diff options
Diffstat (limited to 'Eigen/src/Core/arch')
-rw-r--r-- | Eigen/src/Core/arch/AltiVec/CMakeLists.txt | 6 | ||||
-rw-r--r-- | Eigen/src/Core/arch/AltiVec/Complex.h | 217 | ||||
-rw-r--r-- | Eigen/src/Core/arch/AltiVec/PacketMath.h | 498 | ||||
-rw-r--r-- | Eigen/src/Core/arch/CMakeLists.txt | 4 | ||||
-rw-r--r-- | Eigen/src/Core/arch/Default/CMakeLists.txt | 6 | ||||
-rw-r--r-- | Eigen/src/Core/arch/Default/Settings.h | 49 | ||||
-rw-r--r-- | Eigen/src/Core/arch/NEON/CMakeLists.txt | 6 | ||||
-rw-r--r-- | Eigen/src/Core/arch/NEON/Complex.h | 259 | ||||
-rw-r--r-- | Eigen/src/Core/arch/NEON/PacketMath.h | 424 | ||||
-rw-r--r-- | Eigen/src/Core/arch/SSE/CMakeLists.txt | 6 | ||||
-rw-r--r-- | Eigen/src/Core/arch/SSE/Complex.h | 436 | ||||
-rw-r--r-- | Eigen/src/Core/arch/SSE/MathFunctions.h | 384 | ||||
-rw-r--r-- | Eigen/src/Core/arch/SSE/PacketMath.h | 632 |
13 files changed, 2927 insertions, 0 deletions
diff --git a/Eigen/src/Core/arch/AltiVec/CMakeLists.txt b/Eigen/src/Core/arch/AltiVec/CMakeLists.txt new file mode 100644 index 000000000..9f8d2e9c4 --- /dev/null +++ b/Eigen/src/Core/arch/AltiVec/CMakeLists.txt @@ -0,0 +1,6 @@ +FILE(GLOB Eigen_Core_arch_AltiVec_SRCS "*.h") + +INSTALL(FILES + ${Eigen_Core_arch_AltiVec_SRCS} + DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/Core/arch/AltiVec COMPONENT Devel +) diff --git a/Eigen/src/Core/arch/AltiVec/Complex.h b/Eigen/src/Core/arch/AltiVec/Complex.h new file mode 100644 index 000000000..68d9a2bff --- /dev/null +++ b/Eigen/src/Core/arch/AltiVec/Complex.h @@ -0,0 +1,217 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2010 Gael Guennebaud <gael.guennebaud@inria.fr> +// +// 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 +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_COMPLEX_ALTIVEC_H +#define EIGEN_COMPLEX_ALTIVEC_H + +namespace Eigen { + +namespace internal { + +static Packet4ui p4ui_CONJ_XOR = vec_mergeh((Packet4ui)p4i_ZERO, (Packet4ui)p4f_ZERO_);//{ 0x00000000, 0x80000000, 0x00000000, 0x80000000 }; +static Packet16uc p16uc_COMPLEX_RE = vec_sld((Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 0), (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 2), 8);//{ 0,1,2,3, 0,1,2,3, 8,9,10,11, 8,9,10,11 }; +static Packet16uc p16uc_COMPLEX_IM = vec_sld((Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 1), (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 3), 8);//{ 4,5,6,7, 4,5,6,7, 12,13,14,15, 12,13,14,15 }; +static Packet16uc p16uc_COMPLEX_REV = vec_sld(p16uc_REVERSE, p16uc_REVERSE, 8);//{ 4,5,6,7, 0,1,2,3, 12,13,14,15, 8,9,10,11 }; +static Packet16uc p16uc_COMPLEX_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_PSET_HI = (Packet16uc) vec_mergeh((Packet4ui) vec_splat((Packet4ui)p16uc_FORWARD, 0), (Packet4ui) vec_splat((Packet4ui)p16uc_FORWARD, 1));//{ 0,1,2,3, 4,5,6,7, 0,1,2,3, 4,5,6,7 }; +static Packet16uc p16uc_PSET_LO = (Packet16uc) vec_mergeh((Packet4ui) vec_splat((Packet4ui)p16uc_FORWARD, 2), (Packet4ui) vec_splat((Packet4ui)p16uc_FORWARD, 3));//{ 8,9,10,11, 12,13,14,15, 8,9,10,11, 12,13,14,15 }; + +//---------- float ---------- +struct Packet2cf +{ + EIGEN_STRONG_INLINE Packet2cf() {} + EIGEN_STRONG_INLINE explicit Packet2cf(const Packet4f& a) : v(a) {} + Packet4f v; +}; + +template<> struct packet_traits<std::complex<float> > : default_packet_traits +{ + typedef Packet2cf type; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size = 2, + + HasAdd = 1, + HasSub = 1, + HasMul = 1, + HasDiv = 1, + HasNegate = 1, + HasAbs = 0, + HasAbs2 = 0, + HasMin = 0, + HasMax = 0, + HasSetLinear = 0 + }; +}; + +template<> struct unpacket_traits<Packet2cf> { typedef std::complex<float> type; enum {size=2}; }; + +template<> EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<float>& from) +{ + Packet2cf res; + /* On AltiVec we cannot load 64-bit registers, so wa have to take care of alignment */ + if((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_PSET_HI); + return res; +} + +template<> EIGEN_STRONG_INLINE Packet2cf padd<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(vec_add(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet2cf psub<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(vec_sub(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet2cf pnegate(const Packet2cf& a) { return Packet2cf(pnegate(a.v)); } +template<> EIGEN_STRONG_INLINE Packet2cf pconj(const Packet2cf& a) { return Packet2cf((Packet4f)vec_xor((Packet4ui)a.v, p4ui_CONJ_XOR)); } + +template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, const Packet2cf& b) +{ + Packet4f v1, v2; + + // Permute and multiply the real parts of a and b + v1 = vec_perm(a.v, a.v, p16uc_COMPLEX_RE); + // Get the imaginary parts of a + v2 = vec_perm(a.v, a.v, p16uc_COMPLEX_IM); + // multiply a_re * b + v1 = vec_madd(v1, b.v, p4f_ZERO); + // multiply a_im * b and get the conjugate result + v2 = vec_madd(v2, b.v, p4f_ZERO); + v2 = (Packet4f) vec_xor((Packet4ui)v2, p4ui_CONJ_XOR); + // permute back to a proper order + v2 = vec_perm(v2, v2, p16uc_COMPLEX_REV); + + return Packet2cf(vec_add(v1, v2)); +} + +template<> EIGEN_STRONG_INLINE Packet2cf pand <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(vec_and(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet2cf por <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(vec_or(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet2cf pxor <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(vec_xor(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet2cf pandnot<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(vec_and(a.v, vec_nor(b.v,b.v))); } + +template<> EIGEN_STRONG_INLINE Packet2cf pload <Packet2cf>(const std::complex<float>* from) { EIGEN_DEBUG_ALIGNED_LOAD return Packet2cf(pload<Packet4f>((const float*)from)); } +template<> EIGEN_STRONG_INLINE Packet2cf ploadu<Packet2cf>(const std::complex<float>* from) { EIGEN_DEBUG_UNALIGNED_LOAD return Packet2cf(ploadu<Packet4f>((const float*)from)); } + +template<> EIGEN_STRONG_INLINE Packet2cf ploaddup<Packet2cf>(const std::complex<float>* from) +{ + return pset1<Packet2cf>(*from); +} + +template<> EIGEN_STRONG_INLINE void pstore <std::complex<float> >(std::complex<float> * to, const Packet2cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((float*)to, from.v); } +template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<float> >(std::complex<float> * to, const Packet2cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((float*)to, from.v); } + +template<> EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::complex<float> * addr) { vec_dstt((float *)addr, DST_CTRL(2,2,32), DST_CHAN); } + +template<> EIGEN_STRONG_INLINE std::complex<float> pfirst<Packet2cf>(const Packet2cf& a) +{ + std::complex<float> EIGEN_ALIGN16 res[2]; + pstore((float *)&res, a.v); + + return res[0]; +} + +template<> EIGEN_STRONG_INLINE Packet2cf preverse(const Packet2cf& a) +{ + Packet4f rev_a; + rev_a = vec_perm(a.v, a.v, p16uc_COMPLEX_REV2); + return Packet2cf(rev_a); +} + +template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(const Packet2cf& a) +{ + Packet4f b; + b = (Packet4f) vec_sld(a.v, a.v, 8); + b = padd(a.v, b); + return pfirst(Packet2cf(b)); +} + +template<> EIGEN_STRONG_INLINE Packet2cf preduxp<Packet2cf>(const Packet2cf* vecs) +{ + Packet4f b1, b2; + + b1 = (Packet4f) vec_sld(vecs[0].v, vecs[1].v, 8); + b2 = (Packet4f) vec_sld(vecs[1].v, vecs[0].v, 8); + b2 = (Packet4f) vec_sld(b2, b2, 8); + b2 = padd(b1, b2); + + return Packet2cf(b2); +} + +template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(const Packet2cf& a) +{ + Packet4f b; + Packet2cf prod; + b = (Packet4f) vec_sld(a.v, a.v, 8); + prod = pmul(a, Packet2cf(b)); + + return pfirst(prod); +} + +template<int Offset> +struct palign_impl<Offset,Packet2cf> +{ + static EIGEN_STRONG_INLINE void run(Packet2cf& first, const Packet2cf& second) + { + if (Offset==1) + { + first.v = vec_sld(first.v, second.v, 8); + } + } +}; + +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 Packet2cf pdiv<Packet2cf>(const Packet2cf& a, const Packet2cf& b) +{ + // TODO optimize it for AltiVec + Packet2cf res = conj_helper<Packet2cf,Packet2cf,false,true>().pmul(a,b); + Packet4f s = vec_madd(b.v, b.v, p4f_ZERO); + return Packet2cf(pdiv(res.v, vec_add(s,vec_perm(s, s, p16uc_COMPLEX_REV)))); +} + +template<> EIGEN_STRONG_INLINE Packet2cf pcplxflip<Packet2cf>(const Packet2cf& x) +{ + return Packet2cf(vec_perm(x.v, x.v, p16uc_COMPLEX_REV)); +} + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_COMPLEX_ALTIVEC_H diff --git a/Eigen/src/Core/arch/AltiVec/PacketMath.h b/Eigen/src/Core/arch/AltiVec/PacketMath.h new file mode 100644 index 000000000..75de19311 --- /dev/null +++ b/Eigen/src/Core/arch/AltiVec/PacketMath.h @@ -0,0 +1,498 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008 Konstantinos Margaritis <markos@codex.gr> +// +// 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 +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_PACKET_MATH_ALTIVEC_H +#define EIGEN_PACKET_MATH_ALTIVEC_H + +namespace Eigen { + +namespace internal { + +#ifndef EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD +#define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 4 +#endif + +#ifndef EIGEN_HAS_FUSE_CJMADD +#define EIGEN_HAS_FUSE_CJMADD 1 +#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 16 +#endif + +typedef __vector float Packet4f; +typedef __vector int Packet4i; +typedef __vector unsigned int Packet4ui; +typedef __vector __bool int Packet4bi; +typedef __vector short int Packet8i; +typedef __vector unsigned char Packet16uc; + +// We don't want to write the same code all the time, but we need to reuse the constants +// and it doesn't really work to declare them global, so we define macros instead + +#define _EIGEN_DECLARE_CONST_FAST_Packet4f(NAME,X) \ + Packet4f p4f_##NAME = (Packet4f) vec_splat_s32(X) + +#define _EIGEN_DECLARE_CONST_FAST_Packet4i(NAME,X) \ + Packet4i p4i_##NAME = 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) \ + Packet4f p4f_##NAME = vreinterpretq_f32_u32(pset1<int>(X)) + +#define _EIGEN_DECLARE_CONST_Packet4i(NAME,X) \ + Packet4i p4i_##NAME = pset1<Packet4i>(X) + +#define DST_CHAN 1 +#define DST_CTRL(size, count, stride) (((size) << 24) | ((count) << 16) | (stride)) + +// Define global static constants: +static Packet4f p4f_COUNTDOWN = { 3.0, 2.0, 1.0, 0.0 }; +static Packet4i p4i_COUNTDOWN = { 3, 2, 1, 0 }; +static Packet16uc p16uc_REVERSE = {12,13,14,15, 8,9,10,11, 4,5,6,7, 0,1,2,3}; +static Packet16uc p16uc_FORWARD = vec_lvsl(0, (float*)0); +static Packet16uc p16uc_DUPLICATE = {0,1,2,3, 0,1,2,3, 4,5,6,7, 4,5,6,7}; + +static _EIGEN_DECLARE_CONST_FAST_Packet4f(ZERO, 0); +static _EIGEN_DECLARE_CONST_FAST_Packet4i(ZERO, 0); +static _EIGEN_DECLARE_CONST_FAST_Packet4i(ONE,1); +static _EIGEN_DECLARE_CONST_FAST_Packet4i(MINUS16,-16); +static _EIGEN_DECLARE_CONST_FAST_Packet4i(MINUS1,-1); +static Packet4f p4f_ONE = vec_ctf(p4i_ONE, 0); +static Packet4f p4f_ZERO_ = (Packet4f) vec_sl((Packet4ui)p4i_MINUS1, (Packet4ui)p4i_MINUS1); + +template<> struct packet_traits<float> : default_packet_traits +{ + typedef Packet4f type; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size=4, + + // FIXME check the Has* + HasSin = 0, + HasCos = 0, + HasLog = 0, + HasExp = 0, + HasSqrt = 0 + }; +}; +template<> struct packet_traits<int> : default_packet_traits +{ + typedef Packet4i type; + enum { + // FIXME check the Has* + Vectorizable = 1, + AlignedOnScalar = 1, + size=4 + }; +}; + +template<> struct unpacket_traits<Packet4f> { typedef float type; enum {size=4}; }; +template<> struct unpacket_traits<Packet4i> { typedef int type; enum {size=4}; }; +/* +inline std::ostream & operator <<(std::ostream & s, const Packet4f & v) +{ + union { + Packet4f v; + float n[4]; + } vt; + vt.v = v; + s << vt.n[0] << ", " << vt.n[1] << ", " << vt.n[2] << ", " << vt.n[3]; + return s; +} + +inline std::ostream & operator <<(std::ostream & s, const Packet4i & v) +{ + union { + Packet4i v; + int n[4]; + } vt; + vt.v = v; + s << vt.n[0] << ", " << vt.n[1] << ", " << vt.n[2] << ", " << vt.n[3]; + return s; +} + +inline std::ostream & operator <<(std::ostream & s, const Packet4ui & v) +{ + union { + Packet4ui v; + unsigned int n[4]; + } vt; + vt.v = v; + s << vt.n[0] << ", " << vt.n[1] << ", " << vt.n[2] << ", " << vt.n[3]; + return s; +} + +inline std::ostream & operator <<(std::ostream & s, const Packetbi & v) +{ + union { + Packet4bi v; + unsigned int n[4]; + } vt; + vt.v = v; + s << vt.n[0] << ", " << vt.n[1] << ", " << vt.n[2] << ", " << vt.n[3]; + return s; +} +*/ +template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float& from) { + // Taken from http://developer.apple.com/hardwaredrivers/ve/alignment.html + float EIGEN_ALIGN16 af[4]; + af[0] = from; + Packet4f vc = vec_ld(0, af); + vc = vec_splat(vc, 0); + return vc; +} + +template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int& from) { + int EIGEN_ALIGN16 ai[4]; + ai[0] = from; + Packet4i