diff options
Diffstat (limited to 'Eigen/src/Core/arch/AltiVec/Complex.h')
-rw-r--r-- | Eigen/src/Core/arch/AltiVec/Complex.h | 352 |
1 files changed, 154 insertions, 198 deletions
diff --git a/Eigen/src/Core/arch/AltiVec/Complex.h b/Eigen/src/Core/arch/AltiVec/Complex.h index 67db2f8ee..f424f11cf 100644 --- a/Eigen/src/Core/arch/AltiVec/Complex.h +++ b/Eigen/src/Core/arch/AltiVec/Complex.h @@ -29,8 +29,54 @@ static Packet2ul p2ul_CONJ_XOR2 = (Packet2ul) vec_sld((Packet4ui) p2d_MZERO, (P //---------- float ---------- struct Packet2cf { - EIGEN_STRONG_INLINE explicit Packet2cf() : v(p4f_ZERO) {} + EIGEN_STRONG_INLINE explicit Packet2cf() {} EIGEN_STRONG_INLINE explicit Packet2cf(const Packet4f& a) : v(a) {} + + EIGEN_STRONG_INLINE Packet2cf pmul(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_PSET32_WODD); + // Get the imaginary parts of a + 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 = reinterpret_cast<Packet4f>(pxor(v2, reinterpret_cast<Packet4f>(p4ui_CONJ_XOR))); + // permute back to a proper order + v2 = vec_perm(v2, v2, p16uc_COMPLEX32_REV); + + return Packet2cf(padd<Packet4f>(v1, v2)); + } + + EIGEN_STRONG_INLINE Packet2cf& operator*=(const Packet2cf& b) { + v = pmul(Packet2cf(*this), b).v; + return *this; + } + EIGEN_STRONG_INLINE Packet2cf operator*(const Packet2cf& b) const { + return Packet2cf(*this) *= b; + } + + EIGEN_STRONG_INLINE Packet2cf& operator+=(const Packet2cf& b) { + v = padd(v, b.v); + return *this; + } + EIGEN_STRONG_INLINE Packet2cf operator+(const Packet2cf& b) const { + return Packet2cf(*this) += b; + } + EIGEN_STRONG_INLINE Packet2cf& operator-=(const Packet2cf& b) { + v = psub(v, b.v); + return *this; + } + EIGEN_STRONG_INLINE Packet2cf operator-(const Packet2cf& b) const { + return Packet2cf(*this) -= b; + } + EIGEN_STRONG_INLINE Packet2cf operator-(void) const { + return Packet2cf(-v); + } + Packet4f v; }; @@ -38,6 +84,7 @@ template<> struct packet_traits<std::complex<float> > : default_packet_traits { typedef Packet2cf type; typedef Packet2cf half; + typedef Packet4f as_real; enum { Vectorizable = 1, AlignedOnScalar = 1, @@ -60,7 +107,7 @@ template<> struct packet_traits<std::complex<float> > : default_packet_traits }; }; -template<> struct unpacket_traits<Packet2cf> { typedef std::complex<float> type; enum {size=2, alignment=Aligned16}; typedef Packet2cf half; }; +template<> struct unpacket_traits<Packet2cf> { typedef std::complex<float> type; enum {size=2, alignment=Aligned16, vectorizable=true, masked_load_available=false, masked_store_available=false}; typedef Packet2cf half; typedef Packet4f as_real; }; template<> EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<float>& from) { @@ -80,16 +127,35 @@ template<> EIGEN_STRONG_INLINE Packet2cf ploaddup<Packet2cf>(const std::complex< template<> EIGEN_STRONG_INLINE void pstore <std::complex<float> >(std::complex<float> * to, const Packet2cf& from) { pstore((float*)to, from.v); } template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<float> >(std::complex<float> * to, const Packet2cf& from) { pstoreu((float*)to, from.v); } +EIGEN_STRONG_INLINE Packet2cf pload2(const std::complex<float>* from0, const std::complex<float>* from1) +{ + Packet4f res0, res1; +#ifdef __VSX__ + __asm__ ("lxsdx %x0,%y1" : "=wa" (res0) : "Z" (*from0)); + __asm__ ("lxsdx %x0,%y1" : "=wa" (res1) : "Z" (*from1)); +#ifdef _BIG_ENDIAN + __asm__ ("xxpermdi %x0, %x1, %x2, 0" : "=wa" (res0) : "wa" (res0), "wa" (res1)); +#else + __asm__ ("xxpermdi %x0, %x2, %x1, 0" : "=wa" (res0) : "wa" (res0), "wa" (res1)); +#endif +#else + *reinterpret_cast<std::complex<float> *>(&res0) = *from0; + *reinterpret_cast<std::complex<float> *>(&res1) = *from1; + res0 = vec_perm(res0, res1, p16uc_TRANSPOSE64_HI); +#endif + return Packet2cf(res0); +} + template<> EIGEN_DEVICE_FUNC inline Packet2cf