synch with main branch

Eigen/src/Cholesky/LDLT.h

@@ -62,7 +62,7 @@ template<typename MatrixType> class LDLT
     typedef Matrix<int, MatrixType::RowsAtCompileTime, 1> IntColVectorType;
     typedef Matrix<int, 1, MatrixType::RowsAtCompileTime> IntRowVectorType;
 
-    /** 
+    /**
     * \brief Default Constructor.
     *
     * The default constructor is useful in cases in which the user intends to
@@ -83,7 +83,7 @@ template<typename MatrixType> class LDLT
     inline TriangularView<MatrixType, UnitLowerTriangular> matrixL(void) const
     { 
       ei_assert(m_isInitialized && "LDLT is not initialized.");
-      return m_matrix; 
+      return m_matrix;
     }
 
     /** \returns a vector of integers, whose size is the number of rows of the matrix being decomposed,
@@ -97,24 +97,24 @@ template<typename MatrixType> class LDLT
     }
 
     /** \returns the coefficients of the diagonal matrix D */
-    inline Diagonal<MatrixType,0> vectorD(void) const 
-    { 
+    inline Diagonal<MatrixType,0> vectorD(void) const
+    {
       ei_assert(m_isInitialized && "LDLT is not initialized.");
-      return m_matrix.diagonal(); 
+      return m_matrix.diagonal();
     }
 
     /** \returns true if the matrix is positive (semidefinite) */
-    inline bool isPositive(void) const 
-    { 
+    inline bool isPositive(void) const
+    {
       ei_assert(m_isInitialized && "LDLT is not initialized.");
-      return m_sign == 1; 
+      return m_sign == 1;
     }
 
     /** \returns true if the matrix is negative (semidefinite) */
-    inline bool isNegative(void) const 
-    { 
+    inline bool isNegative(void) const
+    {
       ei_assert(m_isInitialized && "LDLT is not initialized.");
-      return m_sign == -1; 
+      return m_sign == -1;
     }
 
     template<typename RhsDerived, typename ResultType>

Eigen/src/Core/Block.h

@@ -33,8 +33,8 @@
   * \param MatrixType the type of the object in which we are taking a block
   * \param BlockRows the number of rows of the block we are taking at compile time (optional)
   * \param BlockCols the number of columns of the block we are taking at compile time (optional)
-  * \param _PacketAccess allows to enforce aligned loads and stores if set to ForceAligned.
-  *                      The default is AsRequested. This parameter is internaly used by Eigen
+  * \param _PacketAccess allows to enforce aligned loads and stores if set to \b ForceAligned.
+  *                      The default is \b AsRequested. This parameter is internaly used by Eigen
   *                      in expressions such as \code mat.block() += other; \endcode and most of
   *                      the time this is the only way it is used.
   * \param _DirectAccessStatus \internal used for partial specialization

Eigen/src/Core/Dot.h

@@ -292,154 +292,6 @@ inline typename NumTraits<typename ei_traits<Derived>::Scalar>::Real MatrixBase<
   return ei_sqrt(squaredNorm());
 }
 