vc = vec_ld(0, ai); + vc = vec_splat(vc, 0); + return vc; +} + +template<> EIGEN_STRONG_INLINE Packet4f plset<float>(const float& a) { return vec_add(pset1<Packet4f>(a), p4f_COUNTDOWN); } +template<> EIGEN_STRONG_INLINE Packet4i plset<int>(const int& a) { return vec_add(pset1<Packet4i>(a), p4i_COUNTDOWN); } + +template<> EIGEN_STRONG_INLINE Packet4f padd<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_add(a,b); } +template<> EIGEN_STRONG_INLINE Packet4i padd<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_add(a,b); } + +template<> EIGEN_STRONG_INLINE Packet4f psub<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_sub(a,b); } +template<> EIGEN_STRONG_INLINE Packet4i psub<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_sub(a,b); } + +template<> EIGEN_STRONG_INLINE Packet4f pnegate(const Packet4f& a) { return psub<Packet4f>(p4f_ZERO, a); } +template<> EIGEN_STRONG_INLINE Packet4i pnegate(const Packet4i& a) { return psub<Packet4i>(p4i_ZERO, a); } + +template<> EIGEN_STRONG_INLINE Packet4f pmul<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_madd(a,b,p4f_ZERO); } +/* Commented out: it's actually slower than processing it scalar + * +template<> EIGEN_STRONG_INLINE Packet4i pmul<Packet4i>(const Packet4i& a, const Packet4i& b) +{ + // Detailed in: http://freevec.org/content/32bit_signed_integer_multiplication_altivec + //Set up constants, variables + Packet4i a1, b1, bswap, low_prod, high_prod, prod, prod_, v1sel; + + // Get the absolute values + a1 = vec_abs(a); + b1 = vec_abs(b); + + // Get the signs using xor + Packet4bi sgn = (Packet4bi) vec_cmplt(vec_xor(a, b), p4i_ZERO); + + // Do the multiplication for the asbolute values. + bswap = (Packet4i) vec_rl((Packet4ui) b1, (Packet4ui) p4i_MINUS16 ); + low_prod = vec_mulo((Packet8i) a1, (Packet8i)b1); + high_prod = vec_msum((Packet8i) a1, (Packet8i) bswap, p4i_ZERO); + high_prod = (Packet4i) vec_sl((Packet4ui) high_prod, (Packet4ui) p4i_MINUS16); + prod = vec_add( low_prod, high_prod ); + + // NOR the product and select only the negative elements according to the sign mask + prod_ = vec_nor(prod, prod); + prod_ = vec_sel(p4i_ZERO, prod_, sgn); + + // Add 1 to the result to get the negative numbers + v1sel = vec_sel(p4i_ZERO, p4i_ONE, sgn); + prod_ = vec_add(prod_, v1sel); + + // Merge the results back to the final vector. + prod = vec_sel(prod, prod_, sgn); + + return prod; +} +*/ +template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const Packet4f& b) +{ + Packet4f t, y_0, y_1, res; + + // Altivec does not offer a divide instruction, we have to do a reciprocal approximation + y_0 = vec_re(b); + + // Do one Newton-Raphson iteration to get the needed accuracy + t = vec_nmsub(y_0, b, p4f_ONE); + y_1 = vec_madd(y_0, t, y_0); + + res = vec_madd(a, y_1, p4f_ZERO); + return res; +} + +template<> EIGEN_STRONG_INLINE Packet4i pdiv<Packet4i>(const Packet4i& /*a*/, const Packet4i& /*b*/) +{ eigen_assert(false && "packet integer division are not supported by AltiVec"); + return pset1<Packet4i>(0); +} + +// for some weird raisons, it has to be overloaded for packet of integers +template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) { return vec_madd(a, b, c); } +template<> EIGEN_STRONG_INLINE Packet4i pmadd(const Packet4i& a, const Packet4i& b, const Packet4i& c) { return padd(pmul(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 Packet4i pmin<Packet4i>(const Packet4i& a, const Packet4i& 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 Packet4i pmax<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_max(a, b); } + +// Logical Operations are not supported for float, so we have to reinterpret casts using NEON intrinsics +template<> EIGEN_STRONG_INLINE Packet4f pand<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_and(a, b); } +template<> EIGEN_STRONG_INLINE Packet4i pand<Packet4i>(const Packet4i& a, const Packet4i& 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 Packet4i por<Packet4i>(const Packet4i& a, const Packet4i& 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 Packet4i pxor<Packet4i>(const Packet4i& a, const Packet4i& 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 Packet4i pandnot<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_and(a, vec_nor(b, b)); } + +template<> EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(const float* from) { EIGEN_DEBUG_ALIGNED_LOAD return vec_ld(0, from); } +template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int* from) { EIGEN_DEBUG_ALIGNED_LOAD return vec_ld(0, from); } + +template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from) +{ + EIGEN_DEBUG_ALIGNED_LOAD + // Taken from http://developer.apple.com/hardwaredrivers/ve/alignment.html + Packet16uc MSQ, LSQ; + Packet16uc mask; + MSQ = vec_ld(0, (unsigned char *)from); // most significant quadword + LSQ = vec_ld(15, (unsigned char *)from); // least significant quadword + mask = vec_lvsl(0, from); // create the permute mask + return (Packet4f) vec_perm(MSQ, LSQ, mask); // align the data + +} +template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from) +{ + EIGEN_DEBUG_ALIGNED_LOAD + // Taken from http://developer.apple.com/hardwaredrivers/ve/alignment.html + Packet16uc MSQ, LSQ; + Packet16uc mask; + MSQ = vec_ld(0, (unsigned char *)from); // most significant quadword + LSQ = vec_ld(15, (unsigned char *)from); // least significant quadword + mask = vec_lvsl(0, from); // create the permute mask + return (Packet4i) vec_perm(MSQ, LSQ, mask); // align the data +} + +template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float* from) +{ + Packet4f p; + if((ptrdiff_t(&from) % 16) == 0) p = pload<Packet4f>(from); + else p = ploadu<Packet4f>(from); + return vec_perm(p, p, p16uc_DUPLICATE); +} +template<> EIGEN_STRONG_INLINE Packet4i ploaddup<Packet4i>(const int* from) +{ + Packet4i p; + if((ptrdiff_t(&from) % 16) == 0) p = pload<Packet4i>(from); + else p = ploadu<Packet4i>(from); + return vec_perm(p, p, p16uc_DUPLICATE); +} + +template<> EIGEN_STRONG_INLINE void pstore<float>(float* to, const Packet4f& from) { EIGEN_DEBUG_ALIGNED_STORE vec_st(from, 0, to); } +template<> EIGEN_STRONG_INLINE void pstore<int>(int* to, const Packet4i& from) { EIGEN_DEBUG_ALIGNED_STORE vec_st(from, 0, to); } + +template<> EIGEN_STRONG_INLINE void pstoreu<float>(float* to, const Packet4f& from) +{ + EIGEN_DEBUG_UNALIGNED_STORE + // Taken from http://developer.apple.com/hardwaredrivers/ve/alignment.html + // Warning: not thread safe! + Packet16uc MSQ, LSQ, edges; + Packet16uc edgeAlign, align; + + MSQ = vec_ld(0, (unsigned char *)to); // most significant quadword + LSQ = vec_ld(15, (unsigned char *)to); // least significant quadword + edgeAlign = vec_lvsl(0, to); // permute map to extract edges + edges=vec_perm(LSQ,MSQ,edgeAlign); // extract the edges + align = vec_lvsr( 0, to ); // permute map to misalign data + MSQ = vec_perm(edges,(Packet16uc)from,align); // misalign the data (MSQ) + LSQ = vec_perm((Packet16uc)from,edges,align); // misalign the data (LSQ) + vec_st( LSQ, 15, (unsigned char *)to ); // Store the LSQ part first + vec_st( MSQ, 0, (unsigned char *)to ); // Store the MSQ part +} +template<> EIGEN_STRONG_INLINE void pstoreu<int>(int* to, const Packet4i& from) +{ + EIGEN_DEBUG_UNALIGNED_STORE + // Taken from http://developer.apple.com/hardwaredrivers/ve/alignment.html + // Warning: not thread safe! + Packet16uc MSQ, LSQ, edges; + Packet16uc edgeAlign, align; + + MSQ = vec_ld(0, (unsigned char *)to); // most significant quadword + LSQ = vec_ld(15, (unsigned char *)to); // least significant quadword + edgeAlign = vec_lvsl(0, to); // permute map to extract edges + edges=vec_perm(LSQ, MSQ, edgeAlign); // extract the edges + align = vec_lvsr( 0, to ); // permute map to misalign data + MSQ = vec_perm(edges, (Packet16uc) from, align); // misalign the data (MSQ) + LSQ = vec_perm((Packet16uc) from, edges, align); // misalign the data (LSQ) + vec_st( LSQ, 15, (unsigned char *)to ); // Store the LSQ part first + vec_st( MSQ, 0, (unsigned char *)to ); // Store the MSQ part +} + +template<> EIGEN_STRONG_INLINE void prefetch<float>(const float* addr) { vec_dstt(addr, DST_CTRL(2,2,32), DST_CHAN); } +template<> EIGEN_STRONG_INLINE void prefetch<int>(const int* addr) { vec_dstt(addr, DST_CTRL(2,2,32), DST_CHAN); } + +template<> EIGEN_STRONG_INLINE float pfirst<Packet4f>(const Packet4f& a) { float EIGEN_ALIGN16 x[4]; vec_st(a, 0, x); return x[0]; } +template<> EIGEN_STRONG_INLINE int pfirst<Packet4i>(const Packet4i& a) { int EIGEN_ALIGN16 x[4]; vec_st(a, 0, x); return x[0]; } + +template<> EIGEN_STRONG_INLINE Packet4f preverse(const Packet4f& a) { return (Packet4f)vec_perm((Packet16uc)a,(Packet16uc)a, p16uc_REVERSE); } +template<> EIGEN_STRONG_INLINE Packet4i preverse(const Packet4i& a) { return (Packet4i)vec_perm((Packet16uc)a,(Packet16uc)a, p16uc_REVERSE); } + +template<> EIGEN_STRONG_INLINE Packet4f pabs(const Packet4f& a) { return vec_abs(a); } +template<> EIGEN_STRONG_INLINE Packet4i pabs(const Packet4i& a) { return vec_abs(a); } + +template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a) +{ + Packet4f b, sum; + b = (Packet4f) vec_sld(a, a, 8); + sum = vec_add(a, b); + b = (Packet4f) vec_sld(sum, sum, 4); + sum = vec_add(sum, b); + return pfirst(sum); +} + +template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs) +{ + Packet4f 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] = vec_add(sum[0], sum[1]); + // Lines 2+3 + sum[1] = vec_add(sum[2], sum[3]); + // Add the results + sum[0] = vec_add(sum[0], sum[1]); + + return sum[0]; +} + +template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a) +{ + Packet4i sum; + sum = vec_sums(a, p4i_ZERO); + sum = vec_sld(sum, p4i_ZERO, 12); + return pfirst(sum); +} + +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] = vec_add(sum[0], sum[1]); + // Lines 2+3 + sum[1] = vec_add(sum[2], sum[3]); + // Add the results + sum[0] = vec_add(sum[0], sum[1]); + + return sum[0]; +} + +// Other reduction functions: +// mul +template<> EIGEN_STRONG_INLINE float predux_mul<Packet4f>(const Packet4f& a) +{ + Packet4f prod; + prod = pmul(a, (Packet4f)vec_sld(a, a, 8)); + return pfirst(pmul(prod, (Packet4f)vec_sld(prod, prod, 4))); +} + +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]; +} + +// min +template<> EIGEN_STRONG_INLINE float predux_min<Packet4f>(const Packet4f& a) +{ + Packet4f b, res; + b = vec_min(a, vec_sld(a, a, 8)); + res = vec_min(b, vec_sld(b, b, 4)); + return pfirst(res); +} + +template<> EIGEN_STRONG_INLINE int predux_min<Packet4i>(const Packet4i& a) +{ + Packet4i b, res; + b = vec_min(a, vec_sld(a, a, 8)); + res = vec_min(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 = vec_max(a, vec_sld(a, a, 8)); + res = vec_max(b, vec_sld(b, b, 4)); + return pfirst(res); +} + +template<> EIGEN_STRONG_INLINE int predux_max<Packet4i>(const Packet4i& a) +{ + Packet4i b, res; + b = vec_max(a, vec_sld(a, a, 8)); + res = vec_max(b, vec_sld(b, b, 4)); + return pfirst(res); +} + +template<int Offset> +struct palign_impl<Offset,Packet4f> +{ + static EIGEN_STRONG_INLINE void run(Packet4f& first, const Packet4f& second) + { + if (Offset!=0) + first = vec_sld(first, second, Offset*4); + } +}; + +template<int Offset> +struct palign_impl<Offset,Packet4i> +{ + static EIGEN_STRONG_INLINE void run(Packet4i& first, const Packet4i& second) + { + if (Offset!