diff options
author | 2014-09-22 12:57:06 +0200 | |
---|---|---|
committer | 2014-09-22 12:57:06 +0200 | |
commit | d9e0336a789ba796c1c7bd6d8f7a11794837d03e (patch) | |
tree | 71ed168b16ff4737b96b0741ccd2ac3f53679973 /Eigen/src/Core | |
parent | 333905b0c2e2e377744c7ed326ead34625510530 (diff) | |
parent | de38ff24993af77f188ab97d458d2884eeffd1ba (diff) |
Merged in kmargar/eigen (pull request PR-84)
Add VSX support
Diffstat (limited to 'Eigen/src/Core')
-rw-r--r-- | Eigen/src/Core/arch/AltiVec/Complex.h | 228 | ||||
-rwxr-xr-x | Eigen/src/Core/arch/AltiVec/PacketMath.h | 411 | ||||
-rw-r--r-- | Eigen/src/Core/products/GeneralBlockPanelKernel.h | 4 | ||||
-rw-r--r-- | Eigen/src/Core/util/Constants.h | 6 |
4 files changed, 582 insertions, 67 deletions
diff --git a/Eigen/src/Core/arch/AltiVec/Complex.h b/Eigen/src/Core/arch/AltiVec/Complex.h index 13b874d0c..30e2088ef 100644 --- a/Eigen/src/Core/arch/AltiVec/Complex.h +++ b/Eigen/src/Core/arch/AltiVec/Complex.h @@ -7,23 +7,19 @@ // 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 +#ifndef EIGEN_COMPLEX32_ALTIVEC_H +#define EIGEN_COMPLEX32_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(p16uc_DUPLICATE, (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 3), 8);//{ 4,5,6,7, 4,5,6,7, 12,13,14,15, 12,13,14,15 }; -static Packet16uc p16uc_COMPLEX_REV = vec_sld(p16uc_REVERSE, p16uc_REVERSE, 8);//{ 4,5,6,7, 0,1,2,3, 12,13,14,15, 8,9,10,11 }; -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)p16uc_COMPLEX_RE, (Packet4ui)p16uc_COMPLEX_IM);//{ 0,1,2,3, 4,5,6,7, 0,1,2,3, 4,5,6,7 }; -static Packet16uc p16uc_PSET_LO = (Packet16uc) vec_mergel((Packet4ui)p16uc_COMPLEX_RE, (Packet4ui)p16uc_COMPLEX_IM);//{ 8,9,10,11, 12,13,14,15, 8,9,10,11, 12,13,14,15 }; -static Packet16uc p16uc_COMPLEX_MASK16 = vec_sld((Packet16uc)p4i_ZERO, vec_splat((Packet16uc) vec_abs(p4i_MINUS16), 3), 8);//{ 0,0,0,0, 0,0,0,0, 16,16,16,16, 16,16,16,16}; -static Packet16uc p16uc_COMPLEX_TRANSPOSE_0 = vec_add(p16uc_PSET_HI, p16uc_COMPLEX_MASK16);//{ 0,1,2,3, 4,5,6,7, 16,17,18,19, 20,21,22,23}; -static Packet16uc p16uc_COMPLEX_TRANSPOSE_1 = vec_add(p16uc_PSET_LO, p16uc_COMPLEX_MASK16);//{ 8,9,10,11, 12,13,14,15, 24,25,26,27, 28,29,30,31}; +#ifdef _BIG_ENDIAN +#else +static Packet2ul p2ul_CONJ_XOR1 = (Packet2ul) vec_sld((Packet4ui) p2l_ZERO, (Packet4ui) p2d_ZERO_, 8);//{ 0x8000000000000000, 0x0000000000000000 }; +static Packet2ul p2ul_CONJ_XOR2 = (Packet2ul) vec_sld((Packet4ui) p2d_ZERO_, (Packet4ui) p2l_ZERO, 8);//{ 0x8000000000000000, 0x0000000000000000 }; +#endif //---------- float ---------- struct Packet2cf @@ -65,7 +61,7 @@ template<> EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<flo 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); + res.v = vec_perm(res.v, res.v, p16uc_PSET64_HI); return res; } @@ -95,16 +91,16 @@ template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, con Packet4f v1, v2; // Permute and multiply the real parts of a and b - v1 = vec_perm(a.v, a.v, p16uc_COMPLEX_RE); + v1 = vec_perm(a.v, a.v, p16uc_PSET32_WODD); // Get the imaginary parts of a - v2 = vec_perm(a.v, a.v, p16uc_COMPLEX_IM); + v2 = vec_perm(a.v, a.v, p16uc_PSET32_WEVEN); // 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); + v2 = vec_perm(v2, v2, p16uc_COMPLEX32_REV); return Packet2cf(vec_add(v1, v2)); } @@ -138,7 +134,7 @@ template<> EIGEN_STRONG_INLINE std::complex<float> pfirst<Packet2cf>(const Pack template<> EIGEN_STRONG_INLINE Packet2cf preverse(const Packet2cf& a) { Packet4f rev_a; - rev_a = vec_perm(a.v, a.v, p16uc_COMPLEX_REV2); + rev_a = vec_perm(a.v, a.v, p16uc_COMPLEX32_REV2); return Packet2cf(rev_a); } @@ -153,9 +149,13 @@ template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(const Packe template<> EIGEN_STRONG_INLINE Packet2cf preduxp<Packet2cf>(const Packet2cf* vecs) { Packet4f b1, b2; - +#ifdef _BIG_ENDIAN b1 = (Packet4f) vec_sld(vecs[0].v, vecs[1].v, 8); b2 = (Packet4f) vec_sld(vecs[1].v, vecs[0].v, 8); +#else + b1 = (Packet4f) vec_sld(vecs[1].v, vecs[0].v, 8); + b2 = (Packet4f) vec_sld(vecs[0].v, vecs[1].v, 8); +#endif b2 = (Packet4f) vec_sld(b2, b2, 8); b2 = padd(b1, b2); @@ -179,7 +179,11 @@ struct palign_impl<Offset,Packet2cf> { if (Offset==1) { +#ifdef _BIG_ENDIAN first.v = vec_sld(first.v, second.v, 8); +#else + first.v = vec_sld(second.v, first.v, 8); +#endif } } }; @@ -222,23 +226,203 @@ template<> EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, con // 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)))); + return Packet2cf(pdiv(res.v, vec_add(s,vec_perm(s, s, p16uc_COMPLEX32_REV)))); } template<> EIGEN_STRONG_INLINE Packet2cf pcplxflip<Packet2cf>(const Packet2cf& x) { - return Packet2cf(vec_perm(x.v, x.v, p16uc_COMPLEX_REV)); + return Packet2cf(vec_perm(x.v, x.v, p16uc_COMPLEX32_REV)); } EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet2cf,2>& kernel) { - Packet4f tmp = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_COMPLEX_TRANSPOSE_0); - kernel.packet[1].v = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_COMPLEX_TRANSPOSE_1); + Packet4f tmp = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_HI); + kernel.packet[1].v = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_LO); kernel.packet[0].v = tmp; } +//---------- double ---------- +#ifdef __VSX__ +struct Packet1cd +{ + EIGEN_STRONG_INLINE Packet1cd() {} + EIGEN_STRONG_INLINE explicit Packet1cd(const Packet2d& a) : v(a) {} + Packet2d v; +}; + +template<> struct packet_traits<std::complex<double> > : default_packet_traits +{ + typedef Packet1cd type; + typedef Packet1cd half; + enum { + Vectorizable = 1, + AlignedOnScalar = 0, + size = 1, + HasHalfPacket = 0, + + HasAdd = 1, + HasSub = 1, + HasMul = 1, + HasDiv = 1, + HasNegate = 1, + HasAbs = 0, + HasAbs2 = 0, + HasMin = 0, + HasMax = 0, + HasSetLinear = 0 + }; +}; + +template<> struct unpacket_traits<Packet1cd> { typedef std::complex<double> type; enum {size=1}; typedef Packet1cd half; }; + +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 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 Packet1cd pset1<Packet1cd>(const std::complex<double>& from) +{ /* here we really have to use unaligned loads :( */ return ploadu<Packet1cd>(&from); } + +template<> EIGEN_DEVICE_FUNC inline Packet1cd pgather<std::complex<double>, Packet1cd>(const std::complex<double>* from, DenseIndex stride) +{ + std::complex<double> EIGEN_ALIGN16 af[2]; + af[0] = from[0*stride]; + af[1] = from[1*stride]; + return pload<Packet1cd>(af); +} +template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<double>, Packet1cd>(std::complex<double>* to, const Packet1cd& from, DenseIndex stride) +{ + std::complex<double> EIGEN_ALIGN16 af[2]; + pstore<std::complex<double> >(af, from); + to[0*stride] = af[0]; + to[1*stride] = af[1]; +} + +template<> EIGEN_STRONG_INLINE Packet1cd padd<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_add(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet1cd psub<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_sub(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet1cd pnegate(const Packet1cd& a) { return Packet1cd(pnegate(Packet2d(a.v))); } +template<> EIGEN_STRONG_INLINE Packet1cd pconj(const Packet1cd& a) { return Packet1cd((Packet2d)vec_xor((Packet2d)a.v, (Packet2d)p2ul_CONJ_XOR2)); } + +template<> EIGEN_STRONG_INLINE Packet1cd pmul<Packet1cd>(const Packet1cd& a, const Packet1cd& b) +{ + Packet2d a_re, a_im, v1, v2; + + // Permute and multiply the real parts of a and b + a_re = vec_perm(a.v, a.v, p16uc_PSET64_HI); + // Get the imaginary parts of a + a_im = vec_perm(a.v, a.v, p16uc_PSET64_LO); + // multiply a_re * b + v1 = vec_madd(a_re, b.v, p2d_ZERO); + // multiply a_im * b and get the conjugate result + v2 = vec_madd(a_im, b.v, p2d_ZERO); + v2 = (Packet2d) vec_sld((Packet4ui)v2, (Packet4ui)v2, 8); + v2 = (Packet2d) vec_xor((Packet2d)v2, (Packet2d) p2ul_CONJ_XOR1); + + return