ruby-eigen 0.0.9 → 0.0.10.pre1

data/ext/eigen/eigen3/Eigen/src/Core/SolveTriangular.h

@@ -0,0 +1,260 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SOLVETRIANGULAR_H
+#define EIGEN_SOLVETRIANGULAR_H
+
+namespace Eigen { 
+
+namespace internal {
+
+// Forward declarations:
+// The following two routines are implemented in the products/TriangularSolver*.h files
+template<typename LhsScalar, typename RhsScalar, typename Index, int Side, int Mode, bool Conjugate, int StorageOrder>
+struct triangular_solve_vector;
+
+template <typename Scalar, typename Index, int Side, int Mode, bool Conjugate, int TriStorageOrder, int OtherStorageOrder>
+struct triangular_solve_matrix;
+
+// small helper struct extracting some traits on the underlying solver operation
+template<typename Lhs, typename Rhs, int Side>
+class trsolve_traits
+{
+  private:
+    enum {
+      RhsIsVectorAtCompileTime = (Side==OnTheLeft ? Rhs::ColsAtCompileTime : Rhs::RowsAtCompileTime)==1
+    };
+  public:
+    enum {
+      Unrolling   = (RhsIsVectorAtCompileTime && Rhs::SizeAtCompileTime != Dynamic && Rhs::SizeAtCompileTime <= 8)
+                  ? CompleteUnrolling : NoUnrolling,
+      RhsVectors  = RhsIsVectorAtCompileTime ? 1 : Dynamic
+    };
+};
+
+template<typename Lhs, typename Rhs,
+  int Side, // can be OnTheLeft/OnTheRight
+  int Mode, // can be Upper/Lower | UnitDiag
+  int Unrolling = trsolve_traits<Lhs,Rhs,Side>::Unrolling,
+  int RhsVectors = trsolve_traits<Lhs,Rhs,Side>::RhsVectors
+  >
+struct triangular_solver_selector;
+
+template<typename Lhs, typename Rhs, int Side, int Mode>
+struct triangular_solver_selector<Lhs,Rhs,Side,Mode,NoUnrolling,1>
+{
+  typedef typename Lhs::Scalar LhsScalar;
+  typedef typename Rhs::Scalar RhsScalar;
+  typedef blas_traits<Lhs> LhsProductTraits;
+  typedef typename LhsProductTraits::ExtractType ActualLhsType;
+  typedef Map<Matrix<RhsScalar,Dynamic,1>, Aligned> MappedRhs;
+  static void run(const Lhs& lhs, Rhs& rhs)
+  {
+    ActualLhsType actualLhs = LhsProductTraits::extract(lhs);
+
+    // FIXME find a way to allow an inner stride if packet_traits<Scalar>::size==1
+
+    bool useRhsDirectly = Rhs::InnerStrideAtCompileTime==1 || rhs.innerStride()==1;
+
+    ei_declare_aligned_stack_constructed_variable(RhsScalar,actualRhs,rhs.size(),
+                                                  (useRhsDirectly ? rhs.data() : 0));
+                                                  
+    if(!useRhsDirectly)
+      MappedRhs(actualRhs,rhs.size()) = rhs;
+
+    triangular_solve_vector<LhsScalar, RhsScalar, typename Lhs::Index, Side, Mode, LhsProductTraits::NeedToConjugate,
+                            (int(Lhs::Flags) & RowMajorBit) ? RowMajor : ColMajor>
+      ::run(actualLhs.cols(), actualLhs.data(), actualLhs.outerStride(), actualRhs);
+
+    if(!useRhsDirectly)
+      rhs = MappedRhs(actualRhs, rhs.size());
+  }
+};
+
+// the rhs is a matrix
+template<typename Lhs, typename Rhs, int Side, int Mode>
+struct triangular_solver_selector<Lhs,Rhs,Side,Mode,NoUnrolling,Dynamic>
+{
+  typedef typename Rhs::Scalar Scalar;
+  typedef typename Rhs::Index Index;
+  typedef blas_traits<Lhs> LhsProductTraits;
+  typedef typename LhsProductTraits::DirectLinearAccessType ActualLhsType;
+
+  static void run(const Lhs& lhs, Rhs& rhs)
+  {
+    typename internal::add_const_on_value_type<ActualLhsType>::type actualLhs = LhsProductTraits::extract(lhs);
+
+    const Index size = lhs.rows();
+    const Index othersize = Side==OnTheLeft? rhs.cols() : rhs.rows();
+
+    typedef internal::gemm_blocking_space<(Rhs::Flags&RowMajorBit) ? RowMajor : ColMajor,Scalar,Scalar,
+              Rhs::MaxRowsAtCompileTime, Rhs::MaxColsAtCompileTime, Lhs::MaxRowsAtCompileTime,4> BlockingType;
+
+    BlockingType blocking(rhs.rows(), rhs.cols(), size);
+