=0) + first = vec_sld(first, second, Offset*4); + } +}; + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_PACKET_MATH_ALTIVEC_H diff --git a/Eigen/src/Core/arch/CMakeLists.txt b/Eigen/src/Core/arch/CMakeLists.txt new file mode 100644 index 000000000..8456dec15 --- /dev/null +++ b/Eigen/src/Core/arch/CMakeLists.txt @@ -0,0 +1,4 @@ +ADD_SUBDIRECTORY(SSE) +ADD_SUBDIRECTORY(AltiVec) +ADD_SUBDIRECTORY(NEON) +ADD_SUBDIRECTORY(Default) diff --git a/Eigen/src/Core/arch/Default/CMakeLists.txt b/Eigen/src/Core/arch/Default/CMakeLists.txt new file mode 100644 index 000000000..339c091d1 --- /dev/null +++ b/Eigen/src/Core/arch/Default/CMakeLists.txt @@ -0,0 +1,6 @@ +FILE(GLOB Eigen_Core_arch_Default_SRCS "*.h") + +INSTALL(FILES + ${Eigen_Core_arch_Default_SRCS} + DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/Core/arch/Default COMPONENT Devel +) diff --git a/Eigen/src/Core/arch/Default/Settings.h b/Eigen/src/Core/arch/Default/Settings.h new file mode 100644 index 000000000..097373c84 --- /dev/null +++ b/Eigen/src/Core/arch/Default/Settings.h @@ -0,0 +1,49 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2010 Gael Guennebaud <gael.guennebaud@inria.fr> +// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com> +// +// 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 +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + + +/* All the parameters defined in this file can be specialized in the + * architecture specific files, and/or by the user. + * More to come... */ + +#ifndef EIGEN_DEFAULT_SETTINGS_H +#define EIGEN_DEFAULT_SETTINGS_H + +/** Defines the maximal loop size to enable meta unrolling of loops. + * Note that the value here is expressed in Eigen's own notion of "number of FLOPS", + * it does not correspond to the number of iterations or the number of instructions + */ +#ifndef EIGEN_UNROLLING_LIMIT +#define EIGEN_UNROLLING_LIMIT 100 +#endif + +/** Defines the threshold between a "small" and a "large" matrix. + * This threshold is mainly used to select the proper product implementation. + */ +#ifndef EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD +#define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 8 +#endif + +/** Defines the maximal width of the blocks used in the triangular product and solver + * for vectors (level 2 blas xTRMV and xTRSV). The default is 8. + */ +#ifndef EIGEN_TUNE_TRIANGULAR_PANEL_WIDTH +#define EIGEN_TUNE_TRIANGULAR_PANEL_WIDTH 8 +#endif + + +/** Defines the default number of registers available for that architecture. + * Currently it must be 8 or 16. Other values will fail. + */ +#ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS +#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 8 +#endif + +#endif // EIGEN_DEFAULT_SETTINGS_H diff --git a/Eigen/src/Core/arch/NEON/CMakeLists.txt b/Eigen/src/Core/arch/NEON/CMakeLists.txt new file mode 100644 index 000000000..fd4d4af50 --- /dev/null +++ b/Eigen/src/Core/arch/NEON/CMakeLists.txt @@ -0,0 +1,6 @@ +FILE(GLOB Eigen_Core_arch_NEON_SRCS "*.h") + +INSTALL(FILES + ${Eigen_Core_arch_NEON_SRCS} + DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/Core/arch/NEON COMPONENT Devel +) diff --git a/Eigen/src/Core/arch/NEON/Complex.h b/Eigen/src/Core/arch/NEON/Complex.h new file mode 100644 index 000000000..795b4be73 --- /dev/null +++ b/Eigen/src/Core/arch/NEON/Complex.h @@ -0,0 +1,259 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2010 Gael Guennebaud <gael.guennebaud@inria.fr> +// +// 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 +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_COMPLEX_NEON_H +#define EIGEN_COMPLEX_NEON_H + +namespace Eigen { + +namespace internal { + +static uint32x4_t p4ui_CONJ_XOR = EIGEN_INIT_NEON_PACKET4(0x00000000, 0x80000000, 0x00000000, 0x80000000); +static uint32x2_t p2ui_CONJ_XOR = EIGEN_INIT_NEON_PACKET2(0x00000000, 0x80000000); + +//---------- float ---------- +struct Packet2cf +{ + EIGEN_STRONG_INLINE Packet2cf() {} + EIGEN_STRONG_INLINE explicit Packet2cf(const Packet4f& a) : v(a) {} + Packet4f v; +}; + +template<> struct packet_traits<std::complex<float> > : default_packet_traits +{ + typedef Packet2cf type; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size = 2, + + HasAdd = 1, + HasSub = 1, + HasMul = 1, + HasDiv = 1, + HasNegate = 1, + HasAbs = 0, + HasAbs2 = 0, + HasMin = 0, + HasMax = 0, + HasSetLinear = 0 + }; +}; + +template<> struct unpacket_traits<Packet2cf> { typedef std::complex<float> type; enum {size=2}; }; + +template<> EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<float>& from) +{ + float32x2_t r64; + r64 = vld1_f32((float *)&from); + + return Packet2cf(vcombine_f32(r64, r64)); +} + +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<> EIGEN_STRONG_INLINE Packet2cf pconj(const Packet2cf& a) +{ + Packet4ui b = vreinterpretq_u32_f32(a.v); + return Packet2cf(vreinterpretq_f32_u32(veorq_u32(b, p4ui_CONJ_XOR))); +} + +template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, const Packet2cf& b) +{ + Packet4f v1, v2; + float32x2_t a_lo, a_hi; + + // Get the real values of a | a1_re | a1_re | a2_re | a2_re | + v1 = vcombine_f32(vdup_lane_f32(vget_low_f32(a.v), 0), vdup_lane_f32(vget_high_f32(a.v), 0)); + // Get the real values of a | a1_im | a1_im | a2_im | a2_im | + v2 = vcombine_f32(vdup_lane_f32(vget_low_f32(a.v), 1), vdup_lane_f32(vget_high_f32(a.v), 1)); + // Multiply the real a with b + v1 = vmulq_f32(v1, b.v); + // Multiply the imag a with b + v2 = vmulq_f32(v2, b.v); + // Conjugate v2 + v2 = vreinterpretq_f32_u32(veorq_u32(vreinterpretq_u32_f32(v2), p4ui_CONJ_XOR)); + // Swap real/imag elements in v2. + a_lo = vrev64_f32(vget_low_f32(v2)); + a_hi = vrev64_f32(vget_high_f32(v2)); + v2 = vcombine_f32(a_lo, a_hi); + // Add and return the result + return Packet2cf(vaddq_f32(v1, v2)); +} + +template<> EIGEN_STRONG_INLINE Packet2cf pand <Packet2cf>(const Packet2cf& a, const Packet2cf& b) +{ + return Packet2cf(vreinterpretq_f32_u32(vorrq_u32(vreinterpretq_u32_f32(a.v),vreinterpretq_u32_f32(b.v)))); +} +template<> EIGEN_STRONG_INLINE Packet2cf por <Packet2cf>(const Packet2cf& a, const Packet2cf& b) +{ + return Packet2cf(vreinterpretq_f32_u32(vorrq_u32(vreinterpretq_u32_f32(a.v),vreinterpretq_u32_f32(b.v)))); +} +template<> EIGEN_STRONG_INLINE Packet2cf pxor <Packet2cf>(const Packet2cf& a, const Packet2cf& b) +{ + return Packet2cf(vreinterpretq_f32_u32(veorq_u32(vreinterpretq_u32_f32(a.v),vreinterpretq_u32_f32(b.v)))); +} +template<> EIGEN_STRONG_INLINE Packet2cf pandnot<Packet2cf>(const Packet2cf& a, const Packet2cf& b) +{ + return Packet2cf(vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(a.v),vreinterpretq_u32_f32(b.v)))); +} + +template<> EIGEN_STRONG_INLINE Packet2cf pload<Packet2cf>(const std::complex<float>* from) { EIGEN_DEBUG_ALIGNED_LOAD return Packet2cf(pload<Packet4f>((const float*)from)); } +template<> EIGEN_STRONG_INLINE Packet2cf ploadu<Packet2cf>(const std::complex<float>* from) { EIGEN_DEBUG_UNALIGNED_LOAD return Packet2cf(ploadu<Packet4f>((const float*)from)); } + +template<> EIGEN_STRONG_INLINE Packet2cf ploaddup<Packet2cf>(const std::complex<float>* from) { return pset1<Packet2cf>(*from); } + +template<> EIGEN_STRONG_INLINE void pstore <std::complex<float> >(std::complex<float> * to, const Packet2cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((float*)to, from.v); } +template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<float> >(std::complex<float> * to, const Packet2cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((float*)to, from.v); } + +template<> EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::complex<float> * addr) { __pld((float *)addr); } + +template<> EIGEN_STRONG_INLINE std::complex<float> pfirst<Packet2cf>(const Packet2cf& a) +{ + std::complex<float> EIGEN_ALIGN16 x[2]; + vst1q_f32((float *)x, a.v); + return x[0]; +} + +template<> EIGEN_STRONG_INLINE Packet2cf preverse(const Packet2cf& a) +{ + float32x2_t a_lo, a_hi; + Packet4f a_r128; + + a_lo = vget_low_f32(a.v); + a_hi = vget_high_f32(a.v); + a_r128 = vcombine_f32(a_hi, a_lo); + + return Packet2cf(a_r128); +} + +template<> EIGEN_STRONG_INLINE Packet2cf pcplxflip<Packet2cf>(const Packet2cf& a) +{ + return Packet2cf(vrev64q_f32(a.v)); +} + +template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(const Packet2cf& a) +{ + float32x2_t a1, a2; + std::complex<float> s; + + a1 = vget_low_f32(a.v); + a2 = vget_high_f32(a.v); + a2 = vadd_f32(a1, a2); + vst1_f32((float *)&s, a2); + + return s; +} + +template<> EIGEN_STRONG_INLINE Packet2cf preduxp<Packet2cf>(const Packet2cf* vecs) +{ + Packet4f sum1, sum2, sum; + + // Add the first two 64-bit float32x2_t of vecs[0] + sum1 = vcombine_f32(vget_low_f32(vecs[0].v), vget_low_f32(vecs[1].v)); + sum2 = vcombine_f32(vget_high_f32(vecs[0].v), vget_high_f32(vecs[1].v)); + sum = vaddq_f32(sum1, sum2); + + return Packet2cf(sum); +} + +template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(const Packet2cf& a) +{ + float32x2_t a1, a2, v1, v2, prod; + std::complex<float> s; + + a1 = vget_low_f32(a.v); + a2 = vget_high_f32(a.v); + // Get the real values of a | a1_re | a1_re | a2_re | a2_re | + v1 = vdup_lane_f32(a1, 0); + // Get the real values of a | a1_im | a1_im | a2_im | a2_im | + v2 = vdup_lane_f32(a1, 1); + // Multiply the real a with b + v1 = vmul_f32(v1, a2); + // Multiply the imag a with b + v2 = vmul_f32(v2, a2); + // Conjugate v2 + v2 = vreinterpret_f32_u32(veor_u32(vreinterpret_u32_f32(v2), p2ui_CONJ_XOR)); + // Swap real/imag elements in v2. + v2 = vrev64_f32(v2); + // Add v1, v2 + prod = vadd_f32(v1, v2); + + vst1_f32((float *)&s, prod); + + return s; +} + +template<int Offset> +struct palign_impl<Offset,Packet2cf> +{ + EIGEN_STRONG_INLINE static void run(Packet2cf& first, const Packet2cf& second) + { + if (Offset==1) + { + first.v = vextq_f32(first.v, second.v, 2); + } + } +}; + +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 Packet2cf pdiv<Packet2cf>(const Packet2cf& a, const Packet2cf& b) +{ + // TODO optimize it for AltiVec + Packet2cf res = conj_helper<Packet2cf,Packet2cf,false,true>().pmul(a,b); + Packet4f s, rev_s; + float32x2_t a_lo, a_hi; + + // this computes the norm + s = vmulq_f32(b.v, b.v); + a_lo = vrev64_f32(vget_low_f32(s)); + a_hi = vrev64_f32(vget_high_f32(s)); + rev_s = vcombine_f32(a_lo, a_hi); + + return Packet2cf(pdiv(res.v, vaddq_f32(s,rev_s))); +} + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_COMPLEX_NEON_H diff --git a/Eigen/src/Core/arch/NEON/PacketMath.h b/Eigen/src/Core/arch/NEON/PacketMath.h new file mode 100644 index 000000000..a20250f7c --- /dev/null +++ b/Eigen/src/Core/arch/NEON/PacketMath.h @@ -0,0 +1,424 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2009 Gael Guennebaud <gael.guennebaud@inria.fr> +// Copyright (C) 2010 Konstantinos Margaritis <markos@codex.gr> +// Heavily based on Gael's SSE version. +// +// 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 +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_PACKET_MATH_NEON_H +#define EIGEN_PACKET_MATH_NEON_H + +namespace Eigen { + +namespace internal { + +#ifndef EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD +#define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 8 +#endif + +// FIXME NEON has 16 quad registers, but since the current register allocator +// is so bad, it is much better to reduce it to 8 +#ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS +#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 8 +#endif + +typedef float32x4_t Packet4f; +typedef int32x4_t Packet4i; +typedef uint32x4_t Packet4ui; + +#define _EIGEN_DECLARE_CONST_Packet4f(NAME,X) \ + const Packet4f p4f_##NAME = pset1<Packet4f>(X) + +#define _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(NAME,X) \ + const Packet4f p4f_##NAME = vreinterpretq_f32_u32(pset1<int>(X)) + +#define _EIGEN_DECLARE_CONST_Packet4i(NAME,X) \ + const Packet4i p4i_##NAME = pset1<Packet4i>(X) + +#if defined(__llvm__) && !defined(__clang__) + //Special treatment for Apple's llvm-gcc, its NEON packet types are unions + #define EIGEN_INIT_NEON_PACKET2(X, Y) {{X, Y}} + #define EIGEN_INIT_NEON_PACKET4(X, Y, Z, W) {{X, Y, Z, W}} +#else + //Default initializer for packets + #define EIGEN_INIT_NEON_PACKET2(X, Y) {X, Y} + #define EIGEN_INIT_NEON_PACKET4(X, Y, Z, W) {X, Y, Z, W} +#endif + +#ifndef __pld +#define __pld(x) asm volatile ( " pld [%[addr]]\n" :: [addr] "r" (x) : "cc" ); +#endif + +template<> struct packet_traits<float> : default_packet_traits +{ + typedef Packet4f type; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size = 4, + + HasDiv = 1, + // FIXME check the Has* + HasSin = 0, + HasCos = 0, + HasLog = 0, + HasExp = 0, + HasSqrt = 0 + }; +}; +template<> struct packet_traits<int> : default_packet_traits +{ + typedef Packet4i type; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size=4 + // FIXME check the Has* + }; +}; + +#if EIGEN_GNUC_AT_MOST(4,4) && !defined(__llvm__) +// workaround gcc 4.2, 4.3 and 4.4 compilatin issue +EIGEN_STRONG_INLINE float32x4_t vld1q_f32(const float* x) { return ::vld1q_f32((const float32_t*)x); } +EIGEN_STRONG_INLINE float32x2_t vld1_f32 (const float* x) { return ::vld1_f32 ((const float32_t*)x); } +EIGEN_STRONG_INLINE void vst1q_f32(float* to, float32x4_t from) { ::vst1q_f32((float32_t*)to,from); } +EIGEN_STRONG_INLINE void vst1_f32 (float* to, float32x2_t from) { ::vst1_f32 ((float32_t*)to,from); } +#endif + +template<> struct unpacket_traits<Packet4f> { typedef float type; enum {size=4}; }; +template<> struct unpacket_traits<Packet4i> { typedef int type; enum {size=4}; }; + +template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float& from) { return vdupq_n_f32(from); } +template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int& from) { return vdupq_n_s32(from); } + +template<> EIGEN_STRONG_INLINE Packet4f plset<float>(const float& a) +{ + Packet4f countdown = EIGEN_INIT_NEON_PACKET4(0, 1, 2, 3); + return vaddq_f32(pset1<Packet4f>(a), countdown); +} +template<> EIGEN_STRONG_INLINE Packet4i plset<int>(const int& a) +{ + Packet4i countdown = EIGEN_INIT_NEON_PACKET4(0, 1, 2, 3); + return vaddq_s32(pset1<Packet4i>(a), countdown); +} + +template<> EIGEN_STRONG_INLINE Packet4f padd<Packet4f>(const Packet4f& a, const Packet4f& b) { return vaddq_f32(a,b); } +template<> EIGEN_STRONG_INLINE Packet4i padd<Packet4i>(const Packet4i& a, const Packet4i& b) { return vaddq_s32(a,b); } + +template<> EIGEN_STRONG_INLINE Packet4f psub<Packet4f>(const Packet4f& a, const Packet4f& b) { return vsubq_f32(a,b); } +template<> EIGEN_STRONG_INLINE Packet4i