Packet1cd(vec_add(v1, v2)); +} + +template<> EIGEN_STRONG_INLINE Packet1cd pand <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_and(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet1cd por <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_or(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet1cd pxor <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_xor(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet1cd pandnot<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_and(a.v, vec_nor(b.v,b.v))); } + +template<> EIGEN_STRONG_INLINE Packet1cd ploaddup<Packet1cd>(const std::complex<double>* from) +{ + return pset1<Packet1cd>(*from); +} + +template<> EIGEN_STRONG_INLINE void prefetch<std::complex<double> >(const std::complex<double> * addr) { vec_dstt((long *)addr, DST_CTRL(2,2,32), DST_CHAN); } + +template<> EIGEN_STRONG_INLINE std::complex<double> pfirst<Packet1cd>(const Packet1cd& a) +{ + std::complex<double> EIGEN_ALIGN16 res[2]; + pstore<std::complex<double> >(res, a); + + return res[0]; +} + +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 + { + return internal::pmul(a, pconj(b)); + } +}; + +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 + { + return internal::pmul(pconj(a), b); + } +}; + +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 + { + return pconj(internal::pmul(a, b)); + } +}; + +template<> EIGEN_STRONG_INLINE Packet1cd pdiv<Packet1cd>(const Packet1cd& a, const Packet1cd& b) +{ + // TODO optimize it for AltiVec + Packet1cd res = conj_helper<Packet1cd,Packet1cd,false,true>().pmul(a,b); + Packet2d s = vec_madd(b.v, b.v, p2d_ZERO_); + return Packet1cd(pdiv(res.v, vec_add(s,vec_perm(s, s, p16uc_COMPLEX32_REV)))); +} + +EIGEN_STRONG_INLINE Packet1cd pcplxflip/*<Packet1cd>*/(const Packet1cd& x) +{ + return Packet1cd(preverse(Packet2d(x.v))); +} + +EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet1cd,2>& kernel) +{ + Packet2d tmp = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_HI); + kernel.packet[1].v = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_LO); + kernel.packet[0].v = tmp; +} +#endif // __VSX__ } // end namespace internal } // end namespace Eigen -#endif // EIGEN_COMPLEX_ALTIVEC_H +#endif // EIGEN_COMPLEX32_ALTIVEC_H diff --git a/Eigen/src/Core/arch/AltiVec/PacketMath.h b/Eigen/src/Core/arch/AltiVec/PacketMath.h index b43e8ace3..1b86e1227 100755 --- a/Eigen/src/Core/arch/AltiVec/PacketMath.h +++ b/Eigen/src/Core/arch/AltiVec/PacketMath.h @@ -28,7 +28,7 @@ namespace internal { // 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 +#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 32 #endif typedef __vector float Packet4f; @@ -50,22 +50,20 @@ typedef __vector unsigned char Packet16uc; #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 _EIGEN_DECLARE_CONST_Packet2d(NAME,X) \ + Packet2d p2d_##NAME = pset1<Packet2d>(X) + +#define _EIGEN_DECLARE_CONST_Packet2l(NAME,X) \ + Packet2l p2l_##NAME = pset1<Packet2l>(X) + #define DST_CHAN 1 #define DST_CTRL(size, count, stride) (((size) << 24) | ((count) << 16) | (stride)) -// Define global static constants: -static Packet4f p4f_COUNTDOWN = { 0.0, 1.0, 2.0, 3.0 }; -static Packet4i p4i_COUNTDOWN = { 0, 1, 2, 3 }; -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); //{ 0,1,2,3, 4,5,6,7, 8,9,10,11, 12,13,14,15} -static Packet16uc p16uc_DUPLICATE = { 0,1,2,3, 0,1,2,3, 4,5,6,7, 4,5,6,7}; +// These constants are endian-agnostic static _EIGEN_DECLARE_CONST_FAST_Packet4f(ZERO, 0); //{ 0.0, 0.0, 0.0, 0.0} static _EIGEN_DECLARE_CONST_FAST_Packet4i(ZERO, 0); //{ 0, 0, 0, 0,} static _EIGEN_DECLARE_CONST_FAST_Packet4i(ONE,1); //{ 1, 1, 1, 1} @@ -74,6 +72,44 @@ static _EIGEN_DECLARE_CONST_FAST_Packet4i(MINUS1,-1); //{ -1, -1, -1, -1} static Packet4f p4f_ONE = vec_ctf(p4i_ONE, 0); //{ 1.0, 1.0, 1.0, 1.0} static Packet4f p4f_ZERO_ = (Packet4f) vec_sl((Packet4ui)p4i_MINUS1, (Packet4ui)p4i_MINUS1); //{ 0x80000000, 0x80000000, 0x80000000, 0x80000000} +static Packet4f