pgather<std::complex<float>, Packet2cf>(const std::complex<float>* from, Index stride) { - std::complex<float> EIGEN_ALIGN16 af[2]; + EIGEN_ALIGN16 std::complex<float> af[2]; af[0] = from[0*stride]; af[1] = from[1*stride]; return pload<Packet2cf>(af); } template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<float>, Packet2cf>(std::complex<float>* to, const Packet2cf& from, Index stride) { - std::complex<float> EIGEN_ALIGN16 af[2]; + EIGEN_ALIGN16 std::complex<float> af[2]; pstore<std::complex<float> >((std::complex<float> *) af, from); to[0*stride] = af[0]; to[1*stride] = af[1]; @@ -100,25 +166,6 @@ template<> EIGEN_STRONG_INLINE Packet2cf psub<Packet2cf>(const Packet2cf& a, con 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(pxor<Packet4f>(a.v, reinterpret_cast<Packet4f>(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_PSET32_WODD); - // Get the imaginary parts of a - 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 = reinterpret_cast<Packet4f>(pxor(v2, reinterpret_cast<Packet4f>(p4ui_CONJ_XOR))); - // permute back to a proper order - v2 = vec_perm(v2, v2, p16uc_COMPLEX32_REV); - - return Packet2cf(padd<Packet4f>(v1, v2)); -} - template<> EIGEN_STRONG_INLINE Packet2cf pand <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(pand<Packet4f>(a.v, b.v)); } template<> EIGEN_STRONG_INLINE Packet2cf por <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(por<Packet4f>(a.v, b.v)); } template<> EIGEN_STRONG_INLINE Packet2cf pxor <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(pxor<Packet4f>(a.v, b.v)); } @@ -128,7 +175,7 @@ template<> EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::co template<> EIGEN_STRONG_INLINE std::complex<float> pfirst<Packet2cf>(const Packet2cf& a) { - std::complex<float> EIGEN_ALIGN16 res[2]; + EIGEN_ALIGN16 std::complex<float> res[2]; pstore((float *)&res, a.v); return res[0]; @@ -149,22 +196,6 @@ template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(const Packe return pfirst<Packet2cf>(Packet2cf(b)); } -template<> EIGEN_STRONG_INLINE Packet2cf preduxp<Packet2cf>(const Packet2cf* vecs) -{ - Packet4f b1, b2; -#ifdef _BIG_ENDIAN - b1 = vec_sld(vecs[0].v, vecs[1].v, 8); - b2 = vec_sld(vecs[1].v, vecs[0].v, 8); -#else - b1 = vec_sld(vecs[1].v, vecs[0].v, 8); - b2 = vec_sld(vecs[0].v, vecs[1].v, 8); -#endif - b2 = vec_sld(b2, b2, 8); - b2 = padd<Packet4f>(b1, b2); - - return Packet2cf(b2); -} - template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(const Packet2cf& a) { Packet4f b; @@ -175,77 +206,12 @@ template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(const P return pfirst<Packet2cf>(prod); } -template<int Offset> -struct palign_impl<Offset,Packet2cf> -{ - static EIGEN_STRONG_INLINE void run(Packet2cf& first, const Packet2cf& second) - { - 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 - } - } -}; - -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<> struct conj_helper<Packet4f, Packet2cf, false,false> -{ - EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet4f& x, const Packet2cf& y, const Packet2cf& c) const - { return padd(c, pmul(x,y)); } - - EIGEN_STRONG_INLINE Packet2cf pmul(const Packet4f& x, const Packet2cf& y) const - { return Packet2cf(internal::pmul<Packet4f>(x, y.v)); } -}; - -template<> struct conj_helper<Packet2cf, Packet4f, false,false> -{ - EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet4f& y, const Packet2cf& c) const - { return padd(c, pmul(x,y)); } - - EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& x, const Packet4f& y) const - { return Packet2cf(internal::pmul<Packet4f>(x.v, y)); } -}; +EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet2cf,Packet4f) template<> EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { // TODO optimize it for AltiVec - Packet2cf res = conj_helper<Packet2cf,Packet2cf,false,true>().pmul(a, b); + Packet2cf