psub<Packet4i>(const Packet4i& a, const Packet4i& b) { return vsubq_s32(a,b); } + +template<> EIGEN_STRONG_INLINE Packet4f pnegate(const Packet4f& a) { return vnegq_f32(a); } +template<> EIGEN_STRONG_INLINE Packet4i pnegate(const Packet4i& a) { return vnegq_s32(a); } + +template<> EIGEN_STRONG_INLINE Packet4f pmul<Packet4f>(const Packet4f& a, const Packet4f& b) { return vmulq_f32(a,b); } +template<> EIGEN_STRONG_INLINE Packet4i pmul<Packet4i>(const Packet4i& a, const Packet4i& b) { return vmulq_s32(a,b); } + +template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const Packet4f& b) +{ + Packet4f inv, restep, div; + + // NEON does not offer a divide instruction, we have to do a reciprocal approximation + // However NEON in contrast to other SIMD engines (AltiVec/SSE), offers + // a reciprocal estimate AND a reciprocal step -which saves a few instructions + // vrecpeq_f32() returns an estimate to 1/b, which we will finetune with + // Newton-Raphson and vrecpsq_f32() + inv = vrecpeq_f32(b); + + // This returns a differential, by which we will have to multiply inv to get a better + // approximation of 1/b. + restep = vrecpsq_f32(b, inv); + inv = vmulq_f32(restep, inv); + + // Finally, multiply a by 1/b and get the wanted result of the division. + div = vmulq_f32(a, inv); + + return div; +} +template<> EIGEN_STRONG_INLINE Packet4i pdiv<Packet4i>(const Packet4i& /*a*/, const Packet4i& /*b*/) +{ eigen_assert(false && "packet integer division are not supported by NEON"); + return pset1<Packet4i>(0); +} + +// for some weird raisons, it has to be overloaded for packet of integers +template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) { return vmlaq_f32(c,a,b); } +template<> EIGEN_STRONG_INLINE Packet4i pmadd(const Packet4i& a, const Packet4i& b, const Packet4i& c) { return vmlaq_s32(c,a,b); } + +template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const Packet4f& b) { return vminq_f32(a,b); } +template<> EIGEN_STRONG_INLINE Packet4i pmin<Packet4i>(const Packet4i& a, const Packet4i& b) { return vminq_s32(a,b); } + +template<> EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>(const Packet4f& a, const Packet4f& b) { return vmaxq_f32(a,b); } +template<> EIGEN_STRONG_INLINE Packet4i pmax<Packet4i>(const Packet4i& a, const Packet4i& b) { return vmaxq_s32(a,b); } + +// Logical Operations are not supported for float, so we have to reinterpret casts using NEON intrinsics +template<> EIGEN_STRONG_INLINE Packet4f pand<Packet4f>(const Packet4f& a, const Packet4f& b) +{ + return vreinterpretq_f32_u32(vandq_u32(vreinterpretq_u32_f32(a),vreinterpretq_u32_f32(b))); +} +template<> EIGEN_STRONG_INLINE Packet4i pand<Packet4i>(const Packet4i& a, const Packet4i& b) { return vandq_s32(a,b); } + +template<> EIGEN_STRONG_INLINE Packet4f por<Packet4f>(const Packet4f& a, const Packet4f& b) +{ + return vreinterpretq_f32_u32(vorrq_u32(vreinterpretq_u32_f32(a),vreinterpretq_u32_f32(b))); +} +template<> EIGEN_STRONG_INLINE Packet4i por<Packet4i>(const Packet4i& a, const Packet4i& b) { return vorrq_s32(a,b); } + +template<> EIGEN_STRONG_INLINE Packet4f pxor<Packet4f>(const Packet4f& a, const Packet4f& b) +{ + return vreinterpretq_f32_u32(veorq_u32(vreinterpretq_u32_f32(a),vreinterpretq_u32_f32(b))); +} +template<> EIGEN_STRONG_INLINE Packet4i pxor<Packet4i>(const Packet4i& a, const Packet4i& b) { return veorq_s32(a,b); } + +template<> EIGEN_STRONG_INLINE Packet4f pandnot<Packet4f>(const Packet4f& a, const Packet4f& b) +{ + return vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(a),vreinterpretq_u32_f32(b))); +} +template<> EIGEN_STRONG_INLINE Packet4i pandnot<Packet4i>(const Packet4i& a, const Packet4i& b) { return vbicq_s32(a,b); } + +template<> EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(const float* from) { EIGEN_DEBUG_ALIGNED_LOAD return vld1q_f32(from); } +template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int* from) { EIGEN_DEBUG_ALIGNED_LOAD return vld1q_s32(from); } + +template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from) { EIGEN_DEBUG_UNALIGNED_LOAD return vld1q_f32(from); } +template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from) { EIGEN_DEBUG_UNALIGNED_LOAD return vld1q_s32(from); } + +template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float* from) +{ + float32x2_t lo, hi; + lo = vdup_n_f32(*from); + hi = vdup_n_f32(*(from+1)); + return vcombine_f32(lo, hi); +} +template<> EIGEN_STRONG_INLINE Packet4i ploaddup<Packet4i>(const int* from) +{ + int32x2_t lo, hi; + lo = vdup_n_s32(*from); + hi = vdup_n_s32(*(from+1)); + return vcombine_s32(lo, hi); +} + +template<> EIGEN_STRONG_INLINE void pstore<float>(float* to, const Packet4f& from) { EIGEN_DEBUG_ALIGNED_STORE vst1q_f32(to, from); } +template<> EIGEN_STRONG_INLINE void pstore<int>(int* to, const Packet4i& from) { EIGEN_DEBUG_ALIGNED_STORE vst1q_s32(to, from); } + +template<> EIGEN_STRONG_INLINE void pstoreu<float>(float* to, const Packet4f& from) { EIGEN_DEBUG_UNALIGNED_STORE vst1q_f32(to, from); } +template<> EIGEN_STRONG_INLINE void pstoreu<int>(int* to, const Packet4i& from) { EIGEN_DEBUG_UNALIGNED_STORE vst1q_s32(to, from); } + +template<> EIGEN_STRONG_INLINE void prefetch<float>(const float* addr) { __pld(addr); } +template<> EIGEN_STRONG_INLINE void prefetch<int>(const int* addr) { __pld(addr); } + +// FIXME only store the 2 first elements ? +template<> EIGEN_STRONG_INLINE float pfirst<Packet4f>(const Packet4f& a) { float EIGEN_ALIGN16 x[4]; vst1q_f32(x, a); return x[0]; } +template<> EIGEN_STRONG_INLINE int pfirst<Packet4i>(const Packet4i& a) { int EIGEN_ALIGN16 x[4]; vst1q_s32(x, a); return x[0]; } + +template<> EIGEN_STRONG_INLINE Packet4f preverse(const Packet4f& a) { + float32x2_t a_lo, a_hi; + Packet4f a_r64; + + a_r64 = vrev64q_f32(a); + a_lo = vget_low_f32(a_r64); + a_hi = vget_high_f32(a_r64); + return vcombine_f32(a_hi, a_lo); +} +template<> EIGEN_STRONG_INLINE Packet4i preverse(const Packet4i& a) { + int32x2_t a_lo, a_hi; + Packet4i a_r64; + + a_r64 = vrev64q_s32(a); + a_lo = vget_low_s32(a_r64); + a_hi = vget_high_s32(a_r64); + return vcombine_s32(a_hi, a_lo); +} +template<> EIGEN_STRONG_INLINE Packet4f pabs(const Packet4f& a) { return vabsq_f32(a); } +template<> EIGEN_STRONG_INLINE Packet4i pabs(const Packet4i& a) { return vabsq_s32(a); } + +template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a) +{ + float32x2_t a_lo, a_hi, sum; + float s[2]; + + a_lo = vget_low_f32(a); + a_hi = vget_high_f32(a); + sum = vpadd_f32(a_lo, a_hi); + sum = vpadd_f32(sum, sum); + vst1_f32(s, sum); + + return s[0]; +} + +template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs) +{ + float32x4x2_t vtrn1, vtrn2, res1, res2; + Packet4f sum1, sum2, sum; + + // NEON zip performs interleaving of the supplied vectors. + // We perform two interleaves in a row to acquire the transposed vector + vtrn1 = vzipq_f32(vecs[0], vecs[2]); + vtrn2 = vzipq_f32(vecs[1], vecs[3]); + res1 = vzipq_f32(vtrn1.val[0], vtrn2.val[0]); + res2 = vzipq_f32(vtrn1.val[1], vtrn2.val[1]); + + // Do the addition of the resulting vectors + sum1 = vaddq_f32(res1.val[0], res1.val[1]); + sum2 = vaddq_f32(res2.val[0], res2.val[1]); + sum = vaddq_f32(sum1, sum2); + + return sum; +} + +template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a) +{ + int32x2_t a_lo, a_hi, sum; + int32_t s[2]; + + a_lo = vget_low_s32(a); + a_hi = vget_high_s32(a); + sum = vpadd_s32(a_lo, a_hi); + sum = vpadd_s32(sum, sum); + vst1_s32(s, sum); + + return s[0]; +} + +template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs) +{ + int32x4x2_t vtrn1, vtrn2, res1, res2; + Packet4i sum1, sum2, sum; + + // NEON zip performs interleaving of the supplied vectors. + // We perform two interleaves in a row to acquire the transposed vector + vtrn1 = vzipq_s32(vecs[0], vecs[2]); + vtrn2 = vzipq_s32(vecs[1], vecs[3]); + res1 = vzipq_s32(vtrn1.val[0], vtrn2.val[0]); + res2 = vzipq_s32(vtrn1.val[1], vtrn2.val[1]); + + // Do the addition of the resulting vectors + sum1 = vaddq_s32(res1.val[0], res1.val[1]); + sum2 = vaddq_s32(res2.val[0], res2.val[1]); + sum = vaddq_s32(sum1, sum2); + + return sum; +} + +// Other reduction functions: +// mul +template<> EIGEN_STRONG_INLINE float predux_mul<Packet4f>(const Packet4f& a) +{ + float32x2_t a_lo, a_hi, prod; + float s[2]; + + // Get a_lo = |a1|a2| and a_hi = |a3|a4| + a_lo = vget_low_f32(a); + a_hi = vget_high_f32(a); + // Get the product of a_lo * a_hi -> |a1*a3|a2*a4| + prod = vmul_f32(a_lo, a_hi); + // Multiply prod with its swapped value |a2*a4|a1*a3| + prod = vmul_f32(prod, vrev64_f32(prod)); + vst1_f32(s, prod); + + return s[0]; +} +template<> EIGEN_STRONG_INLINE int predux_mul<Packet4i>(const Packet4i& a) +{ + int32x2_t a_lo, a_hi, prod; + int32_t s[2]; + + // Get a_lo = |a1|a2| and a_hi = |a3|a4| + a_lo = vget_low_s32(a); + a_hi = vget_high_s32(a); + // Get the product of a_lo * a_hi -> |a1*a3|a2*a4| + prod = vmul_s32(a_lo, a_hi); + // Multiply prod with its swapped value |a2*a4|a1*a3| + prod = vmul_s32(prod, vrev64_s32(prod)); + vst1_s32(s, prod); + + return s[0]; +} + +// min +template<> EIGEN_STRONG_INLINE float predux_min<Packet4f>(const Packet4f& a) +{ + float32x2_t a_lo, a_hi, min; + float s[2]; + + a_lo = vget_low_f32(a); + a_hi = vget_high_f32(a); + min = vpmin_f32(a_lo, a_hi); + min = vpmin_f32(min, min); + vst1_f32(s, min); + + return s[0]; +} +template<> EIGEN_STRONG_INLINE int predux_min<Packet4i>(const Packet4i& a) +{ + int32x2_t a_lo, a_hi, min; + int32_t s[2]; + + a_lo = vget_low_s32(a); + a_hi = vget_high_s32(a); + min = vpmin_s32(a_lo, a_hi); + min = vpmin_s32(min, min); + vst1_s32(s, min); + + return s[0]; +} + +// max +template<> EIGEN_STRONG_INLINE float predux_max<Packet4f>(const Packet4f& a) +{ + float32x2_t a_lo, a_hi, max; + float s[2]; + + a_lo = vget_low_f32(a); + a_hi = vget_high_f32(a); + max = vpmax_f32(a_lo, a_hi); + max = vpmax_f32(max, max); + vst1_f32(s, max); + + return s[0]; +} +template<> EIGEN_STRONG_INLINE int predux_max<Packet4i>(const Packet4i& a) +{ + int32x2_t a_lo, a_hi, max; + int32_t s[2]; + + a_lo = vget_low_s32(a); + a_hi = vget_high_s32(a); + max = vpmax_s32(a_lo, a_hi); + max = vpmax_s32(max, max); + vst1_s32(s, max); + + return s[0]; +} + +// this PALIGN_NEON business is to work around a bug in LLVM Clang 3.0 causing incorrect compilation errors, +// see bug 347 and this LLVM bug: http://llvm.org/bugs/show_bug.cgi?id=11074 +#define PALIGN_NEON(Offset,Type,Command) \ +template<>\ +struct palign_impl<Offset,Type>\ +{\ + EIGEN_STRONG_INLINE static void run(Type& first, const Type& second)\ + {\ + if (Offset!=0)\ + first = Command(first, second, Offset);\ + }\ +};\ + +PALIGN_NEON(0,Packet4f,vextq_f32) +PALIGN_NEON(1,Packet4f,vextq_f32) +PALIGN_NEON(2,Packet4f,vextq_f32) +PALIGN_NEON(3,Packet4f,vextq_f32) +PALIGN_NEON(0,Packet4i,vextq_s32) +PALIGN_NEON(1,Packet4i,vextq_s32) +PALIGN_NEON(2,Packet4i,vextq_s32) +PALIGN_NEON(3,Packet4i,vextq_s32) + +#undef PALIGN_NEON + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_PACKET_MATH_NEON_H diff --git a/Eigen/src/Core/arch/SSE/CMakeLists.txt b/Eigen/src/Core/arch/SSE/CMakeLists.txt new file mode 100644 index 000000000..46ea7cc62 --- /dev/null +++ b/Eigen/src/Core/arch/SSE/CMakeLists.txt @@ -0,0 +1,6 @@ +FILE(GLOB Eigen_Core_arch_SSE_SRCS "*.h") + +INSTALL(FILES + ${Eigen_Core_arch_SSE_SRCS} + DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/Core/arch/SSE COMPONENT Devel +) diff --git a/Eigen/src/Core/arch/SSE/Complex.h b/Eigen/src/Core/arch/SSE/Complex.h new file mode 100644 index 000000000..12df98775 --- /dev/null +++ b/Eigen/src/Core/arch/SSE/Complex.h @@ -0,0 +1,436 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2010 Gael Guennebaud <gael.guennebaud@inria.fr> +// +// 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 +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_COMPLEX_SSE_H +#define EIGEN_COMPLEX_SSE_H + +namespace Eigen { + +namespace internal { + +//---------- float ---------- +struct Packet2cf +{ + EIGEN_STRONG_INLINE Packet2cf() {} + EIGEN_STRONG_INLINE explicit Packet2cf(const __m128& a) : v(a) {} + __m128 v; +}; + +template<> struct packet_traits<std::complex<float> > : default_packet_traits +{ + typedef Packet2cf type; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size = 2, + + HasAdd = 1, + HasSub = 1, + HasMul = 1, + HasDiv = 1, + HasNegate = 1, + HasAbs = 0, + HasAbs2 = 0, + HasMin = 0, + HasMax = 0, + HasSetLinear = 0 + }; +}; + +template<> struct unpacket_traits<Packet2cf> { typedef std::complex<float> type; enum {size=2}; }; + +template<> EIGEN_STRONG_INLINE Packet2cf padd<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_add_ps(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet2cf psub<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_sub_ps(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet2cf pnegate(const Packet2cf& a) +{ + const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x80000000,0x80000000,0x80000000,0x80000000)); + return