-/** \returns the \em l2 norm of \c *this using a numerically more stable
-  * algorithm.
-  *
-  * \sa norm(), dot(), squaredNorm(), blueNorm()
-  */
-template<typename Derived>
-inline typename NumTraits<typename ei_traits<Derived>::Scalar>::Real
-MatrixBase<Derived>::stableNorm() const
-{
-  return this->cwise().abs().redux(ei_scalar_hypot_op<RealScalar>());
-}
-
-/** \internal Computes ibeta^iexp by binary expansion of iexp,
-  * exact if ibeta is the machine base */
-template<typename T> inline T bexp(int ibeta, int iexp)
-{
-  T tbeta = T(ibeta);
-  T res = 1.0;
-  int n = iexp;
-  if (n<0)
-  {
-    n = - n;
-    tbeta = 1.0/tbeta;
-  }
-  for(;;)
-  {
-    if ((n % 2)==0)
-      res = res * tbeta;
-    n = n/2;
-    if (n==0) return res;
-    tbeta = tbeta*tbeta;
-  }
-  return res;
-}
-
-/** \returns the \em l2 norm of \c *this using the Blue's algorithm.
-  * A Portable Fortran Program to Find the Euclidean Norm of a Vector,
-  * ACM TOMS, Vol 4, Issue 1, 1978.
-  *
-  * \sa norm(), dot(), squaredNorm(), stableNorm()
-  */
-template<typename Derived>
-inline typename NumTraits<typename ei_traits<Derived>::Scalar>::Real
-MatrixBase<Derived>::blueNorm() const
-{
-  static int nmax;
-  static Scalar b1, b2, s1m, s2m, overfl, rbig, relerr;
-  int n;
-  Scalar ax, abig, amed, asml;
-
-  if(nmax <= 0)
-  {
-    int nbig, ibeta, it, iemin, iemax, iexp;
-    Scalar abig, eps;
-    // This program calculates the machine-dependent constants
-    // bl, b2, slm, s2m, relerr overfl, nmax
-    // from the "basic" machine-dependent numbers
-    // nbig, ibeta, it, iemin, iemax, rbig.
-    // The following define the basic machine-dependent constants.
-    // For portability, the PORT subprograms "ilmaeh" and "rlmach"
-    // are used. For any specific computer, each of the assignment
-    // statements can be replaced
-    nbig  = std::numeric_limits<int>::max();            // largest integer
-    ibeta = NumTraits<Scalar>::Base;                    // base for floating-point numbers
-    it    = NumTraits<Scalar>::Mantissa;                // number of base-beta digits in mantissa
-    iemin = std::numeric_limits<Scalar>::min_exponent;  // minimum exponent
-    iemax = std::numeric_limits<Scalar>::max_exponent;  // maximum exponent
-    rbig  = std::numeric_limits<Scalar>::max();         // largest floating-point number
-
-    // Check the basic machine-dependent constants.
-    if(iemin > 1 - 2*it || 1+it>iemax || (it==2 && ibeta<5)
-      || (it<=4 && ibeta <= 3 ) || it<2)
-    {
-      ei_assert(false && "the algorithm cannot be guaranteed on this computer");
-    }
-    iexp  = -((1-iemin)/2);
-    b1    = bexp<Scalar>(ibeta, iexp);  // lower boundary of midrange
-    iexp  = (iemax + 1 - it)/2;
-    b2    = bexp<Scalar>(ibeta,iexp);   // upper boundary of midrange
-
-    iexp  = (2-iemin)/2;
-    s1m   = bexp<Scalar>(ibeta,iexp);   // scaling factor for lower range
-    iexp  = - ((iemax+it)/2);
-    s2m   = bexp<Scalar>(ibeta,iexp);   // scaling factor for upper range
-
-    overfl  = rbig*s2m;          // overfow boundary for abig
-    eps     = bexp<Scalar>(ibeta, 1-it);
-    relerr  = ei_sqrt(eps);      // tolerance for neglecting asml
-    abig    = 1.0/eps - 1.0;
-    if (Scalar(nbig)>abig)  nmax = abig;  // largest safe n
-    else                    nmax = nbig;
-  }
-  n = size();
-  if(n==0)
-    return 0;
-  asml = Scalar(0);
-  amed = Scalar(0);
-  abig = Scalar(0);
-  for(int j=0; j<n; ++j)
-  {
-    ax = ei_abs(coeff(j));
-    if(ax > b2)      abig += ei_abs2(ax*s2m);
-    else if(ax < b1) asml += ei_abs2(ax*s1m);
-    else             amed += ei_abs2(ax);
-  }
-  if(abig > Scalar(0))
-  {
-    abig = ei_sqrt(abig);
-    if(abig > overfl)
-    {
-      ei_assert(false && "overflow");
-      return rbig;
-    }
-    if(amed > Scalar(0))
-    {
-      abig = abig/s2m;
-      amed = ei_sqrt(amed);
-    }
-    else
-    {
-      return abig/s2m;
-    }
-
-  }
-  else if(asml > Scalar(0))
-  {
-    if (amed > Scalar(0))
-    {
-      abig = ei_sqrt(amed);
-      amed = ei_sqrt(asml) / s1m;
-    }
-    else
-    {
-      return ei_sqrt(asml)/s1m;
-    }
-  }
-  else
-  {
-    return ei_sqrt(amed);
-  }
-  asml = std::min(abig, amed);
-  abig = std::max(abig, amed);
-  if(asml <= abig*relerr)
-    return abig;
-  else
-    return abig * ei_sqrt(Scalar(1) + ei_abs2(asml/abig));
-}
-
 /** \returns an expression of the quotient of *this by its own norm.
   *
   * \only_for_vectors

