struct palign_impl<Offset,Packet4i> { @@ -238,6 +300,31 @@ struct palign_impl<Offset,Packet4i> } }; +/* 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> { @@ -257,6 +344,16 @@ 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 @@ -275,6 +372,15 @@ 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 @@ -288,10 +394,16 @@ template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int& from) { return vec_splats(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, @@ -305,6 +417,17 @@ 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) { @@ -326,6 +449,16 @@ 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]; @@ -344,6 +477,16 @@ template<> EIGEN_DEVICE_FUNC inline void pscatter<int, Packet4i>(int* to, const to[3*stride] = ai[3]; } +template<> EIGEN_DEVICE_FUNC inline void pscatter<float, Packet4f>(float* to, const Packet4f& from, Index stride) +{ + float EIGEN_ALIGN16 ai[4]; + pstore<float>((float *)ai, from); + to[0*stride] = ai[0]; + to[1*stride] = ai[1]; + to[2*stride] = ai[2]; + 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]; @@ -353,52 +496,160 @@ template<> EIGEN_DEVICE_FUNC inline void pscatter<double, Packet2d>(double* to, } 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; + c.v4f[0] = a.v4f[0] + b.v4f[0]; + 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; + c.v4f[0] = a.v4f[0] - b.v4f[0]; + 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; + c.v4f[0] = a.v4f[0] * b.v4f[0]; + 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; + c.v4f[0] = a.v4f[0] / b.v4f[0]; + 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; + c.v4f[0] = -a.v4f[0]; + 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; + res.v4f[0] = vec_madd(a.v4f[0], b.v4f[0], c.v4f[0]); + 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; + res.v4f[0] = pmin(a.v4f[0], b.v4f[0]); + res.v4f[1] = pmin(a.v4f[1], b.v4f[1]); + 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; + res.v4f[0] = pmax(a.v4f[0], b.v4f[0]); + res.v4f[1] = pmax(a.v4f[1], b.v4f[1]); + 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; + res.v4f[0] = pand(a.v4f[0], b.v4f[0]); + res.v4f[1] = pand(a.v4f[1], b.v4f[1]); + 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]); + 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]); + 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; + res.v4f[0] = pandnot(a.v4f[0], b.v4f[0]); + res.v4f[1] = pandnot(a.v4f[1], b.v4f[1]); + return res; +} +template<> EIGEN_STRONG_INLINE Packet4f pround<Packet4f>(const Packet4f& a) +{ + Packet4f res; + res.v4f[0] = vec_round(a.v4f[0]); + 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; + res.v4f[0] = vec_ceil(a.v4f[0]); + 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; + res.v4f[0] = vec_floor(a.v4f[0]); + 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); } @@ -408,6 +659,14 @@ template<> EIGEN_STRONG_INLINE Packet4i ploaddup<Packet4i>(const int* from) return vec_perm(p, p, p16uc_DUPLICATE32_HI); } +template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float* from) +{ + Packet4f p = pload<Packet4f>(from); + p.v4f[1] = vec_splat(p.v4f[0], 1); + p.v4f[0] = vec_splat(p.v4f[0], 0); + return p; +} + template<> EIGEN_STRONG_INLINE Packet2d ploaddup<Packet2d>(const double* from) { Packet2d p = pload<Packet2d>(from); @@ -415,12 +674,15 @@ template<> EIGEN_STRONG_INLINE Packet2d ploaddup<Packet2d>(const double* from) } 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) @@ -433,8 +695,23 @@ 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(const Packet4i& a) { return vec_abs(a); } -template<> EIGEN_STRONG_INLINE Packet2d pabs(const Packet2d& a) { return vec_abs(a); } +template<> EIGEN_STRONG_INLINE Packet4f preverse(const Packet4f& a) +{ + Packet4f