p4f_COUNTDOWN = { 0.0, 1.0, 2.0, 3.0 }; +static Packet4i p4i_COUNTDOWN = { 0, 1, 2, 3 }; + +static Packet16uc p16uc_REVERSE32 = { 12,13,14,15, 8,9,10,11, 4,5,6,7, 0,1,2,3 }; +static Packet16uc p16uc_DUPLICATE32_HI = { 0,1,2,3, 0,1,2,3, 4,5,6,7, 4,5,6,7 }; + +// Handle endianness properly while loading constants +// Define global static constants: +#ifdef _BIG_ENDIAN +static Packet16uc p16uc_FORWARD = vec_lvsl(0, (float*)0); +static Packet16uc p16uc_PSET32_WODD = 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_PSET32_WEVEN = vec_sld(p16uc_DUPLICATE32_HI, (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_HALF64_0_16 = vec_sld((Packet16uc)p4i_ZERO, vec_splat((Packet16uc) vec_abs(p4i_MINUS16), 3), 8); //{ 0,0,0,0, 0,0,0,0, 16,16,16,16, 16,16,16,16}; +#else +// Mask alignment +#define _EIGEN_MASK_ALIGNMENT 0xfffffffffffffff0 +#define _EIGEN_ALIGNED_PTR(x) ((ptrdiff_t)(x) & _EIGEN_MASK_ALIGNMENT) + +static Packet16uc p16uc_FORWARD = p16uc_REVERSE32; +static Packet16uc p16uc_REVERSE64 = { 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 }; +static Packet16uc p16uc_PSET32_WODD = vec_sld((Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 1), (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 3), 8);//{ 0,1,2,3, 0,1,2,3, 8,9,10,11, 8,9,10,11 }; +static Packet16uc p16uc_PSET32_WEVEN = vec_sld((Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 0), (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 2), 8);//{ 4,5,6,7, 4,5,6,7, 12,13,14,15, 12,13,14,15 }; +static Packet16uc p16uc_HALF64_0_16 = vec_sld(vec_splat((Packet16uc) vec_abs(p4i_MINUS16), 0), (Packet16uc)p4i_ZERO, 8); //{ 0,0,0,0, 0,0,0,0, 16,16,16,16, 16,16,16,16}; +#endif // _BIG_ENDIAN + +static Packet16uc p16uc_PSET64_HI = (Packet16uc) vec_mergeh((Packet4ui)p16uc_PSET32_WODD, (Packet4ui)p16uc_PSET32_WEVEN); //{ 0,1,2,3, 4,5,6,7, 0,1,2,3, 4,5,6,7 }; +static Packet16uc p16uc_PSET64_LO = (Packet16uc) vec_mergel((Packet4ui)p16uc_PSET32_WODD, (Packet4ui)p16uc_PSET32_WEVEN); //{ 8,9,10,11, 12,13,14,15, 8,9,10,11, 12,13,14,15 }; +static Packet16uc p16uc_TRANSPOSE64_HI = vec_add(p16uc_PSET64_HI, p16uc_HALF64_0_16); //{ 0,1,2,3, 4,5,6,7, 16,17,18,19, 20,21,22,23}; +static Packet16uc p16uc_TRANSPOSE64_LO = vec_add(p16uc_PSET64_LO, p16uc_HALF64_0_16); //{ 8,9,10,11, 12,13,14,15, 24,25,26,27, 28,29,30,31}; + +static Packet16uc p16uc_COMPLEX32_REV = vec_sld(p16uc_REVERSE32, p16uc_REVERSE32, 8); //{ 4,5,6,7, 0,1,2,3, 12,13,14,15, 8,9,10,11 }; + +#ifdef _BIG_ENDIAN +static Packet16uc p16uc_COMPLEX32_REV2 = vec_sld(p16uc_FORWARD, p16uc_FORWARD, 8); //{ 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 }; +#else +static Packet16uc p16uc_COMPLEX32_REV2 = vec_sld(p16uc_PSET64_HI, p16uc_PSET64_LO, 8); //{ 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 }; +#endif // _BIG_ENDIAN + template<> struct packet_traits<float> : default_packet_traits { typedef Packet4f type; @@ -105,9 +141,22 @@ template<> struct packet_traits<int> : default_packet_traits }; }; + template<> struct unpacket_traits<Packet4f> { typedef float type; enum {size=4}; typedef Packet4f half; }; template<> struct unpacket_traits<Packet4i> { typedef int type; enum {size=4}; typedef Packet4i half; }; -/* + +inline std::ostream & operator <<(std::ostream & s, const Packet16uc & v) +{ + union { + Packet16uc v; + unsigned char n[16]; + } vt; + vt.v = v; + for (int i=0; i< 16; i++) + s << (int)vt.n[i] << ", "; + return s; +} + inline std::ostream & operator <<(std::ostream & s, const Packet4f & v) { union { @@ -140,7 +189,7 @@ inline std::ostream & operator <<(std::ostream & s, const Packet4ui & 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 { @@ -150,14 +199,21 @@ inline std::ostream & operator <<(std::ostream & s, const Packetbi & v) vt.v = v; s << vt.n[0] << ", " << vt.n[1] << ", " << vt.n[2] << ", " << vt.n[3]; return s; -} -*/ +}*/ + + +// Need to define them first or we get specialization after instantiation errors +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 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 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); + Packet4f vc = pload<Packet4f>(af); vc = vec_splat(vc, 0); return vc; } @@ -165,17 +221,15 @@ template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float& from) { template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int& from) { int EIGEN_ALIGN16 ai[4]; ai[0] = from; - Packet4i vc = vec_ld(0, ai); + Packet4i vc = pload<Packet4i>(ai); vc = vec_splat(vc, 0); return vc; } - - template<> EIGEN_STRONG_INLINE void pbroadcast4<Packet4f>(const float *a, Packet4f& a0, Packet4f& a1, Packet4f& a2, Packet4f& a3) { - a3 = vec_ld(0,a); + a3 = pload<Packet4f>(a); a0 = vec_splat(a3, 0); a1 = vec_splat(a3, 1); a2 = vec_splat(a3, 2); @@ -185,7 +239,7 @@ template<> EIGEN_STRONG_INLINE void pbroadcast4<Packet4i>(const int *a, Packet4i& a0, Packet4i& a1, Packet4i& a2, Packet4i& a3) { - a3 = vec_ld(0,a); + a3 = pload<Packet4i>(a); a0 = vec_splat(a3, 0); a1 = vec_splat(a3, 1); a2 = vec_splat(a3, 2); @@ -199,7 +253,7 @@ template<> EIGEN_DEVICE_FUNC inline Packet4f pgather<float, Packet4f>(const floa af[1] = from[1*stride]; af[2] = from[2*stride]; af[3] = from[3*stride]; - return vec_ld(0, af); + return pload<Packet4f>(af); } template<> EIGEN_DEVICE_FUNC inline Packet4i pgather<int, Packet4i>(const int* from, DenseIndex stride) { @@ -208,12 +262,12 @@ template<> EIGEN_DEVICE_FUNC inline Packet4i pgather<int, Packet4i>(const int* f ai[1] = from[1*stride]; ai[2] = from[2*stride]; ai[3] = from[3*stride]; - return vec_ld(0, ai); + return pload<Packet4i>(ai); } template<> EIGEN_DEVICE_FUNC inline void pscatter<float, Packet4f>(float* to, const Packet4f& from, DenseIndex stride) { float EIGEN_ALIGN16 af[4]; - vec_st(from, 0, af); + pstore<float>(af, from); to[0*stride] = af[0]; to[1*stride] = af[1]; to[2*stride] = af[2]; @@ -222,7 +276,7 @@ template<> EIGEN_DEVICE_FUNC inline void pscatter<float, Packet4f>(float* to, co template<> EIGEN_DEVICE_FUNC inline void pscatter<int, Packet4i>(int* to, const Packet4i& from, DenseIndex stride) { int EIGEN_ALIGN16 ai[4]; - vec_st(from, 0, ai); + pstore<int>((int *)ai, from); to[0*stride] = ai[0]; to[1*stride] = ai[1]; to[2*stride] = ai[2]; @@ -283,7 +337,8 @@ template<> EIGEN_STRONG_INLINE Packet4i pmul<Packet4i>(const Packet4i& a, const */ template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const Packet4f& b) { - Packet4f t, y_0, y_1, res; +#ifndef __VSX__ // VSX actually provides a div instruction + Packet4f t, y_0, y_1; // Altivec does not offer a divide instruction, we have to do a reciprocal approximation y_0 = vec_re(b); @@ -292,8 +347,10 @@ template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const 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; + return vec_madd(a, y_1, p4f_ZERO); +#else + return vec_div(a, b); +#endif } template<> EIGEN_STRONG_INLINE Packet4i pdiv<Packet4i>(const Packet4i& /*a*/, const Packet4i& /*b*/) @@ -311,7 +368,6 @@ template<> EIGEN_STRONG_INLINE Packet4i pmin<Packet4i>(const Packet4i& a, const 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); } @@ -324,13 +380,10 @@ template<> EIGEN_STRONG_INLINE Packet4i pxor<Packet4i>(const Packet4i& a, const 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); } - +#ifdef _BIG_ENDIAN 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 @@ -350,25 +403,36 @@ template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from) mask = vec_lvsl(0, from); // create the permute mask return (Packet4i) vec_perm(MSQ, LSQ, mask); // align the data } +#else +// We also need ot redefine little endian loading of Packet4i/Packet4f using VSX +template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from) +{ + EIGEN_DEBUG_ALIGNED_LOAD + return (Packet4i) vec_vsx_ld((long)from & 15, (const int*) _EIGEN_ALIGNED_PTR(from)); +} +template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from) +{ + EIGEN_DEBUG_ALIGNED_LOAD + return (Packet4f) vec_vsx_ld((long)from & 15, (const float*) _EIGEN_ALIGNED_PTR(from)); +} +#endif 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); + return vec_perm(p, p, p16uc_DUPLICATE32_HI); } 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); + return vec_perm(p, p, p16uc_DUPLICATE32_HI); } -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); } - +#ifdef _BIG_ENDIAN template<> EIGEN_STRONG_INLINE void pstoreu<float>(float* to, const Packet4f& from) { EIGEN_DEBUG_UNALIGNED_STORE @@ -405,15 +469,30 @@ template<> EIGEN_STRONG_INLINE void pstoreu<int>(int* to, const Packet4i& f vec_st( LSQ, 15, (unsigned char *)to ); // Store the LSQ part first vec_st( MSQ, 0, (unsigned char *)to ); // Store the MSQ part } +#else +// We also need ot redefine little endian loading of Packet4i/Packet4f using VSX +template<> EIGEN_STRONG_INLINE void pstoreu<int>(int* to, const Packet4i& from) +{ + EIGEN_DEBUG_ALIGNED_STORE + vec_vsx_st(from, (long)to & 15, (int*) _EIGEN_ALIGNED_PTR(to)); +} +template<> EIGEN_STRONG_INLINE void pstoreu<float>(float* to, const Packet4f& from) +{ + EIGEN_DEBUG_ALIGNED_STORE + vec_vsx_st(from, (long)to & 15, (float*) _EIGEN_ALIGNED_PTR(to)); +} +#endif +#ifndef __VSX__ 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); } +#endif 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 preverse(const Packet4f& a) { return (Packet4f)vec_perm((Packet16uc)a,(Packet16uc)a, p16uc_REVERSE32); } +template<> EIGEN_STRONG_INLINE Packet4i preverse(const Packet4i& a) { return (Packet4i)vec_perm((Packet16uc)a,(Packet16uc)a, p16uc_REVERSE32); } 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); } @@ -460,7 +539,11 @@ template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a) { Packet4i sum; sum = vec_sums(a, p4i_ZERO); +#ifdef _BIG_ENDIAN sum = vec_sld(sum, p4i_ZERO, 12); +#else + sum = vec_sld(p4i_ZERO, sum, 4); +#endif return pfirst(sum); } @@ -547,8 +630,25 @@ 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); +#ifdef _BIG_ENDIAN + switch (Offset % 4) { + case 1: + first = vec_sld(first, second, 4); break; + case 2: + first = vec_sld(first, second, 8); break; + case 3: + first = vec_sld(first, second, 12); break; + } +#else + switch (Offset % 4) { + case 1: + first = vec_sld(second, first, 12); break; + case 2: + first = vec_sld(second, first, 8); break; + case 3: + first = vec_sld(second, first, 4); break; + } +#endif } }; @@ -557,8 +657,25 @@ 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); +#ifdef _BIG_ENDIAN + switch (Offset % 4) { + case 1: + first = vec_sld(first, second, 4); break; + case 2: + first = vec_sld(first, second, 8); break; + case 3: + first = vec_sld(first, second, 12); break; + } +#else + switch (Offset % 4) { + case 1: + first = vec_sld(second, first, 12); break; + case 2: + first = vec_sld(second, first, 8); break; + case 3: + first = vec_sld(second, first, 4); break; + } +#endif } }; @@ -588,6 +705,214 @@ ptranspose(PacketBlock<Packet4i,4>& kernel) { kernel.packet[3] = vec_mergel(t1, t3); } + +//---------- double ---------- +#ifdef __VSX__ +typedef __vector double Packet2d; +typedef __vector unsigned long long Packet2ul; +typedef __vector long long Packet2l; + +static Packet2l p2l_ZERO = (Packet2l) p4i_ZERO; +static Packet2d p2d_ONE = { 1.0, 1.0 }; +static Packet2d p2d_ZERO = (Packet2d) p4f_ZERO; +static Packet2d p2d_ZERO_ = { -0.0, -0.0 }; +static Packet2d p2d_COUNTDOWN = (Packet2d) vec_sld((Packet16uc) p2d_ONE, (Packet16uc) p2d_ZERO, 8); + +static EIGEN_STRONG_INLINE Packet2d vec_splat_dbl(Packet2d& a, int index) +{ + switch (index) { + case 0: + return (Packet2d) vec_perm(a, a, p16uc_PSET64_HI); + case 1: + return (Packet2d) vec_perm(a, a, p16uc_PSET64_LO); + } + return a; +} + +template<> struct packet_traits<double> : default_packet_traits +{ + typedef Packet2d type; + typedef Packet2d half; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size=2, + HasHalfPacket = 