Eigen/src/Core/Functors.h

@@ -124,9 +124,6 @@ template<typename Scalar> struct ei_scalar_hypot_op EIGEN_EMPTY_STRUCT {
 //   typedef typename NumTraits<Scalar>::Real result_type;
   EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& _x, const Scalar& _y) const
   {
-//     typedef typename NumTraits<T>::Real RealScalar;
-//     RealScalar _x = ei_abs(x);
-//     RealScalar _y = ei_abs(y);
     Scalar p = std::max(_x, _y);
     Scalar q = std::min(_x, _y);
     Scalar qp = q/p;

Eigen/src/Core/MapBase.h

@@ -173,8 +173,14 @@ template<typename Derived> class MapBase
 
     using Base::operator=;
     using Base::operator*=;
-    using Base::operator+=;
-    using Base::operator-=;
+
+    template<typename Lhs,typename Rhs>
+    Derived& operator+=(const Flagged<Product<Lhs,Rhs,CacheFriendlyProduct>, 0, EvalBeforeNestingBit | EvalBeforeAssigningBit>& other)
+    { return Base::operator+=(other); }
+
+    template<typename Lhs,typename Rhs>
+    Derived& operator-=(const Flagged<Product<Lhs,Rhs,CacheFriendlyProduct>, 0, EvalBeforeNestingBit | EvalBeforeAssigningBit>& other)
+    { return Base::operator-=(other); }
 
     template<typename OtherDerived>
     Derived& operator+=(const MatrixBase<OtherDerived>& other)

Eigen/src/Core/Matrix.h

@@ -571,10 +571,16 @@ class Matrix
     template<typename OtherDerived>
     EIGEN_STRONG_INLINE Matrix& _set(const MatrixBase<OtherDerived>& other)
     {
-      _resize_to_match(other);
-      return Base::operator=(other);
+      _set_selector(other.derived(), typename ei_meta_if<(int(OtherDerived::Flags) & EvalBeforeAssigningBit), ei_meta_true, ei_meta_false>::ret());
+      return *this;
     }
 
+    template<typename OtherDerived>
+    EIGEN_STRONG_INLINE void _set_selector(const OtherDerived& other, const ei_meta_true&) { _set_noalias(other.eval()); }
+
+    template<typename OtherDerived>
+    EIGEN_STRONG_INLINE void _set_selector(const OtherDerived& other, const ei_meta_false&) { _set_noalias(other); }
+
     /** \internal Like _set() but additionally makes the assumption that no aliasing effect can happen (which
       * is the case when creating a new matrix) so one can enforce lazy evaluation.
       *

Eigen/src/Core/MatrixBase.h

@@ -437,6 +437,7 @@ template<typename Derived> class MatrixBase
     RealScalar norm() const;
     RealScalar stableNorm() const;
     RealScalar blueNorm() const;
+    RealScalar hypotNorm() const;
     const PlainMatrixType normalized() const;
     void normalize();
 

Eigen/src/Core/NumTraits.h

@@ -70,9 +70,7 @@ template<> struct NumTraits<float>
     HasFloatingPoint = 1,
     ReadCost = 1,
     AddCost = 1,
-    MulCost = 1,
-    Base = 2,
-    Mantissa = 23
+    MulCost = 1
   };
 };
 
@@ -85,9 +83,7 @@ template<> struct NumTraits<double>
     HasFloatingPoint = 1,
     ReadCost = 1,
     AddCost = 1,
-    MulCost = 1,
-    Base = 2,
-    Mantissa = 52
+    MulCost = 1
   };
 };
 

Eigen/src/Core/Product.h

@@ -287,6 +287,18 @@ MatrixBase<Derived>::operator*(const MatrixBase<OtherDerived> &other) const
   return typename ProductReturnType<Derived,OtherDerived>::Type(derived(), other.derived());
 }
 
