Home | History | Annotate | Download | only in products
      1 // This file is part of Eigen, a lightweight C++ template library
      2 // for linear algebra.
      3 //
      4 // Copyright (C) 2008-2009 Gael Guennebaud <gael.guennebaud (at) inria.fr>
      5 //
      6 // This Source Code Form is subject to the terms of the Mozilla
      7 // Public License v. 2.0. If a copy of the MPL was not distributed
      8 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
      9 
     10 #ifndef EIGEN_SELFADJOINT_MATRIX_VECTOR_H
     11 #define EIGEN_SELFADJOINT_MATRIX_VECTOR_H
     12 
     13 namespace Eigen {
     14 
     15 namespace internal {
     16 
     17 /* Optimized selfadjoint matrix * vector product:
     18  * This algorithm processes 2 columns at onces that allows to both reduce
     19  * the number of load/stores of the result by a factor 2 and to reduce
     20  * the instruction dependency.
     21  */
     22 
     23 template<typename Scalar, typename Index, int StorageOrder, int UpLo, bool ConjugateLhs, bool ConjugateRhs, int Version=Specialized>
     24 struct selfadjoint_matrix_vector_product;
     25 
     26 template<typename Scalar, typename Index, int StorageOrder, int UpLo, bool ConjugateLhs, bool ConjugateRhs, int Version>
     27 struct selfadjoint_matrix_vector_product
     28 
     29 {
     30 static EIGEN_DONT_INLINE void run(
     31   Index size,
     32   const Scalar*  lhs, Index lhsStride,
     33   const Scalar* _rhs, Index rhsIncr,
     34   Scalar* res,
     35   Scalar alpha)
     36 {
     37   typedef typename packet_traits<Scalar>::type Packet;
     38   typedef typename NumTraits<Scalar>::Real RealScalar;
     39   const Index PacketSize = sizeof(Packet)/sizeof(Scalar);
     40 
     41   enum {
     42     IsRowMajor = StorageOrder==RowMajor ? 1 : 0,
     43     IsLower = UpLo == Lower ? 1 : 0,
     44     FirstTriangular = IsRowMajor == IsLower
     45   };
     46 
     47   conj_helper<Scalar,Scalar,NumTraits<Scalar>::IsComplex && EIGEN_LOGICAL_XOR(ConjugateLhs,  IsRowMajor), ConjugateRhs> cj0;
     48   conj_helper<Scalar,Scalar,NumTraits<Scalar>::IsComplex && EIGEN_LOGICAL_XOR(ConjugateLhs, !IsRowMajor), ConjugateRhs> cj1;
     49   conj_helper<Scalar,Scalar,NumTraits<Scalar>::IsComplex, ConjugateRhs> cjd;
     50 
     51   conj_helper<Packet,Packet,NumTraits<Scalar>::IsComplex && EIGEN_LOGICAL_XOR(ConjugateLhs,  IsRowMajor), ConjugateRhs> pcj0;
     52   conj_helper<Packet,Packet,NumTraits<Scalar>::IsComplex && EIGEN_LOGICAL_XOR(ConjugateLhs, !IsRowMajor), ConjugateRhs> pcj1;
     53 
     54   Scalar cjAlpha = ConjugateRhs ? conj(alpha) : alpha;
     55 
     56   // FIXME this copy is now handled outside product_selfadjoint_vector, so it could probably be removed.
     57   // if the rhs is not sequentially stored in memory we copy it to a temporary buffer,
     58   // this is because we need to extract packets
     59   ei_declare_aligned_stack_constructed_variable(Scalar,rhs,size,rhsIncr==1 ? const_cast<Scalar*>(_rhs) : 0);
     60   if (rhsIncr!=1)
     61   {
     62     const Scalar* it = _rhs;
     63     for (Index i=0; i<size; ++i, it+=rhsIncr)
     64       rhs[i] = *it;
     65   }
     66 
     67   Index bound = (std::max)(Index(0),size-8) & 0xfffffffe;
     68   if (FirstTriangular)
     69     bound = size - bound;
     70 
     71   for (Index j=FirstTriangular ? bound : 0;
     72        j<(FirstTriangular ? size : bound);j+=2)
     73   {
     74     register const Scalar* EIGEN_RESTRICT A0 = lhs + j*lhsStride;
     75     register const Scalar* EIGEN_RESTRICT A1 = lhs + (j+1)*lhsStride;
     76 
     77     Scalar t0 = cjAlpha * rhs[j];
     78     Packet ptmp0 = pset1<Packet>(t0);
     79     Scalar t1 = cjAlpha * rhs[j+1];
     80     Packet ptmp1 = pset1<Packet>(t1);
     81 
     82     Scalar t2(0);
     83     Packet ptmp2 = pset1<Packet>(t2);
     84     Scalar t3(0);
     85     Packet ptmp3 = pset1<Packet>(t3);
     86 
     87     size_t starti = FirstTriangular ? 0 : j+2;
     88     size_t endi   = FirstTriangular ? j : size;
     89     size_t alignedStart = (starti) + internal::first_aligned(&res[starti], endi-starti);
     90     size_t alignedEnd = alignedStart + ((endi-alignedStart)/(PacketSize))*(PacketSize);
     91 
     92     // TODO make sure this product is a real * complex and that the rhs is properly conjugated if needed
     93     res[j]   += cjd.pmul(internal::real(A0[j]), t0);
     94     res[j+1] += cjd.pmul(internal::real(A1[j+1]), t1);
     95     if(FirstTriangular)
     96     {
     97       res[j]   += cj0.pmul(A1[j],   t1);
     98       t3       += cj1.pmul(A1[j],   rhs[j]);
     99     }
    100     else
    101     {
    102       res[j+1] += cj0.pmul(A0[j+1],t0);
    103       t2 += cj1.pmul(A0[j+1], rhs[j+1]);
    104     }
    105 
    106     for (size_t i=starti; i<alignedStart; ++i)
    107     {
    108       res[i] += t0 * A0[i] + t1 * A1[i];
    109       t2 += conj(A0[i]) * rhs[i];
    110       t3 += conj(A1[i]) * rhs[i];
    111     }
    112     // Yes this an optimization for gcc 4.3 and 4.4 (=> huge speed up)
    113     // gcc 4.2 does this optimization automatically.