Packet2cf(_mm_xor_ps(a.v,mask)); +} +template<> EIGEN_STRONG_INLINE Packet2cf pconj(const Packet2cf& a) +{ + const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x00000000,0x80000000,0x00000000,0x80000000)); + return Packet2cf(_mm_xor_ps(a.v,mask)); +} + +template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, const Packet2cf& b) +{ + // TODO optimize it for SSE3 and 4 + #ifdef EIGEN_VECTORIZE_SSE3 + return Packet2cf(_mm_addsub_ps(_mm_mul_ps(_mm_moveldup_ps(a.v), b.v), + _mm_mul_ps(_mm_movehdup_ps(a.v), + vec4f_swizzle1(b.v, 1, 0, 3, 2)))); +// return Packet2cf(_mm_addsub_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v), +// _mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3), +// vec4f_swizzle1(b.v, 1, 0, 3, 2)))); + #else + const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x80000000,0x00000000,0x80000000,0x00000000)); + return Packet2cf(_mm_add_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v), + _mm_xor_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3), + vec4f_swizzle1(b.v, 1, 0, 3, 2)), mask))); + #endif +} + +template<> EIGEN_STRONG_INLINE Packet2cf pand <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_and_ps(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet2cf por <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_or_ps(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet2cf pxor <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_xor_ps(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet2cf pandnot<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_andnot_ps(a.v,b.v)); } + +template<> EIGEN_STRONG_INLINE Packet2cf pload <Packet2cf>(const std::complex<float>* from) { EIGEN_DEBUG_ALIGNED_LOAD return Packet2cf(pload<Packet4f>(&real_ref(*from))); } +template<> EIGEN_STRONG_INLINE Packet2cf ploadu<Packet2cf>(const std::complex<float>* from) { EIGEN_DEBUG_UNALIGNED_LOAD return Packet2cf(ploadu<Packet4f>(&real_ref(*from))); } + +template<> EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<float>& from) +{ + Packet2cf res; + #if EIGEN_GNUC_AT_MOST(4,2) + // workaround annoying "may be used uninitialized in this function" warning with gcc 4.2 + res.v = _mm_loadl_pi(_mm_set1_ps(0.0f), reinterpret_cast<const __m64*>(&from)); + #else + res.v = _mm_loadl_pi(res.v, (const __m64*)&from); + #endif + return Packet2cf(_mm_movelh_ps(res.v,res.v)); +} + +template<> EIGEN_STRONG_INLINE Packet2cf ploaddup<Packet2cf>(const std::complex<float>* from) { return pset1<Packet2cf>(*from); } + +template<> EIGEN_STRONG_INLINE void pstore <std::complex<float> >(std::complex<float> * to, const Packet2cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore(&real_ref(*to), from.v); } +template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<float> >(std::complex<float> * to, const Packet2cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(&real_ref(*to), from.v); } + +template<> EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::complex<float> * addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } + +template<> EIGEN_STRONG_INLINE std::complex<float> pfirst<Packet2cf>(const Packet2cf& a) +{ + #if EIGEN_GNUC_AT_MOST(4,3) + // Workaround gcc 4.2 ICE - this is not performance wise ideal, but who cares... + // This workaround also fix invalid code generation with gcc 4.3 + EIGEN_ALIGN16 std::complex<float> res[2]; + _mm_store_ps((float*)res, a.v); + return res[0]; + #else + std::complex<float> res; + _mm_storel_pi((__m64*)&res, a.v); + return res; + #endif +} + +template<> EIGEN_STRONG_INLINE Packet2cf preverse(const Packet2cf& a) { return Packet2cf(_mm_castpd_ps(preverse(_mm_castps_pd(a.v)))); } + +template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(const Packet2cf& a) +{ + return pfirst(Packet2cf(_mm_add_ps(a.v, _mm_movehl_ps(a.v,a.v)))); +} + +template<> EIGEN_STRONG_INLINE Packet2cf preduxp<Packet2cf>(const Packet2cf* vecs) +{ + return Packet2cf(_mm_add_ps(_mm_movelh_ps(vecs[0].v,vecs[1].v), _mm_movehl_ps(vecs[1].v,vecs[0].v))); +} + +template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(const Packet2cf& a) +{ + return pfirst(pmul(a, Packet2cf(_mm_movehl_ps(a.v,a.v)))); +} + +template<int Offset> +struct palign_impl<Offset,Packet2cf> +{ + static EIGEN_STRONG_INLINE void run(Packet2cf& first, const Packet2cf& second) + { + if (Offset==1) + { + first.v = _mm_movehl_ps(first.v, first.v); + first.v = _mm_movelh_ps(first.v, second.v); + } + } +}; + +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 + { + #ifdef EIGEN_VECTORIZE_SSE3 + return internal::pmul(a, pconj(b)); + #else + const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x00000000,0x80000000,0x00000000,0x80000000)); + return Packet2cf(_mm_add_ps(_mm_xor_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v), mask), + _mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3), + vec4f_swizzle1(b.v, 1, 0, 3, 2)))); + #endif + } +}; + +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 + { + #ifdef EIGEN_VECTORIZE_SSE3 + return internal::pmul(pconj(a), b); + #else + const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x00000000,0x80000000,0x00000000,0x80000000)); + return Packet2cf(_mm_add_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v), + _mm_xor_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3), + vec4f_swizzle1(b.v, 1, 0, 3, 2)), mask))); + #endif + } +}; + +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 + { + #ifdef EIGEN_VECTORIZE_SSE3 + return pconj(internal::pmul(a, b)); + #else + const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x00000000,0x80000000,0x00000000,0x80000000)); + return Packet2cf(_mm_sub_ps(_mm_xor_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v), mask), + _mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3), + vec4f_swizzle1(b.v, 1, 0, 3, 2)))); + #endif + } +}; + +template<> struct conj_helper<Packet4f, Packet2cf, false,false> +{ + EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet4f& x, const Packet2cf& y, const Packet2cf& c) const + { return padd(c, pmul(x,y)); } + + EIGEN_STRONG_INLINE Packet2cf pmul(const Packet4f& x, const Packet2cf& y) const + { return Packet2cf(Eigen::internal::pmul(x, y.v)); } +}; + +template<> struct conj_helper<Packet2cf, Packet4f, false,false> +{ + EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet4f& y, const Packet2cf& c) const + { return padd(c, pmul(x,y)); } + + EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& x, const Packet4f& y) const + { return Packet2cf(Eigen::internal::pmul(x.v, y)); } +}; + +template<> EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, const Packet2cf& b) +{ + // TODO optimize it for SSE3 and 4 + Packet2cf res = conj_helper<Packet2cf,Packet2cf,false,true>().pmul(a,b); + __m128 s = _mm_mul_ps(b.v,b.v); + return Packet2cf(_mm_div_ps(res.v,_mm_add_ps(s,_mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(s), 0xb1))))); +} + +EIGEN_STRONG_INLINE Packet2cf pcplxflip/*<Packet2cf>*/(const Packet2cf& x) +{ + return Packet2cf(vec4f_swizzle1(x.v, 1, 0, 3, 2)); +} + + +//---------- double ---------- +struct Packet1cd +{ + EIGEN_STRONG_INLINE Packet1cd() {} + EIGEN_STRONG_INLINE explicit Packet1cd(const __m128d& a) : v(a) {} + __m128d v; +}; + +template<> struct packet_traits<std::complex<double> > : default_packet_traits +{ + typedef Packet1cd type; + enum { + Vectorizable = 1, + AlignedOnScalar = 0, + size = 1, + + HasAdd = 1, + HasSub = 1, + HasMul = 1, + HasDiv = 1, + HasNegate = 1, + HasAbs = 0, + HasAbs2 = 0, + HasMin = 0, + HasMax = 0, + HasSetLinear = 0 + }; +}; + +template<> struct unpacket_traits<Packet1cd> { typedef std::complex<double> type; enum {size=1}; }; + +template<> EIGEN_STRONG_INLINE Packet1cd padd<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_add_pd(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet1cd psub<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_sub_pd(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet1cd pnegate(const Packet1cd& a) { return Packet1cd(pnegate(a.v)); } +template<> EIGEN_STRONG_INLINE Packet1cd pconj(const Packet1cd& a) +{ + const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(0x80000000,0x0,0x0,0x0)); + return Packet1cd(_mm_xor_pd(a.v,mask)); +} + +template<> EIGEN_STRONG_INLINE Packet1cd pmul<Packet1cd>(const Packet1cd& a, const Packet1cd& b) +{ + // TODO optimize it for SSE3 and 4 + #ifdef EIGEN_VECTORIZE_SSE3 + return Packet1cd(_mm_addsub_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 0, 0), b.v), + _mm_mul_pd(vec2d_swizzle1(a.v, 1, 1), + vec2d_swizzle1(b.v, 1, 0)))); + #else + const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(0x0,0x0,0x80000000,0x0)); + return Packet1cd(_mm_add_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 0, 0), b.v), + _mm_xor_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 1, 1), + vec2d_swizzle1(b.v, 1, 0)), mask))); + #endif +} + +template<> EIGEN_STRONG_INLINE Packet1cd pand <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_and_pd(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet1cd por <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_or_pd(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet1cd pxor <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_xor_pd(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet1cd pandnot<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_andnot_pd(a.v,b.v)); } + +// FIXME force unaligned load, this is a temporary fix +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 Packet1cd ploadu<Packet1cd>(const std::complex<double>* from) +{ EIGEN_DEBUG_UNALIGNED_LOAD return Packet1cd(ploadu<Packet2d>((const double*)from)); } +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 Packet1cd ploaddup<Packet1cd>(const std::complex<double>* from) { return pset1<Packet1cd>(*from); } + +// FIXME force unaligned store, this is a temporary fix +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<double> >(std::complex<double> * to, const Packet1cd& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((double*)to, from.v); } + +template<> EIGEN_STRONG_INLINE void prefetch<std::complex<double> >(const std::complex<double> * addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } + +template<> EIGEN_STRONG_INLINE std::complex<double> pfirst<Packet1cd>(const Packet1cd& a) +{ + EIGEN_ALIGN16 double res[2]; + _mm_store_pd(res, a.v); + return std::complex<double>(res[0],res[1]); +} + +template<> EIGEN_STRONG_INLINE Packet1cd preverse(const Packet1cd& a) { return a; } + +template<> EIGEN_STRONG_INLINE std::complex<double> predux<Packet1cd>(const Packet1cd& a) +{ + return pfirst(a); +} + +template<> EIGEN_STRONG_INLINE Packet1cd preduxp<Packet1cd>(const Packet1cd* vecs) +{ + return vecs[0]; +} + +template<> EIGEN_STRONG_INLINE std::complex<double> predux_mul<Packet1cd>(const Packet1cd& a) +{ + return pfirst(a); +} + +template<int Offset> +struct palign_impl<Offset,Packet1cd> +{ + 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... + } +}; + +template<> struct conj_helper<Packet1cd, Packet1cd, false,true> +{ + EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const + { + #ifdef EIGEN_VECTORIZE_SSE3 + return internal::pmul(a, pconj(b)); + #else + const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(0x80000000,0x0,0x0,0x0)); + return Packet1cd(_mm_add_pd(_mm_xor_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 0, 0), b.v), mask), + _mm_mul_pd(vec2d_swizzle1(a.v, 1, 1), + vec2d_swizzle1(b.v, 1, 0)))); + #endif + } +}; + +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); } + + EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const + { + #ifdef EIGEN_VECTORIZE_SSE3 + return internal::pmul(pconj(a), b); + #else + const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(0x80000000,0x0,0x0,0x0)); + return Packet1cd(_mm_add_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 0, 0), b.v), + _mm_xor_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 1, 1), + vec2d_swizzle1(b.v, 1, 0)), mask))); + #endif + } +}; + +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); } + + EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const + { + #ifdef EIGEN_VECTORIZE_SSE3 + return pconj(internal::pmul(a, b)); + #else + const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(0x80000000,0x0,0x0,0x0)); + return Packet1cd(_mm_sub_pd(_mm_xor_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 0, 0), b.v), mask), + _mm_mul_pd(vec2d_swizzle1(a.v, 1, 1), + vec2d_swizzle1(b.v, 1, 0)))); + #endif + } +}; + +template<> struct conj_helper<Packet2d, Packet1cd, false,false> +{ + EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet2d& x, const Packet1cd& y, const Packet1cd& c) const + { return padd(c, pmul(x,y)); } + + EIGEN_STRONG_INLINE Packet1cd pmul(const Packet2d& x, const Packet1cd& y) const + { return Packet1cd(Eigen::internal::pmul(x, y.v)); } +}; + +template<> struct conj_helper<Packet1cd, Packet2d, false,false> +{ + EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet2d& y, const