+/** replaces \c *this by \c *this * \a other.
+  *
+  * \returns a reference to \c *this
+  */
+template<typename Derived>
+template<typename OtherDerived>
+inline Derived &
+MatrixBase<Derived>::operator*=(const MatrixBase<OtherDerived> &other)
+{
+  return derived() = derived() * other.derived();
+}
+
 /***************************************************************************
 * Normal product .coeff() implementation (with meta-unrolling)
 ***************************************************************************/

Eigen/src/Core/StableNorm.h

@@ -0,0 +1,194 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Gael Guennebaud <g.gael@free.fr>
+//
+// Eigen is free software; you can redistribute it and/or
+// modify it under the terms of the GNU Lesser General Public
+// License as published by the Free Software Foundation; either
+// version 3 of the License, or (at your option) any later version.
+//
+// Alternatively, you can redistribute it and/or
+// modify it under the terms of the GNU General Public License as
+// published by the Free Software Foundation; either version 2 of
+// the License, or (at your option) any later version.
+//
+// Eigen is distributed in the hope that it will be useful, but WITHOUT ANY
+// WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+// FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public License or the
+// GNU General Public License for more details.
+//
+// You should have received a copy of the GNU Lesser General Public
+// License and a copy of the GNU General Public License along with
+// Eigen. If not, see <http://www.gnu.org/licenses/>.
+
+#ifndef EIGEN_STABLENORM_H
+#define EIGEN_STABLENORM_H
+
+template<typename ExpressionType, typename Scalar>
+inline void ei_stable_norm_kernel(const ExpressionType& bl, Scalar& ssq, Scalar& scale, Scalar& invScale)
+{
+  Scalar max = bl.cwise().abs().maxCoeff();
+  if (max>scale)
+  {
+    ssq = ssq * ei_abs2(scale/max);
+    scale = max;
+    invScale = Scalar(1)/scale;
+  }
+  // TODO if the max is much much smaller than the current scale,
+  // then we can neglect this sub vector
+  ssq += (bl*invScale).squaredNorm();
+}
+
+/** \returns the \em l2 norm of \c *this avoiding underflow and overflow.
+  * This version use a blockwise two passes algorithm:
+  *  1 - find the absolute largest coefficient \c s
+  *  2 - compute \f$ s \Vert \frac{*this}{s} \Vert \f$ in a standard way
+  *
+  * For architecture/scalar types supporting vectorization, this version
+  * is faster than blueNorm(). Otherwise the blueNorm() is much faster.
+  *
+  * \sa norm(), blueNorm(), hypotNorm()
+  */
+template<typename Derived>
+inline typename NumTraits<typename ei_traits<Derived>::Scalar>::Real
+MatrixBase<Derived>::stableNorm() const
+{
+  const int blockSize = 4096;
+  RealScalar scale = 0;
+  RealScalar invScale;
+  RealScalar ssq = 0; // sum of square
+  enum {
+    Alignment = (int(Flags)&DirectAccessBit) || (int(Flags)&AlignedBit) ? ForceAligned : AsRequested
+  };
+  int n = size();
+  int bi=0;
+  if ((int(Flags)&DirectAccessBit) && !(int(Flags)&AlignedBit))
+  {
+    bi = ei_alignmentOffset(&const_cast_derived().coeffRef(0), n);
+    if (bi>0)
+      ei_stable_norm_kernel(start(bi), ssq, scale, invScale);
+  }
+  for (; bi<n; bi+=blockSize)
+    ei_stable_norm_kernel(VectorBlock<Derived,Dynamic,Alignment>(derived(),bi,std::min(blockSize, n - bi)), ssq, scale, invScale);
+  return scale * ei_sqrt(ssq);
+}
+
+/** \returns the \em l2 norm of \c *this using the Blue's algorithm.
+  * A Portable Fortran Program to Find the Euclidean Norm of a Vector,
+  * ACM TOMS, Vol 4, Issue 1, 1978.
+  *
+  * For architecture/scalar types without vectorization, this version
+  * is much faster than stableNorm(). Otherwise the stableNorm() is faster.
+  *
+  * \sa norm(), stableNorm(), hypotNorm()