+    triangular_solve_matrix<Scalar,Index,Side,Mode,LhsProductTraits::NeedToConjugate,(int(Lhs::Flags) & RowMajorBit) ? RowMajor : ColMajor,
+                               (Rhs::Flags&RowMajorBit) ? RowMajor : ColMajor>
+      ::run(size, othersize, &actualLhs.coeffRef(0,0), actualLhs.outerStride(), &rhs.coeffRef(0,0), rhs.outerStride(), blocking);
+  }
+};
+
+/***************************************************************************
+* meta-unrolling implementation
+***************************************************************************/
+
+template<typename Lhs, typename Rhs, int Mode, int Index, int Size,
+         bool Stop = Index==Size>
+struct triangular_solver_unroller;
+
+template<typename Lhs, typename Rhs, int Mode, int Index, int Size>
+struct triangular_solver_unroller<Lhs,Rhs,Mode,Index,Size,false> {
+  enum {
+    IsLower = ((Mode&Lower)==Lower),
+    I = IsLower ? Index : Size - Index - 1,
+    S = IsLower ? 0     : I+1
+  };
+  static void run(const Lhs& lhs, Rhs& rhs)
+  {
+    if (Index>0)
+      rhs.coeffRef(I) -= lhs.row(I).template segment<Index>(S).transpose()
+                         .cwiseProduct(rhs.template segment<Index>(S)).sum();
+
+    if(!(Mode & UnitDiag))
+      rhs.coeffRef(I) /= lhs.coeff(I,I);
+
+    triangular_solver_unroller<Lhs,Rhs,Mode,Index+1,Size>::run(lhs,rhs);
+  }
+};
+
+template<typename Lhs, typename Rhs, int Mode, int Index, int Size>
+struct triangular_solver_unroller<Lhs,Rhs,Mode,Index,Size,true> {
+  static void run(const Lhs&, Rhs&) {}
+};
+
+template<typename Lhs, typename Rhs, int Mode>
+struct triangular_solver_selector<Lhs,Rhs,OnTheLeft,Mode,CompleteUnrolling,1> {
+  static void run(const Lhs& lhs, Rhs& rhs)
+  { triangular_solver_unroller<Lhs,Rhs,Mode,0,Rhs::SizeAtCompileTime>::run(lhs,rhs); }
+};
+
+template<typename Lhs, typename Rhs, int Mode>
+struct triangular_solver_selector<Lhs,Rhs,OnTheRight,Mode,CompleteUnrolling,1> {
+  static void run(const Lhs& lhs, Rhs& rhs)
+  {
+    Transpose<const Lhs> trLhs(lhs);
+    Transpose<Rhs> trRhs(rhs);
+    
+    triangular_solver_unroller<Transpose<const Lhs>,Transpose<Rhs>,
+                              ((Mode&Upper)==Upper ? Lower : Upper) | (Mode&UnitDiag),
+                              0,Rhs::SizeAtCompileTime>::run(trLhs,trRhs);
+  }
+};
+
+} // end namespace internal
+
+/***************************************************************************
+* TriangularView methods
+***************************************************************************/
+
+/** "in-place" version of TriangularView::solve() where the result is written in \a other
+  *
+  * \warning The parameter is only marked 'const' to make the C++ compiler accept a temporary expression here.
+  * This function will const_cast it, so constness isn't honored here.
+  *
+  * See TriangularView:solve() for the details.
+  */
+template<typename MatrixType, unsigned int Mode>
+template<int Side, typename OtherDerived>
+void TriangularView<MatrixType,Mode>::solveInPlace(const MatrixBase<OtherDerived>& _other) const
+{
+  OtherDerived& other = _other.const_cast_derived();
+  eigen_assert( cols() == rows() && ((Side==OnTheLeft && cols() == other.rows()) || (Side==OnTheRight && cols() == other.cols())) );
+  eigen_assert((!(Mode & ZeroDiag)) && bool(Mode & (Upper|Lower)));
+
+  enum { copy = internal::traits<OtherDerived>::Flags & RowMajorBit  && OtherDerived::IsVectorAtCompileTime };
+  typedef typename internal::conditional<copy,
+    typename internal::plain_matrix_type_column_major<OtherDerived>::type, OtherDerived&>::type OtherCopy;
+  OtherCopy otherCopy(other);
+
+  internal::triangular_solver_selector<MatrixType, typename internal::remove_reference<OtherCopy>::type,
+    Side, Mode>::run(nestedExpression(), otherCopy);
+
+  if (copy)
+    other = otherCopy;
+}
+
+/** \returns the product of the inverse of \c *this with \a other, \a *this being triangular.