0, + + HasDiv = 1, + HasExp = 0, + HasSqrt = 0 + }; +}; + +template<> struct unpacket_traits<Packet2d> { typedef double type; enum {size=2}; typedef Packet2d half; }; + + +inline std::ostream & operator <<(std::ostream & s, const Packet2d & v) +{ + union { + Packet2d v; + double n[2]; + } vt; + vt.v = v; + s << vt.n[0] << ", " << vt.n[1]; + return s; +} + +// Need to define them first or we get specialization after instantiation errors +template<> EIGEN_STRONG_INLINE Packet2d pload<Packet2d>(const double* from) { EIGEN_DEBUG_ALIGNED_LOAD return (Packet2d) vec_ld(0, (const float *) from); } //FIXME + +template<> EIGEN_STRONG_INLINE void pstore<double>(double* to, const Packet2d& from) { EIGEN_DEBUG_ALIGNED_STORE vec_st((Packet4f)from, 0, (float *)to); } + +template<> EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(const double& from) { + double EIGEN_ALIGN16 af[2]; + af[0] = from; + Packet2d vc = pload<Packet2d>(af); + vc = vec_splat_dbl(vc, 0); + return vc; +} +template<> EIGEN_STRONG_INLINE void +pbroadcast4<Packet2d>(const double *a, + Packet2d& a0, Packet2d& a1, Packet2d& a2, Packet2d& a3) +{ + a1 = pload<Packet2d>(a); + a0 = vec_splat_dbl(a1, 0); + a1 = vec_splat_dbl(a1, 1); + a3 = pload<Packet2d>(a+2); + a2 = vec_splat_dbl(a3, 0); + a3 = vec_splat_dbl(a3, 1); +} +template<> EIGEN_DEVICE_FUNC inline Packet2d pgather<double, Packet2d>(const double* from, DenseIndex stride) +{ + double EIGEN_ALIGN16 af[2]; + af[0] = from[0*stride]; + af[1] = from[1*stride]; + return pload<Packet2d>(af); +} +template<> EIGEN_DEVICE_FUNC inline void pscatter<double, Packet2d>(double* to, const Packet2d& from, DenseIndex stride) +{ + double EIGEN_ALIGN16 af[2]; + pstore<double>(af, from); + to[0*stride] = af[0]; + to[1*stride] = af[1]; +} +template<> EIGEN_STRONG_INLINE Packet2d plset<double>(const double& a) { return vec_add(pset1<Packet2d>(a), p2d_COUNTDOWN); } + +template<> EIGEN_STRONG_INLINE Packet2d padd<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_add(a,b); } + +template<> EIGEN_STRONG_INLINE Packet2d psub<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_sub(a,b); } + +template<> EIGEN_STRONG_INLINE Packet2d pnegate(const Packet2d& a) { return psub<Packet2d>(p2d_ZERO, a); } + +template<> EIGEN_STRONG_INLINE Packet2d pconj(const Packet2d& a) { return a; } + +template<> EIGEN_STRONG_INLINE Packet2d pmul<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_madd(a,b,p2d_ZERO); } +template<> EIGEN_STRONG_INLINE Packet2d pdiv<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_div(a,b); } + +// for some weird raisons, it has to be overloaded for packet of integers +template<> EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& b, const Packet2d& c) { return vec_madd(a, b, c); } + +template<> EIGEN_STRONG_INLINE Packet2d pmin<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_min(a, b); } + +template<> EIGEN_STRONG_INLINE Packet2d pmax<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_max(a, b); } + +template<> EIGEN_STRONG_INLINE Packet2d pand<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_and(a, b); } + +template<> EIGEN_STRONG_INLINE Packet2d por<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_or(a, b); } + +template<> EIGEN_STRONG_INLINE Packet2d pxor<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_xor(a, b); } + +template<> EIGEN_STRONG_INLINE Packet2d pandnot<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_and(a, vec_nor(b, b)); } + +template<> EIGEN_STRONG_INLINE Packet2d ploadu<Packet2d>(const double* from) +{ + EIGEN_DEBUG_ALIGNED_LOAD + return (Packet2d) vec_vsx_ld((long)from & 15, (const float*) _EIGEN_ALIGNED_PTR(from)); +} +template<> EIGEN_STRONG_INLINE Packet2d ploaddup<Packet2d>(const double* from) +{ + Packet2d p; + if((ptrdiff_t(from) % 16) == 0) p = pload<Packet2d>(from); + else p = ploadu<Packet2d>(from); + return vec_perm(p, p, p16uc_PSET64_HI); +} + +template<> EIGEN_STRONG_INLINE void