+  */
+template<typename Derived>
+inline typename NumTraits<typename ei_traits<Derived>::Scalar>::Real
+MatrixBase<Derived>::blueNorm() const
+{
+  static int nmax = -1;
+  static RealScalar b1, b2, s1m, s2m, overfl, rbig, relerr;
+  if(nmax <= 0)
+  {
+    int nbig, ibeta, it, iemin, iemax, iexp;
+    RealScalar abig, eps;
+    // This program calculates the machine-dependent constants
+    // bl, b2, slm, s2m, relerr overfl, nmax
+    // from the "basic" machine-dependent numbers
+    // nbig, ibeta, it, iemin, iemax, rbig.
+    // The following define the basic machine-dependent constants.
+    // For portability, the PORT subprograms "ilmaeh" and "rlmach"
+    // are used. For any specific computer, each of the assignment
+    // statements can be replaced
+    nbig  = std::numeric_limits<int>::max();            // largest integer
+    ibeta = std::numeric_limits<RealScalar>::radix;         // base for floating-point numbers
+    it    = std::numeric_limits<RealScalar>::digits;        // number of base-beta digits in mantissa
+    iemin = std::numeric_limits<RealScalar>::min_exponent;  // minimum exponent
+    iemax = std::numeric_limits<RealScalar>::max_exponent;  // maximum exponent
+    rbig  = std::numeric_limits<RealScalar>::max();         // largest floating-point number
+
+    // Check the basic machine-dependent constants.
+    if(iemin > 1 - 2*it || 1+it>iemax || (it==2 && ibeta<5)
+      || (it<=4 && ibeta <= 3 ) || it<2)
+    {
+      ei_assert(false && "the algorithm cannot be guaranteed on this computer");
+    }
+    iexp  = -((1-iemin)/2);
+    b1    = std::pow(double(ibeta),iexp);  // lower boundary of midrange
+    iexp  = (iemax + 1 - it)/2;
+    b2    = std::pow(double(ibeta),iexp);   // upper boundary of midrange
+
+    iexp  = (2-iemin)/2;
+    s1m   = std::pow(double(ibeta),iexp);   // scaling factor for lower range
+    iexp  = - ((iemax+it)/2);
+    s2m   = std::pow(double(ibeta),iexp);   // scaling factor for upper range
+
+    overfl  = rbig*s2m;             // overfow boundary for abig
+    eps     = std::pow(double(ibeta), 1-it);
+    relerr  = ei_sqrt(eps);         // tolerance for neglecting asml
+    abig    = 1.0/eps - 1.0;
+    if (RealScalar(nbig)>abig)  nmax = abig;  // largest safe n
+    else                        nmax = nbig;
+  }
+  int n = size();
+  RealScalar ab2 = b2 / RealScalar(n);
+  RealScalar asml = RealScalar(0);
+  RealScalar amed = RealScalar(0);
+  RealScalar abig = RealScalar(0);
+  for(int j=0; j<n; ++j)
+  {
+    RealScalar ax = ei_abs(coeff(j));
+    if(ax > ab2)     abig += ei_abs2(ax*s2m);
+    else if(ax < b1) asml += ei_abs2(ax*s1m);
+    else             amed += ei_abs2(ax);
+  }
+  if(abig > RealScalar(0))
+  {
+    abig = ei_sqrt(abig);
+    if(abig > overfl)
+    {
+      ei_assert(false && "overflow");
+      return rbig;
+    }
+    if(amed > RealScalar(0))
+    {
+      abig = abig/s2m;
+      amed = ei_sqrt(amed);
+    }
+    else
+      return abig/s2m;
+  }
+  else if(asml > RealScalar(0))
+  {
+    if (amed > RealScalar(0))
+    {
+      abig = ei_sqrt(amed);
+      amed = ei_sqrt(asml) / s1m;
+    }
+    else
+      return ei_sqrt(asml)/s1m;
+  }
+  else
+    return ei_sqrt(amed);
+  asml = std::min(abig, amed);
+  abig = std::max(abig, amed);
+  if(asml <= abig*relerr)
+    return abig;
+  else
+    return abig * ei_sqrt(RealScalar(1) + ei_abs2(asml/abig));
+}
+
+/** \returns the \em l2 norm of \c *this avoiding undeflow and overflow.
+  * This version use a concatenation of hypot() calls, and it is very slow.
+  *
+  * \sa norm(), stableNorm()
+  */
+template<typename Derived>
+inline typename NumTraits<typename ei_traits<Derived>::Scalar>::Real
+MatrixBase<Derived>::hypotNorm() const
+{
+  return this->cwise().abs().redux(ei_scalar_hypot_op<RealScalar>());
+}
+
+#endif // EIGEN_STABLENORM_H