res = pmul(a, pconj(b)); Packet4f s = pmul<Packet4f>(b.v, b.v); return Packet2cf(pdiv(res.v, padd<Packet4f>(s, vec_perm(s, s, p16uc_COMPLEX32_REV)))); } @@ -262,6 +228,11 @@ EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet2cf,2>& kernel) kernel.packet[0].v = tmp; } +template<> EIGEN_STRONG_INLINE Packet2cf pcmp_eq(const Packet2cf& a, const Packet2cf& b) { + Packet4f eq = reinterpret_cast<Packet4f>(vec_cmpeq(a.v,b.v)); + return Packet2cf(vec_and(eq, vec_perm(eq, eq, p16uc_COMPLEX32_REV))); +} + #ifdef __VSX__ template<> EIGEN_STRONG_INLINE Packet2cf pblend(const Selector<2>& ifPacket, const Packet2cf& thenPacket, const Packet2cf& elsePacket) { Packet2cf result; @@ -270,12 +241,62 @@ template<> EIGEN_STRONG_INLINE Packet2cf pblend(const Selector<2>& ifPacket, con } #endif +template<> EIGEN_STRONG_INLINE Packet2cf psqrt<Packet2cf>(const Packet2cf& a) +{ + return psqrt_complex<Packet2cf>(a); +} + //---------- double ---------- #ifdef __VSX__ struct Packet1cd { EIGEN_STRONG_INLINE Packet1cd() {} EIGEN_STRONG_INLINE explicit Packet1cd(const Packet2d& a) : v(a) {} + + EIGEN_STRONG_INLINE Packet1cd pmul(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 = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4ui>(v2), reinterpret_cast<Packet4ui>(v2), 8)); + v2 = pxor(v2, reinterpret_cast<Packet2d>(p2ul_CONJ_XOR1)); + + return Packet1cd(padd<Packet2d>(v1, v2)); + } + + EIGEN_STRONG_INLINE Packet1cd& operator*=(const Packet1cd& b) { + v = pmul(Packet1cd(*this), b).v; + return *this; + } + EIGEN_STRONG_INLINE Packet1cd operator*(const Packet1cd& b) const { + return Packet1cd(*this) *= b; + } + + EIGEN_STRONG_INLINE Packet1cd& operator+=(const Packet1cd& b) { + v = padd(v, b.v); + return *this; + } + EIGEN_STRONG_INLINE Packet1cd operator+(const Packet1cd& b) const { + return Packet1cd(*this) += b; + } + EIGEN_STRONG_INLINE Packet1cd& operator-=(const Packet1cd& b) { + v = psub(v, b.v); + return *this; + } + EIGEN_STRONG_INLINE Packet1cd operator-(const Packet1cd& b) const { + return Packet1cd(*this) -= b; + } + EIGEN_STRONG_INLINE Packet1cd operator-(void) const { + return Packet1cd(-v); + } + Packet2d v; }; @@ -283,6 +304,7 @@ template<> struct packet_traits<std::complex<double> > : default_packet_traits { typedef Packet1cd type; typedef Packet1cd half; + typedef Packet2d as_real; enum { Vectorizable = 1, AlignedOnScalar = 0, @@ -302,7 +324,7 @@ template<> struct packet_traits<std::complex<double> > : default_packet_traits }; }; -template<> struct unpacket_traits<Packet1cd> { typedef std::complex<double> type; enum {size=1, alignment=Aligned16}; typedef Packet1cd half; }; +template<> struct unpacket_traits<Packet1cd> { typedef std::complex<double> type; enum {size=1, alignment=Aligned16, vectorizable=true, masked_load_available=false, masked_store_available=false}; typedef Packet1cd half; typedef Packet2d as_real; }; template<> EIGEN_STRONG_INLINE Packet1cd pload <Packet1cd>(const std::complex<double>* from) { return Packet1cd(pload<Packet2d>((const double*)from)); } template<> EIGEN_STRONG_INLINE Packet1cd ploadu<Packet1cd>(const std::complex<double>* from) { return Packet1cd(ploadu<Packet2d>((const double*)from)); } @@ -312,19 +334,13 @@ template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<double> >(std::complex< 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, Index stride) +template<> EIGEN_DEVICE_FUNC inline Packet1cd pgather<std::complex<double>, Packet1cd>(const std::complex<double>* from, Index) { - std::complex<double> EIGEN_ALIGN16 af[2]; - af[0] = from[0*stride]; - af[1] = from[1*stride]; - return pload<Packet1cd>(af); + return pload<Packet1cd>(from); } -template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<double>, Packet1cd>(std::complex<double>* to, const Packet1cd& from, Index