Packet1cd& c) const + { return padd(c, pmul(x,y)); } + + EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& x, const Packet2d& y) const + { return Packet1cd(Eigen::internal::pmul(x.v, y)); } +}; + +template<> EIGEN_STRONG_INLINE Packet1cd pdiv<Packet1cd>(const Packet1cd& a, const Packet1cd& b) +{ + // TODO optimize it for SSE3 and 4 + Packet1cd res = conj_helper<Packet1cd,Packet1cd,false,true>().pmul(a,b); + __m128d s = _mm_mul_pd(b.v,b.v); + return Packet1cd(_mm_div_pd(res.v, _mm_add_pd(s,_mm_shuffle_pd(s, s, 0x1)))); +} + +EIGEN_STRONG_INLINE Packet1cd pcplxflip/*<Packet1cd>*/(const Packet1cd& x) +{ + return Packet1cd(preverse(x.v)); +} + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_COMPLEX_SSE_H diff --git a/Eigen/src/Core/arch/SSE/MathFunctions.h b/Eigen/src/Core/arch/SSE/MathFunctions.h new file mode 100644 index 000000000..3f41a4e26 --- /dev/null +++ b/Eigen/src/Core/arch/SSE/MathFunctions.h @@ -0,0 +1,384 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2007 Julien Pommier +// Copyright (C) 2009 Gael Guennebaud <gael.guennebaud@inria.fr> +// +// 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 +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +/* The sin, cos, exp, and log functions of this file come from + * Julien Pommier's sse math library: http://gruntthepeon.free.fr/ssemath/ + */ + +#ifndef EIGEN_MATH_FUNCTIONS_SSE_H +#define EIGEN_MATH_FUNCTIONS_SSE_H + +namespace Eigen { + +namespace internal { + +template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED +Packet4f plog<Packet4f>(const Packet4f& _x) +{ + Packet4f x = _x; + _EIGEN_DECLARE_CONST_Packet4f(1 , 1.0f); + _EIGEN_DECLARE_CONST_Packet4f(half, 0.5f); + _EIGEN_DECLARE_CONST_Packet4i(0x7f, 0x7f); + + _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(inv_mant_mask, ~0x7f800000); + + /* the smallest non denormalized float number */ + _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(min_norm_pos, 0x00800000); + + /* natural logarithm computed for 4 simultaneous float + return NaN for x <= 0 + */ + _EIGEN_DECLARE_CONST_Packet4f(cephes_SQRTHF, 0.707106781186547524f); + _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p0, 7.0376836292E-2f); + _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p1, - 1.1514610310E-1f); + _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p2, 1.1676998740E-1f); + _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p3, - 1.2420140846E-1f); + _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p4, + 1.4249322787E-1f); + _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p5, - 1.6668057665E-1f); + _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p6, + 2.0000714765E-1f); + _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p7, - 2.4999993993E-1f); + _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p8, + 3.3333331174E-1f); + _EIGEN_DECLARE_CONST_Packet4f(cephes_log_q1, -2.12194440e-4f); + _EIGEN_DECLARE_CONST_Packet4f(cephes_log_q2, 0.693359375f); + + + Packet4i emm0; + + Packet4f invalid_mask = _mm_cmple_ps(x, _mm_setzero_ps()); + + x = pmax(x, p4f_min_norm_pos); /* cut off denormalized stuff */ + emm0 = _mm_srli_epi32(_mm_castps_si128(x), 23); + + /* keep only the fractional part */ + x = _mm_and_ps(x, p4f_inv_mant_mask); + x = _mm_or_ps(x, p4f_half); + + emm0 = _mm_sub_epi32(emm0, p4i_0x7f); + Packet4f e = padd(_mm_cvtepi32_ps(emm0), p4f_1); + + /* part2: + if( x < SQRTHF ) { + e -= 1; + x = x + x - 1.0; + } else { x = x - 1.0; } + */ + Packet4f mask = _mm_cmplt_ps(x, p4f_cephes_SQRTHF); + Packet4f tmp = _mm_and_ps(x, mask); + x = psub(x, p4f_1); + e = psub(e, _mm_and_ps(p4f_1, mask)); + x = padd(x, tmp); + + Packet4f x2 = pmul(x,x); + Packet4f x3 = pmul(x2,x); + + Packet4f y, y1, y2; + y = pmadd(p4f_cephes_log_p0, x, p4f_cephes_log_p1); + y1 = pmadd(p4f_cephes_log_p3, x, p4f_cephes_log_p4); + y2 = pmadd(p4f_cephes_log_p6, x, p4f_cephes_log_p7); + y = pmadd(y , x, p4f_cephes_log_p2); + y1 = pmadd(y1, x, p4f_cephes_log_p5); + y2 = pmadd(y2, x, p4f_cephes_log_p8); + y = pmadd(y, x3, y1); + y = pmadd(y, x3, y2); + y = pmul(y, x3); + + y1 = pmul(e, p4f_cephes_log_q1); + tmp = pmul(x2, p4f_half); + y = padd(y, y1); + x = psub(x, tmp); + y2 = pmul(e, p4f_cephes_log_q2); + x = padd(x, y); + x = padd(x, y2); + return _mm_or_ps(x, invalid_mask); // negative arg will be NAN +} + +template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED +Packet4f pexp<Packet4f>(const Packet4f& _x) +{ + Packet4f x = _x; + _EIGEN_DECLARE_CONST_Packet4f(1 , 1.0f); + _EIGEN_DECLARE_CONST_Packet4f(half, 0.5f); + _EIGEN_DECLARE_CONST_Packet4i(0x7f, 0x7f); + + + _EIGEN_DECLARE_CONST_Packet4f(exp_hi, 88.3762626647950f); + _EIGEN_DECLARE_CONST_Packet4f(exp_lo, -88.3762626647949f); + + _EIGEN_DECLARE_CONST_Packet4f(cephes_LOG2EF, 1.44269504088896341f); + _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_C1, 0.693359375f); + _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_C2, -2.12194440e-4f); + + _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p0, 1.9875691500E-4f); + _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p1, 1.3981999507E-3f); + _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p2, 8.3334519073E-3f); + _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p3, 4.1665795894E-2f); + _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p4, 1.6666665459E-1f); + _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p5, 5.0000001201E-1f); + + Packet4f tmp = _mm_setzero_ps(), 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); + + /* how to perform a floorf with SSE: just below */ + emm0 = _mm_cvttps_epi32(fx); + tmp = _mm_cvtepi32_ps(emm0); + /* if greater, substract 1 */ + Packet4f mask = _mm_cmpgt_ps(tmp, fx); + mask = _mm_and_ps(mask, p4f_1); + fx = psub(tmp, mask); + + 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 = _mm_cvttps_epi32(fx); + emm0 = _mm_add_epi32(emm0, p4i_0x7f); + emm0 = _mm_slli_epi32(emm0, 23); + return pmul(y, _mm_castsi128_ps(emm0)); +} + +/* evaluation of 4 sines at onces, using SSE2 intrinsics. + + The code is the exact rewriting of the cephes sinf function. + Precision is excellent as long as x < 8192 (I did not bother to + take into account the special handling they have for greater values + -- it does not return garbage for arguments over 8192, though, but + the extra precision is missing). + + Note that it is such that sinf((float)M_PI) = 8.74e-8, which is the + surprising but correct result. +*/ + +template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED +Packet4f psin<Packet4f>(const Packet4f& _x) +{ + Packet4f x = _x; + _EIGEN_DECLARE_CONST_Packet4f(1 , 1.0f); + _EIGEN_DECLARE_CONST_Packet4f(half, 0.5f); + + _EIGEN_DECLARE_CONST_Packet4i(1, 1); + _EIGEN_DECLARE_CONST_Packet4i(not1, ~1); + _EIGEN_DECLARE_CONST_Packet4i(2, 2); + _EIGEN_DECLARE_CONST_Packet4i(4, 4); + + _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(sign_mask, 0x80000000); + + _EIGEN_DECLARE_CONST_Packet4f(minus_cephes_DP1,-0.78515625f); + _EIGEN_DECLARE_CONST_Packet4f(minus_cephes_DP2, -2.4187564849853515625e-4f); + _EIGEN_DECLARE_CONST_Packet4f(minus_cephes_DP3, -3.77489497744594108e-8f); + _EIGEN_DECLARE_CONST_Packet4f(sincof_p0, -1.9515295891E-4f); + _EIGEN_DECLARE_CONST_Packet4f(sincof_p1, 8.3321608736E-3f); + _EIGEN_DECLARE_CONST_Packet4f(sincof_p2, -1.6666654611E-1f); + _EIGEN_DECLARE_CONST_Packet4f(coscof_p0, 2.443315711809948E-005f); + _EIGEN_DECLARE_CONST_Packet4f(coscof_p1, -1.388731625493765E-003f); + _EIGEN_DECLARE_CONST_Packet4f(coscof_p2, 4.166664568298827E-002f); + _EIGEN_DECLARE_CONST_Packet4f(cephes_FOPI, 1.27323954473516f); // 4 / M_PI + + Packet4f xmm1, xmm2 = _mm_setzero_ps(), xmm3, sign_bit, y; + + Packet4i emm0, emm2; + sign_bit = x; + /* take the absolute value */ + x = pabs(x); + + /* take the modulo */ + + /* extract the sign bit (upper one) */ + sign_bit = _mm_and_ps(sign_bit, p4f_sign_mask); + + /* scale by 4/Pi */ + y = pmul(x, p4f_cephes_FOPI); + + /* store the integer part of y in mm0 */ + emm2 = _mm_cvttps_epi32(y); + /* j=(j+1) & (~1) (see the cephes sources) */ + emm2 = _mm_add_epi32(emm2, p4i_1); + emm2 = _mm_and_si128(emm2, p4i_not1); + y = _mm_cvtepi32_ps(emm2); + /* get the swap sign flag */ + emm0 = _mm_and_si128(emm2, p4i_4); + emm0 = _mm_slli_epi32(emm0, 29); + /* get the polynom selection mask + there is one polynom for 0 <= x <= Pi/4 + and another one for Pi/4<x<=Pi/2 + + Both branches will be computed. + */ + emm2 = _mm_and_si128(emm2, p4i_2); + emm2 = _mm_cmpeq_epi32(emm2, _mm_setzero_si128()); + + Packet4f swap_sign_bit = _mm_castsi128_ps(emm0); + Packet4f poly_mask = _mm_castsi128_ps(emm2); + sign_bit = _mm_xor_ps(sign_bit, swap_sign_bit); + + /* The magic pass: "Extended precision modular arithmetic" + x = ((x - y * DP1) - y * DP2) - y * DP3; */ + xmm1 = pmul(y, p4f_minus_cephes_DP1); + xmm2 = pmul(y, p4f_minus_cephes_DP2); + xmm3 = pmul(y, p4f_minus_cephes_DP3); + x = padd(x, xmm1); + x = padd(x, xmm2); + x = padd(x, xmm3); + + /* Evaluate the first polynom (0 <= x <= Pi/4) */ + y = p4f_coscof_p0; + Packet4f z = _mm_mul_ps(x,x); + + y = pmadd(y, z, p4f_coscof_p1); + y = pmadd(y, z, p4f_coscof_p2); + y = pmul(y, z); + y = pmul(y, z); + Packet4f tmp = pmul(z, p4f_half); + y = psub(y, tmp); + y = padd(y, p4f_1); + + /* Evaluate the second polynom (Pi/4 <= x <= 0) */ + + Packet4f y2 = p4f_sincof_p0; + y2 = pmadd(y2, z, p4f_sincof_p1); + y2 = pmadd(y2, z, p4f_sincof_p2); + y2 = pmul(y2, z); + y2 = pmul(y2, x); + y2 = padd(y2, x); + + /* select the correct result from the two polynoms */ + y2 = _mm_and_ps(poly_mask, y2); + y = _mm_andnot_ps(poly_mask, y); + y = _mm_or_ps(y,y2); + /* update the sign */ + return _mm_xor_ps(y, sign_bit); +} + +/* almost the same as psin */ +template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED +Packet4f pcos<Packet4f>(const Packet4f& _x) +{ + Packet4f x = _x; + _EIGEN_DECLARE_CONST_Packet4f(1 , 1.0f); + _EIGEN_DECLARE_CONST_Packet4f(half, 0.5f); + + _EIGEN_DECLARE_CONST_Packet4i(1, 1); + _EIGEN_DECLARE_CONST_Packet4i(not1, ~1); + _EIGEN_DECLARE_CONST_Packet4i(2, 2); + _EIGEN_DECLARE_CONST_Packet4i(4, 4); + + _EIGEN_DECLARE_CONST_Packet4f(minus_cephes_DP1,-0.78515625f); + _EIGEN_DECLARE_CONST_Packet4f(minus_cephes_DP2, -2.4187564849853515625e-4f); + _EIGEN_DECLARE_CONST_Packet4f(minus_cephes_DP3, -3.77489497744594108e-8f); + _EIGEN_DECLARE_CONST_Packet4f(sincof_p0, -1.9515295891E-4f); + _EIGEN_DECLARE_CONST_Packet4f(sincof_p1, 8.3321608736E-3f); + _EIGEN_DECLARE_CONST_Packet4f(sincof_p2, -1.6666654611E-1f); + _EIGEN_DECLARE_CONST_Packet4f(coscof_p0, 2.443315711809948E-005f); + _EIGEN_DECLARE_CONST_Packet4f(coscof_p1, -1.388731625493765E-003f); + _EIGEN_DECLARE_CONST_Packet4f(coscof_p2, 4.166664568298827E-002f); + _EIGEN_DECLARE_CONST_Packet4f(cephes_FOPI, 1.27323954473516f); // 4 / M_PI + + Packet4f xmm1, xmm2 = _mm_setzero_ps(), xmm3, y; + Packet4i emm0, emm2; + + x = pabs(x); + + /* scale by 4/Pi */ + y = pmul(x, p4f_cephes_FOPI); + + /* get the integer part of y */ + emm2 = _mm_cvttps_epi32(y); + /* j=(j+1) & (~1) (see the cephes sources) */ + emm2 = _mm_add_epi32(emm2, p4i_1); + emm2 = _mm_and_si128(emm2, p4i_not1); + y = _mm_cvtepi32_ps(emm2); + + emm2 = _mm_sub_epi32(emm2, p4i_2); + + /* get the swap sign flag */ + emm0 = _mm_andnot_si128(emm2, p4i_4); + emm0 = _mm_slli_epi32(emm0, 29); + /* get the polynom selection mask */ + emm2 = _mm_and_si128(emm2, p4i_2); + emm2 = _mm_cmpeq_epi32(emm2, _mm_setzero_si128()); + + Packet4f sign_bit = _mm_castsi128_ps(emm0); + Packet4f poly_mask = _mm_castsi128_ps(emm2); + + /* The magic pass: "Extended precision modular arithmetic" + x = ((x - y * DP1) - y * DP2) - y * DP3; */ + xmm1 = pmul(y, p4f_minus_cephes_DP1); + xmm2 = pmul(y, p4f_minus_cephes_DP2); + xmm3 = pmul(y, p4f_minus_cephes_DP3); + x = padd(x, xmm1); + x = padd(x, xmm2); + x = padd(x, xmm3); + + /* Evaluate the first polynom (0 <= x <= Pi/4) */ + y = p4f_coscof_p0; + Packet4f z = pmul(x,x); + + y = pmadd(y,z,p4f_coscof_p1); + y = pmadd(y,z,p4f_coscof_p2); + y = pmul(y, z); + y = pmul(y, z); + Packet4f tmp = _mm_mul_ps(z, p4f_half); + y = psub(y, tmp); + y = padd(y, p4f_1); + + /* Evaluate the second polynom (Pi/4 <= x <= 0) */ + Packet4f y2 = p4f_sincof_p0; + y2 = pmadd(y2, z, p4f_sincof_p1); + y2 = pmadd(y2, z, p4f_sincof_p2); + y2 = pmul(y2, z); + y2 = pmadd(y2, x, x); + + /* select the correct result from the two polynoms */ + y2 = _mm_and_ps(poly_mask, y2); + y = _mm_andnot_ps(poly_mask, y); + y = _mm_or_ps(y,y2); + + /* update the sign */ + return _mm_xor_ps(y, sign_bit); +} + +// This is based on Quake3's fast inverse square root. +// For detail see here: http://www.beyond3d.com/content/articles/8/ +template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED +Packet4f psqrt<Packet4f>(const Packet4f& _x) +{ + Packet4f half = pmul(_x, pset1<Packet4f>(.5f)); + + /* select only the inverse sqrt of non-zero inputs */ + Packet4f non_zero_mask = _mm_cmpgt_ps(_x, pset1<Packet4f>(std::numeric_limits<float>::epsilon())); + Packet4f x = _mm_and_ps(non_zero_mask, _mm_rsqrt_ps(_x)); + + x = pmul(x, psub(pset1<Packet4f>(1.5f), pmul(half, pmul(x,x)))); + return pmul(_x,x); +} + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_MATH_FUNCTIONS_SSE_H diff --git a/Eigen/src/Core/arch/SSE/PacketMath.h b/Eigen/src/Core/arch/SSE/PacketMath.h new file mode 100644 index 000000000..10d918219 --- /dev/null +++ b/Eigen/src/Core/arch/SSE/PacketMath.h @@ -0,0 +1,632 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2009 Gael Guennebaud <gael.guennebaud@inria.fr> +// +// 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 +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_PACKET_MATH_SSE_H +#define EIGEN_PACKET_MATH_SSE_H + +namespace Eigen { + +namespace internal { + +#ifndef EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD +#define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 8 +#endif + +#ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS +#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS (2*sizeof(void*)) +#endif + +typedef __m128 Packet4f; +typedef __m128i Packet4i; +typedef __m128d Packet2d; + +template<> struct is_arithmetic<__m128> { enum { value = true }; }; +template<> struct is_arithmetic<__m128i> { enum { value = true }; }; +template<> struct is_arithmetic<__m128d> { enum { value = true }; }; + +#define vec4f_swizzle1(v,p,q,r,s) \ + (_mm_castsi128_ps(_mm_shuffle_epi32( _mm_castps_si128(v), ((s)<<6|(r)<<4|(q)<<2|(p))))) + +#define vec4i_swizzle1(v,p,q,r,s) \ + (_mm_shuffle_epi32( v, ((s)<<6|(r)<<4|(q)<<2|(p)))) + +#define vec2d_swizzle1(v,p,q) \ + (_mm_castsi128_pd(_mm_shuffle_epi32( _mm_castpd_si128(v), ((q*2+1)<<6|(q*2)<<4|(p*2+1)<<2|(p*2))))) + +#define vec4f_swizzle2(a,b,p,q,r,s) \ + (_mm_shuffle_ps( (a), (b), ((s)<<6|(r)<<4|(q)<<2|(p)))) + +#define vec4i_swizzle2(a,b,p,q,r,s) \ + (_mm_castps_si128( (_mm_shuffle_ps( _mm_castsi128_ps(a), _mm_castsi128_ps(b), ((s)<<6|(r)<<4|(q)<<2|(p)))))) + +#define _EIGEN_DECLARE_CONST_Packet4f(NAME,X) \ + const Packet4f p4f_##NAME = pset1<Packet4f>(X) + +#define _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(NAME,X) \ + const Packet4f p4f_##NAME = _mm_castsi128_ps(pset1<Packet4i>(X)) + +#define _EIGEN_DECLARE_CONST_Packet4i(NAME,X) \ + const Packet4i p4i_##NAME = pset1<Packet4i>(X) + + +template<> struct packet_traits<float> : default_packet_traits +{ + typedef Packet4f type; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size=4, + + HasDiv = 1, + HasSin = EIGEN_FAST_MATH, + HasCos = EIGEN_FAST_MATH, + HasLog = 1, + HasExp = 1, + HasSqrt = 1 + }; +}; +template<> struct packet_traits<double> : default_packet_traits +{ + typedef Packet2d type; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size=2, + + HasDiv = 1 + }; +}; +template<> struct packet_traits<int> : default_packet_traits +{ + typedef Packet4i type; + enum { + // FIXME check the Has* + Vectorizable = 1, + AlignedOnScalar = 1, + size=4 + }; +}; + +template<> struct unpacket_traits<Packet4f> { typedef float type; enum {size=4}; }; +template<> struct unpacket_traits<Packet2d> { typedef double type; enum {size=2}; }; +template<> struct unpacket_traits<Packet4i> { typedef int type; enum {size=4}; }; + +#if defined(_MSC_VER) && (_MSC_VER==1500) +// Workaround MSVC 9 internal compiler error. +// TODO: It has been detected with win64 builds (amd64), so let's check whether it also happens in 32bits+SSE mode +// TODO: let's check whether there does not exist a better fix, like adding a pset0() function. (it crashed on pset1(0)). +template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float& from) { return _mm_set_ps(from,from,from,from); } +template<> EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(const double& from) { return _mm_set_pd(from,from); } +template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int& from) { return _mm_set_epi32(from,from,from,from); } +#else +template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float& from) { return _mm_set1_ps(from); } +template<> EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(const double& from) { return _mm_set1_pd(from); } +template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int& from) { return _mm_set1_epi32(from); } +#endif + +template<> EIGEN_STRONG_INLINE Packet4f plset<float>(const float& a) { return _mm_add_ps(pset1<Packet4f>(a), _mm_set_ps(3,2,1,0)); } +template<> EIGEN_STRONG_INLINE Packet2d plset<double>(const double& a) { return _mm_add_pd(pset1<Packet2d>(a),_mm_set_pd(1,0)); } +template<> EIGEN_STRONG_INLINE Packet4i plset<int>(const int& a) { return _mm_add_epi32(pset1<Packet4i>(a),_mm_set_epi32(3,2,1,0)); } + +template<> EIGEN_STRONG_INLINE Packet4f padd<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_add_ps(a,b); } +template<> EIGEN_STRONG_INLINE Packet2d padd<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_add_pd(a,b); } +template<> EIGEN_STRONG_INLINE Packet4i padd<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_add_epi32(a,b); } + +template<> EIGEN_STRONG_INLINE Packet4f psub<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_sub_ps(a,b); } +template<> EIGEN_STRONG_INLINE Packet2d psub<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_sub_pd(a,b); } +template<> EIGEN_STRONG_INLINE Packet4i psub<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_sub_epi32(a,b); } + +template<> EIGEN_STRONG_INLINE Packet4f pnegate(const Packet4f& a) +{ + const Packet4f mask = _mm_castsi128_ps(_mm_setr_epi32(0x80000000,0x80000000,0x80000000,0x80000000)); + return _mm_xor_ps(a,mask); +} +template<> EIGEN_STRONG_INLINE Packet2d pnegate(const Packet2d& a) +{ + const Packet2d mask = _mm_castsi128_pd(_mm_setr_epi32(0x0,0x80000000,0x0,0x80000000)); + return _mm_xor_pd(a,mask); +} +template<> EIGEN_STRONG_INLINE Packet4i pnegate(const Packet4i& a) +{ + return psub(_mm_setr_epi32(0,0,0,0), a); +} + +template<> EIGEN_STRONG_INLINE Packet4f pmul<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_mul_ps(a,b); } +template<> EIGEN_STRONG_INLINE Packet2d pmul<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_mul_pd(a,b); } +template<> EIGEN_STRONG_INLINE Packet4i pmul<Packet4i>(const Packet4i& a, const Packet4i& b) +{ +#ifdef EIGEN_VECTORIZE_SSE4_1 + return _mm_mullo_epi32(a,b); +#else + // this version is slightly faster than 4 scalar products + return vec4i_swizzle1( + vec4i_swizzle2( + _mm_mul_epu32(a,b), + _mm_mul_epu32(vec4i_swizzle1(a,1,0,3,2), + vec4i_swizzle1(b,1,0,3,2)), + 0,2,0,2), + 0,2,1,3); +#endif +} + +template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_div_ps(a,b); } +template<> EIGEN_STRONG_INLINE Packet2d pdiv<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_div_pd(a,b); } +template<> EIGEN_STRONG_INLINE Packet4i pdiv<Packet4i>(const Packet4i& /*a*/, const Packet4i& /*b*/) +{ eigen_assert(false && "packet integer division are not supported by SSE"); + return pset1<Packet4i>(0); +} + +// for some weird raisons, it has to be overloaded for packet of integers +template<> EIGEN_STRONG_INLINE Packet4i pmadd(const Packet4i& a, const Packet4i& b, const Packet4i& c) { return padd(pmul(a,b), c); } + +template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_min_ps(a,b); } +template<> EIGEN_STRONG_INLINE Packet2d pmin<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_min_pd(a,b); } +template<> EIGEN_STRONG_INLINE Packet4i pmin<Packet4i>(const Packet4i& a, const Packet4i& b) +{ + // after some bench, this version *is* faster than a scalar implementation + Packet4i mask = _mm_cmplt_epi32(a,b); + return _mm_or_si128(_mm_and_si128(mask,a),_mm_andnot_si128(mask,b)); +} + +template<> EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_max_ps(a,b); } +template<> EIGEN_STRONG_INLINE Packet2d pmax<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_max_pd(a,b); } +template<> EIGEN_STRONG_INLINE Packet4i pmax<Packet4i>(const Packet4i& a, const Packet4i& b) +{ + // after some bench, this version *is* faster than a scalar implementation + Packet4i mask = _mm_cmpgt_epi32(a,b); + return _mm_or_si128(_mm_and_si128(mask,a),_mm_andnot_si128(mask,b)); +} + +template<> EIGEN_STRONG_INLINE Packet4f pand<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_and_ps(a,b); } +template<> EIGEN_STRONG_INLINE Packet2d pand<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_and_pd(a,b); } +template<> EIGEN_STRONG_INLINE Packet4i pand<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_and_si128(a,b); } + +template<> EIGEN_STRONG_INLINE Packet4f por<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_or_ps(a,b); } +template<> EIGEN_STRONG_INLINE Packet2d por<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_or_pd(a,b); } +template<> EIGEN_STRONG_INLINE Packet4i por<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_or_si128(a,b); } + +template<> EIGEN_STRONG_INLINE Packet4f pxor<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_xor_ps(a,b); } +template<> EIGEN_STRONG_INLINE Packet2d pxor<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_xor_pd(a,b); } +template<> EIGEN_STRONG_INLINE Packet4i pxor<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_xor_si128(a,b); } + +template<> EIGEN_STRONG_INLINE Packet4f pandnot<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_andnot_ps(a,b); } +template<> EIGEN_STRONG_INLINE Packet2d pandnot<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_andnot_pd(a,b); } +template<> EIGEN_STRONG_INLINE Packet4i pandnot<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_andnot_si128(a,b); } + +template<> EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(const float* from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm_load_ps(from); } +template<> EIGEN_STRONG_INLINE Packet2d pload<Packet2d>(const double* from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm_load_pd(from); } +template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int* from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm_load_si128(reinterpret_cast<const Packet4i*>(from)); } + +#if defined(_MSC_VER) + template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from) { + EIGEN_DEBUG_UNALIGNED_LOAD + #if (_MSC_VER==1600) + // NOTE Some version of MSVC10 generates bad code when using _mm_loadu_ps + // (i.e., it does not generate an unaligned load!! + // TODO On most architectures this version should also be faster than a single _mm_loadu_ps + // so we could also enable it for MSVC08 but first we have to make this later does not generate crap when doing so... + __m128 res = _mm_loadl_pi(_mm_set1_ps(0.0f), (const __m64*)(from)); + res = _mm_loadh_pi(res, (const __m64*)(from+2)); + return res; + #else + return _mm_loadu_ps(from); + #endif + } + template<> EIGEN_STRONG_INLINE Packet2d ploadu<Packet2d>(const double* from) { EIGEN_DEBUG_UNALIGNED_LOAD return _mm_loadu_pd(from); } + template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from) { EIGEN_DEBUG_UNALIGNED_LOAD return _mm_loadu_si128(reinterpret_cast<const Packet4i*>(from)); } +#else +// Fast unaligned loads. Note that here we cannot directly use intrinsics: this would +// require pointer casting to incompatible pointer types and leads to invalid code +// because of the strict aliasing rule. The "dummy" stuff are required to enforce +// a correct instruction dependency. +// TODO: do the same for MSVC (ICC is compatible) +// NOTE: with the code below, MSVC's compiler crashes! + +#if defined(__GNUC__) && defined(__i386__) + // bug 195: gcc/i386 emits weird x87 fldl/fstpl instructions for _mm_load_sd + #define EIGEN_AVOID_CUSTOM_UNALIGNED_LOADS 1 +#elif defined(__clang__) + // bug 201: Segfaults in __mm_loadh_pd with clang 2.8 + #define EIGEN_AVOID_CUSTOM_UNALIGNED_LOADS 1 +#else + #define EIGEN_AVOID_CUSTOM_UNALIGNED_LOADS 0 +#endif + +template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from) +{ + EIGEN_DEBUG_UNALIGNED_LOAD +#if EIGEN_AVOID_CUSTOM_UNALIGNED_LOADS + return _mm_loadu_ps(from); +#else + __m128d res; + res = _mm_load_sd((const double*)(from)) ; + res = _mm_loadh_pd(res, (const double*)(from+2)) ; + return _mm_castpd_ps(res); +#endif +} +template<> EIGEN_STRONG_INLINE Packet2d ploadu<Packet2d>(const double* from) +{ + EIGEN_DEBUG_UNALIGNED_LOAD +#if EIGEN_AVOID_CUSTOM_UNALIGNED_LOADS + return _mm_loadu_pd(from); +#else + __m128d res; + res = _mm_load_sd(from) ; + res = _mm_loadh_pd(res,from+1); + return res; +#endif +} +template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from) +{ + EIGEN_DEBUG_UNALIGNED_LOAD +#if EIGEN_AVOID_CUSTOM_UNALIGNED_LOADS + return _mm_loadu_si128(reinterpret_cast<const Packet4i*>(from)); +#else + __m128d res; + res = _mm_load_sd((const double*)(from)) ; + res = _mm_loadh_pd(res, (const double*)(from+2)) ; + return _mm_castpd_si128(res); +#endif +} +#endif + +template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float* from) +{ + return vec4f_swizzle1(_mm_castpd_ps(_mm_load_sd(reinterpret_cast<const double*>(from))), 