    114     const Scalar* EIGEN_RESTRICT a0It  = A0  + alignedStart;
    115     const Scalar* EIGEN_RESTRICT a1It  = A1  + alignedStart;
    116     const Scalar* EIGEN_RESTRICT rhsIt = rhs + alignedStart;
    117           Scalar* EIGEN_RESTRICT resIt = res + alignedStart;
    118     for (size_t i=alignedStart; i<alignedEnd; i+=PacketSize)
    119     {
    120       Packet A0i = ploadu<Packet>(a0It);  a0It  += PacketSize;
    121       Packet A1i = ploadu<Packet>(a1It);  a1It  += PacketSize;
    122       Packet Bi  = ploadu<Packet>(rhsIt); rhsIt += PacketSize; // FIXME should be aligned in most cases
    123       Packet Xi  = pload <Packet>(resIt);
    124 
    125       Xi    = pcj0.pmadd(A0i,ptmp0, pcj0.pmadd(A1i,ptmp1,Xi));
    126       ptmp2 = pcj1.pmadd(A0i,  Bi, ptmp2);
    127       ptmp3 = pcj1.pmadd(A1i,  Bi, ptmp3);
    128       pstore(resIt,Xi); resIt += PacketSize;
    129     }
    130     for (size_t i=alignedEnd; i<endi; i++)
    131     {
    132       res[i] += cj0.pmul(A0[i], t0) + cj0.pmul(A1[i],t1);
    133       t2 += cj1.pmul(A0[i], rhs[i]);
    134       t3 += cj1.pmul(A1[i], rhs[i]);
    135     }
    136 
    137     res[j]   += alpha * (t2 + predux(ptmp2));
    138     res[j+1] += alpha * (t3 + predux(ptmp3));
    139   }
    140   for (Index j=FirstTriangular ? 0 : bound;j<(FirstTriangular ? bound : size);j++)
    141   {
    142     register const Scalar* EIGEN_RESTRICT A0 = lhs + j*lhsStride;
    143 
    144     Scalar t1 = cjAlpha * rhs[j];
    145     Scalar t2(0);
    146     // TODO make sure this product is a real * complex and that the rhs is properly conjugated if needed
    147     res[j] += cjd.pmul(internal::real(A0[j]), t1);
    148     for (Index i=FirstTriangular ? 0 : j+1; i<(FirstTriangular ? j : size); i++)
    149     {
    150       res[i] += cj0.pmul(A0[i], t1);
    151       t2 += cj1.pmul(A0[i], rhs[i]);
    152     }
    153     res[j] += alpha * t2;
    154   }
    155 }
    156 };
    157 
    158 } // end namespace internal
    159 
    160 /***************************************************************************
    161 * Wrapper to product_selfadjoint_vector
    162 ***************************************************************************/
    163 
    164 namespace internal {
    165 template<typename Lhs, int LhsMode, typename Rhs>
    166 struct traits<SelfadjointProductMatrix<Lhs,LhsMode,false,Rhs,0,true> >
    167   : traits<ProductBase<SelfadjointProductMatrix<Lhs,LhsMode,false,Rhs,0,true>, Lhs, Rhs> >
    168 {};
    169 }
    170 
    171 template<typename Lhs, int LhsMode, typename Rhs>
    172 struct SelfadjointProductMatrix<Lhs,LhsMode,false,Rhs,0,true>
    173   : public ProductBase<SelfadjointProductMatrix<Lhs,LhsMode,false,Rhs,0,true>, Lhs, Rhs >
    174 {
    175   EIGEN_PRODUCT_PUBLIC_INTERFACE(SelfadjointProductMatrix)
    176 
    177   enum {
    178     LhsUpLo = LhsMode&(Upper|Lower)
    179   };
    180 
    181   SelfadjointProductMatrix(const Lhs& lhs, const Rhs& rhs) : Base(lhs,rhs) {}
    182 
    183   template<typename Dest> void scaleAndAddTo(Dest& dest, Scalar alpha) const
    184   {
    185     typedef typename Dest::Scalar ResScalar;
    186     typedef typename Base::RhsScalar RhsScalar;
    187     typedef Map<Matrix<ResScalar,Dynamic,1>, Aligned> MappedDest;
    188 
    189     eigen_assert(dest.rows()==m_lhs.rows() && dest.cols()==m_rhs.cols());
    190 
    191     typename internal::add_const_on_value_type<ActualLhsType>::type lhs = LhsBlasTraits::extract(m_lhs);
    192     typename internal::add_const_on_value_type<ActualRhsType>::type rhs = RhsBlasTraits::extract(m_rhs);
    193 
    194     Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(m_lhs)