pstoreu<double>(double* to, const Packet2d& from) +{ + EIGEN_DEBUG_ALIGNED_STORE + vec_vsx_st((Packet4f)from, (long)to & 15, (float*) _EIGEN_ALIGNED_PTR(to)); +} + +template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { vec_dstt((const float *) addr, DST_CTRL(2,2,32), DST_CHAN); } + +template<> EIGEN_STRONG_INLINE double pfirst<Packet2d>(const Packet2d& a) { double EIGEN_ALIGN16 x[2]; pstore(x, a); return x[0]; } + +template<> EIGEN_STRONG_INLINE Packet2d preverse(const Packet2d& a) { return (Packet2d)vec_perm((Packet16uc)a,(Packet16uc)a, p16uc_REVERSE64); } + +template<> EIGEN_STRONG_INLINE Packet2d pabs(const Packet2d& a) { return vec_abs(a); } + +template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a) +{ + Packet2d b, sum; + b = (Packet2d) vec_sld((Packet4ui) a, (Packet4ui)a, 8); + sum = vec_add(a, b); + return pfirst(sum); +} + +template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs) +{ + Packet2d v[2], sum; + + v[0] = vec_add(vecs[0], (Packet2d) vec_sld((Packet4ui) vecs[0], (Packet4ui) vecs[0], 8)); + v[1] = vec_add(vecs[1], (Packet2d) vec_sld((Packet4ui) vecs[1], (Packet4ui) vecs[1], 8)); + + sum = (Packet2d) vec_sld((Packet4ui) v[1], (Packet4ui) v[0], 8); + + return sum; +} +// Other reduction functions: +// mul +template<> EIGEN_STRONG_INLINE double predux_mul<Packet2d>(const Packet2d& a) +{ + return pfirst(pmul(a, (Packet2d)vec_sld((Packet4ui) a, (Packet4ui) a, 8))); +} + +// min +template<> EIGEN_STRONG_INLINE double predux_min<Packet2d>(const Packet2d& a) +{ + return pfirst(vec_min(a, (Packet2d) vec_sld((Packet4ui) a, (Packet4ui) a, 8))); +} + +// max +template<> EIGEN_STRONG_INLINE double predux_max<Packet2d>(const Packet2d& a) +{ + return pfirst(vec_max(a, (Packet2d) vec_sld((Packet4ui) a, (Packet4ui) a, 8))); +} + +template<int Offset> +struct palign_impl<Offset,Packet2d> +{ + static EIGEN_STRONG_INLINE void run(Packet2d& first, const Packet2d& second) + { + if (Offset == 1) +#ifdef _BIG_ENDIAN + first = (Packet2d) vec_sld((Packet4ui) first, (Packet4ui) second, 8); +#else + first = (Packet2d) vec_sld((Packet4ui) second, (Packet4ui) first, 8); +#endif + } +}; + +EIGEN_DEVICE_FUNC inline void +ptranspose(PacketBlock<Packet2d,2>& kernel) { + Packet2d t0, t1; + t0 = vec_perm(kernel.packet[0], kernel.packet[1], p16uc_TRANSPOSE64_HI); + t1 = vec_perm(kernel.packet[0], kernel.packet[1], p16uc_TRANSPOSE64_LO); + kernel.packet[0] = t0; + kernel.packet[1] = t1; +} + +#endif // __VSX__ } // end namespace internal } // end namespace Eigen diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h index 7da52c2e8..7b2ed6728 100644 --- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h +++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h @@ -182,7 +182,7 @@ public: nr = 4, // register block size along the M direction (currently, this one cannot be modified) -#if defined(EIGEN_HAS_FUSED_MADD) && !defined(EIGEN_VECTORIZE_ALTIVEC) +#if defined(EIGEN_HAS_FUSED_MADD) && !defined(EIGEN_VECTORIZE_ALTIVEC) && !defined(EIGEN_VECTORIZE_VSX) // we assume 16 registers mr = 3*LhsPacketSize, #else @@ -290,7 +290,7 @@ public: NumberOfRegisters = EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS, nr = 4, -#if defined(EIGEN_HAS_FUSED_MADD) && !defined(EIGEN_VECTORIZE_ALTIVEC) +#if defined(EIGEN_HAS_FUSED_MADD) && !defined(EIGEN_VECTORIZE_ALTIVEC) && !defined(EIGEN_VECTORIZE_VSX) // we assume 16 registers mr = 3*LhsPacketSize, #else diff --git a/Eigen/src/Core/util/Constants.h b/Eigen/src/Core/util/Constants.h index 86817a989..2c9fb443d 100644 --- a/Eigen/src/Core/util/Constants.h +++ b/Eigen/src/Core/util/Constants.h @@ -422,10 +422,16 @@ namespace Architecture Generic = 0x0, SSE = 0x1, AltiVec = 0x2, + VSX = 0x3, + NEON = 0x4, #if defined EIGEN_VECTORIZE_SSE Target = SSE #elif defined EIGEN_VECTORIZE_ALTIVEC Target = AltiVec +#elif defined EIGEN_VECTORIZE_VSX + Target = VSX +#elif defined EIGEN_VECTORIZE_NEON + Target = NEON #else Target = Generic #endif |