0, 0, 1, 1); +} +template<> EIGEN_STRONG_INLINE Packet2d ploaddup<Packet2d>(const double* from) +{ return pset1<Packet2d>(from[0]); } +template<> EIGEN_STRONG_INLINE Packet4i ploaddup<Packet4i>(const int* from) +{ + Packet4i tmp; + tmp = _mm_loadl_epi64(reinterpret_cast<const Packet4i*>(from)); + return vec4i_swizzle1(tmp, 0, 0, 1, 1); +} + +template<> EIGEN_STRONG_INLINE void pstore<float>(float* to, const Packet4f& from) { EIGEN_DEBUG_ALIGNED_STORE _mm_store_ps(to, from); } +template<> EIGEN_STRONG_INLINE void pstore<double>(double* to, const Packet2d& from) { EIGEN_DEBUG_ALIGNED_STORE _mm_store_pd(to, from); } +template<> EIGEN_STRONG_INLINE void pstore<int>(int* to, const Packet4i& from) { EIGEN_DEBUG_ALIGNED_STORE _mm_store_si128(reinterpret_cast<Packet4i*>(to), from); } + +template<> EIGEN_STRONG_INLINE void pstoreu<double>(double* to, const Packet2d& from) { + EIGEN_DEBUG_UNALIGNED_STORE + _mm_storel_pd((to), from); + _mm_storeh_pd((to+1), from); +} +template<> EIGEN_STRONG_INLINE void pstoreu<float>(float* to, const Packet4f& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(reinterpret_cast<double*>(to), _mm_castps_pd(from)); } +template<> EIGEN_STRONG_INLINE void pstoreu<int>(int* to, const Packet4i& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(reinterpret_cast<double*>(to), _mm_castsi128_pd(from)); } + +// some compilers might be tempted to perform multiple moves instead of using a vector path. +template<> EIGEN_STRONG_INLINE void pstore1<Packet4f>(float* to, const float& a) +{ + Packet4f pa = _mm_set_ss(a); + pstore(to, vec4f_swizzle1(pa,0,0,0,0)); +} +// some compilers might be tempted to perform multiple moves instead of using a vector path. +template<> EIGEN_STRONG_INLINE void pstore1<Packet2d>(double* to, const double& a) +{ + Packet2d pa = _mm_set_sd(a); + pstore(to, vec2d_swizzle1(pa,0,0)); +} + +template<> EIGEN_STRONG_INLINE void prefetch<float>(const float* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } +template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } +template<> EIGEN_STRONG_INLINE void prefetch<int>(const int* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } + +#if defined(_MSC_VER) && defined(_WIN64) && !defined(__INTEL_COMPILER) +// The temporary variable fixes an internal compilation error in vs <= 2008 and a wrong-result bug in vs 2010 +// Direct of the struct members fixed bug #62. +template<> EIGEN_STRONG_INLINE float pfirst<Packet4f>(const Packet4f& a) { return a.m128_f32[0]; } +template<> EIGEN_STRONG_INLINE double pfirst<Packet2d>(const Packet2d& a) { return a.m128d_f64[0]; } +template<> EIGEN_STRONG_INLINE int pfirst<Packet4i>(const Packet4i& a) { int x = _mm_cvtsi128_si32(a); return x; } +#elif defined(_MSC_VER) && !defined(__INTEL_COMPILER) +// The temporary variable fixes an internal compilation error in vs <= 2008 and a wrong-result bug in vs 2010 +template<> EIGEN_STRONG_INLINE float pfirst<Packet4f>(const Packet4f& a) { float x = _mm_cvtss_f32(a); return x; } +template<> EIGEN_STRONG_INLINE double pfirst<Packet2d>(const Packet2d& a) { double x = _mm_cvtsd_f64(a); return x; } +template<> EIGEN_STRONG_INLINE int pfirst<Packet4i>(const Packet4i& a) { int x = _mm_cvtsi128_si32(a); return x; } +#else +template<> EIGEN_STRONG_INLINE float pfirst<Packet4f>(const Packet4f& a) { return _mm_cvtss_f32(a); } +template<> EIGEN_STRONG_INLINE double pfirst<Packet2d>(const Packet2d& a) { return _mm_cvtsd_f64(a); } +template<> EIGEN_STRONG_INLINE int pfirst<Packet4i>(const Packet4i& a) { return _mm_cvtsi128_si32(a); } +#endif + +template<> EIGEN_STRONG_INLINE Packet4f preverse(const Packet4f& a) +{ return _mm_shuffle_ps(a,a,0x1B); } +template<> EIGEN_STRONG_INLINE Packet2d preverse(const Packet2d& a) +{ return _mm_shuffle_pd(a,a,0x1); } +template<> EIGEN_STRONG_INLINE Packet4i preverse(const Packet4i& a) +{ return _mm_shuffle_epi32(a,0x1B); } + + +template<> EIGEN_STRONG_INLINE Packet4f pabs(const Packet4f& a) +{ + const Packet4f mask = _mm_castsi128_ps(_mm_setr_epi32(0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF)); + return _mm_and_ps(a,mask); +} +template<> EIGEN_STRONG_INLINE Packet2d pabs(const Packet2d& a) +{ + const Packet2d mask = _mm_castsi128_pd(_mm_setr_epi32(0xFFFFFFFF,0x7FFFFFFF,0xFFFFFFFF,0x7FFFFFFF)); + return _mm_and_pd(a,mask); +} +template<> EIGEN_STRONG_INLINE Packet4i pabs(const Packet4i& a) +{ + #ifdef EIGEN_VECTORIZE_SSSE3 + return _mm_abs_epi32(a); + #else + Packet4i aux = _mm_srai_epi32(a,31); + return _mm_sub_epi32(_mm_xor_si128(a,aux),aux); + #endif +} + +EIGEN_STRONG_INLINE void punpackp(Packet4f* vecs) +{ + vecs[1] = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(vecs[0]), 0x55)); + vecs[2] = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(vecs[0]), 0xAA)); + vecs[3] = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(vecs[0]), 0xFF)); + vecs[0] = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(vecs[0]), 0x00)); +} + +#ifdef EIGEN_VECTORIZE_SSE3 +// TODO implement SSE2 versions as well as integer versions +template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs) +{ + return _mm_hadd_ps(_mm_hadd_ps(vecs[0], vecs[1]),_mm_hadd_ps(vecs[2], vecs[3])); +} +template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs) +{ + return _mm_hadd_pd(vecs[0], vecs[1]); +} +// SSSE3 version: +// EIGEN_STRONG_INLINE Packet4i preduxp(const Packet4i* vecs) +// { +// return _mm_hadd_epi32(_mm_hadd_epi32(vecs[0], vecs[1]),_mm_hadd_epi32(vecs[2], vecs[3])); +// } + +template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a) +{ + Packet4f tmp0 = _mm_hadd_ps(a,a); + return pfirst(_mm_hadd_ps(tmp0, tmp0)); +} + +template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a) { return pfirst(_mm_hadd_pd(a, a)); } + +// SSSE3 version: +// EIGEN_STRONG_INLINE float predux(const Packet4i& a) +// { +// Packet4i tmp0 = _mm_hadd_epi32(a,a); +// return pfirst(_mm_hadd_epi32(tmp0, tmp0)); +// } +#else +// SSE2 versions +template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a) +{ + Packet4f tmp = _mm_add_ps(a, _mm_movehl_ps(a,a)); + return pfirst(_mm_add_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1))); +} +template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a) +{ + return pfirst(_mm_add_sd(a, _mm_unpackhi_pd(a,a))); +} + +template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs) +{ + Packet4f tmp0, tmp1, tmp2; + tmp0 = _mm_unpacklo_ps(vecs[0], vecs[1]); + tmp1 = _mm_unpackhi_ps(vecs[0], vecs[1]); + tmp2 = _mm_unpackhi_ps(vecs[2], vecs[3]); + tmp0 = _mm_add_ps(tmp0, tmp1); + tmp1 = _mm_unpacklo_ps(vecs[2], vecs[3]); + tmp1 = _mm_add_ps(tmp1, tmp2); + tmp2 = _mm_movehl_ps(tmp1, tmp0); + tmp0 = _mm_movelh_ps(tmp0, tmp1); + return _mm_add_ps(tmp0, tmp2); +} + +template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs) +{ + return _mm_add_pd(_mm_unpacklo_pd(vecs[0], vecs[1]), _mm_unpackhi_pd(vecs[0], vecs[1])); +} +#endif // SSE3 + +template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a) +{ + Packet4i tmp = _mm_add_epi32(a, _mm_unpackhi_epi64(a,a)); + return pfirst(tmp) + pfirst(_mm_shuffle_epi32(tmp, 1)); +} + +template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs) +{ + Packet4i tmp0, tmp1, tmp2; + tmp0 = _mm_unpacklo_epi32(vecs[0], vecs[1]); + tmp1 = _mm_unpackhi_epi32(vecs[0], vecs[1]); + tmp2 = _mm_unpackhi_epi32(vecs[2], vecs[3]); + tmp0 = _mm_add_epi32(tmp0, tmp1); + tmp1 = _mm_unpacklo_epi32(vecs[2], vecs[3]); + tmp1 = _mm_add_epi32(tmp1, tmp2); + tmp2 = _mm_unpacklo_epi64(tmp0, tmp1); + tmp0 = _mm_unpackhi_epi64(tmp0, tmp1); + return _mm_add_epi32(tmp0, tmp2); +} + +// Other reduction functions: + +// mul +template<> EIGEN_STRONG_INLINE float predux_mul<Packet4f>(const Packet4f& a) +{ + Packet4f tmp = _mm_mul_ps(a, _mm_movehl_ps(a,a)); + return pfirst(_mm_mul_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1))); +} +template<> EIGEN_STRONG_INLINE double predux_mul<Packet2d>(const Packet2d& a) +{ + return pfirst(_mm_mul_sd(a, _mm_unpackhi_pd(a,a))); +} +template<> EIGEN_STRONG_INLINE int predux_mul<Packet4i>(const Packet4i& a) +{ + // after some experiments, it is seems this is the fastest way to implement it + // for GCC (eg., reusing pmul is very slow !) + // TODO try to call _mm_mul_epu32 directly + EIGEN_ALIGN16 int aux[4]; + pstore(aux, a); + return (aux[0] * aux[1]) * (aux[2] * aux[3]);; +} + +// min +template<> EIGEN_STRONG_INLINE float predux_min<Packet4f>(const Packet4f& a) +{ + Packet4f tmp = _mm_min_ps(a, _mm_movehl_ps(a,a)); + return pfirst(_mm_min_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1))); +} +template<> EIGEN_STRONG_INLINE double predux_min<Packet2d>(const Packet2d& a) +{ + return pfirst(_mm_min_sd(a, _mm_unpackhi_pd(a,a))); +} +template<> EIGEN_STRONG_INLINE int predux_min<Packet4i>(const Packet4i& a) +{ + // after some experiments, it is seems this is the fastest way to implement it + // for GCC (eg., it does not like using std::min after the pstore !!) + EIGEN_ALIGN16 int aux[4]; + pstore(aux, a); + register int aux0 = aux[0]<aux[1] ? aux[0] : aux[1]; + register int aux2 = aux[2]<aux[3] ? aux[2] : aux[3]; + return aux0<aux2 ? aux0 : aux2; +} + +// max +template<> EIGEN_STRONG_INLINE float predux_max<Packet4f>(const Packet4f& a) +{ + Packet4f tmp = _mm_max_ps(a, _mm_movehl_ps(a,a)); + return pfirst(_mm_max_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1))); +} +template<> EIGEN_STRONG_INLINE double predux_max<Packet2d>(const Packet2d& a) +{ + return pfirst(_mm_max_sd(a, _mm_unpackhi_pd(a,a))); +} +template<> EIGEN_STRONG_INLINE int predux_max<Packet4i>(const Packet4i& a) +{ + // after some experiments, it is seems this is the fastest way to implement it + // for GCC (eg., it does not like using std::min after the pstore !!) + EIGEN_ALIGN16 int aux[4]; + pstore(aux, a); + register int aux0 = aux[0]>aux[1] ? aux[0] : aux[1]; + register int aux2 = aux[2]>aux[3] ? aux[2] : aux[3]; + return aux0>aux2 ? aux0 : aux2; +} + +#if (defined __GNUC__) +// template <> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) +// { +// Packet4f res = b; +// asm("mulps %[a], %[b] \n\taddps %[c], %[b]" : [b] "+x" (res) : [a] "x" (a), [c] "x" (c)); +// return res; +// } +// EIGEN_STRONG_INLINE Packet4i _mm_alignr_epi8(const Packet4i& a, const Packet4i& b, const int i) +// { +// Packet4i res = a; +// asm("palignr %[i], %[a], %[b] " : [b] "+x" (res) : [a] "x" (a), [i] "i" (i)); +// return res; +// } +#endif + +#ifdef EIGEN_VECTORIZE_SSSE3 +// SSSE3 versions +template<int Offset> +struct palign_impl<Offset,Packet4f> +{ + static EIGEN_STRONG_INLINE void run(Packet4f& first, const Packet4f& second) + { + if (Offset!=0) + first = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(second), _mm_castps_si128(first), Offset*4)); + } +}; + +template<int Offset> +struct palign_impl<Offset,Packet4i> +{ + static EIGEN_STRONG_INLINE void run(Packet4i& first, const Packet4i& second) + { + if (Offset!=0) + first = _mm_alignr_epi8(second,first, Offset*4); + } +}; + +template<int Offset> +struct palign_impl<Offset,Packet2d> +{ + static EIGEN_STRONG_INLINE void run(Packet2d& first, const Packet2d& second) + { + if (Offset==1) + first = _mm_castsi128_pd(_mm_alignr_epi8(_mm_castpd_si128(second), _mm_castpd_si128(first), 8)); + } +}; +#else +// SSE2 versions +template<int Offset> +struct palign_impl<Offset,Packet4f> +{ + static EIGEN_STRONG_INLINE void run(Packet4f& first, const Packet4f& second) + { + if (Offset==1) + { + first = _mm_move_ss(first,second); + first = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(first),0x39)); + } + else if (Offset==2) + { + first = _mm_movehl_ps(first,first); + first = _mm_movelh_ps(first,second); + } + else if (Offset==3) + { + first = _mm_move_ss(first,second); + first = _mm_shuffle_ps(first,second,0x93); + } + } +}; + +template<int Offset> +struct palign_impl<Offset,Packet4i> +{ + static EIGEN_STRONG_INLINE void run(Packet4i& first, const Packet4i& second) + { + if (Offset==1) + { + first = _mm_castps_si128(_mm_move_ss(_mm_castsi128_ps(first),_mm_castsi128_ps(second))); + first = _mm_shuffle_epi32(first,0x39); + } + else if (Offset==2) + { + first = _mm_castps_si128(_mm_movehl_ps(_mm_castsi128_ps(first),_mm_castsi128_ps(first))); + first = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(first),_mm_castsi128_ps(second))); + } + else if (Offset==3) + { + first = _mm_castps_si128(_mm_move_ss(_mm_castsi128_ps(first),_mm_castsi128_ps(second))); + first = _mm_castps_si128(_mm_shuffle_ps(_mm_castsi128_ps(first),_mm_castsi128_ps(second),0x93)); + } + } +}; + +template<int Offset> +struct palign_impl<Offset,Packet2d> +{ + static EIGEN_STRONG_INLINE void run(Packet2d& first, const Packet2d& second) + { + if (Offset==1) + { + first = _mm_castps_pd(_mm_movehl_ps(_mm_castpd_ps(first),_mm_castpd_ps(first))); + first = _mm_castps_pd(_mm_movelh_ps(_mm_castpd_ps(first),_mm_castpd_ps(second))); + } + } +}; +#endif + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_PACKET_MATH_SSE_H |