Eigen/src/QR/SelfAdjointEigenSolver.h

@@ -383,18 +383,24 @@ static void ei_tridiagonal_qr_step(RealScalar* diag, RealScalar* subdiag, int st
       int kn1 = (k+1)*n;
       #endif
       // let's do the product manually to avoid the need of temporaries...
-      for (int i=0; i<n; ++i)
-      {
-        #ifdef EIGEN_DEFAULT_TO_ROW_MAJOR
-        Scalar matrixQ_i_k = matrixQ[i*n+k];
-        matrixQ[i*n+k]   = c * matrixQ_i_k - s * matrixQ[i*n+k+1];
-        matrixQ[i*n+k+1] = s * matrixQ_i_k + c * matrixQ[i*n+k+1];
-        #else
-        Scalar matrixQ_i_k = matrixQ[i+kn];
-        matrixQ[i+kn]  = c * matrixQ_i_k - s * matrixQ[i+kn1];
-        matrixQ[i+kn1] = s * matrixQ_i_k + c * matrixQ[i+kn1];
-        #endif
-      }
+      Matrix<Scalar,Dynamic,1> aux = Map<Matrix<Scalar,Dynamic,1>, ForceAligned >(matrixQ+kn,n);
+      Map<Matrix<Scalar,Dynamic,1>, ForceAligned >(matrixQ+kn,n)
+        = Map<Matrix<Scalar,Dynamic,1>, ForceAligned >(matrixQ+kn,n) * c - s * Map<Matrix<Scalar,Dynamic,1> >(matrixQ+kn1,n);
+
+      Map<Matrix<Scalar,Dynamic,1>, ForceAligned >(matrixQ+kn1,n)
+        = Map<Matrix<Scalar,Dynamic,1>, ForceAligned >(matrixQ+kn1,n) * c - s * aux;//Map<Matrix<Scalar,Dynamic,1> >(matrixQ+kn,n);
+//       for (int i=0; i<n; ++i)
+//       {
+//         #ifdef EIGEN_DEFAULT_TO_ROW_MAJOR
+//         Scalar matrixQ_i_k = matrixQ[i*n+k];
+//         matrixQ[i*n+k]   = c * matrixQ_i_k - s * matrixQ[i*n+k+1];
+//         matrixQ[i*n+k+1] = s * matrixQ_i_k + c * matrixQ[i*n+k+1];
+//         #else
+//         Scalar matrixQ_i_k = matrixQ[i+kn];
+//         matrixQ[i+kn]  = c * matrixQ_i_k - s * matrixQ[i+kn1];
+//         matrixQ[i+kn1] = s * matrixQ_i_k + c * matrixQ[i+kn1];
+//         #endif
+//       }
     }
   }
 }

Eigen/src/SVD/SVD.h

@@ -125,7 +125,7 @@ template<typename MatrixType> class SVD
     {
       return (b >= Scalar(0.0) ? ei_abs(a) : -ei_abs(a));
     }
-    
+
   protected:
     /** \internal */
     MatrixUType m_matU;
@@ -254,11 +254,14 @@ void SVD<MatrixType>::compute(const MatrixType& matrix)
     if (g != Scalar(0.0))
     {
       g = Scalar(1.0)/g;
-      for (j=l; j<n; j++)
+      if (m-l)
       {
-        s = A.col(i).end(m-l).dot(A.col(j).end(m-l));
-        f = (s/A(i,i))*g;
-        A.col(j).end(m-i) += f * A.col(i).end(m-i);
+        for (j=l; j<n; j++)
+        {
+          s = A.col(i).end(m-l).dot(A.col(j).end(m-l));
+          f = (s/A(i,i))*g;
+          A.col(j).end(m-i) += f * A.col(i).end(m-i);
+        }
       }
       A.col(i).end(m-i) *= g;
     }