+  *
+  * This function computes the inverse-matrix matrix product inverse(\c *this) * \a other if
+  * \a Side==OnTheLeft (the default), or the right-inverse-multiply  \a other * inverse(\c *this) if
+  * \a Side==OnTheRight.
+  *
+  * The matrix \c *this must be triangular and invertible (i.e., all the coefficients of the
+  * diagonal must be non zero). It works as a forward (resp. backward) substitution if \c *this
+  * is an upper (resp. lower) triangular matrix.
+  *
+  * Example: \include MatrixBase_marked.cpp
+  * Output: \verbinclude MatrixBase_marked.out
+  *
+  * This function returns an expression of the inverse-multiply and can works in-place if it is assigned
+  * to the same matrix or vector \a other.
+  *
+  * For users coming from BLAS, this function (and more specifically solveInPlace()) offer
+  * all the operations supported by the \c *TRSV and \c *TRSM BLAS routines.
+  *
+  * \sa TriangularView::solveInPlace()
+  */
+template<typename Derived, unsigned int Mode>
+template<int Side, typename Other>
+const internal::triangular_solve_retval<Side,TriangularView<Derived,Mode>,Other>
+TriangularView<Derived,Mode>::solve(const MatrixBase<Other>& other) const
+{
+  return internal::triangular_solve_retval<Side,TriangularView,Other>(*this, other.derived());
+}
+
+namespace internal {
+
+
+template<int Side, typename TriangularType, typename Rhs>
+struct traits<triangular_solve_retval<Side, TriangularType, Rhs> >
+{
+  typedef typename internal::plain_matrix_type_column_major<Rhs>::type ReturnType;
+};
+
+template<int Side, typename TriangularType, typename Rhs> struct triangular_solve_retval
+ : public ReturnByValue<triangular_solve_retval<Side, TriangularType, Rhs> >
+{
+  typedef typename remove_all<typename Rhs::Nested>::type RhsNestedCleaned;
+  typedef ReturnByValue<triangular_solve_retval> Base;
+  typedef typename Base::Index Index;
+
+  triangular_solve_retval(const TriangularType& tri, const Rhs& rhs)
+    : m_triangularMatrix(tri), m_rhs(rhs)
+  {}
+
+  inline Index rows() const { return m_rhs.rows(); }
+  inline Index cols() const { return m_rhs.cols(); }
+
+  template<typename Dest> inline void evalTo(Dest& dst) const
+  {
+    if(!(is_same<RhsNestedCleaned,Dest>::value && extract_data(dst) == extract_data(m_rhs)))
+      dst = m_rhs;
+    m_triangularMatrix.template solveInPlace<Side>(dst);
+  }
+
+  protected:
+    const TriangularType& m_triangularMatrix;
+    typename Rhs::Nested m_rhs;
+};
+
+} // namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_SOLVETRIANGULAR_H

data/ext/eigen/eigen3/Eigen/src/Core/StableNorm.h

@@ -0,0 +1,203 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_STABLENORM_H
+#define EIGEN_STABLENORM_H
+
+namespace Eigen { 
+
+namespace internal {
+
+template<typename ExpressionType, typename Scalar>
+inline void stable_norm_kernel(const ExpressionType& bl, Scalar& ssq, Scalar& scale, Scalar& invScale)
+{
+  using std::max;
+  Scalar maxCoeff = bl.cwiseAbs().maxCoeff();
+  
+  if (maxCoeff>scale)
+  {
+    ssq = ssq * numext::abs2(scale/maxCoeff);
+    Scalar tmp = Scalar(1)/maxCoeff;
+    if(tmp > NumTraits<Scalar>::highest())
+    {
+      invScale = NumTraits<Scalar>::highest();
+      scale = Scalar(1)/invScale;
+    }
+    else
+    {
+      scale = maxCoeff;
+      invScale = tmp;
+    }
+  }
+  
+  // TODO if the maxCoeff is much much smaller than the current scale,
+  // then we can neglect this sub vector
+  if(scale>Scalar(0)) // if scale==0, then bl is 0 
+    ssq += (bl*invScale).squaredNorm();
+}
+
+template<typename Derived>
+inline typename NumTraits<typename traits<Derived>::Scalar>::Real
+blueNorm_impl(const EigenBase<Derived>& _vec)
+{
+  typedef typename Derived::RealScalar RealScalar;  