rev; + rev.v4f[0] = preverse<Packet2d>(a.v4f[1]); + rev.v4f[1] = preverse<Packet2d>(a.v4f[0]); + 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; + res.v4f[0] = pabs(a.v4f[0]); + res.v4f[1] = pabs(a.v4f[1]); + return res; +} template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a) { @@ -453,6 +730,13 @@ template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a) sum = padd<Packet2d>(a, b); return pfirst(sum); } +template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a) +{ + Packet2d sum; + sum = padd<Packet2d>(a.v4f[0], a.v4f[1]); + double first = predux<Packet2d>(sum); + return static_cast<float>(first); +} template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs) { @@ -493,6 +777,21 @@ template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs) 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) @@ -507,6 +806,12 @@ 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) { @@ -521,6 +826,14 @@ 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; + b = pmin<Packet2d>(a.v4f[0], a.v4f[1]); + res = pmin<Packet2d>(b, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(b), reinterpret_cast<Packet4i>(b), 8))); + return static_cast<float>(pfirst(res)); +} + // max template<> EIGEN_STRONG_INLINE int predux_max<Packet4i>(const Packet4i& a) { @@ -536,6 +849,14 @@ 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; + b = pmax<Packet2d>(a.v4f[0], a.v4f[1]); + res = pmax<Packet2d>(b, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(b), reinterpret_cast<Packet4i>(b), 8))); + 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]); @@ -556,12 +877,61 @@ ptranspose(PacketBlock<Packet2d,2>& kernel) { kernel.packet[1] = t1; } +/* Split the Packet4f PacketBlock into 4 Packet2d PacketBlocks and transpose each one + */ +EIGEN_DEVICE_FUNC inline void +ptranspose(PacketBlock<Packet4f,4>& kernel) { + PacketBlock<Packet2d,2> t0,t1,t2,t3; + // copy top-left 2x2 Packet2d block + t0.packet[0] = kernel.packet[0].v4f[0]; + t0.packet[1] = kernel.packet[1].v4f[0]; + + // copy top-right 2x2 Packet2d block + t1.packet[0] = kernel.packet[0].v4f[1]; + t1.packet[1] = kernel.packet[1].v4f[1]; + + // copy bottom-left 2x2 Packet2d block + t2.packet[0] = kernel.packet[2].v4f[0]; + t2.packet[1] = kernel.packet[3].v4f[0]; + + // copy bottom-right 2x2 Packet2d block + t3.packet[0] = kernel.packet[2].v4f[1]; + t3.packet[1] = kernel.packet[3].v4f[1]; + + // Transpose all 2x2 blocks + ptranspose(t0); + ptranspose(t1); + ptranspose(t2); + ptranspose(t3); + + // Copy back transposed blocks, but exchange t1 and t2 due to transposition + kernel.packet[0].v4f[0] = t0.packet[0]; + kernel.packet[0].v4f[1] = t2.packet[0]; + kernel.packet[1].v4f[0] = t0.packet[1]; + kernel.packet[1].v4f[1] = t2.packet[1]; + kernel.packet[2].v4f[0] = t1.packet[0]; + kernel.packet[2].v4f[1] = t3.packet[0]; + kernel.packet[3].v4f[0] = t1.packet[1]; + 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] }; + Packet2ul mask_hi = vec_cmpeq(select_hi, reinterpret_cast<Packet2ul>(p2l_ONE)); + Packet2ul mask_lo = vec_cmpeq(select_lo, reinterpret_cast<Packet2ul>(p2l_ONE)); + Packet4f result; + result.v4f[0] = vec_sel(elsePacket.v4f[0], thenPacket.v4f[0], mask_hi); + result.v4f[1] = vec_sel(elsePacket.v4f[1], thenPacket.v4f[1], mask_lo); + 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)); diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h index d577fd70d..ece64df15 100644 --- a/Eigen/src/Core/util/Macros.h +++ b/Eigen/src/Core/util/Macros.h @@ -13,7 +13,7 @@ #define EIGEN_WORLD_VERSION 3 #define EIGEN_MAJOR_VERSION 3 -#define EIGEN_MINOR_VERSION 0 +#define EIGEN_MINOR_VERSION 90 #define EIGEN_VERSION_AT_LEAST(x,y,z) (EIGEN_WORLD_VERSION>x || (EIGEN_WORLD_VERSION>=x && \ (EIGEN_MAJOR_VERSION>y || (EIGEN_MAJOR_VERSION>=y && \ |