    195                                * RhsBlasTraits::extractScalarFactor(m_rhs);
    196 
    197     enum {
    198       EvalToDest = (Dest::InnerStrideAtCompileTime==1),
    199       UseRhs = (_ActualRhsType::InnerStrideAtCompileTime==1)
    200     };
    201 
    202     internal::gemv_static_vector_if<ResScalar,Dest::SizeAtCompileTime,Dest::MaxSizeAtCompileTime,!EvalToDest> static_dest;
    203     internal::gemv_static_vector_if<RhsScalar,_ActualRhsType::SizeAtCompileTime,_ActualRhsType::MaxSizeAtCompileTime,!UseRhs> static_rhs;
    204 
    205     ei_declare_aligned_stack_constructed_variable(ResScalar,actualDestPtr,dest.size(),
    206                                                   EvalToDest ? dest.data() : static_dest.data());
    207 
    208     ei_declare_aligned_stack_constructed_variable(RhsScalar,actualRhsPtr,rhs.size(),
    209         UseRhs ? const_cast<RhsScalar*>(rhs.data()) : static_rhs.data());
    210 
    211     if(!EvalToDest)
    212     {
    213       #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN
    214       int size = dest.size();
    215       EIGEN_DENSE_STORAGE_CTOR_PLUGIN
    216       #endif
    217       MappedDest(actualDestPtr, dest.size()) = dest;
    218     }
    219 
    220     if(!UseRhs)
    221     {
    222       #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN
    223       int size = rhs.size();
    224       EIGEN_DENSE_STORAGE_CTOR_PLUGIN
    225       #endif
    226       Map<typename _ActualRhsType::PlainObject>(actualRhsPtr, rhs.size()) = rhs;
    227     }
    228 
    229 
    230     internal::selfadjoint_matrix_vector_product<Scalar, Index, (internal::traits<_ActualLhsType>::Flags&RowMajorBit) ? RowMajor : ColMajor, int(LhsUpLo), bool(LhsBlasTraits::NeedToConjugate), bool(RhsBlasTraits::NeedToConjugate)>::run
    231       (
    232         lhs.rows(),                             // size
    233         &lhs.coeffRef(0,0),  lhs.outerStride(), // lhs info
    234         actualRhsPtr, 1,                        // rhs info
    235         actualDestPtr,                          // result info
    236         actualAlpha                             // scale factor
    237       );
    238 
    239     if(!EvalToDest)
    240       dest = MappedDest(actualDestPtr, dest.size());
    241   }
    242 };
    243 
    244 namespace internal {
    245 template<typename Lhs, typename Rhs, int RhsMode>
    246 struct traits<SelfadjointProductMatrix<Lhs,0,true,Rhs,RhsMode,false> >
    247   : traits<ProductBase<SelfadjointProductMatrix<Lhs,0,true,Rhs,RhsMode,false>, Lhs, Rhs> >
    248 {};
    249 }
    250 
    251 template<typename Lhs, typename Rhs, int RhsMode>
    252 struct SelfadjointProductMatrix<Lhs,0,true,Rhs,RhsMode,false>
    253   : public ProductBase<SelfadjointProductMatrix<Lhs,0,true,Rhs,RhsMode,false>, Lhs, Rhs >
    254 {
    255   EIGEN_PRODUCT_PUBLIC_INTERFACE(SelfadjointProductMatrix)
    256 
    257   enum {
    258     RhsUpLo = RhsMode&(Upper|Lower)
    259   };
    260 
    261   SelfadjointProductMatrix(const Lhs& lhs, const Rhs& rhs) : Base(lhs,rhs) {}
    262 
    263   template<typename Dest> void scaleAndAddTo(Dest& dest, Scalar alpha) const
    264   {
    265     // let's simply transpose the product
    266     Transpose<Dest> destT(dest);
    267     SelfadjointProductMatrix<Transpose<const Rhs>, int(RhsUpLo)==Upper ? Lower : Upper, false,
    268                              Transpose<const Lhs>, 0, true>(m_rhs.transpose(), m_lhs.transpose()).scaleAndAddTo(destT, alpha);
    269   }
    270 };
    271 
    272 } // end namespace Eigen
    273 
    274 #endif // EIGEN_SELFADJOINT_MATRIX_VECTOR_H
    275