+  typedef typename Derived::Index Index;
+  using std::pow;
+  using std::min;
+  using std::max;
+  using std::sqrt;
+  using std::abs;
+  const Derived& vec(_vec.derived());
+  static bool initialized = false;
+  static RealScalar b1, b2, s1m, s2m, overfl, rbig, relerr;
+  if(!initialized)
+  {
+    int ibeta, it, iemin, iemax, iexp;
+    RealScalar eps;
+    // This program calculates the machine-dependent constants
+    // bl, b2, slm, s2m, relerr overfl
+    // 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
+    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
+
+    iexp  = -((1-iemin)/2);
+    b1    = RealScalar(pow(RealScalar(ibeta),RealScalar(iexp)));    // lower boundary of midrange
+    iexp  = (iemax + 1 - it)/2;
+    b2    = RealScalar(pow(RealScalar(ibeta),RealScalar(iexp)));    // upper boundary of midrange
+
+    iexp  = (2-iemin)/2;
+    s1m   = RealScalar(pow(RealScalar(ibeta),RealScalar(iexp)));    // scaling factor for lower range
+    iexp  = - ((iemax+it)/2);
+    s2m   = RealScalar(pow(RealScalar(ibeta),RealScalar(iexp)));    // scaling factor for upper range
+
+    overfl  = rbig*s2m;                                             // overflow boundary for abig
+    eps     = RealScalar(pow(double(ibeta), 1-it));
+    relerr  = sqrt(eps);                                            // tolerance for neglecting asml
+    initialized = true;
+  }
+  Index n = vec.size();
+  RealScalar ab2 = b2 / RealScalar(n);
+  RealScalar asml = RealScalar(0);
+  RealScalar amed = RealScalar(0);
+  RealScalar abig = RealScalar(0);
+  for(typename Derived::InnerIterator it(vec, 0); it; ++it)
+  {
+    RealScalar ax = abs(it.value());
+    if(ax > ab2)     abig += numext::abs2(ax*s2m);
+    else if(ax < b1) asml += numext::abs2(ax*s1m);
+    else             amed += numext::abs2(ax);
+  }
+  if(abig > RealScalar(0))
+  {
+    abig = sqrt(abig);
+    if(abig > overfl)
+    {
+      return rbig;
+    }
+    if(amed > RealScalar(0))
+    {
+      abig = abig/s2m;
+      amed = sqrt(amed);
+    }
+    else
+      return abig/s2m;
+  }
+  else if(asml > RealScalar(0))
+  {
+    if (amed > RealScalar(0))
+    {
+      abig = sqrt(amed);
+      amed = sqrt(asml) / s1m;
+    }
+    else
+      return sqrt(asml)/s1m;
+  }
+  else
+    return sqrt(amed);
+  asml = (min)(abig, amed);
+  abig = (max)(abig, amed);
+  if(asml <= abig*relerr)
+    return abig;
+  else
+    return abig * sqrt(RealScalar(1) + numext::abs2(asml/abig));
+}
+
+} // end namespace internal
+
+/** \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 internal::traits<Derived>::Scalar>::Real
+MatrixBase<Derived>::stableNorm() const
+{
+  using std::min;
+  using std::sqrt;
+  const Index blockSize = 4096;
+  RealScalar scale(0);
+  RealScalar invScale(1);
+  RealScalar ssq(0); // sum of square
+  enum {
+    Alignment = (int(Flags)&DirectAccessBit) || (int(Flags)&AlignedBit) ? 1 : 0
+  };
+  Index n = size();
+  Index bi = internal::first_aligned(derived());
+  if (bi>0)
+    internal::stable_norm_kernel(this->head(bi), ssq, scale, invScale);
+  for (; bi<n; bi+=blockSize)
+    internal::stable_norm_kernel(this->segment(bi,(min)(blockSize, n - bi)).template forceAlignedAccessIf<Alignment>(), ssq, scale, invScale);
+  return scale * 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 internal::traits<Derived>::Scalar>::Real
+MatrixBase<Derived>::blueNorm() const
+{
+  return internal::blueNorm_impl(*this);
+}
+
+/** \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 internal::traits<Derived>::Scalar>::Real
+MatrixBase<Derived>::hypotNorm() const
+{
+  return this->cwiseAbs().redux(internal::scalar_hypot_op<RealScalar>());
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_STABLENORM_H