stride) +template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<double>, Packet1cd>(std::complex<double>* to, const Packet1cd& from, Index) { - std::complex<double> EIGEN_ALIGN16 af[2]; - pstore<std::complex<double> >(af, from); - to[0*stride] = af[0]; - to[1*stride] = af[1]; + pstore<std::complex<double> >(to, from); } template<> EIGEN_STRONG_INLINE Packet1cd padd<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(a.v + b.v); } @@ -332,24 +348,6 @@ template<> EIGEN_STRONG_INLINE Packet1cd psub<Packet1cd>(const Packet1cd& a, con 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(pxor(a.v, reinterpret_cast<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 = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4ui>(v2), reinterpret_cast<Packet4ui>(v2), 8)); - v2 = pxor(v2, reinterpret_cast<Packet2d>(p2ul_CONJ_XOR1)); - - return Packet1cd(padd<Packet2d>(v1, v2)); -} - template<> EIGEN_STRONG_INLINE Packet1cd pand <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(pand(a.v,b.v)); } template<> EIGEN_STRONG_INLINE Packet1cd por <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(por(a.v,b.v)); } template<> EIGEN_STRONG_INLINE Packet1cd pxor <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(pxor(a.v,b.v)); } @@ -361,7 +359,7 @@ template<> EIGEN_STRONG_INLINE void prefetch<std::complex<double> >(const std::c template<> EIGEN_STRONG_INLINE std::complex<double> pfirst<Packet1cd>(const Packet1cd& a) { - std::complex<double> EIGEN_ALIGN16 res[2]; + EIGEN_ALIGN16 std::complex<double> res[2]; pstore<std::complex<double> >(res, a); return res[0]; @@ -370,74 +368,15 @@ template<> EIGEN_STRONG_INLINE std::complex<double> pfirst<Packet1cd>(const Pac template<> EIGEN_STRONG_INLINE Packet1cd preverse(const Packet1cd& a) { return a; } template<> EIGEN_STRONG_INLINE 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<> struct conj_helper<Packet2d, Packet1cd, false,false> -{ - EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet2d& x, const Packet1cd& y, const Packet1cd& c) const - { return padd(c, pmul(x,y)); } - - EIGEN_STRONG_INLINE Packet1cd pmul(const Packet2d& x, const Packet1cd& y) const - { return Packet1cd(internal::pmul<Packet2d>(x, y.v)); } -}; - -template<> struct conj_helper<Packet1cd, Packet2d, false,false> -{ - EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet2d& y, const Packet1cd& c) const - { return padd(c, pmul(x,y)); } - - EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& x, const Packet2d& y) const - { return Packet1cd(internal::pmul<Packet2d>(x.v, y)); } -}; +EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet1cd,Packet2d) 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); + Packet1cd res = pmul(a,pconj(b)); Packet2d s = pmul<Packet2d>(b.v, b.v); return Packet1cd(pdiv(res.v, padd<Packet2d>(s, vec_perm(s, s, p16uc_REVERSE64)))); } @@ -453,6 +392,23 @@ EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet1cd,2>& kernel) kernel.packet[1].v = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_LO); kernel.packet[0].v = tmp; } + +template<> EIGEN_STRONG_INLINE Packet1cd pcmp_eq(const Packet1cd& a, const Packet1cd& b) { + // Compare real and imaginary parts of a and b to get the mask vector: + // [re(a)==re(b), im(a)==im(b)] + Packet2d eq = reinterpret_cast<Packet2d>(vec_cmpeq(a.v,b.v)); + // Swap real/imag elements in the mask in to get: + // [im(a)==im(b), re(a)==re(b)] + Packet2d eq_swapped = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4ui>(eq), reinterpret_cast<Packet4ui>(eq), 8)); + // Return re(a)==re(b) & im(a)==im(b) by computing bitwise AND of eq and eq_swapped + return Packet1cd(vec_and(eq, eq_swapped)); +} + +template<> EIGEN_STRONG_INLINE Packet1cd psqrt<Packet1cd>(const Packet1cd& a) +{ + return psqrt_complex<Packet1cd>(a); +} + #endif // __VSX__ } // end namespace internal |